7 files changed, 336 insertions, 66 deletions

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format.h b/src/gallium/auxiliary/gallivm/lp_bld_format.h
index 73ab6de3f2..bb1298ed3f 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_format.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_format.h
@@ -80,4 +80,15 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
                          LLVMValueRef *rgba);
 
 
+void
+lp_build_fetch_rgba_soa(LLVMBuilderRef builder,
+                        const struct util_format_description *format_desc,
+                        struct lp_type type,
+                        LLVMValueRef base_ptr,
+                        LLVMValueRef offsets,
+                        LLVMValueRef i,
+                        LLVMValueRef j,
+                        LLVMValueRef *rgba);
+
+
 #endif /* !LP_BLD_FORMAT_H */
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c
index c9fbe0ad10..433134b809 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c
@@ -27,10 +27,14 @@
 
 
 #include "util/u_format.h"
+#include "util/u_memory.h"
+#include "util/u_string.h"
 
 #include "lp_bld_type.h"
 #include "lp_bld_const.h"
 #include "lp_bld_conv.h"
+#include "lp_bld_sample.h" /* for lp_build_gather */
+#include "lp_bld_init.h"
 #include "lp_bld_format.h"
 
 
@@ -80,6 +84,24 @@ lp_build_format_swizzle_soa(const struct util_format_description *format_desc,
 }
 
 
+/**
+ * Unpack several pixels in SoA.
+ *
+ * It takes a vector of packed pixels:
+ *
+ *   packed = {P0, P1, P2, P3, ..., Pn}
+ *
+ * And will produce four vectors:
+ *
+ *   red    = {R0, R1, R2, R3, ..., Rn}
+ *   green  = {G0, G1, G2, G3, ..., Gn}
+ *   blue   = {B0, B1, B2, B3, ..., Bn}
+ *   alpha  = {A0, A1, A2, A3, ..., An}
+ *
+ * It requires that a packed pixel fits into an element of the output
+ * channels. The common case is when converting pixel with a depth of 32 bit or
+ * less into floats.
+ */
 void
 lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
                          const struct util_format_description *format_desc,
@@ -91,11 +113,13 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
    unsigned start;
    unsigned chan;
 
-   /* FIXME: Support more formats */
-   assert(format_desc->is_bitmask);
+   /* FIXME: Support more pixel formats */
    assert(format_desc->block.width == 1);
    assert(format_desc->block.height == 1);
-   assert(format_desc->block.bits <= 32);
+   assert(format_desc->block.bits <= type.width);
+   /* FIXME: Support more output types */
+   assert(type.floating);
+   assert(type.width == 32);
 
    /* Decode the input vector components */
    start = 0;
@@ -108,18 +132,32 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
 
       switch(format_desc->channel[chan].type) {
       case UTIL_FORMAT_TYPE_VOID:
-         input = NULL;
+         input = lp_build_undef(type);
          break;
 
       case UTIL_FORMAT_TYPE_UNSIGNED:
-         if(type.floating) {
-            if(start)
-               input = LLVMBuildLShr(builder, input, lp_build_const_int_vec(type, start), "");
-            if(stop < format_desc->block.bits) {
-               unsigned mask = ((unsigned long long)1 << width) - 1;
-               input = LLVMBuildAnd(builder, input, lp_build_const_int_vec(type, mask), "");
-            }
+         /*
+          * Align the LSB
+          */
+
+         if (start) {
+            input = LLVMBuildLShr(builder, input, lp_build_const_int_vec(type, start), "");
+         }
+
+         /*
+          * Zero the MSBs
+          */
 
