gallivm: Bring aos format back to life.

Useful for fetching vertices for formats that are straight arrays.

This reverts commit aa364d091e7e2ef2296fb25f92efc79a8c88f77d.

6 files changed, 698 insertions, 1 deletions

diff --git a/src/gallium/auxiliary/Makefile b/src/gallium/auxiliary/Makefile
index 7d300d42059..f8e65cf6c61 100644
--- a/src/gallium/auxiliary/Makefile
+++ b/src/gallium/auxiliary/Makefile
@@ -149,6 +149,7 @@ GALLIVM_SOURCES = \
         gallivm/lp_bld_conv.c \
         gallivm/lp_bld_debug.c \
         gallivm/lp_bld_flow.c \
+        gallivm/lp_bld_format_aos.c \
         gallivm/lp_bld_format_soa.c \
         gallivm/lp_bld_init.c \
         gallivm/lp_bld_intr.c \
diff --git a/src/gallium/auxiliary/SConscript b/src/gallium/auxiliary/SConscript
index fc20a8bcbb5..db3a1e73114 100644
--- a/src/gallium/auxiliary/SConscript
+++ b/src/gallium/auxiliary/SConscript
@@ -198,6 +198,7 @@ if env['llvm']:
     'gallivm/lp_bld_conv.c',
     'gallivm/lp_bld_debug.c',
     'gallivm/lp_bld_flow.c',
+    'gallivm/lp_bld_format_aos.c',
     'gallivm/lp_bld_format_soa.c',
     'gallivm/lp_bld_intr.c',
     'gallivm/lp_bld_logic.c',
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c b/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c
new file mode 100644
index 00000000000..e55ac6faedd
--- /dev/null
+++ b/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c
@@ -0,0 +1,380 @@
+/**************************************************************************
+ *
+ * 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
+ * AoS pixel format manipulation.
+ *
+ * @author Jose Fonseca <jfonseca@vmware.com>
+ */
+
+
+#include "util/u_cpu_detect.h"
+#include "util/u_format.h"
+
+#include "lp_bld_type.h"
+#include "lp_bld_const.h"
+#include "lp_bld_swizzle.h"
+#include "lp_bld_format.h"
+
+
+/**
+ * Unpack a single pixel into its RGBA components.
+ *
+ * @param packed integer.
+ *
+ * @return RGBA in a 4 floats vector.
+ */
+LLVMValueRef
+lp_build_unpack_rgba_aos(LLVMBuilderRef builder,
+                         const struct util_format_description *desc,
+                         LLVMValueRef packed)
+{
+   LLVMTypeRef type;
+   LLVMValueRef shifted, casted, scaled, masked;
+   LLVMValueRef shifts[4];
+   LLVMValueRef masks[4];
+   LLVMValueRef scales[4];
+   LLVMValueRef swizzles[4];
+   LLVMValueRef aux[4];
+   bool normalized;
+   int empty_channel;
+   unsigned shift;
+   unsigned i;
+
+   /* FIXME: Support more formats */
+   assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
+   assert(desc->block.width == 1);
+   assert(desc->block.height == 1);
+   assert(desc->block.bits <= 32);
+
+   type = LLVMIntType(desc->block.bits);
+
+   /* Do the intermediate integer computations with 32bit integers since it
+    * matches floating point size */
+   if (desc->block.bits < 32)
+      packed = LLVMBuildZExt(builder, packed, LLVMInt32Type(), "");
+
+   /* Broadcast the packed value to all four channels */
+   packed = LLVMBuildInsertElement(builder,
+                                   LLVMGetUndef(LLVMVectorType(LLVMInt32Type(), 4)),
+                                   packed,
+                                   LLVMConstNull(LLVMInt32Type()),
+                                   "");
+   packed = LLVMBuildShuffleVector(builder,
+                                   packed,
+                                   LLVMGetUndef(LLVMVectorType(LLVMInt32Type(), 4)),
+                                   LLVMConstNull(LLVMVectorType(LLVMInt32Type(), 4)),
+                                   "");
+
+   /* Initialize vector constants */
+   normalized = FALSE;
+   empty_channel = -1;
+   shift = 0;
