AMDGPU/GlobalISel: Implement LLT version of allowsMisalignedMemoryAccesses

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

@@ -262,7 +262,7 @@ static bool isLoadStoreSizeLegal(const GCNSubtarget &ST,
 
   unsigned RegSize = Ty.getSizeInBits();
   unsigned MemSize = Query.MMODescrs[0].SizeInBits;
-  unsigned Align = Query.MMODescrs[0].AlignInBits;
+  unsigned AlignBits = Query.MMODescrs[0].AlignInBits;
   unsigned AS = Query.Types[1].getAddressSpace();
 
   // All of these need to be custom lowered to cast the pointer operand.
@@ -305,9 +305,10 @@ static bool isLoadStoreSizeLegal(const GCNSubtarget &ST,
 
   assert(RegSize >= MemSize);
 
-  if (Align < MemSize) {
+  if (AlignBits < MemSize) {
     const SITargetLowering *TLI = ST.getTargetLowering();
-    if (!TLI->allowsMisalignedMemoryAccessesImpl(MemSize, AS, Align / 8))
+    if (!TLI->allowsMisalignedMemoryAccessesImpl(MemSize, AS,
+                                                 Align(AlignBits / 8)))
       return false;
   }
 
@@ -954,10 +955,10 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
 
     // Split vector extloads.
     unsigned MemSize = Query.MMODescrs[0].SizeInBits;
-    unsigned Align = Query.MMODescrs[0].AlignInBits;
+    unsigned AlignBits = Query.MMODescrs[0].AlignInBits;
 
     if (MemSize < DstTy.getSizeInBits())
-      MemSize = std::max(MemSize, Align);
+      MemSize = std::max(MemSize, AlignBits);
 
     if (DstTy.isVector() && DstTy.getSizeInBits() > MemSize)
       return true;
@@ -979,9 +980,10 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
         return true;
     }
 
-    if (Align < MemSize) {
+    if (AlignBits < MemSize) {
       const SITargetLowering *TLI = ST.getTargetLowering();
-      return !TLI->allowsMisalignedMemoryAccessesImpl(MemSize, AS, Align / 8);
+      return !TLI->allowsMisalignedMemoryAccessesImpl(MemSize, AS,
+                                                      Align(AlignBits / 8));
     }
 
     return false;

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -1384,7 +1384,7 @@ bool SITargetLowering::canMergeStoresTo(unsigned AS, EVT MemVT,
 }
 
 bool SITargetLowering::allowsMisalignedMemoryAccessesImpl(
-    unsigned Size, unsigned AddrSpace, unsigned Align,
+    unsigned Size, unsigned AddrSpace, Align Alignment,
     MachineMemOperand::Flags Flags, bool *IsFast) const {
   if (IsFast)
     *IsFast = false;
@@ -1394,7 +1394,7 @@ bool SITargetLowering::allowsMisalignedMemoryAccessesImpl(
     // ds_read/write_b64 require 8-byte alignment, but we can do a 4 byte
     // aligned, 8 byte access in a single operation using ds_read2/write2_b32
     // with adjacent offsets.
-    bool AlignedBy4 = (Align % 4 == 0);
+    bool AlignedBy4 = Alignment >= Align(4);
     if (IsFast)
       *IsFast = AlignedBy4;
 
@@ -1407,7 +1407,7 @@ bool SITargetLowering::allowsMisalignedMemoryAccessesImpl(
   if (!Subtarget->hasUnalignedScratchAccess() &&
       (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS ||
        AddrSpace == AMDGPUAS::FLAT_ADDRESS)) {
-    bool AlignedBy4 = Align >= 4;
+    bool AlignedBy4 = Alignment >= Align(4);
     if (IsFast)
       *IsFast = AlignedBy4;
 
@@ -1422,7 +1422,7 @@ bool SITargetLowering::allowsMisalignedMemoryAccessesImpl(
       // 2-byte alignment is worse than 1 unless doing a 2-byte accesss.
       *IsFast = (AddrSpace == AMDGPUAS::CONSTANT_ADDRESS ||
                  AddrSpace == AMDGPUAS::CONSTANT_ADDRESS_32BIT) ?
-        Align >= 4 : Align != 2;
+        Alignment >= Align(4) : Alignment != Align(2);
     }
 
     return true;
@@ -1438,12 +1438,12 @@ bool SITargetLowering::allowsMisalignedMemoryAccessesImpl(
   if (IsFast)
     *IsFast = true;
 
-  return Size >= 32 && Align >= 4;
+  return Size >= 32 && Alignment >= Align(4);
 }
 
 bool SITargetLowering::allowsMisalignedMemoryAccesses(
-    EVT VT, unsigned AddrSpace, unsigned Align, MachineMemOperand::Flags Flags,
+    EVT VT, unsigned AddrSpace, unsigned Alignment,
-    bool *IsFast) const {
+    MachineMemOperand::Flags Flags, bool *IsFast) const {
   if (IsFast)
     *IsFast = false;
 
@@ -1457,7 +1457,7 @@ bool SITargetLowering::allowsMisalignedMemoryAccesses(
   }
 
   return allowsMisalignedMemoryAccessesImpl(VT.getSizeInBits(), AddrSpace,
-                                            Align, Flags, IsFast);
+                                            Align(Alignment), Flags, IsFast);
 }
 
 EVT SITargetLowering::getOptimalMemOpType(

llvm/lib/Target/AMDGPU/SIISelLowering.h

@@ -260,12 +260,22 @@ public:
                         const SelectionDAG &DAG) const override;
 
   bool allowsMisalignedMemoryAccessesImpl(
-      unsigned Size, unsigned AS, unsigned Align,
+      unsigned Size, unsigned AddrSpace, Align Alignment,
       MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
       bool *IsFast = nullptr) const;
 
   bool allowsMisalignedMemoryAccesses(
-      EVT VT, unsigned AS, unsigned Align,
+      LLT Ty, unsigned AddrSpace, Align Alignment,
+      MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
+      bool *IsFast = nullptr) const override {
+    if (IsFast)
+      *IsFast = false;
+    return allowsMisalignedMemoryAccessesImpl(Ty.getSizeInBits(), AddrSpace,
+                                              Alignment, Flags, IsFast);
+  }
+
+  bool allowsMisalignedMemoryAccesses(
+      EVT VT, unsigned AS, unsigned Alignment,
       MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
       bool *IsFast = nullptr) const override;