gallium/draw: initial code to properly support llvm in the draw module

code generate big chunks of the vertex pipeline in order to speed up
software vertex processing.

12 files changed, 1601 insertions, 18 deletions

diff --git a/src/gallium/auxiliary/SConscript b/src/gallium/auxiliary/SConscript
index b531ad2dbd..51c4a0cbbe 100644
--- a/src/gallium/auxiliary/SConscript
+++ b/src/gallium/auxiliary/SConscript
@@ -194,6 +194,8 @@ if drawllvm:
     'gallivm/lp_bld_swizzle.c',
     'gallivm/lp_bld_tgsi_soa.c',
     'gallivm/lp_bld_type.c',
+    'draw/draw_llvm.c',
+    'draw/draw_pt_fetch_shade_pipeline_llvm.c'
     ]
 
 gallium = env.ConvenienceLibrary(
diff --git a/src/gallium/auxiliary/draw/draw_context.c b/src/gallium/auxiliary/draw/draw_context.c
index d5ddc4a6a9..6fa73ad56b 100644
--- a/src/gallium/auxiliary/draw/draw_context.c
+++ b/src/gallium/auxiliary/draw/draw_context.c
@@ -44,6 +44,18 @@ struct draw_context *draw_create( void )
    if (draw == NULL)
       goto fail;
 
+   if (!draw_init(draw))
+      goto fail;
+
+   return draw;
+
+fail:
+   draw_destroy( draw );
+   return NULL;
+}
+
+boolean draw_init(struct draw_context *draw)
+{
    ASSIGN_4V( draw->plane[0], -1,  0,  0, 1 );
    ASSIGN_4V( draw->plane[1],  1,  0,  0, 1 );
    ASSIGN_4V( draw->plane[2],  0, -1,  0, 1 );
@@ -57,22 +69,18 @@ struct draw_context *draw_create( void )
 
 
    if (!draw_pipeline_init( draw ))
-      goto fail;
+      return FALSE;
 
    if (!draw_pt_init( draw ))
-      goto fail;
+      return FALSE;
 
    if (!draw_vs_init( draw ))
-      goto fail;
+      return FALSE;
 
    if (!draw_gs_init( draw ))
-      goto fail;
+      return FALSE;
 
-   return draw;
-
-fail:
-   draw_destroy( draw );   
-   return NULL;
+   return TRUE;
 }
 
 
diff --git a/src/gallium/auxiliary/draw/draw_context.h b/src/gallium/auxiliary/draw/draw_context.h
index acd81b9712..d42e400318 100644
--- a/src/gallium/auxiliary/draw/draw_context.h
+++ b/src/gallium/auxiliary/draw/draw_context.h
@@ -40,6 +40,9 @@
 
 #include "pipe/p_state.h"
 
+#ifdef DRAW_LLVM
+#include <llvm-c/ExecutionEngine.h>
+#endif
 
 struct pipe_context;
 struct draw_context;
@@ -197,6 +200,11 @@ boolean draw_need_pipeline(const struct draw_context *draw,
                            const struct pipe_rasterizer_state *rasterizer,
                            unsigned prim );
 
-
+#ifdef DRAW_LLVM
+/*******************************************************************************
+ * LLVM integration
+ */
+struct draw_context *draw_create_with_llvm(LLVMExecutionEngineRef engine);
+#endif
 
