draw: add aos vertex shader varient

6 files changed, 2247 insertions, 0 deletions

diff --git a/src/gallium/auxiliary/draw/Makefile b/src/gallium/auxiliary/draw/Makefile
index 84877994fbf..9a88ecc0708 100644
--- a/src/gallium/auxiliary/draw/Makefile
+++ b/src/gallium/auxiliary/draw/Makefile
@@ -35,6 +35,8 @@ C_SOURCES = \
 	draw_vertex.c \
 	draw_vs.c \
 	draw_vs_varient.c \
+	draw_vs_aos.c \
+	draw_vs_aos_io.c \
 	draw_vs_exec.c \
 	draw_vs_llvm.c \
 	draw_vs_sse.c 
diff --git a/src/gallium/auxiliary/draw/draw_vs.h b/src/gallium/auxiliary/draw/draw_vs.h
index 6bfc2c8d752..5a8d0da06d6 100644
--- a/src/gallium/auxiliary/draw/draw_vs.h
+++ b/src/gallium/auxiliary/draw/draw_vs.h
@@ -162,6 +162,16 @@ struct draw_vertex_shader *
 draw_create_vs_llvm(struct draw_context *draw,
 		    const struct pipe_shader_state *templ);
 
+
+
+struct draw_vs_varient_key;
+struct draw_vertex_shader;
+
+struct draw_vs_varient *draw_vs_varient_aos_sse( struct draw_vertex_shader *vs,
+                                                 const struct draw_vs_varient_key *key );
+
+
+
 /********************************************************************************
  * Helpers for vs implementations that don't do their own fetch/emit varients.
  * Means these can be shared between shaders.
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c
new file mode 100644
index 00000000000..620f5e3592f
--- /dev/null
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.c
@@ -0,0 +1,1739 @@
+/*
+ * Mesa 3-D graphics library
+ * Version:  6.3
+ *
+ * Copyright (C) 1999-2004  Brian Paul   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, sublicense,
+ * 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 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 NONINFRINGEMENT.  IN NO EVENT SHALL
+ * BRIAN PAUL 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.
+ */
+
+/**
+ * Translate tgsi vertex programs to x86/x87/SSE/SSE2 machine code
+ * using the rtasm runtime assembler.  Based on the old
+ * t_vb_arb_program_sse.c
+ */
+
+
+#include "pipe/p_util.h"
+#include "pipe/p_shader_tokens.h"
+#include "tgsi/util/tgsi_parse.h"
+#include "tgsi/util/tgsi_util.h"
+#include "tgsi/exec/tgsi_exec.h"
+#include "tgsi/util/tgsi_dump.h"
+
+#include "draw_vs.h"
+#include "draw_vs_aos.h"
+
+#include "rtasm/rtasm_x86sse.h"
+
+#ifdef PIPE_ARCH_X86
+
+
+#define DISASSEM 0
+
+
+
+
+
+static INLINE boolean eq( struct x86_reg a,
+			    struct x86_reg b )
+{
+   return (a.file == b.file &&
+	   a.idx == b.idx &&
+	   a.mod == b.mod &&
+	   a.disp == b.disp);
+}
+      
+
+static struct x86_reg get_reg_ptr(struct aos_compilation *cp,
+                                  unsigned file,
+				  unsigned idx )
+{
+   struct x86_reg ptr = cp->machine_EDX;
+
+   switch (file) {
+   case TGSI_FILE_INPUT:
+      return x86_make_disp(ptr, Offset(struct aos_machine, input[idx]));
+
+   case TGSI_FILE_OUTPUT:
+      return x86_make_disp(ptr, Offset(struct aos_machine, output[idx]));
+
+   case TGSI_FILE_TEMPORARY:
+      return x86_make_disp(ptr, Offset(struct aos_machine, temp[idx]));
+
+   case TGSI_FILE_IMMEDIATE:
+      return x86_make_disp(ptr, Offset(struct aos_machine, immediate[idx]));
+
+   case TGSI_FILE_CONSTANT:       
+      return x86_make_disp(ptr, Offset(struct aos_machine, constant[idx]));
+
+   case AOS_FILE_INTERNAL:
+      return x86_make_disp(ptr, Offset(struct aos_machine, immediate[idx]));
+
+   default:
+      ERROR(cp, "unknown reg file");
+      return x86_make_reg(0,0);
+   }
+}
+		
+
+struct x86_reg aos_get_internal( struct aos_compilation *cp,
+                                 unsigned imm )
+{
+   return get_reg_ptr( cp,
+                       AOS_FILE_INTERNAL, 
+                       imm + 1 );
+}
+
+static void spill( struct aos_compilation *cp, unsigned idx )
+{
+   if (!cp->xmm[idx].dirty ||
+       (cp->xmm[idx].file != TGSI_FILE_INPUT && /* inputs are fetched into xmm & set dirty */
+        cp->xmm[idx].file != TGSI_FILE_OUTPUT &&
+        cp->xmm[idx].file != TGSI_FILE_TEMPORARY)) {
+      ERROR(cp, "invalid spill");
+      return;
+   }
+   else {
+      struct x86_reg oldval = get_reg_ptr(cp,
+                                          cp->xmm[idx].file,
+                                          cp->xmm[idx].idx);
+      
+      assert(cp->xmm[idx].dirty);
+      sse_movups(cp->func, oldval, x86_make_reg(file_XMM, idx));
+      cp->xmm[idx].dirty = 0;
+   }
+}
+
+struct x86_reg aos_get_xmm_reg( struct aos_compilation *cp )
+{
+   unsigned i;
+   unsigned oldest = 0;
+
+   for (i = 0; i < 8; i++) 
+      if (cp->xmm[i].last_used < cp->xmm[oldest].last_used)
+	 oldest = i;
+
+   /* Need to write out the old value?
+    */
+   if (cp->xmm[oldest].dirty) 
+      spill(cp, oldest);
+
+   assert(cp->xmm[oldest].last_used != cp->insn_counter);
+
+   cp->xmm[oldest].file = TGSI_FILE_NULL;
+   cp->xmm[oldest].idx = 0;
+   cp->xmm[oldest].last_used = cp->insn_counter;
+   return x86_make_reg(file_XMM, oldest);
+}
+
+void aos_release_xmm_reg( struct aos_compilation *cp,
+                          unsigned idx )
+{
+   cp->xmm[idx].file = TGSI_FILE_NULL;
+   cp->xmm[idx].idx = 0;
+   cp->xmm[idx].dirty = 0;
+   cp->xmm[idx].last_used = 0;
+}
+
+static void invalidate_xmm( struct aos_compilation *cp, 
+			    unsigned file, unsigned idx )
+{
+   unsigned i;
+
+   /* Invalidate any old copy of this register in XMM0-7.  
+    */
+   for (i = 0; i < 8; i++) {
+      if (cp->xmm[i].file == file && cp->xmm[i].idx == idx) {
+
+         if (cp->xmm[i].dirty) 
+            spill(cp, i);
+         
+         aos_release_xmm_reg(cp, i);
+         break;
+      }
+   }
+
+   for (; i < 8; i++) {
+      if (cp->xmm[i].file == file && cp->xmm[i].idx == idx) {
+         assert(0);
+      }
+   }
+}
+      
+
+void aos_adopt_xmm_reg( struct aos_compilation *cp,
+                        struct x86_reg reg,
+                        unsigned file,
+                        unsigned idx,
+                        unsigned dirty )
+{
+   if (reg.file != file_XMM) {
+      assert(0);
+      return;
+   }
+
+   invalidate_xmm(cp, file, idx);
+   cp->xmm[reg.idx].file = file;
+   cp->xmm[reg.idx].idx = idx;
+   cp->xmm[reg.idx].dirty = dirty;
+}
+
+
+
+static struct x86_reg aos_get_shader_reg_ptr( struct aos_compilation *cp, 
+                                              unsigned file,
+                                              unsigned idx )
+{
+   invalidate_xmm( cp, file, idx );
+   return get_reg_ptr( cp, file, idx );
+}
+
+
+/* As above, but return a pointer.  Note - this pointer may alias
+ * those returned by get_arg_ptr().
+ */
+static struct x86_reg get_dst_ptr( struct aos_compilation *cp, 
+                                   const struct tgsi_full_dst_register *dst )
+{
+   return aos_get_shader_reg_ptr( cp, dst->DstRegister.File, dst->DstRegister.Index );
+}
+
+
+
+
+
+/* Return an XMM reg if the argument is resident, otherwise return a
+ * base+offset pointer to the saved value.