+   for (i = 0; i < 4; ++i) {
+      unsigned bits = desc->channel[i].size;
+
+      if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
+         shifts[i] = LLVMGetUndef(LLVMInt32Type());
+         masks[i] = LLVMConstNull(LLVMInt32Type());
+         scales[i] =  LLVMConstNull(LLVMFloatType());
+         empty_channel = i;
+      }
+      else {
+         unsigned mask = (1 << bits) - 1;
+
+         assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
+         assert(bits < 32);
+
+         shifts[i] = LLVMConstInt(LLVMInt32Type(), shift, 0);
+         masks[i] = LLVMConstInt(LLVMInt32Type(), mask, 0);
+
+         if (desc->channel[i].normalized) {
+            scales[i] = LLVMConstReal(LLVMFloatType(), 1.0/mask);
+            normalized = TRUE;
+         }
+         else
+            scales[i] =  LLVMConstReal(LLVMFloatType(), 1.0);
+      }
+
+      shift += bits;
+   }
+
+   shifted = LLVMBuildLShr(builder, packed, LLVMConstVector(shifts, 4), "");
+   masked = LLVMBuildAnd(builder, shifted, LLVMConstVector(masks, 4), "");
+   /* UIToFP can't be expressed in SSE2 */
+   casted = LLVMBuildSIToFP(builder, masked, LLVMVectorType(LLVMFloatType(), 4), "");
+
+   if (normalized)
+      scaled = LLVMBuildMul(builder, casted, LLVMConstVector(scales, 4), "");
+   else
+      scaled = casted;
+
+   for (i = 0; i < 4; ++i)
+      aux[i] = LLVMGetUndef(LLVMFloatType());
+
+   for (i = 0; i < 4; ++i) {
+      enum util_format_swizzle swizzle = desc->swizzle[i];
+
+      switch (swizzle) {
+      case UTIL_FORMAT_SWIZZLE_X:
+      case UTIL_FORMAT_SWIZZLE_Y:
+      case UTIL_FORMAT_SWIZZLE_Z:
+      case UTIL_FORMAT_SWIZZLE_W:
+         swizzles[i] = LLVMConstInt(LLVMInt32Type(), swizzle, 0);
+         break;
+      case UTIL_FORMAT_SWIZZLE_0:
+         assert(empty_channel >= 0);
+         swizzles[i] = LLVMConstInt(LLVMInt32Type(), empty_channel, 0);
+         break;
+      case UTIL_FORMAT_SWIZZLE_1:
+         swizzles[i] = LLVMConstInt(LLVMInt32Type(), 4, 0);
+         aux[0] = LLVMConstReal(LLVMFloatType(), 1.0);
+         break;
+      case UTIL_FORMAT_SWIZZLE_NONE:
+         swizzles[i] = LLVMGetUndef(LLVMFloatType());
+         assert(0);
+         break;
+      }
+   }
+
+   return LLVMBuildShuffleVector(builder, scaled, LLVMConstVector(aux, 4), LLVMConstVector(swizzles, 4), "");
+}
+
+
+/**
+ * Take a vector with packed pixels and unpack into a rgba8 vector.
+ *
+ * Formats with bit depth smaller than 32bits are accepted, but they must be
+ * padded to 32bits.
+ */
+LLVMValueRef
+lp_build_unpack_rgba8_aos(LLVMBuilderRef builder,
+                          const struct util_format_description *desc,
+                          struct lp_type type,
+                          LLVMValueRef packed)
+{
+   struct lp_build_context bld;
+   bool rgba8;
+   LLVMValueRef res;
+   unsigned i;
+
+   lp_build_context_init(&bld, builder, type);
+
+   /* FIXME: Support more formats */
+   assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
+   assert(desc->block.width == 1);
+   assert(desc->block.height == 1);
+   assert(desc->block.bits <= 32);
+
+   assert(!type.floating);
+   assert(!type.fixed);
+   assert(type.norm);
+   assert(type.width == 8);
+   assert(type.length % 4 == 0);
+
+   rgba8 = TRUE;
+   for(i = 0; i < 4; ++i) {
+      assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED ||
+             desc->channel[i].type == UTIL_FORMAT_TYPE_VOID);
+      if(desc->channel[0].size != 8)
+         rgba8 = FALSE;
+   }
+
+   if(rgba8) {
+      /*
+       * The pixel is already in a rgba8 format variant. All it is necessary
+       * is to swizzle the channels.
