Merge branch 'master' of ssh://git.freedesktop.org/git/mesa/mesa into pipe-video

Conflicts:
	src/gallium/include/pipe/p_format.h

86 files changed, 4105 insertions, 2443 deletions

diff --git a/src/gallium/drivers/galahad/glhd_context.c b/src/gallium/drivers/galahad/glhd_context.c
index ff6d2aa00ab..50f66079c2a 100644
--- a/src/gallium/drivers/galahad/glhd_context.c
+++ b/src/gallium/drivers/galahad/glhd_context.c
@@ -641,7 +641,7 @@ galahad_set_index_buffer(struct pipe_context *_pipe,
          break;
       default:
          glhd_warn("index buffer %p has unrecognized index size %d",
-               _ib->buffer, _ib->index_size);
+                   (void *) _ib->buffer, _ib->index_size);
          break;
       }
    }
@@ -1013,7 +1013,7 @@ galahad_context_create(struct pipe_screen *_screen, struct pipe_context *pipe)
 
    glhd_pipe->pipe = pipe;
 
-   glhd_warn("Created context %p", glhd_pipe);
+   glhd_warn("Created context %p", (void *) glhd_pipe);
 
    return &glhd_pipe->base;
 }
diff --git a/src/gallium/drivers/galahad/glhd_screen.c b/src/gallium/drivers/galahad/glhd_screen.c
index 288941b1066..b6cc41d908b 100644
--- a/src/gallium/drivers/galahad/glhd_screen.c
+++ b/src/gallium/drivers/galahad/glhd_screen.c
@@ -370,7 +370,7 @@ galahad_screen_create(struct pipe_screen *screen)
 
    glhd_screen->screen = screen;
 
-   glhd_warn("Created screen %p", glhd_screen);
+   glhd_warn("Created screen %p", (void *) glhd_screen);
 
    return &glhd_screen->base;
 }
diff --git a/src/gallium/drivers/llvmpipe/Makefile b/src/gallium/drivers/llvmpipe/Makefile
index 55b877b4ab9..669e42e3003 100644
--- a/src/gallium/drivers/llvmpipe/Makefile
+++ b/src/gallium/drivers/llvmpipe/Makefile
@@ -28,8 +28,6 @@ C_SOURCES = \
 	lp_scene_queue.c \
 	lp_screen.c \
 	lp_setup.c \
-	lp_setup_coef.c \
-	lp_setup_coef_intrin.c \
 	lp_setup_line.c \
 	lp_setup_point.c \
 	lp_setup_tri.c \
@@ -38,6 +36,7 @@ C_SOURCES = \
 	lp_state_clip.c \
 	lp_state_derived.c \
 	lp_state_fs.c \
+	lp_state_setup.c \
 	lp_state_gs.c \
 	lp_state_rasterizer.c \
 	lp_state_sampler.c \
@@ -63,14 +62,14 @@ PROGS := lp_test_format	\
 # Need this for the lp_test_*.o files
 CLEAN_EXTRA = *.o
 
+include ../../Makefile.template
+
 lp_test_sincos.o : sse_mathfun.h
 
 PROGS_DEPS := ../../auxiliary/libgallium.a
 
-include ../../Makefile.template
-
 lp_tile_soa.c: lp_tile_soa.py ../../auxiliary/util/u_format_parse.py ../../auxiliary/util/u_format_pack.py ../../auxiliary/util/u_format.csv
-	python lp_tile_soa.py ../../auxiliary/util/u_format.csv > $@
+	$(PYTHON2) $(PYTHON_FLAGS) lp_tile_soa.py ../../auxiliary/util/u_format.csv > $@
 
 LDFLAGS += $(LLVM_LDFLAGS)
 LIBS += -L../../auxiliary/ -lgallium libllvmpipe.a $(LLVM_LIBS) $(GL_LIB_DEPS)
diff --git a/src/gallium/drivers/llvmpipe/SConscript b/src/gallium/drivers/llvmpipe/SConscript
index 650435f0f19..49950153a4f 100644
--- a/src/gallium/drivers/llvmpipe/SConscript
+++ b/src/gallium/drivers/llvmpipe/SConscript
@@ -27,13 +27,7 @@ env.Depends('lp_tile_soa.c', [
 ])
 
 
-# Only enable SSSE3 for lp_tile_soa_sse3.c
-ssse3_env = env.Clone()
-if env['gcc'] \
-   and distutils.version.LooseVersion(env['CCVERSION']) >= distutils.version.LooseVersion('4.3') \
-   and env['machine'] in ('x86', 'x86_64') :
-    ssse3_env.Append(CCFLAGS = ['-mssse3'])
-lp_tile_soa_os = ssse3_env.SharedObject('lp_tile_soa.c')
+lp_tile_soa_os = env.SharedObject('lp_tile_soa.c')
 
 
 llvmpipe = env.ConvenienceLibrary(
@@ -64,13 +58,12 @@ llvmpipe = env.ConvenienceLibrary(
 		'lp_setup_line.c',
 		'lp_setup_point.c',
 		'lp_setup_tri.c',
-		'lp_setup_coef.c',
-		'lp_setup_coef_intrin.c',
 		'lp_setup_vbuf.c',
 		'lp_state_blend.c',
 		'lp_state_clip.c',
 		'lp_state_derived.c',
 		'lp_state_fs.c',
+		'lp_state_setup.c',
 		'lp_state_gs.c',
 		'lp_state_rasterizer.c',
 		'lp_state_sampler.c',
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_alpha.c b/src/gallium/drivers/llvmpipe/lp_bld_alpha.c
index e28efe778f9..e50643790c8 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_alpha.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_alpha.c
@@ -48,7 +48,8 @@ lp_build_alpha_test(LLVMBuilderRef builder,
                     struct lp_type type,
                     struct lp_build_mask_context *mask,
                     LLVMValueRef alpha,
-                    LLVMValueRef ref)
+                    LLVMValueRef ref,
+                    boolean do_branch)
 {
    struct lp_build_context bld;
    LLVMValueRef test;
@@ -60,4 +61,7 @@ lp_build_alpha_test(LLVMBuilderRef builder,
    lp_build_name(test, "alpha_mask");
 
    lp_build_mask_update(mask, test);
+
+   if (do_branch)
+      lp_build_mask_check(mask);
 }
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_alpha.h b/src/gallium/drivers/llvmpipe/lp_bld_alpha.h
index 44603b418c0..27ca8aad4d4 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_alpha.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_alpha.h
@@ -48,7 +48,8 @@ lp_build_alpha_test(LLVMBuilderRef builder,
                     struct lp_type type,
                     struct lp_build_mask_context *mask,
                     LLVMValueRef alpha,
-                    LLVMValueRef ref);
+                    LLVMValueRef ref,
+                    boolean do_branch);
 
 
 #endif /* !LP_BLD_ALPHA_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_depth.c b/src/gallium/drivers/llvmpipe/lp_bld_depth.c
index 7eabe0508de..7eb76d4fb31 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_depth.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_depth.c
@@ -1,6 +1,6 @@
 /**************************************************************************
  *
- * Copyright 2009 VMware, Inc.
+ * Copyright 2009-2010 VMware, Inc.
  * All Rights Reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
@@ -53,15 +53,8 @@
  *  ... ... ... ... ... ... ... ... ...
  *
  *
- * Stencil test:
- * Two-sided stencil test is supported but probably not as efficient as
- * it could be.  Currently, we use if/then/else constructs to do the
- * operations for front vs. back-facing polygons.  We could probably do
- * both the front and back arithmetic then use a Select() instruction to
- * choose the result depending on polyon orientation.  We'd have to
- * measure performance both ways and see which is better.
- *
  * @author Jose Fonseca <[email protected]>
+ * @author Brian Paul <[email protected]>
  */
 
 #include "pipe/p_state.h"
@@ -71,6 +64,7 @@
 #include "gallivm/lp_bld_arit.h"
 #include "gallivm/lp_bld_bitarit.h"
 #include "gallivm/lp_bld_const.h"
+#include "gallivm/lp_bld_conv.h"
 #include "gallivm/lp_bld_logic.h"
 #include "gallivm/lp_bld_flow.h"
 #include "gallivm/lp_bld_intr.h"
@@ -310,8 +304,13 @@ lp_depth_type(const struct util_format_description *format_desc,
    }
    else if(format_desc->channel[swizzle].type == UTIL_FORMAT_TYPE_UNSIGNED) {
       assert(format_desc->block.bits <= 32);
-      if(format_desc->channel[swizzle].normalized)
-         type.norm = TRUE;
+      assert(format_desc->channel[swizzle].normalized);
+      if (format_desc->channel[swizzle].size < format_desc->block.bits) {
+         /* Prefer signed integers when possible, as SSE has less support
+          * for unsigned comparison;
+          */
+         type.sign = TRUE;
+      }
    }
    else
       assert(0);
@@ -333,7 +332,7 @@ lp_depth_type(const struct util_format_description *format_desc,
  */
 static boolean
 get_z_shift_and_mask(const struct util_format_description *format_desc,
-                     unsigned *shift, unsigned *mask)
+                     unsigned *shift, unsigned *width, unsigned *mask)
 {
    const unsigned total_bits = format_desc->block.bits;
    unsigned z_swizzle;
@@ -349,12 +348,14 @@ get_z_shift_and_mask(const struct util_format_description *format_desc,
    if (z_swizzle == UTIL_FORMAT_SWIZZLE_NONE)
       return FALSE;
 
+   *width = format_desc->channel[z_swizzle].size;
+
    padding_right = 0;
    for (chan = 0; chan < z_swizzle; ++chan)
       padding_right += format_desc->channel[chan].size;
 
    padding_left =
-      total_bits - (padding_right + format_desc->channel[z_swizzle].size);
+      total_bits - (padding_right + *width);
 
    if (padding_left || padding_right) {
       unsigned long long mask_left = (1ULL << (total_bits - padding_left)) - 1;
@@ -365,7 +366,7 @@ get_z_shift_and_mask(const struct util_format_description *format_desc,
       *mask = 0xffffffff;
    }
 
-   *shift = padding_left;
+   *shift = padding_right;
 
    return TRUE;
 }
@@ -409,7 +410,7 @@ get_s_shift_and_mask(const struct util_format_description *format_desc,
  * \param maskvalue is the depth test mask.
  * \param counter is a pointer of the uint32 counter.
  */
-static void
+void
 lp_build_occlusion_count(LLVMBuilderRef builder,
                          struct lp_type type,
                          LLVMValueRef maskvalue,
@@ -446,33 +447,58 @@ lp_build_occlusion_count(LLVMBuilderRef builder,
  * \param format_desc  description of the depth/stencil surface
  * \param mask  the alive/dead pixel mask for the quad (vector)
  * \param stencil_refs  the front/back stencil ref values (scalar)
- * \param z_src  the incoming depth/stencil values (a 2x2 quad)
+ * \param z_src  the incoming depth/stencil values (a 2x2 quad, float32)
  * \param zs_dst_ptr  pointer to depth/stencil values in framebuffer
- * \param facing  contains float value indicating front/back facing polygon
+ * \param facing  contains boolean value indicating front/back facing polygon
  */
 void
 lp_build_depth_stencil_test(LLVMBuilderRef builder,
                             const struct pipe_depth_state *depth,
                             const struct pipe_stencil_state stencil[2],
-                            struct lp_type type,
+                            struct lp_type z_src_type,
                             const struct util_format_description *format_desc,
                             struct lp_build_mask_context *mask,
                             LLVMValueRef stencil_refs[2],
                             LLVMValueRef z_src,
                             LLVMValueRef zs_dst_ptr,
                             LLVMValueRef face,
-                            LLVMValueRef counter)
+                            LLVMValueRef *zs_value,
+                            boolean do_branch)
 {
-   struct lp_build_context bld;
-   struct lp_build_context sbld;
+   struct lp_type z_type;
+   struct lp_build_context z_bld;
+   struct lp_build_context s_bld;
    struct lp_type s_type;
+   unsigned z_shift = 0, z_width = 0, z_mask = 0;
    LLVMValueRef zs_dst, z_dst = NULL;
    LLVMValueRef stencil_vals = NULL;
    LLVMValueRef z_bitmask = NULL, stencil_shift = NULL;
    LLVMValueRef z_pass = NULL, s_pass_mask = NULL;
-   LLVMValueRef orig_mask = mask->value;
+   LLVMValueRef orig_mask = lp_build_mask_value(mask);
    LLVMValueRef front_facing = NULL;
 
+
+   /*
+    * Depths are expected to be between 0 and 1, even if they are stored in
+    * floats. Setting these bits here will ensure that the lp_build_conv() call
+    * below won't try to unnecessarily clamp the incoming values.
+    */
+   if(z_src_type.floating) {
+      z_src_type.sign = FALSE;
+      z_src_type.norm = TRUE;
+   }
+   else {
+      assert(!z_src_type.sign);
+      assert(z_src_type.norm);
+   }
+
+   /* Pick the depth type. */
+   z_type = lp_depth_type(format_desc, z_src_type.width*z_src_type.length);
+
+   /* FIXME: Cope with a depth test type with a different bit width. */
+   assert(z_type.width == z_src_type.width);
+   assert(z_type.length == z_src_type.length);
+
    /* Sanity checking */
    {
       const unsigned z_swizzle = format_desc->swizzle[0];
@@ -493,8 +519,8 @@ lp_build_depth_stencil_test(LLVMBuilderRef builder,
       }
 
       assert(z_swizzle < 4);
-      assert(format_desc->block.bits == type.width);
-      if (type.floating) {
+      assert(format_desc->block.bits == z_type.width);
+      if (z_type.floating) {
          assert(z_swizzle == 0);
          assert(format_desc->channel[z_swizzle].type ==
                 UTIL_FORMAT_TYPE_FLOAT);
@@ -505,54 +531,56 @@ lp_build_depth_stencil_test(LLVMBuilderRef builder,
          assert(format_desc->channel[z_swizzle].type ==
                 UTIL_FORMAT_TYPE_UNSIGNED);
          assert(format_desc->channel[z_swizzle].normalized);
-         assert(!type.fixed);
-         assert(!type.sign);
-         assert(type.norm);
+         assert(!z_type.fixed);
       }
    }
 
 
    /* Setup build context for Z vals */
-   lp_build_context_init(&bld, builder, type);
+   lp_build_context_init(&z_bld, builder, z_type);
 
    /* Setup build context for stencil vals */
-   s_type = lp_type_int_vec(type.width);
-   lp_build_context_init(&sbld, builder, s_type);
+   s_type = lp_type_int_vec(z_type.width);
+   lp_build_context_init(&s_bld, builder, s_type);
 
    /* Load current z/stencil value from z/stencil buffer */
+   zs_dst_ptr = LLVMBuildBitCast(builder,
+                                 zs_dst_ptr,
+                                 LLVMPointerType(z_bld.vec_type, 0), "");
    zs_dst = LLVMBuildLoad(builder, zs_dst_ptr, "");
 
-   lp_build_name(zs_dst, "zsbufval");
+   lp_build_name(zs_dst, "zs_dst");
 
 
    /* Compute and apply the Z/stencil bitmasks and shifts.
     */
    {
-      unsigned z_shift, z_mask;
       unsigned s_shift, s_mask;
 
-      if (get_z_shift_and_mask(format_desc, &z_shift, &z_mask)) {
-         if (z_shift) {
-            LLVMValueRef shift = lp_build_const_int_vec(type, z_shift);
-            z_src = LLVMBuildLShr(builder, z_src, shift, "");
-         }
-
+      if (get_z_shift_and_mask(format_desc, &z_shift, &z_width, &z_mask)) {
          if (z_mask != 0xffffffff) {
-            LLVMValueRef mask = lp_build_const_int_vec(type, z_mask);
-            z_src = LLVMBuildAnd(builder, z_src, mask, "");
-            z_dst = LLVMBuildAnd(builder, zs_dst, mask, "");
-            z_bitmask = mask;  /* used below */
+            z_bitmask = lp_build_const_int_vec(z_type, z_mask);
          }
-         else {
+
+         /*
+          * Align the framebuffer Z 's LSB to the right.
+          */
+         if (z_shift) {
+            LLVMValueRef shift = lp_build_const_int_vec(z_type, z_shift);
+            z_dst = LLVMBuildLShr(builder, zs_dst, shift, "z_dst");
+         } else if (z_bitmask) {
+	    /* TODO: Instead of loading a mask from memory and ANDing, it's
+	     * probably faster to just shake the bits with two shifts. */
+            z_dst = LLVMBuildAnd(builder, zs_dst, z_bitmask, "z_dst");
+         } else {
             z_dst = zs_dst;
+            lp_build_name(z_dst, "z_dst");
          }
-
-         lp_build_name(z_dst, "zsbuf.z");
       }
 
       if (get_s_shift_and_mask(format_desc, &s_shift, &s_mask)) {
          if (s_shift) {
-            LLVMValueRef shift = lp_build_const_int_vec(type, s_shift);
+            LLVMValueRef shift = lp_build_const_int_vec(s_type, s_shift);
             stencil_vals = LLVMBuildLShr(builder, zs_dst, shift, "");
             stencil_shift = shift;  /* used below */
          }
@@ -561,48 +589,85 @@ lp_build_depth_stencil_test(LLVMBuilderRef builder,
          }
 
          if (s_mask != 0xffffffff) {
-            LLVMValueRef mask = lp_build_const_int_vec(type, s_mask);
+            LLVMValueRef mask = lp_build_const_int_vec(s_type, s_mask);
             stencil_vals = LLVMBuildAnd(builder, stencil_vals, mask, "");
          }
 
-         lp_build_name(stencil_vals, "stencil");
+         lp_build_name(stencil_vals, "s_dst");
       }
    }
 
    if (stencil[0].enabled) {
 
       if (face) {
-         LLVMValueRef zero = LLVMConstReal(LLVMFloatType(), 0.0);
+         LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0);
 
-         /* front_facing = face > 0.0 ? ~0 : 0 */
-         front_facing = LLVMBuildFCmp(builder, LLVMRealUGT, face, zero, "");
+         /* front_facing = face != 0 ? ~0 : 0 */
+         front_facing = LLVMBuildICmp(builder, LLVMIntNE, face, zero, "");
          front_facing = LLVMBuildSExt(builder, front_facing,
-                                      LLVMIntType(bld.type.length*bld.type.width),
+                                      LLVMIntType(s_bld.type.length*s_bld.type.width),
                                       "");
          front_facing = LLVMBuildBitCast(builder, front_facing,
-                                         bld.int_vec_type, "");
+                                         s_bld.int_vec_type, "");
       }
 
       /* convert scalar stencil refs into vectors */
-      stencil_refs[0] = lp_build_broadcast_scalar(&bld, stencil_refs[0]);
-      stencil_refs[1] = lp_build_broadcast_scalar(&bld, stencil_refs[1]);
+      stencil_refs[0] = lp_build_broadcast_scalar(&s_bld, stencil_refs[0]);
+      stencil_refs[1] = lp_build_broadcast_scalar(&s_bld, stencil_refs[1]);
 
-      s_pass_mask = lp_build_stencil_test(&sbld, stencil,
+      s_pass_mask = lp_build_stencil_test(&s_bld, stencil,
                                           stencil_refs, stencil_vals,
                                           front_facing);
 
       /* apply stencil-fail operator */
       {
-         LLVMValueRef s_fail_mask = lp_build_andnot(&bld, orig_mask, s_pass_mask);
-         stencil_vals = lp_build_stencil_op(&sbld, stencil, S_FAIL_OP,
+         LLVMValueRef s_fail_mask = lp_build_andnot(&s_bld, orig_mask, s_pass_mask);
+         stencil_vals = lp_build_stencil_op(&s_bld, stencil, S_FAIL_OP,
                                             stencil_refs, stencil_vals,
                                             s_fail_mask, front_facing);
       }
    }
 
    if (depth->enabled) {
+      /*
+       * Convert fragment Z to the desired type, aligning the LSB to the right.
+       */
+
+      assert(z_type.width == z_src_type.width);
+      assert(z_type.length == z_src_type.length);
+      assert(lp_check_value(z_src_type, z_src));
+      if (z_src_type.floating) {
+         /*
+          * Convert from floating point values
+          */
+
+         if (!z_type.floating) {
+            z_src = lp_build_clamped_float_to_unsigned_norm(builder,
+                                                            z_src_type,
+                                                            z_width,
+                                                            z_src);
+         }
+      } else {
+         /*
+          * Convert from unsigned normalized values.
+          */
+
+         assert(!z_src_type.sign);
+         assert(!z_src_type.fixed);
+         assert(z_src_type.norm);
+         assert(!z_type.floating);
+         if (z_src_type.width > z_width) {
+            LLVMValueRef shift = lp_build_const_int_vec(z_src_type,
+                                                        z_src_type.width - z_width);
+            z_src = LLVMBuildLShr(builder, z_src, shift, "");
+         }
+      }
+      assert(lp_check_value(z_type, z_src));
+
+      lp_build_name(z_src, "z_src");
+
       /* compare src Z to dst Z, returning 'pass' mask */
-      z_pass = lp_build_cmp(&bld, depth->func, z_src, z_dst);
+      z_pass = lp_build_cmp(&z_bld, depth->func, z_src, z_dst);
 
       if (!stencil[0].enabled) {
          /* We can potentially skip all remaining operations here, but only
@@ -610,28 +675,28 @@ lp_build_depth_stencil_test(LLVMBuilderRef builder,
           * buffer values.  Don't need to update Z buffer values.
           */
          lp_build_mask_update(mask, z_pass);
+
+         if (do_branch) {
+            lp_build_mask_check(mask);
+            do_branch = FALSE;
+         }
       }
 
       if (depth->writemask) {
-         LLVMValueRef zselectmask = mask->value;
+         LLVMValueRef zselectmask;
 
          /* mask off bits that failed Z test */
-         zselectmask = LLVMBuildAnd(builder, zselectmask, z_pass, "");
+         zselectmask = LLVMBuildAnd(builder, orig_mask, z_pass, "");
 
          /* mask off bits that failed stencil test */
          if (s_pass_mask) {
             zselectmask = LLVMBuildAnd(builder, zselectmask, s_pass_mask, "");
          }
 
-         /* if combined Z/stencil format, mask off the stencil bits */
-         if (z_bitmask) {
-            zselectmask = LLVMBuildAnd(builder, zselectmask, z_bitmask, "");
-         }
-
          /* Mix the old and new Z buffer values.
-          * z_dst[i] = (zselectmask[i] & z_src[i]) | (~zselectmask[i] & z_dst[i])
+          * z_dst[i] = zselectmask[i] ? z_src[i] : z_dst[i]
           */
-         z_dst = lp_build_select_bitwise(&bld, zselectmask, z_src, z_dst);
+         z_dst = lp_build_select(&z_bld, zselectmask, z_src, z_dst);
       }
 
       if (stencil[0].enabled) {
@@ -639,14 +704,14 @@ lp_build_depth_stencil_test(LLVMBuilderRef builder,
          LLVMValueRef z_fail_mask, z_pass_mask;
 
          /* apply Z-fail operator */
-         z_fail_mask = lp_build_andnot(&bld, orig_mask, z_pass);
-         stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_FAIL_OP,
+         z_fail_mask = lp_build_andnot(&z_bld, orig_mask, z_pass);
+         stencil_vals = lp_build_stencil_op(&s_bld, stencil, Z_FAIL_OP,
                                             stencil_refs, stencil_vals,
                                             z_fail_mask, front_facing);
 
          /* apply Z-pass operator */
-         z_pass_mask = LLVMBuildAnd(bld.builder, orig_mask, z_pass, "");
-         stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_PASS_OP,
+         z_pass_mask = LLVMBuildAnd(z_bld.builder, orig_mask, z_pass, "");
+         stencil_vals = lp_build_stencil_op(&s_bld, stencil, Z_PASS_OP,
                                             stencil_refs, stencil_vals,
                                             z_pass_mask, front_facing);
       }
@@ -655,17 +720,19 @@ lp_build_depth_stencil_test(LLVMBuilderRef builder,
       /* No depth test: apply Z-pass operator to stencil buffer values which
        * passed the stencil test.
        */
-      s_pass_mask = LLVMBuildAnd(bld.builder, orig_mask, s_pass_mask, "");
-      stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_PASS_OP,
+      s_pass_mask = LLVMBuildAnd(s_bld.builder, orig_mask, s_pass_mask, "");
+      stencil_vals = lp_build_stencil_op(&s_bld, stencil, Z_PASS_OP,
                                          stencil_refs, stencil_vals,
                                          s_pass_mask, front_facing);
    }
 
-   /* The Z bits are already in the right place but we may need to shift the
-    * stencil bits before ORing Z with Stencil to make the final pixel value.
-    */
+   /* Put Z and ztencil bits in the right place */
+   if (z_dst && z_shift) {
+      LLVMValueRef shift = lp_build_const_int_vec(z_type, z_shift);
+      z_dst = LLVMBuildShl(builder, z_dst, shift, "");
+   }
    if (stencil_vals && stencil_shift)
-      stencil_vals = LLVMBuildShl(bld.builder, stencil_vals,
+      stencil_vals = LLVMBuildShl(s_bld.builder, stencil_vals,
                                   stencil_shift, "");
 
    /* Finally, merge/store the z/stencil values */
@@ -673,13 +740,13 @@ lp_build_depth_stencil_test(LLVMBuilderRef builder,
        (stencil[0].enabled && stencil[0].writemask)) {
 
       if (z_dst && stencil_vals)
-         zs_dst = LLVMBuildOr(bld.builder, z_dst, stencil_vals, "");
+         zs_dst = LLVMBuildOr(z_bld.builder, z_dst, stencil_vals, "");
       else if (z_dst)
          zs_dst = z_dst;
       else
          zs_dst = stencil_vals;
 
-      LLVMBuildStore(builder, zs_dst, zs_dst_ptr);
+      *zs_value = zs_dst;
    }
 
    if (s_pass_mask)
@@ -688,6 +755,47 @@ lp_build_depth_stencil_test(LLVMBuilderRef builder,
    if (depth->enabled && stencil[0].enabled)
       lp_build_mask_update(mask, z_pass);
 
-   if (counter)
-      lp_build_occlusion_count(builder, type, mask->value, counter);
+   if (do_branch)
+      lp_build_mask_check(mask);
+
+}
+
+
+void
+lp_build_depth_write(LLVMBuilderRef builder,
+                     const struct util_format_description *format_desc,
+                     LLVMValueRef zs_dst_ptr,
+                     LLVMValueRef zs_value)
+{
+   zs_dst_ptr = LLVMBuildBitCast(builder, zs_dst_ptr,
+                                 LLVMPointerType(LLVMTypeOf(zs_value), 0), "");
+
+   LLVMBuildStore(builder, zs_value, zs_dst_ptr);
+}
+
+
+void
+lp_build_deferred_depth_write(LLVMBuilderRef builder,
+                              struct lp_type z_src_type,
+                              const struct util_format_description *format_desc,
+                              struct lp_build_mask_context *mask,
+                              LLVMValueRef zs_dst_ptr,
+                              LLVMValueRef zs_value)
+{
+   struct lp_type z_type;
+   struct lp_build_context z_bld;
+   LLVMValueRef z_dst;
+
+   /* XXX: pointlessly redo type logic:
+    */
+   z_type = lp_depth_type(format_desc, z_src_type.width*z_src_type.length);
+   lp_build_context_init(&z_bld, builder, z_type);
+
+   zs_dst_ptr = LLVMBuildBitCast(builder, zs_dst_ptr,
+                                 LLVMPointerType(z_bld.vec_type, 0), "");
+
+   z_dst = LLVMBuildLoad(builder, zs_dst_ptr, "zsbufval");
+   z_dst = lp_build_select(&z_bld, lp_build_mask_value(mask), zs_value, z_dst);
+
+   LLVMBuildStore(builder, z_dst, zs_dst_ptr);
 }
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_depth.h b/src/gallium/drivers/llvmpipe/lp_bld_depth.h
index e257a5bd7d0..a54ef3a711e 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_depth.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_depth.h
@@ -61,7 +61,27 @@ lp_build_depth_stencil_test(LLVMBuilderRef builder,
                             LLVMValueRef zs_src,
                             LLVMValueRef zs_dst_ptr,
                             LLVMValueRef facing,
-                            LLVMValueRef counter);
+                            LLVMValueRef *zs_value,
+                            boolean do_branch);
 
+void
+lp_build_depth_write(LLVMBuilderRef builder,
+                     const struct util_format_description *format_desc,
+                     LLVMValueRef zs_dst_ptr,
+                     LLVMValueRef zs_value);
+
+void
+lp_build_deferred_depth_write(LLVMBuilderRef builder,
+                              struct lp_type z_src_type,
+                              const struct util_format_description *format_desc,
+                              struct lp_build_mask_context *mask,
+                              LLVMValueRef zs_dst_ptr,
+                              LLVMValueRef zs_value);
+
+void
+lp_build_occlusion_count(LLVMBuilderRef builder,
+                         struct lp_type type,
+                         LLVMValueRef maskvalue,
+                         LLVMValueRef counter);
 
 #endif /* !LP_BLD_DEPTH_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_interp.c b/src/gallium/drivers/llvmpipe/lp_bld_interp.c
index 2a374f8c390..c9da8900d0c 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_interp.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_interp.c
@@ -206,7 +206,7 @@ coeffs_init(struct lp_build_interp_soa_context *bld,
             dadq2 = LLVMBuildFAdd(builder, dadq, dadq, "");
 
             /*
-             * a = a0 + x * dadx + y * dady
+             * a = a0 + (x * dadx + y * dady)
              */
 
             if (attrib == 0 && chan == 0) {
@@ -219,11 +219,11 @@ coeffs_init(struct lp_build_interp_soa_context *bld,
                a = a0;
                if (interp != LP_INTERP_CONSTANT &&
                    interp != LP_INTERP_FACING) {
-                  LLVMValueRef tmp;
-                  tmp = LLVMBuildFMul(builder, bld->x, dadx, "");
-                  a = LLVMBuildFAdd(builder, a, tmp, "");
-                  tmp = LLVMBuildFMul(builder, bld->y, dady, "");
-                  a = LLVMBuildFAdd(builder, a, tmp, "");
+                  LLVMValueRef ax, ay, axy;
+                  ax = LLVMBuildFMul(builder, bld->x, dadx, "");
+                  ay = LLVMBuildFMul(builder, bld->y, dady, "");
+                  axy = LLVMBuildFAdd(builder, ax, ay, "");
+                  a = LLVMBuildFAdd(builder, a, axy, "");
                }
             }
 
@@ -272,7 +272,10 @@ coeffs_init(struct lp_build_interp_soa_context *bld,
  * This is called when we move from one quad to the next.
  */
 static void
-attribs_update(struct lp_build_interp_soa_context *bld, int quad_index)
+attribs_update(struct lp_build_interp_soa_context *bld,
+               int quad_index,
+               int start,
+               int end)
 {
    struct lp_build_context *coeff_bld = &bld->coeff_bld;
    LLVMValueRef shuffle = lp_build_const_int_vec(coeff_bld->type, quad_index);
@@ -282,7 +285,7 @@ attribs_update(struct lp_build_interp_soa_context *bld, int quad_index)
 
    assert(quad_index < 4);
 
-   for(attrib = 0; attrib < bld->num_attribs; ++attrib) {
+   for(attrib = start; attrib < end; ++attrib) {
       const unsigned mask = bld->mask[attrib];
       const unsigned interp = bld->interp[attrib];
       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
@@ -350,6 +353,14 @@ attribs_update(struct lp_build_interp_soa_context *bld, int quad_index)
                }
 #endif
 
+               if (attrib == 0 && chan == 2) {
+                  /* FIXME: Depth values can exceed 1.0, due to the fact that
+                   * setup interpolation coefficients refer to (0,0) which causes
+                   * precision loss. So we must clamp to 1.0 here to avoid artifacts
+                   */
+                  a = lp_build_min(coeff_bld, a, coeff_bld->one);
+               }
+
                attrib_name(a, attrib, chan, "");
             }
             bld->attribs[attrib][chan] = a;
@@ -434,8 +445,6 @@ lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
    pos_init(bld, x0, y0);
 
    coeffs_init(bld, a0_ptr, dadx_ptr, dady_ptr);
-
-   attribs_update(bld, 0);
 }
 
 
@@ -443,10 +452,20 @@ lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
  * Advance the position and inputs to the given quad within the block.
  */
 void
-lp_build_interp_soa_update(struct lp_build_interp_soa_context *bld,
-                           int quad_index)
+lp_build_interp_soa_update_inputs(struct lp_build_interp_soa_context *bld,
+                                  int quad_index)
+{
+   assert(quad_index < 4);
+
+   attribs_update(bld, quad_index, 1, bld->num_attribs);
+}
+
+void
+lp_build_interp_soa_update_pos(struct lp_build_interp_soa_context *bld,
+                                  int quad_index)
 {
    assert(quad_index < 4);
 
-   attribs_update(bld, quad_index);
+   attribs_update(bld, quad_index, 0, 1);
 }
+
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_interp.h b/src/gallium/drivers/llvmpipe/lp_bld_interp.h
index 3054030f739..a7ebdd1bfa2 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_interp.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_interp.h
@@ -46,7 +46,31 @@
 
 #include "tgsi/tgsi_exec.h"
 
-#include "lp_setup.h"
+/**
+ * Describes how to compute the interpolation coefficients (a0, dadx, dady)
+ * from the vertices passed into our triangle/line/point functions by the
+ * draw module.
+ *
+ * Vertices are treated as an array of float[4] values, indexed by
+ * src_index.
+ *
+ * LP_INTERP_COLOR is translated to either LP_INTERP_CONSTANT or
+ * LINEAR depending on flatshade state.
+ */
+enum lp_interp {
+   LP_INTERP_CONSTANT,
+   LP_INTERP_COLOR,
+   LP_INTERP_LINEAR,
+   LP_INTERP_PERSPECTIVE,
+   LP_INTERP_POSITION,
+   LP_INTERP_FACING
+};
+
+struct lp_shader_input {
+   ushort interp:4;       /* enum lp_interp */
+   ushort usage_mask:4;   /* bitmask of TGSI_WRITEMASK_x flags */
+   ushort src_index:8;    /* where to find values in incoming vertices */
+};
 
 
 struct lp_build_interp_soa_context
@@ -89,7 +113,11 @@ lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
                          LLVMValueRef y);
 
 void
-lp_build_interp_soa_update(struct lp_build_interp_soa_context *bld,
+lp_build_interp_soa_update_inputs(struct lp_build_interp_soa_context *bld,
+                           int quad_index);
+
+void
+lp_build_interp_soa_update_pos(struct lp_build_interp_soa_context *bld,
                            int quad_index);
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_context.c b/src/gallium/drivers/llvmpipe/lp_context.c
index 39f2c6085ef..763432ed712 100644
--- a/src/gallium/drivers/llvmpipe/lp_context.c
+++ b/src/gallium/drivers/llvmpipe/lp_context.c
@@ -82,6 +82,8 @@ static void llvmpipe_destroy( struct pipe_context *pipe )
       }
    }
 
+   lp_delete_setup_variants(llvmpipe);
+
    align_free( llvmpipe );
 }
 
@@ -108,6 +110,7 @@ llvmpipe_create_context( struct pipe_screen *screen, void *priv )
    memset(llvmpipe, 0, sizeof *llvmpipe);
 
    make_empty_list(&llvmpipe->fs_variants_list);
+   make_empty_list(&llvmpipe->setup_variants_list);
 
    llvmpipe->pipe.winsys = screen->winsys;
    llvmpipe->pipe.screen = screen;
diff --git a/src/gallium/drivers/llvmpipe/lp_context.h b/src/gallium/drivers/llvmpipe/lp_context.h
index 34fa20e204a..db09c95b272 100644
--- a/src/gallium/drivers/llvmpipe/lp_context.h
+++ b/src/gallium/drivers/llvmpipe/lp_context.h
@@ -39,6 +39,7 @@
 #include "lp_jit.h"
 #include "lp_setup.h"
 #include "lp_state_fs.h"
+#include "lp_state_setup.h"
 
 
 struct llvmpipe_vbuf_render;
@@ -48,6 +49,7 @@ struct lp_fragment_shader;
 struct lp_vertex_shader;
 struct lp_blend_state;
 struct lp_setup_context;
+struct lp_setup_variant;
 struct lp_velems_state;
 
 struct llvmpipe_context {
@@ -105,12 +107,9 @@ struct llvmpipe_context {
    /** Which vertex shader output slot contains point size */
    int psize_slot;
 
-   /** Fragment shader input interpolation info */
-   unsigned num_inputs;
-   struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
-
    /** The tiling engine */
    struct lp_setup_context *setup;
+   struct lp_setup_variant setup_variant;
 
    /** The primitive drawing context */
    struct draw_context *draw;
@@ -120,6 +119,9 @@ struct llvmpipe_context {
 
    struct lp_fs_variant_list_item fs_variants_list;
    unsigned nr_fs_variants;
+
+   struct lp_setup_variant_list_item setup_variants_list;
+   unsigned nr_setup_variants;
 };
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_flush.h b/src/gallium/drivers/llvmpipe/lp_flush.h
index bb538b2bd83..3626ce4a86c 100644
--- a/src/gallium/drivers/llvmpipe/lp_flush.h
+++ b/src/gallium/drivers/llvmpipe/lp_flush.h
@@ -32,6 +32,7 @@
 
 struct pipe_context;
 struct pipe_fence_handle;
+struct pipe_resource;
 
 void
 llvmpipe_flush(struct pipe_context *pipe,
diff --git a/src/gallium/drivers/llvmpipe/lp_jit.c b/src/gallium/drivers/llvmpipe/lp_jit.c
index 04b12dedccf..c540f9b3628 100644
--- a/src/gallium/drivers/llvmpipe/lp_jit.c
+++ b/src/gallium/drivers/llvmpipe/lp_jit.c
@@ -36,7 +36,6 @@
 #include <llvm-c/Transforms/Scalar.h>
 
 #include "util/u_memory.h"
-#include "util/u_cpu_detect.h"
 #include "gallivm/lp_bld_init.h"
 #include "gallivm/lp_bld_debug.h"
 #include "lp_screen.h"
@@ -162,9 +161,6 @@ lp_jit_init_globals(struct llvmpipe_screen *screen)
 void
 lp_jit_screen_cleanup(struct llvmpipe_screen *screen)
 {
-   if(screen->engine)
-      LLVMDisposeExecutionEngine(screen->engine);
-
    if(screen->pass)
       LLVMDisposePassManager(screen->pass);
 }
@@ -190,13 +186,7 @@ lp_jit_screen_init(struct llvmpipe_screen *screen)
       LLVMAddCFGSimplificationPass(screen->pass);
       LLVMAddPromoteMemoryToRegisterPass(screen->pass);
       LLVMAddConstantPropagationPass(screen->pass);
-      if(util_cpu_caps.has_sse4_1) {
-         /* FIXME: There is a bug in this pass, whereby the combination of fptosi
-          * and sitofp (necessary for trunc/floor/ceil/round implementation)
-          * somehow becomes invalid code.
-          */
-         LLVMAddInstructionCombiningPass(screen->pass);
-      }
+      LLVMAddInstructionCombiningPass(screen->pass);
       LLVMAddGVNPass(screen->pass);
    } else {
       /* We need at least this pass to prevent the backends to fail in
diff --git a/src/gallium/drivers/llvmpipe/lp_jit.h b/src/gallium/drivers/llvmpipe/lp_jit.h
index 16e04fce0cd..114f21f2d16 100644
--- a/src/gallium/drivers/llvmpipe/lp_jit.h
+++ b/src/gallium/drivers/llvmpipe/lp_jit.h
@@ -144,7 +144,7 @@ typedef void
 (*lp_jit_frag_func)(const struct lp_jit_context *context,
                     uint32_t x,
                     uint32_t y,
-                    float facing,
+                    uint32_t facing,
                     const void *a0,
                     const void *dadx,
                     const void *dady,
diff --git a/src/gallium/drivers/llvmpipe/lp_limits.h b/src/gallium/drivers/llvmpipe/lp_limits.h
index d1c431475d8..2538164ffaa 100644
--- a/src/gallium/drivers/llvmpipe/lp_limits.h
+++ b/src/gallium/drivers/llvmpipe/lp_limits.h
@@ -72,4 +72,14 @@
  */
 #define LP_MAX_SHADER_VARIANTS 1024
 
+/**
+ * Max number of setup variants that will be kept around.
+ *
+ * These are determined by the combination of the fragment shader
+ * input signature and a small amount of rasterization state (eg
+ * flatshading).  It is likely that many active fragment shaders will
+ * share the same setup variant.
+ */
+#define LP_MAX_SETUP_VARIANTS 64
+
 #endif /* LP_LIMITS_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_rast.c b/src/gallium/drivers/llvmpipe/lp_rast.c
index db9b2f9b128..d358a983943 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast.c
@@ -334,7 +334,7 @@ lp_rast_shade_tile(struct lp_rasterizer_task *task,
 {
    const struct lp_scene *scene = task->scene;
    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
-   const struct lp_rast_state *state = inputs->state;
+   const struct lp_rast_state *state = task->state;
    struct lp_fragment_shader_variant *variant = state->variant;
    const unsigned tile_x = task->x, tile_y = task->y;
    unsigned x, y;
@@ -365,10 +365,10 @@ lp_rast_shade_tile(struct lp_rasterizer_task *task,
          BEGIN_JIT_CALL(state);
          variant->jit_function[RAST_WHOLE]( &state->jit_context,
                                             tile_x + x, tile_y + y,
-                                            inputs->facing,
-                                            inputs->a0,
-                                            inputs->dadx,
-                                            inputs->dady,
+                                            inputs->frontfacing,
+                                            GET_A0(inputs),
+                                            GET_DADX(inputs),
+                                            GET_DADY(inputs),
                                             color,
                                             depth,
                                             0xffff,
@@ -414,7 +414,7 @@ lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
                          unsigned x, unsigned y,
                          unsigned mask)
 {
-   const struct lp_rast_state *state = inputs->state;
+   const struct lp_rast_state *state = task->state;
    struct lp_fragment_shader_variant *variant = state->variant;
    const struct lp_scene *scene = task->scene;
    uint8_t *color[PIPE_MAX_COLOR_BUFS];
@@ -446,10 +446,10 @@ lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
    BEGIN_JIT_CALL(state);
    variant->jit_function[RAST_EDGE_TEST](&state->jit_context,
                                          x, y,
-                                         inputs->facing,
-                                         inputs->a0,
-                                         inputs->dadx,
-                                         inputs->dady,
+                                         inputs->frontfacing,
+                                         GET_A0(inputs),
+                                         GET_DADX(inputs),
+                                         GET_DADY(inputs),
                                          color,
                                          depth,
                                          mask,
@@ -490,6 +490,14 @@ lp_rast_end_query(struct lp_rasterizer_task *task,
 }
 
 
+void
+lp_rast_set_state(struct lp_rasterizer_task *task,
+                  const union lp_rast_cmd_arg arg)
+{
+   task->state = arg.state;
+}
+
+
 
 /**
  * Set top row and left column of the tile's pixels to white.  For debugging.
@@ -602,6 +610,7 @@ static lp_rast_cmd_func dispatch[LP_RAST_OP_MAX] =
    lp_rast_shade_tile_opaque,
    lp_rast_begin_query,
    lp_rast_end_query,
+   lp_rast_set_state,
 };
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_rast.h b/src/gallium/drivers/llvmpipe/lp_rast.h
index df0bea04b9a..a64c152cf83 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast.h
@@ -78,30 +78,28 @@ struct lp_rast_state {
  * These pointers point into the bin data buffer.
  */
 struct lp_rast_shader_inputs {
-   float facing;     /** Positive for front-facing, negative for back-facing */
-   unsigned disable:1;  /** Partially binned, disable this command */
-   unsigned opaque:1;   /** Is opaque */
-
-   float (*a0)[4];
-   float (*dadx)[4];
-   float (*dady)[4];
-
-   const struct lp_rast_state *state;
+   unsigned frontfacing:1;      /** True for front-facing */
+   unsigned disable:1;          /** Partially binned, disable this command */
+   unsigned opaque:1;           /** Is opaque */
+   unsigned pad0:29;            /* wasted space */
+   unsigned stride;             /* how much to advance data between a0, dadx, dady */
+   unsigned pad2;               /* wasted space */
+   unsigned pad3;               /* wasted space */
+   /* followed by a0, dadx, dady and planes[] */
 };
 
-
+/* Note: the order of these values is important as they are loaded by
+ * sse code in rasterization:
+ */
 struct lp_rast_plane {
-   /* one-pixel sized trivial accept offsets for each plane */
-   int ei;
-
-   /* one-pixel sized trivial reject offsets for each plane */
-   int eo;
-
    /* edge function values at minx,miny ?? */
    int c;
 
    int dcdx;
    int dcdy;
+
+   /* one-pixel sized trivial reject offsets for each plane */
+   int eo;
 };
 
 /**
@@ -111,17 +109,24 @@ struct lp_rast_plane {
  * Objects of this type are put into the lp_setup_context::data buffer.
  */
 struct lp_rast_triangle {
-   /* inputs for the shader */
-   struct lp_rast_shader_inputs inputs;
-
 #ifdef DEBUG
    float v[3][2];
+   float pad0;
+   float pad1;
 #endif
 
-   struct lp_rast_plane plane[8]; /* NOTE: may allocate fewer planes */
+   /* inputs for the shader */
+   struct lp_rast_shader_inputs inputs;
+   /* planes are also allocated here */
 };
 
 
+#define GET_A0(inputs) ((float (*)[4])((inputs)+1))
+#define GET_DADX(inputs) ((float (*)[4])((char *)((inputs) + 1) + (inputs)->stride))
+#define GET_DADY(inputs) ((float (*)[4])((char *)((inputs) + 1) + 2 * (inputs)->stride))
+#define GET_PLANES(tri) ((struct lp_rast_plane *)((char *)(&(tri)->inputs + 1) + 3 * (tri)->inputs.stride))
+
+
 
 struct lp_rasterizer *
 lp_rast_create( unsigned num_threads );
@@ -152,6 +157,7 @@ union lp_rast_cmd_arg {
       uint32_t value;
       uint32_t mask;
    } clear_zstencil;
+   const struct lp_rast_state *state;
    struct lp_fence *fence;
    struct llvmpipe_query *query_obj;
 };
@@ -243,8 +249,9 @@ lp_rast_arg_null( void )
 #define LP_RAST_OP_SHADE_TILE_OPAQUE 0xe
 #define LP_RAST_OP_BEGIN_QUERY       0xf
 #define LP_RAST_OP_END_QUERY         0x10
+#define LP_RAST_OP_SET_STATE         0x11
 
-#define LP_RAST_OP_MAX               0x11
+#define LP_RAST_OP_MAX               0x12
 #define LP_RAST_OP_MASK              0xff
 
 void
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_debug.c b/src/gallium/drivers/llvmpipe/lp_rast_debug.c
index 6f4ba1c6fef..64ac616f629 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_debug.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_debug.c
@@ -12,6 +12,7 @@ static INLINE int u_bit_scan(unsigned *mask)
 struct tile {
    int coverage;
    int overdraw;
+   const struct lp_rast_state *state;
    char data[TILE_SIZE][TILE_SIZE];
 };
 
@@ -47,6 +48,7 @@ static const char *cmd_names[LP_RAST_OP_MAX] =
    "shade_tile_opaque",
    "begin_query",
    "end_query",
+   "set_state",
 };
 
 static const char *cmd_name(unsigned cmd)
@@ -56,31 +58,31 @@ static const char *cmd_name(unsigned cmd)
 }
 
 static const struct lp_fragment_shader_variant *
-get_variant(  const struct cmd_block *block,
-              int k )
+get_variant( const struct lp_rast_state *state,
+             const struct cmd_block *block,
+             int k )
 {
    if (block->cmd[k] == LP_RAST_OP_SHADE_TILE ||
-       block->cmd[k] == LP_RAST_OP_SHADE_TILE_OPAQUE)
-      return  block->arg[k].shade_tile->state->variant;
-
-   if (block->cmd[k] == LP_RAST_OP_TRIANGLE_1 ||
+       block->cmd[k] == LP_RAST_OP_SHADE_TILE_OPAQUE ||
+       block->cmd[k] == LP_RAST_OP_TRIANGLE_1 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_2 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_3 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_4 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_5 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_6 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_7)
-      return block->arg[k].triangle.tri->inputs.state->variant;
+      return state->variant;
 
    return NULL;
 }
 
 
 static boolean
-is_blend( const struct cmd_block *block,
+is_blend( const struct lp_rast_state *state,
+          const struct cmd_block *block,
           int k )
 {
-   const struct lp_fragment_shader_variant *variant = get_variant(block, k);
+   const struct lp_fragment_shader_variant *variant = get_variant(state, block, k);
 
    if (variant)
       return  variant->key.blend.rt[0].blend_enable;
@@ -93,6 +95,7 @@ is_blend( const struct cmd_block *block,
 static void
 debug_bin( const struct cmd_bin *bin )
 {
+   const struct lp_rast_state *state = NULL;
    const struct cmd_block *head = bin->head;
    int i, j = 0;
 
@@ -100,9 +103,12 @@ debug_bin( const struct cmd_bin *bin )
                 
    while (head) {
       for (i = 0; i < head->count; i++, j++) {
+         if (head->cmd[i] == LP_RAST_OP_SET_STATE)
+            state = head->arg[i].state;
+
          debug_printf("%d: %s %s\n", j,
                       cmd_name(head->cmd[i]),
-                      is_blend(head, i) ? "blended" : "");
+                      is_blend(state, head, i) ? "blended" : "");
       }
       head = head->next;
    }
@@ -134,7 +140,7 @@ debug_shade_tile(int x, int y,
                  char val)
 {
    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
-   boolean blend = inputs->state->variant->key.blend.rt[0].blend_enable;
+   boolean blend = tile->state->variant->key.blend.rt[0].blend_enable;
    unsigned i,j;
 
    if (inputs->disable)
@@ -172,11 +178,12 @@ debug_triangle(int tilex, int tiley,
 {
    const struct lp_rast_triangle *tri = arg.triangle.tri;
    unsigned plane_mask = arg.triangle.plane_mask;
+   const struct lp_rast_plane *tri_plane = GET_PLANES(tri);
    struct lp_rast_plane plane[8];
    int x, y;
    int count = 0;
    unsigned i, nr_planes = 0;
-   boolean blend = tri->inputs.state->variant->key.blend.rt[0].blend_enable;
+   boolean blend = tile->state->variant->key.blend.rt[0].blend_enable;
 
    if (tri->inputs.disable) {
       /* This triangle was partially binned and has been disabled */
@@ -184,7 +191,7 @@ debug_triangle(int tilex, int tiley,
    }
 
    while (plane_mask) {
-      plane[nr_planes] = tri->plane[u_bit_scan(&plane_mask)];
+      plane[nr_planes] = tri_plane[u_bit_scan(&plane_mask)];
       plane[nr_planes].c = (plane[nr_planes].c +
                             plane[nr_planes].dcdy * tiley -
                             plane[nr_planes].dcdx * tilex);
@@ -233,15 +240,19 @@ do_debug_bin( struct tile *tile,
    memset(tile->data, ' ', sizeof tile->data);
    tile->coverage = 0;
    tile->overdraw = 0;
+   tile->state = NULL;
 
    for (block = bin->head; block; block = block->next) {
       for (k = 0; k < block->count; k++, j++) {
-         boolean blend = is_blend(block, k);
+         boolean blend = is_blend(tile->state, block, k);
          char val = get_label(j);
          int count = 0;
             
          if (print_cmds)
             debug_printf("%c: %15s", val, cmd_name(block->cmd[k]));
+
+         if (block->cmd[k] == LP_RAST_OP_SET_STATE)
+            tile->state = block->arg[k].state;
          
          if (block->cmd[k] == LP_RAST_OP_CLEAR_COLOR ||
              block->cmd[k] == LP_RAST_OP_CLEAR_ZSTENCIL)
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_priv.h b/src/gallium/drivers/llvmpipe/lp_rast_priv.h
index 104000a040c..b30408f097b 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_priv.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast_priv.h
@@ -77,6 +77,7 @@ struct cmd_bin;
 struct lp_rasterizer_task
 {
    const struct cmd_bin *bin;
+   const struct lp_rast_state *state;
 
    struct lp_scene *scene;
    unsigned x, y;          /**< Pos of this tile in framebuffer, in pixels */
@@ -244,7 +245,7 @@ lp_rast_shade_quads_all( struct lp_rasterizer_task *task,
                          unsigned x, unsigned y )
 {
    const struct lp_scene *scene = task->scene;
-   const struct lp_rast_state *state = inputs->state;
+   const struct lp_rast_state *state = task->state;
    struct lp_fragment_shader_variant *variant = state->variant;
    uint8_t *color[PIPE_MAX_COLOR_BUFS];
    void *depth;
@@ -260,10 +261,10 @@ lp_rast_shade_quads_all( struct lp_rasterizer_task *task,
    BEGIN_JIT_CALL(state);
    variant->jit_function[RAST_WHOLE]( &state->jit_context,
                                       x, y,
-                                      inputs->facing,
-                                      inputs->a0,
-                                      inputs->dadx,
-                                      inputs->dady,
+                                      inputs->frontfacing,
+                                      GET_A0(inputs),
+                                      GET_DADX(inputs),
+                                      GET_DADY(inputs),
                                       color,
                                       depth,
                                       0xffff,
@@ -298,6 +299,10 @@ void lp_rast_triangle_4_16( struct lp_rasterizer_task *,
                             const union lp_rast_cmd_arg );
 
 void
+lp_rast_set_state(struct lp_rasterizer_task *task,
+                  const union lp_rast_cmd_arg arg);
+ 
+void
 lp_debug_bin( const struct cmd_bin *bin );
 
 #endif
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri.c b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
index f870a187db5..042c315635e 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
@@ -239,6 +239,208 @@ sign_bits4(const __m128i *cstep, int cdiff)
    return _mm_movemask_epi8(result);
 }
 
