From ffe58739da9eee2e99682747cc8f26e412c87430 Mon Sep 17 00:00:00 2001
From: michal <michal@michal-laptop.(none)>
Date: Thu, 5 Jul 2007 19:55:38 +0200
Subject: [PATCH] Softpipe: import TGSI tree. Not hooked-up yet.

---
 src/mesa/pipe/tgsi/Makefile            |    3 +
 src/mesa/pipe/tgsi/core/Makefile       |    3 +
 src/mesa/pipe/tgsi/core/tgsi_build.c   | 1315 ++++++++++++++
 src/mesa/pipe/tgsi/core/tgsi_build.h   |  309 ++++
 src/mesa/pipe/tgsi/core/tgsi_core.h    |   12 +
 src/mesa/pipe/tgsi/core/tgsi_dump.c    |  871 ++++++++++
 src/mesa/pipe/tgsi/core/tgsi_dump.h    |   22 +
 src/mesa/pipe/tgsi/core/tgsi_exec.c    | 2193 ++++++++++++++++++++++++
 src/mesa/pipe/tgsi/core/tgsi_exec.h    |  137 ++
 src/mesa/pipe/tgsi/core/tgsi_parse.c   |  284 +++
 src/mesa/pipe/tgsi/core/tgsi_parse.h   |  120 ++
 src/mesa/pipe/tgsi/core/tgsi_token.h   | 1045 +++++++++++
 src/mesa/pipe/tgsi/core/tgsi_util.c    |  263 +++
 src/mesa/pipe/tgsi/core/tgsi_util.h    |   70 +
 src/mesa/pipe/tgsi/mesa/Makefile       |    3 +
 src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.c |  499 ++++++
 src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.h |   25 +
 src/mesa/pipe/tgsi/mesa/tgsi_mesa.h    |    8 +
 src/mesa/pipe/tgsi/tgsi_platform.h     |   18 +
 src/mesa/sources                       |   19 +-
 src/mesa/state_tracker/st_cb_program.c |   14 +-
 src/mesa/state_tracker/st_texobj.c     |    3 +-
 22 files changed, 7229 insertions(+), 7 deletions(-)
 create mode 100644 src/mesa/pipe/tgsi/Makefile
 create mode 100644 src/mesa/pipe/tgsi/core/Makefile
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_build.c
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_build.h
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_core.h
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_dump.c
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_dump.h
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_exec.c
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_exec.h
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_parse.c
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_parse.h
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_token.h
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_util.c
 create mode 100644 src/mesa/pipe/tgsi/core/tgsi_util.h
 create mode 100644 src/mesa/pipe/tgsi/mesa/Makefile
 create mode 100644 src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.c
 create mode 100644 src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.h
 create mode 100644 src/mesa/pipe/tgsi/mesa/tgsi_mesa.h
 create mode 100644 src/mesa/pipe/tgsi/tgsi_platform.h

diff --git a/src/mesa/pipe/tgsi/Makefile b/src/mesa/pipe/tgsi/Makefile
new file mode 100644
index 00000000000..12a8bd0409e
--- /dev/null
+++ b/src/mesa/pipe/tgsi/Makefile
@@ -0,0 +1,3 @@
+default:
+	cd ../.. ; make
+
diff --git a/src/mesa/pipe/tgsi/core/Makefile b/src/mesa/pipe/tgsi/core/Makefile
new file mode 100644
index 00000000000..eb8b14e0e89
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/Makefile
@@ -0,0 +1,3 @@
+default:
+	cd ../../.. ; make
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_build.c b/src/mesa/pipe/tgsi/core/tgsi_build.c
new file mode 100644
index 00000000000..2a482a78dc3
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_build.c
@@ -0,0 +1,1315 @@
+#include "tgsi_platform.h"
+#include "tgsi_core.h"
+
+/*
+ * version
+ */
+
+struct tgsi_version
+tgsi_build_version( void )
+{
+   struct tgsi_version  version;
+
+   version.MajorVersion = 1;
+   version.MinorVersion = 1;
+   version.Padding = 0;
+
+   return version;
+}
+
+/*
+ * header
+ */
+
+struct tgsi_header
+tgsi_build_header( void )
+{
+   struct tgsi_header header;
+
+   header.HeaderSize = 1;
+   header.BodySize = 0;
+
+   return header;
+}
+
+static void
+header_headersize_grow( struct tgsi_header *header )
+{
+   assert (header->HeaderSize < 0xFF);
+   assert (header->BodySize == 0);
+
+   header->HeaderSize++;
+}
+
+static void
+header_bodysize_grow( struct tgsi_header *header )
+{
+   assert (header->BodySize < 0xFFFFFF);
+
+   header->BodySize++;
+}
+
+struct tgsi_processor
+tgsi_default_processor( void )
+{
+   struct tgsi_processor processor;
+
+   processor.Processor = TGSI_PROCESSOR_FRAGMENT;
+   processor.Padding = 0;
+
+   return processor;
+}
+
+struct tgsi_processor
+tgsi_build_processor(
+   GLuint type,
+   struct tgsi_header *header )
+{
+   struct tgsi_processor processor;
+
+   processor = tgsi_default_processor();
+   processor.Processor = type;
+
+   header_headersize_grow( header );
+
+   return processor;
+}
+
+/*
+ * declaration
+ */
+
+struct tgsi_declaration
+tgsi_default_declaration( void )
+{
+   struct tgsi_declaration declaration;
+
+   declaration.Type = TGSI_TOKEN_TYPE_DECLARATION;
+   declaration.Size = 1;
+   declaration.File = TGSI_FILE_NULL;
+   declaration.Declare = TGSI_DECLARE_RANGE;
+   declaration.Interpolate = 0;
+   declaration.Padding = 0;
+   declaration.Extended = 0;
+
+   return declaration;
+}
+
+struct tgsi_declaration
+tgsi_build_declaration(
+   GLuint file,
+   GLuint declare,
+   GLuint interpolate,
+   struct tgsi_header *header )
+{
+   struct tgsi_declaration declaration;
+
+   assert (file <= TGSI_FILE_IMMEDIATE);
+   assert (declare <= TGSI_DECLARE_MASK);
+
+   declaration = tgsi_default_declaration();
+   declaration.File = file;
+   declaration.Declare = declare;
+   declaration.Interpolate = interpolate;
+
+   header_bodysize_grow( header );
+
+   return declaration;
+}
+
+static void
+declaration_grow(
+   struct tgsi_declaration *declaration,
+   struct tgsi_header *header )
+{
+   assert (declaration->Size < 0xFF);
+
+   declaration->Size++;
+
+   header_bodysize_grow( header );
+}
+
+struct tgsi_full_declaration
+tgsi_default_full_declaration( void )
+{
+   struct tgsi_full_declaration  full_declaration;
+
+   full_declaration.Declaration  = tgsi_default_declaration();
+   full_declaration.Interpolation = tgsi_default_declaration_interpolation();
+
+   return full_declaration;
+}
+
+GLuint
+tgsi_build_full_declaration(
+   const struct   tgsi_full_declaration *full_decl,
+   struct tgsi_token *tokens,
+   struct tgsi_header *header,
+   GLuint maxsize )
+{
+   GLuint size = 0;
+   struct tgsi_declaration *declaration;
+
+   if( maxsize <= size )
+     return 0;
+   declaration = (struct tgsi_declaration *) &tokens[size];
+   size++;
+
+   *declaration = tgsi_build_declaration(
+      full_decl->Declaration.File,
+      full_decl->Declaration.Declare,
+      full_decl->Declaration.Interpolate,
+      header );
+
+   switch( full_decl->Declaration.Declare )  {
+   case  TGSI_DECLARE_RANGE:
+   {
+      struct tgsi_declaration_range *dr;
+
+      if( maxsize <=   size )
+         return 0;
+      dr = (struct tgsi_declaration_range *) &tokens[size];
+      size++;
+
+      *dr = tgsi_build_declaration_range(
+         full_decl->u.DeclarationRange.First,
+         full_decl->u.DeclarationRange.Last,
+         declaration,
+         header );
+      break;
+    }
+
+   case  TGSI_DECLARE_MASK:
+   {
+      struct  tgsi_declaration_mask *dm;
+
+      if( maxsize <=   size )
+         return 0;
+      dm = (struct tgsi_declaration_mask  *) &tokens[size];
+      size++;
+
+      *dm = tgsi_build_declaration_mask(
+         full_decl->u.DeclarationMask.Mask,
+         declaration,
+         header );
+      break;
+   }
+
+   default:
+      assert( 0 );
+   }
+
+   if( full_decl->Declaration.Interpolate ) {
+      struct tgsi_declaration_interpolation *di;
+
+      if( maxsize <= size )
+         return  0;
+      di = (struct tgsi_declaration_interpolation *) &tokens[size];
+      size++;
+
+      *di = tgsi_build_declaration_interpolation(
+         full_decl->Interpolation.Interpolate,
+         declaration,
+         header );
+   }
+
+   return size;
+}
+
+struct tgsi_declaration_range
+tgsi_build_declaration_range(
+   GLuint first,
+   GLuint last,
+   struct tgsi_declaration *declaration,
+   struct tgsi_header *header )
+{
+   struct tgsi_declaration_range declaration_range;
+
+   assert (last >=   first);
+   assert (last <=   0xFFFF);
+
+   declaration_range.First = first;
+   declaration_range.Last = last;
+
+   declaration_grow( declaration, header );
+
+   return declaration_range;
+}
+
+struct tgsi_declaration_mask
+tgsi_build_declaration_mask(
+   GLuint mask,
+   struct tgsi_declaration *declaration,
+   struct tgsi_header *header )
+{
+   struct tgsi_declaration_mask  declaration_mask;
+
+   declaration_mask.Mask = mask;
+
+   declaration_grow( declaration, header );
+
+   return declaration_mask;
+}
+
+struct tgsi_declaration_interpolation
+tgsi_default_declaration_interpolation( void )
+{
+   struct tgsi_declaration_interpolation di;
+
+   di.Interpolate = TGSI_INTERPOLATE_CONSTANT;
+   di.Padding = 0;
+
+   return di;
+}
+
+struct tgsi_declaration_interpolation
+tgsi_build_declaration_interpolation(
+   GLuint interpolate,
+   struct tgsi_declaration *declaration,
+   struct tgsi_header *header )
+{
+   struct tgsi_declaration_interpolation di;
+
+   assert( interpolate <= TGSI_INTERPOLATE_PERSPECTIVE );
+
+   di = tgsi_default_declaration_interpolation();
+   di.Interpolate = interpolate;
+
+   declaration_grow( declaration, header );
+
+   return di;
+}
+
+/*
+ * immediate
+ */
+
+struct tgsi_immediate
+tgsi_default_immediate( void )
+{
+   struct tgsi_immediate immediate;
+
+   immediate.Type = TGSI_TOKEN_TYPE_IMMEDIATE;
+   immediate.Size = 1;
+   immediate.DataType = TGSI_IMM_FLOAT32;
+   immediate.Padding = 0;
+   immediate.Extended = 0;
+
+   return immediate;
+}
+
+struct tgsi_immediate
+tgsi_build_immediate(
+   struct tgsi_header *header )
+{
+   struct tgsi_immediate immediate;
+
+   immediate = tgsi_default_immediate();
+
+   header_bodysize_grow( header  );
+
+   return immediate;
+}
+
+struct tgsi_full_immediate
+tgsi_default_full_immediate( void )
+{
+   struct tgsi_full_immediate fullimm;
+
+   fullimm.Immediate = tgsi_default_immediate();
+   fullimm.u.Pointer = (void *) 0;
+
+   return fullimm;
+}
+
+static void
+immediate_grow(
+   struct tgsi_immediate *immediate,
+   struct tgsi_header *header )
+{
+   assert( immediate->Size < 0xFF );
+
+   immediate->Size++;
+
+   header_bodysize_grow( header );
+}
+
+struct tgsi_immediate_float32
+tgsi_build_immediate_float32(
+   GLfloat value,
+   struct tgsi_immediate *immediate,
+   struct tgsi_header *header )
+{
+   struct tgsi_immediate_float32 immediate_float32;
+
+   immediate_float32.Float = value;
+
+   immediate_grow( immediate, header );
+
+   return immediate_float32;
+}
+
+GLuint
+tgsi_build_full_immediate(
+   const struct tgsi_full_immediate *full_imm,
+   struct tgsi_token *tokens,
+   struct tgsi_header *header,
+   GLuint maxsize )
+{
+   GLuint size = 0,  i;
+   struct tgsi_immediate *immediate;
+
+   if( maxsize <= size )
+      return 0;
+   immediate = (struct tgsi_immediate *) &tokens[size];
+   size++;
+
+   *immediate = tgsi_build_immediate( header );
+
+   for( i = 0; i < full_imm->Immediate.Size - 1; i++ ) {
+      struct tgsi_immediate_float32 *if32;
+
+      if( maxsize <= size )
+         return  0;
+      if32 = (struct tgsi_immediate_float32 *) &tokens[size];
+      size++;
+
+      *if32  = tgsi_build_immediate_float32(
+         full_imm->u.ImmediateFloat32[i].Float,
+         immediate,
+         header );
+   }
+
+   return size;
+}
+
+/*
+ * instruction
+ */
+
+struct tgsi_instruction
+tgsi_default_instruction( void )
+{
+   struct tgsi_instruction instruction;
+
+   instruction.Type = TGSI_TOKEN_TYPE_INSTRUCTION;
+   instruction.Size = 1;
+   instruction.Opcode = TGSI_OPCODE_MOV;
+   instruction.Saturate = TGSI_SAT_NONE;
+   instruction.NumDstRegs = 1;
+   instruction.NumSrcRegs = 1;
+   instruction.Padding  = 0;
+   instruction.Extended = 0;
+
+   return instruction;
+}
+
+struct tgsi_instruction
+tgsi_build_instruction(
+   GLuint opcode,
+   GLuint saturate,
+   GLuint num_dst_regs,
+   GLuint num_src_regs,
+   struct tgsi_header *header )
+{
+   struct tgsi_instruction instruction;
+
+   assert (opcode <= TGSI_OPCODE_LAST);
+   assert (saturate <= TGSI_SAT_MINUS_PLUS_ONE);
+   assert (num_dst_regs <= 3);
+   assert (num_src_regs <= 15);
+
+   instruction = tgsi_default_instruction();
+   instruction.Opcode = opcode;
+   instruction.Saturate = saturate;
+   instruction.NumDstRegs = num_dst_regs;
+   instruction.NumSrcRegs = num_src_regs;
+
+   header_bodysize_grow( header );
+
+   return instruction;
+}
+
+static void
+instruction_grow(
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header )
+{
+   assert (instruction->Size <   0xFF);
+
+   instruction->Size++;
+
+   header_bodysize_grow( header );
+}
+
+struct tgsi_full_instruction
+tgsi_default_full_instruction( void )
+{
+   struct tgsi_full_instruction full_instruction;
+   GLuint i;
+
+   full_instruction.Instruction = tgsi_default_instruction();
+   full_instruction.InstructionExtNv = tgsi_default_instruction_ext_nv();
+   full_instruction.InstructionExtLabel = tgsi_default_instruction_ext_label();
+   full_instruction.InstructionExtTexture = tgsi_default_instruction_ext_texture();
+   for( i = 0;  i < TGSI_FULL_MAX_DST_REGISTERS; i++ ) {
+      full_instruction.FullDstRegisters[i] = tgsi_default_full_dst_register();
+   }
+   for( i = 0;  i < TGSI_FULL_MAX_SRC_REGISTERS; i++ ) {
+      full_instruction.FullSrcRegisters[i] = tgsi_default_full_src_register();
+   }
+
+   return full_instruction;
+}
+
+GLuint
+tgsi_build_full_instruction(
+   const struct tgsi_full_instruction *full_inst,
+   struct  tgsi_token *tokens,
+   struct  tgsi_header *header,
+   GLuint  maxsize )
+{
+   GLuint size = 0;
+   GLuint i;
+   struct tgsi_instruction *instruction;
+   struct tgsi_token *prev_token;
+
+   if( maxsize <= size )
+      return 0;
+   instruction = (struct tgsi_instruction *) &tokens[size];
+   size++;
+
+   *instruction = tgsi_build_instruction(
+      full_inst->Instruction.Opcode,
+      full_inst->Instruction.Saturate,
+      full_inst->Instruction.NumDstRegs,
+      full_inst->Instruction.NumSrcRegs,
+      header );
+   prev_token = (struct tgsi_token  *) instruction;
+
+   if( tgsi_compare_instruction_ext_nv(
+         full_inst->InstructionExtNv,
+         tgsi_default_instruction_ext_nv() ) ) {
+      struct tgsi_instruction_ext_nv *instruction_ext_nv;
+
+      if( maxsize <= size )
+         return 0;
+      instruction_ext_nv =
+         (struct  tgsi_instruction_ext_nv *) &tokens[size];
+      size++;
+
+      *instruction_ext_nv  = tgsi_build_instruction_ext_nv(
+         full_inst->InstructionExtNv.Precision,
+         full_inst->InstructionExtNv.CondDstIndex,
+         full_inst->InstructionExtNv.CondFlowIndex,
+         full_inst->InstructionExtNv.CondMask,
+         full_inst->InstructionExtNv.CondSwizzleX,
+         full_inst->InstructionExtNv.CondSwizzleY,
+         full_inst->InstructionExtNv.CondSwizzleZ,
+         full_inst->InstructionExtNv.CondSwizzleW,
+         full_inst->InstructionExtNv.CondDstUpdate,
+         full_inst->InstructionExtNv.CondFlowEnable,
+         prev_token,
+         instruction,
+         header );
+      prev_token = (struct tgsi_token  *) instruction_ext_nv;
+   }
+
+   if( tgsi_compare_instruction_ext_label(
+         full_inst->InstructionExtLabel,
+         tgsi_default_instruction_ext_label() ) ) {
+      struct tgsi_instruction_ext_label *instruction_ext_label;
+
+      if( maxsize <= size )
+         return 0;
+      instruction_ext_label =
+         (struct  tgsi_instruction_ext_label *) &tokens[size];
+      size++;
+
+      *instruction_ext_label = tgsi_build_instruction_ext_label(
+         full_inst->InstructionExtLabel.Label,
+         full_inst->InstructionExtLabel.Target,
+         prev_token,
+         instruction,
+         header );
+      prev_token = (struct tgsi_token  *) instruction_ext_label;
+   }
+
+   if( tgsi_compare_instruction_ext_texture(
+         full_inst->InstructionExtTexture,
+         tgsi_default_instruction_ext_texture() ) ) {
+      struct tgsi_instruction_ext_texture *instruction_ext_texture;
+
+      if( maxsize <= size )
+         return 0;
+      instruction_ext_texture =
+         (struct  tgsi_instruction_ext_texture *) &tokens[size];
+      size++;
+
+      *instruction_ext_texture = tgsi_build_instruction_ext_texture(
+         full_inst->InstructionExtTexture.Texture,
+         prev_token,
+         instruction,
+         header   );
+      prev_token = (struct tgsi_token  *) instruction_ext_texture;
+   }
+
+   for( i = 0;  i <   full_inst->Instruction.NumDstRegs; i++ ) {
+      const struct tgsi_full_dst_register *reg = &full_inst->FullDstRegisters[i];
+      struct tgsi_dst_register *dst_register;
+      struct tgsi_token *prev_token;
+
+      if( maxsize <= size )
+         return 0;
+      dst_register = (struct tgsi_dst_register *) &tokens[size];
+      size++;
+
+      *dst_register = tgsi_build_dst_register(
+         reg->DstRegister.File,
+         reg->DstRegister.WriteMask,
+         reg->DstRegister.Index,
+         instruction,
+         header );
+      prev_token = (struct tgsi_token  *) dst_register;
+
+      if( tgsi_compare_dst_register_ext_concode(
+            reg->DstRegisterExtConcode,
+            tgsi_default_dst_register_ext_concode() ) ) {
+         struct tgsi_dst_register_ext_concode *dst_register_ext_concode;
+
+         if( maxsize <= size )
+            return 0;
+         dst_register_ext_concode =
+            (struct  tgsi_dst_register_ext_concode *) &tokens[size];
+         size++;
+
+         *dst_register_ext_concode =   tgsi_build_dst_register_ext_concode(
+            reg->DstRegisterExtConcode.CondMask,
+            reg->DstRegisterExtConcode.CondSwizzleX,
+            reg->DstRegisterExtConcode.CondSwizzleY,
+            reg->DstRegisterExtConcode.CondSwizzleZ,
+            reg->DstRegisterExtConcode.CondSwizzleW,
+            reg->DstRegisterExtConcode.CondSrcIndex,
+            prev_token,
+            instruction,
+            header );
+         prev_token = (struct tgsi_token  *) dst_register_ext_concode;
+      }
+
+      if( tgsi_compare_dst_register_ext_modulate(
+            reg->DstRegisterExtModulate,
+            tgsi_default_dst_register_ext_modulate() ) ) {
+         struct tgsi_dst_register_ext_modulate *dst_register_ext_modulate;
+
+         if( maxsize <= size )
+            return 0;
+         dst_register_ext_modulate =
+            (struct  tgsi_dst_register_ext_modulate *) &tokens[size];
+         size++;
+
+         *dst_register_ext_modulate = tgsi_build_dst_register_ext_modulate(
+            reg->DstRegisterExtModulate.Modulate,
+            prev_token,
+            instruction,
+            header );
+         prev_token = (struct tgsi_token  *) dst_register_ext_modulate;
+      }
+   }
+
+   for( i = 0;  i < full_inst->Instruction.NumSrcRegs; i++ ) {
+      const struct tgsi_full_src_register *reg = &full_inst->FullSrcRegisters[i];
+      struct tgsi_src_register *src_register;
+      struct tgsi_token *prev_token;
+
+      if( maxsize <= size )
+         return 0;
+      src_register = (struct tgsi_src_register *)  &tokens[size];
+      size++;
+
+      *src_register = tgsi_build_src_register(
+         reg->SrcRegister.File,
+         reg->SrcRegister.SwizzleX,
+         reg->SrcRegister.SwizzleY,
+         reg->SrcRegister.SwizzleZ,
+         reg->SrcRegister.SwizzleW,
+         reg->SrcRegister.Negate,
+         reg->SrcRegister.Indirect,
+         reg->SrcRegister.Dimension,
+         reg->SrcRegister.Index,
+         instruction,
+         header );
+      prev_token = (struct tgsi_token  *) src_register;
+
+      if( tgsi_compare_src_register_ext_swz(
+            reg->SrcRegisterExtSwz,
+            tgsi_default_src_register_ext_swz() ) ) {
+         struct tgsi_src_register_ext_swz *src_register_ext_swz;
+
+         if( maxsize <= size )
+            return 0;
+         src_register_ext_swz =
+            (struct  tgsi_src_register_ext_swz *) &tokens[size];
+         size++;
+
+         *src_register_ext_swz = tgsi_build_src_register_ext_swz(
+            reg->SrcRegisterExtSwz.ExtSwizzleX,
+            reg->SrcRegisterExtSwz.ExtSwizzleY,
+            reg->SrcRegisterExtSwz.ExtSwizzleZ,
+            reg->SrcRegisterExtSwz.ExtSwizzleW,
+            reg->SrcRegisterExtSwz.NegateX,
+            reg->SrcRegisterExtSwz.NegateY,
+            reg->SrcRegisterExtSwz.NegateZ,
+            reg->SrcRegisterExtSwz.NegateW,
+            reg->SrcRegisterExtSwz.ExtDivide,
+            prev_token,
+            instruction,
+            header );
+         prev_token = (struct tgsi_token  *) src_register_ext_swz;
+      }
+
+      if( tgsi_compare_src_register_ext_mod(
+            reg->SrcRegisterExtMod,
+            tgsi_default_src_register_ext_mod() ) ) {
+         struct tgsi_src_register_ext_mod *src_register_ext_mod;
+
+         if( maxsize <= size )
+            return 0;
+         src_register_ext_mod =
+            (struct  tgsi_src_register_ext_mod *) &tokens[size];
+         size++;
+
+         *src_register_ext_mod = tgsi_build_src_register_ext_mod(
+            reg->SrcRegisterExtMod.Complement,
+            reg->SrcRegisterExtMod.Bias,
+            reg->SrcRegisterExtMod.Scale2X,
+            reg->SrcRegisterExtMod.Absolute,
+            reg->SrcRegisterExtMod.Negate,
+            prev_token,
+            instruction,
+            header );
+         prev_token = (struct tgsi_token  *) src_register_ext_mod;
+      }
+
+      if( reg->SrcRegister.Indirect ) {
+         struct  tgsi_src_register *ind;
+
+         if( maxsize <= size )
+            return 0;
+         ind = (struct tgsi_src_register *) &tokens[size];
+         size++;
+
+         *ind = tgsi_build_src_register(
+            reg->SrcRegisterInd.File,
+            reg->SrcRegisterInd.SwizzleX,
+            reg->SrcRegisterInd.SwizzleY,
+            reg->SrcRegisterInd.SwizzleZ,
+            reg->SrcRegisterInd.SwizzleW,
+            reg->SrcRegisterInd.Negate,
+            reg->SrcRegisterInd.Indirect,
+            reg->SrcRegisterInd.Dimension,
+            reg->SrcRegisterInd.Index,
+            instruction,
+            header );
+      }
+
+      if( reg->SrcRegister.Dimension ) {
+         struct  tgsi_dimension *dim;
+
+         assert( !reg->SrcRegisterDim.Dimension );
+
+         if( maxsize <= size )
+            return 0;
+         dim = (struct tgsi_dimension *) &tokens[size];
+         size++;
+
+         *dim = tgsi_build_dimension(
+            reg->SrcRegisterDim.Indirect,
+            reg->SrcRegisterDim.Index,
+            instruction,
+            header );
+
+         if( reg->SrcRegisterDim.Indirect ) {
+            struct tgsi_src_register *ind;
+
+            if( maxsize <= size )
+               return 0;
+            ind = (struct tgsi_src_register *) &tokens[size];
+            size++;
+
+            *ind = tgsi_build_src_register(
+               reg->SrcRegisterDimInd.File,
+               reg->SrcRegisterDimInd.SwizzleX,
+               reg->SrcRegisterDimInd.SwizzleY,
+               reg->SrcRegisterDimInd.SwizzleZ,
+               reg->SrcRegisterDimInd.SwizzleW,
+               reg->SrcRegisterDimInd.Negate,
+               reg->SrcRegisterDimInd.Indirect,
+               reg->SrcRegisterDimInd.Dimension,
+               reg->SrcRegisterDimInd.Index,
+               instruction,
+               header );
+         }
+      }
+   }
+
+   return size;
+}
+
+struct tgsi_instruction_ext_nv
+tgsi_default_instruction_ext_nv( void )
+{
+   struct tgsi_instruction_ext_nv instruction_ext_nv;
+
+   instruction_ext_nv.Type = TGSI_INSTRUCTION_EXT_TYPE_NV;
+   instruction_ext_nv.Precision = TGSI_PRECISION_DEFAULT;
+   instruction_ext_nv.CondDstIndex = 0;
+   instruction_ext_nv.CondFlowIndex = 0;
+   instruction_ext_nv.CondMask = TGSI_CC_TR;
+   instruction_ext_nv.CondSwizzleX = TGSI_SWIZZLE_X;
+   instruction_ext_nv.CondSwizzleY = TGSI_SWIZZLE_Y;
+   instruction_ext_nv.CondSwizzleZ = TGSI_SWIZZLE_Z;
+   instruction_ext_nv.CondSwizzleW = TGSI_SWIZZLE_W;
+   instruction_ext_nv.CondDstUpdate = 0;
+   instruction_ext_nv.CondFlowEnable = 0;
+   instruction_ext_nv.Padding = 0;
+   instruction_ext_nv.Extended = 0;
+
+   return instruction_ext_nv;
+}
+
+union token_u32
+{
+   GLuint   u32;
+};
+
+GLuint
+tgsi_compare_instruction_ext_nv(
+   struct tgsi_instruction_ext_nv a,
+   struct tgsi_instruction_ext_nv b )
+{
+   a.Padding = b.Padding = 0;
+   a.Extended = b.Extended = 0;
+   return ((union token_u32 *) &a)->u32 != ((union token_u32 *) &b)->u32;
+}
+
+struct tgsi_instruction_ext_nv
+tgsi_build_instruction_ext_nv(
+   GLuint precision,
+   GLuint cond_dst_index,
+   GLuint cond_flow_index,
+   GLuint cond_mask,
+   GLuint cond_swizzle_x,
+   GLuint cond_swizzle_y,
+   GLuint cond_swizzle_z,
+   GLuint cond_swizzle_w,
+   GLuint cond_dst_update,
+   GLuint cond_flow_update,
+   struct tgsi_token *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header )
+{
+   struct tgsi_instruction_ext_nv instruction_ext_nv;
+
+   instruction_ext_nv = tgsi_default_instruction_ext_nv();
+   instruction_ext_nv.Precision = precision;
+   instruction_ext_nv.CondDstIndex = cond_dst_index;
+   instruction_ext_nv.CondFlowIndex = cond_flow_index;
+   instruction_ext_nv.CondMask = cond_mask;
+   instruction_ext_nv.CondSwizzleX = cond_swizzle_x;
+   instruction_ext_nv.CondSwizzleY = cond_swizzle_y;
+   instruction_ext_nv.CondSwizzleZ = cond_swizzle_z;
+   instruction_ext_nv.CondSwizzleW = cond_swizzle_w;
+   instruction_ext_nv.CondDstUpdate = cond_dst_update;
+   instruction_ext_nv.CondFlowEnable = cond_flow_update;
+
+   prev_token->Extended = 1;
+   instruction_grow( instruction, header );
+
+   return instruction_ext_nv;
+}
+
+struct tgsi_instruction_ext_label
+tgsi_default_instruction_ext_label( void )
+{
+   struct tgsi_instruction_ext_label instruction_ext_label;
+
+   instruction_ext_label.Type = TGSI_INSTRUCTION_EXT_TYPE_LABEL;
+   instruction_ext_label.Label = 0;
+   instruction_ext_label.Target = 0;
+   instruction_ext_label.Padding = 0;
+   instruction_ext_label.Extended = 0;
+
+   return instruction_ext_label;
+}
+
+GLuint
+tgsi_compare_instruction_ext_label(
+   struct tgsi_instruction_ext_label a,
+   struct tgsi_instruction_ext_label b )
+{
+   a.Padding = b.Padding = 0;
+   a.Extended = b.Extended = 0;
+   return *(GLuint *) &a != *(GLuint *) &b;
+}
+
+struct tgsi_instruction_ext_label
+tgsi_build_instruction_ext_label(