+ */
+struct x86_reg aos_get_shader_reg( struct aos_compilation *cp, 
+                                   unsigned file,
+                                   unsigned idx )
+{
+   unsigned i;
+
+   for (i = 0; i < 8; i++) {
+      if (cp->xmm[i].file == file &&
+	  cp->xmm[i].idx  == idx) 
+      {
+	 cp->xmm[i].last_used = cp->insn_counter;
+	 return x86_make_reg(file_XMM, i);
+      }
+   }
+
+   /* If not found in the XMM register file, return an indirect
+    * reference to the in-memory copy:
+    */
+   return get_reg_ptr( cp, file, idx );
+}
+
+
+
+
+
+/* Emulate pshufd insn in regular SSE, if necessary:
+ */
+static void emit_pshufd( struct aos_compilation *cp,
+			 struct x86_reg dst,
+			 struct x86_reg arg0,
+			 ubyte shuf )
+{
+   if (cp->have_sse2) {
+      sse2_pshufd(cp->func, dst, arg0, shuf);
+   }
+   else {
+      if (!eq(dst, arg0)) 
+	 sse_movups(cp->func, dst, arg0);
+
+      sse_shufps(cp->func, dst, dst, shuf);
+   }
+}
+
+
+
+
+/* Helper for writemask:
+ */
+static boolean emit_shuf_copy1( struct aos_compilation *cp,
+				  struct x86_reg dst,
+				  struct x86_reg arg0,
+				  struct x86_reg arg1,
+				  ubyte shuf )
+{
+   struct x86_reg tmp = aos_get_xmm_reg(cp);
+   sse_movups(cp->func, dst, arg1);
+   emit_pshufd(cp, dst, dst, shuf);
+   emit_pshufd(cp, tmp, arg0, shuf);
+
+   sse_movss(cp->func, dst, tmp);
+
+   emit_pshufd(cp, dst, dst, shuf);
+
+   aos_release_xmm_reg(cp, tmp.idx);
+   return TRUE;
+}
+
+
+/* Helper for writemask:
+ */
+static boolean emit_shuf_copy2( struct aos_compilation *cp,
+				  struct x86_reg dst,
+				  struct x86_reg arg0,
+				  struct x86_reg arg1,
+				  ubyte shuf )
+{
+   struct x86_reg tmp = aos_get_xmm_reg(cp);
+   emit_pshufd(cp, dst, arg1, shuf);
+   emit_pshufd(cp, tmp, arg0, shuf);
+
+   sse_shufps(cp->func, dst, tmp, SHUF(X, Y, Z, W));
+
+   emit_pshufd(cp, dst, dst, shuf);
+
+   aos_release_xmm_reg(cp, tmp.idx);
+   return TRUE;
+}
+
+#define SSE_SWIZZLE_NOOP ((0<<0) | (1<<2) | (2<<4) | (3<<6))
+
+
+/* Locate a source register and perform any required (simple) swizzle.  
+ * 
+ * Just fail on complex swizzles at this point.
+ */
+static struct x86_reg fetch_src( struct aos_compilation *cp, 
+                                 const struct tgsi_full_src_register *src ) 
+{
+   struct x86_reg arg0 = aos_get_shader_reg(cp, 
+                                            src->SrcRegister.File, 
+                                            src->SrcRegister.Index);
+   unsigned i;
+   unsigned swz = 0;
+   unsigned negs = 0;
+   unsigned abs = 0;
+
+   for (i = 0; i < 4; i++) {
+      unsigned swizzle = tgsi_util_get_full_src_register_extswizzle( src, i );
+      unsigned neg = tgsi_util_get_full_src_register_sign_mode( src, i );
+
+      switch (swizzle) {
+      case TGSI_EXTSWIZZLE_ZERO:
+      case TGSI_EXTSWIZZLE_ONE:
+         ERROR(cp, "not supporting full swizzles yet in tgsi_aos_sse2");
+         break;
+
+      default:
+         swz |= (swizzle & 0x3) << (i * 2);
+         break;
+      }
+
+      switch (neg) {
+      case TGSI_UTIL_SIGN_TOGGLE:
+         negs |= (1<<i);
+         break;
+         
+      case TGSI_UTIL_SIGN_KEEP:
+         break;
+
+      case TGSI_UTIL_SIGN_CLEAR:
+         abs |= (1<<i);
+         break;
+
+      default:
+         ERROR(cp, "unsupported sign-mode");
+         break;
+      }
+   }
+
+   if (swz != SSE_SWIZZLE_NOOP || negs != 0 || abs != 0) {
+      struct x86_reg dst = aos_get_xmm_reg(cp);
+
+      if (swz != SSE_SWIZZLE_NOOP) {
+         emit_pshufd(cp, dst, arg0, swz);
+         arg0 = dst;
+      }
+
+      if (negs) {
+         struct x86_reg imm_negs = aos_get_internal(cp, IMM_NEGS);
+         struct x86_reg tmp = aos_get_xmm_reg(cp);
+
+         /* Load 1,-1,0,0
+          * Use neg as arg to pshufd
+          * Multiply
+          */
+         emit_pshufd(cp, tmp, imm_negs, 
+                     SHUF((negs & 1) ? 1 : 0,
+                          (negs & 2) ? 1 : 0,
+                          (negs & 4) ? 1 : 0,
+                          (negs & 8) ? 1 : 0));
+         sse_mulps(cp->func, dst, arg0);
+
+         aos_release_xmm_reg(cp, tmp.idx);
+         arg0 = dst;
+      }
+
+      if (abs && abs != 0xf) {
+         ERROR(cp, "unsupported partial abs");
+      }
+
+      if (abs) {
+         struct x86_reg neg = aos_get_internal(cp, IMM_NEGS);
+         struct x86_reg tmp = aos_get_xmm_reg(cp);
+
+         sse_movups(cp->func, tmp, arg0);
+         sse_mulps(cp->func, tmp, neg);
+         sse_maxps(cp->func, dst, arg0);
+
+         aos_release_xmm_reg(cp, tmp.idx);
+         arg0 = dst;
+      }
+   }
+      
+   return arg0;
+}
+
+static void x87_fld_src( struct aos_compilation *cp, 
+                         const struct tgsi_full_src_register *src,
+                         unsigned channel ) 
+{
+   struct x86_reg arg0 = aos_get_shader_reg_ptr(cp, 
+                                                src->SrcRegister.File, 
+                                                src->SrcRegister.Index);
+
+   unsigned swizzle = tgsi_util_get_full_src_register_extswizzle( src, channel );
+   unsigned neg = tgsi_util_get_full_src_register_sign_mode( src, channel );
+
+   switch (swizzle) {
+   case TGSI_EXTSWIZZLE_ZERO:
+      x87_fldz( cp->func );
+      break;
+
+   case TGSI_EXTSWIZZLE_ONE:
+      x87_fld1( cp->func );
+      break;
+
+   default:
+      x87_fld( cp->func, x86_make_disp(arg0, (swizzle & 3) * sizeof(float)) );
+      break;
+   }
+   
+
+   switch (neg) {
+   case TGSI_UTIL_SIGN_TOGGLE:
+      /* Flip the sign:
+       */
+      x87_fchs( cp->func );
+      break;
+         
+   case TGSI_UTIL_SIGN_KEEP:
+      break;
+
+   case TGSI_UTIL_SIGN_CLEAR:
+      x87_fabs( cp->func );
+      break;
+
+   case TGSI_UTIL_SIGN_SET:
+      x87_fabs( cp->func );
+      x87_fchs( cp->func );
+      break;
+
+   default:
+      ERROR(cp, "unsupported sign-mode");
+      break;
+   }
+}
+
+
+
+
+
+
+/* Used to implement write masking.  This and most of the other instructions
+ * here would be easier to implement if there had been a translation
+ * to a 2 argument format (dst/arg0, arg1) at the shader level before
+ * attempting to translate to x86/sse code.