+         if (stop < format_desc->block.bits) {
+            unsigned mask = ((unsigned long long)1 << width) - 1;
+            input = LLVMBuildAnd(builder, input, lp_build_const_int_vec(type, mask), "");
+         }
+
+         /*
+          * Type conversion
+          */
+
+         if (type.floating) {
             if(format_desc->channel[chan].normalized)
                input = lp_build_unsigned_norm_to_float(builder, width, type, input);
             else
@@ -130,10 +168,71 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
             assert(0);
             input = lp_build_undef(type);
          }
+
+         break;
+
+      case UTIL_FORMAT_TYPE_SIGNED:
+         /*
+          * Align the sign bit first.
+          */
+
+         if (stop < type.width) {
+            unsigned bits = type.width - stop;
+            LLVMValueRef bits_val = lp_build_const_int_vec(type, bits);
+            input = LLVMBuildShl(builder, input, bits_val, "");
+         }
+
+         /*
+          * Align the LSB (with an arithmetic shift to preserve the sign)
+          */
+
+         if (format_desc->channel[chan].size < type.width) {
+            unsigned bits = type.width - format_desc->channel[chan].size;
+            LLVMValueRef bits_val = lp_build_const_int_vec(type, bits);
+            input = LLVMBuildAShr(builder, input, bits_val, "");
+         }
+
+         /*
+          * Type conversion
+          */
+
+         if (type.floating) {
+            input = LLVMBuildSIToFP(builder, input, lp_build_vec_type(type), "");
+            if (format_desc->channel[chan].normalized) {
+               double scale = 1.0 / ((1 << (format_desc->channel[chan].size - 1)) - 1);
+               LLVMValueRef scale_val = lp_build_const_vec(type, scale);
+               input = LLVMBuildMul(builder, input, scale_val, "");
+            }
+         }
+         else {
+            /* FIXME */
+            assert(0);
+            input = lp_build_undef(type);
+         }
+
+         break;
+
+      case UTIL_FORMAT_TYPE_FLOAT:
+         if (type.floating) {
+            assert(start == 0);
+            assert(stop == 32);
+            assert(type.width == 32);
+            input = LLVMBuildBitCast(builder, input, lp_build_vec_type(type), "");
+         }
+         else {
+            /* FIXME */
+            assert(0);
+            input = lp_build_undef(type);
+         }
+         break;
+
+      case UTIL_FORMAT_TYPE_FIXED:
+         assert(0);
+         input = lp_build_undef(type);
          break;
 
       default:
-         /* fall through */
+         assert(0);
          input = lp_build_undef(type);
          break;
       }
@@ -145,3 +244,139 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
 
    lp_build_format_swizzle_soa(format_desc, type, inputs, rgba);
 }
+
+
+/**
+ * Fetch a pixel into a SoA.
+ *
+ * i and j are the sub-block pixel coordinates.
+ */
+void
+lp_build_fetch_rgba_soa(LLVMBuilderRef builder,
+                        const struct util_format_description *format_desc,
+                        struct lp_type type,
+                        LLVMValueRef base_ptr,
+                        LLVMValueRef offset,
+                        LLVMValueRef i,
+                        LLVMValueRef j,
+                        LLVMValueRef *rgba)
+{
+
+   if (format_desc->block.width == 1 &&
+       format_desc->block.height == 1 &&
+       format_desc->block.bits <= type.width)
+   {
+      /*
+       * The packed pixel fits into an element of the destination format. Put
+       * the packed pixels into a vector and estract each component for all
+       * vector elements in parallel.
+       */
+
+      LLVMValueRef packed;
+
+      /*
+       * gather the texels from the texture
+       */
+      packed = lp_build_gather(builder,
+                               type.length,
+                               format_desc->block.bits,
+                               type.width,
+                               base_ptr, offset);
+
+      /*
+       * convert texels to float rgba
+       */
+      lp_build_unpack_rgba_soa(builder,
+                               format_desc,
+                               type,
+                               packed, rgba);
+   }
+   else {
+      /*
+       * Fallback to calling util_format_description::fetch_float for each
+       * pixel.
+       *
+       * This is definitely not the most efficient way of fetching pixels, as
+       * we miss the opportunity to do vectorization, but this it is a
+       * convenient for formats or scenarios for which there was no opportunity
+       * or incentive to optimize.
+       */
+
+      LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)));
+      char name[256];
+      LLVMValueRef function;
+      LLVMValueRef tmp;
+      unsigned k, chan;
+
+      assert(type.floating);
+
+      util_snprintf(name, sizeof name, "util_format_%s_fetch_float", format_desc->short_name);
+
+      /*
+       * Declare and bind format_desc->fetch_float().
+       */
+
+      function = LLVMGetNamedFunction(module, name);
+      if (!function) {
+         LLVMTypeRef ret_type;
+         LLVMTypeRef arg_types[4];
+         LLVMTypeRef function_type;
+
+         ret_type = LLVMVoidType();
+         arg_types[0] = LLVMPointerType(LLVMFloatType(), 0);
+         arg_types[1] = LLVMPointerType(LLVMInt8Type(), 0);
+         arg_types[3] = arg_types[2] = LLVMIntType(sizeof(unsigned) * 8);
+         function_type = LLVMFunctionType(ret_type, arg_types, Elements(arg_types), 0);
+         function = LLVMAddFunction(module, name, function_type);
+
+         LLVMSetFunctionCallConv(function, LLVMCCallConv);
+         LLVMSetLinkage(function, LLVMExternalLinkage);
+
+         assert(LLVMIsDeclaration(function));
+
+         LLVMAddGlobalMapping(lp_build_engine, function, format_desc->fetch_float);
+      }
+
+      for (chan = 0; chan < 4; ++chan) {
+         rgba[chan] = lp_build_undef(type);
+      }
+
+      tmp = LLVMBuildArrayAlloca(builder,
+                                 LLVMFloatType(),
+                                 LLVMConstInt(LLVMInt32Type(), 4, 0),
+                                 "");
+
+      /*
+       * Invoke format_desc->fetch_float() for each pixel and insert the result
+       * in the SoA vectors.
+       */
+
+      for(k = 0; k < type.length; ++k) {
+         LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), k, 0);
+         LLVMValueRef offset_elem;
+         LLVMValueRef ptr;
+         LLVMValueRef i_elem, j_elem;
+         LLVMValueRef args[4];
+
+         offset_elem = LLVMBuildExtractElement(builder, offset, index, "");
+         ptr = LLVMBuildGEP(builder, base_ptr, &offset_elem, 1, "");
+
+         i_elem = LLVMBuildExtractElement(builder, i, index, "");
+         j_elem = LLVMBuildExtractElement(builder, j, index, "");
+
+         args[0] = tmp;
+         args[1] = ptr;
+         args[2] = i_elem;
+         args[3] = j_elem;
+
+         LLVMBuildCall(builder, function, args, 4, "");
+
+         for (chan = 0; chan < 4; ++chan) {
+            LLVMValueRef chan_val = LLVMConstInt(LLVMInt32Type(), chan, 0),
+            tmp_chan = LLVMBuildGEP(builder, tmp, &chan_val, 1, "");
+            tmp_chan = LLVMBuildLoad(builder, tmp_chan, "");
+            rgba[chan] = LLVMBuildInsertElement(builder, rgba[chan], tmp_chan, index, "");
+         }
+      }
+   }
+}
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_init.cpp b/src/gallium/auxiliary/gallivm/lp_bld_init.c
index 067397a520..de07c222a3 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_init.cpp
+++ b/src/gallium/auxiliary/gallivm/lp_bld_init.c
@@ -26,33 +26,42 @@
  **************************************************************************/
 
