Merge branch 'master' into asm-shader-rework-2

Conflicts:
	src/mesa/shader/lex.yy.c
	src/mesa/shader/program_parse.tab.c
	src/mesa/shader/program_parse.tab.h

60 files changed, 2313 insertions, 350 deletions

diff --git a/src/gallium/drivers/cell/spu/spu_exec.c b/src/gallium/drivers/cell/spu/spu_exec.c
index 6db9501128..0eaae2e451 100644
--- a/src/gallium/drivers/cell/spu/spu_exec.c
+++ b/src/gallium/drivers/cell/spu/spu_exec.c
@@ -1150,10 +1150,6 @@ exec_instruction(
       ASSERT (0);
       break;
 
-   case TGSI_OPCODE_CND0:
-      ASSERT (0);
-      break;
-
    case TGSI_OPCODE_DP2A:
       ASSERT (0);
       break;
diff --git a/src/gallium/drivers/i915simple/i915_flush.c b/src/gallium/drivers/i915simple/i915_flush.c
index b223dadab6..1582168eba 100644
--- a/src/gallium/drivers/i915simple/i915_flush.c
+++ b/src/gallium/drivers/i915simple/i915_flush.c
@@ -45,6 +45,7 @@ static void i915_flush( struct pipe_context *pipe,
 
    draw_flush(i915->draw);
 
+#if 0
    /* Do we need to emit an MI_FLUSH command to flush the hardware
     * caches?
     */
@@ -63,6 +64,13 @@ static void i915_flush( struct pipe_context *pipe,
       }
       OUT_BATCH( flush );
    }
+#endif
+
+#if 0
+   if (i915->batch->map == i915->batch->ptr) {
+      return;
+   }
+#endif
 
    /* If there are no flags, just flush pending commands to hardware:
     */
diff --git a/src/gallium/drivers/i915simple/i915_prim_vbuf.c b/src/gallium/drivers/i915simple/i915_prim_vbuf.c
index 508f4560e4..b3a7774fd6 100644
--- a/src/gallium/drivers/i915simple/i915_prim_vbuf.c
+++ b/src/gallium/drivers/i915simple/i915_prim_vbuf.c
@@ -44,6 +44,7 @@
 #include "pipe/p_inlines.h"
 #include "util/u_math.h"
 #include "util/u_memory.h"
+#include "util/u_fifo.h"
 
 #include "i915_context.h"
 #include "i915_reg.h"
@@ -76,8 +77,13 @@ struct i915_vbuf_render {
    size_t vbo_size;
    size_t vbo_offset;
    void *vbo_ptr;
-   size_t vbo_alloc_size;
    size_t vbo_max_used;
+
+   /* stuff for the pool */
+   struct util_fifo *pool_fifo;
+   unsigned pool_used;
+   unsigned pool_buffer_size;
+   boolean pool_not_used;
 };
 
 
@@ -106,33 +112,72 @@ i915_vbuf_render_get_vertex_info(struct vbuf_render *render)
 }
 
 static boolean
+i915_vbuf_render_reserve(struct i915_vbuf_render *i915_render, size_t size)
+{
+   struct i915_context *i915 = i915_render->i915;
+
+   if (i915_render->vbo_size < size + i915_render->vbo_offset)
+      return FALSE;
+
+   if (i915->vbo_flushed)
+      return FALSE;
+
+   return TRUE;
+}
+
+static void
+i915_vbuf_render_new_buf(struct i915_vbuf_render *i915_render, size_t size)
+{
+   struct i915_context *i915 = i915_render->i915;
+   struct intel_winsys *iws = i915->iws;
+
+   if (i915_render->vbo) {
+      if (i915_render->pool_not_used)
+         iws->buffer_destroy(iws, i915_render->vbo);
+      else
+         u_fifo_add(i915_render->pool_fifo, i915_render->vbo);
+      i915_render->vbo = NULL;
+   }
+
+   i915->vbo_flushed = 0;
+
+   i915_render->vbo_size = MAX2(size, i915_render->pool_buffer_size);
+   i915_render->vbo_offset = 0;
+
+   if (i915_render->vbo_size != i915_render->pool_buffer_size) {
+      i915_render->pool_not_used = TRUE;
+      i915_render->vbo = iws->buffer_create(iws, i915_render->vbo_size, 64,
+            INTEL_NEW_VERTEX);
+   } else {
+      i915_render->pool_not_used = FALSE;
+
+      if (i915_render->pool_used >= 2) {
+         FLUSH_BATCH(NULL);
+         i915->vbo_flushed = 0;
+         i915_render->pool_used = 0;
+      }
+      u_fifo_pop(i915_render->pool_fifo, (void**)&i915_render->vbo);
+   }
+}
+
+static boolean
 i915_vbuf_render_allocate_vertices(struct vbuf_render *render,
                                    ushort vertex_size,
                                    ushort nr_vertices)
 {
    struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
    struct i915_context *i915 = i915_render->i915;
-   struct intel_winsys *iws = i915->iws;
    size_t size = (size_t)vertex_size * (size_t)nr_vertices;
 
    /* FIXME: handle failure */
    assert(!i915->vbo);
 
-   if (i915_render->vbo_size > size + i915_render->vbo_offset && !i915->vbo_flushed) {
-   } else {
-      i915->vbo_flushed = 0;
-      if (i915_render->vbo) {
-         iws->buffer_destroy(iws, i915_render->vbo);
-         i915_render->vbo = NULL;
-      }
-   }
+   if (!i915_vbuf_render_reserve(i915_render, size)) {
 
-   if (!i915_render->vbo) {
-      i915_render->vbo_size = MAX2(size, i915_render->vbo_alloc_size);
-      i915_render->vbo_offset = 0;
-      i915_render->vbo = iws->buffer_create(iws, i915_render->vbo_size, 64,
-                                            INTEL_NEW_VERTEX);
+      if (i915->vbo_flushed)
+         i915_render->pool_used = 0;
 
+      i915_vbuf_render_new_buf(i915_render, size);
    }
 
    i915_render->vertex_size = vertex_size;
@@ -504,6 +549,7 @@ i915_vbuf_render_create(struct i915_context *i915)
 {
    struct i915_vbuf_render *i915_render = CALLOC_STRUCT(i915_vbuf_render);
    struct intel_winsys *iws = i915->iws;
+   int i;
 
    i915_render->i915 = i915;
    
@@ -524,14 +570,24 @@ i915_vbuf_render_create(struct i915_context *i915)
    i915_render->base.release_vertices = i915_vbuf_render_release_vertices;
    i915_render->base.destroy = i915_vbuf_render_destroy;
 
-   i915_render->vbo_alloc_size = 128 * 4096;
-   i915_render->vbo_size = i915_render->vbo_alloc_size;
+
+   i915_render->vbo = NULL;
+   i915_render->vbo_size = 0;
    i915_render->vbo_offset = 0;
-   i915_render->vbo = iws->buffer_create(iws, i915_render->vbo_size, 64,
-                                         INTEL_NEW_VERTEX);
+
+   i915_render->pool_used = FALSE;
+   i915_render->pool_buffer_size = 128 * 4096;
+   i915_render->pool_fifo = u_fifo_create(6);
+   for (i = 0; i < 6; i++)
+      u_fifo_add(i915_render->pool_fifo,
+                 iws->buffer_create(iws, i915_render->pool_buffer_size, 64,
+                                    INTEL_NEW_VERTEX));
+
+#if 0
    /* TODO JB: is this realy needed? */
    i915_render->vbo_ptr = iws->buffer_map(iws, i915_render->vbo, TRUE);
    iws->buffer_unmap(iws, i915_render->vbo);
+#endif
 
    return &i915_render->base;
 }
diff --git a/src/gallium/drivers/i915simple/intel_winsys.h b/src/gallium/drivers/i915simple/intel_winsys.h
index e123c32d88..42c5e7470e 100644
--- a/src/gallium/drivers/i915simple/intel_winsys.h
+++ b/src/gallium/drivers/i915simple/intel_winsys.h
@@ -69,8 +69,8 @@ struct intel_batchbuffer {
     * each dword written.
     */
    /*{@*/
-   void *map;
-   void *ptr;
+   uint8_t *map;
+   uint8_t *ptr;
    size_t size;
 
    size_t relocs;
@@ -150,6 +150,17 @@ struct intel_winsys {
    void (*buffer_unmap)(struct intel_winsys *iws,
                         struct intel_buffer *buffer);
 
+   /**
+    * Write to a buffer.
+    *
+    * Arguments follows pwrite(2)
+    */
+   int (*buffer_write)(struct intel_winsys *iws,
+                       struct intel_buffer *dst,
+                       const void *src,
+                       size_t size,
+                       size_t offset);
+
    void (*buffer_destroy)(struct intel_winsys *iws,
                           struct intel_buffer *buffer);
    /*@}*/
diff --git a/src/gallium/drivers/llvmpipe/Makefile b/src/gallium/drivers/llvmpipe/Makefile
index 6e63a0c2b7..06c586e6bb 100644
--- a/src/gallium/drivers/llvmpipe/Makefile
+++ b/src/gallium/drivers/llvmpipe/Makefile
@@ -15,9 +15,11 @@ C_SOURCES = \
 	lp_bld_depth.c \
 	lp_bld_flow.c \
 	lp_bld_format_aos.c \
+	lp_bld_format_soa.c \
 	lp_bld_interp.c \
 	lp_bld_intr.c \
 	lp_bld_logic.c \
+	lp_bld_sample_soa.c \
 	lp_bld_swizzle.c \
 	lp_bld_struct.c \
 	lp_bld_tgsi_soa.c \
@@ -44,7 +46,8 @@ C_SOURCES = \
 	lp_state_vs.c \
 	lp_surface.c \
 	lp_tex_cache.c \
-	lp_tex_sample.c \
+	lp_tex_sample_c.c \
+	lp_tex_sample_llvm.c \
 	lp_texture.c \
 	lp_tile_cache.c \
 	lp_tile_soa.c
diff --git a/src/gallium/drivers/llvmpipe/SConscript b/src/gallium/drivers/llvmpipe/SConscript
index 6bceb84da4..dea4b703c4 100644
--- a/src/gallium/drivers/llvmpipe/SConscript
+++ b/src/gallium/drivers/llvmpipe/SConscript
@@ -2,8 +2,12 @@ Import('*')
 
 env = env.Clone()
 
+env.Tool('llvm')
+if env.has_key('LLVM_VERSION') is False:
+    print 'warning: LLVM not found: not building llvmpipe'
+    Return()
+
 env.Tool('udis86')
-env.ParseConfig('llvm-config --cppflags')
 
 llvmpipe = env.ConvenienceLibrary(
 	target = 'llvmpipe',
@@ -19,8 +23,10 @@ llvmpipe = env.ConvenienceLibrary(
 		'lp_bld_depth.c',
 		'lp_bld_flow.c',
 		'lp_bld_format_aos.c',
+		'lp_bld_format_soa.c',
 		'lp_bld_interp.c',
 		'lp_bld_intr.c',
+		'lp_bld_sample_soa.c',
 		'lp_bld_struct.c',
 		'lp_bld_logic.c',
 		'lp_bld_swizzle.c',
@@ -48,7 +54,8 @@ llvmpipe = env.ConvenienceLibrary(
 		'lp_state_vs.c',
 		'lp_surface.c',
 		'lp_tex_cache.c',
-		'lp_tex_sample.c',
+		'lp_tex_sample_c.c',
+		'lp_tex_sample_llvm.c',
 		'lp_texture.c',
 		'lp_tile_cache.c',
 		'lp_tile_soa.c',
@@ -57,8 +64,6 @@ llvmpipe = env.ConvenienceLibrary(
 
 env = env.Clone()
 
-env['LINK'] = env['CXX']
-env.ParseConfig('llvm-config --libs jit interpreter nativecodegen bitwriter')
 env.Prepend(LIBS = [llvmpipe] + auxiliaries)
 
 env.Program(
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.c b/src/gallium/drivers/llvmpipe/lp_bld_arit.c
index 09a57ff33d..be7442d00a 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_arit.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.c
@@ -502,6 +502,31 @@ lp_build_div(struct lp_build_context *bld,
 }
 
 
+LLVMValueRef
+lp_build_lerp(struct lp_build_context *bld,
+              LLVMValueRef x,
+              LLVMValueRef v0,
+              LLVMValueRef v1)
+{
+   return lp_build_add(bld, v0, lp_build_mul(bld, x, lp_build_sub(bld, v1, v0)));
+}
+
+
+LLVMValueRef
+lp_build_lerp_2d(struct lp_build_context *bld,
+                 LLVMValueRef x,
+                 LLVMValueRef y,
+                 LLVMValueRef v00,
+                 LLVMValueRef v01,
+                 LLVMValueRef v10,
+                 LLVMValueRef v11)
+{
+   LLVMValueRef v0 = lp_build_lerp(bld, x, v00, v01);
+   LLVMValueRef v1 = lp_build_lerp(bld, x, v10, v11);
+   return lp_build_lerp(bld, y, v0, v1);
+}
+
+
 /**
  * Generate min(a, b)
  * Do checks for special cases.
@@ -566,20 +591,31 @@ lp_build_abs(struct lp_build_context *bld,
              LLVMValueRef a)
 {
    const union lp_type type = bld->type;
+   LLVMTypeRef vec_type = lp_build_vec_type(type);
 
    if(!type.sign)
       return a;
 
-   /* XXX: is this really necessary? */
+   if(type.floating) {
+      /* Mask out the sign bit */
+      LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
+      LLVMValueRef mask = lp_build_int_const_scalar(type, ((unsigned long long)1 << type.width) - 1);
+      a = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
+      a = LLVMBuildAnd(bld->builder, a, mask, "");
+      a = LLVMBuildBitCast(bld->builder, a, vec_type, "");
+      return a;
+   }
+
 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
-   if(!type.floating && type.width*type.length == 128) {
-      LLVMTypeRef vec_type = lp_build_vec_type(type);
-      if(type.width == 8)
+   if(type.width*type.length == 128) {
+      switch(type.width) {
+      case 8:
          return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.b.128", vec_type, a);
-      if(type.width == 16)
+      case 16:
          return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.w.128", vec_type, a);
-      if(type.width == 32)
+      case 32:
          return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.d.128", vec_type, a);
+      }
    }
 #endif
 
@@ -587,6 +623,89 @@ lp_build_abs(struct lp_build_context *bld,
 }
 
 
+enum lp_build_round_sse41_mode
+{
+   LP_BUILD_ROUND_SSE41_NEAREST = 0,
+   LP_BUILD_ROUND_SSE41_FLOOR = 1,
+   LP_BUILD_ROUND_SSE41_CEIL = 2,
+   LP_BUILD_ROUND_SSE41_TRUNCATE = 3
+};
+
+
+static INLINE LLVMValueRef
+lp_build_round_sse41(struct lp_build_context *bld,
+                     LLVMValueRef a,
+                     enum lp_build_round_sse41_mode mode)
+{
+   const union lp_type type = bld->type;
+   LLVMTypeRef vec_type = lp_build_vec_type(type);
+   const char *intrinsic;
+
+   assert(type.floating);
+   assert(type.width*type.length == 128);
+
+   switch(type.width) {
+   case 32:
+      intrinsic = "llvm.x86.sse41.round.ps";
+      break;
+   case 64:
+      intrinsic = "llvm.x86.sse41.round.pd";
+      break;
+   default:
+      assert(0);
+      return bld->undef;
+   }
+
+   return lp_build_intrinsic_binary(bld->builder, intrinsic, vec_type, a,
+                                    LLVMConstInt(LLVMInt32Type(), mode, 0));
+}
+
+
+LLVMValueRef
+lp_build_floor(struct lp_build_context *bld,
+               LLVMValueRef a)
+{
+   const union lp_type type = bld->type;
+
+   assert(type.floating);
+
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+   return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
+#endif
+
+   /* FIXME */
+   assert(0);
+   return bld->undef;
+}
+
+
+/**
+ * Convert to integer, through whichever rounding method that's fastest,
+ * typically truncating to zero.
+ */
+LLVMValueRef
+lp_build_int(struct lp_build_context *bld,
+             LLVMValueRef a)
+{
+   const union lp_type type = bld->type;
+   LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
+
+   assert(type.floating);
+
+   return LLVMBuildFPToSI(bld->builder, a, int_vec_type, "");
+}
+
+
+LLVMValueRef
+lp_build_ifloor(struct lp_build_context *bld,
+                LLVMValueRef a)
+{
+   a = lp_build_floor(bld, a);
+   a = lp_build_int(bld, a);
+   return a;
+}
+
+
 LLVMValueRef
 lp_build_sqrt(struct lp_build_context *bld,
               LLVMValueRef a)
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.h b/src/gallium/drivers/llvmpipe/lp_bld_arit.h
index fc8cb25966..383c3c3313 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_arit.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.h
@@ -72,6 +72,26 @@ lp_build_div(struct lp_build_context *bld,
              LLVMValueRef b);
 
 LLVMValueRef
+lp_build_lerp(struct lp_build_context *bld,
+              LLVMValueRef x,
+              LLVMValueRef v0,
+              LLVMValueRef v1);
+
+/**
+ * Bilinear interpolation.
+ *
+ * Values indices are in v_{yx}.
+ */
+LLVMValueRef
+lp_build_lerp_2d(struct lp_build_context *bld,
+                 LLVMValueRef x,
+                 LLVMValueRef y,
+                 LLVMValueRef v00,
+                 LLVMValueRef v01,
+                 LLVMValueRef v10,
+                 LLVMValueRef v11);
+
+LLVMValueRef
 lp_build_min(struct lp_build_context *bld,
              LLVMValueRef a,
              LLVMValueRef b);
@@ -86,6 +106,18 @@ lp_build_abs(struct lp_build_context *bld,
              LLVMValueRef a);
 