+       */
+
+      unsigned char swizzles[4];
+      boolean zeros[4]; /* bitwise AND mask */
+      boolean ones[4]; /* bitwise OR mask */
+      boolean swizzles_needed = FALSE;
+      boolean zeros_needed = FALSE;
+      boolean ones_needed = FALSE;
+
+      for(i = 0; i < 4; ++i) {
+         enum util_format_swizzle swizzle = desc->swizzle[i];
+
+         /* Initialize with the no-op case */
+         swizzles[i] = util_cpu_caps.little_endian ? 3 - i : i;
+         zeros[i] = TRUE;
+         ones[i] = FALSE;
+
+         switch (swizzle) {
+         case UTIL_FORMAT_SWIZZLE_X:
+         case UTIL_FORMAT_SWIZZLE_Y:
+         case UTIL_FORMAT_SWIZZLE_Z:
+         case UTIL_FORMAT_SWIZZLE_W:
+            if(swizzle != swizzles[i]) {
+               swizzles[i] = swizzle;
+               swizzles_needed = TRUE;
+            }
+            break;
+         case UTIL_FORMAT_SWIZZLE_0:
+            zeros[i] = FALSE;
+            zeros_needed = TRUE;
+            break;
+         case UTIL_FORMAT_SWIZZLE_1:
+            ones[i] = TRUE;
+            ones_needed = TRUE;
+            break;
+         case UTIL_FORMAT_SWIZZLE_NONE:
+            assert(0);
+            break;
+         }
+      }
+
+      res = packed;
+
+      if(swizzles_needed)
+         res = lp_build_swizzle1_aos(&bld, res, swizzles);
+
+      if(zeros_needed) {
+         /* Mask out zero channels */
+         LLVMValueRef mask = lp_build_const_mask_aos(type, zeros);
+         res = LLVMBuildAnd(builder, res, mask, "");
+      }
+
+      if(ones_needed) {
+         /* Or one channels */
+         LLVMValueRef mask = lp_build_const_mask_aos(type, ones);
+         res = LLVMBuildOr(builder, res, mask, "");
+      }
+   }
+   else {
+      /* FIXME */
+      assert(0);
+      res = lp_build_undef(type);
+   }
+
+   return res;
+}
+
+
+/**
+ * Pack a single pixel.
+ *
+ * @param rgba 4 float vector with the unpacked components.
+ *
+ * XXX: This is mostly for reference and testing -- operating a single pixel at
+ * a time is rarely if ever needed.
