llvmpipe: Specialize arithmetic operations.

4 files changed, 420 insertions, 244 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.c b/src/gallium/drivers/llvmpipe/lp_bld_arit.c
index b5b4148ac2..f45b7d82f1 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_arit.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.c
@@ -53,7 +53,7 @@
 LLVMTypeRef
 lp_build_elem_type(union lp_type type)
 {
-   if (type.kind == LP_TYPE_FLOAT) {
+   if (type.floating) {
       assert(type.sign);
       switch(type.width) {
          case 32:
@@ -90,9 +90,15 @@ lp_build_vec_type(union lp_type type)
 boolean
 lp_check_elem_type(union lp_type type, LLVMTypeRef elem_type) 
 {
-   LLVMTypeKind elem_kind = LLVMGetTypeKind(elem_type);
+   LLVMTypeKind elem_kind;
 
-   if (type.kind == LP_TYPE_FLOAT) {
+   assert(elem_type);
+   if(!elem_type)
+      return FALSE;
+
+   elem_kind = LLVMGetTypeKind(elem_type);
+
+   if (type.floating) {
       switch(type.width) {
       case 32:
          if(elem_kind != LLVMFloatTypeKind)
@@ -124,6 +130,10 @@ lp_check_vec_type(union lp_type type, LLVMTypeRef vec_type)
 {
    LLVMTypeRef elem_type;
 
+   assert(vec_type);
+   if(!vec_type)
+      return FALSE;
+
    if(LLVMGetTypeKind(vec_type) != LLVMVectorTypeKind)
       return FALSE;
 
@@ -136,6 +146,73 @@ lp_check_vec_type(union lp_type type, LLVMTypeRef vec_type)
 }
 
 
+boolean
+lp_check_value(union lp_type type, LLVMValueRef val) 
+{
+   LLVMTypeRef vec_type;
+
+   assert(val);
+   if(!val)
+      return FALSE;
+
+   vec_type = LLVMTypeOf(val);
+
+   return lp_check_vec_type(type, vec_type);
+}
+
+
+LLVMValueRef
+lp_build_undef(union lp_type type)
+{
+   LLVMTypeRef vec_type = lp_build_vec_type(type);
+   return LLVMGetUndef(vec_type);
+}
+               
+
+LLVMValueRef
+lp_build_zero(union lp_type type)
+{
+   LLVMTypeRef vec_type = lp_build_vec_type(type);
+   return LLVMConstNull(vec_type);
+}
+               
+
+LLVMValueRef
+lp_build_one(union lp_type type)
+{
+   LLVMTypeRef elem_type;
+   LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
+   unsigned i;
+
+   assert(type.length < LP_MAX_VECTOR_LENGTH);
+
+   elem_type = lp_build_elem_type(type);
+
+   if(type.floating)
+      elems[0] = LLVMConstReal(elem_type, 1.0);
+   else if(type.fixed)
+      elems[0] = LLVMConstInt(elem_type, 1LL << (type.width/2), 0);
+   else if(!type.norm)
+      elems[0] = LLVMConstInt(elem_type, 1, 0);
+   else {
+      /* special case' -- 1.0 for normalized types is more easily attained if
+       * we start with a vector consisting of all bits set */
+      LLVMTypeRef vec_type = LLVMVectorType(elem_type, type.length);
+      LLVMValueRef vec = LLVMConstAllOnes(vec_type);
+
+      if(type.sign)
+         vec = LLVMConstLShr(vec, LLVMConstInt(LLVMInt32Type(), 1, 0));
+
+      return vec;
+   }
+
+   for(i = 1; i < type.length; ++i)
+      elems[i] = elems[0];
+
+   return LLVMConstVector(elems, type.length);
+}
+               
+
 LLVMValueRef
 lp_build_const_aos(union lp_type type, 
                    double r, double g, double b, double a, 
@@ -154,20 +231,18 @@ lp_build_const_aos(union lp_type type,
    if(swizzle == NULL)
       swizzle = default_swizzle;
 
-   if(type.kind == LP_TYPE_FLOAT) {
-      for(i = 0; i < type.length; i += 4) {
-         elems[i + swizzle[0]] = LLVMConstReal(elem_type, r);
-         elems[i + swizzle[1]] = LLVMConstReal(elem_type, g);
-         elems[i + swizzle[2]] = LLVMConstReal(elem_type, b);
-         elems[i + swizzle[3]] = LLVMConstReal(elem_type, a);
-      }
+   if(type.floating) {
+      elems[swizzle[0]] = LLVMConstReal(elem_type, r);
+      elems[swizzle[1]] = LLVMConstReal(elem_type, g);
+      elems[swizzle[2]] = LLVMConstReal(elem_type, b);
+      elems[swizzle[3]] = LLVMConstReal(elem_type, a);
    }
    else {
       unsigned shift;
       long long llscale;
       double dscale;
 
