Index: lib/Transforms/Scalar/InferAddressSpaces.cpp =================================================================== --- lib/Transforms/Scalar/InferAddressSpaces.cpp +++ lib/Transforms/Scalar/InferAddressSpaces.cpp @@ -448,6 +448,12 @@ return ConstantExpr::getBitCast(CE->getOperand(0), TargetType); } + if (CE->getOpcode() == Instruction::BitCast) { + if (Value *NewOperand = ValueWithNewAddrSpace.lookup(CE->getOperand(0))) + return ConstantExpr::getBitCast(cast(NewOperand), TargetType); + return ConstantExpr::getAddrSpaceCast(CE, TargetType); + } + if (CE->getOpcode() == Instruction::Select) { Constant *Src0 = CE->getOperand(1); Constant *Src1 = CE->getOperand(2); Index: test/Transforms/InferAddressSpaces/NVPTX/bug31948.ll =================================================================== --- /dev/null +++ test/Transforms/InferAddressSpaces/NVPTX/bug31948.ll @@ -0,0 +1,24 @@ +; RUN: opt -S -mtriple=nvptx64-nvidia-cuda -infer-address-spaces %s | FileCheck %s + +target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64" + +%struct.bar = type { float, float* } + +@var1 = local_unnamed_addr addrspace(3) externally_initialized global %struct.bar undef, align 8 + +; CHECK-LABEL: @bug31948( +; CHECK: %tmp = load float*, float* addrspace(3)* getelementptr inbounds (%struct.bar, %struct.bar addrspace(3)* @var1, i64 0, i32 1), align 8 +; CHECK: %tmp1 = load float, float* %tmp, align 4 +; CHECK: store float %conv1, float* %tmp, align 4 +; CHECK: store i32 32, i32 addrspace(3)* addrspacecast (i32* bitcast (float** getelementptr (%struct.bar, %struct.bar* addrspacecast (%struct.bar addrspace(3)* @var1 to %struct.bar*), i64 0, i32 1) to i32*) to i32 addrspace(3)*), align 4 +define void @bug31948(float %a, float* nocapture readnone %x, float* nocapture readnone %y) local_unnamed_addr #0 { +entry: + %tmp = load float*, float** getelementptr (%struct.bar, %struct.bar* addrspacecast (%struct.bar addrspace(3)* @var1 to %struct.bar*), i64 0, i32 1), align 8 + %tmp1 = load float, float* %tmp, align 4 + %conv1 = fadd float %tmp1, 1.000000e+00 + store float %conv1, float* %tmp, align 4 + store i32 32, i32* bitcast (float** getelementptr (%struct.bar, %struct.bar* addrspacecast (%struct.bar addrspace(3)* @var1 to %struct.bar*), i64 0, i32 1) to i32*), align 4 + ret void +} + +attributes #0 = { norecurse nounwind } Index: test/Transforms/InferAddressSpaces/NVPTX/lit.local.cfg =================================================================== --- /dev/null +++ test/Transforms/InferAddressSpaces/NVPTX/lit.local.cfg @@ -0,0 +1,2 @@ +if not 'NVPTX' in config.root.targets: + config.unsupported = True