 LLVMValueRef
+lp_build_floor(struct lp_build_context *bld,
+               LLVMValueRef a);
+
+LLVMValueRef
+lp_build_int(struct lp_build_context *bld,
+             LLVMValueRef a);
+
+LLVMValueRef
+lp_build_ifloor(struct lp_build_context *bld,
+                LLVMValueRef a);
+
+LLVMValueRef
 lp_build_sqrt(struct lp_build_context *bld,
               LLVMValueRef a);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_debug.c b/src/gallium/drivers/llvmpipe/lp_bld_debug.c
index 30925b5f41..59d8f492e6 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_debug.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_debug.c
@@ -30,10 +30,27 @@
 #include <udis86.h>
 #endif
 
+#include "util/u_math.h"
 #include "util/u_debug.h"
 #include "lp_bld_debug.h"
 
 
+/**
+ * Check alignment.
+ *
+ * It is important that this check is not implemented as a macro or inlined
+ * function, as the compiler assumptions in respect to alignment of global
+ * and stack variables would often make the check a no op, defeating the
+ * whole purpose of the exercise.
+ */
+boolean
+lp_check_alignment(const void *ptr, unsigned alignment)
+{
+   assert(util_is_pot(alignment));
+   return ((uintptr_t)ptr & (alignment - 1)) == 0;
+}
+
+
 void
 lp_disassemble(const void* func)
 {
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_debug.h b/src/gallium/drivers/llvmpipe/lp_bld_debug.h
index ecdafef76d..583e6132b4 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_debug.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_debug.h
@@ -53,6 +53,10 @@ lp_build_name(LLVMValueRef val, const char *format, ...)
 }
 
 
+boolean
+lp_check_alignment(const void *ptr, unsigned alignment);
+
+
 void
 lp_disassemble(const void* func);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_depth.c b/src/gallium/drivers/llvmpipe/lp_bld_depth.c
index 2cd6e6b921..3f88a14b5d 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_depth.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_depth.c
@@ -179,12 +179,13 @@ lp_build_depth_test(LLVMBuilderRef builder,
       padding_right = 0;
       for(chan = 0; chan < z_swizzle; ++chan)
          padding_right += format_desc->channel[chan].size;
-      padding_left = format_desc->block.bits - format_desc->channel[z_swizzle].size;
+      padding_left = format_desc->block.bits -
+                     (padding_right + format_desc->channel[z_swizzle].size);
 
       if(padding_left || padding_right) {
-         const long long mask_left = ((long long)1 << (format_desc->block.bits - padding_left)) - 1;
-         const long long mask_right = ((long long)1 << (padding_right)) - 1;
-         z_bitmask = lp_build_int_const_scalar(type, mask_left & mask_right);
+         const unsigned long long mask_left = ((unsigned long long)1 << (format_desc->block.bits - padding_left)) - 1;
+         const unsigned long long mask_right = ((unsigned long long)1 << (padding_right)) - 1;
+         z_bitmask = lp_build_int_const_scalar(type, mask_left ^ mask_right);
       }
 
       if(padding_left)
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_flow.c b/src/gallium/drivers/llvmpipe/lp_bld_flow.c
index 9d99e1a9d9..69ed014ff3 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_flow.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_flow.c
@@ -32,59 +32,261 @@
  */
 
 #include "util/u_debug.h"
+#include "util/u_memory.h"
 
 #include "lp_bld_type.h"
 #include "lp_bld_flow.h"
 
 
+#define LP_BUILD_FLOW_MAX_VARIABLES 32
+#define LP_BUILD_FLOW_MAX_DEPTH 32
+
+
+/**
+ * Enumeration of all possible flow constructs.
+ */
+enum lp_build_flow_construct_kind {
+   lP_BUILD_FLOW_SCOPE,
+   LP_BUILD_FLOW_SKIP,
+};
+
+
+/**
+ * Variable declaration scope.
+ */
+struct lp_build_flow_scope
+{
+   /** Number of variables declared in this scope */
+   unsigned num_variables;
+};
+
+
+/**
+ * Early exit. Useful to skip to the end of a function or block when
+ * the execution mask becomes zero or when there is an error condition.
+ */
+struct lp_build_flow_skip
+{
+   /** Block to skip to */
+   LLVMBasicBlockRef block;
+
+   /** Number of variables declared at the beginning */
+   unsigned num_variables;
+
+   LLVMValueRef *phi;
+};
+
+
+/**
+ * Union of all possible flow constructs' data
+ */
+union lp_build_flow_construct_data
+{
+   struct lp_build_flow_scope scope;
+   struct lp_build_flow_skip skip;
+};
+
+
+/**
+ * Element of the flow construct stack.
+ */
+struct lp_build_flow_construct
+{
+   enum lp_build_flow_construct_kind kind;
+   union lp_build_flow_construct_data data;
+};
+
+
+/**
+ * All necessary data to generate LLVM control flow constructs.
+ *
+ * Besides keeping track of the control flow construct themselves we also
+ * need to keep track of variables in order to generate SSA Phi values.
+ */
+struct lp_build_flow_context
+{
+   LLVMBuilderRef builder;
+
+   /**
+    * Control flow stack.
+    */
+   struct lp_build_flow_construct constructs[LP_BUILD_FLOW_MAX_DEPTH];
+   unsigned num_constructs;
+
+   /**
+    * Variable stack
+    */
+   LLVMValueRef *variables[LP_BUILD_FLOW_MAX_VARIABLES];
+   unsigned num_variables;
+};
+
+
+struct lp_build_flow_context *
+lp_build_flow_create(LLVMBuilderRef builder)
+{
+   struct lp_build_flow_context *flow;
+
+   flow = CALLOC_STRUCT(lp_build_flow_context);
+   if(!flow)
+      return NULL;
+
+   flow->builder = builder;
+
+   return flow;
+}
+
+
 void
-lp_build_mask_begin(struct lp_build_mask_context *mask,
-                    LLVMBuilderRef builder,
-                    union lp_type type,
-                    LLVMValueRef value)
+lp_build_flow_destroy(struct lp_build_flow_context *flow)
 {
-   memset(mask, 0, sizeof *mask);
+   assert(flow->num_constructs == 0);
+   assert(flow->num_variables == 0);
+   FREE(flow);
+}
 
-   mask->builder = builder;
-   mask->reg_type = LLVMIntType(type.width * type.length);
-   mask->value = value;
+
+static union lp_build_flow_construct_data *
+lp_build_flow_push(struct lp_build_flow_context *flow,
+                   enum lp_build_flow_construct_kind kind)
+{
+   assert(flow->num_constructs < LP_BUILD_FLOW_MAX_DEPTH);
+   if(flow->num_constructs >= LP_BUILD_FLOW_MAX_DEPTH)
+      return NULL;
+
+   flow->constructs[flow->num_constructs].kind = kind;
+   return &flow->constructs[flow->num_constructs++].data;
+}
+
+
+static union lp_build_flow_construct_data *
+lp_build_flow_peek(struct lp_build_flow_context *flow,
+                   enum lp_build_flow_construct_kind kind)
+{
+   assert(flow->num_constructs);
+   if(!flow->num_constructs)
+      return NULL;
+
+   assert(flow->constructs[flow->num_constructs - 1].kind == kind);
+   if(flow->constructs[flow->num_constructs - 1].kind != kind)
+      return NULL;
+
+   return &flow->constructs[flow->num_constructs - 1].data;
 }
 
 
+static union lp_build_flow_construct_data *
+lp_build_flow_pop(struct lp_build_flow_context *flow,
+                  enum lp_build_flow_construct_kind kind)
+{
+   assert(flow->num_constructs);
+   if(!flow->num_constructs)
+      return NULL;
+
+   assert(flow->constructs[flow->num_constructs - 1].kind == kind);
+   if(flow->constructs[flow->num_constructs - 1].kind != kind)
+      return NULL;
+
+   return &flow->constructs[--flow->num_constructs].data;
+}
+
+
+/**
+ * Begin a variable scope.
+ *
+ *
+ */
 void
-lp_build_mask_update(struct lp_build_mask_context *mask,
-                     LLVMValueRef value)
+lp_build_flow_scope_begin(struct lp_build_flow_context *flow)
 {
+   struct lp_build_flow_scope *scope;
 
-   LLVMValueRef cond;
-   LLVMBasicBlockRef current_block;
-   LLVMBasicBlockRef next_block;
-   LLVMBasicBlockRef new_block;
+   scope = &lp_build_flow_push(flow, lP_BUILD_FLOW_SCOPE)->scope;
+   if(!scope)
+      return;
 
-   if(mask->value)
-      mask->value = LLVMBuildAnd(mask->builder, mask->value, value, "");
-   else
-      mask->value = value;
+   scope->num_variables = 0;
+}
 
-   /* FIXME: disabled until we have proper control flow helpers */
-#if 0
-   cond = LLVMBuildICmp(mask->builder,
-                        LLVMIntEQ,
-                        LLVMBuildBitCast(mask->builder, mask->value, mask->reg_type, ""),
-                        LLVMConstNull(mask->reg_type),
-                        "");
 
-   current_block = LLVMGetInsertBlock(mask->builder);
+/**
+ * Declare a variable.
+ *
+ * A variable is a named entity which can have different LLVMValueRef's at
+ * different points of the program. This is relevant for control flow because
+ * when there are mutiple branches to a same location we need to replace
+ * the variable's value with a Phi function as explained in
+ * http://en.wikipedia.org/wiki/Static_single_assignment_form .
+ *
+ * We keep track of variables by keeping around a pointer to where their
+ * current.
+ *
+ * There are a few cautions to observe:
+ *
+ * - Variable's value must not be NULL. If there is no initial value then
+ *   LLVMGetUndef() should be used.
+ *
+ * - Variable's value must be kept up-to-date. If the variable is going to be
+ *   modified by a function then a pointer should be passed so that its value
+ *   is accurate. Failure to do this will cause some of the variables'
+ *   transient values to be lost, leading to wrong results.
+ *
+ * - A program should be written from top to bottom, by always appending
+ *   instructions to the bottom with a single LLVMBuilderRef. Inserting and/or
+ *   modifying existing statements will most likely lead to wrong results.
+ *
+ */
+void
+lp_build_flow_scope_declare(struct lp_build_flow_context *flow,
+                            LLVMValueRef *variable)
+{
+   struct lp_build_flow_scope *scope;
+
+   scope = &lp_build_flow_peek(flow, lP_BUILD_FLOW_SCOPE)->scope;
+   if(!scope)
+      return;
 
-   if(!mask->skip_block) {
-      LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
-      mask->skip_block = LLVMAppendBasicBlock(function, "skip");
+   assert(*variable);
+   if(!*variable)
+      return;
+
+   assert(flow->num_variables < LP_BUILD_FLOW_MAX_VARIABLES);
+   if(flow->num_variables >= LP_BUILD_FLOW_MAX_VARIABLES)
+      return;
+
+   flow->variables[flow->num_variables++] = variable;
+   ++scope->num_variables;
+}
+
+
+void
+lp_build_flow_scope_end(struct lp_build_flow_context *flow)
+{
+   struct lp_build_flow_scope *scope;
+
+   scope = &lp_build_flow_pop(flow, lP_BUILD_FLOW_SCOPE)->scope;
+   if(!scope)
+      return;
 
-      mask->phi = LLVMBuildPhi(mask->builder, LLVMTypeOf(mask->value), "");
+   assert(flow->num_variables >= scope->num_variables);
+   if(flow->num_variables < scope->num_variables) {
+      flow->num_variables = 0;
+      return;
    }
 
+   flow->num_variables -= scope->num_variables;
+}
+
+
+static LLVMBasicBlockRef
+lp_build_flow_insert_block(struct lp_build_flow_context *flow)
+{
+   LLVMBasicBlockRef current_block;
+   LLVMBasicBlockRef next_block;
+   LLVMBasicBlockRef new_block;
+
+   current_block = LLVMGetInsertBlock(flow->builder);
+
    next_block = LLVMGetNextBasicBlock(current_block);
-   assert(next_block);
    if(next_block) {
       new_block = LLVMInsertBasicBlock(next_block, "");
    }
@@ -93,30 +295,148 @@ lp_build_mask_update(struct lp_build_mask_context *mask,
       new_block = LLVMAppendBasicBlock(function, "");
    }
 
-   LLVMAddIncoming(mask->phi, &mask->value, &current_block, 1);
-   LLVMBuildCondBr(mask->builder, cond, mask->skip_block, new_block);
+   return new_block;
+}
+
+void
+lp_build_flow_skip_begin(struct lp_build_flow_context *flow)
+{
+   struct lp_build_flow_skip *skip;
+   LLVMBuilderRef builder;
+   unsigned i;
+
+   skip = &lp_build_flow_push(flow, LP_BUILD_FLOW_SKIP)->skip;
+   if(!skip)
+      return;
+
+   skip->block = lp_build_flow_insert_block(flow);
+   skip->num_variables = flow->num_variables;
+   if(!skip->num_variables) {
+      skip->phi = NULL;
+      return;
+   }
 
-   LLVMPositionBuilderAtEnd(mask->builder, new_block);
-#endif
+   skip->phi = MALLOC(skip->num_variables * sizeof *skip->phi);
+   if(!skip->phi) {
+      skip->num_variables = 0;
+      return;
+   }
+
+   builder = LLVMCreateBuilder();
+   LLVMPositionBuilderAtEnd(builder, skip->block);
+
+   for(i = 0; i < skip->num_variables; ++i)
+      skip->phi[i] = LLVMBuildPhi(builder, LLVMTypeOf(*flow->variables[i]), "");
+
+   LLVMDisposeBuilder(builder);
 }
 
 
-LLVMValueRef
-lp_build_mask_end(struct lp_build_mask_context *mask)
+void
+lp_build_flow_skip_cond_break(struct lp_build_flow_context *flow,
+                              LLVMValueRef cond)
 {
-   if(mask->skip_block) {
-      LLVMBasicBlockRef current_block = LLVMGetInsertBlock(mask->builder);
+   struct lp_build_flow_skip *skip;
+   LLVMBasicBlockRef current_block;
+   LLVMBasicBlockRef new_block;
+   unsigned i;
 
-      LLVMAddIncoming(mask->phi, &mask->value, &current_block, 1);
-      LLVMBuildBr(mask->builder, mask->skip_block);
+   skip = &lp_build_flow_peek(flow, LP_BUILD_FLOW_SKIP)->skip;
+   if(!skip)
+      return;
 
-      LLVMPositionBuilderAtEnd(mask->builder, mask->skip_block);
+   current_block = LLVMGetInsertBlock(flow->builder);
 
-      mask->value = mask->phi;
-      mask->phi = NULL;
-      mask->skip_block = NULL;
+   new_block = lp_build_flow_insert_block(flow);
+
+   for(i = 0; i < skip->num_variables; ++i) {
+      assert(*flow->variables[i]);
+      LLVMAddIncoming(skip->phi[i], flow->variables[i], &current_block, 1);
    }
 
+   LLVMBuildCondBr(flow->builder, cond, skip->block, new_block);
+
+   LLVMPositionBuilderAtEnd(flow->builder, new_block);
+ }
+
+
+void
+lp_build_flow_skip_end(struct lp_build_flow_context *flow)
+{
+   struct lp_build_flow_skip *skip;
+   LLVMBasicBlockRef current_block;
+   unsigned i;
+
+   skip = &lp_build_flow_pop(flow, LP_BUILD_FLOW_SKIP)->skip;
+   if(!skip)
+      return;
+
+   current_block = LLVMGetInsertBlock(flow->builder);
+
+   for(i = 0; i < skip->num_variables; ++i) {
+      assert(*flow->variables[i]);
+      LLVMAddIncoming(skip->phi[i], flow->variables[i], &current_block, 1);
+      *flow->variables[i] = skip->phi[i];
+   }
+
+   LLVMBuildBr(flow->builder, skip->block);
+   LLVMPositionBuilderAtEnd(flow->builder, skip->block);
+
+   FREE(skip->phi);
+}
+
+
+static void
+lp_build_mask_check(struct lp_build_mask_context *mask)
+{
+   LLVMBuilderRef builder = mask->flow->builder;
+   LLVMValueRef cond;
+
+   cond = LLVMBuildICmp(builder,
+                        LLVMIntEQ,
+                        LLVMBuildBitCast(builder, mask->value, mask->reg_type, ""),
+                        LLVMConstNull(mask->reg_type),
+                        "");
+
+   lp_build_flow_skip_cond_break(mask->flow, cond);
+}
+
+
+void
+lp_build_mask_begin(struct lp_build_mask_context *mask,
+                    struct lp_build_flow_context *flow,
+                    union lp_type type,
+                    LLVMValueRef value)
+{
+   memset(mask, 0, sizeof *mask);
+
+   mask->flow = flow;
+   mask->reg_type = LLVMIntType(type.width * type.length);
+   mask->value = value;
+
+   lp_build_flow_scope_begin(flow);
+   lp_build_flow_scope_declare(flow, &mask->value);
+   lp_build_flow_skip_begin(flow);
+
+   lp_build_mask_check(mask);
+}
+
+
+void
+lp_build_mask_update(struct lp_build_mask_context *mask,
+                     LLVMValueRef value)
+{
+   mask->value = LLVMBuildAnd( mask->flow->builder, mask->value, value, "");
+
+   lp_build_mask_check(mask);
+}
+
+
+LLVMValueRef
+lp_build_mask_end(struct lp_build_mask_context *mask)
+{
+   lp_build_flow_skip_end(mask->flow);
+   lp_build_flow_scope_end(mask->flow);
    return mask->value;
 }
 
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_flow.h b/src/gallium/drivers/llvmpipe/lp_bld_flow.h
index 1b634ff038..9d76e3064d 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_flow.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_flow.h
@@ -41,23 +41,49 @@
 union lp_type;
 
 
+struct lp_build_flow_context;
+
+
+struct lp_build_flow_context *
+lp_build_flow_create(LLVMBuilderRef builder);
+
+void
+lp_build_flow_destroy(struct lp_build_flow_context *flow);
+
+void
+lp_build_flow_scope_begin(struct lp_build_flow_context *flow);
+
+void
+lp_build_flow_scope_declare(struct lp_build_flow_context *flow,
+                            LLVMValueRef *variable);
+
+void
+lp_build_flow_scope_end(struct lp_build_flow_context *flow);
+
+void
+lp_build_flow_skip_begin(struct lp_build_flow_context *flow);
+
+void
+lp_build_flow_skip_cond_break(struct lp_build_flow_context *flow,
+                              LLVMValueRef cond);
+
+void
+lp_build_flow_skip_end(struct lp_build_flow_context *flow);
+
+
 struct lp_build_mask_context
 {
-   LLVMBuilderRef builder;
+   struct lp_build_flow_context *flow;
 
