Merge branch 'master' into r300-compiler

There were additional non-textual conflicts.

Conflicts:
	src/gallium/drivers/r300/r300_tgsi_to_rc.c
	src/mesa/drivers/dri/r300/compiler/r3xx_vertprog.c
	src/mesa/drivers/dri/r300/compiler/radeon_program.c
	src/mesa/drivers/dri/r300/compiler/radeon_program_alu.c

57 files changed, 2741 insertions, 726 deletions

diff --git a/src/gallium/drivers/llvmpipe/Makefile b/src/gallium/drivers/llvmpipe/Makefile
index 6e63a0c2b7..cd7b6356d2 100644
--- a/src/gallium/drivers/llvmpipe/Makefile
+++ b/src/gallium/drivers/llvmpipe/Makefile
@@ -3,6 +3,8 @@ include $(TOP)/configs/current
 
 LIBNAME = llvmpipe
 
+CFLAGS += -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS
+
 C_SOURCES = \
 	lp_bld_alpha.c \
 	lp_bld_arit.c \
@@ -15,9 +17,11 @@ 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 +48,8 @@ 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/README b/src/gallium/drivers/llvmpipe/README
index 498d21dea6..89d08834a3 100644
--- a/src/gallium/drivers/llvmpipe/README
+++ b/src/gallium/drivers/llvmpipe/README
@@ -8,13 +8,16 @@ Done so far is:
 
  - the whole fragment pipeline is code generated in a single function
  
+   - input interpolation
+   
    - depth testing
  
+   - texture sampling (not all state/formats are supported) 
+   
    - fragment shader TGSI translation
      - same level of support as the TGSI SSE2 exec machine, with the exception
        we don't fallback to TGSI interpretation when an unsupported opcode is
        found, but just ignore it
-     - texture sampling via an intrinsic call
      - done in SoA layout
      - input interpolation also code generated
  
@@ -28,16 +31,17 @@ Done so far is:
      any width and length
    - not all operations are implemented for these types yet though
 
-Most mesa/progs/demos/* work. Speed is on par with Keith's softpipe-opt branch,
-which includes hand written fast implementations for common cases.
+Most mesa/progs/demos/* work. 
 
 To do (probably by this order):
 
  - code generate stipple and stencil testing
 
- - code generate texture sampling
+ - translate the remaining bits of texture sampling state
 
  - translate TGSI control flow instructions, and all other remaining opcodes
+ 
+ - integrate with the draw module for VS code generation
 
  - code generate the triangle setup and rasterization
 
@@ -93,7 +97,7 @@ Alternatively, you can build it with GNU make, if you prefer, by invoking it as
 
   make linux-llvm
 
-but the rest of these instructions assume scons is used.
+but the rest of these instructions assume that scons is used.
 
 
 Using
@@ -108,6 +112,9 @@ or
 
   export LD_LIBRARY_PATH=$PWD/build/linux-x86-debug/lib:$LD_LIBRARY_PATH
 
+For performance evaluation pass debug=no to scons, and use the corresponding
+lib directory without the "-debug" suffix.
+
 
 Unit testing
 ============
@@ -119,7 +126,7 @@ build/linux-???-debug/gallium/drivers/llvmpipe:
  - lp_test_conv: SIMD vector conversion
  - lp_test_format: pixel unpacking/packing
 
-Some of this tests can output results and benchmarks to a tab-seperated-file
+Some of this tests can output results and benchmarks to a tab-separated-file
 for posterior analysis, e.g.:
 
   build/linux-x86_64-debug/gallium/drivers/llvmpipe/lp_test_blend -o blend.tsv
@@ -133,10 +140,10 @@ Development Notes
   at the top of the lp_bld_*.c functions.  
 
 - All lp_bld_*.[ch] are isolated from the rest of the driver, and could/may be 
-  put in a standalone Gallium state -> LLVM IR translation module.
+  put in a stand-alone Gallium state -> LLVM IR translation module.
 
 - 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
   generation. See 
   http://npcontemplation.blogspot.com/2008/06/secret-of-llvm-c-bindings.html
-  for a standalone example.
+  for a stand-alone example.
diff --git a/src/gallium/drivers/llvmpipe/SConscript b/src/gallium/drivers/llvmpipe/SConscript
index 5c29bdac56..f4a9a3b22e 100644
--- a/src/gallium/drivers/llvmpipe/SConscript
+++ b/src/gallium/drivers/llvmpipe/SConscript
@@ -3,7 +3,7 @@ Import('*')
 env = env.Clone()
 
 env.Tool('llvm')
-if 'LLVM_VERSION' not in env:
+if not env.has_key('LLVM_VERSION'):
     print 'warning: LLVM not found: not building llvmpipe'
     Return()
 
@@ -23,8 +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',
@@ -52,7 +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',
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_alpha.c b/src/gallium/drivers/llvmpipe/lp_bld_alpha.c
index 49c2f911af..2b4bc5c819 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_alpha.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_alpha.c
@@ -45,7 +45,7 @@
 void
 lp_build_alpha_test(LLVMBuilderRef builder,
                     const struct pipe_alpha_state *state,
-                    union lp_type type,
+                    struct lp_type type,
                     struct lp_build_mask_context *mask,
                     LLVMValueRef alpha,
                     LLVMValueRef ref)
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_alpha.h b/src/gallium/drivers/llvmpipe/lp_bld_alpha.h
index 9dbcdb4daa..634575670d 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_alpha.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_alpha.h
@@ -38,14 +38,14 @@
 #include <llvm-c/Core.h>  
 
 struct pipe_alpha_state;
-union lp_type;
+struct lp_type;
 struct lp_build_mask_context;
 
 
 void
 lp_build_alpha_test(LLVMBuilderRef builder,
                     const struct pipe_alpha_state *state,
-                    union lp_type type,
+                    struct lp_type type,
                     struct lp_build_mask_context *mask,
                     LLVMValueRef alpha,
                     LLVMValueRef ref);
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.c b/src/gallium/drivers/llvmpipe/lp_bld_arit.c
index 09a57ff33d..0b115fc9b0 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_arit.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.c
@@ -65,7 +65,7 @@ lp_build_min_simple(struct lp_build_context *bld,
                     LLVMValueRef a,
                     LLVMValueRef b)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    const char *intrinsic = NULL;
    LLVMValueRef cond;
 
@@ -113,7 +113,7 @@ lp_build_max_simple(struct lp_build_context *bld,
                     LLVMValueRef a,
                     LLVMValueRef b)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    const char *intrinsic = NULL;
    LLVMValueRef cond;
 
@@ -159,7 +159,7 @@ LLVMValueRef
 lp_build_comp(struct lp_build_context *bld,
               LLVMValueRef a)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
 
    if(a == bld->one)
       return bld->zero;
@@ -188,7 +188,7 @@ lp_build_add(struct lp_build_context *bld,
              LLVMValueRef a,
              LLVMValueRef b)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    LLVMValueRef res;
 
    if(a == bld->zero)
@@ -241,7 +241,7 @@ lp_build_sub(struct lp_build_context *bld,
              LLVMValueRef a,
              LLVMValueRef b)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    LLVMValueRef res;
 
    if(b == bld->zero)
@@ -405,7 +405,7 @@ lp_build_mul(struct lp_build_context *bld,
              LLVMValueRef a,
              LLVMValueRef b)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
 
    if(a == bld->zero)
       return bld->zero;
@@ -477,7 +477,7 @@ lp_build_div(struct lp_build_context *bld,
              LLVMValueRef a,
              LLVMValueRef b)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
 
    if(a == bld->zero)
       return bld->zero;
@@ -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.
@@ -565,21 +590,32 @@ LLVMValueRef
 lp_build_abs(struct lp_build_context *bld,
              LLVMValueRef a)
 {
-   const union lp_type type = bld->type;
+   const struct 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
 
@@ -588,10 +624,188 @@ lp_build_abs(struct lp_build_context *bld,
 
 
 LLVMValueRef
+lp_build_sgn(struct lp_build_context *bld,
+             LLVMValueRef a)
+{
+   const struct lp_type type = bld->type;
+   LLVMTypeRef vec_type = lp_build_vec_type(type);
+   LLVMValueRef cond;
+   LLVMValueRef res;
+
+   /* Handle non-zero case */
+   if(!type.sign) {
+      /* if not zero then sign must be positive */
+      res = bld->one;
+   }
+   else if(type.floating) {
+      /* Take the sign bit and add it to 1 constant */
+      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));
+      LLVMValueRef sign;
+      LLVMValueRef one;
+      sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
+      sign = LLVMBuildAnd(bld->builder, sign, mask, "");
+      one = LLVMConstBitCast(bld->one, int_vec_type);
+      res = LLVMBuildOr(bld->builder, sign, one, "");
+      res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
+   }
+   else
+   {
+      LLVMValueRef minus_one = lp_build_const_scalar(type, -1.0);
+      cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, bld->zero);
+      res = lp_build_select(bld, cond, bld->one, minus_one);
+   }
+
+   /* Handle zero */
+   cond = lp_build_cmp(bld, PIPE_FUNC_EQUAL, a, bld->zero);
+   res = lp_build_select(bld, cond, bld->zero, bld->one);
+
+   return res;
+}
+
+
+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 struct 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_round(struct lp_build_context *bld,
+               LLVMValueRef a)
+{
+   const struct 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_NEAREST);
+#endif
+
+   /* FIXME */
+   assert(0);
+   return bld->undef;
+}
+
+
+LLVMValueRef
+lp_build_floor(struct lp_build_context *bld,
+               LLVMValueRef a)
+{
+   const struct 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;
+}
+
+
+LLVMValueRef
+lp_build_ceil(struct lp_build_context *bld,
+              LLVMValueRef a)
+{
+   const struct 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_CEIL);
+#endif
+
+   /* FIXME */
+   assert(0);
+   return bld->undef;
+}
+
+
+LLVMValueRef
+lp_build_trunc(struct lp_build_context *bld,
+               LLVMValueRef a)
+{
+   const struct 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_TRUNCATE);
+#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 struct 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)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    LLVMTypeRef vec_type = lp_build_vec_type(type);
    char intrinsic[32];
 
@@ -609,7 +823,7 @@ LLVMValueRef
 lp_build_rcp(struct lp_build_context *bld,
              LLVMValueRef a)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
 
    if(a == bld->zero)
       return bld->undef;
@@ -640,7 +854,7 @@ LLVMValueRef
 lp_build_rsqrt(struct lp_build_context *bld,
                LLVMValueRef a)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
 
    assert(type.floating);
 
@@ -661,7 +875,7 @@ LLVMValueRef
 lp_build_cos(struct lp_build_context *bld,
               LLVMValueRef a)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    LLVMTypeRef vec_type = lp_build_vec_type(type);
    char intrinsic[32];
 
@@ -681,7 +895,7 @@ LLVMValueRef
 lp_build_sin(struct lp_build_context *bld,
               LLVMValueRef a)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    LLVMTypeRef vec_type = lp_build_vec_type(type);
    char intrinsic[32];
 
@@ -752,7 +966,7 @@ lp_build_polynomial(struct lp_build_context *bld,
                     const double *coeffs,
                     unsigned num_coeffs)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    LLVMValueRef res = NULL;
    unsigned i;
 
@@ -800,7 +1014,7 @@ lp_build_exp2_approx(struct lp_build_context *bld,
                      LLVMValueRef *p_frac_part,
                      LLVMValueRef *p_exp2)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    LLVMTypeRef vec_type = lp_build_vec_type(type);
    LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
    LLVMValueRef ipart = NULL;
@@ -893,7 +1107,7 @@ lp_build_log2_approx(struct lp_build_context *bld,
                      LLVMValueRef *p_floor_log2,
                      LLVMValueRef *p_log2)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    LLVMTypeRef vec_type = lp_build_vec_type(type);
    LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.h b/src/gallium/drivers/llvmpipe/lp_bld_arit.h
index fc8cb25966..d68a97c4b8 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_arit.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.h
@@ -40,7 +40,7 @@
 #include <llvm-c/Core.h>  
 
 
-union lp_type type;
+struct lp_type type;
 struct lp_build_context;
 
 
@@ -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 +106,34 @@ lp_build_abs(struct lp_build_context *bld,
              LLVMValueRef a);
 
 LLVMValueRef
+lp_build_sgn(struct lp_build_context *bld,
+             LLVMValueRef a);
+
+LLVMValueRef
+lp_build_round(struct lp_build_context *bld,
+               LLVMValueRef a);
+
+LLVMValueRef
+lp_build_floor(struct lp_build_context *bld,
+               LLVMValueRef a);
+
+LLVMValueRef
+lp_build_ceil(struct lp_build_context *bld,
+              LLVMValueRef a);
+
+LLVMValueRef
+lp_build_trunc(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_blend.h b/src/gallium/drivers/llvmpipe/lp_bld_blend.h
index d19e18846c..da272e549f 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_blend.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_blend.h
@@ -46,7 +46,7 @@
 
 
 struct pipe_blend_state;
-union lp_type;
+struct lp_type;
 struct lp_build_context;
 