 
-#include <llvm/Config/config.h>
-#include <llvm/Target/TargetSelect.h>
-#include <llvm/Target/TargetOptions.h>
-
-#include "pipe/p_config.h"
-
+#include "pipe/p_compiler.h"
+#include "util/u_debug.h"
 #include "lp_bld_init.h"
 
 
-extern "C" void LLVMLinkInJIT();
+LLVMModuleRef lp_build_module = NULL;
+LLVMExecutionEngineRef lp_build_engine = NULL;
+LLVMModuleProviderRef lp_build_provider = NULL;
+LLVMTargetDataRef lp_build_target = NULL;
 
 
-extern "C" void
+void
 lp_build_init(void)
 {
-#if defined(PIPE_OS_WINDOWS) && defined(PIPE_ARCH_X86)
-   /*
-    * This is mis-detected on some hardware / software combinations.
-    */
-   llvm::StackAlignment = 4;
-   llvm::RealignStack = true;
-#endif
-
-   /* Same as LLVMInitializeNativeTarget(); */
-   llvm::InitializeNativeTarget();
+   LLVMInitializeNativeTarget();
 
    LLVMLinkInJIT();
+
+   if (!lp_build_module)
+      lp_build_module = LLVMModuleCreateWithName("gallivm");
+
+   if (!lp_build_provider)
+      lp_build_provider = LLVMCreateModuleProviderForExistingModule(lp_build_module);
+
+   if (!lp_build_engine) {
+      char *error = NULL;
+
+      if (LLVMCreateJITCompiler(&lp_build_engine, lp_build_provider, 1, &error)) {
+         _debug_printf("%s\n", error);
+         LLVMDisposeMessage(error);
+         assert(0);
+      }
+   }
+
+   if (!lp_build_target)
+      lp_build_target = LLVMGetExecutionEngineTargetData(lp_build_engine);
 }
 
 
@@ -64,6 +73,6 @@ lp_build_init(void)
  */
 #if defined(_MSC_VER) && defined(_DEBUG)
 #include <crtdefs.h>
-extern "C" _CRTIMP void __cdecl
+_CRTIMP void __cdecl
 _invalid_parameter_noinfo(void) {}
 #endif
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_init.h b/src/gallium/auxiliary/gallivm/lp_bld_init.h
index 07f50d1c43..0ec2afcd1b 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_init.h
+++ b/src/gallium/auxiliary/gallivm/lp_bld_init.h
@@ -30,18 +30,18 @@
 #define LP_BLD_INIT_H
 
 
-#ifdef __cplusplus
-extern "C" {
-#endif
+#include "lp_bld.h"
+#include <llvm-c/ExecutionEngine.h>
 