 #endif /* DRAW_CONTEXT_H */
diff --git a/src/gallium/auxiliary/draw/draw_llvm.c b/src/gallium/auxiliary/draw/draw_llvm.c
new file mode 100644
index 0000000000..6b0ddfd064
--- /dev/null
+++ b/src/gallium/auxiliary/draw/draw_llvm.c
@@ -0,0 +1,311 @@
+#include "draw_llvm.h"
+
+#include "draw_context.h"
+#include "draw_vs.h"
+
+#include "gallivm/lp_bld_arit.h"
+#include "gallivm/lp_bld_interp.h"
+#include "gallivm/lp_bld_struct.h"
+#include "gallivm/lp_bld_type.h"
+#include "gallivm/lp_bld_flow.h"
+#include "gallivm/lp_bld_debug.h"
+#include "gallivm/lp_bld_tgsi.h"
+
+#include "util/u_cpu_detect.h"
+
+#include <llvm-c/Transforms/Scalar.h>
+
+static void
+init_globals(struct draw_llvm *llvm)
+{
+    LLVMTypeRef vertex_header;
+    LLVMTypeRef texture_type;
+
+   /* struct vertex_header */
+   {
+      LLVMTypeRef elem_types[3];
+
+      elem_types[0]  = LLVMIntType(32);
+      elem_types[1]  = LLVMArrayType(LLVMFloatType(), 4);
+      elem_types[2]  = LLVMArrayType(elem_types[1], 0);
+
+      vertex_header = LLVMStructType(elem_types, Elements(elem_types), 0);
+
+      /* these are bit-fields and we can't take address of them
+      LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
+                             llvm->target, vertex_header,
+                             DRAW_JIT_VERTEX_CLIPMASK);
+      LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
+                             llvm->target, vertex_header,
+                             DRAW_JIT_VERTEX_EDGEFLAG);
+      LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
+                             llvm->target, vertex_header,
+                             DRAW_JIT_VERTEX_PAD);
+      LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
+                             llvm->target, vertex_header,
+                             DRAW_JIT_VERTEX_VERTEX_ID);
+      */
+      LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
+                             llvm->target, vertex_header,
+                             DRAW_JIT_VERTEX_CLIP);
+      LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
+                             llvm->target, vertex_header,
+                             DRAW_JIT_VERTEX_DATA);
+
+      LP_CHECK_STRUCT_SIZE(struct vertex_header,
+                           llvm->target, vertex_header);
+
+      LLVMAddTypeName(llvm->module, "vertex_header", vertex_header);
+
+      llvm->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
+   }
+      /* struct draw_jit_texture */
+   {
+      LLVMTypeRef elem_types[4];
+
+      elem_types[DRAW_JIT_TEXTURE_WIDTH]  = LLVMInt32Type();
+      elem_types[DRAW_JIT_TEXTURE_HEIGHT] = LLVMInt32Type();
+      elem_types[DRAW_JIT_TEXTURE_STRIDE] = LLVMInt32Type();
+      elem_types[DRAW_JIT_TEXTURE_DATA]   = LLVMPointerType(LLVMInt8Type(), 0);
+
+      texture_type = LLVMStructType(elem_types, Elements(elem_types), 0);
+
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
+                             llvm->target, texture_type,
+                             DRAW_JIT_TEXTURE_WIDTH);
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
+                             llvm->target, texture_type,
+                             DRAW_JIT_TEXTURE_HEIGHT);
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, stride,
+                             llvm->target, texture_type,
+                             DRAW_JIT_TEXTURE_STRIDE);
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, data,
+                             llvm->target, texture_type,
+                             DRAW_JIT_TEXTURE_DATA);
+      LP_CHECK_STRUCT_SIZE(struct draw_jit_texture,
+                           llvm->target, texture_type);
+
+      LLVMAddTypeName(llvm->module, "texture", texture_type);
+   }
+
+
+   /* struct draw_jit_context */
+   {
+      LLVMTypeRef elem_types[3];
+      LLVMTypeRef context_type;
+
+      elem_types[0] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
+      elem_types[1] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
+      elem_types[2] = LLVMArrayType(texture_type, PIPE_MAX_SAMPLERS); /* textures */
+
+      context_type = LLVMStructType(elem_types, Elements(elem_types), 0);
+
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
+                             llvm->target, context_type, 0);
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, gs_constants,
+                             llvm->target, context_type, 1);
+      LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
+                             llvm->target, context_type,
+                             DRAW_JIT_CONTEXT_TEXTURES_INDEX);
+      LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
+                           llvm->target, context_type);
+
+      LLVMAddTypeName(llvm->module, "context", context_type);
+
+      llvm->context_ptr_type = LLVMPointerType(context_type, 0);
+   }
+}
+
+struct draw_llvm *
+draw_llvm_create(struct draw_context *draw)
+{
+   struct draw_llvm *llvm = CALLOC_STRUCT( draw_llvm );
+
+   util_cpu_detect();
+
+   llvm->draw = draw;
+   llvm->engine = draw->engine;
+
+   debug_assert(llvm->engine);
+
+   llvm->module = LLVMModuleCreateWithName("draw_llvm");
+   llvm->provider = LLVMCreateModuleProviderForExistingModule(llvm->module);
+
+   LLVMAddModuleProvider(llvm->engine, llvm->provider);
+
+   llvm->target = LLVMGetExecutionEngineTargetData(llvm->engine);
+
+   llvm->pass = LLVMCreateFunctionPassManager(llvm->provider);
+   LLVMAddTargetData(llvm->target, llvm->pass);
+   /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
+    * but there are more on SVN. */
+   /* TODO: Add more passes */
+   LLVMAddConstantPropagationPass(llvm->pass);
+   if(util_cpu_caps.has_sse4_1) {
+      /* FIXME: There is a bug in this pass, whereby the combination of fptosi
+       * and sitofp (necessary for trunc/floor/ceil/round implementation)
+       * somehow becomes invalid code.
+       */
+      LLVMAddInstructionCombiningPass(llvm->pass);
+   }
+   LLVMAddPromoteMemoryToRegisterPass(llvm->pass);
+   LLVMAddGVNPass(llvm->pass);
+   LLVMAddCFGSimplificationPass(llvm->pass);
+
+   init_globals(llvm);
+
+
+#if 1
+   LLVMDumpModule(llvm->module);
+#endif
+
+   return llvm;
+}
+
+void
+draw_llvm_destroy(struct draw_llvm *llvm)
+{
+   free(llvm);
+}
+
+void
+draw_llvm_prepare(struct draw_llvm *llvm)
+{
+}
+
+
+struct draw_context *draw_create_with_llvm(LLVMExecutionEngineRef engine)
+{
+   struct draw_context *draw = CALLOC_STRUCT( draw_context );
+   if (draw == NULL)
+      goto fail;
+   draw->engine = engine;
+
+   if (!draw_init(draw))
+      goto fail;
+
+   return draw;
+
+fail:
+   draw_destroy( draw );
+   return NULL;
+}
+
+static void
+generate_vs(struct draw_llvm *llvm,
+            LLVMBuilderRef builder,
+            LLVMValueRef context_ptr,
+            LLVMValueRef io)
+{
+   const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
+   struct lp_type vs_type = lp_type_float(32);
+   LLVMValueRef vs_consts;
+   const LLVMValueRef (*inputs)[NUM_CHANNELS];
+   LLVMValueRef (*outputs)[NUM_CHANNELS];
+
+   lp_build_tgsi_soa(builder,
+                     tokens,
+                     vs_type,
+                     NULL /*struct lp_build_mask_context *mask*/,
+                     vs_consts,
+                     NULL /*pos*/,
+                     inputs,
+                     outputs,
+                     NULL/*sampler*/);
+}
+
+void
+draw_llvm_generate(struct draw_llvm *llvm)
+{
+   LLVMTypeRef arg_types[5];
+   LLVMTypeRef func_type;
+   LLVMValueRef context_ptr;
+   LLVMBasicBlockRef block;
+   LLVMBuilderRef builder;
+   LLVMValueRef function;
+   LLVMValueRef start, end, count, stride, step;
+   LLVMValueRef io_ptr;
+   unsigned i;
+   unsigned chan;
+   struct lp_build_context bld;
+   struct lp_build_loop_state lp_loop;
+   struct lp_type vs_type = lp_type_float(32);
+
+   arg_types[0] = llvm->context_ptr_type;           /* context */
+   arg_types[1] = llvm->vertex_header_ptr_type;     /* vertex_header */
+   arg_types[2] = LLVMInt32Type();                  /* start */
+   arg_types[3] = LLVMInt32Type();                  /* count */
+   arg_types[4] = LLVMInt32Type();                  /* stride */
+
+   func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
+
+   function = LLVMAddFunction(llvm->module, "draw_llvm_shader", func_type);
+   LLVMSetFunctionCallConv(function, LLVMCCallConv);
+   for(i = 0; i < Elements(arg_types); ++i)
+      if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
+         LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
+
+   context_ptr  = LLVMGetParam(function, 0);
+   io_ptr       = LLVMGetParam(function, 1);