 
@@ -74,7 +74,7 @@ lp_build_blend_func(struct lp_build_context *bld,
 LLVMValueRef
 lp_build_blend_aos(LLVMBuilderRef builder,
                    const struct pipe_blend_state *blend,
-                   union lp_type type,
+                   struct lp_type type,
                    LLVMValueRef src,
                    LLVMValueRef dst,
                    LLVMValueRef const_,
@@ -84,7 +84,7 @@ lp_build_blend_aos(LLVMBuilderRef builder,
 void
 lp_build_blend_soa(LLVMBuilderRef builder,
                    const struct pipe_blend_state *blend,
-                   union lp_type type,
+                   struct lp_type type,
                    LLVMValueRef src[4],
                    LLVMValueRef dst[4],
                    LLVMValueRef const_[4],
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c b/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c
index c11a9398f8..d14f468ba9 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_blend_aos.c
@@ -303,7 +303,7 @@ lp_build_blend_func(struct lp_build_context *bld,
 LLVMValueRef
 lp_build_blend_aos(LLVMBuilderRef builder,
                    const struct pipe_blend_state *blend,
-                   union lp_type type,
+                   struct lp_type type,
                    LLVMValueRef src,
                    LLVMValueRef dst,
                    LLVMValueRef const_,
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c b/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c
index b92254a7d6..9511299d55 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_blend_soa.c
@@ -199,7 +199,7 @@ lp_build_blend_soa_factor(struct lp_build_blend_soa_context *bld,
 void
 lp_build_blend_soa(LLVMBuilderRef builder,
                    const struct pipe_blend_state *blend,
-                   union lp_type type,
+                   struct lp_type type,
                    LLVMValueRef src[4],
                    LLVMValueRef dst[4],
                    LLVMValueRef con[4],
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_const.c b/src/gallium/drivers/llvmpipe/lp_bld_const.c
index 21487365ea..c8eaa8c394 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_const.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_const.c
@@ -42,7 +42,7 @@
 
 
 unsigned
-lp_mantissa(union lp_type type)
+lp_mantissa(struct lp_type type)
 {
    assert(type.floating);
 
@@ -72,7 +72,7 @@ lp_mantissa(union lp_type type)
  * Same as lp_const_scale(), but in terms of shifts.
  */
 unsigned
-lp_const_shift(union lp_type type)
+lp_const_shift(struct lp_type type)
 {
    if(type.floating)
       return 0;
@@ -86,7 +86,7 @@ lp_const_shift(union lp_type type)
 
 
 unsigned
-lp_const_offset(union lp_type type)
+lp_const_offset(struct lp_type type)
 {
    if(type.floating || type.fixed)
       return 0;
@@ -104,7 +104,7 @@ lp_const_offset(union lp_type type)
  * else for the fixed points types and normalized integers.
  */
 double
-lp_const_scale(union lp_type type)
+lp_const_scale(struct lp_type type)
 {
    unsigned long long llscale;
    double dscale;
@@ -122,7 +122,7 @@ lp_const_scale(union lp_type type)
  * Minimum value representable by the type.
  */
 double
-lp_const_min(union lp_type type)
+lp_const_min(struct lp_type type)
 {
    unsigned bits;
 
@@ -158,7 +158,7 @@ lp_const_min(union lp_type type)
  * Maximum value representable by the type.
  */
 double
-lp_const_max(union lp_type type)
+lp_const_max(struct lp_type type)
 {
    unsigned bits;
 
@@ -190,7 +190,7 @@ lp_const_max(union lp_type type)
 
 
 double
-lp_const_eps(union lp_type type)
+lp_const_eps(struct lp_type type)
 {
    if (type.floating) {
       switch(type.width) {
@@ -211,7 +211,7 @@ lp_const_eps(union lp_type type)
 
 
 LLVMValueRef
-lp_build_undef(union lp_type type)
+lp_build_undef(struct lp_type type)
 {
    LLVMTypeRef vec_type = lp_build_vec_type(type);
    return LLVMGetUndef(vec_type);
@@ -219,7 +219,7 @@ lp_build_undef(union lp_type type)
                
 
 LLVMValueRef
-lp_build_zero(union lp_type type)
+lp_build_zero(struct lp_type type)
 {
    LLVMTypeRef vec_type = lp_build_vec_type(type);
    return LLVMConstNull(vec_type);
@@ -227,7 +227,7 @@ lp_build_zero(union lp_type type)
                
 
 LLVMValueRef
-lp_build_one(union lp_type type)
+lp_build_one(struct lp_type type)
 {
    LLVMTypeRef elem_type;
    LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
@@ -269,7 +269,7 @@ lp_build_one(union lp_type type)
                
 
 LLVMValueRef
-lp_build_const_scalar(union lp_type type,
+lp_build_const_scalar(struct lp_type type,
                       double val)
 {
    LLVMTypeRef elem_type = lp_build_elem_type(type);
@@ -295,7 +295,7 @@ lp_build_const_scalar(union lp_type type,
 
 
 LLVMValueRef
-lp_build_int_const_scalar(union lp_type type,
+lp_build_int_const_scalar(struct lp_type type,
                           long long val)
 {
    LLVMTypeRef elem_type = lp_build_int_elem_type(type);
@@ -312,7 +312,7 @@ lp_build_int_const_scalar(union lp_type type,
 
 
 LLVMValueRef
-lp_build_const_aos(union lp_type type, 
+lp_build_const_aos(struct lp_type type, 
                    double r, double g, double b, double a, 
                    const unsigned char *swizzle)
 {
@@ -352,8 +352,8 @@ lp_build_const_aos(union lp_type type,
 
 
 LLVMValueRef
-lp_build_const_mask_aos(union lp_type type,
-                        boolean cond[4])
+lp_build_const_mask_aos(struct lp_type type,
+                        const boolean cond[4])
 {
    LLVMTypeRef elem_type = LLVMIntType(type.width);
    LLVMValueRef masks[LP_MAX_VECTOR_LENGTH];
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_const.h b/src/gallium/drivers/llvmpipe/lp_bld_const.h
index 1934530ea3..ffb302f736 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_const.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_const.h
@@ -42,67 +42,67 @@
 #include <pipe/p_compiler.h>
 
 
-union lp_type type;
+struct lp_type type;
 
 
 unsigned
-lp_mantissa(union lp_type type);
+lp_mantissa(struct lp_type type);
 
 
 unsigned
-lp_const_shift(union lp_type type);
+lp_const_shift(struct lp_type type);
 
 
 unsigned
-lp_const_offset(union lp_type type);
+lp_const_offset(struct lp_type type);
 
 
 double
-lp_const_scale(union lp_type type);
+lp_const_scale(struct lp_type type);
 
 double
-lp_const_min(union lp_type type);
+lp_const_min(struct lp_type type);
 
 
 double
-lp_const_max(union lp_type type);
+lp_const_max(struct lp_type type);
 
 
 double
-lp_const_eps(union lp_type type);
+lp_const_eps(struct lp_type type);
 
 
 LLVMValueRef
-lp_build_undef(union lp_type type);
+lp_build_undef(struct lp_type type);
 
 
 LLVMValueRef
-lp_build_zero(union lp_type type);
+lp_build_zero(struct lp_type type);
 
 
 LLVMValueRef
-lp_build_one(union lp_type type);
+lp_build_one(struct lp_type type);
 
 
 LLVMValueRef
-lp_build_const_scalar(union lp_type type,
+lp_build_const_scalar(struct lp_type type,
                       double val);
 
 
 LLVMValueRef
-lp_build_int_const_scalar(union lp_type type,
+lp_build_int_const_scalar(struct lp_type type,
                           long long val);
 
 
 LLVMValueRef
-lp_build_const_aos(union lp_type type, 
+lp_build_const_aos(struct lp_type type, 
                    double r, double g, double b, double a, 
                    const unsigned char *swizzle);
 
 
 LLVMValueRef
-lp_build_const_mask_aos(union lp_type type,
-                        boolean cond[4]);
+lp_build_const_mask_aos(struct lp_type type,
+                        const boolean cond[4]);
 
 
 #endif /* !LP_BLD_CONST_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_conv.c b/src/gallium/drivers/llvmpipe/lp_bld_conv.c
index c8954c8a34..186cac70f6 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_conv.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_conv.c
@@ -86,7 +86,7 @@
  */
 LLVMValueRef
 lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
-                                        union lp_type src_type,
+                                        struct lp_type src_type,
                                         unsigned dst_width,
                                         LLVMValueRef src)
 {
@@ -122,7 +122,7 @@ lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
       int shift = dst_width - n;
       res = LLVMBuildShl(builder, res, lp_build_int_const_scalar(src_type, shift), "");
 
-      /* Fill in the empty lower bits for added precision? */
+      /* TODO: Fill in the empty lower bits for additional precision? */
 #if 0
       {
          LLVMValueRef msb;
@@ -152,7 +152,7 @@ lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
 LLVMValueRef
 lp_build_unsigned_norm_to_float(LLVMBuilderRef builder,
                                 unsigned src_width,
-                                union lp_type dst_type,
+                                struct lp_type dst_type,
                                 LLVMValueRef src)
 {
    LLVMTypeRef vec_type = lp_build_vec_type(dst_type);
@@ -244,12 +244,12 @@ lp_build_const_pack_shuffle(unsigned n)
  * Expand the bit width.
  *
  * This will only change the number of bits the values are represented, not the
- * values themselved.
+ * values themselves.
  */
 static void
 lp_build_expand(LLVMBuilderRef builder,
-               union lp_type src_type,
-               union lp_type dst_type,
+               struct lp_type src_type,
+               struct lp_type dst_type,
                LLVMValueRef src,
                LLVMValueRef *dst, unsigned num_dsts)
 {
@@ -266,7 +266,7 @@ lp_build_expand(LLVMBuilderRef builder,
    dst[0] = src;
 
    while(src_type.width < dst_type.width) {
-      union lp_type new_type = src_type;
+      struct lp_type new_type = src_type;
       LLVMTypeRef new_vec_type;
 
       new_type.width *= 2;
@@ -314,8 +314,8 @@ lp_build_expand(LLVMBuilderRef builder,
  */
 static LLVMValueRef
 lp_build_pack2(LLVMBuilderRef builder,
-               union lp_type src_type,
-               union lp_type dst_type,
+               struct lp_type src_type,
+               struct lp_type dst_type,
                boolean clamped,
                LLVMValueRef lo,
                LLVMValueRef hi)
@@ -391,11 +391,11 @@ lp_build_pack2(LLVMBuilderRef builder,
  * TODO: Handle saturation consistently.
  */
 static LLVMValueRef
-lp_build_trunc(LLVMBuilderRef builder,
-               union lp_type src_type,
-               union lp_type dst_type,
-               boolean clamped,
-               const LLVMValueRef *src, unsigned num_srcs)
+lp_build_pack(LLVMBuilderRef builder,
+              struct lp_type src_type,
+              struct lp_type dst_type,
+              boolean clamped,
+              const LLVMValueRef *src, unsigned num_srcs)
 {
    LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
    unsigned i;
@@ -410,7 +410,7 @@ lp_build_trunc(LLVMBuilderRef builder,
       tmp[i] = src[i];
 
    while(src_type.width > dst_type.width) {
-      union lp_type new_type = src_type;
+      struct lp_type new_type = src_type;
 
       new_type.width /= 2;
       new_type.length *= 2;
@@ -442,12 +442,12 @@ lp_build_trunc(LLVMBuilderRef builder,
  */
 void
 lp_build_conv(LLVMBuilderRef builder,
-              union lp_type src_type,
-              union lp_type dst_type,
+              struct lp_type src_type,
+              struct lp_type dst_type,
               const LLVMValueRef *src, unsigned num_srcs,
               LLVMValueRef *dst, unsigned num_dsts)
 {
-   union lp_type tmp_type;
+   struct lp_type tmp_type;
    LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
    unsigned num_tmps;
    unsigned i;
@@ -470,7 +470,7 @@ lp_build_conv(LLVMBuilderRef builder,
     * Clamp if necessary
     */
 
-   if(src_type.value != dst_type.value) {
+   if(memcmp(&src_type, &dst_type, sizeof src_type) != 0) {
       struct lp_build_context bld;
       double src_min = lp_const_min(src_type);
       double dst_min = lp_const_min(dst_type);
@@ -565,7 +565,7 @@ lp_build_conv(LLVMBuilderRef builder,
 
    if(tmp_type.width > dst_type.width) {
       assert(num_dsts == 1);
-      tmp[0] = lp_build_trunc(builder, tmp_type, dst_type, TRUE, tmp, num_tmps);
+      tmp[0] = lp_build_pack(builder, tmp_type, dst_type, TRUE, tmp, num_tmps);
       tmp_type.width = dst_type.width;
       tmp_type.length = dst_type.length;
       num_tmps = 1;
@@ -656,8 +656,8 @@ lp_build_conv(LLVMBuilderRef builder,
  */
 void
 lp_build_conv_mask(LLVMBuilderRef builder,
-                   union lp_type src_type,
-                   union lp_type dst_type,
+                   struct lp_type src_type,
+                   struct lp_type dst_type,
                    const LLVMValueRef *src, unsigned num_srcs,
                    LLVMValueRef *dst, unsigned num_dsts)
 {
@@ -689,7 +689,7 @@ lp_build_conv_mask(LLVMBuilderRef builder,
 
    if(src_type.width > dst_type.width) {
       assert(num_dsts == 1);
-      dst[0] = lp_build_trunc(builder, src_type, dst_type, TRUE, src, num_srcs);
+      dst[0] = lp_build_pack(builder, src_type, dst_type, TRUE, src, num_srcs);
    }
    else if(src_type.width < dst_type.width) {
       assert(num_srcs == 1);
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_conv.h b/src/gallium/drivers/llvmpipe/lp_bld_conv.h
index 05c1ef2a10..ca378804d2 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_conv.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_conv.h
@@ -40,33 +40,33 @@
 #include <llvm-c/Core.h>  
 