    LLVMTypeRef reg_type;
 
    LLVMValueRef value;
-
-   LLVMValueRef phi;
-
-   LLVMBasicBlockRef skip_block;
 };
 
 
 void
 lp_build_mask_begin(struct lp_build_mask_context *mask,
-                    LLVMBuilderRef builder,
+                    struct lp_build_flow_context *flow,
                     union lp_type type,
                     LLVMValueRef value);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_format.h b/src/gallium/drivers/llvmpipe/lp_bld_format.h
index 01c8a752d1..5ee0656093 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_format.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_format.h
@@ -31,20 +31,14 @@
 
 /**
  * @file
- * LLVM IR building helpers interfaces.
- *
- * We use LLVM-C bindings for now. They are not documented, but follow the C++
- * interfaces very closely, and appear to be complete enough for code
- * genration. See
- * http://npcontemplation.blogspot.com/2008/06/secret-of-llvm-c-bindings.html
- * for a standalone example.
+ * Pixel format helpers.
  */
 
 #include <llvm-c/Core.h>  
- 
-#include "pipe/p_format.h"
 
+#include "pipe/p_format.h"
 
+struct util_format_description;
 union lp_type;
 
 
@@ -56,9 +50,9 @@ union lp_type;
  * @return RGBA in a 4 floats vector.
  */
 LLVMValueRef
-lp_build_unpack_rgba(LLVMBuilderRef builder,
-                     enum pipe_format format, 
-                     LLVMValueRef packed);
+lp_build_unpack_rgba_aos(LLVMBuilderRef builder,
+                         enum pipe_format format,
+                         LLVMValueRef packed);
 
 
 /**
@@ -67,9 +61,9 @@ lp_build_unpack_rgba(LLVMBuilderRef builder,
  * @param rgba 4 float vector with the unpacked components.
  */
 LLVMValueRef
-lp_build_pack_rgba(LLVMBuilderRef builder,
-                   enum pipe_format format,
-                   LLVMValueRef rgba);
+lp_build_pack_rgba_aos(LLVMBuilderRef builder,
+                       enum pipe_format format,
+                       LLVMValueRef rgba);
 
 
 /**
@@ -81,9 +75,9 @@ lp_build_pack_rgba(LLVMBuilderRef builder,
  * @return RGBA in a 4 floats vector.
  */
 LLVMValueRef
-lp_build_load_rgba(LLVMBuilderRef builder,
-                   enum pipe_format format, 
-                   LLVMValueRef ptr);
+lp_build_load_rgba_aos(LLVMBuilderRef builder,
+                       enum pipe_format format,
+                       LLVMValueRef ptr);
 
 
 /**
@@ -92,10 +86,34 @@ lp_build_load_rgba(LLVMBuilderRef builder,
  * @param rgba 4 float vector with the unpacked components.
  */
 void 
-lp_build_store_rgba(LLVMBuilderRef builder,
-                    enum pipe_format format,
-                    LLVMValueRef ptr,
-                    LLVMValueRef rgba);
+lp_build_store_rgba_aos(LLVMBuilderRef builder,
+                        enum pipe_format format,
+                        LLVMValueRef ptr,
+                        LLVMValueRef rgba);
 
+LLVMValueRef
+lp_build_gather(LLVMBuilderRef builder,
+                unsigned length,
+                unsigned src_width,
+                unsigned dst_width,
+                LLVMValueRef base_ptr,
+                LLVMValueRef offsets);
+
+
+void
+lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
+                         const struct util_format_description *format_desc,
+                         union lp_type type,
+                         LLVMValueRef packed,
+                         LLVMValueRef *rgba);
+
+
+void
+lp_build_load_rgba_soa(LLVMBuilderRef builder,
+                       const struct util_format_description *format_desc,
+                       union lp_type type,
+                       LLVMValueRef base_ptr,
+                       LLVMValueRef offsets,
+                       LLVMValueRef *rgba);
 
 #endif /* !LP_BLD_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_format_aos.c b/src/gallium/drivers/llvmpipe/lp_bld_format_aos.c
index dcbc0076c7..b9b5d84bed 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_format_aos.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_format_aos.c
@@ -32,9 +32,9 @@
 
 
 LLVMValueRef
-lp_build_unpack_rgba(LLVMBuilderRef builder,
-                     enum pipe_format format,
-                     LLVMValueRef packed)
+lp_build_unpack_rgba_aos(LLVMBuilderRef builder,
+                         enum pipe_format format,
+                         LLVMValueRef packed)
 {
    const struct util_format_description *desc;
    LLVMTypeRef type;
@@ -152,9 +152,9 @@ lp_build_unpack_rgba(LLVMBuilderRef builder,
 
 
 LLVMValueRef
-lp_build_pack_rgba(LLVMBuilderRef builder,
-                   enum pipe_format format,
-                   LLVMValueRef rgba)
+lp_build_pack_rgba_aos(LLVMBuilderRef builder,
+                       enum pipe_format format,
+                       LLVMValueRef rgba)
 {
    const struct util_format_description *desc;
    LLVMTypeRef type;
@@ -250,9 +250,9 @@ lp_build_pack_rgba(LLVMBuilderRef builder,
 
 
 LLVMValueRef
-lp_build_load_rgba(LLVMBuilderRef builder,
-                   enum pipe_format format,
-                   LLVMValueRef ptr)
+lp_build_load_rgba_aos(LLVMBuilderRef builder,
+                       enum pipe_format format,
+                       LLVMValueRef ptr)
 {
    const struct util_format_description *desc;
    LLVMTypeRef type;
@@ -272,15 +272,15 @@ lp_build_load_rgba(LLVMBuilderRef builder,
 
    packed = LLVMBuildLoad(builder, ptr, "");
 
-   return lp_build_unpack_rgba(builder, format, packed);
+   return lp_build_unpack_rgba_aos(builder, format, packed);
 }
 
 
 void
-lp_build_store_rgba(LLVMBuilderRef builder,
-                    enum pipe_format format,
-                    LLVMValueRef ptr,
-                    LLVMValueRef rgba)
+lp_build_store_rgba_aos(LLVMBuilderRef builder,
+                        enum pipe_format format,
+                        LLVMValueRef ptr,
+                        LLVMValueRef rgba)
 {
    const struct util_format_description *desc;
    LLVMTypeRef type;
@@ -294,7 +294,7 @@ lp_build_store_rgba(LLVMBuilderRef builder,
 
    type = LLVMIntType(desc->block.bits);
 
-   packed = lp_build_pack_rgba(builder, format, rgba);
+   packed = lp_build_pack_rgba_aos(builder, format, rgba);
 
    ptr = LLVMBuildBitCast(builder, ptr, LLVMPointerType(type, 0), "");
 