+
+#define NR_PLANES 3
+
+
+
+
+
+
+
+void
+lp_rast_triangle_3_16(struct lp_rasterizer_task *task,
+                      const union lp_rast_cmd_arg arg)
+{
+   const struct lp_rast_triangle *tri = arg.triangle.tri;
+   const struct lp_rast_plane *plane = GET_PLANES(tri);
+   int x = (arg.triangle.plane_mask & 0xff) + task->x;
+   int y = (arg.triangle.plane_mask >> 8) + task->y;
+   unsigned i, j;
+
+   struct { unsigned mask:16; unsigned i:8; unsigned j:8; } out[16];
+   unsigned nr = 0;
+
+   __m128i p0 = _mm_load_si128((__m128i *)&plane[0]); /* c, dcdx, dcdy, eo */
+   __m128i p1 = _mm_load_si128((__m128i *)&plane[1]); /* c, dcdx, dcdy, eo */
+   __m128i p2 = _mm_load_si128((__m128i *)&plane[2]); /* c, dcdx, dcdy, eo */
+   __m128i zero = _mm_setzero_si128();
+
+   __m128i c;
+   __m128i dcdx;
+   __m128i dcdy;
+   __m128i rej4;
+
+   __m128i dcdx2;
+   __m128i dcdx3;
+   
+   __m128i span_0;                /* 0,dcdx,2dcdx,3dcdx for plane 0 */
+   __m128i span_1;                /* 0,dcdx,2dcdx,3dcdx for plane 1 */
+   __m128i span_2;                /* 0,dcdx,2dcdx,3dcdx for plane 2 */
+   __m128i unused;
+   
+   transpose4_epi32(&p0, &p1, &p2, &zero,
+                    &c, &dcdx, &dcdy, &rej4);
+
+   /* Adjust dcdx;
+    */
+   dcdx = _mm_sub_epi32(zero, dcdx);
+
+   c = _mm_add_epi32(c, mm_mullo_epi32(dcdx, _mm_set1_epi32(x)));
+   c = _mm_add_epi32(c, mm_mullo_epi32(dcdy, _mm_set1_epi32(y)));
+   rej4 = _mm_slli_epi32(rej4, 2);
+
+   dcdx2 = _mm_add_epi32(dcdx, dcdx);
+   dcdx3 = _mm_add_epi32(dcdx2, dcdx);
+
+   transpose4_epi32(&zero, &dcdx, &dcdx2, &dcdx3,
+                    &span_0, &span_1, &span_2, &unused);
+
+   for (i = 0; i < 4; i++) {
+      __m128i cx = c;
+
+      for (j = 0; j < 4; j++) {
+         __m128i c4rej = _mm_add_epi32(cx, rej4);
+         __m128i rej_masks = _mm_srai_epi32(c4rej, 31);
+
+         /* if (is_zero(rej_masks)) */
+         if (_mm_movemask_epi8(rej_masks) == 0) {
+            __m128i c0_0 = _mm_add_epi32(SCALAR_EPI32(cx, 0), span_0);
+            __m128i c1_0 = _mm_add_epi32(SCALAR_EPI32(cx, 1), span_1);
+            __m128i c2_0 = _mm_add_epi32(SCALAR_EPI32(cx, 2), span_2);
+
+            __m128i c_0 = _mm_or_si128(_mm_or_si128(c0_0, c1_0), c2_0);
+
+            __m128i c0_1 = _mm_add_epi32(c0_0, SCALAR_EPI32(dcdy, 0));
+            __m128i c1_1 = _mm_add_epi32(c1_0, SCALAR_EPI32(dcdy, 1));
+            __m128i c2_1 = _mm_add_epi32(c2_0, SCALAR_EPI32(dcdy, 2));
+
+            __m128i c_1 = _mm_or_si128(_mm_or_si128(c0_1, c1_1), c2_1);
+            __m128i c_01 = _mm_packs_epi32(c_0, c_1);
+
+            __m128i c0_2 = _mm_add_epi32(c0_1, SCALAR_EPI32(dcdy, 0));
+            __m128i c1_2 = _mm_add_epi32(c1_1, SCALAR_EPI32(dcdy, 1));
+            __m128i c2_2 = _mm_add_epi32(c2_1, SCALAR_EPI32(dcdy, 2));
+
+            __m128i c_2 = _mm_or_si128(_mm_or_si128(c0_2, c1_2), c2_2);
+
+            __m128i c0_3 = _mm_add_epi32(c0_2, SCALAR_EPI32(dcdy, 0));
+            __m128i c1_3 = _mm_add_epi32(c1_2, SCALAR_EPI32(dcdy, 1));
+            __m128i c2_3 = _mm_add_epi32(c2_2, SCALAR_EPI32(dcdy, 2));
+
+            __m128i c_3 = _mm_or_si128(_mm_or_si128(c0_3, c1_3), c2_3);
+            __m128i c_23 = _mm_packs_epi32(c_2, c_3);
+            __m128i c_0123 = _mm_packs_epi16(c_01, c_23);
+
+            unsigned mask = _mm_movemask_epi8(c_0123);
+
+            out[nr].i = i;
+            out[nr].j = j;
+            out[nr].mask = mask;
+            if (mask != 0xffff)
+               nr++;
+         }
+         cx = _mm_add_epi32(cx, _mm_slli_epi32(dcdx, 2));
+      }
+
+      c = _mm_add_epi32(c, _mm_slli_epi32(dcdy, 2));
+   }
+
+   for (i = 0; i < nr; i++)
+      lp_rast_shade_quads_mask(task,
+                               &tri->inputs,
+                               x + 4 * out[i].j,
+                               y + 4 * out[i].i,
+                               0xffff & ~out[i].mask);
+}
+
+
+
+
+
+void
+lp_rast_triangle_3_4(struct lp_rasterizer_task *task,
+                     const union lp_rast_cmd_arg arg)
+{
+   const struct lp_rast_triangle *tri = arg.triangle.tri;
+   const struct lp_rast_plane *plane = GET_PLANES(tri);
+   int x = (arg.triangle.plane_mask & 0xff) + task->x;
+   int y = (arg.triangle.plane_mask >> 8) + task->y;
+
+   __m128i p0 = _mm_load_si128((__m128i *)&plane[0]); /* c, dcdx, dcdy, eo */
+   __m128i p1 = _mm_load_si128((__m128i *)&plane[1]); /* c, dcdx, dcdy, eo */
+   __m128i p2 = _mm_load_si128((__m128i *)&plane[2]); /* c, dcdx, dcdy, eo */
+   __m128i zero = _mm_setzero_si128();
+
+   __m128i c;
+   __m128i dcdx;
+   __m128i dcdy;
+
+   __m128i dcdx2;
+   __m128i dcdx3;
+   
+   __m128i span_0;                /* 0,dcdx,2dcdx,3dcdx for plane 0 */
+   __m128i span_1;                /* 0,dcdx,2dcdx,3dcdx for plane 1 */
+   __m128i span_2;                /* 0,dcdx,2dcdx,3dcdx for plane 2 */
+   __m128i unused;
+   
+   transpose4_epi32(&p0, &p1, &p2, &zero,
+                    &c, &dcdx, &dcdy, &unused);
+
+   /* Adjust dcdx;
+    */
+   dcdx = _mm_sub_epi32(zero, dcdx);
+
+   c = _mm_add_epi32(c, mm_mullo_epi32(dcdx, _mm_set1_epi32(x)));
+   c = _mm_add_epi32(c, mm_mullo_epi32(dcdy, _mm_set1_epi32(y)));
+
+   dcdx2 = _mm_add_epi32(dcdx, dcdx);
+   dcdx3 = _mm_add_epi32(dcdx2, dcdx);
+
+   transpose4_epi32(&zero, &dcdx, &dcdx2, &dcdx3,
+                    &span_0, &span_1, &span_2, &unused);
+
+
+   {
+      __m128i c0_0 = _mm_add_epi32(SCALAR_EPI32(c, 0), span_0);
+      __m128i c1_0 = _mm_add_epi32(SCALAR_EPI32(c, 1), span_1);
+      __m128i c2_0 = _mm_add_epi32(SCALAR_EPI32(c, 2), span_2);
+      
+      __m128i c_0 = _mm_or_si128(_mm_or_si128(c0_0, c1_0), c2_0);
+
+      __m128i c0_1 = _mm_add_epi32(c0_0, SCALAR_EPI32(dcdy, 0));
+      __m128i c1_1 = _mm_add_epi32(c1_0, SCALAR_EPI32(dcdy, 1));
+      __m128i c2_1 = _mm_add_epi32(c2_0, SCALAR_EPI32(dcdy, 2));
+
+      __m128i c_1 = _mm_or_si128(_mm_or_si128(c0_1, c1_1), c2_1);
+      __m128i c_01 = _mm_packs_epi32(c_0, c_1);
+
+      __m128i c0_2 = _mm_add_epi32(c0_1, SCALAR_EPI32(dcdy, 0));
+      __m128i c1_2 = _mm_add_epi32(c1_1, SCALAR_EPI32(dcdy, 1));
+      __m128i c2_2 = _mm_add_epi32(c2_1, SCALAR_EPI32(dcdy, 2));
+
+      __m128i c_2 = _mm_or_si128(_mm_or_si128(c0_2, c1_2), c2_2);
+
+      __m128i c0_3 = _mm_add_epi32(c0_2, SCALAR_EPI32(dcdy, 0));
+      __m128i c1_3 = _mm_add_epi32(c1_2, SCALAR_EPI32(dcdy, 1));
+      __m128i c2_3 = _mm_add_epi32(c2_2, SCALAR_EPI32(dcdy, 2));
+
+      __m128i c_3 = _mm_or_si128(_mm_or_si128(c0_3, c1_3), c2_3);
+      __m128i c_23 = _mm_packs_epi32(c_2, c_3);
+      __m128i c_0123 = _mm_packs_epi16(c_01, c_23);
+
+      unsigned mask = _mm_movemask_epi8(c_0123);
+
+      if (mask != 0xffff)
+         lp_rast_shade_quads_mask(task,
+                                  &tri->inputs,
+                                  x,
+                                  y,
+                                  0xffff & ~mask);
+   }
+}
+
+#undef NR_PLANES
 #endif
 
 
@@ -254,8 +456,8 @@ sign_bits4(const __m128i *cstep, int cdiff)
 
 #define TAG(x) x##_3
 #define NR_PLANES 3
-#define TRI_4 lp_rast_triangle_3_4
-#define TRI_16 lp_rast_triangle_3_16
+/*#define TRI_4 lp_rast_triangle_3_4*/
+/*#define TRI_16 lp_rast_triangle_3_16*/
 #include "lp_rast_tri_tmp.h"
 
 #define TAG(x) x##_4
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
index 2f032295126..4825d651c04 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
@@ -82,7 +82,8 @@ TAG(do_block_16)(struct lp_rasterizer_task *task,
       const int dcdx = -plane[j].dcdx * 4;
       const int dcdy = plane[j].dcdy * 4;
       const int cox = plane[j].eo * 4;
-      const int cio = plane[j].ei * 4 - 1;
+      const int ei = plane[j].dcdy - plane[j].dcdx - plane[j].eo;
+      const int cio = ei * 4 - 1;
 
       build_masks(c[j] + cox,
 		  cio - cox,
@@ -156,6 +157,7 @@ TAG(lp_rast_triangle)(struct lp_rasterizer_task *task,
 {
    const struct lp_rast_triangle *tri = arg.triangle.tri;
    unsigned plane_mask = arg.triangle.plane_mask;
+   const struct lp_rast_plane *tri_plane = GET_PLANES(tri);
    const int x = task->x, y = task->y;
    struct lp_rast_plane plane[NR_PLANES];
    int c[NR_PLANES];
@@ -172,7 +174,7 @@ TAG(lp_rast_triangle)(struct lp_rasterizer_task *task,
 
    while (plane_mask) {
       int i = ffs(plane_mask) - 1;
-      plane[j] = tri->plane[i];
+      plane[j] = tri_plane[i];
       plane_mask &= ~(1 << i);
       c[j] = plane[j].c + plane[j].dcdy * y - plane[j].dcdx * x;
 
@@ -180,7 +182,8 @@ TAG(lp_rast_triangle)(struct lp_rasterizer_task *task,
 	 const int dcdx = -plane[j].dcdx * 16;
 	 const int dcdy = plane[j].dcdy * 16;
 	 const int cox = plane[j].eo * 16;
-	 const int cio = plane[j].ei * 16 - 1;
+         const int ei = plane[j].dcdy - plane[j].dcdx - plane[j].eo;
+         const int cio = ei * 16 - 1;
 
 	 build_masks(c[j] + cox,
 		     cio - cox,
@@ -255,7 +258,7 @@ TRI_16(struct lp_rasterizer_task *task,
        const union lp_rast_cmd_arg arg)
 {
    const struct lp_rast_triangle *tri = arg.triangle.tri;
-   const struct lp_rast_plane *plane = tri->plane;
+   const struct lp_rast_plane *plane = GET_PLANES(tri);
    unsigned mask = arg.triangle.plane_mask;
    unsigned outmask, partial_mask;
    unsigned j;
@@ -328,7 +331,7 @@ TRI_4(struct lp_rasterizer_task *task,
       const union lp_rast_cmd_arg arg)
 {
    const struct lp_rast_triangle *tri = arg.triangle.tri;
-   const struct lp_rast_plane *plane = tri->plane;
+   const struct lp_rast_plane *plane = GET_PLANES(tri);
    unsigned mask = arg.triangle.plane_mask;
    const int x = task->x + (mask & 0xff);
    const int y = task->y + (mask >> 8);
diff --git a/src/gallium/drivers/llvmpipe/lp_scene.c b/src/gallium/drivers/llvmpipe/lp_scene.c
index 8b504f23a33..a4fdf7cff36 100644
--- a/src/gallium/drivers/llvmpipe/lp_scene.c
+++ b/src/gallium/drivers/llvmpipe/lp_scene.c
@@ -203,7 +203,9 @@ lp_scene_end_rasterization(struct lp_scene *scene )
    for (i = 0; i < scene->tiles_x; i++) {
       for (j = 0; j < scene->tiles_y; j++) {
          struct cmd_bin *bin = lp_scene_get_bin(scene, i, j);
-         bin->head = bin->tail = NULL;
+         bin->head = NULL;
+         bin->tail = NULL;
+         bin->last_state = NULL;
       }
    }
 
diff --git a/src/gallium/drivers/llvmpipe/lp_scene.h b/src/gallium/drivers/llvmpipe/lp_scene.h
index dbef7692e42..622c522f11a 100644
--- a/src/gallium/drivers/llvmpipe/lp_scene.h
+++ b/src/gallium/drivers/llvmpipe/lp_scene.h
@@ -41,6 +41,7 @@
 #include "lp_debug.h"
 
 struct lp_scene_queue;
+struct lp_rast_state;
 
 /* We're limited to 2K by 2K for 32bit fixed point rasterization.
  * Will need a 64-bit version for larger framebuffers.
@@ -94,6 +95,7 @@ struct data_block {
 struct cmd_bin {
    ushort x;
    ushort y;
+   const struct lp_rast_state *last_state;       /* most recent state set in bin */
    struct cmd_block *head;
    struct cmd_block *tail;
 };
@@ -297,7 +299,7 @@ lp_scene_bin_command( struct lp_scene *scene,
 
    assert(x < scene->tiles_x);
    assert(y < scene->tiles_y);
-   assert(cmd <= LP_RAST_OP_END_QUERY);
+   assert(cmd < LP_RAST_OP_MAX);
 
    if (tail == NULL || tail->count == CMD_BLOCK_MAX) {
       tail = lp_scene_new_cmd_block( scene, bin );
@@ -318,6 +320,30 @@ lp_scene_bin_command( struct lp_scene *scene,
 }
 
 
+static INLINE boolean
+lp_scene_bin_cmd_with_state( struct lp_scene *scene,
+                             unsigned x, unsigned y,
+                             const struct lp_rast_state *state,
+                             unsigned cmd,
+                             union lp_rast_cmd_arg arg )
+{
+   struct cmd_bin *bin = lp_scene_get_bin(scene, x, y);
+
+   if (state != bin->last_state) {
+      bin->last_state = state;
+      if (!lp_scene_bin_command(scene, x, y,
+                                LP_RAST_OP_SET_STATE,
+                                lp_rast_arg_state(state)))
+         return FALSE;
+   }
+
+   if (!lp_scene_bin_command( scene, x, y, cmd, arg ))
+      return FALSE;
+
+   return TRUE;
+}
+
+
 /* Add a command to all active bins.
  */
 static INLINE boolean
diff --git a/src/gallium/drivers/llvmpipe/lp_screen.c b/src/gallium/drivers/llvmpipe/lp_screen.c
index 96633d93654..ad0ea75b3a3 100644
--- a/src/gallium/drivers/llvmpipe/lp_screen.c
+++ b/src/gallium/drivers/llvmpipe/lp_screen.c
@@ -158,6 +158,8 @@ llvmpipe_get_param(struct pipe_screen *screen, enum pipe_cap param)
    case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT:
    case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER:
       return 0;
+   case PIPE_CAP_PRIMITIVE_RESTART:
+      return 1;
    case PIPE_CAP_DEPTHSTENCIL_CLEAR_SEPARATE:
       return 1;
    case PIPE_CAP_DEPTH_CLAMP:
diff --git a/src/gallium/drivers/llvmpipe/lp_setup.c b/src/gallium/drivers/llvmpipe/lp_setup.c
index e96f012f1bb..6118434d3d3 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup.c
@@ -526,14 +526,12 @@ lp_setup_set_point_state( struct lp_setup_context *setup,
 }
 
 void
-lp_setup_set_fs_inputs( struct lp_setup_context *setup,
-                        const struct lp_shader_input *input,
-                        unsigned nr )
+lp_setup_set_setup_variant( struct lp_setup_context *setup,
+			    const struct lp_setup_variant *variant)
 {
-   LP_DBG(DEBUG_SETUP, "%s %p %u\n", __FUNCTION__, (void *) input, nr);
-
-   memcpy( setup->fs.input, input, nr * sizeof input[0] );
-   setup->fs.nr_inputs = nr;
+   LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
+   
+   setup->setup.variant = variant;
 }
 
 void
@@ -921,6 +919,13 @@ lp_setup_update_state( struct lp_setup_context *setup,
       setup->psize = lp->psize_slot;
 
       assert(lp->dirty == 0);
+
+      assert(lp->setup_variant.key.size == 
+	     setup->setup.variant->key.size);
+
+      assert(memcmp(&lp->setup_variant.key,
+		    &setup->setup.variant->key,
+		    setup->setup.variant->key.size) == 0);
    }
 
    if (update_scene) {
diff --git a/src/gallium/drivers/llvmpipe/lp_setup.h b/src/gallium/drivers/llvmpipe/lp_setup.h
index 25dab78f64f..ebb18f81344 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup.h
+++ b/src/gallium/drivers/llvmpipe/lp_setup.h
@@ -33,28 +33,6 @@
 struct draw_context;
 struct vertex_info;
 
-enum lp_interp {
-   LP_INTERP_CONSTANT,
-   LP_INTERP_LINEAR,
-   LP_INTERP_PERSPECTIVE,
-   LP_INTERP_POSITION,
-   LP_INTERP_FACING
-};
-
-
-/**
- * Describes how to compute the interpolation coefficients (a0, dadx, dady)
- * from the vertices passed into our triangle/line/point functions by the
- * draw module.
- *
- * Vertices are treated as an array of float[4] values, indexed by
- * src_index.
- */
-struct lp_shader_input {
-   enum lp_interp interp;       /* how to interpolate values */
-   unsigned src_index;          /* where to find values in incoming vertices */
-   unsigned usage_mask;         /* bitmask of TGSI_WRITEMASK_x flags */
-};
 
 struct pipe_resource;
 struct pipe_query;
@@ -66,7 +44,7 @@ struct lp_fragment_shader_variant;
 struct lp_jit_context;
 struct llvmpipe_query;
 struct pipe_fence_handle;
-
+struct lp_setup_variant;
 
 struct lp_setup_context *
 lp_setup_create( struct pipe_context *pipe,
@@ -111,9 +89,8 @@ lp_setup_set_point_state( struct lp_setup_context *setup,
                           uint sprite_coord_origin);
 
 void
-lp_setup_set_fs_inputs( struct lp_setup_context *setup,
-                        const struct lp_shader_input *interp,
-                        unsigned nr );
+lp_setup_set_setup_variant( struct lp_setup_context *setup,
+			    const struct lp_setup_variant *variant );
 
 void
 lp_setup_set_fs_variant( struct lp_setup_context *setup,
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_coef.c b/src/gallium/drivers/llvmpipe/lp_setup_coef.c
deleted file mode 100644
index 8dc2688ddb6..00000000000
--- a/src/gallium/drivers/llvmpipe/lp_setup_coef.c
+++ /dev/null
@@ -1,279 +0,0 @@
-/**************************************************************************
- *
- * Copyright 2010, VMware.
- * All Rights Reserved.
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the
- * "Software"), to deal in the Software without restriction, including
- * without limitation the rights to use, copy, modify, merge, publish,
- * distribute, sub license, and/or sell copies of the Software, and to
- * permit persons to whom the Software is furnished to do so, subject to
- * the following conditions:
- *
- * The above copyright notice and this permission notice (including the
- * next paragraph) shall be included in all copies or substantial portions
- * of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
- * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
- * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
- * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
- * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
- * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
- * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
- *
- **************************************************************************/
-
-/*
- * Binning code for triangles
- */
-
-#include "util/u_math.h"
-#include "util/u_memory.h"
-#include "lp_perf.h"
-#include "lp_setup_context.h"
-#include "lp_setup_coef.h"
-#include "lp_rast.h"
-#include "lp_state_fs.h"
-
-#if !defined(PIPE_ARCH_SSE)
-
-/**
- * Compute a0 for a constant-valued coefficient (GL_FLAT shading).
- */
-static void constant_coef( struct lp_rast_shader_inputs *inputs,
-                           unsigned slot,
-			   const float value,
-                           unsigned i )
-{
-   inputs->a0[slot][i] = value;
-   inputs->dadx[slot][i] = 0.0f;
-   inputs->dady[slot][i] = 0.0f;
-}
-
-
-
-static void linear_coef( struct lp_rast_shader_inputs *inputs,
-                         const struct lp_tri_info *info,
-                         unsigned slot,
-                         unsigned vert_attr,
-                         unsigned i)
-{
-   float a0 = info->v0[vert_attr][i];
-   float a1 = info->v1[vert_attr][i];
-   float a2 = info->v2[vert_attr][i];
-
-   float da01 = a0 - a1;
-   float da20 = a2 - a0;
-   float dadx = (da01 * info->dy20_ooa - info->dy01_ooa * da20);
-   float dady = (da20 * info->dx01_ooa - info->dx20_ooa * da01);
-
-   inputs->dadx[slot][i] = dadx;
-   inputs->dady[slot][i] = dady;
-
-   /* calculate a0 as the value which would be sampled for the
-    * fragment at (0,0), taking into account that we want to sample at
-    * pixel centers, in other words (0.5, 0.5).
-    *
-    * this is neat but unfortunately not a good way to do things for
-    * triangles with very large values of dadx or dady as it will
-    * result in the subtraction and re-addition from a0 of a very
-    * large number, which means we'll end up loosing a lot of the
-    * fractional bits and precision from a0.  the way to fix this is
-    * to define a0 as the sample at a pixel center somewhere near vmin
-    * instead - i'll switch to this later.
-    */
-   inputs->a0[slot][i] = a0 - (dadx * info->x0_center +
-				   dady * info->y0_center);
-}
-
-
-/**
- * Compute a0, dadx and dady for a perspective-corrected interpolant,
- * for a triangle.
- * We basically multiply the vertex value by 1/w before computing
- * the plane coefficients (a0, dadx, dady).
- * Later, when we compute the value at a particular fragment position we'll
- * divide the interpolated value by the interpolated W at that fragment.
- */
-static void perspective_coef( struct lp_rast_shader_inputs *inputs,
-                              const struct lp_tri_info *info,
-                              unsigned slot,
-			      unsigned vert_attr,
-                              unsigned i)
-{
-   /* premultiply by 1/w  (v[0][3] is always 1/w):
-    */
-   float a0 = info->v0[vert_attr][i] * info->v0[0][3];
-   float a1 = info->v1[vert_attr][i] * info->v1[0][3];
-   float a2 = info->v2[vert_attr][i] * info->v2[0][3];
-   float da01 = a0 - a1;
-   float da20 = a2 - a0;
-   float dadx = da01 * info->dy20_ooa - info->dy01_ooa * da20;
-   float dady = da20 * info->dx01_ooa - info->dx20_ooa * da01;
-
-   inputs->dadx[slot][i] = dadx;
-   inputs->dady[slot][i] = dady;
-   inputs->a0[slot][i] = a0 - (dadx * info->x0_center +
-				   dady * info->y0_center);
-}
-
-
-/**
- * Special coefficient setup for gl_FragCoord.
- * X and Y are trivial
- * Z and W are copied from position_coef which should have already been computed.
- * We could do a bit less work if we'd examine gl_FragCoord's swizzle mask.
- */
-static void
-setup_fragcoord_coef(struct lp_rast_shader_inputs *inputs,
-                     const struct lp_tri_info *info,
-                     unsigned slot,
-                     unsigned usage_mask)
-{
-   /*X*/
-   if (usage_mask & TGSI_WRITEMASK_X) {
-      inputs->a0[slot][0] = 0.0;
-      inputs->dadx[slot][0] = 1.0;
-      inputs->dady[slot][0] = 0.0;
-   }
-
-   /*Y*/
-   if (usage_mask & TGSI_WRITEMASK_Y) {
-      inputs->a0[slot][1] = 0.0;
-      inputs->dadx[slot][1] = 0.0;
-      inputs->dady[slot][1] = 1.0;
-   }
-
-   /*Z*/
-   if (usage_mask & TGSI_WRITEMASK_Z) {
-      linear_coef(inputs, info, slot, 0, 2);
-   }
-
-   /*W*/
-   if (usage_mask & TGSI_WRITEMASK_W) {
-      linear_coef(inputs, info, slot, 0, 3);
-   }
-}
-
-
-/**
- * Setup the fragment input attribute with the front-facing value.
- * \param frontface  is the triangle front facing?
- */
-static void setup_facing_coef( struct lp_rast_shader_inputs *inputs,
-                               unsigned slot,
-                               boolean frontface,
-                               unsigned usage_mask)
-{
-   /* convert TRUE to 1.0 and FALSE to -1.0 */
-   if (usage_mask & TGSI_WRITEMASK_X)
-      constant_coef( inputs, slot, 2.0f * frontface - 1.0f, 0 );
-
-   if (usage_mask & TGSI_WRITEMASK_Y)
-      constant_coef( inputs, slot, 0.0f, 1 ); /* wasted */
-
-   if (usage_mask & TGSI_WRITEMASK_Z)
-      constant_coef( inputs, slot, 0.0f, 2 ); /* wasted */
-
-   if (usage_mask & TGSI_WRITEMASK_W)
-      constant_coef( inputs, slot, 0.0f, 3 ); /* wasted */
-}
-
-
-/**
- * Compute the tri->coef[] array dadx, dady, a0 values.
- */
-void lp_setup_tri_coef( struct lp_setup_context *setup,
-			struct lp_rast_shader_inputs *inputs,
-                        const float (*v0)[4],
-                        const float (*v1)[4],
-                        const float (*v2)[4],
-                        boolean frontfacing)
-{
-   unsigned fragcoord_usage_mask = TGSI_WRITEMASK_XYZ;
-   unsigned slot;
-   unsigned i;
-   struct lp_tri_info info;
-   float dx01 = v0[0][0] - v1[0][0];
-   float dy01 = v0[0][1] - v1[0][1];
-   float dx20 = v2[0][0] - v0[0][0];
-   float dy20 = v2[0][1] - v0[0][1];
-   float oneoverarea = 1.0f / (dx01 * dy20 - dx20 * dy01);
-
-   info.v0 = v0;
-   info.v1 = v1;
-   info.v2 = v2;
-   info.frontfacing = frontfacing;
-   info.x0_center = v0[0][0] - setup->pixel_offset;
-   info.y0_center = v0[0][1] - setup->pixel_offset;
-   info.dx01_ooa  = dx01 * oneoverarea;
-   info.dx20_ooa  = dx20 * oneoverarea;
-   info.dy01_ooa  = dy01 * oneoverarea;
-   info.dy20_ooa  = dy20 * oneoverarea;
-
-
-   /* setup interpolation for all the remaining attributes:
-    */
-   for (slot = 0; slot < setup->fs.nr_inputs; slot++) {
-      unsigned vert_attr = setup->fs.input[slot].src_index;
-      unsigned usage_mask = setup->fs.input[slot].usage_mask;
-
-      switch (setup->fs.input[slot].interp) {
-      case LP_INTERP_CONSTANT:
-         if (setup->flatshade_first) {
-            for (i = 0; i < NUM_CHANNELS; i++)
-               if (usage_mask & (1 << i))
-                  constant_coef(inputs, slot+1, info.v0[vert_attr][i], i);
-         }
-         else {
-            for (i = 0; i < NUM_CHANNELS; i++)
-               if (usage_mask & (1 << i))
-                  constant_coef(inputs, slot+1, info.v2[vert_attr][i], i);
-         }
-         break;
-
-      case LP_INTERP_LINEAR:
-         for (i = 0; i < NUM_CHANNELS; i++)
-            if (usage_mask & (1 << i))
-               linear_coef(inputs, &info, slot+1, vert_attr, i);
-         break;
-
-      case LP_INTERP_PERSPECTIVE:
-         for (i = 0; i < NUM_CHANNELS; i++)
-            if (usage_mask & (1 << i))
-               perspective_coef(inputs, &info, slot+1, vert_attr, i);
-         fragcoord_usage_mask |= TGSI_WRITEMASK_W;
-         break;
-
-      case LP_INTERP_POSITION:
-         /*
-          * The generated pixel interpolators will pick up the coeffs from
-          * slot 0, so all need to ensure that the usage mask is covers all
-          * usages.
-          */
-         fragcoord_usage_mask |= usage_mask;
-         break;
-
-      case LP_INTERP_FACING:
-         setup_facing_coef(inputs, slot+1, info.frontfacing, usage_mask);
-         break;
-
-      default:
-         assert(0);
-      }
-   }
-
-   /* The internal position input is in slot zero:
-    */
-   setup_fragcoord_coef(inputs, &info, 0, fragcoord_usage_mask);
-}
-
-#else
-extern void lp_setup_coef_dummy(void);
-void lp_setup_coef_dummy(void)
-{
-}
-
-#endif
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_coef.h b/src/gallium/drivers/llvmpipe/lp_setup_coef.h
deleted file mode 100644
index 87a3255ccc6..00000000000
--- a/src/gallium/drivers/llvmpipe/lp_setup_coef.h
+++ /dev/null
@@ -1,64 +0,0 @@
-/**************************************************************************
- *
- * Copyright 2010 VMware, Inc.
- * All Rights Reserved.
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the
- * "Software"), to deal in the Software without restriction, including
- * without limitation the rights to use, copy, modify, merge, publish,
- * distribute, sub license, and/or sell copies of the Software, and to
- * permit persons to whom the Software is furnished to do so, subject to
- * the following conditions:
- *
- * The above copyright notice and this permission notice (including the
- * next paragraph) shall be included in all copies or substantial portions
- * of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
- * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
- * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
- * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
- * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
- * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
- * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
- *
- **************************************************************************/
-
-
-/**
- * The setup code is concerned with point/line/triangle setup and
- * putting commands/data into the bins.
- */
-
-
-#ifndef LP_SETUP_COEF_H
-#define LP_SETUP_COEF_H
-
-
-struct lp_tri_info {
-
-   float x0_center;
-   float y0_center;
-
-   /* turn these into an aligned float[4] */
-   float dy01_ooa;
-   float dy20_ooa;
-   float dx01_ooa;
-   float dx20_ooa;
-
-   const float (*v0)[4];
-   const float (*v1)[4];
-   const float (*v2)[4];
-
-   boolean frontfacing;		/* remove eventually */
-};
-
-void lp_setup_tri_coef( struct lp_setup_context *setup,
-			struct lp_rast_shader_inputs *inputs,
-                        const float (*v0)[4],
-                        const float (*v1)[4],
-                        const float (*v2)[4],
-                        boolean frontfacing);
-
-#endif
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_coef_intrin.c b/src/gallium/drivers/llvmpipe/lp_setup_coef_intrin.c
deleted file mode 100644
index 3742fd672b2..00000000000
--- a/src/gallium/drivers/llvmpipe/lp_setup_coef_intrin.c
+++ /dev/null
@@ -1,228 +0,0 @@
-/**************************************************************************
- *
- * Copyright 2010 VMware.
- * All Rights Reserved.
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the
- * "Software"), to deal in the Software without restriction, including
- * without limitation the rights to use, copy, modify, merge, publish,
- * distribute, sub license, and/or sell copies of the Software, and to
- * permit persons to whom the Software is furnished to do so, subject to
- * the following conditions:
- *
- * The above copyright notice and this permission notice (including the
- * next paragraph) shall be included in all copies or substantial portions
- * of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
- * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
- * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
- * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
- * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
- * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
- * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
- *
- **************************************************************************/
-
-/*
- * Binning code for triangles
- */
-
-#include "util/u_math.h"
-#include "util/u_memory.h"
-#include "lp_perf.h"
-#include "lp_setup_context.h"
-#include "lp_setup_coef.h"
-#include "lp_rast.h"
-
-#if defined(PIPE_ARCH_SSE)
-#include <emmintrin.h>
-
-
-static void constant_coef4( struct lp_rast_shader_inputs *inputs,
-			    const struct lp_tri_info *info,
-			    unsigned slot,
-			    const float *attr)
-{
-   *(__m128 *)inputs->a0[slot]   = *(__m128 *)attr;
-   *(__m128 *)inputs->dadx[slot] = _mm_set1_ps(0.0);
-   *(__m128 *)inputs->dady[slot] = _mm_set1_ps(0.0);
-}
-
-
-
-/**
- * Setup the fragment input attribute with the front-facing value.
- * \param frontface  is the triangle front facing?
- */
-static void setup_facing_coef( struct lp_rast_shader_inputs *inputs,
-			       const struct lp_tri_info *info,
-			       unsigned slot )
-{
-   /* XXX: just pass frontface directly to the shader, don't bother
-    * treating it as an input.
-    */
-   __m128 a0 = _mm_setr_ps(info->frontfacing ? 1.0 : -1.0,
-			   0, 0, 0);
-
-   *(__m128 *)inputs->a0[slot]   = a0;
-   *(__m128 *)inputs->dadx[slot] = _mm_set1_ps(0.0);
-   *(__m128 *)inputs->dady[slot] = _mm_set1_ps(0.0);
-}
-
-
-
-static void calc_coef4( struct lp_rast_shader_inputs *inputs,
-			const struct lp_tri_info *info,
-			unsigned slot,
-			__m128 a0,
-			__m128 a1,
-			__m128 a2)
-{
-   __m128 da01          = _mm_sub_ps(a0, a1);
-   __m128 da20          = _mm_sub_ps(a2, a0);
-
-   __m128 da01_dy20_ooa = _mm_mul_ps(da01, _mm_set1_ps(info->dy20_ooa));
-   __m128 da20_dy01_ooa = _mm_mul_ps(da20, _mm_set1_ps(info->dy01_ooa));   
-   __m128 dadx          = _mm_sub_ps(da01_dy20_ooa, da20_dy01_ooa);
-
-   __m128 da01_dx20_ooa = _mm_mul_ps(da01, _mm_set1_ps(info->dx20_ooa));
-   __m128 da20_dx01_ooa = _mm_mul_ps(da20, _mm_set1_ps(info->dx01_ooa));
-   __m128 dady          = _mm_sub_ps(da20_dx01_ooa, da01_dx20_ooa);
-
-   __m128 dadx_x0       = _mm_mul_ps(dadx, _mm_set1_ps(info->x0_center));
-   __m128 dady_y0       = _mm_mul_ps(dady, _mm_set1_ps(info->y0_center));
-   __m128 attr_v0       = _mm_add_ps(dadx_x0, dady_y0);
-   __m128 attr_0        = _mm_sub_ps(a0, attr_v0);
-
-   *(__m128 *)inputs->a0[slot]   = attr_0;
-   *(__m128 *)inputs->dadx[slot] = dadx;
-   *(__m128 *)inputs->dady[slot] = dady;
-}
-
-
-static void linear_coef( struct lp_rast_shader_inputs *inputs,
-                         const struct lp_tri_info *info,
-                         unsigned slot,
-                         unsigned vert_attr)
-{
-   __m128 a0 = *(const __m128 *)info->v0[vert_attr];
-   __m128 a1 = *(const __m128 *)info->v1[vert_attr];
-   __m128 a2 = *(const __m128 *)info->v2[vert_attr];
-
-   calc_coef4(inputs, info, slot, a0, a1, a2);
-}
-
-
-
-/**
- * Compute a0, dadx and dady for a perspective-corrected interpolant,
- * for a triangle.
- * We basically multiply the vertex value by 1/w before computing
- * the plane coefficients (a0, dadx, dady).
- * Later, when we compute the value at a particular fragment position we'll
- * divide the interpolated value by the interpolated W at that fragment.
- */
-static void perspective_coef( struct lp_rast_shader_inputs *inputs,
-                              const struct lp_tri_info *info,
-                              unsigned slot,
-			      unsigned vert_attr)
-{
-   /* premultiply by 1/w  (v[0][3] is always 1/w):
-    */
-   __m128 a0 = *(const __m128 *)info->v0[vert_attr];
-   __m128 a1 = *(const __m128 *)info->v1[vert_attr];
-   __m128 a2 = *(const __m128 *)info->v2[vert_attr];
-
-   __m128 a0_oow = _mm_mul_ps(a0, _mm_set1_ps(info->v0[0][3]));
-   __m128 a1_oow = _mm_mul_ps(a1, _mm_set1_ps(info->v1[0][3]));
-   __m128 a2_oow = _mm_mul_ps(a2, _mm_set1_ps(info->v2[0][3]));
-
-   calc_coef4(inputs, info, slot, a0_oow, a1_oow, a2_oow);
-}
-
-
-
-
-
-/**
- * Compute the inputs-> dadx, dady, a0 values.
- */
-void lp_setup_tri_coef( struct lp_setup_context *setup,
-			struct lp_rast_shader_inputs *inputs,
-                        const float (*v0)[4],
-                        const float (*v1)[4],
-                        const float (*v2)[4],
-                        boolean frontfacing)
-{
-   unsigned slot;
-   struct lp_tri_info info;
-   float dx01 = v0[0][0] - v1[0][0];
-   float dy01 = v0[0][1] - v1[0][1];
-   float dx20 = v2[0][0] - v0[0][0];
-   float dy20 = v2[0][1] - v0[0][1];
-   float oneoverarea = 1.0f / (dx01 * dy20 - dx20 * dy01);
-
-   info.v0 = v0;
-   info.v1 = v1;
-   info.v2 = v2;
-   info.frontfacing = frontfacing;
-   info.x0_center = v0[0][0] - setup->pixel_offset;
-   info.y0_center = v0[0][1] - setup->pixel_offset;
-   info.dx01_ooa  = dx01 * oneoverarea;
-   info.dx20_ooa  = dx20 * oneoverarea;
-   info.dy01_ooa  = dy01 * oneoverarea;
-   info.dy20_ooa  = dy20 * oneoverarea;
-
-
-   /* The internal position input is in slot zero:
-    */
-   linear_coef(inputs, &info, 0, 0);
-
-   /* setup interpolation for all the remaining attributes:
-    */
-   for (slot = 0; slot < setup->fs.nr_inputs; slot++) {
-      unsigned vert_attr = setup->fs.input[slot].src_index;
-
-      switch (setup->fs.input[slot].interp) {
-      case LP_INTERP_CONSTANT:
-         if (setup->flatshade_first) {
-	    constant_coef4(inputs, &info, slot+1, info.v0[vert_attr]);
-         }
-         else {
-	    constant_coef4(inputs, &info, slot+1, info.v2[vert_attr]);
-         }
-         break;
-
-      case LP_INTERP_LINEAR:
-	 linear_coef(inputs, &info, slot+1, vert_attr);
-         break;
-
-      case LP_INTERP_PERSPECTIVE:
-	 perspective_coef(inputs, &info, slot+1, vert_attr);
-         break;
-
-      case LP_INTERP_POSITION:
-         /*
-          * The generated pixel interpolators will pick up the coeffs from
-          * slot 0.
-          */
-         break;
-
-      case LP_INTERP_FACING:
-         setup_facing_coef(inputs, &info, slot+1);
-         break;
-
-      default:
-         assert(0);
-      }
-   }
-}
-
-#else
-extern void lp_setup_coef_dummy(void);
-void lp_setup_coef_dummy(void)
-{
-}
-#endif
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_context.h b/src/gallium/drivers/llvmpipe/lp_setup_context.h
index e7b425ebcbc..dc2533bedc5 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_context.h
+++ b/src/gallium/drivers/llvmpipe/lp_setup_context.h
@@ -39,6 +39,7 @@
 #include "lp_rast.h"
 #include "lp_tile_soa.h"        /* for TILE_SIZE */
 #include "lp_scene.h"
+#include "lp_bld_interp.h"	/* for struct lp_shader_input */
 
 #include "draw/draw_vbuf.h"
 #include "util/u_rect.h"
@@ -49,6 +50,8 @@
 #define LP_SETUP_NEW_SCISSOR     0x08
 
 
+struct lp_setup_variant;
+
 
 /** Max number of scenes */
 #define MAX_SCENES 2
@@ -118,9 +121,6 @@ struct lp_setup_context
    } state;
    
    struct {
-      struct lp_shader_input input[PIPE_MAX_ATTRIBS];
-      unsigned nr_inputs;
-
       const struct lp_rast_state *stored; /**< what's in the scene */
       struct lp_rast_state current;  /**< currently set state */
       struct pipe_resource *current_tex[PIPE_MAX_SAMPLERS];
@@ -139,6 +139,10 @@ struct lp_setup_context
    } blend_color;
 
 
+   struct {
+      const struct lp_setup_variant *variant;
+   } setup;
+
    unsigned dirty;   /**< bitmask of LP_SETUP_NEW_x bits */
 
    void (*point)( struct lp_setup_context *,
@@ -181,7 +185,7 @@ lp_setup_print_vertex(struct lp_setup_context *setup,
 
 struct lp_rast_triangle *
 lp_setup_alloc_triangle(struct lp_scene *scene,
-                        unsigned nr_inputs,
+                        unsigned num_inputs,
                         unsigned nr_planes,
                         unsigned *tri_size);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_line.c b/src/gallium/drivers/llvmpipe/lp_setup_line.c
index 4d7d6235b07..827413bb33e 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_line.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_line.c
@@ -35,6 +35,7 @@
 #include "lp_setup_context.h"
 #include "lp_rast.h"
 #include "lp_state_fs.h"
+#include "lp_state_setup.h"
 
 #define NUM_CHANNELS 4
 
@@ -46,6 +47,10 @@ struct lp_line_info {
 
    const float (*v1)[4];
    const float (*v2)[4];
+
+   float (*a0)[4];
+   float (*dadx)[4];
+   float (*dady)[4];
 };
 
 
@@ -53,14 +58,14 @@ struct lp_line_info {
  * Compute a0 for a constant-valued coefficient (GL_FLAT shading).
  */
 static void constant_coef( struct lp_setup_context *setup,
-                           struct lp_rast_triangle *tri,
+                           struct lp_line_info *info,
                            unsigned slot,
                            const float value,
                            unsigned i )
 {
-   tri->inputs.a0[slot][i] = value;
-   tri->inputs.dadx[slot][i] = 0.0f;
-   tri->inputs.dady[slot][i] = 0.0f;
+   info->a0[slot][i] = value;
+   info->dadx[slot][i] = 0.0f;
+   info->dady[slot][i] = 0.0f;
 }
 
 
@@ -69,7 +74,6 @@ static void constant_coef( struct lp_setup_context *setup,
  * for a triangle.
  */
 static void linear_coef( struct lp_setup_context *setup,
-                         struct lp_rast_triangle *tri,
                          struct lp_line_info *info,
                          unsigned slot,
                          unsigned vert_attr,
@@ -82,10 +86,10 @@ static void linear_coef( struct lp_setup_context *setup,
    float dadx = da21 * info->dx * info->oneoverarea;
    float dady = da21 * info->dy * info->oneoverarea;
 
-   tri->inputs.dadx[slot][i] = dadx;
-   tri->inputs.dady[slot][i] = dady;  
+   info->dadx[slot][i] = dadx;
+   info->dady[slot][i] = dady;  
    
-   tri->inputs.a0[slot][i] = (a1 -
+   info->a0[slot][i] = (a1 -
                               (dadx * (info->v1[0][0] - setup->pixel_offset) +
                                dady * (info->v1[0][1] - setup->pixel_offset)));
 }
@@ -100,7 +104,6 @@ static void linear_coef( struct lp_setup_context *setup,
  * divide the interpolated value by the interpolated W at that fragment.
  */
 static void perspective_coef( struct lp_setup_context *setup,
-                              struct lp_rast_triangle *tri,
                               struct lp_line_info *info,
                               unsigned slot,
                               unsigned vert_attr,
@@ -115,43 +118,42 @@ static void perspective_coef( struct lp_setup_context *setup,
    float dadx = da21 * info->dx * info->oneoverarea;
    float dady = da21 * info->dy * info->oneoverarea;
 
-   tri->inputs.dadx[slot][i] = dadx;
-   tri->inputs.dady[slot][i] = dady;
+   info->dadx[slot][i] = dadx;
+   info->dady[slot][i] = dady;
    
-   tri->inputs.a0[slot][i] = (a1 -
-                              (dadx * (info->v1[0][0] - setup->pixel_offset) +
-                               dady * (info->v1[0][1] - setup->pixel_offset)));
+   info->a0[slot][i] = (a1 -
+                        (dadx * (info->v1[0][0] - setup->pixel_offset) +
+                         dady * (info->v1[0][1] - setup->pixel_offset)));
 }
 
 static void
 setup_fragcoord_coef( struct lp_setup_context *setup,
-                      struct lp_rast_triangle *tri,
                       struct lp_line_info *info,
                       unsigned slot,
                       unsigned usage_mask)
 {
    /*X*/
    if (usage_mask & TGSI_WRITEMASK_X) {
-      tri->inputs.a0[slot][0] = 0.0;
-      tri->inputs.dadx[slot][0] = 1.0;
-      tri->inputs.dady[slot][0] = 0.0;
+      info->a0[slot][0] = 0.0;
+      info->dadx[slot][0] = 1.0;
+      info->dady[slot][0] = 0.0;
    }
 
    /*Y*/
    if (usage_mask & TGSI_WRITEMASK_Y) {
-      tri->inputs.a0[slot][1] = 0.0;
-      tri->inputs.dadx[slot][1] = 0.0;
-      tri->inputs.dady[slot][1] = 1.0;
+      info->a0[slot][1] = 0.0;
+      info->dadx[slot][1] = 0.0;
+      info->dady[slot][1] = 1.0;
    }
 
    /*Z*/
    if (usage_mask & TGSI_WRITEMASK_Z) {
-      linear_coef(setup, tri, info, slot, 0, 2);
+      linear_coef(setup, info, slot, 0, 2);
    }
 
    /*W*/
    if (usage_mask & TGSI_WRITEMASK_W) {
-      linear_coef(setup, tri, info, slot, 0, 3);
+      linear_coef(setup, info, slot, 0, 3);
    }
 }
 
@@ -159,43 +161,43 @@ setup_fragcoord_coef( struct lp_setup_context *setup,
  * Compute the tri->coef[] array dadx, dady, a0 values.
  */
 static void setup_line_coefficients( struct lp_setup_context *setup,
-                                     struct lp_rast_triangle *tri,
                                      struct lp_line_info *info)
 {
+   const struct lp_setup_variant_key *key = &setup->setup.variant->key;
    unsigned fragcoord_usage_mask = TGSI_WRITEMASK_XYZ;
    unsigned slot;
 
    /* setup interpolation for all the remaining attributes:
     */
-   for (slot = 0; slot < setup->fs.nr_inputs; slot++) {
-      unsigned vert_attr = setup->fs.input[slot].src_index;
-      unsigned usage_mask = setup->fs.input[slot].usage_mask;
+   for (slot = 0; slot < key->num_inputs; slot++) {
+      unsigned vert_attr = key->inputs[slot].src_index;
+      unsigned usage_mask = key->inputs[slot].usage_mask;
       unsigned i;
            
-      switch (setup->fs.input[slot].interp) {
+      switch (key->inputs[slot].interp) {
       case LP_INTERP_CONSTANT:
-         if (setup->flatshade_first) {
+         if (key->flatshade_first) {
             for (i = 0; i < NUM_CHANNELS; i++)
                if (usage_mask & (1 << i))
-                  constant_coef(setup, tri, slot+1, info->v1[vert_attr][i], i);
+                  constant_coef(setup, info, slot+1, info->v1[vert_attr][i], i);
          }
          else {
             for (i = 0; i < NUM_CHANNELS; i++)
                if (usage_mask & (1 << i))
-                  constant_coef(setup, tri, slot+1, info->v2[vert_attr][i], i);
+                  constant_coef(setup, info, slot+1, info->v2[vert_attr][i], i);
          }
          break;
 
       case LP_INTERP_LINEAR:
          for (i = 0; i < NUM_CHANNELS; i++)
             if (usage_mask & (1 << i))
-               linear_coef(setup, tri, info, slot+1, vert_attr, i);
+               linear_coef(setup, info, slot+1, vert_attr, i);
          break;
 
       case LP_INTERP_PERSPECTIVE:
          for (i = 0; i < NUM_CHANNELS; i++)
             if (usage_mask & (1 << i))
-               perspective_coef(setup, tri, info, slot+1, vert_attr, i);
+               perspective_coef(setup, info, slot+1, vert_attr, i);
          fragcoord_usage_mask |= TGSI_WRITEMASK_W;
          break;
 
@@ -211,7 +213,7 @@ static void setup_line_coefficients( struct lp_setup_context *setup,
       case LP_INTERP_FACING:
          for (i = 0; i < NUM_CHANNELS; i++)
             if (usage_mask & (1 << i))
-               constant_coef(setup, tri, slot+1, 1.0, i);
+               constant_coef(setup, info, slot+1, 1.0, i);
          break;
 
       default:
@@ -221,7 +223,7 @@ static void setup_line_coefficients( struct lp_setup_context *setup,
 
    /* The internal position input is in slot zero:
     */
-   setup_fragcoord_coef(setup, tri, info, 0,
+   setup_fragcoord_coef(setup, info, 0,
                         fragcoord_usage_mask);
 }
 
@@ -241,14 +243,15 @@ print_line(struct lp_setup_context *setup,
            const float (*v1)[4],
            const float (*v2)[4])
 {
+   const struct lp_setup_variant_key *key = &setup->setup.variant->key;
    uint i;
 
    debug_printf("llvmpipe line\n");
-   for (i = 0; i < 1 + setup->fs.nr_inputs; i++) {
+   for (i = 0; i < 1 + key->num_inputs; i++) {
       debug_printf("  v1[%d]:  %f %f %f %f\n", i,
                    v1[i][0], v1[i][1], v1[i][2], v1[i][3]);
    }
-   for (i = 0; i < 1 + setup->fs.nr_inputs; i++) {
+   for (i = 0; i < 1 + key->num_inputs; i++) {
       debug_printf("  v2[%d]:  %f %f %f %f\n", i,
                    v2[i][0], v2[i][1], v2[i][2], v2[i][3]);
    }
@@ -275,7 +278,9 @@ try_setup_line( struct lp_setup_context *setup,
                const float (*v2)[4])
 {
    struct lp_scene *scene = setup->scene;
+   const struct lp_setup_variant_key *key = &setup->setup.variant->key;
    struct lp_rast_triangle *line;
+   struct lp_rast_plane *plane;
    struct lp_line_info info;
    float width = MAX2(1.0, setup->line_width);
    struct u_rect bbox;
@@ -475,7 +480,7 @@ try_setup_line( struct lp_setup_context *setup,
       else {
          /* do intersection test */
          float xintersect = fracf(v2[0][0]) + y2diff * dxdy;
-         draw_end = (xintersect < 1.0 && xintersect > 0.0);
+         draw_end = (xintersect < 1.0 && xintersect >= 0.0);
       }
 
       /* Are we already drawing start/end?
@@ -513,7 +518,7 @@ try_setup_line( struct lp_setup_context *setup,
             x_offset_end = y_offset_end * dxdy;
          }
       }
- 
+
       /* x/y positions in fixed point */
       x[0] = subpixel_snap(v1[0][0] + x_offset     - setup->pixel_offset) - fixed_width/2;
       x[1] = subpixel_snap(v2[0][0] + x_offset_end - setup->pixel_offset) - fixed_width/2;
@@ -567,7 +572,7 @@ try_setup_line( struct lp_setup_context *setup,
    u_rect_find_intersection(&setup->draw_region, &bbox);
 
    line = lp_setup_alloc_triangle(scene,
-                                  setup->fs.nr_inputs,
+                                  key->num_inputs,
                                   nr_planes,
                                   &tri_bytes);
    if (!line)
@@ -581,33 +586,35 @@ try_setup_line( struct lp_setup_context *setup,
 #endif
 
    /* calculate the deltas */
-   line->plane[0].dcdy = x[0] - x[1];
-   line->plane[1].dcdy = x[1] - x[2];
-   line->plane[2].dcdy = x[2] - x[3];
-   line->plane[3].dcdy = x[3] - x[0];
+   plane = GET_PLANES(line);
+   plane[0].dcdy = x[0] - x[1];
+   plane[1].dcdy = x[1] - x[2];
+   plane[2].dcdy = x[2] - x[3];
+   plane[3].dcdy = x[3] - x[0];
 
-   line->plane[0].dcdx = y[0] - y[1];
-   line->plane[1].dcdx = y[1] - y[2];
-   line->plane[2].dcdx = y[2] - y[3];
-   line->plane[3].dcdx = y[3] - y[0];
+   plane[0].dcdx = y[0] - y[1];
+   plane[1].dcdx = y[1] - y[2];
+   plane[2].dcdx = y[2] - y[3];
+   plane[3].dcdx = y[3] - y[0];
 
 
    /* Setup parameter interpolants:
     */
-   setup_line_coefficients( setup, line, &info); 
+   info.a0 = GET_A0(&line->inputs);
+   info.dadx = GET_DADX(&line->inputs);
+   info.dady = GET_DADY(&line->inputs);
+   setup_line_coefficients(setup, &info); 
 
-   line->inputs.facing = 1.0F;
-   line->inputs.state = setup->fs.stored;
+   line->inputs.frontfacing = TRUE;
    line->inputs.disable = FALSE;
    line->inputs.opaque = FALSE;
 
    for (i = 0; i < 4; i++) {
-      struct lp_rast_plane *plane = &line->plane[i];
 
       /* half-edge constants, will be interated over the whole render
        * target.
        */
-      plane->c = plane->dcdx * x[i] - plane->dcdy * y[i];
+      plane[i].c = plane[i].dcdx * x[i] - plane[i].dcdy * y[i];
 
       
       /* correct for top-left vs. bottom-left fill convention.  
@@ -623,38 +630,34 @@ try_setup_line( struct lp_setup_context *setup,
        * to its usual method, in which case it will probably want
        * to use the opposite, top-left convention.
        */         
-      if (plane->dcdx < 0) {
+      if (plane[i].dcdx < 0) {
          /* both fill conventions want this - adjust for left edges */
-         plane->c++;            
+         plane[i].c++;            
       }
-      else if (plane->dcdx == 0) {
+      else if (plane[i].dcdx == 0) {
          if (setup->pixel_offset == 0) {
             /* correct for top-left fill convention:
              */
-            if (plane->dcdy > 0) plane->c++;
+            if (plane[i].dcdy > 0) plane[i].c++;
          }
          else {
             /* correct for bottom-left fill convention:
              */
-            if (plane->dcdy < 0) plane->c++;
+            if (plane[i].dcdy < 0) plane[i].c++;
          }
       }
 