+   start        = LLVMGetParam(function, 2);
+   count        = LLVMGetParam(function, 3);
+   stride       = LLVMGetParam(function, 4);
+
+   lp_build_name(context_ptr, "context");
+   lp_build_name(io_ptr, "io");
+   lp_build_name(start, "start");
+   lp_build_name(count, "count");
+   lp_build_name(stride, "stride");
+
+   /*
+    * Function body
+    */
+
+   block = LLVMAppendBasicBlock(function, "entry");
+   builder = LLVMCreateBuilder();
+   LLVMPositionBuilderAtEnd(builder, block);
+
+   lp_build_context_init(&bld, builder, vs_type);
+
+   end = lp_build_add(&bld, start, count);
+
+   step = LLVMConstInt(LLVMInt32Type(), 1, 0);
+   lp_build_loop_begin(builder, start, &lp_loop);
+   {
+      LLVMValueRef io = LLVMBuildGEP(builder, io_ptr, &lp_loop.counter, 1, "");
+
+      generate_vs(llvm,
+                  builder,
+                  context_ptr,
+                  io);
+   }
+   lp_build_loop_end(builder, end, step, &lp_loop);
+
+
+   LLVMBuildRetVoid(builder);
+
+   LLVMDisposeBuilder(builder);
+
+   /*
+    * Translate the LLVM IR into machine code.
+    */
+
+#ifdef DEBUG
+   if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
+      LLVMDumpValue(function);
+      assert(0);
+   }
+#endif
+
+   LLVMRunFunctionPassManager(llvm->pass, function);
+
+   if (1) {
+      LLVMDumpValue(function);
+      debug_printf("\n");
+   }
+
+   llvm->jit_func = (draw_jit_vert_func)LLVMGetPointerToGlobal(llvm->draw->engine, function);
+
+   if (1)
+      lp_disassemble(llvm->jit_func);
+}
diff --git a/src/gallium/auxiliary/draw/draw_llvm.h b/src/gallium/auxiliary/draw/draw_llvm.h
new file mode 100644
index 0000000000..0a1845f1fb
--- /dev/null
+++ b/src/gallium/auxiliary/draw/draw_llvm.h
@@ -0,0 +1,154 @@
+#ifndef DRAW_LLVM_H
+#define DRAW_LLVM_H
+
+#include "draw/draw_private.h"
+
+#include "pipe/p_context.h"
+
+#include <llvm-c/Core.h>
+#include <llvm-c/Analysis.h>
+#include <llvm-c/Target.h>
+#include <llvm-c/ExecutionEngine.h>
+
+struct draw_jit_texture
+{
+   uint32_t width;
+   uint32_t height;
+   uint32_t stride;
+   const void *data;
+};
+
+enum {
+   DRAW_JIT_TEXTURE_WIDTH = 0,
+   DRAW_JIT_TEXTURE_HEIGHT,
+   DRAW_JIT_TEXTURE_STRIDE,
+   DRAW_JIT_TEXTURE_DATA
+};
+
+enum {
+   DRAW_JIT_VERTEX_VERTEX_ID = 0,
+   DRAW_JIT_VERTEX_CLIP,
+   DRAW_JIT_VERTEX_DATA
+};
+
+/**
+ * This structure is passed directly to the generated vertex shader.
+ *
+ * It contains the derived state.
+ *
+ * Changes here must be reflected in the draw_jit_context_* macros.
+ * Changes to the ordering should be avoided.
+ *
+ * Only use types with a clear size and padding here, in particular prefer the
+ * stdint.h types to the basic integer types.
+ */
+struct draw_jit_context
+{
+   const float *vs_constants;
+   const float *gs_constants;
+
+
+   struct draw_jit_texture textures[PIPE_MAX_SAMPLERS];
+};
+
+
+#define draw_jit_context_vs_constants(_builder, _ptr) \
+   lp_build_struct_get(_builder, _ptr, 0, "vs_constants")
+
+#define draw_jit_context_gs_constants(_builder, _ptr) \
+   lp_build_struct_get(_builder, _ptr, 1, "gs_constants")
+
+#define DRAW_JIT_CONTEXT_TEXTURES_INDEX 2
+
+#define draw_jit_context_textures(_builder, _ptr) \
+   lp_build_struct_get_ptr(_builder, _ptr, DRAW_JIT_CONTEXT_TEXTURES_INDEX, "textures")
+
+/* we are construction a function of the form:
+
+struct vertex_header {
+   uint32 vertex_id;
+
+   float clip[4];
+   float data[][4];
+};
+
+struct draw_jit_context
+{
+   const float *vs_constants;
+   const float *gs_constants;
+
+   struct draw_jit_texture textures[PIPE_MAX_SAMPLERS];
+   const void *vbuffers;
+};
+
+void
+draw_shader(struct draw_jit_context *context,
+            struct vertex_header *io,
+            unsigned start,
+            unsigned count,
+            unsigned stride)
+{
+  // do a fetch and a run vertex shader
+  for (int i = 0; i < count; ++i) {
+    struct vertex_header *header = &io[i];
+    header->vertex_id = 0xffff;
+    // follows code-genarted fetch/translate section
+    // for each vertex_element ...
+    codegened_translate(header->data[num_element],
+                       context->vertex_elements[num_element],
+                       context->vertex_buffers,
+                       context->vbuffers);
+
+    codegened_vertex_shader(header->data, context->vs_constants);
+  }
+
+  for (int i = 0; i < count; i += context->primitive_size) {
+     struct vertex_header *prim[MAX_PRIMITIVE_SIZE];
+     for (int j = 0; j < context->primitive_size; ++j) {
+       header[j] = &io[i + j];
+     }
+     codegened_geometry_shader(prim, gs_constants);
+  }
+}
+*/
+
+typedef void
+(*draw_jit_vert_func)(struct draw_jit_context *context,
+                      struct vertex_header *io,
+                      unsigned start,
+                      unsigned count,
+                      unsigned stride);
+
+struct draw_llvm {
+   struct draw_context *draw;
+
+   struct draw_jit_context jit_context;
+
+   draw_jit_vert_func jit_func;
+
+   LLVMModuleRef module;
+   LLVMExecutionEngineRef engine;
+   LLVMModuleProviderRef provider;
+   LLVMTargetDataRef target;
+   LLVMPassManagerRef pass;
+
+   LLVMTypeRef context_ptr_type;
+   LLVMTypeRef vertex_header_ptr_type;
+};
+
+
+struct draw_llvm *
+draw_llvm_create(struct draw_context *draw);
+
+void
+draw_llvm_destroy(struct draw_llvm *llvm);
+
+void
+draw_llvm_prepare(struct draw_llvm *llvm);
+
+/* generates the draw jit function */
+void
+draw_llvm_generate(struct draw_llvm *llvm);
+
+
+#endif
diff --git a/src/gallium/auxiliary/draw/draw_llvm_translate.c b/src/gallium/auxiliary/draw/draw_llvm_translate.c
new file mode 100644
index 0000000000..588e97b29c
--- /dev/null
+++ b/src/gallium/auxiliary/draw/draw_llvm_translate.c
@@ -0,0 +1,653 @@
+
+
+
+#include "util/u_memory.h"
+#include "pipe/p_state.h"
+#include "translate.h"
+
+
+#define DRAW_DBG 0
+
+typedef void (*fetch_func)(const void *ptr, float *attrib);
+typedef void (*emit_func)(const float *attrib, void *ptr);
+
+
+
+struct translate_generic {
+   struct translate translate;
+
+   struct {
+      enum translate_element_type type;
+
+      fetch_func fetch;
+      unsigned buffer;
+      unsigned input_offset;
+      unsigned instance_divisor;
+
+      emit_func emit;
+      unsigned output_offset;
+
+      char *input_ptr;
+      unsigned input_stride;
+
+   } attrib[PIPE_MAX_ATTRIBS];
+
+   unsigned nr_attrib;
+};
+
+
+static struct translate_generic *translate_generic( struct translate *translate )
+{
+   return (struct translate_generic *)translate;
+}
+
+/**
+ * Fetch a float[4] vertex attribute from memory, doing format/type
+ * conversion as needed.
+ *
+ * This is probably needed/dupliocated elsewhere, eg format
+ * conversion, texture sampling etc.
+ */
+#define ATTRIB( NAME, SZ, TYPE, FROM, TO )		\
+static void						\
+fetch_##NAME(const void *ptr, float *attrib)		\
+{							\
+   const float defaults[4] = { 0.0f,0.0f,0.0f,1.0f };	\
+   unsigned i;						\
+							\
+   for (i = 0; i < SZ; i++) {				\
+      attrib[i] = FROM(i);				\
+   }							\
+							\
+   for (; i < 4; i++) {					\
+      attrib[i] = defaults[i];				\
+   }							\
+}							\
+							\
+static void						\
+emit_##NAME(const float *attrib, void *ptr)		\
+{  \
+   unsigned i;						\
+   TYPE *out = (TYPE *)ptr;				\
+							\
+   for (i = 0; i < SZ; i++) {				\
+      out[i] = TO(attrib[i]);				\
+   }							\
+}
+
+{
+
+   return conv = instr(builder, bc, "");
+}
+
+static INLINE LLVMValueRef
+from_64_float(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMDoubleType() , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   return LLVMBuildFPTrunc(builder, l, LLVMFloatType(), "");
+}
+
+static INLINE LLVMValueRef
+from_32_float(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMFloatType() , "");
+   return LLVMBuildLoad(builder, bc, "");
+}
+
+static INLINE LLVMValueRef
+from_8_uscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, val, "");
+   return LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+}
+
+static INLINE LLVMValueRef
+from_16_uscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMIntType(16) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   return LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+}
+
+static INLINE LLVMValueRef
+from_32_uscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMIntType(32) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   return LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+}
+
+static INLINE LLVMValueRef
+from_8_sscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, val, "");
+   return LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+}
+
+static INLINE LLVMValueRef
