Merge branch 'gallium_draw_llvm'

16 files changed, 1934 insertions, 24 deletions

diff --git a/src/gallium/auxiliary/SConscript b/src/gallium/auxiliary/SConscript
index ae749acc87e..ca5fab6742a 100644
--- a/src/gallium/auxiliary/SConscript
+++ b/src/gallium/auxiliary/SConscript
@@ -213,6 +213,9 @@ if drawllvm:
     'gallivm/lp_bld_swizzle.c',
     'gallivm/lp_bld_tgsi_soa.c',
     'gallivm/lp_bld_type.c',
+    'draw/draw_llvm.c',
+    'draw/draw_pt_fetch_shade_pipeline_llvm.c',
+    'draw/draw_llvm_translate.c'
     ]
 
 gallium = env.ConvenienceLibrary(
diff --git a/src/gallium/auxiliary/draw/draw_context.c b/src/gallium/auxiliary/draw/draw_context.c
index bb0988543f5..b6574a9fea1 100644
--- a/src/gallium/auxiliary/draw/draw_context.c
+++ b/src/gallium/auxiliary/draw/draw_context.c
@@ -44,6 +44,18 @@ struct draw_context *draw_create( void )
    if (draw == NULL)
       goto fail;
 
+   if (!draw_init(draw))
+      goto fail;
+
+   return draw;
+
+fail:
+   draw_destroy( draw );
+   return NULL;
+}
+
+boolean draw_init(struct draw_context *draw)
+{
    ASSIGN_4V( draw->plane[0], -1,  0,  0, 1 );
    ASSIGN_4V( draw->plane[1],  1,  0,  0, 1 );
    ASSIGN_4V( draw->plane[2],  0, -1,  0, 1 );
@@ -57,22 +69,18 @@ struct draw_context *draw_create( void )
 
 
    if (!draw_pipeline_init( draw ))
-      goto fail;
+      return FALSE;
 
    if (!draw_pt_init( draw ))
-      goto fail;
+      return FALSE;
 
    if (!draw_vs_init( draw ))
-      goto fail;
+      return FALSE;
 
    if (!draw_gs_init( draw ))
-      goto fail;
+      return FALSE;
 
-   return draw;
-
-fail:
-   draw_destroy( draw );   
-   return NULL;
+   return TRUE;
 }
 
 
diff --git a/src/gallium/auxiliary/draw/draw_context.h b/src/gallium/auxiliary/draw/draw_context.h
index acd81b9712d..d42e4003183 100644
--- a/src/gallium/auxiliary/draw/draw_context.h
+++ b/src/gallium/auxiliary/draw/draw_context.h
@@ -40,6 +40,9 @@
 
 #include "pipe/p_state.h"
 
+#ifdef DRAW_LLVM
+#include <llvm-c/ExecutionEngine.h>
+#endif
 
 struct pipe_context;
 struct draw_context;
@@ -197,6 +200,11 @@ boolean draw_need_pipeline(const struct draw_context *draw,
                            const struct pipe_rasterizer_state *rasterizer,
                            unsigned prim );
 
-
+#ifdef DRAW_LLVM
+/*******************************************************************************
+ * LLVM integration
+ */
+struct draw_context *draw_create_with_llvm(LLVMExecutionEngineRef engine);
+#endif
 