-      plane->dcdx *= FIXED_ONE;
-      plane->dcdy *= FIXED_ONE;
+      plane[i].dcdx *= FIXED_ONE;
+      plane[i].dcdy *= FIXED_ONE;
 
       /* find trivial reject offsets for each edge for a single-pixel
        * sized block.  These will be scaled up at each recursive level to
        * match the active blocksize.  Scaling in this way works best if
        * the blocks are square.
        */
-      plane->eo = 0;
-      if (plane->dcdx < 0) plane->eo -= plane->dcdx;
-      if (plane->dcdy > 0) plane->eo += plane->dcdy;
-
-      /* Calculate trivial accept offsets from the above.
-       */
-      plane->ei = plane->dcdy - plane->dcdx - plane->eo;
+      plane[i].eo = 0;
+      if (plane[i].dcdx < 0) plane[i].eo -= plane[i].dcdx;
+      if (plane[i].dcdy > 0) plane[i].eo += plane[i].dcdy;
    }
 
 
@@ -677,29 +680,25 @@ try_setup_line( struct lp_setup_context *setup,
     * these planes elsewhere.
     */
    if (nr_planes == 8) {
-      line->plane[4].dcdx = -1;
-      line->plane[4].dcdy = 0;
-      line->plane[4].c = 1-bbox.x0;
-      line->plane[4].ei = 0;
-      line->plane[4].eo = 1;
-
-      line->plane[5].dcdx = 1;
-      line->plane[5].dcdy = 0;
-      line->plane[5].c = bbox.x1+1;
-      line->plane[5].ei = -1;
-      line->plane[5].eo = 0;
-
-      line->plane[6].dcdx = 0;
-      line->plane[6].dcdy = 1;
-      line->plane[6].c = 1-bbox.y0;
-      line->plane[6].ei = 0;
-      line->plane[6].eo = 1;
-
-      line->plane[7].dcdx = 0;
-      line->plane[7].dcdy = -1;
-      line->plane[7].c = bbox.y1+1;
-      line->plane[7].ei = -1;
-      line->plane[7].eo = 0;
+      plane[4].dcdx = -1;
+      plane[4].dcdy = 0;
+      plane[4].c = 1-bbox.x0;
+      plane[4].eo = 1;
+
+      plane[5].dcdx = 1;
+      plane[5].dcdy = 0;
+      plane[5].c = bbox.x1+1;
+      plane[5].eo = 0;
+
+      plane[6].dcdx = 0;
+      plane[6].dcdy = 1;
+      plane[6].c = 1-bbox.y0;
+      plane[6].eo = 1;
+
+      plane[7].dcdx = 0;
+      plane[7].dcdy = -1;
+      plane[7].c = bbox.y1+1;
+      plane[7].eo = 0;
    }
 
    return lp_setup_bin_triangle(setup, line, &bbox, nr_planes);
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_point.c b/src/gallium/drivers/llvmpipe/lp_setup_point.c
index 31d85f43c2f..146f1bd07ca 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_point.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_point.c
@@ -35,6 +35,7 @@
 #include "lp_perf.h"
 #include "lp_rast.h"
 #include "lp_state_fs.h"
+#include "lp_state_setup.h"
 #include "tgsi/tgsi_scan.h"
 
 #define NUM_CHANNELS 4
@@ -45,6 +46,10 @@ struct point_info {
    int dx01, dx12;
 
    const float (*v0)[4];
+
+   float (*a0)[4];
+   float (*dadx)[4];
+   float (*dady)[4];
 };   
 
 
@@ -53,20 +58,19 @@ struct point_info {
  */
 static void
 constant_coef(struct lp_setup_context *setup,
-              struct lp_rast_triangle *point,
+              struct point_info *info,
               unsigned slot,
               const float value,
               unsigned i)
 {
-   point->inputs.a0[slot][i] = value;
-   point->inputs.dadx[slot][i] = 0.0f;
-   point->inputs.dady[slot][i] = 0.0f;
+   info->a0[slot][i] = value;
+   info->dadx[slot][i] = 0.0f;
+   info->dady[slot][i] = 0.0f;
 }
 
 
 static void
 point_persp_coeff(struct lp_setup_context *setup,
-                  struct lp_rast_triangle *point,
                   const struct point_info *info,
                   unsigned slot,
                   unsigned i)
@@ -82,9 +86,9 @@ point_persp_coeff(struct lp_setup_context *setup,
 
    assert(i < 4);
 
-   point->inputs.a0[slot][i] = info->v0[slot][i]*w0;
-   point->inputs.dadx[slot][i] = 0.0f;
-   point->inputs.dady[slot][i] = 0.0f;
+   info->a0[slot][i] = info->v0[slot][i]*w0;
+   info->dadx[slot][i] = 0.0f;
+   info->dady[slot][i] = 0.0f;
 }
 
 
@@ -98,7 +102,6 @@ point_persp_coeff(struct lp_setup_context *setup,
  */
 static void
 texcoord_coef(struct lp_setup_context *setup,
-              struct lp_rast_triangle *point,
               const struct point_info *info,
               unsigned slot,
               unsigned i,
@@ -115,14 +118,14 @@ texcoord_coef(struct lp_setup_context *setup,
       float x0 = info->v0[0][0] - setup->pixel_offset;
       float y0 = info->v0[0][1] - setup->pixel_offset;
 
-      point->inputs.dadx[slot][0] = dadx;
-      point->inputs.dady[slot][0] = dady;
-      point->inputs.a0[slot][0] = 0.5 - (dadx * x0 + dady * y0);
+      info->dadx[slot][0] = dadx;
+      info->dady[slot][0] = dady;
+      info->a0[slot][0] = 0.5 - (dadx * x0 + dady * y0);
 
       if (perspective) {
-         point->inputs.dadx[slot][0] *= w0;
-         point->inputs.dady[slot][0] *= w0;
-         point->inputs.a0[slot][0] *= w0;
+         info->dadx[slot][0] *= w0;
+         info->dady[slot][0] *= w0;
+         info->a0[slot][0] *= w0;
       }
    }
    else if (i == 1) {
@@ -135,25 +138,25 @@ texcoord_coef(struct lp_setup_context *setup,
          dady = -dady;
       }
 
-      point->inputs.dadx[slot][1] = dadx;
-      point->inputs.dady[slot][1] = dady;
-      point->inputs.a0[slot][1] = 0.5 - (dadx * x0 + dady * y0);
+      info->dadx[slot][1] = dadx;
+      info->dady[slot][1] = dady;
+      info->a0[slot][1] = 0.5 - (dadx * x0 + dady * y0);
 
       if (perspective) {
-         point->inputs.dadx[slot][1] *= w0;
-         point->inputs.dady[slot][1] *= w0;
-         point->inputs.a0[slot][1] *= w0;
+         info->dadx[slot][1] *= w0;
+         info->dady[slot][1] *= w0;
+         info->a0[slot][1] *= w0;
       }
    }
    else if (i == 2) {
-      point->inputs.a0[slot][2] = 0.0f;
-      point->inputs.dadx[slot][2] = 0.0f;
-      point->inputs.dady[slot][2] = 0.0f;
+      info->a0[slot][2] = 0.0f;
+      info->dadx[slot][2] = 0.0f;
+      info->dady[slot][2] = 0.0f;
    }
    else {
-      point->inputs.a0[slot][3] = perspective ? w0 : 1.0f;
-      point->inputs.dadx[slot][3] = 0.0f;
-      point->inputs.dady[slot][3] = 0.0f;
+      info->a0[slot][3] = perspective ? w0 : 1.0f;
+      info->dadx[slot][3] = 0.0f;
+      info->dady[slot][3] = 0.0f;
    }
 }
 
@@ -166,33 +169,32 @@ texcoord_coef(struct lp_setup_context *setup,
  */
 static void
 setup_point_fragcoord_coef(struct lp_setup_context *setup,
-                           struct lp_rast_triangle *point,
-                           const struct point_info *info,
+                           struct point_info *info,
                            unsigned slot,
                            unsigned usage_mask)
 {
    /*X*/
    if (usage_mask & TGSI_WRITEMASK_X) {
-      point->inputs.a0[slot][0] = 0.0;
-      point->inputs.dadx[slot][0] = 1.0;
-      point->inputs.dady[slot][0] = 0.0;
+      info->a0[slot][0] = 0.0;
+      info->dadx[slot][0] = 1.0;
+      info->dady[slot][0] = 0.0;
    }
 
    /*Y*/
    if (usage_mask & TGSI_WRITEMASK_Y) {
-      point->inputs.a0[slot][1] = 0.0;
-      point->inputs.dadx[slot][1] = 0.0;
-      point->inputs.dady[slot][1] = 1.0;
+      info->a0[slot][1] = 0.0;
+      info->dadx[slot][1] = 0.0;
+      info->dady[slot][1] = 1.0;
    }
 
    /*Z*/
    if (usage_mask & TGSI_WRITEMASK_Z) {
-      constant_coef(setup, point, slot, info->v0[0][2], 2);
+      constant_coef(setup, info, slot, info->v0[0][2], 2);
    }
 
    /*W*/
    if (usage_mask & TGSI_WRITEMASK_W) {
-      constant_coef(setup, point, slot, info->v0[0][3], 3);
+      constant_coef(setup, info, slot, info->v0[0][3], 3);
    }
 }
 
@@ -202,20 +204,20 @@ setup_point_fragcoord_coef(struct lp_setup_context *setup,
  */
 static void   
 setup_point_coefficients( struct lp_setup_context *setup,
-                          struct lp_rast_triangle *point,
-                          const struct point_info *info)
+                          struct point_info *info)
 {
+   const struct lp_setup_variant_key *key = &setup->setup.variant->key;
    const struct lp_fragment_shader *shader = setup->fs.current.variant->shader;
    unsigned fragcoord_usage_mask = TGSI_WRITEMASK_XYZ;
    unsigned slot;
 
    /* setup interpolation for all the remaining attributes:
     */
-   for (slot = 0; slot < setup->fs.nr_inputs; slot++) {
-      enum lp_interp interp = setup->fs.input[slot].interp;
+   for (slot = 0; slot < key->num_inputs; slot++) {
+      unsigned vert_attr = key->inputs[slot].src_index;
+      unsigned usage_mask = key->inputs[slot].usage_mask;
+      enum lp_interp interp = key->inputs[slot].interp;
       boolean perspective = !!(interp == LP_INTERP_PERSPECTIVE);
-      unsigned vert_attr = setup->fs.input[slot].src_index;
-      unsigned usage_mask = setup->fs.input[slot].usage_mask;
       unsigned i;
 
       if (perspective & usage_mask) {
@@ -239,8 +241,8 @@ setup_point_coefficients( struct lp_setup_context *setup,
          /* check if the sprite coord flag is set for this attribute.
           * If so, set it up so it up so x and y vary from 0 to 1.
           */
-         if (shader->info.input_semantic_name[slot] == TGSI_SEMANTIC_GENERIC) {
-            unsigned semantic_index = shader->info.input_semantic_index[slot];
+         if (shader->info.base.input_semantic_name[slot] == TGSI_SEMANTIC_GENERIC) {
+            unsigned semantic_index = shader->info.base.input_semantic_index[slot];
             /* Note that sprite_coord enable is a bitfield of
              * PIPE_MAX_SHADER_OUTPUTS bits.
              */
@@ -248,7 +250,7 @@ setup_point_coefficients( struct lp_setup_context *setup,
                 (setup->sprite_coord_enable & (1 << semantic_index))) {
                for (i = 0; i < NUM_CHANNELS; i++) {
                   if (usage_mask & (1 << i)) {
-                     texcoord_coef(setup, point, info, slot + 1, i,
+                     texcoord_coef(setup, info, slot + 1, i,
                                    setup->sprite_coord_origin,
                                    perspective);
                   }
@@ -261,10 +263,10 @@ setup_point_coefficients( struct lp_setup_context *setup,
          for (i = 0; i < NUM_CHANNELS; i++) {
             if (usage_mask & (1 << i)) {
                if (perspective) {
-                  point_persp_coeff(setup, point, info, slot+1, i);
+                  point_persp_coeff(setup, info, slot+1, i);
                }
                else {
-                  constant_coef(setup, point, slot+1, info->v0[vert_attr][i], i);
+                  constant_coef(setup, info, slot+1, info->v0[vert_attr][i], i);
                }
             }
          }
@@ -273,7 +275,7 @@ setup_point_coefficients( struct lp_setup_context *setup,
       case LP_INTERP_FACING:
          for (i = 0; i < NUM_CHANNELS; i++)
             if (usage_mask & (1 << i))
-               constant_coef(setup, point, slot+1, 1.0, i);
+               constant_coef(setup, info, slot+1, 1.0, i);
          break;
 
       default:
@@ -284,7 +286,7 @@ setup_point_coefficients( struct lp_setup_context *setup,
 
    /* The internal position input is in slot zero:
     */
-   setup_point_fragcoord_coef(setup, point, info, 0,
+   setup_point_fragcoord_coef(setup, info, 0,
                               fragcoord_usage_mask);
 }
 
@@ -301,6 +303,7 @@ try_setup_point( struct lp_setup_context *setup,
                  const float (*v0)[4] )
 {
    /* x/y positions in fixed point */
+   const struct lp_setup_variant_key *key = &setup->setup.variant->key;
    const int sizeAttr = setup->psize;
    const float size
       = (setup->point_size_per_vertex && sizeAttr > 0) ? v0[sizeAttr][0]
@@ -352,7 +355,7 @@ try_setup_point( struct lp_setup_context *setup,
    u_rect_find_intersection(&setup->draw_region, &bbox);
 
    point = lp_setup_alloc_triangle(scene,
-                                   setup->fs.nr_inputs,
+                                   key->num_inputs,
                                    nr_planes,
                                    &bytes);
    if (!point)
@@ -368,40 +371,40 @@ try_setup_point( struct lp_setup_context *setup,
    info.dx12 = fixed_width;
    info.dy01 = fixed_width;
    info.dy12 = 0;
+   info.a0 = GET_A0(&point->inputs);
+   info.dadx = GET_DADX(&point->inputs);
+   info.dady = GET_DADY(&point->inputs);
    
    /* Setup parameter interpolants:
     */
-   setup_point_coefficients(setup, point, &info);
+   setup_point_coefficients(setup, &info);
 
-   point->inputs.facing = 1.0F;
-   point->inputs.state = setup->fs.stored;
+   point->inputs.frontfacing = TRUE;
    point->inputs.disable = FALSE;
    point->inputs.opaque = FALSE;
 
    {
-      point->plane[0].dcdx = -1;
-      point->plane[0].dcdy = 0;
-      point->plane[0].c = 1-bbox.x0;
-      point->plane[0].ei = 0;
-      point->plane[0].eo = 1;
-
-      point->plane[1].dcdx = 1;
-      point->plane[1].dcdy = 0;
-      point->plane[1].c = bbox.x1+1;
-      point->plane[1].ei = -1;
-      point->plane[1].eo = 0;
-
-      point->plane[2].dcdx = 0;
-      point->plane[2].dcdy = 1;
-      point->plane[2].c = 1-bbox.y0;
-      point->plane[2].ei = 0;
-      point->plane[2].eo = 1;
-
-      point->plane[3].dcdx = 0;
-      point->plane[3].dcdy = -1;
-      point->plane[3].c = bbox.y1+1;
-      point->plane[3].ei = -1;
-      point->plane[3].eo = 0;
+      struct lp_rast_plane *plane = GET_PLANES(point);
+
+      plane[0].dcdx = -1;
+      plane[0].dcdy = 0;
+      plane[0].c = 1-bbox.x0;
+      plane[0].eo = 1;
+
+      plane[1].dcdx = 1;
+      plane[1].dcdy = 0;
+      plane[1].c = bbox.x1+1;
+      plane[1].eo = 0;
+
+      plane[2].dcdx = 0;
+      plane[2].dcdy = 1;
+      plane[2].c = 1-bbox.y0;
+      plane[2].eo = 1;
+
+      plane[3].dcdx = 0;
+      plane[3].dcdy = -1;
+      plane[3].c = bbox.y1+1;
+      plane[3].eo = 0;
    }
 
    return lp_setup_bin_triangle(setup, point, &bbox, nr_planes);
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_tri.c b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
index 8fd034666c3..4ab0b72a574 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
@@ -32,15 +32,18 @@
 #include "util/u_math.h"
 #include "util/u_memory.h"
 #include "util/u_rect.h"
+#include "util/u_sse.h"
 #include "lp_perf.h"
 #include "lp_setup_context.h"
-#include "lp_setup_coef.h"
 #include "lp_rast.h"
 #include "lp_state_fs.h"
+#include "lp_state_setup.h"
 
 #define NUM_CHANNELS 4
 
-
+#if defined(PIPE_ARCH_SSE)
+#include <emmintrin.h>
+#endif
    
 static INLINE int
 subpixel_snap(float a)
@@ -65,7 +68,7 @@ fixed_to_float(int a)
  * immediately after it.
  * The memory is allocated from the per-scene pool, not per-tile.
  * \param tri_size  returns number of bytes allocated
- * \param nr_inputs  number of fragment shader inputs
+ * \param num_inputs  number of fragment shader inputs
  * \return pointer to triangle space
  */
 struct lp_rast_triangle *
@@ -75,22 +78,23 @@ lp_setup_alloc_triangle(struct lp_scene *scene,
                         unsigned *tri_size)
 {
    unsigned input_array_sz = NUM_CHANNELS * (nr_inputs + 1) * sizeof(float);
+   unsigned plane_sz = nr_planes * sizeof(struct lp_rast_plane);
    struct lp_rast_triangle *tri;
-   unsigned tri_bytes, bytes;
-   char *inputs;
 
-   tri_bytes = align(Offset(struct lp_rast_triangle, plane[nr_planes]), 16);
-   bytes = tri_bytes + (3 * input_array_sz);
+   *tri_size = (sizeof(struct lp_rast_triangle) +
+                3 * input_array_sz +
+                plane_sz);
 
-   tri = lp_scene_alloc_aligned( scene, bytes, 16 );
+   tri = lp_scene_alloc_aligned( scene, *tri_size, 16 );
+   if (tri == NULL)
+      return NULL;
 
-   if (tri) {
-      inputs = ((char *)tri) + tri_bytes;
-      tri->inputs.a0   = (float (*)[4]) inputs;
-      tri->inputs.dadx = (float (*)[4]) (inputs + input_array_sz);
-      tri->inputs.dady = (float (*)[4]) (inputs + 2 * input_array_sz);
+   tri->inputs.stride = input_array_sz;
 
-      *tri_size = bytes;
+   {
+      char *a = (char *)tri;
+      char *b = (char *)&GET_PLANES(tri)[nr_planes];
+      assert(b - a == *tri_size);
    }
 
    return tri;
@@ -101,25 +105,26 @@ lp_setup_print_vertex(struct lp_setup_context *setup,
                       const char *name,
                       const float (*v)[4])
 {
+   const struct lp_setup_variant_key *key = &setup->setup.variant->key;
    int i, j;
 
    debug_printf("   wpos (%s[0]) xyzw %f %f %f %f\n",
                 name,
                 v[0][0], v[0][1], v[0][2], v[0][3]);
 
-   for (i = 0; i < setup->fs.nr_inputs; i++) {
-      const float *in = v[setup->fs.input[i].src_index];
+   for (i = 0; i < key->num_inputs; i++) {
+      const float *in = v[key->inputs[i].src_index];
 
       debug_printf("  in[%d] (%s[%d]) %s%s%s%s ",
                    i, 
-                   name, setup->fs.input[i].src_index,
-                   (setup->fs.input[i].usage_mask & 0x1) ? "x" : " ",
-                   (setup->fs.input[i].usage_mask & 0x2) ? "y" : " ",
-                   (setup->fs.input[i].usage_mask & 0x4) ? "z" : " ",
-                   (setup->fs.input[i].usage_mask & 0x8) ? "w" : " ");
+                   name, key->inputs[i].src_index,
+                   (key->inputs[i].usage_mask & 0x1) ? "x" : " ",
+                   (key->inputs[i].usage_mask & 0x2) ? "y" : " ",
+                   (key->inputs[i].usage_mask & 0x4) ? "z" : " ",
+                   (key->inputs[i].usage_mask & 0x8) ? "w" : " ");
 
       for (j = 0; j < 4; j++)
-         if (setup->fs.input[i].usage_mask & (1<<j))
+         if (key->inputs[i].usage_mask & (1<<j))
             debug_printf("%.5f ", in[j]);
 
       debug_printf("\n");
@@ -200,14 +205,16 @@ lp_setup_whole_tile(struct lp_setup_context *setup,
       }
 
       LP_COUNT(nr_shade_opaque_64);
-      return lp_scene_bin_command( scene, tx, ty,
-                                   LP_RAST_OP_SHADE_TILE_OPAQUE,
-                                   lp_rast_arg_inputs(inputs) );
+      return lp_scene_bin_cmd_with_state( scene, tx, ty,
+                                          setup->fs.stored,
+                                          LP_RAST_OP_SHADE_TILE_OPAQUE,
+                                          lp_rast_arg_inputs(inputs) );
    } else {
       LP_COUNT(nr_shade_64);
-      return lp_scene_bin_command( scene, tx, ty,
-                                   LP_RAST_OP_SHADE_TILE,
-                                   lp_rast_arg_inputs(inputs) );
+      return lp_scene_bin_cmd_with_state( scene, tx, ty,
+                                          setup->fs.stored, 
+                                          LP_RAST_OP_SHADE_TILE,
+                                          lp_rast_arg_inputs(inputs) );
    }
 }
 
@@ -225,12 +232,13 @@ do_triangle_ccw(struct lp_setup_context *setup,
 		boolean frontfacing )
 {
    struct lp_scene *scene = setup->scene;
+   const struct lp_setup_variant_key *key = &setup->setup.variant->key;
    struct lp_rast_triangle *tri;
-   int x[3];
-   int y[3];
+   struct lp_rast_plane *plane;
+   int x[4];
+   int y[4];
    struct u_rect bbox;
    unsigned tri_bytes;
-   int i;
    int nr_planes = 3;
 
    if (0)
@@ -247,10 +255,12 @@ do_triangle_ccw(struct lp_setup_context *setup,
    x[0] = subpixel_snap(v0[0][0] - setup->pixel_offset);
    x[1] = subpixel_snap(v1[0][0] - setup->pixel_offset);
    x[2] = subpixel_snap(v2[0][0] - setup->pixel_offset);
+   x[3] = 0;
    y[0] = subpixel_snap(v0[0][1] - setup->pixel_offset);
    y[1] = subpixel_snap(v1[0][1] - setup->pixel_offset);
    y[2] = subpixel_snap(v2[0][1] - setup->pixel_offset);
-
+   y[3] = 0;
+   
 
    /* Bounding rectangle (in pixels) */
    {
@@ -288,13 +298,13 @@ do_triangle_ccw(struct lp_setup_context *setup,
    u_rect_find_intersection(&setup->draw_region, &bbox);
 
    tri = lp_setup_alloc_triangle(scene,
-                                 setup->fs.nr_inputs,
+                                 key->num_inputs,
                                  nr_planes,
                                  &tri_bytes);
    if (!tri)
       return FALSE;
 
-#ifdef DEBUG
+#if 0
    tri->v[0][0] = v0[0][0];
    tri->v[1][0] = v1[0][0];
    tri->v[2][0] = v2[0][0];
@@ -303,79 +313,172 @@ do_triangle_ccw(struct lp_setup_context *setup,
    tri->v[2][1] = v2[0][1];
 #endif
 
-   tri->plane[0].dcdy = x[0] - x[1];
-   tri->plane[1].dcdy = x[1] - x[2];
-   tri->plane[2].dcdy = x[2] - x[0];
-
-   tri->plane[0].dcdx = y[0] - y[1];
-   tri->plane[1].dcdx = y[1] - y[2];
-   tri->plane[2].dcdx = y[2] - y[0];
-
    LP_COUNT(nr_tris);
 
    /* Setup parameter interpolants:
     */
-   lp_setup_tri_coef( setup, &tri->inputs, v0, v1, v2, frontfacing );
-
-   tri->inputs.facing = frontfacing ? 1.0F : -1.0F;
+   setup->setup.variant->jit_function( v0,
+				       v1,
+				       v2,
+				       frontfacing,
+				       GET_A0(&tri->inputs),
+				       GET_DADX(&tri->inputs),
+				       GET_DADY(&tri->inputs) );
+
+   tri->inputs.frontfacing = frontfacing;
    tri->inputs.disable = FALSE;
    tri->inputs.opaque = setup->fs.current.variant->opaque;
-   tri->inputs.state = setup->fs.stored;
 
-  
-   for (i = 0; i < 3; i++) {
-      struct lp_rast_plane *plane = &tri->plane[i];
+   if (0)
+      lp_dump_setup_coef(&setup->setup.variant->key,
+			 (const float (*)[4])GET_A0(&tri->inputs),
+			 (const float (*)[4])GET_DADX(&tri->inputs),
+			 (const float (*)[4])GET_DADY(&tri->inputs));
+
+   plane = GET_PLANES(tri);
+
+#if defined(PIPE_ARCH_SSE)
+   {
+      __m128i vertx, verty;
+      __m128i shufx, shufy;
+      __m128i dcdx, dcdy, c;
+      __m128i unused;
+      __m128i dcdx_neg_mask;
+      __m128i dcdy_neg_mask;
+      __m128i dcdx_zero_mask;
+      __m128i top_left_flag;
+      __m128i c_inc_mask, c_inc;
+      __m128i eo, p0, p1, p2;
+      __m128i zero = _mm_setzero_si128();
 
-      /* half-edge constants, will be interated over the whole render
-       * target.
+      vertx = _mm_loadu_si128((__m128i *)x); /* vertex x coords */
+      verty = _mm_loadu_si128((__m128i *)y); /* vertex y coords */
+
+      shufx = _mm_shuffle_epi32(vertx, _MM_SHUFFLE(3,0,2,1));
+      shufy = _mm_shuffle_epi32(verty, _MM_SHUFFLE(3,0,2,1));
+
+      dcdx = _mm_sub_epi32(verty, shufy);
+      dcdy = _mm_sub_epi32(vertx, shufx);
+
+      dcdx_neg_mask = _mm_srai_epi32(dcdx, 31);
+      dcdx_zero_mask = _mm_cmpeq_epi32(dcdx, zero);
+      dcdy_neg_mask = _mm_srai_epi32(dcdy, 31);
+
+      top_left_flag = _mm_set1_epi32((setup->pixel_offset == 0) ? ~0 : 0);
+
+      c_inc_mask = _mm_or_si128(dcdx_neg_mask,
+                                _mm_and_si128(dcdx_zero_mask,
+                                              _mm_xor_si128(dcdy_neg_mask,
+                                                            top_left_flag)));
+
+      c_inc = _mm_srli_epi32(c_inc_mask, 31);
+
+      c = _mm_sub_epi32(mm_mullo_epi32(dcdx, vertx),
+                        mm_mullo_epi32(dcdy, verty));
+
+      c = _mm_add_epi32(c, c_inc);
+
+      /* Scale up to match c:
        */
-      plane->c = plane->dcdx * x[i] - plane->dcdy * y[i];
-
-      /* correct for top-left vs. bottom-left fill convention.  
-       *
-       * note that we're overloading gl_rasterization_rules to mean
-       * both (0.5,0.5) pixel centers *and* bottom-left filling
-       * convention.
-       *
-       * GL actually has a top-left filling convention, but GL's
-       * notion of "top" differs from gallium's...
-       *
-       * Also, sometimes (in FBO cases) GL will render upside down
-       * to its usual method, in which case it will probably want
-       * to use the opposite, top-left convention.
-       */         
-      if (plane->dcdx < 0) {
-         /* both fill conventions want this - adjust for left edges */
-         plane->c++;            
-      }
-      else if (plane->dcdx == 0) {
-         if (setup->pixel_offset == 0) {
-            /* correct for top-left fill convention:
-             */
-            if (plane->dcdy > 0) plane->c++;
+      dcdx = _mm_slli_epi32(dcdx, FIXED_ORDER);
+      dcdy = _mm_slli_epi32(dcdy, FIXED_ORDER);
+
+      /* Calculate trivial reject values:
+       */
+      eo = _mm_sub_epi32(_mm_andnot_si128(dcdy_neg_mask, dcdy),
+                         _mm_and_si128(dcdx_neg_mask, dcdx));
+
+      /* ei = _mm_sub_epi32(_mm_sub_epi32(dcdy, dcdx), eo); */
+
+      /* Pointless transpose which gets undone immediately in
+       * rasterization:
+       */
+      transpose4_epi32(&c, &dcdx, &dcdy, &eo,
+                       &p0, &p1, &p2, &unused);
+
+      _mm_store_si128((__m128i *)&plane[0], p0);
+      _mm_store_si128((__m128i *)&plane[1], p1);
+      _mm_store_si128((__m128i *)&plane[2], p2);
+   }
+#else
+   {
+      int i;
+      plane[0].dcdy = x[0] - x[1];
+      plane[1].dcdy = x[1] - x[2];
+      plane[2].dcdy = x[2] - x[0];
+      plane[0].dcdx = y[0] - y[1];
+      plane[1].dcdx = y[1] - y[2];
+      plane[2].dcdx = y[2] - y[0];
+  
+      for (i = 0; i < 3; i++) {
+         /* half-edge constants, will be interated over the whole render
+          * target.
+          */
+         plane[i].c = plane[i].dcdx * x[i] - plane[i].dcdy * y[i];
+
+         /* correct for top-left vs. bottom-left fill convention.  
+          *
+          * note that we're overloading gl_rasterization_rules to mean
+          * both (0.5,0.5) pixel centers *and* bottom-left filling
+          * convention.
+          *
+          * GL actually has a top-left filling convention, but GL's
+          * notion of "top" differs from gallium's...
+          *
+          * Also, sometimes (in FBO cases) GL will render upside down
+          * to its usual method, in which case it will probably want
+          * to use the opposite, top-left convention.
+          */         
+         if (plane[i].dcdx < 0) {
+            /* both fill conventions want this - adjust for left edges */
+            plane[i].c++;            
          }
-         else {
-            /* correct for bottom-left fill convention:
-             */
-            if (plane->dcdy < 0) plane->c++;
+         else if (plane[i].dcdx == 0) {
+            if (setup->pixel_offset == 0) {
+               /* correct for top-left fill convention:
+                */
+               if (plane[i].dcdy > 0) plane[i].c++;
+            }
+            else {
+               /* correct for bottom-left fill convention:
+                */
+               if (plane[i].dcdy < 0) plane[i].c++;
+            }
          }
-      }
 
-      plane->dcdx *= FIXED_ONE;
-      plane->dcdy *= FIXED_ONE;
+         plane[i].dcdx *= FIXED_ONE;
+         plane[i].dcdy *= FIXED_ONE;
 
-      /* find trivial reject offsets for each edge for a single-pixel
-       * sized block.  These will be scaled up at each recursive level to
-       * match the active blocksize.  Scaling in this way works best if
-       * the blocks are square.
-       */
-      plane->eo = 0;
-      if (plane->dcdx < 0) plane->eo -= plane->dcdx;
-      if (plane->dcdy > 0) plane->eo += plane->dcdy;
+         /* find trivial reject offsets for each edge for a single-pixel
+          * sized block.  These will be scaled up at each recursive level to
+          * match the active blocksize.  Scaling in this way works best if
+          * the blocks are square.
+          */
+         plane[i].eo = 0;
+         if (plane[i].dcdx < 0) plane[i].eo -= plane[i].dcdx;
+         if (plane[i].dcdy > 0) plane[i].eo += plane[i].dcdy;
+      }
+   }
+#endif
 
-      /* Calculate trivial accept offsets from the above.
-       */
-      plane->ei = plane->dcdy - plane->dcdx - plane->eo;
+   if (0) {
+      debug_printf("p0: %08x/%08x/%08x/%08x\n",
+                   plane[0].c,
+                   plane[0].dcdx,
+                   plane[0].dcdy,
+                   plane[0].eo);
+      
+      debug_printf("p1: %08x/%08x/%08x/%08x\n",
+                   plane[1].c,
+                   plane[1].dcdx,
+                   plane[1].dcdy,
+                   plane[1].eo);
+      
+      debug_printf("p0: %08x/%08x/%08x/%08x\n",
+                   plane[2].c,
+                   plane[2].dcdx,
+                   plane[2].dcdy,
+                   plane[2].eo);
    }
 
 
@@ -398,29 +501,25 @@ do_triangle_ccw(struct lp_setup_context *setup,
     * these planes elsewhere.
     */
    if (nr_planes == 7) {
-      tri->plane[3].dcdx = -1;
-      tri->plane[3].dcdy = 0;
-      tri->plane[3].c = 1-bbox.x0;
-      tri->plane[3].ei = 0;
-      tri->plane[3].eo = 1;
-
-      tri->plane[4].dcdx = 1;
-      tri->plane[4].dcdy = 0;
-      tri->plane[4].c = bbox.x1+1;
-      tri->plane[4].ei = -1;
-      tri->plane[4].eo = 0;
-
-      tri->plane[5].dcdx = 0;
-      tri->plane[5].dcdy = 1;
-      tri->plane[5].c = 1-bbox.y0;
-      tri->plane[5].ei = 0;
-      tri->plane[5].eo = 1;
-
-      tri->plane[6].dcdx = 0;
-      tri->plane[6].dcdy = -1;
-      tri->plane[6].c = bbox.y1+1;
-      tri->plane[6].ei = -1;
-      tri->plane[6].eo = 0;
+      plane[3].dcdx = -1;
+      plane[3].dcdy = 0;
+      plane[3].c = 1-bbox.x0;
+      plane[3].eo = 1;
+
+      plane[4].dcdx = 1;
+      plane[4].dcdy = 0;
+      plane[4].c = bbox.x1+1;
+      plane[4].eo = 0;
+
+      plane[5].dcdx = 0;
+      plane[5].dcdy = 1;
+      plane[5].c = 1-bbox.y0;
+      plane[5].eo = 1;
+
+      plane[6].dcdx = 0;
+      plane[6].dcdy = -1;
+      plane[6].c = bbox.y1+1;
+      plane[6].eo = 0;
    }
 
    return lp_setup_bin_triangle( setup, tri, &bbox, nr_planes );
@@ -491,37 +590,40 @@ lp_setup_bin_triangle( struct lp_setup_context *setup,
          {
             /* Triangle is contained in a single 4x4 stamp:
              */
-
-            return lp_scene_bin_command( scene, ix0, iy0,
-                                         LP_RAST_OP_TRIANGLE_3_4,
-                                         lp_rast_arg_triangle(tri, mask) );
+            return lp_scene_bin_cmd_with_state( scene, ix0, iy0,
+                                                setup->fs.stored,
+                                                LP_RAST_OP_TRIANGLE_3_4,
+                                                lp_rast_arg_triangle(tri, mask) );
          }
 
          if (sz < 16)
          {
             /* Triangle is contained in a single 16x16 block:
              */
-            return lp_scene_bin_command( scene, ix0, iy0,
-                                         LP_RAST_OP_TRIANGLE_3_16,
-                                         lp_rast_arg_triangle(tri, mask) );
+            return lp_scene_bin_cmd_with_state( scene, ix0, iy0,
+                                                setup->fs.stored,
+                                                LP_RAST_OP_TRIANGLE_3_16,
+                                                lp_rast_arg_triangle(tri, mask) );
          }
       }
       else if (nr_planes == 4 && sz < 16) 
       {
-         return lp_scene_bin_command( scene, ix0, iy0,
-                                      LP_RAST_OP_TRIANGLE_4_16,
-                                      lp_rast_arg_triangle(tri, mask) );
+         return lp_scene_bin_cmd_with_state(scene, ix0, iy0,
+                                            setup->fs.stored,
+                                            LP_RAST_OP_TRIANGLE_4_16,
+                                            lp_rast_arg_triangle(tri, mask) );
       }
 
 
       /* Triangle is contained in a single tile:
        */
-      return lp_scene_bin_command( scene, ix0, iy0,
-                                   lp_rast_tri_tab[nr_planes], 
-                                   lp_rast_arg_triangle(tri, (1<<nr_planes)-1) );
+      return lp_scene_bin_cmd_with_state( scene, ix0, iy0, setup->fs.stored,
+                                          lp_rast_tri_tab[nr_planes], 
+                                          lp_rast_arg_triangle(tri, (1<<nr_planes)-1) );
    }
    else
    {
+      struct lp_rast_plane *plane = GET_PLANES(tri);
       int c[MAX_PLANES];
       int ei[MAX_PLANES];
       int eo[MAX_PLANES];
@@ -535,14 +637,17 @@ lp_setup_bin_triangle( struct lp_setup_context *setup,
       int iy1 = bbox->y1 / TILE_SIZE;
       
       for (i = 0; i < nr_planes; i++) {
-         c[i] = (tri->plane[i].c + 
-                 tri->plane[i].dcdy * iy0 * TILE_SIZE - 
-                 tri->plane[i].dcdx * ix0 * TILE_SIZE);
-
-         ei[i] = tri->plane[i].ei << TILE_ORDER;
-         eo[i] = tri->plane[i].eo << TILE_ORDER;
-         xstep[i] = -(tri->plane[i].dcdx << TILE_ORDER);
-         ystep[i] = tri->plane[i].dcdy << TILE_ORDER;
+         c[i] = (plane[i].c + 
+                 plane[i].dcdy * iy0 * TILE_SIZE - 
+                 plane[i].dcdx * ix0 * TILE_SIZE);
+
+         ei[i] = (plane[i].dcdy - 
+                  plane[i].dcdx - 
+                  plane[i].eo) << TILE_ORDER;
+
+         eo[i] = plane[i].eo << TILE_ORDER;
+         xstep[i] = -(plane[i].dcdx << TILE_ORDER);
+         ystep[i] = plane[i].dcdy << TILE_ORDER;
       }
 
 
@@ -584,9 +689,11 @@ lp_setup_bin_triangle( struct lp_setup_context *setup,
                 */
                int count = util_bitcount(partial);
                in = TRUE;
-               if (!lp_scene_bin_command( scene, x, y,
-                                          lp_rast_tri_tab[count], 
-                                          lp_rast_arg_triangle(tri, partial) ))
+               
+               if (!lp_scene_bin_cmd_with_state( scene, x, y,
+                                                 setup->fs.stored,
+                                                 lp_rast_tri_tab[count], 
+                                                 lp_rast_arg_triangle(tri, partial) ))
                   goto fail;
 
                LP_COUNT(nr_partially_covered_64);
diff --git a/src/gallium/drivers/llvmpipe/lp_state.h b/src/gallium/drivers/llvmpipe/lp_state.h
index 86313e1c484..7893e9cdc0c 100644
--- a/src/gallium/drivers/llvmpipe/lp_state.h
+++ b/src/gallium/drivers/llvmpipe/lp_state.h
@@ -97,6 +97,9 @@ llvmpipe_set_framebuffer_state(struct pipe_context *,
 void
 llvmpipe_update_fs(struct llvmpipe_context *lp);
 
+void 
+llvmpipe_update_setup(struct llvmpipe_context *lp);
+
 void
 llvmpipe_update_derived(struct llvmpipe_context *llvmpipe);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_state_derived.c b/src/gallium/drivers/llvmpipe/lp_state_derived.c
index bb059d04599..0f5f7369e04 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_derived.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_derived.c
@@ -50,12 +50,13 @@ compute_vertex_info(struct llvmpipe_context *llvmpipe)
 {
    const struct lp_fragment_shader *lpfs = llvmpipe->fs;
    struct vertex_info *vinfo = &llvmpipe->vertex_info;
-   struct lp_shader_input *inputs = llvmpipe->inputs;
    unsigned vs_index;
    uint i;
 
    /*
-    * Match FS inputs against VS outputs, emitting the necessary attributes.
+    * Match FS inputs against VS outputs, emitting the necessary
+    * attributes.  Could cache these structs and look them up with a
+    * combination of fragment shader, vertex shader ids.
     */
 
    vinfo->num_attribs = 0;
@@ -66,72 +67,18 @@ compute_vertex_info(struct llvmpipe_context *llvmpipe)
 
    draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE, vs_index);
 
-   for (i = 0; i < lpfs->info.num_inputs; i++) {
+   for (i = 0; i < lpfs->info.base.num_inputs; i++) {
       /*
        * Search for each input in current vs output:
        */
 
       vs_index = draw_find_shader_output(llvmpipe->draw,
-                                         lpfs->info.input_semantic_name[i],
-                                         lpfs->info.input_semantic_index[i]);
-      if (vs_index < 0) {
-         /*
-          * This can happen with sprite coordinates - the vertex
-          * shader doesn't need to provide an output as we generate
-          * them internally.  However, lets keep pretending that there
-          * is something there to not confuse other code.
-          */
-         vs_index = 0;
-      }
-
-      /* This can be pre-computed, except for flatshade:
-       */
-      inputs[i].usage_mask = lpfs->info.input_usage_mask[i];
-
-      switch (lpfs->info.input_interpolate[i]) {
-      case TGSI_INTERPOLATE_CONSTANT:
-         inputs[i].interp = LP_INTERP_CONSTANT;
-         break;
-      case TGSI_INTERPOLATE_LINEAR:
-         inputs[i].interp = LP_INTERP_LINEAR;
-         break;
-      case TGSI_INTERPOLATE_PERSPECTIVE:
-         inputs[i].interp = LP_INTERP_PERSPECTIVE;
-         break;
-      default:
-         assert(0);
-         break;
-      }
-
-      switch (lpfs->info.input_semantic_name[i]) {
-      case TGSI_SEMANTIC_FACE:
-         inputs[i].interp = LP_INTERP_FACING;
-         break;
-      case TGSI_SEMANTIC_POSITION:
-         /* Position was already emitted above
-          */
-         inputs[i].interp = LP_INTERP_POSITION;
-         inputs[i].src_index = 0;
-         continue;
-      case TGSI_SEMANTIC_COLOR:
-         /* Colors are linearly inputs[i].interpolated in the fragment shader
-          * even when flatshading is active.  This just tells the
-          * setup module to use coefficients with ddx==0 and
-          * ddy==0.
-          */
-         if (llvmpipe->rasterizer->flatshade)
-            inputs[i].interp = LP_INTERP_CONSTANT;
-         break;
-
-      default:
-         break;
-      }
+                                         lpfs->info.base.input_semantic_name[i],
+                                         lpfs->info.base.input_semantic_index[i]);
 
       /*
        * Emit the requested fs attribute for all but position.
        */
-
-      inputs[i].src_index = vinfo->num_attribs;
       draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE, vs_index);
    }
 
@@ -145,15 +92,8 @@ compute_vertex_info(struct llvmpipe_context *llvmpipe)
       draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_CONSTANT, vs_index);
    }
 
-   llvmpipe->num_inputs = lpfs->info.num_inputs;
-
    draw_compute_vertex_size(vinfo);
-
    lp_setup_set_vertex_info(llvmpipe->setup, vinfo);
-
-   lp_setup_set_fs_inputs(llvmpipe->setup,
-                          inputs,
-                          lpfs->info.num_inputs);
 }
 
 
@@ -190,6 +130,10 @@ void llvmpipe_update_derived( struct llvmpipe_context *llvmpipe )
                           LP_NEW_QUERY))
       llvmpipe_update_fs( llvmpipe );
 
+   if (llvmpipe->dirty & (LP_NEW_FS |
+			  LP_NEW_RASTERIZER))
+      llvmpipe_update_setup( llvmpipe );
+
    if (llvmpipe->dirty & LP_NEW_BLEND_COLOR)
       lp_setup_set_blend_color(llvmpipe->setup,
                                &llvmpipe->blend_color);
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.c b/src/gallium/drivers/llvmpipe/lp_state_fs.c
index 3ce8be5a0a9..9fbedac165f 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_fs.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs.c
@@ -99,74 +99,12 @@
 
 
 #include <llvm-c/Analysis.h>
+#include <llvm-c/BitWriter.h>
 
 
 static unsigned fs_no = 0;
 
 
-/**
- * Generate the depth /stencil test code.
- */
-static void
-generate_depth_stencil(LLVMBuilderRef builder,
-                       const struct lp_fragment_shader_variant_key *key,
-                       struct lp_type src_type,
-                       struct lp_build_mask_context *mask,
-                       LLVMValueRef stencil_refs[2],
-                       LLVMValueRef src,
-                       LLVMValueRef dst_ptr,
-                       LLVMValueRef facing,
-                       LLVMValueRef counter)
-{
-   const struct util_format_description *format_desc;
-   struct lp_type dst_type;
-
-   if (!key->depth.enabled && !key->stencil[0].enabled && !key->stencil[1].enabled)
-      return;
-
-   format_desc = util_format_description(key->zsbuf_format);
-   assert(format_desc);
-
-   /*
-    * Depths are expected to be between 0 and 1, even if they are stored in
-    * floats. Setting these bits here will ensure that the lp_build_conv() call
-    * below won't try to unnecessarily clamp the incoming values.
-    */
-   if(src_type.floating) {
-      src_type.sign = FALSE;
-      src_type.norm = TRUE;
-   }
-   else {
-      assert(!src_type.sign);
-      assert(src_type.norm);
-   }
-
-   /* Pick the depth type. */
-   dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
-
-   /* FIXME: Cope with a depth test type with a different bit width. */
-   assert(dst_type.width == src_type.width);
-   assert(dst_type.length == src_type.length);
-
-   /* Convert fragment Z from float to integer */
-   lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
-
-   dst_ptr = LLVMBuildBitCast(builder,
-                              dst_ptr,
-                              LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
-   lp_build_depth_stencil_test(builder,
-                               &key->depth,
-                               key->stencil,
-                               dst_type,
-                               format_desc,
-                               mask,
-                               stencil_refs,
-                               src,
-                               dst_ptr,
-                               facing,
-                               counter);
-}
-
 
 /**
  * Expand the relevent bits of mask_input to a 4-dword mask for the 
@@ -248,6 +186,26 @@ generate_quad_mask(LLVMBuilderRef builder,
 }
 
 
+#define EARLY_DEPTH_TEST  0x1
+#define LATE_DEPTH_TEST   0x2
+#define EARLY_DEPTH_WRITE 0x4
+#define LATE_DEPTH_WRITE  0x8
+
+static int
+find_output_by_semantic( const struct tgsi_shader_info *info,
+			 unsigned semantic,
+			 unsigned index )
+{
+   int i;
+
+   for (i = 0; i < info->num_outputs; i++)
+      if (info->output_semantic_name[i] == semantic &&
+	  info->output_semantic_index[i] == index)
+	 return i;
+
+   return -1;
+}
+
 
 /**
  * Generate the fragment shader, depth/stencil test, and alpha tests.
@@ -255,14 +213,13 @@ generate_quad_mask(LLVMBuilderRef builder,
  * \param partial_mask  if 1, do mask_input testing
  */
 static void
-generate_fs(struct llvmpipe_context *lp,
-            struct lp_fragment_shader *shader,
+generate_fs(struct lp_fragment_shader *shader,
             const struct lp_fragment_shader_variant_key *key,
             LLVMBuilderRef builder,
             struct lp_type type,
             LLVMValueRef context_ptr,
             unsigned i,
-            const struct lp_build_interp_soa_context *interp,
+            struct lp_build_interp_soa_context *interp,
             struct lp_build_sampler_soa *sampler,
             LLVMValueRef *pmask,
             LLVMValueRef (*color)[4],
@@ -272,18 +229,52 @@ generate_fs(struct llvmpipe_context *lp,
             LLVMValueRef mask_input,
             LLVMValueRef counter)
 {
+   const struct util_format_description *zs_format_desc = NULL;
    const struct tgsi_token *tokens = shader->base.tokens;
    LLVMTypeRef vec_type;
    LLVMValueRef consts_ptr;
    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
-   LLVMValueRef z = interp->pos[2];
+   LLVMValueRef z;
+   LLVMValueRef zs_value = NULL;
    LLVMValueRef stencil_refs[2];
-   struct lp_build_flow_context *flow;
    struct lp_build_mask_context mask;
-   boolean early_depth_stencil_test;
+   boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
+                            shader->info.base.num_inputs < 3 &&
+                            shader->info.base.num_instructions < 8);
    unsigned attrib;
    unsigned chan;
    unsigned cbuf;
+   unsigned depth_mode;
+
+   if (key->depth.enabled ||
+       key->stencil[0].enabled ||
+       key->stencil[1].enabled) {
+
+      zs_format_desc = util_format_description(key->zsbuf_format);
+      assert(zs_format_desc);
+
+      if (!shader->info.base.writes_z) {
+         if (key->alpha.enabled || shader->info.base.uses_kill)
+            /* With alpha test and kill, can do the depth test early
+             * and hopefully eliminate some quads.  But need to do a
+             * special deferred depth write once the final mask value
+             * is known.
+             */
+            depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
+         else
+            depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
+      }
+      else {
+         depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
+      }
+
+      if (!(key->depth.enabled && key->depth.writemask) &&
+          !(key->stencil[0].enabled && key->stencil[0].writemask))
+         depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
+   }
+   else {
+      depth_mode = 0;
+   }
 
    assert(i < 4);
 
@@ -294,20 +285,14 @@ generate_fs(struct llvmpipe_context *lp,
 
    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 
-   flow = lp_build_flow_create(builder);
-
    memset(outputs, 0, sizeof outputs);
 
-   lp_build_flow_scope_begin(flow);
-
    /* Declare the color and z variables */
    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
-	 color[cbuf][chan] = LLVMGetUndef(vec_type);
-	 lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
+	 color[cbuf][chan] = lp_build_alloca(builder, vec_type, "color");
       }
    }
-   lp_build_flow_scope_declare(flow, &z);
 
    /* do triangle edge testing */
    if (partial_mask) {
@@ -319,74 +304,126 @@ generate_fs(struct llvmpipe_context *lp,
    }
 
    /* 'mask' will control execution based on quad's pixel alive/killed state */
-   lp_build_mask_begin(&mask, flow, type, *pmask);
-
-   early_depth_stencil_test =
-      (key->depth.enabled || key->stencil[0].enabled) &&
-      !key->alpha.enabled &&
-      !shader->info.uses_kill &&
-      !shader->info.writes_z;
-
-   if (early_depth_stencil_test)
-      generate_depth_stencil(builder, key,
-                             type, &mask,
-                             stencil_refs, z, depth_ptr, facing, counter);
+   lp_build_mask_begin(&mask, builder, type, *pmask);
+
+   if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
+      lp_build_mask_check(&mask);
+
+   lp_build_interp_soa_update_pos(interp, i);
+   z = interp->pos[2];
+
+   if (depth_mode & EARLY_DEPTH_TEST) {
+      lp_build_depth_stencil_test(builder,
+                                  &key->depth,
+                                  key->stencil,
+                                  type,
+                                  zs_format_desc,
+                                  &mask,
+                                  stencil_refs,
+                                  z,
+                                  depth_ptr, facing,
+                                  &zs_value,
+                                  !simple_shader);
+
+      if (depth_mode & EARLY_DEPTH_WRITE) {
+         lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
+      }
+   }
 
+   lp_build_interp_soa_update_inputs(interp, i);
+   
+   /* Build the actual shader */
    lp_build_tgsi_soa(builder, tokens, type, &mask,
                      consts_ptr, interp->pos, interp->inputs,
-                     outputs, sampler, &shader->info);
+                     outputs, sampler, &shader->info.base);
 
-   /* loop over fragment shader outputs/results */
-   for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
-      for(chan = 0; chan < NUM_CHANNELS; ++chan) {
-         if(outputs[attrib][chan]) {
-            LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
-            lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
-
-            switch (shader->info.output_semantic_name[attrib]) {
-            case TGSI_SEMANTIC_COLOR:
-               {
-                  unsigned cbuf = shader->info.output_semantic_index[attrib];
-
-                  lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
-
-                  /* Alpha test */
-		  /* XXX: should only test the final assignment to alpha */
-                  if (cbuf == 0 && chan == 3 && key->alpha.enabled) {
-                     LLVMValueRef alpha = out;
-                     LLVMValueRef alpha_ref_value;
-                     alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
-                     alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
-                     lp_build_alpha_test(builder, key->alpha.func, type,
-                                         &mask, alpha, alpha_ref_value);
-                  }
-
-		  color[cbuf][chan] = out;
-                  break;
-               }
-
-            case TGSI_SEMANTIC_POSITION:
-               if(chan == 2)
-                  z = out;
-               break;
-            }
-         }
+
+   /* Alpha test */
+   if (key->alpha.enabled) {
+      int color0 = find_output_by_semantic(&shader->info.base,
+                                           TGSI_SEMANTIC_COLOR,
+                                           0);
+
+      if (color0 != -1 && outputs[color0][3]) {
+         LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
+         LLVMValueRef alpha_ref_value;
+
+         alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
+         alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
+
+         lp_build_alpha_test(builder, key->alpha.func, type,
+                             &mask, alpha, alpha_ref_value,
+                             (depth_mode & LATE_DEPTH_TEST) != 0);
       }
    }
 
-   if (!early_depth_stencil_test)
-      generate_depth_stencil(builder, key,
-                             type, &mask,
-                             stencil_refs, z, depth_ptr, facing, counter);
+   /* Late Z test */
+   if (depth_mode & LATE_DEPTH_TEST) { 
+      int pos0 = find_output_by_semantic(&shader->info.base,
+                                         TGSI_SEMANTIC_POSITION,
+                                         0);
+         
+      if (pos0 != -1 && outputs[pos0][2]) {
+         z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
+      }
 
-   lp_build_mask_end(&mask);
+      lp_build_depth_stencil_test(builder,
+                                  &key->depth,
+                                  key->stencil,
+                                  type,
+                                  zs_format_desc,
+                                  &mask,
+                                  stencil_refs,
+                                  z,
+                                  depth_ptr, facing,
+                                  &zs_value,
+                                  !simple_shader);
+      /* Late Z write */
+      if (depth_mode & LATE_DEPTH_WRITE) {
+         lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
+      }
+   }
+   else if ((depth_mode & EARLY_DEPTH_TEST) &&
+            (depth_mode & LATE_DEPTH_WRITE))
+   {
+      /* Need to apply a reduced mask to the depth write.  Reload the
+       * depth value, update from zs_value with the new mask value and
+       * write that out.
+       */
+      lp_build_deferred_depth_write(builder,
+                                    type,
+                                    zs_format_desc,
+                                    &mask,
+                                    depth_ptr,
+                                    zs_value);
+   }
 
-   lp_build_flow_scope_end(flow);
 
-   lp_build_flow_destroy(flow);
+   /* Color write  */
+   for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
+   {
+      if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR &&
+          shader->info.base.output_semantic_index[attrib] < key->nr_cbufs)
+      {
+         unsigned cbuf = shader->info.base.output_semantic_index[attrib];
+         for(chan = 0; chan < NUM_CHANNELS; ++chan) {
+            if(outputs[attrib][chan]) {
+               /* XXX: just initialize outputs to point at colors[] and
+                * skip this.
+                */
+               LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
+               lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
+               LLVMBuildStore(builder, out, color[cbuf][chan]);
+            }
+         }
+      }
+   }
 