+from_16_sscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMIntType(16) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   return LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+}
+
+static INLINE LLVMValueRef
+from_32_sscaled(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMIntType(32) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   return LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+}
+
+
+static INLINE LLVMValueRef
+from_8_unorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, val, "");
+   LLVMValueRef uscaled = LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(builder, 255.));
+}
+
+static INLINE LLVMValueRef
+from_16_unorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMIntType(16) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   LLVMValueRef uscaled = LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(builder, 65535.));
+}
+
+static INLINE LLVMValueRef
+from_32_unorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMIntType(32) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   LLVMValueRef uscaled = LLVMBuildUIToFP(builder, l, LLVMFloatType(), "");
+
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(builder, 4294967295.));
+}
+
+static INLINE LLVMValueRef
+from_8_snorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, val, "");
+   LLVMValueRef uscaled = LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(builder, 127.0));
+}
+
+static INLINE LLVMValueRef
+from_16_snorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMIntType(16) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   LLVMValueRef uscaled = LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(builder, 32767.0f));
+}
+
+static INLINE LLVMValueRef
+from_32_snorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMIntType(32) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   LLVMValueRef uscaled = LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(builder, 2147483647.0));
+}
+
+static INLINE LLVMValueRef
+from_32_fixed(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef bc = LLVMBuildBitCast(builder, val,
+                                      LLVMIntType(32) , "");
+   LLVMValueRef l = LLVMBuildLoad(builder, bc, "");
+   LLVMValueRef uscaled = LLVMBuildSIToFP(builder, l, LLVMFloatType(), "");
+
+   return LLVMBuildFDiv(builder, uscaled,
+                        LLVMConstReal(builder, 65536.0));
+}
+
+static INLINE LLVMValueRef
+to_64_float(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPExt(builder, l, LLVMDoubleType(), "");
+}
+
+static INLINE LLVMValueRef
+to_32_float(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   return LLVMBuildLoad(builder, fp, "");
+}
+
+atic INLINE LLVMValueRef
+to_8_uscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToUI(builder, l, LLVMIntType(8), "");
+}
+
+static INLINE LLVMValueRef
+to_16_uscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToUI(builder, l, LLVMIntType(16), "");
+}
+
+static INLINE LLVMValueRef
+to_32_uscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToUI(builder, l, LLVMIntType(32), "");
+}
+
+static INLINE LLVMValueRef
+to_8_sscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToSI(builder, l, LLVMIntType(8), "");
+}
+
+static INLINE LLVMValueRef
+to_16_sscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToSI(builder, l, LLVMIntType(16), "");
+}
+
+static INLINE LLVMValueRef
+to_32_sscaled(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   return LLVMBuildFPToSI(builder, l, LLVMIntType(32), "");
+}
+
+static INLINE LLVMValueRef
+to_8_unorm(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToUI(builder, l, LLVMIntType(8), "");
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(builder, 255.));
+}
+
+static INLINE LLVMValueRef
+to_16_unorm(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToUI(builder, l, LLVMIntType(32), "");
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(builder, 65535.));
+}
+
+static INLINE LLVMValueRef
+to_32_unorm(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToUI(builder, l, LLVMIntType(32), "");
+
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(builder, 4294967295.));
+}
+
+static INLINE LLVMValueRef
+to_8_snorm(LLVMBuilderRef builder, LLVMValueRef val)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, val, "");
+   LLVMValueRef uscaled = LLVMBuildFPToSI(builder, l, LLVMIntType(8), "");
+   return LLVMBuildFMUL(builder, uscaled,
+                        LLVMConstReal(builder, 127.0));
+}
+
+static INLINE LLVMValueRef
+to_16_snorm(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToSI(builder, l, LLVMIntType(16), "");
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(builder, 32767.0f));
+}
+
+static INLINE LLVMValueRef
+to_32_snorm(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToSI(builder, l, LLVMIntType(32), "");
+
+   return LLVMBuildFMUL(builder, uscaled,
+                        LLVMConstReal(builder, 2147483647.0));
+}
+
+static INLINE LLVMValueRef
+to_32_fixed(LLVMBuilderRef builder, LLVMValueRef fp)
+{
+   LLVMValueRef l = LLVMBuildLoad(builder, fp, "");
+   LLVMValueRef uscaled = LLVMBuildFPToSI(builder, l, LLVMIntType(32), "");
+
+   return LLVMBuildFMul(builder, uscaled,
+                        LLVMConstReal(builder, 65536.0));
+}
+
+static LLVMValueRef
+fetch(LLVMValueRef ptr, int val_size, int nr_components,
+     LLVMValueRef res)
+{
+   int i;
+   int offset = 0;
+
+   for (i = 0; i < nr_components; ++i) {
+      LLVMValueRef src_index = LLVMConstInt(LLVMInt32Type(), offset, 0);
+      LLVMValueRef dst_index = LLVMConstInt(LLVMInt32Type(), i, 0);
+      //getelementptr i8* ptr, i64 offset
+      LLVMValueRef src_tmp = LLVMBuildGEP(builder, ptr, &src_index, 1, "");
+      //getelementptr float* res, i64 i
+      LLVMValueRef res_tmp = LLVMBuildGEP(builder, res, &dst_index, 1, "");
+      //bitcast i8* src, to res_type*
+      //load res_type src
+      //convert res_type src to float
+      //store float src, float *dst src
+      offset += val_size;
+   }
+}
+
+
+static void
+fetch_B8G8R8A8_UNORM(const void *ptr, float *attrib)
+{
+   attrib[2] = FROM_8_UNORM(0);
+   attrib[1] = FROM_8_UNORM(1);
+   attrib[0] = FROM_8_UNORM(2);
+   attrib[3] = FROM_8_UNORM(3);
+}
+
+static void
+emit_B8G8R8A8_UNORM( const float *attrib, void *ptr)
+{
+   ubyte *out = (ubyte *)ptr;
+   out[2] = TO_8_UNORM(attrib[0]);
+   out[1] = TO_8_UNORM(attrib[1]);
+   out[0] = TO_8_UNORM(attrib[2]);
+   out[3] = TO_8_UNORM(attrib[3]);
+}
+
+static void
+fetch_NULL( const void *ptr, float *attrib )
+{
+   attrib[0] = 0;
+   attrib[1] = 0;
+   attrib[2] = 0;
+   attrib[3] = 1;
+}
+
+static void
+emit_NULL( const float *attrib, void *ptr )
+{
+   /* do nothing is the only sensible option */
+}
+
+typedef LLVMValueRef (*from_func)(LLVMBuilderRef, LLVMValueRef);
+typedef  LLVMValueRef (*to_func)(LLVMBuilderRef, LLVMValueRef);
+
+struct draw_llvm_translate {
+   int format;
+   from_func from;
+   to_func to;
+   LLVMTypeRef type;
+   int num_components;
+} translates[] =
+{
+   {PIPE_FORMAT_R64_FLOAT,          from_64_float, to_64_float, LLVMDoubleType(), 1},
+   {PIPE_FORMAT_R64G64_FLOAT,       from_64_float, to_64_float, LLVMDoubleType(), 2},
+   {PIPE_FORMAT_R64G64B64_FLOAT,    from_64_float, to_64_float, LLVMDoubleType(), 3},
+   {PIPE_FORMAT_R64G64B64A64_FLOAT, from_64_float, to_64_float, LLVMDoubleType(), 4},
+
+   {PIPE_FORMAT_R32_FLOAT,          from_32_float, to_32_float, LLVMFloatType(), 1},
+   {PIPE_FORMAT_R32G32_FLOAT,       from_32_float, to_32_float, LLVMFloatType(), 2},
+   {PIPE_FORMAT_R32G32B32_FLOAT,    from_32_float, to_32_float, LLVMFloatType(), 3},
+   {PIPE_FORMAT_R32G32B32A32_FLOAT, from_32_float, to_32_float, LLVMFloatType(), 4},
+
+   {PIPE_FORMAT_R32_UNORM,          from_32_unorm, to_32_unorm, LLVMIntType(32), 1},
+   {PIPE_FORMAT_R32G32_UNORM,       from_32_unorm, to_32_unorm, LLVMIntType(32), 2},
+   {PIPE_FORMAT_R32G32B32_UNORM,    from_32_unorm, to_32_unorm, LLVMIntType(32), 3},
+   {PIPE_FORMAT_R32G32B32A32_UNORM, from_32_unorm, to_32_unorm, LLVMIntType(32), 4},
+
+   {PIPE_FORMAT_R32_USCALED,          from_32_uscaled, to_32_uscaled, LLVMIntType(32), 1},
+   {PIPE_FORMAT_R32G32_USCALED,       from_32_uscaled, to_32_uscaled, LLVMIntType(32), 2},
+   {PIPE_FORMAT_R32G32B32_USCALED,    from_32_uscaled, to_32_uscaled, LLVMIntType(32), 3},
+   {PIPE_FORMAT_R32G32B32A32_USCALED, from_32_uscaled, to_32_uscaled, LLVMIntType(32), 4},
+
+   {PIPE_FORMAT_R32_SNORM,          from_32_snorm, to_32_snorm, LLVMIntType(32), 1},
+   {PIPE_FORMAT_R32G32_SNORM,       from_32_snorm, to_32_snorm, LLVMIntType(32), 2},
+   {PIPE_FORMAT_R32G32B32_SNORM,    from_32_snorm, to_32_snorm, LLVMIntType(32), 3},