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_format_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_format_soa.c
new file mode 100644
index 0000000000..569e8d10a3
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_bld_format_soa.c
@@ -0,0 +1,208 @@
+/**************************************************************************
+ *
+ * Copyright 2009 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 VMWARE 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_format.h"
+
+#include "lp_bld_type.h"
+#include "lp_bld_const.h"
+#include "lp_bld_conv.h"
+#include "lp_bld_format.h"
+
+
+/**
+ * Gather elements from scatter positions in memory into a single vector.
+ *
+ * @param src_width src element width
+ * @param dst_width result element width (source will be expanded to fit)
+ * @param length length of the offsets,
+ * @param base_ptr base pointer, should be a i8 pointer type.
+ * @param offsets vector with offsets
+ */
+LLVMValueRef
+lp_build_gather(LLVMBuilderRef builder,
+                unsigned length,
+                unsigned src_width,
+                unsigned dst_width,
+                LLVMValueRef base_ptr,
+                LLVMValueRef offsets)
+{
+   LLVMTypeRef src_type = LLVMIntType(src_width);
+   LLVMTypeRef src_ptr_type = LLVMPointerType(src_type, 0);
+   LLVMTypeRef dst_elem_type = LLVMIntType(dst_width);
+   LLVMTypeRef dst_vec_type = LLVMVectorType(dst_elem_type, length);
+   LLVMValueRef res;
+   unsigned i;
+
+   res = LLVMGetUndef(dst_vec_type);
+   for(i = 0; i < length; ++i) {
+      LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
+      LLVMValueRef elem_offset;
+      LLVMValueRef elem_ptr;
+      LLVMValueRef elem;
+
+      elem_offset = LLVMBuildExtractElement(builder, offsets, index, "");
+      elem_ptr = LLVMBuildGEP(builder, base_ptr, &elem_offset, 1, "");
+      elem_ptr = LLVMBuildBitCast(builder, elem_ptr, src_ptr_type, "");
+      elem = LLVMBuildLoad(builder, elem_ptr, "");
+
+      assert(src_width <= dst_width);
+      if(src_width > dst_width)
+         elem = LLVMBuildTrunc(builder, elem, dst_elem_type, "");
+      if(src_width < dst_width)
+         elem = LLVMBuildZExt(builder, elem, dst_elem_type, "");
+
+      res = LLVMBuildInsertElement(builder, res, elem, index, "");
+   }
+
+   return res;
+}
+
+
+static LLVMValueRef
+lp_build_format_swizzle(union lp_type type,
+                        const LLVMValueRef *inputs,
+                        enum util_format_swizzle swizzle)
+{
+   switch (swizzle) {
+   case UTIL_FORMAT_SWIZZLE_X:
+   case UTIL_FORMAT_SWIZZLE_Y:
+   case UTIL_FORMAT_SWIZZLE_Z:
+   case UTIL_FORMAT_SWIZZLE_W:
+      return inputs[swizzle];
+   case UTIL_FORMAT_SWIZZLE_0:
+      return lp_build_zero(type);
+   case UTIL_FORMAT_SWIZZLE_1:
+      return lp_build_one(type);
+   case UTIL_FORMAT_SWIZZLE_NONE:
+      return lp_build_undef(type);
+   default:
+      assert(0);
+      return lp_build_undef(type);
+   }
+}
+
+
+void
+lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
+                         const struct util_format_description *format_desc,
+                         union lp_type type,
+                         LLVMValueRef packed,
+                         LLVMValueRef *rgba)
+{
+   LLVMValueRef inputs[4];
+   unsigned start;
+   unsigned chan;
+
+   /* FIXME: Support more formats */
+   assert(format_desc->layout == UTIL_FORMAT_LAYOUT_ARITH);
+   assert(format_desc->block.width == 1);
+   assert(format_desc->block.height == 1);
+   assert(format_desc->block.bits <= 32);
+
+   /* Decode the input vector components */
+   start = 0;
+   for (chan = 0; chan < 4; ++chan) {
+      unsigned width = format_desc->channel[chan].size;
+      unsigned stop = start + width;
+      LLVMValueRef input;
+
+      input = packed;
+
+      switch(format_desc->channel[chan].type) {
+      case UTIL_FORMAT_TYPE_VOID:
+         input = NULL;
+         break;
+
+      case UTIL_FORMAT_TYPE_UNSIGNED:
+         if(type.floating) {
+            if(start)
+               input = LLVMBuildLShr(builder, input, lp_build_int_const_scalar(type, start), "");
+            if(stop < format_desc->block.bits) {
+               unsigned mask = ((unsigned long long)1 << width) - 1;
+               input = LLVMBuildAnd(builder, input, lp_build_int_const_scalar(type, mask), "");
+            }
+
+            if(format_desc->channel[chan].normalized)
+               input = lp_build_unsigned_norm_to_float(builder, width, type, input);
+            else
+               input = LLVMBuildFPToSI(builder, input, lp_build_vec_type(type), "");
+         }
+         else {
+            /* FIXME */
+            assert(0);
+            input = lp_build_undef(type);
+         }
+         break;
+
+      default:
+         /* fall through */
+         input = lp_build_undef(type);
+         break;
+      }
+
+      inputs[chan] = input;
+
+      start = stop;
+   }
+
+   if(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
+      enum util_format_swizzle swizzle = format_desc->swizzle[0];
+      LLVMValueRef depth = lp_build_format_swizzle(type, inputs, swizzle);
+      rgba[2] = rgba[1] = rgba[0] = depth;
+      rgba[3] = lp_build_one(type);
+   }
+   else {
+      for (chan = 0; chan < 4; ++chan) {
+         enum util_format_swizzle swizzle = format_desc->swizzle[chan];
+         rgba[chan] = lp_build_format_swizzle(type, inputs, swizzle);
+      }
+   }
+}
+
+
+void
+lp_build_load_rgba_soa(LLVMBuilderRef builder,
+                       const struct util_format_description *format_desc,
+                       union lp_type type,
+                       LLVMValueRef base_ptr,
+                       LLVMValueRef offsets,
+                       LLVMValueRef *rgba)
+{
+   LLVMValueRef packed;
+
+   assert(format_desc->layout == UTIL_FORMAT_LAYOUT_ARITH);
+   assert(format_desc->block.width == 1);
+   assert(format_desc->block.height == 1);
+   assert(format_desc->block.bits <= 32);
+
+   packed = lp_build_gather(builder,
+                            type.length, format_desc->block.bits, type.width,
+                            base_ptr, offsets);
+
+   lp_build_unpack_rgba_soa(builder, format_desc, type, packed, rgba);
+}
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_sample.h b/src/gallium/drivers/llvmpipe/lp_bld_sample.h
new file mode 100644
index 0000000000..6f565af76d
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_bld_sample.h
@@ -0,0 +1,135 @@
+/**************************************************************************
+ *
+ * Copyright 2009 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 VMWARE 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.
+ *
+ **************************************************************************/
+
+/**
+ * @file
+ * Texture sampling.
+ *
+ * @author Jose Fonseca <jfonseca@vmware.com>
+ */
+
+#ifndef LP_BLD_SAMPLE_H
+#define LP_BLD_SAMPLE_H
+
+
+#include <llvm-c/Core.h>
+
+struct pipe_texture;
+struct pipe_sampler_state;
+union lp_type;
+
+
+/**
+ * Sampler static state.
+ *
+ * These are the bits of state from pipe_texture and pipe_sampler_state that
+ * are embedded in the generated code.
+ */
+struct lp_sampler_static_state
+{
+   /* pipe_texture's state */
+   enum pipe_format format;
+   unsigned target:2;
+   unsigned pot_width:1;
+   unsigned pot_height:1;
+   unsigned pot_depth:1;
+
+   /* pipe_sampler_state's state */
+   unsigned wrap_s:3;
+   unsigned wrap_t:3;
+   unsigned wrap_r:3;
+   unsigned min_img_filter:2;
+   unsigned min_mip_filter:2;
+   unsigned mag_img_filter:2;
+   unsigned compare_mode:1;
+   unsigned compare_func:3;
+   unsigned normalized_coords:1;
+   unsigned prefilter:4;
+};
+
+
+/**
+ * Sampler dynamic state.
+ *
+ * These are the bits of state from pipe_texture and pipe_sampler_state that
+ * are computed in runtime.
+ *
+ * There are obtained through callbacks, as we don't want to tie the texture
+ * sampling code generation logic to any particular texture layout or pipe
+ * driver.
+ */
+struct lp_sampler_dynamic_state
+{
+
+   /** Obtain the base texture width. */
+   LLVMValueRef
+   (*width)( struct lp_sampler_dynamic_state *state,
+             LLVMBuilderRef builder,
+             unsigned unit);
+
+   /** Obtain the base texture height. */
+   LLVMValueRef
+   (*height)( struct lp_sampler_dynamic_state *state,
+              LLVMBuilderRef builder,
+              unsigned unit);
+
+   LLVMValueRef
+   (*stride)( struct lp_sampler_dynamic_state *state,
+              LLVMBuilderRef builder,
+              unsigned unit);
+
+   LLVMValueRef
+   (*data_ptr)( struct lp_sampler_dynamic_state *state,
+                LLVMBuilderRef builder,
+                unsigned unit);
+
+};
+
+
+/**
+ * Derive the sampler static state.
+ */
+void
+lp_sampler_static_state(struct lp_sampler_static_state *state,
+                        const struct pipe_texture *texture,
+                        const struct pipe_sampler_state *sampler);
+
+
+void
+lp_build_sample_soa(LLVMBuilderRef builder,
+                    const struct lp_sampler_static_state *static_state,
+                    struct lp_sampler_dynamic_state *dynamic_state,
+                    union lp_type fp_type,
+                    unsigned unit,
+                    unsigned num_coords,
+                    const LLVMValueRef *coords,
+                    LLVMValueRef lodbias,
+                    LLVMValueRef *texel);
+
+
+
+#endif /* LP_BLD_SAMPLE_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c
new file mode 100644
index 0000000000..2913b17d3f
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c
@@ -0,0 +1,404 @@
+/**************************************************************************
+ *
+ * Copyright 2009 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 VMWARE 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.
+ *
+ **************************************************************************/
+
+/**
+ * @file
+ * Texture sampling.
+ *
+ * @author Jose Fonseca <jfonseca@vmware.com>
+ */
+
+#include "pipe/p_defines.h"
+#include "pipe/p_state.h"
+#include "util/u_debug.h"
+#include "util/u_memory.h"
+#include "util/u_math.h"
+#include "util/u_format.h"
+#include "lp_bld_debug.h"
+#include "lp_bld_type.h"
+#include "lp_bld_const.h"
+#include "lp_bld_arit.h"
+#include "lp_bld_logic.h"
+#include "lp_bld_swizzle.h"
+#include "lp_bld_format.h"
+#include "lp_bld_sample.h"
+
+
+void
+lp_sampler_static_state(struct lp_sampler_static_state *state,
+                        const struct pipe_texture *texture,
+                        const struct pipe_sampler_state *sampler)
+{
+   memset(state, 0, sizeof *state);
+
+   if(!texture)
+      return;
+
+   if(!sampler)
+      return;
+
+   state->format            = texture->format;
+   state->target            = texture->target;
+   state->pot_width         = util_is_pot(texture->width[0]);
+   state->pot_height        = util_is_pot(texture->height[0]);
+   state->pot_depth         = util_is_pot(texture->depth[0]);
+
+   state->wrap_s            = sampler->wrap_s;
+   state->wrap_t            = sampler->wrap_t;
+   state->wrap_r            = sampler->wrap_r;
+   state->min_img_filter    = sampler->min_img_filter;
+   state->min_mip_filter    = sampler->min_mip_filter;
+   state->mag_img_filter    = sampler->mag_img_filter;
+   if(sampler->compare_mode) {
+      state->compare_mode      = sampler->compare_mode;
+      state->compare_func      = sampler->compare_func;
+   }
+   state->normalized_coords = sampler->normalized_coords;
+   state->prefilter         = sampler->prefilter;
+}
+
+
+
+/**
+ * Keep all information for sampling code generation in a single place.
+ */
+struct lp_build_sample_context
+{
+   LLVMBuilderRef builder;
+
+   const struct lp_sampler_static_state *static_state;
+
+   struct lp_sampler_dynamic_state *dynamic_state;
+
+   const struct util_format_description *format_desc;
+
+   /** Incoming coordinates type and build context */
+   union lp_type coord_type;
+   struct lp_build_context coord_bld;
+
+   /** Integer coordinates */
+   union lp_type int_coord_type;
+   struct lp_build_context int_coord_bld;
+
+   /** Output texels type and build context */
+   union lp_type texel_type;
+   struct lp_build_context texel_bld;
+};
+
+
+static void
+lp_build_sample_texel(struct lp_build_sample_context *bld,
+                      LLVMValueRef x,
+                      LLVMValueRef y,
+                      LLVMValueRef y_stride,
+                      LLVMValueRef data_ptr,
+                      LLVMValueRef *texel)
+{
+   struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
+   LLVMValueRef x_stride;
+   LLVMValueRef offset;
+
+   x_stride = lp_build_const_scalar(bld->int_coord_type, bld->format_desc->block.bits/8);
+
+   if(bld->format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
+      LLVMValueRef x_lo, x_hi;
+      LLVMValueRef y_lo, y_hi;
+      LLVMValueRef x_stride_lo, x_stride_hi;
+      LLVMValueRef y_stride_lo, y_stride_hi;
+      LLVMValueRef x_offset_lo, x_offset_hi;
+      LLVMValueRef y_offset_lo, y_offset_hi;
+      LLVMValueRef offset_lo, offset_hi;
+
+      x_lo = LLVMBuildAnd(bld->builder, x, int_coord_bld->one, "");
+      y_lo = LLVMBuildAnd(bld->builder, y, int_coord_bld->one, "");
+
+      x_hi = LLVMBuildLShr(bld->builder, x, int_coord_bld->one, "");
+      y_hi = LLVMBuildLShr(bld->builder, y, int_coord_bld->one, "");
+
+      x_stride_lo = x_stride;
+      y_stride_lo = lp_build_const_scalar(bld->int_coord_type, 2*bld->format_desc->block.bits/8);
+
+      x_stride_hi = lp_build_const_scalar(bld->int_coord_type, 4*bld->format_desc->block.bits/8);
+      y_stride_hi = LLVMBuildShl(bld->builder, y_stride, int_coord_bld->one, "");
+
+      x_offset_lo = lp_build_mul(int_coord_bld, x_lo, x_stride_lo);
+      y_offset_lo = lp_build_mul(int_coord_bld, y_lo, y_stride_lo);
+      offset_lo = lp_build_add(int_coord_bld, x_offset_lo, y_offset_lo);
+
+      x_offset_hi = lp_build_mul(int_coord_bld, x_hi, x_stride_hi);
+      y_offset_hi = lp_build_mul(int_coord_bld, y_hi, y_stride_hi);
+      offset_hi = lp_build_add(int_coord_bld, x_offset_hi, y_offset_hi);
+
+      offset = lp_build_add(int_coord_bld, offset_hi, offset_lo);
+   }
+   else {
+      LLVMValueRef x_offset;
+      LLVMValueRef y_offset;
+
+      x_offset = lp_build_mul(int_coord_bld, x, x_stride);
+      y_offset = lp_build_mul(int_coord_bld, y, y_stride);
+
+      offset = lp_build_add(int_coord_bld, x_offset, y_offset);
+   }
+
+   lp_build_load_rgba_soa(bld->builder,
+                          bld->format_desc,
+                          bld->texel_type,
+                          data_ptr,
+                          offset,
+                          texel);
+}
+
+
+static LLVMValueRef
+lp_build_sample_wrap(struct lp_build_sample_context *bld,
+                     LLVMValueRef coord,
+                     LLVMValueRef length,
+                     boolean is_pot,
+                     unsigned wrap_mode)
+{
+   struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
+   LLVMValueRef length_minus_one;
+
+   length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
+
+   switch(wrap_mode) {
+   case PIPE_TEX_WRAP_REPEAT:
+      if(is_pot)
+         coord = LLVMBuildAnd(bld->builder, coord, length_minus_one, "");
+      else
+         /* Signed remainder won't give the right results for negative
+          * dividends but unsigned remainder does.*/
+         coord = LLVMBuildURem(bld->builder, coord, length, "");
+      break;
+
+   case PIPE_TEX_WRAP_CLAMP:
+      coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
+      coord = lp_build_min(int_coord_bld, coord, length_minus_one);
+      break;
+
+   case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
+   case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
+   case PIPE_TEX_WRAP_MIRROR_REPEAT:
+   case PIPE_TEX_WRAP_MIRROR_CLAMP:
+   case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
+   case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
+   default:
+      assert(0);
+   }
+
+   return coord;
+}
+
+
+static void
+lp_build_sample_2d_nearest_soa(struct lp_build_sample_context *bld,
+                               LLVMValueRef s,
+                               LLVMValueRef t,
+                               LLVMValueRef width,
+                               LLVMValueRef height,
+                               LLVMValueRef stride,
+                               LLVMValueRef data_ptr,
+                               LLVMValueRef *texel)
+{
+   LLVMValueRef x;
+   LLVMValueRef y;
+
+   x = lp_build_ifloor(&bld->coord_bld, s);
+   y = lp_build_ifloor(&bld->coord_bld, t);
+
+   x = lp_build_sample_wrap(bld, x, width,  bld->static_state->pot_width,  bld->static_state->wrap_s);
+   y = lp_build_sample_wrap(bld, y, height, bld->static_state->pot_height, bld->static_state->wrap_t);
+
+   lp_build_sample_texel(bld, x, y, stride, data_ptr, texel);
+}
+
+
+static void
+lp_build_sample_2d_linear_soa(struct lp_build_sample_context *bld,
+                              LLVMValueRef s,
+                              LLVMValueRef t,
+                              LLVMValueRef width,
+                              LLVMValueRef height,
+                              LLVMValueRef stride,
+                              LLVMValueRef data_ptr,
+                              LLVMValueRef *texel)
+{
+   LLVMValueRef half;
+   LLVMValueRef s_ipart;
+   LLVMValueRef t_ipart;
+   LLVMValueRef s_fpart;
+   LLVMValueRef t_fpart;
+   LLVMValueRef x0, x1;
+   LLVMValueRef y0, y1;
+   LLVMValueRef neighbors[2][2][4];
+   unsigned chan;
+
+   half = lp_build_const_scalar(bld->coord_type, 0.5);
+   s = lp_build_sub(&bld->coord_bld, s, half);
+   t = lp_build_sub(&bld->coord_bld, t, half);
+
+   s_ipart = lp_build_floor(&bld->coord_bld, s);
+   t_ipart = lp_build_floor(&bld->coord_bld, t);
+
+   s_fpart = lp_build_sub(&bld->coord_bld, s, s_ipart);
+   t_fpart = lp_build_sub(&bld->coord_bld, t, t_ipart);
+
+   x0 = lp_build_int(&bld->coord_bld, s_ipart);
+   y0 = lp_build_int(&bld->coord_bld, t_ipart);
+
+   x0 = lp_build_sample_wrap(bld, x0, width,  bld->static_state->pot_width,  bld->static_state->wrap_s);
+   y0 = lp_build_sample_wrap(bld, y0, height, bld->static_state->pot_height, bld->static_state->wrap_t);
+
+   x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one);
+   y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one);
+
+   x1 = lp_build_sample_wrap(bld, x1, width,  bld->static_state->pot_width,  bld->static_state->wrap_s);
+   y1 = lp_build_sample_wrap(bld, y1, height, bld->static_state->pot_height, bld->static_state->wrap_t);
+
+   lp_build_sample_texel(bld, x0, y0, stride, data_ptr, neighbors[0][0]);
+   lp_build_sample_texel(bld, x1, y0, stride, data_ptr, neighbors[0][1]);
+   lp_build_sample_texel(bld, x0, y1, stride, data_ptr, neighbors[1][0]);
+   lp_build_sample_texel(bld, x1, y1, stride, data_ptr, neighbors[1][1]);
+
+   /* TODO: Don't interpolate missing channels */
+   for(chan = 0; chan < 4; ++chan) {
+      texel[chan] = lp_build_lerp_2d(&bld->texel_bld,
+                                     s_fpart, t_fpart,
+                                     neighbors[0][0][chan],
+                                     neighbors[0][1][chan],
+                                     neighbors[1][0][chan],
+                                     neighbors[1][1][chan]);
+   }
+}
+
+
+static void
+lp_build_sample_compare(struct lp_build_sample_context *bld,
+                        LLVMValueRef p,
+                        LLVMValueRef *texel)
+{
+   struct lp_build_context *texel_bld = &bld->texel_bld;
+   LLVMValueRef res;
+   unsigned chan;
+
+   if(!bld->static_state->compare_mode)
+      return;
+
+   /* TODO: Compare before swizzling, to avoid redundant computations */
+   res = NULL;
+   for(chan = 0; chan < 4; ++chan) {
+      LLVMValueRef cmp;
+      cmp = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]);
+      cmp = lp_build_select(texel_bld, cmp, texel_bld->one, texel_bld->zero);
+
+      if(res)
+         res = lp_build_add(texel_bld, res, cmp);
+      else
+         res = cmp;
+   }
+
+   assert(res);
+   res = lp_build_mul(texel_bld, res, lp_build_const_scalar(texel_bld->type, 0.25));
+
+   /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
+   for(chan = 0; chan < 3; ++chan)
+      texel[chan] = res;
+   texel[3] = texel_bld->one;
+}
+
+
+void
+lp_build_sample_soa(LLVMBuilderRef builder,
+                    const struct lp_sampler_static_state *static_state,
+                    struct lp_sampler_dynamic_state *dynamic_state,
+                    union lp_type type,
+                    unsigned unit,
+                    unsigned num_coords,
+                    const LLVMValueRef *coords,
+                    LLVMValueRef lodbias,
+                    LLVMValueRef *texel)
+{
+   struct lp_build_sample_context bld;
+   LLVMValueRef width;
+   LLVMValueRef height;
+   LLVMValueRef stride;
+   LLVMValueRef data_ptr;
+   LLVMValueRef s;
+   LLVMValueRef t;
+   LLVMValueRef p;
+
+   /* Setup our build context */
+   memset(&bld, 0, sizeof bld);
+   bld.builder = builder;
+   bld.static_state = static_state;
+   bld.dynamic_state = dynamic_state;
+   bld.format_desc = util_format_description(static_state->format);
+   bld.coord_type = type;
+   bld.int_coord_type = lp_int_type(type);
+   bld.texel_type = type;
+   lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
+   lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
+   lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
+
+   /* Get the dynamic state */
+   width = dynamic_state->width(dynamic_state, builder, unit);
+   height = dynamic_state->height(dynamic_state, builder, unit);
+   stride = dynamic_state->stride(dynamic_state, builder, unit);
+   data_ptr = dynamic_state->data_ptr(dynamic_state, builder, unit);
+
+   s = coords[0];
+   t = coords[1];
+   p = coords[2];
+
+   width = lp_build_broadcast_scalar(&bld.int_coord_bld, width);
+   height = lp_build_broadcast_scalar(&bld.int_coord_bld, height);
+   stride = lp_build_broadcast_scalar(&bld.int_coord_bld, stride);
+
+   if(static_state->target == PIPE_TEXTURE_1D)
+      t = bld.coord_bld.zero;
+
+   if(static_state->normalized_coords) {
+      LLVMTypeRef coord_vec_type = lp_build_vec_type(bld.coord_type);
+      LLVMValueRef fp_width = LLVMBuildSIToFP(builder, width, coord_vec_type, "");
+      LLVMValueRef fp_height = LLVMBuildSIToFP(builder, height, coord_vec_type, "");
+      s = lp_build_mul(&bld.coord_bld, s, fp_width);
+      t = lp_build_mul(&bld.coord_bld, t, fp_height);
+   }
+
+   switch (static_state->min_img_filter) {
+   case PIPE_TEX_FILTER_NEAREST:
+      lp_build_sample_2d_nearest_soa(&bld, s, t, width, height, stride, data_ptr, texel);
+      break;
+   case PIPE_TEX_FILTER_LINEAR:
+   case PIPE_TEX_FILTER_ANISO:
+      lp_build_sample_2d_linear_soa(&bld, s, t, width, height, stride, data_ptr, texel);
+      break;
+   }
+
+   lp_build_sample_compare(&bld, p, texel);
+}
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_struct.c b/src/gallium/drivers/llvmpipe/lp_bld_struct.c
index 14d2b10df9..3998ac374f 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_struct.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_struct.c
@@ -42,17 +42,30 @@
 
 
 LLVMValueRef
+lp_build_struct_get_ptr(LLVMBuilderRef builder,
+                        LLVMValueRef ptr,
+                        unsigned member,
+                        const char *name)
+{
+   LLVMValueRef indices[2];
+   LLVMValueRef member_ptr;
+   indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
+   indices[1] = LLVMConstInt(LLVMInt32Type(), member, 0);
+   member_ptr = LLVMBuildGEP(builder, ptr, indices, Elements(indices), "");
+   lp_build_name(member_ptr, "%s.%s_ptr", LLVMGetValueName(ptr), name);
+   return member_ptr;
+}
+
+
+LLVMValueRef
 lp_build_struct_get(LLVMBuilderRef builder,
                     LLVMValueRef ptr,
                     unsigned member,
                     const char *name)
 {
-   LLVMValueRef indices[2];
    LLVMValueRef member_ptr;
    LLVMValueRef res;
-   indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
-   indices[1] = LLVMConstInt(LLVMInt32Type(), member, 0);
-   member_ptr = LLVMBuildGEP(builder, ptr, indices, Elements(indices), "");
+   member_ptr = lp_build_struct_get_ptr(builder, ptr, member, name);
    res = LLVMBuildLoad(builder, member_ptr, "");
    lp_build_name(res, "%s.%s", LLVMGetValueName(ptr), name);
    return res;
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_struct.h b/src/gallium/drivers/llvmpipe/lp_bld_struct.h
index cbefdc9f81..740392f561 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_struct.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_struct.h
@@ -53,6 +53,18 @@
              offsetof(_ctype, _cmember))
 
 
+/**
+ * Get value pointer to a structure member.
+ */
+LLVMValueRef
+lp_build_struct_get_ptr(LLVMBuilderRef builder,
+                        LLVMValueRef ptr,
+                        unsigned member,
+                        const char *name);
+
+/**
+ * Get the value of a structure member.
+ */
 LLVMValueRef
 lp_build_struct_get(LLVMBuilderRef builder,
                     LLVMValueRef ptr,
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h b/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h
index 912db24aec..10c251c416 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h
@@ -44,14 +44,30 @@ struct lp_build_context;
 struct lp_build_mask_context;
 
 
-typedef void
-(*lp_emit_fetch_texel_soa_callback)( LLVMBuilderRef builder,
-                                void *context,
-                                unsigned unit,
-                                unsigned num_coords,
-                                const LLVMValueRef *coords,
-                                LLVMValueRef lodbias,
-                                LLVMValueRef *texel);
+/**
+ * Sampler code generation interface.
+ *
+ * Although texture sampling is a requirement for TGSI translation, it is
+ * a very different problem with several different approaches to it. This
+ * structure establishes an interface for texture sampling code generation, so
+ * that we can easily use different texture sampling strategies.
+ */
+struct lp_build_sampler_soa
+{
+   void
+   (*destroy)( struct lp_build_sampler_soa *sampler );
+
+   void
+   (*emit_fetch_texel)( struct lp_build_sampler_soa *sampler,
+                        LLVMBuilderRef builder,
+                        union lp_type type,
+                        unsigned unit,
+                        unsigned num_coords,
+                        const LLVMValueRef *coords,
+                        LLVMValueRef lodbias,
+                        LLVMValueRef *texel);
+};
+
 
 void
 lp_build_tgsi_soa(LLVMBuilderRef builder,
@@ -62,8 +78,7 @@ lp_build_tgsi_soa(LLVMBuilderRef builder,
                   const LLVMValueRef *pos,
                   const LLVMValueRef (*inputs)[4],
                   LLVMValueRef (*outputs)[4],
-                  lp_emit_fetch_texel_soa_callback emit_fetch_texel,
-                  void *emit_fetch_texel_context);
+                  struct lp_build_sampler_soa *sampler);
 
 
 #endif /* LP_BLD_TGSI_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c
index bce26607f9..3ce379de12 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c
@@ -88,8 +88,7 @@ struct lp_build_tgsi_soa_context
    const LLVMValueRef (*inputs)[NUM_CHANNELS];
    LLVMValueRef (*outputs)[NUM_CHANNELS];
 
-   lp_emit_fetch_texel_soa_callback emit_fetch_texel;
-   void *emit_fetch_texel_context;
+   struct lp_build_sampler_soa *sampler;
 
    LLVMValueRef immediates[LP_MAX_IMMEDIATES][NUM_CHANNELS];
    LLVMValueRef temps[LP_MAX_TEMPS][NUM_CHANNELS];
@@ -289,8 +288,11 @@ emit_tex( struct lp_build_tgsi_soa_context *bld,
          coords[i] = lp_build_mul(&bld->base, coords[i], oow);
    }
 