+ */
+LLVMValueRef
+lp_build_pack_rgba_aos(LLVMBuilderRef builder,
+                       const struct util_format_description *desc,
+                       LLVMValueRef rgba)
+{
+   LLVMTypeRef type;
+   LLVMValueRef packed = NULL;
+   LLVMValueRef swizzles[4];
+   LLVMValueRef shifted, casted, scaled, unswizzled;
+   LLVMValueRef shifts[4];
+   LLVMValueRef scales[4];
+   bool normalized;
+   unsigned shift;
+   unsigned i, j;
+
+   assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
+   assert(desc->block.width == 1);
+   assert(desc->block.height == 1);
+
+   type = LLVMIntType(desc->block.bits);
+
+   /* Unswizzle the color components into the source vector. */
+   for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j) {
+         if (desc->swizzle[j] == i)
+            break;
+      }
+      if (j < 4)
+         swizzles[i] = LLVMConstInt(LLVMInt32Type(), j, 0);
+      else
+         swizzles[i] = LLVMGetUndef(LLVMInt32Type());
+   }
+
+   unswizzled = LLVMBuildShuffleVector(builder, rgba,
+                                       LLVMGetUndef(LLVMVectorType(LLVMFloatType(), 4)),
+                                       LLVMConstVector(swizzles, 4), "");
+
+   normalized = FALSE;
+   shift = 0;
+   for (i = 0; i < 4; ++i) {
+      unsigned bits = desc->channel[i].size;
+
+      if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
+         shifts[i] = LLVMGetUndef(LLVMInt32Type());
+         scales[i] =  LLVMGetUndef(LLVMFloatType());
+      }
+      else {
+         unsigned mask = (1 << bits) - 1;
+
+         assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
+         assert(bits < 32);
+
+         shifts[i] = LLVMConstInt(LLVMInt32Type(), shift, 0);
+
+         if (desc->channel[i].normalized) {
+            scales[i] = LLVMConstReal(LLVMFloatType(), mask);
+            normalized = TRUE;
+         }
+         else
+            scales[i] =  LLVMConstReal(LLVMFloatType(), 1.0);
+      }
+
+      shift += bits;
+   }
+
+   if (normalized)
+      scaled = LLVMBuildMul(builder, unswizzled, LLVMConstVector(scales, 4), "");
+   else
+      scaled = unswizzled;
+
+   casted = LLVMBuildFPToSI(builder, scaled, LLVMVectorType(LLVMInt32Type(), 4), "");
+
+   shifted = LLVMBuildShl(builder, casted, LLVMConstVector(shifts, 4), "");
+   
+   /* Bitwise or all components */
+   for (i = 0; i < 4; ++i) {
+      if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
+         LLVMValueRef component = LLVMBuildExtractElement(builder, shifted, LLVMConstInt(LLVMInt32Type(), i, 0), "");
+         if (packed)
+            packed = LLVMBuildOr(builder, packed, component, "");
+         else
+            packed = component;
+      }
+   }
+
+   if (!packed)
+      packed = LLVMGetUndef(LLVMInt32Type());
+
+   if (desc->block.bits < 32)
+      packed = LLVMBuildTrunc(builder, packed, type, "");
+
+   return packed;
+}
diff --git a/src/gallium/drivers/llvmpipe/Makefile b/src/gallium/drivers/llvmpipe/Makefile
index fad9e3482fa..bfa9f02bc56 100644
--- a/src/gallium/drivers/llvmpipe/Makefile
+++ b/src/gallium/drivers/llvmpipe/Makefile
