AMDGPU: Handle addrspacecast of constant 32-bit to flat

[arsenm]

Summary

I accidentally made this work on the GlobalISel path, and there's no
real reason not to handle this.

[arsenm]

33b45ee44b1f32ffdbc995e6fec806271b4b3ba4

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

if (SrcAS == AMDGPUAS::LOCAL_ADDRESS ||
       SDValue NonNull
         = DAG.getSetCC(SL, MVT::i1, Src, SegmentNullPtr, ISD::SETNE);
 
       return DAG.getNode(ISD::SELECT, SL, MVT::i64, NonNull, CvtPtr,
                          FlatNullPtr);
     }
   }
 
+  if (SrcAS == AMDGPUAS::CONSTANT_ADDRESS_32BIT &&
+      Op.getValueType() == MVT::i64) {
+    const SIMachineFunctionInfo *Info =
+        DAG.getMachineFunction().getInfo<SIMachineFunctionInfo>();
+    SDValue Hi = DAG.getConstant(Info->get32BitAddressHighBits(), SL, MVT::i32);
+    SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Src, Hi);
+    return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
+  }
+
   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(

llvm/test/CodeGen/AMDGPU/addrspacecast.ll

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
 }
 
+; GCN-LABEL: {{^}}use_constant32bit_to_flat_addrspacecast_0:
+; GCN: s_load_dword [[PTR:s[0-9]+]],
+; GCN: v_mov_b32_e32 v[[HI:[0-9]+]], 0
+; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], [[PTR]]
+; GCN: flat_load_dword v{{[0-9]+}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
+define amdgpu_kernel void @use_constant32bit_to_flat_addrspacecast_0(i32 addrspace(6)* %ptr) #0 {
+  %stof = addrspacecast i32 addrspace(6)* %ptr to i32*
+  %load = load volatile i32, i32* %stof
+  ret void
+}
+
+; GCN-LABEL: {{^}}use_constant32bit_to_flat_addrspacecast_1:
+; GCN: s_load_dword [[PTR:s[0-9]+]],
+; GCN: v_mov_b32_e32 v[[HI:[0-9]+]], 0xffff8000
+; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], [[PTR]]
+; GCN: flat_load_dword v{{[0-9]+}}, v{{\[}}[[LO]]:[[HI]]{{\]}}
+define amdgpu_kernel void @use_constant32bit_to_flat_addrspacecast_1(i32 addrspace(6)* %ptr) #3 {
+  %stof = addrspacecast i32 addrspace(6)* %ptr to i32*
+  %load = load volatile i32, i32* %stof
+  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 }
+attributes #3 = { nounwind "amdgpu-32bit-address-high-bits"="0xffff8000" }

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
 }