 #endif /* DRAW_CONTEXT_H */
diff --git a/src/gallium/auxiliary/draw/draw_gs.c b/src/gallium/auxiliary/draw/draw_gs.c
index 7069aa6b181..131deed43e4 100644
--- a/src/gallium/auxiliary/draw/draw_gs.c
+++ b/src/gallium/auxiliary/draw/draw_gs.c
@@ -342,10 +342,10 @@ void draw_geometry_shader_delete(struct draw_geometry_shader *shader)
 void draw_geometry_shader_prepare(struct draw_geometry_shader *shader,
                                   struct draw_context *draw)
 {
-    if (shader->machine->Tokens != shader->state.tokens) {
-       tgsi_exec_machine_bind_shader(shader->machine,
-                                     shader->state.tokens,
-                                     draw->gs.num_samplers,
-                                     draw->gs.samplers);
-    }
+   if (shader && shader->machine->Tokens != shader->state.tokens) {
+      tgsi_exec_machine_bind_shader(shader->machine,
+                                    shader->state.tokens,
+                                    draw->gs.num_samplers,
+                                    draw->gs.samplers);
+   }
 }
diff --git a/src/gallium/auxiliary/draw/draw_llvm.c b/src/gallium/auxiliary/draw/draw_llvm.c
new file mode 100644
index 00000000000..26f756c6f8a
--- /dev/null
+++ b/src/gallium/auxiliary/draw/draw_llvm.c
@@ -0,0 +1,697 @@
+#include "draw_llvm.h"
+
+#include "draw_context.h"
+#include "draw_vs.h"
+
+#include "gallivm/lp_bld_arit.h"
+#include "gallivm/lp_bld_interp.h"
+#include "gallivm/lp_bld_struct.h"
+#include "gallivm/lp_bld_type.h"
+#include "gallivm/lp_bld_flow.h"
+#include "gallivm/lp_bld_debug.h"
+#include "gallivm/lp_bld_tgsi.h"
+#include "gallivm/lp_bld_printf.h"
+
+#include "util/u_cpu_detect.h"
+#include "tgsi/tgsi_dump.h"
+
+#include <llvm-c/Transforms/Scalar.h>
+
+#define DEBUG_STORE 0
+
+
+/* generates the draw jit function */
+static void
+draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var);
+
+static void
+init_globals(struct draw_llvm *llvm)
+{
+   LLVMTypeRef texture_type;
+
+   /* struct draw_jit_texture */
+   {
+      LLVMTypeRef elem_types[4];
+
+      elem_types[DRAW_JIT_TEXTURE_WIDTH]  = LLVMInt32Type();
+      elem_types[DRAW_JIT_TEXTURE_HEIGHT] = LLVMInt32Type();
+      elem_types[DRAW_JIT_TEXTURE_STRIDE] = LLVMInt32Type();
+      elem_types[DRAW_JIT_TEXTURE_DATA]   = LLVMPointerType(LLVMInt8Type(), 0);
+
+      texture_type = LLVMStructType(elem_types, Elements(elem_types), 0);
+
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
+                             llvm->target, texture_type,
+                             DRAW_JIT_TEXTURE_WIDTH);
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
+                             llvm->target, texture_type,
+                             DRAW_JIT_TEXTURE_HEIGHT);
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, stride,
+                             llvm->target, texture_type,
+                             DRAW_JIT_TEXTURE_STRIDE);
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, data,
+                             llvm->target, texture_type,
+                             DRAW_JIT_TEXTURE_DATA);
+      LP_CHECK_STRUCT_SIZE(struct draw_jit_texture,
+                           llvm->target, texture_type);
+
+      LLVMAddTypeName(llvm->module, "texture", texture_type);
+   }
+
+
+   /* struct draw_jit_context */
+   {
+      LLVMTypeRef elem_types[3];
+      LLVMTypeRef context_type;
+
+      elem_types[0] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
+      elem_types[1] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
+      elem_types[2] = LLVMArrayType(texture_type, PIPE_MAX_SAMPLERS); /* textures */
+
+      context_type = LLVMStructType(elem_types, Elements(elem_types), 0);
+
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
+                             llvm->target, context_type, 0);
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, gs_constants,
+                             llvm->target, context_type, 1);
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
+                             llvm->target, context_type,
+                             DRAW_JIT_CONTEXT_TEXTURES_INDEX);
+      LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
+                           llvm->target, context_type);
+
+      LLVMAddTypeName(llvm->module, "draw_jit_context", context_type);
+
+      llvm->context_ptr_type = LLVMPointerType(context_type, 0);
+   }
+   {
+      LLVMTypeRef buffer_ptr = LLVMPointerType(LLVMIntType(8), 0);
+      llvm->buffer_ptr_type = LLVMPointerType(buffer_ptr, 0);
+   }
+}
+
+static LLVMTypeRef
+create_vertex_header(struct draw_llvm *llvm, int data_elems)
+{
+   /* struct vertex_header */
+   LLVMTypeRef elem_types[3];
+   LLVMTypeRef vertex_header;
+   char struct_name[24];
+
+   snprintf(struct_name, 23, "vertex_header%d", data_elems);
+
+   elem_types[0]  = LLVMIntType(32);
+   elem_types[1]  = LLVMArrayType(LLVMFloatType(), 4);
+   elem_types[2]  = LLVMArrayType(elem_types[1], data_elems);
+
+   vertex_header = LLVMStructType(elem_types, Elements(elem_types), 0);
+
+   /* these are bit-fields and we can't take address of them
+      LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
+      llvm->target, vertex_header,
+      DRAW_JIT_VERTEX_CLIPMASK);
+      LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
+      llvm->target, vertex_header,
+      DRAW_JIT_VERTEX_EDGEFLAG);
+      LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
+      llvm->target, vertex_header,
+      DRAW_JIT_VERTEX_PAD);
+      LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
+      llvm->target, vertex_header,
+      DRAW_JIT_VERTEX_VERTEX_ID);
+   */
+   LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
+                          llvm->target, vertex_header,
+                          DRAW_JIT_VERTEX_CLIP);
+   LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
+                          llvm->target, vertex_header,
+                          DRAW_JIT_VERTEX_DATA);
+
+   LLVMAddTypeName(llvm->module, struct_name, vertex_header);
+
+   return LLVMPointerType(vertex_header, 0);
+}
+
+struct draw_llvm *
+draw_llvm_create(struct draw_context *draw)
+{
+   struct draw_llvm *llvm = CALLOC_STRUCT( draw_llvm );
+
+   util_cpu_detect();
+
+   llvm->draw = draw;
+   llvm->engine = draw->engine;
+
+   debug_assert(llvm->engine);
+
+   llvm->module = LLVMModuleCreateWithName("draw_llvm");
+   llvm->provider = LLVMCreateModuleProviderForExistingModule(llvm->module);
+
+   LLVMAddModuleProvider(llvm->engine, llvm->provider);
+
+   llvm->target = LLVMGetExecutionEngineTargetData(llvm->engine);
+
+   llvm->pass = LLVMCreateFunctionPassManager(llvm->provider);
+   LLVMAddTargetData(llvm->target, llvm->pass);
+   /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
+    * but there are more on SVN. */
+   /* TODO: Add more passes */
+   LLVMAddConstantPropagationPass(llvm->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(llvm->pass);
+   }
+   LLVMAddPromoteMemoryToRegisterPass(llvm->pass);
+   LLVMAddGVNPass(llvm->pass);
+   LLVMAddCFGSimplificationPass(llvm->pass);
+
+   init_globals(llvm);
+
+
+#if 1
+   LLVMDumpModule(llvm->module);
+#endif
+
+   return llvm;
+}
+
+void
+draw_llvm_destroy(struct draw_llvm *llvm)
+{
+   free(llvm);
+}
+
+struct draw_llvm_variant *
+draw_llvm_prepare(struct draw_llvm *llvm, int num_inputs)
+{
+   struct draw_llvm_variant *variant = MALLOC(sizeof(struct draw_llvm_variant));
+
+   draw_llvm_make_variant_key(llvm, &variant->key);
+
+   llvm->vertex_header_ptr_type = create_vertex_header(llvm, num_inputs);
+
+   draw_llvm_generate(llvm, variant);
+
+   return variant;
+}
+
+
+struct draw_context *draw_create_with_llvm(LLVMExecutionEngineRef engine)
+{
+   struct draw_context *draw = CALLOC_STRUCT( draw_context );
+   if (draw == NULL)
+      goto fail;
+   draw->engine = engine;
+
+   if (!draw_init(draw))
+      goto fail;
+
+   return draw;
+
+fail:
+   draw_destroy( draw );
+   return NULL;
+}
+
+static void
+generate_vs(struct draw_llvm *llvm,
+            LLVMBuilderRef builder,
+            LLVMValueRef (*outputs)[NUM_CHANNELS],
+            const LLVMValueRef (*inputs)[NUM_CHANNELS],
+            LLVMValueRef context_ptr)
+{
+   const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
+   struct lp_type vs_type;
+   LLVMValueRef consts_ptr = draw_jit_context_vs_constants(builder, context_ptr);
+
+   memset(&vs_type, 0, sizeof vs_type);
+   vs_type.floating = TRUE; /* floating point values */
+   vs_type.sign = TRUE;     /* values are signed */
+   vs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
+   vs_type.width = 32;      /* 32-bit float */
+   vs_type.length = 4;      /* 4 elements per vector */
+#if 0
+   num_vs = 4;              /* number of vertices per block */
+#endif
+
+   tgsi_dump(tokens, 0);
+   lp_build_tgsi_soa(builder,
+                     tokens,
+                     vs_type,
+                     NULL /*struct lp_build_mask_context *mask*/,
+                     consts_ptr,
+                     NULL /*pos*/,
+                     inputs,
+                     outputs,
+                     NULL/*sampler*/);
+}
+
+#if DEBUG_STORE
+static void print_vectorf(LLVMBuilderRef builder,
+                         LLVMValueRef vec)
+{
+   LLVMValueRef val[4];
+   val[0] = LLVMBuildExtractElement(builder, vec,
+                                    LLVMConstInt(LLVMInt32Type(), 0, 0), "");
+   val[1] = LLVMBuildExtractElement(builder, vec,
+                                    LLVMConstInt(LLVMInt32Type(), 1, 0), "");
+   val[2] = LLVMBuildExtractElement(builder, vec,
+                                    LLVMConstInt(LLVMInt32Type(), 2, 0), "");
+   val[3] = LLVMBuildExtractElement(builder, vec,
+                                    LLVMConstInt(LLVMInt32Type(), 3, 0), "");
+   lp_build_printf(builder, "vector = [%f, %f, %f, %f]\n",
+                   val[0], val[1], val[2], val[3]);
+}
+#endif
+
+static void
+generate_fetch(LLVMBuilderRef builder,
+               LLVMValueRef vbuffers_ptr,
+               LLVMValueRef *res,
+               struct pipe_vertex_element *velem,
+               struct pipe_vertex_buffer *vbuf,
+               LLVMValueRef index)
+{
+   LLVMValueRef indices = LLVMConstInt(LLVMInt64Type(), velem->vertex_buffer_index, 0);
+   LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
+                                           &indices, 1, "");
+   LLVMValueRef stride = LLVMBuildMul(builder,
+                                      LLVMConstInt(LLVMInt32Type(), vbuf->stride, 0),
+                                      index, "");
+
+   vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer");
+
+   stride = LLVMBuildAdd(builder, stride,
+                         LLVMConstInt(LLVMInt32Type(), vbuf->buffer_offset, 0),
+                         "");
+   stride = LLVMBuildAdd(builder, stride,
+                         LLVMConstInt(LLVMInt32Type(), velem->src_offset, 0),
+                         "");
+
+   /*lp_build_printf(builder, "vbuf index = %d, stride is %d\n", indices, stride);*/
+   vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, "");
+
+   *res = draw_llvm_translate_from(builder, vbuffer_ptr, velem->src_format);
+}
+
+static LLVMValueRef
+aos_to_soa(LLVMBuilderRef builder,
+           LLVMValueRef val0,
+           LLVMValueRef val1,
+           LLVMValueRef val2,
+           LLVMValueRef val3,
+           LLVMValueRef channel)
+{
+   LLVMValueRef ex, res;
+
+   ex = LLVMBuildExtractElement(builder, val0,