-      if(type.kind == LP_TYPE_FIXED)
+      if(type.fixed)
          shift = type.width/2;
       else if(type.norm)
          shift = type.sign ? type.width - 1 : type.width;
@@ -178,14 +253,15 @@ lp_build_const_aos(union lp_type type,
       dscale = (double)llscale;
       assert((long long)dscale == llscale);
 
-      for(i = 0; i < type.length; i += 4) {
-         elems[i + swizzle[0]] = LLVMConstInt(elem_type, r*dscale + 0.5, 0);
-         elems[i + swizzle[1]] = LLVMConstInt(elem_type, g*dscale + 0.5, 0);
-         elems[i + swizzle[2]] = LLVMConstInt(elem_type, b*dscale + 0.5, 0);
-         elems[i + swizzle[3]] = LLVMConstInt(elem_type, a*dscale + 0.5, 0);
-      }
+      elems[swizzle[0]] = LLVMConstInt(elem_type, r*dscale + 0.5, 0);
+      elems[swizzle[1]] = LLVMConstInt(elem_type, g*dscale + 0.5, 0);
+      elems[swizzle[2]] = LLVMConstInt(elem_type, b*dscale + 0.5, 0);
+      elems[swizzle[3]] = LLVMConstInt(elem_type, a*dscale + 0.5, 0);
    }
 
+   for(i = 4; i < type.length; ++i)
+      elems[i] = elems[i % 4];
+
    return LLVMConstVector(elems, type.length);
 }
                
@@ -219,165 +295,261 @@ lp_build_intrinsic_binary(LLVMBuilderRef builder,
 }
 
 
-LLVMValueRef
-lp_build_add(LLVMBuilderRef builder,
-             LLVMValueRef a,
-             LLVMValueRef b,
-             LLVMValueRef zero)
+static LLVMValueRef
+lp_build_min_simple(struct lp_build_context *bld,
+                    LLVMValueRef a,
+                    LLVMValueRef b)
 {
-   if(a == zero)
-      return b;
-   else if(b == zero)
-      return a;
-   else if(LLVMIsConstant(a) && LLVMIsConstant(b))
-      return LLVMConstAdd(a, b);
+   const union lp_type type = bld->type;
+   const char *intrinsic = NULL;
+   LLVMValueRef cond;
+
+   /* TODO: optimize the constant case */
+
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+   if(type.width * type.length == 128) {
+      if(type.floating)
+         if(type.width == 32)
+            intrinsic = "llvm.x86.sse.min.ps";
+         if(type.width == 64)
+            intrinsic = "llvm.x86.sse2.min.pd";
+      else {
+         if(type.width == 8 && !type.sign)
+            intrinsic = "llvm.x86.sse2.pminu.b";
+         if(type.width == 16 && type.sign)
+            intrinsic = "llvm.x86.sse2.pmins.w";
+      }
+   }
+#endif
+   
+   if(intrinsic)
+      return lp_build_intrinsic_binary(bld->builder, intrinsic, a, b);
+
+   if(type.floating)
+      cond = LLVMBuildFCmp(bld->builder, LLVMRealULT, a, b, "");
    else
-      return LLVMBuildAdd(builder, a, b, "");
+      cond = LLVMBuildICmp(bld->builder, type.sign ? LLVMIntSLT : LLVMIntULT, a, b, "");
+   return LLVMBuildSelect(bld->builder, cond, a, b, "");
 }
 