-   *pmask = mask.value;
+   if (counter)
+      lp_build_occlusion_count(builder, type,
+                               lp_build_mask_value(&mask), counter);
 
+   *pmask = lp_build_mask_end(&mask);
 }
 
 
@@ -407,10 +444,10 @@ generate_blend(const struct pipe_blend_state *blend,
                LLVMValueRef context_ptr,
                LLVMValueRef mask,
                LLVMValueRef *src,
-               LLVMValueRef dst_ptr)
+               LLVMValueRef dst_ptr,
+               boolean do_branch)
 {
    struct lp_build_context bld;
-   struct lp_build_flow_context *flow;
    struct lp_build_mask_context mask_ctx;
    LLVMTypeRef vec_type;
    LLVMValueRef const_ptr;
@@ -421,10 +458,9 @@ generate_blend(const struct pipe_blend_state *blend,
 
    lp_build_context_init(&bld, builder, type);
 
-   flow = lp_build_flow_create(builder);
-
-   /* we'll use this mask context to skip blending if all pixels are dead */
-   lp_build_mask_begin(&mask_ctx, flow, type, mask);
+   lp_build_mask_begin(&mask_ctx, builder, type, mask);
+   if (do_branch)
+      lp_build_mask_check(&mask_ctx);
 
    vec_type = lp_build_vec_type(type);
 
@@ -457,7 +493,6 @@ generate_blend(const struct pipe_blend_state *blend,
    }
 
    lp_build_mask_end(&mask_ctx);
-   lp_build_flow_destroy(flow);
 }
 
 
@@ -468,13 +503,13 @@ generate_blend(const struct pipe_blend_state *blend,
  * 2x2 pixels.
  */
 static void
-generate_fragment(struct llvmpipe_context *lp,
+generate_fragment(struct llvmpipe_screen *screen,
                   struct lp_fragment_shader *shader,
                   struct lp_fragment_shader_variant *variant,
                   unsigned partial_mask)
 {
-   struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
    const struct lp_fragment_shader_variant_key *key = &variant->key;
+   struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
    char func_name[256];
    struct lp_type fs_type;
    struct lp_type blend_type;
@@ -502,11 +537,24 @@ generate_fragment(struct llvmpipe_context *lp,
    LLVMValueRef blend_mask;
    LLVMValueRef function;
    LLVMValueRef facing;
+   const struct util_format_description *zs_format_desc;
    unsigned num_fs;
    unsigned i;
    unsigned chan;
    unsigned cbuf;
 
+   /* Adjust color input interpolation according to flatshade state:
+    */
+   memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
+   for (i = 0; i < shader->info.base.num_inputs; i++) {
+      if (inputs[i].interp == LP_INTERP_COLOR) {
+	 if (key->flatshade)
+	    inputs[i].interp = LP_INTERP_CONSTANT;
+	 else
+	    inputs[i].interp = LP_INTERP_LINEAR;
+      }
+   }
+
 
    /* TODO: actually pick these based on the fs and color buffer
     * characteristics. */
@@ -542,12 +590,12 @@ generate_fragment(struct llvmpipe_context *lp,
    arg_types[0] = screen->context_ptr_type;            /* context */
    arg_types[1] = LLVMInt32Type();                     /* x */
    arg_types[2] = LLVMInt32Type();                     /* y */
-   arg_types[3] = LLVMFloatType();                     /* facing */
+   arg_types[3] = LLVMInt32Type();                     /* facing */
    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
-   arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
+   arg_types[8] = LLVMPointerType(LLVMInt8Type(), 0);  /* depth */
    arg_types[9] = LLVMInt32Type();                     /* mask_input */
    arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
 
@@ -558,7 +606,6 @@ generate_fragment(struct llvmpipe_context *lp,
 
    variant->function[partial_mask] = function;
 
-
    /* XXX: need to propagate noalias down into color param now we are
     * passing a pointer-to-pointer?
     */
@@ -606,8 +653,8 @@ generate_fragment(struct llvmpipe_context *lp,
     * already included in the shader key.
     */
    lp_build_interp_soa_init(&interp, 
-                            lp->num_inputs,
-                            lp->inputs,
+                            shader->info.base.num_inputs,
+                            inputs,
                             builder, fs_type,
                             a0_ptr, dadx_ptr, dady_ptr,
                             x, y);
@@ -616,17 +663,18 @@ generate_fragment(struct llvmpipe_context *lp,
    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 
    /* loop over quads in the block */
+   zs_format_desc = util_format_description(key->zsbuf_format);
+
    for(i = 0; i < num_fs; ++i) {
-      LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
+      LLVMValueRef depth_offset = LLVMConstInt(LLVMInt32Type(),
+                                               i*fs_type.length*zs_format_desc->block.bits/8,
+                                               0);
       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
       LLVMValueRef depth_ptr_i;
 
-      if(i != 0)
-         lp_build_interp_soa_update(&interp, i);
-
-      depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
+      depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
 
-      generate_fs(lp, shader, key,
+      generate_fs(shader, key,
                   builder,
                   fs_type,
                   context_ptr,
@@ -660,9 +708,18 @@ generate_fragment(struct llvmpipe_context *lp,
        * Convert the fs's output color and mask to fit to the blending type. 
        */
       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
+         LLVMValueRef fs_color_vals[LP_MAX_VECTOR_LENGTH];
+         
+         for (i = 0; i < num_fs; i++) {
+            fs_color_vals[i] =
+               LLVMBuildLoad(builder, fs_out_color[cbuf][chan][i], "fs_color_vals");
+         }
+
 	 lp_build_conv(builder, fs_type, blend_type,
-		       fs_out_color[cbuf][chan], num_fs,
+                       fs_color_vals,
+                       num_fs,
 		       &blend_in_color[chan], 1);
+
 	 lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
       }
 
@@ -685,14 +742,23 @@ generate_fragment(struct llvmpipe_context *lp,
       /*
        * Blending.
        */
-      generate_blend(&key->blend,
-                     rt,
-		     builder,
-		     blend_type,
-		     context_ptr,
-		     blend_mask,
-		     blend_in_color,
-		     color_ptr);
+      {
+         /* Could the 4x4 have been killed?
+          */
+         boolean do_branch = ((key->depth.enabled || key->stencil[0].enabled) &&
+                              !key->alpha.enabled &&
+                              !shader->info.base.uses_kill);
+
+         generate_blend(&key->blend,
+                        rt,
+                        builder,
+                        blend_type,
+                        context_ptr,
+                        blend_mask,
+                        blend_in_color,
+                        color_ptr,
+                        do_branch);
+      }
    }
 
 #ifdef PIPE_ARCH_X86
@@ -717,12 +783,17 @@ generate_fragment(struct llvmpipe_context *lp,
    /* Apply optimizations to LLVM IR */
    LLVMRunFunctionPassManager(screen->pass, function);
 
-   if (gallivm_debug & GALLIVM_DEBUG_IR) {
+   if ((gallivm_debug & GALLIVM_DEBUG_IR) || (LP_DEBUG & DEBUG_FS)) {
       /* Print the LLVM IR to stderr */
       lp_debug_dump_value(function);
       debug_printf("\n");
    }
 
+   /* Dump byte code to a file */
+   if (0) {
+      LLVMWriteBitcodeToFile(lp_build_module, "llvmpipe.bc");
+   }
+
    /*
     * Translate the LLVM IR into machine code.
     */
@@ -731,7 +802,7 @@ generate_fragment(struct llvmpipe_context *lp,
 
       variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
 
-      if (gallivm_debug & GALLIVM_DEBUG_ASM) {
+      if ((gallivm_debug & GALLIVM_DEBUG_ASM) || (LP_DEBUG & DEBUG_FS)) {
          lp_disassemble(f);
       }
       lp_func_delete_body(function);
@@ -832,7 +903,7 @@ lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
 }
 
 static struct lp_fragment_shader_variant *
-generate_variant(struct llvmpipe_context *lp,
+generate_variant(struct llvmpipe_screen *screen,
                  struct lp_fragment_shader *shader,
                  const struct lp_fragment_shader_variant_key *key)
 {
@@ -870,7 +941,7 @@ generate_variant(struct llvmpipe_context *lp,
          !key->stencil[0].enabled &&
          !key->alpha.enabled &&
          !key->depth.enabled &&
-         !shader->info.uses_kill
+         !shader->info.base.uses_kill
          ? TRUE : FALSE;
 
 
@@ -878,11 +949,11 @@ generate_variant(struct llvmpipe_context *lp,
       lp_debug_fs_variant(variant);
    }
 
-   generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
+   generate_fragment(screen, shader, variant, RAST_EDGE_TEST);
 
    if (variant->opaque) {
       /* Specialized shader, which doesn't need to read the color buffer. */
-      generate_fragment(lp, shader, variant, RAST_WHOLE);
+      generate_fragment(screen, shader, variant, RAST_WHOLE);
    } else {
       variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
    }
@@ -898,6 +969,7 @@ llvmpipe_create_fs_state(struct pipe_context *pipe,
    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
    struct lp_fragment_shader *shader;
    int nr_samplers;
+   int i;
 
    shader = CALLOC_STRUCT(lp_fragment_shader);
    if (!shader)
@@ -907,7 +979,7 @@ llvmpipe_create_fs_state(struct pipe_context *pipe,
    make_empty_list(&shader->variants);
 
    /* get/save the summary info for this shader */
-   tgsi_scan_shader(templ->tokens, &shader->info);
+   lp_build_tgsi_info(templ->tokens, &shader->info);
 
    /* we need to keep a local copy of the tokens */
    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
@@ -919,18 +991,58 @@ llvmpipe_create_fs_state(struct pipe_context *pipe,
       return NULL;
    }
 
-   nr_samplers = shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
+   nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
 
    shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
 				     sampler[nr_samplers]);
 
+   for (i = 0; i < shader->info.base.num_inputs; i++) {
+      shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
+
+      switch (shader->info.base.input_interpolate[i]) {
+      case TGSI_INTERPOLATE_CONSTANT:
+	 shader->inputs[i].interp = LP_INTERP_CONSTANT;
+	 break;
+      case TGSI_INTERPOLATE_LINEAR:
+	 shader->inputs[i].interp = LP_INTERP_LINEAR;
+	 break;
+      case TGSI_INTERPOLATE_PERSPECTIVE:
+	 shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
+	 break;
+      default:
+	 assert(0);
+	 break;
+      }
+
+      switch (shader->info.base.input_semantic_name[i]) {
+      case TGSI_SEMANTIC_COLOR:
+         /* Colors may be either linearly or constant interpolated in
+	  * the fragment shader, but that information isn't available
+	  * here.  Mark color inputs and fix them up later.
+          */
+	 shader->inputs[i].interp = LP_INTERP_COLOR;
+         break;
+      case TGSI_SEMANTIC_FACE:
+	 shader->inputs[i].interp = LP_INTERP_FACING;
+	 break;
+      case TGSI_SEMANTIC_POSITION:
+	 /* Position was already emitted above
+	  */
+	 shader->inputs[i].interp = LP_INTERP_POSITION;
+	 shader->inputs[i].src_index = 0;
+	 continue;
+      }
+
+      shader->inputs[i].src_index = i+1;
+   }
+
    if (LP_DEBUG & DEBUG_TGSI) {
       unsigned attrib;
       debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader->no, (void *) shader);
       tgsi_dump(templ->tokens, 0);
       debug_printf("usage masks:\n");
-      for (attrib = 0; attrib < shader->info.num_inputs; ++attrib) {
-         unsigned usage_mask = shader->info.input_usage_mask[attrib];
+      for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
+         unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
          debug_printf("  IN[%u].%s%s%s%s\n",
                       attrib,
                       usage_mask & TGSI_WRITEMASK_X ? "x" : "",
@@ -1159,10 +1271,10 @@ make_variant_key(struct llvmpipe_context *lp,
 
    /* This value will be the same for all the variants of a given shader:
     */
-   key->nr_samplers = shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
+   key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
 
    for(i = 0; i < key->nr_samplers; ++i) {
-      if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
+      if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
          lp_sampler_static_state(&key->sampler[i],
 				 lp->fragment_sampler_views[i],
 				 lp->sampler[i]);
@@ -1177,6 +1289,7 @@ make_variant_key(struct llvmpipe_context *lp,
 void 
 llvmpipe_update_fs(struct llvmpipe_context *lp)
 {
+   struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
    struct lp_fragment_shader *shader = lp->fs;
    struct lp_fragment_shader_variant_key key;
    struct lp_fragment_shader_variant *variant = NULL;
@@ -1217,7 +1330,7 @@ llvmpipe_update_fs(struct llvmpipe_context *lp)
       }
       t0 = os_time_get();
 
-      variant = generate_variant(lp, shader, &key);
+      variant = generate_variant(screen, shader, &key);
 
       t1 = os_time_get();
       dt = t1 - t0;
@@ -1237,6 +1350,10 @@ llvmpipe_update_fs(struct llvmpipe_context *lp)
 
 
 
+
+
+
+
 void
 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
 {
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.h b/src/gallium/drivers/llvmpipe/lp_state_fs.h
index 4999b8dca1a..7d58c4936c7 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_fs.h
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs.h
@@ -34,6 +34,8 @@
 #include "pipe/p_state.h"
 #include "tgsi/tgsi_scan.h" /* for tgsi_shader_info */
 #include "gallivm/lp_bld_sample.h" /* for struct lp_sampler_static_state */
+#include "gallivm/lp_bld_tgsi.h" /* for lp_tgsi_info */
+#include "lp_bld_interp.h" /* for struct lp_shader_input */
 
 
 struct tgsi_token;
@@ -96,7 +98,7 @@ struct lp_fragment_shader
 {
    struct pipe_shader_state base;
 
-   struct tgsi_shader_info info;
+   struct lp_tgsi_info info;
 
    struct lp_fs_variant_list_item variants;
 
@@ -107,6 +109,9 @@ struct lp_fragment_shader
    unsigned no;
    unsigned variants_created;
    unsigned variants_cached;
+
+   /** Fragment shader input interpolation info */
+   struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
 };
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_state_sampler.c b/src/gallium/drivers/llvmpipe/lp_state_sampler.c
index 17a4a0ed02d..1dd866195d3 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_sampler.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_sampler.c
@@ -246,9 +246,9 @@ llvmpipe_prepare_vertex_sampling(struct llvmpipe_context *lp,
                                  struct pipe_sampler_view **views)
 {
    unsigned i;
-   uint32_t row_stride[DRAW_MAX_TEXTURE_LEVELS];
-   uint32_t img_stride[DRAW_MAX_TEXTURE_LEVELS];
-   const void *data[DRAW_MAX_TEXTURE_LEVELS];
+   uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS];
+   uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS];
+   const void *data[PIPE_MAX_TEXTURE_LEVELS];
 
    assert(num <= PIPE_MAX_VERTEX_SAMPLERS);
    if (!num)
diff --git a/src/gallium/drivers/llvmpipe/lp_state_setup.c b/src/gallium/drivers/llvmpipe/lp_state_setup.c
new file mode 100644
index 00000000000..2c8b8b9a928
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_state_setup.c
@@ -0,0 +1,759 @@
+/**************************************************************************
+ *
+ * Copyright 2010 VMware.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+
+
+#include "util/u_math.h"
+#include "util/u_memory.h"
+#include "util/u_simple_list.h"
+#include "os/os_time.h"
+#include "gallivm/lp_bld_debug.h"
+#include "gallivm/lp_bld_init.h"
+#include "gallivm/lp_bld_intr.h"
+#include <llvm-c/Analysis.h>	/* for LLVMVerifyFunction */
+
+#include "lp_perf.h"
+#include "lp_debug.h"
+#include "lp_flush.h"
+#include "lp_screen.h"
+#include "lp_context.h"
+#include "lp_setup_context.h"
+#include "lp_rast.h"
+#include "lp_state.h"
+#include "lp_state_fs.h"
+#include "lp_state_setup.h"
+
+
+
+/* currently organized to interpolate full float[4] attributes even
+ * when some elements are unused.  Later, can pack vertex data more
+ * closely.
+ */
+
+
+struct lp_setup_args
+{
+   /* Function arguments:
+    */
+   LLVMValueRef v0;
+   LLVMValueRef v1;
+   LLVMValueRef v2;
+   LLVMValueRef facing;		/* boolean */
+   LLVMValueRef a0;
+   LLVMValueRef dadx;
+   LLVMValueRef dady;
+
+   /* Derived:
+    */
+   LLVMValueRef x0_center;
+   LLVMValueRef y0_center;
+   LLVMValueRef dy20_ooa;
+   LLVMValueRef dy01_ooa;
+   LLVMValueRef dx20_ooa;
+   LLVMValueRef dx01_ooa;
+};
+
+static LLVMTypeRef type4f(void)
+{
+   return LLVMVectorType(LLVMFloatType(), 4);
+}
+
+
+/* Equivalent of _mm_setr_ps(a,b,c,d)
+ */
+static LLVMValueRef vec4f(LLVMBuilderRef bld,
+			  LLVMValueRef a, LLVMValueRef b, LLVMValueRef c, LLVMValueRef d,
+			  const char *name)
+{
+   LLVMValueRef i0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
+   LLVMValueRef i1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
+   LLVMValueRef i2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
+   LLVMValueRef i3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
+
+   LLVMValueRef res = LLVMGetUndef(type4f());
+
+   res = LLVMBuildInsertElement(bld, res, a, i0, "");
+   res = LLVMBuildInsertElement(bld, res, b, i1, "");
+   res = LLVMBuildInsertElement(bld, res, c, i2, "");
+   res = LLVMBuildInsertElement(bld, res, d, i3, name);
+
+   return res;
+}
+
+/* Equivalent of _mm_set1_ps(a)
+ */
+static LLVMValueRef vec4f_from_scalar(LLVMBuilderRef bld,
+				      LLVMValueRef a,
+				      const char *name)
+{
+   LLVMValueRef res = LLVMGetUndef(type4f());
+   int i;
+
+   for(i = 0; i < 4; ++i) {
+      LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
+      res = LLVMBuildInsertElement(bld, res, a, index, i == 3 ? name : "");
+   }
+
+   return res;
+}
+
+static void
+store_coef(LLVMBuilderRef builder,
+	   struct lp_setup_args *args,
+	   unsigned slot,
+	   LLVMValueRef a0,
+	   LLVMValueRef dadx,
+	   LLVMValueRef dady)
+{
+   LLVMValueRef idx = LLVMConstInt(LLVMInt32Type(), slot, 0);
+   
+   LLVMBuildStore(builder,
+		  a0, 
+		  LLVMBuildGEP(builder, args->a0, &idx, 1, ""));
+
+   LLVMBuildStore(builder,
+		  dadx, 
+		  LLVMBuildGEP(builder, args->dadx, &idx, 1, ""));
+
+   LLVMBuildStore(builder,
+		  dady, 
+		  LLVMBuildGEP(builder, args->dady, &idx, 1, ""));
+}
+
+
+
+static void 
+emit_constant_coef4( LLVMBuilderRef builder,
+		     struct lp_setup_args *args,
+		     unsigned slot,
+		     LLVMValueRef vert,
+		     unsigned attr)
+{
+   LLVMValueRef zero      = LLVMConstReal(LLVMFloatType(), 0.0);
+   LLVMValueRef zerovec   = vec4f_from_scalar(builder, zero, "zero");
+   LLVMValueRef idx       = LLVMConstInt(LLVMInt32Type(), attr, 0);
+   LLVMValueRef attr_ptr  = LLVMBuildGEP(builder, vert, &idx, 1, "attr_ptr");
+   LLVMValueRef vert_attr = LLVMBuildLoad(builder, attr_ptr, "vert_attr");
+
+   store_coef(builder, args, slot, vert_attr, zerovec, zerovec);
+}
+
+
+
+/**
+ * Setup the fragment input attribute with the front-facing value.
+ * \param frontface  is the triangle front facing?
+ */
+static void 
+emit_facing_coef( LLVMBuilderRef builder,
+		  struct lp_setup_args *args,
+		  unsigned slot )
+{
+   LLVMValueRef a0_0 = args->facing;
+   LLVMValueRef a0_0f = LLVMBuildSIToFP(builder, a0_0, LLVMFloatType(), "");
+   LLVMValueRef zero = LLVMConstReal(LLVMFloatType(), 0.0);
+   LLVMValueRef a0      = vec4f(builder, a0_0f, zero, zero, zero, "facing");
+   LLVMValueRef zerovec = vec4f_from_scalar(builder, zero, "zero");
+
+   store_coef(builder, args, slot, a0, zerovec, zerovec);
+}
+
+
+static LLVMValueRef
+vert_attrib(LLVMBuilderRef b,
+	    LLVMValueRef vert,
+	    int attr,
+	    int elem,
+	    const char *name)
+{
+   LLVMValueRef idx[2];
+   idx[0] = LLVMConstInt(LLVMInt32Type(), attr, 0);
+   idx[1] = LLVMConstInt(LLVMInt32Type(), elem, 0);
+   return LLVMBuildLoad(b, LLVMBuildGEP(b, vert, idx, 2, ""), name);
+}
+
+
+
+static void 
+emit_coef4( LLVMBuilderRef b,
+	    struct lp_setup_args *args,
+	    unsigned slot,
+	    LLVMValueRef a0,
+	    LLVMValueRef a1,
+	    LLVMValueRef a2)
+{
+   LLVMValueRef dy20_ooa = args->dy20_ooa;
+   LLVMValueRef dy01_ooa = args->dy01_ooa;
+   LLVMValueRef dx20_ooa = args->dx20_ooa;
+   LLVMValueRef dx01_ooa = args->dx01_ooa;
+   LLVMValueRef x0_center = args->x0_center;
+   LLVMValueRef y0_center = args->y0_center;
+
+   /* XXX: using fsub, fmul on vector types -- does this work??
+    */
+   LLVMValueRef da01 = LLVMBuildFSub(b, a0, a1, "da01");
+   LLVMValueRef da20 = LLVMBuildFSub(b, a2, a0, "da20");
+
+   /* Calculate dadx (vec4f)
+    */
+   LLVMValueRef da01_dy20_ooa = LLVMBuildFMul(b, da01, dy20_ooa, "da01_dy20_ooa");
+   LLVMValueRef da20_dy01_ooa = LLVMBuildFMul(b, da20, dy01_ooa, "da20_dy01_ooa");
+   LLVMValueRef dadx          = LLVMBuildFSub(b, da01_dy20_ooa, da20_dy01_ooa, "dadx");
+
+   /* Calculate dady (vec4f)
+    */
+   LLVMValueRef da01_dx20_ooa = LLVMBuildFMul(b, da01, dx20_ooa, "da01_dx20_ooa");
+   LLVMValueRef da20_dx01_ooa = LLVMBuildFMul(b, da20, dx01_ooa, "da20_dx01_ooa");
+   LLVMValueRef dady          = LLVMBuildFSub(b, da20_dx01_ooa, da01_dx20_ooa, "dady");
+
+   /* Calculate a0 - the attribute value at the origin
+    */
+   LLVMValueRef dadx_x0       = LLVMBuildFMul(b, dadx, x0_center, "dadx_x0"); 
+   LLVMValueRef dady_y0       = LLVMBuildFMul(b, dady, y0_center, "dady_y0"); 
+   LLVMValueRef attr_v0       = LLVMBuildFAdd(b, dadx_x0, dady_y0, "attr_v0"); 
+   LLVMValueRef attr_0        = LLVMBuildFSub(b, a0, attr_v0, "attr_0"); 
+
+   store_coef(b, args, slot, attr_0, dadx, dady);
+}
+
+
+static void 
+emit_linear_coef( LLVMBuilderRef b,
+		  struct lp_setup_args *args,
+		  unsigned slot,
+		  unsigned vert_attr)
+{
+   LLVMValueRef idx = LLVMConstInt(LLVMInt32Type(), vert_attr, 0);
+
+   LLVMValueRef a0 = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v0, &idx, 1, ""), "v0a");
+   LLVMValueRef a1 = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v1, &idx, 1, ""), "v1a");
+   LLVMValueRef a2 = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v2, &idx, 1, ""), "v2a");
+
+   emit_coef4(b, args, slot, a0, a1, a2);
+}
+
+
+
+/**
+ * Compute a0, dadx and dady for a perspective-corrected interpolant,
+ * for a triangle.
+ * We basically multiply the vertex value by 1/w before computing
+ * the plane coefficients (a0, dadx, dady).
+ * Later, when we compute the value at a particular fragment position we'll
+ * divide the interpolated value by the interpolated W at that fragment.
+ */
+static void 
+emit_perspective_coef( LLVMBuilderRef b,
+		       struct lp_setup_args *args,
+		       unsigned slot,
+		       unsigned vert_attr)
+{
+   /* premultiply by 1/w  (v[0][3] is always 1/w):
+    */
+   LLVMValueRef idx = LLVMConstInt(LLVMInt32Type(), vert_attr, 0);
+
+   LLVMValueRef v0a = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v0, &idx, 1, ""), "v0a");
+   LLVMValueRef v1a = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v1, &idx, 1, ""), "v1a");
+   LLVMValueRef v2a = LLVMBuildLoad(b, LLVMBuildGEP(b, args->v2, &idx, 1, ""), "v2a");
+
+   LLVMValueRef v0_oow = vec4f_from_scalar(b, vert_attrib(b, args->v0, 0, 3, ""), "v0_oow");
+   LLVMValueRef v1_oow = vec4f_from_scalar(b, vert_attrib(b, args->v1, 0, 3, ""), "v1_oow");
+   LLVMValueRef v2_oow = vec4f_from_scalar(b, vert_attrib(b, args->v2, 0, 3, ""), "v2_oow");
+
+   LLVMValueRef v0_oow_v0a = LLVMBuildFMul(b, v0a, v0_oow, "v0_oow_v0a");
+   LLVMValueRef v1_oow_v1a = LLVMBuildFMul(b, v1a, v1_oow, "v1_oow_v1a");
+   LLVMValueRef v2_oow_v2a = LLVMBuildFMul(b, v2a, v2_oow, "v2_oow_v2a");
+
+   emit_coef4(b, args, slot, v0_oow_v0a, v1_oow_v1a, v2_oow_v2a);
+}
+
+
+static void
+emit_position_coef( LLVMBuilderRef builder,
+		    struct lp_setup_args *args,
+		    int slot, int attrib )
+{
+   emit_linear_coef(builder, args, slot, attrib);
+}
+
+
+
+
+/**
+ * Compute the inputs-> dadx, dady, a0 values.
+ */
+static void 
+emit_tri_coef( LLVMBuilderRef builder,
+	       const struct lp_setup_variant_key *key,
+	       struct lp_setup_args *args )
+{
+   unsigned slot;
+
+   /* The internal position input is in slot zero:
+    */
+   emit_position_coef(builder, args, 0, 0);
+
+   /* setup interpolation for all the remaining attributes:
+    */
+   for (slot = 0; slot < key->num_inputs; slot++) {
+      unsigned vert_attr = key->inputs[slot].src_index;
+
+      switch (key->inputs[slot].interp) {
+      case LP_INTERP_CONSTANT:
+	 if (key->flatshade_first) {
+	    emit_constant_coef4(builder, args, slot+1, args->v0, vert_attr);
+	 }
+	 else {
+	    emit_constant_coef4(builder, args, slot+1, args->v2, vert_attr);
+	 }
+	 break;
+
+      case LP_INTERP_LINEAR:
+	 emit_linear_coef(builder, args, slot+1, vert_attr);
+         break;
+
+      case LP_INTERP_PERSPECTIVE:
+	 emit_perspective_coef(builder, args, slot+1, vert_attr);
+         break;
+
+      case LP_INTERP_POSITION:
+         /*
+          * The generated pixel interpolators will pick up the coeffs from
+          * slot 0.
+          */
+         break;
+
+      case LP_INTERP_FACING:
+         emit_facing_coef(builder, args, slot+1);
+         break;
+
+      default:
+         assert(0);
+      }
+   }
+}
+
+
+/* XXX: This is generic code, share with fs/vs codegen:
+ */
+static lp_jit_setup_triangle
+finalize_function(struct llvmpipe_screen *screen,
+		  LLVMBuilderRef builder,
+		  LLVMValueRef function)
+{
+   void *f;
+
+   /* Verify the LLVM IR.  If invalid, dump and abort */
+#ifdef DEBUG
+   if (LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
+      if (1)
+         lp_debug_dump_value(function);
+      abort();
+   }
+#endif
+
+   /* Apply optimizations to LLVM IR */
+   LLVMRunFunctionPassManager(screen->pass, function);
+
+   if (gallivm_debug & GALLIVM_DEBUG_IR)
+   {
+      /* Print the LLVM IR to stderr */
+      lp_debug_dump_value(function);
+      debug_printf("\n");
+   }
+
+   /*
+    * Translate the LLVM IR into machine code.
+    */
+   f = LLVMGetPointerToGlobal(screen->engine, function);
+
+   if (gallivm_debug & GALLIVM_DEBUG_ASM)
+   {
+      lp_disassemble(f);
+   }
+
+   lp_func_delete_body(function);
+
+   return f;
+}
+
+/* XXX: Generic code:
+ */
+static void
+lp_emit_emms(LLVMBuilderRef builder)
+{
+#ifdef PIPE_ARCH_X86
+   /* Avoid corrupting the FPU stack on 32bit OSes. */
+   lp_build_intrinsic(builder, "llvm.x86.mmx.emms", LLVMVoidType(), NULL, 0);
+#endif
+}
+
+
+/* XXX: generic code:
+ */
+static void
+set_noalias(LLVMBuilderRef builder,
+	    LLVMValueRef function,
+	    const LLVMTypeRef *arg_types,
+	    int nr_args)
+{
+   int i;
+   for(i = 0; i < Elements(arg_types); ++i)
+      if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
+         LLVMAddAttribute(LLVMGetParam(function, i),
+			  LLVMNoAliasAttribute);
+}
+
+static void
+init_args(LLVMBuilderRef b, 
+	  struct lp_setup_args *args,
+	  const struct lp_setup_variant *variant)
+{
+   LLVMValueRef v0_x = vert_attrib(b, args->v0, 0, 0, "v0_x");
+   LLVMValueRef v0_y = vert_attrib(b, args->v0, 0, 1, "v0_y");
+
+   LLVMValueRef v1_x = vert_attrib(b, args->v1, 0, 0, "v1_x");
+   LLVMValueRef v1_y = vert_attrib(b, args->v1, 0, 1, "v1_y");
+
+   LLVMValueRef v2_x = vert_attrib(b, args->v2, 0, 0, "v2_x");
+   LLVMValueRef v2_y = vert_attrib(b, args->v2, 0, 1, "v2_y");
+
+   LLVMValueRef pixel_center = LLVMConstReal(LLVMFloatType(),
+					     variant->key.pixel_center_half ? 0.5 : 0);
+
+   LLVMValueRef x0_center = LLVMBuildFSub(b, v0_x, pixel_center, "x0_center" );
+   LLVMValueRef y0_center = LLVMBuildFSub(b, v0_y, pixel_center, "y0_center" );
+   
+   LLVMValueRef dx01 = LLVMBuildFSub(b, v0_x, v1_x, "dx01");
+   LLVMValueRef dy01 = LLVMBuildFSub(b, v0_y, v1_y, "dy01");
+   LLVMValueRef dx20 = LLVMBuildFSub(b, v2_x, v0_x, "dx20");
+   LLVMValueRef dy20 = LLVMBuildFSub(b, v2_y, v0_y, "dy20");
+
+   LLVMValueRef one  = LLVMConstReal(LLVMFloatType(), 1.0);
+   LLVMValueRef e    = LLVMBuildFMul(b, dx01, dy20, "e");
+   LLVMValueRef f    = LLVMBuildFMul(b, dx20, dy01, "f");
+   LLVMValueRef ooa  = LLVMBuildFDiv(b, one, LLVMBuildFSub(b, e, f, ""), "ooa");
+
+   LLVMValueRef dy20_ooa = LLVMBuildFMul(b, dy20, ooa, "dy20_ooa");
+   LLVMValueRef dy01_ooa = LLVMBuildFMul(b, dy01, ooa, "dy01_ooa");
+   LLVMValueRef dx20_ooa = LLVMBuildFMul(b, dx20, ooa, "dx20_ooa");
+   LLVMValueRef dx01_ooa = LLVMBuildFMul(b, dx01, ooa, "dx01_ooa");
+
+   args->dy20_ooa  = vec4f_from_scalar(b, dy20_ooa, "dy20_ooa_4f");
+   args->dy01_ooa  = vec4f_from_scalar(b, dy01_ooa, "dy01_ooa_4f");
+
+   args->dx20_ooa  = vec4f_from_scalar(b, dx20_ooa, "dx20_ooa_4f");
+   args->dx01_ooa  = vec4f_from_scalar(b, dx01_ooa, "dx01_ooa_4f");
+
+   args->x0_center = vec4f_from_scalar(b, x0_center, "x0_center_4f");
+   args->y0_center = vec4f_from_scalar(b, y0_center, "y0_center_4f");
+}
+
+/**
+ * Generate the runtime callable function for the coefficient calculation.
+ *
+ */
+static struct lp_setup_variant *
+generate_setup_variant(struct llvmpipe_screen *screen,
+		       struct lp_setup_variant_key *key)
+{
+   struct lp_setup_variant *variant = NULL;
+   struct lp_setup_args args;
+   char func_name[256];
+   LLVMTypeRef vec4f_type;
+   LLVMTypeRef func_type;
+   LLVMTypeRef arg_types[7];
+   LLVMBasicBlockRef block;
+   LLVMBuilderRef builder;
+   int64_t t0, t1;
+
+   if (0)
+      goto fail;
+
+   variant = CALLOC_STRUCT(lp_setup_variant);
+   if (variant == NULL)
+      goto fail;
+
+   if (LP_DEBUG & DEBUG_COUNTERS) {
+      t0 = os_time_get();
+   }
+
+   memcpy(&variant->key, key, key->size);
+   variant->list_item_global.base = variant;
+
+   util_snprintf(func_name, sizeof(func_name), "fs%u_setup%u",
+		 0,
+		 variant->no);
+
+   /* Currently always deal with full 4-wide vertex attributes from
+    * the vertices.
+    */
+
+   vec4f_type = LLVMVectorType(LLVMFloatType(), 4);
+
+   arg_types[0] = LLVMPointerType(vec4f_type, 0);        /* v0 */
+   arg_types[1] = LLVMPointerType(vec4f_type, 0);        /* v1 */
+   arg_types[2] = LLVMPointerType(vec4f_type, 0);        /* v2 */
+   arg_types[3] = LLVMInt32Type();			/* facing */
+   arg_types[4] = LLVMPointerType(vec4f_type, 0);	/* a0, aligned */
+   arg_types[5] = LLVMPointerType(vec4f_type, 0);	/* dadx, aligned */
+   arg_types[6] = LLVMPointerType(vec4f_type, 0);	/* dady, aligned */
+
+   func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
+
+   variant->function = LLVMAddFunction(screen->module, func_name, func_type);
+   if (!variant->function)
+      goto fail;
+
+   LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
+
+   args.v0       = LLVMGetParam(variant->function, 0);
+   args.v1       = LLVMGetParam(variant->function, 1);
+   args.v2       = LLVMGetParam(variant->function, 2);
+   args.facing   = LLVMGetParam(variant->function, 3);
+   args.a0       = LLVMGetParam(variant->function, 4);
+   args.dadx     = LLVMGetParam(variant->function, 5);
+   args.dady     = LLVMGetParam(variant->function, 6);
+
+   lp_build_name(args.v0, "in_v0");
+   lp_build_name(args.v1, "in_v1");
+   lp_build_name(args.v2, "in_v2");
+   lp_build_name(args.facing, "in_facing");
+   lp_build_name(args.a0, "out_a0");
+   lp_build_name(args.dadx, "out_dadx");
+   lp_build_name(args.dady, "out_dady");
+
+   /*
+    * Function body
+    */
+   block = LLVMAppendBasicBlock(variant->function, "entry");
+   builder = LLVMCreateBuilder();
+   LLVMPositionBuilderAtEnd(builder, block);
+
+   set_noalias(builder, variant->function, arg_types, Elements(arg_types));
+   init_args(builder, &args, variant);
+   emit_tri_coef(builder, &variant->key, &args);
+
+   lp_emit_emms(builder);
+   LLVMBuildRetVoid(builder);
+   LLVMDisposeBuilder(builder);
+
+   variant->jit_function = finalize_function(screen, builder,
+					     variant->function);
+   if (!variant->jit_function)
+      goto fail;
+
+   /*
+    * Update timing information:
+    */
+   if (LP_DEBUG & DEBUG_COUNTERS) {
+      t1 = os_time_get();
+      LP_COUNT_ADD(llvm_compile_time, t1 - t0);
+      LP_COUNT_ADD(nr_llvm_compiles, 1);
+   }
+   
+   return variant;
+
+fail:
+   if (variant) {
+      if (variant->function) {
+	 if (variant->jit_function)
+	    LLVMFreeMachineCodeForFunction(screen->engine,
+					   variant->function);
+	 LLVMDeleteFunction(variant->function);
+      }
+      FREE(variant);
+   }
+   
+   return NULL;
+}
+
+
+
+static void
+lp_make_setup_variant_key(struct llvmpipe_context *lp,
+			  struct lp_setup_variant_key *key)
+{
+   struct lp_fragment_shader *fs = lp->fs;
+   unsigned i;
+
+   assert(sizeof key->inputs[0] == sizeof(ushort));
+   
+   key->num_inputs = fs->info.base.num_inputs;
+   key->flatshade_first = lp->rasterizer->flatshade_first;
+   key->pixel_center_half = lp->rasterizer->gl_rasterization_rules;
+   key->size = Offset(struct lp_setup_variant_key,
+		      inputs[key->num_inputs]);
+   key->pad = 0;
+
+   memcpy(key->inputs, fs->inputs, key->num_inputs * sizeof key->inputs[0]);
+   for (i = 0; i < key->num_inputs; i++) {
+      if (key->inputs[i].interp == LP_INTERP_COLOR) {
+	 if (lp->rasterizer->flatshade)
+	    key->inputs[i].interp = LP_INTERP_CONSTANT;
+	 else
+	    key->inputs[i].interp = LP_INTERP_LINEAR;
+      }
+   }
+
+}
+
+
+static void
+remove_setup_variant(struct llvmpipe_context *lp,
+		     struct lp_setup_variant *variant)
+{
+   struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
+
+   if (gallivm_debug & GALLIVM_DEBUG_IR) {
+      debug_printf("llvmpipe: del setup_variant #%u total %u\n",
+		   variant->no, lp->nr_setup_variants);
+   }
+
+   if (variant->function) {
+      if (variant->jit_function)
+	 LLVMFreeMachineCodeForFunction(screen->engine,
+					variant->function);
+      LLVMDeleteFunction(variant->function);
+   }
+
+   remove_from_list(&variant->list_item_global);
+   lp->nr_setup_variants--;
+   FREE(variant);
+}
+
+
+
+/* When the number of setup variants exceeds a threshold, cull a
+ * fraction (currently a quarter) of them.
+ */
+static void
+cull_setup_variants(struct llvmpipe_context *lp)
+{
+   struct pipe_context *pipe = &lp->pipe;
+   int i;
+
+   /*
+    * XXX: we need to flush the context until we have some sort of reference
+    * counting in fragment shaders as they may still be binned
+    * Flushing alone might not be sufficient we need to wait on it too.
+    */
+   llvmpipe_finish(pipe, __FUNCTION__);
+
+   for (i = 0; i < LP_MAX_SETUP_VARIANTS / 4; i++) {
+      struct lp_setup_variant_list_item *item = last_elem(&lp->setup_variants_list);
+      remove_setup_variant(lp, item->base);
+   }
+}
+
+
+/**
+ * Update fragment/vertex shader linkage state.  This is called just
+ * prior to drawing something when some fragment-related state has
+ * changed.
+ */
+void 
+llvmpipe_update_setup(struct llvmpipe_context *lp)
+{
+   struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
+
+   struct lp_setup_variant_key *key = &lp->setup_variant.key;
+   struct lp_setup_variant *variant = NULL;
+   struct lp_setup_variant_list_item *li;
+
+   lp_make_setup_variant_key(lp, key);
+
+   foreach(li, &lp->setup_variants_list) {
+      if(li->base->key.size == key->size &&
+	 memcmp(&li->base->key, key, key->size) == 0) {
+         variant = li->base;
+         break;
+      }
+   }
+
+   if (variant) {
+      move_to_head(&lp->setup_variants_list, &variant->list_item_global);
+   }
+   else {
+      if (lp->nr_setup_variants >= LP_MAX_SETUP_VARIANTS) {
+	 cull_setup_variants(lp);
+      }
+
+      variant = generate_setup_variant(screen, key);
+      insert_at_head(&lp->setup_variants_list, &variant->list_item_global);
+      lp->nr_setup_variants++;
+   }
+
+   lp_setup_set_setup_variant(lp->setup,
+			      variant);
+}
+
+void
+lp_delete_setup_variants(struct llvmpipe_context *lp)
+{
+   struct lp_setup_variant_list_item *li;
+   li = first_elem(&lp->setup_variants_list);
+   while(!at_end(&lp->setup_variants_list, li)) {
+      struct lp_setup_variant_list_item *next = next_elem(li);
+      remove_setup_variant(lp, li->base);
+      li = next;
+   }
+}
+
+void
+lp_dump_setup_coef( const struct lp_setup_variant_key *key,
+		    const float (*sa0)[4],
+		    const float (*sdadx)[4],
+		    const float (*sdady)[4])
+{
+   int i, slot;
+
+   for (i = 0; i < NUM_CHANNELS; i++) {
+      float a0   = sa0  [0][i];
+      float dadx = sdadx[0][i];
+      float dady = sdady[0][i];
+
+      debug_printf("POS.%c: a0 = %f, dadx = %f, dady = %f\n",
+		   "xyzw"[i],
+		   a0, dadx, dady);
+   }
+
+   for (slot = 0; slot < key->num_inputs; slot++) {
+      unsigned usage_mask = key->inputs[slot].usage_mask;
+      for (i = 0; i < NUM_CHANNELS; i++) {
+	 if (usage_mask & (1 << i)) {
+	    float a0   = sa0  [1 + slot][i];
+	    float dadx = sdadx[1 + slot][i];
+	    float dady = sdady[1 + slot][i];
+
+	    debug_printf("IN[%u].%c: a0 = %f, dadx = %f, dady = %f\n",
+			 slot,
+			 "xyzw"[i],
+			 a0, dadx, dady);
+	 }
+      }
+   }
+}
diff --git a/src/gallium/drivers/llvmpipe/lp_state_setup.h b/src/gallium/drivers/llvmpipe/lp_state_setup.h
new file mode 100644
index 00000000000..b0c81baa75f
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_state_setup.h
@@ -0,0 +1,80 @@
+#ifndef LP_STATE_SETUP_H
+#define LP_STATE_SETUP_H
+
+#include "lp_bld_interp.h"
+
+
+struct llvmpipe_context;
+struct lp_setup_variant;
+
+struct lp_setup_variant_list_item
+{
+   struct lp_setup_variant *base;
+   struct lp_setup_variant_list_item *next, *prev;
+};
+
+
+struct lp_setup_variant_key {   
+   unsigned num_inputs:8;
+   unsigned flatshade_first:1;
+   unsigned pixel_center_half:1;
+   unsigned pad:7;
+   unsigned size:16;
+   struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
+};
+
+
+typedef void (*lp_jit_setup_triangle)( const float (*v0)[4],
+				       const float (*v1)[4],
+				       const float (*v2)[4],
+				       boolean front_facing,
+				       float (*a0)[4],
+				       float (*dadx)[4],
+				       float (*dady)[4] );
+
+
+
+
+/* At this stage, for a given variant key, we create a
+ * draw_vertex_info struct telling the draw module how to format the
+ * vertices, and an llvm-generated function which calculates the
+ * attribute interpolants (a0, dadx, dady) from three of those
+ * vertices.
+ */
+struct lp_setup_variant {
+   struct lp_setup_variant_key key;
+   
+   struct lp_setup_variant_list_item list_item_global;
+
+   /* XXX: this is a pointer to the LLVM IR.  Once jit_function is
+    * generated, we never need to use the IR again - need to find a
+    * way to release this data without destroying the generated
+    * assembly.
+    */
+   LLVMValueRef function;
+
+   /* The actual generated setup function:
+    */
+   lp_jit_setup_triangle jit_function;
+
+   unsigned no;
+};
+
+void lp_setup_tri_fallback( const float (*v0)[4],
+			    const float (*v1)[4],
+			    const float (*v2)[4],
+			    boolean front_facing,
+			    float (*a0)[4],
+			    float (*dadx)[4],
+			    float (*dady)[4],
+			    const struct lp_setup_variant_key *key );
+
+void lp_delete_setup_variants(struct llvmpipe_context *lp);
+
+void
+lp_dump_setup_coef( const struct lp_setup_variant_key *key,
+		    const float (*sa0)[4],
+		    const float (*sdadx)[4],
+		    const float (*sdady)[4]);
+
+#endif
diff --git a/src/gallium/drivers/llvmpipe/lp_test_round.c b/src/gallium/drivers/llvmpipe/lp_test_round.c
index 57b0ee57767..816518e5081 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_round.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_round.c
@@ -75,10 +75,7 @@ add_test(LLVMModuleRef module, const char *name, lp_func_t lp_func)
    LLVMValueRef ret;
    struct lp_build_context bld;
 
-   bld.builder = builder;
-   bld.type.floating = 1;
-   bld.type.width = 32;
-   bld.type.length = 4;
+   lp_build_context_init(&bld, builder, lp_float32_vec4_type());
 
    LLVMSetFunctionCallConv(func, LLVMCCallConv);
 
@@ -100,9 +97,10 @@ printv(char* string, v4sf value)
            f[0], f[1], f[2], f[3]);
 }
 
-static void
+static boolean
 compare(v4sf x, v4sf y)
 {
+   boolean success = TRUE;
    float *xp = (float *) &x;
    float *yp = (float *) &y;
    if (xp[0] != yp[0] ||
@@ -110,7 +108,9 @@ compare(v4sf x, v4sf y)
        xp[2] != yp[2] ||
        xp[3] != yp[3]) {
       printf(" Incorrect result! ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ \n");
+      success = FALSE;
    }
+   return success;
 }
 
 
@@ -171,9 +171,12 @@ test_round(unsigned verbose, FILE *fp)
       LLVMDumpModule(module);
 
    for (i = 0; i < 3; i++) {
+      /* NOTE: There are several acceptable rules for x.5 rounding: ceiling,
+       * nearest even, etc. So we avoid testing such corner cases here.
+       */
       v4sf xvals[3] = {
          {-10.0, -1, 0, 12.0},
-         {-1.5, -0.25, 1.25, 2.5},
+         {-1.49, -0.25, 1.25, 2.51},
          {-0.99, -0.01, 0.01, 0.99}
       };
       v4sf x = xvals[i];
@@ -191,7 +194,7 @@ test_round(unsigned verbose, FILE *fp)
       y = round_func(x);
       printv("C round(x)   ", ref);
       printv("LLVM round(x)", y);
-      compare(ref, y);
+      success = success && compare(ref, y);
 
       refp[0] = trunc(xp[0]);
       refp[1] = trunc(xp[1]);
@@ -200,7 +203,7 @@ test_round(unsigned verbose, FILE *fp)
       y = trunc_func(x);
       printv("C trunc(x)   ", ref);
       printv("LLVM trunc(x)", y);
-      compare(ref, y);
+      success = success && compare(ref, y);
 
       refp[0] = floor(xp[0]);
       refp[1] = floor(xp[1]);
@@ -209,7 +212,7 @@ test_round(unsigned verbose, FILE *fp)
       y = floor_func(x);
       printv("C floor(x)   ", ref);
       printv("LLVM floor(x)", y);
-      compare(ref, y);
+      success = success && compare(ref, y);
 
       refp[0] = ceil(xp[0]);
       refp[1] = ceil(xp[1]);
@@ -218,7 +221,7 @@ test_round(unsigned verbose, FILE *fp)
       y = ceil_func(x);
       printv("C ceil(x)    ", ref);
       printv("LLVM ceil(x) ", y);
-      compare(ref, y);
+      success = success && compare(ref, y);
    }
 
    LLVMFreeMachineCodeForFunction(engine, test_round);
@@ -247,11 +250,7 @@ test_round(unsigned verbose, FILE *fp)
 boolean
 test_all(unsigned verbose, FILE *fp)
 {
-   boolean success = TRUE;
-
-   test_round(verbose, fp);
-
-   return success;
+   return test_round(verbose, fp);
 }
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_test_sincos.c b/src/gallium/drivers/llvmpipe/lp_test_sincos.c
index 7ab357f162e..79939b1a393 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_sincos.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_sincos.c
@@ -72,10 +72,7 @@ add_sincos_test(LLVMModuleRef module, boolean sin)
    LLVMValueRef ret;
    struct lp_build_context bld;
 
-   bld.builder = builder;
-   bld.type.floating = 1;
-   bld.type.width = 32;
-   bld.type.length = 4;
+   lp_build_context_init(&bld, builder, lp_float32_vec4_type());
 
    LLVMSetFunctionCallConv(func, LLVMCCallConv);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_tile_soa.py b/src/gallium/drivers/llvmpipe/lp_tile_soa.py
index 2ba39052aba..8df7b236fe0 100644
--- a/src/gallium/drivers/llvmpipe/lp_tile_soa.py
+++ b/src/gallium/drivers/llvmpipe/lp_tile_soa.py
@@ -1,6 +1,6 @@
 #!/usr/bin/env python
 
-'''
+CopyRight = '''
 /**************************************************************************
  *
  * Copyright 2009 VMware, Inc.
@@ -289,172 +289,141 @@ def generate_format_write(format, src_channel, src_native_type, src_suffix):
     print
     
 
-def generate_ssse3():
+def generate_sse2():
     print '''
 #if defined(PIPE_ARCH_SSE)
 
 #include "util/u_sse.h"
 
-static void
-lp_tile_b8g8r8a8_unorm_swizzle_4ub_ssse3(uint8_t *dst,
-                                         const uint8_t *src, unsigned src_stride,
-                                         unsigned x0, unsigned y0)
+static ALWAYS_INLINE void 
+swz4( const __m128i * restrict x, 
+      const __m128i * restrict y, 
+      const __m128i * restrict z, 
+      const __m128i * restrict w, 
+      __m128i * restrict a, 
+      __m128i * restrict b, 
+      __m128i * restrict c, 
+      __m128i * restrict d)
+{
+   __m128i i, j, k, l;
+   __m128i m, n, o, p;
+   __m128i e, f, g, h;
+
+   m = _mm_unpacklo_epi8(*x,*y);
+   n = _mm_unpackhi_epi8(*x,*y);
+   o = _mm_unpacklo_epi8(*z,*w);
+   p = _mm_unpackhi_epi8(*z,*w);
+
+   i = _mm_unpacklo_epi16(m,n);
+   j = _mm_unpackhi_epi16(m,n);
+   k = _mm_unpacklo_epi16(o,p);
+   l = _mm_unpackhi_epi16(o,p);
+
+   e = _mm_unpacklo_epi8(i,j);
+   f = _mm_unpackhi_epi8(i,j);
+   g = _mm_unpacklo_epi8(k,l);
+   h = _mm_unpackhi_epi8(k,l);
+
+   *a = _mm_unpacklo_epi64(e,g);
+   *b = _mm_unpackhi_epi64(e,g);
+   *c = _mm_unpacklo_epi64(f,h);
+   *d = _mm_unpackhi_epi64(f,h);
+}
+
+static ALWAYS_INLINE void
+unswz4( const __m128i * restrict a, 
+        const __m128i * restrict b, 
+        const __m128i * restrict c, 
+        const __m128i * restrict d, 
+        __m128i * restrict x, 
+        __m128i * restrict y, 
+        __m128i * restrict z, 
+        __m128i * restrict w)
 {
+   __m128i i, j, k, l;
+   __m128i m, n, o, p;
+
+   i = _mm_unpacklo_epi8(*a,*b);
+   j = _mm_unpackhi_epi8(*a,*b);
+   k = _mm_unpacklo_epi8(*c,*d);
+   l = _mm_unpackhi_epi8(*c,*d);
+
+   m = _mm_unpacklo_epi16(i,k);
+   n = _mm_unpackhi_epi16(i,k);
+   o = _mm_unpacklo_epi16(j,l);
+   p = _mm_unpackhi_epi16(j,l);
+
+   *x = _mm_unpacklo_epi64(m,n);
+   *y = _mm_unpackhi_epi64(m,n);
+   *z = _mm_unpacklo_epi64(o,p);
+   *w = _mm_unpackhi_epi64(o,p);
+}
 