-   bld->emit_fetch_texel(bld->base.builder, bld->emit_fetch_texel_context,
-                         unit, num_coords, coords, lodbias, texel);
+   bld->sampler->emit_fetch_texel(bld->sampler,
+                                  bld->base.builder,
+                                  bld->base.type,
+                                  unit, num_coords, coords, lodbias,
+                                  texel);
 
    FOR_EACH_DST0_ENABLED_CHANNEL( inst, i ) {
       emit_store( bld, inst, 0, i, texel[i] );
@@ -687,10 +689,6 @@ emit_instruction(
       return 0;
       break;
 
-   case TGSI_OPCODE_CND0:
-      return 0;
-      break;
-
    case TGSI_OPCODE_DP2A:
       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );  /* xmm0 = src[0].x */
       tmp1 = emit_fetch( bld, inst, 1, CHAN_X );  /* xmm1 = src[1].x */
@@ -1287,8 +1285,7 @@ lp_build_tgsi_soa(LLVMBuilderRef builder,
                   const LLVMValueRef *pos,
                   const LLVMValueRef (*inputs)[NUM_CHANNELS],
                   LLVMValueRef (*outputs)[NUM_CHANNELS],
-                  lp_emit_fetch_texel_soa_callback emit_fetch_texel,
-                  void *emit_fetch_texel_context)
+                  struct lp_build_sampler_soa *sampler)
 {
    struct lp_build_tgsi_soa_context bld;
    struct tgsi_parse_context parse;
@@ -1303,8 +1300,7 @@ lp_build_tgsi_soa(LLVMBuilderRef builder,
    bld.inputs = inputs;
    bld.outputs = outputs;
    bld.consts_ptr = consts_ptr;
-   bld.emit_fetch_texel = emit_fetch_texel;
-   bld.emit_fetch_texel_context = emit_fetch_texel_context;
+   bld.sampler = sampler;
 
    tgsi_parse_init( &parse, tokens );
 
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_type.c b/src/gallium/drivers/llvmpipe/lp_bld_type.c
index 8e0026fd97..577644b7ab 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_type.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_type.c
@@ -157,6 +157,17 @@ lp_build_int_vec_type(union lp_type type)
 }
 
 
+union lp_type
+lp_int_type(union lp_type type)
+{
+   union lp_type int_type;
+   int_type.value = 0;
+   int_type.width = type.width;
+   int_type.length = type.length;
+   return int_type;
+}
+
+
 void
 lp_build_context_init(struct lp_build_context *bld,
                       LLVMBuilderRef builder,
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_type.h b/src/gallium/drivers/llvmpipe/lp_bld_type.h
index 3ce566be64..9933e0b45c 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_type.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_type.h
@@ -165,6 +165,10 @@ LLVMTypeRef
 lp_build_int_vec_type(union lp_type type);
 
 
+union lp_type
+lp_int_type(union lp_type type);
+
+
 void
 lp_build_context_init(struct lp_build_context *bld,
                       LLVMBuilderRef builder,
diff --git a/src/gallium/drivers/llvmpipe/lp_jit.c b/src/gallium/drivers/llvmpipe/lp_jit.c
index d288460a1b..9465f763d5 100644
--- a/src/gallium/drivers/llvmpipe/lp_jit.c
+++ b/src/gallium/drivers/llvmpipe/lp_jit.c
@@ -44,15 +44,47 @@
 static void
 lp_jit_init_globals(struct llvmpipe_screen *screen)
 {
-   /* struct lp_jit_context */
+   LLVMTypeRef texture_type;
+
+   /* struct lp_jit_texture */
    {
       LLVMTypeRef elem_types[4];
+
+      elem_types[LP_JIT_TEXTURE_WIDTH]  = LLVMInt32Type();
+      elem_types[LP_JIT_TEXTURE_HEIGHT] = LLVMInt32Type();
+      elem_types[LP_JIT_TEXTURE_STRIDE] = LLVMInt32Type();
+      elem_types[LP_JIT_TEXTURE_DATA]   = LLVMPointerType(LLVMInt8Type(), 0);
+
+      texture_type = LLVMStructType(elem_types, Elements(elem_types), 0);
+
+      LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, width,
+                             screen->target, texture_type,
+                             LP_JIT_TEXTURE_WIDTH);
+      LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, height,
+                             screen->target, texture_type,
+                             LP_JIT_TEXTURE_HEIGHT);
+      LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, stride,
+                             screen->target, texture_type,
+                             LP_JIT_TEXTURE_STRIDE);
+      LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, data,
+                             screen->target, texture_type,
+                             LP_JIT_TEXTURE_DATA);
+      LP_CHECK_STRUCT_SIZE(struct lp_jit_texture,
+                           screen->target, texture_type);
+
+      LLVMAddTypeName(screen->module, "texture", texture_type);
+   }
+
+   /* struct lp_jit_context */
+   {
+      LLVMTypeRef elem_types[5];
       LLVMTypeRef context_type;
 
       elem_types[0] = LLVMPointerType(LLVMFloatType(), 0); /* constants */
       elem_types[1] = LLVMPointerType(LLVMInt8Type(), 0);  /* samplers */
       elem_types[2] = LLVMFloatType();                     /* alpha_ref_value */
       elem_types[3] = LLVMPointerType(LLVMInt8Type(), 0);  /* blend_color */
+      elem_types[4] = LLVMArrayType(texture_type, PIPE_MAX_SAMPLERS); /* textures */
 
       context_type = LLVMStructType(elem_types, Elements(elem_types), 0);
 
@@ -64,6 +96,9 @@ lp_jit_init_globals(struct llvmpipe_screen *screen)
                              screen->target, context_type, 2);
       LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, blend_color,
                              screen->target, context_type, 3);
+      LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, textures,
+                             screen->target, context_type,
+                             LP_JIT_CONTEXT_TEXTURES_INDEX);
       LP_CHECK_STRUCT_SIZE(struct lp_jit_context,
                            screen->target, context_type);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_jit.h b/src/gallium/drivers/llvmpipe/lp_jit.h
index a7fb60f9f5..c3e3e1af67 100644
--- a/src/gallium/drivers/llvmpipe/lp_jit.h
+++ b/src/gallium/drivers/llvmpipe/lp_jit.h
@@ -38,11 +38,31 @@
 
 #include "lp_bld_struct.h"
 
+#include "pipe/p_state.h"
+
 
 struct tgsi_sampler;
 struct llvmpipe_screen;
 
 
+struct lp_jit_texture
+{
+   uint32_t width;
+   uint32_t height;
+   uint32_t stride;
+   const void *data;
+};
+
+
+enum {
+   LP_JIT_TEXTURE_WIDTH = 0,
+   LP_JIT_TEXTURE_HEIGHT,
+   LP_JIT_TEXTURE_STRIDE,
+   LP_JIT_TEXTURE_DATA
+};
+
+
+
 /**
  * This structure is passed directly to the generated fragment shader.
  *
@@ -65,6 +85,8 @@ struct lp_jit_context
 
    /* TODO: blend constant color */
    uint8_t *blend_color;
+
+   struct lp_jit_texture textures[PIPE_MAX_SAMPLERS];
 };
 
 
@@ -80,6 +102,11 @@ struct lp_jit_context
 #define lp_jit_context_blend_color(_builder, _ptr) \
    lp_build_struct_get(_builder, _ptr, 3, "blend_color")
 
+#define LP_JIT_CONTEXT_TEXTURES_INDEX 4
+
+#define lp_jit_context_textures(_builder, _ptr) \
+   lp_build_struct_get_ptr(_builder, _ptr, LP_JIT_CONTEXT_TEXTURES_INDEX, "textures")
+
 
 typedef void
 (*lp_jit_frag_func)(struct lp_jit_context *context,
diff --git a/src/gallium/drivers/llvmpipe/lp_setup.c b/src/gallium/drivers/llvmpipe/lp_setup.c
index d145f6d6bb..f9e254efca 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup.c
@@ -44,6 +44,7 @@
 #include "pipe/p_thread.h"
 #include "util/u_math.h"
 #include "util/u_memory.h"
+#include "lp_bld_debug.h"
 #include "lp_tile_cache.h"
 #include "lp_tile_soa.h"
 
@@ -164,10 +165,12 @@ shade_quads(struct llvmpipe_context *llvmpipe,
 
    /* TODO: blend color */
 
-   assert((((uintptr_t)mask) & 0xf) == 0);
-   assert((((uintptr_t)depth) & 0xf) == 0);
-   assert((((uintptr_t)color) & 0xf) == 0);
-   assert((((uintptr_t)llvmpipe->jit_context.blend_color) & 0xf) == 0);
+   /* XXX: This will most likely fail on 32bit x86 without -mstackrealign */
+   assert(lp_check_alignment(mask, 16));
+
+   assert(lp_check_alignment(depth, 16));
+   assert(lp_check_alignment(color, 16));
+   assert(lp_check_alignment(llvmpipe->jit_context.blend_color, 16));
 
    /* run shader */
    fs->current->jit_function( &llvmpipe->jit_context,
diff --git a/src/gallium/drivers/llvmpipe/lp_state.h b/src/gallium/drivers/llvmpipe/lp_state.h
index fb10329887..7b26ce61a3 100644
--- a/src/gallium/drivers/llvmpipe/lp_state.h
+++ b/src/gallium/drivers/llvmpipe/lp_state.h
@@ -36,6 +36,7 @@
 #include "pipe/p_state.h"
 #include "tgsi/tgsi_scan.h"
 #include "lp_jit.h"
+#include "lp_bld_sample.h" /* for struct lp_sampler_static_state */
 
 
 #define LP_NEW_VIEWPORT      0x1
@@ -57,16 +58,20 @@
 
 struct tgsi_sampler;
 struct vertex_info;
-
+struct pipe_context;
+struct llvmpipe_context;
 
 struct lp_fragment_shader;
 
 
 struct lp_fragment_shader_variant_key
 {
+   enum pipe_format zsbuf_format;
    struct pipe_depth_state depth;
    struct pipe_alpha_state alpha;
    struct pipe_blend_state blend;
+
+   struct lp_sampler_static_state sampler[PIPE_MAX_SAMPLERS];
 };
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_state_derived.c b/src/gallium/drivers/llvmpipe/lp_state_derived.c
index 6fbb057937..e87976b9f3 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_derived.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_derived.c
@@ -250,7 +250,9 @@ void llvmpipe_update_derived( struct llvmpipe_context *llvmpipe )
 
    if (llvmpipe->dirty & (LP_NEW_FS |
                           LP_NEW_BLEND |
-                          LP_NEW_DEPTH_STENCIL_ALPHA))
+                          LP_NEW_DEPTH_STENCIL_ALPHA |
+                          LP_NEW_SAMPLER |
+                          LP_NEW_TEXTURE))
       llvmpipe_update_fs( llvmpipe );
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.c b/src/gallium/drivers/llvmpipe/lp_state_fs.c
index 94170bd716..618cf1ffb8 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_fs.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs.c
@@ -85,6 +85,7 @@
 #include "lp_context.h"
 #include "lp_state.h"
 #include "lp_quad.h"
+#include "lp_tex_sample.h"
 
 
 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
@@ -130,9 +131,8 @@ generate_pos0(LLVMBuilderRef builder,
  * Generate the depth test.
  */
 static void
-generate_depth(struct llvmpipe_context *lp,
-               LLVMBuilderRef builder,
-               const struct pipe_depth_state *state,
+generate_depth(LLVMBuilderRef builder,
+               const struct lp_fragment_shader_variant_key *key,
                union lp_type src_type,
                struct lp_build_mask_context *mask,
                LLVMValueRef src,
@@ -141,10 +141,10 @@ generate_depth(struct llvmpipe_context *lp,
    const struct util_format_description *format_desc;
    union lp_type dst_type;
 
-   if(!lp->framebuffer.zsbuf)
+   if(!key->depth.enabled)
       return;
 
-   format_desc = util_format_description(lp->framebuffer.zsbuf->format);
+   format_desc = util_format_description(key->zsbuf_format);
    assert(format_desc);
 
    /* Pick the depth type. */
@@ -164,7 +164,7 @@ generate_depth(struct llvmpipe_context *lp,
 #endif
 
    lp_build_depth_test(builder,
-                       state,
+                       &key->depth,
                        dst_type,
                        format_desc,
                        mask,
@@ -173,107 +173,6 @@ generate_depth(struct llvmpipe_context *lp,
 }
 
 
-struct build_fetch_texel_context
-{
-   LLVMValueRef context_ptr;
-
-   LLVMValueRef samplers_ptr;
-
-   /** Coords/texels store */
-   LLVMValueRef store_ptr;
-};
-
-
-void PIPE_CDECL
-lp_fetch_texel_soa( struct tgsi_sampler **samplers,
-                    uint32_t unit,
-                    float *store )
-{
-   struct tgsi_sampler *sampler = samplers[unit];
-
-#if 0
-   uint j;
-
-   debug_printf("%s sampler: %p (%p) store: %p\n",
-                __FUNCTION__,
-                sampler, *sampler,
-                store );
-
-   debug_printf("lodbias %f\n", store[12]);
-
-   for (j = 0; j < 4; j++)
-      debug_printf("sample %d texcoord %f %f\n",
-                   j,
-                   store[0+j],
-                   store[4+j]);
-#endif
-
-   {
-      float rgba[NUM_CHANNELS][QUAD_SIZE];
-      sampler->get_samples(sampler,
-                           &store[0],
-                           &store[4],
-                           &store[8],
-                           0.0f, /*store[12],  lodbias */
-                           rgba);
-      memcpy(store, rgba, sizeof rgba);
-   }
-
-#if 0
-   for (j = 0; j < 4; j++)
-      debug_printf("sample %d result %f %f %f %f\n",
-                   j,
-                   store[0+j],
-                   store[4+j],
-                   store[8+j],
-                   store[12+j]);
-#endif
-}
-
-
-static void
-emit_fetch_texel( LLVMBuilderRef builder,
-                  void *context,
-                  unsigned unit,
-                  unsigned num_coords,
-                  const LLVMValueRef *coords,
-                  LLVMValueRef lodbias,
-                  LLVMValueRef *texel)
-{
-   struct build_fetch_texel_context *bld = context;
-   LLVMTypeRef vec_type = LLVMTypeOf(coords[0]);
-   LLVMValueRef args[3];
-   unsigned i;
-
-   if(!bld->samplers_ptr)
-      bld->samplers_ptr = lp_jit_context_samplers(builder, bld->context_ptr);
-
-   if(!bld->store_ptr)
-      bld->store_ptr = LLVMBuildArrayAlloca(builder,
-                                            vec_type,
-                                            LLVMConstInt(LLVMInt32Type(), 4, 0),
-                                            "texel_store");
-
-   for (i = 0; i < num_coords; i++) {
-      LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
-      LLVMValueRef coord_ptr = LLVMBuildGEP(builder, bld->store_ptr, &index, 1, "");
-      LLVMBuildStore(builder, coords[i], coord_ptr);
-   }
-
-   args[0] = bld->samplers_ptr;
-   args[1] = LLVMConstInt(LLVMInt32Type(), unit, 0);
-   args[2] = bld->store_ptr;
-
-   lp_build_intrinsic(builder, "fetch_texel", LLVMVoidType(), args, 3);
-
-   for (i = 0; i < NUM_CHANNELS; ++i) {
-      LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
-      LLVMValueRef texel_ptr = LLVMBuildGEP(builder, bld->store_ptr, &index, 1, "");
-      texel[i] = LLVMBuildLoad(builder, texel_ptr, "");
-   }
-}
-
-
 /**
  * Generate the fragment shader, depth/stencil test, and alpha tests.
  */
@@ -286,7 +185,7 @@ generate_fs(struct llvmpipe_context *lp,
             LLVMValueRef context_ptr,
             unsigned i,
             const struct lp_build_interp_soa_context *interp,
-            struct build_fetch_texel_context *sampler,
+            struct lp_build_sampler_soa *sampler,
             LLVMValueRef *pmask,
             LLVMValueRef *color,
             LLVMValueRef depth_ptr)
@@ -298,6 +197,7 @@ generate_fs(struct llvmpipe_context *lp,
    LLVMValueRef consts_ptr;
    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
    LLVMValueRef z = interp->pos[2];
+   struct lp_build_flow_context *flow;
    struct lp_build_mask_context mask;
    boolean early_depth_test;
    unsigned attrib;
@@ -309,25 +209,35 @@ generate_fs(struct llvmpipe_context *lp,
 
    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 
-   lp_build_mask_begin(&mask, builder, type, *pmask);
+   flow = lp_build_flow_create(builder);
+
+   memset(outputs, 0, sizeof outputs);
+
+   lp_build_flow_scope_begin(flow);
+
+   /* Declare the color and z variables */
+   for(chan = 0; chan < NUM_CHANNELS; ++chan) {
+      color[chan] = LLVMGetUndef(vec_type);
+      lp_build_flow_scope_declare(flow, &color[chan]);
+   }
+   lp_build_flow_scope_declare(flow, &z);
+
+   lp_build_mask_begin(&mask, flow, type, *pmask);
 
    early_depth_test =
-      lp->depth_stencil->depth.enabled &&
-      lp->framebuffer.zsbuf &&
-      !lp->depth_stencil->alpha.enabled &&
-      !lp->fs->info.uses_kill &&
-      !lp->fs->info.writes_z;
+      key->depth.enabled &&
+      !key->alpha.enabled &&
+      !shader->info.uses_kill &&
+      !shader->info.writes_z;
 
    if(early_depth_test)
-      generate_depth(lp, builder, &key->depth,
+      generate_depth(builder, key,
                      type, &mask,
                      z, depth_ptr);
 
-   memset(outputs, 0, sizeof outputs);
-
    lp_build_tgsi_soa(builder, tokens, type, &mask,
                      consts_ptr, interp->pos, interp->inputs,
-                     outputs, emit_fetch_texel, sampler);
+                     outputs, sampler);
 