 
-LLVMValueRef
-lp_build_sub(LLVMBuilderRef builder,
-             LLVMValueRef a,
-             LLVMValueRef b,
-             LLVMValueRef zero)
+static LLVMValueRef
+lp_build_max_simple(struct lp_build_context *bld,
+                    LLVMValueRef a,
+                    LLVMValueRef b)
 {
-   if(b == zero)
-      return a;
-   else if(a == b)
-      return zero;
-   else if(LLVMIsConstant(a) && LLVMIsConstant(b))
-      return LLVMConstSub(a, b);
+   const union lp_type type = bld->type;
+   const char *intrinsic = NULL;
+   LLVMValueRef cond;
+
+   /* TODO: optimize the constant case */
+
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+   if(type.width * type.length == 128) {
+      if(type.floating)
+         if(type.width == 32)
+            intrinsic = "llvm.x86.sse.max.ps";
+         if(type.width == 64)
+            intrinsic = "llvm.x86.sse2.max.pd";
+      else {
+         if(type.width == 8 && !type.sign)
+            intrinsic = "llvm.x86.sse2.pmaxu.b";
+         if(type.width == 16 && type.sign)
+            intrinsic = "llvm.x86.sse2.pmaxs.w";
+      }
+   }
+#endif
+
+   if(intrinsic)
+      return lp_build_intrinsic_binary(bld->builder, intrinsic, a, b);
+
+   if(type.floating)
+      cond = LLVMBuildFCmp(bld->builder, LLVMRealULT, a, b, "");
    else
-      return LLVMBuildSub(builder, a, b, "");
+      cond = LLVMBuildICmp(bld->builder, type.sign ? LLVMIntSLT : LLVMIntULT, a, b, "");
+   return LLVMBuildSelect(bld->builder, cond, b, a, "");
 }
 
 
 LLVMValueRef
-lp_build_mul(LLVMBuilderRef builder,
-             LLVMValueRef a,
-             LLVMValueRef b,
-             LLVMValueRef zero,
-             LLVMValueRef one)
+lp_build_comp(struct lp_build_context *bld,
+              LLVMValueRef a)
 {
-   if(a == zero)
-      return zero;
-   else if(a == one)
-      return b;
-   else if(b == zero)
-      return zero;
-   else if(b == one)
-      return a;
-   else if(LLVMIsConstant(a) && LLVMIsConstant(b))
-      return LLVMConstMul(a, b);
+   const union lp_type type = bld->type;
+
+   if(a == bld->one)
+      return bld->zero;
+   if(a == bld->zero)
+      return bld->one;
+
+   if(type.norm && !type.floating && !type.fixed && !type.sign) {
+      if(LLVMIsConstant(a))
+         return LLVMConstNot(a);
+      else
+         return LLVMBuildNot(bld->builder, a, "");
+   }
+
+   if(LLVMIsConstant(a))
+      return LLVMConstSub(bld->one, a);
    else
-      return LLVMBuildMul(builder, a, b, "");
+      return LLVMBuildSub(bld->builder, bld->one, a, "");
 }
 
 
 LLVMValueRef
-lp_build_min(LLVMBuilderRef builder,
+lp_build_add(struct lp_build_context *bld,
              LLVMValueRef a,
              LLVMValueRef b)
 {
-   /* TODO: optimize the constant case */
-
-#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+   const union lp_type type = bld->type;
+   LLVMValueRef res;
 
-   return lp_build_intrinsic_binary(builder, "llvm.x86.sse.min.ps", a, b);
+   if(a == bld->zero)
+      return b;
+   if(b == bld->zero)
+      return a;
+   if(a == bld->undef || b == bld->undef)
+      return bld->undef;
 
-#else
+   if(bld->type.norm) {
+      const char *intrinsic = NULL;
 
-   LLVMValueRef cond = LLVMBuildFCmp(values->builder, LLVMRealULT, a, b, "");
-   return LLVMBuildSelect(values->builder, cond, a, b, "");
+      if(a == bld->one || b == bld->one)
+        return bld->one;
 
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+      if(type.width * type.length == 128 &&
+         !type.floating && !type.fixed) {
+         if(type.width == 8)
+            intrinsic = type.sign ? "llvm.x86.sse2.adds.b" : "llvm.x86.sse2.addus.b";
+         if(type.width == 16)
+            intrinsic = type.sign ? "llvm.x86.sse2.adds.w" : "llvm.x86.sse2.addus.w";
+      }
 #endif
+   
+      if(intrinsic)
+         return lp_build_intrinsic_binary(bld->builder, intrinsic, a, b);
+   }
+
+   if(LLVMIsConstant(a) && LLVMIsConstant(b))
+      res = LLVMConstAdd(a, b);
+   else
+      res = LLVMBuildAdd(bld->builder, a, b, "");
+
+   if(bld->type.norm && (bld->type.floating || bld->type.fixed))
+      res = lp_build_min_simple(bld, res, bld->one);
+
+   return res;
 }
 
 
 LLVMValueRef
-lp_build_max(LLVMBuilderRef builder,
+lp_build_sub(struct lp_build_context *bld,
              LLVMValueRef a,
              LLVMValueRef b)
 {
-   /* TODO: optimize the constant case */
-
-#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+   const union lp_type type = bld->type;
+   LLVMValueRef res;
 