@@ -47,7 +47,7 @@ C_SOURCES = \
 
 CPP_SOURCES = \
 
-PROGS := \
+PROGS := lp_test_format	\
 	 lp_test_blend	\
 	 lp_test_conv	\
 	 lp_test_printf
diff --git a/src/gallium/drivers/llvmpipe/SConscript b/src/gallium/drivers/llvmpipe/SConscript
index 0b827281ff9..b9e9826e2a3 100644
--- a/src/gallium/drivers/llvmpipe/SConscript
+++ b/src/gallium/drivers/llvmpipe/SConscript
@@ -73,6 +73,7 @@ if env['platform'] != 'embedded':
     env.Prepend(LIBS = [llvmpipe] + gallium)
 
     tests = [
+        'format',
         'blend',
         'conv',
     ]
diff --git a/src/gallium/drivers/llvmpipe/lp_test_format.c b/src/gallium/drivers/llvmpipe/lp_test_format.c
new file mode 100644
index 00000000000..fb595893bd0
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_test_format.c
@@ -0,0 +1,314 @@
+/**************************************************************************
+ *
+ * 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 <stdlib.h>
+#include <stdio.h>
+
+#include "gallivm/lp_bld.h"
+#include <llvm-c/Analysis.h>
+#include <llvm-c/ExecutionEngine.h>
+#include <llvm-c/Target.h>
+#include <llvm-c/Transforms/Scalar.h>
+
+#include "util/u_cpu_detect.h"
+#include "util/u_format.h"
+
+#include "gallivm/lp_bld_format.h"
+#include "lp_test.h"
+
+
+struct pixel_test_case
+{
+   enum pipe_format format;
+   uint32_t packed;
+   double unpacked[4];
+};
+
+
+struct pixel_test_case test_cases[] =
+{
+   {PIPE_FORMAT_B5G6R5_UNORM,   0x0000, {0.0, 0.0, 0.0, 1.0}},
+   {PIPE_FORMAT_B5G6R5_UNORM,   0x001f, {0.0, 0.0, 1.0, 1.0}},
+   {PIPE_FORMAT_B5G6R5_UNORM,   0x07e0, {0.0, 1.0, 0.0, 1.0}},
+   {PIPE_FORMAT_B5G6R5_UNORM,   0xf800, {1.0, 0.0, 0.0, 1.0}},
+   {PIPE_FORMAT_B5G6R5_UNORM,   0xffff, {1.0, 1.0, 1.0, 1.0}},
+
+   {PIPE_FORMAT_B5G5R5A1_UNORM, 0x0000, {0.0, 0.0, 0.0, 0.0}},
+   {PIPE_FORMAT_B5G5R5A1_UNORM, 0x001f, {0.0, 0.0, 1.0, 0.0}},
+   {PIPE_FORMAT_B5G5R5A1_UNORM, 0x03e0, {0.0, 1.0, 0.0, 0.0}},
+   {PIPE_FORMAT_B5G5R5A1_UNORM, 0x7c00, {1.0, 0.0, 0.0, 0.0}},
+   {PIPE_FORMAT_B5G5R5A1_UNORM, 0x8000, {0.0, 0.0, 0.0, 1.0}},
+   {PIPE_FORMAT_B5G5R5A1_UNORM, 0xffff, {1.0, 1.0, 1.0, 1.0}},
+
+   {PIPE_FORMAT_B8G8R8A8_UNORM, 0x00000000, {0.0, 0.0, 0.0, 0.0}},
+   {PIPE_FORMAT_B8G8R8A8_UNORM, 0x000000ff, {0.0, 0.0, 1.0, 0.0}},
+   {PIPE_FORMAT_B8G8R8A8_UNORM, 0x0000ff00, {0.0, 1.0, 0.0, 0.0}},
+   {PIPE_FORMAT_B8G8R8A8_UNORM, 0x00ff0000, {1.0, 0.0, 0.0, 0.0}},
+   {PIPE_FORMAT_B8G8R8A8_UNORM, 0xff000000, {0.0, 0.0, 0.0, 1.0}},
+   {PIPE_FORMAT_B8G8R8A8_UNORM, 0xffffffff, {1.0, 1.0, 1.0, 1.0}},
+
+#if 0
+   {PIPE_FORMAT_R8G8B8A8_UNORM, 0x00000000, {0.0, 0.0, 0.0, 0.0}},
+   {PIPE_FORMAT_R8G8B8A8_UNORM, 0x000000ff, {0.0, 0.0, 0.0, 1.0}},
+   {PIPE_FORMAT_R8G8B8A8_UNORM, 0x0000ff00, {0.0, 0.0, 1.0, 0.0}},
+   {PIPE_FORMAT_R8G8B8A8_UNORM, 0x00ff0000, {0.0, 1.0, 0.0, 0.0}},
+   {PIPE_FORMAT_R8G8B8A8_UNORM, 0xff000000, {1.0, 0.0, 0.0, 0.0}},
+   {PIPE_FORMAT_R8G8B8A8_UNORM, 0xffffffff, {1.0, 1.0, 1.0, 1.0}},
+#endif
+
+   {PIPE_FORMAT_A8R8G8B8_UNORM, 0x00000000, {0.0, 0.0, 0.0, 0.0}},
+   {PIPE_FORMAT_A8R8G8B8_UNORM, 0x000000ff, {0.0, 0.0, 0.0, 1.0}},