+   GLuint label,
+   GLuint target,
+   struct tgsi_token  *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header )
+{
+   struct tgsi_instruction_ext_label instruction_ext_label;
+
+   instruction_ext_label = tgsi_default_instruction_ext_label();
+   instruction_ext_label.Label = label;
+   instruction_ext_label.Target = target;
+
+   prev_token->Extended = 1;
+   instruction_grow( instruction, header );
+
+   return instruction_ext_label;
+}
+
+struct tgsi_instruction_ext_texture
+tgsi_default_instruction_ext_texture( void )
+{
+   struct tgsi_instruction_ext_texture instruction_ext_texture;
+
+   instruction_ext_texture.Type = TGSI_INSTRUCTION_EXT_TYPE_TEXTURE;
+   instruction_ext_texture.Texture = TGSI_TEXTURE_UNKNOWN;
+   instruction_ext_texture.Padding = 0;
+   instruction_ext_texture.Extended = 0;
+
+   return instruction_ext_texture;
+}
+
+GLuint
+tgsi_compare_instruction_ext_texture(
+   struct tgsi_instruction_ext_texture a,
+   struct tgsi_instruction_ext_texture b )
+{
+   a.Padding = b.Padding = 0;
+   a.Extended = b.Extended = 0;
+   return *(GLuint *) &a != *(GLuint *) &b;
+}
+
+struct tgsi_instruction_ext_texture
+tgsi_build_instruction_ext_texture(
+   GLuint texture,
+   struct tgsi_token *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header )
+{
+   struct tgsi_instruction_ext_texture instruction_ext_texture;
+
+   instruction_ext_texture = tgsi_default_instruction_ext_texture();
+   instruction_ext_texture.Texture = texture;
+
+   prev_token->Extended = 1;
+   instruction_grow( instruction, header );
+
+   return instruction_ext_texture;
+}
+
+struct tgsi_src_register
+tgsi_default_src_register( void )
+{
+   struct tgsi_src_register src_register;
+
+   src_register.File = TGSI_FILE_NULL;
+   src_register.SwizzleX = TGSI_SWIZZLE_X;
+   src_register.SwizzleY = TGSI_SWIZZLE_Y;
+   src_register.SwizzleZ = TGSI_SWIZZLE_Z;
+   src_register.SwizzleW = TGSI_SWIZZLE_W;
+   src_register.Negate = 0;
+   src_register.Indirect = 0;
+   src_register.Dimension = 0;
+   src_register.Index = 0;
+   src_register.Extended = 0;
+
+   return src_register;
+}
+
+struct tgsi_src_register
+tgsi_build_src_register(
+   GLuint file,
+   GLuint swizzle_x,
+   GLuint swizzle_y,
+   GLuint swizzle_z,
+   GLuint swizzle_w,
+   GLuint negate,
+   GLuint indirect,
+   GLuint dimension,
+   GLint index,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header )
+{
+   struct tgsi_src_register   src_register;
+
+   assert( file <= TGSI_FILE_IMMEDIATE );
+   assert( swizzle_x <= TGSI_SWIZZLE_W );
+   assert( swizzle_y <= TGSI_SWIZZLE_W );
+   assert( swizzle_z <= TGSI_SWIZZLE_W );
+   assert( swizzle_w <= TGSI_SWIZZLE_W );
+   assert( negate <= 1 );
+   assert( index >= -0x8000 && index <= 0x7FFF );
+
+   src_register = tgsi_default_src_register();
+   src_register.File = file;
+   src_register.SwizzleX = swizzle_x;
+   src_register.SwizzleY = swizzle_y;
+   src_register.SwizzleZ = swizzle_z;
+   src_register.SwizzleW = swizzle_w;
+   src_register.Negate = negate;
+   src_register.Indirect = indirect;
+   src_register.Dimension = dimension;
+   src_register.Index = index;
+
+   instruction_grow( instruction, header );
+
+   return src_register;
+}
+
+struct tgsi_full_src_register
+tgsi_default_full_src_register( void )
+{
+   struct tgsi_full_src_register full_src_register;
+
+   full_src_register.SrcRegister = tgsi_default_src_register();
+   full_src_register.SrcRegisterExtSwz = tgsi_default_src_register_ext_swz();
+   full_src_register.SrcRegisterExtMod = tgsi_default_src_register_ext_mod();
+   full_src_register.SrcRegisterInd = tgsi_default_src_register();
+   full_src_register.SrcRegisterDim = tgsi_default_dimension();
+   full_src_register.SrcRegisterDimInd = tgsi_default_src_register();
+
+   return full_src_register;
+}
+
+struct tgsi_src_register_ext_swz
+tgsi_default_src_register_ext_swz( void )
+{
+   struct tgsi_src_register_ext_swz src_register_ext_swz;
+
+   src_register_ext_swz.Type = TGSI_SRC_REGISTER_EXT_TYPE_SWZ;
+   src_register_ext_swz.ExtSwizzleX = TGSI_EXTSWIZZLE_X;
+   src_register_ext_swz.ExtSwizzleY = TGSI_EXTSWIZZLE_Y;
+   src_register_ext_swz.ExtSwizzleZ = TGSI_EXTSWIZZLE_Z;
+   src_register_ext_swz.ExtSwizzleW = TGSI_EXTSWIZZLE_W;
+   src_register_ext_swz.NegateX = 0;
+   src_register_ext_swz.NegateY = 0;
+   src_register_ext_swz.NegateZ = 0;
+   src_register_ext_swz.NegateW = 0;
+   src_register_ext_swz.ExtDivide = TGSI_EXTSWIZZLE_ONE;
+   src_register_ext_swz.Padding = 0;
+   src_register_ext_swz.Extended = 0;
+
+   return src_register_ext_swz;
+}
+
+GLuint
+tgsi_compare_src_register_ext_swz(
+   struct tgsi_src_register_ext_swz a,
+   struct tgsi_src_register_ext_swz b )
+{
+   a.Padding = b.Padding = 0;
+   a.Extended = b.Extended = 0;
+   return *(GLuint *) &a != *(GLuint *) &b;
+}
+
+struct tgsi_src_register_ext_swz
+tgsi_build_src_register_ext_swz(
+   GLuint ext_swizzle_x,
+   GLuint ext_swizzle_y,
+   GLuint ext_swizzle_z,
+   GLuint ext_swizzle_w,
+   GLuint negate_x,
+   GLuint negate_y,
+   GLuint negate_z,
+   GLuint negate_w,
+   GLuint ext_divide,
+   struct tgsi_token *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header )
+{
+   struct tgsi_src_register_ext_swz src_register_ext_swz;
+
+   assert (ext_swizzle_x <= TGSI_EXTSWIZZLE_ONE);
+   assert (ext_swizzle_y <= TGSI_EXTSWIZZLE_ONE);
+   assert (ext_swizzle_z <= TGSI_EXTSWIZZLE_ONE);
+   assert (ext_swizzle_w <= TGSI_EXTSWIZZLE_ONE);
+   assert (negate_x <= 1);
+   assert (negate_y <= 1);
+   assert (negate_z <= 1);
+   assert (negate_w <= 1);
+   assert (ext_divide <= TGSI_EXTSWIZZLE_ONE);
+
+   src_register_ext_swz = tgsi_default_src_register_ext_swz();
+   src_register_ext_swz.ExtSwizzleX = ext_swizzle_x;
+   src_register_ext_swz.ExtSwizzleY = ext_swizzle_y;
+   src_register_ext_swz.ExtSwizzleZ = ext_swizzle_z;
+   src_register_ext_swz.ExtSwizzleW = ext_swizzle_w;
+   src_register_ext_swz.NegateX = negate_x;
+   src_register_ext_swz.NegateY = negate_y;
+   src_register_ext_swz.NegateZ = negate_z;
+   src_register_ext_swz.NegateW = negate_w;
+   src_register_ext_swz.ExtDivide = ext_divide;
+
+   prev_token->Extended = 1;
+   instruction_grow( instruction, header );
+
+   return src_register_ext_swz;
+}
+
+struct tgsi_src_register_ext_mod
+tgsi_default_src_register_ext_mod( void )
+{
+   struct tgsi_src_register_ext_mod src_register_ext_mod;
+
+   src_register_ext_mod.Type = TGSI_SRC_REGISTER_EXT_TYPE_MOD;
+   src_register_ext_mod.Complement = 0;
+   src_register_ext_mod.Bias = 0;
+   src_register_ext_mod.Scale2X = 0;
+   src_register_ext_mod.Absolute = 0;
+   src_register_ext_mod.Negate = 0;
+   src_register_ext_mod.Padding = 0;
+   src_register_ext_mod.Extended = 0;
+
+   return src_register_ext_mod;
+}
+
+GLuint
+tgsi_compare_src_register_ext_mod(
+   struct tgsi_src_register_ext_mod a,
+   struct tgsi_src_register_ext_mod b )
+{
+   a.Padding = b.Padding = 0;
+   a.Extended = b.Extended = 0;
+   return *(GLuint *) &a != *(GLuint *) &b;
+}
+
+struct tgsi_src_register_ext_mod
+tgsi_build_src_register_ext_mod(
+   GLuint  complement,
+   GLuint  bias,
+   GLuint  scale_2x,
+   GLuint  absolute,
+   GLuint  negate,
+   struct  tgsi_token *prev_token,
+   struct  tgsi_instruction *instruction,
+   struct  tgsi_header *header )
+{
+   struct tgsi_src_register_ext_mod src_register_ext_mod;
+
+   assert (complement <= 1);
+   assert (bias <= 1);
+   assert (scale_2x <= 1);
+   assert (absolute <= 1);
+   assert (negate <= 1);
+
+   src_register_ext_mod = tgsi_default_src_register_ext_mod();
+   src_register_ext_mod.Complement = complement;
+   src_register_ext_mod.Bias = bias;
+   src_register_ext_mod.Scale2X = scale_2x;
+   src_register_ext_mod.Absolute = absolute;
+   src_register_ext_mod.Negate = negate;
+
+   prev_token->Extended = 1;
+   instruction_grow( instruction, header );
+
+   return src_register_ext_mod;
+}
+
+struct tgsi_dimension
+tgsi_default_dimension( void )
+{
+   struct tgsi_dimension dimension;
+
+   dimension.Indirect = 0;
+   dimension.Dimension = 0;
+   dimension.Padding = 0;
+   dimension.Index = 0;
+   dimension.Extended = 0;
+
+   return dimension;
+}
+
+struct tgsi_dimension
+tgsi_build_dimension(
+   GLuint indirect,
+   GLuint index,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header )
+{
+   struct tgsi_dimension dimension;
+
+   dimension = tgsi_default_dimension();
+   dimension.Indirect = indirect;
+   dimension.Index = index;
+
+   instruction_grow( instruction, header );
+
+   return dimension;
+}
+
+struct tgsi_dst_register
+tgsi_default_dst_register( void )
+{
+   struct tgsi_dst_register dst_register;
+
+   dst_register.File = TGSI_FILE_NULL;
+   dst_register.WriteMask = TGSI_WRITEMASK_XYZW;
+   dst_register.Indirect = 0;
+   dst_register.Dimension = 0;
+   dst_register.Index = 0;
+   dst_register.Padding = 0;
+   dst_register.Extended = 0;
+
+   return dst_register;
+}
+
+struct tgsi_dst_register
+tgsi_build_dst_register(
+   GLuint file,
+   GLuint mask,
+   GLint index,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header )
+{
+   struct tgsi_dst_register dst_register;
+
+   assert (file <= TGSI_FILE_IMMEDIATE);
+   assert (mask <= TGSI_WRITEMASK_XYZW);
+   assert (index >= -32768 && index <= 32767);
+
+   dst_register = tgsi_default_dst_register();
+   dst_register.File = file;
+   dst_register.WriteMask = mask;
+   dst_register.Index = index;
+
+   instruction_grow( instruction, header );
+
+   return dst_register;
+}
+
+struct tgsi_full_dst_register
+tgsi_default_full_dst_register( void )
+{
+   struct tgsi_full_dst_register full_dst_register;
+
+   full_dst_register.DstRegister = tgsi_default_dst_register();
+   full_dst_register.DstRegisterExtConcode =
+      tgsi_default_dst_register_ext_concode();
+   full_dst_register.DstRegisterExtModulate =
+      tgsi_default_dst_register_ext_modulate();
+
+   return full_dst_register;
+}
+
+struct tgsi_dst_register_ext_concode
+tgsi_default_dst_register_ext_concode( void )
+{
+   struct tgsi_dst_register_ext_concode dst_register_ext_concode;
+
+   dst_register_ext_concode.Type = TGSI_DST_REGISTER_EXT_TYPE_CONDCODE;
+   dst_register_ext_concode.CondMask = TGSI_CC_TR;
+   dst_register_ext_concode.CondSwizzleX = TGSI_SWIZZLE_X;
+   dst_register_ext_concode.CondSwizzleY = TGSI_SWIZZLE_Y;
+   dst_register_ext_concode.CondSwizzleZ = TGSI_SWIZZLE_Z;
+   dst_register_ext_concode.CondSwizzleW = TGSI_SWIZZLE_W;
+   dst_register_ext_concode.CondSrcIndex = 0;
+   dst_register_ext_concode.Padding = 0;
+   dst_register_ext_concode.Extended = 0;
+
+   return dst_register_ext_concode;
+}
+
+GLuint
+tgsi_compare_dst_register_ext_concode(
+   struct tgsi_dst_register_ext_concode a,
+   struct tgsi_dst_register_ext_concode b )
+{
+   a.Padding = b.Padding = 0;
+   a.Extended = b.Extended = 0;
+   return *(GLuint *) &a != *(GLuint *) &b;
+}
+
+struct tgsi_dst_register_ext_concode
+tgsi_build_dst_register_ext_concode(
+   GLuint  cc,
+   GLuint  swizzle_x,
+   GLuint  swizzle_y,
+   GLuint  swizzle_z,
+   GLuint  swizzle_w,
+   GLint index,
+   struct  tgsi_token *prev_token,
+   struct  tgsi_instruction *instruction,
+   struct  tgsi_header *header )
+{
+   struct tgsi_dst_register_ext_concode dst_register_ext_concode;
+
+   assert (cc <= TGSI_CC_FL);
+   assert (swizzle_x <= TGSI_SWIZZLE_W);
+   assert (swizzle_y <= TGSI_SWIZZLE_W);
+   assert (swizzle_z <= TGSI_SWIZZLE_W);
+   assert (swizzle_w <= TGSI_SWIZZLE_W);
+   assert (index >= -32768 && index <= 32767);
+
+   dst_register_ext_concode = tgsi_default_dst_register_ext_concode();
+   dst_register_ext_concode.CondMask = cc;
+   dst_register_ext_concode.CondSwizzleX = swizzle_x;
+   dst_register_ext_concode.CondSwizzleY = swizzle_y;
+   dst_register_ext_concode.CondSwizzleZ = swizzle_z;
+   dst_register_ext_concode.CondSwizzleW = swizzle_w;
+   dst_register_ext_concode.CondSrcIndex = index;
+
+   prev_token->Extended = 1;
+   instruction_grow( instruction, header );
+
+   return dst_register_ext_concode;
+}
+
+struct tgsi_dst_register_ext_modulate
+tgsi_default_dst_register_ext_modulate( void )
+{
+   struct tgsi_dst_register_ext_modulate dst_register_ext_modulate;
+
+   dst_register_ext_modulate.Type = TGSI_DST_REGISTER_EXT_TYPE_MODULATE;
+   dst_register_ext_modulate.Modulate = TGSI_MODULATE_1X;
+   dst_register_ext_modulate.Padding = 0;
+   dst_register_ext_modulate.Extended = 0;
+
+   return dst_register_ext_modulate;
+}
+
+GLuint
+tgsi_compare_dst_register_ext_modulate(
+   struct tgsi_dst_register_ext_modulate a,
+   struct tgsi_dst_register_ext_modulate b )
+{
+   a.Padding = b.Padding = 0;
+   a.Extended = b.Extended = 0;
+   return *(GLuint *) &a != *(GLuint *) &b;
+}
+
+struct tgsi_dst_register_ext_modulate
+tgsi_build_dst_register_ext_modulate(
+   GLuint modulate,
+   struct tgsi_token  *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header )
+{
+   struct tgsi_dst_register_ext_modulate dst_register_ext_modulate;
+
+   assert (modulate <=  TGSI_MODULATE_EIGHTH);
+
+   dst_register_ext_modulate = tgsi_default_dst_register_ext_modulate();
+   dst_register_ext_modulate.Modulate = modulate;
+
+   prev_token->Extended = 1;
+   instruction_grow( instruction, header );
+
+   return dst_register_ext_modulate;
+}
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_build.h b/src/mesa/pipe/tgsi/core/tgsi_build.h
new file mode 100644
index 00000000000..db259565193
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_build.h
@@ -0,0 +1,309 @@
+#if !defined TGSI_BUILD_H
+#define TGSI_BUILD_H
+
+#if defined __cplusplus
+extern "C" {
+#endif // defined __cplusplus
+
+/*
+ * version
+ */
+
+struct tgsi_version
+tgsi_build_version( void );
+
+/*
+ * header
+ */
+
+struct tgsi_header
+tgsi_build_header( void );
+
+struct tgsi_processor
+tgsi_default_processor( void );
+
+struct tgsi_processor
+tgsi_build_processor(
+   GLuint processor,
+   struct tgsi_header *header );
+
+/*
+ * declaration
+ */
+
+struct tgsi_declaration
+tgsi_default_declaration( void );
+
+struct tgsi_declaration
+tgsi_build_declaration(
+   GLuint file,
+   GLuint declare,
+   GLuint interpolate,
+   struct tgsi_header *header );
+
+struct tgsi_full_declaration
+tgsi_default_full_declaration( void );
+
+GLuint
+tgsi_build_full_declaration(
+   const struct tgsi_full_declaration *full_decl,
+   struct tgsi_token *tokens,
+   struct tgsi_header *header,
+   GLuint maxsize );
+
+struct tgsi_declaration_range
+tgsi_build_declaration_range(
+   GLuint first,
+   GLuint last,
+   struct tgsi_declaration *declaration,
+   struct tgsi_header *header );
+
+struct tgsi_declaration_mask
+tgsi_build_declaration_mask(
+   GLuint mask,
+   struct tgsi_declaration *declaration,
+   struct tgsi_header *header );
+
+struct tgsi_declaration_interpolation
+tgsi_default_declaration_interpolation( void );
+
+struct tgsi_declaration_interpolation
+tgsi_build_declaration_interpolation(
+   GLuint interpolate,
+   struct tgsi_declaration *declaration,
+   struct tgsi_header *header );
+
+/*
+ * immediate
+ */
+
+struct tgsi_immediate
+tgsi_default_immediate( void );
+
+struct tgsi_immediate
+tgsi_build_immediate(
+   struct tgsi_header *header );
+
+struct tgsi_full_immediate
+tgsi_default_full_immediate( void );
+
+struct tgsi_immediate_float32
+tgsi_build_immediate_float32(
+   GLfloat value,
+   struct tgsi_immediate *immediate,
+   struct tgsi_header *header );
+
+GLuint
+tgsi_build_full_immediate(
+   const struct tgsi_full_immediate *full_imm,
+   struct tgsi_token *tokens,
+   struct tgsi_header *header,
+   GLuint maxsize );
+
+/*
+ * instruction
+ */
+
+struct tgsi_instruction
+tgsi_default_instruction( void );
+
+struct tgsi_instruction
+tgsi_build_instruction(
+   GLuint opcode,
+   GLuint saturate,
+   GLuint num_dst_regs,
+   GLuint num_src_regs,
+   struct tgsi_header *header );
+
+struct tgsi_full_instruction
+tgsi_default_full_instruction( void );
+
+GLuint
+tgsi_build_full_instruction(
+   const struct tgsi_full_instruction *full_inst,
+   struct tgsi_token *tokens,
+   struct tgsi_header *header,
+   GLuint maxsize );
+
+struct tgsi_instruction_ext_nv
+tgsi_default_instruction_ext_nv( void );
+
+GLuint
+tgsi_compare_instruction_ext_nv(
+   struct tgsi_instruction_ext_nv a,
+   struct tgsi_instruction_ext_nv b );
+
+struct tgsi_instruction_ext_nv
+tgsi_build_instruction_ext_nv(
+   GLuint precision,
+   GLuint cond_dst_index,
+   GLuint cond_flow_index,
+   GLuint cond_mask,
+   GLuint cond_swizzle_x,
+   GLuint cond_swizzle_y,
+   GLuint cond_swizzle_z,
+   GLuint cond_swizzle_w,
+   GLuint cond_dst_update,
+   GLuint cond_flow_update,
+   struct tgsi_token *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header );
+
+struct tgsi_instruction_ext_label
+tgsi_default_instruction_ext_label( void );
+
+GLuint
+tgsi_compare_instruction_ext_label(
+   struct tgsi_instruction_ext_label a,
+   struct tgsi_instruction_ext_label b );
+
+struct tgsi_instruction_ext_label
+tgsi_build_instruction_ext_label(
+   GLuint label,
+   GLuint target,
+   struct tgsi_token *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header );
+
+struct tgsi_instruction_ext_texture
+tgsi_default_instruction_ext_texture( void );
+
+GLuint
+tgsi_compare_instruction_ext_texture(
+   struct tgsi_instruction_ext_texture a,
+   struct tgsi_instruction_ext_texture b );
+
+struct tgsi_instruction_ext_texture
+tgsi_build_instruction_ext_texture(
+   GLuint texture,
+   struct tgsi_token *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header );
+
+struct tgsi_src_register
+tgsi_default_src_register( void );
+
+struct tgsi_src_register
+tgsi_build_src_register(
+   GLuint file,
+   GLuint swizzle_x,
+   GLuint swizzle_y,
+   GLuint swizzle_z,
+   GLuint swizzle_w,
+   GLuint negate,
+   GLuint indirect,
+   GLuint dimension,
+   GLint index,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header );
+
+struct tgsi_full_src_register
+tgsi_default_full_src_register( void );
+
+struct tgsi_src_register_ext_swz
+tgsi_default_src_register_ext_swz( void );
+
+GLuint
+tgsi_compare_src_register_ext_swz(
+   struct tgsi_src_register_ext_swz a,
+   struct tgsi_src_register_ext_swz b );
+
+struct tgsi_src_register_ext_swz
+tgsi_build_src_register_ext_swz(
+   GLuint ext_swizzle_x,
+   GLuint ext_swizzle_y,
+   GLuint ext_swizzle_z,
+   GLuint ext_swizzle_w,
+   GLuint negate_x,
+   GLuint negate_y,
+   GLuint negate_z,
+   GLuint negate_w,
+   GLuint ext_divide,
+   struct tgsi_token *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header );
+
+struct tgsi_src_register_ext_mod
+tgsi_default_src_register_ext_mod( void );
+
+GLuint
+tgsi_compare_src_register_ext_mod(
+   struct tgsi_src_register_ext_mod a,
+   struct tgsi_src_register_ext_mod b );
+
+struct tgsi_src_register_ext_mod
+tgsi_build_src_register_ext_mod(
+   GLuint complement,
+   GLuint bias,
+   GLuint scale_2x,
+   GLuint absolute,
+   GLuint negate,
+   struct tgsi_token *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header );
+
+struct tgsi_dimension
+tgsi_default_dimension( void );
+
+struct tgsi_dimension
+tgsi_build_dimension(
+   GLuint indirect,
+   GLuint index,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header );
+
+struct tgsi_dst_register
+tgsi_default_dst_register( void );
+
+struct tgsi_dst_register
+tgsi_build_dst_register(
+   GLuint file,
+   GLuint mask,
+   GLint index,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header );
+
+struct tgsi_full_dst_register
+tgsi_default_full_dst_register( void );
+
+struct tgsi_dst_register_ext_concode
+tgsi_default_dst_register_ext_concode( void );
+
+GLuint
+tgsi_compare_dst_register_ext_concode(
+   struct tgsi_dst_register_ext_concode a,
+   struct tgsi_dst_register_ext_concode b );
+
+struct tgsi_dst_register_ext_concode
+tgsi_build_dst_register_ext_concode(
+   GLuint cc,
+   GLuint swizzle_x,
+   GLuint swizzle_y,
+   GLuint swizzle_z,
+   GLuint swizzle_w,
+   GLint index,
+   struct tgsi_token *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header );
+
+struct tgsi_dst_register_ext_modulate
+tgsi_default_dst_register_ext_modulate( void );
+
+GLuint
+tgsi_compare_dst_register_ext_modulate(
+   struct tgsi_dst_register_ext_modulate a,
+   struct tgsi_dst_register_ext_modulate b );
+
+struct tgsi_dst_register_ext_modulate
+tgsi_build_dst_register_ext_modulate(
+   GLuint modulate,
+   struct tgsi_token *prev_token,
+   struct tgsi_instruction *instruction,
+   struct tgsi_header *header );
+
+#if defined __cplusplus
+} // extern "C"
+#endif // defined __cplusplus
+
+#endif // !defined TGSI_BUILD_H
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_core.h b/src/mesa/pipe/tgsi/core/tgsi_core.h
new file mode 100644
index 00000000000..1f5f00a38ea
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_core.h
@@ -0,0 +1,12 @@
+#if !defined TGSI_CORE_H
+#define TGSI_CORE_H
+
+#include "tgsi_token.h"
+#include "tgsi_parse.h"
+#include "tgsi_build.h"
+#include "tgsi_exec.h"
+#include "tgsi_dump.h"
+#include "tgsi_util.h"
+
+#endif // !defined TGSI_CORE_H
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_dump.c b/src/mesa/pipe/tgsi/core/tgsi_dump.c
new file mode 100644
index 00000000000..fecb246ab1a
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_dump.c
@@ -0,0 +1,871 @@
+#include "tgsi_platform.h"
+#include "tgsi_core.h"
+
+struct text_dump
+{
+    FILE *file;
+    GLuint tabs;
+};
+
+static void
+text_dump_write(
+   struct text_dump *dump,
+   const void *buffer,
+   GLuint size )
+{
+   fwrite( buffer, size, 1, dump->file );
+}
+
+static void
+text_dump_str(
+   struct text_dump *dump,
+   const char *str )
+{
+   GLuint i;
+   GLuint len = strlen( str );
+
+   for( i = 0; i < len; i++ ) {
+      text_dump_write( dump, &str[i], 1 );
+
+      if( str[i] == '\n' ) {
+         GLuint i;
+
+         for( i = 0; i < dump->tabs; i++ ) {
+            text_dump_write( dump, "    ", 4 );
+         }
+      }
+   }
+}
+
+static void
+text_dump_chr(
+   struct text_dump *dump,
+   const char chr )
+{
+   char str[2];
+
+   str[0] = chr;
+   str[1] = '\0';
+   text_dump_str( dump, str );
+}
+
+static void
+text_dump_uix(
+   struct text_dump *dump,
+   const GLuint ui)
+{
+   char str[36];
+
+   sprintf( str, "0x%x", ui );
+   text_dump_str( dump, str );
+}
+
+static void
+text_dump_uid(
+   struct text_dump *dump,
+   const GLuint ui )
+{
+   char str[16];
+
+   sprintf( str, "%u", ui );
+   text_dump_str( dump, str );
+}
+
+static void
+text_dump_sid(
+   struct text_dump *dump,
+   const GLint si )
+{
+   char str[16];
+
+   sprintf( str, "%d", si );
+   text_dump_str( dump, str );
+}
+
+static void
+text_dump_flt(
+   struct text_dump *dump,
+   const GLfloat f )
+{
+   char str[48];
+
+   sprintf( str, "%40.6f", f );
+   text_dump_str( dump, str );
+}
+
+static void
+text_dump_enum(
+   struct text_dump *dump,
+   const GLuint e,
+   const char **enums,
+   const GLuint enums_count )
+{
+   if( e >= enums_count ) {
+      text_dump_uid( dump, e );
+   }
+   else {
+      text_dump_str( dump, enums[e] );
+   }
+}
+
+static void
+text_dump_tab(
+   struct text_dump *dump )
+{
+   dump->tabs++;
+}
+
+static void
+text_dump_untab(
+   struct text_dump *dump )
+{
+   assert( dump->tabs > 0 );
+
+   --dump->tabs;
+}
+
+#define TXT(S)          text_dump_str( &dump, S )
+#define CHR(C)          text_dump_chr( &dump, C )
+#define UIX(I)          text_dump_uix( &dump, I )
+#define UID(I)          text_dump_uid( &dump, I )
+#define SID(I)          text_dump_sid( &dump, I )
+#define FLT(F)          text_dump_flt( &dump, F )
+#define TAB()           text_dump_tab( &dump )
+#define UNT()           text_dump_untab( &dump )
+#define ENM(E,ENUMS)    text_dump_enum( &dump, E, ENUMS, sizeof( ENUMS ) / sizeof( *ENUMS ) )
+
+static const char *TGSI_PROCESSOR_TYPES[] =
+{
+   "PROCESSOR_FRAGMENT",
+   "PROCESSOR_VERTEX"
+};
+
+static const char *TGSI_TOKEN_TYPES[] =
+{
+   "TOKEN_TYPE_DECLARATION",
+   "TOKEN_TYPE_IMMEDIATE",
+   "TOKEN_TYPE_INSTRUCTION"
+};
+
+static const char *TGSI_FILES[] =
+{
+   "FILE_NULL",
+   "FILE_CONSTANT",
+   "FILE_INPUT",
+   "FILE_OUTPUT",
+   "FILE_TEMPORARY",
+   "FILE_SAMPLER",
+   "FILE_ADDRESS",
+   "FILE_IMMEDIATE"
+};
+
+static const char *TGSI_DECLARES[] =
+{
+   "DECLARE_RANGE",
+   "DECLARE_MASK"
+};
+
+static const char *TGSI_INTERPOLATES[] =
+{
+   "INTERPOLATE_CONSTANT",
+   "INTERPOLATE_LINEAR",
+   "INTERPOLATE_PERSPECTIVE"
+};
+
+static const char *TGSI_IMMS[] =
+{
+   "IMM_FLOAT32"
+};
+
+static const char *TGSI_OPCODES[] =
+{
+   "OPCODE_MOV",
+   "OPCODE_LIT",
+   "OPCODE_RCP",
+   "OPCODE_RSQ",
+   "OPCODE_EXP",
+   "OPCODE_LOG",
+   "OPCODE_MUL",
+   "OPCODE_ADD",
+   "OPCODE_DP3",
+   "OPCODE_DP4",
+   "OPCODE_DST",
+   "OPCODE_MIN",
+   "OPCODE_MAX",
+   "OPCODE_SLT",
+   "OPCODE_SGE",
+   "OPCODE_MAD",
+   "OPCODE_SUB",
+   "OPCODE_LERP",
+   "OPCODE_CND",
+   "OPCODE_CND0",
+   "OPCODE_DOT2ADD",
+   "OPCODE_INDEX",
+   "OPCODE_NEGATE",
+   "OPCODE_FRAC",
+   "OPCODE_CLAMP",
+   "OPCODE_FLOOR",
+   "OPCODE_ROUND",
+   "OPCODE_EXPBASE2",
+   "OPCODE_LOGBASE2",
+   "OPCODE_POWER",
+   "OPCODE_CROSSPRODUCT",
+   "OPCODE_MULTIPLYMATRIX",
+   "OPCODE_ABS",
+   "OPCODE_RCC",
+   "OPCODE_DPH",
+   "OPCODE_COS",
+   "OPCODE_DDX",
+   "OPCODE_DDY",
+   "OPCODE_KIL",
+   "OPCODE_PK2H",
+   "OPCODE_PK2US",
+   "OPCODE_PK4B",
+   "OPCODE_PK4UB",
+   "OPCODE_RFL",
+   "OPCODE_SEQ",
+   "OPCODE_SFL",
+   "OPCODE_SGT",
+   "OPCODE_SIN",
+   "OPCODE_SLE",
+   "OPCODE_SNE",
+   "OPCODE_STR",
+   "OPCODE_TEX",
+   "OPCODE_TXD",
+   "OPCODE_UP2H",
+   "OPCODE_UP2US",
+   "OPCODE_UP4B",
+   "OPCODE_UP4UB",
+   "OPCODE_X2D",
+   "OPCODE_ARA",
+   "OPCODE_ARR",
+   "OPCODE_BRA",
+   "OPCODE_CAL",
+   "OPCODE_RET",
+   "OPCODE_SSG",
+   "OPCODE_CMP",
+   "OPCODE_SCS",
+   "OPCODE_TXB",
+   "OPCODE_NRM",
+   "OPCODE_DIV",
+   "OPCODE_DP2",
+   "OPCODE_TXL",
+   "OPCODE_BRK",
+   "OPCODE_IF",
+   "OPCODE_LOOP",
+   "OPCODE_REP",
+   "OPCODE_ELSE",
+   "OPCODE_ENDIF",
+   "OPCODE_ENDLOOP",
+   "OPCODE_ENDREP",
+   "OPCODE_PUSHA",
+   "OPCODE_POPA",
+   "OPCODE_CEIL",
+   "OPCODE_I2F",
+   "OPCODE_NOT",
+   "OPCODE_TRUNC",
+   "OPCODE_SHL",
+   "OPCODE_SHR",
+   "OPCODE_AND",
+   "OPCODE_OR",
+   "OPCODE_MOD",
+   "OPCODE_XOR",
+   "OPCODE_SAD",
+   "OPCODE_TXF",
+   "OPCODE_TXQ",
+   "OPCODE_CONT",
+   "OPCODE_EMIT",
+   "OPCODE_ENDPRIM"
+};
+
+static const char *TGSI_SATS[] =
+{
+   "SAT_NONE",
+   "SAT_ZERO_ONE",
+   "SAT_MINUS_PLUS_ONE"
+};
+
+static const char *TGSI_INSTRUCTION_EXTS[] =
+{
+   "INSTRUCTION_EXT_TYPE_NV",