-   return lp_build_intrinsic_binary(builder, "llvm.x86.sse.max.ps", a, b);
+   if(b == bld->zero)
+      return a;
+   if(a == bld->undef || b == bld->undef)
+      return bld->undef;
+   if(a == b)
+      return bld->zero;
 
-#else
+   if(bld->type.norm) {
+      const char *intrinsic = NULL;
 
-   LLVMValueRef cond = LLVMBuildFCmp(values->builder, LLVMRealULT, a, b, "");
-   return LLVMBuildSelect(values->builder, cond, b, a, "");
+      if(b == bld->one)
+        return bld->zero;
 
+#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
+      if(type.width * type.length == 128 &&
+         !type.floating && !type.fixed) {
+         if(type.width == 8)
+            intrinsic = type.sign ? "llvm.x86.sse2.subs.b" : "llvm.x86.sse2.subus.b";
+         if(type.width == 16)
+            intrinsic = type.sign ? "llvm.x86.sse2.subs.w" : "llvm.x86.sse2.subus.w";
+      }
 #endif
+   
+      if(intrinsic)
+         return lp_build_intrinsic_binary(bld->builder, intrinsic, a, b);
+   }
+
+   if(LLVMIsConstant(a) && LLVMIsConstant(b))
+      res = LLVMConstSub(a, b);
+   else
+      res = LLVMBuildSub(bld->builder, a, b, "");
+
+   if(bld->type.norm && (bld->type.floating || bld->type.fixed))
+      res = lp_build_max_simple(bld, res, bld->zero);
+
+   return res;
 }
 
 
 LLVMValueRef
-lp_build_add_sat(LLVMBuilderRef builder,
-                 LLVMValueRef a,
-                 LLVMValueRef b,
-                 LLVMValueRef zero,
-                 LLVMValueRef one)
+lp_build_mul(struct lp_build_context *bld,
+             LLVMValueRef a,
+             LLVMValueRef b)
 {
-   if(a == zero)
+   if(a == bld->zero)
+      return bld->zero;
+   if(a == bld->one)
       return b;
-   else if(b == zero)
+   if(b == bld->zero)
+      return bld->zero;
+   if(b == bld->one)
       return a;
-   else if(a == one || b == one)
-      return one;
-   else
-      return lp_build_min(builder, lp_build_add(builder, a, b, zero), one);
-}
+   if(a == bld->undef || b == bld->undef)
+      return bld->undef;
 
-LLVMValueRef
-lp_build_sub_sat(LLVMBuilderRef builder,
-                 LLVMValueRef a,
-                 LLVMValueRef b,
-                 LLVMValueRef zero,
-                 LLVMValueRef one)
-{
-   if(b == zero)
-      return a;
-   else if(b == one)
-      return zero;
-   else
-      return lp_build_max(builder, lp_build_sub(builder, a, b, zero), zero);
+   if(LLVMIsConstant(a) && LLVMIsConstant(b))
+      return LLVMConstMul(a, b);
+
+   return LLVMBuildMul(bld->builder, a, b, "");
 }
 
+
 LLVMValueRef
-lp_build_min_sat(LLVMBuilderRef builder,
-                 LLVMValueRef a,
-                 LLVMValueRef b,
-                 LLVMValueRef zero,
-                 LLVMValueRef one)
+lp_build_min(struct lp_build_context *bld,
+             LLVMValueRef a,
+             LLVMValueRef b)
 {
-   if(a == zero || b == zero)
-      return zero;
-   else if(a == one)
-      return b;
-   else if(b == one)
-      return a;
-   else
-      return lp_build_min(builder, a, b);
+   if(a == bld->undef || b == bld->undef)
+      return bld->undef;
+
+   if(bld->type.norm) {
+      if(a == bld->zero || b == bld->zero)
+         return bld->zero;
+      if(a == bld->one)
+         return b;
+      if(b == bld->one)
+         return a;
+   }
+
+   return lp_build_min_simple(bld, a, b);
 }
 