+static void
+lp_tile_b8g8r8a8_unorm_swizzle_4ub_sse2(uint8_t * restrict dst,
+                                        const uint8_t * restrict src, unsigned src_stride,
+                                        unsigned x0, unsigned y0)
+{
+   __m128i *dst128 = (__m128i *) dst;
    unsigned x, y;
-   __m128i *pdst = (__m128i*) dst;
-   const uint8_t *ysrc0 = src + y0*src_stride + x0*sizeof(uint32_t);
-   unsigned int tile_stridex = src_stride*(TILE_VECTOR_HEIGHT - 1) - sizeof(uint32_t)*TILE_VECTOR_WIDTH;
-   unsigned int tile_stridey = src_stride*TILE_VECTOR_HEIGHT;
-
-   const __m128i shuffle00 = _mm_setr_epi8(0x02,0x06,0xff,0xff,0x0a,0x0e,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
-   const __m128i shuffle01 = _mm_setr_epi8(0x01,0x05,0xff,0xff,0x09,0x0d,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
-   const __m128i shuffle02 = _mm_setr_epi8(0x00,0x04,0xff,0xff,0x08,0x0c,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
-   const __m128i shuffle03 = _mm_setr_epi8(0x03,0x07,0xff,0xff,0x0b,0x0f,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
-
-   const __m128i shuffle10 = _mm_setr_epi8(0xff,0xff,0x02,0x06,0xff,0xff,0x0a,0x0e,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
-   const __m128i shuffle11 = _mm_setr_epi8(0xff,0xff,0x01,0x05,0xff,0xff,0x09,0x0d,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
-   const __m128i shuffle12 = _mm_setr_epi8(0xff,0xff,0x00,0x04,0xff,0xff,0x08,0x0c,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
-   const __m128i shuffle13 = _mm_setr_epi8(0xff,0xff,0x03,0x07,0xff,0xff,0x0b,0x0f,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
-
-   const __m128i shuffle20 = _mm_setr_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0x02,0x06,0xff,0xff,0x0a,0x0e,0xff,0xff);
-   const __m128i shuffle21 = _mm_setr_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0x01,0x05,0xff,0xff,0x09,0x0d,0xff,0xff);
-   const __m128i shuffle22 = _mm_setr_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0x00,0x04,0xff,0xff,0x08,0x0c,0xff,0xff);
-   const __m128i shuffle23 = _mm_setr_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0x03,0x07,0xff,0xff,0x0b,0x0f,0xff,0xff);
-
-   const __m128i shuffle30 = _mm_setr_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0x02,0x06,0xff,0xff,0x0a,0x0e);
-   const __m128i shuffle31 = _mm_setr_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0x01,0x05,0xff,0xff,0x09,0x0d);
-   const __m128i shuffle32 = _mm_setr_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0x00,0x04,0xff,0xff,0x08,0x0c);
-   const __m128i shuffle33 = _mm_setr_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0x03,0x07,0xff,0xff,0x0b,0x0f);
-
-   for (y = 0; y < TILE_SIZE; y += TILE_VECTOR_HEIGHT) {
-      __m128i line0 = *(__m128i*)ysrc0;
-      const uint8_t *ysrc = ysrc0 + src_stride;
-      ysrc0 += tile_stridey;
-
-      for (x = 0; x < TILE_SIZE; x += TILE_VECTOR_WIDTH) {
-         __m128i r, g, b, a, line1;
-         line1 = *(__m128i*)ysrc;
-         PIPE_READ_WRITE_BARRIER();
-         ysrc += src_stride;
-         r = _mm_shuffle_epi8(line0, shuffle00);
-         g = _mm_shuffle_epi8(line0, shuffle01);
-         b = _mm_shuffle_epi8(line0, shuffle02);
-         a = _mm_shuffle_epi8(line0, shuffle03);
-
-         line0 = *(__m128i*)ysrc;
-         PIPE_READ_WRITE_BARRIER();
-         ysrc += src_stride;
-         r = _mm_or_si128(r, _mm_shuffle_epi8(line1, shuffle10));
-         g = _mm_or_si128(g, _mm_shuffle_epi8(line1, shuffle11));
-         b = _mm_or_si128(b, _mm_shuffle_epi8(line1, shuffle12));
-         a = _mm_or_si128(a, _mm_shuffle_epi8(line1, shuffle13));
-
-         line1 = *(__m128i*)ysrc;
-         PIPE_READ_WRITE_BARRIER();
-         ysrc -= tile_stridex;
-         r = _mm_or_si128(r, _mm_shuffle_epi8(line0, shuffle20));
-         g = _mm_or_si128(g, _mm_shuffle_epi8(line0, shuffle21));
-         b = _mm_or_si128(b, _mm_shuffle_epi8(line0, shuffle22));
-         a = _mm_or_si128(a, _mm_shuffle_epi8(line0, shuffle23));
-
-         if (x + 1 < TILE_SIZE) {
-            line0 = *(__m128i*)ysrc;
-            ysrc += src_stride;
-         }
-
-         PIPE_READ_WRITE_BARRIER();
-         r = _mm_or_si128(r, _mm_shuffle_epi8(line1, shuffle30));
-         g = _mm_or_si128(g, _mm_shuffle_epi8(line1, shuffle31));
-         b = _mm_or_si128(b, _mm_shuffle_epi8(line1, shuffle32));
-         a = _mm_or_si128(a, _mm_shuffle_epi8(line1, shuffle33));
-
-         *pdst++ = r;
-         *pdst++ = g;
-         *pdst++ = b;
-         *pdst++ = a;
+   
+   src += y0 * src_stride;
+   src += x0 * sizeof(uint32_t);
+
+   for (y = 0; y < TILE_SIZE; y += 4) {
+      const uint8_t *src_row = src;
+
+      for (x = 0; x < TILE_SIZE; x += 4) {
+         swz4((const __m128i *) (src_row + 0 * src_stride),
+              (const __m128i *) (src_row + 1 * src_stride),
+              (const __m128i *) (src_row + 2 * src_stride),
+              (const __m128i *) (src_row + 3 * src_stride),
+              dst128 + 2,     /* b */
+              dst128 + 1,     /* g */
+              dst128 + 0,     /* r */
+              dst128 + 3);    /* a */
+
+         dst128 += 4;
+         src_row += sizeof(__m128i);
       }
-   }
 
+      src += 4 * src_stride;
+   }
 }
 
 static void
-lp_tile_b8g8r8a8_unorm_unswizzle_4ub_ssse3(const uint8_t *src,
-                                          uint8_t *dst, unsigned dst_stride,
+lp_tile_b8g8r8a8_unorm_unswizzle_4ub_sse2(const uint8_t * restrict src,
+                                          uint8_t * restrict dst, unsigned dst_stride,
                                           unsigned x0, unsigned y0)
 {
    unsigned int x, y;
-   const __m128i *psrc = (__m128i*) src;
-   const __m128i *end = (__m128i*) (src + (y0 + TILE_SIZE - 1)*dst_stride + (x0 + TILE_SIZE - 1)*sizeof(uint32_t));
-   uint8_t *pdst = dst + y0 * dst_stride + x0 * sizeof(uint32_t);
-   __m128i c0 = *psrc++;
-   __m128i c1;
-
-   const __m128i shuffle00 = _mm_setr_epi8(0xff,0xff,0x00,0xff,0xff,0xff,0x01,0xff,0xff,0xff,0x04,0xff,0xff,0xff,0x05,0xff);
-   const __m128i shuffle01 = _mm_setr_epi8(0xff,0xff,0x02,0xff,0xff,0xff,0x03,0xff,0xff,0xff,0x06,0xff,0xff,0xff,0x07,0xff);
-   const __m128i shuffle02 = _mm_setr_epi8(0xff,0xff,0x08,0xff,0xff,0xff,0x09,0xff,0xff,0xff,0x0c,0xff,0xff,0xff,0x0d,0xff);
-   const __m128i shuffle03 = _mm_setr_epi8(0xff,0xff,0x0a,0xff,0xff,0xff,0x0b,0xff,0xff,0xff,0x0e,0xff,0xff,0xff,0x0f,0xff);
-
-   const __m128i shuffle10 = _mm_setr_epi8(0xff,0x00,0xff,0xff,0xff,0x01,0xff,0xff,0xff,0x04,0xff,0xff,0xff,0x05,0xff,0xff);
-   const __m128i shuffle11 = _mm_setr_epi8(0xff,0x02,0xff,0xff,0xff,0x03,0xff,0xff,0xff,0x06,0xff,0xff,0xff,0x07,0xff,0xff);
-   const __m128i shuffle12 = _mm_setr_epi8(0xff,0x08,0xff,0xff,0xff,0x09,0xff,0xff,0xff,0x0c,0xff,0xff,0xff,0x0d,0xff,0xff);
-   const __m128i shuffle13 = _mm_setr_epi8(0xff,0x0a,0xff,0xff,0xff,0x0b,0xff,0xff,0xff,0x0e,0xff,0xff,0xff,0x0f,0xff,0xff);
-
-   const __m128i shuffle20 = _mm_setr_epi8(0x00,0xff,0xff,0xff,0x01,0xff,0xff,0xff,0x04,0xff,0xff,0xff,0x05,0xff,0xff,0xff);
-   const __m128i shuffle21 = _mm_setr_epi8(0x02,0xff,0xff,0xff,0x03,0xff,0xff,0xff,0x06,0xff,0xff,0xff,0x07,0xff,0xff,0xff);
-   const __m128i shuffle22 = _mm_setr_epi8(0x08,0xff,0xff,0xff,0x09,0xff,0xff,0xff,0x0c,0xff,0xff,0xff,0x0d,0xff,0xff,0xff);
-   const __m128i shuffle23 = _mm_setr_epi8(0x0a,0xff,0xff,0xff,0x0b,0xff,0xff,0xff,0x0e,0xff,0xff,0xff,0x0f,0xff,0xff,0xff);
-
-   const __m128i shuffle30 = _mm_setr_epi8(0xff,0xff,0xff,0x00,0xff,0xff,0xff,0x01,0xff,0xff,0xff,0x04,0xff,0xff,0xff,0x05);
-   const __m128i shuffle31 = _mm_setr_epi8(0xff,0xff,0xff,0x02,0xff,0xff,0xff,0x03,0xff,0xff,0xff,0x06,0xff,0xff,0xff,0x07);
-   const __m128i shuffle32 = _mm_setr_epi8(0xff,0xff,0xff,0x08,0xff,0xff,0xff,0x09,0xff,0xff,0xff,0x0c,0xff,0xff,0xff,0x0d);
-   const __m128i shuffle33 = _mm_setr_epi8(0xff,0xff,0xff,0x0a,0xff,0xff,0xff,0x0b,0xff,0xff,0xff,0x0e,0xff,0xff,0xff,0x0f);
-
-   for (y = 0; y < TILE_SIZE; y += TILE_VECTOR_HEIGHT) {
-      __m128i *tile = (__m128i*) pdst;
-      pdst += dst_stride * TILE_VECTOR_HEIGHT;
-      for (x = 0; x < TILE_SIZE; x += TILE_VECTOR_WIDTH) {
-         uint8_t *linep = (uint8_t*) (tile++);
-         __m128i line0, line1, line2, line3;
-
-         c1 = *psrc++; /* r */
-         PIPE_READ_WRITE_BARRIER();
-         line0 = _mm_shuffle_epi8(c0, shuffle00);
-         line1 = _mm_shuffle_epi8(c0, shuffle01);
-         line2 = _mm_shuffle_epi8(c0, shuffle02);
-         line3 = _mm_shuffle_epi8(c0, shuffle03);
-
-         c0 = *psrc++; /* g */
-         PIPE_READ_WRITE_BARRIER();
-         line0 = _mm_or_si128(line0, _mm_shuffle_epi8(c1, shuffle10));
-         line1 = _mm_or_si128(line1, _mm_shuffle_epi8(c1, shuffle11));
-         line2 = _mm_or_si128(line2, _mm_shuffle_epi8(c1, shuffle12));
-         line3 = _mm_or_si128(line3, _mm_shuffle_epi8(c1, shuffle13));
-
-         c1 = *psrc++; /* b */
-         PIPE_READ_WRITE_BARRIER();
-         line0 = _mm_or_si128(line0, _mm_shuffle_epi8(c0, shuffle20));
-         line1 = _mm_or_si128(line1, _mm_shuffle_epi8(c0, shuffle21));
-         line2 = _mm_or_si128(line2, _mm_shuffle_epi8(c0, shuffle22));
-         line3 = _mm_or_si128(line3, _mm_shuffle_epi8(c0, shuffle23));
-
-         if (psrc != end)
-                 c0 = *psrc++; /* a */
-         PIPE_READ_WRITE_BARRIER();
-         line0 = _mm_or_si128(line0, _mm_shuffle_epi8(c1, shuffle30));
-         line1 = _mm_or_si128(line1, _mm_shuffle_epi8(c1, shuffle31));
-         line2 = _mm_or_si128(line2, _mm_shuffle_epi8(c1, shuffle32));
-         line3 = _mm_or_si128(line3, _mm_shuffle_epi8(c1, shuffle33));
-
-         *(__m128i*) (linep) = line0;
-         *(__m128i*) (((char*)linep) + dst_stride) = line1;
-         *(__m128i*) (((char*)linep) + 2 * dst_stride) = line2;
-         *(__m128i*) (((char*)linep) + 3 * dst_stride) = line3;
+   const __m128i *src128 = (const __m128i *) src;
+   
+   dst += y0 * dst_stride;
+   dst += x0 * sizeof(uint32_t);
+   
+   for (y = 0; y < TILE_SIZE; y += 4) {
+      const uint8_t *dst_row = dst;
+
+      for (x = 0; x < TILE_SIZE; x += 4) {
+         unswz4( &src128[2],     /* b */
+                 &src128[1],     /* g */
+                 &src128[0],     /* r */
+                 &src128[3],     /* a */
+                 (__m128i *) (dst_row + 0 * dst_stride),
+                 (__m128i *) (dst_row + 1 * dst_stride),
+                 (__m128i *) (dst_row + 2 * dst_stride),
+                 (__m128i *) (dst_row + 3 * dst_stride));
+
+         src128 += 4;
+         dst_row += sizeof(__m128i);;
       }
+
+      dst += 4 * dst_stride;
    }
 }
 
-#endif /* PIPE_ARCH_SSSE3 */
+#endif /* PIPE_ARCH_SSE */
 '''
 
 
@@ -479,7 +448,7 @@ def generate_swizzle(formats, dst_channel, dst_native_type, dst_suffix):
             func_name = 'lp_tile_%s_swizzle_%s' % (format.short_name(), dst_suffix)
             if format.name == 'PIPE_FORMAT_B8G8R8A8_UNORM':
                 print '#ifdef PIPE_ARCH_SSE'
-                print '      func = util_cpu_caps.has_ssse3 ? %s_ssse3 : %s;' % (func_name, func_name)
+                print '      func = util_cpu_caps.has_sse2 ? %s_sse2 : %s;' % (func_name, func_name)
                 print '#else'
                 print '      func = %s;' % (func_name,)
                 print '#endif'
@@ -517,7 +486,7 @@ def generate_unswizzle(formats, src_channel, src_native_type, src_suffix):
             func_name = 'lp_tile_%s_unswizzle_%s' % (format.short_name(), src_suffix)
             if format.name == 'PIPE_FORMAT_B8G8R8A8_UNORM':
                 print '#ifdef PIPE_ARCH_SSE'
-                print '      func = util_cpu_caps.has_ssse3 ? %s_ssse3 : %s;' % (func_name, func_name)
+                print '      func = util_cpu_caps.has_sse2 ? %s_sse2 : %s;' % (func_name, func_name)
                 print '#else'
                 print '      func = %s;' % (func_name,)
                 print '#endif'
@@ -541,7 +510,7 @@ def main():
     print '/* This file is autogenerated by lp_tile_soa.py from u_format.csv. Do not edit directly. */'
     print
     # This will print the copyright message on the top of this file
-    print __doc__.strip()
+    print CopyRight.strip()
     print
     print '#include "pipe/p_compiler.h"'
     print '#include "util/u_format.h"'
@@ -577,7 +546,7 @@ def main():
     print '};'
     print
 
-    generate_ssse3()
+    generate_sse2()
 
     channel = Channel(UNSIGNED, True, 8)
     native_type = 'uint8_t'
diff --git a/src/gallium/drivers/nouveau/nouveau_screen.c b/src/gallium/drivers/nouveau/nouveau_screen.c
index ebb21a6e5a3..a9426df686f 100644
--- a/src/gallium/drivers/nouveau/nouveau_screen.c
+++ b/src/gallium/drivers/nouveau/nouveau_screen.c
@@ -236,7 +236,7 @@ nouveau_screen_init(struct nouveau_screen *screen, struct nouveau_device *dev)
 	int ret;
 
 	ret = nouveau_channel_alloc(dev, 0xbeef0201, 0xbeef0202,
-				    &screen->channel);
+				    512*1024, &screen->channel);
 	if (ret)
 		return ret;
 	screen->device = dev;
diff --git a/src/gallium/drivers/nv50/nv50_context.h b/src/gallium/drivers/nv50/nv50_context.h
index ac69c7848e9..bf6a577188b 100644
--- a/src/gallium/drivers/nv50/nv50_context.h
+++ b/src/gallium/drivers/nv50/nv50_context.h
@@ -26,6 +26,9 @@
 #define NOUVEAU_MSG(fmt, args...) \
 	fprintf(stderr, "nouveau: "fmt, ##args);
 
+#define nouveau_bo_tile_layout(nvbo) \
+	((nvbo)->tile_flags & NOUVEAU_BO_TILE_LAYOUT_MASK)
+
 /* Constant buffer assignment */
 #define NV50_CB_PMISC		0
 #define NV50_CB_PVP		1
diff --git a/src/gallium/drivers/nv50/nv50_surface.c b/src/gallium/drivers/nv50/nv50_surface.c
index 3f3166261b1..f70c138fe1a 100644
--- a/src/gallium/drivers/nv50/nv50_surface.c
+++ b/src/gallium/drivers/nv50/nv50_surface.c
@@ -92,7 +92,7 @@ nv50_surface_set(struct nv50_screen *screen, struct pipe_surface *ps, int dst)
  		return 1;
 	}
 
- 	if (!bo->tile_flags) {
+	if (!nouveau_bo_tile_layout(bo)) {
  		BEGIN_RING(chan, eng2d, mthd, 2);
  		OUT_RING  (chan, format);
  		OUT_RING  (chan, 1);
diff --git a/src/gallium/drivers/nv50/nv50_transfer.c b/src/gallium/drivers/nv50/nv50_transfer.c
index f973cf24b98..0cc2f4a837f 100644
--- a/src/gallium/drivers/nv50/nv50_transfer.c
+++ b/src/gallium/drivers/nv50/nv50_transfer.c
@@ -45,7 +45,7 @@ nv50_transfer_rect_m2mf(struct pipe_screen *pscreen,
 
 	WAIT_RING (chan, 14);
 
-	if (!src_bo->tile_flags) {
+	if (!nouveau_bo_tile_layout(src_bo)) {
 		BEGIN_RING(chan, m2mf,
 			NV50_MEMORY_TO_MEMORY_FORMAT_LINEAR_IN, 1);
 		OUT_RING  (chan, 1);
@@ -64,7 +64,7 @@ nv50_transfer_rect_m2mf(struct pipe_screen *pscreen,
 		OUT_RING  (chan, sz); /* copying only 1 zslice per call */
 	}
 
-	if (!dst_bo->tile_flags) {
+	if (!nouveau_bo_tile_layout(dst_bo)) {
 		BEGIN_RING(chan, m2mf,
 			NV50_MEMORY_TO_MEMORY_FORMAT_LINEAR_OUT, 1);
 		OUT_RING  (chan, 1);
@@ -95,14 +95,14 @@ nv50_transfer_rect_m2mf(struct pipe_screen *pscreen,
 			NV04_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 2);
 		OUT_RELOCl(chan, src_bo, src_offset, src_reloc);
 		OUT_RELOCl(chan, dst_bo, dst_offset, dst_reloc);
-		if (src_bo->tile_flags) {
+		if (nouveau_bo_tile_layout(src_bo)) {
 			BEGIN_RING(chan, m2mf,
 				NV50_MEMORY_TO_MEMORY_FORMAT_TILING_POSITION_IN, 1);
 			OUT_RING  (chan, (sy << 16) | (sx * cpp));
 		} else {
 			src_offset += (line_count * src_pitch);
 		}
-		if (dst_bo->tile_flags) {
+		if (nouveau_bo_tile_layout(dst_bo)) {
 			BEGIN_RING(chan, m2mf,
 				NV50_MEMORY_TO_MEMORY_FORMAT_TILING_POSITION_OUT, 1);
 			OUT_RING  (chan, (dy << 16) | (dx * cpp));
@@ -280,7 +280,7 @@ nv50_upload_sifc(struct nv50_context *nv50,
 
 	MARK_RING (chan, 32, 2); /* flush on lack of space or relocs */
 
-	if (bo->tile_flags) {
+	if (nouveau_bo_tile_layout(bo)) {
 		BEGIN_RING(chan, eng2d, NV50_2D_DST_FORMAT, 5);
 		OUT_RING  (chan, dst_format);
 		OUT_RING  (chan, 0);
diff --git a/src/gallium/drivers/r300/r300_debug.c b/src/gallium/drivers/r300/r300_debug.c
index 145a7985da3..f78fe34790c 100644
--- a/src/gallium/drivers/r300/r300_debug.c
+++ b/src/gallium/drivers/r300/r300_debug.c
@@ -29,6 +29,7 @@
 static const struct debug_named_value debug_options[] = {
     { "fp", DBG_FP, "Log fragment program compilation" },
     { "vp", DBG_VP, "Log vertex program compilation" },
+    { "pstat", DBG_P_STAT, "Log vertex/fragment program stats" },
     { "draw", DBG_DRAW, "Log draw calls" },
     { "swtcl", DBG_SWTCL, "Log SWTCL-specific info" },
     { "rsblock", DBG_RS_BLOCK, "Log rasterizer registers" },
diff --git a/src/gallium/drivers/r300/r300_fs.c b/src/gallium/drivers/r300/r300_fs.c
index d9d4a9304df..c91532eb7b3 100644
--- a/src/gallium/drivers/r300/r300_fs.c
+++ b/src/gallium/drivers/r300/r300_fs.c
@@ -378,7 +378,8 @@ static void r300_translate_fragment_shader(
     /* Setup the compiler. */
     memset(&compiler, 0, sizeof(compiler));
     rc_init(&compiler.Base);
-    compiler.Base.Debug = DBG_ON(r300, DBG_FP);
+    DBG_ON(r300, DBG_FP) ? compiler.Base.Debug |= RC_DBG_LOG : 0;
+    DBG_ON(r300, DBG_P_STAT) ? compiler.Base.Debug |= RC_DBG_STATS : 0;
 
     compiler.code = &shader->code;
     compiler.state = shader->compare_state;
@@ -395,7 +396,7 @@ static void r300_translate_fragment_shader(
 
     find_output_registers(&compiler, shader);
 
-    if (compiler.Base.Debug) {
+    if (compiler.Base.Debug & RC_DBG_LOG) {
         DBG(r300, DBG_FP, "r300: Initial fragment program\n");
         tgsi_dump(tokens, 0);
     }
diff --git a/src/gallium/drivers/r300/r300_screen.c b/src/gallium/drivers/r300/r300_screen.c
index 7f41ff0e2ec..d445df408dd 100644
--- a/src/gallium/drivers/r300/r300_screen.c
+++ b/src/gallium/drivers/r300/r300_screen.c
@@ -124,6 +124,9 @@ static int r300_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
         case PIPE_CAP_INDEP_BLEND_FUNC:
         case PIPE_CAP_DEPTH_CLAMP: /* XXX implemented, but breaks Regnum Online */
         case PIPE_CAP_DEPTHSTENCIL_CLEAR_SEPARATE:
+        case PIPE_CAP_SHADER_STENCIL_EXPORT:
+        case PIPE_CAP_STREAM_OUTPUT:
+        case PIPE_CAP_PRIMITIVE_RESTART:
             return 0;
 
         /* Texturing. */
@@ -153,8 +156,8 @@ static int r300_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
         case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER:
             return 0;
         default:
-            fprintf(stderr, "r300: Implementation error: Bad param %d\n",
-                param);
+            debug_printf("r300: Warning: Unknown CAP %d in get_param.\n",
+                         param);
             return 0;
     }
 }
@@ -264,8 +267,8 @@ static float r300_get_paramf(struct pipe_screen* pscreen, enum pipe_cap param)
         case PIPE_CAP_MAX_TEXTURE_LOD_BIAS:
             return 16.0f;
         default:
-            fprintf(stderr, "r300: Implementation error: Bad paramf %d\n",
-                param);
+            debug_printf("r300: Warning: Unknown CAP %d in get_paramf.\n",
+                         param);
             return 0.0f;
     }
 }
diff --git a/src/gallium/drivers/r300/r300_screen.h b/src/gallium/drivers/r300/r300_screen.h
index dc2bc7e8279..8b7f1fab61b 100644
--- a/src/gallium/drivers/r300/r300_screen.h
+++ b/src/gallium/drivers/r300/r300_screen.h
@@ -102,6 +102,7 @@ r300_winsys_screen(struct pipe_screen *screen) {
 #define DBG_NO_CBZB     (1 << 21)
 /* Statistics. */
 #define DBG_STATS       (1 << 24)
+#define DBG_P_STAT      (1 << 25)
 /*@}*/
 
 static INLINE boolean SCREEN_DBG_ON(struct r300_screen * screen, unsigned flags)
diff --git a/src/gallium/drivers/r300/r300_state_derived.c b/src/gallium/drivers/r300/r300_state_derived.c
index 904736ef06d..50366e32c2b 100644
--- a/src/gallium/drivers/r300/r300_state_derived.c
+++ b/src/gallium/drivers/r300/r300_state_derived.c
@@ -620,14 +620,20 @@ static uint32_t r300_get_border_color(enum pipe_format format,
             }
             break;
 
+        case 16:
+            r = ((float_to_ubyte(border_swizzled[2]) & 0xff) << 0) |
+                ((float_to_ubyte(border_swizzled[1]) & 0xff) << 8) |
+                ((float_to_ubyte(border_swizzled[0]) & 0xff) << 16) |
+                ((float_to_ubyte(border_swizzled[3]) & 0xff) << 24);
+            break;
+
+        case 8:
         default:
-            /* I think the fat formats (16, 32) are specified
-             * as the 8-bit ones. I am not sure how compressed formats
-             * work here. */
             r = ((float_to_ubyte(border_swizzled[0]) & 0xff) << 0) |
                 ((float_to_ubyte(border_swizzled[1]) & 0xff) << 8) |
                 ((float_to_ubyte(border_swizzled[2]) & 0xff) << 16) |
                 ((float_to_ubyte(border_swizzled[3]) & 0xff) << 24);
+            break;
     }
 
     return r;
diff --git a/src/gallium/drivers/r300/r300_vs.c b/src/gallium/drivers/r300/r300_vs.c
index e2b9af9d018..65696555ac3 100644
--- a/src/gallium/drivers/r300/r300_vs.c
+++ b/src/gallium/drivers/r300/r300_vs.c
@@ -202,7 +202,8 @@ void r300_translate_vertex_shader(struct r300_context *r300,
     memset(&compiler, 0, sizeof(compiler));
     rc_init(&compiler.Base);
 
-    compiler.Base.Debug = DBG_ON(r300, DBG_VP);
+    DBG_ON(r300, DBG_VP) ? compiler.Base.Debug |= RC_DBG_LOG : 0;
+    DBG_ON(r300, DBG_P_STAT) ? compiler.Base.Debug |= RC_DBG_STATS : 0;
     compiler.code = &vs->code;
     compiler.UserData = vs;
     compiler.Base.is_r500 = r300->screen->caps.is_r500;
@@ -214,7 +215,7 @@ void r300_translate_vertex_shader(struct r300_context *r300,
     compiler.Base.max_alu_insts = r300->screen->caps.is_r500 ? 1024 : 256;
     compiler.Base.remove_unused_constants = TRUE;
 
-    if (compiler.Base.Debug) {
+    if (compiler.Base.Debug & RC_DBG_LOG) {
         DBG(r300, DBG_VP, "r300: Initial vertex program\n");
         tgsi_dump(vs->state.tokens, 0);
     }
diff --git a/src/gallium/drivers/r600/Makefile b/src/gallium/drivers/r600/Makefile
index 0dc9663de6a..436de9c4dbd 100644
--- a/src/gallium/drivers/r600/Makefile
+++ b/src/gallium/drivers/r600/Makefile
@@ -20,6 +20,8 @@ C_SOURCES = \
 	r600_video_context.c \
 	r700_asm.c \
 	evergreen_state.c \
-	eg_asm.c
+	eg_asm.c \
+	r600_translate.c \
+	r600_state_common.c
 
 include ../../Makefile.template
diff --git a/src/gallium/drivers/r600/SConscript b/src/gallium/drivers/r600/SConscript
index b6d9b7d8d22..bf0ad8571ba 100644
--- a/src/gallium/drivers/r600/SConscript
+++ b/src/gallium/drivers/r600/SConscript
@@ -18,6 +18,7 @@ r600 = env.ConvenienceLibrary(
     source = [
         'r600_asm.c',
         'r600_buffer.c',
+        'r600_blit.c',
         'r600_helper.c',
         'r600_pipe.c',
         'r600_query.c',
diff --git a/src/gallium/drivers/r600/eg_asm.c b/src/gallium/drivers/r600/eg_asm.c
index 52b7189e9e5..c30f09c394b 100644
--- a/src/gallium/drivers/r600/eg_asm.c
+++ b/src/gallium/drivers/r600/eg_asm.c
@@ -36,8 +36,13 @@ int eg_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf)
 	case (EG_V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3):
 	case (EG_V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3):
 		bc->bytecode[id++] = S_SQ_CF_ALU_WORD0_ADDR(cf->addr >> 1) |
-		  S_SQ_CF_ALU_WORD0_KCACHE_MODE0(cf->kcache0_mode);
+			S_SQ_CF_ALU_WORD0_KCACHE_MODE0(cf->kcache0_mode) |
+			S_SQ_CF_ALU_WORD0_KCACHE_BANK0(cf->kcache0_bank) |
+			S_SQ_CF_ALU_WORD0_KCACHE_BANK1(cf->kcache1_bank);
 		bc->bytecode[id++] = S_SQ_CF_ALU_WORD1_CF_INST(cf->inst >> 3) |
+			S_SQ_CF_ALU_WORD1_KCACHE_MODE1(cf->kcache1_mode) |
+			S_SQ_CF_ALU_WORD1_KCACHE_ADDR0(cf->kcache0_addr) |
+			S_SQ_CF_ALU_WORD1_KCACHE_ADDR1(cf->kcache1_addr) |
 					S_SQ_CF_ALU_WORD1_BARRIER(1) |
 					S_SQ_CF_ALU_WORD1_COUNT((cf->ndw / 2) - 1);
 		break;
diff --git a/src/gallium/drivers/r600/eg_state_inlines.h b/src/gallium/drivers/r600/eg_state_inlines.h
index 5a4e00aac38..be81c28b43f 100644
--- a/src/gallium/drivers/r600/eg_state_inlines.h
+++ b/src/gallium/drivers/r600/eg_state_inlines.h
@@ -25,6 +25,7 @@
 
 #include "util/u_format.h"
 #include "evergreend.h"
+#include "r600_formats.h"
 
 static INLINE uint32_t r600_translate_blend_function(int blend_func)
 {
@@ -523,32 +524,32 @@ static INLINE uint32_t r600_translate_vertex_data_type(enum pipe_format format)
 		case 16:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_030008_FMT_16_FLOAT;
+				result = FMT_16_FLOAT;
 				break;
 			case 2:
-				result = V_030008_FMT_16_16_FLOAT;
+				result = FMT_16_16_FLOAT;
 				break;
 			case 3:
-				result = V_030008_FMT_16_16_16_FLOAT;
+				result = FMT_16_16_16_FLOAT;
 				break;
 			case 4:
-				result = V_030008_FMT_16_16_16_16_FLOAT;
+				result = FMT_16_16_16_16_FLOAT;
 				break;
 			}
 			break;
 		case 32:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_030008_FMT_32_FLOAT;
+				result = FMT_32_FLOAT;
 				break;
 			case 2:
-				result = V_030008_FMT_32_32_FLOAT;
+				result = FMT_32_32_FLOAT;
 				break;
 			case 3:
-				result = V_030008_FMT_32_32_32_FLOAT;
+				result = FMT_32_32_32_FLOAT;
 				break;
 			case 4:
-				result = V_030008_FMT_32_32_32_32_FLOAT;
+				result = FMT_32_32_32_32_FLOAT;
 				break;
 			}
 			break;
@@ -564,48 +565,48 @@ static INLINE uint32_t r600_translate_vertex_data_type(enum pipe_format format)
 		case 8:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_030008_FMT_8;
+				result = FMT_8;
 				break;
 			case 2:
-				result = V_030008_FMT_8_8;
+				result = FMT_8_8;
 				break;
 			case 3:
 //				result = V_038008_FMT_8_8_8; /* fails piglit draw-vertices test */
 //				break;
 			case 4:
-				result = V_030008_FMT_8_8_8_8;
+				result = FMT_8_8_8_8;
 				break;
 			}
 			break;
 		case 16:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_030008_FMT_16;
+				result = FMT_16;
 				break;
 			case 2:
-				result = V_030008_FMT_16_16;
+				result = FMT_16_16;
 				break;
 			case 3:
 //				result = V_038008_FMT_16_16_16; /* fails piglit draw-vertices test */
 //				break;
 			case 4:
-				result = V_030008_FMT_16_16_16_16;
+				result = FMT_16_16_16_16;
 				break;
 			}
 			break;
 		case 32:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_030008_FMT_32;
+				result = FMT_32;
 				break;
 			case 2:
-				result = V_030008_FMT_32_32;
+				result = FMT_32_32;
 				break;
 			case 3:
-				result = V_030008_FMT_32_32_32;
+				result = FMT_32_32_32;
 				break;
 			case 4:
-				result = V_030008_FMT_32_32_32_32;
+				result = FMT_32_32_32_32;
 				break;
 			}
 			break;
diff --git a/src/gallium/drivers/r600/evergreen_state.c b/src/gallium/drivers/r600/evergreen_state.c
index 0fd1d399518..72223485067 100644
--- a/src/gallium/drivers/r600/evergreen_state.c
+++ b/src/gallium/drivers/r600/evergreen_state.c
@@ -39,6 +39,7 @@
 #include <util/u_pack_color.h>
 #include <util/u_memory.h>
 #include <util/u_inlines.h>
+#include <util/u_framebuffer.h>
 #include <pipebuffer/pb_buffer.h>
 #include "r600.h"
 #include "evergreend.h"
@@ -136,20 +137,6 @@ static void *evergreen_create_blend_state(struct pipe_context *ctx,
 	return rstate;
 }
 
-static void evergreen_bind_blend_state(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_blend *blend = (struct r600_pipe_blend *)state;
-	struct r600_pipe_state *rstate;
-
-	if (state == NULL)
-		return;
-	rstate = &blend->rstate;
-	rctx->states[rstate->id] = rstate;
-	rctx->cb_target_mask = blend->cb_target_mask;
-	r600_context_pipe_state_set(&rctx->ctx, rstate);
-}
-
 static void *evergreen_create_dsa_state(struct pipe_context *ctx,
 				   const struct pipe_depth_stencil_alpha_state *state)
 {
@@ -241,6 +228,7 @@ static void *evergreen_create_rs_state(struct pipe_context *ctx,
 	struct r600_pipe_state *rstate;
 	unsigned tmp;
 	unsigned prov_vtx = 1, polygon_dual_mode;
+	unsigned clip_rule;
 
 	if (rs == NULL) {
 		return NULL;
@@ -250,6 +238,8 @@ static void *evergreen_create_rs_state(struct pipe_context *ctx,
 	rs->flatshade = state->flatshade;
 	rs->sprite_coord_enable = state->sprite_coord_enable;
 
+	clip_rule = state->scissor ? 0xAAAA : 0xFFFF;
+
 	/* offset */
 	rs->offset_units = state->offset_units;
 	rs->offset_scale = state->offset_scale * 12.0f;
@@ -257,7 +247,7 @@ static void *evergreen_create_rs_state(struct pipe_context *ctx,
 	rstate->id = R600_PIPE_STATE_RASTERIZER;
 	if (state->flatshade_first)
 		prov_vtx = 0;
-	tmp = 0x00000001;
+	tmp = S_0286D4_FLAT_SHADE_ENA(1);
 	if (state->sprite_coord_enable) {
 		tmp |= S_0286D4_PNT_SPRITE_ENA(1) |
 			S_0286D4_PNT_SPRITE_OVRD_X(2) |
@@ -299,39 +289,10 @@ static void *evergreen_create_rs_state(struct pipe_context *ctx,
 	r600_pipe_state_add_reg(rstate, R_028C18_PA_CL_GB_HORZ_DISC_ADJ, 0x3F800000, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_028B7C_PA_SU_POLY_OFFSET_CLAMP, 0x0, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_028C08_PA_SU_VTX_CNTL, 0x00000005, 0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate, R_02820C_PA_SC_CLIPRECT_RULE, clip_rule, 0xFFFFFFFF, NULL);
 	return rstate;
 }
 
-static void evergreen_bind_rs_state(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_rasterizer *rs = (struct r600_pipe_rasterizer *)state;
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-
-	if (state == NULL)
-		return;
-
-	rctx->flatshade = rs->flatshade;
-	rctx->sprite_coord_enable = rs->sprite_coord_enable;
-	rctx->rasterizer = rs;
-
-	rctx->states[rs->rstate.id] = &rs->rstate;
-	r600_context_pipe_state_set(&rctx->ctx, &rs->rstate);
-}
-
-static void evergreen_delete_rs_state(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_rasterizer *rs = (struct r600_pipe_rasterizer *)state;
-
-	if (rctx->rasterizer == rs) {
-		rctx->rasterizer = NULL;
-	}
-	if (rctx->states[rs->rstate.id] == &rs->rstate) {
-		rctx->states[rs->rstate.id] = NULL;
-	}
-	free(rs);
-}
-
 static void *evergreen_create_sampler_state(struct pipe_context *ctx,
 					const struct pipe_sampler_state *state)
 {
@@ -372,28 +333,6 @@ static void *evergreen_create_sampler_state(struct pipe_context *ctx,
 	return rstate;
 }
 
-static void *evergreen_create_vertex_elements(struct pipe_context *ctx,
-				unsigned count,
-				const struct pipe_vertex_element *elements)
-{
-	struct r600_vertex_element *v = CALLOC_STRUCT(r600_vertex_element);
-
-	assert(count < 32);
-	v->count = count;
-	v->refcount = 1;
-	memcpy(v->elements, elements, count * sizeof(struct pipe_vertex_element));
-	return v;
-}
-
-static void evergreen_sampler_view_destroy(struct pipe_context *ctx,
-				      struct pipe_sampler_view *state)
-{
-	struct r600_pipe_sampler_view *resource = (struct r600_pipe_sampler_view *)state;
-
-	pipe_resource_reference(&state->texture, NULL);
-	FREE(resource);
-}
-
 static struct pipe_sampler_view *evergreen_create_sampler_view(struct pipe_context *ctx,
 							struct pipe_resource *texture,
 							const struct pipe_sampler_view *state)
@@ -446,7 +385,7 @@ static struct pipe_sampler_view *evergreen_create_sampler_view(struct pipe_conte
 		bo[0] = rbuffer->bo;
 		bo[1] = rbuffer->bo;
 	}
-	pitch = align(tmp->pitch[0] / tmp->bpt, 8);
+	pitch = align(tmp->pitch_in_pixels[0], 8);
 
 	/* FIXME properly handle first level != 0 */
 	r600_pipe_state_add_reg(rstate, R_030000_RESOURCE0_WORD0,
@@ -464,7 +403,6 @@ static struct pipe_sampler_view *evergreen_create_sampler_view(struct pipe_conte
 	r600_pipe_state_add_reg(rstate, R_030010_RESOURCE0_WORD4,
 				word4 | S_030010_NUM_FORMAT_ALL(V_030010_SQ_NUM_FORMAT_NORM) |
 				S_030010_SRF_MODE_ALL(V_030010_SFR_MODE_NO_ZERO) |
-				S_030010_REQUEST_SIZE(1) |
 				S_030010_BASE_LEVEL(state->first_level), 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_030014_RESOURCE0_WORD5,
 				S_030014_LAST_LEVEL(state->last_level) |
@@ -481,32 +419,42 @@ static struct pipe_sampler_view *evergreen_create_sampler_view(struct pipe_conte
 static void evergreen_set_vs_sampler_view(struct pipe_context *ctx, unsigned count,
 					struct pipe_sampler_view **views)
 {
-	/* TODO */
-	assert(1);
-}
-
-static void evergreen_set_ps_sampler_view(struct pipe_context *ctx, unsigned count,
-					struct pipe_sampler_view **views)
-{
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
 	struct r600_pipe_sampler_view **resource = (struct r600_pipe_sampler_view **)views;
 
 	for (int i = 0; i < count; i++) {
 		if (resource[i]) {
-			evergreen_context_pipe_state_set_ps_resource(&rctx->ctx, &resource[i]->state, i);
+			evergreen_context_pipe_state_set_vs_resource(&rctx->ctx, &resource[i]->state, i + PIPE_MAX_ATTRIBS);
 		}
 	}
 }
 
-static void evergreen_bind_state(struct pipe_context *ctx, void *state)
+static void evergreen_set_ps_sampler_view(struct pipe_context *ctx, unsigned count,
+					struct pipe_sampler_view **views)
 {
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_state *rstate = (struct r600_pipe_state *)state;
-
-	if (state == NULL)
-		return;
-	rctx->states[rstate->id] = rstate;
-	r600_context_pipe_state_set(&rctx->ctx, rstate);
+	struct r600_pipe_sampler_view **resource = (struct r600_pipe_sampler_view **)views;
+	int i;
+
+	for (i = 0; i < count; i++) {
+		if (&rctx->ps_samplers.views[i]->base != views[i]) {
+			if (resource[i])
+				evergreen_context_pipe_state_set_ps_resource(&rctx->ctx, &resource[i]->state, i);
+			else
+				evergreen_context_pipe_state_set_ps_resource(&rctx->ctx, NULL, i);
+
+			pipe_sampler_view_reference(
+				(struct pipe_sampler_view **)&rctx->ps_samplers.views[i],
+				views[i]);
+		}
+	}
+	for (i = count; i < NUM_TEX_UNITS; i++) {
+		if (rctx->ps_samplers.views[i]) {
+			evergreen_context_pipe_state_set_ps_resource(&rctx->ctx, NULL, i);
+			pipe_sampler_view_reference((struct pipe_sampler_view **)&rctx->ps_samplers.views[i], NULL);
+		}
+	}
+	rctx->ps_samplers.n_views = count;
 }
 
 static void evergreen_bind_ps_sampler(struct pipe_context *ctx, unsigned count, void **states)
@@ -514,6 +462,10 @@ static void evergreen_bind_ps_sampler(struct pipe_context *ctx, unsigned count,
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
 	struct r600_pipe_state **rstates = (struct r600_pipe_state **)states;
 
+
+	memcpy(rctx->ps_samplers.samplers, states, sizeof(void*) * count);
+	rctx->ps_samplers.n_samplers = count;
+
 	for (int i = 0; i < count; i++) {
 		evergreen_context_pipe_state_set_ps_sampler(&rctx->ctx, rstates[i], i);
 	}
@@ -524,37 +476,11 @@ static void evergreen_bind_vs_sampler(struct pipe_context *ctx, unsigned count,
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
 	struct r600_pipe_state **rstates = (struct r600_pipe_state **)states;
 
-	/* TODO implement */
 	for (int i = 0; i < count; i++) {
 		evergreen_context_pipe_state_set_vs_sampler(&rctx->ctx, rstates[i], i);
 	}
 }
 
-static void evergreen_delete_state(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_state *rstate = (struct r600_pipe_state *)state;
-
-	if (rctx->states[rstate->id] == rstate) {
-		rctx->states[rstate->id] = NULL;
-	}
-	for (int i = 0; i < rstate->nregs; i++) {
-		r600_bo_reference(rctx->radeon, &rstate->regs[i].bo, NULL);
-	}
-	free(rstate);
-}
-
-static void evergreen_delete_vertex_element(struct pipe_context *ctx, void *state)
-{
-	struct r600_vertex_element *v = (struct r600_vertex_element*)state;
-
-	if (v == NULL)
-		return;
-	if (--v->refcount)
-		return;
-	free(v);
-}
-
 static void evergreen_set_clip_state(struct pipe_context *ctx,
 				const struct pipe_clip_state *state)
 {
@@ -590,19 +516,6 @@ static void evergreen_set_clip_state(struct pipe_context *ctx,
 	r600_context_pipe_state_set(&rctx->ctx, rstate);
 }
 
-static void evergreen_bind_vertex_elements(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_vertex_element *v = (struct r600_vertex_element*)state;
-
-	evergreen_delete_vertex_element(ctx, rctx->vertex_elements);
-	rctx->vertex_elements = v;
-	if (v) {
-		v->refcount++;
-//		rctx->vs_rebuild = TRUE;
-	}
-}
-
 static void evergreen_set_polygon_stipple(struct pipe_context *ctx,
 					 const struct pipe_poly_stipple *state)
 {
@@ -626,18 +539,6 @@ static void evergreen_set_scissor_state(struct pipe_context *ctx,
 	tl = S_028240_TL_X(state->minx) | S_028240_TL_Y(state->miny);
 	br = S_028244_BR_X(state->maxx) | S_028244_BR_Y(state->maxy);
 	r600_pipe_state_add_reg(rstate,
-				R_028030_PA_SC_SCREEN_SCISSOR_TL, tl,
-				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_028034_PA_SC_SCREEN_SCISSOR_BR, br,
-				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_028204_PA_SC_WINDOW_SCISSOR_TL, tl,
-				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_028208_PA_SC_WINDOW_SCISSOR_BR, br,
-				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
 				R_028210_PA_SC_CLIPRECT_0_TL, tl,
 				0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate,
@@ -661,15 +562,6 @@ static void evergreen_set_scissor_state(struct pipe_context *ctx,
 	r600_pipe_state_add_reg(rstate,
 				R_02822C_PA_SC_CLIPRECT_3_BR, br,
 				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_028200_PA_SC_WINDOW_OFFSET, 0x00000000,
-				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_02820C_PA_SC_CLIPRECT_RULE, 0x0000FFFF,
-				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_028230_PA_SC_EDGERULE, 0xAAAAAAAA,
-				0xFFFFFFFF, NULL);
 
 	free(rctx->states[R600_PIPE_STATE_SCISSOR]);
 	rctx->states[R600_PIPE_STATE_SCISSOR] = rstate;
@@ -733,6 +625,7 @@ static void evergreen_cb(struct r600_pipe_context *rctx, struct r600_pipe_state
 {
 	struct r600_resource_texture *rtex;
 	struct r600_resource *rbuffer;
+	struct r600_surface *surf;
 	unsigned level = state->cbufs[cb]->level;
 	unsigned pitch, slice;
 	unsigned color_info;
@@ -740,14 +633,15 @@ static void evergreen_cb(struct r600_pipe_context *rctx, struct r600_pipe_state
 	const struct util_format_description *desc;
 	struct r600_bo *bo[3];
 
+	surf = (struct r600_surface *)state->cbufs[cb];
 	rtex = (struct r600_resource_texture*)state->cbufs[cb]->texture;
 	rbuffer = &rtex->resource;
 	bo[0] = rbuffer->bo;
 	bo[1] = rbuffer->bo;
 	bo[2] = rbuffer->bo;
 
-	pitch = (rtex->pitch[level] / rtex->bpt) / 8 - 1;
-	slice = (rtex->pitch[level] / rtex->bpt) * state->cbufs[cb]->height / 64 - 1;
+	pitch = rtex->pitch_in_pixels[level] / 8 - 1;
+	slice = rtex->pitch_in_pixels[level] * surf->aligned_height / 64 - 1;
 	ntype = 0;
 	desc = util_format_description(rtex->resource.base.b.format);
 	if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB)
@@ -794,22 +688,25 @@ static void evergreen_db(struct r600_pipe_context *rctx, struct r600_pipe_state
 {
 	struct r600_resource_texture *rtex;
 	struct r600_resource *rbuffer;
+	struct r600_surface *surf;
 	unsigned level;
 	unsigned pitch, slice, format, stencil_format;
 
 	if (state->zsbuf == NULL)
 		return;
 
+	level = state->zsbuf->level;
+
+	surf = (struct r600_surface *)state->zsbuf;
 	rtex = (struct r600_resource_texture*)state->zsbuf->texture;
 	rtex->tiled = 1;
-	rtex->array_mode = 2;
+	rtex->array_mode[level] = 2;
 	rtex->tile_type = 1;
 	rtex->depth = 1;
 	rbuffer = &rtex->resource;
 
-	level = state->zsbuf->level;
-	pitch = (rtex->pitch[level] / rtex->bpt) / 8 - 1;
-	slice = (rtex->pitch[level] / rtex->bpt) * state->zsbuf->height / 64 - 1;
+	pitch = rtex->pitch_in_pixels[level] / 8 - 1;
+	slice = rtex->pitch_in_pixels[level] * surf->aligned_height / 64 - 1;
 	format = r600_translate_dbformat(state->zsbuf->texture->format);
 	stencil_format = r600_translate_stencilformat(state->zsbuf->texture->format);
 
@@ -821,7 +718,7 @@ static void evergreen_db(struct r600_pipe_context *rctx, struct r600_pipe_state
 	if (stencil_format) {
 		uint32_t stencil_offset;
 
-		stencil_offset = ((state->zsbuf->height * rtex->pitch[level]) + 255) & ~255;
+		stencil_offset = ((surf->aligned_height * rtex->pitch_in_bytes[level]) + 255) & ~255;
 		r600_pipe_state_add_reg(rstate, R_02804C_DB_STENCIL_READ_BASE,
 					(state->zsbuf->offset + stencil_offset + r600_bo_offset(rbuffer->bo)) >> 8, 0xFFFFFFFF, rbuffer->bo);
 		r600_pipe_state_add_reg(rstate, R_028054_DB_STENCIL_WRITE_BASE,
@@ -833,7 +730,7 @@ static void evergreen_db(struct r600_pipe_context *rctx, struct r600_pipe_state
 				S_028044_FORMAT(stencil_format), 0xFFFFFFFF, rbuffer->bo);
 
 	r600_pipe_state_add_reg(rstate, R_028040_DB_Z_INFO,
-				S_028040_ARRAY_MODE(rtex->array_mode) | S_028040_FORMAT(format),
+				S_028040_ARRAY_MODE(rtex->array_mode[level]) | S_028040_FORMAT(format),
 				0xFFFFFFFF, rbuffer->bo);
 	r600_pipe_state_add_reg(rstate, R_028058_DB_DEPTH_SIZE,
 				S_028058_PITCH_TILE_MAX(pitch),
@@ -855,14 +752,9 @@ static void evergreen_set_framebuffer_state(struct pipe_context *ctx,
 
 	/* unreference old buffer and reference new one */
 	rstate->id = R600_PIPE_STATE_FRAMEBUFFER;
-	for (int i = 0; i < rctx->framebuffer.nr_cbufs; i++) {
-		pipe_surface_reference(&rctx->framebuffer.cbufs[i], NULL);
-	}
-	for (int i = 0; i < state->nr_cbufs; i++) {
-		pipe_surface_reference(&rctx->framebuffer.cbufs[i], state->cbufs[i]);
-	}
-	pipe_surface_reference(&rctx->framebuffer.zsbuf, state->zsbuf);
-	rctx->framebuffer = *state;
+
+	util_copy_framebuffer_state(&rctx->framebuffer, state);
+
 	rctx->pframebuffer = &rctx->framebuffer;
 
 	/* build states */
@@ -895,6 +787,24 @@ static void evergreen_set_framebuffer_state(struct pipe_context *ctx,
 	r600_pipe_state_add_reg(rstate,
 				R_028254_PA_SC_VPORT_SCISSOR_0_BR, br,
 				0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate,
+				R_028030_PA_SC_SCREEN_SCISSOR_TL, tl,
+				0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate,
+				R_028034_PA_SC_SCREEN_SCISSOR_BR, br,
+				0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate,
+				R_028204_PA_SC_WINDOW_SCISSOR_TL, tl,
+				0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate,
+				R_028208_PA_SC_WINDOW_SCISSOR_BR, br,
+				0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate,
+				R_028200_PA_SC_WINDOW_OFFSET, 0x00000000,
+				0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate,
+				R_028230_PA_SC_EDGERULE, 0xAAAAAAAA,
+				0xFFFFFFFF, NULL);
 
 	r600_pipe_state_add_reg(rstate, R_028238_CB_TARGET_MASK,
 				0x00000000, target_mask, NULL);
@@ -910,40 +820,6 @@ static void evergreen_set_framebuffer_state(struct pipe_context *ctx,
 	r600_context_pipe_state_set(&rctx->ctx, rstate);
 }
 
-static void evergreen_set_index_buffer(struct pipe_context *ctx,
-				  const struct pipe_index_buffer *ib)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-
-	if (ib) {
-		pipe_resource_reference(&rctx->index_buffer.buffer, ib->buffer);
-		memcpy(&rctx->index_buffer, ib, sizeof(rctx->index_buffer));
-	} else {
-		pipe_resource_reference(&rctx->index_buffer.buffer, NULL);
-		memset(&rctx->index_buffer, 0, sizeof(rctx->index_buffer));
-	}
-
-	/* TODO make this more like a state */
-}
-
-static void evergreen_set_vertex_buffers(struct pipe_context *ctx, unsigned count,
-					const struct pipe_vertex_buffer *buffers)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-
-	for (int i = 0; i < rctx->nvertex_buffer; i++) {
-		pipe_resource_reference(&rctx->vertex_buffer[i].buffer, NULL);
-	}
-	memcpy(rctx->vertex_buffer, buffers, sizeof(struct pipe_vertex_buffer) * count);
-	for (int i = 0; i < count; i++) {
-		rctx->vertex_buffer[i].buffer = NULL;
-		if (r600_buffer_is_user_buffer(buffers[i].buffer))
-			rctx->any_user_vbs = TRUE;
-		pipe_resource_reference(&rctx->vertex_buffer[i].buffer, buffers[i].buffer);
-	}
-	rctx->nvertex_buffer = count;
-}
-
 static void evergreen_set_constant_buffer(struct pipe_context *ctx, uint shader, uint index,
 					struct pipe_resource *buffer)
 {
@@ -979,84 +855,31 @@ static void evergreen_set_constant_buffer(struct pipe_context *ctx, uint shader,
 	}
 }
 