+ */
+static void store_dest( struct aos_compilation *cp, 
+                        const struct tgsi_full_dst_register *reg,
+                        struct x86_reg result )
+{
+   if (reg->DstRegister.WriteMask == 0) 
+   {
+      return;
+   }
+   else if (reg->DstRegister.WriteMask == TGSI_WRITEMASK_XYZW)
+   {
+      if (result.file == file_XMM) {
+         aos_adopt_xmm_reg(cp, 
+                           result, 
+                           reg->DstRegister.File,
+                           reg->DstRegister.Index,
+                           TRUE);
+      }
+      else {
+         struct x86_reg dst = aos_get_xmm_reg(cp);
+         aos_adopt_xmm_reg(cp, 
+                           dst, 
+                           reg->DstRegister.File,
+                           reg->DstRegister.Index,
+                           TRUE);
+         sse_movups(cp->func, dst, result);
+      }
+   }
+   else
+   {
+      /* Previous value of the dest register:
+       */
+      struct x86_reg old_dst = aos_get_shader_reg(cp, 
+                                                  reg->DstRegister.File,
+                                                  reg->DstRegister.Index);
+
+
+      /* Alloc an xmm reg to hold the new value of the dest register:
+       */
+      struct x86_reg dst = aos_get_xmm_reg(cp);
+
+      aos_adopt_xmm_reg(cp, 
+                        dst, 
+                        reg->DstRegister.File,
+                        reg->DstRegister.Index,
+                        TRUE );
+
+      switch (reg->DstRegister.WriteMask) {
+      case TGSI_WRITEMASK_X:
+         if (result.file == file_XMM) {
+            sse_movups(cp->func, dst, old_dst);
+            sse_movss(cp->func, dst, result);
+         }
+         else {
+            struct x86_reg tmp = aos_get_xmm_reg(cp);
+            sse_movups(cp->func, dst, old_dst);
+            sse_movss(cp->func, tmp, result);
+            sse_movss(cp->func, dst, tmp);
+            aos_release_xmm_reg(cp, tmp.idx);
+         }
+         break;
+
+      case TGSI_WRITEMASK_XY:
+         sse_movups(cp->func, dst, old_dst);
+         sse_shufps(cp->func, dst, result, SHUF(X, Y, Z, W));
+         break;
+
+      case TGSI_WRITEMASK_ZW: 
+         sse_movups(cp->func, dst, result);
+         sse_shufps(cp->func, dst, old_dst, SHUF(X, Y, Z, W));
+         break;
+
+      case TGSI_WRITEMASK_YZW: 
+         if (old_dst.file == file_XMM) {
+            sse_movups(cp->func, dst, result);
+            sse_movss(cp->func, dst, old_dst);
+         }
+         else {
+            struct x86_reg tmp = aos_get_xmm_reg(cp);      
+            sse_movups(cp->func, dst, result);
+            sse_movss(cp->func, tmp, old_dst);
+            sse_movss(cp->func, dst, tmp);
+            aos_release_xmm_reg(cp, tmp.idx);
+         }
+         break;
+
+      case TGSI_WRITEMASK_Y:
+         emit_shuf_copy1(cp, dst, result, old_dst, SHUF(Y,X,Z,W));
+         break;
+
+      case TGSI_WRITEMASK_Z: 
+         emit_shuf_copy1(cp, dst, result, old_dst, SHUF(Z,Y,X,W));
+         break;
+
+      case TGSI_WRITEMASK_W: 
+         emit_shuf_copy1(cp, dst, result, old_dst, SHUF(W,Y,Z,X));
+         break;
+
+      case TGSI_WRITEMASK_XZ:
+         emit_shuf_copy2(cp, dst, result, old_dst, SHUF(X,Z,Y,W));
+         break;
+
+      case TGSI_WRITEMASK_XW: 
+         emit_shuf_copy2(cp, dst, result, old_dst, SHUF(X,W,Z,Y));
+
+      case TGSI_WRITEMASK_YZ:      
+         emit_shuf_copy2(cp, dst, result, old_dst, SHUF(Z,Y,X,W));
+         break;
+
+      case TGSI_WRITEMASK_YW:
+         emit_shuf_copy2(cp, dst, result, old_dst, SHUF(W,Y,Z,X));
+         break;
+
+      case TGSI_WRITEMASK_XZW:
+         emit_shuf_copy1(cp, dst, old_dst, result, SHUF(Y,X,Z,W));
+         break;
+
+      case TGSI_WRITEMASK_XYW: 
+         emit_shuf_copy1(cp, dst, old_dst, result, SHUF(Z,Y,X,W));
+         break;
+
+      case TGSI_WRITEMASK_XYZ: 
+         emit_shuf_copy1(cp, dst, old_dst, result, SHUF(W,Y,Z,X));
+         break;
+
+      default:
+         assert(0);             /* not possible */
+         break;
+      }
+   }
+}
+
+
+static void x87_fst_or_nop( struct x86_function *func,
+                            unsigned writemask,
+                            unsigned channel,
+                            struct x86_reg ptr )
+{
+   if (writemask & (1<<channel)) 
+      x87_fst( func, x86_make_disp(ptr, channel * sizeof(float)) );
+}
+
+static void x87_fstp_or_pop( struct x86_function *func,
+                             unsigned writemask,
+                             unsigned channel,
+                             struct x86_reg ptr )
+{
+   if (writemask & (1<<channel)) 
+      x87_fstp( func, x86_make_disp(ptr, channel * sizeof(float)) );
+   else
+      x87_fstp( func, x86_make_reg( file_x87, 0 ));
+}
+
+
+
+/* 
+ */
+static void x87_fstp_dest4( struct aos_compilation *cp,
+                            const struct tgsi_full_dst_register *dst )
+{
+   struct x86_reg ptr = get_dst_ptr(cp, dst); 
+   unsigned writemask = dst->DstRegister.WriteMask;
+
+   x87_fst_or_nop(cp->func, writemask, 0, ptr);
+   x87_fst_or_nop(cp->func, writemask, 1, ptr);
+   x87_fst_or_nop(cp->func, writemask, 2, ptr);
+   x87_fstp_or_pop(cp->func, writemask, 3, ptr);
+}
+
+/* Save current x87 state and put it into single precision mode.
+ */
+static void save_fpu_state( struct aos_compilation *cp )
+{
+#if 0
+   x87_fnstcw( cp->func, x86_make_disp(regEDX, get_offset(m, &m->fpucntl_restore)));
+   x87_fldcw( cp->func, );
+#endif
+}
+
+static void restore_fpu_state( struct aos_compilation *cp )
+{
+#if 0
+   x87_fnclex(cp->func);
+   x87_fldcw(cp->func, x86_make_disp(regEDX, get_offset(m, &m->fpucntl_restore)));
+#endif
+}
+
+static void set_fpu_round_neg_inf( struct aos_compilation *cp )
+{
+#if 0
+   if (cp->fpucntl != RND_NEG_FPU) {
+      struct x86_reg regEDX = x86_make_reg(file_REG32, reg_DX);
+      struct arb_vp_machine *m = NULL;
+
+      cp->fpucntl = RND_NEG_FPU;
+      x87_fnclex(cp->func);
+      x87_fldcw(cp->func, x86_make_disp(regEDX, get_offset(m, &m->fpucntl_rnd_neg)));
+   }
+#endif
+}
+
+static void set_fpu_round_nearest( struct aos_compilation *cp )
+{
+#if 0
+#endif
+}
+
+
+static void emit_x87_ex2( struct aos_compilation *cp )
+{
+   struct x86_reg st0 = x86_make_reg(file_x87, 0);
+   struct x86_reg st1 = x86_make_reg(file_x87, 1);
+   struct x86_reg st3 = x86_make_reg(file_x87, 3);
+
+   set_fpu_round_neg_inf( cp );
+
+   x87_fld(cp->func, st0); /* a a */
+   x87_fprndint( cp->func );	/* int(a) a */
+   x87_fld(cp->func, st0); /* int(a) int(a) a */
+   x87_fstp(cp->func, st3); /* int(a) a int(a)*/
+   x87_fsubp(cp->func, st1); /* frac(a) int(a) */
+   x87_f2xm1(cp->func);    /* (2^frac(a))-1 int(a)*/
+   x87_fld1(cp->func);    /* 1 (2^frac(a))-1 int(a)*/
+   x87_faddp(cp->func, st1);	/* 2^frac(a) int(a) */
+   x87_fscale(cp->func);	/* 2^a */
+}
+
+
+
+/**
+ * The traditional instructions.  All operate on internal registers
+ * and ignore write masks and swizzling issues.
+ */
+
+static boolean emit_ABS( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+   struct x86_reg neg = aos_get_internal(cp, IMM_NEGS);
+
+   sse_movups(cp->func, dst, arg0);
+   sse_mulps(cp->func, dst, neg);
+   sse_maxps(cp->func, dst, arg0);
+   
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+static boolean emit_ADD( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+
+   sse_movups(cp->func, dst, arg0);
+   sse_addps(cp->func, dst, arg1);
+
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+static boolean emit_COS( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   x87_fld_src(cp, &op->FullSrcRegisters[0], 0);
+   x87_fcos(cp->func);
+   x87_fstp_dest4(cp, &op->FullDstRegisters[0]);
+   return TRUE;
+}
+
+
+/* The dotproduct instructions don't really do that well in sse:
+ */
+static boolean emit_DP3( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+   struct x86_reg tmp = aos_get_xmm_reg(cp); 
+
+   sse_movups(cp->func, dst, arg0);
+   sse_mulps(cp->func, dst, arg1);
+   
+   /* Now the hard bit: sum the first 3 values:
+    */ 
+   sse_movhlps(cp->func, tmp, dst);
+   sse_addss(cp->func, dst, tmp); /* a*x+c*z, b*y, ?, ? */
+   emit_pshufd(cp, tmp, dst, SHUF(Y,X,W,Z));
+   sse_addss(cp->func, dst, tmp);
+   sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X));
+
+   aos_release_xmm_reg(cp, tmp.idx);
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+
+
+static boolean emit_DP4( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+   struct x86_reg tmp = aos_get_xmm_reg(cp);      
+
+   sse_movups(cp->func, dst, arg0);
+   sse_mulps(cp->func, dst, arg1);
+   
+   /* Now the hard bit: sum the values:
+    */ 
+   sse_movhlps(cp->func, tmp, dst);
+   sse_addps(cp->func, dst, tmp); /* a*x+c*z, b*y+d*w, a*x+c*z, b*y+d*w */
+   emit_pshufd(cp, tmp, dst, SHUF(Y,X,W,Z));
+   sse_addss(cp->func, dst, tmp);
+   sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X));
+
+   aos_release_xmm_reg(cp, tmp.idx);
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+static boolean emit_DPH( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+   struct x86_reg tmp = aos_get_xmm_reg(cp);
+
+   sse_movups(cp->func, dst, arg0);
+   sse_mulps(cp->func, dst, arg1);
+
+   /* Now the hard bit: sum the values (from DP3):
+    */ 
+   sse_movhlps(cp->func, tmp, dst);
+   sse_addss(cp->func, dst, tmp); /* a*x+c*z, b*y, ?, ? */
+   emit_pshufd(cp, tmp, dst, SHUF(Y,X,W,Z));
+   sse_addss(cp->func, dst, tmp);
+   emit_pshufd(cp, tmp, arg1, SHUF(W,W,W,W));
+   sse_addss(cp->func, dst, tmp);
+   sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X));
+
+   aos_release_xmm_reg(cp, tmp.idx);
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+static boolean emit_DST( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+    struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+    struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+    struct x86_reg dst = aos_get_xmm_reg(cp);
+    struct x86_reg tmp = aos_get_xmm_reg(cp);
+    struct x86_reg ones = aos_get_internal(cp, IMM_ONES);
+
+/*    dst[0] = 1.0     * 1.0F; */
+/*    dst[1] = arg0[1] * arg1[1]; */
+/*    dst[2] = arg0[2] * 1.0; */
+/*    dst[3] = 1.0     * arg1[3]; */
+
+    emit_shuf_copy2(cp, dst, arg0, ones, SHUF(X,W,Z,Y));
+    emit_shuf_copy2(cp, tmp, arg1, ones, SHUF(X,Z,Y,W));
+    sse_mulps(cp->func, dst, tmp);
+
+    aos_release_xmm_reg(cp, tmp.idx);
+    store_dest(cp, &op->FullDstRegisters[0], dst);
+    return TRUE;
+}
+
+static boolean emit_LG2( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   x87_fld1(cp->func);		/* 1 */
+   x87_fld_src(cp, &op->FullSrcRegisters[0], 0);	/* a0 1 */
+   x87_fyl2x(cp->func);	/* log2(a0) */
+   x87_fstp_dest4(cp, &op->FullDstRegisters[0]);
+   return TRUE;
+}
+
+
+static boolean emit_EX2( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   x87_fld_src(cp, &op->FullSrcRegisters[0], 0);
+
+   emit_x87_ex2(cp);
+
+   x87_fstp_dest4(cp, &op->FullDstRegisters[0]);
+   return TRUE;
+}
+
+static boolean emit_EXP( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+    struct x86_reg dst = get_dst_ptr(cp, &op->FullDstRegisters[0]); 
+    struct x86_reg st0 = x86_make_reg(file_x87, 0);
+    struct x86_reg st1 = x86_make_reg(file_x87, 1);
+    struct x86_reg st3 = x86_make_reg(file_x87, 3);
+    unsigned writemask = op->FullDstRegisters[0].DstRegister.WriteMask;
+
+    /* CAUTION: dst may alias arg0!
