rewrite the way we handle ir in llvm code

introduce intermediate step gallivm_ir before compiling it
down to the final llvm ir.

11 files changed, 1079 insertions, 970 deletions

diff --git a/src/mesa/pipe/cell/ppu/cell_state_fs.c b/src/mesa/pipe/cell/ppu/cell_state_fs.c
index 96a52273b0..3f46a87d18 100644
--- a/src/mesa/pipe/cell/ppu/cell_state_fs.c
+++ b/src/mesa/pipe/cell/ppu/cell_state_fs.c
@@ -67,7 +67,7 @@ cell_create_fs_state(struct pipe_context *pipe,
 #endif
 
 #ifdef MESA_LLVM
-   state->llvm_prog = gallivm_from_tgsi(state->shader.tokens, GALLIVM_FS);
+   state->llvm_prog = 0;
    if (!gallivm_global_cpu_engine()) {
       gallivm_cpu_engine_create(state->llvm_prog);
    }
diff --git a/src/mesa/pipe/draw/draw_vertex_shader.c b/src/mesa/pipe/draw/draw_vertex_shader.c
index 5ca93aa615..9567283ff5 100644
--- a/src/mesa/pipe/draw/draw_vertex_shader.c
+++ b/src/mesa/pipe/draw/draw_vertex_shader.c
@@ -113,7 +113,16 @@ run_vertex_program(struct draw_context *draw,
    draw->vertex_fetch.fetch_func( draw, machine, elts, count );
 
    /* run shader */
-#if defined(__i386__) || defined(__386__)
+#ifdef MESA_LLVM
+   if (1) {
+   struct gallivm_prog  *prog  = draw->vertex_shader->llvm_prog;
+   gallivm_prog_exec(prog,
+                     machine->Inputs,
+                     machine->Outputs,
+                     machine->Consts,
+                     12, 12, 12);
+   } else
+#elif defined(__i386__) || defined(__386__)
    if (draw->use_sse) {
       /* SSE */
       /* cast away const */
@@ -212,13 +221,7 @@ draw_vertex_shader_queue_flush(struct draw_context *draw)
     */
    draw_update_vertex_fetch( draw );
 
-//   debug_printf( " q(%d) ", draw->vs.queue_nr );
-#ifdef MESA_LLVM
-   if (draw->vertex_shader->llvm_prog) {
-      draw_vertex_shader_queue_flush_llvm(draw);
-      return;
-   }
-#endif
+//   fprintf(stderr, " q(%d) ", draw->vs.queue_nr );
 
    /* run vertex shader on vertex cache entries, four per invokation */
    for (i = 0; i < draw->vs.queue_nr; i += 4) {
@@ -260,7 +263,13 @@ draw_create_vertex_shader(struct draw_context *draw,
    vs->state = shader;
 
 #ifdef MESA_LLVM
-   vs->llvm_prog = gallivm_from_tgsi(shader->tokens, GALLIVM_VS);
+   struct gallivm_ir *ir = gallivm_ir_new(GALLIVM_VS);
+   gallivm_ir_set_layout(ir, GALLIVM_SOA);
+   gallivm_ir_set_components(ir, 4);
+   gallivm_ir_fill_from_tgsi(ir, shader->tokens);
+   vs->llvm_prog = gallivm_ir_compile(ir);
+   gallivm_ir_delete(ir);
+
    draw->engine = gallivm_global_cpu_engine();
    if (!draw->engine) {
       draw->engine = gallivm_cpu_engine_create(vs->llvm_prog);
diff --git a/src/mesa/pipe/llvm/Makefile b/src/mesa/pipe/llvm/Makefile
index 1e3ae988df..10ff7aacae 100644
--- a/src/mesa/pipe/llvm/Makefile
+++ b/src/mesa/pipe/llvm/Makefile
@@ -9,6 +9,7 @@ GALLIVM_SOURCES = \
         gallivm.cpp  \
         instructions.cpp  \
         loweringpass.cpp \
+        tgsitollvm.cpp \
         storage.cpp
 
 INC_SOURCES = gallivm_builtins.cpp llvm_base_shader.cpp
diff --git a/src/mesa/pipe/llvm/gallivm.cpp b/src/mesa/pipe/llvm/gallivm.cpp
index 46e11c185a..cf9b0f6406 100644
--- a/src/mesa/pipe/llvm/gallivm.cpp
+++ b/src/mesa/pipe/llvm/gallivm.cpp
@@ -32,17 +32,17 @@
 #ifdef MESA_LLVM
 
 #include "gallivm.h"
+#include "gallivm_p.h"
 
 #include "instructions.h"
 #include "loweringpass.h"
 #include "storage.h"
+#include "tgsitollvm.h"
 
 #include "pipe/p_context.h"
 #include "pipe/p_shader_tokens.h"
-#include "pipe/tgsi/util/tgsi_parse.h"
+
 #include "pipe/tgsi/exec/tgsi_exec.h"
-#include "pipe/tgsi/util/tgsi_util.h"
-#include "pipe/tgsi/util/tgsi_build.h"
 #include "pipe/tgsi/util/tgsi_dump.h"
 
 #include <llvm/Module.h>
@@ -64,38 +64,23 @@
 #include <llvm/Analysis/LoopPass.h>
 #include <llvm/Target/TargetData.h>
 #include <llvm/Bitcode/ReaderWriter.h>
+#include <llvm/Transforms/Utils/Cloning.h>
 
 #include <sstream>
 #include <fstream>
 #include <iostream>
 
-struct gallivm_interpolate {
-   int attrib;
-   int chan;
-   int type;
-};
-
-struct gallivm_prog {
-   llvm::Module *module;
-   void *function;
-   int   num_consts;
-   int   id;
-   enum gallivm_shader_type type;
-
-   struct gallivm_interpolate interpolators[32*4]; //FIXME: this might not be enough for some shaders
-   int   num_interp;
-};
-
 struct gallivm_cpu_engine {
    llvm::ExecutionEngine *engine;
 };
 
-using namespace llvm;
-#include "llvm_base_shader.cpp"
-
 static int GLOBAL_ID = 0;
 
-static inline void AddStandardCompilePasses(PassManager &PM) {
+using namespace llvm;
+
+static inline
+void AddStandardCompilePasses(PassManager &PM)
+{
    PM.add(new LoweringPass());
    PM.add(createVerifierPass());                  // Verify that input is correct
 
@@ -152,691 +137,16 @@ static inline void AddStandardCompilePasses(PassManager &PM) {
    PM.add(createConstantMergePass());        // Merge dup global constants
 }
 