-static void *evergreen_create_shader_state(struct pipe_context *ctx,
-					const struct pipe_shader_state *state)
-{
-	struct r600_pipe_shader *shader =  CALLOC_STRUCT(r600_pipe_shader);
-	int r;
-
-	r =  r600_pipe_shader_create(ctx, shader, state->tokens);
-	if (r) {
-		return NULL;
-	}
-	return shader;
-}
-
-static void evergreen_bind_ps_shader(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-
-	/* TODO delete old shader */
-	rctx->ps_shader = (struct r600_pipe_shader *)state;
-}
-
-static void evergreen_bind_vs_shader(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-
-	/* TODO delete old shader */
-	rctx->vs_shader = (struct r600_pipe_shader *)state;
-}
-
-static void evergreen_delete_ps_shader(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_shader *shader = (struct r600_pipe_shader *)state;
-
-	if (rctx->ps_shader == shader) {
-		rctx->ps_shader = NULL;
-	}
-	/* TODO proper delete */
-	free(shader);
-}
-
-static void evergreen_delete_vs_shader(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_shader *shader = (struct r600_pipe_shader *)state;
-
-	if (rctx->vs_shader == shader) {
-		rctx->vs_shader = NULL;
-	}
-	/* TODO proper delete */
-	free(shader);
-}
-
 void evergreen_init_state_functions(struct r600_pipe_context *rctx)
 {
 	rctx->context.create_blend_state = evergreen_create_blend_state;
 	rctx->context.create_depth_stencil_alpha_state = evergreen_create_dsa_state;
-	rctx->context.create_fs_state = evergreen_create_shader_state;
+	rctx->context.create_fs_state = r600_create_shader_state;
 	rctx->context.create_rasterizer_state = evergreen_create_rs_state;
 	rctx->context.create_sampler_state = evergreen_create_sampler_state;
 	rctx->context.create_sampler_view = evergreen_create_sampler_view;
-	rctx->context.create_vertex_elements_state = evergreen_create_vertex_elements;
-	rctx->context.create_vs_state = evergreen_create_shader_state;
-	rctx->context.bind_blend_state = evergreen_bind_blend_state;
-	rctx->context.bind_depth_stencil_alpha_state = evergreen_bind_state;
+	rctx->context.create_vertex_elements_state = r600_create_vertex_elements;
+	rctx->context.create_vs_state = r600_create_shader_state;
+	rctx->context.bind_blend_state = r600_bind_blend_state;
+	rctx->context.bind_depth_stencil_alpha_state = r600_bind_state;
 	rctx->context.bind_fragment_sampler_states = evergreen_bind_ps_sampler;
-	rctx->context.bind_fs_state = evergreen_bind_ps_shader;
-	rctx->context.bind_rasterizer_state = evergreen_bind_rs_state;
-	rctx->context.bind_vertex_elements_state = evergreen_bind_vertex_elements;
+	rctx->context.bind_fs_state = r600_bind_ps_shader;
+	rctx->context.bind_rasterizer_state = r600_bind_rs_state;
+	rctx->context.bind_vertex_elements_state = r600_bind_vertex_elements;
 	rctx->context.bind_vertex_sampler_states = evergreen_bind_vs_sampler;
-	rctx->context.bind_vs_state = evergreen_bind_vs_shader;
-	rctx->context.delete_blend_state = evergreen_delete_state;
-	rctx->context.delete_depth_stencil_alpha_state = evergreen_delete_state;
-	rctx->context.delete_fs_state = evergreen_delete_ps_shader;
-	rctx->context.delete_rasterizer_state = evergreen_delete_rs_state;
-	rctx->context.delete_sampler_state = evergreen_delete_state;
-	rctx->context.delete_vertex_elements_state = evergreen_delete_vertex_element;
-	rctx->context.delete_vs_state = evergreen_delete_vs_shader;
+	rctx->context.bind_vs_state = r600_bind_vs_shader;
+	rctx->context.delete_blend_state = r600_delete_state;
+	rctx->context.delete_depth_stencil_alpha_state = r600_delete_state;
+	rctx->context.delete_fs_state = r600_delete_ps_shader;
+	rctx->context.delete_rasterizer_state = r600_delete_rs_state;
+	rctx->context.delete_sampler_state = r600_delete_state;
+	rctx->context.delete_vertex_elements_state = r600_delete_vertex_element;
+	rctx->context.delete_vs_state = r600_delete_vs_shader;
 	rctx->context.set_blend_color = evergreen_set_blend_color;
 	rctx->context.set_clip_state = evergreen_set_clip_state;
 	rctx->context.set_constant_buffer = evergreen_set_constant_buffer;
@@ -1066,11 +889,11 @@ void evergreen_init_state_functions(struct r600_pipe_context *rctx)
 	rctx->context.set_sample_mask = evergreen_set_sample_mask;
 	rctx->context.set_scissor_state = evergreen_set_scissor_state;
 	rctx->context.set_stencil_ref = evergreen_set_stencil_ref;
-	rctx->context.set_vertex_buffers = evergreen_set_vertex_buffers;
-	rctx->context.set_index_buffer = evergreen_set_index_buffer;
+	rctx->context.set_vertex_buffers = r600_set_vertex_buffers;
+	rctx->context.set_index_buffer = r600_set_index_buffer;
 	rctx->context.set_vertex_sampler_views = evergreen_set_vs_sampler_view;
 	rctx->context.set_viewport_state = evergreen_set_viewport_state;
-	rctx->context.sampler_view_destroy = evergreen_sampler_view_destroy;
+	rctx->context.sampler_view_destroy = r600_sampler_view_destroy;
 }
 
 void evergreen_init_config(struct r600_pipe_context *rctx)
@@ -1353,6 +1176,12 @@ void evergreen_draw(struct pipe_context *ctx, const struct pipe_draw_info *info)
 	struct r600_draw rdraw;
 	struct r600_pipe_state vgt;
 	struct r600_drawl draw;
+	boolean translate = FALSE;
+
+	if (rctx->vertex_elements->incompatible_layout) {
+		r600_begin_vertex_translate(rctx);
+		translate = TRUE;
+	}
 
 	if (rctx->any_user_vbs) {
 		r600_upload_user_buffers(rctx);
@@ -1429,11 +1258,11 @@ void evergreen_draw(struct pipe_context *ctx, const struct pipe_draw_info *info)
 			vertex_buffer->buffer_offset +
 			r600_bo_offset(rbuffer->bo);
 
-		format = r600_translate_vertex_data_type(rctx->vertex_elements->elements[i].src_format);
+		format = r600_translate_vertex_data_type(rctx->vertex_elements->hw_format[i]);
 
 		word2 = format | S_030008_STRIDE(vertex_buffer->stride);
 
-		word3 = r600_translate_vertex_data_swizzle(rctx->vertex_elements->elements[i].src_format);
+		word3 = r600_translate_vertex_data_swizzle(rctx->vertex_elements->hw_format[i]);
 
 		r600_pipe_state_add_reg(rstate, R_030000_RESOURCE0_WORD0, offset, 0xFFFFFFFF, rbuffer->bo);
 		r600_pipe_state_add_reg(rstate, R_030004_RESOURCE0_WORD1, rbuffer->size - offset - 1, 0xFFFFFFFF, NULL);
@@ -1514,6 +1343,9 @@ void evergreen_draw(struct pipe_context *ctx, const struct pipe_draw_info *info)
 	}
 	evergreen_context_draw(&rctx->ctx, &rdraw);
 
+	if (translate)
+		r600_end_vertex_translate(rctx);
+
 	pipe_resource_reference(&draw.index_buffer, NULL);
 }
 
@@ -1522,23 +1354,39 @@ void evergreen_pipe_shader_ps(struct pipe_context *ctx, struct r600_pipe_shader
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
 	struct r600_pipe_state *rstate = &shader->rstate;
 	struct r600_shader *rshader = &shader->shader;
-	unsigned i, tmp, exports_ps, num_cout, spi_ps_in_control_0, spi_input_z;
-	boolean have_pos = FALSE, have_face = FALSE;
+	unsigned i, tmp, exports_ps, num_cout, spi_ps_in_control_0, spi_input_z, spi_ps_in_control_1;
+	int pos_index = -1, face_index = -1;
+	int ninterp = 0;
+	boolean have_linear = FALSE, have_centroid = FALSE, have_perspective = FALSE;
+	unsigned spi_baryc_cntl;
 
 	/* clear previous register */
 	rstate->nregs = 0;
 
 	for (i = 0; i < rshader->ninput; i++) {
 		tmp = S_028644_SEMANTIC(r600_find_vs_semantic_index(&rctx->vs_shader->shader, rshader, i));
+		/* evergreen NUM_INTERP only contains values interpolated into the LDS,
+		   POSITION goes via GPRs from the SC so isn't counted */
 		if (rshader->input[i].name == TGSI_SEMANTIC_POSITION)
-			have_pos = TRUE;
+			pos_index = i;
+		else if (rshader->input[i].name == TGSI_SEMANTIC_FACE)
+			face_index = i;
+		else {
+			if (rshader->input[i].interpolate == TGSI_INTERPOLATE_LINEAR ||
+			    rshader->input[i].interpolate == TGSI_INTERPOLATE_PERSPECTIVE)
+				ninterp++;
+			if (rshader->input[i].interpolate == TGSI_INTERPOLATE_LINEAR)
+				have_linear = TRUE;
+			if (rshader->input[i].interpolate == TGSI_INTERPOLATE_PERSPECTIVE)
+				have_perspective = TRUE;
+			if (rshader->input[i].centroid)
+				have_centroid = TRUE;
+		}
 		if (rshader->input[i].name == TGSI_SEMANTIC_COLOR ||
 		    rshader->input[i].name == TGSI_SEMANTIC_BCOLOR ||
 		    rshader->input[i].name == TGSI_SEMANTIC_POSITION) {
 			tmp |= S_028644_FLAT_SHADE(rshader->flat_shade);
 		}
-		if (rshader->input[i].name == TGSI_SEMANTIC_FACE)
-			have_face = TRUE;
 		if (rshader->input[i].name == TGSI_SEMANTIC_GENERIC &&
 			rctx->sprite_coord_enable & (1 << rshader->input[i].sid)) {
 			tmp |= S_028644_PT_SPRITE_TEX(1);
@@ -1546,17 +1394,23 @@ void evergreen_pipe_shader_ps(struct pipe_context *ctx, struct r600_pipe_shader
 		r600_pipe_state_add_reg(rstate, R_028644_SPI_PS_INPUT_CNTL_0 + i * 4, tmp, 0xFFFFFFFF, NULL);
 	}
 	for (i = 0; i < rshader->noutput; i++) {
-		if (rshader->input[i].name == TGSI_SEMANTIC_POSITION)
+		if (rshader->output[i].name == TGSI_SEMANTIC_POSITION)
 			r600_pipe_state_add_reg(rstate,
 						R_02880C_DB_SHADER_CONTROL,
 						S_02880C_Z_EXPORT_ENABLE(1),
 						S_02880C_Z_EXPORT_ENABLE(1), NULL);
+		if (rshader->output[i].name == TGSI_SEMANTIC_STENCIL)
+			r600_pipe_state_add_reg(rstate,
+						R_02880C_DB_SHADER_CONTROL,
+						S_02880C_STENCIL_EXPORT_ENABLE(1),
+						S_02880C_STENCIL_EXPORT_ENABLE(1), NULL);
 	}
 
 	exports_ps = 0;
 	num_cout = 0;
 	for (i = 0; i < rshader->noutput; i++) {
-		if (rshader->output[i].name == TGSI_SEMANTIC_POSITION)
+		if (rshader->output[i].name == TGSI_SEMANTIC_POSITION ||
+		    rshader->output[i].name == TGSI_SEMANTIC_STENCIL)
 			exports_ps |= 1;
 		else if (rshader->output[i].name == TGSI_SEMANTIC_COLOR) {
 			num_cout++;
@@ -1568,19 +1422,49 @@ void evergreen_pipe_shader_ps(struct pipe_context *ctx, struct r600_pipe_shader
 		exports_ps = 2;
 	}
 
-	spi_ps_in_control_0 = S_0286CC_NUM_INTERP(rshader->ninput) |
-				S_0286CC_PERSP_GRADIENT_ENA(1);
+	if (ninterp == 0) {
+		ninterp = 1;
+		have_perspective = TRUE;
+	}
+
+	spi_ps_in_control_0 = S_0286CC_NUM_INTERP(ninterp) |
+		              S_0286CC_PERSP_GRADIENT_ENA(have_perspective) |
+		              S_0286CC_LINEAR_GRADIENT_ENA(have_linear);
 	spi_input_z = 0;
-	if (have_pos) {
-		spi_ps_in_control_0 |=  S_0286CC_POSITION_ENA(1);
+	if (pos_index != -1) {
+		spi_ps_in_control_0 |=  S_0286CC_POSITION_ENA(1) |
+			S_0286CC_POSITION_CENTROID(rshader->input[pos_index].centroid) |
+			S_0286CC_POSITION_ADDR(rshader->input[pos_index].gpr);
 		spi_input_z |= 1;
 	}
+
+	spi_ps_in_control_1 = 0;
+	if (face_index != -1) {
+		spi_ps_in_control_1 |= S_0286D0_FRONT_FACE_ENA(1) |
+			S_0286D0_FRONT_FACE_ADDR(rshader->input[face_index].gpr);
+	}
+
+	spi_baryc_cntl = 0;
+	if (have_perspective)
+		spi_baryc_cntl |= S_0286E0_PERSP_CENTER_ENA(1) |
+				  S_0286E0_PERSP_CENTROID_ENA(have_centroid);
+	if (have_linear)
+		spi_baryc_cntl |= S_0286E0_LINEAR_CENTER_ENA(1) |
+			          S_0286E0_LINEAR_CENTROID_ENA(have_centroid);
+				
 	r600_pipe_state_add_reg(rstate, R_0286CC_SPI_PS_IN_CONTROL_0,
 				spi_ps_in_control_0, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_0286D0_SPI_PS_IN_CONTROL_1,
-				S_0286D0_FRONT_FACE_ENA(have_face), 0xFFFFFFFF, NULL);
+				spi_ps_in_control_1, 0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate, R_0286E4_SPI_PS_IN_CONTROL_2,
+				0, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_0286D8_SPI_INPUT_Z, spi_input_z, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate,
+				R_0286E0_SPI_BARYC_CNTL,
+				spi_baryc_cntl,
+				0xFFFFFFFF, NULL);
+
+	r600_pipe_state_add_reg(rstate,
 				R_028840_SQ_PGM_START_PS,
 				(r600_bo_offset(shader->bo)) >> 8, 0xFFFFFFFF, shader->bo);
 	r600_pipe_state_add_reg(rstate,
@@ -1595,11 +1479,6 @@ void evergreen_pipe_shader_ps(struct pipe_context *ctx, struct r600_pipe_shader
 	r600_pipe_state_add_reg(rstate,
 				R_02884C_SQ_PGM_EXPORTS_PS,
 				exports_ps, 0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_0286E0_SPI_BARYC_CNTL,
-				S_0286E0_PERSP_CENTROID_ENA(1) |
-				S_0286E0_LINEAR_CENTROID_ENA(1),
-				0xFFFFFFFF, NULL);
 
 	if (rshader->uses_kill) {
 		/* only set some bits here, the other bits are set in the dsa state */
diff --git a/src/gallium/drivers/r600/evergreend.h b/src/gallium/drivers/r600/evergreend.h
index 9971dded782..8e96f9355e6 100644
--- a/src/gallium/drivers/r600/evergreend.h
+++ b/src/gallium/drivers/r600/evergreend.h
@@ -686,6 +686,9 @@
 #define   S_02880C_Z_EXPORT_ENABLE(x)                  (((x) & 0x1) << 0)
 #define   G_02880C_Z_EXPORT_ENABLE(x)                  (((x) >> 0) & 0x1)
 #define   C_02880C_Z_EXPORT_ENABLE                     0xFFFFFFFE
+#define   S_02880C_STENCIL_EXPORT_ENABLE(x)            (((x) & 0x1) << 1)
+#define   G_02880C_STENCIL_EXPORT_ENABLE(x)            (((x) >> 1) & 0x1)
+#define   C_02880C_STENCIL_EXPORT_ENABLE               0xFFFFFFFD
 #define   S_02880C_Z_ORDER(x)                          (((x) & 0x3) << 4)
 #define   G_02880C_Z_ORDER(x)                          (((x) >> 4) & 0x3)
 #define   C_02880C_Z_ORDER                             0xFFFFFCFF
@@ -984,9 +987,6 @@
 #define   S_030010_ENDIAN_SWAP(x)                      (((x) & 0x3) << 12)
 #define   G_030010_ENDIAN_SWAP(x)                      (((x) >> 12) & 0x3)
 #define   C_030010_ENDIAN_SWAP                         0xFFFFCFFF
-#define   S_030010_REQUEST_SIZE(x)                     (((x) & 0x3) << 14)
-#define   G_030010_REQUEST_SIZE(x)                     (((x) >> 14) & 0x3)
-#define   C_030010_REQUEST_SIZE                        0xFFFF3FFF
 #define   S_030010_DST_SEL_X(x)                        (((x) & 0x7) << 16)
 #define   G_030010_DST_SEL_X(x)                        (((x) >> 16) & 0x7)
 #define   C_030010_DST_SEL_X                           0xFFF8FFFF
@@ -1050,45 +1050,6 @@
 #define   S_030008_DATA_FORMAT(x)                      (((x) & 0x3F) << 20)
 #define   G_030008_DATA_FORMAT(x)                      (((x) >> 20) & 0x3F)
 #define   C_030008_DATA_FORMAT                         0xFC0FFFFF
-#define     V_030008_FMT_INVALID                     0x00000000
-#define     V_030008_FMT_8                           0x00000001
-#define     V_030008_FMT_4_4                         0x00000002
-#define     V_030008_FMT_3_3_2                       0x00000003
-#define     V_030008_FMT_16                          0x00000005
-#define     V_030008_FMT_16_FLOAT                    0x00000006
-#define     V_030008_FMT_8_8                         0x00000007
-#define     V_030008_FMT_5_6_5                       0x00000008
-#define     V_030008_FMT_6_5_5                       0x00000009
-#define     V_030008_FMT_1_5_5_5                     0x0000000A
-#define     V_030008_FMT_4_4_4_4                     0x0000000B
-#define     V_030008_FMT_5_5_5_1                     0x0000000C
-#define     V_030008_FMT_32                          0x0000000D
-#define     V_030008_FMT_32_FLOAT                    0x0000000E
-#define     V_030008_FMT_16_16                       0x0000000F
-#define     V_030008_FMT_16_16_FLOAT                 0x00000010
-#define     V_030008_FMT_8_24                        0x00000011
-#define     V_030008_FMT_8_24_FLOAT                  0x00000012
-#define     V_030008_FMT_24_8                        0x00000013
-#define     V_030008_FMT_24_8_FLOAT                  0x00000014
-#define     V_030008_FMT_10_11_11                    0x00000015
-#define     V_030008_FMT_10_11_11_FLOAT              0x00000016
-#define     V_030008_FMT_11_11_10                    0x00000017
-#define     V_030008_FMT_11_11_10_FLOAT              0x00000018
-#define     V_030008_FMT_2_10_10_10                  0x00000019
-#define     V_030008_FMT_8_8_8_8                     0x0000001A
-#define     V_030008_FMT_10_10_10_2                  0x0000001B
-#define     V_030008_FMT_X24_8_32_FLOAT              0x0000001C
-#define     V_030008_FMT_32_32                       0x0000001D
-#define     V_030008_FMT_32_32_FLOAT                 0x0000001E
-#define     V_030008_FMT_16_16_16_16                 0x0000001F
-#define     V_030008_FMT_16_16_16_16_FLOAT           0x00000020
-#define     V_030008_FMT_32_32_32_32                 0x00000022
-#define     V_030008_FMT_32_32_32_32_FLOAT           0x00000023
-#define     V_030008_FMT_8_8_8                       0x0000002c
-#define     V_030008_FMT_16_16_16                    0x0000002d
-#define     V_030008_FMT_16_16_16_FLOAT              0x0000002e
-#define     V_030008_FMT_32_32_32                    0x0000002f
-#define     V_030008_FMT_32_32_32_FLOAT              0x00000030
 #define   S_030008_NUM_FORMAT_ALL(x)                   (((x) & 0x3) << 26)
 #define   G_030008_NUM_FORMAT_ALL(x)                   (((x) >> 26) & 0x3)
 #define   C_030008_NUM_FORMAT_ALL                      0xF3FFFFFF
diff --git a/src/gallium/drivers/r600/r600.h b/src/gallium/drivers/r600/r600.h
index 24e25cec0db..62d983269f5 100644
--- a/src/gallium/drivers/r600/r600.h
+++ b/src/gallium/drivers/r600/r600.h
@@ -99,15 +99,22 @@ enum chip_class {
 	EVERGREEN,
 };
 
+struct r600_tiling_info {
+	unsigned num_channels;
+	unsigned num_banks;
+	unsigned group_bytes;
+};
+
 enum radeon_family r600_get_family(struct radeon *rw);
 enum chip_class r600_get_family_class(struct radeon *radeon);
+struct r600_tiling_info *r600_get_tiling_info(struct radeon *radeon);
 
 /* r600_bo.c */
 struct r600_bo;
 struct r600_bo *r600_bo(struct radeon *radeon,
 				  unsigned size, unsigned alignment, unsigned usage);
 struct r600_bo *r600_bo_handle(struct radeon *radeon,
-					 unsigned handle);
+			       unsigned handle, unsigned *array_mode);
 void *r600_bo_map(struct radeon *radeon, struct r600_bo *bo, unsigned usage, void *ctx);
 void r600_bo_unmap(struct radeon *radeon, struct r600_bo *bo);
 void r600_bo_reference(struct radeon *radeon, struct r600_bo **dst,
diff --git a/src/gallium/drivers/r600/r600_asm.c b/src/gallium/drivers/r600/r600_asm.c
index d13da0ef638..8a7f3ce575c 100644
--- a/src/gallium/drivers/r600/r600_asm.c
+++ b/src/gallium/drivers/r600/r600_asm.c
@@ -701,9 +701,14 @@ static int r600_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf)
 	case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU << 3):
 	case (V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU_PUSH_BEFORE << 3):
 		bc->bytecode[id++] = S_SQ_CF_ALU_WORD0_ADDR(cf->addr >> 1) |
-			S_SQ_CF_ALU_WORD0_KCACHE_MODE0(cf->kcache0_mode);
+			S_SQ_CF_ALU_WORD0_KCACHE_MODE0(cf->kcache0_mode) |
+			S_SQ_CF_ALU_WORD0_KCACHE_BANK0(cf->kcache0_bank) |
+			S_SQ_CF_ALU_WORD0_KCACHE_BANK1(cf->kcache1_bank);
 
 		bc->bytecode[id++] = S_SQ_CF_ALU_WORD1_CF_INST(cf->inst >> 3) |
+			S_SQ_CF_ALU_WORD1_KCACHE_MODE1(cf->kcache1_mode) |
+			S_SQ_CF_ALU_WORD1_KCACHE_ADDR0(cf->kcache0_addr) |
+			S_SQ_CF_ALU_WORD1_KCACHE_ADDR1(cf->kcache1_addr) |
 					S_SQ_CF_ALU_WORD1_BARRIER(1) |
 					S_SQ_CF_ALU_WORD1_USES_WATERFALL(bc->chiprev == 0 ? cf->r6xx_uses_waterfall : 0) |
 					S_SQ_CF_ALU_WORD1_COUNT((cf->ndw / 2) - 1);
@@ -871,3 +876,39 @@ int r600_bc_build(struct r600_bc *bc)
 	}
 	return 0;
 }
+
+void r600_bc_clear(struct r600_bc *bc)
+{
+	struct r600_bc_cf *cf = NULL, *next_cf;
+
+	free(bc->bytecode);
+	bc->bytecode = NULL;
+
+	LIST_FOR_EACH_ENTRY_SAFE(cf, next_cf, &bc->cf, list) {
+		struct r600_bc_alu *alu = NULL, *next_alu;
+		struct r600_bc_tex *tex = NULL, *next_tex;
+		struct r600_bc_tex *vtx = NULL, *next_vtx;
+
+		LIST_FOR_EACH_ENTRY_SAFE(alu, next_alu, &cf->alu, list) {
+			free(alu);
+		}
+
+		LIST_INITHEAD(&cf->alu);
+
+		LIST_FOR_EACH_ENTRY_SAFE(tex, next_tex, &cf->tex, list) {
+			free(tex);
+		}
+
+		LIST_INITHEAD(&cf->tex);
+
+		LIST_FOR_EACH_ENTRY_SAFE(vtx, next_vtx, &cf->vtx, list) {
+			free(vtx);
+		}
+
+		LIST_INITHEAD(&cf->vtx);
+
+		free(cf);
+	}
+
+	LIST_INITHEAD(&cf->list);
+}
diff --git a/src/gallium/drivers/r600/r600_asm.h b/src/gallium/drivers/r600/r600_asm.h
index bebc7c15b00..25cda16837d 100644
--- a/src/gallium/drivers/r600/r600_asm.h
+++ b/src/gallium/drivers/r600/r600_asm.h
@@ -132,6 +132,11 @@ struct r600_bc_cf {
 	unsigned			pop_count;
 	unsigned			cf_addr; /* control flow addr */
 	unsigned			kcache0_mode;
+	unsigned			kcache1_mode;
+	unsigned			kcache0_addr;
+	unsigned			kcache1_addr;
+	unsigned			kcache0_bank;
+	unsigned			kcache1_bank;
 	unsigned			r6xx_uses_waterfall;
 	struct list_head		alu;
 	struct list_head		tex;
@@ -185,6 +190,7 @@ int eg_bc_cf_build(struct r600_bc *bc, struct r600_bc_cf *cf);
 
 /* r600_asm.c */
 int r600_bc_init(struct r600_bc *bc, enum radeon_family family);
+void r600_bc_clear(struct r600_bc *bc);
 int r600_bc_add_alu(struct r600_bc *bc, const struct r600_bc_alu *alu);
 int r600_bc_add_literal(struct r600_bc *bc, const u32 *value);
 int r600_bc_add_vtx(struct r600_bc *bc, const struct r600_bc_vtx *vtx);
diff --git a/src/gallium/drivers/r600/r600_blit.c b/src/gallium/drivers/r600/r600_blit.c
index 4bf44a171af..50d47060c1a 100644
--- a/src/gallium/drivers/r600/r600_blit.c
+++ b/src/gallium/drivers/r600/r600_blit.c
@@ -22,12 +22,22 @@
  */
 #include <util/u_surface.h>
 #include <util/u_blitter.h>
+#include <util/u_format.h>
 #include "r600_pipe.h"
 
-static void r600_blitter_save_states(struct pipe_context *ctx)
+enum r600_blitter_op /* bitmask */
+{
+    R600_CLEAR         = 1,
+    R600_CLEAR_SURFACE = 2,
+    R600_COPY          = 4
+};
+
+static void r600_blitter_begin(struct pipe_context *ctx, enum r600_blitter_op op)
 {
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
 
+	r600_context_queries_suspend(&rctx->ctx);
+
 	util_blitter_save_blend(rctx->blitter, rctx->states[R600_PIPE_STATE_BLEND]);
 	util_blitter_save_depth_stencil_alpha(rctx->blitter, rctx->states[R600_PIPE_STATE_DSA]);
 	if (rctx->states[R600_PIPE_STATE_STENCIL_REF]) {
@@ -47,25 +57,34 @@ static void r600_blitter_save_states(struct pipe_context *ctx)
 
 	rctx->vertex_elements = NULL;
 
-	/* TODO queries */
+	if (op & (R600_CLEAR_SURFACE | R600_COPY))
+		util_blitter_save_framebuffer(rctx->blitter, &rctx->framebuffer);
+
+	if (op & R600_COPY) {
+		util_blitter_save_fragment_sampler_states(
+			rctx->blitter, rctx->ps_samplers.n_samplers,
+			(void**)rctx->ps_samplers.samplers);
+
+		util_blitter_save_fragment_sampler_views(
+			rctx->blitter, rctx->ps_samplers.n_views,
+			(struct pipe_sampler_view**)rctx->ps_samplers.views);
+	}
+
+}
+
+static void r600_blitter_end(struct pipe_context *ctx)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+	r600_context_queries_resume(&rctx->ctx);
 }
 
 int r600_blit_uncompress_depth(struct pipe_context *ctx, struct r600_resource_texture *texture)
 {
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct pipe_framebuffer_state fb = *rctx->pframebuffer;
 	struct pipe_surface *zsurf, *cbsurf;
 	int level = 0;
 	float depth = 1.0f;
 
-	r600_context_queries_suspend(&rctx->ctx);
-	for (int i = 0; i < fb.nr_cbufs; i++) {
-		fb.cbufs[i] = NULL;
-		pipe_surface_reference(&fb.cbufs[i], rctx->pframebuffer->cbufs[i]);
-	}
-	fb.zsbuf = NULL;
-	pipe_surface_reference(&fb.zsbuf, rctx->pframebuffer->zsbuf);
-
 	zsurf = ctx->screen->get_tex_surface(ctx->screen, &texture->resource.base.b, 0, level, 0,
 					     PIPE_BIND_DEPTH_STENCIL);
 
@@ -73,22 +92,17 @@ int r600_blit_uncompress_depth(struct pipe_context *ctx, struct r600_resource_te
 			(struct pipe_resource*)texture->flushed_depth_texture,
 			0, level, 0, PIPE_BIND_RENDER_TARGET);
 
-	r600_blitter_save_states(ctx);
-	util_blitter_save_framebuffer(rctx->blitter, &fb);
-
 	if (rctx->family == CHIP_RV610 || rctx->family == CHIP_RV630 ||
-		rctx->family == CHIP_RV620 || rctx->family == CHIP_RV635)
+	    rctx->family == CHIP_RV620 || rctx->family == CHIP_RV635)
 		depth = 0.0f;
 
+	r600_blitter_begin(ctx, R600_CLEAR_SURFACE);
 	util_blitter_custom_depth_stencil(rctx->blitter, zsurf, cbsurf, rctx->custom_dsa_flush, depth);
+	r600_blitter_end(ctx);
 
 	pipe_surface_reference(&zsurf, NULL);
 	pipe_surface_reference(&cbsurf, NULL);
-	for (int i = 0; i < fb.nr_cbufs; i++) {
-		pipe_surface_reference(&fb.cbufs[i], NULL);
-	}
-	pipe_surface_reference(&fb.zsbuf, NULL);
-	r600_context_queries_resume(&rctx->ctx);
+
 
 	return 0;
 }
@@ -99,12 +113,11 @@ static void r600_clear(struct pipe_context *ctx, unsigned buffers,
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
 	struct pipe_framebuffer_state *fb = &rctx->framebuffer;
 
-	r600_context_queries_suspend(&rctx->ctx);
-	r600_blitter_save_states(ctx);
+	r600_blitter_begin(ctx, R600_CLEAR);
 	util_blitter_clear(rctx->blitter, fb->width, fb->height,
 				fb->nr_cbufs, buffers, rgba, depth,
 				stencil);
-	r600_context_queries_resume(&rctx->ctx);
+	r600_blitter_end(ctx);
 }
 
 static void r600_clear_render_target(struct pipe_context *ctx,
@@ -114,13 +127,11 @@ static void r600_clear_render_target(struct pipe_context *ctx,
 				     unsigned width, unsigned height)
 {
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct pipe_framebuffer_state *fb = &rctx->framebuffer;
 
-	r600_context_queries_suspend(&rctx->ctx);
-	util_blitter_save_framebuffer(rctx->blitter, fb);
+	r600_blitter_begin(ctx, R600_CLEAR_SURFACE);
 	util_blitter_clear_render_target(rctx->blitter, dst, rgba,
 					 dstx, dsty, width, height);
-	r600_context_queries_resume(&rctx->ctx);
+	r600_blitter_end(ctx);
 }
 
 static void r600_clear_depth_stencil(struct pipe_context *ctx,
@@ -132,16 +143,34 @@ static void r600_clear_depth_stencil(struct pipe_context *ctx,
 				     unsigned width, unsigned height)
 {
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct pipe_framebuffer_state *fb = &rctx->framebuffer;
 
-	r600_context_queries_suspend(&rctx->ctx);
-	util_blitter_save_framebuffer(rctx->blitter, fb);
+	r600_blitter_begin(ctx, R600_CLEAR_SURFACE);
 	util_blitter_clear_depth_stencil(rctx->blitter, dst, clear_flags, depth, stencil,
 					 dstx, dsty, width, height);
-	r600_context_queries_resume(&rctx->ctx);
+	r600_blitter_end(ctx);
 }
 
 
+
+/* Copy a block of pixels from one surface to another using HW. */
+static void r600_hw_copy_region(struct pipe_context *ctx,
+                                struct pipe_resource *dst,
+                                struct pipe_subresource subdst,
+                                unsigned dstx, unsigned dsty, unsigned dstz,
+                                struct pipe_resource *src,
+                                struct pipe_subresource subsrc,
+                                unsigned srcx, unsigned srcy, unsigned srcz,
+                                unsigned width, unsigned height)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+
+	r600_blitter_begin(ctx, R600_COPY);
+	util_blitter_copy_region(rctx->blitter, dst, subdst, dstx, dsty, dstz,
+				 src, subsrc, srcx, srcy, srcz, width, height,
+				 TRUE);
+	r600_blitter_end(ctx);
+}
+
 static void r600_resource_copy_region(struct pipe_context *ctx,
 				      struct pipe_resource *dst,
 				      struct pipe_subresource subdst,
@@ -151,8 +180,16 @@ static void r600_resource_copy_region(struct pipe_context *ctx,
 				      unsigned srcx, unsigned srcy, unsigned srcz,
 				      unsigned width, unsigned height)
 {
-	util_resource_copy_region(ctx, dst, subdst, dstx, dsty, dstz,
-				  src, subsrc, srcx, srcy, srcz, width, height);
+	boolean is_depth;
+	/* there is something wrong with depth resource copies at the moment so avoid them for now */
+	is_depth = util_format_get_component_bits(src->format, UTIL_FORMAT_COLORSPACE_ZS, 0) != 0;
+	if (is_depth)
+		util_resource_copy_region(ctx, dst, subdst, dstx, dsty, dstz,
+					  src, subsrc, srcx, srcy, srcz, width, height);
+	else
+		r600_hw_copy_region(ctx, dst, subdst, dstx, dsty, dstz,
+				    src, subsrc, srcx, srcy, srcz, width, height);
+
 }
 
 void r600_init_blit_functions(struct r600_pipe_context *rctx)
diff --git a/src/gallium/drivers/r600/r600_buffer.c b/src/gallium/drivers/r600/r600_buffer.c
index 2bfa4e22fec..455aa2e81f6 100644
--- a/src/gallium/drivers/r600/r600_buffer.c
+++ b/src/gallium/drivers/r600/r600_buffer.c
@@ -227,7 +227,7 @@ struct pipe_resource *r600_buffer_from_handle(struct pipe_screen *screen,
 	struct r600_resource *rbuffer;
 	struct r600_bo *bo = NULL;
 
-	bo = r600_bo_handle(rw, whandle->handle);
+	bo = r600_bo_handle(rw, whandle->handle, NULL);
 	if (bo == NULL) {
 		return NULL;
 	}
diff --git a/src/gallium/drivers/r600/r600_formats.h b/src/gallium/drivers/r600/r600_formats.h
new file mode 100644
index 00000000000..0c91a212384
--- /dev/null
+++ b/src/gallium/drivers/r600/r600_formats.h
@@ -0,0 +1,56 @@
+#ifndef R600_FORMATS_H
+#define R600_FORMATS_H
+
+/* list of formats from R700 ISA document - apply across GPUs in different registers */
+#define     FMT_INVALID                     0x00000000
+#define     FMT_8                           0x00000001
+#define     FMT_4_4                         0x00000002
+#define     FMT_3_3_2                       0x00000003
+#define     FMT_16                          0x00000005
+#define     FMT_16_FLOAT                    0x00000006
+#define     FMT_8_8                         0x00000007
+#define     FMT_5_6_5                       0x00000008
+#define     FMT_6_5_5                       0x00000009
+#define     FMT_1_5_5_5                     0x0000000A
+#define     FMT_4_4_4_4                     0x0000000B
+#define     FMT_5_5_5_1                     0x0000000C
+#define     FMT_32                          0x0000000D
+#define     FMT_32_FLOAT                    0x0000000E
+#define     FMT_16_16                       0x0000000F
+#define     FMT_16_16_FLOAT                 0x00000010
+#define     FMT_8_24                        0x00000011
+#define     FMT_8_24_FLOAT                  0x00000012
+#define     FMT_24_8                        0x00000013
+#define     FMT_24_8_FLOAT                  0x00000014
+#define     FMT_10_11_11                    0x00000015
+#define     FMT_10_11_11_FLOAT              0x00000016
+#define     FMT_11_11_10                    0x00000017
+#define     FMT_11_11_10_FLOAT              0x00000018
+#define     FMT_2_10_10_10                  0x00000019
+#define     FMT_8_8_8_8                     0x0000001A
+#define     FMT_10_10_10_2                  0x0000001B
+#define     FMT_X24_8_32_FLOAT              0x0000001C
+#define     FMT_32_32                       0x0000001D
+#define     FMT_32_32_FLOAT                 0x0000001E
+#define     FMT_16_16_16_16                 0x0000001F
+#define     FMT_16_16_16_16_FLOAT           0x00000020
+#define     FMT_32_32_32_32                 0x00000022
+#define     FMT_32_32_32_32_FLOAT           0x00000023
+#define     FMT_1                           0x00000025
+#define     FMT_GB_GR                       0x00000027
+#define     FMT_BG_RG                       0x00000028
+#define     FMT_32_AS_8                     0x00000029
+#define     FMT_32_AS_8_8                   0x0000002a
+#define     FMT_5_9_9_9_SHAREDEXP           0x0000002b
+#define     FMT_8_8_8                       0x0000002c
+#define     FMT_16_16_16                    0x0000002d
+#define     FMT_16_16_16_FLOAT              0x0000002e
+#define     FMT_32_32_32                    0x0000002f
+#define     FMT_32_32_32_FLOAT              0x00000030
+#define     FMT_BC1                         0x00000031
+#define     FMT_BC2                         0x00000032
+#define     FMT_BC3                         0x00000033
+#define     FMT_BC4                         0x00000034
+#define     FMT_BC5                         0x00000035
+
+#endif
diff --git a/src/gallium/drivers/r600/r600_opcodes.h b/src/gallium/drivers/r600/r600_opcodes.h
index 0cf9c1c401c..4f9b39a7fdc 100644
--- a/src/gallium/drivers/r600/r600_opcodes.h
+++ b/src/gallium/drivers/r600/r600_opcodes.h
@@ -233,12 +233,6 @@
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_CEIL                      0x00000012
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RNDNE                     0x00000013
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLOOR                     0x00000014
-/* same up to here */
-/*
-#define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA                      0x00000015
-#define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_FLOOR                0x00000016
-#define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_INT                  0x00000018
-*/
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_ASHR_INT                  0x00000015
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LSHR_INT                  0x00000016
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_LSHL_INT                  0x00000017
@@ -336,9 +330,11 @@
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIPSQRT_CLAMPED_64      0x00000098
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_SQRT_64                   0x00000099
 /* TODO Fill in more ALU */
+#define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_INT_FLOOR          0x000000B1
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4                      0x000000BE
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4_IEEE                 0x000000BF
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_CUBE                      0x000000C0
+#define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_INT                  0x000000CC
 
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INTERP_XY                      0x000000D6
 #define     EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INTERP_ZW                      0x000000D7
diff --git a/src/gallium/drivers/r600/r600_pipe.c b/src/gallium/drivers/r600/r600_pipe.c
index a9fc31d9fbd..87ec8cd2e83 100644
--- a/src/gallium/drivers/r600/r600_pipe.c
+++ b/src/gallium/drivers/r600/r600_pipe.c
@@ -83,9 +83,16 @@ static void r600_destroy_context(struct pipe_context *context)
 		free(rctx->states[i]);
 	}
 
+	util_blitter_destroy(rctx->blitter);
+
 	u_upload_destroy(rctx->upload_vb);
 	u_upload_destroy(rctx->upload_ib);
 
+	if (rctx->tran.translate_cache)
+		translate_cache_destroy(rctx->tran.translate_cache);
+
+	FREE(rctx->ps_resource);
+	FREE(rctx->vs_resource);
 	FREE(rctx);
 }
 
@@ -172,6 +179,24 @@ static struct pipe_context *r600_create_context(struct pipe_screen *screen, void
 		return NULL;
 	}
 
+	rctx->tran.translate_cache = translate_cache_create();
+	if (rctx->tran.translate_cache == NULL) {
+		FREE(rctx);
+		return NULL;
+	}
+	
+	rctx->vs_resource = CALLOC(R600_RESOURCE_ARRAY_SIZE, sizeof(struct r600_pipe_state));
+	if (!rctx->vs_resource) {
+		FREE(rctx);
+		return NULL;
+	}
+
+	rctx->ps_resource = CALLOC(R600_RESOURCE_ARRAY_SIZE, sizeof(struct r600_pipe_state));
+	if (!rctx->ps_resource) {
+		FREE(rctx);
+		return NULL;
+	}
+
 	class = r600_get_family_class(rctx->radeon);
 	if (class == R600 || class == R700)
 		rctx->custom_dsa_flush = r600_create_db_flush_dsa(rctx);
@@ -243,6 +268,7 @@ static int r600_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
 	case PIPE_CAP_INDEP_BLEND_ENABLE:
 	case PIPE_CAP_DEPTHSTENCIL_CLEAR_SEPARATE:
 	case PIPE_CAP_DEPTH_CLAMP:
+	case PIPE_CAP_SHADER_STENCIL_EXPORT:
 		return 1;
 
 	/* Unsupported features (boolean caps). */
@@ -258,7 +284,7 @@ static int r600_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
 		return 14;
 	case PIPE_CAP_MAX_VERTEX_TEXTURE_UNITS:
 		/* FIXME allow this once infrastructure is there */
-		return 0;
+		return 16;
 	case PIPE_CAP_MAX_TEXTURE_IMAGE_UNITS:
 	case PIPE_CAP_MAX_COMBINED_SAMPLERS:
 		return 16;
@@ -430,5 +456,7 @@ struct pipe_screen *r600_screen_create(struct radeon *radeon)
 	r600_init_screen_texture_functions(&rscreen->screen);
 	r600_init_screen_resource_functions(&rscreen->screen);
 
+	rscreen->tiling_info = r600_get_tiling_info(radeon);
+
 	return &rscreen->screen;
 }
diff --git a/src/gallium/drivers/r600/r600_pipe.h b/src/gallium/drivers/r600/r600_pipe.h
index c46029a5617..1c691f6b764 100644
--- a/src/gallium/drivers/r600/r600_pipe.h
+++ b/src/gallium/drivers/r600/r600_pipe.h
@@ -30,6 +30,7 @@
 #include <pipe/p_screen.h>
 #include <pipe/p_context.h>
 #include <util/u_math.h>
+#include "translate/translate_cache.h"
 #include "r600.h"
 #include "r600_public.h"
 #include "r600_shader.h"
@@ -58,6 +59,7 @@ enum r600_pipe_state_id {
 struct r600_screen {
 	struct pipe_screen		screen;
 	struct radeon			*radeon;
+	struct r600_tiling_info		*tiling_info;
 };
 
 struct r600_pipe_sampler_view {
@@ -81,8 +83,10 @@ struct r600_pipe_blend {
 struct r600_vertex_element
 {
 	unsigned			count;
-	unsigned			refcount;
-	struct pipe_vertex_element	elements[32];
+	struct pipe_vertex_element	elements[PIPE_MAX_ATTRIBS];
+	enum pipe_format		hw_format[PIPE_MAX_ATTRIBS];
+	unsigned			hw_format_size[PIPE_MAX_ATTRIBS];
+	boolean incompatible_layout;
 };
 
 struct r600_pipe_shader {
@@ -92,6 +96,29 @@ struct r600_pipe_shader {
 	struct r600_vertex_element	vertex_elements;
 };
 
+/* needed for blitter save */
+#define NUM_TEX_UNITS 16
+
+struct r600_textures_info {
+	struct r600_pipe_sampler_view   *views[NUM_TEX_UNITS];
+	unsigned                        n_views;
+	void				*samplers[NUM_TEX_UNITS];
+	unsigned                        n_samplers;
+};
+
+struct r600_translate_context {
+	/* Translate cache for incompatible vertex offset/stride/format fallback. */
+	struct translate_cache *translate_cache;
+
+	/* The vertex buffer slot containing the translated buffer. */
+	unsigned vb_slot;
+	/* Saved and new vertex element state. */
+	void *saved_velems, *new_velems;
+};
+
+#define R600_CONSTANT_ARRAY_SIZE 256
+#define R600_RESOURCE_ARRAY_SIZE 160
+
 struct r600_pipe_context {
 	struct pipe_context		context;
 	struct blitter_context		*blitter;
@@ -112,12 +139,8 @@ struct r600_pipe_context {
 	struct pipe_stencil_ref		stencil_ref;
 	struct pipe_viewport_state	viewport;
 	struct pipe_clip_state		clip;
-	unsigned			vs_nconst;
-	unsigned			ps_nconst;
-	struct r600_pipe_state		vs_const[256];
-	struct r600_pipe_state		ps_const[256];
-	struct r600_pipe_state		vs_resource[160];
-	struct r600_pipe_state		ps_resource[160];
+	struct r600_pipe_state		*vs_resource;
+	struct r600_pipe_state		*ps_resource;
 	struct r600_pipe_state		config;
 	struct r600_pipe_shader 	*ps_shader;
 	struct r600_pipe_shader 	*vs_shader;
@@ -130,6 +153,11 @@ struct r600_pipe_context {
 	struct u_upload_mgr		*upload_vb;
 	struct u_upload_mgr		*upload_ib;
 	unsigned			any_user_vbs;
+	struct r600_textures_info       ps_samplers;
+
+	unsigned vb_max_index;
+	struct r600_translate_context tran;
+
 };
 
 struct r600_drawl {
@@ -180,6 +208,7 @@ void r600_init_context_resource_functions(struct r600_pipe_context *r600);
 /* r600_shader.c */
 int r600_pipe_shader_update(struct pipe_context *ctx, struct r600_pipe_shader *shader);
 int r600_pipe_shader_create(struct pipe_context *ctx, struct r600_pipe_shader *shader, const struct tgsi_token *tokens);
+void r600_pipe_shader_destroy(struct pipe_context *ctx, struct r600_pipe_shader *shader);
 int r600_find_vs_semantic_index(struct r600_shader *vs,
 				struct r600_shader *ps, int id);
 
@@ -187,10 +216,6 @@ int r600_find_vs_semantic_index(struct r600_shader *vs,
 void r600_init_state_functions(struct r600_pipe_context *rctx);
 void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info);
 void r600_init_config(struct r600_pipe_context *rctx);
-void r600_translate_index_buffer(struct r600_pipe_context *r600,
-					struct pipe_resource **index_buffer,
-					unsigned *index_size,
-					unsigned *start, unsigned count);
 void *r600_create_db_flush_dsa(struct r600_pipe_context *rctx);
 /* r600_helper.h */
 int r600_conv_pipe_prim(unsigned pprim, unsigned *prim);
@@ -201,6 +226,38 @@ uint32_t r600_translate_texformat(enum pipe_format format,
 				  const unsigned char *swizzle_view, 
 				  uint32_t *word4_p, uint32_t *yuv_format_p);
 
+/* r600_translate.c */
+void r600_begin_vertex_translate(struct r600_pipe_context *rctx);
+void r600_end_vertex_translate(struct r600_pipe_context *rctx);
+void r600_translate_index_buffer(struct r600_pipe_context *r600,
+				 struct pipe_resource **index_buffer,
+				 unsigned *index_size,
+				 unsigned *start, unsigned count);
+
+/* r600_state_common.c */
+void r600_set_index_buffer(struct pipe_context *ctx,
+			   const struct pipe_index_buffer *ib);
+void r600_set_vertex_buffers(struct pipe_context *ctx, unsigned count,
+			     const struct pipe_vertex_buffer *buffers);
+void *r600_create_vertex_elements(struct pipe_context *ctx,
+				  unsigned count,
+				  const struct pipe_vertex_element *elements);
+void r600_delete_vertex_element(struct pipe_context *ctx, void *state);
+void r600_bind_blend_state(struct pipe_context *ctx, void *state);
+void r600_bind_rs_state(struct pipe_context *ctx, void *state);
+void r600_delete_rs_state(struct pipe_context *ctx, void *state);
+void r600_sampler_view_destroy(struct pipe_context *ctx,
+			       struct pipe_sampler_view *state);
+void r600_bind_state(struct pipe_context *ctx, void *state);
+void r600_delete_state(struct pipe_context *ctx, void *state);
+void r600_bind_vertex_elements(struct pipe_context *ctx, void *state);
+
+void *r600_create_shader_state(struct pipe_context *ctx,
+			       const struct pipe_shader_state *state);
+void r600_bind_ps_shader(struct pipe_context *ctx, void *state);
+void r600_bind_vs_shader(struct pipe_context *ctx, void *state);
+void r600_delete_ps_shader(struct pipe_context *ctx, void *state);
+void r600_delete_vs_shader(struct pipe_context *ctx, void *state);
 /*
  * common helpers
  */
diff --git a/src/gallium/drivers/r600/r600_resource.h b/src/gallium/drivers/r600/r600_resource.h
index 323960960d6..d152285815c 100644
--- a/src/gallium/drivers/r600/r600_resource.h
+++ b/src/gallium/drivers/r600/r600_resource.h
@@ -25,6 +25,9 @@
 
 #include "util/u_transfer.h"
 
+/* flag to indicate a resource is to be used as a transfer so should not be tiled */
+#define R600_RESOURCE_FLAG_TRANSFER     PIPE_RESOURCE_FLAG_DRV_PRIV
+
 /* Texture transfer. */
 struct r600_transfer {
 	/* Base class. */
@@ -49,14 +52,14 @@ struct r600_resource {
 
 struct r600_resource_texture {
 	struct r600_resource		resource;
-	unsigned long			offset[PIPE_MAX_TEXTURE_LEVELS];
-	unsigned long			pitch[PIPE_MAX_TEXTURE_LEVELS];
-	unsigned long			layer_size[PIPE_MAX_TEXTURE_LEVELS];
-	unsigned long			pitch_override;
-	unsigned long			bpt;
-	unsigned long			size;
+	unsigned			offset[PIPE_MAX_TEXTURE_LEVELS];
+	unsigned			pitch_in_bytes[PIPE_MAX_TEXTURE_LEVELS];
+	unsigned			pitch_in_pixels[PIPE_MAX_TEXTURE_LEVELS];
+	unsigned			layer_size[PIPE_MAX_TEXTURE_LEVELS];
+	unsigned			array_mode[PIPE_MAX_TEXTURE_LEVELS];
+	unsigned			pitch_override;
+	unsigned			size;
 	unsigned			tiled;
-	unsigned			array_mode;
 	unsigned			tile_type;
 	unsigned			depth;
 	unsigned			dirty;
@@ -124,4 +127,9 @@ void* r600_texture_transfer_map(struct pipe_context *ctx,
 void r600_texture_transfer_unmap(struct pipe_context *ctx,
 				 struct pipe_transfer* transfer);
 
+struct r600_surface {
+	struct pipe_surface base;
+	unsigned aligned_height;
+};
+
 #endif
diff --git a/src/gallium/drivers/r600/r600_shader.c b/src/gallium/drivers/r600/r600_shader.c
index 366d5d9c351..4106587398b 100644
--- a/src/gallium/drivers/r600/r600_shader.c
+++ b/src/gallium/drivers/r600/r600_shader.c
@@ -107,24 +107,28 @@ static void r600_pipe_shader_ps(struct pipe_context *ctx, struct r600_pipe_shade
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
 	struct r600_pipe_state *rstate = &shader->rstate;
 	struct r600_shader *rshader = &shader->shader;
-	unsigned i, tmp, exports_ps, num_cout, spi_ps_in_control_0, spi_input_z;
-	boolean have_pos = FALSE, have_face = FALSE;
+	unsigned i, tmp, exports_ps, num_cout, spi_ps_in_control_0, spi_input_z, spi_ps_in_control_1;
+	int pos_index = -1, face_index = -1;
 
 	/* clear previous register */
 	rstate->nregs = 0;
 
 	for (i = 0; i < rshader->ninput; i++) {
 		tmp = S_028644_SEMANTIC(r600_find_vs_semantic_index(&rctx->vs_shader->shader, rshader, i));
-		tmp |= S_028644_SEL_CENTROID(1);
+		if (rshader->input[i].centroid)
+			tmp |= S_028644_SEL_CENTROID(1);
+		if (rshader->input[i].interpolate == TGSI_INTERPOLATE_LINEAR)
+			tmp |= S_028644_SEL_LINEAR(1);
+
 		if (rshader->input[i].name == TGSI_SEMANTIC_POSITION)
-			have_pos = TRUE;
+			pos_index = i;
 		if (rshader->input[i].name == TGSI_SEMANTIC_COLOR ||
 		    rshader->input[i].name == TGSI_SEMANTIC_BCOLOR ||
 		    rshader->input[i].name == TGSI_SEMANTIC_POSITION) {
 			tmp |= S_028644_FLAT_SHADE(rshader->flat_shade);
 		}
 		if (rshader->input[i].name == TGSI_SEMANTIC_FACE)
-			have_face = TRUE;
+			face_index = i;
 		if (rshader->input[i].name == TGSI_SEMANTIC_GENERIC &&
 			rctx->sprite_coord_enable & (1 << rshader->input[i].sid)) {
 			tmp |= S_028644_PT_SPRITE_TEX(1);
@@ -137,12 +141,17 @@ static void r600_pipe_shader_ps(struct pipe_context *ctx, struct r600_pipe_shade
 						R_02880C_DB_SHADER_CONTROL,
 						S_02880C_Z_EXPORT_ENABLE(1),
 						S_02880C_Z_EXPORT_ENABLE(1), NULL);
+		if (rshader->output[i].name == TGSI_SEMANTIC_STENCIL)
+			r600_pipe_state_add_reg(rstate,
+						R_02880C_DB_SHADER_CONTROL,
+						S_02880C_STENCIL_REF_EXPORT_ENABLE(1),
+						S_02880C_STENCIL_REF_EXPORT_ENABLE(1), NULL);
 	}
 
 	exports_ps = 0;
 	num_cout = 0;
 	for (i = 0; i < rshader->noutput; i++) {
-		if (rshader->output[i].name == TGSI_SEMANTIC_POSITION)
+		if (rshader->output[i].name == TGSI_SEMANTIC_POSITION || rshader->output[i].name == TGSI_SEMANTIC_STENCIL)
 			exports_ps |= 1;
 		else if (rshader->output[i].name == TGSI_SEMANTIC_COLOR) {
 			num_cout++;
@@ -157,13 +166,22 @@ static void r600_pipe_shader_ps(struct pipe_context *ctx, struct r600_pipe_shade
 	spi_ps_in_control_0 = S_0286CC_NUM_INTERP(rshader->ninput) |
 				S_0286CC_PERSP_GRADIENT_ENA(1);
 	spi_input_z = 0;
-	if (have_pos) {
-		spi_ps_in_control_0 |=  S_0286CC_POSITION_ENA(1) |
-					S_0286CC_BARYC_SAMPLE_CNTL(1);
+	if (pos_index != -1) {
+		spi_ps_in_control_0 |= (S_0286CC_POSITION_ENA(1) |
+					S_0286CC_POSITION_CENTROID(rshader->input[pos_index].centroid) |
+					S_0286CC_POSITION_ADDR(rshader->input[pos_index].gpr) |
+					S_0286CC_BARYC_SAMPLE_CNTL(1));
 		spi_input_z |= 1;
 	}
+
+	spi_ps_in_control_1 = 0;
+	if (face_index != -1) {
+		spi_ps_in_control_1 |= S_0286D0_FRONT_FACE_ENA(1) |
+			S_0286D0_FRONT_FACE_ADDR(rshader->input[face_index].gpr);
+	}
+
 	r600_pipe_state_add_reg(rstate, R_0286CC_SPI_PS_IN_CONTROL_0, spi_ps_in_control_0, 0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate, R_0286D0_SPI_PS_IN_CONTROL_1, S_0286D0_FRONT_FACE_ENA(have_face), 0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate, R_0286D0_SPI_PS_IN_CONTROL_1, spi_ps_in_control_1, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_0286D8_SPI_INPUT_Z, spi_input_z, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate,
 				R_028840_SQ_PGM_START_PS,
@@ -252,7 +270,7 @@ static int r600_shader_update(struct pipe_context *ctx, struct r600_pipe_shader
 	}
 	rshader->vertex_elements = *rctx->vertex_elements;
 	for (i = 0; i < rctx->vertex_elements->count; i++) {
-		resource_format[nresources++] = rctx->vertex_elements->elements[i].src_format;
+		resource_format[nresources++] = rctx->vertex_elements->hw_format[i];
 	}
 	r600_bo_reference(rctx->radeon, &rshader->bo, NULL);
 	LIST_FOR_EACH_ENTRY(cf, &bc->cf, list) {
@@ -320,6 +338,18 @@ int r600_pipe_shader_create(struct pipe_context *ctx, struct r600_pipe_shader *s
 	return 0;
 }
 