+     */
+    x87_fld_src(cp, &op->FullSrcRegisters[0], 0);	/* arg0.x */
+    x87_fld(cp->func, st0); /* arg arg */
+
+    /* by default, fpu is setup to round-to-nearest.  We want to
+     * change this now, and track the state through to the end of the
+     * generated function so that it isn't repeated unnecessarily.
+     * Alternately, could subtract .5 to get round to -inf behaviour.
+     */
+    set_fpu_round_neg_inf( cp );
+    x87_fprndint( cp->func );	/* flr(a) a */
+    x87_fld(cp->func, st0); /* flr(a) flr(a) a */
+    x87_fld1(cp->func);    /* 1 floor(a) floor(a) a */
+    x87_fst_or_nop(cp->func, writemask, 3, dst);  /* stack unchanged */
+
+    x87_fscale(cp->func);  /* 2^floor(a) floor(a) a */
+    x87_fst(cp->func, st3); /* 2^floor(a) floor(a) a 2^floor(a)*/
+
+    x87_fstp_or_pop(cp->func, writemask, 0, dst); /* flr(a) a 2^flr(a) */
+
+    x87_fsubrp(cp->func, st1); /* frac(a) 2^flr(a) */
+
+    x87_fst_or_nop(cp->func, writemask, 1, dst);    /* frac(a) 2^flr(a) */
+
+    x87_f2xm1(cp->func);    /* (2^frac(a))-1 2^flr(a)*/
+    x87_fld1(cp->func);    /* 1 (2^frac(a))-1 2^flr(a)*/
+    x87_faddp(cp->func, st1);	/* 2^frac(a) 2^flr(a) */
+    x87_fmulp(cp->func, st1);	/* 2^a */
+    
+    x87_fstp_or_pop(cp->func, writemask, 2, dst);    
+
+/*    dst[0] = 2^floor(tmp); */
+/*    dst[1] = frac(tmp); */
+/*    dst[2] = 2^floor(tmp) * 2^frac(tmp); */
+/*    dst[3] = 1.0F; */
+    return TRUE;
+}
+
+static boolean emit_LOG( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+    struct x86_reg dst = get_dst_ptr(cp, &op->FullDstRegisters[0]); 
+    struct x86_reg st0 = x86_make_reg(file_x87, 0);
+    struct x86_reg st1 = x86_make_reg(file_x87, 1);
+    struct x86_reg st2 = x86_make_reg(file_x87, 2);
+    unsigned writemask = op->FullDstRegisters[0].DstRegister.WriteMask;
+ 
+    /* CAUTION: dst may alias arg0!
+     */
+    x87_fld_src(cp, &op->FullSrcRegisters[0], 0);	/* arg0.x */
+    x87_fabs(cp->func);	/* |arg0.x| */
+    x87_fxtract(cp->func);	/* mantissa(arg0.x), exponent(arg0.x) */
+    x87_fst(cp->func, st2);	/* mantissa, exponent, mantissa */
+    x87_fld1(cp->func);	/* 1, mantissa, exponent, mantissa */
+    x87_fyl2x(cp->func); 	/* log2(mantissa), exponent, mantissa */
+    x87_fadd(cp->func, st0, st1);	/* e+l2(m), e, m  */
+    
+    x87_fstp_or_pop(cp->func, writemask, 2, dst); /* e, m */
+
+    x87_fld1(cp->func);	/* 1, e, m */
+    x87_fsub(cp->func, st1, st0);	/* 1, e-1, m */
+
+    x87_fstp_or_pop(cp->func, writemask, 3, dst); /* e-1,m */
+    x87_fstp_or_pop(cp->func, writemask, 0, dst);	/* m */
+
+    x87_fadd(cp->func, st0, st0);	/* 2m */
+
+    x87_fstp_or_pop( cp->func, writemask, 1, dst );
+
+    return TRUE;
+}
+
+static boolean emit_FLR( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   struct x86_reg dst = get_dst_ptr(cp, &op->FullDstRegisters[0]); 
+   unsigned writemask = op->FullDstRegisters[0].DstRegister.WriteMask;
+   int i;
+
+   set_fpu_round_neg_inf( cp );
+
+   /* Load all sources first to avoid aliasing
+    */
+   for (i = 0; i < 4; i++) {
+      if (writemask & (1<<i)) {
+         x87_fld_src(cp, &op->FullSrcRegisters[0], i);
+      }
+   }
+
+   for (i = 0; i < 4; i++) {
+      if (writemask & (1<<i)) {
+         x87_fprndint( cp->func );   
+         x87_fstp(cp->func, x86_make_disp(dst, i*4));
+      }
+   }
+
+   return TRUE;
+}
+
+
+static boolean emit_RND( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   struct x86_reg dst = get_dst_ptr(cp, &op->FullDstRegisters[0]); 
+   unsigned writemask = op->FullDstRegisters[0].DstRegister.WriteMask;
+   int i;
+
+   set_fpu_round_nearest( cp );
+
+   /* Load all sources first to avoid aliasing
+    */
+   for (i = 0; i < 4; i++) {
+      if (writemask & (1<<i)) {
+         x87_fld_src(cp, &op->FullSrcRegisters[0], i);
+      }
+   }
+
+   for (i = 0; i < 4; i++) {
+      if (writemask & (1<<i)) {
+         x87_fprndint( cp->func );   
+         x87_fstp(cp->func, x86_make_disp(dst, i*4));
+      }
+   }
+
+   return TRUE;
+}
+
+
+static boolean emit_FRC( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   struct x86_reg dst = get_dst_ptr(cp, &op->FullDstRegisters[0]); 
+   struct x86_reg st0 = x86_make_reg(file_x87, 0);
+   struct x86_reg st1 = x86_make_reg(file_x87, 1);
+   unsigned writemask = op->FullDstRegisters[0].DstRegister.WriteMask;
+   int i;
+
+   set_fpu_round_neg_inf( cp );
+
+   /* suck all the source values onto the stack before writing out any
+    * dst, which may alias...
+    */
+   for (i = 0; i < 4; i++) {
+      if (writemask & (1<<i)) {
+         x87_fld_src(cp, &op->FullSrcRegisters[0], i);   
+      }
+   }
+
+   for (i = 0; i < 4; i++) {
+      if (writemask & (1<<i)) {
+         x87_fld(cp->func, st0);     /* a a */
+         x87_fprndint( cp->func );   /* flr(a) a */
+         x87_fsubrp(cp->func, st1);  /* frc(a) */
+         x87_fstp(cp->func, x86_make_disp(dst, i*4));
+      }
+   }
+
+   return TRUE;
+}
+
+
+
+static boolean emit_LIT( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg dst = get_dst_ptr(cp, &op->FullDstRegisters[0]); 
+   struct x86_reg st1 = x86_make_reg(file_x87, 1);
+   unsigned fixup1, fixup2;
+   unsigned writemask = op->FullDstRegisters[0].DstRegister.WriteMask;
+
+
+   /* Load the interesting parts of arg0:
+    */
+   x87_fld_src(cp, &op->FullSrcRegisters[0], 3);
+   x87_fld_src(cp, &op->FullSrcRegisters[0], 1);
+   x87_fld_src(cp, &op->FullSrcRegisters[0], 0);
+   
+
+   if (writemask & TGSI_WRITEMASK_XW) {
+      x87_fld1(cp->func);
+      x87_fst_or_nop(cp->func, writemask, 0, dst);
+      x87_fstp_or_pop(cp->func, writemask, 3, dst);
+   }
+
+   if (writemask & TGSI_WRITEMASK_YZ) {
+      
+      /* Pre-zero destinations, may be overwritten later...  fixme. 