+   {PIPE_FORMAT_R32G32B32A32_SNORM, from_32_snorm, to_32_snorm, LLVMIntType(32), 4},
+
+   {PIPE_FORMAT_R32_SSCALED,          from_32_sscaled, to_32_sscaled, LLVMIntType(32), 1},
+   {PIPE_FORMAT_R32G32_SSCALED,       from_32_sscaled, to_32_sscaled, LLVMIntType(32), 2},
+   {PIPE_FORMAT_R32G32B32_SSCALED,    from_32_sscaled, to_32_sscaled, LLVMIntType(32), 3},
+   {PIPE_FORMAT_R32G32B32A32_SSCALED, from_32_sscaled, to_32_sscaled, LLVMIntType(32), 4},
+
+   {PIPE_FORMAT_R16_UNORM,          from_16_unorm, to_16_unorm, LLVMIntType(16), 1},
+   {PIPE_FORMAT_R16G16_UNORM,       from_16_unorm, to_16_unorm, LLVMIntType(16), 2},
+   {PIPE_FORMAT_R16G16B16_UNORM,    from_16_unorm, to_16_unorm, LLVMIntType(16), 3},
+   {PIPE_FORMAT_R16G16B16A16_UNORM, from_16_unorm, to_16_unorm, LLVMIntType(16), 4},
+
+   {PIPE_FORMAT_R16_USCALED,          from_16_uscaled, to_16_uscaled, LLVMIntType(16), 1},
+   {PIPE_FORMAT_R16G16_USCALED,       from_16_uscaled, to_16_uscaled, LLVMIntType(16), 2},
+   {PIPE_FORMAT_R16G16B16_USCALED,    from_16_uscaled, to_16_uscaled, LLVMIntType(16), 3},
+   {PIPE_FORMAT_R16G16B16A16_USCALED, from_16_uscaled, to_16_uscaled, LLVMIntType(16), 4},
+
+   {PIPE_FORMAT_R16_SNORM,          from_16_snorm, to_16_snorm, LLVMIntType(16), 1},
+   {PIPE_FORMAT_R16G16_SNORM,       from_16_snorm, to_16_snorm, LLVMIntType(16), 2},
+   {PIPE_FORMAT_R16G16B16_SNORM,    from_16_snorm, to_16_snorm, LLVMIntType(16), 3},
+   {PIPE_FORMAT_R16G16B16A16_SNORM, from_16_snorm, to_16_snorm, LLVMIntType(16), 4},
+
+   {PIPE_FORMAT_R16_SSCALED,          from_16_sscaled, to_16_sscaled, LLVMIntType(16), 1},
+   {PIPE_FORMAT_R16G16_SSCALED,       from_16_sscaled, to_16_sscaled, LLVMIntType(16), 2},
+   {PIPE_FORMAT_R16G16B16_SSCALED,    from_16_sscaled, to_16_sscaled, LLVMIntType(16), 3},
+   {PIPE_FORMAT_R16G16B16A16_SSCALED, from_16_sscaled, to_16_sscaled, LLVMIntType(16), 4},
+
+   {PIPE_FORMAT_R8_UNORM,       from_8_unorm, to_8_unorm, LLVMIntType(8), 1},
+   {PIPE_FORMAT_R8G8_UNORM,     from_8_unorm, to_8_unorm, LLVMIntType(8), 2},
+   {PIPE_FORMAT_R8G8B8_UNORM,   from_8_unorm, to_8_unorm, LLVMIntType(8), 3},
+   {PIPE_FORMAT_R8G8B8A8_UNORM, from_8_unorm, to_8_unorm, LLVMIntType(8), 4},
+
+   {PIPE_FORMAT_R8_USCALED,       from_8_uscaled, to_8_uscaled, LLVMIntType(8), 1},
+   {PIPE_FORMAT_R8G8_USCALED,     from_8_uscaled, to_8_uscaled, LLVMIntType(8), 2},
+   {PIPE_FORMAT_R8G8B8_USCALED,   from_8_uscaled, to_8_uscaled, LLVMIntType(8), 3},
+   {PIPE_FORMAT_R8G8B8A8_USCALED, from_8_uscaled, to_8_uscaled, LLVMIntType(8), 4},
+
+   {PIPE_FORMAT_R8_SNORM,       from_8_snorm, to_8_snorm, LLVMIntType(8), 1},
+   {PIPE_FORMAT_R8G8_SNORM,     from_8_snorm, to_8_snorm, LLVMIntType(8), 2},
+   {PIPE_FORMAT_R8G8B8_SNORM,   from_8_snorm, to_8_snorm, LLVMIntType(8), 3},
+   {PIPE_FORMAT_R8G8B8A8_SNORM, from_8_snorm, to_8_snorm, LLVMIntType(8), 4},
+
+   {PIPE_FORMAT_R8_SSCALED,       from_8_sscaled, to_8_sscaled, LLVMIntType(8), 1},
+   {PIPE_FORMAT_R8G8_SSCALED,     from_8_sscaled, to_8_sscaled, LLVMIntType(8), 2},
+   {PIPE_FORMAT_R8G8B8_SSCALED,   from_8_sscaled, to_8_sscaled, LLVMIntType(8), 3},
+   {PIPE_FORMAT_R8G8B8A8_SSCALED, from_8_sscaled, to_8_sscaled, LLVMIntType(8), 4},
+
+   {PIPE_FORMAT_R32_FIXED,          from_32_fixed, to_32_fixed, LLVMIntType(32), 1},
+   {PIPE_FORMAT_R32G32_FIXED,       from_32_fixed, to_32_fixed, LLVMIntType(32), 2},
+   {PIPE_FORMAT_R32G32B32_FIXED,    from_32_fixed, to_32_fixed, LLVMIntType(32), 3},
+   {PIPE_FORMAT_R32G32B32A32_FIXED, from_32_fixed, to_32_fixed, LLVMIntType(32), 4},
+
+   {PIPE_FORMAT_A8R8G8B8_UNORM, from_8_unorm, to_8_unorm, LLVMIntType(8), 4},
+   {PIPE_FORMAT_B8G8R8A8_UNORM, from_8_unorm, to_8_unorm, LLVMIntType(), 4},
+};
+
+/**
+ * Fetch vertex attributes for 'count' vertices.
+ */
+static void PIPE_CDECL generic_run_elts( struct translate *translate,
+                                         const unsigned *elts,
+                                         unsigned count,
+                                         unsigned instance_id,
+                                         void *output_buffer )
+{
+   struct translate_generic *tg = translate_generic(translate);
+   char *vert = output_buffer;
+   unsigned nr_attrs = tg->nr_attrib;
+   unsigned attr;
+   unsigned i;
+
+   /* loop over vertex attributes (vertex shader inputs)
+    */
+   for (i = 0; i < count; i++) {
+      unsigned elt = *elts++;
+
+      for (attr = 0; attr < nr_attrs; attr++) {
+	 float data[4];
+         const char *src;
+
+	 char *dst = (vert +
+		      tg->attrib[attr].output_offset);
+
+         if (tg->attrib[attr].instance_divisor) {
+            src = tg->attrib[attr].input_ptr +
+                  tg->attrib[attr].input_stride *
+                  (instance_id / tg->attrib[attr].instance_divisor);
+         } else {
+            src = tg->attrib[attr].input_ptr +
+                  tg->attrib[attr].input_stride * elt;
+         }
+
+	 tg->attrib[attr].fetch( src, data );
+
+         if (0) debug_printf("vert %d/%d attr %d: %f %f %f %f\n",
+                             i, elt, attr, data[0], data[1], data[2], data[3]);
+
+	 tg->attrib[attr].emit( data, dst );
+      }
+
+      vert += tg->translate.key.output_stride;
+   }
+}
+
+
+
+static void PIPE_CDECL generic_run( struct translate *translate,
+                                    unsigned start,
+                                    unsigned count,
+                                    unsigned instance_id,
+                                    void *output_buffer )
+{
+   struct translate_generic *tg = translate_generic(translate);
+   char *vert = output_buffer;
+   unsigned nr_attrs = tg->nr_attrib;
+   unsigned attr;
+   unsigned i;
+
+   /* loop over vertex attributes (vertex shader inputs)
+    */
+   for (i = 0; i < count; i++) {
+      unsigned elt = start + i;
+
+      for (attr = 0; attr < nr_attrs; attr++) {
+	 float data[4];
+
+	 char *dst = (vert +
+		      tg->attrib[attr].output_offset);
+
+         if (tg->attrib[attr].type == TRANSLATE_ELEMENT_NORMAL) {
+            const char *src;
+
+            if (tg->attrib[attr].instance_divisor) {
+               src = tg->attrib[attr].input_ptr +
+                     tg->attrib[attr].input_stride *
+                     (instance_id / tg->attrib[attr].instance_divisor);
+            } else {
+               src = tg->attrib[attr].input_ptr +
+                     tg->attrib[attr].input_stride * elt;
+            }
+
+            tg->attrib[attr].fetch( src, data );
+         } else {
+            data[0] = (float)instance_id;
+         }
+
+         if (0) debug_printf("vert %d attr %d: %f %f %f %f\n",
+                             i, attr, data[0], data[1], data[2], data[3]);
+
+	 tg->attrib[attr].emit( data, dst );
+      }
+
+      vert += tg->translate.key.output_stride;
+   }
+}
+
+
+
+static void generic_set_buffer( struct translate *translate,
+				unsigned buf,
+				const void *ptr,
+				unsigned stride )
+{
+   struct translate_generic *tg = translate_generic(translate);
+   unsigned i;
+
+   for (i = 0; i < tg->nr_attrib; i++) {
+      if (tg->attrib[i].buffer == buf) {
+	 tg->attrib[i].input_ptr = ((char *)ptr +
+				    tg->attrib[i].input_offset);
+	 tg->attrib[i].input_stride = stride;
+      }
+   }
+}
+
+
+static void generic_release( struct translate *translate )
+{
+   /* Refcount?
+    */
+   FREE(translate);
+}
+
+struct translate *translate_generic_create( const struct translate_key *key )
+{
+   struct translate_generic *tg = CALLOC_STRUCT(translate_generic);
+   unsigned i;
+
+   if (tg == NULL)
+      return NULL;
+
+   tg->translate.key = *key;
+   tg->translate.release = generic_release;
+   tg->translate.set_buffer = generic_set_buffer;
+   tg->translate.run_elts = generic_run_elts;
+   tg->translate.run = generic_run;
+
+   for (i = 0; i < key->nr_elements; i++) {
+      tg->attrib[i].type = key->element[i].type;
+
+      tg->attrib[i].fetch = get_fetch_func(key->element[i].input_format);
+      tg->attrib[i].buffer = key->element[i].input_buffer;
+      tg->attrib[i].input_offset = key->element[i].input_offset;
+      tg->attrib[i].instance_divisor = key->element[i].instance_divisor;
+
+      tg->attrib[i].emit = get_emit_func(key->element[i].output_format);
+      tg->attrib[i].output_offset = key->element[i].output_offset;
+
+   }
+
+   tg->nr_attrib = key->nr_elements;
+
+
+   return &tg->translate;
+}
diff --git a/src/gallium/auxiliary/draw/draw_private.h b/src/gallium/auxiliary/draw/draw_private.h
index 1e6e01af9e..7e24e5fd6f 100644
--- a/src/gallium/auxiliary/draw/draw_private.h
+++ b/src/gallium/auxiliary/draw/draw_private.h
@@ -46,6 +46,10 @@
 