 
-union lp_type type;
+struct lp_type type;
 
 
 LLVMValueRef
 lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
-                                        union lp_type src_type,
+                                        struct lp_type src_type,
                                         unsigned dst_width,
                                         LLVMValueRef src);
 
 LLVMValueRef
 lp_build_unsigned_norm_to_float(LLVMBuilderRef builder,
                                 unsigned src_width,
-                                union lp_type dst_type,
+                                struct lp_type dst_type,
                                 LLVMValueRef src);
 
 
 void
 lp_build_conv(LLVMBuilderRef builder,
-              union lp_type src_type,
-              union lp_type dst_type,
+              struct lp_type src_type,
+              struct lp_type dst_type,
               const LLVMValueRef *srcs, unsigned num_srcs,
               LLVMValueRef *dsts, unsigned num_dsts);
 
 void
 lp_build_conv_mask(LLVMBuilderRef builder,
-                   union lp_type src_type,
-                   union lp_type dst_type,
+                   struct lp_type src_type,
+                   struct lp_type dst_type,
                    const LLVMValueRef *src, unsigned num_srcs,
                    LLVMValueRef *dst, unsigned num_dsts);
 
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..21c665c4d4 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_depth.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_depth.c
@@ -71,11 +71,11 @@
 /**
  * Return a type appropriate for depth/stencil testing.
  */
-union lp_type
+struct lp_type
 lp_depth_type(const struct util_format_description *format_desc,
               unsigned length)
 {
-   union lp_type type;
+   struct lp_type type;
    unsigned swizzle;
 
    assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS);
@@ -85,7 +85,7 @@ lp_depth_type(const struct util_format_description *format_desc,
    swizzle = format_desc->swizzle[0];
    assert(swizzle < 4);
 
-   type.value = 0;
+   memset(&type, 0, sizeof type);
    type.width = format_desc->block.bits;
 
    if(format_desc->channel[swizzle].type == UTIL_FORMAT_TYPE_FLOAT) {
@@ -114,7 +114,7 @@ lp_depth_type(const struct util_format_description *format_desc,
 void
 lp_build_depth_test(LLVMBuilderRef builder,
                     const struct pipe_depth_state *state,
-                    union lp_type type,
+                    struct lp_type type,
                     const struct util_format_description *format_desc,
                     struct lp_build_mask_context *mask,
                     LLVMValueRef src,
@@ -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)
@@ -210,5 +211,6 @@ lp_build_depth_test(LLVMBuilderRef builder,
       LLVMBuildStore(builder, dst, dst_ptr);
    }
 
+   /* FIXME */
    assert(!state->occlusion_count);
 }
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_depth.h b/src/gallium/drivers/llvmpipe/lp_bld_depth.h
index 5d2e042fcc..79d6981bb5 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_depth.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_depth.h
@@ -41,11 +41,11 @@
  
 struct pipe_depth_state;
 struct util_format_description;
-union lp_type;
+struct lp_type;
 struct lp_build_mask_context;
 
 
-union lp_type
+struct lp_type
 lp_depth_type(const struct util_format_description *format_desc,
               unsigned length);
 
@@ -53,7 +53,7 @@ lp_depth_type(const struct util_format_description *format_desc,
 void
 lp_build_depth_test(LLVMBuilderRef builder,
                     const struct pipe_depth_state *state,
-                    union lp_type type,
+                    struct lp_type type,
                     const struct util_format_description *format_desc,
                     struct lp_build_mask_context *mask,
                     LLVMValueRef src,
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_flow.c b/src/gallium/drivers/llvmpipe/lp_bld_flow.c
index 9d99e1a9d9..dcc25fbff8 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)
+{
+   struct lp_build_flow_skip *skip;
+   LLVMBasicBlockRef current_block;
+   LLVMBasicBlockRef new_block;
+   unsigned i;
+
+   skip = &lp_build_flow_peek(flow, LP_BUILD_FLOW_SKIP)->skip;
+   if(!skip)
+      return;
+
+   current_block = LLVMGetInsertBlock(flow->builder);
+
+   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)
 {
-   if(mask->skip_block) {
-      LLVMBasicBlockRef current_block = LLVMGetInsertBlock(mask->builder);
+   struct lp_build_flow_skip *skip;
+   LLVMBasicBlockRef current_block;
+   unsigned i;
 
-      LLVMAddIncoming(mask->phi, &mask->value, &current_block, 1);
-      LLVMBuildBr(mask->builder, mask->skip_block);
+   skip = &lp_build_flow_pop(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;
+   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,
+                    struct 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..e61999ff06 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_flow.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_flow.h
@@ -38,27 +38,53 @@
 #include <llvm-c/Core.h>  
 
 
-union lp_type;
+struct 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,
-                    union lp_type type,
+                    struct lp_build_flow_context *flow,
+                    struct 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..6d3f692619 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_format.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_format.h
@@ -31,21 +31,15 @@
 
 /**
  * @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"
 
-union lp_type;
+struct util_format_description;
+struct 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,
+                         struct lp_type type,
+                         LLVMValueRef packed,
+                         LLVMValueRef *rgba);
+
+
+void
+lp_build_load_rgba_soa(LLVMBuilderRef builder,
+                       const struct util_format_description *format_desc,
+                       struct 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..b5ff434e1a
--- /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(struct 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,
+                         struct 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,
+                       struct 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_interp.c b/src/gallium/drivers/llvmpipe/lp_bld_interp.c
index cfe20a0d75..338dbca6d1 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_interp.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_interp.c
@@ -292,7 +292,7 @@ void
 lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
                          const struct tgsi_token *tokens,
                          LLVMBuilderRef builder,
-                         union lp_type type,
+                         struct lp_type type,
                          LLVMValueRef a0_ptr,
                          LLVMValueRef dadx_ptr,
                          LLVMValueRef dady_ptr,
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_interp.h b/src/gallium/drivers/llvmpipe/lp_bld_interp.h
index 9194f6233a..9c57a10879 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_interp.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_interp.h
@@ -83,7 +83,7 @@ void
 lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
                          const struct tgsi_token *tokens,
                          LLVMBuilderRef builder,
-                         union lp_type type,
+                         struct lp_type type,
                          LLVMValueRef a0_ptr,
                          LLVMValueRef dadx_ptr,
                          LLVMValueRef dady_ptr,
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_logic.c b/src/gallium/drivers/llvmpipe/lp_bld_logic.c
index 8631efd6c3..6b6f820769 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_logic.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_logic.c
@@ -45,7 +45,7 @@ lp_build_cmp(struct lp_build_context *bld,
              LLVMValueRef a,
              LLVMValueRef b)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    LLVMTypeRef vec_type = lp_build_vec_type(type);
    LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
    LLVMValueRef zeros = LLVMConstNull(int_vec_type);
@@ -301,7 +301,7 @@ lp_build_select(struct lp_build_context *bld,
                 LLVMValueRef a,
                 LLVMValueRef b)
 {
-   union lp_type type = bld->type;
+   struct lp_type type = bld->type;
    LLVMValueRef res;
 
    if(a == b)
@@ -313,8 +313,6 @@ lp_build_select(struct lp_build_context *bld,
       b = LLVMBuildBitCast(bld->builder, b, int_vec_type, "");
    }
 
-   /* TODO: On SSE4 we could do this with a single instruction -- PBLENDVB */
-
    a = LLVMBuildAnd(bld->builder, a, mask, "");
 
    /* This often gets translated to PANDN, but sometimes the NOT is
@@ -339,9 +337,9 @@ LLVMValueRef
 lp_build_select_aos(struct lp_build_context *bld,
                     LLVMValueRef a,
                     LLVMValueRef b,
-                    boolean cond[4])
+                    const boolean cond[4])
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    const unsigned n = type.length;
    unsigned i, j;
 
@@ -376,9 +374,9 @@ lp_build_select_aos(struct lp_build_context *bld,
 
       return LLVMBuildShuffleVector(bld->builder, a, b, LLVMConstVector(shuffles, n), "");
    }
+   else {
 #if 0
-   else if(0) {
-      /* FIXME: Unfortunately select of vectors do not work */
+      /* XXX: Unfortunately select of vectors do not work */
       /* Use a select */
       LLVMTypeRef elem_type = LLVMInt1Type();
       LLVMValueRef cond[LP_MAX_VECTOR_LENGTH];
@@ -388,10 +386,9 @@ lp_build_select_aos(struct lp_build_context *bld,
             cond[j + i] = LLVMConstInt(elem_type, cond[i] ? 1 : 0, 0);
 
       return LLVMBuildSelect(bld->builder, LLVMConstVector(cond, n), a, b, "");
-   }
-#endif
-   else {
+#else
       LLVMValueRef mask = lp_build_const_mask_aos(type, cond);
       return lp_build_select(bld, mask, a, b);
+#endif
    }
 }
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_logic.h b/src/gallium/drivers/llvmpipe/lp_bld_logic.h
index 29b9e1c45b..a4ee7723b5 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_logic.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_logic.h
@@ -42,7 +42,7 @@
 #include "pipe/p_defines.h" /* For PIPE_FUNC_xxx */
 
 
-union lp_type type;
+struct lp_type type;
 struct lp_build_context;
 
 
@@ -66,7 +66,7 @@ LLVMValueRef
 lp_build_select_aos(struct lp_build_context *bld,
                     LLVMValueRef a,
                     LLVMValueRef b,
-                    boolean cond[4]);
+                    const boolean cond[4]);
 
 
 #endif /* !LP_BLD_LOGIC_H */
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..403d0e4836
--- /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;
+struct 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,
+                    struct 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..8ca1be6f1b
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_bld_sample_soa.c
@@ -0,0 +1,416 @@
+/**************************************************************************
+ *
+ * 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 */
+   struct lp_type coord_type;
+   struct lp_build_context coord_bld;
+
+   /** Integer coordinates */
+   struct lp_type int_coord_type;
+   struct lp_build_context int_coord_bld;
+
+   /** Output texels type and build context */
+   struct 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:
+      /* FIXME */
+      _debug_printf("warning: failed to translate texture wrap mode %u\n", wrap_mode);
+      coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
+      coord = lp_build_min(int_coord_bld, coord, length_minus_one);
+      break;
+
+   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,
+                    struct 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;
+   default:
+      assert(0);
+   }
+
+   /* FIXME: respect static_state->min_mip_filter */;
+   /* FIXME: respect static_state->mag_img_filter */;
+   /* FIXME: respect static_state->prefilter */;
+
+   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_swizzle.c b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.c
index ac7eed9379..64e81f7b1f 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_swizzle.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.c
@@ -64,7 +64,7 @@ LLVMValueRef
 lp_build_broadcast_scalar(struct lp_build_context *bld,
                           LLVMValueRef scalar)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    LLVMValueRef res;
    unsigned i;
 
@@ -83,7 +83,7 @@ lp_build_broadcast_aos(struct lp_build_context *bld,
                        LLVMValueRef a,
                        unsigned channel)
 {
-   const union lp_type type = bld->type;
+   const struct lp_type type = bld->type;
    const unsigned n = type.length;
    unsigned i, j;
 
@@ -115,7 +115,7 @@ lp_build_broadcast_aos(struct lp_build_context *bld,
        *   YY00 YY00 .... YY00
        *   YYYY YYYY .... YYYY  <= output
        */
-      union lp_type type4 = type;
+      struct lp_type type4 = type;
       const char shifts[4][2] = {
          { 1,  2},
          {-1,  2},
@@ -161,7 +161,7 @@ lp_build_broadcast_aos(struct lp_build_context *bld,
 LLVMValueRef
 lp_build_swizzle1_aos(struct lp_build_context *bld,
                       LLVMValueRef a,
-                      unsigned char swizzle[4])
+                      const unsigned char swizzle[4])
 {
    const unsigned n = bld->type.length;
    unsigned i, j;
@@ -192,7 +192,7 @@ LLVMValueRef
 lp_build_swizzle2_aos(struct lp_build_context *bld,
                       LLVMValueRef a,
                       LLVMValueRef b,
-                      unsigned char swizzle[4])
+                      const unsigned char swizzle[4])
 {
    const unsigned n = bld->type.length;
    unsigned i, j;
@@ -201,11 +201,12 @@ lp_build_swizzle2_aos(struct lp_build_context *bld,
       return lp_build_swizzle1_aos(bld, a, swizzle);
 
    if(a == b) {
-      swizzle[0] %= 4;
-      swizzle[1] %= 4;
-      swizzle[2] %= 4;
-      swizzle[3] %= 4;
-      return lp_build_swizzle1_aos(bld, a, swizzle);
+      unsigned char swizzle1[4];
+      swizzle1[0] = swizzle[0] % 4;
+      swizzle1[1] = swizzle[1] % 4;
+      swizzle1[2] = swizzle[2] % 4;
+      swizzle1[3] = swizzle[3] % 4;
+      return lp_build_swizzle1_aos(bld, a, swizzle1);
    }
 
    if(swizzle[0] % 4 == 0 &&
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_swizzle.h b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.h
index d7dd6a8a60..1f6da80448 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_swizzle.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_swizzle.h
@@ -40,7 +40,7 @@
 #include <llvm-c/Core.h>  
 
 
-union lp_type type;
+struct lp_type type;
 struct lp_build_context;
 