+   "INSTRUCTION_EXT_TYPE_LABEL",
+   "INSTRUCTION_EXT_TYPE_TEXTURE"
+};
+
+static const char *TGSI_PRECISIONS[] =
+{
+   "PRECISION_DEFAULT",
+   "TGSI_PRECISION_FLOAT32",
+   "TGSI_PRECISION_FLOAT16",
+   "TGSI_PRECISION_FIXED12"
+};
+
+static const char *TGSI_CCS[] =
+{
+   "CC_GT",
+   "CC_EQ",
+   "CC_LT",
+   "CC_UN",
+   "CC_GE",
+   "CC_LE",
+   "CC_NE",
+   "CC_TR",
+   "CC_FL"
+};
+
+static const char *TGSI_SWIZZLES[] =
+{
+   "SWIZZLE_X",
+   "SWIZZLE_Y",
+   "SWIZZLE_Z",
+   "SWIZZLE_W"
+};
+
+static const char *TGSI_TEXTURES[] =
+{
+   "TEXTURE_UNKNOWN",
+   "TEXTURE_1D",
+   "TEXTURE_2D",
+   "TEXTURE_3D",
+   "TEXTURE_CUBE",
+   "TEXTURE_RECT",
+   "TEXTURE_SHADOW1D",
+   "TEXTURE_SHADOW2D",
+   "TEXTURE_SHADOWRECT"
+};
+
+static const char *TGSI_SRC_REGISTER_EXTS[] =
+{
+   "SRC_REGISTER_EXT_TYPE_SWZ",
+   "SRC_REGISTER_EXT_TYPE_MOD"
+};
+
+static const char *TGSI_EXTSWIZZLES[] =
+{
+   "EXTSWIZZLE_X",
+   "EXTSWIZZLE_Y",
+   "EXTSWIZZLE_Z",
+   "EXTSWIZZLE_W",
+   "EXTSWIZZLE_ZERO",
+   "EXTSWIZZLE_ONE"
+};
+
+static const char *TGSI_WRITEMASKS[] =
+{
+   "0",
+   "WRITEMASK_X",
+   "WRITEMASK_Y",
+   "WRITEMASK_XY",
+   "WRITEMASK_Z",
+   "WRITEMASK_XZ",
+   "WRITEMASK_YZ",
+   "WRITEMASK_XYZ",
+   "WRITEMASK_W",
+   "WRITEMASK_XW",
+   "WRITEMASK_YW",
+   "WRITEMASK_XYW",
+   "WRITEMASK_ZW",
+   "WRITEMASK_XZW",
+   "WRITEMASK_YZW",
+   "WRITEMASK_XYZW"
+};
+
+static const char *TGSI_DST_REGISTER_EXTS[] =
+{
+   "DST_REGISTER_EXT_TYPE_CONDCODE",
+   "DST_REGISTER_EXT_TYPE_MODULATE"
+};
+
+static const char *TGSI_MODULATES[] =
+{
+   "MODULATE_1X",
+   "MODULATE_2X",
+   "MODULATE_4X",
+   "MODULATE_8X",
+   "MODULATE_HALF",
+   "MODULATE_QUARTER",
+   "MODULATE_EIGHTH"
+};
+
+void
+tgsi_dump(
+   const struct tgsi_token *tokens,
+   GLuint flags )
+{
+   struct text_dump dump;
+   struct tgsi_parse_context parse;
+   struct tgsi_full_instruction fi;
+   struct tgsi_full_declaration fd;
+   GLuint ignored = !(flags & TGSI_DUMP_NO_IGNORED);
+   GLuint deflt = !(flags & TGSI_DUMP_NO_DEFAULT);
+
+   {
+      static GLuint counter = 0;
+      char buffer[64];
+
+      sprintf( buffer, "sbir-dump-%.4u.txt", counter++ );
+      dump.file = fopen( buffer, "wt" );
+      dump.tabs = 0;
+   }
+
+   tgsi_parse_init( &parse, tokens );
+
+   TXT( "sbir-dump" );
+
+   CHR( '\n' );
+   TXT( "\nMajorVersion: " );
+   UID( parse.FullVersion.Version.MajorVersion );
+   TXT( "\nMinorVersion: " );
+   UID( parse.FullVersion.Version.MinorVersion );
+
+   CHR( '\n' );
+   TXT( "\nHeaderSize: " );
+   UID( parse.FullHeader.Header.HeaderSize );
+   TXT( "\nBodySize  : " );
+   UID( parse.FullHeader.Header.BodySize );
+   TXT( "\nProcessor : " );
+   ENM( parse.FullHeader.Processor.Processor, TGSI_PROCESSOR_TYPES );
+
+   fi = tgsi_default_full_instruction();
+   fd = tgsi_default_full_declaration();
+
+   while( !tgsi_parse_end_of_tokens( &parse ) ) {
+      GLuint i;
+
+      tgsi_parse_token( &parse );
+
+      CHR( '\n' );
+      TXT( "\nType       : " );
+      ENM( parse.FullToken.Token.Type, TGSI_TOKEN_TYPES );
+      if( ignored ) {
+         TXT( "\nSize       : " );
+         UID( parse.FullToken.Token.Size );
+         if( deflt || parse.FullToken.Token.Extended ) {
+            TXT( "\nExtended   : " );
+            UID( parse.FullToken.Token.Extended );
+         }
+      }
+
+      switch( parse.FullToken.Token.Type ) {
+      case TGSI_TOKEN_TYPE_DECLARATION:
+         {
+            struct tgsi_full_declaration *decl = &parse.FullToken.FullDeclaration;
+
+            TXT( "\nFile       : " );
+            ENM( decl->Declaration.File, TGSI_FILES );
+            TXT( "\nDeclare    : " );
+            ENM( decl->Declaration.Declare, TGSI_DECLARES );
+            if( deflt || fd.Declaration.Interpolate != decl->Declaration.Interpolate ) {
+               TXT( "\nInterpolate: " );
+               UID( decl->Declaration.Interpolate );
+            }
+            if( ignored ) {
+               TXT( "\nPadding    : " );
+               UIX( decl->Declaration.Padding );
+            }
+
+            CHR( '\n' );
+            switch( decl->Declaration.Declare ) {
+            case TGSI_DECLARE_RANGE:
+               TXT( "\nFirst: " );
+               UID( decl->u.DeclarationRange.First );
+               TXT( "\nLast : " );
+               UID( decl->u.DeclarationRange.Last );
+               break;
+
+            case TGSI_DECLARE_MASK:
+               TXT( "\nMask: " );
+               UIX( decl->u.DeclarationMask.Mask );
+               break;
+
+            default:
+               assert( 0 );
+            }
+
+            if( decl->Declaration.Interpolate ) {
+               CHR( '\n' );
+               TXT( "\nInterpolate: " );
+               ENM( decl->Interpolation.Interpolate, TGSI_INTERPOLATES );
+               if( ignored ) {
+                  TXT( "\nPadding    : " );
+                  UIX( decl->Interpolation.Padding );
+               }
+            }
+         }
+         break;
+
+      case TGSI_TOKEN_TYPE_IMMEDIATE:
+         TXT( "\nDataType   : " );
+         ENM( parse.FullToken.FullImmediate.Immediate.DataType, TGSI_IMMS );
+         if( ignored ) {
+            TXT( "\nPadding    : " );
+            UIX( parse.FullToken.FullImmediate.Immediate.Padding );
+         }
+
+         for( i = 0; i < parse.FullToken.FullImmediate.Immediate.Size - 1; i++ ) {
+            CHR( '\n' );
+            switch( parse.FullToken.FullImmediate.Immediate.DataType ) {
+            case TGSI_IMM_FLOAT32:
+               TXT( "\nFloat: " );
+               FLT( parse.FullToken.FullImmediate.u.ImmediateFloat32[i].Float );
+               break;
+
+            default:
+               assert( 0 );
+            }
+         }
+         break;
+
+      case TGSI_TOKEN_TYPE_INSTRUCTION:
+         {
+            struct tgsi_full_instruction *inst = &parse.FullToken.FullInstruction;
+
+            TXT( "\nOpcode     : " );
+            ENM( inst->Instruction.Opcode, TGSI_OPCODES );
+            if( deflt || fi.Instruction.Saturate != inst->Instruction.Saturate ) {
+               TXT( "\nSaturate   : " );
+               ENM( inst->Instruction.Saturate, TGSI_SATS );
+            }
+            if( deflt || fi.Instruction.NumDstRegs != inst->Instruction.NumDstRegs ) {
+               TXT( "\nNumDstRegs : " );
+               UID( inst->Instruction.NumDstRegs );
+            }
+            if( deflt || fi.Instruction.NumSrcRegs != inst->Instruction.NumSrcRegs ) {
+               TXT( "\nNumSrcRegs : " );
+               UID( inst->Instruction.NumSrcRegs );
+            }
+            if( ignored ) {
+               TXT( "\nPadding    : " );
+               UIX( inst->Instruction.Padding );
+            }
+
+            if( deflt || tgsi_compare_instruction_ext_nv( inst->InstructionExtNv, fi.InstructionExtNv)) {
+               CHR( '\n' );
+               TXT( "\nType          : " );
+               ENM( inst->InstructionExtNv.Type, TGSI_INSTRUCTION_EXTS );
+               if( deflt || fi.InstructionExtNv.Precision != inst->InstructionExtNv.Precision ) {
+                  TXT( "\nPrecision     : " );
+                  ENM( inst->InstructionExtNv.Precision, TGSI_PRECISIONS );
+               }
+               if( deflt || fi.InstructionExtNv.CondDstIndex != inst->InstructionExtNv.CondDstIndex ) {
+                  TXT( "\nCondDstIndex  : " );
+                  UID( inst->InstructionExtNv.CondDstIndex );
+               }
+               if( deflt || fi.InstructionExtNv.CondFlowIndex != inst->InstructionExtNv.CondFlowIndex ) {
+                  TXT( "\nCondFlowIndex : " );
+                  UID( inst->InstructionExtNv.CondFlowIndex );
+               }
+               if( deflt || fi.InstructionExtNv.CondMask != inst->InstructionExtNv.CondMask ) {
+                  TXT( "\nCondMask      : " );
+                  ENM( inst->InstructionExtNv.CondMask, TGSI_CCS );
+               }
+               if( deflt || fi.InstructionExtNv.CondSwizzleX != inst->InstructionExtNv.CondSwizzleX ) {
+                  TXT( "\nCondSwizzleX  : " );
+                  ENM( inst->InstructionExtNv.CondSwizzleX, TGSI_SWIZZLES );
+               }
+               if( deflt || fi.InstructionExtNv.CondSwizzleY != inst->InstructionExtNv.CondSwizzleY ) {
+                  TXT( "\nCondSwizzleY  : " );
+                  ENM( inst->InstructionExtNv.CondSwizzleY, TGSI_SWIZZLES );
+               }
+               if( deflt || fi.InstructionExtNv.CondSwizzleZ != inst->InstructionExtNv.CondSwizzleZ ) {
+                  TXT( "\nCondSwizzleZ  : " );
+                  ENM( inst->InstructionExtNv.CondSwizzleZ, TGSI_SWIZZLES );
+               }
+               if( deflt || fi.InstructionExtNv.CondSwizzleW != inst->InstructionExtNv.CondSwizzleW ) {
+                  TXT( "\nCondSwizzleW  : " );
+                  ENM( inst->InstructionExtNv.CondSwizzleW, TGSI_SWIZZLES );
+               }
+               if( deflt || fi.InstructionExtNv.CondDstUpdate != inst->InstructionExtNv.CondDstUpdate ) {
+                  TXT( "\nCondDstUpdate : " );
+                  UID( inst->InstructionExtNv.CondDstUpdate );
+               }
+               if( deflt || fi.InstructionExtNv.CondFlowEnable != inst->InstructionExtNv.CondFlowEnable ) {
+                  TXT( "\nCondFlowEnable: " );
+                  UID( inst->InstructionExtNv.CondFlowEnable );
+               }
+               if( ignored ) {
+                  TXT( "\nPadding       : " );
+                  UIX( inst->InstructionExtNv.Padding );
+                  if( deflt || fi.InstructionExtNv.Extended != inst->InstructionExtNv.Extended ) {
+                     TXT( "\nExtended      : " );
+                     UID( inst->InstructionExtNv.Extended );
+                  }
+               }
+            }
+
+            if( deflt || tgsi_compare_instruction_ext_label( inst->InstructionExtLabel, fi.InstructionExtLabel ) ) {
+               CHR( '\n' );
+               TXT( "\nType    : " );
+               ENM( inst->InstructionExtLabel.Type, TGSI_INSTRUCTION_EXTS );
+               if( deflt || fi.InstructionExtLabel.Label != inst->InstructionExtLabel.Label ) {
+                  TXT( "\nLabel   : " );
+                  UID( inst->InstructionExtLabel.Label );
+               }
+               if( deflt || fi.InstructionExtLabel.Target != inst->InstructionExtLabel.Target ) {
+                  TXT( "\nTarget  : " );
+                  UID( inst->InstructionExtLabel.Target );
+               }
+               if( ignored ) {
+                  TXT( "\nPadding : " );
+                  UIX( inst->InstructionExtLabel.Padding );
+                  if( deflt || fi.InstructionExtLabel.Extended != inst->InstructionExtLabel.Extended ) {
+                     TXT( "\nExtended: " );
+                     UID( inst->InstructionExtLabel.Extended );
+                  }
+               }
+            }
+
+            if( deflt || tgsi_compare_instruction_ext_texture( inst->InstructionExtTexture, fi.InstructionExtTexture ) ) {
+               CHR( '\n' );
+               TXT( "\nType    : " );
+               ENM( inst->InstructionExtTexture.Type, TGSI_INSTRUCTION_EXTS );
+               if( deflt || fi.InstructionExtTexture.Texture != inst->InstructionExtTexture.Texture ) {
+                  TXT( "\nTexture : " );
+                  ENM( inst->InstructionExtTexture.Texture, TGSI_TEXTURES );
+               }
+               if( ignored ) {
+                  TXT( "\nPadding : " );
+                  UIX( inst->InstructionExtTexture.Padding );
+                  if( deflt || fi.InstructionExtTexture.Extended != inst->InstructionExtTexture.Extended ) {
+                     TXT( "\nExtended: " );
+                     UID( inst->InstructionExtTexture.Extended );
+                  }
+               }
+            }
+
+            for( i = 0; i < inst->Instruction.NumDstRegs; i++ ) {
+               struct tgsi_full_dst_register *dst = &inst->FullDstRegisters[i];
+               struct tgsi_full_dst_register *fd = &fi.FullDstRegisters[i];
+
+               CHR( '\n' );
+               TXT( "\nFile     : " );
+               ENM( dst->DstRegister.File, TGSI_FILES );
+               if( deflt || fd->DstRegister.WriteMask != dst->DstRegister.WriteMask ) {
+                  TXT( "\nWriteMask: " );
+                  ENM( dst->DstRegister.WriteMask, TGSI_WRITEMASKS );
+               }
+               if( ignored ) {
+                  if( deflt || fd->DstRegister.Indirect != dst->DstRegister.Indirect ) {
+                     TXT( "\nIndirect : " );
+                     UID( dst->DstRegister.Indirect );
+                  }
+                  if( deflt || fd->DstRegister.Dimension != dst->DstRegister.Dimension ) {
+                     TXT( "\nDimension: " );
+                     UID( dst->DstRegister.Dimension );
+                  }
+               }
+               if( deflt || fd->DstRegister.Index != dst->DstRegister.Index ) {
+                  TXT( "\nIndex    : " );
+                  SID( dst->DstRegister.Index );
+               }
+               if( ignored ) {
+                  TXT( "\nPadding  : " );
+                  UIX( dst->DstRegister.Padding );
+                  if( deflt || fd->DstRegister.Extended != dst->DstRegister.Extended ) {
+                     TXT( "\nExtended : " );
+                     UID( dst->DstRegister.Extended );
+                  }
+               }
+
+               if( deflt || tgsi_compare_dst_register_ext_concode( dst->DstRegisterExtConcode, fd->DstRegisterExtConcode ) ) {
+                  CHR( '\n' );
+                  TXT( "\nType        : " );
+                  ENM( dst->DstRegisterExtConcode.Type, TGSI_DST_REGISTER_EXTS );
+                  if( deflt || fd->DstRegisterExtConcode.CondMask != dst->DstRegisterExtConcode.CondMask ) {
+                     TXT( "\nCondMask    : " );
+                     ENM( dst->DstRegisterExtConcode.CondMask, TGSI_CCS );
+                  }
+                  if( deflt || fd->DstRegisterExtConcode.CondSwizzleX != dst->DstRegisterExtConcode.CondSwizzleX ) {
+                     TXT( "\nCondSwizzleX: " );
+                     ENM( dst->DstRegisterExtConcode.CondSwizzleX, TGSI_SWIZZLES );
+                  }
+                  if( deflt || fd->DstRegisterExtConcode.CondSwizzleY != dst->DstRegisterExtConcode.CondSwizzleY ) {
+                     TXT( "\nCondSwizzleY: " );
+                     ENM( dst->DstRegisterExtConcode.CondSwizzleY, TGSI_SWIZZLES );
+                  }
+                  if( deflt || fd->DstRegisterExtConcode.CondSwizzleZ != dst->DstRegisterExtConcode.CondSwizzleZ ) {
+                     TXT( "\nCondSwizzleZ: " );
+                     ENM( dst->DstRegisterExtConcode.CondSwizzleZ, TGSI_SWIZZLES );
+                  }
+                  if( deflt || fd->DstRegisterExtConcode.CondSwizzleW != dst->DstRegisterExtConcode.CondSwizzleW ) {
+                     TXT( "\nCondSwizzleW: " );
+                     ENM( dst->DstRegisterExtConcode.CondSwizzleW, TGSI_SWIZZLES );
+                  }
+                  if( deflt || fd->DstRegisterExtConcode.CondSrcIndex != dst->DstRegisterExtConcode.CondSrcIndex ) {
+                     TXT( "\nCondSrcIndex: " );
+                     UID( dst->DstRegisterExtConcode.CondSrcIndex );
+                  }
+                  if( ignored ) {
+                     TXT( "\nPadding     : " );
+                     UIX( dst->DstRegisterExtConcode.Padding );
+                     if( deflt || fd->DstRegisterExtConcode.Extended != dst->DstRegisterExtConcode.Extended ) {
+                        TXT( "\nExtended    : " );
+                        UID( dst->DstRegisterExtConcode.Extended );
+                     }
+                  }
+               }
+
+               if( deflt || tgsi_compare_dst_register_ext_modulate( dst->DstRegisterExtModulate, fd->DstRegisterExtModulate ) ) {
+                  CHR( '\n' );
+                  TXT( "\nType    : " );
+                  ENM( dst->DstRegisterExtModulate.Type, TGSI_DST_REGISTER_EXTS );
+                  if( deflt || fd->DstRegisterExtModulate.Modulate != dst->DstRegisterExtModulate.Modulate ) {
+                     TXT( "\nModulate: " );
+                     ENM( dst->DstRegisterExtModulate.Modulate, TGSI_MODULATES );
+                  }
+                  if( ignored ) {
+                     TXT( "\nPadding : " );
+                     UIX( dst->DstRegisterExtModulate.Padding );
+                     if( deflt || fd->DstRegisterExtModulate.Extended != dst->DstRegisterExtModulate.Extended ) {
+                        TXT( "\nExtended: " );
+                        UID( dst->DstRegisterExtModulate.Extended );
+                     }
+                  }
+               }
+            }
+
+            for( i = 0; i < inst->Instruction.NumSrcRegs; i++ ) {
+               struct tgsi_full_src_register *src = &inst->FullSrcRegisters[i];
+               struct tgsi_full_src_register *fs = &fi.FullSrcRegisters[i];
+
+               CHR( '\n' );
+               TXT( "\nFile     : ");
+               ENM( src->SrcRegister.File, TGSI_FILES );
+               if( deflt || fs->SrcRegister.SwizzleX != src->SrcRegister.SwizzleX ) {
+                  TXT( "\nSwizzleX : " );
+                  ENM( src->SrcRegister.SwizzleX, TGSI_SWIZZLES );
+               }
+               if( deflt || fs->SrcRegister.SwizzleY != src->SrcRegister.SwizzleY ) {
+                  TXT( "\nSwizzleY : " );
+                  ENM( src->SrcRegister.SwizzleY, TGSI_SWIZZLES );
+               }
+               if( deflt || fs->SrcRegister.SwizzleZ != src->SrcRegister.SwizzleZ ) {
+                  TXT( "\nSwizzleZ : " );
+                  ENM( src->SrcRegister.SwizzleZ, TGSI_SWIZZLES );
+               }
+               if( deflt || fs->SrcRegister.SwizzleW != src->SrcRegister.SwizzleW ) {
+                  TXT( "\nSwizzleW : " );
+                  ENM( src->SrcRegister.SwizzleW, TGSI_SWIZZLES );
+               }
+               if( deflt || fs->SrcRegister.Negate != src->SrcRegister.Negate ) {
+                  TXT( "\nNegate   : " );
+                  UID( src->SrcRegister.Negate );
+               }
+               if( ignored ) {
+                  if( deflt || fs->SrcRegister.Indirect != src->SrcRegister.Indirect ) {
+                     TXT( "\nIndirect : " );
+                     UID( src->SrcRegister.Indirect );
+                  }
+                  if( deflt || fs->SrcRegister.Dimension != src->SrcRegister.Dimension ) {
+                     TXT( "\nDimension: " );
+                     UID( src->SrcRegister.Dimension );
+                  }
+               }
+               if( deflt || fs->SrcRegister.Index != src->SrcRegister.Index ) {
+                  TXT( "\nIndex    : " );
+                  SID( src->SrcRegister.Index );
+               }
+               if( ignored ) {
+                  if( deflt || fs->SrcRegister.Extended != src->SrcRegister.Extended ) {
+                     TXT( "\nExtended : " );
+                     UID( src->SrcRegister.Extended );
+                  }
+               }
+
+               if( deflt || tgsi_compare_src_register_ext_swz( src->SrcRegisterExtSwz, fs->SrcRegisterExtSwz ) ) {
+                  CHR( '\n' );
+                  TXT( "\nType       : " );
+                  ENM( src->SrcRegisterExtSwz.Type, TGSI_SRC_REGISTER_EXTS );
+                  if( deflt || fs->SrcRegisterExtSwz.ExtSwizzleX != src->SrcRegisterExtSwz.ExtSwizzleX ) {
+                     TXT( "\nExtSwizzleX: " );
+                     ENM( src->SrcRegisterExtSwz.ExtSwizzleX, TGSI_EXTSWIZZLES );
+                  }
+                  if( deflt || fs->SrcRegisterExtSwz.ExtSwizzleY != src->SrcRegisterExtSwz.ExtSwizzleY ) {
+                     TXT( "\nExtSwizzleY: " );
+                     ENM( src->SrcRegisterExtSwz.ExtSwizzleY, TGSI_EXTSWIZZLES );
+                  }
+                  if( deflt || fs->SrcRegisterExtSwz.ExtSwizzleZ != src->SrcRegisterExtSwz.ExtSwizzleZ ) {
+                     TXT( "\nExtSwizzleZ: " );
+                     ENM( src->SrcRegisterExtSwz.ExtSwizzleZ, TGSI_EXTSWIZZLES );
+                  }
+                  if( deflt || fs->SrcRegisterExtSwz.ExtSwizzleW != src->SrcRegisterExtSwz.ExtSwizzleW ) {
+                     TXT( "\nExtSwizzleW: " );
+                     ENM( src->SrcRegisterExtSwz.ExtSwizzleW, TGSI_EXTSWIZZLES );
+                  }
+                  if( deflt || fs->SrcRegisterExtSwz.NegateX != src->SrcRegisterExtSwz.NegateX ) {
+                     TXT( "\nNegateX   : " );
+                     UID( src->SrcRegisterExtSwz.NegateX );
+                  }
+                  if( deflt || fs->SrcRegisterExtSwz.NegateY != src->SrcRegisterExtSwz.NegateY ) {
+                     TXT( "\nNegateY   : " );
+                     UID( src->SrcRegisterExtSwz.NegateY );
+                  }
+                  if( deflt || fs->SrcRegisterExtSwz.NegateZ != src->SrcRegisterExtSwz.NegateZ ) {
+                     TXT( "\nNegateZ   : " );
+                     UID( src->SrcRegisterExtSwz.NegateZ );
+                  }
+                  if( deflt || fs->SrcRegisterExtSwz.NegateW != src->SrcRegisterExtSwz.NegateW ) {
+                     TXT( "\nNegateW   : " );
+                     UID( src->SrcRegisterExtSwz.NegateW );
+                  }
+                  if( deflt || fs->SrcRegisterExtSwz.ExtDivide != src->SrcRegisterExtSwz.ExtDivide ) {
+                     TXT( "\nExtDivide  : " );
+                     ENM( src->SrcRegisterExtSwz.ExtDivide, TGSI_EXTSWIZZLES );
+                  }
+                  if( ignored ) {
+                     TXT( "\nPadding   : " );
+                     UIX( src->SrcRegisterExtSwz.Padding );
+                     if( deflt || fs->SrcRegisterExtSwz.Extended != src->SrcRegisterExtSwz.Extended ) {
+                        TXT( "\nExtended   : " );
+                        UID( src->SrcRegisterExtSwz.Extended );
+                     }
+                  }
+               }
+
+               if( deflt || tgsi_compare_src_register_ext_mod( src->SrcRegisterExtMod, fs->SrcRegisterExtMod ) ) {
+                  CHR( '\n' );
+                  TXT( "\nType     : " );
+                  ENM( src->SrcRegisterExtMod.Type, TGSI_SRC_REGISTER_EXTS );
+                  if( deflt || fs->SrcRegisterExtMod.Complement != src->SrcRegisterExtMod.Complement ) {
+                     TXT( "\nComplement: " );
+                     UID( src->SrcRegisterExtMod.Complement );
+                  }
+                  if( deflt || fs->SrcRegisterExtMod.Bias != src->SrcRegisterExtMod.Bias ) {
+                     TXT( "\nBias     : " );
+                     UID( src->SrcRegisterExtMod.Bias );
+                  }
+                  if( deflt || fs->SrcRegisterExtMod.Scale2X != src->SrcRegisterExtMod.Scale2X ) {
+                     TXT( "\nScale2X   : " );
+                     UID( src->SrcRegisterExtMod.Scale2X );
+                  }
+                  if( deflt || fs->SrcRegisterExtMod.Absolute != src->SrcRegisterExtMod.Absolute ) {
+                     TXT( "\nAbsolute  : " );
+                     UID( src->SrcRegisterExtMod.Absolute );
+                  }
+                  if( deflt || fs->SrcRegisterExtMod.Negate != src->SrcRegisterExtMod.Negate ) {
+                     TXT( "\nNegate   : " );
+                     UID( src->SrcRegisterExtMod.Negate );
+                  }
+                  if( ignored ) {
+                     TXT( "\nPadding   : " );
+                     UIX( src->SrcRegisterExtMod.Padding );
+                     if( deflt || fs->SrcRegisterExtMod.Extended != src->SrcRegisterExtMod.Extended ) {
+                        TXT( "\nExtended  : " );
+                        UID( src->SrcRegisterExtMod.Extended );
+                     }
+                  }
+               }
+            }
+         }
+         break;
+
+      default:
+         assert( 0 );
+      }
+   }
+
+   tgsi_parse_free( &parse );
+
+   fclose( dump.file );
+}
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_dump.h b/src/mesa/pipe/tgsi/core/tgsi_dump.h
new file mode 100644
index 00000000000..dc34a840de6
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_dump.h
@@ -0,0 +1,22 @@
+#if !defined TGSI_DUMP_H
+#define TGSI_DUMP_H
+
+#if defined __cplusplus
+extern "C" {
+#endif // defined __cplusplus
+
+#define TGSI_DUMP_VERBOSE       0
+#define TGSI_DUMP_NO_IGNORED    1
+#define TGSI_DUMP_NO_DEFAULT    2
+
+void
+tgsi_dump(
+   const struct tgsi_token *tokens,
+   GLuint flags );
+
+#if defined __cplusplus
+} // extern "C"
+#endif // defined __cplusplus
+
+#endif // !defined TGSI_DUMP_H
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_exec.c b/src/mesa/pipe/tgsi/core/tgsi_exec.c
new file mode 100644
index 00000000000..189a41167a9
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_exec.c
@@ -0,0 +1,2193 @@
+#include "tgsi_platform.h"
+#include "tgsi_core.h"
+
+#define TILE_BOTTOM_LEFT  0
+#define TILE_BOTTOM_RIGHT 1
+#define TILE_TOP_LEFT     2
+#define TILE_TOP_RIGHT    3
+
+#define TEMP_0_I           TGSI_EXEC_TEMP_00000000_I
+#define TEMP_0_C           TGSI_EXEC_TEMP_00000000_C
+#define TEMP_7F_I          TGSI_EXEC_TEMP_7FFFFFFF_I
+#define TEMP_7F_C          TGSI_EXEC_TEMP_7FFFFFFF_C
+#define TEMP_80_I          TGSI_EXEC_TEMP_80000000_I
+#define TEMP_80_C          TGSI_EXEC_TEMP_80000000_C
+#define TEMP_FF_I          TGSI_EXEC_TEMP_FFFFFFFF_I
+#define TEMP_FF_C          TGSI_EXEC_TEMP_FFFFFFFF_C
+#define TEMP_1_I           TGSI_EXEC_TEMP_ONE_I
+#define TEMP_1_C           TGSI_EXEC_TEMP_ONE_C
+#define TEMP_2_I           TGSI_EXEC_TEMP_TWO_I
+#define TEMP_2_C           TGSI_EXEC_TEMP_TWO_C
+#define TEMP_128_I         TGSI_EXEC_TEMP_128_I
+#define TEMP_128_C         TGSI_EXEC_TEMP_128_C
+#define TEMP_M128_I        TGSI_EXEC_TEMP_MINUS_128_I
+#define TEMP_M128_C        TGSI_EXEC_TEMP_MINUS_128_C
+#define TEMP_KILMASK_I     TGSI_EXEC_TEMP_KILMASK_I
+#define TEMP_KILMASK_C     TGSI_EXEC_TEMP_KILMASK_C
+#define TEMP_OUTPUT_I      TGSI_EXEC_TEMP_OUTPUT_I
+#define TEMP_OUTPUT_C      TGSI_EXEC_TEMP_OUTPUT_C
+#define TEMP_PRIMITIVE_I   TGSI_EXEC_TEMP_PRIMITIVE_I
+#define TEMP_PRIMITIVE_C   TGSI_EXEC_TEMP_PRIMITIVE_C
+#define TEMP_R0            TGSI_EXEC_TEMP_R0
+
+#define FOR_EACH_CHANNEL(CHAN)\
+   for (CHAN = 0; CHAN < 4; CHAN++)
+
+#define IS_CHANNEL_ENABLED(INST, CHAN)\
