Fix a bug where scalarrepl would discard offset if type would match.

In practice this can only happen on code with already undefined behavior,
but this is still a good thing to handle correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@46539 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
index fc4a6aa278d71869e1ff181a0bcd111cafe3764a..c80f050730a04073d6974643527058b0e9c46bc3 100644 (file)
--- a/lib/Transforms/Scalar/ScalarReplAggregates.cpp
+++ b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
@@ -1078,7 +1078,7 @@ void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) {
     if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
       // The load is a bit extract from NewAI shifted right by Offset bits.
       Value *NV = new LoadInst(NewAI, LI->getName(), LI);
-      if (NV->getType() == LI->getType()) {
+      if (NV->getType() == LI->getType() && Offset == 0) {
         // We win, no conversion needed.
       } else if (const VectorType *PTy = dyn_cast<VectorType>(NV->getType())) {
         // If the result alloca is a vector type, this is either an element
@@ -1232,7 +1232,7 @@ void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) {
       SI->eraseFromParent();
       
     } else if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
-       ConvertUsesToScalar(CI, NewAI, Offset);
+      ConvertUsesToScalar(CI, NewAI, Offset);
       CI->eraseFromParent();
     } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
       const PointerType *AggPtrTy = 

diff --git a/test/Transforms/ScalarRepl/2008-01-29-PromoteBug.ll b/test/Transforms/ScalarRepl/2008-01-29-PromoteBug.ll
new file mode 100644 (file)
index 0000000..0675a96
--- /dev/null
+++ b/test/Transforms/ScalarRepl/2008-01-29-PromoteBug.ll
@@ -0,0 +1,21 @@
+; RUN: llvm-as < %s | opt -scalarrepl -instcombine | llvm-dis | grep {ret i8 17}
+; rdar://5707076
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
+target triple = "i386-apple-darwin9.1.0"
+       %struct.T = type <{ i8, [3 x i8] }>
+
+define i8 @f() {
+entry:
+       %s = alloca [1 x %struct.T], align 4            ; <[1 x %struct.T]*> [#uses=2]
+       %T3 = bitcast [1 x %struct.T]* %s to i32*
+       store i32 -61184, i32* %T3
+
+       %tmp16 = getelementptr [1 x %struct.T]* %s, i32 0, i32 0                ; <%struct.T*> [#uses=1]
+       %tmp17 = getelementptr %struct.T* %tmp16, i32 0, i32 1          ; <[3 x i8]*> [#uses=1]
+       %tmp1718 = bitcast [3 x i8]* %tmp17 to i32*             ; <i32*> [#uses=1]
+       %tmp19 = load i32* %tmp1718, align 4            ; <i32> [#uses=1]
+       %mask = and i32 %tmp19, 16777215                ; <i32> [#uses=2]
+       %mask2324 = trunc i32 %mask to i8               ; <i8> [#uses=1]
+       ret i8 %mask2324
+}
+