 
 LLVMValueRef
-lp_build_max_sat(LLVMBuilderRef builder,
-                 LLVMValueRef a,
-                 LLVMValueRef b,
-                 LLVMValueRef zero,
-                 LLVMValueRef one)
+lp_build_max(struct lp_build_context *bld,
+             LLVMValueRef a,
+             LLVMValueRef b)
 {
-   if(a == zero)
-      return b;
-   else if(b == zero)
-      return a;
-   else if(a == one || b == one)
-      return one;
-   else
-      return lp_build_max(builder, a, b);
+   if(a == bld->undef || b == bld->undef)
+      return bld->undef;
+
+   if(bld->type.norm) {
+      if(a == bld->one || b == bld->one)
+         return bld->one;
+      if(a == bld->zero)
+         return b;
+      if(b == bld->zero)
+         return a;
+   }
+
+   return lp_build_max_simple(bld, a, b);
 }
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_arit.h b/src/gallium/drivers/llvmpipe/lp_bld_arit.h
index f9a61be516..795b816507 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_arit.h
+++ b/src/gallium/drivers/llvmpipe/lp_bld_arit.h
@@ -66,7 +66,13 @@ union lp_type {
        * Integer. floating-point, or fixed point as established by the
        * lp_build_type_kind enum above.
        */
-      unsigned kind:2;
+      unsigned floating:1;
+
+      /** 
+       * Integer. floating-point, or fixed point as established by the
+       * lp_build_type_kind enum above.
+       */
+      unsigned fixed:1;
       
       /** 
        * Whether it can represent negative values or not.
@@ -79,9 +85,11 @@ union lp_type {
        * Whether values are normalized to fit [0, 1] interval, or [-1, 1] interval for
        * signed types.
        *
-       * For integer types
+       * For integer types it means the representable integer range should be
+       * interpreted as the interval above.
        *
-       * It makes no sense to use this with fixed point values.
+       * For floating and fixed point formats it means the values should be
+       * clamped to the interval above.
        */
       unsigned norm:1;
 
@@ -123,10 +131,27 @@ boolean
 lp_check_vec_type(union lp_type type, LLVMTypeRef vec_type);
 
 
+boolean
+lp_check_value(union lp_type type, LLVMValueRef val);
+
+
 /*
  * Constants
  */
 
+
+LLVMValueRef
+lp_build_undef(union lp_type type);
+
+
+LLVMValueRef
+lp_build_zero(union lp_type type);
+
+
+LLVMValueRef
+lp_build_one(union lp_type type);
+
+
 LLVMValueRef
 lp_build_const_aos(union lp_type type, 
                    double r, double g, double b, double a, 
@@ -136,66 +161,52 @@ lp_build_const_aos(union lp_type type,
  * Basic arithmetic
  */
 
-LLVMValueRef
-lp_build_add(LLVMBuilderRef builder,
-             LLVMValueRef a,
-             LLVMValueRef b,
-             LLVMValueRef zero);
 
-LLVMValueRef
-lp_build_sub(LLVMBuilderRef builder,
-             LLVMValueRef a,
-             LLVMValueRef b,
-             LLVMValueRef zero);
+/**
+ */
+struct lp_build_context
+{
+   LLVMBuilderRef builder;
+   
+   union lp_type type;
+
+   LLVMValueRef undef;
+   LLVMValueRef zero;
+   LLVMValueRef one;
+};
 
+
+/**
+ * Complement, i.e., 1 - a.
+ */
 LLVMValueRef
-lp_build_mul(LLVMBuilderRef builder,
-             LLVMValueRef a,
-             LLVMValueRef b,
-             LLVMValueRef zero,
-             LLVMValueRef one);
+lp_build_comp(struct lp_build_context *bld,
+              LLVMValueRef a);
 
 LLVMValueRef
-lp_build_min(LLVMBuilderRef builder,
+lp_build_add(struct lp_build_context *bld,
              LLVMValueRef a,
              LLVMValueRef b);
 
 LLVMValueRef
-lp_build_max(LLVMBuilderRef builder,
+lp_build_sub(struct lp_build_context *bld,
              LLVMValueRef a,
              LLVMValueRef b);
 
-/*
- * Satured arithmetic
- */
-
 LLVMValueRef
-lp_build_add_sat(LLVMBuilderRef builder,
-                 LLVMValueRef a,
-                 LLVMValueRef b,
-                 LLVMValueRef zero,
-                 LLVMValueRef one);
-
-LLVMValueRef
-lp_build_sub_sat(LLVMBuilderRef builder,
-                 LLVMValueRef a,
-                 LLVMValueRef b,
-                 LLVMValueRef zero,
-                 LLVMValueRef one);
+lp_build_mul(struct lp_build_context *bld,
+             LLVMValueRef a,
+             LLVMValueRef b);
 
 LLVMValueRef
-lp_build_min_sat(LLVMBuilderRef builder,
-                 LLVMValueRef a,
-                 LLVMValueRef b,
-                 LLVMValueRef zero,
-                 LLVMValueRef one);
+lp_build_min(struct lp_build_context *bld,
+             LLVMValueRef a,
+             LLVMValueRef b);
 