+                                channel, "");
+   res = LLVMBuildInsertElement(builder,
+                                LLVMConstNull(LLVMTypeOf(val0)),
+                                ex,
+                                LLVMConstInt(LLVMInt32Type(), 0, 0),
+                                "");
+
+   ex = LLVMBuildExtractElement(builder, val1,
+                                channel, "");
+   res = LLVMBuildInsertElement(builder,
+                                res, ex,
+                                LLVMConstInt(LLVMInt32Type(), 1, 0),
+                                "");
+
+   ex = LLVMBuildExtractElement(builder, val2,
+                                channel, "");
+   res = LLVMBuildInsertElement(builder,
+                                res, ex,
+                                LLVMConstInt(LLVMInt32Type(), 2, 0),
+                                "");
+
+   ex = LLVMBuildExtractElement(builder, val3,
+                                channel, "");
+   res = LLVMBuildInsertElement(builder,
+                                res, ex,
+                                LLVMConstInt(LLVMInt32Type(), 3, 0),
+                                "");
+
+   return res;
+}
+
+static void
+soa_to_aos(LLVMBuilderRef builder,
+           LLVMValueRef soa[NUM_CHANNELS],
+           LLVMValueRef aos[NUM_CHANNELS])
+{
+   LLVMValueRef comp;
+   int i = 0;
+
+   debug_assert(NUM_CHANNELS == 4);
+
+   aos[0] = LLVMConstNull(LLVMTypeOf(soa[0]));
+   aos[1] = aos[2] = aos[3] = aos[0];
+
+   for (i = 0; i < NUM_CHANNELS; ++i) {
+      LLVMValueRef channel = LLVMConstInt(LLVMInt32Type(), i, 0);
+
+      comp = LLVMBuildExtractElement(builder, soa[i],
+                                     LLVMConstInt(LLVMInt32Type(), 0, 0), "");
+      aos[0] = LLVMBuildInsertElement(builder, aos[0], comp, channel, "");
+
+      comp = LLVMBuildExtractElement(builder, soa[i],
+                                     LLVMConstInt(LLVMInt32Type(), 1, 0), "");
+      aos[1] = LLVMBuildInsertElement(builder, aos[1], comp, channel, "");
+
+      comp = LLVMBuildExtractElement(builder, soa[i],
+                                     LLVMConstInt(LLVMInt32Type(), 2, 0), "");
+      aos[2] = LLVMBuildInsertElement(builder, aos[2], comp, channel, "");
+
+      comp = LLVMBuildExtractElement(builder, soa[i],
+                                     LLVMConstInt(LLVMInt32Type(), 3, 0), "");
+      aos[3] = LLVMBuildInsertElement(builder, aos[3], comp, channel, "");
+
+   }
+}
+
+static void
+convert_to_soa(LLVMBuilderRef builder,
+               LLVMValueRef (*aos)[NUM_CHANNELS],
+               LLVMValueRef (*soa)[NUM_CHANNELS],
+               int num_attribs)
+{
+   int i;
+
+   debug_assert(NUM_CHANNELS == 4);
+
+   for (i = 0; i < num_attribs; ++i) {
+      LLVMValueRef val0 = aos[i][0];
+      LLVMValueRef val1 = aos[i][1];
+      LLVMValueRef val2 = aos[i][2];
+      LLVMValueRef val3 = aos[i][3];
+
+      soa[i][0] = aos_to_soa(builder, val0, val1, val2, val3,
+                             LLVMConstInt(LLVMInt32Type(), 0, 0));
+      soa[i][1] = aos_to_soa(builder, val0, val1, val2, val3,
+                             LLVMConstInt(LLVMInt32Type(), 1, 0));
+      soa[i][2] = aos_to_soa(builder, val0, val1, val2, val3,
+                             LLVMConstInt(LLVMInt32Type(), 2, 0));
+      soa[i][3] = aos_to_soa(builder, val0, val1, val2, val3,
+                             LLVMConstInt(LLVMInt32Type(), 3, 0));
+   }
+}
+
+static void
+store_aos(LLVMBuilderRef builder,
+          LLVMValueRef io_ptr,
+          LLVMValueRef index,
+          LLVMValueRef value)
+{
+   LLVMValueRef id_ptr = draw_jit_header_id(builder, io_ptr);
+   LLVMValueRef data_ptr = draw_jit_header_data(builder, io_ptr);
+   LLVMValueRef indices[3];
+
+   indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
+   indices[1] = index;
+   indices[2] = LLVMConstInt(LLVMInt32Type(), 0, 0);
+
+   /* undefined vertex */
+   LLVMBuildStore(builder, LLVMConstInt(LLVMInt32Type(),
+                                        0xffff, 0), id_ptr);
+
+#if DEBUG_STORE
+   lp_build_printf(builder, "    ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
+#endif
+#if 0
+   /*lp_build_printf(builder, " ---- %p storing at %d (%p)  ", io_ptr, index, data_ptr);
+     print_vectorf(builder, value);*/
+   data_ptr = LLVMBuildBitCast(builder, data_ptr,
+                               LLVMPointerType(LLVMArrayType(LLVMVectorType(LLVMFloatType(), 4), 0), 0),
+                               "datavec");
+   data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 2, "");
+
+   LLVMBuildStore(builder, value, data_ptr);
+#else
+   {
+      LLVMValueRef x, y, z, w;
+      LLVMValueRef idx0, idx1, idx2, idx3;
+      LLVMValueRef gep0, gep1, gep2, gep3;
+      data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
+
+      idx0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
+      idx1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
+      idx2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
+      idx3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
+
+      x = LLVMBuildExtractElement(builder, value,
+                                  idx0, "");
+      y = LLVMBuildExtractElement(builder, value,
+                                  idx1, "");
+      z = LLVMBuildExtractElement(builder, value,
+                                  idx2, "");
+      w = LLVMBuildExtractElement(builder, value,
+                                  idx3, "");
+
+      gep0 = LLVMBuildGEP(builder, data_ptr, &idx0, 1, "");
+      gep1 = LLVMBuildGEP(builder, data_ptr, &idx1, 1, "");
+      gep2 = LLVMBuildGEP(builder, data_ptr, &idx2, 1, "");
+      gep3 = LLVMBuildGEP(builder, data_ptr, &idx3, 1, "");
+
+      /*lp_build_printf(builder, "##### x = %f (%p), y = %f (%p), z = %f (%p), w = %f (%p)\n",
+        x, gep0, y, gep1, z, gep2, w, gep3);*/
+      LLVMBuildStore(builder, x, gep0);
+      LLVMBuildStore(builder, y, gep1);
+      LLVMBuildStore(builder, z, gep2);
+      LLVMBuildStore(builder, w, gep3);
+   }
+#endif
+}
+
+static void
+store_aos_array(LLVMBuilderRef builder,
+                LLVMValueRef io_ptr,
+                LLVMValueRef aos[NUM_CHANNELS],
+                int attrib,
+                int num_outputs)
+{
+   LLVMValueRef attr_index = LLVMConstInt(LLVMInt32Type(), attrib, 0);
+   LLVMValueRef ind0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
+   LLVMValueRef ind1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
+   LLVMValueRef ind2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
+   LLVMValueRef ind3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
+   LLVMValueRef io0_ptr, io1_ptr, io2_ptr, io3_ptr;
+
+   debug_assert(NUM_CHANNELS == 4);
+
+   io0_ptr = LLVMBuildGEP(builder, io_ptr,
+                          &ind0, 1, "");
+   io1_ptr = LLVMBuildGEP(builder, io_ptr,
+                          &ind1, 1, "");
+   io2_ptr = LLVMBuildGEP(builder, io_ptr,
+                          &ind2, 1, "");
+   io3_ptr = LLVMBuildGEP(builder, io_ptr,
+                          &ind3, 1, "");
+
+#if DEBUG_STORE
+   lp_build_printf(builder, "   io = %p, indexes[%d, %d, %d, %d]\n",
+                   io_ptr, ind0, ind1, ind2, ind3);
+#endif
+
+   store_aos(builder, io0_ptr, attr_index, aos[0]);
+   store_aos(builder, io1_ptr, attr_index, aos[1]);
+   store_aos(builder, io2_ptr, attr_index, aos[2]);
+   store_aos(builder, io3_ptr, attr_index, aos[3]);
+}
+
+static void
+convert_to_aos(LLVMBuilderRef builder,
+               LLVMValueRef io,
+               LLVMValueRef (*outputs)[NUM_CHANNELS],
+               int num_outputs,
+               int max_vertices)
+{
+   unsigned chan, attrib;
+
+#if DEBUG_STORE
+   lp_build_printf(builder, "   # storing begin\n");
+#endif
+   for (attrib = 0; attrib < num_outputs; ++attrib) {
+      LLVMValueRef soa[4];
+      LLVMValueRef aos[4];
+      for(chan = 0; chan < NUM_CHANNELS; ++chan) {
+         if(outputs[attrib][chan]) {
+            LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
+            lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
+            /*lp_build_printf(builder, "output %d : %d ",
+                            LLVMConstInt(LLVMInt32Type(), attrib, 0),
+                            LLVMConstInt(LLVMInt32Type(), chan, 0));
+              print_vectorf(builder, out);*/
+            soa[chan] = out;
+         } else
+            soa[chan] = 0;
+      }
+      soa_to_aos(builder, soa, aos);
+      store_aos_array(builder,
+                      io,
+                      aos,
+                      attrib,
+                      num_outputs);
+   }
+#if DEBUG_STORE
+   lp_build_printf(builder, "   # storing end\n");
+#endif
+}
+
+static void
+draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
+{
+   LLVMTypeRef arg_types[6];
+   LLVMTypeRef func_type;
+   LLVMValueRef context_ptr;
+   LLVMBasicBlockRef block;
+   LLVMBuilderRef builder;
+   LLVMValueRef start, end, count, stride, step, io_itr;
+   LLVMValueRef io_ptr, vbuffers_ptr;
+   struct draw_context *draw = llvm->draw;
+   unsigned i, j;
+   struct lp_build_context bld;
+   struct lp_build_loop_state lp_loop;
+   struct lp_type vs_type = lp_type_float_vec(32);
+   const int max_vertices = 4;
+   LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
+
+   arg_types[0] = llvm->context_ptr_type;           /* context */
+   arg_types[1] = llvm->vertex_header_ptr_type;     /* vertex_header */
+   arg_types[2] = llvm->buffer_ptr_type;            /* vbuffers */
+   arg_types[3] = LLVMInt32Type();                  /* start */
+   arg_types[4] = LLVMInt32Type();                  /* count */
+   arg_types[5] = LLVMInt32Type();                  /* stride */
+
+   func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
+
+   variant->function = LLVMAddFunction(llvm->module, "draw_llvm_shader", func_type);
+   LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
+   for(i = 0; i < Elements(arg_types); ++i)
+      if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
+         LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
+
+   context_ptr  = LLVMGetParam(variant->function, 0);
+   io_ptr       = LLVMGetParam(variant->function, 1);
+   vbuffers_ptr = LLVMGetParam(variant->function, 2);
+   start        = LLVMGetParam(variant->function, 3);
+   count        = LLVMGetParam(variant->function, 4);
+   stride       = LLVMGetParam(variant->function, 5);
+
+   lp_build_name(context_ptr, "context");
+   lp_build_name(io_ptr, "io");
+   lp_build_name(vbuffers_ptr, "vbuffers");
+   lp_build_name(start, "start");
+   lp_build_name(count, "count");
+   lp_build_name(stride, "stride");
+
+   /*
+    * Function body
+    */
+
+   block = LLVMAppendBasicBlock(variant->function, "entry");
+   builder = LLVMCreateBuilder();
+   LLVMPositionBuilderAtEnd(builder, block);
+
+   lp_build_context_init(&bld, builder, vs_type);
+
+   end = lp_build_add(&bld, start, count);
+
+   step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
+
+#if DEBUG_STORE
+   lp_build_printf(builder, "start = %d, end = %d, step = %d\n",
+                   start, end, step);
+#endif
+   lp_build_loop_begin(builder, start, &lp_loop);
+   {
+      LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
+      LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
+      LLVMValueRef io;
+      const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
+
+      io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
