Fix an overaggressive SimplifyDemandedBits optimization on urem. This

fixes the 254.gap regression on x86 and the 403.gcc regression on x86-64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50537 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Scalar/InstructionCombining.cpp b/lib/Transforms/Scalar/InstructionCombining.cpp
index 507e6f2a854d4069cd675b5a391ed28b81d18e4a..ce052f724e84b98996219c610f322d8bab60e068 100644 (file)
--- a/lib/Transforms/Scalar/InstructionCombining.cpp
+++ b/lib/Transforms/Scalar/InstructionCombining.cpp
@@ -1762,11 +1762,12 @@ bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
 
     APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
     APInt AllOnes = APInt::getAllOnesValue(BitWidth);
-    ComputeMaskedBits(I->getOperand(0), AllOnes,
-                      KnownZero2, KnownOne2, Depth+1);
+    if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
+                             KnownZero2, KnownOne2, Depth+1))
+      return true;
+
     uint32_t Leaders = KnownZero2.countLeadingOnes();
-    APInt HighZeros = APInt::getHighBitsSet(BitWidth, Leaders);
-    if (SimplifyDemandedBits(I->getOperand(1), ~HighZeros,
+    if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
                              KnownZero2, KnownOne2, Depth+1))
       return true;
 

diff --git a/test/Transforms/InstCombine/urem-simplify-bug.ll b/test/Transforms/InstCombine/urem-simplify-bug.ll
new file mode 100644 (file)
index 0000000..15956f2
--- /dev/null
+++ b/test/Transforms/InstCombine/urem-simplify-bug.ll
@@ -0,0 +1,32 @@
+; RUN: llvm-as < %s | opt -instcombine | llvm-dis | grep {= or i32 %x, -5      }
+
+@.str = internal constant [5 x i8] c"foo\0A\00"                ; <[5 x i8]*> [#uses=1]
+@.str1 = internal constant [5 x i8] c"bar\0A\00"               ; <[5 x i8]*> [#uses=1]
+
+define i32 @main() nounwind  {
+entry:
+       %x = call i32 @func_11( ) nounwind              ; <i32> [#uses=1]
+       %tmp3 = or i32 %x, -5           ; <i32> [#uses=1]
+       %tmp5 = urem i32 251, %tmp3             ; <i32> [#uses=1]
+       %tmp6 = icmp ne i32 %tmp5, 0            ; <i1> [#uses=1]
+       %tmp67 = zext i1 %tmp6 to i32           ; <i32> [#uses=1]
+       %tmp9 = urem i32 %tmp67, 95             ; <i32> [#uses=1]
+       %tmp10 = and i32 %tmp9, 1               ; <i32> [#uses=1]
+       %tmp12 = icmp eq i32 %tmp10, 0          ; <i1> [#uses=1]
+       br i1 %tmp12, label %bb14, label %bb
+
+bb:            ; preds = %entry
+       br label %bb15
+
+bb14:          ; preds = %entry
+       br label %bb15
+
+bb15:          ; preds = %bb14, %bb
+       %iftmp.0.0 = phi i8* [ getelementptr ([5 x i8]* @.str1, i32 0, i32 0), %bb14 ], [ getelementptr ([5 x i8]* @.str, i32 0, i32 0), %bb ]          ; <i8*> [#uses=1]
+       %tmp17 = call i32 (i8*, ...)* @printf( i8* %iftmp.0.0 ) nounwind                ; <i32> [#uses=0]
+       ret i32 0
+}
+
+declare i32 @func_11()
+
+declare i32 @printf(i8*, ...) nounwind