 
-void
-lp_build_init(void);
+extern LLVMModuleRef lp_build_module;
+extern LLVMExecutionEngineRef lp_build_engine;
+extern LLVMModuleProviderRef lp_build_provider;
+extern LLVMTargetDataRef lp_build_target;
 
 
-#ifdef __cplusplus
-}
-#endif
+void
+lp_build_init(void);
 
 
 #endif /* !LP_BLD_INIT_H */
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_printf.c b/src/gallium/auxiliary/gallivm/lp_bld_printf.c
index e4fa2c264c..153ba5b15b 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_printf.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_printf.c
@@ -54,6 +54,7 @@ lp_get_printf_arg_count(const char *fmt)
 	       p += 3;
                continue;
 	    }
+	    /* fallthrough */
 	 default:
 	    count ++;
       }
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample.c b/src/gallium/auxiliary/gallivm/lp_bld_sample.c
index bb76ad4c6b..0744be6b41 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample.c
@@ -157,9 +157,10 @@ lp_build_gather(LLVMBuilderRef builder,
 
 
 /**
- * Compute the offset of a pixel.
+ * Compute the offset of a pixel block.
  *
- * x, y, z, y_stride, z_stride are vectors
+ * x, y, z, y_stride, z_stride are vectors, and they refer to pixel blocks, as
+ * per format description, and not individual pixels.
  */
 LLVMValueRef
 lp_build_sample_offset(struct lp_build_context *bld,
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
index 5b56f2cf3b..395eaaba26 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
@@ -211,7 +211,7 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
    const int dims = texture_dims(bld->static_state->target);
    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
    LLVMValueRef offset;
-   LLVMValueRef packed;
+   LLVMValueRef i, j;
    LLVMValueRef use_border = NULL;
 
    /* use_border = x < 0 || x >= width || y < 0 || y >= height */
@@ -249,6 +249,43 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
    }
 
    /*
+    * Describe the coordinates in terms of pixel blocks.
+    *
+    * TODO: pixel blocks are power of two. LLVM should convert rem/div to
+    * bit arithmetic. Verify this.
+    */
+
+   if (bld->format_desc->block.width == 1) {
+      i = bld->uint_coord_bld.zero;
+   }
+   else {
+      LLVMValueRef block_width = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.width);
+      i = LLVMBuildURem(bld->builder, x, block_width, "");
+      x = LLVMBuildUDiv(bld->builder, x, block_width, "");
+   }
+
+   if (bld->format_desc->block.height == 1) {
+      j = bld->uint_coord_bld.zero;
+   }
+   else {
+      LLVMValueRef block_height = lp_build_const_int_vec(bld->uint_coord_bld.type, bld->format_desc->block.height);
+      j = LLVMBuildURem(bld->builder, y, block_height, "");
+      y = LLVMBuildUDiv(bld->builder, y, block_height, "");
+   }
+
+   /* convert x,y,z coords to linear offset from start of texture, in bytes */
+   offset = lp_build_sample_offset(&bld->uint_coord_bld,
+                                   bld->format_desc,
+                                   x, y, z, y_stride, z_stride);
+
+   lp_build_fetch_rgba_soa(bld->builder,
+                           bld->format_desc,
+                           bld->texel_type,
+                           data_ptr, offset,
+                           i, j,
+                           texel);
+
+   /*
     * Note: if we find an app which frequently samples the texture border
     * we might want to implement a true conditional here to avoid sampling
     * the texture whenever possible (since that's quite a bit of code).
@@ -263,30 +300,6 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
     * the texel color results with the border color.
     */
 
-   /* convert x,y,z coords to linear offset from start of texture, in bytes */
-   offset = lp_build_sample_offset(&bld->uint_coord_bld,
-                                   bld->format_desc,
-                                   x, y, z, y_stride, z_stride);
-
-   assert(bld->format_desc->block.width == 1);
-   assert(bld->format_desc->block.height == 1);
-   assert(bld->format_desc->block.bits <= bld->texel_type.width);
-
-   /* gather the texels from the texture */
-   packed = lp_build_gather(bld->builder,
-                            bld->texel_type.length,
-                            bld->format_desc->block.bits,
-                            bld->texel_type.width,
-                            data_ptr, offset);
-
-   texel[0] = texel[1] = texel[2] = texel[3] = NULL;
-
-   /* convert texels to float rgba */
-   lp_build_unpack_rgba_soa(bld->builder,
-                            bld->format_desc,
-                            bld->texel_type,
-                            packed, texel);
-
    if (use_border) {
       /* select texel color or border color depending on use_border */
       int chan;