+   {PIPE_FORMAT_A8R8G8B8_UNORM, 0x0000ff00, {1.0, 0.0, 0.0, 0.0}},
+   {PIPE_FORMAT_A8R8G8B8_UNORM, 0x00ff0000, {0.0, 1.0, 0.0, 0.0}},
+   {PIPE_FORMAT_A8R8G8B8_UNORM, 0xff000000, {0.0, 0.0, 1.0, 0.0}},
+   {PIPE_FORMAT_A8R8G8B8_UNORM, 0xffffffff, {1.0, 1.0, 1.0, 1.0}},
+};
+
+
+void
+write_tsv_header(FILE *fp)
+{
+   fprintf(fp,
+           "result\t"
+           "format\n");
+
+   fflush(fp);
+}
+
+
+static void
+write_tsv_row(FILE *fp,
+              const struct util_format_description *desc,
+              boolean success)
+{
+   fprintf(fp, "%s\t", success ? "pass" : "fail");
+
+   fprintf(fp, "%s\n", desc->name);
+
+   fflush(fp);
+}
+
+
+typedef void (*load_ptr_t)(const uint32_t packed, float *);
+
+
+static LLVMValueRef
+add_load_rgba_test(LLVMModuleRef module,
+                   const struct util_format_description *desc)
+{
+   LLVMTypeRef args[2];
+   LLVMValueRef func;
+   LLVMValueRef packed;
+   LLVMValueRef rgba_ptr;
+   LLVMBasicBlockRef block;
+   LLVMBuilderRef builder;
+   LLVMValueRef rgba;
+
+   args[0] = LLVMInt32Type();
+   args[1] = LLVMPointerType(LLVMVectorType(LLVMFloatType(), 4), 0);
+
+   func = LLVMAddFunction(module, "load", LLVMFunctionType(LLVMVoidType(), args, 2, 0));
+   LLVMSetFunctionCallConv(func, LLVMCCallConv);
+   packed = LLVMGetParam(func, 0);
+   rgba_ptr = LLVMGetParam(func, 1);
+
+   block = LLVMAppendBasicBlock(func, "entry");
+   builder = LLVMCreateBuilder();
+   LLVMPositionBuilderAtEnd(builder, block);
+
+   if(desc->block.bits < 32)
+      packed = LLVMBuildTrunc(builder, packed, LLVMIntType(desc->block.bits), "");
+
+   rgba = lp_build_unpack_rgba_aos(builder, desc, packed);
+
+   LLVMBuildStore(builder, rgba, rgba_ptr);
+
+   LLVMBuildRetVoid(builder);
+
+   LLVMDisposeBuilder(builder);
+   return func;
+}
+
+
+typedef void (*store_ptr_t)(uint32_t *, const float *);
+
+
+static LLVMValueRef
+add_store_rgba_test(LLVMModuleRef module,
+                    const struct util_format_description *desc)
+{
+   LLVMTypeRef args[2];
+   LLVMValueRef func;
+   LLVMValueRef packed_ptr;
+   LLVMValueRef rgba_ptr;
+   LLVMBasicBlockRef block;
+   LLVMBuilderRef builder;
+   LLVMValueRef rgba;
+   LLVMValueRef packed;
+
+   args[0] = LLVMPointerType(LLVMInt32Type(), 0);
+   args[1] = LLVMPointerType(LLVMVectorType(LLVMFloatType(), 4), 0);
+
+   func = LLVMAddFunction(module, "store", LLVMFunctionType(LLVMVoidType(), args, 2, 0));
+   LLVMSetFunctionCallConv(func, LLVMCCallConv);
+   packed_ptr = LLVMGetParam(func, 0);
+   rgba_ptr = LLVMGetParam(func, 1);
+
+   block = LLVMAppendBasicBlock(func, "entry");
+   builder = LLVMCreateBuilder();
+   LLVMPositionBuilderAtEnd(builder, block);
+
+   rgba = LLVMBuildLoad(builder, rgba_ptr, "");
+
+   packed = lp_build_pack_rgba_aos(builder, desc, rgba);
+
+   if(desc->block.bits < 32)
+      packed = LLVMBuildZExt(builder, packed, LLVMInt32Type(), "");
+
+   LLVMBuildStore(builder, packed, packed_ptr);
+
+   LLVMBuildRetVoid(builder);
+
+   LLVMDisposeBuilder(builder);
+   return func;
+}
+
+
+PIPE_ALIGN_STACK
+static boolean