 
@@ -73,7 +73,7 @@ lp_build_broadcast_aos(struct lp_build_context *bld,
 LLVMValueRef
 lp_build_swizzle1_aos(struct lp_build_context *bld,
                       LLVMValueRef a,
-                      unsigned char swizzle[4]);
+                      const unsigned char swizzle[4]);
 
 
 /**
@@ -85,7 +85,7 @@ LLVMValueRef
 lp_build_swizzle2_aos(struct lp_build_context *bld,
                       LLVMValueRef a,
                       LLVMValueRef b,
-                      unsigned char swizzle[4]);
+                      const unsigned char swizzle[4]);
 
 
 #endif /* !LP_BLD_SWIZZLE_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h b/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h
index 912db24aec..eddb7a83fa 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_tgsi.h
@@ -39,31 +39,46 @@
 
 
 struct tgsi_token;
-union lp_type;
+struct lp_type;
 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,
+                        struct lp_type type,
+                        unsigned unit,
+                        unsigned num_coords,
+                        const LLVMValueRef *coords,
+                        LLVMValueRef lodbias,
+                        LLVMValueRef *texel);
+};
+
 
 void
 lp_build_tgsi_soa(LLVMBuilderRef builder,
                   const struct tgsi_token *tokens,
-                  union lp_type type,
+                  struct lp_type type,
                   struct lp_build_mask_context *mask,
                   LLVMValueRef consts_ptr,
                   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 d4d18febec..adc81569ed 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_tgsi_soa.c
@@ -78,6 +78,11 @@
 #define CHAN_Z 2
 #define CHAN_W 3
 
+#define QUAD_TOP_LEFT     0
+#define QUAD_TOP_RIGHT    1
+#define QUAD_BOTTOM_LEFT  2
+#define QUAD_BOTTOM_RIGHT 3
+
 
 struct lp_build_tgsi_soa_context
 {
@@ -88,8 +93,7 @@ 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];
@@ -98,6 +102,51 @@ struct lp_build_tgsi_soa_context
 };
 
 
+static const unsigned char
+swizzle_left[4] = {
+   QUAD_TOP_LEFT,     QUAD_TOP_LEFT,
+   QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_LEFT
+};
+
+static const unsigned char
+swizzle_right[4] = {
+   QUAD_TOP_RIGHT,    QUAD_TOP_RIGHT,
+   QUAD_BOTTOM_RIGHT, QUAD_BOTTOM_RIGHT
+};
+
+static const unsigned char
+swizzle_top[4] = {
+   QUAD_TOP_LEFT,     QUAD_TOP_RIGHT,
+   QUAD_TOP_LEFT,     QUAD_TOP_RIGHT
+};
+
+static const unsigned char
+swizzle_bottom[4] = {
+   QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_RIGHT,
+   QUAD_BOTTOM_LEFT,  QUAD_BOTTOM_RIGHT
+};
+
+
+static LLVMValueRef
+emit_ddx(struct lp_build_tgsi_soa_context *bld,
+         LLVMValueRef src)
+{
+   LLVMValueRef src_left  = lp_build_swizzle1_aos(&bld->base, src, swizzle_left);
+   LLVMValueRef src_right = lp_build_swizzle1_aos(&bld->base, src, swizzle_right);
+   return lp_build_sub(&bld->base, src_right, src_left);
+}
+
+
+static LLVMValueRef
+emit_ddy(struct lp_build_tgsi_soa_context *bld,
+         LLVMValueRef src)
+{
+   LLVMValueRef src_top    = lp_build_swizzle1_aos(&bld->base, src, swizzle_top);
+   LLVMValueRef src_bottom = lp_build_swizzle1_aos(&bld->base, src, swizzle_bottom);
+   return lp_build_sub(&bld->base, src_top, src_bottom);
+}
+
+
 /**
  * Register fetch.
  */
@@ -168,6 +217,7 @@ emit_fetch(
       break;
 
    case TGSI_UTIL_SIGN_SET:
+      /* TODO: Use bitwese OR for floating point */
       res = lp_build_abs( &bld->base, res );
       res = LLVMBuildNeg( bld->base.builder, res, "" );
       break;
@@ -185,6 +235,36 @@ emit_fetch(
 
 
 /**
+ * Register fetch with derivatives.
+ */
+static void
+emit_fetch_deriv(
+   struct lp_build_tgsi_soa_context *bld,
+   const struct tgsi_full_instruction *inst,
+   unsigned index,
+   const unsigned chan_index,
+   LLVMValueRef *res,
+   LLVMValueRef *ddx,
+   LLVMValueRef *ddy)
+{
+   LLVMValueRef src;
+
+   src = emit_fetch(bld, inst, index, chan_index);
+
+   if(res)
+      *res = src;
+
+   /* TODO: use interpolation coeffs for inputs */
+
+   if(ddx)
+      *ddx = emit_ddx(bld, src);
+
+   if(ddy)
+      *ddy = emit_ddy(bld, src);
+}
+
+
+/**
  * Register store.
  */
 static void
@@ -239,17 +319,18 @@ emit_store(
  * High-level instruction translators.
  */
 
+
 static void
 emit_tex( struct lp_build_tgsi_soa_context *bld,
           const struct tgsi_full_instruction *inst,
           boolean apply_lodbias,
-          boolean projected)
+          boolean projected,
+          LLVMValueRef *texel)
 {
    const uint unit = inst->FullSrcRegisters[1].SrcRegister.Index;
    LLVMValueRef lodbias;
    LLVMValueRef oow;
    LLVMValueRef coords[3];
-   LLVMValueRef texel[4];
    unsigned num_coords;
    unsigned i;
 
@@ -289,12 +370,11 @@ 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);
-
-   FOR_EACH_DST0_ENABLED_CHANNEL( inst, i ) {
-      emit_store( bld, inst, 0, i, texel[i] );
-   }
+   bld->sampler->emit_fetch_texel(bld->sampler,
+                                  bld->base.builder,
+                                  bld->base.type,
+                                  unit, num_coords, coords, lodbias,
+                                  texel);
 }
 
 
@@ -347,14 +427,6 @@ emit_kil(
 }
 
 
-static void
-emit_kilp(
-   struct lp_build_tgsi_soa_context *bld )
-{
-   /* XXX todo / fix me */
-}
-
-
 /**
  * Check if inst src/dest regs use indirect addressing into temporary
  * register file.
@@ -382,25 +454,35 @@ indirect_temp_reference(const struct tgsi_full_instruction *inst)
 static int
 emit_instruction(
    struct lp_build_tgsi_soa_context *bld,
-   struct tgsi_full_instruction *inst )
+   const struct tgsi_full_instruction *inst,
+   const struct tgsi_opcode_info *info)
 {
    unsigned chan_index;
    LLVMValueRef src0, src1, src2;
    LLVMValueRef tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
-   LLVMValueRef dst0;
+   LLVMValueRef res;
+   LLVMValueRef dst0[NUM_CHANNELS];
 
    /* we can't handle indirect addressing into temp register file yet */
    if (indirect_temp_reference(inst))
       return FALSE;
 
+   assert(info->num_dst <= 1);
+   if(info->num_dst) {
+      FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+         dst0[chan_index] = bld->base.undef;
+      }
+   }
+
    switch (inst->Instruction.Opcode) {
 #if 0
    case TGSI_OPCODE_ARL:
+      /* FIXME */
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          tmp0 = emit_fetch( bld, inst, 0, chan_index );
          emit_flr(bld, 0, 0);
          emit_f2it( bld, 0 );
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = tmp0;
       }
       break;
 #endif
@@ -408,19 +490,17 @@ emit_instruction(
    case TGSI_OPCODE_MOV:
    case TGSI_OPCODE_SWZ:
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         tmp0 = emit_fetch( bld, inst, 0, chan_index );
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = emit_fetch( bld, inst, 0, chan_index );
       }
       break;
 
    case TGSI_OPCODE_LIT:
       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) {
-         emit_store( bld, inst, 0, CHAN_X, bld->base.one);
+         dst0[CHAN_X] = bld->base.one;
       }
       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
          src0 = emit_fetch( bld, inst, 0, CHAN_X );
-         dst0 = lp_build_max( &bld->base, src0, bld->base.zero);
-         emit_store( bld, inst, 0, CHAN_Y, dst0);
+         dst0[CHAN_Y] = lp_build_max( &bld->base, src0, bld->base.zero);
       }
       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
          /* XMM[1] = SrcReg[0].yyyy */
@@ -432,20 +512,19 @@ emit_instruction(
          tmp1 = lp_build_pow( &bld->base, tmp1, tmp2);
          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
          tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero);
-         dst0 = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero);
-         emit_store( bld, inst, 0, CHAN_Z, dst0);
+         dst0[CHAN_Z] = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero);
       }
       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) {
-         emit_store( bld, inst, 0, CHAN_W, bld->base.one);
+         dst0[CHAN_W] = bld->base.one;
       }
       break;
 
    case TGSI_OPCODE_RCP:
    /* TGSI_OPCODE_RECIP */
       src0 = emit_fetch( bld, inst, 0, CHAN_X );
-      dst0 = lp_build_rcp(&bld->base, src0);
+      res = lp_build_rcp(&bld->base, src0);
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, dst0 );
+         dst0[chan_index] = res;
       }
       break;
 
@@ -453,9 +532,9 @@ emit_instruction(
    /* TGSI_OPCODE_RECIPSQRT */
       src0 = emit_fetch( bld, inst, 0, CHAN_X );
       src0 = lp_build_abs(&bld->base, src0);
-      dst0 = lp_build_rsqrt(&bld->base, src0);
+      res = lp_build_rsqrt(&bld->base, src0);
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, dst0 );
+         dst0[chan_index] = res;
       }
       break;
 
@@ -479,16 +558,15 @@ emit_instruction(
          lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2);
 
          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
-            emit_store( bld, inst, 0, CHAN_X, tmp0);
+            dst0[CHAN_X] = tmp0;
          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
-            emit_store( bld, inst, 0, CHAN_Y, tmp1);
+            dst0[CHAN_Y] = tmp1;
          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
-            emit_store( bld, inst, 0, CHAN_Z, tmp2);
+            dst0[CHAN_Z] = tmp2;
       }
       /* dst.w = 1.0 */
       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
-         tmp0 = bld->base.one;
-         emit_store( bld, inst, 0, CHAN_W, tmp0);
+         dst0[CHAN_W] = bld->base.one;
       }
       break;
 
@@ -514,20 +592,18 @@ emit_instruction(
 
          /* dst.x = floor(lg2(abs(src.x))) */
          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
-            emit_store( bld, inst, 0, CHAN_X, tmp0);
+            dst0[CHAN_X] = tmp0;
          /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */
          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) {
-            tmp1 = lp_build_div( &bld->base, src0, tmp1);
-            emit_store( bld, inst, 0, CHAN_Y, tmp1);
+            dst0[CHAN_Y] = lp_build_div( &bld->base, src0, tmp1);
          }
          /* dst.z = lg2(abs(src.x)) */
          if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
-            emit_store( bld, inst, 0, CHAN_Z, tmp2);
+            dst0[CHAN_Z] = tmp2;
       }
       /* dst.w = 1.0 */
       if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
-         tmp0 = bld->base.one;
-         emit_store( bld, inst, 0, CHAN_W, tmp0);
+         dst0[CHAN_W] = bld->base.one;
       }
       break;
 
@@ -535,8 +611,7 @@ emit_instruction(
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          src0 = emit_fetch( bld, inst, 0, chan_index );
          src1 = emit_fetch( bld, inst, 1, chan_index );
-         dst0 = lp_build_mul(&bld->base, src0, src1);
-         emit_store( bld, inst, 0, chan_index, dst0);
+         dst0[chan_index] = lp_build_mul(&bld->base, src0, src1);
       }
       break;
 
@@ -544,8 +619,7 @@ emit_instruction(
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          src0 = emit_fetch( bld, inst, 0, chan_index );
          src1 = emit_fetch( bld, inst, 1, chan_index );
-         dst0 = lp_build_add(&bld->base, src0, src1);
-         emit_store( bld, inst, 0, chan_index, dst0);
+         dst0[chan_index] = lp_build_add(&bld->base, src0, src1);
       }
       break;
 
@@ -563,7 +637,7 @@ emit_instruction(
       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = tmp0;
       }
       break;
 
@@ -585,28 +659,24 @@ emit_instruction(
       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = tmp0;
       }
       break;
 
    case TGSI_OPCODE_DST:
       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
-         tmp0 = bld->base.one;
-         emit_store( bld, inst, 0, CHAN_X, tmp0);
+         dst0[CHAN_X] = bld->base.one;
       }
       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
          tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
          tmp1 = emit_fetch( bld, inst, 1, CHAN_Y );
-         tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
-         emit_store( bld, inst, 0, CHAN_Y, tmp0);
+         dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp0, tmp1);
       }
       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
-         tmp0 = emit_fetch( bld, inst, 0, CHAN_Z );
-         emit_store( bld, inst, 0, CHAN_Z, tmp0);
+         dst0[CHAN_Z] = emit_fetch( bld, inst, 0, CHAN_Z );
       }
       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
-         tmp0 = emit_fetch( bld, inst, 1, CHAN_W );
-         emit_store( bld, inst, 0, CHAN_W, tmp0);
+         dst0[CHAN_W] = emit_fetch( bld, inst, 1, CHAN_W );
       }
       break;
 