    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
@@ -368,12 +278,16 @@ generate_fs(struct llvmpipe_context *lp,
    }
 
    if(!early_depth_test)
-      generate_depth(lp, builder, &key->depth,
+      generate_depth(builder, key,
                      type, &mask,
                      z, depth_ptr);
 
    lp_build_mask_end(&mask);
 
+   lp_build_flow_scope_end(flow);
+
+   lp_build_flow_destroy(flow);
+
    *pmask = mask.value;
 
 }
@@ -392,6 +306,8 @@ generate_blend(const struct pipe_blend_state *blend,
                LLVMValueRef dst_ptr)
 {
    struct lp_build_context bld;
+   struct lp_build_flow_context *flow;
+   struct lp_build_mask_context mask_ctx;
    LLVMTypeRef vec_type;
    LLVMTypeRef int_vec_type;
    LLVMValueRef const_ptr;
@@ -400,11 +316,14 @@ generate_blend(const struct pipe_blend_state *blend,
    LLVMValueRef res[4];
    unsigned chan;
 
+   lp_build_context_init(&bld, builder, type);
+
+   flow = lp_build_flow_create(builder);
+   lp_build_mask_begin(&mask_ctx, flow, type, mask);
+
    vec_type = lp_build_vec_type(type);
    int_vec_type = lp_build_int_vec_type(type);
 
-   lp_build_context_init(&bld, builder, type);
-
    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
    const_ptr = LLVMBuildBitCast(builder, const_ptr,
                                 LLVMPointerType(vec_type, 0), "");
@@ -422,11 +341,16 @@ generate_blend(const struct pipe_blend_state *blend,
    lp_build_blend_soa(builder, blend, type, src, dst, con, res);
 
    for(chan = 0; chan < 4; ++chan) {
-      LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
-      lp_build_name(res[chan], "res.%c", "rgba"[chan]);
-      res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
-      LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
+      if(blend->colormask & (1 << chan)) {
+         LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
+         lp_build_name(res[chan], "res.%c", "rgba"[chan]);
+         res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
+         LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
+      }
    }
+
+   lp_build_mask_end(&mask_ctx);
+   lp_build_flow_destroy(flow);
 }
 
 
@@ -462,7 +386,7 @@ generate_fragment(struct llvmpipe_context *lp,
    LLVMBuilderRef builder;
    LLVMValueRef x0;
    LLVMValueRef y0;
-   struct build_fetch_texel_context sampler;
+   struct lp_build_sampler_soa *sampler;
    struct lp_build_interp_soa_context interp;
    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
    LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
@@ -586,8 +510,13 @@ generate_fragment(struct llvmpipe_context *lp,
                             a0_ptr, dadx_ptr, dady_ptr,
                             x0, y0, 2, 0);
 
-   memset(&sampler, 0, sizeof sampler);
-   sampler.context_ptr = context_ptr;
+#if 0
+   /* C texture sampling */
+   sampler = lp_c_sampler_soa_create(context_ptr);
+#else
+   /* code generated texture sampling */
+   sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
+#endif
 
    for(i = 0; i < num_fs; ++i) {
       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
@@ -606,7 +535,7 @@ generate_fragment(struct llvmpipe_context *lp,
                   context_ptr,
                   i,
                   &interp,
-                  &sampler,
+                  sampler,
                   &fs_mask[i],
                   out_color,
                   depth_ptr_i);
@@ -615,6 +544,8 @@ generate_fragment(struct llvmpipe_context *lp,
          fs_out_color[chan][i] = out_color[chan];
    }
 
+   sampler->destroy(sampler);
+
    /* 
     * Convert the fs's output color and mask to fit to the blending type. 
     */
@@ -765,18 +696,45 @@ llvmpipe_set_constant_buffer(struct pipe_context *pipe,
  */
 static void
 make_variant_key(struct llvmpipe_context *lp,
+                 struct lp_fragment_shader *shader,
                  struct lp_fragment_shader_variant_key *key)
 {
+   unsigned i;
+
    memset(key, 0, sizeof *key);
 
-   memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
+   if(lp->framebuffer.zsbuf &&
+      lp->depth_stencil->depth.enabled) {
+      key->zsbuf_format = lp->framebuffer.zsbuf->format;
+      memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
+   }
 
    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
    if(key->alpha.enabled)
       key->alpha.func = lp->depth_stencil->alpha.func;
    /* alpha.ref_value is passed in jit_context */
 
-   memcpy(&key->blend, lp->blend, sizeof key->blend);
+   if(lp->framebuffer.cbufs[0]) {
+      const struct util_format_description *format_desc;
+      unsigned chan;
+
+      memcpy(&key->blend, lp->blend, sizeof key->blend);
+
+      format_desc = util_format_description(lp->framebuffer.cbufs[0]->format);
+      assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
+             format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
+
+      /* mask out color channels not present in the color buffer */
+      for(chan = 0; chan < 4; ++chan) {
+         enum util_format_swizzle swizzle = format_desc->swizzle[chan];
+         if(swizzle > 4)
+            key->blend.colormask &= ~(1 << chan);
+      }
+   }
+
+   for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
+      if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
+         lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]);
 }
 
 
@@ -787,7 +745,7 @@ llvmpipe_update_fs(struct llvmpipe_context *lp)
    struct lp_fragment_shader_variant_key key;
    struct lp_fragment_shader_variant *variant;
 
-   make_variant_key(lp, &key);
+   make_variant_key(lp, shader, &key);
 
    variant = shader->variants;
    while(variant) {
diff --git a/src/gallium/drivers/llvmpipe/lp_state_sampler.c b/src/gallium/drivers/llvmpipe/lp_state_sampler.c
index 4fef541b1e..c69d90c723 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_sampler.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_sampler.c
@@ -98,6 +98,16 @@ llvmpipe_set_sampler_textures(struct pipe_context *pipe,
 
       pipe_texture_reference(&llvmpipe->texture[i], tex);
       lp_tex_tile_cache_set_texture(llvmpipe->tex_cache[i], tex);
+
+      if(tex) {
+         struct llvmpipe_texture *lp_tex = llvmpipe_texture(tex);
+         struct lp_jit_texture *jit_tex = &llvmpipe->jit_context.textures[i];
+         jit_tex->width = tex->width[0];
+         jit_tex->height = tex->height[0];
+         jit_tex->stride = lp_tex->stride[0];
+         if(!lp_tex->dt)
+            jit_tex->data = lp_tex->data;
+      }
    }
 
    llvmpipe->num_textures = num;
diff --git a/src/gallium/drivers/llvmpipe/lp_test_format.c b/src/gallium/drivers/llvmpipe/lp_test_format.c
index 1d192355ee..d8455e5649 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_format.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_format.c
@@ -119,7 +119,7 @@ add_load_rgba_test(LLVMModuleRef module,
 
    lp_build_loop_begin(builder, LLVMConstInt(LLVMInt32Type(), 1, 0), &loop);
 
-   rgba = lp_build_load_rgba(builder, format, ptr);
+   rgba = lp_build_load_rgba_aos(builder, format, ptr);
    LLVMBuildStore(builder, rgba, rgba_ptr);
 
    lp_build_loop_end(builder, LLVMConstInt(LLVMInt32Type(), 4, 0), NULL, &loop);
@@ -160,7 +160,7 @@ add_store_rgba_test(LLVMModuleRef module,
 
    rgba = LLVMBuildLoad(builder, rgba_ptr, "");
 
-   lp_build_store_rgba(builder, format, ptr, rgba);
+   lp_build_store_rgba_aos(builder, format, ptr, rgba);
 
    LLVMBuildRetVoid(builder);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_tex_cache.c b/src/gallium/drivers/llvmpipe/lp_tex_cache.c
index 23a94b5b0d..773e848242 100644
--- a/src/gallium/drivers/llvmpipe/lp_tex_cache.c
+++ b/src/gallium/drivers/llvmpipe/lp_tex_cache.c
@@ -154,7 +154,7 @@ lp_tex_tile_cache_validate_texture(struct llvmpipe_tex_tile_cache *tc)
       if (lpt->timestamp != tc->timestamp) {
          /* texture was modified, invalidate all cached tiles */
          uint i;
-         _debug_printf("INV %d %d\n", tc->timestamp, lpt->timestamp);
+         debug_printf("INV %d %d\n", tc->timestamp, lpt->timestamp);
          for (i = 0; i < NUM_ENTRIES; i++) {
             tc->entries[i].addr.bits.invalid = 1;
          }
diff --git a/src/gallium/drivers/llvmpipe/lp_tex_sample.h b/src/gallium/drivers/llvmpipe/lp_tex_sample.h
index 628ec3f1ef..9ad1bde956 100644
--- a/src/gallium/drivers/llvmpipe/lp_tex_sample.h
+++ b/src/gallium/drivers/llvmpipe/lp_tex_sample.h
@@ -29,10 +29,13 @@
 #define LP_TEX_SAMPLE_H
 
 
+#include <llvm-c/Core.h>
+
 #include "tgsi/tgsi_exec.h"
 
 
 struct llvmpipe_tex_tile_cache;
+struct lp_sampler_static_state;
 
 
 /**
@@ -75,4 +78,24 @@ lp_get_samples(struct tgsi_sampler *tgsi_sampler,
                float rgba[NUM_CHANNELS][QUAD_SIZE]);
 
 
+/**
+ * Texture sampling code generator that just calls lp_get_samples C function
+ * for the actual sampling computation.
+ *
+ * @param context_ptr LLVM value with the pointer to the struct lp_jit_context.
+ */
+struct lp_build_sampler_soa *
+lp_c_sampler_soa_create(LLVMValueRef context_ptr);
+
+
+/**
+ * Pure-LLVM texture sampling code generator.
+ *
+ * @param context_ptr LLVM value with the pointer to the struct lp_jit_context.
+ */
+struct lp_build_sampler_soa *
+lp_llvm_sampler_soa_create(const struct lp_sampler_static_state *key,
+                           LLVMValueRef context_ptr);
+
+
 #endif /* LP_TEX_SAMPLE_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_tex_sample.c b/src/gallium/drivers/llvmpipe/lp_tex_sample_c.c
index 94eb6dad5a..9a876f404d 100644
--- a/src/gallium/drivers/llvmpipe/lp_tex_sample.c
+++ b/src/gallium/drivers/llvmpipe/lp_tex_sample_c.c
@@ -1578,3 +1578,136 @@ out:
    tgsi_sampler->get_samples( tgsi_sampler, s, t, p, lodbias, rgba );
 }
 
+
+void PIPE_CDECL
+lp_fetch_texel_soa( struct tgsi_sampler **samplers,
+                    uint32_t unit,
+                    float *store )
+{
+   struct tgsi_sampler *sampler = samplers[unit];
+
+#if 0
+   uint j;
+
+   debug_printf("%s sampler: %p (%p) store: %p\n",
+                __FUNCTION__,
+                sampler, *sampler,
+                store );
+
+   debug_printf("lodbias %f\n", store[12]);
+
+   for (j = 0; j < 4; j++)
+      debug_printf("sample %d texcoord %f %f\n",
+                   j,
+                   store[0+j],
+                   store[4+j]);
+#endif
+
+   {
+      float rgba[NUM_CHANNELS][QUAD_SIZE];
+      sampler->get_samples(sampler,
+                           &store[0],
+                           &store[4],
+                           &store[8],
+                           0.0f, /*store[12],  lodbias */
+                           rgba);
+      memcpy(store, rgba, sizeof rgba);
+   }
+
+#if 0
+   for (j = 0; j < 4; j++)
+      debug_printf("sample %d result %f %f %f %f\n",
+                   j,
+                   store[0+j],
+                   store[4+j],
+                   store[8+j],
+                   store[12+j]);
+#endif
+}
+
+
+#include "lp_bld_type.h"
+#include "lp_bld_intr.h"
+#include "lp_bld_tgsi.h"
+
+
+struct lp_c_sampler_soa
+{
+   struct lp_build_sampler_soa base;
+
+   LLVMValueRef context_ptr;
+
+   LLVMValueRef samplers_ptr;
+
+   /** Coords/texels store */
+   LLVMValueRef store_ptr;
+};
+
+
+static void
+lp_c_sampler_soa_destroy(struct lp_build_sampler_soa *sampler)
+{
+   FREE(sampler);
+}
+
+
+static void
+lp_c_sampler_soa_emit_fetch_texel(struct lp_build_sampler_soa *_sampler,
+                                  LLVMBuilderRef builder,
+                                  union lp_type type,
+                                  unsigned unit,
+                                  unsigned num_coords,
+                                  const LLVMValueRef *coords,
+                                  LLVMValueRef lodbias,
+                                  LLVMValueRef *texel)
+{
+   struct lp_c_sampler_soa *sampler = (struct lp_c_sampler_soa *)_sampler;
+   LLVMTypeRef vec_type = LLVMTypeOf(coords[0]);
+   LLVMValueRef args[3];
+   unsigned i;
+
+   if(!sampler->samplers_ptr)
+      sampler->samplers_ptr = lp_jit_context_samplers(builder, sampler->context_ptr);
+
+   if(!sampler->store_ptr)
+      sampler->store_ptr = LLVMBuildArrayAlloca(builder,
+                                            vec_type,
+                                            LLVMConstInt(LLVMInt32Type(), 4, 0),
+                                            "texel_store");
+
+   for (i = 0; i < num_coords; i++) {
+      LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
+      LLVMValueRef coord_ptr = LLVMBuildGEP(builder, sampler->store_ptr, &index, 1, "");
+      LLVMBuildStore(builder, coords[i], coord_ptr);
+   }
+
+   args[0] = sampler->samplers_ptr;
+   args[1] = LLVMConstInt(LLVMInt32Type(), unit, 0);
+   args[2] = sampler->store_ptr;
+
+   lp_build_intrinsic(builder, "fetch_texel", LLVMVoidType(), args, 3);
+
+   for (i = 0; i < NUM_CHANNELS; ++i) {
+      LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
+      LLVMValueRef texel_ptr = LLVMBuildGEP(builder, sampler->store_ptr, &index, 1, "");
+      texel[i] = LLVMBuildLoad(builder, texel_ptr, "");
+   }
+}
+
+
+struct lp_build_sampler_soa *
+lp_c_sampler_soa_create(LLVMValueRef context_ptr)
+{
+   struct lp_c_sampler_soa *sampler;
+
+   sampler = CALLOC_STRUCT(lp_c_sampler_soa);
+   if(!sampler)
+      return NULL;
+
+   sampler->base.destroy = lp_c_sampler_soa_destroy;
+   sampler->base.emit_fetch_texel = lp_c_sampler_soa_emit_fetch_texel;
+   sampler->context_ptr = context_ptr;
+
+   return &sampler->base;
+}
+
diff --git a/src/gallium/drivers/llvmpipe/lp_tex_sample_llvm.c b/src/gallium/drivers/llvmpipe/lp_tex_sample_llvm.c
new file mode 100644
index 0000000000..7d31705d01
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_tex_sample_llvm.c
@@ -0,0 +1,196 @@
+/**************************************************************************
+ * 
+ * Copyright 2009 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 VMWARE 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.
+ * 
+ **************************************************************************/
+
+/**
+ * Texture sampling code generation
+ *
+ * This file is nothing more than ugly glue between three largely independent
+ * entities:
+ * - TGSI -> LLVM translation (i.e., lp_build_tgsi_soa)
+ * - texture sampling code generation (i.e., lp_build_sample_soa)
+ * - LLVM pipe driver
+ *
+ * All interesting code is in the functions mentioned above. There is really
+ * nothing to see here.
+ *
+ * @author Jose Fonseca <jfonseca@vmware.com>
+ */
+
+#include "pipe/p_defines.h"
+#include "pipe/p_shader_tokens.h"
+#include "lp_bld_debug.h"
+#include "lp_bld_type.h"
+#include "lp_bld_intr.h"
+#include "lp_bld_sample.h"
+#include "lp_bld_tgsi.h"
+#include "lp_state.h"
+#include "lp_tex_sample.h"
+
+
+/**
+ * This provides the bridge between the sampler state store in lp_jit_context
+ * and lp_jit_texture and the sampler code generator. It provides the
+ * texture layout information required by the texture sampler code generator
+ * in terms of the state stored in lp_jit_context and lp_jit_texture in runtime.
+ */
+struct llvmpipe_sampler_dynamic_state
+{
+   struct lp_sampler_dynamic_state base;
+
+   const struct lp_sampler_static_state *static_state;
+
+   LLVMValueRef context_ptr;
+};
+
+
+/**
+ * This is the bridge between our sampler and the TGSI translator.
+ */
+struct lp_llvm_sampler_soa
+{
+   struct lp_build_sampler_soa base;
+
+   struct llvmpipe_sampler_dynamic_state dynamic_state;
+};
+
+
+/**
+ * Fetch the specified member of the lp_jit_texture structure.
+ *
+ * @sa http://llvm.org/docs/GetElementPtr.html
+ */
+static LLVMValueRef
+lp_llvm_texture_member(struct lp_sampler_dynamic_state *base,
+                       LLVMBuilderRef builder,
+                       unsigned unit,
+                       unsigned member_index,
+                       const char *member_name)
+{
+   struct llvmpipe_sampler_dynamic_state *state = (struct llvmpipe_sampler_dynamic_state *)base;
+   LLVMValueRef indices[4];
+   LLVMValueRef ptr;
+   LLVMValueRef res;
+
+   assert(unit < PIPE_MAX_SAMPLERS);
+
+   /* context[0] */
+   indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
+   /* context[0].textures */
+   indices[1] = LLVMConstInt(LLVMInt32Type(), LP_JIT_CONTEXT_TEXTURES_INDEX, 0);
+   /* context[0].textures[unit] */
+   indices[2] = LLVMConstInt(LLVMInt32Type(), unit, 0);
+   /* context[0].textures[unit].member */
+   indices[3] = LLVMConstInt(LLVMInt32Type(), member_index, 0);
+
+   ptr = LLVMBuildGEP(builder, state->context_ptr, indices, Elements(indices), "");
+
+   res = LLVMBuildLoad(builder, ptr, "");
+
+   lp_build_name(res, "context.texture%u.%s", unit, member_name);
+
+   return res;
+}
+
+
+/**
+ * Helper macro to instantiate the functions that generate the code to fetch
+ * the members of lp_jit_texture to fulfill the sampler code generator requests.
+ *
+ * This complexity is the price we have to pay to keep the texture sampler code
+ * generator a reusable module without dependencies to llvmpipe internals.
+ */
+#define LP_LLVM_TEXTURE_MEMBER(_name, _index) \
+   static LLVMValueRef \
+   lp_llvm_texture_##_name( struct lp_sampler_dynamic_state *base, \
+                            LLVMBuilderRef builder, \
+                            unsigned unit) \
+   { \
+      return lp_llvm_texture_member(base, builder, unit, _index, #_name ); \
+   }
+
+
+LP_LLVM_TEXTURE_MEMBER(width,    LP_JIT_TEXTURE_WIDTH)
+LP_LLVM_TEXTURE_MEMBER(height,   LP_JIT_TEXTURE_HEIGHT)
+LP_LLVM_TEXTURE_MEMBER(stride,   LP_JIT_TEXTURE_STRIDE)
+LP_LLVM_TEXTURE_MEMBER(data_ptr, LP_JIT_TEXTURE_DATA)
+
+
+static void
+lp_llvm_sampler_soa_destroy(struct lp_build_sampler_soa *sampler)
+{
+   FREE(sampler);
+}
+
+
+static void
+lp_llvm_sampler_soa_emit_fetch_texel(struct lp_build_sampler_soa *base,
+                                     LLVMBuilderRef builder,
+                                     union lp_type type,
+                                     unsigned unit,
+                                     unsigned num_coords,
+                                     const LLVMValueRef *coords,
+                                     LLVMValueRef lodbias,
+                                     LLVMValueRef *texel)
+{
+   struct lp_llvm_sampler_soa *sampler = (struct lp_llvm_sampler_soa *)base;
+
+   assert(unit < PIPE_MAX_SAMPLERS);
+
+   lp_build_sample_soa(builder,
+                       &sampler->dynamic_state.static_state[unit],
+                       &sampler->dynamic_state.base,
+                       type,
+                       unit,
+                       num_coords,
+                       coords,
+                       lodbias,
+                       texel);
+}
+
+
+struct lp_build_sampler_soa *
+lp_llvm_sampler_soa_create(const struct lp_sampler_static_state *static_state,
+                           LLVMValueRef context_ptr)
+{
+   struct lp_llvm_sampler_soa *sampler;
+
+   sampler = CALLOC_STRUCT(lp_llvm_sampler_soa);
+   if(!sampler)
+      return NULL;
+
+   sampler->base.destroy = lp_llvm_sampler_soa_destroy;
+   sampler->base.emit_fetch_texel = lp_llvm_sampler_soa_emit_fetch_texel;
+   sampler->dynamic_state.base.width = lp_llvm_texture_width;
+   sampler->dynamic_state.base.height = lp_llvm_texture_height;
+   sampler->dynamic_state.base.stride = lp_llvm_texture_stride;
+   sampler->dynamic_state.base.data_ptr = lp_llvm_texture_data_ptr;
+   sampler->dynamic_state.static_state = static_state;
+   sampler->dynamic_state.context_ptr = context_ptr;
+
+   return &sampler->base;
+}
+
diff --git a/src/gallium/drivers/nv50/nv50_miptree.c b/src/gallium/drivers/nv50/nv50_miptree.c
index 03b9243b82..93479a0314 100644
--- a/src/gallium/drivers/nv50/nv50_miptree.c
+++ b/src/gallium/drivers/nv50/nv50_miptree.c
@@ -148,6 +148,7 @@ nv50_miptree_blanket(struct pipe_screen *pscreen, const struct pipe_texture *pt,
 	mt->image_nr = 1;
 	mt->level[0].pitch = *stride;
 	mt->level[0].image_offset = CALLOC(1, sizeof(unsigned));
+	mt->level[0].tile_mode = bo->tile_mode;
 