+   ((INST).FullDstRegisters[0].DstRegister.WriteMask & (1 << (CHAN)))
+
+#define IS_CHANNEL_ENABLED2(INST, CHAN)\
+   ((INST).FullDstRegisters[1].DstRegister.WriteMask & (1 << (CHAN)))
+
+#define FOR_EACH_ENABLED_CHANNEL(INST, CHAN)\
+   FOR_EACH_CHANNEL( CHAN )\
+      if (IS_CHANNEL_ENABLED( INST, CHAN ))
+
+#define FOR_EACH_ENABLED_CHANNEL2(INST, CHAN)\
+   FOR_EACH_CHANNEL( CHAN )\
+      if (IS_CHANNEL_ENABLED2( INST, CHAN ))
+
+#define CHAN_X  0
+#define CHAN_Y  1
+#define CHAN_Z  2
+#define CHAN_W  3
+
+void
+tgsi_exec_machine_init(
+   struct tgsi_exec_machine *mach,
+   struct tgsi_token *tokens )
+{
+   GLuint i;
+   struct tgsi_parse_context parse;
+
+   mach->Tokens = tokens;
+
+   tgsi_parse_init (&parse, mach->Tokens);
+   mach->Processor = parse.FullHeader.Processor.Processor;
+   tgsi_parse_free (&parse);
+
+   mach->Temps = (struct tgsi_exec_vector *) tgsi_align_128bit( mach->_Temps);
+   mach->Addrs = &mach->Temps[TGSI_EXEC_NUM_TEMPS];
+
+#if XXX_SSE
+    tgsi_emit_sse (tokens,
+                   &mach->Function);
+#endif
+
+   /* Setup constants. */
+   for( i = 0; i < 4; i++ ) {
+      mach->Temps[TEMP_0_I].xyzw[TEMP_0_C].u[i] = 0x00000000;
+      mach->Temps[TEMP_7F_I].xyzw[TEMP_7F_C].u[i] = 0x7FFFFFFF;
+      mach->Temps[TEMP_80_I].xyzw[TEMP_80_C].u[i] = 0x80000000;
+      mach->Temps[TEMP_FF_I].xyzw[TEMP_FF_C].u[i] = 0xFFFFFFFF;
+      mach->Temps[TEMP_1_I].xyzw[TEMP_1_C].f[i] = 1.0f;
+      mach->Temps[TEMP_2_I].xyzw[TEMP_2_C].f[i] = 2.0f;
+      mach->Temps[TEMP_128_I].xyzw[TEMP_128_C].f[i] = 128.0f;
+      mach->Temps[TEMP_M128_I].xyzw[TEMP_M128_C].f[i] = -128.0f;
+   }
+}
+
+void
+tgsi_exec_prepare(
+   struct tgsi_exec_machine *mach,
+   struct tgsi_exec_labels *labels )
+{
+   struct tgsi_parse_context parse;
+
+   mach->ImmLimit = 0;
+   labels->count = 0;
+
+   tgsi_parse_init( &parse, mach->Tokens );
+   while( !tgsi_parse_end_of_tokens( &parse ) ) {
+      GLuint pointer = parse.Position;
+      GLuint i;
+      tgsi_parse_token( &parse );
+      switch( parse.FullToken.Token.Type ) {
+      case TGSI_TOKEN_TYPE_DECLARATION:
+         break;
+      case TGSI_TOKEN_TYPE_IMMEDIATE:
+         assert( (parse.FullToken.FullImmediate.Immediate.Size - 1) % 4 == 0 );
+         assert( mach->ImmLimit + (parse.FullToken.FullImmediate.Immediate.Size - 1) / 4 <= 256 );
+         for( i = 0; i < parse.FullToken.FullImmediate.Immediate.Size - 1; i++ ) {
+            mach->Imms[mach->ImmLimit + i / 4][i % 4] = parse.FullToken.FullImmediate.u.ImmediateFloat32[i].Float;
+         }
+         mach->ImmLimit += (parse.FullToken.FullImmediate.Immediate.Size - 1) / 4;
+         break;
+      case TGSI_TOKEN_TYPE_INSTRUCTION:
+         if( parse.FullToken.FullInstruction.InstructionExtLabel.Label &&
+             parse.FullToken.FullInstruction.InstructionExtLabel.Target ) {
+            assert( labels->count < 128 );
+            labels->labels[labels->count][0] = parse.FullToken.FullInstruction.InstructionExtLabel.Label;
+            labels->labels[labels->count][1] = pointer;
+            labels->count++;
+         }
+         break;
+      default:
+         assert( 0 );
+      }
+   }
+   tgsi_parse_free (&parse);
+}
+
+void
+tgsi_exec_machine_run(
+   struct tgsi_exec_machine *mach )
+{
+   struct tgsi_exec_labels labels;
+
+   tgsi_exec_prepare( mach, &labels );
+   tgsi_exec_machine_run2( mach, &labels );
+}
+
+static void
+micro_abs(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = (GLfloat) fabs( (GLdouble) src->f[0] );
+   dst->f[1] = (GLfloat) fabs( (GLdouble) src->f[1] );
+   dst->f[2] = (GLfloat) fabs( (GLdouble) src->f[2] );
+   dst->f[3] = (GLfloat) fabs( (GLdouble) src->f[3] );
+}
+
+static void
+micro_add(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] + src1->f[0];
+   dst->f[1] = src0->f[1] + src1->f[1];
+   dst->f[2] = src0->f[2] + src1->f[2];
+   dst->f[3] = src0->f[3] + src1->f[3];
+}
+
+static void
+micro_iadd(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] + src1->i[0];
+   dst->i[1] = src0->i[1] + src1->i[1];
+   dst->i[2] = src0->i[2] + src1->i[2];
+   dst->i[3] = src0->i[3] + src1->i[3];
+}
+
+static void
+micro_and(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] & src1->u[0];
+   dst->u[1] = src0->u[1] & src1->u[1];
+   dst->u[2] = src0->u[2] & src1->u[2];
+   dst->u[3] = src0->u[3] & src1->u[3];
+}
+
+static void
+micro_ceil(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = (GLfloat) ceil( (GLdouble) src->f[0] );
+   dst->f[1] = (GLfloat) ceil( (GLdouble) src->f[1] );
+   dst->f[2] = (GLfloat) ceil( (GLdouble) src->f[2] );
+   dst->f[3] = (GLfloat) ceil( (GLdouble) src->f[3] );
+}
+
+static void
+micro_cos(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = (GLfloat) cos( (GLdouble) src->f[0] );
+   dst->f[1] = (GLfloat) cos( (GLdouble) src->f[1] );
+   dst->f[2] = (GLfloat) cos( (GLdouble) src->f[2] );
+   dst->f[3] = (GLfloat) cos( (GLdouble) src->f[3] );
+}
+
+static void
+micro_div(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] / src1->f[0];
+   dst->f[1] = src0->f[1] / src1->f[1];
+   dst->f[2] = src0->f[2] / src1->f[2];
+   dst->f[3] = src0->f[3] / src1->f[3];
+}
+
+static void
+micro_udiv(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] / src1->u[0];
+   dst->u[1] = src0->u[1] / src1->u[1];
+   dst->u[2] = src0->u[2] / src1->u[2];
+   dst->u[3] = src0->u[3] / src1->u[3];
+}
+
+static void
+micro_eq(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1,
+   const union tgsi_exec_channel *src2,
+   const union tgsi_exec_channel *src3 )
+{
+   dst->f[0] = src0->f[0] == src1->f[0] ? src2->f[0] : src3->f[0];
+   dst->f[1] = src0->f[1] == src1->f[1] ? src2->f[1] : src3->f[1];
+   dst->f[2] = src0->f[2] == src1->f[2] ? src2->f[2] : src3->f[2];
+   dst->f[3] = src0->f[3] == src1->f[3] ? src2->f[3] : src3->f[3];
+}
+
+static void
+micro_ieq(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1,
+   const union tgsi_exec_channel *src2,
+   const union tgsi_exec_channel *src3 )
+{
+   dst->i[0] = src0->i[0] == src1->i[0] ? src2->i[0] : src3->i[0];
+   dst->i[1] = src0->i[1] == src1->i[1] ? src2->i[1] : src3->i[1];
+   dst->i[2] = src0->i[2] == src1->i[2] ? src2->i[2] : src3->i[2];
+   dst->i[3] = src0->i[3] == src1->i[3] ? src2->i[3] : src3->i[3];
+}
+
+static void
+micro_exp2(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = (GLfloat) pow( 2.0, (GLdouble) src->f[0] );
+   dst->f[1] = (GLfloat) pow( 2.0, (GLdouble) src->f[1] );
+   dst->f[2] = (GLfloat) pow( 2.0, (GLdouble) src->f[2] );
+   dst->f[3] = (GLfloat) pow( 2.0, (GLdouble) src->f[3] );
+}
+
+static void
+micro_f2it(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->i[0] = (GLint) src->f[0];
+   dst->i[1] = (GLint) src->f[1];
+   dst->i[2] = (GLint) src->f[2];
+   dst->i[3] = (GLint) src->f[3];
+}
+
+static void
+micro_f2ut(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->u[0] = (GLuint) src->f[0];
+   dst->u[1] = (GLuint) src->f[1];
+   dst->u[2] = (GLuint) src->f[2];
+   dst->u[3] = (GLuint) src->f[3];
+}
+
+static void
+micro_flr(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = (GLfloat) floor( (GLdouble) src->f[0] );
+   dst->f[1] = (GLfloat) floor( (GLdouble) src->f[1] );
+   dst->f[2] = (GLfloat) floor( (GLdouble) src->f[2] );
+   dst->f[3] = (GLfloat) floor( (GLdouble) src->f[3] );
+}
+
+static void
+micro_frc(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = src->f[0] - (GLfloat) floor( (GLdouble) src->f[0] );
+   dst->f[1] = src->f[1] - (GLfloat) floor( (GLdouble) src->f[1] );
+   dst->f[2] = src->f[2] - (GLfloat) floor( (GLdouble) src->f[2] );
+   dst->f[3] = src->f[3] - (GLfloat) floor( (GLdouble) src->f[3] );
+}
+
+static void
+micro_i2f(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = (GLfloat) src->i[0];
+   dst->f[1] = (GLfloat) src->i[1];
+   dst->f[2] = (GLfloat) src->i[2];
+   dst->f[3] = (GLfloat) src->i[3];
+}
+
+static void
+micro_lg2(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = (GLfloat) log( (GLdouble) src->f[0] ) * 1.442695f;
+   dst->f[1] = (GLfloat) log( (GLdouble) src->f[1] ) * 1.442695f;
+   dst->f[2] = (GLfloat) log( (GLdouble) src->f[2] ) * 1.442695f;
+   dst->f[3] = (GLfloat) log( (GLdouble) src->f[3] ) * 1.442695f;
+}
+
+static void
+micro_lt(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1,
+   const union tgsi_exec_channel *src2,
+   const union tgsi_exec_channel *src3 )
+{
+   dst->f[0] = src0->f[0] < src1->f[0] ? src2->f[0] : src3->f[0];
+   dst->f[1] = src0->f[1] < src1->f[1] ? src2->f[1] : src3->f[1];
+   dst->f[2] = src0->f[2] < src1->f[2] ? src2->f[2] : src3->f[2];
+   dst->f[3] = src0->f[3] < src1->f[3] ? src2->f[3] : src3->f[3];
+}
+
+static void
+micro_ilt(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1,
+   const union tgsi_exec_channel *src2,
+   const union tgsi_exec_channel *src3 )
+{
+   dst->i[0] = src0->i[0] < src1->i[0] ? src2->i[0] : src3->i[0];
+   dst->i[1] = src0->i[1] < src1->i[1] ? src2->i[1] : src3->i[1];
+   dst->i[2] = src0->i[2] < src1->i[2] ? src2->i[2] : src3->i[2];
+   dst->i[3] = src0->i[3] < src1->i[3] ? src2->i[3] : src3->i[3];
+}
+
+static void
+micro_ult(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1,
+   const union tgsi_exec_channel *src2,
+   const union tgsi_exec_channel *src3 )
+{
+   dst->u[0] = src0->u[0] < src1->u[0] ? src2->u[0] : src3->u[0];
+   dst->u[1] = src0->u[1] < src1->u[1] ? src2->u[1] : src3->u[1];
+   dst->u[2] = src0->u[2] < src1->u[2] ? src2->u[2] : src3->u[2];
+   dst->u[3] = src0->u[3] < src1->u[3] ? src2->u[3] : src3->u[3];
+}
+
+static void
+micro_max(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] > src1->f[0] ? src0->f[0] : src1->f[0];
+   dst->f[1] = src0->f[1] > src1->f[1] ? src0->f[1] : src1->f[1];
+   dst->f[2] = src0->f[2] > src1->f[2] ? src0->f[2] : src1->f[2];
+   dst->f[3] = src0->f[3] > src1->f[3] ? src0->f[3] : src1->f[3];
+}
+
+static void
+micro_imax(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] > src1->i[0] ? src0->i[0] : src1->i[0];
+   dst->i[1] = src0->i[1] > src1->i[1] ? src0->i[1] : src1->i[1];
+   dst->i[2] = src0->i[2] > src1->i[2] ? src0->i[2] : src1->i[2];
+   dst->i[3] = src0->i[3] > src1->i[3] ? src0->i[3] : src1->i[3];
+}
+
+static void
+micro_umax(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] > src1->u[0] ? src0->u[0] : src1->u[0];
+   dst->u[1] = src0->u[1] > src1->u[1] ? src0->u[1] : src1->u[1];
+   dst->u[2] = src0->u[2] > src1->u[2] ? src0->u[2] : src1->u[2];
+   dst->u[3] = src0->u[3] > src1->u[3] ? src0->u[3] : src1->u[3];
+}
+
+static void
+micro_min(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] < src1->f[0] ? src0->f[0] : src1->f[0];
+   dst->f[1] = src0->f[1] < src1->f[1] ? src0->f[1] : src1->f[1];
+   dst->f[2] = src0->f[2] < src1->f[2] ? src0->f[2] : src1->f[2];
+   dst->f[3] = src0->f[3] < src1->f[3] ? src0->f[3] : src1->f[3];
+}
+
+static void
+micro_imin(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] < src1->i[0] ? src0->i[0] : src1->i[0];
+   dst->i[1] = src0->i[1] < src1->i[1] ? src0->i[1] : src1->i[1];
+   dst->i[2] = src0->i[2] < src1->i[2] ? src0->i[2] : src1->i[2];
+   dst->i[3] = src0->i[3] < src1->i[3] ? src0->i[3] : src1->i[3];
+}
+
+static void
+micro_umin(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] < src1->u[0] ? src0->u[0] : src1->u[0];
+   dst->u[1] = src0->u[1] < src1->u[1] ? src0->u[1] : src1->u[1];
+   dst->u[2] = src0->u[2] < src1->u[2] ? src0->u[2] : src1->u[2];
+   dst->u[3] = src0->u[3] < src1->u[3] ? src0->u[3] : src1->u[3];
+}
+
+static void
+micro_umod(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] % src1->u[0];
+   dst->u[1] = src0->u[1] % src1->u[1];
+   dst->u[2] = src0->u[2] % src1->u[2];
+   dst->u[3] = src0->u[3] % src1->u[3];
+}
+
+static void
+micro_mul(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] * src1->f[0];
+   dst->f[1] = src0->f[1] * src1->f[1];
+   dst->f[2] = src0->f[2] * src1->f[2];
+   dst->f[3] = src0->f[3] * src1->f[3];
+}
+
+static void
+micro_imul(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] * src1->i[0];
+   dst->i[1] = src0->i[1] * src1->i[1];
+   dst->i[2] = src0->i[2] * src1->i[2];
+   dst->i[3] = src0->i[3] * src1->i[3];
+}
+
+static void
+micro_imul64(
+   union tgsi_exec_channel *dst0,
+   union tgsi_exec_channel *dst1,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst1->i[0] = src0->i[0] * src1->i[0];
+   dst1->i[1] = src0->i[1] * src1->i[1];
+   dst1->i[2] = src0->i[2] * src1->i[2];
+   dst1->i[3] = src0->i[3] * src1->i[3];
+   dst0->i[0] = 0;
+   dst0->i[1] = 0;
+   dst0->i[2] = 0;
+   dst0->i[3] = 0;
+}
+
+static void
+micro_umul64(
+   union tgsi_exec_channel *dst0,
+   union tgsi_exec_channel *dst1,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst1->u[0] = src0->u[0] * src1->u[0];
+   dst1->u[1] = src0->u[1] * src1->u[1];
+   dst1->u[2] = src0->u[2] * src1->u[2];
+   dst1->u[3] = src0->u[3] * src1->u[3];
+   dst0->u[0] = 0;
+   dst0->u[1] = 0;
+   dst0->u[2] = 0;
+   dst0->u[3] = 0;
+}
+
+static void
+micro_movc(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1,
+   const union tgsi_exec_channel *src2 )
+{
+   dst->u[0] = src0->u[0] ? src1->u[0] : src2->u[0];
+   dst->u[1] = src0->u[1] ? src1->u[1] : src2->u[1];
+   dst->u[2] = src0->u[2] ? src1->u[2] : src2->u[2];
+   dst->u[3] = src0->u[3] ? src1->u[3] : src2->u[3];
+}
+
+static void
+micro_neg(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = -src->f[0];
+   dst->f[1] = -src->f[1];
+   dst->f[2] = -src->f[2];
+   dst->f[3] = -src->f[3];
+}
+
+static void
+micro_ineg(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->i[0] = -src->i[0];
+   dst->i[1] = -src->i[1];
+   dst->i[2] = -src->i[2];
+   dst->i[3] = -src->i[3];
+}
+
+static void
+micro_not(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->u[0] = ~src->u[0];
+   dst->u[1] = ~src->u[1];
+   dst->u[2] = ~src->u[2];
+   dst->u[3] = ~src->u[3];
+}
+
+static void
+micro_or(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] | src1->u[0];
+   dst->u[1] = src0->u[1] | src1->u[1];
+   dst->u[2] = src0->u[2] | src1->u[2];
+   dst->u[3] = src0->u[3] | src1->u[3];
+}
+
+static void
+micro_pow(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->f[0] = (GLfloat) pow( (GLdouble) src0->f[0], (GLdouble) src1->f[0] );
+   dst->f[1] = (GLfloat) pow( (GLdouble) src0->f[1], (GLdouble) src1->f[1] );
+   dst->f[2] = (GLfloat) pow( (GLdouble) src0->f[2], (GLdouble) src1->f[2] );
+   dst->f[3] = (GLfloat) pow( (GLdouble) src0->f[3], (GLdouble) src1->f[3] );
+}
+
+static void
+micro_rnd(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = (GLfloat) floor( (GLdouble) (src->f[0] + 0.5f) );
+   dst->f[1] = (GLfloat) floor( (GLdouble) (src->f[1] + 0.5f) );
+   dst->f[2] = (GLfloat) floor( (GLdouble) (src->f[2] + 0.5f) );
+   dst->f[3] = (GLfloat) floor( (GLdouble) (src->f[3] + 0.5f) );
+}
+
+static void
+micro_shl(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] << src1->i[0];
+   dst->i[1] = src0->i[1] << src1->i[1];
+   dst->i[2] = src0->i[2] << src1->i[2];
+   dst->i[3] = src0->i[3] << src1->i[3];
+}
+
+static void
+micro_ishr(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->i[0] = src0->i[0] >> src1->i[0];
+   dst->i[1] = src0->i[1] >> src1->i[1];
+   dst->i[2] = src0->i[2] >> src1->i[2];
+   dst->i[3] = src0->i[3] >> src1->i[3];
+}
+
+static void
+micro_ushr(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] >> src1->u[0];
+   dst->u[1] = src0->u[1] >> src1->u[1];
+   dst->u[2] = src0->u[2] >> src1->u[2];
+   dst->u[3] = src0->u[3] >> src1->u[3];
+}
+
+static void
+micro_sin(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = (GLfloat) sin( (GLdouble) src->f[0] );
+   dst->f[1] = (GLfloat) sin( (GLdouble) src->f[1] );
+   dst->f[2] = (GLfloat) sin( (GLdouble) src->f[2] );
+   dst->f[3] = (GLfloat) sin( (GLdouble) src->f[3] );
+}
+
+static void
+micro_sqrt( union tgsi_exec_channel *dst,
+            const union tgsi_exec_channel *src )
+{
+   dst->f[0] = (GLfloat) sqrt( (GLdouble) src->f[0] );
+   dst->f[1] = (GLfloat) sqrt( (GLdouble) src->f[1] );
+   dst->f[2] = (GLfloat) sqrt( (GLdouble) src->f[2] );
+   dst->f[3] = (GLfloat) sqrt( (GLdouble) src->f[3] );
+}
+
+static void
+micro_sub(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->f[0] = src0->f[0] - src1->f[0];
+   dst->f[1] = src0->f[1] - src1->f[1];
+   dst->f[2] = src0->f[2] - src1->f[2];
+   dst->f[3] = src0->f[3] - src1->f[3];
+}
+
+static void
+micro_u2f(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src )
+{
+   dst->f[0] = (GLfloat) src->u[0];
+   dst->f[1] = (GLfloat) src->u[1];
+   dst->f[2] = (GLfloat) src->u[2];
+   dst->f[3] = (GLfloat) src->u[3];
+}
+
+static void
+micro_xor(
+   union tgsi_exec_channel *dst,
+   const union tgsi_exec_channel *src0,
+   const union tgsi_exec_channel *src1 )
+{
+   dst->u[0] = src0->u[0] ^ src1->u[0];
+   dst->u[1] = src0->u[1] ^ src1->u[1];
+   dst->u[2] = src0->u[2] ^ src1->u[2];
+   dst->u[3] = src0->u[3] ^ src1->u[3];
+}
+
+static void
+fetch_src_file_channel(
+   const struct tgsi_exec_machine *mach,
+   const GLuint file,
+   const GLuint swizzle,
+   const union tgsi_exec_channel *index,
+   union tgsi_exec_channel *chan )
+{
+   switch( swizzle ) {
+   case TGSI_EXTSWIZZLE_X:
+   case TGSI_EXTSWIZZLE_Y:
+   case TGSI_EXTSWIZZLE_Z:
+   case TGSI_EXTSWIZZLE_W:
+      switch( file ) {
+      case TGSI_FILE_CONSTANT:
+         chan->f[0] = mach->Consts[index->i[0]][swizzle];
+         chan->f[1] = mach->Consts[index->i[1]][swizzle];
+         chan->f[2] = mach->Consts[index->i[2]][swizzle];
+         chan->f[3] = mach->Consts[index->i[3]][swizzle];
+         break;
+
+      case TGSI_FILE_INPUT:
+         chan->u[0] = mach->Inputs[index->i[0]].xyzw[swizzle].u[0];
+         chan->u[1] = mach->Inputs[index->i[1]].xyzw[swizzle].u[1];
+         chan->u[2] = mach->Inputs[index->i[2]].xyzw[swizzle].u[2];
+         chan->u[3] = mach->Inputs[index->i[3]].xyzw[swizzle].u[3];
+         break;
+
+      case TGSI_FILE_TEMPORARY:
+         chan->u[0] = mach->Temps[index->i[0]].xyzw[swizzle].u[0];
+         chan->u[1] = mach->Temps[index->i[1]].xyzw[swizzle].u[1];
+         chan->u[2] = mach->Temps[index->i[2]].xyzw[swizzle].u[2];
+         chan->u[3] = mach->Temps[index->i[3]].xyzw[swizzle].u[3];
+         break;
+
+      case TGSI_FILE_IMMEDIATE:
+         assert( index->i[0] < (GLint) mach->ImmLimit );
+         chan->f[0] = mach->Imms[index->i[0]][swizzle];
+         assert( index->i[1] < (GLint) mach->ImmLimit );
+         chan->f[1] = mach->Imms[index->i[1]][swizzle];
+         assert( index->i[2] < (GLint) mach->ImmLimit );
+         chan->f[2] = mach->Imms[index->i[2]][swizzle];
+         assert( index->i[3] < (GLint) mach->ImmLimit );
+         chan->f[3] = mach->Imms[index->i[3]][swizzle];
+         break;
+
+      case TGSI_FILE_ADDRESS:
+         chan->u[0] = mach->Addrs[index->i[0]].xyzw[swizzle].u[0];
+         chan->u[1] = mach->Addrs[index->i[1]].xyzw[swizzle].u[1];
+         chan->u[2] = mach->Addrs[index->i[2]].xyzw[swizzle].u[2];
+         chan->u[3] = mach->Addrs[index->i[3]].xyzw[swizzle].u[3];
+         break;
+
+      default:
+         assert( 0 );
+      }
+      break;
+
+   case TGSI_EXTSWIZZLE_ZERO:
+      *chan = mach->Temps[TEMP_0_I].xyzw[TEMP_0_C];
+      break;
+
+   case TGSI_EXTSWIZZLE_ONE:
+      *chan = mach->Temps[TEMP_1_I].xyzw[TEMP_1_C];
+      break;
+
+   default:
+      assert( 0 );
+   }
+}
+
+static void
+fetch_source(
+   const struct tgsi_exec_machine *mach,
+   union tgsi_exec_channel *chan,
+   const struct tgsi_full_src_register *reg,
+   const GLuint chan_index )
+{
+   union tgsi_exec_channel index;
+   GLuint swizzle;
+
+   index.i[0] =
+   index.i[1] =
+   index.i[2] =
+   index.i[3] = reg->SrcRegister.Index;
+
+   if (reg->SrcRegister.Indirect) {
+      union tgsi_exec_channel index2;
+      union tgsi_exec_channel indir_index;
+
+      index2.i[0] =
+      index2.i[1] =
+      index2.i[2] =
+      index2.i[3] = reg->SrcRegisterInd.Index;
+
+      swizzle = tgsi_util_get_src_register_swizzle( &reg->SrcRegisterInd, CHAN_X );
+      fetch_src_file_channel(
+         mach,
+         reg->SrcRegisterInd.File,
+         swizzle,
+         &index2,
+         &indir_index );
+
+      index.i[0] += indir_index.i[0];
+      index.i[1] += indir_index.i[1];
+      index.i[2] += indir_index.i[2];
+      index.i[3] += indir_index.i[3];
+   }
+
+   if( reg->SrcRegister.Dimension ) {
+      switch( reg->SrcRegister.File ) {
+      case TGSI_FILE_INPUT:
+         index.i[0] *= 17;
+         index.i[1] *= 17;
+         index.i[2] *= 17;
+         index.i[3] *= 17;
+         break;
+      case TGSI_FILE_CONSTANT:
+         index.i[0] *= 4096;
+         index.i[1] *= 4096;
+         index.i[2] *= 4096;
+         index.i[3] *= 4096;
+         break;
+      default:
+         assert( 0 );
+      }
+
+      index.i[0] += reg->SrcRegisterDim.Index;
+      index.i[1] += reg->SrcRegisterDim.Index;
+      index.i[2] += reg->SrcRegisterDim.Index;
+      index.i[3] += reg->SrcRegisterDim.Index;
+
+      if (reg->SrcRegisterDim.Indirect) {
+         union tgsi_exec_channel index2;
+         union tgsi_exec_channel indir_index;
+
+         index2.i[0] =
+         index2.i[1] =
+         index2.i[2] =
+         index2.i[3] = reg->SrcRegisterDimInd.Index;
+
+         swizzle = tgsi_util_get_src_register_swizzle( &reg->SrcRegisterDimInd, CHAN_X );
+         fetch_src_file_channel(
+            mach,
+            reg->SrcRegisterDimInd.File,
+            swizzle,
+            &index2,
+            &indir_index );
+
+         index.i[0] += indir_index.i[0];
+         index.i[1] += indir_index.i[1];
+         index.i[2] += indir_index.i[2];
+         index.i[3] += indir_index.i[3];
+      }
+   }
+
+   swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index );
+   fetch_src_file_channel(
+      mach,
+      reg->SrcRegister.File,
+      swizzle,
+      &index,
+      chan );
+
+   switch (tgsi_util_get_full_src_register_sign_mode( reg, chan_index )) {
+   case TGSI_UTIL_SIGN_CLEAR:
+      micro_abs( chan, chan );
+      break;
+
+   case TGSI_UTIL_SIGN_SET:
+      micro_abs( chan, chan );
+      micro_neg( chan, chan );
+      break;
+
+   case TGSI_UTIL_SIGN_TOGGLE:
+      micro_neg( chan, chan );
+      break;
+
+   case TGSI_UTIL_SIGN_KEEP:
+      break;
+   }
+}
+
+static void
+store_dest(
+   struct tgsi_exec_machine *mach,
+   const union tgsi_exec_channel *chan,
+   const struct tgsi_full_dst_register *reg,
+   const struct tgsi_full_instruction *inst,
+   GLuint chan_index )
+{
+   union tgsi_exec_channel *dst;
+
+   switch( reg->DstRegister.File ) {
+   case TGSI_FILE_NULL:
+      return;
+
+   case TGSI_FILE_OUTPUT:
+      dst = &mach->Outputs[mach->Temps[TEMP_OUTPUT_I].xyzw[TEMP_OUTPUT_C].u[0] + reg->DstRegister.Index].xyzw[chan_index];
+      break;
+
+   case TGSI_FILE_TEMPORARY:
+      dst = &mach->Temps[reg->DstRegister.Index].xyzw[chan_index];
+      break;
+
+   case TGSI_FILE_ADDRESS:
+      dst = &mach->Addrs[reg->DstRegister.Index].xyzw[chan_index];
+      break;
+
+   default:
+      assert( 0 );
+   }
+
+   switch (inst->Instruction.Saturate)
+   {
+   case TGSI_SAT_NONE:
+      *dst = *chan;
+      break;
+
+   case TGSI_SAT_ZERO_ONE:
+      micro_lt( dst, chan, &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], chan );
+      micro_lt( dst, chan, &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], chan, &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C] );
+      break;
+
+   case TGSI_SAT_MINUS_PLUS_ONE:
+      assert( 0 );
+      break;
+
+   default:
+      assert( 0 );
+   }
+}
+
+#define FETCH(VAL,INDEX,CHAN)\
+    fetch_source (mach, VAL, &inst->FullSrcRegisters[INDEX], CHAN)
+
+#define STORE(VAL,INDEX,CHAN)\
+    store_dest (mach, VAL, &inst->FullDstRegisters[INDEX], inst, CHAN)
+
+static void
+exec_kil (struct tgsi_exec_machine *mach,
+          const struct tgsi_full_instruction *inst)
+{
+    GLuint uniquemask;
+    GLuint chan_index;
+    GLuint kilmask = 0;
+    union tgsi_exec_channel r[1];
+
+    /* This mask stores component bits that were already tested. Note that
+     * we test if the value is less than zero, so 1.0 and 0.0 need not to be
+     * tested. */
+    uniquemask = (1 << TGSI_EXTSWIZZLE_ZERO) | (1 << TGSI_EXTSWIZZLE_ONE);
+
+    for (chan_index = 0; chan_index < 4; chan_index++)
+    {
+        GLuint swizzle;
+        GLuint i;
+
+        /* unswizzle channel */
+        swizzle = tgsi_util_get_full_src_register_extswizzle (
+                        &inst->FullSrcRegisters[0],
+                        chan_index);
+
+        /* check if the component has not been already tested */
+        if (uniquemask & (1 << swizzle))
+            continue;
+        uniquemask |= 1 << swizzle;
+
+        FETCH(&r[0], 0, chan_index);
+        for (i = 0; i < 4; i++)
+            if (r[0].f[i] < 0.0f)
+                kilmask |= 1 << (i * 4);
+    }
+
+    mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] |= kilmask;
+}
+
+#if MESA
+/*
+ * Fetch a texel using S texture coordinate.
+ */
+static void
+fetch_texel_1d( GLcontext *ctx,
+                struct tgsi_sampler_state *sampler,
+                const union tgsi_exec_channel *s,
+                GLuint unit,
+                union tgsi_exec_channel *r,
+                union tgsi_exec_channel *g,
+                union tgsi_exec_channel *b,
+                union tgsi_exec_channel *a )
+{
+    SWcontext *swrast = SWRAST_CONTEXT(ctx);
+    GLuint fragment_index;
+    GLfloat stpq[4][4];
+    GLfloat lambdas[4];
+    GLchan rgba[4][4];
+
+    for (fragment_index = 0; fragment_index < 4; fragment_index++)
+    {
+        stpq[fragment_index][0] = s->f[fragment_index];
+    }
+
+    if (sampler->NeedLambda)
+    {
+        GLfloat dsdx = s->f[TILE_BOTTOM_RIGHT] - s->f[TILE_BOTTOM_LEFT];
+        GLfloat dsdy = s->f[TILE_TOP_LEFT]     - s->f[TILE_BOTTOM_LEFT];
+
+        GLfloat rho, lambda;
+
+        dsdx = FABSF(dsdx);
+        dsdy = FABSF(dsdy);
+
+        rho = MAX2(dsdx, dsdy) * sampler->ImageWidth;
+
+        lambda = LOG2(rho);
+
+        if (sampler->NeedLodBias)
+            lambda += sampler->LodBias;
+
+        if (sampler->NeedLambdaClamp)
+            lambda = CLAMP(lambda, sampler->MinLod, sampler->MaxLod);
+
+        /* XXX: Use the same lambda value throughout the tile.  Could
+         * end up with four unique values by recalculating partial
+         * derivs in the other row and column, and calculating lambda
+         * using the dx and dy values appropriate for each fragment in
+         * the tile.