@@ -614,8 +684,7 @@ emit_instruction(
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          src0 = emit_fetch( bld, inst, 0, chan_index );
          src1 = emit_fetch( bld, inst, 1, chan_index );
-         dst0 = lp_build_min( &bld->base, src0, src1 );
-         emit_store( bld, inst, 0, chan_index, dst0);
+         dst0[chan_index] = lp_build_min( &bld->base, src0, src1 );
       }
       break;
 
@@ -623,8 +692,7 @@ emit_instruction(
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          src0 = emit_fetch( bld, inst, 0, chan_index );
          src1 = emit_fetch( bld, inst, 1, chan_index );
-         dst0 = lp_build_max( &bld->base, src0, src1 );
-         emit_store( bld, inst, 0, chan_index, dst0);
+         dst0[chan_index] = lp_build_max( &bld->base, src0, src1 );
       }
       break;
 
@@ -634,8 +702,7 @@ emit_instruction(
          src0 = emit_fetch( bld, inst, 0, chan_index );
          src1 = emit_fetch( bld, inst, 1, chan_index );
          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, src1 );
-         dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
-         emit_store( bld, inst, 0, chan_index, dst0);
+         dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
       }
       break;
 
@@ -645,8 +712,7 @@ emit_instruction(
          src0 = emit_fetch( bld, inst, 0, chan_index );
          src1 = emit_fetch( bld, inst, 1, chan_index );
          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 );
-         dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
-         emit_store( bld, inst, 0, chan_index, dst0);
+         dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
       }
       break;
 
@@ -658,7 +724,7 @@ emit_instruction(
          tmp2 = emit_fetch( bld, inst, 2, chan_index );
          tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
          tmp0 = lp_build_add( &bld->base, tmp0, tmp2);
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = tmp0;
       }
       break;
 
@@ -666,8 +732,7 @@ emit_instruction(
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          tmp0 = emit_fetch( bld, inst, 0, chan_index );
          tmp1 = emit_fetch( bld, inst, 1, chan_index );
-         tmp0 = lp_build_sub( &bld->base, tmp0, tmp1);
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = lp_build_sub( &bld->base, tmp0, tmp1);
       }
       break;
 
@@ -678,13 +743,19 @@ emit_instruction(
          src2 = emit_fetch( bld, inst, 2, chan_index );
          tmp0 = lp_build_sub( &bld->base, src1, src2 );
          tmp0 = lp_build_mul( &bld->base, src0, tmp0 );
-         dst0 = lp_build_add( &bld->base, tmp0, src2 );
-         emit_store( bld, inst, 0, chan_index, dst0 );
+         dst0[chan_index] = lp_build_add( &bld->base, tmp0, src2 );
       }
       break;
 
    case TGSI_OPCODE_CND:
-      return 0;
+      FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+         src0 = emit_fetch( bld, inst, 0, chan_index );
+         src1 = emit_fetch( bld, inst, 1, chan_index );
+         src2 = emit_fetch( bld, inst, 2, chan_index );
+         tmp1 = lp_build_const_scalar(bld->base.type, 0.5);
+         tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src2, tmp1);
+         dst0[chan_index] = lp_build_select( &bld->base, tmp0, src0, src1 );
+      }
       break;
 
    case TGSI_OPCODE_DP2A:
@@ -698,45 +769,49 @@ emit_instruction(
       tmp1 = emit_fetch( bld, inst, 2, CHAN_X );  /* xmm1 = src[2].x */
       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, tmp0);  /* dest[ch] = xmm0 */
+         dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
       }
       break;
 
-#if 0
    case TGSI_OPCODE_FRC:
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         tmp0 = emit_fetch( bld, inst, 0, chan_index );
-         emit_frc( bld, 0, 0 );
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         src0 = emit_fetch( bld, inst, 0, chan_index );
+         tmp0 = lp_build_floor(&bld->base, src0);
+         tmp0 = lp_build_sub(&bld->base, tmp0, src0);
+         dst0[chan_index] = tmp0;
       }
       break;
 
    case TGSI_OPCODE_CLAMP:
-      return 0;
+      FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+         tmp0 = emit_fetch( bld, inst, 0, chan_index );
+         src1 = emit_fetch( bld, inst, 1, chan_index );
+         src2 = emit_fetch( bld, inst, 2, chan_index );
+         tmp0 = lp_build_max(&bld->base, tmp0, src1);
+         tmp0 = lp_build_min(&bld->base, tmp0, src2);
+         dst0[chan_index] = tmp0;
+      }
       break;
 
    case TGSI_OPCODE_FLR:
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          tmp0 = emit_fetch( bld, inst, 0, chan_index );
-         emit_flr( bld, 0, 0 );
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = lp_build_floor(&bld->base, tmp0);
       }
       break;
 
    case TGSI_OPCODE_ROUND:
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          tmp0 = emit_fetch( bld, inst, 0, chan_index );
-         emit_rnd( bld, 0, 0 );
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = lp_build_round(&bld->base, tmp0);
       }
       break;
-#endif
 
    case TGSI_OPCODE_EX2: {
       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
       tmp0 = lp_build_exp2( &bld->base, tmp0);
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = tmp0;
       }
       break;
    }
@@ -745,16 +820,16 @@ emit_instruction(
       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
       tmp0 = lp_build_log2( &bld->base, tmp0);
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = tmp0;
       }
       break;
 
    case TGSI_OPCODE_POW:
       src0 = emit_fetch( bld, inst, 0, CHAN_X );
       src1 = emit_fetch( bld, inst, 1, CHAN_X );
-      dst0 = lp_build_pow( &bld->base, src0, src1 );
+      res = lp_build_pow( &bld->base, src0, src1 );
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, dst0 );
+         dst0[chan_index] = res;
       }
       break;
 
@@ -775,7 +850,7 @@ emit_instruction(
          tmp5 = tmp3;
          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
          tmp2 = lp_build_sub( &bld->base, tmp2, tmp5);
-         emit_store( bld, inst, 0, CHAN_X, tmp2);
+         dst0[CHAN_X] = tmp2;
       }
       if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
           IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
@@ -786,31 +861,30 @@ emit_instruction(
          tmp3 = lp_build_mul( &bld->base, tmp3, tmp2);
          tmp1 = lp_build_mul( &bld->base, tmp1, tmp5);
          tmp3 = lp_build_sub( &bld->base, tmp3, tmp1);
-         emit_store( bld, inst, 0, CHAN_Y, tmp3);
+         dst0[CHAN_Y] = tmp3;
       }
       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
          tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
          tmp0 = lp_build_mul( &bld->base, tmp0, tmp2);
          tmp5 = lp_build_sub( &bld->base, tmp5, tmp0);
-         emit_store( bld, inst, 0, CHAN_Z, tmp5);
+         dst0[CHAN_Z] = tmp5;
       }
       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
-         tmp0 = bld->base.one;
-         emit_store( bld, inst, 0, CHAN_W, tmp0);
+         dst0[CHAN_W] = bld->base.one;
       }
       break;
 
    case TGSI_OPCODE_ABS:
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          tmp0 = emit_fetch( bld, inst, 0, chan_index );
-         tmp0 = lp_build_abs( &bld->base, tmp0 ) ;
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = lp_build_abs( &bld->base, tmp0 );
       }
       break;
 
    case TGSI_OPCODE_RCC:
+      /* deprecated? */
+      assert(0);
       return 0;
-      break;
 
    case TGSI_OPCODE_DPH:
       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
@@ -827,7 +901,7 @@ emit_instruction(
       tmp1 = emit_fetch( bld, inst, 1, CHAN_W );
       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = tmp0;
       }
       break;
 
@@ -835,25 +909,27 @@ emit_instruction(
       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
       tmp0 = lp_build_cos( &bld->base, tmp0 );
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = tmp0;
       }
       break;
 
    case TGSI_OPCODE_DDX:
-      return 0;
+      FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+         emit_fetch_deriv( bld, inst, 0, chan_index, NULL, &dst0[chan_index], NULL);
+      }
       break;
 
    case TGSI_OPCODE_DDY:
-      return 0;
+      FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+         emit_fetch_deriv( bld, inst, 0, chan_index, NULL, NULL, &dst0[chan_index]);
+      }
       break;
 
-#if 0
    case TGSI_OPCODE_KILP:
       /* predicated kill */
-      emit_kilp( bld );
-      return 0; /* XXX fix me */
+      /* FIXME */
+      return 0;
       break;
-#endif
 
    case TGSI_OPCODE_KIL:
       /* conditional kill */
@@ -885,13 +961,14 @@ emit_instruction(
          src0 = emit_fetch( bld, inst, 0, chan_index );
          src1 = emit_fetch( bld, inst, 1, chan_index );
          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 );
-         dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
-         emit_store( bld, inst, 0, chan_index, dst0);
+         dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
       }
       break;
 
    case TGSI_OPCODE_SFL:
-      return 0;
+      FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+         dst0[chan_index] = bld->base.zero;
+      }
       break;
 
    case TGSI_OPCODE_SGT:
@@ -899,8 +976,7 @@ emit_instruction(
          src0 = emit_fetch( bld, inst, 0, chan_index );
          src1 = emit_fetch( bld, inst, 1, chan_index );
          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 );
-         dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
-         emit_store( bld, inst, 0, chan_index, dst0);
+         dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
       }
       break;
 
@@ -908,7 +984,7 @@ emit_instruction(
       tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
       tmp0 = lp_build_sin( &bld->base, tmp0 );
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = tmp0;
       }
       break;
 
@@ -917,8 +993,7 @@ emit_instruction(
          src0 = emit_fetch( bld, inst, 0, chan_index );
          src1 = emit_fetch( bld, inst, 1, chan_index );
          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LEQUAL, src0, src1 );
-         dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
-         emit_store( bld, inst, 0, chan_index, dst0);
+         dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
       }
       break;
 
@@ -927,85 +1002,99 @@ emit_instruction(
          src0 = emit_fetch( bld, inst, 0, chan_index );
          src1 = emit_fetch( bld, inst, 1, chan_index );
          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 );
-         dst0 = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
-         emit_store( bld, inst, 0, chan_index, dst0);
+         dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
       }
       break;
 
    case TGSI_OPCODE_STR:
-      return 0;
+      FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+         dst0[chan_index] = bld->base.one;
+      }
       break;
 
    case TGSI_OPCODE_TEX:
-      emit_tex( bld, inst, FALSE, FALSE );
+      emit_tex( bld, inst, FALSE, FALSE, dst0 );
       break;
 
    case TGSI_OPCODE_TXD:
+      /* FIXME */
       return 0;
       break;
 
    case TGSI_OPCODE_UP2H:
+      /* deprecated */
+      assert (0);
       return 0;
       break;
 
    case TGSI_OPCODE_UP2US:
+      /* deprecated */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_UP4B:
+      /* deprecated */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_UP4UB:
+      /* deprecated */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_X2D:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_ARA:
+      /* deprecated */
+      assert(0);
       return 0;
       break;
 
 #if 0
    case TGSI_OPCODE_ARR:
+      /* FIXME */
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          tmp0 = emit_fetch( bld, inst, 0, chan_index );
          emit_rnd( bld, 0, 0 );
          emit_f2it( bld, 0 );
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = tmp0;
       }
       break;
 #endif
 
    case TGSI_OPCODE_BRA:
+      /* deprecated */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_CAL:
+      /* FIXME */
       return 0;
       break;
 
-#if 0
    case TGSI_OPCODE_RET:
-      emit_ret( bld );
+      /* FIXME */
+      return 0;
       break;
-#endif
 
    case TGSI_OPCODE_END:
       break;
 
-#if 0
    case TGSI_OPCODE_SSG:
    /* TGSI_OPCODE_SGN */
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          tmp0 = emit_fetch( bld, inst, 0, chan_index );
-         emit_sgn( bld, 0, 0 );
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = lp_build_sgn( &bld->base, tmp0 );
       }
       break;
-#endif
 
    case TGSI_OPCODE_CMP:
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
@@ -1013,34 +1102,29 @@ emit_instruction(
          src1 = emit_fetch( bld, inst, 1, chan_index );
          src2 = emit_fetch( bld, inst, 2, chan_index );
          tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero );
-         dst0 = lp_build_select( &bld->base, tmp0, src1, src2);
-         emit_store( bld, inst, 0, chan_index, dst0);
+         dst0[chan_index] = lp_build_select( &bld->base, tmp0, src1, src2);
       }
       break;
 
    case TGSI_OPCODE_SCS:
       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
-         tmp0 = lp_build_cos( &bld->base, tmp0 );
-         emit_store( bld, inst, 0, CHAN_X, tmp0);
+         dst0[CHAN_X] = lp_build_cos( &bld->base, tmp0 );
       }
       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
          tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
-         tmp0 = lp_build_sin( &bld->base, tmp0 );
-         emit_store( bld, inst, 0, CHAN_Y, tmp0);
+         dst0[CHAN_Y] = lp_build_sin( &bld->base, tmp0 );
       }
       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
-         tmp0 = bld->base.zero;
-         emit_store( bld, inst, 0, CHAN_Z, tmp0);
+         dst0[CHAN_Z] = bld->base.zero;
       }
       IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
-         tmp0 = bld->base.one;
-         emit_store( bld, inst, 0, CHAN_W, tmp0);
+         dst0[CHAN_W] = bld->base.one;
       }
       break;
 
    case TGSI_OPCODE_TXB:
-      emit_tex( bld, inst, TRUE, FALSE );
+      emit_tex( bld, inst, TRUE, FALSE, dst0 );
       break;
 
    case TGSI_OPCODE_NRM:
@@ -1099,38 +1183,35 @@ emit_instruction(
 
             /* dst.x = xmm1 * src.x */
             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
-               tmp4 = lp_build_mul( &bld->base, tmp4, tmp1);
-               emit_store(bld, inst, 0, CHAN_X, tmp4);
+               dst0[CHAN_X] = lp_build_mul( &bld->base, tmp4, tmp1);
             }
 