+      io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
+#if DEBUG_STORE
+      lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
+                      io_itr, io, lp_loop.counter);
+#endif
+      for (i = 0; i < NUM_CHANNELS; ++i) {
+         LLVMValueRef true_index = LLVMBuildAdd(
+            builder,
+            lp_loop.counter,
+            LLVMConstInt(LLVMInt32Type(), i, 0), "");
+         for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
+            struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
+            struct pipe_vertex_buffer *vbuf = &draw->pt.vertex_buffer[
+               velem->vertex_buffer_index];
+            generate_fetch(builder, vbuffers_ptr,
+                           &aos_attribs[j][i], velem, vbuf, true_index);
+         }
+      }
+      convert_to_soa(builder, aos_attribs, inputs,
+                     draw->pt.nr_vertex_elements);
+
+      ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
+      generate_vs(llvm,
+                  builder,
+                  outputs,
+                  ptr_aos,
+                  context_ptr);
+
+      convert_to_aos(builder, io, outputs,
+                     draw->vs.vertex_shader->info.num_outputs,
+                     max_vertices);
+   }
+   lp_build_loop_end_cond(builder, end, step, LLVMIntUGE, &lp_loop);
+
+   LLVMBuildRetVoid(builder);
+
+   LLVMDisposeBuilder(builder);
+
+   /*
+    * Translate the LLVM IR into machine code.
+    */
+
+#ifdef DEBUG
+   if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
+      LLVMDumpValue(variant->function);
+      assert(0);
+   }
+#endif
+
+   LLVMRunFunctionPassManager(llvm->pass, variant->function);
+
+   if (1) {
+      LLVMDumpValue(variant->function);
+      debug_printf("\n");
+   }
+   variant->jit_func = (draw_jit_vert_func)LLVMGetPointerToGlobal(llvm->draw->engine, variant->function);
+
+   if (1)
+      lp_disassemble(variant->jit_func);
+}
+
+void
+draw_llvm_make_variant_key(struct draw_llvm *llvm,
+                           struct draw_llvm_variant_key *key)
+{
+   key->nr_vertex_buffers = llvm->draw->pt.nr_vertex_buffers;
+   key->nr_vertex_elements = llvm->draw->pt.nr_vertex_elements;
+
+   memcpy(key->vertex_buffer,
+          llvm->draw->pt.vertex_buffer,
+          sizeof(struct pipe_vertex_buffer) * PIPE_MAX_ATTRIBS);
+
+   memcpy(key->vertex_element,
+          llvm->draw->pt.vertex_element,
+          sizeof(struct pipe_vertex_element) * PIPE_MAX_ATTRIBS);
+
+   memcpy(&key->vs,
+          &llvm->draw->vs.vertex_shader->state,
+          sizeof(struct pipe_shader_state));
+}
diff --git a/src/gallium/auxiliary/draw/draw_llvm.h b/src/gallium/auxiliary/draw/draw_llvm.h
new file mode 100644
index 00000000000..774eb16d90c
--- /dev/null
+++ b/src/gallium/auxiliary/draw/draw_llvm.h
@@ -0,0 +1,186 @@
+#ifndef DRAW_LLVM_H
+#define DRAW_LLVM_H
+
+#include "draw/draw_private.h"
+
+#include "pipe/p_context.h"
+
+#include <llvm-c/Core.h>
+#include <llvm-c/Analysis.h>
+#include <llvm-c/Target.h>
+#include <llvm-c/ExecutionEngine.h>
+
+struct draw_jit_texture
+{
+   uint32_t width;
+   uint32_t height;
+   uint32_t stride;
+   const void *data;
+};
+
+enum {
+   DRAW_JIT_TEXTURE_WIDTH = 0,
+   DRAW_JIT_TEXTURE_HEIGHT,
+   DRAW_JIT_TEXTURE_STRIDE,
+   DRAW_JIT_TEXTURE_DATA
+};
+
+enum {
+   DRAW_JIT_VERTEX_VERTEX_ID = 0,
+   DRAW_JIT_VERTEX_CLIP,
+   DRAW_JIT_VERTEX_DATA
+};
+
+/**
+ * This structure is passed directly to the generated vertex shader.
+ *
+ * It contains the derived state.
+ *
+ * Changes here must be reflected in the draw_jit_context_* macros.
+ * Changes to the ordering should be avoided.
+ *
+ * Only use types with a clear size and padding here, in particular prefer the
+ * stdint.h types to the basic integer types.
+ */
+struct draw_jit_context
+{
+   const float *vs_constants;
+   const float *gs_constants;
+
+
+   struct draw_jit_texture textures[PIPE_MAX_SAMPLERS];
+};
+
+
+#define draw_jit_context_vs_constants(_builder, _ptr) \
+   lp_build_struct_get(_builder, _ptr, 0, "vs_constants")
+
+#define draw_jit_context_gs_constants(_builder, _ptr) \
+   lp_build_struct_get(_builder, _ptr, 1, "gs_constants")
+
+#define DRAW_JIT_CONTEXT_TEXTURES_INDEX 2
+
+#define draw_jit_context_textures(_builder, _ptr) \
+   lp_build_struct_get_ptr(_builder, _ptr, DRAW_JIT_CONTEXT_TEXTURES_INDEX, "textures")
+
+
+
+#define draw_jit_header_id(_builder, _ptr)              \
+   lp_build_struct_get_ptr(_builder, _ptr, 0, "id")
+
+#define draw_jit_header_clip(_builder, _ptr) \
+   lp_build_struct_get(_builder, _ptr, 1, "clip")
+
+#define draw_jit_header_data(_builder, _ptr)            \
+   lp_build_struct_get_ptr(_builder, _ptr, 2, "data")
+
+/* we are construction a function of the form:
+
+struct vertex_header {
+   uint32 vertex_id;
+
+   float clip[4];
+   float data[][4];
+};
+
+struct draw_jit_context
+{
+   const float *vs_constants;
+   const float *gs_constants;
+
+   struct draw_jit_texture textures[PIPE_MAX_SAMPLERS];
+};
+
+void
+draw_shader(struct draw_jit_context *context,
+            struct vertex_header *io,
+            const void *vbuffers[PIPE_MAX_ATTRIBS],
+            unsigned start,
+            unsigned count,
+            unsigned stride)
+{
+  // do a fetch and a run vertex shader
+  for (int i = 0; i < count; ++i) {
+    struct vertex_header *header = &io[i];
+    header->vertex_id = 0xffff;
+    // follows code-genarted fetch/translate section
+    // for each vertex_element ...
+    codegened_translate(header->data[num_element],
+                       context->vertex_elements[num_element],
+                       context->vertex_buffers,
+                       context->vbuffers);
+
+    codegened_vertex_shader(header->data, context->vs_constants);
+  }
+
+  for (int i = 0; i < count; i += context->primitive_size) {
+     struct vertex_header *prim[MAX_PRIMITIVE_SIZE];
+     for (int j = 0; j < context->primitive_size; ++j) {
+       header[j] = &io[i + j];
+     }
+     codegened_geometry_shader(prim, gs_constants);
+  }
+}
+*/
+
+typedef void
+(*draw_jit_vert_func)(struct draw_jit_context *context,
+                      struct vertex_header *io,
+                      const char *vbuffers[PIPE_MAX_ATTRIBS],
+                      unsigned start,
+                      unsigned count,
+                      unsigned stride);
+
+struct draw_llvm {
+   struct draw_context *draw;
+
+   struct draw_jit_context jit_context;
+
+   LLVMModuleRef module;
+   LLVMExecutionEngineRef engine;
+   LLVMModuleProviderRef provider;
+   LLVMTargetDataRef target;
+   LLVMPassManagerRef pass;
+
+   LLVMTypeRef context_ptr_type;
+   LLVMTypeRef vertex_header_ptr_type;
+   LLVMTypeRef buffer_ptr_type;
+};
+
+
+struct draw_llvm_variant_key
+{
+   struct pipe_vertex_buffer  vertex_buffer[PIPE_MAX_ATTRIBS];
+   unsigned nr_vertex_buffers;
+   struct pipe_vertex_element vertex_element[PIPE_MAX_ATTRIBS];
+   unsigned nr_vertex_elements;
+   struct pipe_shader_state   vs;
+};
+
+struct draw_llvm_variant
+{
+   struct draw_llvm_variant_key key;
+   LLVMValueRef function;
+   draw_jit_vert_func jit_func;
+
+   struct draw_llvm_variant *next;
+};
+
+struct draw_llvm *
+draw_llvm_create(struct draw_context *draw);
+
+void
+draw_llvm_destroy(struct draw_llvm *llvm);
+
+struct draw_llvm_variant *
+draw_llvm_prepare(struct draw_llvm *llvm, int num_inputs);
+
+void
+draw_llvm_make_variant_key(struct draw_llvm *llvm,
+                           struct draw_llvm_variant_key *key);
+
+LLVMValueRef
+draw_llvm_translate_from(LLVMBuilderRef builder,
+                         LLVMValueRef vbuffer,
+                         enum pipe_format from_format);
+#endif
diff --git a/src/gallium/auxiliary/draw/draw_llvm_translate.c b/src/gallium/auxiliary/draw/draw_llvm_translate.c
new file mode 100644
index 00000000000..b29ebdec3a9
--- /dev/null
+++ b/src/gallium/auxiliary/draw/draw_llvm_translate.c
@@ -0,0 +1,484 @@
+#include "draw_private.h"
+#include "draw_context.h"
+
+#include "draw_llvm.h"
+
+#include "gallivm/lp_bld_arit.h"
+#include "gallivm/lp_bld_interp.h"
+#include "gallivm/lp_bld_struct.h"
+#include "gallivm/lp_bld_type.h"
+#include "gallivm/lp_bld_flow.h"
+#include "gallivm/lp_bld_debug.h"
+#include "gallivm/lp_bld_tgsi.h"
+#include "gallivm/lp_bld_printf.h"
+
+#include "util/u_memory.h"
+#include "pipe/p_state.h"
+
+
+#define DRAW_DBG 0
+
+static  LLVMValueRef
+from_64_float(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMPointerType(LLVMDoubleType(), 0) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   return LLVMBuildFPTrunc(builder, l, LLVMFloatType(), "");
+}
+
+static LLVMValueRef
+from_32_float(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMPointerType(LLVMFloatType(), 0) , "");
+   return LLVMBuildLoad(builder, bc, "");
+}
+
+static INLINE LLVMValueRef
+from_8_uscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, val, "");
+   return LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+}
+
+static INLINE LLVMValueRef
+from_16_uscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMPointerType(LLVMIntType(16), 0) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   return LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+}
+
+static INLINE LLVMValueRef
+from_32_uscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMPointerType(LLVMIntType(32), 0) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   return LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+}
+
+static INLINE LLVMValueRef
+from_8_sscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, val, "");
+   return LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+}
+
+static INLINE LLVMValueRef
+from_16_sscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMPointerType(LLVMIntType(16), 0) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   return LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+}
+
+static INLINE LLVMValueRef
+from_32_sscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMPointerType(LLVMIntType(32), 0) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   return LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+}
+
+
+static INLINE LLVMValueRef
+from_8_unorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, val, "");
+   LLVMValueRef uscaled = LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 255.), "");
+}
+
+static INLINE LLVMValueRef
+from_16_unorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMPointerType(LLVMIntType(16), 0) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   LLVMValueRef uscaled = LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 65535.), "");
+}
+
+static INLINE LLVMValueRef
+from_32_unorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMPointerType(LLVMIntType(32), 0) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   LLVMValueRef uscaled = LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 4294967295.), "");
+}
+
+static INLINE LLVMValueRef
+from_8_snorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, val, "");
+   LLVMValueRef uscaled = LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 127.0), "");
+}
+
+static INLINE LLVMValueRef
+from_16_snorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMPointerType(LLVMIntType(16), 0) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   LLVMValueRef uscaled = LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 32767.0f), "");
+}
+
+static INLINE LLVMValueRef