+         */
+        lambdas[0] =
+        lambdas[1] =
+        lambdas[2] = 
+        lambdas[3] = lambda;
+    }
+
+    /* XXX use a float-valued TextureSample routine here!!! */
+    swrast->TextureSample[unit] (ctx,
+                                 ctx->Texture.Unit[unit]._Current,
+                                 4,
+                                 (const GLfloat (*)[4])stpq,
+                                 lambdas,
+                                 rgba);
+
+    for (fragment_index = 0; fragment_index < 4; fragment_index++)
+    {
+        r->f[fragment_index] = CHAN_TO_FLOAT(rgba[fragment_index][0]);
+        g->f[fragment_index] = CHAN_TO_FLOAT(rgba[fragment_index][1]);
+        b->f[fragment_index] = CHAN_TO_FLOAT(rgba[fragment_index][2]);
+        a->f[fragment_index] = CHAN_TO_FLOAT(rgba[fragment_index][3]);
+    }
+}
+
+/*
+ * Fetch a texel using ST texture coordinates.
+ */
+static void
+fetch_texel_2d( GLcontext *ctx,
+                struct tgsi_sampler_state *sampler,
+                const union tgsi_exec_channel *s,
+                const union tgsi_exec_channel *t,
+                GLuint unit,
+                union tgsi_exec_channel *r,
+                union tgsi_exec_channel *g,
+                union tgsi_exec_channel *b,
+                union tgsi_exec_channel *a )
+{
+   SWcontext *swrast = SWRAST_CONTEXT( ctx );
+   GLuint fragment_index;
+   GLfloat stpq[4][4];
+   GLfloat lambdas[4];
+   GLchan rgba[4][4];
+
+   for (fragment_index = 0; fragment_index < 4; fragment_index++) {
+      stpq[fragment_index][0] = s->f[fragment_index];
+      stpq[fragment_index][1] = t->f[fragment_index];
+   }
+
+   if (sampler->NeedLambda) {
+      GLfloat dsdx = s->f[TILE_BOTTOM_RIGHT] - s->f[TILE_BOTTOM_LEFT];
+      GLfloat dsdy = s->f[TILE_TOP_LEFT]     - s->f[TILE_BOTTOM_LEFT];
+
+      GLfloat dtdx = t->f[TILE_BOTTOM_RIGHT] - t->f[TILE_BOTTOM_LEFT];
+      GLfloat dtdy = t->f[TILE_TOP_LEFT]     - t->f[TILE_BOTTOM_LEFT];
+
+      GLfloat maxU, maxV, rho, lambda;
+
+      dsdx = FABSF( dsdx );
+      dsdy = FABSF( dsdy );
+      dtdx = FABSF( dtdx );
+      dtdy = FABSF( dtdy );
+
+      maxU = MAX2( dsdx, dsdy ) * sampler->ImageWidth;
+      maxV = MAX2( dtdx, dtdy ) * sampler->ImageHeight;
+
+      rho = MAX2( maxU, maxV );
+
+      lambda = LOG2( rho );
+
+      if (sampler->NeedLodBias)
+	 lambda += sampler->LodBias;
+
+      if (sampler->NeedLambdaClamp)
+         lambda = CLAMP(
+		     lambda,
+		     sampler->MinLod,
+		     sampler->MaxLod );
+
+      /* XXX: Use the same lambda value throughout the tile.  Could
+	 * end up with four unique values by recalculating partial
+	 * derivs in the other row and column, and calculating lambda
+	 * using the dx and dy values appropriate for each fragment in
+	 * the tile.
+	 */
+      lambdas[0] =
+      lambdas[1] =
+      lambdas[2] = 
+      lambdas[3] = lambda;
+   }
+
+   /* XXX use a float-valued TextureSample routine here!!! */
+   swrast->TextureSample[unit](
+      ctx,
+      ctx->Texture.Unit[unit]._Current,
+      4,
+      (const GLfloat (*)[4]) stpq,
+      lambdas,
+      rgba );
+
+   for (fragment_index = 0; fragment_index < 4; fragment_index++) {
+      r->f[fragment_index] = CHAN_TO_FLOAT( rgba[fragment_index][0] );
+      g->f[fragment_index] = CHAN_TO_FLOAT( rgba[fragment_index][1] );
+      b->f[fragment_index] = CHAN_TO_FLOAT( rgba[fragment_index][2] );
+      a->f[fragment_index] = CHAN_TO_FLOAT( rgba[fragment_index][3] );
+   }
+}
+
+/*
+ * Fetch a texel using STR texture coordinates.
+ */
+static void
+fetch_texel_3d( GLcontext *ctx,
+                struct tgsi_sampler_state *sampler,
+                const union tgsi_exec_channel *s,
+                const union tgsi_exec_channel *t,
+                const union tgsi_exec_channel *p,
+                GLuint unit,
+                union tgsi_exec_channel *r,
+                union tgsi_exec_channel *g,
+                union tgsi_exec_channel *b,
+                union tgsi_exec_channel *a )
+{
+    SWcontext *swrast = SWRAST_CONTEXT(ctx);
+    GLuint fragment_index;
+    GLfloat stpq[4][4];
+    GLfloat lambdas[4];
+    GLchan rgba[4][4];
+
+    for (fragment_index = 0; fragment_index < 4; fragment_index++)
+    {
+        stpq[fragment_index][0] = s->f[fragment_index];
+        stpq[fragment_index][1] = t->f[fragment_index];
+        stpq[fragment_index][2] = p->f[fragment_index];
+    }
+
+    if (sampler->NeedLambda)
+    {
+        GLfloat dsdx = s->f[TILE_BOTTOM_RIGHT] - s->f[TILE_BOTTOM_LEFT];
+        GLfloat dsdy = s->f[TILE_TOP_LEFT]     - s->f[TILE_BOTTOM_LEFT];
+
+        GLfloat dtdx = t->f[TILE_BOTTOM_RIGHT] - t->f[TILE_BOTTOM_LEFT];
+        GLfloat dtdy = t->f[TILE_TOP_LEFT]     - t->f[TILE_BOTTOM_LEFT];
+
+        GLfloat dpdx = p->f[TILE_BOTTOM_RIGHT] - p->f[TILE_BOTTOM_LEFT];
+        GLfloat dpdy = p->f[TILE_TOP_LEFT]     - p->f[TILE_BOTTOM_LEFT];
+
+        GLfloat maxU, maxV, maxW, rho, lambda;
+
+        dsdx = FABSF(dsdx);
+        dsdy = FABSF(dsdy);
+        dtdx = FABSF(dtdx);
+        dtdy = FABSF(dtdy);
+        dpdx = FABSF(dpdx);
+        dpdy = FABSF(dpdy);
+
+        maxU = MAX2(dsdx, dsdy) * sampler->ImageWidth;
+        maxV = MAX2(dtdx, dtdy) * sampler->ImageHeight;
+        maxW = MAX2(dpdx, dpdy) * sampler->ImageDepth;
+
+        rho = MAX2(maxU, MAX2(maxV, maxW));
+
+        lambda = LOG2(rho);
+
+        if (sampler->NeedLodBias)
+            lambda += sampler->LodBias;
+
+        if (sampler->NeedLambdaClamp)
+            lambda = CLAMP(lambda, sampler->MinLod, sampler->MaxLod);
+
+        /* XXX: Use the same lambda value throughout the tile.  Could
+         * end up with four unique values by recalculating partial
+         * derivs in the other row and column, and calculating lambda
+         * using the dx and dy values appropriate for each fragment in
+         * the tile.
+         */
+        lambdas[0] =
+        lambdas[1] =
+        lambdas[2] = 
+        lambdas[3] = lambda;
+    }
+
+    /* XXX use a float-valued TextureSample routine here!!! */
+    swrast->TextureSample[unit] (ctx,
+                                 ctx->Texture.Unit[unit]._Current,
+                                 4,
+                                 (const GLfloat (*)[4])stpq,
+                                 lambdas,
+                                 rgba);
+
+    for (fragment_index = 0; fragment_index < 4; fragment_index++)
+    {
+        r->f[fragment_index] = CHAN_TO_FLOAT(rgba[fragment_index][0]);
+        g->f[fragment_index] = CHAN_TO_FLOAT(rgba[fragment_index][1]);
+        b->f[fragment_index] = CHAN_TO_FLOAT(rgba[fragment_index][2]);
+        a->f[fragment_index] = CHAN_TO_FLOAT(rgba[fragment_index][3]);
+    }
+}
+#endif
+
+static GLuint
+map_label(
+   GLuint label,
+   struct tgsi_exec_labels *labels )
+{
+   GLuint i;
+
+   for( i = 0; i < labels->count; i++ ) {
+      if( labels->labels[i][0] == label ) {
+         return labels->labels[i][1];
+      }
+   }
+   assert( 0 );
+   return 0;
+}
+
+static void
+exec_instruction(
+   struct tgsi_exec_machine *mach,
+   const struct tgsi_full_instruction *inst,
+   struct tgsi_exec_labels *labels,
+   GLuint *programCounter )
+{
+#if MESA
+    GET_CURRENT_CONTEXT(ctx);
+#endif
+    GLuint chan_index;
+    union tgsi_exec_channel r[8];
+
+   switch (inst->Instruction.Opcode) {
+   case TGSI_OPCODE_ARL:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 FETCH( &r[0], 0, chan_index );
+	 micro_f2it( &r[0], &r[0] );
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_MOV:
+   /* TGSI_OPCODE_SWZ */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_LIT:
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_X )) {
+	 STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_X );
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_Y ) || IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+	 FETCH( &r[0], 0, CHAN_X );
+	 if (IS_CHANNEL_ENABLED( *inst, CHAN_Y )) {
+	    micro_max( &r[0], &r[0], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C] );
+	    STORE( &r[0], 0, CHAN_Y );
+	 }
+
+	 if (IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+	    FETCH( &r[1], 0, CHAN_Y );
+	    micro_max( &r[1], &r[1], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C] );
+
+	    FETCH( &r[2], 0, CHAN_W );
+	    micro_min( &r[2], &r[2], &mach->Temps[TEMP_128_I].xyzw[TEMP_128_C] );
+	    micro_max( &r[2], &r[2], &mach->Temps[TEMP_M128_I].xyzw[TEMP_M128_C] );
+	    micro_pow( &r[1], &r[1], &r[2] );
+	    micro_lt( &r[0], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], &r[0], &r[1], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C] );
+	    STORE( &r[0], 0, CHAN_Z );
+	 }
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_W )) {
+	 STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
+      }
+      break;
+
+   case TGSI_OPCODE_RCP:
+   /* TGSI_OPCODE_RECIP */
+      FETCH( &r[0], 0, CHAN_X );
+      micro_div( &r[0], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], &r[0] );
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_RSQ:
+   /* TGSI_OPCODE_RECIPSQRT */
+      FETCH( &r[0], 0, CHAN_X );
+      micro_sqrt( &r[0], &r[0] );
+      micro_div( &r[0], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], &r[0] );
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_EXP:
+      assert (0);
+      break;
+
+    case TGSI_OPCODE_LOG:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_MUL:
+        FOR_EACH_ENABLED_CHANNEL( *inst, chan_index )
+        {
+            FETCH(&r[0], 0, chan_index);
+            FETCH(&r[1], 1, chan_index);
+
+            micro_mul( &r[0], &r[0], &r[1] );
+
+            STORE(&r[0], 0, chan_index);
+        }
+        break;
+
+   case TGSI_OPCODE_ADD:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_add( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_DP3:
+   /* TGSI_OPCODE_DOT3 */
+      FETCH( &r[0], 0, CHAN_X );
+      FETCH( &r[1], 1, CHAN_X );
+      micro_mul( &r[0], &r[0], &r[1] );
+
+      FETCH( &r[1], 0, CHAN_Y );
+      FETCH( &r[2], 1, CHAN_Y );
+      micro_mul( &r[1], &r[1], &r[2] );
+      micro_add( &r[0], &r[0], &r[1] );
+
+      FETCH( &r[1], 0, CHAN_Z );
+      FETCH( &r[2], 1, CHAN_Z );
+      micro_mul( &r[1], &r[1], &r[2] );
+      micro_add( &r[0], &r[0], &r[1] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_DP4:
+    /* TGSI_OPCODE_DOT4 */
+        FETCH(&r[0], 0, CHAN_X);
+        FETCH(&r[1], 1, CHAN_X);
+
+        micro_mul( &r[0], &r[0], &r[1] );
+
+        FETCH(&r[1], 0, CHAN_Y);
+        FETCH(&r[2], 1, CHAN_Y);
+
+        micro_mul( &r[1], &r[1], &r[2] );
+        micro_add( &r[0], &r[0], &r[1] );
+
+        FETCH(&r[1], 0, CHAN_Z);
+        FETCH(&r[2], 1, CHAN_Z);
+
+        micro_mul( &r[1], &r[1], &r[2] );
+        micro_add( &r[0], &r[0], &r[1] );
+
+        FETCH(&r[1], 0, CHAN_W);
+        FETCH(&r[2], 1, CHAN_W);
+
+        micro_mul( &r[1], &r[1], &r[2] );
+        micro_add( &r[0], &r[0], &r[1] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_DST:
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_X )) {
+	 STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_X );
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_Y )) {
+	 FETCH( &r[0], 0, CHAN_Y );
+	 FETCH( &r[1], 1, CHAN_Y);
+	 micro_mul( &r[0], &r[0], &r[1] );
+	 STORE( &r[0], 0, CHAN_Y );
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+	 FETCH( &r[0], 0, CHAN_Z );
+	 STORE( &r[0], 0, CHAN_Z );
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_W )) {
+	 FETCH( &r[0], 1, CHAN_W );
+	 STORE( &r[0], 0, CHAN_W );
+      }
+      break;
+
+    case TGSI_OPCODE_MIN:
+        FOR_EACH_ENABLED_CHANNEL( *inst, chan_index )
+        {
+            FETCH(&r[0], 0, chan_index);
+            FETCH(&r[1], 1, chan_index);
+
+            micro_lt( &r[0], &r[0], &r[1], &r[0], &r[1] );
+
+            STORE(&r[0], 0, chan_index);
+        }
+        break;
+
+    case TGSI_OPCODE_MAX:
+        FOR_EACH_ENABLED_CHANNEL( *inst, chan_index )
+        {
+            FETCH(&r[0], 0, chan_index);
+            FETCH(&r[1], 1, chan_index);
+
+            micro_lt( &r[0], &r[0], &r[1], &r[1], &r[0] );
+
+            STORE(&r[0], 0, chan_index);
+        }
+        break;
+
+   case TGSI_OPCODE_SLT:
+   /* TGSI_OPCODE_SETLT */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_lt( &r[0], &r[0], &r[1], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_SGE:
+    /* TGSI_OPCODE_SETGE */
+        FOR_EACH_ENABLED_CHANNEL( *inst, chan_index )
+        {
+            FETCH(&r[0], 0, chan_index);
+            FETCH(&r[1], 1, chan_index);
+
+            micro_lt( &r[0], &r[0], &r[1], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C] );
+
+            STORE(&r[0], 0, chan_index);
+        }
+        break;
+
+   case TGSI_OPCODE_MAD:
+   /* TGSI_OPCODE_MADD */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_mul( &r[0], &r[0], &r[1] );
+         FETCH( &r[1], 2, chan_index );
+         micro_add( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_SUB:
+        FOR_EACH_ENABLED_CHANNEL( *inst, chan_index )
+        {
+            FETCH(&r[0], 0, chan_index);
+            FETCH(&r[1], 1, chan_index);
+
+            micro_sub( &r[0], &r[0], &r[1] );
+
+            STORE(&r[0], 0, chan_index);
+        }
+        break;
+
+    case TGSI_OPCODE_LERP:
+    /* TGSI_OPCODE_LRP */
+        FOR_EACH_ENABLED_CHANNEL( *inst, chan_index )
+        {
+            FETCH(&r[0], 0, chan_index);
+            FETCH(&r[1], 1, chan_index);
+            FETCH(&r[2], 2, chan_index);
+
+            micro_sub( &r[1], &r[1], &r[2] );
+            micro_mul( &r[0], &r[0], &r[1] );
+            micro_add( &r[0], &r[0], &r[2] );
+
+            STORE(&r[0], 0, chan_index);
+        }
+        break;
+
+    case TGSI_OPCODE_CND:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_CND0:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_DOT2ADD:
+    /* TGSI_OPCODE_DP2A */
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_INDEX:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_NEGATE:
+        assert (0);
+        break;
+
+   case TGSI_OPCODE_FRAC:
+   /* TGSI_OPCODE_FRC */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_frc( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_CLAMP:
+        assert (0);
+        break;
+
+   case TGSI_OPCODE_FLOOR:
+   /* TGSI_OPCODE_FLR */
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_flr( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_ROUND:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_rnd( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_EXPBASE2:
+    /* TGSI_OPCODE_EX2 */
+        FETCH(&r[0], 0, CHAN_X);
+
+        micro_pow( &r[0], &mach->Temps[TEMP_2_I].xyzw[TEMP_2_C], &r[0] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_LOGBASE2:
+   /* TGSI_OPCODE_LG2 */
+      FETCH( &r[0], 0, CHAN_X );
+      micro_lg2( &r[0], &r[0] );
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_POWER:
+    /* TGSI_OPCODE_POW */
+        FETCH(&r[0], 0, CHAN_X);
+        FETCH(&r[1], 1, CHAN_X);
+
+        micro_pow( &r[0], &r[0], &r[1] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_CROSSPRODUCT:
+    /* TGSI_OPCODE_XPD */
+        FETCH(&r[0], 0, CHAN_Y);
+        FETCH(&r[1], 1, CHAN_Z);
+
+        micro_mul( &r[2], &r[0], &r[1] );
+
+        FETCH(&r[3], 0, CHAN_Z);
+        FETCH(&r[4], 1, CHAN_Y);
+
+        micro_mul( &r[5], &r[3], &r[4] );
+        micro_sub( &r[2], &r[2], &r[5] );
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_X )) {
+	 STORE( &r[2], 0, CHAN_X );
+      }
+
+        FETCH(&r[2], 1, CHAN_X);
+
+        micro_mul( &r[3], &r[3], &r[2] );
+
+        FETCH(&r[5], 0, CHAN_X);
+
+        micro_mul( &r[1], &r[1], &r[5] );
+        micro_sub( &r[3], &r[3], &r[1] );
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_Y )) {
+	 STORE( &r[3], 0, CHAN_Y );
+      }
+
+        micro_mul( &r[5], &r[5], &r[4] );
+        micro_mul( &r[0], &r[0], &r[2] );
+        micro_sub( &r[5], &r[5], &r[0] );
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+	 STORE( &r[5], 0, CHAN_Z );
+      }
+
+      if (IS_CHANNEL_ENABLED( *inst, CHAN_W )) {
+	 STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
+      }
+        break;
+
+    case TGSI_OPCODE_MULTIPLYMATRIX:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_ABS:
+        FOR_EACH_ENABLED_CHANNEL( *inst, chan_index )
+        {
+            FETCH(&r[0], 0, chan_index);
+
+            micro_abs( &r[0], &r[0] );
+
+            STORE(&r[0], 0, chan_index);
+        }
+        break;
+
+    case TGSI_OPCODE_RCC:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_DPH:
+        FETCH(&r[0], 0, CHAN_X);
+        FETCH(&r[1], 1, CHAN_X);
+
+        micro_mul( &r[0], &r[0], &r[1] );
+
+        FETCH(&r[1], 0, CHAN_Y);
+        FETCH(&r[2], 1, CHAN_Y);
+
+        micro_mul( &r[1], &r[1], &r[2] );
+        micro_add( &r[0], &r[0], &r[1] );
+
+        FETCH(&r[1], 0, CHAN_Z);
+        FETCH(&r[2], 1, CHAN_Z);
+
+        micro_mul( &r[1], &r[1], &r[2] );
+        micro_add( &r[0], &r[0], &r[1] );
+
+        FETCH(&r[1], 1, CHAN_W);
+
+        micro_add( &r[0], &r[0], &r[1] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_COS:
+        FETCH(&r[0], 0, CHAN_X);
+
+        micro_cos( &r[0], &r[0] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_DDX:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_DDY:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_KIL:
+        exec_kil (mach, inst);
+        break;
+
+    case TGSI_OPCODE_PK2H:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_PK2US:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_PK4B:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_PK4UB:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_RFL:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_SEQ:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_SFL:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_SGT:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_SIN:
+        FETCH(&r[0], 0, CHAN_X);
+
+        micro_sin( &r[0], &r[0] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_SLE:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_SNE:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_STR:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_TEX:
+
+        switch (inst->InstructionExtTexture.Texture)
+        {
+        case TGSI_TEXTURE_1D:
+
+            FETCH(&r[0], 0, CHAN_X);
+
+            switch (inst->FullSrcRegisters[0].SrcRegisterExtSwz.ExtDivide)
+            {
+            case TGSI_EXTSWIZZLE_W:
+                FETCH(&r[1], 0, CHAN_W);
+                micro_div( &r[0], &r[0], &r[1] );
+                break;
+
+            case TGSI_EXTSWIZZLE_ONE:
+                break;
+
+            default:
+                assert (0);
+            }
+#if MESA
+            fetch_texel_1d (ctx,
+                            &mach->Samplers[inst->FullSrcRegisters[1].SrcRegister.Index],
+                            &r[0],
+                            inst->FullSrcRegisters[1].SrcRegister.Index,
+                            &r[0], &r[1], &r[2], &r[3]);
+#endif
+            break;
+
+        case TGSI_TEXTURE_2D:
+        case TGSI_TEXTURE_RECT:
+
+            FETCH(&r[0], 0, CHAN_X);
+            FETCH(&r[1], 0, CHAN_Y);
+
+            switch (inst->FullSrcRegisters[0].SrcRegisterExtSwz.ExtDivide)
+            {
+            case TGSI_EXTSWIZZLE_W:
+                FETCH(&r[2], 0, CHAN_W);
+                micro_div( &r[0], &r[0], &r[2] );
+                micro_div( &r[1], &r[1], &r[2] );
+                break;
+
+            case TGSI_EXTSWIZZLE_ONE:
+                break;
+
+            default:
+                assert (0);
+            }
+
+#if MESA
+            fetch_texel_2d (ctx,
+                            &mach->Samplers[inst->FullSrcRegisters[1].SrcRegister.Index],
+                            &r[0], &r[1],
+                            inst->FullSrcRegisters[1].SrcRegister.Index,
+                            &r[0], &r[1], &r[2], &r[3]);
+#endif
+            break;
+
+        case TGSI_TEXTURE_3D:
+        case TGSI_TEXTURE_CUBE:
+
+            FETCH(&r[0], 0, CHAN_X);
+            FETCH(&r[1], 0, CHAN_Y);
+            FETCH(&r[2], 0, CHAN_Z);
+
+            switch (inst->FullSrcRegisters[0].SrcRegisterExtSwz.ExtDivide)
+            {
+            case TGSI_EXTSWIZZLE_W:
+                FETCH(&r[3], 0, CHAN_W);
+                micro_div( &r[0], &r[0], &r[3] );
+                micro_div( &r[1], &r[1], &r[3] );
+                micro_div( &r[2], &r[2], &r[3] );
+                break;
+
+            case TGSI_EXTSWIZZLE_ONE:
+                break;
+
+            default:
+                assert (0);
+            }
+
+#if MESA
+            fetch_texel_3d (ctx,
+                            &mach->Samplers[inst->FullSrcRegisters[1].SrcRegister.Index],
+                            &r[0], &r[1], &r[2],
+                            inst->FullSrcRegisters[1].SrcRegister.Index,
+                            &r[0], &r[1], &r[2], &r[3]);
+#endif
+            break;
+
+        default:
+            assert (0);
+        }
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+	 STORE( &r[chan_index], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_TXD:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_UP2H:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_UP2US:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_UP4B:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_UP4UB:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_X2D:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_ARA:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_ARR:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_BRA:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_CAL:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_RET:
+        /* XXX: end of shader! */
+        /*assert (0);*/
+        break;
+
+    case TGSI_OPCODE_SSG:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_CMP:
+        FOR_EACH_ENABLED_CHANNEL( *inst, chan_index )
+        {
+            FETCH(&r[0], 0, chan_index);
+            FETCH(&r[1], 1, chan_index);
+            FETCH(&r[2], 2, chan_index);
+
+            micro_lt( &r[0], &r[0], &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], &r[1], &r[2] );
+
+            STORE(&r[0], 0, chan_index);
+        }
+        break;
+
+   case TGSI_OPCODE_SCS:
+      if( IS_CHANNEL_ENABLED( *inst, CHAN_X ) || IS_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
+         FETCH( &r[0], 0, CHAN_X );
+      }
+      if( IS_CHANNEL_ENABLED( *inst, CHAN_X ) ) {
+         micro_cos( &r[1], &r[0] );
+         STORE( &r[1], 0, CHAN_X );
+      }
+      if( IS_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
+         micro_sin( &r[1], &r[0] );
+         STORE( &r[1], 0, CHAN_Y );
+      }
+      if( IS_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
+         STORE( &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], 0, CHAN_Z );
+      }
+      if( IS_CHANNEL_ENABLED( *inst, CHAN_W ) ) {
+         STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
+      }
+      break;
+
+    case TGSI_OPCODE_TXB:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_NRM:
+        assert (0);
+        break;
+
+   case TGSI_OPCODE_DIV:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_DP2:
+      FETCH( &r[0], 0, CHAN_X );
+      FETCH( &r[1], 1, CHAN_X );
+      micro_mul( &r[0], &r[0], &r[1] );
+
+      FETCH( &r[1], 0, CHAN_Y );
+      FETCH( &r[2], 1, CHAN_Y );
+      micro_mul( &r[1], &r[1], &r[2] );
+      micro_add( &r[0], &r[0], &r[1] );
+
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_TXL:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_BRK:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_IF:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_LOOP:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_REP:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_ELSE:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_ENDIF:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_ENDLOOP:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_ENDREP:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_PUSHA:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_POPA:
+        assert (0);
+        break;
+
+   case TGSI_OPCODE_CEIL:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_ceil( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_I2F:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_i2f( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_NOT:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         micro_not( &r[0], &r[0] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_TRUNC:
+        assert (0);
+        break;
+
+   case TGSI_OPCODE_SHL:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_shl( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SHR:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_ishr( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_AND:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_and( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_OR:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_or( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_MOD:
+        assert (0);
+        break;
+
+   case TGSI_OPCODE_XOR:
+      FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( &r[0], 0, chan_index );
+         FETCH( &r[1], 1, chan_index );
+         micro_xor( &r[0], &r[0], &r[1] );
+         STORE( &r[0], 0, chan_index );
+      }
+      break;
+
+    case TGSI_OPCODE_SAD:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_TXF:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_TXQ:
+        assert (0);
+        break;
+
+    case TGSI_OPCODE_CONT:
+        assert (0);
+        break;
+
+   case TGSI_OPCODE_EMIT:
+      mach->Temps[TEMP_OUTPUT_I].xyzw[TEMP_OUTPUT_C].u[0] += 16;
+      mach->Primitives[mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0]]++;
+      break;
+
+   case TGSI_OPCODE_ENDPRIM:
+      mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0]++;
+      mach->Primitives[mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0]] = 0;
+      break;
+
+   default:
+      assert( 0 );
+    }
+}
+
+
+#if !defined(XSTDCALL) 
+#if defined(WIN32)
+#define XSTDCALL __stdcall
+#else
+#define XSTDCALL
+#endif
+#endif
+
+typedef void (XSTDCALL *fp_function) (const struct tgsi_exec_vector *input,
+				     struct tgsi_exec_vector *output,
+				     GLfloat (*constant)[4],
+				     struct tgsi_exec_vector *temporary);
+
+void
+tgsi_exec_machine_run2(
+   struct tgsi_exec_machine *mach,
+   struct tgsi_exec_labels *labels )
+{
+#if MESA
+    GET_CURRENT_CONTEXT(ctx);
+    GLuint i;
+#endif
+
+#if XXX_SSE
+    fp_function function;
+
+    mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] = 0;
+
+    function = (fp_function) x86_get_func (&mach->Function);
+
+    function (mach->Inputs,
+              mach->Outputs,
+              mach->Consts,
+              mach->Temps);
+#else
+   struct tgsi_parse_context parse;
+
+   mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] = 0;
+   mach->Temps[TEMP_OUTPUT_I].xyzw[TEMP_OUTPUT_C].u[0] = 0;
+
+   if( mach->Processor == TGSI_PROCESSOR_GEOMETRY ) {
+      mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0] = 0;
+      mach->Primitives[0] = 0;
+   }
+
+   tgsi_parse_init( &parse, mach->Tokens );
+   while( !tgsi_parse_end_of_tokens( &parse ) ) {
+      tgsi_parse_token( &parse );
+      switch( parse.FullToken.Token.Type ) {
+      case TGSI_TOKEN_TYPE_DECLARATION:
+         break;
+      case TGSI_TOKEN_TYPE_IMMEDIATE:
+         break;
+      case TGSI_TOKEN_TYPE_INSTRUCTION:
+         exec_instruction( mach, &parse.FullToken.FullInstruction, labels, &parse.Position );
+         break;
+      default:
+         assert( 0 );
+      }
+   }
+   tgsi_parse_free (&parse);
+#endif
+
+#if MESA
+   if (mach->Processor == TGSI_PROCESSOR_FRAGMENT) {
+      /*
+       * Scale back depth component.
+       */
+      for (i = 0; i < 4; i++)
+         mach->Outputs[0].xyzw[2].f[i] *= ctx->DrawBuffer->_DepthMaxF;
+   }
+#endif
+}
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_exec.h b/src/mesa/pipe/tgsi/core/tgsi_exec.h
new file mode 100644
index 00000000000..abce6ee924a
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_exec.h
@@ -0,0 +1,137 @@
+#if !defined TGSI_EXEC_H
+#define TGSI_EXEC_H
+
+#if 0
+#include "x86/rtasm/x86sse.h"
+#endif
+
+#if defined __cplusplus
+extern "C" {
+#endif // defined __cplusplus
+
+union tgsi_exec_channel
+{
+   GLfloat  f[4];
+   GLint    i[4];
+   GLuint   u[4];
+};
+
+struct tgsi_exec_vector
+{
+   union tgsi_exec_channel xyzw[4];
+};
+
+struct tgsi_sampler_state
+{
+    GLboolean   NeedLambda;
+    GLboolean   NeedLodBias;        /* if NeedLambda */
+    GLboolean   NeedLambdaClamp;    /* if NeedLambda */
+    GLfloat     LodBias;            /* if NeedLodBias */
+    GLfloat     MinLod;             /* if NeedLambdaClamp */
+    GLfloat     MaxLod;             /* if NeedLambdaClamp */
+    GLfloat     ImageWidth;
+    GLfloat     ImageHeight;
+    GLfloat     ImageDepth;
+};
+
+struct tgsi_exec_labels
+{
+   GLuint   labels[128][2];
+   GLuint   count;
+};
+
+#define TGSI_EXEC_TEMP_00000000_I   32
+#define TGSI_EXEC_TEMP_00000000_C   0
+
+#define TGSI_EXEC_TEMP_7FFFFFFF_I   32
+#define TGSI_EXEC_TEMP_7FFFFFFF_C   1
+
+#define TGSI_EXEC_TEMP_80000000_I   32
+#define TGSI_EXEC_TEMP_80000000_C   2
+
+#define TGSI_EXEC_TEMP_FFFFFFFF_I   32
+#define TGSI_EXEC_TEMP_FFFFFFFF_C   3
+
+#define TGSI_EXEC_TEMP_ONE_I        33
+#define TGSI_EXEC_TEMP_ONE_C        0
+
+#define TGSI_EXEC_TEMP_TWO_I        33
+#define TGSI_EXEC_TEMP_TWO_C        1
+
+#define TGSI_EXEC_TEMP_128_I        33
+#define TGSI_EXEC_TEMP_128_C        2
+
+#define TGSI_EXEC_TEMP_MINUS_128_I  33
+#define TGSI_EXEC_TEMP_MINUS_128_C  3
+
+#define TGSI_EXEC_TEMP_KILMASK_I    34
+#define TGSI_EXEC_TEMP_KILMASK_C    0
+
+#define TGSI_EXEC_TEMP_OUTPUT_I     34
+#define TGSI_EXEC_TEMP_OUTPUT_C     1
+
+#define TGSI_EXEC_TEMP_PRIMITIVE_I  34
+#define TGSI_EXEC_TEMP_PRIMITIVE_C  2
+
+#define TGSI_EXEC_TEMP_R0           35
+
+#define TGSI_EXEC_NUM_TEMPS   (32 + 4)
+#define TGSI_EXEC_NUM_ADDRS   1
+
+struct tgsi_exec_machine
+{
+   /*
+    * 32 program temporaries
+    * 4  internal temporaries
+    * 1  address
+    * 1  temporary of padding to align to 16 bytes
+    */
+   struct tgsi_exec_vector       _Temps[TGSI_EXEC_NUM_TEMPS + TGSI_EXEC_NUM_ADDRS + 1];
+
+   /*
+    * This will point to _Temps after aligning to 16B boundary.