             /* dst.y = xmm1 * src.y */
             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
-               tmp5 = lp_build_mul( &bld->base, tmp5, tmp1);
-               emit_store(bld, inst, 0, CHAN_Y, tmp5);
+               dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp5, tmp1);
             }
 
             /* dst.z = xmm1 * src.z */
             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
-               tmp6 = lp_build_mul( &bld->base, tmp6, tmp1);
-               emit_store(bld, inst, 0, CHAN_Z, tmp6);
+               dst0[CHAN_Z] = lp_build_mul( &bld->base, tmp6, tmp1);
             }
 
             /* dst.w = xmm1 * src.w */
             if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) {
-               tmp7 = lp_build_mul( &bld->base, tmp7, tmp1);
-               emit_store(bld, inst, 0, CHAN_W, tmp7);
+               dst0[CHAN_W] = lp_build_mul( &bld->base, tmp7, tmp1);
             }
          }
 
-         /* dst0.w = 1.0 */
+         /* dst.w = 1.0 */
          if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) {
-            tmp0 = bld->base.one;
-            emit_store(bld, inst, 0, CHAN_W, tmp0);
+            dst0[CHAN_W] = bld->base.one;
          }
       }
       break;
 
    case TGSI_OPCODE_DIV:
+      /* deprecated */
+      assert( 0 );
       return 0;
       break;
 
@@ -1143,118 +1224,157 @@ emit_instruction(
       tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);              /* xmm1 = xmm1 * xmm2 */
       tmp0 = lp_build_add( &bld->base, tmp0, tmp1);              /* xmm0 = xmm0 + xmm1 */
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
-         emit_store( bld, inst, 0, chan_index, tmp0);  /* dest[ch] = xmm0 */
+         dst0[chan_index] = tmp0;  /* dest[ch] = xmm0 */
       }
       break;
 
    case TGSI_OPCODE_TXL:
-      emit_tex( bld, inst, TRUE, FALSE );
+      emit_tex( bld, inst, TRUE, FALSE, dst0 );
       break;
 
    case TGSI_OPCODE_TXP:
-      emit_tex( bld, inst, FALSE, TRUE );
+      emit_tex( bld, inst, FALSE, TRUE, dst0 );
       break;
       
    case TGSI_OPCODE_BRK:
+      /* FIXME */
       return 0;
       break;
 
    case TGSI_OPCODE_IF:
+      /* FIXME */
       return 0;
       break;
 
    case TGSI_OPCODE_BGNFOR:
+      /* deprecated */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_REP:
+      /* deprecated */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_ELSE:
+      /* FIXME */
       return 0;
       break;
 
    case TGSI_OPCODE_ENDIF:
+      /* FIXME */
       return 0;
       break;
 
    case TGSI_OPCODE_ENDFOR:
+      /* deprecated */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_ENDREP:
+      /* deprecated */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_PUSHA:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_POPA:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_CEIL:
-      return 0;
+      FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+         tmp0 = emit_fetch( bld, inst, 0, chan_index );
+         dst0[chan_index] = lp_build_ceil(&bld->base, tmp0);
+      }
       break;
 
    case TGSI_OPCODE_I2F:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_NOT:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
-#if 0
    case TGSI_OPCODE_TRUNC:
       FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
          tmp0 = emit_fetch( bld, inst, 0, chan_index );
-         emit_f2it( bld, 0 );
-         emit_i2f( bld, 0 );
-         emit_store( bld, inst, 0, chan_index, tmp0);
+         dst0[chan_index] = lp_build_trunc(&bld->base, tmp0);
       }
       break;
-#endif
 
    case TGSI_OPCODE_SHL:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_SHR:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_AND:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_OR:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_MOD:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_XOR:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_SAD:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_TXF:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_TXQ:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
    case TGSI_OPCODE_CONT:
+      /* deprecated? */
+      assert(0);
       return 0;
       break;
 
@@ -1266,10 +1386,28 @@ emit_instruction(
       return 0;
       break;
 
+   case TGSI_OPCODE_NOISE1:
+   case TGSI_OPCODE_NOISE2:
+   case TGSI_OPCODE_NOISE3:
+   case TGSI_OPCODE_NOISE4:
+      FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+         dst0[chan_index] = bld->base.zero;
+      }
+      break;
+
+   case TGSI_OPCODE_NOP:
+      break;
+
    default:
       return 0;
    }
    
+   if(info->num_dst) {
+      FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+         emit_store( bld, inst, 0, chan_index, dst0[chan_index]);
+      }
+   }
+
    return 1;
 }
 
@@ -1277,14 +1415,13 @@ emit_instruction(
 void
 lp_build_tgsi_soa(LLVMBuilderRef builder,
                   const struct tgsi_token *tokens,
-                  union lp_type type,
+                  struct lp_type type,
                   struct lp_build_mask_context *mask,
                   LLVMValueRef consts_ptr,
                   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;
@@ -1299,8 +1436,7 @@ 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 );
 
@@ -1309,16 +1445,18 @@ lp_build_tgsi_soa(LLVMBuilderRef builder,
 
       switch( parse.FullToken.Token.Type ) {
       case TGSI_TOKEN_TYPE_DECLARATION:
-         /* Input already interpolated */
+         /* Inputs already interpolated */
          break;
 
       case TGSI_TOKEN_TYPE_INSTRUCTION:
-         if (!emit_instruction( &bld, &parse.FullToken.FullInstruction )) {
+         {
             unsigned opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
             const struct tgsi_opcode_info *info = tgsi_get_opcode_info(opcode);
-	    _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
-	                  info ? info->mnemonic : "<invalid>");
-	 }
+            if (!emit_instruction( &bld, &parse.FullToken.FullInstruction, info ))
+               _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
+                             info ? info->mnemonic : "<invalid>");
+         }
+
          break;
 
       case TGSI_TOKEN_TYPE_IMMEDIATE:
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_type.c b/src/gallium/drivers/llvmpipe/lp_bld_type.c
index 8e0026fd97..606243d6c5 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_type.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_type.c
@@ -33,7 +33,7 @@
 
 
 LLVMTypeRef
-lp_build_elem_type(union lp_type type)
+lp_build_elem_type(struct lp_type type)
 {
    if (type.floating) {
       switch(type.width) {
@@ -55,7 +55,7 @@ lp_build_elem_type(union lp_type type)
 
 
 LLVMTypeRef
-lp_build_vec_type(union lp_type type)
+lp_build_vec_type(struct lp_type type)
 {
    LLVMTypeRef elem_type = lp_build_elem_type(type);
    return LLVMVectorType(elem_type, type.length);
@@ -69,7 +69,7 @@ lp_build_vec_type(union lp_type type)
  * type and check for identity.
  */
 boolean
-lp_check_elem_type(union lp_type type, LLVMTypeRef elem_type) 
+lp_check_elem_type(struct lp_type type, LLVMTypeRef elem_type) 
 {
    LLVMTypeKind elem_kind;
 
@@ -107,7 +107,7 @@ lp_check_elem_type(union lp_type type, LLVMTypeRef elem_type)
 
 
 boolean
-lp_check_vec_type(union lp_type type, LLVMTypeRef vec_type) 
+lp_check_vec_type(struct lp_type type, LLVMTypeRef vec_type) 
 {
    LLVMTypeRef elem_type;
 
@@ -128,7 +128,7 @@ lp_check_vec_type(union lp_type type, LLVMTypeRef vec_type)
 
 
 boolean
-lp_check_value(union lp_type type, LLVMValueRef val) 
+lp_check_value(struct lp_type type, LLVMValueRef val) 
 {
    LLVMTypeRef vec_type;
 
@@ -143,24 +143,36 @@ lp_check_value(union lp_type type, LLVMValueRef val)
 
 
 LLVMTypeRef
-lp_build_int_elem_type(union lp_type type)
+lp_build_int_elem_type(struct lp_type type)
 {
    return LLVMIntType(type.width);
 }
 
 
 LLVMTypeRef
-lp_build_int_vec_type(union lp_type type)
+lp_build_int_vec_type(struct lp_type type)
 {
    LLVMTypeRef elem_type = lp_build_int_elem_type(type);
    return LLVMVectorType(elem_type, type.length);
 }
 
 
+struct lp_type
+lp_int_type(struct lp_type type)
+{
+   struct lp_type int_type;
+
+   memset(&int_type, 0, sizeof int_type);
+   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,
-                      union lp_type type)
+                      struct lp_type type)
 {
    bld->builder = builder;
    bld->type = type;
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_type.h b/src/gallium/drivers/llvmpipe/lp_bld_type.h
index 3ce566be64..ee5ca3483c 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_type.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_type.h
@@ -56,58 +56,55 @@
  * on the types used for intermediate computations, such as signed vs unsigned,
  * normalized values, or fixed point.
  */
-union lp_type {
-   struct {
-      /** 
-       * Floating-point. Cannot be used with fixed. Integer numbers are
-       * represented by this zero.
-       */
-      unsigned floating:1;
-
-      /** 
-       * Fixed-point. Cannot be used with floating. Integer numbers are
-       * represented by this zero.
-       */
-      unsigned fixed:1;
-      
-      /** 
-       * Whether it can represent negative values or not.
-       *
-       * If this is not set for floating point, it means that all values are
-       * assumed to be positive.
-       */
-      unsigned sign:1;
-
-      /**
-       * Whether values are normalized to fit [0, 1] interval, or [-1, 1]
-       * interval for signed types.
-       *
-       * For integer types it means the representable integer range should be
-       * interpreted as the interval above.
-       *
-       * For floating and fixed point formats it means the values should be
-       * clamped to the interval above.
-       */
-      unsigned norm:1;
-
-      /**
-       * Element width.
-       *
-       * For fixed point values, the fixed point is assumed to be at half the
-       * width.
-       */
-      unsigned width:14;
-
-      /** 
-       * Vector length.
-       *
-       * width*length should be a power of two greater or equal to eight.
-       *
-       * @sa LP_MAX_VECTOR_LENGTH
-       */
-      unsigned length:14;
-   };
-   uint32_t value;
+struct lp_type {
+   /**
+    * Floating-point. Cannot be used with fixed. Integer numbers are
+    * represented by this zero.
+    */
+   unsigned floating:1;
+
+   /**
+    * Fixed-point. Cannot be used with floating. Integer numbers are
+    * represented by this zero.
+    */
+   unsigned fixed:1;
+
+   /**
+    * Whether it can represent negative values or not.
+    *
+    * If this is not set for floating point, it means that all values are
+    * assumed to be positive.
+    */
+   unsigned sign:1;
+
+   /**
+    * Whether values are normalized to fit [0, 1] interval, or [-1, 1]
+    * interval for signed types.
+    *
+    * For integer types it means the representable integer range should be
+    * interpreted as the interval above.
+    *
+    * For floating and fixed point formats it means the values should be
+    * clamped to the interval above.
+    */
+   unsigned norm:1;
+
+   /**
+    * Element width.
+    *
+    * For fixed point values, the fixed point is assumed to be at half the
+    * width.
+    */
+   unsigned width:14;
+
+   /**
+    * Vector length.
+    *
+    * width*length should be a power of two greater or equal to eight.
+    *
+    * @sa LP_MAX_VECTOR_LENGTH
+    */
+   unsigned length:14;
 };
 
 
@@ -124,7 +121,7 @@ struct lp_build_context
     * This not only describes the input/output LLVM types, but also whether
     * to normalize/clamp the results.
     */
-   union lp_type type;
+   struct lp_type type;
 
    /** Same as lp_build_undef(type) */
    LLVMValueRef undef;
@@ -138,37 +135,41 @@ struct lp_build_context
 
 
 LLVMTypeRef
-lp_build_elem_type(union lp_type type);
+lp_build_elem_type(struct lp_type type);
 
 
 LLVMTypeRef
-lp_build_vec_type(union lp_type type);
+lp_build_vec_type(struct lp_type type);
 
 
 boolean
-lp_check_elem_type(union lp_type type, LLVMTypeRef elem_type);
+lp_check_elem_type(struct lp_type type, LLVMTypeRef elem_type);
 
 
 boolean
-lp_check_vec_type(union lp_type type, LLVMTypeRef vec_type);
+lp_check_vec_type(struct lp_type type, LLVMTypeRef vec_type);
 
 
 boolean
-lp_check_value(union lp_type type, LLVMValueRef val);
+lp_check_value(struct lp_type type, LLVMValueRef val);
 
 
 LLVMTypeRef
-lp_build_int_elem_type(union lp_type type);
+lp_build_int_elem_type(struct lp_type type);
 
 
 LLVMTypeRef
-lp_build_int_vec_type(union lp_type type);
+lp_build_int_vec_type(struct lp_type type);
+
+
+struct lp_type
+lp_int_type(struct lp_type type);
 
 
 void
 lp_build_context_init(struct lp_build_context *bld,
                       LLVMBuilderRef builder,
-                      union lp_type type);
+                      struct lp_type type);
 
