Make some existing optimizations that would only trigger on scalars

also apply to vectors.  This allows us to compile this:

#include <emmintrin.h>
__m128i a(__m128 a, __m128 b) { return a==a & b==b; }
__m128i b(__m128 a, __m128 b) { return a!=a | b!=b; }

to:

_a:
	cmpordps	%xmm1, %xmm0
	ret
_b:
	cmpunordps	%xmm1, %xmm0
	ret

with clang instead of to a ton of horrible code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@76863 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Scalar/InstructionCombining.cpp b/lib/Transforms/Scalar/InstructionCombining.cpp
index c3055ae482d17dc56b3445df33eb82f701977041..000243ae64f0e1ae96382a60679561de0d22c863 100644 (file)
--- a/lib/Transforms/Scalar/InstructionCombining.cpp
+++ b/lib/Transforms/Scalar/InstructionCombining.cpp
@@ -3948,6 +3948,13 @@ Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
         return new FCmpInst(*Context, FCmpInst::FCMP_ORD, 
                             LHS->getOperand(0), RHS->getOperand(0));
       }
+    
+    // Handle vector zeros.  This occurs because the canonical form of
+    // "fcmp ord x,x" is "fcmp ord x, 0".
+    if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
+        isa<ConstantAggregateZero>(RHS->getOperand(1)))
+      return new FCmpInst(*Context, FCmpInst::FCMP_ORD, 
+                          LHS->getOperand(0), RHS->getOperand(0));
     return 0;
   }
   
@@ -4242,7 +4249,8 @@ Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
     if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
       if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
         const Type *SrcTy = Op0C->getOperand(0)->getType();
-        if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
+        if (SrcTy == Op1C->getOperand(0)->getType() &&
+            SrcTy->isIntOrIntVector() &&
             // Only do this if the casts both really cause code to be generated.
             ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0), 
                               I.getType(), TD) &&
@@ -4901,7 +4909,8 @@ Instruction *InstCombiner::visitOr(BinaryOperator &I) {
         if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
             !isa<ICmpInst>(Op1C->getOperand(0))) {
           const Type *SrcTy = Op0C->getOperand(0)->getType();
-          if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
+          if (SrcTy == Op1C->getOperand(0)->getType() &&
+              SrcTy->isIntOrIntVector() &&
               // Only do this if the casts both really cause code to be
               // generated.
               ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0), 
@@ -4937,6 +4946,15 @@ Instruction *InstCombiner::visitOr(BinaryOperator &I) {
             return new FCmpInst(*Context, FCmpInst::FCMP_UNO, 
                                 LHS->getOperand(0), RHS->getOperand(0));
           }
+        
+        // Handle vector zeros.  This occurs because the canonical form of
+        // "fcmp uno x,x" is "fcmp uno x, 0".
+        if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
+            isa<ConstantAggregateZero>(RHS->getOperand(1)))
+          return new FCmpInst(*Context, FCmpInst::FCMP_UNO, 
+                              LHS->getOperand(0), RHS->getOperand(0));
+        
+        
       } else {
         Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
         FCmpInst::Predicate Op0CC, Op1CC;

diff --git a/test/Transforms/InstCombine/vector-casts.ll b/test/Transforms/InstCombine/vector-casts.ll
index c6d1eaa1ff025b72080f372fd74111fbbfa443e9..e6171852061cc080531538b246666ced5e980813 100644 (file)
--- a/test/Transforms/InstCombine/vector-casts.ll
+++ b/test/Transforms/InstCombine/vector-casts.ll
@@ -5,7 +5,7 @@ define <2 x i1> @test1(<2 x i64> %a) {
   %t = trunc <2 x i64> %a to <2 x i1>
   ret <2 x i1> %t
 
-; CHECK: define <2 x i1> @test1
+; CHECK: @test1
 ; CHECK:   and <2 x i64> %a, <i64 1, i64 1>
 ; CHECK:   icmp ne <2 x i64> %tmp, zeroinitializer
 }
@@ -16,7 +16,36 @@ define <2 x i64> @test2(<2 x i64> %a) {
   %t = ashr <2 x i64> %b, <i64 1, i64 1>
   ret <2 x i64> %t
 
-; CHECK: define <2 x i64> @test2
+; CHECK: @test2
 ; CHECK:   and <2 x i64> %a, <i64 65535, i64 65535>
 ; CHECK:   lshr <2 x i64> %b, <i64 1, i64 1>
 }
+
+
+
+define <2 x i64> @test3(<4 x float> %a, <4 x float> %b) nounwind readnone {
+entry:
+       %cmp = fcmp ord <4 x float> %a, zeroinitializer 
+       %sext = sext <4 x i1> %cmp to <4 x i32> 
+       %cmp4 = fcmp ord <4 x float> %b, zeroinitializer
+       %sext5 = sext <4 x i1> %cmp4 to <4 x i32>
+       %and = and <4 x i32> %sext, %sext5
+       %conv = bitcast <4 x i32> %and to <2 x i64>
+       ret <2 x i64> %conv
+        
+; CHECK: @test3
+; CHECK:   fcmp ord <4 x float> %a, %b
+}
+
+define <2 x i64> @test4(<4 x float> %a, <4 x float> %b) nounwind readnone {
+entry:
+       %cmp = fcmp uno <4 x float> %a, zeroinitializer
+       %sext = sext <4 x i1> %cmp to <4 x i32>
+       %cmp4 = fcmp uno <4 x float> %b, zeroinitializer
+       %sext5 = sext <4 x i1> %cmp4 to <4 x i32>
+       %or = or <4 x i32> %sext, %sext5
+       %conv = bitcast <4 x i32> %or to <2 x i64>
+       ret <2 x i64> %conv
+; CHECK: @test4
+; CHECK:   fcmp uno <4 x float> %a, %b
+}