+    */
+   struct tgsi_exec_vector       *Temps;
+   struct tgsi_exec_vector       *Addrs;
+
+   struct tgsi_sampler_state     Samplers[16];
+
+   GLfloat                       Imms[256][4];
+   GLuint                        ImmLimit;
+   GLfloat                       (*Consts)[4];
+   const struct tgsi_exec_vector *Inputs;
+   struct tgsi_exec_vector       *Outputs;
+   struct tgsi_token             *Tokens;
+   GLuint                        Processor;
+
+   GLuint                        *Primitives;
+#if XXX_SSE
+   struct x86_function           Function;
+#endif
+};
+
+void
+tgsi_exec_machine_init(
+   struct tgsi_exec_machine *mach,
+   struct tgsi_token *tokens );
+
+void
+tgsi_exec_prepare(
+   struct tgsi_exec_machine *mach,
+   struct tgsi_exec_labels *labels );
+
+void
+tgsi_exec_machine_run(
+   struct tgsi_exec_machine *mach );
+
+void
+tgsi_exec_machine_run2(
+   struct tgsi_exec_machine *mach,
+   struct tgsi_exec_labels *labels );
+
+#if defined __cplusplus
+} // extern "C"
+#endif // defined __cplusplus
+
+#endif // !defined TGSI_EXEC_H
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_parse.c b/src/mesa/pipe/tgsi/core/tgsi_parse.c
new file mode 100644
index 00000000000..df34fa7d7ef
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_parse.c
@@ -0,0 +1,284 @@
+#include "tgsi_platform.h"
+#include "tgsi_core.h"
+
+void
+tgsi_full_token_init(
+   union tgsi_full_token *full_token )
+{
+   full_token->Token.Type = TGSI_TOKEN_TYPE_DECLARATION;
+}
+
+void
+tgsi_full_token_free(
+   union tgsi_full_token *full_token )
+{
+   if( full_token->Token.Type == TGSI_TOKEN_TYPE_IMMEDIATE )
+      free( full_token->FullImmediate.u.Pointer );
+}
+
+GLuint
+tgsi_parse_init(
+   struct tgsi_parse_context *ctx,
+   const struct tgsi_token *tokens )
+{
+   ctx->FullVersion.Version = *(struct tgsi_version *) &tokens[0];
+   if( ctx->FullVersion.Version.MajorVersion > 1 ) {
+      return TGSI_PARSE_ERROR;
+   }
+
+   ctx->FullHeader.Header = *(struct tgsi_header *) &tokens[1];
+   if( ctx->FullHeader.Header.HeaderSize >= 2 ) {
+      ctx->FullHeader.Processor = *(struct tgsi_processor *) &tokens[2];
+   }
+   else {
+      ctx->FullHeader.Processor = tgsi_default_processor();
+   }
+
+   ctx->Tokens = tokens;
+   ctx->Position = 1 + ctx->FullHeader.Header.HeaderSize;
+
+   tgsi_full_token_init( &ctx->FullToken );
+
+   return TGSI_PARSE_OK;
+}
+
+void
+tgsi_parse_free(
+   struct tgsi_parse_context *ctx )
+{
+   tgsi_full_token_free( &ctx->FullToken );
+}
+
+GLuint
+tgsi_parse_end_of_tokens(
+   struct tgsi_parse_context *ctx )
+{
+   return ctx->Position >=
+      1 + ctx->FullHeader.Header.HeaderSize + ctx->FullHeader.Header.BodySize;
+}
+
+static void
+next_token(
+   struct tgsi_parse_context *ctx,
+   void *token )
+{
+   assert( !tgsi_parse_end_of_tokens( ctx ) );
+
+   *(struct tgsi_token *) token = ctx->Tokens[ctx->Position++];
+}
+
+void
+tgsi_parse_token(
+   struct tgsi_parse_context *ctx )
+{
+   struct tgsi_token token;
+   GLuint i;
+
+   tgsi_full_token_free( &ctx->FullToken );
+   tgsi_full_token_init( &ctx->FullToken );
+
+   next_token( ctx, &token );
+
+   switch( token.Type ) {
+   case TGSI_TOKEN_TYPE_DECLARATION:
+   {
+      struct tgsi_full_declaration *decl = &ctx->FullToken.FullDeclaration;
+
+      *decl = tgsi_default_full_declaration();
+      decl->Declaration = *(struct tgsi_declaration *) &token;
+
+      switch( decl->Declaration.Type ) {
+      case TGSI_DECLARE_RANGE:
+         next_token( ctx, &decl->u.DeclarationRange );
+         break;
+
+      case TGSI_DECLARE_MASK:
+         next_token( ctx, &decl->u.DeclarationMask );
+         break;
+
+      default:
+         assert (0);
+      }
+
+      if( decl->Declaration.Interpolate ) {
+         next_token( ctx, &decl->Interpolation );
+      }
+
+      break;
+   }
+
+   case TGSI_TOKEN_TYPE_IMMEDIATE:
+   {
+      struct tgsi_full_immediate *imm = &ctx->FullToken.FullImmediate;
+
+      *imm = tgsi_default_full_immediate();
+      imm->Immediate = *(struct tgsi_immediate *) &token;
+
+      assert( !imm->Immediate.Extended );
+
+      switch (imm->Immediate.DataType) {
+      case TGSI_IMM_FLOAT32:
+         imm->u.Pointer = malloc(
+            sizeof( struct tgsi_immediate_float32 ) * (imm->Immediate.Size - 1) );
+         for( i = 0; i < imm->Immediate.Size - 1; i++ ) {
+            next_token( ctx, &imm->u.ImmediateFloat32[i] );
+         }
+         break;
+
+      default:
+         assert( 0 );
+      }
+
+      break;
+   }
+
+   case TGSI_TOKEN_TYPE_INSTRUCTION:
+   {
+      struct tgsi_full_instruction *inst = &ctx->FullToken.FullInstruction;
+      GLuint extended;
+
+      *inst = tgsi_default_full_instruction();
+      inst->Instruction = *(struct tgsi_instruction *) &token;
+
+      extended = inst->Instruction.Extended;
+
+      while( extended ) {
+         struct tgsi_src_register_ext token;
+
+         next_token( ctx, &token );
+
+         switch( token.Type ) {
+         case TGSI_INSTRUCTION_EXT_TYPE_NV:
+            inst->InstructionExtNv =
+               *(struct tgsi_instruction_ext_nv *) &token;
+            break;
+
+         case TGSI_INSTRUCTION_EXT_TYPE_LABEL:
+            inst->InstructionExtLabel =
+               *(struct tgsi_instruction_ext_label *) &token;
+            break;
+
+         case TGSI_INSTRUCTION_EXT_TYPE_TEXTURE:
+            inst->InstructionExtTexture =
+               *(struct tgsi_instruction_ext_texture *) &token;
+            break;
+
+         default:
+            assert( 0 );
+         }
+
+         extended = token.Extended;
+      }
+
+      assert( inst->Instruction.NumDstRegs <= TGSI_FULL_MAX_DST_REGISTERS );
+
+      for(  i = 0; i < inst->Instruction.NumDstRegs; i++ ) {
+         GLuint extended;
+
+         next_token( ctx, &inst->FullDstRegisters[i].DstRegister );
+
+         /*
+          * No support for indirect or multi-dimensional addressing.
+          */
+         assert( !inst->FullDstRegisters[i].DstRegister.Indirect );
+         assert( !inst->FullDstRegisters[i].DstRegister.Dimension );
+
+         extended = inst->FullDstRegisters[i].DstRegister.Extended;
+
+         while( extended ) {
+            struct tgsi_src_register_ext token;
+
+            next_token( ctx, &token );
+
+            switch( token.Type ) {
+            case TGSI_DST_REGISTER_EXT_TYPE_CONDCODE:
+               inst->FullDstRegisters[i].DstRegisterExtConcode =
+                  *(struct tgsi_dst_register_ext_concode *) &token;
+               break;
+
+            case TGSI_DST_REGISTER_EXT_TYPE_MODULATE:
+               inst->FullDstRegisters[i].DstRegisterExtModulate =
+                  *(struct tgsi_dst_register_ext_modulate *) &token;
+               break;
+
+            default:
+               assert( 0 );
+            }
+
+            extended = token.Extended;
+         }
+      }
+
+      assert( inst->Instruction.NumSrcRegs <= TGSI_FULL_MAX_SRC_REGISTERS );
+
+      for( i = 0; i < inst->Instruction.NumSrcRegs; i++ ) {
+         GLuint extended;
+
+         next_token( ctx, &inst->FullSrcRegisters[i].SrcRegister );
+
+         extended = inst->FullSrcRegisters[i].SrcRegister.Extended;
+
+         while( extended ) {
+            struct tgsi_src_register_ext token;
+
+            next_token( ctx, &token );
+
+            switch( token.Type ) {
+            case TGSI_SRC_REGISTER_EXT_TYPE_SWZ:
+               inst->FullSrcRegisters[i].SrcRegisterExtSwz =
+                  *(struct tgsi_src_register_ext_swz *) &token;
+               break;
+
+            case TGSI_SRC_REGISTER_EXT_TYPE_MOD:
+               inst->FullSrcRegisters[i].SrcRegisterExtMod =
+                  *(struct tgsi_src_register_ext_mod *) &token;
+               break;
+
+            default:
+               assert( 0 );
+            }
+
+            extended = token.Extended;
+         }
+
+         if( inst->FullSrcRegisters[i].SrcRegister.Indirect ) {
+            next_token( ctx, &inst->FullSrcRegisters[i].SrcRegisterInd );
+
+            /*
+             * No support for indirect or multi-dimensional addressing.
+             */
+            assert( !inst->FullSrcRegisters[i].SrcRegisterInd.Indirect );
+            assert( !inst->FullSrcRegisters[i].SrcRegisterInd.Dimension );
+            assert( !inst->FullSrcRegisters[i].SrcRegisterInd.Extended );
+         }
+
+         if( inst->FullSrcRegisters[i].SrcRegister.Dimension ) {
+            next_token( ctx, &inst->FullSrcRegisters[i].SrcRegisterDim );
+
+            /*
+             * No support for multi-dimensional addressing.
+             */
+            assert( !inst->FullSrcRegisters[i].SrcRegisterDim.Dimension );
+            assert( !inst->FullSrcRegisters[i].SrcRegisterDim.Extended );
+
+            if( inst->FullSrcRegisters[i].SrcRegisterDim.Indirect ) {
+               next_token( ctx, &inst->FullSrcRegisters[i].SrcRegisterDimInd );
+
+               /*
+               * No support for indirect or multi-dimensional addressing.
+               */
+               assert( !inst->FullSrcRegisters[i].SrcRegisterInd.Indirect );
+               assert( !inst->FullSrcRegisters[i].SrcRegisterInd.Dimension );
+               assert( !inst->FullSrcRegisters[i].SrcRegisterInd.Extended );
+            }
+         }
+      }
+
+      break;
+   }
+
+   default:
+      assert( 0 );
+   }
+}
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_parse.h b/src/mesa/pipe/tgsi/core/tgsi_parse.h
new file mode 100644
index 00000000000..61ad0669b17
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_parse.h
@@ -0,0 +1,120 @@
+#if !defined TGSI_PARSE_H
+#define TGSI_PARSE_H
+
+#if defined __cplusplus
+extern "C" {
+#endif // defined __cplusplus
+
+struct tgsi_full_version
+{
+   struct tgsi_version  Version;
+};
+
+struct tgsi_full_header
+{
+   struct tgsi_header      Header;
+   struct tgsi_processor   Processor;
+};
+
+struct tgsi_full_dst_register
+{
+   struct tgsi_dst_register               DstRegister;
+   struct tgsi_dst_register_ext_concode   DstRegisterExtConcode;
+   struct tgsi_dst_register_ext_modulate  DstRegisterExtModulate;
+};
+
+struct tgsi_full_src_register
+{
+   struct tgsi_src_register         SrcRegister;
+   struct tgsi_src_register_ext_swz SrcRegisterExtSwz;
+   struct tgsi_src_register_ext_mod SrcRegisterExtMod;
+   struct tgsi_src_register         SrcRegisterInd;
+   struct tgsi_dimension            SrcRegisterDim;
+   struct tgsi_src_register         SrcRegisterDimInd;
+};
+
+struct tgsi_full_declaration
+{
+   struct tgsi_declaration Declaration;
+   union
+   {
+      struct tgsi_declaration_range DeclarationRange;
+      struct tgsi_declaration_mask  DeclarationMask;
+   } u;
+   struct tgsi_declaration_interpolation  Interpolation;
+};
+
+struct tgsi_full_immediate
+{
+   struct tgsi_immediate   Immediate;
+   union
+   {
+      void                          *Pointer;
+      struct tgsi_immediate_float32 *ImmediateFloat32;
+   } u;
+};
+
+#define TGSI_FULL_MAX_DST_REGISTERS 2
+#define TGSI_FULL_MAX_SRC_REGISTERS 3
+
+struct tgsi_full_instruction
+{
+   struct tgsi_instruction             Instruction;
+   struct tgsi_instruction_ext_nv      InstructionExtNv;
+   struct tgsi_instruction_ext_label   InstructionExtLabel;
+   struct tgsi_instruction_ext_texture InstructionExtTexture;
+   struct tgsi_full_dst_register       FullDstRegisters[TGSI_FULL_MAX_DST_REGISTERS];
+   struct tgsi_full_src_register       FullSrcRegisters[TGSI_FULL_MAX_SRC_REGISTERS];
+};
+
+union tgsi_full_token
+{
+   struct tgsi_token             Token;
+   struct tgsi_full_declaration  FullDeclaration;
+   struct tgsi_full_immediate    FullImmediate;
+   struct tgsi_full_instruction  FullInstruction;
+};
+
+void
+tgsi_full_token_init(
+   union tgsi_full_token *full_token );
+
+void
+tgsi_full_token_free(
+   union tgsi_full_token *full_token );
+
+struct tgsi_parse_context
+{
+   const struct tgsi_token    *Tokens;
+   GLuint                     Position;
+   struct tgsi_full_version   FullVersion;
+   struct tgsi_full_header    FullHeader;
+   union tgsi_full_token      FullToken;
+};
+
+#define TGSI_PARSE_OK      0
+#define TGSI_PARSE_ERROR   1
+
+GLuint
+tgsi_parse_init(
+   struct tgsi_parse_context *ctx,
+   const struct tgsi_token *tokens );
+
+void
+tgsi_parse_free(
+   struct tgsi_parse_context *ctx );
+
+GLuint
+tgsi_parse_end_of_tokens(
+   struct tgsi_parse_context *ctx );
+
+void
+tgsi_parse_token(
+   struct tgsi_parse_context *ctx );
+
+#if defined __cplusplus
+} // extern "C"
+#endif // defined __cplusplus
+
+#endif // !defined TGSI_PARSE_H
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_token.h b/src/mesa/pipe/tgsi/core/tgsi_token.h
new file mode 100644
index 00000000000..becdd480fbd
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_token.h
@@ -0,0 +1,1045 @@
+#if !defined TGSI_TOKEN_H
+#define TGSI_TOKEN_H
+
+#if defined __cplusplus
+extern "C" {
+#endif // defined __cplusplus
+
+struct tgsi_version
+{
+   GLuint MajorVersion  : 8;
+   GLuint MinorVersion  : 8;
+   GLuint Padding       : 16;
+};
+
+struct tgsi_header
+{
+   GLuint HeaderSize : 8;
+   GLuint BodySize   : 24;
+};
+
+#define TGSI_PROCESSOR_FRAGMENT  0
+#define TGSI_PROCESSOR_VERTEX    1
+#define TGSI_PROCESSOR_GEOMETRY  2
+
+struct tgsi_processor
+{
+   GLuint Processor  : 4;  /* TGSI_PROCESSOR_ */
+   GLuint Padding    : 28;
+};
+
+#define TGSI_TOKEN_TYPE_DECLARATION    0
+#define TGSI_TOKEN_TYPE_IMMEDIATE      1
+#define TGSI_TOKEN_TYPE_INSTRUCTION    2
+
+struct tgsi_token
+{
+   GLuint Type       : 4;  /* TGSI_TOKEN_TYPE_ */
+   GLuint Size       : 8;  /* UINT */
+   GLuint Padding    : 19;
+   GLuint Extended   : 1;  /* BOOL */
+};
+
+#define TGSI_FILE_NULL        0
+#define TGSI_FILE_CONSTANT    1
+#define TGSI_FILE_INPUT       2
+#define TGSI_FILE_OUTPUT      3
+#define TGSI_FILE_TEMPORARY   4
+#define TGSI_FILE_SAMPLER     5
+#define TGSI_FILE_ADDRESS     6
+#define TGSI_FILE_IMMEDIATE   7
+
+#define TGSI_DECLARE_RANGE    0
+#define TGSI_DECLARE_MASK     1
+
+struct tgsi_declaration
+{
+   GLuint Type          : 4;  /* TGSI_TOKEN_TYPE_DECLARATION */
+   GLuint Size          : 8;  /* UINT */
+   GLuint File          : 4;  /* TGSI_FILE_ */
+   GLuint Declare       : 4;  /* TGSI_DECLARE_ */
+   GLuint Interpolate   : 1;  /* BOOL */
+   GLuint Padding       : 10;
+   GLuint Extended      : 1;  /* BOOL */
+};
+
+struct tgsi_declaration_range
+{
+   GLuint First   : 16; /* UINT */
+   GLuint Last    : 16; /* UINT */
+};
+
+struct tgsi_declaration_mask
+{
+   GLuint Mask : 32; /* UINT */
+};
+
+#define TGSI_INTERPOLATE_CONSTANT      0
+#define TGSI_INTERPOLATE_LINEAR        1
+#define TGSI_INTERPOLATE_PERSPECTIVE   2
+
+struct tgsi_declaration_interpolation
+{
+   GLuint Interpolate   : 4;  /* TGSI_INTERPOLATE_ */
+   GLuint Padding       : 28;
+};
+
+#define TGSI_IMM_FLOAT32   0
+
+struct tgsi_immediate
+{
+   GLuint Type       : 4;  /* TGSI_TOKEN_TYPE_IMMEDIATE */
+   GLuint Size       : 8;  /* UINT */
+   GLuint DataType   : 4;  /* TGSI_IMM_ */
+   GLuint Padding    : 15;
+   GLuint Extended   : 1;  /* BOOL */
+};
+
+struct tgsi_immediate_float32
+{
+   GLfloat Float;
+};
+
+/*
+ * GL_NV_vertex_program
+ */
+#define TGSI_OPCODE_ARL                 0
+#define TGSI_OPCODE_MOV                 1
+#define TGSI_OPCODE_LIT                 2
+#define TGSI_OPCODE_RCP                 3
+#define TGSI_OPCODE_RSQ                 4
+#define TGSI_OPCODE_EXP                 5
+#define TGSI_OPCODE_LOG                 6
+#define TGSI_OPCODE_MUL                 7
+#define TGSI_OPCODE_ADD                 8
+#define TGSI_OPCODE_DP3                 9
+#define TGSI_OPCODE_DP4                 10
+#define TGSI_OPCODE_DST                 11
+#define TGSI_OPCODE_MIN                 12
+#define TGSI_OPCODE_MAX                 13
+#define TGSI_OPCODE_SLT                 14
+#define TGSI_OPCODE_SGE                 15
+#define TGSI_OPCODE_MAD                 16
+
+/*
+ * GL_ATI_fragment_shader
+ */
+/* TGSI_OPCODE_MOV */
+/* TGSI_OPCODE_ADD */
+/* TGSI_OPCODE_MUL */
+#define TGSI_OPCODE_SUB                 17
+#define TGSI_OPCODE_DOT3                TGSI_OPCODE_DP3
+#define TGSI_OPCODE_DOT4                TGSI_OPCODE_DP4
+/* TGSI_OPCODE_MAD */
+#define TGSI_OPCODE_LERP                18
+#define TGSI_OPCODE_CND                 19
+#define TGSI_OPCODE_CND0                20
+#define TGSI_OPCODE_DOT2ADD             21
+
+/*
+ * GL_EXT_vertex_shader
+ */
+#define TGSI_OPCODE_INDEX               22
+#define TGSI_OPCODE_NEGATE              23
+/* TGSI_OPCODE_DOT3 */
+/* TGSI_OPCODE_DOT4 */
+/* TGSI_OPCODE_MUL */
+/* TGSI_OPCODE_ADD */
+#define TGSI_OPCODE_MADD                TGSI_OPCODE_MAD
+#define TGSI_OPCODE_FRAC                24
+/* TGSI_OPCODE_MAX */
+/* TGSI_OPCODE_MIN */
+#define TGSI_OPCODE_SETGE               TGSI_OPCODE_SGE
+#define TGSI_OPCODE_SETLT               TGSI_OPCODE_SLT
+#define TGSI_OPCODE_CLAMP               25
+#define TGSI_OPCODE_FLOOR               26
+#define TGSI_OPCODE_ROUND               27
+#define TGSI_OPCODE_EXPBASE2            28
+#define TGSI_OPCODE_LOGBASE2            29
+#define TGSI_OPCODE_POWER               30
+#define TGSI_OPCODE_RECIP               TGSI_OPCODE_RCP
+#define TGSI_OPCODE_RECIPSQRT           TGSI_OPCODE_RSQ
+/* TGSI_OPCODE_SUB */
+#define TGSI_OPCODE_CROSSPRODUCT        31
+#define TGSI_OPCODE_MULTIPLYMATRIX      32
+/* TGSI_OPCODE_MOV */
+
+/*
+ * GL_NV_vertex_program1_1
+ */
+/* TGSI_OPCODE_ARL */
+/* TGSI_OPCODE_MOV */
+/* TGSI_OPCODE_LIT */
+#define TGSI_OPCODE_ABS                 33
+/* TGSI_OPCODE_RCP */
+/* TGSI_OPCODE_RSQ */
+/* TGSI_OPCODE_EXP */
+/* TGSI_OPCODE_LOG */
+#define TGSI_OPCODE_RCC                 34
+/* TGSI_OPCODE_MUL */
+/* TGSI_OPCODE_ADD */
+/* TGSI_OPCODE_DP3 */
+/* TGSI_OPCODE_DP4 */
+/* TGSI_OPCODE_DST */
+/* TGSI_OPCODE_MIN */
+/* TGSI_OPCODE_MAX */
+/* TGSI_OPCODE_SLT */
+/* TGSI_OPCODE_SGE */
+#define TGSI_OPCODE_DPH                 35
+/* TGSI_OPCODE_SUB */
+/* TGSI_OPCODE_MAD */
+
+/*
+ * GL_NV_fragment_program
+ */
+/* TGSI_OPCODE_ADD */
+#define TGSI_OPCODE_COS                 36
+#define TGSI_OPCODE_DDX                 37
+#define TGSI_OPCODE_DDY                 38
+/* TGSI_OPCODE_DP3 */
+/* TGSI_OPCODE_DP4 */
+/* TGSI_OPCODE_DST */
+#define TGSI_OPCODE_EX2                 TGSI_OPCODE_EXPBASE2
+#define TGSI_OPCODE_FLR                 TGSI_OPCODE_FLOOR
+#define TGSI_OPCODE_FRC                 TGSI_OPCODE_FRAC
+#define TGSI_OPCODE_KIL                 39
+#define TGSI_OPCODE_LG2                 TGSI_OPCODE_LOGBASE2
+/* TGSI_OPCODE_LIT */
+#define TGSI_OPCODE_LRP                 TGSI_OPCODE_LERP
+/* TGSI_OPCODE_MAD */
+/* TGSI_OPCODE_MAX */
+/* TGSI_OPCODE_MIN */
+/* TGSI_OPCODE_MOV */
+/* TGSI_OPCODE_MUL */
+#define TGSI_OPCODE_PK2H                40
+#define TGSI_OPCODE_PK2US               41
+#define TGSI_OPCODE_PK4B                42
+#define TGSI_OPCODE_PK4UB               43
+#define TGSI_OPCODE_POW                 TGSI_OPCODE_POWER
+/* TGSI_OPCODE_RCP */
+#define TGSI_OPCODE_RFL                 44
+/* TGSI_OPCODE_RSQ */
+#define TGSI_OPCODE_SEQ                 45
+#define TGSI_OPCODE_SFL                 46
+/* TGSI_OPCODE_SGE */
+#define TGSI_OPCODE_SGT                 47
+#define TGSI_OPCODE_SIN                 48
+#define TGSI_OPCODE_SLE                 49
+/* TGSI_OPCODE_SLT */
+#define TGSI_OPCODE_SNE                 50
+#define TGSI_OPCODE_STR                 51
+/* TGSI_OPCODE_SUB */
+#define TGSI_OPCODE_TEX                 52
+#define TGSI_OPCODE_TXD                 53
+/* TGSI_OPCODE_TXP - use TGSI_OPCODE_TEX */
+#define TGSI_OPCODE_UP2H                54
+#define TGSI_OPCODE_UP2US               55
+#define TGSI_OPCODE_UP4B                56
+#define TGSI_OPCODE_UP4UB               57
+#define TGSI_OPCODE_X2D                 58
+
+/*
+ * GL_NV_vertex_program2
+ */
+/* TGSI_OPCODE_ABS */
+/* TGSI_OPCODE_ADD */
+#define TGSI_OPCODE_ARA                 59
+/* TGSI_OPCODE_ARL */
+#define TGSI_OPCODE_ARR                 60
+#define TGSI_OPCODE_BRA                 61
+#define TGSI_OPCODE_CAL                 62
+/* TGSI_OPCODE_COS */
+/* TGSI_OPCODE_DP3 */
+/* TGSI_OPCODE_DP4 */
+/* TGSI_OPCODE_DPH */
+/* TGSI_OPCODE_DST */
+/* TGSI_OPCODE_EX2 */
+/* TGSI_OPCODE_EXP */
+/* TGSI_OPCODE_FLR */
+/* TGSI_OPCODE_FRC */
+/* TGSI_OPCODE_LG2 */
+/* TGSI_OPCODE_LIT */
+/* TGSI_OPCODE_LOG */
+/* TGSI_OPCODE_MAD */
+/* TGSI_OPCODE_MAX */
+/* TGSI_OPCODE_MIN */
+/* TGSI_OPCODE_MOV */
+/* TGSI_OPCODE_MUL */
+/* TGSI_OPCODE_RCC */
+/* TGSI_OPCODE_RCP */
+#define TGSI_OPCODE_RET                 63
+/* TGSI_OPCODE_RSQNV */
+/* TGSI_OPCODE_SEQ */
+/* TGSI_OPCODE_SFL */
+/* TGSI_OPCODE_SGE */
+/* TGSI_OPCODE_SGT */
+/* TGSI_OPCODE_SIN */
+/* TGSI_OPCODE_SLE */
+/* TGSI_OPCODE_SLT */
+/* TGSI_OPCODE_SNE */
+#define TGSI_OPCODE_SSG                 64
+/* TGSI_OPCODE_STR */
+/* TGSI_OPCODE_SUB */
+
+/*
+ * GL_ARB_vertex_program
+ */
+/* TGSI_OPCODE_ABS */
+/* TGSI_OPCODE_ADD */
+/* TGSI_OPCODE_ARL */
+/* TGSI_OPCODE_DP3 */
+/* TGSI_OPCODE_DP4 */
+/* TGSI_OPCODE_DPH */
+/* TGSI_OPCODE_DST */
+/* TGSI_OPCODE_EX2 */
+/* TGSI_OPCODE_EXP */
+/* TGSI_OPCODE_FLR */
+/* TGSI_OPCODE_FRC */
+/* TGSI_OPCODE_LG2 */
+/* TGSI_OPCODE_LIT */
+/* TGSI_OPCODE_LOG */
+/* TGSI_OPCODE_MAD */
+/* TGSI_OPCODE_MAX */
+/* TGSI_OPCODE_MIN */
+/* TGSI_OPCODE_MOV */
+/* TGSI_OPCODE_MUL */
+/* TGSI_OPCODE_POW */
+/* TGSI_OPCODE_RCP */
+/* TGSI_OPCODE_RSQ */
+/* TGSI_OPCODE_SGE */
+/* TGSI_OPCODE_SLT */
+/* TGSI_OPCODE_SUB */
+#define TGSI_OPCODE_SWZ                 TGSI_OPCODE_MOV
+#define TGSI_OPCODE_XPD                 TGSI_OPCODE_CROSSPRODUCT
+
+/*
+ * GL_ARB_fragment_program
+ */
+/* TGSI_OPCODE_ABS */
+/* TGSI_OPCODE_ADD */
+#define TGSI_OPCODE_CMP                 65
+/* TGSI_OPCODE_COS */
+/* TGSI_OPCODE_DP3 */
+/* TGSI_OPCODE_DP4 */
+/* TGSI_OPCODE_DPH */
+/* TGSI_OPCODE_DST */
+/* TGSI_OPCODE_EX2 */
+/* TGSI_OPCODE_FLR */
+/* TGSI_OPCODE_FRC */
+/* TGSI_OPCODE_LG2 */
+/* TGSI_OPCODE_LIT */
+/* TGSI_OPCODE_LRP */
+/* TGSI_OPCODE_MAD */
+/* TGSI_OPCODE_MAX */
+/* TGSI_OPCODE_MIN */
+/* TGSI_OPCODE_MOV */
+/* TGSI_OPCODE_MUL */
+/* TGSI_OPCODE_POW */
+/* TGSI_OPCODE_RCP */
+/* TGSI_OPCODE_RSQ */
+#define TGSI_OPCODE_SCS                 66
+/* TGSI_OPCODE_SGE */
+/* TGSI_OPCODE_SIN */
+/* TGSI_OPCODE_SLT */
+/* TGSI_OPCODE_SUB */