-static inline void
-add_interpolator(struct gallivm_prog *prog,
-                 struct gallivm_interpolate *interp)
-{
-   prog->interpolators[prog->num_interp] = *interp;
-   ++prog->num_interp;
-}
-
-static void
-translate_declaration(struct gallivm_prog *prog,
-                      llvm::Module *module,
-                      Storage *storage,
-                      struct tgsi_full_declaration *decl,
-                      struct tgsi_full_declaration *fd)
-{
-   if (decl->Declaration.File == TGSI_FILE_INPUT) {
-      unsigned first, last, mask;
-      uint interp_method;
-
-      assert(decl->Declaration.Declare == TGSI_DECLARE_RANGE);
-
-      first = decl->u.DeclarationRange.First;
-      last = decl->u.DeclarationRange.Last;
-      mask = decl->Declaration.UsageMask;
-
-      /* Do not touch WPOS.xy */
-      if (first == 0) {
-         mask &= ~TGSI_WRITEMASK_XY;
-         if (mask == TGSI_WRITEMASK_NONE) {
-            first++;
-            if (first > last) {
-               return;
-            }
-         }
-      }
-
-      interp_method = decl->Interpolation.Interpolate;
-
-      if (mask == TGSI_WRITEMASK_XYZW) {
-         unsigned i, j;
-
-         for (i = first; i <= last; i++) {
-            for (j = 0; j < NUM_CHANNELS; j++) {
-               //interp( mach, i, j );
-               struct gallivm_interpolate interp;
-               interp.type = interp_method;
-               interp.attrib = i;
-               interp.chan = j;
-               add_interpolator(prog, &interp);
-            }
-         }
-      } else {
-         unsigned i, j;
-         for( j = 0; j < NUM_CHANNELS; j++ ) {
-            if( mask & (1 << j) ) {
-               for( i = first; i <= last; i++ ) {
-                  struct gallivm_interpolate interp;
-                  interp.type = interp_method;
-                  interp.attrib = i;
-                  interp.chan = j;
-                  add_interpolator(prog, &interp);
-               }
-            }
-         }
-      }
-   }
-}
-
-
-static void
-translate_immediate(Storage *storage,
-                    struct tgsi_full_immediate *imm)
-{
-   float vec[4];
-   int i;
-   for (i = 0; i < imm->Immediate.Size - 1; ++i) {
-      switch( imm->Immediate.DataType ) {
-      case TGSI_IMM_FLOAT32:
-         vec[i] = imm->u.ImmediateFloat32[i].Float;
-         break;
-      default:
-         assert( 0 );
-      }
-   }
-   storage->addImmediate(vec);
-}
-
-static inline llvm::Value *
-swizzleVector(llvm::Value *val, struct tgsi_full_src_register *src,
-              Storage *storage)
-{
-   int swizzle = 0;
-   int start = 1000;
-   const int NO_SWIZZLE = TGSI_SWIZZLE_X * 1000 + TGSI_SWIZZLE_Y * 100 +
-                          TGSI_SWIZZLE_Z * 10 + TGSI_SWIZZLE_W;
-   for (int k = 0; k < 4; ++k) {
-      swizzle += tgsi_util_get_full_src_register_extswizzle(src, k) * start;
-      start /= 10;
-   }
-   if (swizzle != NO_SWIZZLE) {
-      /*fprintf(stderr, "XXXXXXXX swizzle = %d\n", swizzle);*/
-      val = storage->shuffleVector(val, swizzle);
-   }
-   return val;
-}
-
-static void
-translate_instruction(llvm::Module *module,
-                      Storage *storage,
-                      Instructions *instr,
-                      struct tgsi_full_instruction *inst,
-                      struct tgsi_full_instruction *fi,
-                      unsigned instno)
-{
-   llvm::Value *inputs[4];
-   inputs[0] = 0;
-   inputs[1] = 0;
-   inputs[2] = 0;
-   inputs[3] = 0;
-
-   for (int i = 0; i < inst->Instruction.NumSrcRegs; ++i) {
-      struct tgsi_full_src_register *src = &inst->FullSrcRegisters[i];
-      llvm::Value *val = 0;
-      llvm::Value *indIdx = 0;
-
-      if (src->SrcRegister.Indirect) {
-         indIdx = storage->addrElement(src->SrcRegisterInd.Index);
-         indIdx = storage->extractIndex(indIdx);
-      }
-      if (src->SrcRegister.File == TGSI_FILE_CONSTANT) {
-         val = storage->constElement(src->SrcRegister.Index, indIdx);
-      } else if (src->SrcRegister.File == TGSI_FILE_INPUT) {
-         val = storage->inputElement(src->SrcRegister.Index, indIdx);
-      } else if (src->SrcRegister.File == TGSI_FILE_TEMPORARY) {
-         val = storage->tempElement(src->SrcRegister.Index);
-      } else if (src->SrcRegister.File == TGSI_FILE_OUTPUT) {
-         val = storage->outputElement(src->SrcRegister.Index, indIdx);
-      } else if (src->SrcRegister.File == TGSI_FILE_IMMEDIATE) {
-         val = storage->immediateElement(src->SrcRegister.Index);
-      } else {
-         fprintf(stderr, "ERROR: not supported llvm source %d\n", src->SrcRegister.File);
-         return;
-      }
-
-      inputs[i] = swizzleVector(val, src, storage);
-   }
-
-   /*if (inputs[0])
-     instr->printVector(inputs[0]);
-     if (inputs[1])
-     instr->printVector(inputs[1]);*/
-   llvm::Value *out = 0;
-   switch (inst->Instruction.Opcode) {
-   case TGSI_OPCODE_ARL: {
-      out = instr->arl(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_MOV: {
-      out = inputs[0];
-   }
-      break;
-   case TGSI_OPCODE_LIT: {
-      out = instr->lit(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_RCP: {
-      out = instr->rcp(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_RSQ: {
-      out = instr->rsq(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_EXP:
-      break;
-   case TGSI_OPCODE_LOG:
-      break;
-   case TGSI_OPCODE_MUL: {
-      out = instr->mul(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_ADD: {
-      out = instr->add(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_DP3: {
-      out = instr->dp3(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_DP4: {
-      out = instr->dp4(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_DST: {
-      out = instr->dst(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_MIN: {
-      out = instr->min(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_MAX: {
-      out = instr->max(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_SLT: {
-      out = instr->slt(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_SGE: {
-      out = instr->sge(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_MAD: {
-      out = instr->madd(inputs[0], inputs[1], inputs[2]);
-   }
-      break;
-   case TGSI_OPCODE_SUB: {
-      out = instr->sub(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_LERP: {
-      out = instr->lerp(inputs[0], inputs[1], inputs[2]);
-   }
-      break;
-   case TGSI_OPCODE_CND:
-      break;
-   case TGSI_OPCODE_CND0:
-      break;
-   case TGSI_OPCODE_DOT2ADD:
-      break;
-   case TGSI_OPCODE_INDEX:
-      break;
-   case TGSI_OPCODE_NEGATE:
-      break;
-   case TGSI_OPCODE_FRAC: {
-      out = instr->frc(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_CLAMP:
-      break;
-   case TGSI_OPCODE_FLOOR: {
-      out = instr->floor(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_ROUND:
-      break;
-   case TGSI_OPCODE_EXPBASE2: {
-      out = instr->ex2(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_LOGBASE2: {
-      out = instr->lg2(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_POWER: {
-      out = instr->pow(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_CROSSPRODUCT: {
-      out = instr->cross(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_MULTIPLYMATRIX:
-      break;
-   case TGSI_OPCODE_ABS: {
-      out = instr->abs(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_RCC:
-      break;
-   case TGSI_OPCODE_DPH: {
-      out = instr->dph(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_COS: {
-      out = instr->cos(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_DDX:
-      break;
-   case TGSI_OPCODE_DDY:
-      break;
-   case TGSI_OPCODE_KILP: {
-      out = instr->kilp(inputs[0]);
-      storage->setKilElement(out);
-      return;
-   }
-      break;
-   case TGSI_OPCODE_PK2H:
-      break;
-   case TGSI_OPCODE_PK2US:
-      break;
-   case TGSI_OPCODE_PK4B:
-      break;
-   case TGSI_OPCODE_PK4UB:
-      break;
-   case TGSI_OPCODE_RFL:
-      break;
-   case TGSI_OPCODE_SEQ:
-      break;
-   case TGSI_OPCODE_SFL:
-      break;
-   case TGSI_OPCODE_SGT: {
-      out = instr->sgt(inputs[0], inputs[1]);
-   }
-      break;
-   case TGSI_OPCODE_SIN: {
-      out = instr->sin(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_SLE:
-      break;
-   case TGSI_OPCODE_SNE:
-      break;
-   case TGSI_OPCODE_STR:
-      break;
-   case TGSI_OPCODE_TEX:
-      break;
-   case TGSI_OPCODE_TXD:
-      break;
-   case TGSI_OPCODE_UP2H:
-      break;
-   case TGSI_OPCODE_UP2US:
-      break;
-   case TGSI_OPCODE_UP4B:
-      break;
-   case TGSI_OPCODE_UP4UB:
-      break;
-   case TGSI_OPCODE_X2D:
-      break;
-   case TGSI_OPCODE_ARA:
-      break;
-   case TGSI_OPCODE_ARR:
-      break;
-   case TGSI_OPCODE_BRA:
-      break;
-   case TGSI_OPCODE_CAL: {
-      instr->cal(inst->InstructionExtLabel.Label, storage->inputPtr());
-      return;
-   }
-      break;
-   case TGSI_OPCODE_RET: {
-      instr->end();
-      return;
-   }
-      break;
-   case TGSI_OPCODE_SSG:
-      break;
-   case TGSI_OPCODE_CMP: {
-      out = instr->cmp(inputs[0], inputs[1], inputs[2]);
-   }
-      break;
-   case TGSI_OPCODE_SCS: {
-      out = instr->scs(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_TXB:
-      break;
-   case TGSI_OPCODE_NRM:
-      break;
-   case TGSI_OPCODE_DIV:
-      break;
-   case TGSI_OPCODE_DP2:
-      break;
-   case TGSI_OPCODE_TXL:
-      break;
-   case TGSI_OPCODE_BRK: {
-      instr->brk();
-      return;
-   }
-      break;
-   case TGSI_OPCODE_IF: {
-      instr->ifop(inputs[0]);
-      storage->setCurrentBlock(instr->currentBlock());
-      return;  //just update the state
-   }
-      break;
-   case TGSI_OPCODE_LOOP:
-      break;
-   case TGSI_OPCODE_REP:
-      break;
-   case TGSI_OPCODE_ELSE: {
-      instr->elseop();
-      storage->setCurrentBlock(instr->currentBlock());
-      return; //only state update
-   }
-      break;
-   case TGSI_OPCODE_ENDIF: {
-      instr->endif();
-      storage->setCurrentBlock(instr->currentBlock());
-      return; //just update the state
-   }
-      break;
-   case TGSI_OPCODE_ENDLOOP:
-      break;
-   case TGSI_OPCODE_ENDREP:
-      break;
-   case TGSI_OPCODE_PUSHA:
-      break;
-   case TGSI_OPCODE_POPA:
-      break;
-   case TGSI_OPCODE_CEIL:
-      break;
-   case TGSI_OPCODE_I2F:
-      break;
-   case TGSI_OPCODE_NOT:
-      break;
-   case TGSI_OPCODE_TRUNC: {
-      out = instr->trunc(inputs[0]);
-   }
-      break;
-   case TGSI_OPCODE_SHL:
-      break;
-   case TGSI_OPCODE_SHR:
-      break;
-   case TGSI_OPCODE_AND:
-      break;
-   case TGSI_OPCODE_OR:
-      break;
-   case TGSI_OPCODE_MOD:
-      break;
-   case TGSI_OPCODE_XOR:
-      break;
-   case TGSI_OPCODE_SAD:
-      break;
-   case TGSI_OPCODE_TXF:
-      break;
-   case TGSI_OPCODE_TXQ:
-      break;
-   case TGSI_OPCODE_CONT:
-      break;
-   case TGSI_OPCODE_EMIT:
-      break;
-   case TGSI_OPCODE_ENDPRIM:
-      break;
-   case TGSI_OPCODE_BGNLOOP2: {
-      instr->beginLoop();
-      storage->setCurrentBlock(instr->currentBlock());
-      return;
-   }
-      break;
-   case TGSI_OPCODE_BGNSUB: {
-      instr->bgnSub(instno);
-      storage->setCurrentBlock(instr->currentBlock());
-      storage->pushTemps();
-      return;
-   }
-      break;
-   case TGSI_OPCODE_ENDLOOP2: {
-      instr->endLoop();
-      storage->setCurrentBlock(instr->currentBlock());
-      return;
-   }
-      break;
-   case TGSI_OPCODE_ENDSUB: {
-      instr->endSub();
-      storage->setCurrentBlock(instr->currentBlock());
-      storage->popArguments();
-      storage->popTemps();
-      return;
-   }
-      break;
-   case TGSI_OPCODE_NOISE1:
-      break;
-   case TGSI_OPCODE_NOISE2:
-      break;
-   case TGSI_OPCODE_NOISE3:
-      break;
-   case TGSI_OPCODE_NOISE4:
-      break;
-   case TGSI_OPCODE_NOP:
-      break;
-   case TGSI_OPCODE_TEXBEM:
-      break;
-   case TGSI_OPCODE_TEXBEML:
-      break;
-   case TGSI_OPCODE_TEXREG2AR:
-      break;
-   case TGSI_OPCODE_TEXM3X2PAD:
-      break;
-   case TGSI_OPCODE_TEXM3X2TEX:
-      break;
-   case TGSI_OPCODE_TEXM3X3PAD:
-      break;
-   case TGSI_OPCODE_TEXM3X3TEX:
-      break;
-   case TGSI_OPCODE_TEXM3X3SPEC:
-      break;
-   case TGSI_OPCODE_TEXM3X3VSPEC:
-      break;
-   case TGSI_OPCODE_TEXREG2GB:
-      break;
-   case TGSI_OPCODE_TEXREG2RGB:
-      break;
-   case TGSI_OPCODE_TEXDP3TEX:
-      break;
-   case TGSI_OPCODE_TEXDP3:
-      break;
-   case TGSI_OPCODE_TEXM3X3:
-      break;
-   case TGSI_OPCODE_TEXM3X2DEPTH:
-      break;
-   case TGSI_OPCODE_TEXDEPTH:
-      break;
-   case TGSI_OPCODE_BEM:
-      break;
-   case TGSI_OPCODE_M4X3:
-      break;
-   case TGSI_OPCODE_M3X4:
-      break;
-   case TGSI_OPCODE_M3X3:
-      break;
-   case TGSI_OPCODE_M3X2:
-      break;
-   case TGSI_OPCODE_NRM4:
-      break;
-   case TGSI_OPCODE_CALLNZ:
-      break;
-   case TGSI_OPCODE_IFC:
-      break;
-   case TGSI_OPCODE_BREAKC:
-      break;
-   case TGSI_OPCODE_KIL:
-      break;
-   case TGSI_OPCODE_END:
-      instr->end();
-      return;
-      break;
-   default:
-      fprintf(stderr, "ERROR: Unknown opcode %d\n",
-              inst->Instruction.Opcode);
-      assert(0);
-      break;
-   }
-
-   if (!out) {
-      fprintf(stderr, "ERROR: unsupported opcode %d\n",
-              inst->Instruction.Opcode);
-      assert(!"Unsupported opcode");
-   }
-
-   /* # not sure if we need this */
-   switch( inst->Instruction.Saturate ) {
-   case TGSI_SAT_NONE:
-      break;
-   case TGSI_SAT_ZERO_ONE:
-      /*TXT( "_SAT" );*/
-      break;
-   case TGSI_SAT_MINUS_PLUS_ONE:
-      /*TXT( "_SAT[-1,1]" );*/
-      break;
-   default:
-      assert( 0 );
-   }
-
-   /* store results  */
-   for (int i = 0; i < inst->Instruction.NumDstRegs; ++i) {
-      struct tgsi_full_dst_register *dst = &inst->FullDstRegisters[i];
-
-      if (dst->DstRegister.File == TGSI_FILE_OUTPUT) {
-         storage->setOutputElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
-      } else if (dst->DstRegister.File == TGSI_FILE_TEMPORARY) {
-         storage->setTempElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
-      } else if (dst->DstRegister.File == TGSI_FILE_ADDRESS) {
-         storage->setAddrElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
-      } else {
-         fprintf(stderr, "ERROR: unsupported LLVM destination!");
-         assert(!"wrong destination");
-      }
-   }
-}
-
-static llvm::Module *
-tgsi_to_llvm(struct gallivm_prog *prog, const struct tgsi_token *tokens)
-{
-   llvm::Module *mod = createBaseShader();
-   struct tgsi_parse_context parse;
-   struct tgsi_full_instruction fi;
-   struct tgsi_full_declaration fd;
-   unsigned instno = 0;
-   Function* shader = mod->getFunction("execute_shader");
-   std::ostringstream stream;
-   if (prog->type == GALLIVM_VS) {
-      stream << "vs_shader";
-   } else {
-      stream << "fs_shader";
-   }
-   stream << prog->id;
-   std::string func_name = stream.str();
-   shader->setName(func_name.c_str());
-
-   Function::arg_iterator args = shader->arg_begin();
-   Value *ptr_INPUT = args++;
-   ptr_INPUT->setName("input");
-
-   BasicBlock *label_entry = new BasicBlock("entry", shader, 0);
-
-   tgsi_parse_init(&parse, tokens);
-
-   fi = tgsi_default_full_instruction();
-   fd = tgsi_default_full_declaration();
-   Storage storage(label_entry, ptr_INPUT);
-   Instructions instr(mod, shader, label_entry, &storage);
-   while(!tgsi_parse_end_of_tokens(&parse)) {
-      tgsi_parse_token(&parse);
-
-      switch (parse.FullToken.Token.Type) {
-      case TGSI_TOKEN_TYPE_DECLARATION:
-         translate_declaration(prog, mod, &storage,
-                               &parse.FullToken.FullDeclaration,
-                               &fd);
-         break;
-
-      case TGSI_TOKEN_TYPE_IMMEDIATE:
-         translate_immediate(&storage,
-                             &parse.FullToken.FullImmediate);
-         break;
-
-      case TGSI_TOKEN_TYPE_INSTRUCTION:
-         translate_instruction(mod, &storage, &instr,
-                               &parse.FullToken.FullInstruction,
-                               &fi, instno);
-         ++instno;
-         break;
-
-      default:
-         assert(0);
-      }
-   }
-
-   tgsi_parse_free(&parse);
-
-   prog->num_consts = storage.numConsts();
-   return mod;
-}
-
-/*!
-  Translates the TGSI tokens into LLVM format. Translated representation
-  is stored in the gallivm_prog and returned.
-  After calling this function the gallivm_prog can either be used with a custom
-  code generator to generate machine code for the GPU which the code generator
-  addresses or it can be jit compiled with gallivm_cpu_jit_compile and executed
-  with gallivm_prog_exec to run the module on the CPU.
- */
-struct gallivm_prog *
-gallivm_from_tgsi(const struct tgsi_token *tokens, enum gallivm_shader_type type)
-{
-   std::cout << "Creating llvm from: " <<std::endl;
-   ++GLOBAL_ID;
-   struct gallivm_prog *gallivm =
-      (struct gallivm_prog *)calloc(1, sizeof(struct gallivm_prog));
-   gallivm->id = GLOBAL_ID;
-   gallivm->type = type;
-   tgsi_dump(tokens, 0);
-
-   llvm::Module *mod = tgsi_to_llvm(gallivm, tokens);
-   gallivm->module = mod;
-   gallivm_prog_dump(gallivm, 0);
-
-   /* Run optimization passes over it */
-   PassManager passes;
-   passes.add(new TargetData(mod));
-   AddStandardCompilePasses(passes);
-   passes.run(*mod);
-
-   gallivm->module = mod;
-
-   gallivm_prog_dump(gallivm, 0);
-
-   return gallivm;
-}
-
-
 void gallivm_prog_delete(struct gallivm_prog *prog)
 {
-   llvm::Module *mod = static_cast<llvm::Module*>(prog->module);
-   delete mod;
+   delete prog->module;
    prog->module = 0;
    prog->function = 0;
    free(prog);
 }
 
-typedef void (*vertex_shader_runner)(float (*ainputs)[PIPE_MAX_SHADER_INPUTS][4],
-                                     float (*dests)[PIPE_MAX_SHADER_INPUTS][4],
+typedef void (*vertex_shader_runner)(void *ainputs,
+                                     void *dests,
                                      float (*aconsts)[4],
                                      int num_vertices,
                                      int num_inputs,
@@ -850,8 +160,8 @@ typedef void (*vertex_shader_runner)(float (*ainputs)[PIPE_MAX_SHADER_INPUTS][4]
   function.
  */
 int gallivm_prog_exec(struct gallivm_prog *prog,
-                      float (*inputs)[PIPE_MAX_SHADER_INPUTS][4],
-                      float (*dests)[PIPE_MAX_SHADER_INPUTS][4],
+                      struct tgsi_exec_vector       *inputs,
+                      struct tgsi_exec_vector       *dests,
                       float (*consts)[4],
                       int num_vertices,
                       int num_inputs,
@@ -943,18 +253,15 @@ int gallivm_fragment_shader_exec(struct gallivm_prog *prog,
                  samplers);
 }
 
-void gallivm_prog_dump(struct gallivm_prog *prog, const char *file_prefix)
+void gallivm_ir_dump(struct gallivm_ir *ir, const char *file_prefix)
 {
-   llvm::Module *mod;
-   if (!prog || !prog->module)
+   if (!ir || !ir->module)
       return;
 