 	nouveau_bo_ref(bo, &mt->base.bo);
 	return &mt->base.base;
diff --git a/src/gallium/drivers/nv50/nv50_program.c b/src/gallium/drivers/nv50/nv50_program.c
index 289c3485e0..4a838529de 100644
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -1106,10 +1106,10 @@ convert_to_long(struct nv50_pc *pc, struct nv50_program_exec *e)
 		m = 0xffff7fff;
 		break;
 	case 0x8:
-		/* INTERP */
-		m = ~0x02000000;
-		if (e->inst[0] & 0x02000000)
-			q = 0x00020000;
+		/* INTERP (move centroid, perspective and flat bits) */
+		m = ~0x03000100;
+		q = (e->inst[0] & (3 << 24)) >> (24 - 16);
+		q |= (e->inst[0] & (1 << 8)) << (18 - 8);
 		break;
 	case 0x9:
 		/* RCP */
@@ -1495,6 +1495,7 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union tgsi_full_token *tok)
 		}
 		break;
 	case TGSI_OPCODE_MOV:
+	case TGSI_OPCODE_SWZ:
 		for (c = 0; c < 4; c++) {
 			if (!(mask & (1 << c)))
 				continue;
diff --git a/src/gallium/drivers/nv50/nv50_state_validate.c b/src/gallium/drivers/nv50/nv50_state_validate.c
index 344c2cf6dd..d294356f75 100644
--- a/src/gallium/drivers/nv50/nv50_state_validate.c
+++ b/src/gallium/drivers/nv50/nv50_state_validate.c
@@ -66,7 +66,8 @@ nv50_state_validate_fb(struct nv50_context *nv50)
 			so_data(so, NV50TCL_RT_FORMAT_X8R8G8B8_UNORM);
 			break;
 		}
-		so_data(so, bo->tile_mode << 4);
+		so_data(so, nv50_miptree(pt)->
+				level[fb->cbufs[i]->level].tile_mode << 4);
 		so_data(so, 0x00000000);
 
 		so_method(so, tesla, 0x1224, 1);
@@ -110,7 +111,8 @@ nv50_state_validate_fb(struct nv50_context *nv50)
 			so_data(so, NV50TCL_ZETA_FORMAT_S8Z24_UNORM);
 			break;
 		}
-		so_data(so, bo->tile_mode << 4);
+		so_data(so, nv50_miptree(pt)->
+				level[fb->zsbuf->level].tile_mode << 4);
 		so_data(so, 0x00000000);
 
 		so_method(so, tesla, 0x1538, 1);
diff --git a/src/gallium/drivers/nv50/nv50_surface.c b/src/gallium/drivers/nv50/nv50_surface.c
index b266324f58..6bf6f773b0 100644
--- a/src/gallium/drivers/nv50/nv50_surface.c
+++ b/src/gallium/drivers/nv50/nv50_surface.c
@@ -60,13 +60,13 @@ nv50_surface_set(struct nv50_screen *screen, struct pipe_surface *ps, int dst)
  	format = nv50_format(ps->format);
  	if (format < 0)
  		return 1;
-  
+
  	if (!bo->tile_flags) {
  		BEGIN_RING(chan, eng2d, mthd, 2);
  		OUT_RING  (chan, format);
  		OUT_RING  (chan, 1);
  		BEGIN_RING(chan, eng2d, mthd + 0x14, 5);
- 		OUT_RING  (chan, mt->level[0].pitch);
+		OUT_RING  (chan, mt->level[ps->level].pitch);
  		OUT_RING  (chan, ps->width);
  		OUT_RING  (chan, ps->height);
  		OUT_RELOCh(chan, bo, ps->offset, flags);
@@ -75,7 +75,7 @@ nv50_surface_set(struct nv50_screen *screen, struct pipe_surface *ps, int dst)
  		BEGIN_RING(chan, eng2d, mthd, 5);
  		OUT_RING  (chan, format);
  		OUT_RING  (chan, 0);
- 		OUT_RING  (chan, bo->tile_mode << 4);
+		OUT_RING  (chan, mt->level[ps->level].tile_mode << 4);
  		OUT_RING  (chan, 1);
  		OUT_RING  (chan, 0);
  		BEGIN_RING(chan, eng2d, mthd + 0x18, 4);
diff --git a/src/gallium/drivers/nv50/nv50_vbo.c b/src/gallium/drivers/nv50/nv50_vbo.c
index 36f1b24b2f..eeed148c7b 100644
--- a/src/gallium/drivers/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nv50/nv50_vbo.c
@@ -178,7 +178,7 @@ nv50_draw_elements_inline_u08(struct nv50_context *nv50, uint8_t *map,
 
 		BEGIN_RING(chan, tesla, 0x400015f0, nr >> 1);
 		for (i = 0; i < nr; i += 2)
-			OUT_RING  (chan, (map[1] << 16) | map[0]);
+			OUT_RING  (chan, (map[i + 1] << 16) | map[i]);
 
 		count -= nr;
 		map += nr;
@@ -207,7 +207,7 @@ nv50_draw_elements_inline_u16(struct nv50_context *nv50, uint16_t *map,
 
 		BEGIN_RING(chan, tesla, 0x400015f0, nr >> 1);
 		for (i = 0; i < nr; i += 2)
-			OUT_RING  (chan, (map[1] << 16) | map[0]);
+			OUT_RING  (chan, (map[i + 1] << 16) | map[i]);
 
 		count -= nr;
 		map += nr;
@@ -313,18 +313,18 @@ nv50_vbo_static_attrib(struct nv50_context *nv50, unsigned attrib,
 		so_data  (so, fui(v[3]));
 		break;
 	case 3:
-		so_method(so, tesla, NV50TCL_VTX_ATTR_3F_X(attrib), 4);
+		so_method(so, tesla, NV50TCL_VTX_ATTR_3F_X(attrib), 3);
 		so_data  (so, fui(v[0]));
 		so_data  (so, fui(v[1]));
 		so_data  (so, fui(v[2]));
 		break;
 	case 2:
-		so_method(so, tesla, NV50TCL_VTX_ATTR_2F_X(attrib), 4);
+		so_method(so, tesla, NV50TCL_VTX_ATTR_2F_X(attrib), 2);
 		so_data  (so, fui(v[0]));
 		so_data  (so, fui(v[1]));
 		break;
 	case 1:
-		so_method(so, tesla, NV50TCL_VTX_ATTR_1F(attrib), 4);
+		so_method(so, tesla, NV50TCL_VTX_ATTR_1F(attrib), 1);
 		so_data  (so, fui(v[0]));
 		break;
 	default:
diff --git a/src/gallium/drivers/r300/Makefile b/src/gallium/drivers/r300/Makefile
index d7a2c8c462..93c2152edc 100644
--- a/src/gallium/drivers/r300/Makefile
+++ b/src/gallium/drivers/r300/Makefile
@@ -9,6 +9,7 @@ C_SOURCES = \
 	r300_chipset.c \
 	r300_clear.c \
 	r300_context.c \
+	r300_debug.c \
 	r300_emit.c \
 	r300_flush.c \
 	r300_fs.c \
diff --git a/src/gallium/drivers/r300/r300_context.c b/src/gallium/drivers/r300/r300_context.c
index da67bc29b8..9cc455135d 100644
--- a/src/gallium/drivers/r300/r300_context.c
+++ b/src/gallium/drivers/r300/r300_context.c
@@ -22,6 +22,9 @@
 
 #include "r300_context.h"
 
+#include "r300_flush.h"
+#include "r300_state_invariant.h"
+
 static boolean r300_draw_range_elements(struct pipe_context* pipe,
                                         struct pipe_buffer* indexBuffer,
                                         unsigned indexSize,
@@ -146,6 +149,8 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
     r300->context.winsys = (struct pipe_winsys*)r300_winsys;
     r300->context.screen = r300_screen(screen);
 
+    r300_init_debug(r300);
+
     r300->context.destroy = r300_destroy_context;
 
     r300->context.clear = r300_clear;
diff --git a/src/gallium/drivers/r300/r300_context.h b/src/gallium/drivers/r300/r300_context.h
index f78492d4aa..6c5914baa3 100644
--- a/src/gallium/drivers/r300/r300_context.h
+++ b/src/gallium/drivers/r300/r300_context.h
@@ -211,10 +211,7 @@ struct r300_vertex_format {
     int fs_tab[16];
 };
 
-static struct pipe_viewport_state r300_viewport_identity = {
-    .scale = {1.0, 1.0, 1.0, 1.0},
-    .translate = {0.0, 0.0, 0.0, 0.0},
-};
+extern struct pipe_viewport_state r300_viewport_identity;
 
 struct r300_context {
     /* Parent class */
@@ -275,6 +272,9 @@ struct r300_context {
     uint32_t dirty_state;
     /* Flag indicating whether or not the HW is dirty. */
     uint32_t dirty_hw;
+
+    /** Combination of DBG_xxx flags */
+    unsigned debug;
 };
 
 /* Convenience cast wrapper. */
@@ -288,4 +288,40 @@ struct draw_stage* r300_draw_stage(struct r300_context* r300);
 void r300_init_state_functions(struct r300_context* r300);
 void r300_init_surface_functions(struct r300_context* r300);
 
+/* Debug functionality. */
+
+/**
+ * Debug flags to disable/enable certain groups of debugging outputs.
+ *
+ * \note These may be rather coarse, and the grouping may be impractical.
+ * If you find, while debugging the driver, that a different grouping
+ * of these flags would be beneficial, just feel free to change them
+ * but make sure to update the documentation in r300_debug.c to reflect
+ * those changes.
+ */
+/*@{*/
+#define DBG_HELP    0x0000001
+#define DBG_FP      0x0000002
+#define DBG_VP      0x0000004
+#define DBG_CS      0x0000008
+#define DBG_DRAW    0x0000010
+/*@}*/
+
+static INLINE boolean DBG_ON(struct r300_context * ctx, unsigned flags)
+{
+    return (ctx->debug & flags) ? true : false;
+}
+
+static INLINE void DBG(struct r300_context * ctx, unsigned flags, const char * fmt, ...)
+{
+    if (DBG_ON(ctx, flags)) {
+        va_list va;
+        va_start(va, fmt);
+        debug_vprintf(fmt, va);
+        va_end(va);
+    }
+}
+
+void r300_init_debug(struct r300_context * ctx);
+
 #endif /* R300_CONTEXT_H */
diff --git a/src/gallium/drivers/r300/r300_cs.h b/src/gallium/drivers/r300/r300_cs.h
index 71b142c0db..0a7e470363 100644
--- a/src/gallium/drivers/r300/r300_cs.h
+++ b/src/gallium/drivers/r300/r300_cs.h
@@ -49,7 +49,8 @@
     (RADEON_CP_PACKET0 | ((count) << 16) | ((register) >> 2))
 
 #define CS_LOCALS(context) \
-    struct r300_winsys* cs_winsys = context->winsys; \
+    struct r300_context* const cs_context_copy = (context); \
+    struct r300_winsys* cs_winsys = cs_context_copy->winsys; \
     int cs_count = 0;
 
 #define CHECK_CS(size) \
@@ -58,7 +59,7 @@
 #define BEGIN_CS(size) do { \
     CHECK_CS(size); \
     if (VERY_VERBOSE_CS) { \
-        debug_printf("r300: BEGIN_CS, count %d, in %s (%s:%d)\n", \
+        DBG(cs_context_copy, DBG_CS, "r300: BEGIN_CS, count %d, in %s (%s:%d)\n", \
                 size, __FUNCTION__, __FILE__, __LINE__); \
     } \
     cs_winsys->begin_cs(cs_winsys, (size), \
@@ -78,7 +79,7 @@
 
 #define OUT_CS_REG(register, value) do { \
     if (VERY_VERBOSE_REGISTERS) \
-        debug_printf("r300: writing 0x%08X to register 0x%04X\n", \
+        DBG(cs_context_copy, DBG_CS, "r300: writing 0x%08X to register 0x%04X\n", \
             value, register); \
     assert(register); \
     OUT_CS(CP_PACKET0(register, 0)); \
@@ -89,14 +90,14 @@
  * not the actual packet0 count! */
 #define OUT_CS_REG_SEQ(register, count) do { \
     if (VERY_VERBOSE_REGISTERS) \
-        debug_printf("r300: writing register sequence of %d to 0x%04X\n", \
+        DBG(cs_context_copy, DBG_CS, "r300: writing register sequence of %d to 0x%04X\n", \
             count, register); \
     assert(register); \
     OUT_CS(CP_PACKET0(register, ((count) - 1))); \
 } while (0)
 
 #define OUT_CS_RELOC(bo, offset, rd, wd, flags) do { \
-    debug_printf("r300: writing relocation for buffer %p, offset %d, " \
+    DBG(cs_context_copy, DBG_CS, "r300: writing relocation for buffer %p, offset %d, " \
             "domains (%d, %d, %d)\n", \
         bo, offset, rd, wd, flags); \
     assert(bo); \
@@ -107,7 +108,7 @@
 
 #define END_CS do { \
     if (VERY_VERBOSE_CS) { \
-        debug_printf("r300: END_CS in %s (%s:%d)\n", __FUNCTION__, \
+        DBG(cs_context_copy, DBG_CS, "r300: END_CS in %s (%s:%d)\n", __FUNCTION__, \
                 __FILE__, __LINE__); \
     } \
     if (cs_count != 0) \
@@ -117,7 +118,7 @@
 
 #define FLUSH_CS do { \
     if (VERY_VERBOSE_CS) { \
-        debug_printf("r300: FLUSH_CS in %s (%s:%d)\n\n", __FUNCTION__, \
+        DBG(cs_context_copy, DBG_CS, "r300: FLUSH_CS in %s (%s:%d)\n\n", __FUNCTION__, \
                 __FILE__, __LINE__); \
     } \
     cs_winsys->flush_cs(cs_winsys); \
@@ -127,7 +128,7 @@
 
 #define OUT_CS_ONE_REG(register, count) do { \
     if (VERY_VERBOSE_REGISTERS) \
-        debug_printf("r300: writing data sequence of %d to 0x%04X\n", \
+        DBG(cs_context_copy, DBG_CS, "r300: writing data sequence of %d to 0x%04X\n", \
             count, register); \
     assert(register); \
     OUT_CS(CP_PACKET0(register, ((count) - 1)) | RADEON_ONE_REG_WR); \
@@ -141,7 +142,7 @@
 } while (0)
 