+       */
+      x87_fldz(cp->func);
+      x87_fst_or_nop(cp->func, writemask, 1, dst);
+      x87_fstp_or_pop(cp->func, writemask, 2, dst);
+
+
+      /* Check arg0[0]:
+       */
+      x87_fldz(cp->func);		/* 0 a0 a1 a3 */
+      x87_fucomp(cp->func, st1);	/* a0 a1 a3 */
+      x87_fnstsw(cp->func, cp->tmp_EAX);
+      x86_sahf(cp->func);
+      fixup1 = x86_jcc_forward(cp->func, cc_AE); 
+   
+      x87_fstp_or_pop(cp->func, writemask, 1, dst);	/* a1 a3 */
+
+      /* Check arg0[1]:
+       */ 
+      x87_fldz(cp->func);		/* 0 a1 a3 */
+      x87_fucomp(cp->func, st1);	/* a1 a3 */
+      x87_fnstsw(cp->func, cp->tmp_EAX);
+      x86_sahf(cp->func);
+      fixup2 = x86_jcc_forward(cp->func, cc_AE); 
+
+      /* Compute pow(a1, a3)
+       */
+      x87_fyl2x(cp->func);	/* a3*log2(a1) */
+
+      emit_x87_ex2( cp );		/* 2^(a3*log2(a1)) */
+
+      x87_fstp_or_pop(cp->func, writemask, 2, dst);
+   
+      /* Land jumps:
+       */
+      x86_fixup_fwd_jump(cp->func, fixup1);
+      x86_fixup_fwd_jump(cp->func, fixup2);
+   }
+
+   return TRUE;
+}
+
+
+
+static boolean emit_MAX( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+
+   sse_movups(cp->func, dst, arg0);
+   sse_maxps(cp->func, dst, arg1);
+
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+
+static boolean emit_MIN( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+
+   sse_movups(cp->func, dst, arg0);
+   sse_minps(cp->func, dst, arg1);
+
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+static boolean emit_MOV( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+
+   sse_movups(cp->func, dst, arg0);
+
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+static boolean emit_MUL( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+
+   sse_movups(cp->func, dst, arg0);
+   sse_mulps(cp->func, dst, arg1);
+
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+
+static boolean emit_MAD( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg arg2 = fetch_src(cp, &op->FullSrcRegisters[2]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+
+   sse_movups(cp->func, dst, arg0);
+   sse_mulps(cp->func, dst, arg1);
+   sse_addps(cp->func, dst, arg2);
+
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+
+static boolean emit_POW( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   x87_fld_src(cp, &op->FullSrcRegisters[1], 0);  /* a1.x */
+   x87_fld_src(cp, &op->FullSrcRegisters[0], 0);	/* a0.x a1.x */
+   x87_fyl2x(cp->func);	                                /* a1*log2(a0) */
+
+   emit_x87_ex2( cp );		/* 2^(a1*log2(a0)) */
+
+   x87_fstp_dest4(cp, &op->FullDstRegisters[0]);
+   return TRUE;
+}
+
+
+static boolean emit_RCP( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+
+   if (cp->have_sse2) {
+      sse2_rcpss(cp->func, dst, arg0);
+      /* extend precision here...
+       */
+   }
+   else {
+      struct x86_reg ones = aos_get_internal(cp, IMM_ONES);
+      sse_movss(cp->func, dst, ones);
+      sse_divss(cp->func, dst, arg0);
+   }
+
+   sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X));
+
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+static boolean emit_RSQ( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+
+   sse_rsqrtss(cp->func, dst, arg0);
+
+   /* Extend precision here...
+    */
+
+   sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X));
+
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+
+static boolean emit_SGE( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+   struct x86_reg ones = aos_get_internal(cp, IMM_ONES);
+
+   sse_movups(cp->func, dst, arg0);
+   sse_cmpps(cp->func, dst, arg1, cc_NotLessThan);
+   sse_andps(cp->func, dst, ones);
+
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+static boolean emit_SIN( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   x87_fld_src(cp, &op->FullSrcRegisters[0], 0);
+   x87_fsin(cp->func);
+   x87_fstp_dest4(cp, &op->FullDstRegisters[0]);
+   return TRUE;
+}
+
+
+
+static boolean emit_SLT( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+   struct x86_reg ones = aos_get_internal(cp, IMM_ONES);
+   
+   sse_movups(cp->func, dst, arg0);
+   sse_cmpps(cp->func, dst, arg1, cc_LessThan);
+   sse_andps(cp->func, dst, ones);
+
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+static boolean emit_SUB( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+
+   sse_movups(cp->func, dst, arg0);
+   sse_subps(cp->func, dst, arg1);
+
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+
+static boolean emit_XPD( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg arg1 = fetch_src(cp, &op->FullSrcRegisters[1]);
+   struct x86_reg dst = aos_get_xmm_reg(cp);
+   struct x86_reg tmp0 = aos_get_xmm_reg(cp);
+   struct x86_reg tmp1 = aos_get_xmm_reg(cp);
+
+   /* Could avoid tmp0, tmp1 if we overwrote arg0, arg1.  Need a way
+    * to invalidate registers.  This will come with better analysis
+    * (liveness analysis) of the incoming program.
+    */
+   emit_pshufd(cp, dst, arg0, SHUF(Y, Z, X, W));
+   emit_pshufd(cp, tmp1, arg1, SHUF(Z, X, Y, W));
+   sse_mulps(cp->func, dst, tmp1);
+   emit_pshufd(cp, tmp0, arg0, SHUF(Z, X, Y, W));
+   emit_pshufd(cp, tmp1, arg1, SHUF(Y, Z, X, W));
+   sse_mulps(cp->func, tmp0, tmp1);
+   sse_subps(cp->func, dst, tmp0);
+
+/*    dst[0] = arg0[1] * arg1[2] - arg0[2] * arg1[1]; */
+/*    dst[1] = arg0[2] * arg1[0] - arg0[0] * arg1[2]; */
+/*    dst[2] = arg0[0] * arg1[1] - arg0[1] * arg1[0]; */
+/*    dst[3] is undef */
+
+
+   aos_release_xmm_reg(cp, tmp0.idx);
+   aos_release_xmm_reg(cp, tmp1.idx);
+   store_dest(cp, &op->FullDstRegisters[0], dst);
+   return TRUE;
+}
+
+
+
+static boolean
+emit_instruction( struct aos_compilation *cp,
+                  struct tgsi_full_instruction *inst )
+{
+   switch( inst->Instruction.Opcode ) {
+   case TGSI_OPCODE_MOV:
+      return emit_MOV( cp, inst );
+
+   case TGSI_OPCODE_LIT:
+      return emit_LIT(cp, inst);
+
+   case TGSI_OPCODE_RCP:
+      return emit_RCP(cp, inst);
+
+   case TGSI_OPCODE_RSQ:
+      return emit_RSQ(cp, inst);
+
+   case TGSI_OPCODE_EXP:
+      return emit_EXP(cp, inst);
+
+   case TGSI_OPCODE_LOG:
+      return emit_LOG(cp, inst);
+
+   case TGSI_OPCODE_MUL:
+      return emit_MUL(cp, inst);
+
+   case TGSI_OPCODE_ADD:
+      return emit_ADD(cp, inst);
+
+   case TGSI_OPCODE_DP3:
+      return emit_DP3(cp, inst);
+
+   case TGSI_OPCODE_DP4:
+      return emit_DP4(cp, inst);
+
+   case TGSI_OPCODE_DST:
+      return emit_DST(cp, inst);
+
+   case TGSI_OPCODE_MIN:
+      return emit_MIN(cp, inst);
+
+   case TGSI_OPCODE_MAX:
+      return emit_MAX(cp, inst);
+
+   case TGSI_OPCODE_SLT:
+      return emit_SLT(cp, inst);
+
+   case TGSI_OPCODE_SGE:
+      return emit_SGE(cp, inst);
+
+   case TGSI_OPCODE_MAD:
+      return emit_MAD(cp, inst);
+
+   case TGSI_OPCODE_SUB:
+      return emit_SUB(cp, inst);
+ 
+   case TGSI_OPCODE_LERP:
+//      return emit_LERP(cp, inst);
+      return FALSE;
+
+   case TGSI_OPCODE_FRAC:
+      return emit_FRC(cp, inst);
+
+   case TGSI_OPCODE_CLAMP:
+//      return emit_CLAMP(cp, inst);
+      return FALSE;
+
+   case TGSI_OPCODE_FLOOR:
+      return emit_FLR(cp, inst);
+
+   case TGSI_OPCODE_ROUND:
+      return emit_RND(cp, inst);
+
+   case TGSI_OPCODE_EXPBASE2:
+      return emit_EX2(cp, inst);
+
+   case TGSI_OPCODE_LOGBASE2:
+      return emit_LG2(cp, inst);
+
+   case TGSI_OPCODE_POWER:
+      return emit_POW(cp, inst);
+
+   case TGSI_OPCODE_CROSSPRODUCT:
+      return emit_XPD(cp, inst);
+
+   case TGSI_OPCODE_ABS:
+      return emit_ABS(cp, inst);
+
+   case TGSI_OPCODE_DPH:
+      return emit_DPH(cp, inst);
+
+   case TGSI_OPCODE_COS:
+      return emit_COS(cp, inst);
+
+   case TGSI_OPCODE_SIN:
+      return emit_SIN(cp, inst);
+
+   case TGSI_OPCODE_END:
+      return TRUE;
+
+   default:
+      return FALSE;
+   }
+}
+
+static boolean note_immediate( struct aos_compilation *cp,
+                               struct tgsi_full_immediate *imm )
+{
+   unsigned pos = cp->num_immediates++;
+   unsigned j;
+
+   for (j = 0; j < imm->Immediate.Size; j++) {
+      cp->vaos->machine->immediate[pos][j] = imm->u.ImmediateFloat32[j].Float;
+   }
+
+   return TRUE;
+}
+
+
+
+
+static void find_last_write_outputs( struct aos_compilation *cp )
+{
+   struct tgsi_parse_context parse;
+   unsigned this_instruction = 0;
+   unsigned i;
+
+   tgsi_parse_init( &parse, cp->vaos->base.vs->state.tokens );
+
+   while (!tgsi_parse_end_of_tokens( &parse )) {
+      
+      tgsi_parse_token( &parse );
+
+      if (parse.FullToken.Token.Type != TGSI_TOKEN_TYPE_INSTRUCTION) 
+         continue;
+
+      for (i = 0; i < TGSI_FULL_MAX_DST_REGISTERS; i++) {
+         if (parse.FullToken.FullInstruction.FullDstRegisters[i].DstRegister.File ==
+             TGSI_FILE_OUTPUT) 
+         {
+            unsigned idx = parse.FullToken.FullInstruction.FullDstRegisters[i].DstRegister.Index;
+            cp->output_last_write[idx] = this_instruction;
+         }
+      }
+
+      this_instruction++;
+   }
+
+   tgsi_parse_free( &parse );
+}
+
+
+#define ARG_VARIENT    1
+#define ARG_START_ELTS 2
+#define ARG_COUNT      3
+#define ARG_OUTBUF     4
+
+
+static boolean build_vertex_program( struct draw_vs_varient_aos_sse *varient,
+                                     boolean linear )
+{ 
+   struct tgsi_parse_context parse;
+   struct aos_compilation cp;
+   unsigned fixup, label;
+
+   tgsi_parse_init( &parse, varient->base.vs->state.tokens );
+
+   memset(&cp, 0, sizeof(cp));
+
+   cp.insn_counter = 1;
+   cp.vaos = varient;
+   cp.have_sse2 = 1;
+   cp.func = &varient->func[ linear ? 0 : 1 ];
+
+   cp.tmp_EAX       = x86_make_reg(file_REG32, reg_AX);
+   cp.idx_EBX      = x86_make_reg(file_REG32, reg_BX);
+   cp.outbuf_ECX    = x86_make_reg(file_REG32, reg_CX);
+   cp.machine_EDX   = x86_make_reg(file_REG32, reg_DX);
+   cp.count_ESI     = x86_make_reg(file_REG32, reg_SI);
+
+   x86_init_func(cp.func);
+
+   find_last_write_outputs(&cp);
+
+   x86_push(cp.func, cp.idx_EBX);
+   x86_push(cp.func, cp.count_ESI);
+
+
+   /* Load arguments into regs:
+    */
+   x86_mov(cp.func, cp.machine_EDX, x86_fn_arg(cp.func, ARG_VARIENT));
+   x86_mov(cp.func, cp.idx_EBX, x86_fn_arg(cp.func, ARG_START_ELTS));
+   x86_mov(cp.func, cp.count_ESI, x86_fn_arg(cp.func, ARG_COUNT));
+   x86_mov(cp.func, cp.outbuf_ECX, x86_fn_arg(cp.func, ARG_OUTBUF));
+
+
+   /* Compare count to zero and possibly bail.
+    */
+   x86_xor(cp.func, cp.tmp_EAX, cp.tmp_EAX);
+   x86_cmp(cp.func, cp.count_ESI, cp.tmp_EAX);
+   fixup = x86_jcc_forward(cp.func, cc_E);
+
+   /* Dig out the machine pointer from inside the varient arg 
+    */
+   x86_mov(cp.func, cp.machine_EDX, 
+           x86_make_disp(cp.machine_EDX,
+                         Offset( struct draw_vs_varient_aos_sse, machine )));
+
+   save_fpu_state( &cp );
+
+   /* Note address for loop jump 
+    */
+   label = x86_get_label(cp.func);
+   {
+      /* Fetch inputs...  TODO:  fetch lazily...
+       */
+      if (!aos_fetch_inputs( &cp, linear ))
+         goto fail;
+
+      /* Emit the shader:
+       */
+      while( !tgsi_parse_end_of_tokens( &parse ) && !cp.error ) 
+      {
+         tgsi_parse_token( &parse );
+
+         switch (parse.FullToken.Token.Type) {
+         case TGSI_TOKEN_TYPE_IMMEDIATE:
+            if (!note_immediate( &cp, &parse.FullToken.FullImmediate ))
+               goto fail;
+            break;
+
+         case TGSI_TOKEN_TYPE_INSTRUCTION:
+            if (!emit_instruction( &cp, &parse.FullToken.FullInstruction ))
+               goto fail;
+            break;
+         }
+
+         cp.insn_counter++;
+         debug_printf("\n");
+      }
+
+      if (cp.error)
+         goto fail;
+
+      /* Emit output...  TODO: do this eagerly after the last write to a
+       * given output.
+       */
+      if (!aos_emit_outputs( &cp ))
+         goto fail;
+
+
+      /* Next vertex:
+       */
+      x86_lea(cp.func, 
+              cp.outbuf_ECX, 
+              x86_make_disp(cp.outbuf_ECX, 
+                            cp.vaos->base.key.output_stride));
+
+      /* Incr index
+       */   
+      if (linear) {
+         x86_inc(cp.func, cp.idx_EBX);
+      } 
+      else {
+         x86_lea(cp.func, cp.idx_EBX, x86_make_disp(cp.idx_EBX, 4));
+      }
+
+   }
+   /* decr count, loop if not zero
+    */
+   x86_dec(cp.func, cp.count_ESI);
+/*   x86_test(cp.func, cp.count_ESI, cp.count_ESI);  */
+   x86_jcc(cp.func, cc_NZ, label);
+
+   restore_fpu_state(&cp);
+
+   /* Land forward jump here:
+    */
+   x86_fixup_fwd_jump(cp.func, fixup);
+
+   /* Exit mmx state?