-   mod = static_cast<llvm::Module*>(prog->module);
-
    if (file_prefix) {
       std::ostringstream stream;
       stream << file_prefix;
-      stream << prog->id;
+      stream << ir->id;
       stream << ".ll";
       std::string name = stream.str();
       std::ofstream out(name.c_str());
@@ -962,12 +269,12 @@ void gallivm_prog_dump(struct gallivm_prog *prog, const char *file_prefix)
          std::cerr<<"Can't open file : "<<stream.str()<<std::endl;;
          return;
       }
-      out << (*mod);
+      out << (*ir->module);
       out.close();
    } else {
-      const llvm::Module::FunctionListType &funcs = mod->getFunctionList();
+      const llvm::Module::FunctionListType &funcs = ir->module->getFunctionList();
       llvm::Module::FunctionListType::const_iterator itr;
-      std::cout<<"; ---------- Start shader "<<prog->id<<std::endl;
+      std::cout<<"; ---------- Start shader "<<ir->id<<std::endl;
       for (itr = funcs.begin(); itr != funcs.end(); ++itr) {
          const llvm::Function &func = (*itr);
          std::string name = func.getName();
@@ -980,7 +287,7 @@ void gallivm_prog_dump(struct gallivm_prog *prog, const char *file_prefix)
             std::cout<<*found<<std::endl;
          }
       }
-      std::cout<<"; ---------- End shader "<<prog->id<<std::endl;
+      std::cout<<"; ---------- End shader "<<ir->id<<std::endl;
    }
 }
 
@@ -1086,9 +393,63 @@ void gallivm_prog_inputs_interpolate(struct gallivm_prog *prog,
    }
 }
 
-#endif /* MESA_LLVM */
 
+struct gallivm_ir * gallivm_ir_new(enum gallivm_shader_type type)
+{
+   struct gallivm_ir *ir =
+      (struct gallivm_ir *)calloc(1, sizeof(struct gallivm_ir));
+   ++GLOBAL_ID;
+   ir->id   = GLOBAL_ID;
+   ir->type = type;
 
+   return ir;
+}
 
+void gallivm_ir_set_layout(struct gallivm_ir *ir,
+                           enum gallivm_vector_layout layout)
+{
+   ir->layout = layout;
+}
 
+void gallivm_ir_set_components(struct gallivm_ir *ir, int num)
+{
+   ir->num_components = num;
+}
 
+void gallivm_ir_fill_from_tgsi(struct gallivm_ir *ir,
+                               const struct tgsi_token *tokens)
+{
+   std::cout << "Creating llvm from: " <<std::endl;
+   tgsi_dump(tokens, 0);
+
+   llvm::Module *mod = tgsi_to_llvm(ir, tokens);
+   ir->module = mod;
+   gallivm_ir_dump(ir, 0);
+}
+
+void gallivm_ir_delete(struct gallivm_ir *ir)
+{
+   delete ir->module;
+   free(ir);
+}
+
+struct gallivm_prog * gallivm_ir_compile(struct gallivm_ir *ir)
+{
+   struct gallivm_prog *prog =
+      (struct gallivm_prog *)calloc(1, sizeof(struct gallivm_prog));
+   llvm::Module *mod = llvm::CloneModule(ir->module);
+   prog->num_consts = ir->num_consts;
+   memcpy(prog->interpolators, ir->interpolators, sizeof(prog->interpolators));
+   prog->num_interp = ir->num_interp;
+
+   /* Run optimization passes over it */
+   PassManager passes;
+   passes.add(new TargetData(mod));
+   AddStandardCompilePasses(passes);
+   passes.run(*mod);
+   prog->module = mod;
+
+   return prog;
+}
+
+#endif /* MESA_LLVM */
diff --git a/src/mesa/pipe/llvm/gallivm.h b/src/mesa/pipe/llvm/gallivm.h
index 4695de3127..f9f5d5ee74 100644
--- a/src/mesa/pipe/llvm/gallivm.h
+++ b/src/mesa/pipe/llvm/gallivm.h
@@ -43,21 +43,36 @@ extern "C" {
 
 struct tgsi_token;
 
+struct gallivm_ir;
 struct gallivm_prog;
 struct gallivm_cpu_engine;
 struct tgsi_interp_coef;
 struct tgsi_sampler;
+struct tgsi_exec_vector;
 
 enum gallivm_shader_type {
    GALLIVM_VS,
    GALLIVM_FS
 };
 
-struct gallivm_prog *gallivm_from_tgsi(const struct tgsi_token *tokens, enum gallivm_shader_type type);
-void gallivm_prog_delete(struct gallivm_prog *prog);
+enum gallivm_vector_layout {
+   GALLIVM_AOS,
+   GALLIVM_SOA
+};
+
+struct gallivm_ir *gallivm_ir_new(enum gallivm_shader_type type);
+void gallivm_ir_set_layout(struct gallivm_ir *prog,
+                           enum gallivm_vector_layout layout);
+void gallivm_ir_set_components(struct gallivm_ir *prog, int num);
+void gallivm_ir_fill_from_tgsi(struct gallivm_ir *prog,
+                               const struct tgsi_token *tokens);
+void gallivm_ir_delete(struct gallivm_ir *prog);
+
+struct gallivm_prog *gallivm_ir_compile(struct gallivm_ir *ir);
+
 int gallivm_prog_exec(struct gallivm_prog *prog,
-                      float (*inputs)[PIPE_MAX_SHADER_INPUTS][4],
-                      float (*dests)[PIPE_MAX_SHADER_INPUTS][4],
+                      struct tgsi_exec_vector       *inputs,
+                      struct tgsi_exec_vector       *dests,
                       float (*consts)[4],
                       int num_vertices,
                       int num_inputs,
diff --git a/src/mesa/pipe/llvm/gallivm_p.h b/src/mesa/pipe/llvm/gallivm_p.h
new file mode 100644
index 0000000000..2c6e5e8f5f
--- /dev/null
+++ b/src/mesa/pipe/llvm/gallivm_p.h
@@ -0,0 +1,56 @@
+#ifndef GALLIVM_P_H
+#define GALLIVM_P_H
+
+#ifdef MESA_LLVM
+
+namespace llvm {
+   class Module;
+}
+
+#if defined __cplusplus
+extern "C" {
+#endif
+
+enum gallivm_shader_type;
+enum gallivm_vector_layout;
+
+struct gallivm_interpolate {
+   int attrib;
+   int chan;
+   int type;
+};
+
+struct gallivm_ir {
+   llvm::Module *module;
+   int id;
+   enum gallivm_shader_type type;
+   enum gallivm_vector_layout layout;
+   int num_components;
+   int   num_consts;
+
+   //FIXME: this might not be enough for some shaders
+   struct gallivm_interpolate interpolators[32*4];
+   int   num_interp;
+};
+
+struct gallivm_prog {
+   llvm::Module *module;
+   void *function;
+
+   int   id;
+   enum gallivm_shader_type type;
+
+   int   num_consts;
+
+   //FIXME: this might not be enough for some shaders
+   struct gallivm_interpolate interpolators[32*4];
+   int   num_interp;
+};
+
+#endif /* MESA_LLVM */
+
+#if defined __cplusplus
+} // extern "C"
+#endif
+
+#endif
diff --git a/src/mesa/pipe/llvm/llvm_base_shader.cpp b/src/mesa/pipe/llvm/llvm_base_shader.cpp
index b574b550ae..90a25a440a 100644
--- a/src/mesa/pipe/llvm/llvm_base_shader.cpp
+++ b/src/mesa/pipe/llvm/llvm_base_shader.cpp
@@ -104,8 +104,8 @@ Module* createBaseShader() {
     /*isVarArg=*/false);
   
   std::vector<const Type*>FuncTy_18_args;
-  FuncTy_18_args.push_back(PointerTy_12);
-  FuncTy_18_args.push_back(PointerTy_12);
+  FuncTy_18_args.push_back(PointerTy_9);
+  FuncTy_18_args.push_back(PointerTy_9);
   FuncTy_18_args.push_back(PointerTy_7);
   FuncTy_18_args.push_back(IntegerType::get(32));
   FuncTy_18_args.push_back(IntegerType::get(32));
@@ -526,14 +526,23 @@ Module* createBaseShader() {
     BasicBlock* label_forbody_preheader_i = new BasicBlock("forbody.preheader.i",func_run_vertex_shader,0);
     BasicBlock* label_forbody_i = new BasicBlock("forbody.i",func_run_vertex_shader,0);
     BasicBlock* label_from_consts_exit = new BasicBlock("from_consts.exit",func_run_vertex_shader,0);
-    BasicBlock* label_forbody_preheader_91 = new BasicBlock("forbody.preheader",func_run_vertex_shader,0);
-    BasicBlock* label_forbody_92 = new BasicBlock("forbody",func_run_vertex_shader,0);
-    BasicBlock* label_afterfor_93 = new BasicBlock("afterfor",func_run_vertex_shader,0);
     
     // Block entry (label_entry_90)
     AllocaInst* ptr_consts = new AllocaInst(ArrayTy_20, "consts", label_entry_90);
     AllocaInst* ptr_temps = new AllocaInst(ArrayTy_22, "temps", label_entry_90);
     AllocaInst* ptr_args = new AllocaInst(StructTy_struct_ShaderInput, "args", label_entry_90);
+    std::vector<Value*> ptr_tmp_indices;
+    ptr_tmp_indices.push_back(const_int32_29);
+    ptr_tmp_indices.push_back(const_int32_29);
+    Instruction* ptr_tmp = new GetElementPtrInst(ptr_args, ptr_tmp_indices.begin(), ptr_tmp_indices.end(), "tmp", label_entry_90);
+    CastInst* ptr_conv = new BitCastInst(ptr_results, PointerTy_0, "conv", label_entry_90);
+    StoreInst* void_91 = new StoreInst(ptr_conv, ptr_tmp, false, label_entry_90);
+    std::vector<Value*> ptr_tmp2_indices;
+    ptr_tmp2_indices.push_back(const_int32_29);
+    ptr_tmp2_indices.push_back(const_int32_31);
+    Instruction* ptr_tmp2 = new GetElementPtrInst(ptr_args, ptr_tmp2_indices.begin(), ptr_tmp2_indices.end(), "tmp2", label_entry_90);
+    CastInst* ptr_conv4 = new BitCastInst(ptr_inputs, PointerTy_0, "conv4", label_entry_90);
+    StoreInst* void_92 = new StoreInst(ptr_conv4, ptr_tmp2, false, label_entry_90);
     ICmpInst* int1_cmp_i = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_consts, const_int32_29, "cmp.i", label_entry_90);
     new BranchInst(label_forbody_preheader_i, label_from_consts_exit, int1_cmp_i, label_entry_90);
     
@@ -544,17 +553,17 @@ Module* createBaseShader() {
     new BranchInst(label_forbody_i, label_forbody_preheader_i);
     
     // Block forbody.i (label_forbody_i)
-    Argument* fwdref_96 = new Argument(IntegerType::get(32));
+    Argument* fwdref_95 = new Argument(IntegerType::get(32));
     PHINode* int32_i_0_reg2mem_0_i = new PHINode(IntegerType::get(32), "i.0.reg2mem.0.i", label_forbody_i);
     int32_i_0_reg2mem_0_i->reserveOperandSpace(2);
     int32_i_0_reg2mem_0_i->addIncoming(const_int32_29, label_forbody_preheader_i);
-    int32_i_0_reg2mem_0_i->addIncoming(fwdref_96, label_forbody_i);
+    int32_i_0_reg2mem_0_i->addIncoming(fwdref_95, label_forbody_i);
     
-    Argument* fwdref_97 = new Argument(VectorTy_1);
+    Argument* fwdref_96 = new Argument(VectorTy_1);
     PHINode* packed_vec_0_reg2mem_0_i = new PHINode(VectorTy_1, "vec.0.reg2mem.0.i", label_forbody_i);
     packed_vec_0_reg2mem_0_i->reserveOperandSpace(2);
     packed_vec_0_reg2mem_0_i->addIncoming(const_packed_32, label_forbody_preheader_i);
-    packed_vec_0_reg2mem_0_i->addIncoming(fwdref_97, label_forbody_i);
+    packed_vec_0_reg2mem_0_i->addIncoming(fwdref_96, label_forbody_i);
     
     std::vector<Value*> ptr_arraydecay_i_indices;
     ptr_arraydecay_i_indices.push_back(int32_i_0_reg2mem_0_i);
@@ -584,80 +593,40 @@ Module* createBaseShader() {
     ptr_arrayidx34_i_indices.push_back(const_int32_29);
     ptr_arrayidx34_i_indices.push_back(int32_i_0_reg2mem_0_i);
     Instruction* ptr_arrayidx34_i = new GetElementPtrInst(ptr_consts, ptr_arrayidx34_i_indices.begin(), ptr_arrayidx34_i_indices.end(), "arrayidx34.i", label_forbody_i);
-    StoreInst* void_98 = new StoreInst(packed_tmp31_i, ptr_arrayidx34_i, false, label_forbody_i);
-    BinaryOperator* int32_indvar_next8 = BinaryOperator::create(Instruction::Add, int32_i_0_reg2mem_0_i, const_int32_31, "indvar.next8", label_forbody_i);
-    ICmpInst* int1_exitcond9 = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next8, int32_tmp10_i, "exitcond9", label_forbody_i);
-    new BranchInst(label_from_consts_exit, label_forbody_i, int1_exitcond9, label_forbody_i);
+    StoreInst* void_97 = new StoreInst(packed_tmp31_i, ptr_arrayidx34_i, false, label_forbody_i);
+    BinaryOperator* int32_indvar_next_98 = BinaryOperator::create(Instruction::Add, int32_i_0_reg2mem_0_i, const_int32_31, "indvar.next", label_forbody_i);
+    ICmpInst* int1_exitcond_99 = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next_98, int32_tmp10_i, "exitcond", label_forbody_i);
+    new BranchInst(label_from_consts_exit, label_forbody_i, int1_exitcond_99, label_forbody_i);
     
