[GISel] Introduce MachineIRBuilder::(build|materialize)ObjectPtrOffset (#150392)

These functions are for building G_PTR_ADDs when we know that the base
pointer and the result are both valid pointers into (or just after) the
same object. They are similar to SelectionDAG::getObjectPtrOffset.

This PR also changes call sites of the generic (build|materialize)PtrAdd
functions that implement pointer arithmetic to split large memory
accesses to the new functions. Since memory accesses have to fit into an
object in memory, pointer arithmetic to an offset into a large memory
access also yields an address in that object.

Currently, these (build|materialize)ObjectPtrOffset functions only add
"nuw" to the generated G_PTR_ADD, but I intend to introduce an
"inbounds" MIFlag in a later PR (analogous to a concurrent effort in
SDAG: #131862, related: #140017, #141725) that will also be set in the
(build|materialize)ObjectPtrOffset functions.

Most test changes just add "nuw" to G_PTR_ADDs. Exceptions are AMDGPU's
call-outgoing-stack-args.ll, flat-scratch.ll, and freeze.ll tests, where
offsets are now folded into scratch instructions, and cases where the
behavior of the check regeneration script changed, resulting, e.g., in
better checks for "nusw G_PTR_ADD" instructions, matched empty lines,
and the use of "CHECK-NEXT" in MIPS tests.

For SWDEV-516125.

Author: ritter-x2a

llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h

@@ -518,6 +518,21 @@ class LLVM_ABI MachineIRBuilder {
                                   const SrcOp &Op1,
                                   std::optional<unsigned> Flags = std::nullopt);
 
+  /// Build and insert an instruction with appropriate flags for addressing some
+  /// offset of an object, i.e.: \p Res = nuw G_PTR_ADD \p Op0, \p Op1
+  /// The value of \p Op0 must be a pointer into or just after an object, adding
+  /// the value of \p Op1 to it must yield to a pointer into or just after the
+  /// same object.
+  ///
+  /// \pre setBasicBlock or setMI must have been called.
+  /// \pre \p Res and \p Op0 must be generic virtual registers with pointer
+  ///      type.
+  /// \pre \p Op1 must be a generic virtual register with scalar type.
+  ///
+  /// \return a MachineInstrBuilder for the newly created instruction.
+  MachineInstrBuilder buildObjectPtrOffset(const DstOp &Res, const SrcOp &Op0,
+                                           const SrcOp &Op1);
+
   /// Materialize and insert \p Res = G_PTR_ADD \p Op0, (G_CONSTANT \p Value)
   ///
   /// G_PTR_ADD adds \p Value bytes to the pointer specified by \p Op0,
@@ -534,10 +549,29 @@ class LLVM_ABI MachineIRBuilder {
   ///       type as \p Op0 or \p Op0 itself.
   ///
   /// \return a MachineInstrBuilder for the newly created instruction.
-  std::optional<MachineInstrBuilder> materializePtrAdd(Register &Res,
-                                                       Register Op0,
-                                                       const LLT ValueTy,
-                                                       uint64_t Value);
+  std::optional<MachineInstrBuilder>
+  materializePtrAdd(Register &Res, Register Op0, const LLT ValueTy,
+                    uint64_t Value,
+                    std::optional<unsigned> Flags = std::nullopt);
+
+  /// Materialize and insert an instruction with appropriate flags for
+  /// addressing some offset of an object, i.e.:
+  ///   \p Res = nuw G_PTR_ADD \p Op0, (G_CONSTANT \p Value)
+  /// The value of \p Op0 must be a pointer into or just after an object, adding
+  /// \p Value to it must yield to a pointer into or just after the same object.
+  ///
+  /// \pre setBasicBlock or setMI must have been called.
+  /// \pre \p Op0 must be a generic virtual register with pointer type.
+  /// \pre \p ValueTy must be a scalar type.
+  /// \pre \p Res must be 0. This is to detect confusion between
+  ///      materializeObjectPtrOffset() and buildObjectPtrOffset().
+  /// \post \p Res will either be a new generic virtual register of the same
+  ///       type as \p Op0 or \p Op0 itself.
+  ///
+  /// \return a MachineInstrBuilder for the newly created instruction.
+  std::optional<MachineInstrBuilder>
+  materializeObjectPtrOffset(Register &Res, Register Op0, const LLT ValueTy,
+                             uint64_t Value);
 
