AMDGPU: Lower addrspacecast to 32-bit constant

[arsenm]

Summary

Somehow this was missing from the DAG path, but not global isel.

[rampitec]

LGTM

[arsenm]

fda0c8df289db56c4f009a157e9c5bfa51c06044

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

if (SrcAS == AMDGPUAS::LOCAL_ADDRESS ||
         = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Src, Aperture);
 
       return DAG.getNode(ISD::SELECT, SL, MVT::i64, NonNull,
                          DAG.getNode(ISD::BITCAST, SL, MVT::i64, CvtPtr),
                          FlatNullPtr);
     }
   }
 
+  if (ASC->getDestAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT &&
+      Src.getValueType() == MVT::i64)
+    return DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, Src);
+
   // global <-> flat are no-ops and never emitted.
 
   const MachineFunction &MF = DAG.getMachineFunction();
   DiagnosticInfoUnsupported InvalidAddrSpaceCast(
     MF.getFunction(), "invalid addrspacecast", SL.getDebugLoc());
   DAG.getContext()->diagnose(InvalidAddrSpaceCast);
 
   return DAG.getUNDEF(ASC->getValueType(0));

llvm/test/CodeGen/AMDGPU/addrspacecast.ll

define amdgpu_kernel void @store_flat_scratch(i32 addrspace(1)* noalias %out, i32) #0 {
   store volatile i32 %x, i32* %fptr
   ; Dummy call
   call void @llvm.amdgcn.s.barrier() #1
   %reload = load volatile i32, i32* %fptr, align 4
   store volatile i32 %reload, i32 addrspace(1)* %out, align 4
   ret void
 }
 
+; HSA-LABEL: {{^}}use_constant_to_constant32_addrspacecast
+; GFX9: s_load_dwordx2 [[PTRPTR:s\[[0-9]+:[0-9]+\]]], s[4:5], 0x0{{$}}
+; GFX9: s_load_dword [[OFFSET:s[0-9]+]], s[4:5], 0x8{{$}}
+; GFX9: s_load_dwordx2 s{{\[}}[[PTR_LO:[0-9]+]]:[[PTR_HI:[0-9]+]]{{\]}}, [[PTRPTR]], 0x0{{$}}
+; GFX9: s_mov_b32 s[[PTR_HI]], 0{{$}}
+; GFX9: s_add_i32 s[[PTR_LO]], s[[PTR_LO]], [[OFFSET]]
+; GFX9: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0x0{{$}}
+define amdgpu_kernel void @use_constant_to_constant32_addrspacecast(i8 addrspace(4)* addrspace(4)* %ptr.ptr, i32 %offset) #0 {
+  %ptr = load volatile i8 addrspace(4)*, i8 addrspace(4)* addrspace(4)* %ptr.ptr
+  %addrspacecast = addrspacecast i8 addrspace(4)* %ptr to i8 addrspace(6)*
+  %gep = getelementptr i8, i8 addrspace(6)* %addrspacecast, i32 %offset
+  %ptr.cast = bitcast i8 addrspace(6)* %gep to i32 addrspace(6)*
+  %load = load volatile i32, i32 addrspace(6)* %ptr.cast, align 4
+  ret void
+}
+
+; HSA-LABEL: {{^}}use_global_to_constant32_addrspacecast
+; GFX9: s_load_dwordx2 [[PTRPTR:s\[[0-9]+:[0-9]+\]]], s[4:5], 0x0{{$}}
+; GFX9: s_load_dword [[OFFSET:s[0-9]+]], s[4:5], 0x8{{$}}
+; GFX9: s_load_dwordx2 s{{\[}}[[PTR_LO:[0-9]+]]:[[PTR_HI:[0-9]+]]{{\]}}, [[PTRPTR]], 0x0{{$}}
+; GFX9: s_mov_b32 s[[PTR_HI]], 0{{$}}
+; GFX9: s_add_i32 s[[PTR_LO]], s[[PTR_LO]], [[OFFSET]]
+; GFX9: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0x0{{$}}
+define amdgpu_kernel void @use_global_to_constant32_addrspacecast(i8 addrspace(1)* addrspace(4)* %ptr.ptr, i32 %offset) #0 {
+  %ptr = load volatile i8 addrspace(1)*, i8 addrspace(1)* addrspace(4)* %ptr.ptr
+  %addrspacecast = addrspacecast i8 addrspace(1)* %ptr to i8 addrspace(6)*
+  %gep = getelementptr i8, i8 addrspace(6)* %addrspacecast, i32 %offset
+  %ptr.cast = bitcast i8 addrspace(6)* %gep to i32 addrspace(6)*
+  %load = load volatile i32, i32 addrspace(6)* %ptr.cast, align 4
+  ret void
+}
+
 declare void @llvm.amdgcn.s.barrier() #1
 declare i32 @llvm.amdgcn.workitem.id.x() #2
 
 attributes #0 = { nounwind }
 attributes #1 = { nounwind convergent }
 attributes #2 = { nounwind readnone }

llvm/test/CodeGen/AMDGPU/invalid-addrspacecast.ll

 ; RUN: not llc -march=amdgcn -mcpu=bonaire -mattr=-promote-alloca < %s 2>&1 | FileCheck -check-prefix=ERROR %s
 
 ; ERROR: error: <unknown>:0:0: in function use_group_to_global_addrspacecast void (i32 addrspace(3)*): invalid addrspacecast
 define amdgpu_kernel void @use_group_to_global_addrspacecast(i32 addrspace(3)* %ptr) {
   %stof = addrspacecast i32 addrspace(3)* %ptr to i32 addrspace(1)*
   store volatile i32 0, i32 addrspace(1)* %stof
   ret void
 }
 
 ; ERROR: error: <unknown>:0:0: in function use_constant32bit_to_flat_addrspacecast void (i32 addrspace(6)*): invalid addrspacecast
 define amdgpu_kernel void @use_constant32bit_to_flat_addrspacecast(i32 addrspace(6)* %ptr) #0 {
   %stof = addrspacecast i32 addrspace(6)* %ptr to i32*
   store volatile i32 7, i32* %stof
   ret void
 }
+
+; ERROR: error: <unknown>:0:0: in function use_local_to_constant32bit_addrspacecast void (i32 addrspace(3)*): invalid addrspacecast
+define amdgpu_kernel void @use_local_to_constant32bit_addrspacecast(i32 addrspace(3)* %ptr) #0 {
+  %stof = addrspacecast i32 addrspace(3)* %ptr to i32 addrspace(6)*
+  %load = load volatile i32, i32 addrspace(6)* %stof
+  ret void
+}