     // Block from_consts.exit (label_from_consts_exit)
-    std::vector<Value*> ptr_tmp2_indices;
-    ptr_tmp2_indices.push_back(const_int32_29);
-    ptr_tmp2_indices.push_back(const_int32_34);
-    Instruction* ptr_tmp2 = new GetElementPtrInst(ptr_args, ptr_tmp2_indices.begin(), ptr_tmp2_indices.end(), "tmp2", label_from_consts_exit);
-    std::vector<Value*> ptr_arraydecay3_indices;
-    ptr_arraydecay3_indices.push_back(const_int32_29);
-    ptr_arraydecay3_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay3 = new GetElementPtrInst(ptr_consts, ptr_arraydecay3_indices.begin(), ptr_arraydecay3_indices.end(), "arraydecay3", label_from_consts_exit);
-    StoreInst* void_100 = new StoreInst(ptr_arraydecay3, ptr_tmp2, false, label_from_consts_exit);
-    std::vector<Value*> ptr_tmp4_indices;
-    ptr_tmp4_indices.push_back(const_int32_29);
-    ptr_tmp4_indices.push_back(const_int32_33);
-    Instruction* ptr_tmp4 = new GetElementPtrInst(ptr_args, ptr_tmp4_indices.begin(), ptr_tmp4_indices.end(), "tmp4", label_from_consts_exit);
-    std::vector<Value*> ptr_arraydecay5_indices;
-    ptr_arraydecay5_indices.push_back(const_int32_29);
-    ptr_arraydecay5_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay5 = new GetElementPtrInst(ptr_temps, ptr_arraydecay5_indices.begin(), ptr_arraydecay5_indices.end(), "arraydecay5", label_from_consts_exit);
-    StoreInst* void_101 = new StoreInst(ptr_arraydecay5, ptr_tmp4, false, label_from_consts_exit);
-    ICmpInst* int1_cmp_102 = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_vertices, const_int32_29, "cmp", label_from_consts_exit);
-    new BranchInst(label_forbody_preheader_91, label_afterfor_93, int1_cmp_102, label_from_consts_exit);
-    
-    // Block forbody.preheader (label_forbody_preheader_91)
-    std::vector<Value*> ptr_tmp8_indices;
-    ptr_tmp8_indices.push_back(const_int32_29);
-    ptr_tmp8_indices.push_back(const_int32_29);
-    Instruction* ptr_tmp8 = new GetElementPtrInst(ptr_args, ptr_tmp8_indices.begin(), ptr_tmp8_indices.end(), "tmp8", label_forbody_preheader_91);
-    std::vector<Value*> ptr_tmp12_indices;
-    ptr_tmp12_indices.push_back(const_int32_29);
-    ptr_tmp12_indices.push_back(const_int32_31);
-    Instruction* ptr_tmp12 = new GetElementPtrInst(ptr_args, ptr_tmp12_indices.begin(), ptr_tmp12_indices.end(), "tmp12", label_forbody_preheader_91);
-    BinaryOperator* int32_tmp_104 = BinaryOperator::create(Instruction::Add, int32_num_vertices, const_int32_30, "tmp", label_forbody_preheader_91);
-    ICmpInst* int1_tmp6 = new ICmpInst(ICmpInst::ICMP_SLT, int32_tmp_104, const_int32_29, "tmp6", label_forbody_preheader_91);
-    SelectInst* int32_tmp7 = new SelectInst(int1_tmp6, const_int32_31, int32_num_vertices, "tmp7", label_forbody_preheader_91);
-    new BranchInst(label_forbody_92, label_forbody_preheader_91);
-    
-    // Block forbody (label_forbody_92)
-    Argument* fwdref_107 = new Argument(IntegerType::get(32));
-    PHINode* int32_i_0_reg2mem_0_106 = new PHINode(IntegerType::get(32), "i.0.reg2mem.0", label_forbody_92);
-    int32_i_0_reg2mem_0_106->reserveOperandSpace(2);
-    int32_i_0_reg2mem_0_106->addIncoming(const_int32_29, label_forbody_preheader_91);
-    int32_i_0_reg2mem_0_106->addIncoming(fwdref_107, label_forbody_92);
-    
-    std::vector<Value*> ptr_arraydecay11_108_indices;
-    ptr_arraydecay11_108_indices.push_back(int32_i_0_reg2mem_0_106);
-    ptr_arraydecay11_108_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay11_108 = new GetElementPtrInst(ptr_results, ptr_arraydecay11_108_indices.begin(), ptr_arraydecay11_108_indices.end(), "arraydecay11", label_forbody_92);
-    StoreInst* void_109 = new StoreInst(ptr_arraydecay11_108, ptr_tmp8, false, label_forbody_92);
-    std::vector<Value*> ptr_arraydecay16_indices;
-    ptr_arraydecay16_indices.push_back(int32_i_0_reg2mem_0_106);
-    ptr_arraydecay16_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay16 = new GetElementPtrInst(ptr_inputs, ptr_arraydecay16_indices.begin(), ptr_arraydecay16_indices.end(), "arraydecay16", label_forbody_92);
-    StoreInst* void_110 = new StoreInst(ptr_arraydecay16, ptr_tmp12, false, label_forbody_92);
-    CallInst* void_111 = new CallInst(func_execute_shader, ptr_args, "", label_forbody_92);
-    void_111->setCallingConv(CallingConv::C);
-    void_111->setTailCall(false);const ParamAttrsList *void_111_PAL = 0;
-    void_111->setParamAttrs(void_111_PAL);
-    
-    BinaryOperator* int32_indvar_next_112 = BinaryOperator::create(Instruction::Add, int32_i_0_reg2mem_0_106, const_int32_31, "indvar.next", label_forbody_92);
-    ICmpInst* int1_exitcond_113 = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next_112, int32_tmp7, "exitcond", label_forbody_92);
-    new BranchInst(label_afterfor_93, label_forbody_92, int1_exitcond_113, label_forbody_92);
-    
-    // Block afterfor (label_afterfor_93)
-    new ReturnInst(label_afterfor_93);
+    std::vector<Value*> ptr_tmp7_indices;
+    ptr_tmp7_indices.push_back(const_int32_29);
+    ptr_tmp7_indices.push_back(const_int32_34);
+    Instruction* ptr_tmp7 = new GetElementPtrInst(ptr_args, ptr_tmp7_indices.begin(), ptr_tmp7_indices.end(), "tmp7", label_from_consts_exit);
+    std::vector<Value*> ptr_arraydecay8_indices;
+    ptr_arraydecay8_indices.push_back(const_int32_29);
+    ptr_arraydecay8_indices.push_back(const_int32_29);
+    Instruction* ptr_arraydecay8 = new GetElementPtrInst(ptr_consts, ptr_arraydecay8_indices.begin(), ptr_arraydecay8_indices.end(), "arraydecay8", label_from_consts_exit);
+    StoreInst* void_101 = new StoreInst(ptr_arraydecay8, ptr_tmp7, false, label_from_consts_exit);
+    std::vector<Value*> ptr_tmp9_indices;
+    ptr_tmp9_indices.push_back(const_int32_29);
+    ptr_tmp9_indices.push_back(const_int32_33);
+    Instruction* ptr_tmp9 = new GetElementPtrInst(ptr_args, ptr_tmp9_indices.begin(), ptr_tmp9_indices.end(), "tmp9", label_from_consts_exit);
+    std::vector<Value*> ptr_arraydecay10_indices;
+    ptr_arraydecay10_indices.push_back(const_int32_29);
+    ptr_arraydecay10_indices.push_back(const_int32_29);
+    Instruction* ptr_arraydecay10 = new GetElementPtrInst(ptr_temps, ptr_arraydecay10_indices.begin(), ptr_arraydecay10_indices.end(), "arraydecay10", label_from_consts_exit);
+    StoreInst* void_102 = new StoreInst(ptr_arraydecay10, ptr_tmp9, false, label_from_consts_exit);
+    CallInst* void_103 = new CallInst(func_execute_shader, ptr_args, "", label_from_consts_exit);
+    void_103->setCallingConv(CallingConv::C);
+    void_103->setTailCall(false);const ParamAttrsList *void_103_PAL = 0;
+    void_103->setParamAttrs(void_103_PAL);
+    
+    new ReturnInst(label_from_consts_exit);
     
     // Resolve Forward References
-    fwdref_107->replaceAllUsesWith(int32_indvar_next_112); delete fwdref_107;
-    fwdref_97->replaceAllUsesWith(packed_tmp31_i); delete fwdref_97;
-    fwdref_96->replaceAllUsesWith(int32_indvar_next8); delete fwdref_96;
+    fwdref_96->replaceAllUsesWith(packed_tmp31_i); delete fwdref_96;
+    fwdref_95->replaceAllUsesWith(int32_indvar_next_98); delete fwdref_95;
     
   }
   
@@ -668,195 +637,195 @@ Module* createBaseShader() {
     float_x->setName("x");
     Value* float_y = args++;
     float_y->setName("y");
-    Value* ptr_results_116 = args++;
-    ptr_results_116->setName("results");
-    Value* ptr_inputs_117 = args++;
-    ptr_inputs_117->setName("inputs");
-    Value* int32_num_inputs_118 = args++;
-    int32_num_inputs_118->setName("num_inputs");
-    Value* ptr_aconsts_119 = args++;
-    ptr_aconsts_119->setName("aconsts");
-    Value* int32_num_consts_120 = args++;
-    int32_num_consts_120->setName("num_consts");
+    Value* ptr_results_105 = args++;
+    ptr_results_105->setName("results");
+    Value* ptr_inputs_106 = args++;
+    ptr_inputs_106->setName("inputs");
+    Value* int32_num_inputs_107 = args++;
+    int32_num_inputs_107->setName("num_inputs");
+    Value* ptr_aconsts_108 = args++;
+    ptr_aconsts_108->setName("aconsts");
+    Value* int32_num_consts_109 = args++;
+    int32_num_consts_109->setName("num_consts");
     Value* ptr_samplers = args++;
     ptr_samplers->setName("samplers");
     