+from_32_snorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMPointerType(LLVMIntType(32), 0) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   LLVMValueRef uscaled = LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 2147483647.0), "");
+}
+
+static INLINE LLVMValueRef
+from_32_fixed(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMPointerType(LLVMIntType(32), 0) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   LLVMValueRef uscaled = LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 65536.0), "");
+}
+
+static LLVMValueRef
+to_64_float(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPExt(builder, l, LLVMDoubleType(), "");
+}
+
+static LLVMValueRef
+to_32_float(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   return LLVMBuildLoad(builder, fp, "");
+}
+
+static INLINE LLVMValueRef
+to_8_uscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToUI(builder, l, LLVMIntType(8), "");
+}
+
+static INLINE LLVMValueRef
+to_16_uscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToUI(builder, l, LLVMIntType(16), "");
+}
+
+static INLINE LLVMValueRef
+to_32_uscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToUI(builder, l, LLVMIntType(32), "");
+}
+
+static INLINE LLVMValueRef
+to_8_sscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToSI(builder, l, LLVMIntType(8), "");
+}
+
+static INLINE LLVMValueRef
+to_16_sscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToSI(builder, l, LLVMIntType(16), "");
+}
+
+static INLINE LLVMValueRef
+to_32_sscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToSI(builder, l, LLVMIntType(32), "");
+}
+
+static INLINE LLVMValueRef
+to_8_unorm(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToUI(builder, l, LLVMIntType(8), "");
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 255.), "");
+}
+
+static INLINE LLVMValueRef
+to_16_unorm(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToUI(builder, l, LLVMIntType(32), "");
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 65535.), "");
+}
+
+static INLINE LLVMValueRef
+to_32_unorm(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToUI(builder, l, LLVMIntType(32), "");
+
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 4294967295.), "");
+}
+
+static INLINE LLVMValueRef
+to_8_snorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, val, "");
+   LLVMValueRef uscaled = LLVMBuildFPToSI(builder, l, LLVMIntType(8), "");
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 127.0), "");
+}
+
+static INLINE LLVMValueRef
+to_16_snorm(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToSI(builder, l, LLVMIntType(16), "");
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 32767.0f), "");
+}
+
+static INLINE LLVMValueRef
+to_32_snorm(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToSI(builder, l, LLVMIntType(32), "");
+
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 2147483647.0), "");
+}
+
+static INLINE LLVMValueRef
+to_32_fixed(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToSI(builder, l, LLVMIntType(32), "");
+
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(LLVMFloatType(), 65536.0), "");
+}
+
+typedef LLVMValueRef (*from_func)(LLVMBuilderRef, LLVMValueRef);
+typedef  LLVMValueRef (*to_func)(LLVMBuilderRef, LLVMValueRef);
+
+/* so that underneath can avoid function calls which are prohibited
+ * for static initialization we need this conversion */
+enum ll_type {
+   LL_Double,
+   LL_Float,
+   LL_Int32,
+   LL_Int16,
+   LL_Int8
+};
+
+static INLINE LLVMTypeRef
+ll_type_to_llvm(enum ll_type type)
+{
+   switch (type) {
+   case LL_Double:
+      return LLVMDoubleType();
+   case LL_Float:
+      return LLVMFloatType();
+   case LL_Int32:
+      return LLVMInt32Type();
+   case LL_Int16:
+      return LLVMIntType(16);
+   case LL_Int8:
+      return LLVMIntType(8);
+   }
+   return LLVMIntType(8);
+}
+
+static INLINE int
+ll_type_size(enum ll_type type)
+{
+   switch (type) {
+   case LL_Double:
+      return 8;
+   case LL_Float:
+      return 4;
+   case LL_Int32:
+      return 4;
+   case LL_Int16:
+      return 2;
+   case LL_Int8:
+      return 1;
+   }
+   return 1;
+}
+
+struct draw_llvm_translate {
+   int format;
+   from_func from;
+   to_func to;
+   enum ll_type type;
+   int num_components;
+} translates[] =
+{
+   {PIPE_FORMAT_R64_FLOAT,          from_64_float, to_64_float, LL_Double, 1},
+   {PIPE_FORMAT_R64G64_FLOAT,       from_64_float, to_64_float, LL_Double, 2},
+   {PIPE_FORMAT_R64G64B64_FLOAT,    from_64_float, to_64_float, LL_Double, 3},
+   {PIPE_FORMAT_R64G64B64A64_FLOAT, from_64_float, to_64_float, LL_Double, 4},
+   {PIPE_FORMAT_R32_FLOAT,          from_32_float, to_32_float, LL_Float, 1},
+   {PIPE_FORMAT_R32G32_FLOAT,       from_32_float, to_32_float, LL_Float, 2},
+   {PIPE_FORMAT_R32G32B32_FLOAT,    from_32_float, to_32_float, LL_Float, 3},
+   {PIPE_FORMAT_R32G32B32A32_FLOAT, from_32_float, to_32_float, LL_Float, 4},
+
+   {PIPE_FORMAT_R32_UNORM,          from_32_unorm, to_32_unorm, LL_Int32, 1},
+   {PIPE_FORMAT_R32G32_UNORM,       from_32_unorm, to_32_unorm, LL_Int32, 2},
+   {PIPE_FORMAT_R32G32B32_UNORM,    from_32_unorm, to_32_unorm, LL_Int32, 3},
+   {PIPE_FORMAT_R32G32B32A32_UNORM, from_32_unorm, to_32_unorm, LL_Int32, 4},
+
+   {PIPE_FORMAT_R32_USCALED,          from_32_uscaled, to_32_uscaled, LL_Int32, 1},
+   {PIPE_FORMAT_R32G32_USCALED,       from_32_uscaled, to_32_uscaled, LL_Int32, 2},
+   {PIPE_FORMAT_R32G32B32_USCALED,    from_32_uscaled, to_32_uscaled, LL_Int32, 3},
+   {PIPE_FORMAT_R32G32B32A32_USCALED, from_32_uscaled, to_32_uscaled, LL_Int32, 4},
+
+   {PIPE_FORMAT_R32_SNORM,          from_32_snorm, to_32_snorm, LL_Int32, 1},
+   {PIPE_FORMAT_R32G32_SNORM,       from_32_snorm, to_32_snorm, LL_Int32, 2},
+   {PIPE_FORMAT_R32G32B32_SNORM,    from_32_snorm, to_32_snorm, LL_Int32, 3},
+   {PIPE_FORMAT_R32G32B32A32_SNORM, from_32_snorm, to_32_snorm, LL_Int32, 4},
+
+   {PIPE_FORMAT_R32_SSCALED,          from_32_sscaled, to_32_sscaled, LL_Int32, 1},
+   {PIPE_FORMAT_R32G32_SSCALED,       from_32_sscaled, to_32_sscaled, LL_Int32, 2},
+   {PIPE_FORMAT_R32G32B32_SSCALED,    from_32_sscaled, to_32_sscaled, LL_Int32, 3},
+   {PIPE_FORMAT_R32G32B32A32_SSCALED, from_32_sscaled, to_32_sscaled, LL_Int32, 4},
+
+   {PIPE_FORMAT_R16_UNORM,          from_16_unorm, to_16_unorm, LL_Int16, 1},
+   {PIPE_FORMAT_R16G16_UNORM,       from_16_unorm, to_16_unorm, LL_Int16, 2},
+   {PIPE_FORMAT_R16G16B16_UNORM,    from_16_unorm, to_16_unorm, LL_Int16, 3},
+   {PIPE_FORMAT_R16G16B16A16_UNORM, from_16_unorm, to_16_unorm, LL_Int16, 4},
+
+   {PIPE_FORMAT_R16_USCALED,          from_16_uscaled, to_16_uscaled, LL_Int16, 1},
+   {PIPE_FORMAT_R16G16_USCALED,       from_16_uscaled, to_16_uscaled, LL_Int16, 2},
+   {PIPE_FORMAT_R16G16B16_USCALED,    from_16_uscaled, to_16_uscaled, LL_Int16, 3},
+   {PIPE_FORMAT_R16G16B16A16_USCALED, from_16_uscaled, to_16_uscaled, LL_Int16, 4},
+
+   {PIPE_FORMAT_R16_SNORM,          from_16_snorm, to_16_snorm, LL_Int16, 1},
+   {PIPE_FORMAT_R16G16_SNORM,       from_16_snorm, to_16_snorm, LL_Int16, 2},
+   {PIPE_FORMAT_R16G16B16_SNORM,    from_16_snorm, to_16_snorm, LL_Int16, 3},
+   {PIPE_FORMAT_R16G16B16A16_SNORM, from_16_snorm, to_16_snorm, LL_Int16, 4},
+
+   {PIPE_FORMAT_R16_SSCALED,          from_16_sscaled, to_16_sscaled, LL_Int16, 1},
+   {PIPE_FORMAT_R16G16_SSCALED,       from_16_sscaled, to_16_sscaled, LL_Int16, 2},
+   {PIPE_FORMAT_R16G16B16_SSCALED,    from_16_sscaled, to_16_sscaled, LL_Int16, 3},
+   {PIPE_FORMAT_R16G16B16A16_SSCALED, from_16_sscaled, to_16_sscaled, LL_Int16, 4},
+
+   {PIPE_FORMAT_R8_UNORM,       from_8_unorm, to_8_unorm, LL_Int8, 1},
+   {PIPE_FORMAT_R8G8_UNORM,     from_8_unorm, to_8_unorm, LL_Int8, 2},
+   {PIPE_FORMAT_R8G8B8_UNORM,   from_8_unorm, to_8_unorm, LL_Int8, 3},
+   {PIPE_FORMAT_R8G8B8A8_UNORM, from_8_unorm, to_8_unorm, LL_Int8, 4},
+
+   {PIPE_FORMAT_R8_USCALED,       from_8_uscaled, to_8_uscaled, LL_Int8, 1},
+   {PIPE_FORMAT_R8G8_USCALED,     from_8_uscaled, to_8_uscaled, LL_Int8, 2},
+   {PIPE_FORMAT_R8G8B8_USCALED,   from_8_uscaled, to_8_uscaled, LL_Int8, 3},
+   {PIPE_FORMAT_R8G8B8A8_USCALED, from_8_uscaled, to_8_uscaled, LL_Int8, 4},
+
+   {PIPE_FORMAT_R8_SNORM,       from_8_snorm, to_8_snorm, LL_Int8, 1},
+   {PIPE_FORMAT_R8G8_SNORM,     from_8_snorm, to_8_snorm, LL_Int8, 2},
+   {PIPE_FORMAT_R8G8B8_SNORM,   from_8_snorm, to_8_snorm, LL_Int8, 3},
+   {PIPE_FORMAT_R8G8B8A8_SNORM, from_8_snorm, to_8_snorm, LL_Int8, 4},
+
+   {PIPE_FORMAT_R8_SSCALED,       from_8_sscaled, to_8_sscaled, LL_Int8, 1},
+   {PIPE_FORMAT_R8G8_SSCALED,     from_8_sscaled, to_8_sscaled, LL_Int8, 2},
+   {PIPE_FORMAT_R8G8B8_SSCALED,   from_8_sscaled, to_8_sscaled, LL_Int8, 3},
+   {PIPE_FORMAT_R8G8B8A8_SSCALED, from_8_sscaled, to_8_sscaled, LL_Int8, 4},
+
+   {PIPE_FORMAT_R32_FIXED,          from_32_fixed, to_32_fixed, LL_Int32, 1},
+   {PIPE_FORMAT_R32G32_FIXED,       from_32_fixed, to_32_fixed, LL_Int32, 2},
+   {PIPE_FORMAT_R32G32B32_FIXED,    from_32_fixed, to_32_fixed, LL_Int32, 3},
+   {PIPE_FORMAT_R32G32B32A32_FIXED, from_32_fixed, to_32_fixed, LL_Int32, 4},
+
+   {PIPE_FORMAT_A8R8G8B8_UNORM, from_8_unorm, to_8_unorm, LL_Int8, 4},
+   {PIPE_FORMAT_B8G8R8A8_UNORM, from_8_unorm, to_8_unorm, LL_Int8, 4},
+};
+
+
+static LLVMValueRef
+fetch(LLVMBuilderRef builder,
+      LLVMValueRef ptr, int val_size, int nr_components,
+      from_func func)
+{
+   int i;
+   int offset = 0;
+   LLVMValueRef res = LLVMConstNull(
+      LLVMVectorType(LLVMFloatType(), 4));
+   LLVMValueRef defaults[4];
+
+   defaults[0] = LLVMConstReal(LLVMFloatType(), 0);
+   defaults[1] = LLVMConstReal(LLVMFloatType(), 0);
+   defaults[2] = LLVMConstReal(LLVMFloatType(), 0);
+   defaults[3] = LLVMConstReal(LLVMFloatType(), 1);
+
+   for (i = 0; i < nr_components; ++i) {
+      LLVMValueRef src_index = LLVMConstInt(LLVMInt32Type(), offset, 0);
+      LLVMValueRef dst_index = LLVMConstInt(LLVMInt32Type(), i, 0);
+      LLVMValueRef src_tmp;
+      LLVMValueRef component;
+
+      src_tmp = LLVMBuildGEP(builder, ptr, &src_index, 1, "src_tmp");
+
+      /* convert src_tmp to float */
+      component = func(builder, src_tmp);
+
+      /* vec.comp = component */
+      res = LLVMBuildInsertElement(builder,
+                                   res,
+                                   component,
+                                   dst_index, "");
+      offset += val_size;
+   }
+   for (; i < 4; ++i) {
+      LLVMValueRef dst_index = LLVMConstInt(LLVMInt32Type(), i, 0);
+      res = LLVMBuildInsertElement(builder,
+                                   res,
+                                   defaults[i],
+                                   dst_index, "");
+   }
+   return res;
+}
+
+
+LLVMValueRef
+draw_llvm_translate_from(LLVMBuilderRef builder,
+                         LLVMValueRef vbuffer,
+                         enum pipe_format from_format)
+{
+   int i;
+   for (i = 0; i < Elements(translates); ++i) {
+      if (translates[i].format == from_format) {
+         /*LLVMTypeRef type = ll_type_to_llvm(translates[i].type);*/
+         return fetch(builder,
+                      vbuffer,
+                      ll_type_size(translates[i].type),
+                      translates[i].num_components,
+                      translates[i].from);
+      }
+   }
+   return LLVMGetUndef(LLVMVectorType(LLVMFloatType(), 4));
+}
diff --git a/src/gallium/auxiliary/draw/draw_private.h b/src/gallium/auxiliary/draw/draw_private.h
index 1e6e01af9e2..7e24e5fd6fc 100644
--- a/src/gallium/auxiliary/draw/draw_private.h
+++ b/src/gallium/auxiliary/draw/draw_private.h
@@ -46,6 +46,10 @@
 