+    */
+   if (cp.func->need_emms)
+      mmx_emms(cp.func);
+
+   x86_pop(cp.func, cp.count_ESI);
+   x86_pop(cp.func, cp.idx_EBX);
+
+   x86_ret(cp.func);
+
+   tgsi_parse_free( &parse );
+   return !cp.error;
+
+ fail:
+   tgsi_parse_free( &parse );
+   return FALSE;
+}
+
+
+
+static void vaos_set_buffer( struct draw_vs_varient *varient,
+                             unsigned buf,
+                             const void *ptr,
+                             unsigned stride )
+{
+   struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient;
+   unsigned i;
+
+   for (i = 0; i < vaos->base.vs->info.num_inputs; i++) {
+      if (vaos->base.key.element[i].in.buffer == buf) {
+         vaos->machine->attrib[i].input_ptr = ((char *)ptr +
+                                               vaos->base.key.element[i].in.offset);
+         vaos->machine->attrib[i].input_stride = stride;
+      }
+   }
+}
+
+
+static void vaos_destroy( struct draw_vs_varient *varient )
+{
+   struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient;
+
+   if (vaos->machine)
+      align_free( vaos->machine );
+
+   x86_release_func( &vaos->func[0] );
+   x86_release_func( &vaos->func[1] );
+
+   FREE(vaos);
+}
+
+static void vaos_run_elts( struct draw_vs_varient *varient,
+                           const unsigned *elts,
+                           unsigned count,
+                           void *output_buffer )
+{
+   struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient;
+
+   vaos->gen_run_elts( varient,
+                       elts,
+                       count,
+                       output_buffer );
+}
+
+static void vaos_run_linear( struct draw_vs_varient *varient,
+                             unsigned start,
+                             unsigned count,
+                             void *output_buffer )
+{
+   struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient;
+
+   vaos->gen_run_linear( varient,
+                         start,
+                         count,
+                         output_buffer );
+}
+
+
+static void vaos_set_constants( struct draw_vs_varient *varient,
+                                const float (*constants)[4] )
+{
+   struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient;
+
+   memcpy(vaos->machine->constant,
+          constants,
+          (vaos->base.vs->info.file_max[TGSI_FILE_CONSTANT] + 1) * 4 * sizeof(float));
+}
+
+
+static struct draw_vs_varient *varient_aos_sse( struct draw_vertex_shader *vs,
+                                                 const struct draw_vs_varient_key *key )
+{
+   struct draw_vs_varient_aos_sse *vaos = CALLOC_STRUCT(draw_vs_varient_aos_sse);
+
+   if (!vaos)
+      goto fail;
+   
+   vaos->base.key = *key;
+   vaos->base.vs = vs;
+   vaos->base.set_input = vaos_set_buffer;
+   vaos->base.set_constants = vaos_set_constants;
+   vaos->base.destroy = vaos_destroy;
+   vaos->base.run_linear = vaos_run_linear;
+   vaos->base.run_elts = vaos_run_elts;
+
+   vaos->machine = align_malloc( sizeof(struct aos_machine), 16 );
+   if (!vaos->machine)
+      goto fail;
+   
+   memset(vaos->machine, 0, sizeof(struct aos_machine));
+
+   tgsi_dump(vs->state.tokens, 0);
+
+   if (!build_vertex_program( vaos, TRUE ))
+      goto fail;
+
+   if (!build_vertex_program( vaos, FALSE ))
+      goto fail;
+
+   vaos->gen_run_linear = (vsv_run_linear_func)x86_get_func(&vaos->func[0]);
+   if (!vaos->gen_run_linear)
+      goto fail;
+
+   vaos->gen_run_elts = (vsv_run_elts_func)x86_get_func(&vaos->func[1]);
+   if (!vaos->gen_run_elts)
+      goto fail;
+
+   return &vaos->base;
+
+ fail:
+   if (vaos->machine)
+      align_free( vaos->machine );
+
+   if (vaos)
+      x86_release_func( &vaos->func[0] );
+
+   if (vaos)
+      x86_release_func( &vaos->func[1] );
+
+   FREE(vaos);
+   
+   return NULL;
+}
+
+
+struct draw_vs_varient *draw_vs_varient_aos_sse( struct draw_vertex_shader *vs,
+                                                 const struct draw_vs_varient_key *key )
+{
+   struct draw_vs_varient *varient = varient_aos_sse( vs, key );
+
+   if (varient == NULL) {
+      assert(0);
+      varient = draw_vs_varient_generic( vs, key );
+   }
+
+   return varient;
+}
+
+
+
+#endif
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.h b/src/gallium/auxiliary/draw/draw_vs_aos.h
new file mode 100644
index 00000000000..1d8a055a904
--- /dev/null
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.h
@@ -0,0 +1,181 @@
+/**************************************************************************
+ * 
+ * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * 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.
+ * 
+ **************************************************************************/
+
+/* Authors:  Keith Whitwell <[email protected]>
+ */
+
+#ifndef DRAW_VS_AOS_H
+#define DRAW_VS_AOS_H
+
+
+struct tgsi_token;
+struct x86_function;
+
+#include "pipe/p_state.h"
+#include "rtasm/rtasm_x86sse.h"
+
+
+
+
+
+#define X    0
+#define Y    1
+#define Z    2
+#define W    3
+
+#define MAX_INPUTS     PIPE_MAX_ATTRIBS
+#define MAX_OUTPUTS    PIPE_MAX_ATTRIBS
+#define MAX_TEMPS      PIPE_MAX_ATTRIBS /* say */
+#define MAX_CONSTANTS  PIPE_MAX_ATTRIBS /* say */
+#define MAX_IMMEDIATES PIPE_MAX_ATTRIBS /* say */
+#define MAX_INTERNALS  4
+
+#define AOS_FILE_INTERNAL TGSI_FILE_COUNT
+
+/* This is the temporary storage used by all the aos_sse vs varients.
+ * Create one per context and reuse by passing a pointer in at
+ * vs_varient creation??
+ */
+struct aos_machine {
+   float input    [MAX_INPUTS    ][4];
+   float output   [MAX_OUTPUTS   ][4];
+   float temp     [MAX_TEMPS     ][4];
+   float constant [MAX_CONSTANTS ][4]; /* fixme -- should just be a pointer */
+   float immediate[MAX_IMMEDIATES][4]; /* fixme -- should just be a pointer */
+   float internal [MAX_INTERNALS ][4];
+
+   unsigned fpu_round_nearest;
+   unsigned fpu_round_neg_inf;
+
+   struct {
+      const void *input_ptr;
+      unsigned input_stride;
+
+      unsigned output_offset;
+   } attrib[PIPE_MAX_ATTRIBS];
+};
+
+
+
+
+struct aos_compilation {
+   struct x86_function *func;
+   struct draw_vs_varient_aos_sse *vaos;
+
+   unsigned insn_counter;
+   unsigned num_immediates;
+
+   struct {
+      unsigned idx:16;
+      unsigned file:8;
+      unsigned dirty:8;
+      unsigned last_used;
+   } xmm[8];
+
+
+   boolean input_fetched[PIPE_MAX_ATTRIBS];
+   unsigned output_last_write[PIPE_MAX_ATTRIBS];
+
+   boolean have_sse2;
+   boolean error;
+   short fpucntl;
+
+   /* these are actually known values, but putting them in a struct
+    * like this is helpful to keep them in sync across the file.
+    */
+   struct x86_reg tmp_EAX;
+   struct x86_reg idx_EBX;     /* either start+i or &elt[i] */
+   struct x86_reg outbuf_ECX;
+   struct x86_reg machine_EDX;
+   struct x86_reg count_ESI;    /* decrements to zero */
+};
+
+struct x86_reg aos_get_xmm_reg( struct aos_compilation *cp );
+void aos_release_xmm_reg( struct aos_compilation *cp, unsigned idx );
+
+void aos_adopt_xmm_reg( struct aos_compilation *cp,
+                        struct x86_reg reg,
+                        unsigned file,
+                        unsigned idx,
+                        unsigned dirty );
+
+struct x86_reg aos_get_shader_reg( struct aos_compilation *cp, 
+                                   unsigned file,
+                                   unsigned idx );
+
+boolean aos_fetch_inputs( struct aos_compilation *cp,
+                          boolean linear );
+
+boolean aos_emit_outputs( struct aos_compilation *cp );
+
+
+#define IMM_ONES     0              /* 1, 1,1,1 */
+#define IMM_NEGS     1              /* 1,-1,0,0 */
+#define IMM_IDENTITY 2              /* 0, 0,0,1 */
+#define IMM_INV_255  3              /* 1/255, 1/255, 1/255, 1/255 */
+#define IMM_255      4              /* 255, 255, 255, 255 */
+
+struct x86_reg aos_get_internal( struct aos_compilation *cp,
+                                 unsigned imm );
+
+
+#define ERROR(cp, msg)                                                  \
+do {                                                                    \
+   debug_printf("%s: x86 translation failed: %s\n", __FUNCTION__, msg); \
+   cp->error = 1;                                                       \
+   assert(0);                                                           \
+} while (0)
+
+
+
+
+
+
+struct draw_vs_varient_aos_sse {
+   struct draw_vs_varient base;
+   struct draw_context *draw;
+
+#if 0
+   struct {
+      const void *ptr;
+      unsigned stride;
+   } attrib[PIPE_MAX_ATTRIBS];
+#endif
+
+   struct aos_machine *machine; /* XXX: temporarily unshared */
+
+   vsv_run_linear_func gen_run_linear;
+   vsv_run_elts_func gen_run_elts;
+
+
+   struct x86_function func[2];
+};
+
+
+
+#endif 
+
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos_io.c b/src/gallium/auxiliary/draw/draw_vs_aos_io.c
new file mode 100644
index 00000000000..72b2b3d11d2
--- /dev/null
+++ b/src/gallium/auxiliary/draw/draw_vs_aos_io.c
@@ -0,0 +1,314 @@
+/**************************************************************************
+ * 
+ * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * 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 "pipe/p_util.h"
+#include "pipe/p_shader_tokens.h"
+#include "tgsi/util/tgsi_parse.h"
+#include "tgsi/util/tgsi_util.h"
+#include "tgsi/exec/tgsi_exec.h"
+#include "draw_vs.h"
+#include "draw_vs_aos.h"
+
+#include "rtasm/rtasm_x86sse.h"
+
+#ifdef PIPE_ARCH_X86
+
+/* Note - don't yet have to worry about interacting with the code in
+ * draw_vs_aos.c as there is no intermingling of generated code...
+ * That may have to change, we'll see.