   /// Build and insert \p Res = G_PTRMASK \p Op0, \p Op1
   MachineInstrBuilder buildPtrMask(const DstOp &Res, const SrcOp &Op0,

llvm/lib/CodeGen/GlobalISel/CallLowering.cpp

@@ -1009,7 +1009,8 @@ void CallLowering::insertSRetLoads(MachineIRBuilder &MIRBuilder, Type *RetTy,
 
   for (unsigned I = 0; I < NumValues; ++I) {
     Register Addr;
-    MIRBuilder.materializePtrAdd(Addr, DemoteReg, OffsetLLTy, Offsets[I]);
+    MIRBuilder.materializeObjectPtrOffset(Addr, DemoteReg, OffsetLLTy,
+                                          Offsets[I]);
     auto *MMO = MF.getMachineMemOperand(PtrInfo, MachineMemOperand::MOLoad,
                                         MRI.getType(VRegs[I]),
                                         commonAlignment(BaseAlign, Offsets[I]));
@@ -1039,7 +1040,8 @@ void CallLowering::insertSRetStores(MachineIRBuilder &MIRBuilder, Type *RetTy,
 
   for (unsigned I = 0; I < NumValues; ++I) {
     Register Addr;
-    MIRBuilder.materializePtrAdd(Addr, DemoteReg, OffsetLLTy, Offsets[I]);
+    MIRBuilder.materializeObjectPtrOffset(Addr, DemoteReg, OffsetLLTy,
+                                          Offsets[I]);
     auto *MMO = MF.getMachineMemOperand(PtrInfo, MachineMemOperand::MOStore,
                                         MRI.getType(VRegs[I]),
                                         commonAlignment(BaseAlign, Offsets[I]));

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

@@ -1409,7 +1409,7 @@ bool IRTranslator::translateLoad(const User &U, MachineIRBuilder &MIRBuilder) {
       Regs.size() == 1 ? LI.getMetadata(LLVMContext::MD_range) : nullptr;
   for (unsigned i = 0; i < Regs.size(); ++i) {
     Register Addr;
-    MIRBuilder.materializePtrAdd(Addr, Base, OffsetTy, Offsets[i] / 8);
+    MIRBuilder.materializeObjectPtrOffset(Addr, Base, OffsetTy, Offsets[i] / 8);
 
     MachinePointerInfo Ptr(LI.getPointerOperand(), Offsets[i] / 8);
     Align BaseAlign = getMemOpAlign(LI);
@@ -1448,7 +1448,7 @@ bool IRTranslator::translateStore(const User &U, MachineIRBuilder &MIRBuilder) {
 
   for (unsigned i = 0; i < Vals.size(); ++i) {
     Register Addr;
-    MIRBuilder.materializePtrAdd(Addr, Base, OffsetTy, Offsets[i] / 8);
+    MIRBuilder.materializeObjectPtrOffset(Addr, Base, OffsetTy, Offsets[i] / 8);
 
     MachinePointerInfo Ptr(SI.getPointerOperand(), Offsets[i] / 8);
     Align BaseAlign = getMemOpAlign(SI);

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

@@ -4170,7 +4170,7 @@ LegalizerHelper::LegalizeResult LegalizerHelper::lowerLoad(GAnyLoad &LoadMI) {
   auto OffsetCst = MIRBuilder.buildConstant(LLT::scalar(PtrTy.getSizeInBits()),
                                             LargeSplitSize / 8);
   Register PtrAddReg = MRI.createGenericVirtualRegister(PtrTy);
-  auto SmallPtr = MIRBuilder.buildPtrAdd(PtrAddReg, PtrReg, OffsetCst);
+  auto SmallPtr = MIRBuilder.buildObjectPtrOffset(PtrAddReg, PtrReg, OffsetCst);
   auto SmallLoad = MIRBuilder.buildLoadInstr(LoadMI.getOpcode(), AnyExtTy,
                                              SmallPtr, *SmallMMO);
 
@@ -4277,8 +4277,7 @@ LegalizerHelper::LegalizeResult LegalizerHelper::lowerStore(GStore &StoreMI) {
   LLT PtrTy = MRI.getType(PtrReg);
   auto OffsetCst = MIRBuilder.buildConstant(
     LLT::scalar(PtrTy.getSizeInBits()), LargeSplitSize / 8);
-  auto SmallPtr =
-    MIRBuilder.buildPtrAdd(PtrTy, PtrReg, OffsetCst);
+  auto SmallPtr = MIRBuilder.buildObjectPtrOffset(PtrTy, PtrReg, OffsetCst);
 
   MachineMemOperand *LargeMMO =
     MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8);
@@ -5349,7 +5348,8 @@ LegalizerHelper::reduceLoadStoreWidth(GLoadStore &LdStMI, unsigned TypeIdx,
       unsigned ByteOffset = Offset / 8;
       Register NewAddrReg;
 
-      MIRBuilder.materializePtrAdd(NewAddrReg, AddrReg, OffsetTy, ByteOffset);
+      MIRBuilder.materializeObjectPtrOffset(NewAddrReg, AddrReg, OffsetTy,
+                                            ByteOffset);
 
       MachineMemOperand *NewMMO =
           MF.getMachineMemOperand(&MMO, ByteOffset, PartTy);
@@ -9822,7 +9822,7 @@ LegalizerHelper::lowerMemset(MachineInstr &MI, Register Dst, Register Val,
     if (DstOff != 0) {
       auto Offset =
           MIB.buildConstant(LLT::scalar(PtrTy.getSizeInBits()), DstOff);
-      Ptr = MIB.buildPtrAdd(PtrTy, Dst, Offset).getReg(0);
+      Ptr = MIB.buildObjectPtrOffset(PtrTy, Dst, Offset).getReg(0);
     }
 
