DAG: Fix expansion of unaligned FP loads and stores

[arsenm]

Summary

This was trying to scalarizing a scalar FP type,
resulting in an assert.

Fixes unaligned f64 stack stores for AMDGPU.

[nhaehnle]

LGTM

[arsenm]

r342132

lib/CodeGen/SelectionDAG/TargetLowering.cpp

TargetLowering::expandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG) const {
   EVT VT = LD->getValueType(0);
   EVT LoadedVT = LD->getMemoryVT();
   SDLoc dl(LD);
   auto &MF = DAG.getMachineFunction();
 
   if (VT.isFloatingPoint() || VT.isVector()) {
     EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
     if (isTypeLegal(intVT) && isTypeLegal(LoadedVT)) {
-      if (!isOperationLegalOrCustom(ISD::LOAD, intVT)) {
+      if (!isOperationLegalOrCustom(ISD::LOAD, intVT) &&
+          LoadedVT.isVector()) {
         // Scalarize the load and let the individual components be handled.
         SDValue Scalarized = scalarizeVectorLoad(LD, DAG);
         if (Scalarized->getOpcode() == ISD::MERGE_VALUES)
           return std::make_pair(Scalarized.getOperand(0), Scalarized.getOperand(1));
         return std::make_pair(Scalarized.getValue(0), Scalarized.getValue(1));
       }
 
       // Expand to a (misaligned) integer load of the same size,
SDValue TargetLowering::expandUnalignedStore(StoreSDNode *ST,
   assert(ST->getAddressingMode() == ISD::UNINDEXED &&
          "unaligned indexed stores not implemented!");
   SDValue Chain = ST->getChain();
   SDValue Ptr = ST->getBasePtr();
   SDValue Val = ST->getValue();
   EVT VT = Val.getValueType();
   int Alignment = ST->getAlignment();
   auto &MF = DAG.getMachineFunction();
+  EVT MemVT = ST->getMemoryVT();
 
   SDLoc dl(ST);
-  if (ST->getMemoryVT().isFloatingPoint() ||
-      ST->getMemoryVT().isVector()) {
+  if (MemVT.isFloatingPoint() || MemVT.isVector()) {
     EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
     if (isTypeLegal(intVT)) {
-      if (!isOperationLegalOrCustom(ISD::STORE, intVT)) {
+      if (!isOperationLegalOrCustom(ISD::STORE, intVT) &&
+          MemVT.isVector()) {
         // Scalarize the store and let the individual components be handled.
         SDValue Result = scalarizeVectorStore(ST, DAG);
 
         return Result;
       }
       // Expand to a bitconvert of the value to the integer type of the
       // same size, then a (misaligned) int store.
       // FIXME: Does not handle truncating floating point stores!

test/CodeGen/AMDGPU/unaligned-load-store.ll

 ; SI: ds_write_b8
 
 ; SI: ScratchSize: 0{{$}}
 define amdgpu_kernel void @local_store_align1_v16i8(<16 x i8> addrspace(3)* %out) #0 {
   store <16 x i8> zeroinitializer, <16 x i8> addrspace(3)* %out, align 1
   ret void
 }
 
+; SI-LABEL: {{^}}private_load_align1_f64:
+; SI: buffer_load_ubyte
+; SI: buffer_load_ubyte
+; SI: buffer_load_ubyte
+; SI: buffer_load_ubyte
+; SI: buffer_load_ubyte
+; SI: buffer_load_ubyte
+; SI: buffer_load_ubyte
+; SI: buffer_load_ubyte
+define double @private_load_align1_f64(double addrspace(5)* %in) {
+  %x = load double, double addrspace(5)* %in, align 1
+  ret double %x
+}
+
+; SI-LABEL: {{^}}private_store_align1_f64:
+; SI: buffer_store_byte
+; SI: buffer_store_byte
+; SI: buffer_store_byte
+; SI: buffer_store_byte
+; SI: buffer_store_byte
+; SI: buffer_store_byte
+; SI: buffer_store_byte
+; SI: buffer_store_byte
+define void @private_store_align1_f64(double addrspace(5)* %out, double %x) #0 {
+  store double %x, double addrspace(5)* %out, align 1
+  ret void
+}
+
+; SI-LABEL: {{^}}private_load_align4_f64:
+; SI: buffer_load_dword
+; SI: buffer_load_dword
+define double @private_load_align4_f64(double addrspace(5)* %in) {
+  %x = load double, double addrspace(5)* %in, align 4
+  ret double %x
+}
+
+; SI-LABEL: {{^}}private_store_align4_f64:
+; SI: buffer_store_dword
+; SI: buffer_store_dword
+define void @private_store_align4_f64(double addrspace(5)* %out, double %x) #0 {
+  store double %x, double addrspace(5)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}private_load_align2_f64:
+; SI: buffer_load_ushort
+; SI: buffer_load_ushort
+; SI: buffer_load_ushort
+; SI: buffer_load_ushort
+define double @private_load_align2_f64(double addrspace(5)* %in) {
+  %x = load double, double addrspace(5)* %in, align 2
+  ret double %x
+}
+
+; SI-LABEL: {{^}}private_store_align2_f64:
+; SI: buffer_store_short
+; SI: buffer_store_short
+; SI: buffer_store_short
+; SI: buffer_store_short
+define void @private_store_align2_f64(double addrspace(5)* %out, double %x) #0 {
+  store double %x, double addrspace(5)* %out, align 2
+  ret void
+}
+
 attributes #0 = { nounwind }