 #include "tgsi/tgsi_scan.h"
 
+#ifdef DRAW_LLVM
+#include <llvm-c/ExecutionEngine.h>
+#endif
+
 
 struct pipe_context;
 struct draw_vertex_shader;
@@ -237,9 +241,16 @@ struct draw_context
 
    unsigned instance_id;
 
+#ifdef DRAW_LLVM
+   LLVMExecutionEngineRef engine;
+#endif
    void *driver_private;
 };
 
+/*******************************************************************************
+ * Draw common initialization code
+ */
+boolean draw_init(struct draw_context *draw);
 
 /*******************************************************************************
  * Vertex shader code:
diff --git a/src/gallium/auxiliary/draw/draw_pt.c b/src/gallium/auxiliary/draw/draw_pt.c
index a8cdc57ad96..43126c6c881 100644
--- a/src/gallium/auxiliary/draw/draw_pt.c
+++ b/src/gallium/auxiliary/draw/draw_pt.c
@@ -140,7 +140,9 @@ boolean draw_pt_init( struct draw_context *draw )
    if (!draw->pt.middle.fetch_shade_emit)
       return FALSE;
 
-   draw->pt.middle.general = draw_pt_fetch_pipeline_or_emit( draw );
+   draw->pt.middle.general = draw_pt_fetch_pipeline_or_emit_llvm( draw );
+   if (!draw->pt.middle.general)
+      draw->pt.middle.general = draw_pt_fetch_pipeline_or_emit( draw );
    if (!draw->pt.middle.general)
       return FALSE;
 
diff --git a/src/gallium/auxiliary/draw/draw_pt.h b/src/gallium/auxiliary/draw/draw_pt.h
index d5e0d92a605..c2797a759ee 100644
--- a/src/gallium/auxiliary/draw/draw_pt.h
+++ b/src/gallium/auxiliary/draw/draw_pt.h
@@ -147,6 +147,7 @@ struct draw_pt_front_end *draw_pt_varray(struct draw_context *draw);
 struct draw_pt_middle_end *draw_pt_fetch_emit( struct draw_context *draw );
 struct draw_pt_middle_end *draw_pt_middle_fse( struct draw_context *draw );
 struct draw_pt_middle_end *draw_pt_fetch_pipeline_or_emit(struct draw_context *draw);
+struct draw_pt_middle_end *draw_pt_fetch_pipeline_or_emit_llvm(struct draw_context *draw);
 
