[LoongArch] Optimize *W Instructions at MI level

Referring to RISC-V, adding an MI level pass to optimize *W instructions
for LoongArch.

First it removes unneeded sext(addi.w rd, rs, 0) instructions. Either
because the sign extended bits aren't consumed or because the input was
already sign extended by an earlier instruction.

Then:
1. Unless explicit disabled or the target prefers instructions with W suffix,
   it removes the -w suffix from opw instructions whenever all users are
   dependent only on the lower word of the result of the instruction.
   The cases handled are:
   * addi.w because it helps reduce test differences between LA32 and LA64
     w/o being a pessimization.

2. Or if explicit enabled or the target prefers instructions with W suffix,
   it adds the W suffix to the instruction whenever all users are dependent
   only on the lower word of the result of the instruction.
   The cases handled are:
   * add.d/addi.d/sub.d/mul.d.
   * slli.d with imm < 32.
   * ld.d/ld.wu.

Author: heiher

llvm/lib/Target/LoongArch/CMakeLists.txt

@@ -23,6 +23,7 @@ add_llvm_target(LoongArchCodeGen
   LoongArchISelDAGToDAG.cpp
   LoongArchISelLowering.cpp
   LoongArchMCInstLower.cpp
+  LoongArchOptWInstrs.cpp
   LoongArchRegisterInfo.cpp
   LoongArchSubtarget.cpp
   LoongArchTargetMachine.cpp

llvm/lib/Target/LoongArch/LoongArch.h

@@ -35,10 +35,12 @@ bool lowerLoongArchMachineOperandToMCOperand(const MachineOperand &MO,
 
 FunctionPass *createLoongArchExpandAtomicPseudoPass();
 FunctionPass *createLoongArchISelDag(LoongArchTargetMachine &TM);
+FunctionPass *createLoongArchOptWInstrsPass();
 FunctionPass *createLoongArchPreRAExpandPseudoPass();
 FunctionPass *createLoongArchExpandPseudoPass();
 void initializeLoongArchDAGToDAGISelPass(PassRegistry &);
 void initializeLoongArchExpandAtomicPseudoPass(PassRegistry &);
+void initializeLoongArchOptWInstrsPass(PassRegistry &);
 void initializeLoongArchPreRAExpandPseudoPass(PassRegistry &);
 void initializeLoongArchExpandPseudoPass(PassRegistry &);
 } // end namespace llvm

llvm/lib/Target/LoongArch/LoongArch.td

@@ -117,6 +117,9 @@ def FeatureFrecipe
                        "Support frecipe.{s/d} and frsqrte.{s/d} instructions.">;
 def HasFrecipe : Predicate<"Subtarget->hasFrecipe()">;
 
+def TunePreferWInst
+    : SubtargetFeature<"prefer-w-inst", "PreferWInst", "true",
+                       "Prefer instructions with W suffix">;
 
 //===----------------------------------------------------------------------===//
 // Registers, instruction descriptions ...

llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp

@@ -3746,6 +3746,7 @@ static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDValue Val,
 
 static SDValue unpackFromRegLoc(SelectionDAG &DAG, SDValue Chain,
                                 const CCValAssign &VA, const SDLoc &DL,
+                                const ISD::InputArg &In,
                                 const LoongArchTargetLowering &TLI) {
   MachineFunction &MF = DAG.getMachineFunction();
   MachineRegisterInfo &RegInfo = MF.getRegInfo();
@@ -3756,6 +3757,21 @@ static SDValue unpackFromRegLoc(SelectionDAG &DAG, SDValue Chain,
   RegInfo.addLiveIn(VA.getLocReg(), VReg);
   Val = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
 
+  // If input is sign extended from 32 bits, note it for the OptW pass.
+  if (In.isOrigArg()) {
+    Argument *OrigArg = MF.getFunction().getArg(In.getOrigArgIndex());
+    if (OrigArg->getType()->isIntegerTy()) {
+      unsigned BitWidth = OrigArg->getType()->getIntegerBitWidth();
+      // An input zero extended from i31 can also be considered sign extended.
+      if ((BitWidth <= 32 && In.Flags.isSExt()) ||
+          (BitWidth < 32 && In.Flags.isZExt())) {
+        LoongArchMachineFunctionInfo *LAFI =
+            MF.getInfo<LoongArchMachineFunctionInfo>();
+        LAFI->addSExt32Register(VReg);
+      }
+    }
+  }
+
   return convertLocVTToValVT(DAG, Val, VA, DL);
 }
 
@@ -3887,7 +3903,7 @@ SDValue LoongArchTargetLowering::LowerFormalArguments(
     CCValAssign &VA = ArgLocs[i];
     SDValue ArgValue;
     if (VA.isRegLoc())
-      ArgValue = unpackFromRegLoc(DAG, Chain, VA, DL, *this);
+      ArgValue = unpackFromRegLoc(DAG, Chain, VA, DL, Ins[i], *this);
     else
       ArgValue = unpackFromMemLoc(DAG, Chain, VA, DL);
     if (VA.getLocInfo() == CCValAssign::Indirect) {

llvm/lib/Target/LoongArch/LoongArchInstrInfo.cpp

@@ -534,3 +534,9 @@ LoongArchInstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
       {MO_GD_PC_HI, "loongarch-gd-pc-hi"}};
   return ArrayRef(TargetFlags);
 }
+
+// Returns true if this is the sext.w pattern, addi.w rd, rs, 0.
+bool LoongArch::isSEXT_W(const MachineInstr &MI) {
+  return MI.getOpcode() == LoongArch::ADDI_W && MI.getOperand(1).isReg() &&
+         MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0;
+}

llvm/lib/Target/LoongArch/LoongArchInstrInfo.h

@@ -90,6 +90,9 @@ class LoongArchInstrInfo : public LoongArchGenInstrInfo {
 
 namespace LoongArch {
 
+// Returns true if this is the sext.w pattern, addi.w rd, rs, 0.
+bool isSEXT_W(const MachineInstr &MI);
+
 // Mask assignments for floating-point.
 static constexpr unsigned FClassMaskSignalingNaN = 0x001;
 static constexpr unsigned FClassMaskQuietNaN = 0x002;

llvm/lib/Target/LoongArch/LoongArchMachineFunctionInfo.h

@@ -36,6 +36,9 @@ class LoongArchMachineFunctionInfo : public MachineFunctionInfo {
   /// insertIndirectBranch.
   int BranchRelaxationSpillFrameIndex = -1;
 
+  /// Registers that have been sign extended from i32.
+  SmallVector<Register, 8> SExt32Registers;
+
 public:
   LoongArchMachineFunctionInfo(const Function &F,
                                const TargetSubtargetInfo *STI) {}
@@ -62,6 +65,12 @@ class LoongArchMachineFunctionInfo : public MachineFunctionInfo {
   void setBranchRelaxationSpillFrameIndex(int Index) {
     BranchRelaxationSpillFrameIndex = Index;
   }
+
+  void addSExt32Register(Register Reg) { SExt32Registers.push_back(Reg); }
+
+  bool isSExt32Register(Register Reg) const {
+    return is_contained(SExt32Registers, Reg);
+  }
 };
 
 } // end namespace llvm