-    BasicBlock* label_entry_121 = new BasicBlock("entry",func_run_fragment_shader,0);
-    BasicBlock* label_forbody_preheader_i_122 = new BasicBlock("forbody.preheader.i",func_run_fragment_shader,0);
-    BasicBlock* label_forbody_i_123 = new BasicBlock("forbody.i",func_run_fragment_shader,0);
-    BasicBlock* label_from_consts_exit_124 = new BasicBlock("from_consts.exit",func_run_fragment_shader,0);
-    
-    // Block entry (label_entry_121)
-    AllocaInst* ptr_consts_125 = new AllocaInst(ArrayTy_20, "consts", label_entry_121);
-    AllocaInst* ptr_temps_126 = new AllocaInst(ArrayTy_22, "temps", label_entry_121);
-    AllocaInst* ptr_args_127 = new AllocaInst(StructTy_struct_ShaderInput, "args", label_entry_121);
-    std::vector<Value*> ptr_tmp_indices;
-    ptr_tmp_indices.push_back(const_int32_29);
-    ptr_tmp_indices.push_back(const_int32_35);
-    Instruction* ptr_tmp = new GetElementPtrInst(ptr_args_127, ptr_tmp_indices.begin(), ptr_tmp_indices.end(), "tmp", label_entry_121);
-    StoreInst* void_128 = new StoreInst(const_int32_29, ptr_tmp, false, label_entry_121);
-    ICmpInst* int1_cmp_i_129 = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_consts_120, const_int32_29, "cmp.i", label_entry_121);
-    new BranchInst(label_forbody_preheader_i_122, label_from_consts_exit_124, int1_cmp_i_129, label_entry_121);
-    
-    // Block forbody.preheader.i (label_forbody_preheader_i_122)
-    BinaryOperator* int32_tmp_i_131 = BinaryOperator::create(Instruction::Add, int32_num_consts_120, const_int32_30, "tmp.i", label_forbody_preheader_i_122);
-    ICmpInst* int1_tmp9_i_132 = new ICmpInst(ICmpInst::ICMP_SLT, int32_tmp_i_131, const_int32_29, "tmp9.i", label_forbody_preheader_i_122);
-    SelectInst* int32_tmp10_i_133 = new SelectInst(int1_tmp9_i_132, const_int32_31, int32_num_consts_120, "tmp10.i", label_forbody_preheader_i_122);
-    new BranchInst(label_forbody_i_123, label_forbody_preheader_i_122);
-    
-    // Block forbody.i (label_forbody_i_123)
-    Argument* fwdref_136 = new Argument(IntegerType::get(32));
-    PHINode* int32_i_0_reg2mem_0_i_135 = new PHINode(IntegerType::get(32), "i.0.reg2mem.0.i", label_forbody_i_123);
-    int32_i_0_reg2mem_0_i_135->reserveOperandSpace(2);
-    int32_i_0_reg2mem_0_i_135->addIncoming(const_int32_29, label_forbody_preheader_i_122);
-    int32_i_0_reg2mem_0_i_135->addIncoming(fwdref_136, label_forbody_i_123);
-    
-    Argument* fwdref_138 = new Argument(VectorTy_1);
-    PHINode* packed_vec_0_reg2mem_0_i_137 = new PHINode(VectorTy_1, "vec.0.reg2mem.0.i", label_forbody_i_123);
-    packed_vec_0_reg2mem_0_i_137->reserveOperandSpace(2);
-    packed_vec_0_reg2mem_0_i_137->addIncoming(const_packed_32, label_forbody_preheader_i_122);
-    packed_vec_0_reg2mem_0_i_137->addIncoming(fwdref_138, label_forbody_i_123);
-    
-    std::vector<Value*> ptr_arraydecay_i_139_indices;
-    ptr_arraydecay_i_139_indices.push_back(int32_i_0_reg2mem_0_i_135);
-    ptr_arraydecay_i_139_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay_i_139 = new GetElementPtrInst(ptr_aconsts_119, ptr_arraydecay_i_139_indices.begin(), ptr_arraydecay_i_139_indices.end(), "arraydecay.i", label_forbody_i_123);
-    LoadInst* float_tmp5_i_140 = new LoadInst(ptr_arraydecay_i_139, "tmp5.i", false, label_forbody_i_123);
-    InsertElementInst* packed_tmp7_i_141 = new InsertElementInst(packed_vec_0_reg2mem_0_i_137, float_tmp5_i_140, const_int32_29, "tmp7.i", label_forbody_i_123);
-    std::vector<Value*> ptr_arrayidx12_i_142_indices;
-    ptr_arrayidx12_i_142_indices.push_back(int32_i_0_reg2mem_0_i_135);
-    ptr_arrayidx12_i_142_indices.push_back(const_int32_31);
-    Instruction* ptr_arrayidx12_i_142 = new GetElementPtrInst(ptr_aconsts_119, ptr_arrayidx12_i_142_indices.begin(), ptr_arrayidx12_i_142_indices.end(), "arrayidx12.i", label_forbody_i_123);
-    LoadInst* float_tmp13_i_143 = new LoadInst(ptr_arrayidx12_i_142, "tmp13.i", false, label_forbody_i_123);
-    InsertElementInst* packed_tmp15_i_144 = new InsertElementInst(packed_tmp7_i_141, float_tmp13_i_143, const_int32_31, "tmp15.i", label_forbody_i_123);
-    std::vector<Value*> ptr_arrayidx20_i_145_indices;
-    ptr_arrayidx20_i_145_indices.push_back(int32_i_0_reg2mem_0_i_135);
-    ptr_arrayidx20_i_145_indices.push_back(const_int32_33);
-    Instruction* ptr_arrayidx20_i_145 = new GetElementPtrInst(ptr_aconsts_119, ptr_arrayidx20_i_145_indices.begin(), ptr_arrayidx20_i_145_indices.end(), "arrayidx20.i", label_forbody_i_123);
-    LoadInst* float_tmp21_i_146 = new LoadInst(ptr_arrayidx20_i_145, "tmp21.i", false, label_forbody_i_123);
-    InsertElementInst* packed_tmp23_i_147 = new InsertElementInst(packed_tmp15_i_144, float_tmp21_i_146, const_int32_33, "tmp23.i", label_forbody_i_123);
-    std::vector<Value*> ptr_arrayidx28_i_148_indices;
-    ptr_arrayidx28_i_148_indices.push_back(int32_i_0_reg2mem_0_i_135);
-    ptr_arrayidx28_i_148_indices.push_back(const_int32_34);
-    Instruction* ptr_arrayidx28_i_148 = new GetElementPtrInst(ptr_aconsts_119, ptr_arrayidx28_i_148_indices.begin(), ptr_arrayidx28_i_148_indices.end(), "arrayidx28.i", label_forbody_i_123);
-    LoadInst* float_tmp29_i_149 = new LoadInst(ptr_arrayidx28_i_148, "tmp29.i", false, label_forbody_i_123);
-    InsertElementInst* packed_tmp31_i_150 = new InsertElementInst(packed_tmp23_i_147, float_tmp29_i_149, const_int32_34, "tmp31.i", label_forbody_i_123);
-    std::vector<Value*> ptr_arrayidx34_i_151_indices;
-    ptr_arrayidx34_i_151_indices.push_back(const_int32_29);
-    ptr_arrayidx34_i_151_indices.push_back(int32_i_0_reg2mem_0_i_135);
-    Instruction* ptr_arrayidx34_i_151 = new GetElementPtrInst(ptr_consts_125, ptr_arrayidx34_i_151_indices.begin(), ptr_arrayidx34_i_151_indices.end(), "arrayidx34.i", label_forbody_i_123);
-    StoreInst* void_152 = new StoreInst(packed_tmp31_i_150, ptr_arrayidx34_i_151, false, label_forbody_i_123);
-    BinaryOperator* int32_indvar_next7 = BinaryOperator::create(Instruction::Add, int32_i_0_reg2mem_0_i_135, const_int32_31, "indvar.next7", label_forbody_i_123);
-    ICmpInst* int1_exitcond8 = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next7, int32_tmp10_i_133, "exitcond8", label_forbody_i_123);
-    new BranchInst(label_from_consts_exit_124, label_forbody_i_123, int1_exitcond8, label_forbody_i_123);
-    
-    // Block from_consts.exit (label_from_consts_exit_124)
+    BasicBlock* label_entry_110 = new BasicBlock("entry",func_run_fragment_shader,0);
+    BasicBlock* label_forbody_preheader_i_111 = new BasicBlock("forbody.preheader.i",func_run_fragment_shader,0);
+    BasicBlock* label_forbody_i_112 = new BasicBlock("forbody.i",func_run_fragment_shader,0);
+    BasicBlock* label_from_consts_exit_113 = new BasicBlock("from_consts.exit",func_run_fragment_shader,0);
+    
+    // Block entry (label_entry_110)
+    AllocaInst* ptr_consts_114 = new AllocaInst(ArrayTy_20, "consts", label_entry_110);
+    AllocaInst* ptr_temps_115 = new AllocaInst(ArrayTy_22, "temps", label_entry_110);
+    AllocaInst* ptr_args_116 = new AllocaInst(StructTy_struct_ShaderInput, "args", label_entry_110);
+    std::vector<Value*> ptr_tmp_117_indices;
+    ptr_tmp_117_indices.push_back(const_int32_29);
+    ptr_tmp_117_indices.push_back(const_int32_35);
+    Instruction* ptr_tmp_117 = new GetElementPtrInst(ptr_args_116, ptr_tmp_117_indices.begin(), ptr_tmp_117_indices.end(), "tmp", label_entry_110);
+    StoreInst* void_118 = new StoreInst(const_int32_29, ptr_tmp_117, false, label_entry_110);
+    ICmpInst* int1_cmp_i_119 = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_consts_109, const_int32_29, "cmp.i", label_entry_110);
+    new BranchInst(label_forbody_preheader_i_111, label_from_consts_exit_113, int1_cmp_i_119, label_entry_110);
+    
+    // Block forbody.preheader.i (label_forbody_preheader_i_111)
+    BinaryOperator* int32_tmp_i_121 = BinaryOperator::create(Instruction::Add, int32_num_consts_109, const_int32_30, "tmp.i", label_forbody_preheader_i_111);
+    ICmpInst* int1_tmp9_i_122 = new ICmpInst(ICmpInst::ICMP_SLT, int32_tmp_i_121, const_int32_29, "tmp9.i", label_forbody_preheader_i_111);
+    SelectInst* int32_tmp10_i_123 = new SelectInst(int1_tmp9_i_122, const_int32_31, int32_num_consts_109, "tmp10.i", label_forbody_preheader_i_111);
+    new BranchInst(label_forbody_i_112, label_forbody_preheader_i_111);
+    
+    // Block forbody.i (label_forbody_i_112)
+    Argument* fwdref_126 = new Argument(IntegerType::get(32));
+    PHINode* int32_i_0_reg2mem_0_i_125 = new PHINode(IntegerType::get(32), "i.0.reg2mem.0.i", label_forbody_i_112);
+    int32_i_0_reg2mem_0_i_125->reserveOperandSpace(2);
+    int32_i_0_reg2mem_0_i_125->addIncoming(const_int32_29, label_forbody_preheader_i_111);
+    int32_i_0_reg2mem_0_i_125->addIncoming(fwdref_126, label_forbody_i_112);
+    
+    Argument* fwdref_128 = new Argument(VectorTy_1);
+    PHINode* packed_vec_0_reg2mem_0_i_127 = new PHINode(VectorTy_1, "vec.0.reg2mem.0.i", label_forbody_i_112);
+    packed_vec_0_reg2mem_0_i_127->reserveOperandSpace(2);
+    packed_vec_0_reg2mem_0_i_127->addIncoming(const_packed_32, label_forbody_preheader_i_111);
+    packed_vec_0_reg2mem_0_i_127->addIncoming(fwdref_128, label_forbody_i_112);
+    
+    std::vector<Value*> ptr_arraydecay_i_129_indices;
+    ptr_arraydecay_i_129_indices.push_back(int32_i_0_reg2mem_0_i_125);
+    ptr_arraydecay_i_129_indices.push_back(const_int32_29);
+    Instruction* ptr_arraydecay_i_129 = new GetElementPtrInst(ptr_aconsts_108, ptr_arraydecay_i_129_indices.begin(), ptr_arraydecay_i_129_indices.end(), "arraydecay.i", label_forbody_i_112);
+    LoadInst* float_tmp5_i_130 = new LoadInst(ptr_arraydecay_i_129, "tmp5.i", false, label_forbody_i_112);
+    InsertElementInst* packed_tmp7_i_131 = new InsertElementInst(packed_vec_0_reg2mem_0_i_127, float_tmp5_i_130, const_int32_29, "tmp7.i", label_forbody_i_112);
+    std::vector<Value*> ptr_arrayidx12_i_132_indices;
+    ptr_arrayidx12_i_132_indices.push_back(int32_i_0_reg2mem_0_i_125);
+    ptr_arrayidx12_i_132_indices.push_back(const_int32_31);
+    Instruction* ptr_arrayidx12_i_132 = new GetElementPtrInst(ptr_aconsts_108, ptr_arrayidx12_i_132_indices.begin(), ptr_arrayidx12_i_132_indices.end(), "arrayidx12.i", label_forbody_i_112);
+    LoadInst* float_tmp13_i_133 = new LoadInst(ptr_arrayidx12_i_132, "tmp13.i", false, label_forbody_i_112);
+    InsertElementInst* packed_tmp15_i_134 = new InsertElementInst(packed_tmp7_i_131, float_tmp13_i_133, const_int32_31, "tmp15.i", label_forbody_i_112);
+    std::vector<Value*> ptr_arrayidx20_i_135_indices;
+    ptr_arrayidx20_i_135_indices.push_back(int32_i_0_reg2mem_0_i_125);
+    ptr_arrayidx20_i_135_indices.push_back(const_int32_33);
+    Instruction* ptr_arrayidx20_i_135 = new GetElementPtrInst(ptr_aconsts_108, ptr_arrayidx20_i_135_indices.begin(), ptr_arrayidx20_i_135_indices.end(), "arrayidx20.i", label_forbody_i_112);
+    LoadInst* float_tmp21_i_136 = new LoadInst(ptr_arrayidx20_i_135, "tmp21.i", false, label_forbody_i_112);
+    InsertElementInst* packed_tmp23_i_137 = new InsertElementInst(packed_tmp15_i_134, float_tmp21_i_136, const_int32_33, "tmp23.i", label_forbody_i_112);
+    std::vector<Value*> ptr_arrayidx28_i_138_indices;
+    ptr_arrayidx28_i_138_indices.push_back(int32_i_0_reg2mem_0_i_125);
+    ptr_arrayidx28_i_138_indices.push_back(const_int32_34);
+    Instruction* ptr_arrayidx28_i_138 = new GetElementPtrInst(ptr_aconsts_108, ptr_arrayidx28_i_138_indices.begin(), ptr_arrayidx28_i_138_indices.end(), "arrayidx28.i", label_forbody_i_112);
+    LoadInst* float_tmp29_i_139 = new LoadInst(ptr_arrayidx28_i_138, "tmp29.i", false, label_forbody_i_112);
+    InsertElementInst* packed_tmp31_i_140 = new InsertElementInst(packed_tmp23_i_137, float_tmp29_i_139, const_int32_34, "tmp31.i", label_forbody_i_112);
+    std::vector<Value*> ptr_arrayidx34_i_141_indices;
+    ptr_arrayidx34_i_141_indices.push_back(const_int32_29);
+    ptr_arrayidx34_i_141_indices.push_back(int32_i_0_reg2mem_0_i_125);
+    Instruction* ptr_arrayidx34_i_141 = new GetElementPtrInst(ptr_consts_114, ptr_arrayidx34_i_141_indices.begin(), ptr_arrayidx34_i_141_indices.end(), "arrayidx34.i", label_forbody_i_112);
+    StoreInst* void_142 = new StoreInst(packed_tmp31_i_140, ptr_arrayidx34_i_141, false, label_forbody_i_112);
+    BinaryOperator* int32_indvar_next7 = BinaryOperator::create(Instruction::Add, int32_i_0_reg2mem_0_i_125, const_int32_31, "indvar.next7", label_forbody_i_112);
+    ICmpInst* int1_exitcond8 = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next7, int32_tmp10_i_123, "exitcond8", label_forbody_i_112);
+    new BranchInst(label_from_consts_exit_113, label_forbody_i_112, int1_exitcond8, label_forbody_i_112);
+    
+    // Block from_consts.exit (label_from_consts_exit_113)
     std::vector<Value*> ptr_tmp3_indices;
     ptr_tmp3_indices.push_back(const_int32_29);
     ptr_tmp3_indices.push_back(const_int32_34);
-    Instruction* ptr_tmp3 = new GetElementPtrInst(ptr_args_127, ptr_tmp3_indices.begin(), ptr_tmp3_indices.end(), "tmp3", label_from_consts_exit_124);
+    Instruction* ptr_tmp3 = new GetElementPtrInst(ptr_args_116, ptr_tmp3_indices.begin(), ptr_tmp3_indices.end(), "tmp3", label_from_consts_exit_113);
     std::vector<Value*> ptr_arraydecay4_indices;
     ptr_arraydecay4_indices.push_back(const_int32_29);
     ptr_arraydecay4_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay4 = new GetElementPtrInst(ptr_consts_125, ptr_arraydecay4_indices.begin(), ptr_arraydecay4_indices.end(), "arraydecay4", label_from_consts_exit_124);
-    StoreInst* void_154 = new StoreInst(ptr_arraydecay4, ptr_tmp3, false, label_from_consts_exit_124);
+    Instruction* ptr_arraydecay4 = new GetElementPtrInst(ptr_consts_114, ptr_arraydecay4_indices.begin(), ptr_arraydecay4_indices.end(), "arraydecay4", label_from_consts_exit_113);
+    StoreInst* void_144 = new StoreInst(ptr_arraydecay4, ptr_tmp3, false, label_from_consts_exit_113);
     std::vector<Value*> ptr_tmp5_indices;
     ptr_tmp5_indices.push_back(const_int32_29);
     ptr_tmp5_indices.push_back(const_int32_33);
-    Instruction* ptr_tmp5 = new GetElementPtrInst(ptr_args_127, ptr_tmp5_indices.begin(), ptr_tmp5_indices.end(), "tmp5", label_from_consts_exit_124);
+    Instruction* ptr_tmp5 = new GetElementPtrInst(ptr_args_116, ptr_tmp5_indices.begin(), ptr_tmp5_indices.end(), "tmp5", label_from_consts_exit_113);