 
 
diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline_llvm.c b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline_llvm.c
new file mode 100644
index 00000000000..d2ed0eb3d18
--- /dev/null
+++ b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline_llvm.c
@@ -0,0 +1,464 @@
+/**************************************************************************
+ *
+ * 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 TUNGSTEN GRAPHICS 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 "draw/draw_context.h"
+#include "draw/draw_vbuf.h"
+#include "draw/draw_vertex.h"
+#include "draw/draw_pt.h"
+#include "draw/draw_vs.h"
+#include "draw/draw_gs.h"
+#include "draw/draw_llvm.h"
+
+#include "translate/translate.h"
+
+
+struct llvm_middle_end {
+   struct draw_pt_middle_end base;
+   struct draw_context *draw;
+
+   struct pt_emit *emit;
+   struct pt_fetch *fetch;
+   struct pt_post_vs *post_vs;
+
+
+   unsigned vertex_data_offset;
+   unsigned vertex_size;
+   unsigned prim;
+   unsigned opt;
+
+   struct draw_llvm *llvm;
+   struct draw_llvm_variant *variants;
+   struct draw_llvm_variant *current_variant;
+};
+
+
+static void
+llvm_middle_end_prepare( struct draw_pt_middle_end *middle,
+                         unsigned prim,
+                         unsigned opt,
+                         unsigned *max_vertices )
+{
+   struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
+   struct draw_context *draw = fpme->draw;
+   struct draw_vertex_shader *vs = draw->vs.vertex_shader;
+   struct draw_geometry_shader *gs = draw->gs.geometry_shader;
+   struct draw_llvm_variant_key key;
+   struct draw_llvm_variant *variant = NULL;
+   unsigned i;
+   unsigned instance_id_index = ~0;
+
+   /* Add one to num_outputs because the pipeline occasionally tags on
+    * an additional texcoord, eg for AA lines.
+    */
+   unsigned nr = MAX2( vs->info.num_inputs,
+		       vs->info.num_outputs + 1 );
+
+   /* Scan for instanceID system value.
+    */
+   for (i = 0; i < vs->info.num_inputs; i++) {
+      if (vs->info.input_semantic_name[i] == TGSI_SEMANTIC_INSTANCEID) {
+         instance_id_index = i;
+         break;
+      }
+   }
+
+   fpme->prim = prim;
+   fpme->opt = opt;
+
+   /* Always leave room for the vertex header whether we need it or
+    * not.  It's hard to get rid of it in particular because of the
+    * viewport code in draw_pt_post_vs.c.
+    */
+   fpme->vertex_size = sizeof(struct vertex_header) + nr * 4 * sizeof(float);
+
+
+
+   draw_pt_fetch_prepare( fpme->fetch,
+                          vs->info.num_inputs,
+                          fpme->vertex_size,
+                          instance_id_index );
+   if (opt & PT_SHADE) {
+      vs->prepare(vs, draw);
+      draw_geometry_shader_prepare(gs, draw);
+   }
+
+
+   /* XXX: it's not really gl rasterization rules we care about here,
+    * but gl vs dx9 clip spaces.
+    */
+   draw_pt_post_vs_prepare( fpme->post_vs,
+			    (boolean)draw->bypass_clipping,
+			    (boolean)(draw->identity_viewport),
+			    (boolean)draw->rasterizer->gl_rasterization_rules,
+			    (draw->vs.edgeflag_output ? true : false) );
+
+   if (!(opt & PT_PIPELINE)) {
+      draw_pt_emit_prepare( fpme->emit,
+			    prim,
+                            max_vertices );
+
+      *max_vertices = MAX2( *max_vertices,
+                            DRAW_PIPE_MAX_VERTICES );
+   }
+   else {
+      *max_vertices = DRAW_PIPE_MAX_VERTICES;
+   }
+
+   /* return even number */
+   *max_vertices = *max_vertices & ~1;
+
+   draw_llvm_make_variant_key(fpme->llvm, &key);
+
+   variant = fpme->variants;
+   while(variant) {
+      if(memcmp(&variant->key, &key, sizeof key) == 0)
+         break;
+
+      variant = variant->next;
+   }
+
+   if (!variant) {
+      variant = draw_llvm_prepare(fpme->llvm, nr);
+      variant->next = fpme->variants;
+      fpme->variants = variant;
+   }
+   fpme->current_variant = variant;
+
+   /*XXX we only support one constant buffer */
+   fpme->llvm->jit_context.vs_constants =
+      draw->pt.user.vs_constants[0];
+   fpme->llvm->jit_context.gs_constants =
+      draw->pt.user.gs_constants[0];
+}
+
+
+
+static void llvm_middle_end_run( struct draw_pt_middle_end *middle,
+                                 const unsigned *fetch_elts,
+                                 unsigned fetch_count,
+                                 const ushort *draw_elts,
+                                 unsigned draw_count )
+{
+   struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
+   struct draw_context *draw = fpme->draw;
+   struct draw_vertex_shader *vshader = draw->vs.vertex_shader;
+   struct draw_geometry_shader *gshader = draw->gs.geometry_shader;
+   unsigned opt = fpme->opt;
+   unsigned alloc_count = align( fetch_count, 4 );
+
+   struct vertex_header *pipeline_verts =
+      (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
+
+   if (!pipeline_verts) {
+      /* Not much we can do here - just skip the rendering.
+       */
+      assert(0);
+      return;
+   }
+
+   /* Fetch into our vertex buffer
+    */
+   draw_pt_fetch_run( fpme->fetch,
+		      fetch_elts,
+		      fetch_count,
+		      (char *)pipeline_verts );
+
+   /* Run the shader, note that this overwrites the data[] parts of
+    * the pipeline verts.  If there is no shader, eg if
+    * bypass_vs_clip_and_viewport, then the inputs == outputs, and are
+    * already in the correct place.*/
+   if (opt & PT_SHADE)
+   {
+      vshader->run_linear(vshader,
+                          (const float (*)[4])pipeline_verts->data,
+                          (      float (*)[4])pipeline_verts->data,
+                          draw->pt.user.vs_constants,
+                          fetch_count,
+                          fpme->vertex_size,
+                          fpme->vertex_size);
+      if (gshader)
+         draw_geometry_shader_run(gshader,
+                                  (const float (*)[4])pipeline_verts->data,
+                                  (      float (*)[4])pipeline_verts->data,
+                                  draw->pt.user.gs_constants,
+                                  fetch_count,
+                                  fpme->vertex_size,
+                                  fpme->vertex_size);
+   }
+
+   if (draw_pt_post_vs_run( fpme->post_vs,
+			    pipeline_verts,
+			    fetch_count,
+			    fpme->vertex_size ))
+   {
+      opt |= PT_PIPELINE;
+   }
+
+   /* Do we need to run the pipeline?
+    */
+   if (opt & PT_PIPELINE) {
+      draw_pipeline_run( fpme->draw,
+                         fpme->prim,
+                         pipeline_verts,
+                         fetch_count,
+                         fpme->vertex_size,
+                         draw_elts,
+                         draw_count );
+   }
+   else {
+      draw_pt_emit( fpme->emit,
+		    (const float (*)[4])pipeline_verts->data,
+		    fetch_count,
+		    fpme->vertex_size,
+		    draw_elts,
+		    draw_count );
+   }
+
+
+   FREE(pipeline_verts);
+}
+
+
+static void llvm_middle_end_linear_run( struct draw_pt_middle_end *middle,
+                                       unsigned start,
+                                       unsigned count)
+{
+   struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
+   struct draw_context *draw = fpme->draw;
+   unsigned opt = fpme->opt;
+   unsigned alloc_count = align( count, 4 );
+
+   struct vertex_header *pipeline_verts =
+      (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
+
+   if (!pipeline_verts) {
+      /* Not much we can do here - just skip the rendering.
+       */
+      assert(0);
+      return;
+   }
+
+#if 0
+   debug_printf("#### Pipeline = %p (data = %p)\n",
+                pipeline_verts, pipeline_verts->data);
+#endif
+   fpme->current_variant->jit_func( &fpme->llvm->jit_context,
+                                    pipeline_verts,
+                                    (const char **)draw->pt.user.vbuffer,
+                                    start,
+                                    count,
+                                    fpme->vertex_size );
+
+   if (draw_pt_post_vs_run( fpme->post_vs,
+			    pipeline_verts,
+			    count,
+			    fpme->vertex_size ))
+   {
+      opt |= PT_PIPELINE;
+   }
+
+   /* Do we need to run the pipeline?
+    */
+   if (opt & PT_PIPELINE) {
+      draw_pipeline_run_linear( fpme->draw,
+                                fpme->prim,
+                                pipeline_verts,
+                                count,
+                                fpme->vertex_size);
+   }
+   else {
+      draw_pt_emit_linear( fpme->emit,
+                           (const float (*)[4])pipeline_verts->data,
+                           fpme->vertex_size,
+                           count );
+   }
+
+   FREE(pipeline_verts);
+}
+
+
+
+static boolean
+llvm_middle_end_linear_run_elts( struct draw_pt_middle_end *middle,
+                                 unsigned start,
+                                 unsigned count,
+                                 const ushort *draw_elts,
+                                 unsigned draw_count )
+{
+   struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
+   struct draw_context *draw = fpme->draw;
+   struct draw_vertex_shader *shader = draw->vs.vertex_shader;
+   struct draw_geometry_shader *geometry_shader = draw->gs.geometry_shader;
+   unsigned opt = fpme->opt;
+   unsigned alloc_count = align( count, 4 );
+
+   struct vertex_header *pipeline_verts =
+      (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
+
+   if (!pipeline_verts)
+      return FALSE;
+
+   /* Fetch into our vertex buffer
+    */
+   draw_pt_fetch_run_linear( fpme->fetch,
+                             start,
+                             count,
+                             (char *)pipeline_verts );
+
+   /* Run the shader, note that this overwrites the data[] parts of
+    * the pipeline verts.  If there is no shader, ie if
+    * bypass_vs_clip_and_viewport, then the inputs == outputs, and are
+    * already in the correct place.
+    */
+   if (opt & PT_SHADE)
+   {
+      shader->run_linear(shader,
+			 (const float (*)[4])pipeline_verts->data,
+			 (      float (*)[4])pipeline_verts->data,
+			 draw->pt.user.vs_constants,
+			 count,
+			 fpme->vertex_size,
+			 fpme->vertex_size);
+
+      if (geometry_shader)
+         draw_geometry_shader_run(geometry_shader,
+                                  (const float (*)[4])pipeline_verts->data,
+                                  (      float (*)[4])pipeline_verts->data,
+                                  draw->pt.user.gs_constants,
+                                  count,
+                                  fpme->vertex_size,
+                                  fpme->vertex_size);
+   }
+
+   if (draw_pt_post_vs_run( fpme->post_vs,
+			    pipeline_verts,
+			    count,
+			    fpme->vertex_size ))
+   {
+      opt |= PT_PIPELINE;
+   }
+
+   /* Do we need to run the pipeline?
+    */
+   if (opt & PT_PIPELINE) {
+      draw_pipeline_run( fpme->draw,
+                         fpme->prim,
+                         pipeline_verts,
+                         count,
+                         fpme->vertex_size,
+                         draw_elts,
+                         draw_count );
+   }
+   else {
+      draw_pt_emit( fpme->emit,
+		    (const float (*)[4])pipeline_verts->data,
+		    count,
+		    fpme->vertex_size,
+		    draw_elts,
+		    draw_count );
+   }
+
+   FREE(pipeline_verts);
+   return TRUE;
+}
+
+
+
+static void llvm_middle_end_finish( struct draw_pt_middle_end *middle )
+{
+   /* nothing to do */
+}
+
+static void llvm_middle_end_destroy( struct draw_pt_middle_end *middle )
+{
+   struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
+
+   if (fpme->fetch)
+      draw_pt_fetch_destroy( fpme->fetch );
+
+   if (fpme->emit)
+      draw_pt_emit_destroy( fpme->emit );
+
+   if (fpme->post_vs)
+      draw_pt_post_vs_destroy( fpme->post_vs );
+
+   if (fpme->llvm)
+      draw_llvm_destroy( fpme->llvm );
+
+   FREE(middle);
+}
+
+
+struct draw_pt_middle_end *draw_pt_fetch_pipeline_or_emit_llvm( struct draw_context *draw )
+{
+   struct llvm_middle_end *fpme = 0;
+
+   if (!draw->engine)
+      return NULL;
+
+   fpme = CALLOC_STRUCT( llvm_middle_end );
+   if (!fpme)
+      goto fail;
+
+   fpme->base.prepare         = llvm_middle_end_prepare;
+   fpme->base.run             = llvm_middle_end_run;
+   fpme->base.run_linear      = llvm_middle_end_linear_run;
+   fpme->base.run_linear_elts = llvm_middle_end_linear_run_elts;
+   fpme->base.finish          = llvm_middle_end_finish;
+   fpme->base.destroy         = llvm_middle_end_destroy;
+
+   fpme->draw = draw;
+
+   fpme->fetch = draw_pt_fetch_create( draw );
+   if (!fpme->fetch)
+      goto fail;
+
+   fpme->post_vs = draw_pt_post_vs_create( draw );
+   if (!fpme->post_vs)
+      goto fail;
+
+   fpme->emit = draw_pt_emit_create( draw );
+   if (!fpme->emit)
+      goto fail;
+
+   fpme->llvm = draw_llvm_create(draw);
+   if (!fpme->llvm)
+      goto fail;
+
+   fpme->variants = NULL;
+   fpme->current_variant = NULL;
+
+   return &fpme->base;
+
+ fail:
+   if (fpme)
+      llvm_middle_end_destroy( &fpme->base );
+
+   return NULL;
+}
diff --git a/src/gallium/auxiliary/draw/draw_pt_post_vs.c b/src/gallium/auxiliary/draw/draw_pt_post_vs.c
index 9728d5c2bdf..5525dfc748d 100644
--- a/src/gallium/auxiliary/draw/draw_pt_post_vs.c
+++ b/src/gallium/auxiliary/draw/draw_pt_post_vs.c
@@ -108,6 +108,11 @@ static boolean post_vs_cliptest_viewport_gl( struct pt_post_vs *pvs,
    for (j = 0; j < count; j++) {
       float *position = out->data[pos];
 