 
 #endif /* !LP_BLD_TYPE_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_clear.c b/src/gallium/drivers/llvmpipe/lp_clear.c
index 580cca5b46..bdcff94b9b 100644
--- a/src/gallium/drivers/llvmpipe/lp_clear.c
+++ b/src/gallium/drivers/llvmpipe/lp_clear.c
@@ -67,6 +67,7 @@ llvmpipe_clear(struct pipe_context *pipe, unsigned buffers, const float *rgba,
          util_pack_color(rgba, ps->format, &cv);
          lp_tile_cache_clear(llvmpipe->cbuf_cache[i], rgba, cv);
       }
+      llvmpipe->dirty_render_cache = TRUE;
    }
 
    if (buffers & PIPE_CLEAR_DEPTHSTENCIL) {
diff --git a/src/gallium/drivers/llvmpipe/lp_context.c b/src/gallium/drivers/llvmpipe/lp_context.c
index 233d1df0e1..a4b2bd8c2a 100644
--- a/src/gallium/drivers/llvmpipe/lp_context.c
+++ b/src/gallium/drivers/llvmpipe/lp_context.c
@@ -141,8 +141,6 @@ llvmpipe_is_texture_referenced( struct pipe_context *pipe,
          return PIPE_REFERENCED_FOR_WRITE;
    }
    
-   /* FIXME: we also need to do the same for the texture cache */
-   
    return PIPE_UNREFERENCED;
 }
 
diff --git a/src/gallium/drivers/llvmpipe/lp_jit.c b/src/gallium/drivers/llvmpipe/lp_jit.c
index d288460a1b..b4a22ff4a9 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);
 
@@ -117,7 +152,7 @@ lp_jit_screen_init(struct llvmpipe_screen *screen)
    screen->provider = LLVMCreateModuleProviderForExistingModule(screen->module);
 
    if (LLVMCreateJITCompiler(&screen->engine, screen->provider, 1, &error)) {
-      fprintf(stderr, "%s\n", error);
+      _debug_printf("%s\n", error);
       LLVMDisposeMessage(error);
       abort();
    }
diff --git a/src/gallium/drivers/llvmpipe/lp_jit.h b/src/gallium/drivers/llvmpipe/lp_jit.h
index a7fb60f9f5..58f716ede2 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.
  *
@@ -60,11 +80,12 @@ struct lp_jit_context
 
    struct tgsi_sampler **samplers;
 
-   /* TODO: alpha reference value */
    float alpha_ref_value;
 
-   /* TODO: blend constant color */
+   /* FIXME: store (also?) in floats */
    uint8_t *blend_color;
+
+   struct lp_jit_texture textures[PIPE_MAX_SAMPLERS];
 };
 
 
@@ -80,6 +101,11 @@ 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_screen.c b/src/gallium/drivers/llvmpipe/lp_screen.c
index 125035771e..ff7ef8658a 100644
--- a/src/gallium/drivers/llvmpipe/lp_screen.c
+++ b/src/gallium/drivers/llvmpipe/lp_screen.c
@@ -27,8 +27,6 @@
 
 
 #include "util/u_memory.h"
-#include "util/u_simple_screen.h"
-#include "pipe/internal/p_winsys_screen.h"
 #include "pipe/p_defines.h"
 #include "pipe/p_screen.h"
 
@@ -67,8 +65,6 @@ llvmpipe_get_param(struct pipe_screen *screen, int param)
       return 1;
    case PIPE_CAP_GLSL:
       return 1;
-   case PIPE_CAP_S3TC:
-      return 0;
    case PIPE_CAP_ANISOTROPIC_FILTER:
       return 0;
    case PIPE_CAP_POINT_SPRITE:
@@ -196,8 +192,7 @@ static void
 llvmpipe_destroy_screen( struct pipe_screen *_screen )
 {
    struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
-
-   struct pipe_winsys *winsys = _screen->winsys;
+   struct llvmpipe_winsys *winsys = screen->winsys;
 
    lp_jit_screen_cleanup(screen);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_setup.c b/src/gallium/drivers/llvmpipe/lp_setup.c
index d145f6d6bb..2d2fc19a65 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"
 
@@ -162,12 +163,12 @@ shade_quads(struct llvmpipe_context *llvmpipe,
    else
       depth = NULL;
 
-   /* TODO: blend color */
+   /* XXX: This will most likely fail on 32bit x86 without -mstackrealign */
+   assert(lp_check_alignment(mask, 16));
 
-   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);
+   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..30fb41ea65 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_derived.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_derived.c
@@ -93,6 +93,23 @@ llvmpipe_get_vertex_info(struct llvmpipe_context *llvmpipe)
       vinfo->num_attribs = 0;
       for (i = 0; i < lpfs->info.num_inputs; i++) {
          int src;
+         enum interp_mode interp;
+
+         switch (lpfs->info.input_interpolate[i]) {
+         case TGSI_INTERPOLATE_CONSTANT:
+            interp = INTERP_CONSTANT;
+            break;
+         case TGSI_INTERPOLATE_LINEAR:
+            interp = INTERP_LINEAR;
+            break;
+         case TGSI_INTERPOLATE_PERSPECTIVE:
+            interp = INTERP_PERSPECTIVE;
+            break;
+         default:
+            assert(0);
+            interp = INTERP_LINEAR;
+         }
+
          switch (lpfs->info.input_semantic_name[i]) {
          case TGSI_SEMANTIC_POSITION:
             src = draw_find_vs_output(llvmpipe->draw,
@@ -108,7 +125,7 @@ llvmpipe_get_vertex_info(struct llvmpipe_context *llvmpipe)
 
          case TGSI_SEMANTIC_FOG:
             src = draw_find_vs_output(llvmpipe->draw, TGSI_SEMANTIC_FOG, 0);
-            draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE, src);
+            draw_emit_vertex_attr(vinfo, EMIT_4F, interp, src);
             break;
 
          case TGSI_SEMANTIC_GENERIC:
@@ -116,7 +133,7 @@ llvmpipe_get_vertex_info(struct llvmpipe_context *llvmpipe)
             /* this includes texcoords and varying vars */
             src = draw_find_vs_output(llvmpipe->draw, TGSI_SEMANTIC_GENERIC,
                                       lpfs->info.input_semantic_index[i]);
-            draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE, src);
+            draw_emit_vertex_attr(vinfo, EMIT_4F, interp, src);
             break;
 
          default:
@@ -250,7 +267,9 @@ 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..9faed5a0b1 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};
@@ -130,21 +131,20 @@ generate_pos0(LLVMBuilderRef builder,
  * Generate the depth test.
  */
 static void
-generate_depth(struct llvmpipe_context *lp,
-               LLVMBuilderRef builder,
-               const struct pipe_depth_state *state,
-               union lp_type src_type,
+generate_depth(LLVMBuilderRef builder,
+               const struct lp_fragment_shader_variant_key *key,
+               struct lp_type src_type,
                struct lp_build_mask_context *mask,
                LLVMValueRef src,
                LLVMValueRef dst_ptr)
 {
    const struct util_format_description *format_desc;
-   union lp_type dst_type;
+   struct 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.
  */
@@ -282,11 +181,11 @@ generate_fs(struct llvmpipe_context *lp,
             struct lp_fragment_shader *shader,
             const struct lp_fragment_shader_variant_key *key,
             LLVMBuilderRef builder,
-            union lp_type type,
+            struct lp_type type,
             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;
 
 }
@@ -385,13 +299,15 @@ generate_fs(struct llvmpipe_context *lp,
 static void
 generate_blend(const struct pipe_blend_state *blend,
                LLVMBuilderRef builder,
-               union lp_type type,
+               struct lp_type type,
                LLVMValueRef context_ptr,
                LLVMValueRef mask,
                LLVMValueRef *src,
                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);
 }
 
 
@@ -440,8 +364,8 @@ generate_fragment(struct llvmpipe_context *lp,
 {
    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
    struct lp_fragment_shader_variant *variant;
-   union lp_type fs_type;
-   union lp_type blend_type;
+   struct lp_type fs_type;
+   struct lp_type blend_type;
    LLVMTypeRef fs_elem_type;
    LLVMTypeRef fs_vec_type;
    LLVMTypeRef fs_int_vec_type;
@@ -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];
@@ -507,7 +431,7 @@ generate_fragment(struct llvmpipe_context *lp,
    /* TODO: actually pick these based on the fs and color buffer
     * characteristics. */
 
-   fs_type.value = 0;
+   memset(&fs_type, 0, sizeof fs_type);
    fs_type.floating = TRUE; /* floating point values */
    fs_type.sign = TRUE;     /* values are signed */
    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
@@ -515,7 +439,7 @@ generate_fragment(struct llvmpipe_context *lp,
    fs_type.length = 4;      /* 4 element per vector */
    num_fs = 4;
 
-   blend_type.value = 0;
+   memset(&blend_type, 0, sizeof blend_type);
    blend_type.floating = FALSE; /* values are integers */
    blend_type.sign = FALSE;     /* values are unsigned */
    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
@@ -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.h b/src/gallium/drivers/llvmpipe/lp_test.h
index 69aaae26e0..a88e110c66 100644
--- a/src/gallium/drivers/llvmpipe/lp_test.h
+++ b/src/gallium/drivers/llvmpipe/lp_test.h
@@ -86,43 +86,43 @@ random_float(void);
 
 
 void
-dump_type(FILE *fp, union lp_type type);
+dump_type(FILE *fp, struct lp_type type);
 
 
 double
-read_elem(union lp_type type, const void *src, unsigned index);
+read_elem(struct lp_type type, const void *src, unsigned index);
 
 
 void
-write_elem(union lp_type type, void *dst, unsigned index, double src);
+write_elem(struct lp_type type, void *dst, unsigned index, double src);
 
 
 void
-random_elem(union lp_type type, void *dst, unsigned index);
+random_elem(struct lp_type type, void *dst, unsigned index);
 
 
 void
-read_vec(union lp_type type, const void *src, double *dst);
+read_vec(struct lp_type type, const void *src, double *dst);
 
 
 void
-write_vec(union lp_type type, void *dst, const double *src);
+write_vec(struct lp_type type, void *dst, const double *src);
 
 
 void
-random_vec(union lp_type type, void *dst);
+random_vec(struct lp_type type, void *dst);
 
 
 boolean
-compare_vec_with_eps(union lp_type type, const void *res, const void *ref, double eps);
+compare_vec_with_eps(struct lp_type type, const void *res, const void *ref, double eps);
 
 
 boolean
-compare_vec(union lp_type type, const void *res, const void *ref);
+compare_vec(struct lp_type type, const void *res, const void *ref);
 
 
 void
-dump_vec(FILE *fp, union lp_type type, const void *src);
+dump_vec(FILE *fp, struct lp_type type, const void *src);
 
 
 #endif /* !LP_TEST_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_test_blend.c b/src/gallium/drivers/llvmpipe/lp_test_blend.c
index 8dfad468e3..94b661dcba 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_blend.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_blend.c
@@ -80,7 +80,7 @@ static void
 write_tsv_row(FILE *fp,
               const struct pipe_blend_state *blend,
               enum vector_mode mode,
-              union lp_type type,
+              struct lp_type type,
               double cycles,
               boolean success)
 {
@@ -125,7 +125,7 @@ static void
 dump_blend_type(FILE *fp,
                 const struct pipe_blend_state *blend,
                 enum vector_mode mode,
-                union lp_type type)
+                struct lp_type type)
 {
    fprintf(fp, "%s", mode ? "soa" : "aos");
 
@@ -153,7 +153,7 @@ static LLVMValueRef
 add_blend_test(LLVMModuleRef module,
                const struct pipe_blend_state *blend,
                enum vector_mode mode,
-               union lp_type type)
+               struct lp_type type)
 {
    LLVMTypeRef ret_type;
    LLVMTypeRef vec_type;
@@ -467,7 +467,7 @@ test_one(unsigned verbose,
          FILE *fp,
          const struct pipe_blend_state *blend,
          enum vector_mode mode,
-         union lp_type type)
+         struct lp_type type)
 {
    LLVMModuleRef module = NULL;
    LLVMValueRef func = NULL;
@@ -765,10 +765,10 @@ blend_funcs[] = {
 };
 
 
-const union lp_type blend_types[] = {
+const struct lp_type blend_types[] = {
    /* float, fixed,  sign,  norm, width, len */
-   {{  TRUE, FALSE, FALSE,  TRUE,    32,   4 }}, /* f32 x 4 */
-   {{ FALSE, FALSE, FALSE,  TRUE,     8,  16 }}, /* u8n x 16 */
+   {   TRUE, FALSE, FALSE,  TRUE,    32,   4 }, /* f32 x 4 */
+   {  FALSE, FALSE, FALSE,  TRUE,     8,  16 }, /* u8n x 16 */
 };
 