+void
+r600_pipe_shader_destroy(struct pipe_context *ctx, struct r600_pipe_shader *shader)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+
+	r600_bo_reference(rctx->radeon, &shader->bo, NULL);
+
+	r600_bc_clear(&shader->shader.bc);
+
+	/* FIXME: is there more stuff to free? */
+}
+
 /*
  * tgsi -> r600 shader
  */
@@ -339,6 +369,11 @@ struct r600_shader_ctx {
 	u32					*literals;
 	u32					nliterals;
 	u32					max_driver_temp_used;
+	/* needed for evergreen interpolation */
+	boolean                                 input_centroid;
+	boolean                                 input_linear;
+	boolean                                 input_perspective;
+	int					num_interp_gpr;
 };
 
 struct r600_shader_tgsi_instruction {
@@ -371,11 +406,9 @@ static int tgsi_is_supported(struct r600_shader_ctx *ctx)
 	}
 #endif
 	for (j = 0; j < i->Instruction.NumSrcRegs; j++) {
-		if (i->Src[j].Register.Dimension ||
-			i->Src[j].Register.Absolute) {
-			R600_ERR("unsupported src %d (dimension %d|absolute %d)\n", j,
-				 i->Src[j].Register.Dimension,
-				 i->Src[j].Register.Absolute);
+		if (i->Src[j].Register.Dimension) {
+			R600_ERR("unsupported src %d (dimension %d)\n", j,
+				 i->Src[j].Register.Dimension);
 			return -EINVAL;
 		}
 	}
@@ -388,10 +421,33 @@ static int tgsi_is_supported(struct r600_shader_ctx *ctx)
 	return 0;
 }
 
-static int evergreen_interp_alu(struct r600_shader_ctx *ctx, int gpr)
+static int evergreen_interp_alu(struct r600_shader_ctx *ctx, int input)
 {
 	int i, r;
 	struct r600_bc_alu alu;
+	int gpr = 0, base_chan = 0;
+	int ij_index = 0;
+
+	if (ctx->shader->input[input].interpolate == TGSI_INTERPOLATE_PERSPECTIVE) {
+		ij_index = 0;
+		if (ctx->shader->input[input].centroid)
+			ij_index++;
+	} else if (ctx->shader->input[input].interpolate == TGSI_INTERPOLATE_LINEAR) {
+		ij_index = 0;
+		/* if we have perspective add one */
+		if (ctx->input_perspective)  {
+			ij_index++;
+			/* if we have perspective centroid */
+			if (ctx->input_centroid)
+				ij_index++;
+		}
+		if (ctx->shader->input[input].centroid)
+			ij_index++;
+	}
+		
+	/* work out gpr and base_chan from index */
+	gpr = ij_index / 2;
+	base_chan = (2 * (ij_index % 2)) + 1;
 
 	for (i = 0; i < 8; i++) {
 		memset(&alu, 0, sizeof(struct r600_bc_alu));
@@ -402,13 +458,16 @@ static int evergreen_interp_alu(struct r600_shader_ctx *ctx, int gpr)
 			alu.inst = EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INTERP_XY;
 
 		if ((i > 1) && (i < 6)) {
-			alu.dst.sel = ctx->shader->input[gpr].gpr;
+			alu.dst.sel = ctx->shader->input[input].gpr;
 			alu.dst.write = 1;
 		}
 
 		alu.dst.chan = i % 4;
-		alu.src[0].chan = (1 - (i % 2));
-		alu.src[1].sel = V_SQ_ALU_SRC_PARAM_BASE + gpr;
+
+		alu.src[0].sel = gpr;
+		alu.src[0].chan = (base_chan - (i % 2));
+
+		alu.src[1].sel = V_SQ_ALU_SRC_PARAM_BASE + ctx->shader->input[input].lds_pos;
 
 		alu.bank_swizzle_force = SQ_ALU_VEC_210;
 		if ((i % 4) == 3)
@@ -434,6 +493,7 @@ static int tgsi_declaration(struct r600_shader_ctx *ctx)
 		ctx->shader->input[i].name = d->Semantic.Name;
 		ctx->shader->input[i].sid = d->Semantic.Index;
 		ctx->shader->input[i].interpolate = d->Declaration.Interpolate;
+		ctx->shader->input[i].centroid = d->Declaration.Centroid;
 		ctx->shader->input[i].gpr = ctx->file_offset[TGSI_FILE_INPUT] + i;
 		if (ctx->type == TGSI_PROCESSOR_VERTEX) {
 			/* turn input into fetch */
@@ -457,7 +517,12 @@ static int tgsi_declaration(struct r600_shader_ctx *ctx)
 		}
 		if (ctx->type == TGSI_PROCESSOR_FRAGMENT && ctx->bc->chiprev == 2) {
 			/* turn input into interpolate on EG */
-			evergreen_interp_alu(ctx, i);
+			if (ctx->shader->input[i].name != TGSI_SEMANTIC_POSITION) {
+				if (ctx->shader->input[i].interpolate > 0) {
+					ctx->shader->input[i].lds_pos = ctx->shader->nlds++;
+					evergreen_interp_alu(ctx, i);
+				}
+			}
 		}
 		break;
 	case TGSI_FILE_OUTPUT:
@@ -484,6 +549,53 @@ static int r600_get_temp(struct r600_shader_ctx *ctx)
 	return ctx->temp_reg + ctx->max_driver_temp_used++;
 }
 
+/* 
+ * for evergreen we need to scan the shader to find the number of GPRs we need to
+ * reserve for interpolation.
+ *
+ * we need to know if we are going to emit
+ * any centroid inputs
+ * if perspective and linear are required
+*/
+static int evergreen_gpr_count(struct r600_shader_ctx *ctx)
+{
+	int i;
+	int num_baryc;
+
+	ctx->input_linear = FALSE;
+	ctx->input_perspective = FALSE;
+	ctx->input_centroid = FALSE;
+	ctx->num_interp_gpr = 1;
+
+	/* any centroid inputs */
+	for (i = 0; i < ctx->info.num_inputs; i++) {
+		/* skip position/face */
+		if (ctx->info.input_semantic_name[i] == TGSI_SEMANTIC_POSITION ||
+		    ctx->info.input_semantic_name[i] == TGSI_SEMANTIC_FACE)
+			continue;
+		if (ctx->info.input_interpolate[i] == TGSI_INTERPOLATE_LINEAR)
+			ctx->input_linear = TRUE;
+		if (ctx->info.input_interpolate[i] == TGSI_INTERPOLATE_PERSPECTIVE)
+			ctx->input_perspective = TRUE;
+		if (ctx->info.input_centroid[i])
+			ctx->input_centroid = TRUE;
+	}
+
+	num_baryc = 0;
+	/* ignoring sample for now */
+	if (ctx->input_perspective)
+		num_baryc++;
+	if (ctx->input_linear)
+		num_baryc++;
+	if (ctx->input_centroid)
+		num_baryc *= 2;
+
+	ctx->num_interp_gpr += (num_baryc + 1) >> 1;
+
+	/* TODO PULL MODEL and LINE STIPPLE, FIXED PT POS */
+	return ctx->num_interp_gpr;
+}
+
 int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *shader)
 {
 	struct tgsi_full_immediate *immediate;
@@ -530,7 +642,7 @@ int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *s
 		ctx.file_offset[TGSI_FILE_INPUT] = 1;
 	}
 	if (ctx.type == TGSI_PROCESSOR_FRAGMENT && ctx.bc->chiprev == 2) {
-		ctx.file_offset[TGSI_FILE_INPUT] = 1;
+		ctx.file_offset[TGSI_FILE_INPUT] = evergreen_gpr_count(&ctx);
 	}
 	ctx.file_offset[TGSI_FILE_OUTPUT] = ctx.file_offset[TGSI_FILE_INPUT] +
 						ctx.info.file_count[TGSI_FILE_INPUT];
@@ -628,7 +740,14 @@ int r600_shader_from_tgsi(const struct tgsi_token *tokens, struct r600_shader *s
 			} else if (shader->output[i].name == TGSI_SEMANTIC_POSITION) {
 				output[i].array_base = 61;
 				output[i].swizzle_x = 2;
-				output[i].swizzle_y = output[i].swizzle_z = output[i].swizzle_w = 7;
+				output[i].swizzle_y = 7;
+				output[i].swizzle_z = output[i].swizzle_w = 7;
+				output[i].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL;
+			} else if (shader->output[i].name == TGSI_SEMANTIC_STENCIL) {
+				output[i].array_base = 61;
+				output[i].swizzle_x = 7;
+				output[i].swizzle_y = 1;
+				output[i].swizzle_z = output[i].swizzle_w = 7;
 				output[i].type = V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL;
 			} else {
 				R600_ERR("unsupported fragment output name %d\n", shader->output[i].name);
@@ -734,6 +853,7 @@ static int tgsi_src(struct r600_shader_ctx *ctx,
 	if (tgsi_src->Register.Indirect)
 		r600_src->rel = V_SQ_REL_RELATIVE;
 	r600_src->neg = tgsi_src->Register.Negate;
+	r600_src->abs = tgsi_src->Register.Absolute;
 	r600_src->sel += ctx->file_offset[tgsi_src->Register.File];
 	return 0;
 }
@@ -796,6 +916,7 @@ static int tgsi_split_constant(struct r600_shader_ctx *ctx, struct r600_bc_alu_s
 				alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
 				alu.src[0].sel = r600_src[i].sel;
 				alu.src[0].chan = k;
+				alu.src[0].rel = r600_src[i].rel;
 				alu.dst.sel = treg;
 				alu.dst.chan = k;
 				alu.dst.write = 1;
@@ -806,6 +927,7 @@ static int tgsi_split_constant(struct r600_shader_ctx *ctx, struct r600_bc_alu_s
 					return r;
 			}
 			r600_src[i].sel = treg;
+			r600_src[i].rel =0;
 			j--;
 		}
 	}
@@ -1846,7 +1968,7 @@ static int tgsi_tex(struct r600_shader_ctx *ctx)
 			memset(&alu, 0, sizeof(struct r600_bc_alu));
 			alu.inst = CTX_INST(V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV);
 			alu.src[0].sel = src_gpr;
-			alu.src[0].chan = i;
+			alu.src[0].chan = tgsi_chan(&inst->Src[0], i);
 			alu.dst.sel = ctx->temp_reg;
 			alu.dst.chan = i;
 			if (i == 3)
@@ -1866,8 +1988,10 @@ static int tgsi_tex(struct r600_shader_ctx *ctx)
 
 	memset(&tex, 0, sizeof(struct r600_bc_tex));
 	tex.inst = opcode;
-	tex.resource_id = ctx->file_offset[inst->Src[1].Register.File] + inst->Src[1].Register.Index;
-	tex.sampler_id = tex.resource_id;
+	tex.sampler_id = ctx->file_offset[inst->Src[1].Register.File] + inst->Src[1].Register.Index;
+	tex.resource_id = tex.sampler_id;
+	if (ctx->shader->processor_type == TGSI_PROCESSOR_VERTEX)
+		tex.resource_id += PIPE_MAX_ATTRIBS;
 	tex.src_gpr = src_gpr;
 	tex.dst_gpr = ctx->file_offset[inst->Dst[0].Register.File] + inst->Dst[0].Register.Index;
 	tex.dst_sel_x = (inst->Dst[0].Register.WriteMask & 1) ? 0 : 7;
@@ -2175,7 +2299,7 @@ static int tgsi_xpd(struct r600_shader_ctx *ctx)
 static int tgsi_exp(struct r600_shader_ctx *ctx)
 {
 	struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
-	struct r600_bc_alu_src r600_src[3];
+	struct r600_bc_alu_src r600_src[3] = { { 0 } };
 	struct r600_bc_alu alu;
 	int r;
 
@@ -2498,7 +2622,40 @@ static int tgsi_log(struct r600_shader_ctx *ctx)
 }
 
 /* r6/7 only for now */
-static int tgsi_arl(struct r600_shader_ctx *ctx)
+static int tgsi_eg_arl(struct r600_shader_ctx *ctx)
+{
+	struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
+	struct r600_bc_alu alu;
+	int r;
+	
+	memset(&alu, 0, sizeof(struct r600_bc_alu));
+
+	alu.inst = EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_FLT_TO_INT_FLOOR;
+	r = tgsi_src(ctx, &inst->Src[0], &alu.src[0]);
+	if (r)
+		return r;
+	alu.src[0].chan = tgsi_chan(&inst->Src[0], 0);
+	alu.last = 1;
+	alu.dst.chan = 0;
+	alu.dst.sel = ctx->temp_reg;
+	alu.dst.write = 1;
+	r = r600_bc_add_alu_type(ctx->bc, &alu, CTX_INST(V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU));
+	if (r)
+		return r;
+	memset(&alu, 0, sizeof(struct r600_bc_alu));
+	alu.inst = EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOVA_INT;
+	r = tgsi_src(ctx, &inst->Src[0], &alu.src[0]);
+	if (r)
+		return r;
+	alu.src[0].sel = ctx->temp_reg;
+	alu.src[0].chan = 0;
+	alu.last = 1;
+	r = r600_bc_add_alu_type(ctx->bc, &alu, CTX_INST(V_SQ_CF_ALU_WORD1_SQ_CF_INST_ALU));
+	if (r)
+		return r;
+	return 0;
+}
+static int tgsi_r600_arl(struct r600_shader_ctx *ctx)
 {
 	/* TODO from r600c, ar values don't persist between clauses */
 	struct tgsi_full_instruction *inst = &ctx->parse.FullToken.FullInstruction;
@@ -2843,7 +3000,7 @@ static int tgsi_loop_brk_cont(struct r600_shader_ctx *ctx)
 }
 
 static struct r600_shader_tgsi_instruction r600_shader_tgsi_instruction[] = {
-	{TGSI_OPCODE_ARL,	0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_arl},
+	{TGSI_OPCODE_ARL,	0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_r600_arl},
 	{TGSI_OPCODE_MOV,	0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV, tgsi_op2},
 	{TGSI_OPCODE_LIT,	0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_lit},
 
@@ -2924,7 +3081,7 @@ static struct r600_shader_tgsi_instruction r600_shader_tgsi_instruction[] = {
 	{TGSI_OPCODE_NRM,	0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
 	{TGSI_OPCODE_DIV,	0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
 	{TGSI_OPCODE_DP2,	0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
-	{TGSI_OPCODE_TXL,	0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
+	{TGSI_OPCODE_TXL,	0, SQ_TEX_INST_SAMPLE_L, tgsi_tex},
 	{TGSI_OPCODE_BRK,	0, V_SQ_CF_WORD1_SQ_CF_INST_LOOP_BREAK, tgsi_loop_brk_cont},
 	{TGSI_OPCODE_IF,	0, V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_if},
 	/* gap */
@@ -3007,7 +3164,7 @@ static struct r600_shader_tgsi_instruction r600_shader_tgsi_instruction[] = {
 };
 
 static struct r600_shader_tgsi_instruction eg_shader_tgsi_instruction[] = {
-	{TGSI_OPCODE_ARL,	0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
+	{TGSI_OPCODE_ARL,	0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_eg_arl},
 	{TGSI_OPCODE_MOV,	0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_MOV, tgsi_op2},
 	{TGSI_OPCODE_LIT,	0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_lit},
 	{TGSI_OPCODE_RCP,	0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_RECIP_IEEE, tgsi_trans_srcx_replicate},
@@ -3082,7 +3239,7 @@ static struct r600_shader_tgsi_instruction eg_shader_tgsi_instruction[] = {
 	{TGSI_OPCODE_NRM,	0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
 	{TGSI_OPCODE_DIV,	0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
 	{TGSI_OPCODE_DP2,	0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_DOT4, tgsi_dp},
-	{TGSI_OPCODE_TXL,	0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_unsupported},
+	{TGSI_OPCODE_TXL,	0, SQ_TEX_INST_SAMPLE_L, tgsi_tex},
 	{TGSI_OPCODE_BRK,	0, EG_V_SQ_CF_WORD1_SQ_CF_INST_LOOP_BREAK, tgsi_loop_brk_cont},
 	{TGSI_OPCODE_IF,	0, EG_V_SQ_ALU_WORD1_OP2_SQ_OP2_INST_NOP, tgsi_if},
 	/* gap */
diff --git a/src/gallium/drivers/r600/r600_shader.h b/src/gallium/drivers/r600/r600_shader.h
index 6e2620f2012..f8bc5951395 100644
--- a/src/gallium/drivers/r600/r600_shader.h
+++ b/src/gallium/drivers/r600/r600_shader.h
@@ -31,6 +31,8 @@ struct r600_shader_io {
 	unsigned		done;
 	int			sid;
 	unsigned		interpolate;
+	boolean                 centroid;
+	unsigned		lds_pos; /* for evergreen */
 };
 
 struct r600_shader {
@@ -39,6 +41,7 @@ struct r600_shader {
 	boolean			flat_shade;
 	unsigned		ninput;
 	unsigned		noutput;
+	unsigned		nlds;
 	struct r600_shader_io	input[32];
 	struct r600_shader_io	output[32];
 	enum radeon_family	family;
diff --git a/src/gallium/drivers/r600/r600_state.c b/src/gallium/drivers/r600/r600_state.c
index 7ceedf6363b..b3e0d493217 100644
--- a/src/gallium/drivers/r600/r600_state.c
+++ b/src/gallium/drivers/r600/r600_state.c
@@ -37,7 +37,7 @@
 #include <util/u_memory.h>
 #include <util/u_inlines.h>
 #include <util/u_upload_mgr.h>
-#include <util/u_index_modify.h>
+#include <util/u_framebuffer.h>
 #include <pipebuffer/pb_buffer.h>
 #include "r600.h"
 #include "r600d.h"
@@ -98,7 +98,7 @@ static void r600_draw_common(struct r600_drawl *draw)
 			vertex_buffer->buffer_offset +
 			r600_bo_offset(rbuffer->bo);
 
-		format = r600_translate_vertex_data_type(rctx->vertex_elements->elements[i].src_format);
+		format = r600_translate_vertex_data_type(rctx->vertex_elements->hw_format[i]);
 
 		word2 = format | S_038008_STRIDE(vertex_buffer->stride);
 
@@ -181,40 +181,21 @@ static void r600_draw_common(struct r600_drawl *draw)
 	r600_context_draw(&rctx->ctx, &rdraw);
 }
 
-void r600_translate_index_buffer(struct r600_pipe_context *r600,
-					struct pipe_resource **index_buffer,
-					unsigned *index_size,
-					unsigned *start, unsigned count)
-{
-	switch (*index_size) {
-	case 1:
-		util_shorten_ubyte_elts(&r600->context, index_buffer, 0, *start, count);
-		*index_size = 2;
-		*start = 0;
-		break;
-
-	case 2:
-		if (*start % 2 != 0) {
-			util_rebuild_ushort_elts(&r600->context, index_buffer, 0, *start, count);
-			*start = 0;
-		}
-		break;
-
-	case 4:
-		break;
-	}
-}
-
 void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
 {
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
 	struct r600_drawl draw;
+	boolean translate = FALSE;
+
+	if (rctx->vertex_elements->incompatible_layout) {
+		r600_begin_vertex_translate(rctx);
+		translate = TRUE;
+	}
 
 	if (rctx->any_user_vbs) {
 		r600_upload_user_buffers(rctx);
 		rctx->any_user_vbs = FALSE;
 	}
-
 	memset(&draw, 0, sizeof(struct r600_drawl));
 	draw.ctx = ctx;
 	draw.mode = info->mode;
@@ -245,6 +226,9 @@ void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
 	}
 	r600_draw_common(&draw);
 
+	if (translate)
+		r600_end_vertex_translate(rctx);
+
 	pipe_resource_reference(&draw.index_buffer, NULL);
 }
 
@@ -340,20 +324,6 @@ static void *r600_create_blend_state(struct pipe_context *ctx,
 	return rstate;
 }
 
-static void r600_bind_blend_state(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_blend *blend = (struct r600_pipe_blend *)state;
-	struct r600_pipe_state *rstate;
-
-	if (state == NULL)
-		return;
-	rstate = &blend->rstate;
-	rctx->states[rstate->id] = rstate;
-	rctx->cb_target_mask = blend->cb_target_mask;
-	r600_context_pipe_state_set(&rctx->ctx, rstate);
-}
-
 static void *r600_create_dsa_state(struct pipe_context *ctx,
 				   const struct pipe_depth_stencil_alpha_state *state)
 {
@@ -446,6 +416,7 @@ static void *r600_create_rs_state(struct pipe_context *ctx,
 	struct r600_pipe_state *rstate;
 	unsigned tmp;
 	unsigned prov_vtx = 1, polygon_dual_mode;
+	unsigned clip_rule;
 
 	if (rs == NULL) {
 		return NULL;
@@ -455,6 +426,7 @@ static void *r600_create_rs_state(struct pipe_context *ctx,
 	rs->flatshade = state->flatshade;
 	rs->sprite_coord_enable = state->sprite_coord_enable;
 
+	clip_rule = state->scissor ? 0xAAAA : 0xFFFF;
 	/* offset */
 	rs->offset_units = state->offset_units;
 	rs->offset_scale = state->offset_scale * 12.0f;
@@ -462,7 +434,7 @@ static void *r600_create_rs_state(struct pipe_context *ctx,
 	rstate->id = R600_PIPE_STATE_RASTERIZER;
 	if (state->flatshade_first)
 		prov_vtx = 0;
-	tmp = 0x00000001;
+	tmp = S_0286D4_FLAT_SHADE_ENA(1);
 	if (state->sprite_coord_enable) {
 		tmp |= S_0286D4_PNT_SPRITE_ENA(1) |
 			S_0286D4_PNT_SPRITE_OVRD_X(2) |
@@ -496,7 +468,10 @@ static void *r600_create_rs_state(struct pipe_context *ctx,
 	tmp = (unsigned)(state->point_size * 8.0);
 	r600_pipe_state_add_reg(rstate, R_028A00_PA_SU_POINT_SIZE, S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp), 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_028A04_PA_SU_POINT_MINMAX, 0x80000000, 0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate, R_028A08_PA_SU_LINE_CNTL, 0x00000008, 0xFFFFFFFF, NULL);
+
+	tmp = (unsigned)(state->line_width * 8.0);
+	r600_pipe_state_add_reg(rstate, R_028A08_PA_SU_LINE_CNTL, S_028A08_WIDTH(tmp), 0xFFFFFFFF, NULL);
+
 	r600_pipe_state_add_reg(rstate, R_028A0C_PA_SC_LINE_STIPPLE, 0x00000005, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_028A48_PA_SC_MPASS_PS_CNTL, 0x00000000, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_028C00_PA_SC_LINE_CNTL, 0x00000400, 0xFFFFFFFF, NULL);
@@ -505,37 +480,9 @@ static void *r600_create_rs_state(struct pipe_context *ctx,
 	r600_pipe_state_add_reg(rstate, R_028C14_PA_CL_GB_HORZ_CLIP_ADJ, 0x3F800000, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_028C18_PA_CL_GB_HORZ_DISC_ADJ, 0x3F800000, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_028DFC_PA_SU_POLY_OFFSET_CLAMP, 0x00000000, 0xFFFFFFFF, NULL);
-	return rstate;
-}
-
-static void r600_bind_rs_state(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_rasterizer *rs = (struct r600_pipe_rasterizer *)state;
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-
-	if (state == NULL)
-		return;
-
-	rctx->flatshade = rs->flatshade;
-	rctx->sprite_coord_enable = rs->sprite_coord_enable;
-	rctx->rasterizer = rs;
+	r600_pipe_state_add_reg(rstate, R_02820C_PA_SC_CLIPRECT_RULE, clip_rule, 0xFFFFFFFF, NULL);
 
-	rctx->states[rs->rstate.id] = &rs->rstate;
-	r600_context_pipe_state_set(&rctx->ctx, &rs->rstate);
-}
-
-static void r600_delete_rs_state(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_rasterizer *rs = (struct r600_pipe_rasterizer *)state;
-
-	if (rctx->rasterizer == rs) {
-		rctx->rasterizer = NULL;
-	}
-	if (rctx->states[rs->rstate.id] == &rs->rstate) {
-		rctx->states[rs->rstate.id] = NULL;
-	}
-	free(rs);
+	return rstate;
 }
 
 static void *r600_create_sampler_state(struct pipe_context *ctx,
@@ -574,28 +521,6 @@ static void *r600_create_sampler_state(struct pipe_context *ctx,
 	return rstate;
 }
 
-static void *r600_create_vertex_elements(struct pipe_context *ctx,
-				unsigned count,
-				const struct pipe_vertex_element *elements)
-{
-	struct r600_vertex_element *v = CALLOC_STRUCT(r600_vertex_element);
-
-	assert(count < 32);
-	v->count = count;
-	v->refcount = 1;
-	memcpy(v->elements, elements, count * sizeof(struct pipe_vertex_element));
-	return v;
-}
-
-static void r600_sampler_view_destroy(struct pipe_context *ctx,
-				      struct pipe_sampler_view *state)
-{
-	struct r600_pipe_sampler_view *resource = (struct r600_pipe_sampler_view *)state;
-
-	pipe_resource_reference(&state->texture, NULL);
-	FREE(resource);
-}
-
 static struct pipe_sampler_view *r600_create_sampler_view(struct pipe_context *ctx,
 							struct pipe_resource *texture,
 							const struct pipe_sampler_view *state)
@@ -648,7 +573,11 @@ static struct pipe_sampler_view *r600_create_sampler_view(struct pipe_context *c
 		bo[0] = rbuffer->bo;
 		bo[1] = rbuffer->bo;
 	}
-	pitch = align(tmp->pitch[0] / tmp->bpt, 8);
+	pitch = align(tmp->pitch_in_pixels[0], 8);
+	if (tmp->tiled) {
+		array_mode = tmp->array_mode[0];
+		tile_type = tmp->tile_type;
+	}
 
 	/* FIXME properly handle first level != 0 */
 	r600_pipe_state_add_reg(rstate, R_038000_RESOURCE0_WORD0,
@@ -683,32 +612,43 @@ static struct pipe_sampler_view *r600_create_sampler_view(struct pipe_context *c
 static void r600_set_vs_sampler_view(struct pipe_context *ctx, unsigned count,
 					struct pipe_sampler_view **views)
 {
-	/* TODO */
-	assert(1);
-}
-
-static void r600_set_ps_sampler_view(struct pipe_context *ctx, unsigned count,
-					struct pipe_sampler_view **views)
-{
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
 	struct r600_pipe_sampler_view **resource = (struct r600_pipe_sampler_view **)views;
 
 	for (int i = 0; i < count; i++) {
 		if (resource[i]) {
-			r600_context_pipe_state_set_ps_resource(&rctx->ctx, &resource[i]->state, i);
+			r600_context_pipe_state_set_vs_resource(&rctx->ctx, &resource[i]->state, i + PIPE_MAX_ATTRIBS);
 		}
 	}
 }
 
-static void r600_bind_state(struct pipe_context *ctx, void *state)
+static void r600_set_ps_sampler_view(struct pipe_context *ctx, unsigned count,
+					struct pipe_sampler_view **views)
 {
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_state *rstate = (struct r600_pipe_state *)state;
+	struct r600_pipe_sampler_view **resource = (struct r600_pipe_sampler_view **)views;
+	int i;
 
-	if (state == NULL)
-		return;
-	rctx->states[rstate->id] = rstate;
-	r600_context_pipe_state_set(&rctx->ctx, rstate);
+	for (i = 0; i < count; i++) {
+		if (&rctx->ps_samplers.views[i]->base != views[i]) {
+			if (resource[i])
+				r600_context_pipe_state_set_ps_resource(&rctx->ctx, &resource[i]->state, i);
+			else
+				r600_context_pipe_state_set_ps_resource(&rctx->ctx, NULL, i);
+
+			pipe_sampler_view_reference(
+				(struct pipe_sampler_view **)&rctx->ps_samplers.views[i],
+				views[i]);
+
+		}
+	}
+	for (i = count; i < NUM_TEX_UNITS; i++) {
+		if (rctx->ps_samplers.views[i]) {
+			r600_context_pipe_state_set_ps_resource(&rctx->ctx, NULL, i);
+			pipe_sampler_view_reference((struct pipe_sampler_view **)&rctx->ps_samplers.views[i], NULL);
+		}
+	}
+	rctx->ps_samplers.n_views = count;
 }
 
 static void r600_bind_ps_sampler(struct pipe_context *ctx, unsigned count, void **states)
@@ -716,6 +656,9 @@ static void r600_bind_ps_sampler(struct pipe_context *ctx, unsigned count, void
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
 	struct r600_pipe_state **rstates = (struct r600_pipe_state **)states;
 
+	memcpy(rctx->ps_samplers.samplers, states, sizeof(void*) * count);
+	rctx->ps_samplers.n_samplers = count;
+
 	for (int i = 0; i < count; i++) {
 		r600_context_pipe_state_set_ps_sampler(&rctx->ctx, rstates[i], i);
 	}
@@ -726,37 +669,11 @@ static void r600_bind_vs_sampler(struct pipe_context *ctx, unsigned count, void
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
 	struct r600_pipe_state **rstates = (struct r600_pipe_state **)states;
 
-	/* TODO implement */
 	for (int i = 0; i < count; i++) {
 		r600_context_pipe_state_set_vs_sampler(&rctx->ctx, rstates[i], i);
 	}
 }
 
-static void r600_delete_state(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_state *rstate = (struct r600_pipe_state *)state;
-
-	if (rctx->states[rstate->id] == rstate) {
-		rctx->states[rstate->id] = NULL;
-	}
-	for (int i = 0; i < rstate->nregs; i++) {
-		r600_bo_reference(rctx->radeon, &rstate->regs[i].bo, NULL);
-	}
-	free(rstate);
-}
-
-static void r600_delete_vertex_element(struct pipe_context *ctx, void *state)
-{
-	struct r600_vertex_element *v = (struct r600_vertex_element*)state;
-
-	if (v == NULL)
-		return;
-	if (--v->refcount)
-		return;
-	free(v);
-}
-
 static void r600_set_clip_state(struct pipe_context *ctx,
 				const struct pipe_clip_state *state)
 {
@@ -792,19 +709,6 @@ static void r600_set_clip_state(struct pipe_context *ctx,
 	r600_context_pipe_state_set(&rctx->ctx, rstate);
 }
 
-static void r600_bind_vertex_elements(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_vertex_element *v = (struct r600_vertex_element*)state;
-
-	r600_delete_vertex_element(ctx, rctx->vertex_elements);
-	rctx->vertex_elements = v;
-	if (v) {
-		v->refcount++;
-//		rctx->vs_rebuild = TRUE;
-	}
-}
-
 static void r600_set_polygon_stipple(struct pipe_context *ctx,
 					 const struct pipe_poly_stipple *state)
 {
@@ -828,18 +732,6 @@ static void r600_set_scissor_state(struct pipe_context *ctx,
 	tl = S_028240_TL_X(state->minx) | S_028240_TL_Y(state->miny) | S_028240_WINDOW_OFFSET_DISABLE(1);
 	br = S_028244_BR_X(state->maxx) | S_028244_BR_Y(state->maxy);
 	r600_pipe_state_add_reg(rstate,
-				R_028030_PA_SC_SCREEN_SCISSOR_TL, tl,
-				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_028034_PA_SC_SCREEN_SCISSOR_BR, br,
-				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_028204_PA_SC_WINDOW_SCISSOR_TL, tl,
-				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_028208_PA_SC_WINDOW_SCISSOR_BR, br,
-				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
 				R_028210_PA_SC_CLIPRECT_0_TL, tl,
 				0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate,
@@ -863,17 +755,6 @@ static void r600_set_scissor_state(struct pipe_context *ctx,
 	r600_pipe_state_add_reg(rstate,
 				R_02822C_PA_SC_CLIPRECT_3_BR, br,
 				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_028200_PA_SC_WINDOW_OFFSET, 0x00000000,
-				0xFFFFFFFF, NULL);
-	r600_pipe_state_add_reg(rstate,
-				R_02820C_PA_SC_CLIPRECT_RULE, 0x0000FFFF,
-				0xFFFFFFFF, NULL);
-	if (rctx->family >= CHIP_RV770) {
-		r600_pipe_state_add_reg(rstate,
-					R_028230_PA_SC_EDGERULE, 0xAAAAAAAA,
-					0xFFFFFFFF, NULL);
-	}
 
 	free(rctx->states[R600_PIPE_STATE_SCISSOR]);
 	rctx->states[R600_PIPE_STATE_SCISSOR] = rstate;
@@ -937,6 +818,7 @@ static void r600_cb(struct r600_pipe_context *rctx, struct r600_pipe_state *rsta
 {
 	struct r600_resource_texture *rtex;
 	struct r600_resource *rbuffer;
+	struct r600_surface *surf;
 	unsigned level = state->cbufs[cb]->level;
 	unsigned pitch, slice;
 	unsigned color_info;
@@ -944,14 +826,15 @@ static void r600_cb(struct r600_pipe_context *rctx, struct r600_pipe_state *rsta
 	const struct util_format_description *desc;
 	struct r600_bo *bo[3];
 
+	surf = (struct r600_surface *)state->cbufs[cb];
 	rtex = (struct r600_resource_texture*)state->cbufs[cb]->texture;
 	rbuffer = &rtex->resource;
 	bo[0] = rbuffer->bo;
 	bo[1] = rbuffer->bo;
 	bo[2] = rbuffer->bo;
 
-	pitch = (rtex->pitch[level] / rtex->bpt) / 8 - 1;
-	slice = (rtex->pitch[level] / rtex->bpt) * state->cbufs[cb]->height / 64 - 1;
+	pitch = rtex->pitch_in_pixels[level] / 8 - 1;
+	slice = rtex->pitch_in_pixels[level] * surf->aligned_height / 64 - 1;
 	ntype = 0;
 	desc = util_format_description(rtex->resource.base.b.format);
 	if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB)
@@ -961,6 +844,7 @@ static void r600_cb(struct r600_pipe_context *rctx, struct r600_pipe_state *rsta
 	swap = r600_translate_colorswap(rtex->resource.base.b.format);
 	color_info = S_0280A0_FORMAT(format) |
 		S_0280A0_COMP_SWAP(swap) |
+		S_0280A0_ARRAY_MODE(rtex->array_mode[level]) |
 		S_0280A0_BLEND_CLAMP(1) |
 		S_0280A0_NUMBER_TYPE(ntype);
 	if (desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS) 
@@ -996,22 +880,25 @@ static void r600_db(struct r600_pipe_context *rctx, struct r600_pipe_state *rsta
 {
 	struct r600_resource_texture *rtex;
 	struct r600_resource *rbuffer;
+	struct r600_surface *surf;
 	unsigned level;
 	unsigned pitch, slice, format;
 
 	if (state->zsbuf == NULL)
 		return;
 
+	level = state->zsbuf->level;
+
+	surf = (struct r600_surface *)state->zsbuf;
 	rtex = (struct r600_resource_texture*)state->zsbuf->texture;
 	rtex->tiled = 1;
-	rtex->array_mode = 2;
+	rtex->array_mode[level] = 2;
 	rtex->tile_type = 1;
 	rtex->depth = 1;
 	rbuffer = &rtex->resource;
 
-	level = state->zsbuf->level;
-	pitch = (rtex->pitch[level] / rtex->bpt) / 8 - 1;
-	slice = (rtex->pitch[level] / rtex->bpt) * state->zsbuf->height / 64 - 1;
+	pitch = rtex->pitch_in_pixels[level] / 8 - 1;
+	slice = rtex->pitch_in_pixels[level] * surf->aligned_height / 64 - 1;
 	format = r600_translate_dbformat(state->zsbuf->texture->format);
 
 	r600_pipe_state_add_reg(rstate, R_02800C_DB_DEPTH_BASE,
@@ -1021,10 +908,10 @@ static void r600_db(struct r600_pipe_context *rctx, struct r600_pipe_state *rsta
 				0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_028004_DB_DEPTH_VIEW, 0x00000000, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_028010_DB_DEPTH_INFO,
-				S_028010_ARRAY_MODE(rtex->array_mode) | S_028010_FORMAT(format),
+				S_028010_ARRAY_MODE(rtex->array_mode[level]) | S_028010_FORMAT(format),
 				0xFFFFFFFF, rbuffer->bo);
 	r600_pipe_state_add_reg(rstate, R_028D34_DB_PREFETCH_LIMIT,
-				(state->zsbuf->height / 8) - 1, 0xFFFFFFFF, NULL);
+				(surf->aligned_height / 8) - 1, 0xFFFFFFFF, NULL);
 }
 
 static void r600_set_framebuffer_state(struct pipe_context *ctx,
@@ -1039,14 +926,9 @@ static void r600_set_framebuffer_state(struct pipe_context *ctx,
 
 	/* unreference old buffer and reference new one */
 	rstate->id = R600_PIPE_STATE_FRAMEBUFFER;
-	for (int i = 0; i < rctx->framebuffer.nr_cbufs; i++) {
-		pipe_surface_reference(&rctx->framebuffer.cbufs[i], NULL);
-	}
-	for (int i = 0; i < state->nr_cbufs; i++) {
-		pipe_surface_reference(&rctx->framebuffer.cbufs[i], state->cbufs[i]);
-	}
-	pipe_surface_reference(&rctx->framebuffer.zsbuf, state->zsbuf);
-	rctx->framebuffer = *state;
+
+	util_copy_framebuffer_state(&rctx->framebuffer, state);
+	
 	rctx->pframebuffer = &rctx->framebuffer;
 
 	/* build states */
@@ -1070,6 +952,18 @@ static void r600_set_framebuffer_state(struct pipe_context *ctx,
 	br = S_028244_BR_X(state->width) | S_028244_BR_Y(state->height);
 
 	r600_pipe_state_add_reg(rstate,
+				R_028030_PA_SC_SCREEN_SCISSOR_TL, tl,
+				0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate,
+				R_028034_PA_SC_SCREEN_SCISSOR_BR, br,
+				0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate,
+				R_028204_PA_SC_WINDOW_SCISSOR_TL, tl,
+				0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate,
+				R_028208_PA_SC_WINDOW_SCISSOR_BR, br,
+				0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate,
 				R_028240_PA_SC_GENERIC_SCISSOR_TL, tl,
 				0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate,
@@ -1081,6 +975,14 @@ static void r600_set_framebuffer_state(struct pipe_context *ctx,
 	r600_pipe_state_add_reg(rstate,
 				R_028254_PA_SC_VPORT_SCISSOR_0_BR, br,
 				0xFFFFFFFF, NULL);
+	r600_pipe_state_add_reg(rstate,
+				R_028200_PA_SC_WINDOW_OFFSET, 0x00000000,
+				0xFFFFFFFF, NULL);
+	if (rctx->family >= CHIP_RV770) {
+		r600_pipe_state_add_reg(rstate,
+					R_028230_PA_SC_EDGERULE, 0xAAAAAAAA,
+					0xFFFFFFFF, NULL);
+	}
 
 	r600_pipe_state_add_reg(rstate, R_0287A0_CB_SHADER_CONTROL,
 				shader_control, 0xFFFFFFFF, NULL);
@@ -1110,40 +1012,6 @@ static void r600_set_framebuffer_state(struct pipe_context *ctx,
 	r600_context_pipe_state_set(&rctx->ctx, rstate);
 }
 
-static void r600_set_index_buffer(struct pipe_context *ctx,
-				  const struct pipe_index_buffer *ib)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-
-	if (ib) {
-		pipe_resource_reference(&rctx->index_buffer.buffer, ib->buffer);
-		memcpy(&rctx->index_buffer, ib, sizeof(rctx->index_buffer));
-	} else {
-		pipe_resource_reference(&rctx->index_buffer.buffer, NULL);
-		memset(&rctx->index_buffer, 0, sizeof(rctx->index_buffer));
-	}
-
-	/* TODO make this more like a state */
-}
-
-static void r600_set_vertex_buffers(struct pipe_context *ctx, unsigned count,
-					const struct pipe_vertex_buffer *buffers)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-
-	for (int i = 0; i < rctx->nvertex_buffer; i++) {
-		pipe_resource_reference(&rctx->vertex_buffer[i].buffer, NULL);
-	}
-	memcpy(rctx->vertex_buffer, buffers, sizeof(struct pipe_vertex_buffer) * count);
-	for (int i = 0; i < count; i++) {
-		rctx->vertex_buffer[i].buffer = NULL;
-		if (r600_buffer_is_user_buffer(buffers[i].buffer))
-			rctx->any_user_vbs = TRUE;
-		pipe_resource_reference(&rctx->vertex_buffer[i].buffer, buffers[i].buffer);
-	}
-	rctx->nvertex_buffer = count;
-}
-
 static void r600_set_constant_buffer(struct pipe_context *ctx, uint shader, uint index,
 					struct pipe_resource *buffer)
 {
@@ -1179,59 +1047,6 @@ static void r600_set_constant_buffer(struct pipe_context *ctx, uint shader, uint
 	}
 }
 
-static void *r600_create_shader_state(struct pipe_context *ctx,
-					const struct pipe_shader_state *state)
-{
-	struct r600_pipe_shader *shader =  CALLOC_STRUCT(r600_pipe_shader);
-	int r;
-
-	r =  r600_pipe_shader_create(ctx, shader, state->tokens);
-	if (r) {
-		return NULL;
-	}
-	return shader;
-}
-
-static void r600_bind_ps_shader(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-
-	/* TODO delete old shader */
-	rctx->ps_shader = (struct r600_pipe_shader *)state;
-}
-
-static void r600_bind_vs_shader(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-
-	/* TODO delete old shader */
-	rctx->vs_shader = (struct r600_pipe_shader *)state;
-}
-
-static void r600_delete_ps_shader(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_shader *shader = (struct r600_pipe_shader *)state;
-
-	if (rctx->ps_shader == shader) {
-		rctx->ps_shader = NULL;
-	}
-	/* TODO proper delete */
-	free(shader);
-}
-
-static void r600_delete_vs_shader(struct pipe_context *ctx, void *state)
-{
-	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-	struct r600_pipe_shader *shader = (struct r600_pipe_shader *)state;
-
-	if (rctx->vs_shader == shader) {
-		rctx->vs_shader = NULL;
-	}
-	/* TODO proper delete */
-	free(shader);
-}
-
 void r600_init_state_functions(struct r600_pipe_context *rctx)
 {
 	rctx->context.create_blend_state = r600_create_blend_state;
@@ -1480,14 +1295,14 @@ void r600_init_config(struct r600_pipe_context *rctx)
 		r600_pipe_state_add_reg(rstate, R_009830_DB_DEBUG, 0x00000000, 0xFFFFFFFF, NULL);
 		r600_pipe_state_add_reg(rstate, R_009838_DB_WATERMARKS, 0x00420204, 0xFFFFFFFF, NULL);
 		r600_pipe_state_add_reg(rstate, R_0286C8_SPI_THREAD_GROUPING, 0x00000000, 0xFFFFFFFF, NULL);
-		r600_pipe_state_add_reg(rstate, R_028A4C_PA_SC_MODE_CNTL, 0x00514000, 0xFFFFFFFF, NULL);
+		r600_pipe_state_add_reg(rstate, R_028A4C_PA_SC_MODE_CNTL, 0x00514002, 0xFFFFFFFF, NULL);
 	} else {
 		r600_pipe_state_add_reg(rstate, R_008D8C_SQ_DYN_GPR_CNTL_PS_FLUSH_REQ, 0x00000000, 0xFFFFFFFF, NULL);
 		r600_pipe_state_add_reg(rstate, R_009508_TA_CNTL_AUX, 0x07000003, 0xFFFFFFFF, NULL);
 		r600_pipe_state_add_reg(rstate, R_009830_DB_DEBUG, 0x82000000, 0xFFFFFFFF, NULL);
 		r600_pipe_state_add_reg(rstate, R_009838_DB_WATERMARKS, 0x01020204, 0xFFFFFFFF, NULL);
 		r600_pipe_state_add_reg(rstate, R_0286C8_SPI_THREAD_GROUPING, 0x00000001, 0xFFFFFFFF, NULL);
-		r600_pipe_state_add_reg(rstate, R_028A4C_PA_SC_MODE_CNTL, 0x00004010, 0xFFFFFFFF, NULL);
+		r600_pipe_state_add_reg(rstate, R_028A4C_PA_SC_MODE_CNTL, 0x00004012, 0xFFFFFFFF, NULL);
 	}
 	r600_pipe_state_add_reg(rstate, R_0288A8_SQ_ESGS_RING_ITEMSIZE, 0x00000000, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_0288AC_SQ_GSVS_RING_ITEMSIZE, 0x00000000, 0xFFFFFFFF, NULL);
diff --git a/src/gallium/drivers/r600/r600_state_common.c b/src/gallium/drivers/r600/r600_state_common.c
new file mode 100644
index 00000000000..210420e823b
--- /dev/null
+++ b/src/gallium/drivers/r600/r600_state_common.c
@@ -0,0 +1,276 @@
+/*
+ * Copyright 2010 Red Hat Inc.
+ *           2010 Jerome Glisse
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * on the rights to use, copy, modify, merge, publish, distribute, sub
+ * license, and/or sell copies of the Software, and to permit persons to whom
+ * the Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * Authors: Dave Airlie <[email protected]>
+ *          Jerome Glisse <[email protected]>
+ */
+#include <util/u_memory.h>
+#include <util/u_format.h>
+#include <pipebuffer/pb_buffer.h>
+#include "r600_pipe.h"
+
+/* common state between evergreen and r600 */
+void r600_bind_blend_state(struct pipe_context *ctx, void *state)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+	struct r600_pipe_blend *blend = (struct r600_pipe_blend *)state;
+	struct r600_pipe_state *rstate;
+
+	if (state == NULL)
+		return;
+	rstate = &blend->rstate;
+	rctx->states[rstate->id] = rstate;
+	rctx->cb_target_mask = blend->cb_target_mask;
+	r600_context_pipe_state_set(&rctx->ctx, rstate);
+}
+
+void r600_bind_rs_state(struct pipe_context *ctx, void *state)
+{
+	struct r600_pipe_rasterizer *rs = (struct r600_pipe_rasterizer *)state;
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+
+	if (state == NULL)
+		return;
+
+	rctx->flatshade = rs->flatshade;
+	rctx->sprite_coord_enable = rs->sprite_coord_enable;
+	rctx->rasterizer = rs;
+
+	rctx->states[rs->rstate.id] = &rs->rstate;
+	r600_context_pipe_state_set(&rctx->ctx, &rs->rstate);
+}
+
+void r600_delete_rs_state(struct pipe_context *ctx, void *state)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+	struct r600_pipe_rasterizer *rs = (struct r600_pipe_rasterizer *)state;
+
+	if (rctx->rasterizer == rs) {
+		rctx->rasterizer = NULL;
+	}
+	if (rctx->states[rs->rstate.id] == &rs->rstate) {
+		rctx->states[rs->rstate.id] = NULL;
+	}
+	free(rs);
+}
+
+void r600_sampler_view_destroy(struct pipe_context *ctx,
+			       struct pipe_sampler_view *state)
+{
+	struct r600_pipe_sampler_view *resource = (struct r600_pipe_sampler_view *)state;
+
+	pipe_resource_reference(&state->texture, NULL);
+	FREE(resource);
+}
+
+void r600_bind_state(struct pipe_context *ctx, void *state)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+	struct r600_pipe_state *rstate = (struct r600_pipe_state *)state;
+
+	if (state == NULL)
+		return;
+	rctx->states[rstate->id] = rstate;
+	r600_context_pipe_state_set(&rctx->ctx, rstate);
+}
+
+void r600_delete_state(struct pipe_context *ctx, void *state)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+	struct r600_pipe_state *rstate = (struct r600_pipe_state *)state;
+
+	if (rctx->states[rstate->id] == rstate) {
+		rctx->states[rstate->id] = NULL;
+	}
+	for (int i = 0; i < rstate->nregs; i++) {
+		r600_bo_reference(rctx->radeon, &rstate->regs[i].bo, NULL);
+	}
+	free(rstate);
+}
+
+void r600_bind_vertex_elements(struct pipe_context *ctx, void *state)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+	struct r600_vertex_element *v = (struct r600_vertex_element*)state;
+
+	rctx->vertex_elements = v;
+	if (v) {
+//		rctx->vs_rebuild = TRUE;
+	}
+}
+
+void r600_delete_vertex_element(struct pipe_context *ctx, void *state)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+
+	FREE(state);
+
+	if (rctx->vertex_elements == state)
+		rctx->vertex_elements = NULL;
+}
+
+
+void r600_set_index_buffer(struct pipe_context *ctx,
+			   const struct pipe_index_buffer *ib)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+
+	if (ib) {
+		pipe_resource_reference(&rctx->index_buffer.buffer, ib->buffer);
+		memcpy(&rctx->index_buffer, ib, sizeof(rctx->index_buffer));
+	} else {
+		pipe_resource_reference(&rctx->index_buffer.buffer, NULL);
+		memset(&rctx->index_buffer, 0, sizeof(rctx->index_buffer));
+	}
+
+	/* TODO make this more like a state */
+}
+
+void r600_set_vertex_buffers(struct pipe_context *ctx, unsigned count,
+			     const struct pipe_vertex_buffer *buffers)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+	struct pipe_vertex_buffer *vbo;
+	unsigned max_index = (unsigned)-1;
+
+	for (int i = 0; i < rctx->nvertex_buffer; i++) {
+		pipe_resource_reference(&rctx->vertex_buffer[i].buffer, NULL);
+	}
+	memcpy(rctx->vertex_buffer, buffers, sizeof(struct pipe_vertex_buffer) * count);
+
+	for (int i = 0; i < count; i++) {
+		vbo = (struct pipe_vertex_buffer*)&buffers[i];
+
+		rctx->vertex_buffer[i].buffer = NULL;
+		if (r600_buffer_is_user_buffer(buffers[i].buffer))
+			rctx->any_user_vbs = TRUE;
+		pipe_resource_reference(&rctx->vertex_buffer[i].buffer, buffers[i].buffer);
+
+		if (vbo->max_index == ~0) {
+			if (!vbo->stride)
+				vbo->max_index = 1;
+			else
+				vbo->max_index = (vbo->buffer->width0 - vbo->buffer_offset) / vbo->stride;
+		}
+		max_index = MIN2(vbo->max_index, max_index);
+	}
+	rctx->nvertex_buffer = count;
+	rctx->vb_max_index = max_index;
+}
+
+
+#define FORMAT_REPLACE(what, withwhat) \
+	case PIPE_FORMAT_##what: *format = PIPE_FORMAT_##withwhat; break
+
+void *r600_create_vertex_elements(struct pipe_context *ctx,
+				  unsigned count,
+				  const struct pipe_vertex_element *elements)
+{
+	struct r600_vertex_element *v = CALLOC_STRUCT(r600_vertex_element);
+	int i;
+	enum pipe_format *format;
+
+	assert(count < 32);
+	if (!v)
+		return NULL;
+
+	v->count = count;
+	memcpy(v->elements, elements, count * sizeof(struct pipe_vertex_element));
+
+	for (i = 0; i < count; i++) {
+		v->hw_format[i] = v->elements[i].src_format;
+		format = &v->hw_format[i];
+
+		switch (*format) {
+                    FORMAT_REPLACE(R64_FLOAT,           R32_FLOAT);
+                    FORMAT_REPLACE(R64G64_FLOAT,        R32G32_FLOAT);
+                    FORMAT_REPLACE(R64G64B64_FLOAT,     R32G32B32_FLOAT);
+                    FORMAT_REPLACE(R64G64B64A64_FLOAT,  R32G32B32A32_FLOAT);
+		default:;
+		}
+		v->incompatible_layout =
+			v->incompatible_layout ||
+			v->elements[i].src_format != v->hw_format[i] ||
+			v->elements[i].src_offset % 4 != 0;
+
+                v->hw_format_size[i] =
+			align(util_format_get_blocksize(v->hw_format[i]), 4);
+	}
+
+	return v;
+}
+
+void *r600_create_shader_state(struct pipe_context *ctx,
+			       const struct pipe_shader_state *state)
+{
+	struct r600_pipe_shader *shader =  CALLOC_STRUCT(r600_pipe_shader);
+	int r;
+
+	r =  r600_pipe_shader_create(ctx, shader, state->tokens);
+	if (r) {
+		return NULL;
+	}
+	return shader;
+}
+
+void r600_bind_ps_shader(struct pipe_context *ctx, void *state)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+
+	/* TODO delete old shader */
+	rctx->ps_shader = (struct r600_pipe_shader *)state;
+}
+
+void r600_bind_vs_shader(struct pipe_context *ctx, void *state)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+
+	/* TODO delete old shader */
+	rctx->vs_shader = (struct r600_pipe_shader *)state;
+}
+
+void r600_delete_ps_shader(struct pipe_context *ctx, void *state)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+	struct r600_pipe_shader *shader = (struct r600_pipe_shader *)state;
+
+	if (rctx->ps_shader == shader) {
+		rctx->ps_shader = NULL;
+	}
+
+	r600_pipe_shader_destroy(ctx, shader);
+	free(shader);
+}
+
+void r600_delete_vs_shader(struct pipe_context *ctx, void *state)
+{
+	struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
+	struct r600_pipe_shader *shader = (struct r600_pipe_shader *)state;
+
+	if (rctx->vs_shader == shader) {
+		rctx->vs_shader = NULL;
+	}
+
+	r600_pipe_shader_destroy(ctx, shader);
+	free(shader);
+}
diff --git a/src/gallium/drivers/r600/r600_state_inlines.h b/src/gallium/drivers/r600/r600_state_inlines.h
index 29a29d87cd3..1c1978f8abb 100644
--- a/src/gallium/drivers/r600/r600_state_inlines.h
+++ b/src/gallium/drivers/r600/r600_state_inlines.h
@@ -25,6 +25,7 @@
 
 #include "util/u_format.h"
 #include "r600d.h"
+#include "r600_formats.h"
 
 static INLINE uint32_t r600_translate_blend_function(int blend_func)
 {
@@ -352,13 +353,13 @@ static inline uint32_t r600_translate_colorswap(enum pipe_format format)
 		/* 64-bit buffers. */
 	case PIPE_FORMAT_R16G16B16A16_UNORM:
 	case PIPE_FORMAT_R16G16B16A16_SNORM:
-		//		return V_0280A0_COLOR_16_16_16_16;
+		//		return FMT_16_16_16_16;
 	case PIPE_FORMAT_R16G16B16A16_FLOAT:
-		//		return V_0280A0_COLOR_16_16_16_16_FLOAT;
+		//		return FMT_16_16_16_16_FLOAT;
 