     std::vector<Value*> ptr_arraydecay6_indices;
     ptr_arraydecay6_indices.push_back(const_int32_29);
     ptr_arraydecay6_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay6 = new GetElementPtrInst(ptr_temps_126, ptr_arraydecay6_indices.begin(), ptr_arraydecay6_indices.end(), "arraydecay6", label_from_consts_exit_124);
-    StoreInst* void_155 = new StoreInst(ptr_arraydecay6, ptr_tmp5, false, label_from_consts_exit_124);
-    std::vector<Value*> ptr_tmp8_156_indices;
-    ptr_tmp8_156_indices.push_back(const_int32_29);
-    ptr_tmp8_156_indices.push_back(const_int32_31);
-    Instruction* ptr_tmp8_156 = new GetElementPtrInst(ptr_args_127, ptr_tmp8_156_indices.begin(), ptr_tmp8_156_indices.end(), "tmp8", label_from_consts_exit_124);
-    std::vector<Value*> ptr_tmp12_157_indices;
-    ptr_tmp12_157_indices.push_back(const_int32_29);
-    ptr_tmp12_157_indices.push_back(const_int32_29);
-    Instruction* ptr_tmp12_157 = new GetElementPtrInst(ptr_args_127, ptr_tmp12_157_indices.begin(), ptr_tmp12_157_indices.end(), "tmp12", label_from_consts_exit_124);
-    std::vector<Value*> ptr_arraydecay11_158_indices;
-    ptr_arraydecay11_158_indices.push_back(const_int32_29);
-    ptr_arraydecay11_158_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay11_158 = new GetElementPtrInst(ptr_inputs_117, ptr_arraydecay11_158_indices.begin(), ptr_arraydecay11_158_indices.end(), "arraydecay11", label_from_consts_exit_124);
-    StoreInst* void_159 = new StoreInst(ptr_arraydecay11_158, ptr_tmp8_156, false, label_from_consts_exit_124);
-    std::vector<Value*> ptr_arraydecay16_160_indices;
-    ptr_arraydecay16_160_indices.push_back(const_int32_29);
-    ptr_arraydecay16_160_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay16_160 = new GetElementPtrInst(ptr_results_116, ptr_arraydecay16_160_indices.begin(), ptr_arraydecay16_160_indices.end(), "arraydecay16", label_from_consts_exit_124);
-    StoreInst* void_161 = new StoreInst(ptr_arraydecay16_160, ptr_tmp12_157, false, label_from_consts_exit_124);
-    StoreInst* void_162 = new StoreInst(const_int32_29, ptr_tmp, false, label_from_consts_exit_124);
-    CallInst* void_163 = new CallInst(func_execute_shader, ptr_args_127, "", label_from_consts_exit_124);
-    void_163->setCallingConv(CallingConv::C);
-    void_163->setTailCall(false);const ParamAttrsList *void_163_PAL = 0;
-    void_163->setParamAttrs(void_163_PAL);
-    
-    LoadInst* int32_tmp23 = new LoadInst(ptr_tmp, "tmp23", false, label_from_consts_exit_124);
+    Instruction* ptr_arraydecay6 = new GetElementPtrInst(ptr_temps_115, ptr_arraydecay6_indices.begin(), ptr_arraydecay6_indices.end(), "arraydecay6", label_from_consts_exit_113);
+    StoreInst* void_145 = new StoreInst(ptr_arraydecay6, ptr_tmp5, false, label_from_consts_exit_113);
+    std::vector<Value*> ptr_tmp8_indices;
+    ptr_tmp8_indices.push_back(const_int32_29);
+    ptr_tmp8_indices.push_back(const_int32_31);
+    Instruction* ptr_tmp8 = new GetElementPtrInst(ptr_args_116, ptr_tmp8_indices.begin(), ptr_tmp8_indices.end(), "tmp8", label_from_consts_exit_113);
+    std::vector<Value*> ptr_tmp12_indices;
+    ptr_tmp12_indices.push_back(const_int32_29);
+    ptr_tmp12_indices.push_back(const_int32_29);
+    Instruction* ptr_tmp12 = new GetElementPtrInst(ptr_args_116, ptr_tmp12_indices.begin(), ptr_tmp12_indices.end(), "tmp12", label_from_consts_exit_113);
+    std::vector<Value*> ptr_arraydecay11_146_indices;
+    ptr_arraydecay11_146_indices.push_back(const_int32_29);
+    ptr_arraydecay11_146_indices.push_back(const_int32_29);
+    Instruction* ptr_arraydecay11_146 = new GetElementPtrInst(ptr_inputs_106, ptr_arraydecay11_146_indices.begin(), ptr_arraydecay11_146_indices.end(), "arraydecay11", label_from_consts_exit_113);
+    StoreInst* void_147 = new StoreInst(ptr_arraydecay11_146, ptr_tmp8, false, label_from_consts_exit_113);
+    std::vector<Value*> ptr_arraydecay16_indices;
+    ptr_arraydecay16_indices.push_back(const_int32_29);
+    ptr_arraydecay16_indices.push_back(const_int32_29);
+    Instruction* ptr_arraydecay16 = new GetElementPtrInst(ptr_results_105, ptr_arraydecay16_indices.begin(), ptr_arraydecay16_indices.end(), "arraydecay16", label_from_consts_exit_113);
+    StoreInst* void_148 = new StoreInst(ptr_arraydecay16, ptr_tmp12, false, label_from_consts_exit_113);
+    StoreInst* void_149 = new StoreInst(const_int32_29, ptr_tmp_117, false, label_from_consts_exit_113);
+    CallInst* void_150 = new CallInst(func_execute_shader, ptr_args_116, "", label_from_consts_exit_113);
+    void_150->setCallingConv(CallingConv::C);
+    void_150->setTailCall(false);const ParamAttrsList *void_150_PAL = 0;
+    void_150->setParamAttrs(void_150_PAL);
+    
+    LoadInst* int32_tmp23 = new LoadInst(ptr_tmp_117, "tmp23", false, label_from_consts_exit_113);
     std::vector<Value*> ptr_arraydecay11_1_indices;
     ptr_arraydecay11_1_indices.push_back(const_int32_31);
     ptr_arraydecay11_1_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay11_1 = new GetElementPtrInst(ptr_inputs_117, ptr_arraydecay11_1_indices.begin(), ptr_arraydecay11_1_indices.end(), "arraydecay11.1", label_from_consts_exit_124);
-    StoreInst* void_164 = new StoreInst(ptr_arraydecay11_1, ptr_tmp8_156, false, label_from_consts_exit_124);
+    Instruction* ptr_arraydecay11_1 = new GetElementPtrInst(ptr_inputs_106, ptr_arraydecay11_1_indices.begin(), ptr_arraydecay11_1_indices.end(), "arraydecay11.1", label_from_consts_exit_113);
+    StoreInst* void_151 = new StoreInst(ptr_arraydecay11_1, ptr_tmp8, false, label_from_consts_exit_113);
     std::vector<Value*> ptr_arraydecay16_1_indices;
     ptr_arraydecay16_1_indices.push_back(const_int32_31);
     ptr_arraydecay16_1_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay16_1 = new GetElementPtrInst(ptr_results_116, ptr_arraydecay16_1_indices.begin(), ptr_arraydecay16_1_indices.end(), "arraydecay16.1", label_from_consts_exit_124);
-    StoreInst* void_165 = new StoreInst(ptr_arraydecay16_1, ptr_tmp12_157, false, label_from_consts_exit_124);
-    StoreInst* void_166 = new StoreInst(const_int32_29, ptr_tmp, false, label_from_consts_exit_124);
-    CallInst* void_167 = new CallInst(func_execute_shader, ptr_args_127, "", label_from_consts_exit_124);
-    void_167->setCallingConv(CallingConv::C);
-    void_167->setTailCall(false);const ParamAttrsList *void_167_PAL = 0;
-    void_167->setParamAttrs(void_167_PAL);
-    
-    LoadInst* int32_tmp23_1 = new LoadInst(ptr_tmp, "tmp23.1", false, label_from_consts_exit_124);
-    BinaryOperator* int32_shl_1 = BinaryOperator::create(Instruction::Shl, int32_tmp23_1, const_int32_31, "shl.1", label_from_consts_exit_124);
-    BinaryOperator* int32_or_1 = BinaryOperator::create(Instruction::Or, int32_shl_1, int32_tmp23, "or.1", label_from_consts_exit_124);
+    Instruction* ptr_arraydecay16_1 = new GetElementPtrInst(ptr_results_105, ptr_arraydecay16_1_indices.begin(), ptr_arraydecay16_1_indices.end(), "arraydecay16.1", label_from_consts_exit_113);
+    StoreInst* void_152 = new StoreInst(ptr_arraydecay16_1, ptr_tmp12, false, label_from_consts_exit_113);
+    StoreInst* void_153 = new StoreInst(const_int32_29, ptr_tmp_117, false, label_from_consts_exit_113);
+    CallInst* void_154 = new CallInst(func_execute_shader, ptr_args_116, "", label_from_consts_exit_113);
+    void_154->setCallingConv(CallingConv::C);
+    void_154->setTailCall(false);const ParamAttrsList *void_154_PAL = 0;
+    void_154->setParamAttrs(void_154_PAL);
+    
+    LoadInst* int32_tmp23_1 = new LoadInst(ptr_tmp_117, "tmp23.1", false, label_from_consts_exit_113);
+    BinaryOperator* int32_shl_1 = BinaryOperator::create(Instruction::Shl, int32_tmp23_1, const_int32_31, "shl.1", label_from_consts_exit_113);
+    BinaryOperator* int32_or_1 = BinaryOperator::create(Instruction::Or, int32_shl_1, int32_tmp23, "or.1", label_from_consts_exit_113);
     std::vector<Value*> ptr_arraydecay11_2_indices;
     ptr_arraydecay11_2_indices.push_back(const_int32_33);
     ptr_arraydecay11_2_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay11_2 = new GetElementPtrInst(ptr_inputs_117, ptr_arraydecay11_2_indices.begin(), ptr_arraydecay11_2_indices.end(), "arraydecay11.2", label_from_consts_exit_124);
-    StoreInst* void_168 = new StoreInst(ptr_arraydecay11_2, ptr_tmp8_156, false, label_from_consts_exit_124);
+    Instruction* ptr_arraydecay11_2 = new GetElementPtrInst(ptr_inputs_106, ptr_arraydecay11_2_indices.begin(), ptr_arraydecay11_2_indices.end(), "arraydecay11.2", label_from_consts_exit_113);
+    StoreInst* void_155 = new StoreInst(ptr_arraydecay11_2, ptr_tmp8, false, label_from_consts_exit_113);
     std::vector<Value*> ptr_arraydecay16_2_indices;
     ptr_arraydecay16_2_indices.push_back(const_int32_33);
     ptr_arraydecay16_2_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay16_2 = new GetElementPtrInst(ptr_results_116, ptr_arraydecay16_2_indices.begin(), ptr_arraydecay16_2_indices.end(), "arraydecay16.2", label_from_consts_exit_124);
-    StoreInst* void_169 = new StoreInst(ptr_arraydecay16_2, ptr_tmp12_157, false, label_from_consts_exit_124);
-    StoreInst* void_170 = new StoreInst(const_int32_29, ptr_tmp, false, label_from_consts_exit_124);
-    CallInst* void_171 = new CallInst(func_execute_shader, ptr_args_127, "", label_from_consts_exit_124);
-    void_171->setCallingConv(CallingConv::C);
-    void_171->setTailCall(false);const ParamAttrsList *void_171_PAL = 0;
-    void_171->setParamAttrs(void_171_PAL);
-    
-    LoadInst* int32_tmp23_2 = new LoadInst(ptr_tmp, "tmp23.2", false, label_from_consts_exit_124);
-    BinaryOperator* int32_shl_2 = BinaryOperator::create(Instruction::Shl, int32_tmp23_2, const_int32_33, "shl.2", label_from_consts_exit_124);
-    BinaryOperator* int32_or_2 = BinaryOperator::create(Instruction::Or, int32_shl_2, int32_or_1, "or.2", label_from_consts_exit_124);
+    Instruction* ptr_arraydecay16_2 = new GetElementPtrInst(ptr_results_105, ptr_arraydecay16_2_indices.begin(), ptr_arraydecay16_2_indices.end(), "arraydecay16.2", label_from_consts_exit_113);
+    StoreInst* void_156 = new StoreInst(ptr_arraydecay16_2, ptr_tmp12, false, label_from_consts_exit_113);
+    StoreInst* void_157 = new StoreInst(const_int32_29, ptr_tmp_117, false, label_from_consts_exit_113);
+    CallInst* void_158 = new CallInst(func_execute_shader, ptr_args_116, "", label_from_consts_exit_113);
+    void_158->setCallingConv(CallingConv::C);
+    void_158->setTailCall(false);const ParamAttrsList *void_158_PAL = 0;
+    void_158->setParamAttrs(void_158_PAL);
+    
+    LoadInst* int32_tmp23_2 = new LoadInst(ptr_tmp_117, "tmp23.2", false, label_from_consts_exit_113);
+    BinaryOperator* int32_shl_2 = BinaryOperator::create(Instruction::Shl, int32_tmp23_2, const_int32_33, "shl.2", label_from_consts_exit_113);
+    BinaryOperator* int32_or_2 = BinaryOperator::create(Instruction::Or, int32_shl_2, int32_or_1, "or.2", label_from_consts_exit_113);
     std::vector<Value*> ptr_arraydecay11_3_indices;
     ptr_arraydecay11_3_indices.push_back(const_int32_34);
     ptr_arraydecay11_3_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay11_3 = new GetElementPtrInst(ptr_inputs_117, ptr_arraydecay11_3_indices.begin(), ptr_arraydecay11_3_indices.end(), "arraydecay11.3", label_from_consts_exit_124);
-    StoreInst* void_172 = new StoreInst(ptr_arraydecay11_3, ptr_tmp8_156, false, label_from_consts_exit_124);
+    Instruction* ptr_arraydecay11_3 = new GetElementPtrInst(ptr_inputs_106, ptr_arraydecay11_3_indices.begin(), ptr_arraydecay11_3_indices.end(), "arraydecay11.3", label_from_consts_exit_113);
+    StoreInst* void_159 = new StoreInst(ptr_arraydecay11_3, ptr_tmp8, false, label_from_consts_exit_113);
     std::vector<Value*> ptr_arraydecay16_3_indices;
     ptr_arraydecay16_3_indices.push_back(const_int32_34);
     ptr_arraydecay16_3_indices.push_back(const_int32_29);
-    Instruction* ptr_arraydecay16_3 = new GetElementPtrInst(ptr_results_116, ptr_arraydecay16_3_indices.begin(), ptr_arraydecay16_3_indices.end(), "arraydecay16.3", label_from_consts_exit_124);
-    StoreInst* void_173 = new StoreInst(ptr_arraydecay16_3, ptr_tmp12_157, false, label_from_consts_exit_124);
-    StoreInst* void_174 = new StoreInst(const_int32_29, ptr_tmp, false, label_from_consts_exit_124);
-    CallInst* void_175 = new CallInst(func_execute_shader, ptr_args_127, "", label_from_consts_exit_124);
-    void_175->setCallingConv(CallingConv::C);
-    void_175->setTailCall(false);const ParamAttrsList *void_175_PAL = 0;
-    void_175->setParamAttrs(void_175_PAL);
-    
-    LoadInst* int32_tmp23_3 = new LoadInst(ptr_tmp, "tmp23.3", false, label_from_consts_exit_124);
-    BinaryOperator* int32_shl_3 = BinaryOperator::create(Instruction::Shl, int32_tmp23_3, const_int32_34, "shl.3", label_from_consts_exit_124);
-    BinaryOperator* int32_or_3 = BinaryOperator::create(Instruction::Or, int32_shl_3, int32_or_2, "or.3", label_from_consts_exit_124);
-    BinaryOperator* int32_neg = BinaryOperator::create(Instruction::Xor, int32_or_3, const_int32_30, "neg", label_from_consts_exit_124);
-    new ReturnInst(int32_neg, label_from_consts_exit_124);
+    Instruction* ptr_arraydecay16_3 = new GetElementPtrInst(ptr_results_105, ptr_arraydecay16_3_indices.begin(), ptr_arraydecay16_3_indices.end(), "arraydecay16.3", label_from_consts_exit_113);
+    StoreInst* void_160 = new StoreInst(ptr_arraydecay16_3, ptr_tmp12, false, label_from_consts_exit_113);
+    StoreInst* void_161 = new StoreInst(const_int32_29, ptr_tmp_117, false, label_from_consts_exit_113);
+    CallInst* void_162 = new CallInst(func_execute_shader, ptr_args_116, "", label_from_consts_exit_113);
+    void_162->setCallingConv(CallingConv::C);
+    void_162->setTailCall(false);const ParamAttrsList *void_162_PAL = 0;
+    void_162->setParamAttrs(void_162_PAL);
+    
+    LoadInst* int32_tmp23_3 = new LoadInst(ptr_tmp_117, "tmp23.3", false, label_from_consts_exit_113);
+    BinaryOperator* int32_shl_3 = BinaryOperator::create(Instruction::Shl, int32_tmp23_3, const_int32_34, "shl.3", label_from_consts_exit_113);
+    BinaryOperator* int32_or_3 = BinaryOperator::create(Instruction::Or, int32_shl_3, int32_or_2, "or.3", label_from_consts_exit_113);
+    BinaryOperator* int32_neg = BinaryOperator::create(Instruction::Xor, int32_or_3, const_int32_30, "neg", label_from_consts_exit_113);
+    new ReturnInst(int32_neg, label_from_consts_exit_113);
     