 
@@ -788,7 +788,7 @@ test_all(unsigned verbose, FILE *fp)
    const unsigned *alpha_dst_factor;
    struct pipe_blend_state blend;
    enum vector_mode mode;
-   const union lp_type *type;
+   const struct lp_type *type;
    bool success = TRUE;
 
    for(rgb_func = blend_funcs; rgb_func < &blend_funcs[num_funcs]; ++rgb_func) {
@@ -841,27 +841,27 @@ test_some(unsigned verbose, FILE *fp, unsigned long n)
    const unsigned *alpha_dst_factor;
    struct pipe_blend_state blend;
    enum vector_mode mode;
-   const union lp_type *type;
+   const struct lp_type *type;
    unsigned long i;
    bool success = TRUE;
 
    for(i = 0; i < n; ++i) {
-      rgb_func = &blend_funcs[random() % num_funcs];
-      alpha_func = &blend_funcs[random() % num_funcs];
-      rgb_src_factor = &blend_factors[random() % num_factors];
-      alpha_src_factor = &blend_factors[random() % num_factors];
+      rgb_func = &blend_funcs[rand() % num_funcs];
+      alpha_func = &blend_funcs[rand() % num_funcs];
+      rgb_src_factor = &blend_factors[rand() % num_factors];
+      alpha_src_factor = &blend_factors[rand() % num_factors];
       
       do {
-         rgb_dst_factor = &blend_factors[random() % num_factors];
+         rgb_dst_factor = &blend_factors[rand() % num_factors];
       } while(*rgb_dst_factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE);
 
       do {
-         alpha_dst_factor = &blend_factors[random() % num_factors];
+         alpha_dst_factor = &blend_factors[rand() % num_factors];
       } while(*alpha_dst_factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE);
 
-      mode = random() & 1;
+      mode = rand() & 1;
 
-      type = &blend_types[random() % num_types];
+      type = &blend_types[rand() % num_types];
 
       memset(&blend, 0, sizeof blend);
       blend.blend_enable      = 1;
diff --git a/src/gallium/drivers/llvmpipe/lp_test_conv.c b/src/gallium/drivers/llvmpipe/lp_test_conv.c
index e6489834af..9dcf58e5dc 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_conv.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_conv.c
@@ -59,8 +59,8 @@ write_tsv_header(FILE *fp)
 
 static void
 write_tsv_row(FILE *fp,
-              union lp_type src_type,
-              union lp_type dst_type,
+              struct lp_type src_type,
+              struct lp_type dst_type,
               double cycles,
               boolean success)
 {
@@ -80,8 +80,8 @@ write_tsv_row(FILE *fp,
 
 static void
 dump_conv_types(FILE *fp,
-               union lp_type src_type,
-               union lp_type dst_type)
+               struct lp_type src_type,
+               struct lp_type dst_type)
 {
    fprintf(fp, "src_type=");
    dump_type(fp, src_type);
@@ -96,8 +96,8 @@ dump_conv_types(FILE *fp,
 
 static LLVMValueRef
 add_conv_test(LLVMModuleRef module,
-              union lp_type src_type, unsigned num_srcs,
-              union lp_type dst_type, unsigned num_dsts)
+              struct lp_type src_type, unsigned num_srcs,
+              struct lp_type dst_type, unsigned num_dsts)
 {
    LLVMTypeRef args[2];
    LLVMValueRef func;
@@ -145,8 +145,8 @@ add_conv_test(LLVMModuleRef module,
 static boolean
 test_one(unsigned verbose,
          FILE *fp,
-         union lp_type src_type,
-         union lp_type dst_type)
+         struct lp_type src_type,
+         struct lp_type dst_type)
 {
    LLVMModuleRef module = NULL;
    LLVMValueRef func = NULL;
@@ -343,35 +343,35 @@ test_one(unsigned verbose,
 }
 
 
-const union lp_type conv_types[] = {
+const struct lp_type conv_types[] = {
    /* float, fixed,  sign,  norm, width, len */
 
-   {{  TRUE, FALSE,  TRUE,  TRUE,    32,   4 }},
-   {{  TRUE, FALSE,  TRUE, FALSE,    32,   4 }},
-   {{  TRUE, FALSE, FALSE,  TRUE,    32,   4 }},
-   {{  TRUE, FALSE, FALSE, FALSE,    32,   4 }},
+   {   TRUE, FALSE,  TRUE,  TRUE,    32,   4 },
+   {   TRUE, FALSE,  TRUE, FALSE,    32,   4 },
+   {   TRUE, FALSE, FALSE,  TRUE,    32,   4 },
+   {   TRUE, FALSE, FALSE, FALSE,    32,   4 },
 
    /* TODO: test fixed formats too */
 
-   {{ FALSE, FALSE,  TRUE,  TRUE,    16,   8 }},
-   {{ FALSE, FALSE,  TRUE, FALSE,    16,   8 }},
-   {{ FALSE, FALSE, FALSE,  TRUE,    16,   8 }},
-   {{ FALSE, FALSE, FALSE, FALSE,    16,   8 }},
-
-   {{ FALSE, FALSE,  TRUE,  TRUE,    32,   4 }},
-   {{ FALSE, FALSE,  TRUE, FALSE,    32,   4 }},
-   {{ FALSE, FALSE, FALSE,  TRUE,    32,   4 }},
-   {{ FALSE, FALSE, FALSE, FALSE,    32,   4 }},
-
-   {{ FALSE, FALSE,  TRUE,  TRUE,    16,   8 }},
-   {{ FALSE, FALSE,  TRUE, FALSE,    16,   8 }},
-   {{ FALSE, FALSE, FALSE,  TRUE,    16,   8 }},
-   {{ FALSE, FALSE, FALSE, FALSE,    16,   8 }},
-
-   {{ FALSE, FALSE,  TRUE,  TRUE,     8,  16 }},
-   {{ FALSE, FALSE,  TRUE, FALSE,     8,  16 }},
-   {{ FALSE, FALSE, FALSE,  TRUE,     8,  16 }},
-   {{ FALSE, FALSE, FALSE, FALSE,     8,  16 }},
+   {  FALSE, FALSE,  TRUE,  TRUE,    16,   8 },
+   {  FALSE, FALSE,  TRUE, FALSE,    16,   8 },
+   {  FALSE, FALSE, FALSE,  TRUE,    16,   8 },
+   {  FALSE, FALSE, FALSE, FALSE,    16,   8 },
+
+   {  FALSE, FALSE,  TRUE,  TRUE,    32,   4 },
+   {  FALSE, FALSE,  TRUE, FALSE,    32,   4 },
+   {  FALSE, FALSE, FALSE,  TRUE,    32,   4 },
+   {  FALSE, FALSE, FALSE, FALSE,    32,   4 },
+
+   {  FALSE, FALSE,  TRUE,  TRUE,    16,   8 },
+   {  FALSE, FALSE,  TRUE, FALSE,    16,   8 },
+   {  FALSE, FALSE, FALSE,  TRUE,    16,   8 },
+   {  FALSE, FALSE, FALSE, FALSE,    16,   8 },
+
+   {  FALSE, FALSE,  TRUE,  TRUE,     8,  16 },
+   {  FALSE, FALSE,  TRUE, FALSE,     8,  16 },
+   {  FALSE, FALSE, FALSE,  TRUE,     8,  16 },
+   {  FALSE, FALSE, FALSE, FALSE,     8,  16 },
 };
 
 
@@ -381,8 +381,8 @@ const unsigned num_types = sizeof(conv_types)/sizeof(conv_types[0]);
 boolean
 test_all(unsigned verbose, FILE *fp)
 {
-   const union lp_type *src_type;
-   const union lp_type *dst_type;
+   const struct lp_type *src_type;
+   const struct lp_type *dst_type;
    bool success = TRUE;
 
    for(src_type = conv_types; src_type < &conv_types[num_types]; ++src_type) {
@@ -407,16 +407,16 @@ test_all(unsigned verbose, FILE *fp)
 boolean
 test_some(unsigned verbose, FILE *fp, unsigned long n)
 {
-   const union lp_type *src_type;
-   const union lp_type *dst_type;
+   const struct lp_type *src_type;
+   const struct lp_type *dst_type;
    unsigned long i;
    bool success = TRUE;
 
    for(i = 0; i < n; ++i) {
-      src_type = &conv_types[random() % num_types];
+      src_type = &conv_types[rand() % num_types];
       
       do {
-         dst_type = &conv_types[random() % num_types];
+         dst_type = &conv_types[rand() % num_types];
       } while (src_type == dst_type || src_type->norm != dst_type->norm);
 
       if(!test_one(verbose, fp, *src_type, *dst_type))
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_test_main.c b/src/gallium/drivers/llvmpipe/lp_test_main.c
index 49213fb4f0..4592dc0b2d 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_main.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_main.c
@@ -40,7 +40,7 @@
 
 void
 dump_type(FILE *fp,
-          union lp_type type)
+          struct lp_type type)
 {
    fprintf(fp, "%s%s%u%sx%u",
            type.sign ? (type.floating || type.fixed ? "" : "s") : "u",
@@ -52,7 +52,7 @@ dump_type(FILE *fp,
 
 
 double
-read_elem(union lp_type type, const void *src, unsigned index)
+read_elem(struct lp_type type, const void *src, unsigned index)
 {
    double scale = lp_const_scale(type);
    double value;
@@ -115,7 +115,7 @@ read_elem(union lp_type type, const void *src, unsigned index)
 
 
 void
-write_elem(union lp_type type, void *dst, unsigned index, double value)
+write_elem(struct lp_type type, void *dst, unsigned index, double value)
 {
    assert(index < type.length);
    if(!type.sign && value < 0.0)
@@ -184,11 +184,11 @@ write_elem(union lp_type type, void *dst, unsigned index, double value)
 
 
 void
-random_elem(union lp_type type, void *dst, unsigned index)
+random_elem(struct lp_type type, void *dst, unsigned index)
 {
    double value;
    assert(index < type.length);
-   value = (double)random()/(double)RAND_MAX;
+   value = (double)rand()/(double)RAND_MAX;
    if(!type.norm) {
       unsigned long long mask;
       if (type.floating)
@@ -199,17 +199,17 @@ random_elem(union lp_type type, void *dst, unsigned index)
          mask = ((unsigned long long)1 << (type.width - 1)) - 1;
       else
          mask = ((unsigned long long)1 << type.width) - 1;
-      value += (double)(mask & random());
+      value += (double)(mask & rand());
    }
    if(!type.sign)
-      if(random() & 1)
+      if(rand() & 1)
          value = -value;
    write_elem(type, dst, index, value);
 }
 
 
 void
-read_vec(union lp_type type, const void *src, double *dst)
+read_vec(struct lp_type type, const void *src, double *dst)
 {
    unsigned i;
    for (i = 0; i < type.length; ++i)
@@ -218,7 +218,7 @@ read_vec(union lp_type type, const void *src, double *dst)
 
 
 void
-write_vec(union lp_type type, void *dst, const double *src)
+write_vec(struct lp_type type, void *dst, const double *src)
 {
    unsigned i;
    for (i = 0; i < type.length; ++i)
@@ -229,12 +229,12 @@ write_vec(union lp_type type, void *dst, const double *src)
 float
 random_float(void)
 {
-    return (float)((double)random()/(double)RAND_MAX);
+    return (float)((double)rand()/(double)RAND_MAX);
 }
 
 
 void
-random_vec(union lp_type type, void *dst)
+random_vec(struct lp_type type, void *dst)
 {
    unsigned i;
    for (i = 0; i < type.length; ++i)
@@ -243,7 +243,7 @@ random_vec(union lp_type type, void *dst)
 
 
 boolean
-compare_vec_with_eps(union lp_type type, const void *res, const void *ref, double eps)
+compare_vec_with_eps(struct lp_type type, const void *res, const void *ref, double eps)
 {
    unsigned i;
    for (i = 0; i < type.length; ++i) {
@@ -259,7 +259,7 @@ compare_vec_with_eps(union lp_type type, const void *res, const void *ref, doubl
 
 
 boolean
-compare_vec(union lp_type type, const void *res, const void *ref)
+compare_vec(struct lp_type type, const void *res, const void *ref)
 {
    double eps = lp_const_eps(type);
    return compare_vec_with_eps(type, res, ref, eps);
@@ -267,7 +267,7 @@ compare_vec(union lp_type type, const void *res, const void *ref)
 
 
 void
-dump_vec(FILE *fp, union lp_type type, const void *src)
+dump_vec(FILE *fp, struct lp_type type, const void *src)
 {
    unsigned i;
    for (i = 0; i < type.length; ++i) {
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..a1365a045f 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,
+                                  struct 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..d2a6ae21f5
--- /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,
+                                     struct 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/llvmpipe/lp_tile_cache.c b/src/gallium/drivers/llvmpipe/lp_tile_cache.c
index 143afec3d3..2e576e6039 100644
--- a/src/gallium/drivers/llvmpipe/lp_tile_cache.c
+++ b/src/gallium/drivers/llvmpipe/lp_tile_cache.c
@@ -225,11 +225,14 @@ lp_flush_tile_cache(struct llvmpipe_tile_cache *tc)
                            tc->clear_val);
 
             screen->transfer_unmap(screen, pt);
+
+            tile->status = LP_TILE_STATUS_UNDEFINED;
             break;
          }
 
          case LP_TILE_STATUS_DEFINED:
             lp_put_tile_rgba_soa(pt, x, y, tile->color);
+            tile->status = LP_TILE_STATUS_UNDEFINED;
             break;
          }
       }
@@ -257,7 +260,7 @@ lp_get_cached_tile(struct llvmpipe_tile_cache *tc,
 
    case LP_TILE_STATUS_UNDEFINED:
       /* get new tile data from transfer */
-      lp_get_tile_rgba_soa(pt, x, y, tile->color);
+      lp_get_tile_rgba_soa(pt, x & ~(TILE_SIZE - 1), y & ~(TILE_SIZE - 1), tile->color);
       tile->status = LP_TILE_STATUS_DEFINED;
       break;