 		/* 128-bit buffers. */
 	case PIPE_FORMAT_R32G32B32A32_FLOAT:
-		//		return V_0280A0_COLOR_32_32_32_32_FLOAT;
+		//		return FMT_32_32_32_32_FLOAT;
 		return 0;
 	default:
 		R600_ERR("unsupported colorswap format %d\n", format);
@@ -522,32 +523,32 @@ static INLINE uint32_t r600_translate_vertex_data_type(enum pipe_format format)
                 case 16:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_038008_FMT_16_FLOAT;
+				result = FMT_16_FLOAT;
 				break;
 			case 2:
-				result = V_038008_FMT_16_16_FLOAT;
+				result = FMT_16_16_FLOAT;
 				break;
 			case 3:
-				result = V_038008_FMT_16_16_16_FLOAT;
+				result = FMT_16_16_16_FLOAT;
 				break;
 			case 4:
-				result = V_038008_FMT_16_16_16_16_FLOAT;
+				result = FMT_16_16_16_16_FLOAT;
 				break;
 			}
 			break;
                 case 32:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_038008_FMT_32_FLOAT;
+				result = FMT_32_FLOAT;
 				break;
 			case 2:
-				result = V_038008_FMT_32_32_FLOAT;
+				result = FMT_32_32_FLOAT;
 				break;
 			case 3:
-				result = V_038008_FMT_32_32_32_FLOAT;
+				result = FMT_32_32_32_FLOAT;
 				break;
 			case 4:
-				result = V_038008_FMT_32_32_32_32_FLOAT;
+				result = FMT_32_32_32_32_FLOAT;
 				break;
 			}
 			break;
@@ -563,48 +564,48 @@ static INLINE uint32_t r600_translate_vertex_data_type(enum pipe_format format)
                 case 8:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_038008_FMT_8;
+				result = FMT_8;
 				break;
 			case 2:
-				result = V_038008_FMT_8_8;
+				result = FMT_8_8;
 				break;
 			case 3:
-			//	result = V_038008_FMT_8_8_8; /* fails piglit draw-vertices test */
+			//	result = FMT_8_8_8; /* fails piglit draw-vertices test */
 			//	break;
 			case 4:
-				result = V_038008_FMT_8_8_8_8;
+				result = FMT_8_8_8_8;
 				break;
 			}
 			break;
                 case 16:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_038008_FMT_16;
+				result = FMT_16;
 				break;
 			case 2:
-				result = V_038008_FMT_16_16;
+				result = FMT_16_16;
 				break;
 			case 3:
-			//	result = V_038008_FMT_16_16_16; /* fails piglit draw-vertices test */
+			//	result = FMT_16_16_16; /* fails piglit draw-vertices test */
 			//	break;
 			case 4:
-				result = V_038008_FMT_16_16_16_16;
+				result = FMT_16_16_16_16;
 				break;
 			}
 			break;
                 case 32:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_038008_FMT_32;
+				result = FMT_32;
 				break;
 			case 2:
-				result = V_038008_FMT_32_32;
+				result = FMT_32_32;
 				break;
 			case 3:
-				result = V_038008_FMT_32_32_32;
+				result = FMT_32_32_32;
 				break;
 			case 4:
-				result = V_038008_FMT_32_32_32_32;
+				result = FMT_32_32_32_32;
 				break;
 			}
 			break;
diff --git a/src/gallium/drivers/r600/r600_texture.c b/src/gallium/drivers/r600/r600_texture.c
index db88346afbe..4ebd5b754b3 100644
--- a/src/gallium/drivers/r600/r600_texture.c
+++ b/src/gallium/drivers/r600/r600_texture.c
@@ -35,6 +35,7 @@
 #include "r600_resource.h"
 #include "r600_state_inlines.h"
 #include "r600d.h"
+#include "r600_formats.h"
 
 extern struct u_resource_vtbl r600_texture_vtbl;
 
@@ -53,11 +54,30 @@ static void r600_copy_from_tiled_texture(struct pipe_context *ctx, struct r600_t
 				transfer->box.width, transfer->box.height);
 }
 
-static unsigned long r600_texture_get_offset(struct r600_resource_texture *rtex,
+
+/* Copy from a detiled texture to a tiled one. */
+static void r600_copy_into_tiled_texture(struct pipe_context *ctx, struct r600_transfer *rtransfer)
+{
+	struct pipe_transfer *transfer = (struct pipe_transfer*)rtransfer;
+	struct pipe_resource *texture = transfer->resource;
+	struct pipe_subresource subsrc;
+
+	subsrc.face = 0;
+	subsrc.level = 0;
+	ctx->resource_copy_region(ctx, texture, transfer->sr,
+				  transfer->box.x, transfer->box.y, transfer->box.z,
+				  rtransfer->linear_texture, subsrc,
+				  0, 0, 0,
+				  transfer->box.width, transfer->box.height);
+
+	ctx->flush(ctx, 0, NULL);
+}
+
+static unsigned r600_texture_get_offset(struct r600_resource_texture *rtex,
 					unsigned level, unsigned zslice,
 					unsigned face)
 {
-	unsigned long offset = rtex->offset[level];
+	unsigned offset = rtex->offset[level];
 
 	switch (rtex->resource.base.b.target) {
 	case PIPE_TEXTURE_3D:
@@ -72,22 +92,158 @@ static unsigned long r600_texture_get_offset(struct r600_resource_texture *rtex,
 	}
 }
 
-static void r600_setup_miptree(struct r600_resource_texture *rtex, enum chip_class chipc)
+static unsigned r600_get_pixel_alignment(struct pipe_screen *screen,
+					 enum pipe_format format,
+					 unsigned array_mode)
+{
+	struct r600_screen* rscreen = (struct r600_screen *)screen;
+	unsigned pixsize = util_format_get_blocksize(format);
+	int p_align;
+
+	switch(array_mode) {
+	case V_038000_ARRAY_1D_TILED_THIN1:
+		p_align = MAX2(8,
+			       ((rscreen->tiling_info->group_bytes / 8 / pixsize)));
+		break;
+	case V_038000_ARRAY_2D_TILED_THIN1:
+		p_align = MAX2(rscreen->tiling_info->num_banks,
+			       (((rscreen->tiling_info->group_bytes / 8 / pixsize)) *
+				rscreen->tiling_info->num_banks));
+		break;
+	case 0:
+	default:
+		p_align = 64;
+		break;
+	}
+	return p_align;
+}
+
+static unsigned r600_get_height_alignment(struct pipe_screen *screen,
+					  unsigned array_mode)
+{
+	struct r600_screen* rscreen = (struct r600_screen *)screen;
+	int h_align;
+
+	switch (array_mode) {
+	case V_038000_ARRAY_2D_TILED_THIN1:
+		h_align = rscreen->tiling_info->num_channels * 8;
+		break;
+	case V_038000_ARRAY_1D_TILED_THIN1:
+		h_align = 8;
+		break;
+	default:
+		h_align = 1;
+		break;
+	}
+	return h_align;
+}
+
+static unsigned mip_minify(unsigned size, unsigned level)
+{
+	unsigned val;
+	val = u_minify(size, level);
+	if (level > 0)
+		val = util_next_power_of_two(val);
+	return val;
+}
+
+static unsigned r600_texture_get_stride(struct pipe_screen *screen,
+					struct r600_resource_texture *rtex,
+					unsigned level)
+{
+	struct pipe_resource *ptex = &rtex->resource.base.b;
+	struct radeon *radeon = (struct radeon *)screen->winsys;
+	enum chip_class chipc = r600_get_family_class(radeon);
+	unsigned width, stride, tile_width;
+	
+	if (rtex->pitch_override)
+		return rtex->pitch_override;
+
+	width = mip_minify(ptex->width0, level);
+	if (util_format_is_plain(ptex->format)) {
+		tile_width = r600_get_pixel_alignment(screen, ptex->format,
+						      rtex->array_mode[level]);
+		width = align(width, tile_width);
+	}
+	stride = util_format_get_stride(ptex->format, width);
+	if (chipc == EVERGREEN)
+		stride = align(stride, 512);
+	return stride;
+}
+
+static unsigned r600_texture_get_nblocksy(struct pipe_screen *screen,
+					  struct r600_resource_texture *rtex,
+					  unsigned level)
 {
 	struct pipe_resource *ptex = &rtex->resource.base.b;
-	unsigned long w, h, pitch, size, layer_size, i, offset;
+	unsigned height, tile_height;
 
-	rtex->bpt = util_format_get_blocksize(ptex->format);
-	for (i = 0, offset = 0; i <= ptex->last_level; i++) {
-		w = u_minify(ptex->width0, i);
-		h = u_minify(ptex->height0, i);
-		h = util_next_power_of_two(h);
-		pitch = util_format_get_stride(ptex->format, align(w, 64));
-		if (chipc == EVERGREEN)
-			pitch = align(pitch, 512);
+	height = mip_minify(ptex->height0, level);
+	if (util_format_is_plain(ptex->format)) {
+		tile_height = r600_get_height_alignment(screen,
+							rtex->array_mode[level]);
+		height = align(height, tile_height);
+	}
+	return util_format_get_nblocksy(ptex->format, height);
+}
+
+/* Get a width in pixels from a stride in bytes. */
+static unsigned pitch_to_width(enum pipe_format format,
+                                unsigned pitch_in_bytes)
+{
+    return (pitch_in_bytes / util_format_get_blocksize(format)) *
+            util_format_get_blockwidth(format);
+}
+
+static void r600_texture_set_array_mode(struct pipe_screen *screen,
+					struct r600_resource_texture *rtex,
+					unsigned level, unsigned array_mode)
+{
+	struct pipe_resource *ptex = &rtex->resource.base.b;
+
+	switch (array_mode) {
+	case V_0280A0_ARRAY_LINEAR_GENERAL:
+	case V_0280A0_ARRAY_LINEAR_ALIGNED:
+	case V_0280A0_ARRAY_1D_TILED_THIN1:
+	default:
+		rtex->array_mode[level] = array_mode;
+		break;
+	case V_0280A0_ARRAY_2D_TILED_THIN1:
+	{
+		unsigned w, h, tile_height, tile_width;
+
+		tile_height = r600_get_height_alignment(screen, array_mode);
+		tile_width = r600_get_pixel_alignment(screen, ptex->format, array_mode);
+
+		w = mip_minify(ptex->width0, level);
+		h = mip_minify(ptex->height0, level);
+		if (w < tile_width || h < tile_height)
+			rtex->array_mode[level] = V_0280A0_ARRAY_1D_TILED_THIN1;
 		else
-			pitch = align(pitch, 256);
-		layer_size = pitch * h;
+			rtex->array_mode[level] = array_mode;
+	}
+	break;
+	}
+}
+
+static void r600_setup_miptree(struct pipe_screen *screen,
+			       struct r600_resource_texture *rtex,
+			       unsigned array_mode)
+{
+	struct pipe_resource *ptex = &rtex->resource.base.b;
+	struct radeon *radeon = (struct radeon *)screen->winsys;
+	enum chip_class chipc = r600_get_family_class(radeon);
+	unsigned pitch, size, layer_size, i, offset;
+	unsigned nblocksy;
+
+	for (i = 0, offset = 0; i <= ptex->last_level; i++) {
+		r600_texture_set_array_mode(screen, rtex, i, array_mode);
+
+		pitch = r600_texture_get_stride(screen, rtex, i);
+		nblocksy = r600_texture_get_nblocksy(screen, rtex, i);
+
+		layer_size = pitch * nblocksy;
+
 		if (ptex->target == PIPE_TEXTURE_CUBE) {
 			if (chipc >= R700)
 				size = layer_size * 8;
@@ -98,39 +254,69 @@ static void r600_setup_miptree(struct r600_resource_texture *rtex, enum chip_cla
 			size = layer_size * u_minify(ptex->depth0, i);
 		rtex->offset[i] = offset;
 		rtex->layer_size[i] = layer_size;
-		rtex->pitch[i] = pitch;
+		rtex->pitch_in_bytes[i] = pitch;
+		rtex->pitch_in_pixels[i] = pitch_to_width(ptex->format, pitch);
 		offset += size;
 	}
 	rtex->size = offset;
 }
 
-struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
-						const struct pipe_resource *templ)
+static struct r600_resource_texture *
+r600_texture_create_object(struct pipe_screen *screen,
+			   const struct pipe_resource *base,
+			   unsigned array_mode,
+			   unsigned pitch_in_bytes_override,
+			   unsigned max_buffer_size,
+			   struct r600_bo *bo)
 {
 	struct r600_resource_texture *rtex;
 	struct r600_resource *resource;
 	struct radeon *radeon = (struct radeon *)screen->winsys;
 
 	rtex = CALLOC_STRUCT(r600_resource_texture);
-	if (!rtex) {
+	if (rtex == NULL)
 		return NULL;
-	}
+
 	resource = &rtex->resource;
-	resource->base.b = *templ;
+	resource->base.b = *base;
 	resource->base.vtbl = &r600_texture_vtbl;
 	pipe_reference_init(&resource->base.b.reference, 1);
 	resource->base.b.screen = screen;
-	r600_setup_miptree(rtex, r600_get_family_class(radeon));
-
-	/* FIXME alignment 4096 enought ? too much ? */
+	resource->bo = bo;
 	resource->domain = r600_domain_from_usage(resource->base.b.bind);
+	rtex->pitch_override = pitch_in_bytes_override;
+
+	if (array_mode)
+		rtex->tiled = 1;
+	r600_setup_miptree(screen, rtex, array_mode);
+
 	resource->size = rtex->size;
-	resource->bo = r600_bo(radeon, rtex->size, 4096, 0);
-	if (resource->bo == NULL) {
-		FREE(rtex);
-		return NULL;
+
+	if (!resource->bo) {
+		resource->bo = r600_bo(radeon, rtex->size, 4096, 0);
+		if (!resource->bo) {
+			FREE(rtex);
+			return NULL;
+		}
+	}
+	return rtex;
+}
+
+struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
+						const struct pipe_resource *templ)
+{
+	unsigned array_mode = 0;
+
+	if (debug_get_bool_option("R600_FORCE_TILING", FALSE)) {
+		if (!(templ->flags & R600_RESOURCE_FLAG_TRANSFER) &&
+		    !(templ->bind & PIPE_BIND_SCANOUT)) {
+			array_mode = V_038000_ARRAY_2D_TILED_THIN1;
+		}
 	}
-	return &resource->base.b;
+
+	return (struct pipe_resource *)r600_texture_create_object(screen, templ, array_mode,
+								  0, 0, NULL);
+
 }
 
 static void r600_texture_destroy(struct pipe_screen *screen,
@@ -155,24 +341,27 @@ static struct pipe_surface *r600_get_tex_surface(struct pipe_screen *screen,
 						unsigned zslice, unsigned flags)
 {
 	struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
-	struct pipe_surface *surface = CALLOC_STRUCT(pipe_surface);
-	unsigned long offset;
+	struct r600_surface *surface = CALLOC_STRUCT(r600_surface);
+	unsigned offset, tile_height;
 
 	if (surface == NULL)
 		return NULL;
 	offset = r600_texture_get_offset(rtex, level, zslice, face);
-	pipe_reference_init(&surface->reference, 1);
-	pipe_resource_reference(&surface->texture, texture);
-	surface->format = texture->format;
-	surface->width = u_minify(texture->width0, level);
-	surface->height = u_minify(texture->height0, level);
-	surface->offset = offset;
-	surface->usage = flags;
-	surface->zslice = zslice;
-	surface->texture = texture;
-	surface->face = face;
-	surface->level = level;
-	return surface;
+	pipe_reference_init(&surface->base.reference, 1);
+	pipe_resource_reference(&surface->base.texture, texture);
+	surface->base.format = texture->format;
+	surface->base.width = mip_minify(texture->width0, level);
+	surface->base.height = mip_minify(texture->height0, level);
+	surface->base.offset = offset;
+	surface->base.usage = flags;
+	surface->base.zslice = zslice;
+	surface->base.texture = texture;
+	surface->base.face = face;
+	surface->base.level = level;
+
+	tile_height = r600_get_height_alignment(screen, rtex->array_mode[level]);
+	surface->aligned_height = align(surface->base.height, tile_height);
+	return &surface->base;
 }
 
 static void r600_tex_surface_destroy(struct pipe_surface *surface)
@@ -181,44 +370,29 @@ static void r600_tex_surface_destroy(struct pipe_surface *surface)
 	FREE(surface);
 }
 
+
 struct pipe_resource *r600_texture_from_handle(struct pipe_screen *screen,
 					       const struct pipe_resource *templ,
 					       struct winsys_handle *whandle)
 {
 	struct radeon *rw = (struct radeon*)screen->winsys;
-	struct r600_resource_texture *rtex;
-	struct r600_resource *resource;
 	struct r600_bo *bo = NULL;
+	unsigned array_mode = 0;
 
 	/* Support only 2D textures without mipmaps */
 	if ((templ->target != PIPE_TEXTURE_2D && templ->target != PIPE_TEXTURE_RECT) ||
 	      templ->depth0 != 1 || templ->last_level != 0)
 		return NULL;
 
-	rtex = CALLOC_STRUCT(r600_resource_texture);
-	if (rtex == NULL)
-		return NULL;
-
-	bo = r600_bo_handle(rw, whandle->handle);
+	bo = r600_bo_handle(rw, whandle->handle, &array_mode);
 	if (bo == NULL) {
-		FREE(rtex);
 		return NULL;
 	}
 
-	resource = &rtex->resource;
-	resource->base.b = *templ;
-	resource->base.vtbl = &r600_texture_vtbl;
-	pipe_reference_init(&resource->base.b.reference, 1);
-	resource->base.b.screen = screen;
-	resource->bo = bo;
-	rtex->depth = 0;
-	rtex->pitch_override = whandle->stride;
-	rtex->bpt = util_format_get_blocksize(templ->format);
-	rtex->pitch[0] = whandle->stride;
-	rtex->offset[0] = 0;
-	rtex->size = align(rtex->pitch[0] * templ->height0, 64);
-
-	return &resource->base.b;
+	return (struct pipe_resource *)r600_texture_create_object(screen, templ, array_mode,
+								  whandle->stride,
+								  0,
+								  bo);
 }
 
 static unsigned int r600_texture_is_referenced(struct pipe_context *context,
@@ -244,14 +418,14 @@ int r600_texture_depth_flush(struct pipe_context *ctx,
 	resource.format = texture->format;
 	resource.width0 = texture->width0;
 	resource.height0 = texture->height0;
-	resource.depth0 = 0;
+	resource.depth0 = 1;
 	resource.last_level = 0;
 	resource.nr_samples = 0;
 	resource.usage = PIPE_USAGE_DYNAMIC;
 	resource.bind = 0;
-	resource.flags = 0;
+	resource.flags = R600_RESOURCE_FLAG_TRANSFER;
 
-	resource.bind |= PIPE_BIND_RENDER_TARGET;
+	resource.bind |= PIPE_BIND_DEPTH_STENCIL;
 
 	rtex->flushed_depth_texture = (struct r600_resource_texture *)ctx->screen->resource_create(ctx->screen, &resource);
 	if (rtex->flushed_depth_texture == NULL) {
@@ -282,8 +456,6 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
 	trans->transfer.sr = sr;
 	trans->transfer.usage = usage;
 	trans->transfer.box = *box;
-	trans->transfer.stride = rtex->pitch[sr.level];
-	trans->offset = r600_texture_get_offset(rtex, sr.level, box->z, sr.face);
 	if (rtex->depth) {
 		r = r600_texture_depth_flush(ctx, texture);
 		if (r < 0) {
@@ -297,12 +469,12 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
 		resource.format = texture->format;
 		resource.width0 = box->width;
 		resource.height0 = box->height;
-		resource.depth0 = 0;
+		resource.depth0 = 1;
 		resource.last_level = 0;
 		resource.nr_samples = 0;
 		resource.usage = PIPE_USAGE_DYNAMIC;
 		resource.bind = 0;
-		resource.flags = 0;
+		resource.flags = R600_RESOURCE_FLAG_TRANSFER;
 		/* For texture reading, the temporary (detiled) texture is used as
 		 * a render target when blitting from a tiled texture. */
 		if (usage & PIPE_TRANSFER_READ) {
@@ -321,6 +493,9 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
 			FREE(trans);
 			return NULL;
 		}
+
+		trans->transfer.stride =
+		  ((struct r600_resource_texture *)trans->linear_texture)->pitch_in_bytes[0];
 		if (usage & PIPE_TRANSFER_READ) {
 			/* We cannot map a tiled texture directly because the data is
 			 * in a different order, therefore we do detiling using a blit. */
@@ -328,7 +503,10 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
 			/* Always referenced in the blit. */
 			ctx->flush(ctx, 0, NULL);
 		}
+		return &trans->transfer;
 	}
+	trans->transfer.stride = rtex->pitch_in_bytes[sr.level];
+	trans->offset = r600_texture_get_offset(rtex, sr.level, box->z, sr.face);
 	return &trans->transfer;
 }
 
@@ -339,12 +517,12 @@ void r600_texture_transfer_destroy(struct pipe_context *ctx,
 	struct r600_resource_texture *rtex = (struct r600_resource_texture*)transfer->resource;
 
 	if (rtransfer->linear_texture) {
+		if (transfer->usage & PIPE_TRANSFER_WRITE) {
+			r600_copy_into_tiled_texture(ctx, rtransfer);
+		}
 		pipe_resource_reference(&rtransfer->linear_texture, NULL);
 	}
 	if (rtex->flushed_depth_texture) {
-		if (transfer->usage & PIPE_TRANSFER_WRITE) {
-			// TODO
-		}
 		pipe_resource_reference((struct pipe_resource **)&rtex->flushed_depth_texture, NULL);
 	}
 	pipe_resource_reference(&transfer->resource, NULL);
@@ -358,7 +536,7 @@ void* r600_texture_transfer_map(struct pipe_context *ctx,
 	struct r600_bo *bo;
 	enum pipe_format format = transfer->resource->format;
 	struct radeon *radeon = (struct radeon *)ctx->screen->winsys;
-	unsigned long offset = 0;
+	unsigned offset = 0;
 	char *map;
 
 	if (rtransfer->linear_texture) {
@@ -496,15 +674,23 @@ uint32_t r600_translate_texformat(enum pipe_format format,
 	case UTIL_FORMAT_COLORSPACE_ZS:
 		switch (format) {
 		case PIPE_FORMAT_Z16_UNORM:
-			result = V_0280A0_COLOR_16;
+			result = FMT_16;
 			goto out_word4;
+		case PIPE_FORMAT_X24S8_USCALED:
+			word4 |= S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_INT);
 		case PIPE_FORMAT_Z24X8_UNORM:
 		case PIPE_FORMAT_Z24_UNORM_S8_USCALED:
-			result = V_0280A0_COLOR_8_24;
+			result = FMT_8_24;
 			goto out_word4;
+		case PIPE_FORMAT_S8X24_USCALED:
+			word4 |= S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_INT);
 		case PIPE_FORMAT_X8Z24_UNORM:
 		case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
-			result = V_0280A0_COLOR_24_8;
+			result = FMT_24_8;
+			goto out_word4;
+		case PIPE_FORMAT_S8_USCALED:
+			result = V_0280A0_COLOR_8;
+			word4 |= S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_INT);
 			goto out_word4;
 		default:
 			goto out_unknown;
@@ -558,7 +744,7 @@ uint32_t r600_translate_texformat(enum pipe_format format,
 			if (desc->channel[0].size == 5 &&
 			    desc->channel[1].size == 6 &&
 			    desc->channel[2].size == 5) {
-				result = V_0280A0_COLOR_5_6_5;
+				result = FMT_5_6_5;
 				goto out_word4;
 			}
 			goto out_unknown;
@@ -567,14 +753,14 @@ uint32_t r600_translate_texformat(enum pipe_format format,
 			    desc->channel[1].size == 5 &&
 			    desc->channel[2].size == 5 &&
 			    desc->channel[3].size == 1) {
-				result = V_0280A0_COLOR_1_5_5_5;
+				result = FMT_1_5_5_5;
 				goto out_word4;
 			}
 			if (desc->channel[0].size == 10 &&
 			    desc->channel[1].size == 10 &&
 			    desc->channel[2].size == 10 &&
 			    desc->channel[3].size == 2) {
-				result = V_0280A0_COLOR_10_10_10_2;
+				result = FMT_10_10_10_2;
 				goto out_word4;
 			}
 			goto out_unknown;
@@ -605,36 +791,36 @@ uint32_t r600_translate_texformat(enum pipe_format format,
 		case 4:
 			switch (desc->nr_channels) {
 			case 2:
-				result = V_0280A0_COLOR_4_4;
+				result = FMT_4_4;
 				goto out_word4;
 			case 4:
-				result = V_0280A0_COLOR_4_4_4_4;
+				result = FMT_4_4_4_4;
 				goto out_word4;
 			}
 			goto out_unknown;
 		case 8:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_0280A0_COLOR_8;
+				result = FMT_8;
 				goto out_word4;
 			case 2:
-				result = V_0280A0_COLOR_8_8;
+				result = FMT_8_8;
 				goto out_word4;
 			case 4:
-				result = V_0280A0_COLOR_8_8_8_8;
+				result = FMT_8_8_8_8;
 				goto out_word4;
 			}
 			goto out_unknown;
 		case 16:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_0280A0_COLOR_16;
+				result = FMT_16;
 				goto out_word4;
 			case 2:
-				result = V_0280A0_COLOR_16_16;
+				result = FMT_16_16;
 				goto out_word4;
 			case 4:
-				result = V_0280A0_COLOR_16_16_16_16;
+				result = FMT_16_16_16_16;
 				goto out_word4;
 			}
 		}
@@ -645,26 +831,26 @@ uint32_t r600_translate_texformat(enum pipe_format format,
 		case 16:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_0280A0_COLOR_16_FLOAT;
+				result = FMT_16_FLOAT;
 				goto out_word4;
 			case 2:
-				result = V_0280A0_COLOR_16_16_FLOAT;
+				result = FMT_16_16_FLOAT;
 				goto out_word4;
 			case 4:
-				result = V_0280A0_COLOR_16_16_16_16_FLOAT;
+				result = FMT_16_16_16_16_FLOAT;
 				goto out_word4;
 			}
 			goto out_unknown;
 		case 32:
 			switch (desc->nr_channels) {
 			case 1:
-				result = V_0280A0_COLOR_32_FLOAT;
+				result = FMT_32_FLOAT;
 				goto out_word4;
 			case 2:
-				result = V_0280A0_COLOR_32_32_FLOAT;
+				result = FMT_32_32_FLOAT;
 				goto out_word4;
 			case 4:
-				result = V_0280A0_COLOR_32_32_32_32_FLOAT;
+				result = FMT_32_32_32_32_FLOAT;
 				goto out_word4;
 			}
 		}
diff --git a/src/gallium/drivers/r600/r600_translate.c b/src/gallium/drivers/r600/r600_translate.c
new file mode 100644
index 00000000000..9a07cf2073f
--- /dev/null
+++ b/src/gallium/drivers/r600/r600_translate.c
@@ -0,0 +1,211 @@
+/*
+ * Copyright 2010 Red Hat Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * on the rights to use, copy, modify, merge, publish, distribute, sub
+ * license, and/or sell copies of the Software, and to permit persons to whom
+ * the Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * Authors: Dave Airlie <[email protected]>
+ */
+#include "translate/translate_cache.h"
+#include "translate/translate.h"
+#include <pipebuffer/pb_buffer.h>
+#include <util/u_index_modify.h>
+#include "r600_pipe.h"
+
+void r600_begin_vertex_translate(struct r600_pipe_context *rctx)
+{
+	struct pipe_context *pipe = &rctx->context;
+	struct translate_key key = {0};
+	struct translate_element *te;
+	unsigned tr_elem_index[PIPE_MAX_ATTRIBS] = {0};
+	struct translate *tr;
+	struct r600_vertex_element *ve = rctx->vertex_elements;
+	boolean vb_translated[PIPE_MAX_ATTRIBS] = {0};
+	void *vb_map[PIPE_MAX_ATTRIBS] = {0}, *out_map;
+	struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS] = {0}, *out_transfer;
+	struct pipe_resource *out_buffer;
+	unsigned i, num_verts;
+
+	/* Initialize the translate key, i.e. the recipe how vertices should be
+	 * translated. */
+	for (i = 0; i < ve->count; i++) {
+		struct pipe_vertex_buffer *vb =
+			&rctx->vertex_buffer[ve->elements[i].vertex_buffer_index];
+		enum pipe_format output_format = ve->hw_format[i];
+		unsigned output_format_size = ve->hw_format_size[i];
+
+		/* Check for support. */
+		if (ve->elements[i].src_format == ve->hw_format[i] &&
+		    (vb->buffer_offset + ve->elements[i].src_offset) % 4 == 0 &&
+		    vb->stride % 4 == 0) {
+			continue;
+		}
+
+		/* Workaround for translate: output floats instead of halfs. */
+		switch (output_format) {
+		case PIPE_FORMAT_R16_FLOAT:
+			output_format = PIPE_FORMAT_R32_FLOAT;
+			output_format_size = 4;
+			break;
+		case PIPE_FORMAT_R16G16_FLOAT:
+			output_format = PIPE_FORMAT_R32G32_FLOAT;
+			output_format_size = 8;
+			break;
+		case PIPE_FORMAT_R16G16B16_FLOAT:
+			output_format = PIPE_FORMAT_R32G32B32_FLOAT;
+			output_format_size = 12;
+			break;
+		case PIPE_FORMAT_R16G16B16A16_FLOAT:
+			output_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
+			output_format_size = 16;
+			break;
+		default:;
+		}
+
+		/* Add this vertex element. */
+		te = &key.element[key.nr_elements];
+		/*te->type;
+		  te->instance_divisor;*/
+		te->input_buffer = ve->elements[i].vertex_buffer_index;
+		te->input_format = ve->elements[i].src_format;
+		te->input_offset = vb->buffer_offset + ve->elements[i].src_offset;
+		te->output_format = output_format;
+		te->output_offset = key.output_stride;
+
+		key.output_stride += output_format_size;
+		vb_translated[ve->elements[i].vertex_buffer_index] = TRUE;
+		tr_elem_index[i] = key.nr_elements;
+		key.nr_elements++;
+	}
+
+	/* Get a translate object. */
+	tr = translate_cache_find(rctx->tran.translate_cache, &key);
+
+	/* Map buffers we want to translate. */
+	for (i = 0; i < rctx->nvertex_buffer; i++) {
+		if (vb_translated[i]) {
+			struct pipe_vertex_buffer *vb = &rctx->vertex_buffer[i];
+
+			vb_map[i] = pipe_buffer_map(pipe, vb->buffer,
+						    PIPE_TRANSFER_READ, &vb_transfer[i]);
+
+			tr->set_buffer(tr, i, vb_map[i], vb->stride, vb->max_index);
+		}
+	}
+
+	/* Create and map the output buffer. */
+	num_verts = rctx->vb_max_index + 1;
+
+	out_buffer = pipe_buffer_create(&rctx->screen->screen,
+					PIPE_BIND_VERTEX_BUFFER,
+					key.output_stride * num_verts);
+
+	out_map = pipe_buffer_map(pipe, out_buffer, PIPE_TRANSFER_WRITE,
+				  &out_transfer);
+
+	/* Translate. */
+	tr->run(tr, 0, num_verts, 0, out_map);
+
+	/* Unmap all buffers. */
+	for (i = 0; i < rctx->nvertex_buffer; i++) {
+		if (vb_translated[i]) {
+			pipe_buffer_unmap(pipe, rctx->vertex_buffer[i].buffer,
+					  vb_transfer[i]);
+		}
+	}
+
+	pipe_buffer_unmap(pipe, out_buffer, out_transfer);
+
+	/* Setup the new vertex buffer in the first free slot. */
+	for (i = 0; i < PIPE_MAX_ATTRIBS; i++) {
+		struct pipe_vertex_buffer *vb = &rctx->vertex_buffer[i];
+
+		if (!vb->buffer) {
+			pipe_resource_reference(&vb->buffer, out_buffer);
+			vb->buffer_offset = 0;
+			vb->max_index = num_verts - 1;
+			vb->stride = key.output_stride;
+			rctx->tran.vb_slot = i;
+			break;
+		}
+	}
+
+	/* Save and replace vertex elements. */
+	{
+		struct pipe_vertex_element new_velems[PIPE_MAX_ATTRIBS];
+
+		rctx->tran.saved_velems = rctx->vertex_elements;
+
+		for (i = 0; i < ve->count; i++) {
+			if (vb_translated[ve->elements[i].vertex_buffer_index]) {
+				te = &key.element[tr_elem_index[i]];
+				new_velems[i].instance_divisor = ve->elements[i].instance_divisor;
+				new_velems[i].src_format = te->output_format;
+				new_velems[i].src_offset = te->output_offset;
+				new_velems[i].vertex_buffer_index = rctx->tran.vb_slot;
+			} else {
+				memcpy(&new_velems[i], &ve->elements[i],
+				       sizeof(struct pipe_vertex_element));
+			}
+		}
+
+		rctx->tran.new_velems =
+			pipe->create_vertex_elements_state(pipe, ve->count, new_velems);
+		pipe->bind_vertex_elements_state(pipe, rctx->tran.new_velems);
+	}
+
+	pipe_resource_reference(&out_buffer, NULL);
+}
+
+void r600_end_vertex_translate(struct r600_pipe_context *rctx)
+{
+	struct pipe_context *pipe = &rctx->context;
+
+	/* Restore vertex elements. */
+	pipe->bind_vertex_elements_state(pipe, rctx->tran.saved_velems);
+	pipe->delete_vertex_elements_state(pipe, rctx->tran.new_velems);
+
+	/* Delete the now-unused VBO. */
+	pipe_resource_reference(&rctx->vertex_buffer[rctx->tran.vb_slot].buffer,
+				NULL);
+}
+
+void r600_translate_index_buffer(struct r600_pipe_context *r600,
+					struct pipe_resource **index_buffer,
+					unsigned *index_size,
+					unsigned *start, unsigned count)
+{
+	switch (*index_size) {
+	case 1:
+		util_shorten_ubyte_elts(&r600->context, index_buffer, 0, *start, count);
+		*index_size = 2;
+		*start = 0;
+		break;
+
+	case 2:
+		if (*start % 2 != 0) {
+			util_rebuild_ushort_elts(&r600->context, index_buffer, 0, *start, count);
+			*start = 0;
+		}
+		break;
+
+	case 4:
+		break;
+	}
+}
diff --git a/src/gallium/drivers/r600/r600d.h b/src/gallium/drivers/r600/r600d.h
index a96d2ce26c1..a3cb5b86004 100644
--- a/src/gallium/drivers/r600/r600d.h
+++ b/src/gallium/drivers/r600/r600d.h
@@ -667,6 +667,9 @@
 #define   S_02880C_Z_EXPORT_ENABLE(x)                  (((x) & 0x1) << 0)
 #define   G_02880C_Z_EXPORT_ENABLE(x)                  (((x) >> 0) & 0x1)
 #define   C_02880C_Z_EXPORT_ENABLE                     0xFFFFFFFE
+#define   S_02880C_STENCIL_REF_EXPORT_ENABLE(x)        (((x) & 0x1) << 1)
+#define   G_02880C_STENCIL_REF_EXPORT_ENABLE(x)        (((x) >> 1) & 0x1)
+#define   C_02880C_STENCIL_REF_EXPORT_ENABLE           0xFFFFFFFD
 #define   S_02880C_Z_ORDER(x)                          (((x) & 0x3) << 4)
 #define   G_02880C_Z_ORDER(x)                          (((x) >> 4) & 0x3)
 #define   C_02880C_Z_ORDER                             0xFFFFFCFF
@@ -901,6 +904,10 @@
 #define   S_038000_TILE_MODE(x)                        (((x) & 0xF) << 3)
 #define   G_038000_TILE_MODE(x)                        (((x) >> 3) & 0xF)
 #define   C_038000_TILE_MODE                           0xFFFFFF87
+#define     V_038000_ARRAY_LINEAR_GENERAL              0x00000000
+#define     V_038000_ARRAY_LINEAR_ALIGNED              0x00000001
+#define     V_038000_ARRAY_1D_TILED_THIN1              0x00000002
+#define     V_038000_ARRAY_2D_TILED_THIN1              0x00000004
 #define   S_038000_TILE_TYPE(x)                        (((x) & 0x1) << 7)
 #define   G_038000_TILE_TYPE(x)                        (((x) >> 7) & 0x1)
 #define   C_038000_TILE_TYPE                           0xFFFFFF7F
@@ -1025,45 +1032,7 @@
 #define   S_038008_DATA_FORMAT(x)                      (((x) & 0x3F) << 20)
 #define   G_038008_DATA_FORMAT(x)                      (((x) >> 20) & 0x3F)
 #define   C_038008_DATA_FORMAT                         0xFC0FFFFF
-#define     V_038008_FMT_INVALID                     0x00000000
-#define     V_038008_FMT_8                           0x00000001
-#define     V_038008_FMT_4_4                         0x00000002
-#define     V_038008_FMT_3_3_2                       0x00000003
-#define     V_038008_FMT_16                          0x00000005
-#define     V_038008_FMT_16_FLOAT                    0x00000006
-#define     V_038008_FMT_8_8                         0x00000007
-#define     V_038008_FMT_5_6_5                       0x00000008
-#define     V_038008_FMT_6_5_5                       0x00000009
-#define     V_038008_FMT_1_5_5_5                     0x0000000A
-#define     V_038008_FMT_4_4_4_4                     0x0000000B
-#define     V_038008_FMT_5_5_5_1                     0x0000000C
-#define     V_038008_FMT_32                          0x0000000D
-#define     V_038008_FMT_32_FLOAT                    0x0000000E
-#define     V_038008_FMT_16_16                       0x0000000F
-#define     V_038008_FMT_16_16_FLOAT                 0x00000010
-#define     V_038008_FMT_8_24                        0x00000011
-#define     V_038008_FMT_8_24_FLOAT                  0x00000012
-#define     V_038008_FMT_24_8                        0x00000013
-#define     V_038008_FMT_24_8_FLOAT                  0x00000014
-#define     V_038008_FMT_10_11_11                    0x00000015
-#define     V_038008_FMT_10_11_11_FLOAT              0x00000016
-#define     V_038008_FMT_11_11_10                    0x00000017
-#define     V_038008_FMT_11_11_10_FLOAT              0x00000018
-#define     V_038008_FMT_2_10_10_10                  0x00000019
-#define     V_038008_FMT_8_8_8_8                     0x0000001A
-#define     V_038008_FMT_10_10_10_2                  0x0000001B
-#define     V_038008_FMT_X24_8_32_FLOAT              0x0000001C
-#define     V_038008_FMT_32_32                       0x0000001D
-#define     V_038008_FMT_32_32_FLOAT                 0x0000001E
-#define     V_038008_FMT_16_16_16_16                 0x0000001F
-#define     V_038008_FMT_16_16_16_16_FLOAT           0x00000020
-#define     V_038008_FMT_32_32_32_32                 0x00000022
-#define     V_038008_FMT_32_32_32_32_FLOAT           0x00000023
-#define     V_038008_FMT_8_8_8                       0x0000002c
-#define     V_038008_FMT_16_16_16                    0x0000002d
-#define     V_038008_FMT_16_16_16_FLOAT              0x0000002e
-#define     V_038008_FMT_32_32_32                    0x0000002f
-#define     V_038008_FMT_32_32_32_FLOAT              0x00000030
+
 #define   S_038008_NUM_FORMAT_ALL(x)                   (((x) & 0x3) << 26)
 #define   G_038008_NUM_FORMAT_ALL(x)                   (((x) >> 26) & 0x3)
 #define   C_038008_NUM_FORMAT_ALL                      0xF3FFFFFF
diff --git a/src/gallium/drivers/softpipe/sp_fs_exec.c b/src/gallium/drivers/softpipe/sp_fs_exec.c
index 67e2c8f8bc4..346e1b402ba 100644
--- a/src/gallium/drivers/softpipe/sp_fs_exec.c
+++ b/src/gallium/drivers/softpipe/sp_fs_exec.c
@@ -158,9 +158,17 @@ exec_run( const struct sp_fragment_shader *base,
          case TGSI_SEMANTIC_POSITION:
             {
                uint j;
-               for (j = 0; j < 4; j++) {
+
+               for (j = 0; j < 4; j++)
                   quad->output.depth[j] = machine->Outputs[i].xyzw[2].f[j];
-               }
+            }
+            break;
+         case TGSI_SEMANTIC_STENCIL:
+            {
+               uint j;
+
+               for (j = 0; j < 4; j++)
+                  quad->output.stencil[j] = (unsigned)machine->Outputs[i].xyzw[1].f[j];
             }
             break;
          }
diff --git a/src/gallium/drivers/softpipe/sp_fs_sse.c b/src/gallium/drivers/softpipe/sp_fs_sse.c
index daa158df7c4..5b18cd035e3 100644
--- a/src/gallium/drivers/softpipe/sp_fs_sse.c
+++ b/src/gallium/drivers/softpipe/sp_fs_sse.c
@@ -169,9 +169,15 @@ fs_sse_run( const struct sp_fragment_shader *base,
          case TGSI_SEMANTIC_POSITION:
             {
                uint j;
-               for (j = 0; j < 4; j++) {
-                  quad->output.depth[j] = machine->Outputs[0].xyzw[2].f[j];
-               }
+               for (j = 0; j < 4; j++)
+                  quad->output.depth[j] = machine->Outputs[i].xyzw[2].f[j];
+            }
+            break;
+         case TGSI_SEMANTIC_STENCIL:
+            {
+               uint j;
+               for (j = 0; j < 4; j++)
+                  quad->output.stencil[j] = machine->Outputs[i].xyzw[1].f[j];
             }
             break;
          }
diff --git a/src/gallium/drivers/softpipe/sp_quad.h b/src/gallium/drivers/softpipe/sp_quad.h
index a3236bd1169..e745aa80619 100644
--- a/src/gallium/drivers/softpipe/sp_quad.h
+++ b/src/gallium/drivers/softpipe/sp_quad.h
@@ -85,6 +85,7 @@ struct quad_header_output
    /** colors in SOA format (rrrr, gggg, bbbb, aaaa) */
    float color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][QUAD_SIZE];
    float depth[QUAD_SIZE];
+   uint8_t stencil[QUAD_SIZE];
 };
 
 
diff --git a/src/gallium/drivers/softpipe/sp_quad_depth_test.c b/src/gallium/drivers/softpipe/sp_quad_depth_test.c
index e9b92626176..c8f5f89568a 100644
--- a/src/gallium/drivers/softpipe/sp_quad_depth_test.c
+++ b/src/gallium/drivers/softpipe/sp_quad_depth_test.c
@@ -47,6 +47,8 @@ struct depth_data {
    unsigned bzzzz[QUAD_SIZE];  /**< Z values fetched from depth buffer */
    unsigned qzzzz[QUAD_SIZE];  /**< Z values from the quad */
    ubyte stencilVals[QUAD_SIZE];
+   boolean use_shader_stencil_refs;
+   ubyte shader_stencil_refs[QUAD_SIZE];
    struct softpipe_cached_tile *tile;
 };
 
@@ -186,6 +188,33 @@ convert_quad_depth( struct depth_data *data,
 }
 
 
+/**
+ * Compute the depth_data::shader_stencil_refs[] values from the float fragment stencil values.
+ */
+static void
+convert_quad_stencil( struct depth_data *data, 
+                      const struct quad_header *quad )
+{
+   unsigned j;
+
+   data->use_shader_stencil_refs = TRUE;
+   /* Copy quads stencil values
+    */
+   switch (data->format) {
+   case PIPE_FORMAT_Z24X8_UNORM:
+   case PIPE_FORMAT_Z24_UNORM_S8_USCALED:
+   case PIPE_FORMAT_X8Z24_UNORM:
+   case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
+      {
+         for (j = 0; j < QUAD_SIZE; j++) {
+            data->shader_stencil_refs[j] = ((unsigned)(quad->output.stencil[j]));
+         }
+      }
+      break;
+   default:
+      assert(0);
+   }
+}
 
 /**
  * Write data->bzzzz[] values and data->stencilVals into the Z/stencil buffer.
@@ -272,8 +301,14 @@ do_stencil_test(struct depth_data *data,
 {
    unsigned passMask = 0x0;
    unsigned j;
+   ubyte refs[QUAD_SIZE];
 
-   ref &= valMask;
+   for (j = 0; j < QUAD_SIZE; j++) {
+      if (data->use_shader_stencil_refs)
+         refs[j] = data->shader_stencil_refs[j] & valMask;
+      else 
+         refs[j] = ref & valMask;
+   }
 
    switch (func) {
    case PIPE_FUNC_NEVER:
@@ -281,42 +316,42 @@ do_stencil_test(struct depth_data *data,
       break;
    case PIPE_FUNC_LESS:
       for (j = 0; j < QUAD_SIZE; j++) {
-         if (ref < (data->stencilVals[j] & valMask)) {
+         if (refs[j] < (data->stencilVals[j] & valMask)) {
             passMask |= (1 << j);
          }
       }
       break;
    case PIPE_FUNC_EQUAL:
       for (j = 0; j < QUAD_SIZE; j++) {
-         if (ref == (data->stencilVals[j] & valMask)) {
+         if (refs[j] == (data->stencilVals[j] & valMask)) {
             passMask |= (1 << j);
          }
       }
       break;
    case PIPE_FUNC_LEQUAL:
       for (j = 0; j < QUAD_SIZE; j++) {
-         if (ref <= (data->stencilVals[j] & valMask)) {
+         if (refs[j] <= (data->stencilVals[j] & valMask)) {
             passMask |= (1 << j);
          }
       }
       break;
    case PIPE_FUNC_GREATER:
       for (j = 0; j < QUAD_SIZE; j++) {
-         if (ref > (data->stencilVals[j] & valMask)) {
+         if (refs[j] > (data->stencilVals[j] & valMask)) {
             passMask |= (1 << j);
          }
       }
       break;
    case PIPE_FUNC_NOTEQUAL:
       for (j = 0; j < QUAD_SIZE; j++) {
-         if (ref != (data->stencilVals[j] & valMask)) {
+         if (refs[j] != (data->stencilVals[j] & valMask)) {
             passMask |= (1 << j);
          }
       }
       break;
    case PIPE_FUNC_GEQUAL:
       for (j = 0; j < QUAD_SIZE; j++) {
-         if (ref >= (data->stencilVals[j] & valMask)) {
+         if (refs[j] >= (data->stencilVals[j] & valMask)) {
             passMask |= (1 << j);
          }
       }
@@ -348,9 +383,14 @@ apply_stencil_op(struct depth_data *data,
 {
    unsigned j;
    ubyte newstencil[QUAD_SIZE];
+   ubyte refs[QUAD_SIZE];
 
    for (j = 0; j < QUAD_SIZE; j++) {
       newstencil[j] = data->stencilVals[j];
+      if (data->use_shader_stencil_refs)
+         refs[j] = data->shader_stencil_refs[j];
+      else
+         refs[j] = ref;
    }
 
    switch (op) {
@@ -367,7 +407,7 @@ apply_stencil_op(struct depth_data *data,
    case PIPE_STENCIL_OP_REPLACE:
       for (j = 0; j < QUAD_SIZE; j++) {
          if (mask & (1 << j)) {
-            newstencil[j] = ref;
+            newstencil[j] = refs[j];
          }
       }
       break;
@@ -688,8 +728,10 @@ depth_test_quads_fallback(struct quad_stage *qs,
    unsigned i, pass = 0;
    const struct sp_fragment_shader *fs = qs->softpipe->fs;
    boolean interp_depth = !fs->info.writes_z;
+   boolean shader_stencil_ref = fs->info.writes_stencil;
    struct depth_data data;
 
+   data.use_shader_stencil_refs = FALSE;
 
    if (qs->softpipe->depth_stencil->alpha.enabled) {
       nr = alpha_test_quads(qs, quads, nr);
@@ -716,6 +758,9 @@ depth_test_quads_fallback(struct quad_stage *qs,
          }
 
          if (qs->softpipe->depth_stencil->stencil[0].enabled) {
+            if (shader_stencil_ref)
+               convert_quad_stencil(&data, quads[i]);
+            
             depth_stencil_test_quad(qs, &data, quads[i]);
             write_depth_stencil_values(&data, quads[i]);
          }
diff --git a/src/gallium/drivers/softpipe/sp_quad_pipe.c b/src/gallium/drivers/softpipe/sp_quad_pipe.c
index 43b8e88e334..2cfd02a22c6 100644
--- a/src/gallium/drivers/softpipe/sp_quad_pipe.c
+++ b/src/gallium/drivers/softpipe/sp_quad_pipe.c
@@ -47,7 +47,8 @@ sp_build_quad_pipeline(struct softpipe_context *sp)
       sp->framebuffer.zsbuf &&
       !sp->depth_stencil->alpha.enabled &&
       !sp->fs->info.uses_kill &&
-      !sp->fs->info.writes_z;
+      !sp->fs->info.writes_z &&
+      !sp->fs->info.writes_stencil;
 
    sp->quad.first = sp->quad.blend;
 
diff --git a/src/gallium/drivers/softpipe/sp_screen.c b/src/gallium/drivers/softpipe/sp_screen.c
index 25a0a622179..edc2a6dacf2 100644
--- a/src/gallium/drivers/softpipe/sp_screen.c
+++ b/src/gallium/drivers/softpipe/sp_screen.c
@@ -113,8 +113,12 @@ softpipe_get_param(struct pipe_screen *screen, enum pipe_cap param)
       return 1;
    case PIPE_CAP_STREAM_OUTPUT:
       return 1;
+   case PIPE_CAP_PRIMITIVE_RESTART:
+      return 1;
    case PIPE_CAP_DEPTHSTENCIL_CLEAR_SEPARATE:
       return 0;
+   case PIPE_CAP_SHADER_STENCIL_EXPORT:
+      return 1;
    default:
       return 0;
    }
@@ -206,13 +210,6 @@ softpipe_is_format_supported( struct pipe_screen *screen,
       if (format_desc->block.width != 1 ||
           format_desc->block.height != 1)
          return FALSE;
-
-      /*
-       * TODO: Unfortunately we cannot render into anything more than 32 bits
-       * because we encode color clear values into a 32bit word.
-       */
-      if (format_desc->block.bits > 32)
-         return FALSE;
    }
 
    if (bind & PIPE_BIND_DEPTH_STENCIL) {
diff --git a/src/gallium/drivers/softpipe/sp_tex_sample.c b/src/gallium/drivers/softpipe/sp_tex_sample.c
index 088e48f81fe..2eac4c7a82b 100644
--- a/src/gallium/drivers/softpipe/sp_tex_sample.c
+++ b/src/gallium/drivers/softpipe/sp_tex_sample.c
@@ -44,6 +44,9 @@
 #include "sp_tex_tile_cache.h"
 
 
+/** Set to one to help debug texture sampling */
+#define DEBUG_TEX 0
+
 
 /*
  * Return fractional part of 'f'.  Used for computing interpolation weights.
@@ -774,6 +777,18 @@ pot_level_size(unsigned base_pot, unsigned level)
 }
 
 
+static void
+print_sample(const char *function, float rgba[NUM_CHANNELS][QUAD_SIZE])
+{
+   debug_printf("%s %g %g %g %g, %g %g %g %g, %g %g %g %g, %g %g %g %g\n",
+                function,
+                rgba[0][0], rgba[1][0], rgba[2][0], rgba[3][0],
+                rgba[0][1], rgba[1][1], rgba[2][1], rgba[3][1],
+                rgba[0][2], rgba[1][2], rgba[2][2], rgba[3][2],
+                rgba[0][3], rgba[1][3], rgba[2][3], rgba[3][3]);
+}
+
+
 /* Some image-filter fastpaths:
  */
 static INLINE void
@@ -832,6 +847,10 @@ img_filter_2d_linear_repeat_POT(struct tgsi_sampler *tgsi_sampler,
                               tx[2][c], tx[3][c]);
       }
    }
+
+   if (DEBUG_TEX) {
+      print_sample(__FUNCTION__, rgba);
+   }
 }
 
 
@@ -872,6 +891,10 @@ img_filter_2d_nearest_repeat_POT(struct tgsi_sampler *tgsi_sampler,
          rgba[c][j] = out[c];
       }
    }
+
+   if (DEBUG_TEX) {
+      print_sample(__FUNCTION__, rgba);
+   }
 }
 
 
@@ -921,6 +944,10 @@ img_filter_2d_nearest_clamp_POT(struct tgsi_sampler *tgsi_sampler,
          rgba[c][j] = out[c];
       }
    }
+
+   if (DEBUG_TEX) {
+      print_sample(__FUNCTION__, rgba);
+   }
 }
 
 
@@ -957,6 +984,10 @@ img_filter_1d_nearest(struct tgsi_sampler *tgsi_sampler,
          rgba[c][j] = out[c];
       }
    }
+
+   if (DEBUG_TEX) {
+      print_sample(__FUNCTION__, rgba);
+   }
 }
 
 
@@ -997,6 +1028,10 @@ img_filter_2d_nearest(struct tgsi_sampler *tgsi_sampler,
          rgba[c][j] = out[c];
       }
    }
+
+   if (DEBUG_TEX) {
+      print_sample(__FUNCTION__, rgba);
+   }
 }
 
 
@@ -1045,6 +1080,10 @@ img_filter_cube_nearest(struct tgsi_sampler *tgsi_sampler,
          rgba[c][j] = out[c];
       }
    }
+
+   if (DEBUG_TEX) {
+      print_sample(__FUNCTION__, rgba);
+   }
 }
 
 
@@ -1357,6 +1396,10 @@ mip_filter_linear(struct tgsi_sampler *tgsi_sampler,
          }
       }
    }
+
+   if (DEBUG_TEX) {
+      print_sample(__FUNCTION__, rgba);
+   }
 }
 
 
@@ -1402,13 +1445,9 @@ mip_filter_nearest(struct tgsi_sampler *tgsi_sampler,
       samp->min_img_filter(tgsi_sampler, s, t, p, NULL, tgsi_sampler_lod_bias, rgba);
    }
 
-#if 0
-   printf("RGBA %g %g %g %g, %g %g %g %g, %g %g %g %g, %g %g %g %g\n",
-          rgba[0][0], rgba[1][0], rgba[2][0], rgba[3][0],
-          rgba[0][1], rgba[1][1], rgba[2][1], rgba[3][1],
-          rgba[0][2], rgba[1][2], rgba[2][2], rgba[3][2],
-          rgba[0][3], rgba[1][3], rgba[2][3], rgba[3][3]);
-#endif
+   if (DEBUG_TEX) {
+      print_sample(__FUNCTION__, rgba);
+   }
 }
 
 
@@ -1510,6 +1549,10 @@ mip_filter_linear_2d_linear_repeat_POT(
          }
       }
    }
+
+   if (DEBUG_TEX) {
+      print_sample(__FUNCTION__, rgba);
+   }
 }
 
 
diff --git a/src/gallium/drivers/softpipe/sp_tile_cache.h b/src/gallium/drivers/softpipe/sp_tile_cache.h
index 031c7c1ea5c..4151a47c323 100644
--- a/src/gallium/drivers/softpipe/sp_tile_cache.h
+++ b/src/gallium/drivers/softpipe/sp_tile_cache.h
@@ -86,7 +86,7 @@ struct softpipe_tile_cache
    struct softpipe_cached_tile *entries[NUM_ENTRIES];
    uint clear_flags[(MAX_WIDTH / TILE_SIZE) * (MAX_HEIGHT / TILE_SIZE) / 32];
    float clear_color[4];  /**< for color bufs */
-   uint clear_val;        /**< for z+stencil, or packed color clear value */
+   uint clear_val;        /**< for z+stencil */
    boolean depth_stencil; /**< Is the surface a depth/stencil format? */
 
    struct softpipe_cached_tile *tile;  /**< scratch tile for clears */