 #define OUT_CS_INDEX_RELOC(bo, offset, count, rd, wd, flags) do { \
-    debug_printf("r300: writing relocation for index buffer %p," \
+    DBG(cs_context_copy, DBG_CS, "r300: writing relocation for index buffer %p," \
             "offset %d\n", bo, offset); \
     assert(bo); \
     OUT_CS(offset); \
diff --git a/src/gallium/drivers/r300/r300_debug.c b/src/gallium/drivers/r300/r300_debug.c
new file mode 100644
index 0000000000..15308dda1d
--- /dev/null
+++ b/src/gallium/drivers/r300/r300_debug.c
@@ -0,0 +1,88 @@
+/*
+ * Copyright 2009 Nicolai Haehnle <nhaehnle@gmail.com>
+ *
+ * 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
+ * on 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
+ * THE AUTHOR(S) AND/OR THEIR 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 "r300_context.h"
+
+#include <ctype.h>
+
+
+struct debug_option {
+    const char * name;
+    unsigned flag;
+    const char * description;
+};
+
+static struct debug_option debug_options[] = {
+    { "help", DBG_HELP, "Helpful meta-information about the driver" },
+    { "fp", DBG_FP, "Fragment program handling" },
+    { "vp", DBG_VP, "Vertex program handling" },
+    { "cs", DBG_CS, "Command submissions" },
+    { "draw", DBG_DRAW, "Draw and emit" },
+
+    { "all", ~0, "Convenience option that enables all debug flags" },
+
+    /* must be last */
+    { 0, 0, 0 }
+};
+
+void r300_init_debug(struct r300_context * ctx)
+{
+    const char * options = debug_get_option("RADEON_DEBUG", 0);
+    boolean printhint = false;
+
+    if (options) {
+        while(*options) {
+            if (*options == ' ' || *options == ',') {
+                options++;
+                continue;
+            }
+
+            size_t length = strcspn(options, " ,");
+            struct debug_option * opt;
+
+            for(opt = debug_options; opt->name; ++opt) {
+                if (!strncmp(options, opt->name, length)) {
+                    ctx->debug |= opt->flag;
+                    break;
+                }
+            }
+
+            if (!opt->name) {
+                debug_printf("Unknown debug option: %s\n", options);
+                printhint = true;
+            }
+
+            options += length;
+        }
+
+        if (!ctx->debug)
+            printhint = true;
+    }
+
+    if (printhint || ctx->debug & DBG_HELP) {
+        debug_printf("You can enable debug output by setting the RADEON_DEBUG environment variable\n"
+                     "to a comma-separated list of debug options. Available options are:\n");
+        for(struct debug_option * opt = debug_options; opt->name; ++opt) {
+            debug_printf("    %s: %s\n", opt->name, opt->description);
+        }
+    }
+}
diff --git a/src/gallium/drivers/r300/r300_emit.c b/src/gallium/drivers/r300/r300_emit.c
index bd4d59e6f1..1bc35c2486 100644
--- a/src/gallium/drivers/r300/r300_emit.c
+++ b/src/gallium/drivers/r300/r300_emit.c
@@ -25,6 +25,7 @@
 #include "r300_emit.h"
 
 #include "r300_fs.h"
+#include "r300_state_derived.h"
 #include "r300_vs.h"
 
 void r300_emit_blend_state(struct r300_context* r300,
@@ -490,7 +491,7 @@ void r300_emit_vertex_buffer(struct r300_context* r300)
 {
     CS_LOCALS(r300);
 
-    debug_printf("r300: Preparing vertex buffer %p for render, "
+    DBG(r300, DBG_DRAW, "r300: Preparing vertex buffer %p for render, "
             "vertex size %d\n", r300->vbo,
             r300->vertex_info.vinfo.size);
     /* Set the pointer to our vertex buffer. The emitted values are this:
diff --git a/src/gallium/drivers/r300/r300_emit.h b/src/gallium/drivers/r300/r300_emit.h
index 350691d592..c4002b8e5d 100644
--- a/src/gallium/drivers/r300/r300_emit.h
+++ b/src/gallium/drivers/r300/r300_emit.h
@@ -56,6 +56,11 @@ void r500_emit_fragment_program_code(struct r300_context* r300,
 void r300_emit_fb_state(struct r300_context* r300,
                         struct pipe_framebuffer_state* fb);
 
+void r300_emit_query_begin(struct r300_context* r300,
+                           struct r300_query* query);
+void r300_emit_query_end(struct r300_context* r300,
+                         struct r300_query* query);
+
 void r300_emit_rs_state(struct r300_context* r300, struct r300_rs_state* rs);
 
 void r300_emit_rs_block_state(struct r300_context* r300,
diff --git a/src/gallium/drivers/r300/r300_fs.c b/src/gallium/drivers/r300/r300_fs.c
index 36463b9a2e..a0e848a59a 100644
--- a/src/gallium/drivers/r300/r300_fs.c
+++ b/src/gallium/drivers/r300/r300_fs.c
@@ -96,7 +96,7 @@ void r300_translate_fragment_shader(struct r300_context* r300,
 
     memset(&compiler, 0, sizeof(compiler));
     rc_init(&compiler.Base);
-    compiler.Base.Debug = 1;
+    compiler.Base.Debug = DBG_ON(r300, DBG_FP);
 
     compiler.code = &fs->code;
     compiler.is_r500 = r300_screen(r300->context.screen)->caps->is_r500;
diff --git a/src/gallium/drivers/r300/r300_query.c b/src/gallium/drivers/r300/r300_query.c
index 1d5185b417..2880d34877 100644
--- a/src/gallium/drivers/r300/r300_query.c
+++ b/src/gallium/drivers/r300/r300_query.c
@@ -22,6 +22,8 @@
 
 #include "r300_query.h"
 
+#include "r300_emit.h"
+
 static struct pipe_query* r300_create_query(struct pipe_context* pipe,
                                             unsigned query_type)
 {
diff --git a/src/gallium/drivers/r300/r300_render.c b/src/gallium/drivers/r300/r300_render.c
index cd458d019a..d05a736dd9 100644
--- a/src/gallium/drivers/r300/r300_render.c
+++ b/src/gallium/drivers/r300/r300_render.c
@@ -26,6 +26,7 @@
 
 #include "r300_cs.h"
 #include "r300_context.h"
+#include "r300_emit.h"
 #include "r300_reg.h"
 #include "r300_state_derived.h"
 
@@ -34,7 +35,7 @@
 struct r300_render {
     /* Parent class */
     struct vbuf_render base;
-    
+
     /* Pipe context */
     struct r300_context* r300;
 
@@ -77,7 +78,7 @@ static boolean r300_render_allocate_vertices(struct vbuf_render* render,
     if (r300render->vbo && (size > r300render->vbo_alloc_size)) {
         pipe_buffer_reference(&r300render->vbo, NULL);
     }
-    
+
     if (!r300render->vbo) {
         r300render->vbo = pipe_buffer_create(screen,
                                              64,
@@ -184,7 +185,7 @@ static void r300_render_draw_arrays(struct vbuf_render* render,
 
     prepare_render(r300render, count);
 
-    debug_printf("r300: Doing vbuf render, count %d\n", count);
+    DBG(r300, DBG_DRAW, "r300: Doing vbuf render, count %d\n", count);
 
     BEGIN_CS(2);
     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
@@ -233,7 +234,8 @@ static void r300_render_draw(struct vbuf_render* render,
     OUT_CS_INDEX_RELOC(index_buffer, 0, count, RADEON_GEM_DOMAIN_GTT, 0, 0);
     END_CS; */
 
-    BEGIN_CS(2 + (count+1)/2);
+    BEGIN_CS(4 + (count+1)/2);
+    OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, count);
     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, (count+1)/2);
     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
            r300render->hwprim);
diff --git a/src/gallium/drivers/r300/r300_state.c b/src/gallium/drivers/r300/r300_state.c
index c16cadd040..88cb9af6fb 100644
--- a/src/gallium/drivers/r300/r300_state.c
+++ b/src/gallium/drivers/r300/r300_state.c
@@ -20,10 +20,11 @@
  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
 
+#include "util/u_debug.h"
 #include "util/u_math.h"
 #include "util/u_pack_color.h"
 
-#include "util/u_debug.h"
+#include "tgsi/tgsi_parse.h"
 
 #include "pipe/p_config.h"
 #include "pipe/internal/p_winsys_screen.h"
@@ -429,6 +430,9 @@ static void r300_bind_rs_state(struct pipe_context* pipe, void* state)
 
     r300->rs_state = rs;
     r300->dirty_state |= R300_NEW_RASTERIZER;
+    r300->dirty_state |= R300_NEW_RS_BLOCK;
+    r300->dirty_state |= R300_NEW_SCISSOR;
+    r300->dirty_state |= R300_NEW_VIEWPORT;
 }
 
 /* Free rasterizer state. */
diff --git a/src/gallium/drivers/r300/r300_state_derived.c b/src/gallium/drivers/r300/r300_state_derived.c
index c01e61a9b1..5f6b225d34 100644
--- a/src/gallium/drivers/r300/r300_state_derived.c
+++ b/src/gallium/drivers/r300/r300_state_derived.c
@@ -23,6 +23,7 @@
 #include "r300_state_derived.h"
 
 #include "r300_fs.h"
+#include "r300_state_inlines.h"
 #include "r300_vs.h"
 
 /* r300_state_derived: Various bits of state which are dependent upon
@@ -195,13 +196,13 @@ static void r300_vertex_psc(struct r300_context* r300,
      * and not on attrib information. */
     if (r300screen->caps->has_tcl) {
         attrib_count = r300->vs->info.num_inputs;
-        debug_printf("r300: routing %d attribs in psc for vs\n",
+        DBG(r300, DBG_DRAW, "r300: routing %d attribs in psc for vs\n",
                 attrib_count);
     } else {
         attrib_count = vinfo->num_attribs;
-        debug_printf("r300: attrib count: %d\n", attrib_count);
+        DBG(r300, DBG_DRAW, "r300: attrib count: %d\n", attrib_count);
         for (i = 0; i < attrib_count; i++) {
-            debug_printf("r300: attrib: offset %d, interp %d, size %d,"
+            DBG(r300, DBG_DRAW, "r300: attrib: offset %d, interp %d, size %d,"
                    " tab %d\n", vinfo->attrib[i].src_index,
                    vinfo->attrib[i].interp_mode, vinfo->attrib[i].emit,
                    tab[i]);
@@ -299,18 +300,18 @@ static void r300_update_fs_tab(struct r300_context* r300)
     }
 
     /* Now that we know where everything is... */
-    debug_printf("r300: fp input count: %d\n", info->num_inputs);
+    DBG(r300, DBG_DRAW, "r300: fp input count: %d\n", info->num_inputs);
     for (i = 0; i < info->num_inputs; i++) {
         switch (tab[i]) {
             case INTERP_LINEAR:
-                debug_printf("r300: attrib: "
+                DBG(r300, DBG_DRAW, "r300: attrib: "
                         "stack offset %d, color,    tab %d\n",
                         i, cols_emitted);
                 tab[i] = cols_emitted;
                 cols_emitted++;
                 break;
             case INTERP_PERSPECTIVE:
-                debug_printf("r300: attrib: "
+                DBG(r300, DBG_DRAW, "r300: attrib: "
                         "stack offset %d, texcoord, tab %d\n",
                         i, cols + texs);
                 tab[i] = cols + texs;
diff --git a/src/gallium/drivers/r300/r300_state_derived.h b/src/gallium/drivers/r300/r300_state_derived.h
index 63ae8eb8d0..71a4a47b00 100644
--- a/src/gallium/drivers/r300/r300_state_derived.h
+++ b/src/gallium/drivers/r300/r300_state_derived.h
@@ -23,11 +23,7 @@
 #ifndef R300_STATE_DERIVED_H
 #define R300_STATE_DERIVED_H
 
-#include "draw/draw_vertex.h"
-
-#include "r300_context.h"
-#include "r300_reg.h"
-#include "r300_state_inlines.h"
+struct r300_context;
 
 void r300_update_derived_state(struct r300_context* r300);
 
diff --git a/src/gallium/drivers/r300/r300_state_invariant.c b/src/gallium/drivers/r300/r300_state_invariant.c
index 1e92374a4e..3865730d63 100644
--- a/src/gallium/drivers/r300/r300_state_invariant.c
+++ b/src/gallium/drivers/r300/r300_state_invariant.c
@@ -23,6 +23,12 @@
 
 #include "r300_state_invariant.h"
 
+
+struct pipe_viewport_state r300_viewport_identity = {
+    .scale = {1.0, 1.0, 1.0, 1.0},
+    .translate = {0.0, 0.0, 0.0, 0.0},
+};
+
 /* Calculate and emit invariant state. This is data that the 3D engine
  * will probably want at the beginning of every CS, but it's not currently
  * handled by any CSO setup, and in addition it doesn't really change much.
@@ -38,7 +44,9 @@ void r300_emit_invariant_state(struct r300_context* r300)
 
     /*** Graphics Backend (GB) ***/
     /* Various GB enables */
-    OUT_CS_REG(R300_GB_ENABLE, 0x0);
+    OUT_CS_REG(R300_GB_ENABLE, R300_GB_POINT_STUFF_ENABLE |
+                               R300_GB_LINE_STUFF_ENABLE  |
+                               R300_GB_TRIANGLE_STUFF_ENABLE);
     /* Subpixel multisampling for AA
      * These are commented out because glisse's CS checker doesn't like them.
      * I presume these will be re-enabled later.
diff --git a/src/gallium/drivers/r300/r300_surface.h b/src/gallium/drivers/r300/r300_surface.h
index d01f0b143f..f9e98b2ec9 100644
--- a/src/gallium/drivers/r300/r300_surface.h
+++ b/src/gallium/drivers/r300/r300_surface.h
@@ -73,9 +73,9 @@ static struct r300_rs_state rs_clear_state = {
 };
 
 static struct r300_rs_block r3xx_rs_block_clear_state = {
-    .ip[0] = R500_RS_SEL_S(R300_RS_SEL_K0) |
-        R500_RS_SEL_T(R300_RS_SEL_K0) |
-        R500_RS_SEL_R(R300_RS_SEL_K0) |
+    .ip[0] = R500_RS_SEL_S(R300_RS_SEL_C0) |
+        R500_RS_SEL_T(R300_RS_SEL_C0) |
+        R500_RS_SEL_R(R300_RS_SEL_C0) |
         R500_RS_SEL_Q(R300_RS_SEL_K1),
     .inst[0] = R300_RS_INST_COL_CN_WRITE,
     .count = R300_IT_COUNT(0) | R300_IC_COUNT(1) | R300_HIRES_EN,
diff --git a/src/gallium/drivers/r300/r300_tgsi_to_rc.c b/src/gallium/drivers/r300/r300_tgsi_to_rc.c
index 3adbb715f3..d68a104106 100644
--- a/src/gallium/drivers/r300/r300_tgsi_to_rc.c
+++ b/src/gallium/drivers/r300/r300_tgsi_to_rc.c
@@ -53,7 +53,6 @@ static unsigned translate_opcode(unsigned opcode)
         case TGSI_OPCODE_SUB: return OPCODE_SUB;
         case TGSI_OPCODE_LRP: return OPCODE_LRP;
      /* case TGSI_OPCODE_CND: return OPCODE_CND; */
-     /* case TGSI_OPCODE_CND0: return OPCODE_CND0; */
         case TGSI_OPCODE_DP2A: return OPCODE_DP2A;
                                         /* gap */
         case TGSI_OPCODE_FRC: return OPCODE_FRC;
diff --git a/src/gallium/drivers/r300/r300_vs.c b/src/gallium/drivers/r300/r300_vs.c
index 2cb903bba2..12a6e37be6 100644
--- a/src/gallium/drivers/r300/r300_vs.c
+++ b/src/gallium/drivers/r300/r300_vs.c
@@ -116,7 +116,7 @@ void r300_translate_vertex_shader(struct r300_context* r300,
     /* Setup the compiler */
     rc_init(&compiler.Base);
 
-    compiler.Base.Debug = 1;
+    compiler.Base.Debug = DBG_ON(r300, DBG_VP);
     compiler.code = &vs->code;
     compiler.UserData = vs;
 
diff --git a/src/gallium/drivers/softpipe/sp_clear.c b/src/gallium/drivers/softpipe/sp_clear.c
index fa59277438..d3af18e162 100644
--- a/src/gallium/drivers/softpipe/sp_clear.c
+++ b/src/gallium/drivers/softpipe/sp_clear.c
@@ -85,5 +85,7 @@ softpipe_clear(struct pipe_context *pipe, unsigned buffers, const float *rgba,
       /* non-cached surface */
       pipe->surface_fill(pipe, ps, 0, 0, ps->width, ps->height, cv);
 #endif
-      }
+   }
+
+   softpipe->dirty_render_cache = TRUE;
 }
diff --git a/src/gallium/drivers/trace/tr_context.c b/src/gallium/drivers/trace/tr_context.c
index 4ab718f233..ae0af4d055 100644
--- a/src/gallium/drivers/trace/tr_context.c
+++ b/src/gallium/drivers/trace/tr_context.c
@@ -1277,8 +1277,10 @@ trace_context_create(struct pipe_screen *_screen,
    tr_ctx->base.set_sampler_textures = trace_context_set_sampler_textures;
    tr_ctx->base.set_vertex_buffers = trace_context_set_vertex_buffers;
    tr_ctx->base.set_vertex_elements = trace_context_set_vertex_elements;
-   tr_ctx->base.surface_copy = trace_context_surface_copy;
-   tr_ctx->base.surface_fill = trace_context_surface_fill;
+   if (pipe->surface_copy)
+      tr_ctx->base.surface_copy = trace_context_surface_copy;
+   if (pipe->surface_fill)
+      tr_ctx->base.surface_fill = trace_context_surface_fill;
    tr_ctx->base.clear = trace_context_clear;
    tr_ctx->base.flush = trace_context_flush;
    tr_ctx->base.is_texture_referenced = trace_is_texture_referenced;