+ */
+static void emit_load_R32G32B32A32( struct aos_compilation *cp, 			   
+				    struct x86_reg data,
+				    struct x86_reg src_ptr )
+{
+   sse_movups(cp->func, data, src_ptr);
+}
+
+static void emit_load_R32G32B32( struct aos_compilation *cp, 			   
+				 struct x86_reg data,
+				 struct x86_reg src_ptr )
+{
+   sse_movss(cp->func, data, x86_make_disp(src_ptr, 8));
+   sse_shufps(cp->func, data, aos_get_internal( cp, IMM_IDENTITY ), SHUF(X,Y,Z,W) );
+   sse_shufps(cp->func, data, data, SHUF(Y,Z,X,W) );
+   sse_movlps(cp->func, data, src_ptr);
+}
+
+static void emit_load_R32G32( struct aos_compilation *cp, 
+			   struct x86_reg data,
+			   struct x86_reg src_ptr )
+{
+   sse_movups(cp->func, data, aos_get_internal( cp, IMM_IDENTITY ) );
+   sse_movlps(cp->func, data, src_ptr);
+}
+
+
+static void emit_load_R32( struct aos_compilation *cp, 
+			   struct x86_reg data,
+			   struct x86_reg src_ptr )
+{
+   sse_movss(cp->func, data, src_ptr);
+   sse_orps(cp->func, data, aos_get_internal( cp, IMM_IDENTITY ) );
+}
+
+
+static void emit_load_R8G8B8A8_UNORM( struct aos_compilation *cp,
+				       struct x86_reg data,
+				       struct x86_reg src_ptr )
+{
+   sse_movss(cp->func, data, src_ptr);
+   sse2_punpcklbw(cp->func, data, aos_get_internal( cp, IMM_IDENTITY ));
+   sse2_punpcklbw(cp->func, data, aos_get_internal( cp, IMM_IDENTITY ));
+   sse2_cvtdq2ps(cp->func, data, data);
+   sse_mulps(cp->func, data, aos_get_internal(cp, IMM_INV_255));
+}
+
+
+
+static void get_src_ptr( struct x86_function *func,
+                         struct x86_reg src,
+                         struct x86_reg machine,
+                         struct x86_reg elt,
+                         unsigned a )
+{
+   struct x86_reg input_ptr = 
+      x86_make_disp(machine, 
+		    Offset(struct aos_machine, attrib[a].input_ptr));
+
+   struct x86_reg input_stride = 
+      x86_make_disp(machine, 
+		    Offset(struct aos_machine, attrib[a].input_stride));
+
+   /* Calculate pointer to current attrib:
+    */
+   x86_mov(func, src, input_stride);
+   x86_imul(func, src, elt);
+   x86_add(func, src, input_ptr);
+}
+
+
+/* Extended swizzles?  Maybe later.
+ */  
+static void emit_swizzle( struct aos_compilation *cp,
+			  struct x86_reg dest,
+			  struct x86_reg src,
+			  unsigned shuffle )
+{
+   sse_shufps(cp->func, dest, src, shuffle);
+}
+
+
+static boolean load_input( struct aos_compilation *cp,
+                           unsigned idx,
+                           boolean linear )
+{
+   unsigned format = cp->vaos->base.key.element[idx].in.format;
+   struct x86_reg src = cp->tmp_EAX;
+   struct x86_reg dataXMM = aos_get_xmm_reg(cp);
+
+   /* Figure out source pointer address:
+    */
+   get_src_ptr(cp->func, 
+               src, 
+               cp->machine_EDX, 
+               linear ? cp->idx_EBX : x86_deref(cp->idx_EBX),
+               idx);
+
+   src = x86_deref(src);
+
+   aos_adopt_xmm_reg( cp,
+                      dataXMM,
+                      TGSI_FILE_INPUT,
+                      idx,
+                      TRUE );
+
+   switch (format) {
+   case PIPE_FORMAT_R32_FLOAT:
+      emit_load_R32(cp, dataXMM, src);
+      break;
+   case PIPE_FORMAT_R32G32_FLOAT:
+      emit_load_R32G32(cp, dataXMM, src);
+      break;
+   case PIPE_FORMAT_R32G32B32_FLOAT:
+      emit_load_R32G32B32(cp, dataXMM, src);
+      break;
+   case PIPE_FORMAT_R32G32B32A32_FLOAT:
+      emit_load_R32G32B32A32(cp, dataXMM, src);
+      break;
+   case PIPE_FORMAT_B8G8R8A8_UNORM:
+      emit_load_R8G8B8A8_UNORM(cp, dataXMM, src);
+      emit_swizzle(cp, dataXMM, dataXMM, SHUF(Z,Y,X,W));
+      break;
+   case PIPE_FORMAT_R8G8B8A8_UNORM:
+      emit_load_R8G8B8A8_UNORM(cp, dataXMM, src);
+      break;
+   default:
+      ERROR(cp, "unhandled input format");
+      return FALSE;
+   }
+
+   return TRUE;
+}
+
+
+boolean aos_fetch_inputs( struct aos_compilation *cp, boolean linear )
+{
+   unsigned i;
+   
+   for (i = 0; i < cp->vaos->base.vs->info.num_inputs; i++) {
+      if (!load_input( cp, i, linear ))
+         return FALSE;
+      cp->insn_counter++;
+      debug_printf("\n");
+   }
+
+   return TRUE;
+}
+
+
+
+
+
+
+
+static void emit_store_R32G32B32A32( struct aos_compilation *cp, 			   
+				     struct x86_reg dst_ptr,
+				     struct x86_reg dataXMM )
+{
+   sse_movups(cp->func, dst_ptr, dataXMM);
+}
+
+static void emit_store_R32G32B32( struct aos_compilation *cp, 
+				  struct x86_reg dst_ptr,
+				  struct x86_reg dataXMM )
+{
+   sse_movlps(cp->func, dst_ptr, dataXMM);
+   sse_shufps(cp->func, dataXMM, dataXMM, SHUF(Z,Z,Z,Z) ); /* NOTE! destructive */
+   sse_movss(cp->func, x86_make_disp(dst_ptr,8), dataXMM);
+}
+
+static void emit_store_R32G32( struct aos_compilation *cp, 
+			       struct x86_reg dst_ptr,
+			       struct x86_reg dataXMM )
+{
+   sse_movlps(cp->func, dst_ptr, dataXMM);
+}
+
+static void emit_store_R32( struct aos_compilation *cp, 
+			    struct x86_reg dst_ptr,
+			    struct x86_reg dataXMM )
+{
+   sse_movss(cp->func, dst_ptr, dataXMM);
+}
+
+
+
+static void emit_store_R8G8B8A8_UNORM( struct aos_compilation *cp,
+				       struct x86_reg dst_ptr,
+				       struct x86_reg dataXMM )
+{
+   sse_mulps(cp->func, dataXMM, aos_get_internal(cp, IMM_255));
+   sse2_cvtps2dq(cp->func, dataXMM, dataXMM);
+   sse2_packssdw(cp->func, dataXMM, dataXMM);
+   sse2_packuswb(cp->func, dataXMM, dataXMM);
+   sse_movss(cp->func, dst_ptr, dataXMM);
+}
+
+
+
+
+
+static boolean emit_output( struct aos_compilation *cp,
+                            struct x86_reg ptr,
+                            struct x86_reg dataXMM, 
+                            unsigned format )
+{
+   switch (format) {
+   case PIPE_FORMAT_R32_FLOAT:
+      emit_store_R32(cp, ptr, dataXMM);
+      break;
+   case PIPE_FORMAT_R32G32_FLOAT:
+      emit_store_R32G32(cp, ptr, dataXMM);
+      break;
+   case PIPE_FORMAT_R32G32B32_FLOAT:
+      emit_store_R32G32B32(cp, ptr, dataXMM);
+      break;
+   case PIPE_FORMAT_R32G32B32A32_FLOAT:
+      emit_store_R32G32B32A32(cp, ptr, dataXMM);
+      break;
+   case PIPE_FORMAT_B8G8R8A8_UNORM:
+      emit_swizzle(cp, dataXMM, dataXMM, SHUF(Z,Y,X,W));
+      emit_store_R8G8B8A8_UNORM(cp, ptr, dataXMM);
+      break;
+   case PIPE_FORMAT_R8G8B8A8_UNORM:
+      emit_store_R8G8B8A8_UNORM(cp, ptr, dataXMM);
+      break;
+   default:
+      ERROR(cp, "unhandled output format");
+      return FALSE;
+   }
+
+   return TRUE;
+}
+
+
+
+boolean aos_emit_outputs( struct aos_compilation *cp )
+{
+   unsigned i;
+   
+   for (i = 0; i < cp->vaos->base.vs->info.num_inputs; i++) {
+      unsigned format = cp->vaos->base.key.element[i].out.format;
+      unsigned offset = cp->vaos->base.key.element[i].out.offset;
+
+      struct x86_reg data = aos_get_shader_reg( cp, 
+                                                TGSI_FILE_OUTPUT,
+                                                i );
+
+      if (data.file != file_XMM) {
+         struct x86_reg tmp = aos_get_xmm_reg( cp );
+         sse_movups(cp->func, tmp, data);
+         data = tmp;
+      }
+      
+      if (!emit_output( cp, 
+                        x86_make_disp( cp->outbuf_ECX, offset ),
+                        data, 
+                        format ))
+         return FALSE;
+
+      aos_release_xmm_reg( cp, data.idx );
+
+      cp->insn_counter++;
+      debug_printf("\n");
+   }
+
+   return TRUE;
+}
+
+#endif
diff --git a/src/gallium/auxiliary/draw/draw_vs_sse.c b/src/gallium/auxiliary/draw/draw_vs_sse.c
index df94a7e0c7a..0581c3042f7 100644
--- a/src/gallium/auxiliary/draw/draw_vs_sse.c
+++ b/src/gallium/auxiliary/draw/draw_vs_sse.c
@@ -157,6 +157,7 @@ draw_create_vs_sse(struct draw_context *draw,
 
    vs->base.draw = draw;
    vs->base.create_varient = draw_vs_varient_generic;
+//   vs->base.create_varient = draw_vs_varient_aos_sse;
    vs->base.prepare = vs_sse_prepare;
    vs->base.run_linear = vs_sse_run_linear;
    vs->base.delete = vs_sse_delete;