 LLVMValueRef
-lp_build_max_sat(LLVMBuilderRef builder,
-                 LLVMValueRef a,
-                 LLVMValueRef b,
-                 LLVMValueRef zero,
-                 LLVMValueRef one);
+lp_build_max(struct lp_build_context *bld,
+             LLVMValueRef a,
+             LLVMValueRef b);
 
 
 #endif /* !LP_BLD_ARIT_H */
diff --git a/src/gallium/drivers/llvmpipe/lp_bld_blend.c b/src/gallium/drivers/llvmpipe/lp_bld_blend.c
index 48d1e0028a..ce8408b79b 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_blend.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_blend.c
@@ -44,18 +44,14 @@
 
 
 /**
- * We may the same values several times, so we keep them here to avoid
- * recomputing them. Also reusing the values allows us to do simplifications
+ * We may the same bld several times, so we keep them here to avoid
+ * recomputing them. Also reusing the bld allows us to do simplifications
  * that LLVM optimization passes wouldn't normally be able to do.
  */
-struct lp_build_blend_values
+struct lp_build_blend_context
 {
-   LLVMBuilderRef builder;
+   struct lp_build_context base;
    
-   LLVMValueRef undef;
-   LLVMValueRef zero;
-   LLVMValueRef one;
-
    LLVMValueRef src;
    LLVMValueRef dst;
    LLVMValueRef const_;
@@ -73,62 +69,62 @@ struct lp_build_blend_values
 
 
 static LLVMValueRef
-lp_build_blend_factor_unswizzled(struct lp_build_blend_values *values,
+lp_build_blend_factor_unswizzled(struct lp_build_blend_context *bld,
                                  unsigned factor,
                                  boolean alpha)
 {
    switch (factor) {
    case PIPE_BLENDFACTOR_ZERO:
-      return values->zero;
+      return bld->base.zero;
    case PIPE_BLENDFACTOR_ONE:
-      return values->one;
+      return bld->base.one;
    case PIPE_BLENDFACTOR_SRC_COLOR:
    case PIPE_BLENDFACTOR_SRC_ALPHA:
-      return values->src;
+      return bld->src;
    case PIPE_BLENDFACTOR_DST_COLOR:
    case PIPE_BLENDFACTOR_DST_ALPHA:
-      return values->dst;
+      return bld->dst;
    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
       if(alpha)
-         return values->one;
+         return bld->base.one;
       else {
-         if(!values->inv_dst)
-            values->inv_dst = lp_build_sub(values->builder, values->one, values->dst, values->zero);
-         if(!values->saturate)
-            values->saturate = lp_build_min_sat(values->builder, values->src, values->inv_dst, values->zero, values->one);
-         return values->saturate;
+         if(!bld->inv_dst)
+            bld->inv_dst = lp_build_comp(&bld->base, bld->dst);
+         if(!bld->saturate)
+            bld->saturate = lp_build_min(&bld->base, bld->src, bld->inv_dst);
+         return bld->saturate;
       }
    case PIPE_BLENDFACTOR_CONST_COLOR:
    case PIPE_BLENDFACTOR_CONST_ALPHA:
-      return values->const_;
+      return bld->const_;
    case PIPE_BLENDFACTOR_SRC1_COLOR:
    case PIPE_BLENDFACTOR_SRC1_ALPHA:
       /* TODO */
       assert(0);
-      return values->zero;
+      return bld->base.zero;
    case PIPE_BLENDFACTOR_INV_SRC_COLOR:
    case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
-      if(!values->inv_src)
-         values->inv_src = lp_build_sub(values->builder, values->one, values->src, values->zero);
-      return values->inv_src;
+      if(!bld->inv_src)
+         bld->inv_src = lp_build_comp(&bld->base, bld->src);
+      return bld->inv_src;
    case PIPE_BLENDFACTOR_INV_DST_COLOR:
    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
-      if(!values->inv_dst)
-         values->inv_dst = lp_build_sub(values->builder, values->one, values->dst, values->zero);
-      return values->inv_dst;
+      if(!bld->inv_dst)
+         bld->inv_dst = lp_build_comp(&bld->base, bld->dst);
+      return bld->inv_dst;
    case PIPE_BLENDFACTOR_INV_CONST_COLOR:
    case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
-      if(!values->inv_const)
-         values->inv_const = lp_build_sub(values->builder, values->one, values->const_, values->zero);
-      return values->inv_const;
+      if(!bld->inv_const)
+         bld->inv_const = lp_build_comp(&bld->base, bld->const_);
+      return bld->inv_const;
    case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
    case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
       /* TODO */
       assert(0);
-      return values->zero;
+      return bld->base.zero;
    default:
       assert(0);
-      return values->zero;
+      return bld->base.zero;
    }
 }
 
