AMDGPU: Fix AMDGPUPromoteAlloca breaking addrspacecasts

[arsenm]

Summary

The users of the addrspacecast were having their types incorrectly
changed, producing invalid bitcasts between address spaces.

[• tstellarAMD]

LGTM.

[arsenm]

r289307

lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp

for (User *User : Val->users()) {
     if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
       if (!binaryOpIsDerivedFromSameAlloca(BaseAlloca, Val, ICmp, 0, 1))
         return false;
 
       // May need to rewrite constant operands.
       WorkList.push_back(ICmp);
     }
 
+    if (UseInst->getOpcode() == Instruction::AddrSpaceCast) {
+      // Don't collect the users of this.
+      WorkList.push_back(User);
+      continue;
+    }
+
     if (!User->getType()->isPointerTy())
       continue;
 
     if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UseInst)) {
       // Be conservative if an address could be computed outside the bounds of
       // the alloca.
       if (!GEP->isInBounds())
         return false;
if (!Call) {
           CI->setOperand(0, ConstantPointerNull::get(NewTy));
 
         if (isa<ConstantPointerNull>(CI->getOperand(1)))
           CI->setOperand(1, ConstantPointerNull::get(NewTy));
 
         continue;
       }
 
-      // The operand's value should be corrected on its own.
+      // The operand's value should be corrected on its own and we don't want to
+      // touch the users.
       if (isa<AddrSpaceCastInst>(V))
         continue;
 
       Type *EltTy = V->getType()->getPointerElementType();
       PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS);
 
       // FIXME: It doesn't really make sense to try to do this for all
       // instructions.

test/CodeGen/AMDGPU/promote-alloca-addrspacecast.ll

+; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -mcpu=kaveri -amdgpu-promote-alloca < %s | FileCheck %s
+
+; The types of the users of the addrspacecast should not be changed.
+
+; CHECK-LABEL: @invalid_bitcast_addrspace(
+; CHECK: getelementptr inbounds [256 x [1 x i32]], [256 x [1 x i32]] addrspace(3)* @invalid_bitcast_addrspace.data, i32 0, i32 %14
+; CHECK: bitcast [1 x i32] addrspace(3)* %{{[0-9]+}} to half addrspace(3)*
+; CHECK: addrspacecast half addrspace(3)* %{{[0-9]+}} to half addrspace(4)*
+; CHECK: bitcast half addrspace(4)* %{{[0-9]+}} to <2 x i16> addrspace(4)*
+define amdgpu_kernel void @invalid_bitcast_addrspace() #0 {
+entry:
+  %data = alloca [1 x i32], align 4
+  %0 = bitcast [1 x i32]* %data to half*
+  %1 = addrspacecast half* %0 to half addrspace(4)*
+  %2 = bitcast half addrspace(4)* %1 to <2 x i16> addrspace(4)*
+  %3 = load <2 x i16>, <2 x i16> addrspace(4)* %2, align 2
+  %4 = bitcast <2 x i16> %3 to <2 x half>
+  ret void
+}
+
+attributes #0 = { nounwind }