 #include "tgsi/tgsi_scan.h"
 
+#ifdef DRAW_LLVM
+#include <llvm-c/ExecutionEngine.h>
+#endif
+
 
 struct pipe_context;
 struct draw_vertex_shader;
@@ -237,9 +241,16 @@ struct draw_context
 
    unsigned instance_id;
 
+#ifdef DRAW_LLVM
+   LLVMExecutionEngineRef engine;
+#endif
    void *driver_private;
 };
 
+/*******************************************************************************
+ * Draw common initialization code
+ */
+boolean draw_init(struct draw_context *draw);
 
 /*******************************************************************************
  * Vertex shader code:
diff --git a/src/gallium/auxiliary/draw/draw_pt.c b/src/gallium/auxiliary/draw/draw_pt.c
index 341353f628..9b1e319551 100644
--- a/src/gallium/auxiliary/draw/draw_pt.c
+++ b/src/gallium/auxiliary/draw/draw_pt.c
@@ -142,7 +142,9 @@ boolean draw_pt_init( struct draw_context *draw )
    if (!draw->pt.middle.fetch_shade_emit)
       return FALSE;
 
-   draw->pt.middle.general = draw_pt_fetch_pipeline_or_emit( draw );
+   draw->pt.middle.general = draw_pt_fetch_pipeline_or_emit_llvm( draw );
+   if (!draw->pt.middle.general)
+      draw->pt.middle.general = draw_pt_fetch_pipeline_or_emit( draw );
    if (!draw->pt.middle.general)
       return FALSE;
 
diff --git a/src/gallium/auxiliary/draw/draw_pt.h b/src/gallium/auxiliary/draw/draw_pt.h
index d5e0d92a60..c2797a759e 100644
--- a/src/gallium/auxiliary/draw/draw_pt.h
+++ b/src/gallium/auxiliary/draw/draw_pt.h
@@ -147,6 +147,7 @@ struct draw_pt_front_end *draw_pt_varray(struct draw_context *draw);
 struct draw_pt_middle_end *draw_pt_fetch_emit( struct draw_context *draw );
 struct draw_pt_middle_end *draw_pt_middle_fse( struct draw_context *draw );
 struct draw_pt_middle_end *draw_pt_fetch_pipeline_or_emit(struct draw_context *draw);
+struct draw_pt_middle_end *draw_pt_fetch_pipeline_or_emit_llvm(struct draw_context *draw);
 