     MIB.buildStore(Value, Ptr, *StoreMMO);
@@ -9962,15 +9962,15 @@ LegalizerHelper::lowerMemcpy(MachineInstr &MI, Register Dst, Register Src,
       LLT SrcTy = MRI.getType(Src);
       Offset = MIB.buildConstant(LLT::scalar(SrcTy.getSizeInBits()), CurrOffset)
                    .getReg(0);
-      LoadPtr = MIB.buildPtrAdd(SrcTy, Src, Offset).getReg(0);
+      LoadPtr = MIB.buildObjectPtrOffset(SrcTy, Src, Offset).getReg(0);
     }
     auto LdVal = MIB.buildLoad(CopyTy, LoadPtr, *LoadMMO);
 
     // Create the store.
     Register StorePtr = Dst;
     if (CurrOffset != 0) {
       LLT DstTy = MRI.getType(Dst);
-      StorePtr = MIB.buildPtrAdd(DstTy, Dst, Offset).getReg(0);
+      StorePtr = MIB.buildObjectPtrOffset(DstTy, Dst, Offset).getReg(0);
     }
     MIB.buildStore(LdVal, StorePtr, *StoreMMO);
     CurrOffset += CopyTy.getSizeInBytes();
@@ -10060,7 +10060,7 @@ LegalizerHelper::lowerMemmove(MachineInstr &MI, Register Dst, Register Src,
       LLT SrcTy = MRI.getType(Src);
       auto Offset =
           MIB.buildConstant(LLT::scalar(SrcTy.getSizeInBits()), CurrOffset);
-      LoadPtr = MIB.buildPtrAdd(SrcTy, Src, Offset).getReg(0);
+      LoadPtr = MIB.buildObjectPtrOffset(SrcTy, Src, Offset).getReg(0);
     }
     LoadVals.push_back(MIB.buildLoad(CopyTy, LoadPtr, *LoadMMO).getReg(0));
     CurrOffset += CopyTy.getSizeInBytes();
@@ -10078,7 +10078,7 @@ LegalizerHelper::lowerMemmove(MachineInstr &MI, Register Dst, Register Src,
       LLT DstTy = MRI.getType(Dst);
       auto Offset =
           MIB.buildConstant(LLT::scalar(DstTy.getSizeInBits()), CurrOffset);
-      StorePtr = MIB.buildPtrAdd(DstTy, Dst, Offset).getReg(0);
+      StorePtr = MIB.buildObjectPtrOffset(DstTy, Dst, Offset).getReg(0);
     }
     MIB.buildStore(LoadVals[I], StorePtr, *StoreMMO);
     CurrOffset += CopyTy.getSizeInBytes();

llvm/lib/CodeGen/GlobalISel/MachineIRBuilder.cpp

@@ -208,11 +208,18 @@ MachineIRBuilder::buildPtrAdd(const DstOp &Res, const SrcOp &Op0,
   return buildInstr(TargetOpcode::G_PTR_ADD, {Res}, {Op0, Op1}, Flags);
 }
 
+MachineInstrBuilder MachineIRBuilder::buildObjectPtrOffset(const DstOp &Res,
+                                                           const SrcOp &Op0,
+                                                           const SrcOp &Op1) {
+  return buildPtrAdd(Res, Op0, Op1, MachineInstr::MIFlag::NoUWrap);
+}
+
 std::optional<MachineInstrBuilder>
 MachineIRBuilder::materializePtrAdd(Register &Res, Register Op0,
-                                    const LLT ValueTy, uint64_t Value) {
+                                    const LLT ValueTy, uint64_t Value,
+                                    std::optional<unsigned> Flags) {
   assert(Res == 0 && "Res is a result argument");
-  assert(ValueTy.isScalar()  && "invalid offset type");
+  assert(ValueTy.isScalar() && "invalid offset type");
 
   if (Value == 0) {
     Res = Op0;
@@ -221,7 +228,13 @@ MachineIRBuilder::materializePtrAdd(Register &Res, Register Op0,
 
   Res = getMRI()->createGenericVirtualRegister(getMRI()->getType(Op0));
   auto Cst = buildConstant(ValueTy, Value);
-  return buildPtrAdd(Res, Op0, Cst.getReg(0));
+  return buildPtrAdd(Res, Op0, Cst.getReg(0), Flags);
+}
+
+std::optional<MachineInstrBuilder> MachineIRBuilder::materializeObjectPtrOffset(
+    Register &Res, Register Op0, const LLT ValueTy, uint64_t Value) {
+  return materializePtrAdd(Res, Op0, ValueTy, Value,
+                           MachineInstr::MIFlag::NoUWrap);
 }
 
 MachineInstrBuilder MachineIRBuilder::buildMaskLowPtrBits(const DstOp &Res,