     // Resolve Forward References
-    fwdref_138->replaceAllUsesWith(packed_tmp31_i_150); delete fwdref_138;
-    fwdref_136->replaceAllUsesWith(int32_indvar_next7); delete fwdref_136;
+    fwdref_128->replaceAllUsesWith(packed_tmp31_i_140); delete fwdref_128;
+    fwdref_126->replaceAllUsesWith(int32_indvar_next7); delete fwdref_126;
     
   }
   
diff --git a/src/mesa/pipe/llvm/llvm_entry.c b/src/mesa/pipe/llvm/llvm_entry.c
index c3b34584e1..fa50b60e66 100644
--- a/src/mesa/pipe/llvm/llvm_entry.c
+++ b/src/mesa/pipe/llvm/llvm_entry.c
@@ -86,8 +86,8 @@ struct ShaderInput
 
 extern void execute_shader(struct ShaderInput *input);
 
-void run_vertex_shader(float4 (*inputs)[16],
-                       float4 (*results)[16],
+void run_vertex_shader(void *inputs,
+                       void *results,
                        float (*aconsts)[4],
                        int num_vertices,
                        int num_inputs,
@@ -98,16 +98,16 @@ void run_vertex_shader(float4 (*inputs)[16],
    float4  temps[128];//MAX_PROGRAM_TEMPS
 
    struct ShaderInput args;
+   args.dests  = results;
+   args.inputs = inputs;
+
    /*printf("XXX LLVM run_vertex_shader vertices = %d, inputs = %d, attribs = %d, consts = %d\n",
      num_vertices, num_inputs, num_attribs, num_consts);*/
    from_consts(consts, aconsts, num_consts);
    args.consts = consts;
    args.temps = temps;
-   for (int i = 0; i < num_vertices; ++i) {
-      args.dests  = results[i];
-      args.inputs = inputs[i];
-      execute_shader(&args);
-   }
+
+   execute_shader(&args);
 }
 