 
 
diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline_llvm.c b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline_llvm.c
new file mode 100644
index 0000000000..ae5956333e
--- /dev/null
+++ b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline_llvm.c
@@ -0,0 +1,432 @@
+/**************************************************************************
+ *
+ * Copyright 2010 VMWare, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+
+#include "util/u_math.h"
+#include "util/u_memory.h"
+#include "draw/draw_context.h"
+#include "draw/draw_vbuf.h"
+#include "draw/draw_vertex.h"
+#include "draw/draw_pt.h"
+#include "draw/draw_vs.h"
+#include "draw/draw_gs.h"
+#include "draw/draw_llvm.h"
+
+#include "translate/translate.h"
+
+
+struct llvm_middle_end {
+   struct draw_pt_middle_end base;
+   struct draw_context *draw;
+
+   struct pt_emit *emit;
+   struct pt_fetch *fetch;
+   struct pt_post_vs *post_vs;
+
+
+   unsigned vertex_data_offset;
+   unsigned vertex_size;
+   unsigned prim;
+   unsigned opt;
+
+   struct draw_llvm *llvm;
+};
+
+
+static void
+llvm_middle_end_prepare( struct draw_pt_middle_end *middle,
+                         unsigned prim,
+                         unsigned opt,
+                         unsigned *max_vertices )
+{
+   struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
+   struct draw_context *draw = fpme->draw;
+   struct draw_vertex_shader *vs = draw->vs.vertex_shader;
+   struct draw_geometry_shader *gs = draw->gs.geometry_shader;
+   unsigned i;
+   unsigned instance_id_index = ~0;
+
+   /* Add one to num_outputs because the pipeline occasionally tags on
+    * an additional texcoord, eg for AA lines.
+    */
+   unsigned nr = MAX2( vs->info.num_inputs,
+		       vs->info.num_outputs + 1 );
+
+   /* Scan for instanceID system value.
+    */
+   for (i = 0; i < vs->info.num_inputs; i++) {
+      if (vs->info.input_semantic_name[i] == TGSI_SEMANTIC_INSTANCEID) {
+         instance_id_index = i;
+         break;
+      }
+   }
+
+   fpme->prim = prim;
+   fpme->opt = opt;
+
+   /* Always leave room for the vertex header whether we need it or
+    * not.  It's hard to get rid of it in particular because of the
+    * viewport code in draw_pt_post_vs.c.
+    */
+   fpme->vertex_size = sizeof(struct vertex_header) + nr * 4 * sizeof(float);
+
+
+
+   draw_pt_fetch_prepare( fpme->fetch,
+                          vs->info.num_inputs,
+                          fpme->vertex_size,
+                          instance_id_index );
+   if (opt & PT_SHADE) {
+      vs->prepare(vs, draw);
+      draw_geometry_shader_prepare(gs, draw);
+   }
+
+
+   /* XXX: it's not really gl rasterization rules we care about here,
+    * but gl vs dx9 clip spaces.
+    */
+   draw_pt_post_vs_prepare( fpme->post_vs,
+			    (boolean)draw->bypass_clipping,
+			    (boolean)(draw->identity_viewport ||
+			    draw->rasterizer->bypass_vs_clip_and_viewport),
+			    (boolean)draw->rasterizer->gl_rasterization_rules,
+			    (draw->vs.edgeflag_output ? true : false) );
+
+   if (!(opt & PT_PIPELINE)) {
+      draw_pt_emit_prepare( fpme->emit,
+			    prim,
+                            max_vertices );
+
+      *max_vertices = MAX2( *max_vertices,
+                            DRAW_PIPE_MAX_VERTICES );
+   }
+   else {
+      *max_vertices = DRAW_PIPE_MAX_VERTICES;
+   }
+
+   /* return even number */
+   *max_vertices = *max_vertices & ~1;
+
+   draw_llvm_prepare(fpme->llvm);
+}
+
+
+
+static void llvm_middle_end_run( struct draw_pt_middle_end *middle,
+                                 const unsigned *fetch_elts,
+                                 unsigned fetch_count,
+                                 const ushort *draw_elts,
+                                 unsigned draw_count )
+{
+   struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
+   struct draw_context *draw = fpme->draw;
+   struct draw_vertex_shader *vshader = draw->vs.vertex_shader;
+   struct draw_geometry_shader *gshader = draw->gs.geometry_shader;
+   unsigned opt = fpme->opt;
+   unsigned alloc_count = align( fetch_count, 4 );
+
+   struct vertex_header *pipeline_verts =
+      (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
+
+   if (!pipeline_verts) {
+      /* Not much we can do here - just skip the rendering.
+       */
+      assert(0);
+      return;
+   }
+
+   /* Fetch into our vertex buffer
+    */
+   draw_pt_fetch_run( fpme->fetch,
+		      fetch_elts,
+		      fetch_count,
+		      (char *)pipeline_verts );
+
+   /* Run the shader, note that this overwrites the data[] parts of
+    * the pipeline verts.  If there is no shader, eg if
+    * bypass_vs_clip_and_viewport, then the inputs == outputs, and are
+    * already in the correct place.*/
+   if (opt & PT_SHADE)
+   {
+      vshader->run_linear(vshader,
+                          (const float (*)[4])pipeline_verts->data,
+                          (      float (*)[4])pipeline_verts->data,
+                          draw->pt.user.vs_constants,
+                          fetch_count,
+                          fpme->vertex_size,
+                          fpme->vertex_size);
+      if (gshader)
+         draw_geometry_shader_run(gshader,
+                                  (const float (*)[4])pipeline_verts->data,
+                                  (      float (*)[4])pipeline_verts->data,
+                                  draw->pt.user.gs_constants,
+                                  fetch_count,
+                                  fpme->vertex_size,
+                                  fpme->vertex_size);
+   }
+
+   if (draw_pt_post_vs_run( fpme->post_vs,
+			    pipeline_verts,
+			    fetch_count,
+			    fpme->vertex_size ))
+   {
+      opt |= PT_PIPELINE;
+   }
+
+   /* Do we need to run the pipeline?
+    */
+   if (opt & PT_PIPELINE) {
+      draw_pipeline_run( fpme->draw,
+                         fpme->prim,
+                         pipeline_verts,
+                         fetch_count,
+                         fpme->vertex_size,
+                         draw_elts,
+                         draw_count );
+   }
+   else {
+      draw_pt_emit( fpme->emit,
+		    (const float (*)[4])pipeline_verts->data,
+		    fetch_count,
+		    fpme->vertex_size,
+		    draw_elts,
+		    draw_count );
+   }
+
+
+   FREE(pipeline_verts);
+}
+
+
+static void llvm_middle_end_linear_run( struct draw_pt_middle_end *middle,
+                                       unsigned start,
+                                       unsigned count)
+{
+   struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
+   struct draw_context *draw = fpme->draw;
+   unsigned opt = fpme->opt;
+   unsigned alloc_count = align( count, 4 );
+
+   struct vertex_header *pipeline_verts =
+      (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
+
+   if (!pipeline_verts) {
+      /* Not much we can do here - just skip the rendering.
+       */
+      assert(0);
+      return;
+   }
+
+   fpme->llvm->jit_func( &fpme->llvm->jit_context,
+                         pipeline_verts,
+                         start,
+                         count,
+                         fpme->vertex_size );
+
+   if (draw_pt_post_vs_run( fpme->post_vs,
+			    pipeline_verts,
+			    count,
+			    fpme->vertex_size ))
+   {
+      opt |= PT_PIPELINE;
+   }
+
+   /* Do we need to run the pipeline?
+    */
+   if (opt & PT_PIPELINE) {
+      draw_pipeline_run_linear( fpme->draw,
+                                fpme->prim,
+                                pipeline_verts,
+                                count,
+                                fpme->vertex_size);
+   }
+   else {
+      draw_pt_emit_linear( fpme->emit,
+                           (const float (*)[4])pipeline_verts->data,
+                           fpme->vertex_size,
+                           count );
+   }
+
+   FREE(pipeline_verts);
+}
+
+
+
+static boolean
+llvm_middle_end_linear_run_elts( struct draw_pt_middle_end *middle,
+                                 unsigned start,
+                                 unsigned count,
+                                 const ushort *draw_elts,
+                                 unsigned draw_count )
+{
+   struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
+   struct draw_context *draw = fpme->draw;
+   struct draw_vertex_shader *shader = draw->vs.vertex_shader;
+   struct draw_geometry_shader *geometry_shader = draw->gs.geometry_shader;
+   unsigned opt = fpme->opt;
+   unsigned alloc_count = align( count, 4 );
+
+   struct vertex_header *pipeline_verts =
+      (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
+
+   if (!pipeline_verts)
+      return FALSE;
+
+   /* Fetch into our vertex buffer
+    */
+   draw_pt_fetch_run_linear( fpme->fetch,
+                             start,
+                             count,
+                             (char *)pipeline_verts );
+
+   /* Run the shader, note that this overwrites the data[] parts of
+    * the pipeline verts.  If there is no shader, ie if
+    * bypass_vs_clip_and_viewport, then the inputs == outputs, and are
+    * already in the correct place.
+    */
+   if (opt & PT_SHADE)
+   {
+      shader->run_linear(shader,
+			 (const float (*)[4])pipeline_verts->data,
+			 (      float (*)[4])pipeline_verts->data,
+			 draw->pt.user.vs_constants,
+			 count,
+			 fpme->vertex_size,
+			 fpme->vertex_size);
+
+      if (geometry_shader)
+         draw_geometry_shader_run(geometry_shader,
+                                  (const float (*)[4])pipeline_verts->data,
+                                  (      float (*)[4])pipeline_verts->data,
+                                  draw->pt.user.gs_constants,
+                                  count,
+                                  fpme->vertex_size,
+                                  fpme->vertex_size);
+   }
+
+   if (draw_pt_post_vs_run( fpme->post_vs,
+			    pipeline_verts,
+			    count,
+			    fpme->vertex_size ))
+   {
+      opt |= PT_PIPELINE;
+   }
+
+   /* Do we need to run the pipeline?
+    */
+   if (opt & PT_PIPELINE) {
+      draw_pipeline_run( fpme->draw,
+                         fpme->prim,
+                         pipeline_verts,
+                         count,
+                         fpme->vertex_size,
+                         draw_elts,
+                         draw_count );
+   }
+   else {
+      draw_pt_emit( fpme->emit,
+		    (const float (*)[4])pipeline_verts->data,
+		    count,
+		    fpme->vertex_size,
+		    draw_elts,
+		    draw_count );
+   }
+
+   FREE(pipeline_verts);
+   return TRUE;
+}
+
+
+
+static void llvm_middle_end_finish( struct draw_pt_middle_end *middle )
+{
+   /* nothing to do */
+}
+
+static void llvm_middle_end_destroy( struct draw_pt_middle_end *middle )
+{
+   struct llvm_middle_end *fpme = (struct llvm_middle_end *)middle;
+
+   if (fpme->fetch)
+      draw_pt_fetch_destroy( fpme->fetch );
+
+   if (fpme->emit)
+      draw_pt_emit_destroy( fpme->emit );
+
+   if (fpme->post_vs)
+      draw_pt_post_vs_destroy( fpme->post_vs );
+
+   if (fpme->llvm)
+      draw_llvm_destroy( fpme->llvm );
+
+   FREE(middle);
+}
+
+
+struct draw_pt_middle_end *draw_pt_fetch_pipeline_or_emit_llvm( struct draw_context *draw )
+{
+   struct llvm_middle_end *fpme = 0;
+
+   if (!draw->engine)
+      return NULL;
+
+   fpme = CALLOC_STRUCT( llvm_middle_end );
+   if (!fpme)
+      goto fail;
+
+   fpme->base.prepare         = llvm_middle_end_prepare;
+   fpme->base.run             = llvm_middle_end_run;
+   fpme->base.run_linear      = llvm_middle_end_linear_run;
+   fpme->base.run_linear_elts = llvm_middle_end_linear_run_elts;
+   fpme->base.finish          = llvm_middle_end_finish;
+   fpme->base.destroy         = llvm_middle_end_destroy;
+
+   fpme->draw = draw;
+
+   fpme->fetch = draw_pt_fetch_create( draw );
+   if (!fpme->fetch)
+      goto fail;
+
+   fpme->post_vs = draw_pt_post_vs_create( draw );
+   if (!fpme->post_vs)
+      goto fail;
+
+   fpme->emit = draw_pt_emit_create( draw );
+   if (!fpme->emit)
+      goto fail;
+
+   fpme->llvm = draw_llvm_create(draw);
+   if (!fpme->llvm)
+      goto fail;
+
+   return &fpme->base;
+
+ fail:
+   if (fpme)
+      llvm_middle_end_destroy( &fpme->base );
+
+   return NULL;
+}
diff --git a/src/gallium/auxiliary/draw/draw_vs_llvm.c b/src/gallium/auxiliary/draw/draw_vs_llvm.c
index 5f7a645f5d..0c483de407 100644
--- a/src/gallium/auxiliary/draw/draw_vs_llvm.c
+++ b/src/gallium/auxiliary/draw/draw_vs_llvm.c
@@ -40,7 +40,7 @@
 