+/* TGSI_OPCODE_SWZ */
+/* TGSI_OPCODE_XPD */
+/* TGSI_OPCODE_TEX */
+/* TGSI_OPCODE_TXP */
+#define TGSI_OPCODE_TXB                 67
+/* TGSI_OPCODE_KIL */
+
+/*
+ * GL_NV_fragment_program_option
+ */
+/* TGSI_OPCODE_ABS */
+/* TGSI_OPCODE_FLR */
+/* TGSI_OPCODE_FRC */
+/* TGSI_OPCODE_LIT */
+/* TGSI_OPCODE_MOV */
+/* TGSI_OPCODE_DDX */
+/* TGSI_OPCODE_DDY */
+/* TGSI_OPCODE_PK2H */
+/* TGSI_OPCODE_PK2US */
+/* TGSI_OPCODE_PK4B */
+/* TGSI_OPCODE_PK4UB */
+/* TGSI_OPCODE_COS */
+/* TGSI_OPCODE_EX2 */
+/* TGSI_OPCODE_LG2 */
+/* TGSI_OPCODE_RCP */
+/* TGSI_OPCODE_RSQ */
+/* TGSI_OPCODE_SIN */
+/* TGSI_OPCODE_SCS */
+/* TGSI_OPCODE_UP2H */
+/* TGSI_OPCODE_UP2US */
+/* TGSI_OPCODE_UP4B */
+/* TGSI_OPCODE_UP4UB */
+/* TGSI_OPCODE_POW */
+/* TGSI_OPCODE_ADD */
+/* TGSI_OPCODE_DP3 */
+/* TGSI_OPCODE_DP4 */
+/* TGSI_OPCODE_DPH */
+/* TGSI_OPCODE_DST */
+/* TGSI_OPCODE_MAX */
+/* TGSI_OPCODE_MIN */
+/* TGSI_OPCODE_MUL */
+/* TGSI_OPCODE_SGE */
+/* TGSI_OPCODE_SLT */
+/* TGSI_OPCODE_SUB */
+/* TGSI_OPCODE_XPD */
+/* TGSI_OPCODE_RFL */
+/* TGSI_OPCODE_SEQ */
+/* TGSI_OPCODE_SFL */
+/* TGSI_OPCODE_SGT */
+/* TGSI_OPCODE_SLE */
+/* TGSI_OPCODE_SNE */
+/* TGSI_OPCODE_STR */
+/* TGSI_OPCODE_CMP */
+/* TGSI_OPCODE_LRP */
+/* TGSI_OPCODE_MAD */
+/* TGSI_OPCODE_X2D */
+/* TGSI_OPCODE_SWZ */
+/* TGSI_OPCODE_TEX */
+/* TGSI_OPCODE_TXP */
+/* TGSI_OPCODE_TXB */
+/* TGSI_OPCODE_KIL */
+/* TGSI_OPCODE_TXD */
+
+/*
+ * GL_NV_fragment_program2
+ */
+/* TGSI_OPCODE_ABS */
+/* TGSI_OPCODE_FLR */
+/* TGSI_OPCODE_FRC */
+/* TGSI_OPCODE_LIT */
+/* TGSI_OPCODE_MOV */
+/* TGSI_OPCODE_DDX */
+/* TGSI_OPCODE_DDY */
+/* TGSI_OPCODE_PK2H */
+/* TGSI_OPCODE_PK2US */
+/* TGSI_OPCODE_PK4B */
+/* TGSI_OPCODE_PK4UB */
+#define TGSI_OPCODE_NRM                 68
+#define TGSI_OPCODE_DIV                 69
+/* TGSI_OPCODE_COS */
+/* TGSI_OPCODE_EX2 */
+/* TGSI_OPCODE_LG2 */
+/* TGSI_OPCODE_RCP */
+/* TGSI_OPCODE_RSQ */
+/* TGSI_OPCODE_SIN */
+/* TGSI_OPCODE_SCS */
+/* TGSI_OPCODE_UP2H */
+/* TGSI_OPCODE_UP2US */
+/* TGSI_OPCODE_UP4B */
+/* TGSI_OPCODE_UP4UB */
+/* TGSI_OPCODE_POW */
+/* TGSI_OPCODE_ADD */
+/* TGSI_OPCODE_DP3 */
+/* TGSI_OPCODE_DP4 */
+/* TGSI_OPCODE_DPH */
+/* TGSI_OPCODE_DST */
+/* TGSI_OPCODE_MAX */
+/* TGSI_OPCODE_MIN */
+/* TGSI_OPCODE_MUL */
+/* TGSI_OPCODE_SGE */
+/* TGSI_OPCODE_SLT */
+/* TGSI_OPCODE_SUB */
+/* TGSI_OPCODE_XPD */
+/* TGSI_OPCODE_RFL */
+/* TGSI_OPCODE_SEQ */
+/* TGSI_OPCODE_SFL */
+/* TGSI_OPCODE_SGT */
+/* TGSI_OPCODE_SLE */
+/* TGSI_OPCODE_SNE */
+/* TGSI_OPCODE_STR */
+#define TGSI_OPCODE_DP2                 70
+/* TGSI_OPCODE_CMP */
+/* TGSI_OPCODE_LRP */
+/* TGSI_OPCODE_MAD */
+/* TGSI_OPCODE_X2D */
+#define TGSI_OPCODE_DP2A                TGSI_OPCODE_DOT2ADD
+/* TGSI_OPCODE_SWZ */
+/* TGSI_OPCODE_TEX */
+/* TGSI_OPCODE_TXP */
+/* TGSI_OPCODE_TXB */
+#define TGSI_OPCODE_TXL                 71
+/* TGSI_OPCODE_KIL */
+/* TGSI_OPCODE_TXD */
+/* TGSI_OPCODE_CAL */
+/* TGSI_OPCODE_RET */
+#define TGSI_OPCODE_BRK                 72
+#define TGSI_OPCODE_IF                  73
+#define TGSI_OPCODE_LOOP                74
+#define TGSI_OPCODE_REP                 75
+#define TGSI_OPCODE_ELSE                76
+#define TGSI_OPCODE_ENDIF               77
+#define TGSI_OPCODE_ENDLOOP             78
+#define TGSI_OPCODE_ENDREP              79
+
+/*
+ * GL_NV_vertex_program2_option
+ */
+/* TGSI_OPCODE_ARL */
+/* TGSI_OPCODE_ABS */
+/* TGSI_OPCODE_FLR */
+/* TGSI_OPCODE_FRC */
+/* TGSI_OPCODE_LIT */
+/* TGSI_OPCODE_MOV */
+/* TGSI_OPCODE_SSG */
+/* TGSI_OPCODE_EX2 */
+/* TGSI_OPCODE_EXP */
+/* TGSI_OPCODE_LG2 */
+/* TGSI_OPCODE_LOG */
+/* TGSI_OPCODE_RCP */
+/* TGSI_OPCODE_RSQ */
+/* TGSI_OPCODE_COS */
+/* TGSI_OPCODE_RCC */
+/* TGSI_OPCODE_SIN */
+/* TGSI_OPCODE_POW */
+/* TGSI_OPCODE_ADD */
+/* TGSI_OPCODE_DP3 */
+/* TGSI_OPCODE_DP4 */
+/* TGSI_OPCODE_DPH */
+/* TGSI_OPCODE_DST */
+/* TGSI_OPCODE_MAX */
+/* TGSI_OPCODE_MIN */
+/* TGSI_OPCODE_MUL */
+/* TGSI_OPCODE_SGE */
+/* TGSI_OPCODE_SLT */
+/* TGSI_OPCODE_SUB */
+/* TGSI_OPCODE_XPD */
+/* TGSI_OPCODE_SEQ */
+/* TGSI_OPCODE_SFL */
+/* TGSI_OPCODE_SGT */
+/* TGSI_OPCODE_SLE */
+/* TGSI_OPCODE_SNE */
+/* TGSI_OPCODE_STR */
+/* TGSI_OPCODE_MAD */
+/* TGSI_OPCODE_SWZ */
+/* TGSI_OPCODE_ARR */
+/* TGSI_OPCODE_ARA */
+/* TGSI_OPCODE_BRA */
+/* TGSI_OPCODE_CAL */
+/* TGSI_OPCODE_RET */
+
+/*
+ * GL_NV_vertex_program3
+ */
+/* TGSI_OPCODE_ARL */
+/* TGSI_OPCODE_ABS */
+/* TGSI_OPCODE_FLR */
+/* TGSI_OPCODE_FRC */
+/* TGSI_OPCODE_LIT */
+/* TGSI_OPCODE_MOV */
+/* TGSI_OPCODE_SSG */
+/* TGSI_OPCODE_EX2 */
+/* TGSI_OPCODE_EXP */
+/* TGSI_OPCODE_LG2 */
+/* TGSI_OPCODE_LOG */
+/* TGSI_OPCODE_RCP */
+/* TGSI_OPCODE_RSQ */
+/* TGSI_OPCODE_COS */
+/* TGSI_OPCODE_RCC */
+/* TGSI_OPCODE_SIN */
+/* TGSI_OPCODE_POW */
+/* TGSI_OPCODE_ADD */
+/* TGSI_OPCODE_DP3 */
+/* TGSI_OPCODE_DP4 */
+/* TGSI_OPCODE_DPH */
+/* TGSI_OPCODE_DST */
+/* TGSI_OPCODE_MAX */
+/* TGSI_OPCODE_MIN */
+/* TGSI_OPCODE_MUL */
+/* TGSI_OPCODE_SGE */
+/* TGSI_OPCODE_SLT */
+/* TGSI_OPCODE_SUB */
+/* TGSI_OPCODE_XPD */
+/* TGSI_OPCODE_SEQ */
+/* TGSI_OPCODE_SFL */
+/* TGSI_OPCODE_SGT */
+/* TGSI_OPCODE_SLE */
+/* TGSI_OPCODE_SNE */
+/* TGSI_OPCODE_STR */
+/* TGSI_OPCODE_MAD */
+/* TGSI_OPCODE_SWZ */
+/* TGSI_OPCODE_ARR */
+/* TGSI_OPCODE_ARA */
+/* TGSI_OPCODE_BRA */
+/* TGSI_OPCODE_CAL */
+/* TGSI_OPCODE_RET */
+#define TGSI_OPCODE_PUSHA               80
+#define TGSI_OPCODE_POPA                81
+/* TGSI_OPCODE_TEX */
+/* TGSI_OPCODE_TXP */
+/* TGSI_OPCODE_TXB */
+/* TGSI_OPCODE_TXL */
+
+/*
+ * GL_NV_gpu_program4
+ */
+/* TGSI_OPCODE_ABS */
+#define TGSI_OPCODE_CEIL                82
+/* TGSI_OPCODE_FLR */
+/* TGSI_OPCODE_FRC */
+#define TGSI_OPCODE_I2F                 83
+/* TGSI_OPCODE_LIT */
+/* TGSI_OPCODE_MOV */
+#define TGSI_OPCODE_NOT                 84
+/* TGSI_OPCODE_NRM */
+/* TGSI_OPCODE_PK2H */
+/* TGSI_OPCODE_PK2US */
+/* TGSI_OPCODE_PK4B */
+/* TGSI_OPCODE_PK4UB */
+/* TGSI_OPCODE_ROUND */
+/* TGSI_OPCODE_SSG */
+#define TGSI_OPCODE_TRUNC               85
+/* TGSI_OPCODE_COS */
+/* TGSI_OPCODE_EX2 */
+/* TGSI_OPCODE_LG2 */
+/* TGSI_OPCODE_RCC */
+/* TGSI_OPCODE_RCP */
+/* TGSI_OPCODE_RSQ */
+/* TGSI_OPCODE_SCS */
+/* TGSI_OPCODE_SIN */
+/* TGSI_OPCODE_UP2H */
+/* TGSI_OPCODE_UP2US */
+/* TGSI_OPCODE_UP4B */
+/* TGSI_OPCODE_UP4UB */
+/* TGSI_OPCODE_POW */
+/* TGSI_OPCODE_DIV */
+#define TGSI_OPCODE_SHL                 86
+#define TGSI_OPCODE_SHR                 87
+/* TGSI_OPCODE_ADD */
+#define TGSI_OPCODE_AND                 88
+/* TGSI_OPCODE_DP3 */
+/* TGSI_OPCODE_DP4 */
+/* TGSI_OPCODE_DPH */
+/* TGSI_OPCODE_DST */
+/* TGSI_OPCODE_MAX */
+/* TGSI_OPCODE_MIN */
+/* TGSI_OPCODE_MUL */
+#define TGSI_OPCODE_OR                  89
+/* TGSI_OPCODE_RFL */
+/* TGSI_OPCODE_SEQ */
+/* TGSI_OPCODE_SFL */
+/* TGSI_OPCODE_SGE */
+/* TGSI_OPCODE_SGT */
+/* TGSI_OPCODE_SLE */
+/* TGSI_OPCODE_SLT */
+/* TGSI_OPCODE_SNE */
+/* TGSI_OPCODE_STR */
+/* TGSI_OPCODE_SUB */
+/* TGSI_OPCODE_XPD */
+/* TGSI_OPCODE_DP2 */
+#define TGSI_OPCODE_MOD                 90
+#define TGSI_OPCODE_XOR                 91
+/* TGSI_OPCODE_CMP */
+/* TGSI_OPCODE_DP2A */
+/* TGSI_OPCODE_LRP */
+/* TGSI_OPCODE_MAD */
+#define TGSI_OPCODE_SAD                 92
+/* TGSI_OPCODE_X2D */
+/* TGSI_OPCODE_SWZ */
+/* TGSI_OPCODE_TEX */
+/* TGSI_OPCODE_TXB */
+#define TGSI_OPCODE_TXF                 93
+/* TGSI_OPCODE_TXL */
+/* TGSI_OPCODE_TXP */
+#define TGSI_OPCODE_TXQ                 94
+/* TGSI_OPCODE_TXD */
+/* TGSI_OPCODE_CAL */
+/* TGSI_OPCODE_RET */
+/* TGSI_OPCODE_BRK */
+#define TGSI_OPCODE_CONT                95
+/* TGSI_OPCODE_IF */
+/* TGSI_OPCODE_REP */
+/* TGSI_OPCODE_ELSE */
+/* TGSI_OPCODE_ENDIF */
+/* TGSI_OPCODE_ENDREP */
+
+/*
+ * GL_NV_vertex_program4
+ */
+/* Same as GL_NV_gpu_program4 */
+
+/*
+ * GL_NV_fragment_program4
+ */
+/* Same as GL_NV_gpu_program4 */
+/* TGSI_OPCODE_KIL */
+/* TGSI_OPCODE_DDX */
+/* TGSI_OPCODE_DDY */
+
+/*
+ * GL_NV_geometry_program4
+ */
+/* Same as GL_NV_gpu_program4 */
+#define TGSI_OPCODE_EMIT                96
+#define TGSI_OPCODE_ENDPRIM             97
+
+#define TGSI_OPCODE_LAST                98
+
+#define TGSI_SAT_NONE            0  /* do not saturate */
+#define TGSI_SAT_ZERO_ONE        1  /* clamp to [0,1] */
+#define TGSI_SAT_MINUS_PLUS_ONE  2  /* clamp to [-1,1] */
+
+/*
+ * Opcode is the operation code to execute. A given operation defines the
+ * semantics how the source registers (if any) are interpreted and what is
+ * written to the destination registers (if any) as a result of execution.
+ *
+ * NumDstRegs and NumSrcRegs is the number of destination and source registers,
+ * respectively. For a given operation code, those numbers are fixed and are
+ * present here only for convenience.
+ *
+ * If Extended is TRUE, it is now executed.
+ *
+ * Saturate controls how are final results in destination registers modified.
+ */
+
+struct tgsi_instruction
+{
+   GLuint Type       : 4;  /* TGSI_TOKEN_TYPE_INSTRUCTION */
+   GLuint Size       : 8;  /* UINT */
+   GLuint Opcode     : 8;  /* TGSI_OPCODE_ */
+   GLuint Saturate   : 2;  /* TGSI_SAT_ */
+   GLuint NumDstRegs : 2;  /* UINT */
+   GLuint NumSrcRegs : 4;  /* UINT */
+   GLuint Padding    : 3;
+   GLuint Extended   : 1;  /* BOOL */
+};
+
+/*
+ * If tgsi_instruction::Extended is TRUE, tgsi_instruction_ext follows.
+ * 
+ * Then, tgsi_instruction::NumDstRegs of tgsi_dst_register follow.
+ * 
+ * Then, tgsi_instruction::NumSrcRegs of tgsi_src_register follow.
+ *
+ * tgsi_instruction::Size contains the total number of words that make the
+ * instruction, including the instruction word.
+ */
+
+#define TGSI_INSTRUCTION_EXT_TYPE_NV        0
+#define TGSI_INSTRUCTION_EXT_TYPE_LABEL     1
+#define TGSI_INSTRUCTION_EXT_TYPE_TEXTURE   2
+
+struct tgsi_instruction_ext
+{
+   GLuint Type       : 4;  /* TGSI_INSTRUCTION_EXT_TYPE_ */
+   GLuint Padding    : 27;
+   GLuint Extended   : 1;  /* BOOL */
+};
+
+/*
+ * If tgsi_instruction_ext::Type is TGSI_INSTRUCTION_EXT_TYPE_NV, it should
+ * be cast to tgsi_instruction_ext_nv.
+ * 
+ * If tgsi_instruction_ext::Type is TGSI_INSTRUCTION_EXT_TYPE_LABEL, it
+ * should be cast to tgsi_instruction_ext_label.
+ * 
+ * If tgsi_instruction_ext::Type is TGSI_INSTRUCTION_EXT_TYPE_TEXTURE, it
+ * should be cast to tgsi_instruction_ext_texture.
+ * 
+ * If tgsi_instruction_ext::Extended is TRUE, another tgsi_instruction_ext
+ * follows.
+ */
+
+#define TGSI_PRECISION_DEFAULT      0
+#define TGSI_PRECISION_FLOAT32      1
+#define TGSI_PRECISION_FLOAT16      2
+#define TGSI_PRECISION_FIXED12      3
+
+#define TGSI_CC_GT      0
+#define TGSI_CC_EQ      1
+#define TGSI_CC_LT      2
+#define TGSI_CC_UN      3
+#define TGSI_CC_GE      4
+#define TGSI_CC_LE      5
+#define TGSI_CC_NE      6
+#define TGSI_CC_TR      7
+#define TGSI_CC_FL      8
+
+#define TGSI_SWIZZLE_X      0
+#define TGSI_SWIZZLE_Y      1
+#define TGSI_SWIZZLE_Z      2
+#define TGSI_SWIZZLE_W      3
+
+/*
+ * Precision controls the precision at which the operation should be executed.
+ *
+ * CondDstUpdate enables condition code register writes. When this field is
+ * TRUE, CondDstIndex specifies the index of the condition code register to
+ * update.
+ *
+ * CondFlowEnable enables conditional execution of the operation. When this
+ * field is TRUE, CondFlowIndex specifies the index of the condition code
+ * register to test against CondMask with component swizzle controled by
+ * CondSwizzleX, CondSwizzleY, CondSwizzleZ and CondSwizzleW. If the test fails,
+ * the operation is not executed.
+ */
+
+struct tgsi_instruction_ext_nv
+{
+   GLuint Type             : 4;    /* TGSI_INSTRUCTION_EXT_TYPE_NV */
+   GLuint Precision        : 4;    /* TGSI_PRECISION_ */
+   GLuint CondDstIndex     : 4;    /* UINT */
+   GLuint CondFlowIndex    : 4;    /* UINT */
+   GLuint CondMask         : 4;    /* TGSI_CC_ */
+   GLuint CondSwizzleX     : 2;    /* TGSI_SWIZZLE_ */
+   GLuint CondSwizzleY     : 2;    /* TGSI_SWIZZLE_ */
+   GLuint CondSwizzleZ     : 2;    /* TGSI_SWIZZLE_ */
+   GLuint CondSwizzleW     : 2;    /* TGSI_SWIZZLE_ */
+   GLuint CondDstUpdate    : 1;    /* BOOL */
+   GLuint CondFlowEnable   : 1;    /* BOOL */
+   GLuint Padding          : 1;
+   GLuint Extended         : 1;    /* BOOL */
+};
+
+struct tgsi_instruction_ext_label
+{
+   GLuint Type     : 4;    /* TGSI_INSTRUCTION_EXT_TYPE_LABEL */
+   GLuint Label    : 24;   /* UINT */
+   GLuint Target   : 1;    /* BOOL */
+   GLuint Padding  : 2;
+   GLuint Extended : 1;    /* BOOL */
+};
+
+#define TGSI_TEXTURE_UNKNOWN        0
+#define TGSI_TEXTURE_1D             1
+#define TGSI_TEXTURE_2D             2
+#define TGSI_TEXTURE_3D             3
+#define TGSI_TEXTURE_CUBE           4
+#define TGSI_TEXTURE_RECT           5
+#define TGSI_TEXTURE_SHADOW1D       6
+#define TGSI_TEXTURE_SHADOW2D       7
+#define TGSI_TEXTURE_SHADOWRECT     8
+
+struct tgsi_instruction_ext_texture
+{
+   GLuint Type     : 4;    /* TGSI_INSTRUCTION_EXT_TYPE_TEXTURE */
+   GLuint Texture  : 8;    /* TGSI_TEXTURE_ */
+   GLuint Padding  : 19;
+   GLuint Extended : 1;    /* BOOL */
+};
+
+/*
+ * File specifies the register array to access.
+ *
+ * Index specifies the element number of a register in the register file.
+ *
+ * If Indirect is TRUE, Index should be offset by the X component of a source
+ * register that follows. The register can be now fetched into local storage
+ * for further processing.
+ *
+ * If Negate is TRUE, all components of the fetched register are negated.
+ *
+ * The fetched register components are swizzled according to SwizzleX, SwizzleY,
+ * SwizzleZ and SwizzleW.
+ *
+ * If Extended is TRUE, any further modifications to the source register are
+ * made to this temporary storage.
+ */
+
+struct tgsi_src_register
+{
+   GLuint File         : 4;    /* TGSI_FILE_ */
+   GLuint SwizzleX     : 2;    /* TGSI_SWIZZLE_ */
+   GLuint SwizzleY     : 2;    /* TGSI_SWIZZLE_ */
+   GLuint SwizzleZ     : 2;    /* TGSI_SWIZZLE_ */
+   GLuint SwizzleW     : 2;    /* TGSI_SWIZZLE_ */
+   GLuint Negate       : 1;    /* BOOL */
+   GLuint Indirect     : 1;    /* BOOL */
+   GLuint Dimension    : 1;    /* BOOL */
+   GLint  Index        : 16;   /* SINT */
+   GLuint Extended     : 1;    /* BOOL */
+};
+
+/*
+ * If tgsi_src_register::Extended is TRUE, tgsi_src_register_ext follows.
+ * 
+ * Then, if tgsi_src_register::Indirect is TRUE, another tgsi_src_register
+ * follows.
+ *
+ * Then, if tgsi_src_register::Dimension is TRUE, tgsi_dimension follows.
+ */
+
+#define TGSI_SRC_REGISTER_EXT_TYPE_SWZ      0
+#define TGSI_SRC_REGISTER_EXT_TYPE_MOD      1
+
+struct tgsi_src_register_ext
+{
+   GLuint Type     : 4;    /* TGSI_SRC_REGISTER_EXT_TYPE_ */
+   GLuint Padding  : 27;
+   GLuint Extended : 1;    /* BOOL */
+};
+
+/*
+ * If tgsi_src_register_ext::Type is TGSI_SRC_REGISTER_EXT_TYPE_SWZ,
+ * it should be cast to tgsi_src_register_ext_extswz.
+ * 
+ * If tgsi_src_register_ext::Type is TGSI_SRC_REGISTER_EXT_TYPE_MOD,
+ * it should be cast to tgsi_src_register_ext_mod.
+ * 
+ * If tgsi_dst_register_ext::Extended is TRUE, another tgsi_dst_register_ext
+ * follows.
+ */
+
+#define TGSI_EXTSWIZZLE_X       TGSI_SWIZZLE_X
+#define TGSI_EXTSWIZZLE_Y       TGSI_SWIZZLE_Y
+#define TGSI_EXTSWIZZLE_Z       TGSI_SWIZZLE_Z
+#define TGSI_EXTSWIZZLE_W       TGSI_SWIZZLE_W
+#define TGSI_EXTSWIZZLE_ZERO    4
+#define TGSI_EXTSWIZZLE_ONE     5
+
+/*
+ * ExtSwizzleX, ExtSwizzleY, ExtSwizzleZ and ExtSwizzleW swizzle the source
+ * register in an extended manner.
+ *
+ * NegateX, NegateY, NegateZ and NegateW negate individual components of the
+ * source register.
+ *
+ * ExtDivide specifies which component is used to divide all components of the
+ * source register.
+ */
+
+struct tgsi_src_register_ext_swz
+{
+   GLuint Type         : 4;    /* TGSI_SRC_REGISTER_EXT_TYPE_SWZ */
+   GLuint ExtSwizzleX  : 4;    /* TGSI_EXTSWIZZLE_ */
+   GLuint ExtSwizzleY  : 4;    /* TGSI_EXTSWIZZLE_ */
+   GLuint ExtSwizzleZ  : 4;    /* TGSI_EXTSWIZZLE_ */
+   GLuint ExtSwizzleW  : 4;    /* TGSI_EXTSWIZZLE_ */
+   GLuint NegateX      : 1;    /* BOOL */
+   GLuint NegateY      : 1;    /* BOOL */
+   GLuint NegateZ      : 1;    /* BOOL */
+   GLuint NegateW      : 1;    /* BOOL */
+   GLuint ExtDivide    : 4;    /* TGSI_EXTSWIZZLE_ */
+   GLuint Padding      : 3;
+   GLuint Extended     : 1;    /* BOOL */
+};
+
+/*
+ * If Complement is TRUE, the source register is modified by subtracting it
+ * from 1.0.
+ *
+ * If Bias is TRUE, the source register is modified by subtracting 0.5 from it.
+ *
+ * If Scale2X is TRUE, the source register is modified by multiplying it by 2.0.
+ *
+ * If Absolute is TRUE, the source register is modified by removing the sign.
+ *
+ * If Negate is TRUE, the source register is modified by negating it.
+ */
+
+struct tgsi_src_register_ext_mod
+{
+   GLuint Type         : 4;    /* TGSI_SRC_REGISTER_EXT_TYPE_MOD */
+   GLuint Complement   : 1;    /* BOOL */
+   GLuint Bias         : 1;    /* BOOL */
+   GLuint Scale2X      : 1;    /* BOOL */
+   GLuint Absolute     : 1;    /* BOOL */
+   GLuint Negate       : 1;    /* BOOL */
+   GLuint Padding      : 22;
+   GLuint Extended     : 1;    /* BOOL */
+};
+
+struct tgsi_dimension
+{
+   GLuint Indirect     : 1;    /* BOOL */
+   GLuint Dimension    : 1;    /* BOOL */
+   GLuint Padding      : 13;
+   GLint  Index        : 16;   /* SINT */
+   GLuint Extended     : 1;    /* BOOL */
+};
+
+#define TGSI_WRITEMASK_NONE     0x00
+#define TGSI_WRITEMASK_X        0x01
+#define TGSI_WRITEMASK_Y        0x02
+#define TGSI_WRITEMASK_XY       0x03
+#define TGSI_WRITEMASK_Z        0x04
+#define TGSI_WRITEMASK_XZ       0x05
+#define TGSI_WRITEMASK_YZ       0x06
+#define TGSI_WRITEMASK_XYZ      0x07
+#define TGSI_WRITEMASK_W        0x08
+#define TGSI_WRITEMASK_XW       0x09
+#define TGSI_WRITEMASK_YW       0x0A
+#define TGSI_WRITEMASK_XYW      0x0B
+#define TGSI_WRITEMASK_ZW       0x0C
+#define TGSI_WRITEMASK_XZW      0x0D
+#define TGSI_WRITEMASK_YZW      0x0E
+#define TGSI_WRITEMASK_XYZW     0x0F
+
+struct tgsi_dst_register
+{
+   GLuint File         : 4;    /* TGSI_FILE_ */
+   GLuint WriteMask    : 4;    /* TGSI_WRITEMASK_ */
+   GLuint Indirect     : 1;    /* BOOL */
+   GLuint Dimension    : 1;    /* BOOL */
+   GLint  Index        : 16;   /* SINT */
+   GLuint Padding      : 5;
+   GLuint Extended     : 1;    /* BOOL */
+};
+
+/*
+ * If tgsi_dst_register::Extended is TRUE, tgsi_dst_register_ext follows.
+ * 
+ * Then, if tgsi_dst_register::Indirect is TRUE, tgsi_src_register follows.
+ */
+
+#define TGSI_DST_REGISTER_EXT_TYPE_CONDCODE     0
+#define TGSI_DST_REGISTER_EXT_TYPE_MODULATE     1
+
+struct tgsi_dst_register_ext
+{
+   GLuint Type     : 4;    /* TGSI_DST_REGISTER_EXT_TYPE_ */
+   GLuint Padding  : 27;
+   GLuint Extended : 1;    /* BOOL */
+};
+
+/*
+ * If tgsi_dst_register_ext::Type is TGSI_DST_REGISTER_EXT_TYPE_CONDCODE,
+ * it should be cast to tgsi_dst_register_ext_condcode.
+ * 
+ * If tgsi_dst_register_ext::Type is TGSI_DST_REGISTER_EXT_TYPE_MODULATE,
+ * it should be cast to tgsi_dst_register_ext_modulate.
+ * 
+ * If tgsi_dst_register_ext::Extended is TRUE, another tgsi_dst_register_ext
+ * follows.
+ */
+
+struct tgsi_dst_register_ext_concode
+{
+   GLuint Type         : 4;    /* TGSI_DST_REGISTER_EXT_TYPE_CONDCODE */
+   GLuint CondMask     : 4;    /* TGSI_CC_ */
+   GLuint CondSwizzleX : 2;    /* TGSI_SWIZZLE_ */
+   GLuint CondSwizzleY : 2;    /* TGSI_SWIZZLE_ */
+   GLuint CondSwizzleZ : 2;    /* TGSI_SWIZZLE_ */
+   GLuint CondSwizzleW : 2;    /* TGSI_SWIZZLE_ */
+   GLuint CondSrcIndex : 4;    /* UINT */
+   GLuint Padding      : 11;
+   GLuint Extended     : 1;    /* BOOL */
+};
+
+#define TGSI_MODULATE_1X        0
+#define TGSI_MODULATE_2X        1
+#define TGSI_MODULATE_4X        2
+#define TGSI_MODULATE_8X        3
+#define TGSI_MODULATE_HALF      4
+#define TGSI_MODULATE_QUARTER   5
+#define TGSI_MODULATE_EIGHTH    6
+
+struct tgsi_dst_register_ext_modulate
+{
+   GLuint Type     : 4;    /* TGSI_DST_REGISTER_EXT_TYPE_MODULATE */
+   GLuint Modulate : 4;    /* TGSI_MODULATE_ */
+   GLuint Padding  : 23;
+   GLuint Extended : 1;    /* BOOL */
+};
+
+#if defined __cplusplus
+} // extern "C"
+#endif // defined __cplusplus
+
+#endif // !defined TGSI_TOKEN_H
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_util.c b/src/mesa/pipe/tgsi/core/tgsi_util.c
new file mode 100644
index 00000000000..2331affdfdf
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_util.c
@@ -0,0 +1,263 @@
+#include "tgsi_platform.h"
+#include "tgsi_core.h"
+
+void *
+tgsi_align_128bit(
+   void *unaligned )
+{
+   GLuint *ptr, addr;
+
+   ptr = (GLuint *) unaligned;
+   addr = (*(GLuint *) &ptr + 15) & ~15;
+   return *(void **) &addr;
+}
+
+GLuint
+tgsi_util_get_src_register_swizzle(
+   const struct tgsi_src_register *reg,
+   GLuint component )
+{
+   switch( component ) {
+   case 0:
+      return reg->SwizzleX;
+   case 1:
+      return reg->SwizzleY;
+   case 2:
+      return reg->SwizzleZ;
+   case 3:
+      return reg->SwizzleW;
+   default:
+      assert( 0 );
+   }
+   return 0;
+}
+
+GLuint
+tgsi_util_get_src_register_extswizzle(
+   const struct   tgsi_src_register_ext_swz *reg,
+   GLuint component )
+{
+   switch( component ) {
+   case 0:
+      return reg->ExtSwizzleX;
+   case 1:
+      return reg->ExtSwizzleY;
+   case 2:
+      return reg->ExtSwizzleZ;
+   case 3:
+      return reg->ExtSwizzleW;
+   default:
+      assert( 0 );
+   }
+   return 0;
+}
+
+GLuint
+tgsi_util_get_full_src_register_extswizzle(
+   const struct tgsi_full_src_register  *reg,
+   GLuint component )
+{
+   GLuint swizzle;
+
+   /*
+    * First, calculate  the   extended swizzle for a given channel. This will give
+    * us either a channel index into the simple swizzle or  a constant 1 or   0.
+    */
+   swizzle = tgsi_util_get_src_register_extswizzle(
+      &reg->SrcRegisterExtSwz,
+      component );
+
+   assert (TGSI_SWIZZLE_X == TGSI_EXTSWIZZLE_X);
+   assert (TGSI_SWIZZLE_Y == TGSI_EXTSWIZZLE_Y);
+   assert (TGSI_SWIZZLE_Z == TGSI_EXTSWIZZLE_Z);
+   assert (TGSI_SWIZZLE_W == TGSI_EXTSWIZZLE_W);
+   assert (TGSI_EXTSWIZZLE_ZERO > TGSI_SWIZZLE_W);
+   assert (TGSI_EXTSWIZZLE_ONE > TGSI_SWIZZLE_W);
+
+   /*
+    * Second, calculate the simple  swizzle  for   the   unswizzled channel index.