 
diff --git a/src/mesa/pipe/llvm/tgsitollvm.cpp b/src/mesa/pipe/llvm/tgsitollvm.cpp
new file mode 100644
index 0000000000..eb9e1196f1
--- /dev/null
+++ b/src/mesa/pipe/llvm/tgsitollvm.cpp
@@ -0,0 +1,682 @@
+#include "tgsitollvm.h"
+
+#include "gallivm.h"
+#include "gallivm_p.h"
+
+#include "storage.h"
+#include "instructions.h"
+
+#include "pipe/p_shader_tokens.h"
+
+#include "pipe/tgsi/util/tgsi_parse.h"
+#include "pipe/tgsi/exec/tgsi_exec.h"
+#include "pipe/tgsi/util/tgsi_util.h"
+#include "pipe/tgsi/util/tgsi_build.h"
+#include "pipe/tgsi/util/tgsi_dump.h"
+
+
+#include <llvm/Module.h>
+#include <llvm/CallingConv.h>
+#include <llvm/Constants.h>
+#include <llvm/DerivedTypes.h>
+#include <llvm/Instructions.h>
+#include <llvm/ModuleProvider.h>
+#include <llvm/Pass.h>
+#include <llvm/PassManager.h>
+#include <llvm/ParameterAttributes.h>
+#include <llvm/Support/PatternMatch.h>
+#include <llvm/ExecutionEngine/JIT.h>
+#include <llvm/ExecutionEngine/Interpreter.h>
+#include <llvm/ExecutionEngine/GenericValue.h>
+#include <llvm/Support/MemoryBuffer.h>
+#include <llvm/LinkAllPasses.h>
+#include <llvm/Analysis/Verifier.h>
+#include <llvm/Analysis/LoopPass.h>
+#include <llvm/Target/TargetData.h>
+#include <llvm/Bitcode/ReaderWriter.h>
+#include <llvm/Transforms/Utils/Cloning.h>
+
+
+#include <sstream>
+#include <fstream>
+#include <iostream>
+
+using namespace llvm;
+#include "llvm_base_shader.cpp"
+
+static inline void
+add_interpolator(struct gallivm_ir *ir,
+                 struct gallivm_interpolate *interp)
+{
+   ir->interpolators[ir->num_interp] = *interp;
+   ++ir->num_interp;
+}
+
+static void
+translate_declaration(struct gallivm_ir *prog,
+                      llvm::Module *module,
+                      Storage *storage,
+                      struct tgsi_full_declaration *decl,
+                      struct tgsi_full_declaration *fd)
+{
+   if (decl->Declaration.File == TGSI_FILE_INPUT) {
+      unsigned first, last, mask;
+      uint interp_method;
+
+      assert(decl->Declaration.Declare == TGSI_DECLARE_RANGE);
+
+      first = decl->u.DeclarationRange.First;
+      last = decl->u.DeclarationRange.Last;
+      mask = decl->Declaration.UsageMask;
+
+      /* Do not touch WPOS.xy */
+      if (first == 0) {
+         mask &= ~TGSI_WRITEMASK_XY;
+         if (mask == TGSI_WRITEMASK_NONE) {
+            first++;
+            if (first > last) {
+               return;
+            }
+         }
+      }
+
+      interp_method = decl->Interpolation.Interpolate;
+
+      if (mask == TGSI_WRITEMASK_XYZW) {
+         unsigned i, j;
+
+         for (i = first; i <= last; i++) {
+            for (j = 0; j < NUM_CHANNELS; j++) {
+               //interp( mach, i, j );
+               struct gallivm_interpolate interp;
+               interp.type = interp_method;
+               interp.attrib = i;
+               interp.chan = j;
+               add_interpolator(prog, &interp);
+            }
+         }
+      } else {
+         unsigned i, j;
+         for( j = 0; j < NUM_CHANNELS; j++ ) {
+            if( mask & (1 << j) ) {
+               for( i = first; i <= last; i++ ) {
+                  struct gallivm_interpolate interp;
+                  interp.type = interp_method;
+                  interp.attrib = i;
+                  interp.chan = j;
+                  add_interpolator(prog, &interp);
+               }
+            }
+         }
+      }
+   }
+}
+
+
+static void
+translate_immediate(Storage *storage,
+                    struct tgsi_full_immediate *imm)
+{
+   float vec[4];
+   int i;
+   for (i = 0; i < imm->Immediate.Size - 1; ++i) {
+      switch( imm->Immediate.DataType ) {
+      case TGSI_IMM_FLOAT32:
+         vec[i] = imm->u.ImmediateFloat32[i].Float;
+         break;
+      default:
+         assert( 0 );
+      }
+   }
+   storage->addImmediate(vec);
+}
+
+static inline llvm::Value *
+swizzleVector(llvm::Value *val, struct tgsi_full_src_register *src,
+              Storage *storage)
+{
+   int swizzle = 0;
+   int start = 1000;
+   const int NO_SWIZZLE = TGSI_SWIZZLE_X * 1000 + TGSI_SWIZZLE_Y * 100 +
+                          TGSI_SWIZZLE_Z * 10 + TGSI_SWIZZLE_W;
+   for (int k = 0; k < 4; ++k) {
+      swizzle += tgsi_util_get_full_src_register_extswizzle(src, k) * start;
+      start /= 10;
+   }
+   if (swizzle != NO_SWIZZLE) {
+      /*fprintf(stderr, "XXXXXXXX swizzle = %d\n", swizzle);*/
+      val = storage->shuffleVector(val, swizzle);
+   }
+   return val;
+}
+
+static void
+translate_instruction(llvm::Module *module,
+                      Storage *storage,
+                      Instructions *instr,
+                      struct tgsi_full_instruction *inst,
+                      struct tgsi_full_instruction *fi,
+                      unsigned instno)
+{
+   llvm::Value *inputs[4];
+   inputs[0] = 0;
+   inputs[1] = 0;
+   inputs[2] = 0;
+   inputs[3] = 0;
+
+   for (int i = 0; i < inst->Instruction.NumSrcRegs; ++i) {
+      struct tgsi_full_src_register *src = &inst->FullSrcRegisters[i];
+      llvm::Value *val = 0;
+      llvm::Value *indIdx = 0;
+
+      if (src->SrcRegister.Indirect) {
+         indIdx = storage->addrElement(src->SrcRegisterInd.Index);
+         indIdx = storage->extractIndex(indIdx);
+      }
+      if (src->SrcRegister.File == TGSI_FILE_CONSTANT) {
+         val = storage->constElement(src->SrcRegister.Index, indIdx);
+      } else if (src->SrcRegister.File == TGSI_FILE_INPUT) {
+         val = storage->inputElement(src->SrcRegister.Index, indIdx);
+      } else if (src->SrcRegister.File == TGSI_FILE_TEMPORARY) {
+         val = storage->tempElement(src->SrcRegister.Index);
+      } else if (src->SrcRegister.File == TGSI_FILE_OUTPUT) {
+         val = storage->outputElement(src->SrcRegister.Index, indIdx);
+      } else if (src->SrcRegister.File == TGSI_FILE_IMMEDIATE) {
+         val = storage->immediateElement(src->SrcRegister.Index);
+      } else {
+         fprintf(stderr, "ERROR: not supported llvm source %d\n", src->SrcRegister.File);
+         return;
+      }
+
+      inputs[i] = swizzleVector(val, src, storage);
+   }
+
+   /*if (inputs[0])
+     instr->printVector(inputs[0]);
+     if (inputs[1])
+     instr->printVector(inputs[1]);*/
+   llvm::Value *out = 0;
+   switch (inst->Instruction.Opcode) {
+   case TGSI_OPCODE_ARL: {
+      out = instr->arl(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_MOV: {
+      out = inputs[0];
+   }
+      break;
+   case TGSI_OPCODE_LIT: {
+      out = instr->lit(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_RCP: {
+      out = instr->rcp(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_RSQ: {
+      out = instr->rsq(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_EXP:
+      break;
+   case TGSI_OPCODE_LOG:
+      break;
+   case TGSI_OPCODE_MUL: {
+      out = instr->mul(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_ADD: {
+      out = instr->add(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_DP3: {
+      out = instr->dp3(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_DP4: {
+      out = instr->dp4(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_DST: {
+      out = instr->dst(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_MIN: {
+      out = instr->min(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_MAX: {
+      out = instr->max(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_SLT: {
+      out = instr->slt(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_SGE: {
+      out = instr->sge(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_MAD: {
+      out = instr->madd(inputs[0], inputs[1], inputs[2]);
+   }
+      break;
+   case TGSI_OPCODE_SUB: {
+      out = instr->sub(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_LERP: {
+      out = instr->lerp(inputs[0], inputs[1], inputs[2]);
+   }
+      break;
+   case TGSI_OPCODE_CND:
+      break;
+   case TGSI_OPCODE_CND0:
+      break;
+   case TGSI_OPCODE_DOT2ADD:
+      break;
+   case TGSI_OPCODE_INDEX:
+      break;
+   case TGSI_OPCODE_NEGATE:
+      break;
+   case TGSI_OPCODE_FRAC: {
+      out = instr->frc(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_CLAMP:
+      break;
+   case TGSI_OPCODE_FLOOR: {
+      out = instr->floor(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_ROUND:
+      break;
+   case TGSI_OPCODE_EXPBASE2: {
+      out = instr->ex2(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_LOGBASE2: {
+      out = instr->lg2(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_POWER: {
+      out = instr->pow(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_CROSSPRODUCT: {
+      out = instr->cross(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_MULTIPLYMATRIX:
+      break;
+   case TGSI_OPCODE_ABS: {
+      out = instr->abs(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_RCC:
+      break;
+   case TGSI_OPCODE_DPH: {
+      out = instr->dph(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_COS: {
+      out = instr->cos(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_DDX:
+      break;
+   case TGSI_OPCODE_DDY:
+      break;
+   case TGSI_OPCODE_KILP: {
+      out = instr->kilp(inputs[0]);
+      storage->setKilElement(out);
+      return;
+   }
+      break;
+   case TGSI_OPCODE_PK2H:
+      break;
+   case TGSI_OPCODE_PK2US:
+      break;
+   case TGSI_OPCODE_PK4B:
+      break;
+   case TGSI_OPCODE_PK4UB:
+      break;
+   case TGSI_OPCODE_RFL:
+      break;
+   case TGSI_OPCODE_SEQ:
+      break;
+   case TGSI_OPCODE_SFL:
+      break;
+   case TGSI_OPCODE_SGT: {
+      out = instr->sgt(inputs[0], inputs[1]);
+   }
+      break;
+   case TGSI_OPCODE_SIN: {
+      out = instr->sin(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_SLE:
+      break;
+   case TGSI_OPCODE_SNE:
+      break;
+   case TGSI_OPCODE_STR:
+      break;
+   case TGSI_OPCODE_TEX:
+      break;
+   case TGSI_OPCODE_TXD:
+      break;
+   case TGSI_OPCODE_UP2H:
+      break;
+   case TGSI_OPCODE_UP2US:
+      break;
+   case TGSI_OPCODE_UP4B:
+      break;
+   case TGSI_OPCODE_UP4UB:
+      break;
+   case TGSI_OPCODE_X2D:
+      break;
+   case TGSI_OPCODE_ARA:
+      break;
+   case TGSI_OPCODE_ARR:
+      break;
+   case TGSI_OPCODE_BRA:
+      break;
+   case TGSI_OPCODE_CAL: {
+      instr->cal(inst->InstructionExtLabel.Label, storage->inputPtr());
+      return;
+   }
+      break;
+   case TGSI_OPCODE_RET: {
+      instr->end();
+      return;
+   }
+      break;
+   case TGSI_OPCODE_SSG:
+      break;
+   case TGSI_OPCODE_CMP: {
+      out = instr->cmp(inputs[0], inputs[1], inputs[2]);
+   }
+      break;
+   case TGSI_OPCODE_SCS: {
+      out = instr->scs(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_TXB:
+      break;
+   case TGSI_OPCODE_NRM:
+      break;
+   case TGSI_OPCODE_DIV:
+      break;
+   case TGSI_OPCODE_DP2:
+      break;
+   case TGSI_OPCODE_TXL:
+      break;
+   case TGSI_OPCODE_BRK: {
+      instr->brk();
+      return;
+   }
+      break;
+   case TGSI_OPCODE_IF: {
+      instr->ifop(inputs[0]);
+      storage->setCurrentBlock(instr->currentBlock());
+      return;  //just update the state
+   }
+      break;
+   case TGSI_OPCODE_LOOP:
+      break;
+   case TGSI_OPCODE_REP:
+      break;
+   case TGSI_OPCODE_ELSE: {
+      instr->elseop();
+      storage->setCurrentBlock(instr->currentBlock());
+      return; //only state update
+   }
+      break;
+   case TGSI_OPCODE_ENDIF: {
+      instr->endif();
+      storage->setCurrentBlock(instr->currentBlock());
+      return; //just update the state
+   }
+      break;
+   case TGSI_OPCODE_ENDLOOP:
+      break;
+   case TGSI_OPCODE_ENDREP:
+      break;
+   case TGSI_OPCODE_PUSHA:
+      break;
+   case TGSI_OPCODE_POPA:
+      break;
+   case TGSI_OPCODE_CEIL:
+      break;
+   case TGSI_OPCODE_I2F:
+      break;
+   case TGSI_OPCODE_NOT:
+      break;
+   case TGSI_OPCODE_TRUNC: {
+      out = instr->trunc(inputs[0]);
+   }
+      break;
+   case TGSI_OPCODE_SHL:
+      break;
+   case TGSI_OPCODE_SHR:
+      break;
+   case TGSI_OPCODE_AND:
+      break;
+   case TGSI_OPCODE_OR:
+      break;
+   case TGSI_OPCODE_MOD:
+      break;
+   case TGSI_OPCODE_XOR:
+      break;
+   case TGSI_OPCODE_SAD:
+      break;
+   case TGSI_OPCODE_TXF:
+      break;
+   case TGSI_OPCODE_TXQ:
+      break;
+   case TGSI_OPCODE_CONT:
+      break;
+   case TGSI_OPCODE_EMIT:
+      break;
+   case TGSI_OPCODE_ENDPRIM:
+      break;
+   case TGSI_OPCODE_BGNLOOP2: {
+      instr->beginLoop();
+      storage->setCurrentBlock(instr->currentBlock());
+      return;
+   }
+      break;
+   case TGSI_OPCODE_BGNSUB: {
+      instr->bgnSub(instno);
+      storage->setCurrentBlock(instr->currentBlock());
+      storage->pushTemps();
+      return;
+   }
+      break;
+   case TGSI_OPCODE_ENDLOOP2: {
+      instr->endLoop();
+      storage->setCurrentBlock(instr->currentBlock());
+      return;
+   }
+      break;
+   case TGSI_OPCODE_ENDSUB: {
+      instr->endSub();
+      storage->setCurrentBlock(instr->currentBlock());
+      storage->popArguments();
+      storage->popTemps();
+      return;
+   }
+      break;
+   case TGSI_OPCODE_NOISE1:
+      break;
+   case TGSI_OPCODE_NOISE2:
+      break;
+   case TGSI_OPCODE_NOISE3:
+      break;
+   case TGSI_OPCODE_NOISE4:
+      break;
+   case TGSI_OPCODE_NOP:
+      break;
+   case TGSI_OPCODE_TEXBEM:
+      break;
+   case TGSI_OPCODE_TEXBEML:
+      break;
+   case TGSI_OPCODE_TEXREG2AR:
+      break;
+   case TGSI_OPCODE_TEXM3X2PAD:
+      break;
+   case TGSI_OPCODE_TEXM3X2TEX:
+      break;
+   case TGSI_OPCODE_TEXM3X3PAD:
+      break;
+   case TGSI_OPCODE_TEXM3X3TEX:
+      break;
+   case TGSI_OPCODE_TEXM3X3SPEC:
+      break;
+   case TGSI_OPCODE_TEXM3X3VSPEC:
+      break;
+   case TGSI_OPCODE_TEXREG2GB:
+      break;
+   case TGSI_OPCODE_TEXREG2RGB:
+      break;
+   case TGSI_OPCODE_TEXDP3TEX:
+      break;
+   case TGSI_OPCODE_TEXDP3:
+      break;
+   case TGSI_OPCODE_TEXM3X3:
+      break;
+   case TGSI_OPCODE_TEXM3X2DEPTH:
+      break;
+   case TGSI_OPCODE_TEXDEPTH:
+      break;
+   case TGSI_OPCODE_BEM:
+      break;
+   case TGSI_OPCODE_M4X3:
+      break;
+   case TGSI_OPCODE_M3X4:
+      break;
+   case TGSI_OPCODE_M3X3:
+      break;
+   case TGSI_OPCODE_M3X2:
+      break;
+   case TGSI_OPCODE_NRM4:
+      break;
+   case TGSI_OPCODE_CALLNZ:
+      break;
+   case TGSI_OPCODE_IFC:
+      break;
+   case TGSI_OPCODE_BREAKC:
+      break;
+   case TGSI_OPCODE_KIL:
+      break;
+   case TGSI_OPCODE_END:
+      instr->end();
+      return;
+      break;
+   default:
+      fprintf(stderr, "ERROR: Unknown opcode %d\n",
+              inst->Instruction.Opcode);
+      assert(0);
+      break;
+   }
+
+   if (!out) {
+      fprintf(stderr, "ERROR: unsupported opcode %d\n",
+              inst->Instruction.Opcode);
+      assert(!"Unsupported opcode");
+   }
+
+   /* # not sure if we need this */
+   switch( inst->Instruction.Saturate ) {
+   case TGSI_SAT_NONE:
+      break;
+   case TGSI_SAT_ZERO_ONE:
+      /*TXT( "_SAT" );*/
+      break;
+   case TGSI_SAT_MINUS_PLUS_ONE:
+      /*TXT( "_SAT[-1,1]" );*/
+      break;
+   default:
+      assert( 0 );
+   }
+
+   /* store results  */
+   for (int i = 0; i < inst->Instruction.NumDstRegs; ++i) {
+      struct tgsi_full_dst_register *dst = &inst->FullDstRegisters[i];
+
+      if (dst->DstRegister.File == TGSI_FILE_OUTPUT) {
+         storage->setOutputElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
+      } else if (dst->DstRegister.File == TGSI_FILE_TEMPORARY) {
+         storage->setTempElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
+      } else if (dst->DstRegister.File == TGSI_FILE_ADDRESS) {
+         storage->setAddrElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
+      } else {
+         fprintf(stderr, "ERROR: unsupported LLVM destination!");
+         assert(!"wrong destination");
+      }
+   }
+}
+
+llvm::Module *
+tgsi_to_llvm(struct gallivm_ir *ir, const struct tgsi_token *tokens)
+{
+   llvm::Module *mod = createBaseShader();
+   struct tgsi_parse_context parse;
+   struct tgsi_full_instruction fi;
+   struct tgsi_full_declaration fd;
+   unsigned instno = 0;
+   Function* shader = mod->getFunction("execute_shader");
+   std::ostringstream stream;
+   if (ir->type == GALLIVM_VS) {
+      stream << "vs_shader";
+   } else {
+      stream << "fs_shader";
+   }
+   stream << ir->id;
+   std::string func_name = stream.str();
+   shader->setName(func_name.c_str());
+
+   Function::arg_iterator args = shader->arg_begin();
+   Value *ptr_INPUT = args++;
+   ptr_INPUT->setName("input");
+
+   BasicBlock *label_entry = new BasicBlock("entry", shader, 0);
+
+   tgsi_parse_init(&parse, tokens);
+
+   fi = tgsi_default_full_instruction();
+   fd = tgsi_default_full_declaration();
+   Storage storage(label_entry, ptr_INPUT);
+   Instructions instr(mod, shader, label_entry, &storage);
+   while(!tgsi_parse_end_of_tokens(&parse)) {
+      tgsi_parse_token(&parse);
+
+      switch (parse.FullToken.Token.Type) {
+      case TGSI_TOKEN_TYPE_DECLARATION:
+         translate_declaration(ir, mod, &storage,
+                               &parse.FullToken.FullDeclaration,
+                               &fd);
+         break;
+
+      case TGSI_TOKEN_TYPE_IMMEDIATE:
+         translate_immediate(&storage,
+                             &parse.FullToken.FullImmediate);
+         break;
+
+      case TGSI_TOKEN_TYPE_INSTRUCTION:
+         translate_instruction(mod, &storage, &instr,
+                               &parse.FullToken.FullInstruction,
+                               &fi, instno);
+         ++instno;
+         break;
+
+      default:
+         assert(0);
+      }
+   }
+
+   tgsi_parse_free(&parse);
+
+   ir->num_consts = storage.numConsts();
+   return mod;
+}
diff --git a/src/mesa/pipe/llvm/tgsitollvm.h b/src/mesa/pipe/llvm/tgsitollvm.h
new file mode 100644
index 0000000000..073ffb5749
--- /dev/null
+++ b/src/mesa/pipe/llvm/tgsitollvm.h
@@ -0,0 +1,16 @@
+#ifndef TGSITOLLVM_H
+#define TGSITOLLVM_H
+
+
+namespace llvm {
+   class Module;
+}
+
+struct gallivm_ir;
+struct tgsi_token;
+
+
+llvm::Module * tgsi_to_llvm(struct gallivm_ir *ir,
+                            const struct tgsi_token *tokens);
+
+#endif
diff --git a/src/mesa/pipe/softpipe/sp_state_fs.c b/src/mesa/pipe/softpipe/sp_state_fs.c
index 1430be7869..598a70c827 100644
--- a/src/mesa/pipe/softpipe/sp_state_fs.c
+++ b/src/mesa/pipe/softpipe/sp_state_fs.c
@@ -61,7 +61,7 @@ softpipe_create_fs_state(struct pipe_context *pipe,
    }
 
 #ifdef MESA_LLVM
-   state->llvm_prog = gallivm_from_tgsi(state->shader.tokens, GALLIVM_FS);
+   state->llvm_prog = 0;
    if (!gallivm_global_cpu_engine()) {
       gallivm_cpu_engine_create(state->llvm_prog);
    }