 #include "tgsi/tgsi_parse.h"
 
-#ifdef MESA_LLVM
+#ifdef DRAW_LLVM
 
 struct draw_llvm_vertex_shader {
    struct draw_vertex_shader base;
@@ -64,12 +64,8 @@ vs_llvm_run_linear( struct draw_vertex_shader *base,
 		   unsigned input_stride,
 		   unsigned output_stride )
 {
-   struct draw_llvm_vertex_shader *shader =
-      (struct draw_llvm_vertex_shader *)base;
 }
 
-
-
 static void
 vs_llvm_delete( struct draw_vertex_shader *base )
 {
@@ -90,6 +86,7 @@ struct draw_vertex_shader *
 draw_create_vs_llvm(struct draw_context *draw,
 		    const struct pipe_shader_state *templ)
 {
+#if 0
    struct draw_llvm_vertex_shader *vs;
 
    vs = CALLOC_STRUCT( draw_llvm_vertex_shader );
@@ -113,6 +110,8 @@ draw_create_vs_llvm(struct draw_context *draw,
    vs->machine = draw->vs.machine;
 
    return &vs->base;
+#endif
+   return NULL;
 }
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_context.c b/src/gallium/drivers/llvmpipe/lp_context.c
index 9120226de0..3edc62d0c6 100644
--- a/src/gallium/drivers/llvmpipe/lp_context.c
+++ b/src/gallium/drivers/llvmpipe/lp_context.c
@@ -40,6 +40,7 @@
 #include "lp_context.h"
 #include "lp_flush.h"
 #include "lp_perf.h"
+#include "lp_screen.h"
 #include "lp_state.h"
 #include "lp_surface.h"
 #include "lp_query.h"
@@ -105,6 +106,7 @@ struct pipe_context *
 llvmpipe_create_context( struct pipe_screen *screen, void *priv )
 {
    struct llvmpipe_context *llvmpipe;
+   struct llvmpipe_screen *llvmscreen = llvmpipe_screen(screen);
 
    llvmpipe = align_malloc(sizeof(struct llvmpipe_context), 16);
    if (!llvmpipe)
@@ -174,8 +176,8 @@ llvmpipe_create_context( struct pipe_screen *screen, void *priv )
    /*
     * Create drawing context and plug our rendering stage into it.
     */
-   llvmpipe->draw = draw_create();
-   if (!llvmpipe->draw) 
+   llvmpipe->draw = draw_create_with_llvm(llvmscreen->engine);
+   if (!llvmpipe->draw)
       goto fail;
 
    /* FIXME: devise alternative to draw_texture_samplers */