@@ -175,13 +171,13 @@ lp_build_blend_factor_swizzle(unsigned factor)
 
 
 static LLVMValueRef
-lp_build_blend_swizzle(struct lp_build_blend_values *values,
+lp_build_blend_swizzle(struct lp_build_blend_context *bld,
                        LLVMValueRef rgb, 
                        LLVMValueRef alpha, 
                        enum lp_build_blend_swizzle rgb_swizzle,
-                       unsigned alpha_swizzle,
-                       unsigned n)
+                       unsigned alpha_swizzle)
 {
+   const unsigned n = bld->base.type.length;
    LLVMValueRef swizzles[LP_MAX_VECTOR_LENGTH];
    unsigned i, j;
 
@@ -189,14 +185,14 @@ lp_build_blend_swizzle(struct lp_build_blend_values *values,
       if(rgb_swizzle == LP_BUILD_BLEND_SWIZZLE_RGBA)
          return rgb;
 
-      alpha = values->undef;
+      alpha = bld->base.undef;
    }
 
    for(j = 0; j < n; j += 4) {
       for(i = 0; i < 4; ++i) {
          unsigned swizzle;
 
-         if(i == alpha_swizzle && alpha != values->undef) {
+         if(i == alpha_swizzle && alpha != bld->base.undef) {
             /* Take the alpha from the second shuffle argument */
             swizzle = n + j + alpha_swizzle;
          }
@@ -212,55 +208,54 @@ lp_build_blend_swizzle(struct lp_build_blend_values *values,
       }
    }
 
-   return LLVMBuildShuffleVector(values->builder, rgb, alpha, LLVMConstVector(swizzles, n), "");
+   return LLVMBuildShuffleVector(bld->base.builder, rgb, alpha, LLVMConstVector(swizzles, n), "");
 }
 
 
 static LLVMValueRef
-lp_build_blend_factor(struct lp_build_blend_values *values,
+lp_build_blend_factor(struct lp_build_blend_context *bld,
                       LLVMValueRef factor1,
                       unsigned rgb_factor,
                       unsigned alpha_factor,
-                      unsigned alpha_swizzle,
-                      unsigned n)
+                      unsigned alpha_swizzle)
 {
    LLVMValueRef rgb_factor_;
    LLVMValueRef alpha_factor_;
    LLVMValueRef factor2;
    enum lp_build_blend_swizzle rgb_swizzle;
 
-   rgb_factor_   = lp_build_blend_factor_unswizzled(values, rgb_factor,   FALSE);
-   alpha_factor_ = lp_build_blend_factor_unswizzled(values, alpha_factor, TRUE);
+   rgb_factor_   = lp_build_blend_factor_unswizzled(bld, rgb_factor,   FALSE);
+   alpha_factor_ = lp_build_blend_factor_unswizzled(bld, alpha_factor, TRUE);
 
    rgb_swizzle = lp_build_blend_factor_swizzle(rgb_factor);
 
-   factor2 = lp_build_blend_swizzle(values, rgb_factor_, alpha_factor_, rgb_swizzle, alpha_swizzle, n);
+   factor2 = lp_build_blend_swizzle(bld, rgb_factor_, alpha_factor_, rgb_swizzle, alpha_swizzle);
 
-   return lp_build_mul(values->builder, factor1, factor2, values->zero, values->one);
+   return lp_build_mul(&bld->base, factor1, factor2);
 }
 
 
 static LLVMValueRef
-lp_build_blend_func(struct lp_build_blend_values *values,
+lp_build_blend_func(struct lp_build_blend_context *bld,
                     unsigned func,
                     LLVMValueRef term1, 
                     LLVMValueRef term2)
 {
    switch (func) {
    case PIPE_BLEND_ADD:
-      return lp_build_add_sat(values->builder, term1, term2, values->zero, values->one);
+      return lp_build_add(&bld->base, term1, term2);
       break;
    case PIPE_BLEND_SUBTRACT:
-      return lp_build_sub_sat(values->builder, term1, term2, values->zero, values->one);
+      return lp_build_sub(&bld->base, term1, term2);
    case PIPE_BLEND_REVERSE_SUBTRACT:
-      return lp_build_sub_sat(values->builder, term2, term1, values->zero, values->one);
+      return lp_build_sub(&bld->base, term2, term1);
    case PIPE_BLEND_MIN:
-      return lp_build_min_sat(values->builder, term1, term2, values->zero, values->one);
+      return lp_build_min(&bld->base, term1, term2);
    case PIPE_BLEND_MAX:
-      return lp_build_max_sat(values->builder, term1, term2, values->zero, values->one);
+      return lp_build_max(&bld->base, term1, term2);
    default:
       assert(0);
-      return values->zero;
+      return bld->base.zero;
    }
 }
 