+    * Leave the constants intact, they are   not   affected by the   simple swizzle.
+    */
+   if( swizzle <= TGSI_SWIZZLE_W ) {
+      swizzle = tgsi_util_get_src_register_swizzle(
+         &reg->SrcRegister,
+         swizzle );
+   }
+
+   return swizzle;
+}
+
+void
+tgsi_util_set_src_register_swizzle(
+   struct tgsi_src_register *reg,
+   GLuint swizzle,
+   GLuint component )
+{
+   switch( component ) {
+   case 0:
+      reg->SwizzleX = swizzle;
+      break;
+   case 1:
+      reg->SwizzleY = swizzle;
+      break;
+   case 2:
+      reg->SwizzleZ = swizzle;
+      break;
+   case 3:
+      reg->SwizzleW = swizzle;
+      break;
+   default:
+      assert( 0 );
+   }
+}
+
+void
+tgsi_util_set_src_register_extswizzle(
+   struct tgsi_src_register_ext_swz *reg,
+   GLuint swizzle,
+   GLuint component )
+{
+   switch( component ) {
+   case 0:
+      reg->ExtSwizzleX = swizzle;
+      break;
+   case 1:
+      reg->ExtSwizzleY = swizzle;
+      break;
+   case 2:
+      reg->ExtSwizzleZ = swizzle;
+      break;
+   case 3:
+      reg->ExtSwizzleW = swizzle;
+      break;
+   default:
+      assert( 0 );
+   }
+}
+
+GLuint
+tgsi_util_get_src_register_extnegate(
+   const  struct tgsi_src_register_ext_swz *reg,
+   GLuint component )
+{
+   switch( component ) {
+   case 0:
+      return reg->NegateX;
+   case 1:
+      return reg->NegateY;
+   case 2:
+      return reg->NegateZ;
+   case 3:
+      return reg->NegateW;
+   default:
+      assert( 0 );
+   }
+   return 0;
+}
+
+void
+tgsi_util_set_src_register_extnegate(
+   struct tgsi_src_register_ext_swz *reg,
+   GLuint negate,
+   GLuint component )
+{
+   switch( component ) {
+   case 0:
+      reg->NegateX = negate;
+      break;
+   case 1:
+      reg->NegateY = negate;
+      break;
+   case 2:
+      reg->NegateZ = negate;
+      break;
+   case 3:
+      reg->NegateW = negate;
+      break;
+   default:
+      assert( 0 );
+   }
+}
+
+GLuint
+tgsi_util_get_full_src_register_sign_mode(
+   const struct  tgsi_full_src_register *reg,
+   GLuint component )
+{
+   GLuint sign_mode;
+
+   if( reg->SrcRegisterExtMod.Absolute ) {
+      /* Consider only the post-abs negation. */
+
+      if( reg->SrcRegisterExtMod.Negate ) {
+         sign_mode = TGSI_UTIL_SIGN_SET;
+      }
+      else {
+         sign_mode = TGSI_UTIL_SIGN_CLEAR;
+      }
+   }
+   else {
+      /* Accumulate the three negations. */
+
+      GLuint negate;
+
+      negate = reg->SrcRegister.Negate;
+      if( tgsi_util_get_src_register_extnegate( &reg->SrcRegisterExtSwz, component ) ) {
+         negate = !negate;
+      }
+      if( reg->SrcRegisterExtMod.Negate ) {
+         negate = !negate;
+      }
+
+      if( negate ) {
+         sign_mode = TGSI_UTIL_SIGN_TOGGLE;
+      }
+      else {
+         sign_mode = TGSI_UTIL_SIGN_KEEP;
+      }
+   }
+
+   return sign_mode;
+}
+
+void
+tgsi_util_set_full_src_register_sign_mode(
+   struct tgsi_full_src_register *reg,
+   GLuint sign_mode )
+{
+   reg->SrcRegisterExtSwz.NegateX = 0;
+   reg->SrcRegisterExtSwz.NegateY = 0;
+   reg->SrcRegisterExtSwz.NegateZ = 0;
+   reg->SrcRegisterExtSwz.NegateW = 0;
+
+   switch (sign_mode)
+   {
+   case TGSI_UTIL_SIGN_CLEAR:
+      reg->SrcRegister.Negate = 0;
+      reg->SrcRegisterExtMod.Absolute = 1;
+      reg->SrcRegisterExtMod.Negate = 0;
+      break;
+
+   case TGSI_UTIL_SIGN_SET:
+      reg->SrcRegister.Negate = 0;
+      reg->SrcRegisterExtMod.Absolute = 1;
+      reg->SrcRegisterExtMod.Negate = 1;
+      break;
+
+   case TGSI_UTIL_SIGN_TOGGLE:
+      reg->SrcRegister.Negate = 1;
+      reg->SrcRegisterExtMod.Absolute = 0;
+      reg->SrcRegisterExtMod.Negate = 0;
+      break;
+
+   case TGSI_UTIL_SIGN_KEEP:
+      reg->SrcRegister.Negate = 0;
+      reg->SrcRegisterExtMod.Absolute = 0;
+      reg->SrcRegisterExtMod.Negate = 0;
+      break;
+
+   default:
+      assert( 0 );
+   }
+}
+
diff --git a/src/mesa/pipe/tgsi/core/tgsi_util.h b/src/mesa/pipe/tgsi/core/tgsi_util.h
new file mode 100644
index 00000000000..70c48690c5b
--- /dev/null
+++ b/src/mesa/pipe/tgsi/core/tgsi_util.h
@@ -0,0 +1,70 @@
+#if !defined TGSI_UTIL_H
+#define TGSI_UTIL_H
+
+#if defined __cplusplus
+extern "C" {
+#endif // defined __cplusplus
+
+void *
+tgsi_align_128bit(
+   void *unaligned );
+
+GLuint
+tgsi_util_get_src_register_swizzle(
+   const struct tgsi_src_register *reg,
+   GLuint component );
+
+GLuint
+tgsi_util_get_src_register_extswizzle(
+   const struct tgsi_src_register_ext_swz *reg,
+   GLuint component);
+
+GLuint
+tgsi_util_get_full_src_register_extswizzle(
+   const struct tgsi_full_src_register *reg,
+   GLuint component );
+
+void
+tgsi_util_set_src_register_swizzle(
+   struct tgsi_src_register *reg,
+   GLuint swizzle,
+   GLuint component );
+
+void
+tgsi_util_set_src_register_extswizzle(
+   struct tgsi_src_register_ext_swz *reg,
+   GLuint swizzle,
+   GLuint component );
+
+GLuint
+tgsi_util_get_src_register_extnegate(
+   const struct tgsi_src_register_ext_swz *reg,
+   GLuint component );
+
+void
+tgsi_util_set_src_register_extnegate(
+   struct tgsi_src_register_ext_swz *reg,
+   GLuint negate,
+   GLuint component );
+
+#define TGSI_UTIL_SIGN_CLEAR    0   /* Force positive */
+#define TGSI_UTIL_SIGN_SET      1   /* Force negative */
+#define TGSI_UTIL_SIGN_TOGGLE   2   /* Negate */
+#define TGSI_UTIL_SIGN_KEEP     3   /* No change */
+
+GLuint
+tgsi_util_get_full_src_register_sign_mode(
+   const struct tgsi_full_src_register *reg,
+   GLuint component );
+
+void
+tgsi_util_set_full_src_register_sign_mode(
+   struct tgsi_full_src_register *reg,
+   GLuint sign_mode );
+
+#if defined __cplusplus
+} // extern "C"
+#endif // defined __cplusplus
+
+#endif // !defined TGSI_UTIL_H
+
diff --git a/src/mesa/pipe/tgsi/mesa/Makefile b/src/mesa/pipe/tgsi/mesa/Makefile
new file mode 100644
index 00000000000..eb8b14e0e89
--- /dev/null
+++ b/src/mesa/pipe/tgsi/mesa/Makefile
@@ -0,0 +1,3 @@
+default:
+	cd ../../.. ; make
+
diff --git a/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.c b/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.c
new file mode 100644
index 00000000000..e37778fe41e
--- /dev/null
+++ b/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.c
@@ -0,0 +1,499 @@
+#include "tgsi_platform.h"
+#include "tgsi_mesa.h"
+
+/*
+ * Map mesa register file to SBIR register file.
+ */
+static GLuint
+map_register_file( enum register_file file )
+{
+   switch (file) {
+   case PROGRAM_UNDEFINED:
+      return TGSI_FILE_NULL;
+   case PROGRAM_TEMPORARY:
+      return TGSI_FILE_TEMPORARY;
+   //case PROGRAM_LOCAL_PARAM:
+   //case PROGRAM_ENV_PARAM:
+   case PROGRAM_STATE_VAR:
+   //case PROGRAM_NAMED_PARAM:
+   case PROGRAM_CONSTANT:
+      return TGSI_FILE_CONSTANT;
+   case PROGRAM_INPUT:
+      return TGSI_FILE_INPUT;
+   case PROGRAM_OUTPUT:
+      return TGSI_FILE_OUTPUT;
+   case PROGRAM_ADDRESS:
+      return TGSI_FILE_ADDRESS;
+   default:
+      assert (0);
+      return TGSI_FILE_NULL;
+   }
+}
+
+/*
+ * Map mesa register file index to SBIR index.
+ * Take special care when processing input and output indices.
+ */
+static GLuint
+map_register_file_index( GLuint processor,
+			 GLuint file,
+			 GLuint index,
+			 GLuint usage_bitmask )
+{
+   GLuint mapped_index;
+   GLuint i;
+
+   switch (file)
+   {
+   case TGSI_FILE_INPUT:
+      assert (index < 32);
+      assert (usage_bitmask & (1 << index));
+      mapped_index = 0;
+      for (i = 0; i < index; i++) {
+	 if (usage_bitmask & (1 << i))
+	    mapped_index++;
+      }
+      break;
+
+   case TGSI_FILE_OUTPUT:
+      assert (usage_bitmask == 0);
+      if (processor == TGSI_PROCESSOR_FRAGMENT) {
+	 if (index == FRAG_RESULT_DEPR) {
+	    mapped_index = 0;
+	 } else {
+	    assert (index == FRAG_RESULT_COLR);
+	    mapped_index = index + 1;
+	 }
+      } else {
+	 mapped_index = index;
+      }
+      break;
+
+   default:
+      mapped_index = index;
+   }
+
+   return mapped_index;
+}
+
+/*
+ * Map mesa texture target to SBIR texture target.
+ */
+static GLuint
+map_texture_target( GLuint textarget )
+{
+   switch (textarget) {
+   case TEXTURE_1D_INDEX:
+      return TGSI_TEXTURE_1D;
+   case TEXTURE_2D_INDEX:
+      return TGSI_TEXTURE_2D;
+   case TEXTURE_3D_INDEX:
+      return TGSI_TEXTURE_3D;
+   case TEXTURE_CUBE_INDEX:
+      return TGSI_TEXTURE_CUBE;
+   case TEXTURE_RECT_INDEX:
+      return TGSI_TEXTURE_RECT;
+   default:
+      assert (0);
+   }
+   return TGSI_TEXTURE_1D;
+}
+
+static GLuint
+convert_sat( GLuint sat )
+{
+   switch (sat) {
+   case SATURATE_OFF:
+      return TGSI_SAT_NONE;
+   case SATURATE_ZERO_ONE:
+      return TGSI_SAT_ZERO_ONE;
+   case SATURATE_PLUS_MINUS_ONE:
+      return TGSI_SAT_MINUS_PLUS_ONE;
+   default:
+      assert (0);
+      return TGSI_SAT_NONE;
+   }
+}
+
+static GLuint
+convert_writemask( GLuint writemask )
+{
+   assert (WRITEMASK_X == TGSI_WRITEMASK_X);
+   assert (WRITEMASK_Y == TGSI_WRITEMASK_Y);
+   assert (WRITEMASK_Z == TGSI_WRITEMASK_Z);
+   assert (WRITEMASK_W == TGSI_WRITEMASK_W);
+   assert ((writemask & ~TGSI_WRITEMASK_XYZW) == 0);
+
+   return writemask;
+}
+
+static GLboolean
+compile_instruction( struct prog_instruction *inst,
+		     struct tgsi_full_instruction *fullinst,
+		     GLuint inputs_read,
+		     GLuint processor )
+{
+   GLuint i;
+   struct tgsi_full_dst_register *fulldst;
+   struct tgsi_full_src_register *fullsrc;
+
+   *fullinst = tgsi_default_full_instruction ();
+
+   fullinst->Instruction.Saturate = convert_sat (inst->SaturateMode);
+   fullinst->Instruction.NumDstRegs = 1;
+   fullinst->Instruction.NumSrcRegs = _mesa_num_inst_src_regs (inst->Opcode);
+
+   fulldst = &fullinst->FullDstRegisters[0];
+   fulldst->DstRegister.File =
+      map_register_file (inst->DstReg.File);
+   fulldst->DstRegister.Index =
+      map_register_file_index (processor,
+			       fulldst->DstRegister.File,
+			       inst->DstReg.Index,
+			       0);
+   fulldst->DstRegister.WriteMask =
+      convert_writemask (inst->DstReg.WriteMask);
+
+   for (i = 0; i < fullinst->Instruction.NumSrcRegs; i++) {
+      GLuint j;
+
+      fullsrc = &fullinst->FullSrcRegisters[i];
+      fullsrc->SrcRegister.File =
+	 map_register_file (inst->SrcReg[i].File);
+      fullsrc->SrcRegister.Index =
+	 map_register_file_index (processor,
+				  fullsrc->SrcRegister.File,
+				  inst->SrcReg[i].Index,
+				  inputs_read);
+
+      for (j = 0; j < 4; j++) {
+	 GLuint swz;
+
+	 swz = GET_SWZ(inst->SrcReg[i].Swizzle, j);
+	 if (swz > SWIZZLE_W) {
+	    tgsi_util_set_src_register_extswizzle (
+	       &fullsrc->SrcRegisterExtSwz,
+	       swz,
+	       j);
+	 } else {
+	    tgsi_util_set_src_register_swizzle (
+	       &fullsrc->SrcRegister,
+	       swz,
+	       j);
+	 }
+      }
+
+      if (inst->SrcReg[i].NegateBase == NEGATE_XYZW) {
+	 fullsrc->SrcRegister.Negate = 1;
+      } else if (inst->SrcReg[i].NegateBase != NEGATE_NONE) {
+	 if (inst->SrcReg[i].NegateBase & NEGATE_X)
+	    fullsrc->SrcRegisterExtSwz.NegateX = 1;
+	 if (inst->SrcReg[i].NegateBase & NEGATE_Y)
+	    fullsrc->SrcRegisterExtSwz.NegateY = 1;
+	 if (inst->SrcReg[i].NegateBase & NEGATE_Z)
+	    fullsrc->SrcRegisterExtSwz.NegateZ = 1;
+	 if (inst->SrcReg[i].NegateBase & NEGATE_W)
+	    fullsrc->SrcRegisterExtSwz.NegateW = 1;
+      }
+
+      if (inst->SrcReg[i].Abs)
+	 fullsrc->SrcRegisterExtMod.Absolute = 1;
+
+      if (inst->SrcReg[i].NegateAbs)
+	 fullsrc->SrcRegisterExtMod.Negate = 1;
+
+      if (inst->SrcReg[i].RelAddr) {
+	 fullsrc->SrcRegister.Indirect = 1;
+
+	 fullsrc->SrcRegisterInd.File = TGSI_FILE_ADDRESS;
+	 fullsrc->SrcRegisterInd.Index = 0;
+      }
+   }
+
+   switch (inst->Opcode) {
+   case OPCODE_ARL:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_ARL;
+      break;
+   case OPCODE_ABS:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_ABS;
+      break;
+   case OPCODE_ADD:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_ADD;
+      break;
+   case OPCODE_CMP:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_CMP;
+      break;
+   case OPCODE_COS:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_COS;
+      break;
+   case OPCODE_DP3:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_DP3;
+      break;
+   case OPCODE_DP4:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_DP4;
+      break;
+   case OPCODE_DPH:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_DPH;
+      break;
+   case OPCODE_DST:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_DST;
+      break;
+   case OPCODE_EX2:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_EX2;
+      break;
+   case OPCODE_FLR:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_FLR;
+      break;
+   case OPCODE_FRC:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_FRC;
+      break;
+   case OPCODE_KIL:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_KIL;
+      break;
+   case OPCODE_LG2:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_LG2;
+      break;
+   case OPCODE_LIT:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_LIT;
+      break;
+   case OPCODE_LRP:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_LRP;
+      break;
+   case OPCODE_MAD:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_MAD;
+      break;
+   case OPCODE_MAX:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_MAX;
+      break;
+   case OPCODE_MIN:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_MIN;
+      break;
+   case OPCODE_MOV:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_MOV;
+      break;
+   case OPCODE_MUL:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_MUL;
+      break;
+   case OPCODE_POW:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_POW;
+      break;
+   case OPCODE_RCP:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_RCP;
+      break;
+   case OPCODE_RSQ:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_RSQ;
+      tgsi_util_set_full_src_register_sign_mode (&fullinst->FullSrcRegisters[0],
+						 TGSI_UTIL_SIGN_CLEAR);
+      break;
+   case OPCODE_SCS:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_SCS;
+      fulldst->DstRegister.WriteMask &= TGSI_WRITEMASK_XY;
+      break;
+   case OPCODE_SGE:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_SGE;
+      break;
+   case OPCODE_SIN:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_SIN;
+      break;
+   case OPCODE_SLT:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_SLT;
+      break;
+   case OPCODE_SUB:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_SUB;
+      break;
+   case OPCODE_SWZ:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_SWZ;
+      break;
+   case OPCODE_TEX:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_TEX;
+      fullinst->Instruction.NumSrcRegs = 2;
+      fullinst->InstructionExtTexture.Texture = map_texture_target (inst->TexSrcTarget);
+      fullinst->FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_SAMPLER;
+      fullinst->FullSrcRegisters[1].SrcRegister.Index = inst->TexSrcUnit;
+      break;
+   case OPCODE_TXB:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_TXB;
+      fullinst->Instruction.NumSrcRegs = 2;
+      fullinst->InstructionExtTexture.Texture = map_texture_target (inst->TexSrcTarget);
+      fullinst->FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_SAMPLER;
+      fullinst->FullSrcRegisters[1].SrcRegister.Index = inst->TexSrcUnit;
+      break;
+   case OPCODE_TXP:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_TEX;
+      fullinst->Instruction.NumSrcRegs = 2;
+      fullinst->InstructionExtTexture.Texture = map_texture_target (inst->TexSrcTarget);
+      fullinst->FullSrcRegisters[0].SrcRegisterExtSwz.ExtDivide = TGSI_EXTSWIZZLE_W;
+      fullinst->FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_SAMPLER;
+      fullinst->FullSrcRegisters[1].SrcRegister.Index = inst->TexSrcUnit;
+      break;
+   case OPCODE_XPD:
+      fullinst->Instruction.Opcode = TGSI_OPCODE_XPD;
+      fulldst->DstRegister.WriteMask &= TGSI_WRITEMASK_XYZ;
+      break;
+   case OPCODE_END:
+      return GL_TRUE;
+   default:
+      assert (0);
+   }
+
+   return GL_FALSE;
+}
+
+GLboolean
+tgsi_compile_fp_program( const struct gl_fragment_program *program,
+			 struct tgsi_token *tokens,
+			 GLuint max_token_count,
+			 GLuint *token_count )
+{
+   GLuint i, ti;
+   struct tgsi_header *header;
+   struct tgsi_full_declaration fulldecl;
+   struct tgsi_full_instruction fullinst;
+   struct tgsi_full_dst_register *fulldst;
+   struct tgsi_full_src_register *fullsrc;
+   GLuint inputs_read;
+
+   *(struct tgsi_version *) &tokens[0] = tgsi_build_version ();
+
+   header = (struct tgsi_header *) &tokens[1];
+   *header = tgsi_build_header ();
+
+   ti = 2;
+
+   /*
+    * Input 0 is always read, at least implicitly by the instruction generated
+    * above, so mark it as used.
+    */
+   inputs_read = program->Base.InputsRead | 1;
+
+   /*
+    * Declare input attributes.
+    */
+   fulldecl = tgsi_default_full_declaration();
+
+   fulldecl.Declaration.File = TGSI_FILE_INPUT;
+   fulldecl.Declaration.Declare = TGSI_DECLARE_RANGE;
+   fulldecl.Declaration.Interpolate = 1;
+
+   /*
+    * Do not interpolate fragment position.
+    */
+   fulldecl.u.DeclarationRange.First = 0;
+   fulldecl.u.DeclarationRange.Last = 0;
+
+   fulldecl.Interpolation.Interpolate = TGSI_INTERPOLATE_CONSTANT;
+
+   ti += tgsi_build_full_declaration(
+      &fulldecl,
+      &tokens[ti],
+      header,
+      max_token_count - ti );
+
+   /*
+    * Interpolate generic attributes.
+    */
+   fulldecl.u.DeclarationRange.First = 1;
+   fulldecl.u.DeclarationRange.Last = 1;
+   for( i = 1; i < 32; i++ ) {
+      if( inputs_read & (1 << i) ) {
+         fulldecl.u.DeclarationRange.Last++;
+      }
+   }
+
+   fulldecl.Interpolation.Interpolate = TGSI_INTERPOLATE_LINEAR;
+
+   ti += tgsi_build_full_declaration(
+      &fulldecl,
+      &tokens[ti],
+      header,
+      max_token_count - ti );
+
+   /*
+    * Copy input fragment xyz to output xyz.
+    * If the shader writes depth, do not copy the z component.
+    */
+
+   fullinst = tgsi_default_full_instruction ();
+
+   fullinst.Instruction.Opcode = TGSI_OPCODE_MOV;
+   fullinst.Instruction.NumDstRegs = 1;
+   fullinst.Instruction.NumSrcRegs = 1;
+
+   fulldst = &fullinst.FullDstRegisters[0];
+   fulldst->DstRegister.File = TGSI_FILE_OUTPUT;
+   fulldst->DstRegister.Index = 0;
+   if (program->Base.OutputsWritten & (1 << FRAG_RESULT_DEPR)) {
+      fulldst->DstRegister.WriteMask = TGSI_WRITEMASK_XY;
+   } else {
+      fulldst->DstRegister.WriteMask = TGSI_WRITEMASK_XYZ;
+   }
+
+   fullsrc = &fullinst.FullSrcRegisters[0];
+   fullsrc->SrcRegister.File = TGSI_FILE_INPUT;
+   fullsrc->SrcRegister.Index = 0;
+
+   ti += tgsi_build_full_instruction (&fullinst,
+				      &tokens[ti],
+				      header,
+				      max_token_count - ti);
+
+   for( i = 0; i < program->Base.NumInstructions; i++ ) {
+      if (compile_instruction (&program->Base.Instructions[i],
+			       &fullinst,
+			       inputs_read,
+			       TGSI_PROCESSOR_FRAGMENT)) {
+	 assert (i == program->Base.NumInstructions - 1);
+	 tgsi_dump (tokens, TGSI_DUMP_NO_IGNORED | TGSI_DUMP_NO_DEFAULT);
+	 break;
+      }
+
+      ti += tgsi_build_full_instruction (&fullinst,
+					 &tokens[ti],
+					 header,
+					 max_token_count - ti);
+   }
+
+   return GL_TRUE;
+}
+
+GLboolean
+tgsi_compile_vp_program( const struct gl_vertex_program *program,
+			 struct tgsi_token *tokens,
+			 GLuint max_token_count,
+			 GLuint *token_count )
+{
+   GLuint ii, ti;
+   struct tgsi_header *header;
+   struct tgsi_processor *processor;
+   struct tgsi_full_instruction fullinst;
+   GLuint inputs_read = ~0;
+
+   *(struct tgsi_version *) &tokens[0] = tgsi_build_version ();
+
+   header = (struct tgsi_header *) &tokens[1];
+   *header = tgsi_build_header ();
+
+   processor = (struct tgsi_processor *) &tokens[2];
+   *processor = tgsi_build_processor (TGSI_PROCESSOR_VERTEX, header);
+
+   ti = 3;
+
+   for (ii = 0; ii < program->Base.NumInstructions; ii++) {
+      if (compile_instruction (&program->Base.Instructions[ii],
+			       &fullinst,
+			       inputs_read,
+			       TGSI_PROCESSOR_VERTEX)) {
+	 assert (ii == program->Base.NumInstructions - 1);
+	 tgsi_dump (tokens, TGSI_DUMP_NO_IGNORED | TGSI_DUMP_NO_DEFAULT);
+	 break;
+      }
+
+      ti += tgsi_build_full_instruction (&fullinst,
+					 &tokens[ti],
+					 header,
+					 max_token_count - ti);
+   }
+
+   return GL_TRUE;
+}
+
diff --git a/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.h b/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.h
new file mode 100644
index 00000000000..4c1141e5796
--- /dev/null
+++ b/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.h
@@ -0,0 +1,25 @@
+#if !defined MESA_TO_TGSI_H
+#define MESA_TO_TGSI_H
+
+#if defined __cplusplus
+extern "C" {
+#endif // defined __cplusplus
+
+GLboolean
+tgsi_mesa_compile_fp_program(
+   const struct gl_fragment_program *program,
+   struct tgsi_token *tokens,
+   GLuint maxTokens );
+
+GLboolean
+tgsi_mesa_compile_vp_program(
+   const struct gl_vertex_program *program,
+   struct tgsi_token *tokens,
+   GLuint maxTokens );
+
+#if defined __cplusplus
+} // extern "C"
+#endif // defined __cplusplus
+
+#endif // !defined MESA_TO_TGSI_H
+
diff --git a/src/mesa/pipe/tgsi/mesa/tgsi_mesa.h b/src/mesa/pipe/tgsi/mesa/tgsi_mesa.h
new file mode 100644
index 00000000000..0053748fdef
--- /dev/null
+++ b/src/mesa/pipe/tgsi/mesa/tgsi_mesa.h
@@ -0,0 +1,8 @@
+#if !defined TGSI_MESA_H
+#define TGSI_MESA_H
+
+#include "../core/tgsi_core.h"
+#include "mesa_to_tgsi.h"
+
+#endif // !defined TGSI_MESA_H
+
diff --git a/src/mesa/pipe/tgsi/tgsi_platform.h b/src/mesa/pipe/tgsi/tgsi_platform.h
new file mode 100644
index 00000000000..553f0b26a89
--- /dev/null
+++ b/src/mesa/pipe/tgsi/tgsi_platform.h
@@ -0,0 +1,18 @@
+#if !defined TGSI_PLATFORM_H
+#define TGSI_PLATFORM_H
+
+#if defined __cplusplus
+extern "C" {
+#endif // defined __cplusplus
+
+#include "imports.h"
+#include "mtypes.h"
+#include "prog_instruction.h"
+#include "program.h"
+
+#if defined __cplusplus
+} // extern "C"
+#endif // defined __cplusplus
+
+#endif // !defined TGSI_PLATFORM_H
+
diff --git a/src/mesa/sources b/src/mesa/sources
index 643e841b153..b1ff3798a63 100644
--- a/src/mesa/sources
+++ b/src/mesa/sources
@@ -179,6 +179,16 @@ SOFTPIPE_SOURCES = \
 	pipe/softpipe/sp_state_setup.c \
 	pipe/softpipe/sp_state_surface.c
 
+TGSICORE_SOURCES = \
+	pipe/tgsi/core/tgsi_build.c \
+	pipe/tgsi/core/tgsi_dump.c \
+	pipe/tgsi/core/tgsi_exec.c \
+	pipe/tgsi/core/tgsi_parse.c \
+	pipe/tgsi/core/tgsi_util.c
+
+TGSIMESA_SOURCES = \
+	pipe/tgsi/mesa/mesa_to_tgsi.c
+
 STATETRACKER_SOURCES = \
 	state_tracker/st_atom.c \
 	state_tracker/st_atom_alphatest.c \
@@ -344,8 +354,10 @@ SOLO_SOURCES = \
 	$(MATH_SOURCES)		\
 	$(VBO_SOURCES)		\
 	$(VF_SOURCES)		\
-	$(SOFTPIPE_SOURCES)		\
-	$(STATETRACKER_SOURCES)		\
+	$(SOFTPIPE_SOURCES)	\
+	$(TGSICORE_SOURCES)	\
+	$(TGSIMESA_SOURCES)	\
+	$(STATETRACKER_SOURCES)	\
 	$(TNL_SOURCES)		\
 	$(SHADER_SOURCES)	\
 	$(SWRAST_SOURCES)	\
@@ -391,4 +403,5 @@ INCLUDE_DIRS = \
 	-I$(TOP)/src/mesa/shader/grammar \
 	-I$(TOP)/src/mesa/shader/slang \
 	-I$(TOP)/src/mesa/swrast \
-	-I$(TOP)/src/mesa/swrast_setup
+	-I$(TOP)/src/mesa/swrast_setup \
+	-I$(TOP)/src/mesa/pipe/tgsi
diff --git a/src/mesa/state_tracker/st_cb_program.c b/src/mesa/state_tracker/st_cb_program.c
index 67a589ca8d3..327b627722f 100644
--- a/src/mesa/state_tracker/st_cb_program.c
+++ b/src/mesa/state_tracker/st_cb_program.c
@@ -41,6 +41,7 @@
 #include "program.h"
 #include "programopt.h"
 #include "tnl/tnl.h"
+#include "pipe/tgsi/mesa/tgsi_mesa.h"
 
 
 static void st_bind_program( GLcontext *ctx,
@@ -89,7 +90,7 @@ static struct gl_program *st_new_program( GLcontext *ctx,
 
 static void st_delete_program( GLcontext *ctx,
 			       struct gl_program *prog )
-{   
+{
    _mesa_delete_program( ctx, prog );
 }
 
@@ -115,7 +116,7 @@ static void st_program_string_notify( GLcontext *ctx,
 
 	 st->dirty.st |= ST_NEW_FRAGMENT_PROGRAM;
 
-	 p->id = st->program_id++;      
+	 p->id = st->program_id++;
 #if 0
 	 p->param_state = p->Base.Base.Parameters->StateFlags; 
 	 p->translated = 0;
@@ -128,6 +129,14 @@ static void st_program_string_notify( GLcontext *ctx,
 	    _mesa_append_fog_code(ctx, &p->Base);
 	    p->Base.FogOption = GL_NONE;
 	 }
+
+         /* XXX: Not hooked-up yet. */
+	 {
+	    struct tgsi_token tokens[1024];
+
+	    tgsi_mesa_compile_fp_program( prog, tokens, 1024 );
+	    tgsi_dump( tokens, TGSI_DUMP_VERBOSE );
+	 }
       }
    }
    else if (target == GL_VERTEX_PROGRAM_ARB) {
@@ -148,7 +157,6 @@ void st_init_cb_program( struct st_context *st )
     */
    st->ctx->FragmentProgram._MaintainTexEnvProgram = GL_TRUE;
    st->ctx->FragmentProgram._UseTexEnvProgram = GL_TRUE;
-      
 
    assert(functions->ProgramStringNotify == _tnl_program_string); 
    functions->BindProgram = st_bind_program;
diff --git a/src/mesa/state_tracker/st_texobj.c b/src/mesa/state_tracker/st_texobj.c
index 455e89b6bd1..eb5bdb2d08f 100644
--- a/src/mesa/state_tracker/st_texobj.c
+++ b/src/mesa/state_tracker/st_texobj.c
@@ -52,7 +52,8 @@ create_texture_object(struct gl_texture_object *texObj)
    if (!pto)
       return NULL;
 
-   assert(texObj->Complete);
+   /* XXX: Member not defined. Comment-out to get it compile. */
+   /*assert(texObj->Complete);*/
 
    switch (texObj->Target) {
    case GL_TEXTURE_1D:
-- 
2.30.2