+test_format(unsigned verbose, FILE *fp, const struct pixel_test_case *test)
+{
+   LLVMModuleRef module = NULL;
+   LLVMValueRef load = NULL;
+   LLVMValueRef store = NULL;
+   LLVMExecutionEngineRef engine = NULL;
+   LLVMModuleProviderRef provider = NULL;
+   LLVMPassManagerRef pass = NULL;
+   char *error = NULL;
+   const struct util_format_description *desc;
+   load_ptr_t load_ptr;
+   store_ptr_t store_ptr;
+   float unpacked[4];
+   unsigned packed;
+   boolean success;
+   unsigned i;
+
+   desc = util_format_description(test->format);
+   fprintf(stderr, "%s\n", desc->name);
+
+   module = LLVMModuleCreateWithName("test");
+
+   load = add_load_rgba_test(module, desc);
+   store = add_store_rgba_test(module, desc);
+
+   if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) {
+      LLVMDumpModule(module);
+      abort();
+   }
+   LLVMDisposeMessage(error);
+
+   provider = LLVMCreateModuleProviderForExistingModule(module);
+   if (LLVMCreateJITCompiler(&engine, provider, 1, &error)) {
+      fprintf(stderr, "%s\n", error);
+      LLVMDisposeMessage(error);
+      abort();
+   }
+
+#if 0
+   pass = LLVMCreatePassManager();
+   LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass);
+   /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
+    * but there are more on SVN. */
+   LLVMAddConstantPropagationPass(pass);
+   LLVMAddInstructionCombiningPass(pass);
+   LLVMAddPromoteMemoryToRegisterPass(pass);
+   LLVMAddGVNPass(pass);
+   LLVMAddCFGSimplificationPass(pass);
+   LLVMRunPassManager(pass, module);
+#else
+   (void)pass;
+#endif
+
+   load_ptr  = (load_ptr_t) LLVMGetPointerToGlobal(engine, load);
+   store_ptr = (store_ptr_t)LLVMGetPointerToGlobal(engine, store);
+
+   memset(unpacked, 0, sizeof unpacked);
+   packed = 0;
+
+   load_ptr(test->packed, unpacked);
+   store_ptr(&packed, unpacked);
+
+   success = TRUE;
+   if(test->packed != packed)
+      success = FALSE;
+   for(i = 0; i < 4; ++i)
+      if(test->unpacked[i] != unpacked[i])
+         success = FALSE;
+
+   if (!success) {
+      printf("FAILED\n");
+      printf("  Packed: %08x\n", test->packed);
+      printf("          %08x\n", packed);
+      printf("  Unpacked: %f %f %f %f\n", unpacked[0], unpacked[1], unpacked[2], unpacked[3]);
+      printf("            %f %f %f %f\n", test->unpacked[0], test->unpacked[1], test->unpacked[2], test->unpacked[3]);
+      LLVMDumpModule(module);
+   }
+
+   LLVMFreeMachineCodeForFunction(engine, store);
+   LLVMFreeMachineCodeForFunction(engine, load);
+
+   LLVMDisposeExecutionEngine(engine);
+   if(pass)
+      LLVMDisposePassManager(pass);
+
+   if(fp)
+      write_tsv_row(fp, desc, success);
+
+   return success;
+}
+
+
+boolean
+test_all(unsigned verbose, FILE *fp)
+{
+   unsigned i;
+   bool success = TRUE;
+
+   for (i = 0; i < sizeof(test_cases)/sizeof(test_cases[0]); ++i)
+      if(!test_format(verbose, fp, &test_cases[i]))
+        success = FALSE;
+
+   return success;
+}
+
+
+boolean
+test_some(unsigned verbose, FILE *fp, unsigned long n)
+{
+   return test_all(verbose, fp);
+}