@@ -274,35 +269,33 @@ lp_build_blend(LLVMBuilderRef builder,
                LLVMValueRef const_,
                unsigned alpha_swizzle)
 {
-   struct lp_build_blend_values values;
+   struct lp_build_blend_context bld;
    LLVMValueRef src_term;
    LLVMValueRef dst_term;
-   LLVMTypeRef vec_type;
-
-   vec_type = lp_build_vec_type(type);
 
-   /*
-    * Compute constants
-    */
-   memset(&values, 0, sizeof values);
-   values.builder = builder;
-   values.undef = LLVMGetUndef(vec_type);
-   values.zero = LLVMConstNull(vec_type);
-   values.one = lp_build_const_aos(type, 1.0, 1.0, 1.0, 1.0, NULL);
+   /* It makes no sense to blend unless values are normalized */
+   assert(type.norm);
 
-   values.src = src;
-   values.dst = dst;
-   values.const_ = const_;
+   /* Setup build context */
+   memset(&bld, 0, sizeof bld);
+   bld.base.builder = builder;
+   bld.base.type = type;
+   bld.base.undef = lp_build_undef(type);
+   bld.base.zero = lp_build_zero(type);
+   bld.base.one = lp_build_one(type);
+   bld.src = src;
+   bld.dst = dst;
+   bld.const_ = const_;
 
    /* TODO: There are still a few optimization oportunities here. For certain
     * combinations it is possible to reorder the operations and therefor saving
     * some instructions. */
 
-   src_term = lp_build_blend_factor(&values, src, blend->rgb_src_factor, blend->alpha_src_factor, alpha_swizzle, type.length);
-   dst_term = lp_build_blend_factor(&values, dst, blend->rgb_dst_factor, blend->alpha_dst_factor, alpha_swizzle, type.length);
+   src_term = lp_build_blend_factor(&bld, src, blend->rgb_src_factor, blend->alpha_src_factor, alpha_swizzle);
+   dst_term = lp_build_blend_factor(&bld, dst, blend->rgb_dst_factor, blend->alpha_dst_factor, alpha_swizzle);
 
    if(blend->rgb_func == blend->alpha_func) {
-      return lp_build_blend_func(&values, blend->rgb_func, src_term, dst_term);
+      return lp_build_blend_func(&bld, blend->rgb_func, src_term, dst_term);
    }
    else {
       /* Seperate RGB / A functions */
@@ -310,9 +303,9 @@ lp_build_blend(LLVMBuilderRef builder,
       LLVMValueRef rgb;
       LLVMValueRef alpha;
 
-      rgb   = lp_build_blend_func(&values, blend->rgb_func,   src_term, dst_term);
-      alpha = lp_build_blend_func(&values, blend->alpha_func, src_term, dst_term);
+      rgb   = lp_build_blend_func(&bld, blend->rgb_func,   src_term, dst_term);
+      alpha = lp_build_blend_func(&bld, blend->alpha_func, src_term, dst_term);
 
-      return lp_build_blend_swizzle(&values, rgb, alpha, LP_BUILD_BLEND_SWIZZLE_RGBA, alpha_swizzle, type.length);
+      return lp_build_blend_swizzle(&bld, rgb, alpha, LP_BUILD_BLEND_SWIZZLE_RGBA, alpha_swizzle);
    }
 }
diff --git a/src/gallium/drivers/llvmpipe/lp_test_blend.c b/src/gallium/drivers/llvmpipe/lp_test_blend.c
index 3becac1871..5f46bb5f36 100644
--- a/src/gallium/drivers/llvmpipe/lp_test_blend.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_blend.c
@@ -81,7 +81,7 @@ add_blend_test(LLVMModuleRef module,
    LLVMValueRef res;
 
    type.value = 0;
-   type.kind = LP_TYPE_FLOAT;
+   type.floating = TRUE;
    type.sign = TRUE;
    type.norm = TRUE;
    type.width = 32;