+#if 0
+      debug_printf("%d) io = %p, data = %p = [%f, %f, %f, %f]\n",
+                   j, out, position, position[0], position[1], position[2], position[3]);
+#endif
+
       out->clip[0] = position[0];
       out->clip[1] = position[1];
       out->clip[2] = position[2];
diff --git a/src/gallium/auxiliary/draw/draw_vs_llvm.c b/src/gallium/auxiliary/draw/draw_vs_llvm.c
index 5f7a645f5d8..0c483de4071 100644
--- a/src/gallium/auxiliary/draw/draw_vs_llvm.c
+++ b/src/gallium/auxiliary/draw/draw_vs_llvm.c
@@ -40,7 +40,7 @@
 
 #include "tgsi/tgsi_parse.h"
 
-#ifdef MESA_LLVM
+#ifdef DRAW_LLVM
 
 struct draw_llvm_vertex_shader {
    struct draw_vertex_shader base;
@@ -64,12 +64,8 @@ vs_llvm_run_linear( struct draw_vertex_shader *base,
 		   unsigned input_stride,
 		   unsigned output_stride )
 {
-   struct draw_llvm_vertex_shader *shader =
-      (struct draw_llvm_vertex_shader *)base;
 }
 
-
-
 static void
 vs_llvm_delete( struct draw_vertex_shader *base )
 {
@@ -90,6 +86,7 @@ struct draw_vertex_shader *
 draw_create_vs_llvm(struct draw_context *draw,
 		    const struct pipe_shader_state *templ)
 {
+#if 0
    struct draw_llvm_vertex_shader *vs;
 
    vs = CALLOC_STRUCT( draw_llvm_vertex_shader );
@@ -113,6 +110,8 @@ draw_create_vs_llvm(struct draw_context *draw,
    vs->machine = draw->vs.machine;
 
    return &vs->base;
+#endif
+   return NULL;
 }
 
 
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_flow.c b/src/gallium/auxiliary/gallivm/lp_bld_flow.c
index 106fc03e46f..e60ab4f6ba1 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_flow.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_flow.c
@@ -570,6 +570,35 @@ lp_build_loop_end(LLVMBuilderRef builder,
    LLVMPositionBuilderAtEnd(builder, after_block);
 }
 
+void
+lp_build_loop_end_cond(LLVMBuilderRef builder,
+                       LLVMValueRef end,
+                       LLVMValueRef step,
+                       int llvm_cond,
+                       struct lp_build_loop_state *state)
+{
+   LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
+   LLVMValueRef function = LLVMGetBasicBlockParent(block);
+   LLVMValueRef next;
+   LLVMValueRef cond;
+   LLVMBasicBlockRef after_block;
+
+   if (!step)
+      step = LLVMConstInt(LLVMTypeOf(end), 1, 0);
+
+   next = LLVMBuildAdd(builder, state->counter, step, "");
+
+   cond = LLVMBuildICmp(builder, llvm_cond, next, end, "");
+
+   after_block = LLVMAppendBasicBlock(function, "");
+
+   LLVMBuildCondBr(builder, cond, after_block, state->block);
+
+   LLVMAddIncoming(state->counter, &next, &block, 1);
+
+   LLVMPositionBuilderAtEnd(builder, after_block);
+}
+
 
 
 /*
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_flow.h b/src/gallium/auxiliary/gallivm/lp_bld_flow.h
index c2b50e1b602..745838570c8 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_flow.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_flow.h
@@ -124,6 +124,13 @@ lp_build_loop_end(LLVMBuilderRef builder,
                   LLVMValueRef step,
                   struct lp_build_loop_state *state);
 
+void
+lp_build_loop_end_cond(LLVMBuilderRef builder,
+                       LLVMValueRef end,
+                       LLVMValueRef step,
+                       int cond, /* LLVM condition */
+                       struct lp_build_loop_state *state);
+
 
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_context.c b/src/gallium/drivers/llvmpipe/lp_context.c
index 951a695f964..5c476f693f2 100644
--- a/src/gallium/drivers/llvmpipe/lp_context.c
+++ b/src/gallium/drivers/llvmpipe/lp_context.c
@@ -40,13 +40,14 @@
 #include "lp_context.h"
 #include "lp_flush.h"
 #include "lp_perf.h"
+#include "lp_screen.h"
 #include "lp_state.h"
 #include "lp_surface.h"
 #include "lp_query.h"
 #include "lp_setup.h"
 
 
-
+#define USE_DRAW_LLVM 0
 
 
 static void llvmpipe_destroy( struct pipe_context *pipe )
@@ -105,6 +106,7 @@ struct pipe_context *
 llvmpipe_create_context( struct pipe_screen *screen, void *priv )
 {
    struct llvmpipe_context *llvmpipe;
+   struct llvmpipe_screen *llvmscreen = llvmpipe_screen(screen);
 
    llvmpipe = align_malloc(sizeof(struct llvmpipe_context), 16);
    if (!llvmpipe)
@@ -180,8 +182,12 @@ llvmpipe_create_context( struct pipe_screen *screen, void *priv )
    /*
     * Create drawing context and plug our rendering stage into it.
     */
+#if USE_DRAW_LLVM
+   llvmpipe->draw = draw_create_with_llvm(llvmscreen->engine);
+#else
    llvmpipe->draw = draw_create();
-   if (!llvmpipe->draw) 
+#endif
+   if (!llvmpipe->draw)
       goto fail;
 
    /* FIXME: devise alternative to draw_texture_samplers */