test case: inferred different alloca types

Author: inbelic

llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp

@@ -114,10 +114,11 @@ static bool hasUndefSource(AnyMemTransferInst *MI) {
   return isa<AllocaInst>(Src) && Src->hasOneUse();
 }
 
-// Optimistically infer a type from either the Src or Dest.
+// Optimistically infer a type from either the Src or Dest. Prefers the Src
+// over the Dest type.
 //
-// Returns the DefaultTy if unable to infer a type, if inferred types
-// disagree, or, if inferred type does not match the size of load/store.
+// Returns the DefaultTy if unable to infer a type, or, if inferred type does
+// not match the size of load/store.
 static Type *inferType(const DataLayout &DL, IntegerType *DefaultTy, Value *Src,
                        Value *Dest) {
   Type *SrcTy = nullptr;
@@ -129,9 +130,6 @@ static Type *inferType(const DataLayout &DL, IntegerType *DefaultTy, Value *Src,
   if (auto *DestAI = dyn_cast<AllocaInst>(Dest))
     DestTy = DestAI->getAllocatedType();
 
-  if (SrcTy && DestTy && SrcTy != DestTy)
-    return DefaultTy; // Unable to infer common type
-
   Type *InferredTy = SrcTy ? SrcTy : DestTy;
 
   if (InferredTy &&

llvm/test/Transforms/InstCombine/memcpy_alloca.ll

@@ -105,4 +105,29 @@ define void @test8(ptr %src, ptr %dest) {
   ret void
 }
 
+; Ensure we don't use alloca type if they don't agree
+
+define double @test9(ptr %src, ptr %dest) {
+; CHECK-LABEL: @test9(
+; CHECK-NEXT:    %[[TEMP:.*]] = alloca double, align 1
+; CHECK-NEXT:    %[[UNPACK0:.*]] = load i32, ptr %src, align 1
+; CHECK-NEXT:    %[[SRC_GEP:.*]] = getelementptr inbounds nuw i8, ptr %src, i64 4
+; CHECK-NEXT:    %[[UNPACK1:.*]] = load i32, ptr %[[SRC_GEP]], align 1
+; CHECK-NEXT:    store i32 %[[UNPACK0]], ptr %[[TEMP]], align 1
+; CHECK-NEXT:    %[[TEMP_GEP:.*]] = getelementptr inbounds nuw i8, ptr %[[TEMP]], i64 4
+; CHECK-NEXT:    store i32 %[[UNPACK1]], ptr %[[TEMP_GEP]], align 1
+; CHECK-NEXT:    %[[RES:.*]] = load double, ptr %[[TEMP]]
+; CHECK-NEXT:    ret double %[[RES]]
+;
+  %temp = alloca [2 x i32], align 4
+  %out = alloca double, align 1
+
+  call void @llvm.memcpy.p0.p0.i32(ptr %temp, ptr %src, i32 8, i1 false)
+  call void @llvm.memcpy.p0.p0.i32(ptr %out, ptr %temp, i32 8, i1 false)
+
+  %res = load double, ptr %out
+
+  ret double %res
+}
+
 declare void @llvm.memcpy.p0.p0.i64(ptr, ptr, i64, i1)