[LoongArch] Add a pass to rewrite rd to r0 for non-computational instrs whose return values are unused (llvm#94590)

This patch adds a peephole pass `LoongArchDeadRegisterDefinitions`. It
rewrites `rd` to `r0` when `rd` is marked as dead. It may improve the
register allocation and reduce pipeline hazards on CPUs without register
renaming and OOO.

Author: heiher

llvm/lib/Target/LoongArch/CMakeLists.txt

@@ -16,6 +16,7 @@ add_public_tablegen_target(LoongArchCommonTableGen)
 
 add_llvm_target(LoongArchCodeGen
   LoongArchAsmPrinter.cpp
+  LoongArchDeadRegisterDefinitions.cpp
   LoongArchExpandAtomicPseudoInsts.cpp
   LoongArchExpandPseudoInsts.cpp
   LoongArchFrameLowering.cpp

llvm/lib/Target/LoongArch/LoongArch.h

@@ -33,12 +33,14 @@ bool lowerLoongArchMachineOperandToMCOperand(const MachineOperand &MO,
                                              MCOperand &MCOp,
                                              const AsmPrinter &AP);
 
+FunctionPass *createLoongArchDeadRegisterDefinitionsPass();
 FunctionPass *createLoongArchExpandAtomicPseudoPass();
 FunctionPass *createLoongArchISelDag(LoongArchTargetMachine &TM);
 FunctionPass *createLoongArchOptWInstrsPass();
 FunctionPass *createLoongArchPreRAExpandPseudoPass();
 FunctionPass *createLoongArchExpandPseudoPass();
 void initializeLoongArchDAGToDAGISelLegacyPass(PassRegistry &);
+void initializeLoongArchDeadRegisterDefinitionsPass(PassRegistry &);
 void initializeLoongArchExpandAtomicPseudoPass(PassRegistry &);
 void initializeLoongArchOptWInstrsPass(PassRegistry &);
 void initializeLoongArchPreRAExpandPseudoPass(PassRegistry &);

llvm/lib/Target/LoongArch/LoongArchDeadRegisterDefinitions.cpp

@@ -0,0 +1,108 @@
+//=== LoongArchDeadRegisterDefinitions.cpp - Replace dead defs w/ zero reg ===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===---------------------------------------------------------------------===//
+//
+// This pass rewrites Rd to r0 for instrs whose return values are unused.
+//
+//===---------------------------------------------------------------------===//
+
+#include "LoongArch.h"
+#include "LoongArchInstrInfo.h"
+#include "LoongArchSubtarget.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/LiveDebugVariables.h"
+#include "llvm/CodeGen/LiveIntervals.h"
+#include "llvm/CodeGen/LiveStacks.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+
+using namespace llvm;
+#define DEBUG_TYPE "loongarch-dead-defs"
+#define LoongArch_DEAD_REG_DEF_NAME "LoongArch Dead register definitions"
+
+STATISTIC(NumDeadDefsReplaced, "Number of dead definitions replaced");
+
+namespace {
+class LoongArchDeadRegisterDefinitions : public MachineFunctionPass {
+public:
+  static char ID;
+
+  LoongArchDeadRegisterDefinitions() : MachineFunctionPass(ID) {}
+  bool runOnMachineFunction(MachineFunction &MF) override;
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.setPreservesCFG();
+    AU.addRequired<LiveIntervals>();
+    AU.addPreserved<LiveIntervals>();
+    AU.addRequired<LiveIntervals>();
+    AU.addPreserved<SlotIndexes>();
+    AU.addPreserved<LiveDebugVariables>();
+    AU.addPreserved<LiveStacks>();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+  StringRef getPassName() const override { return LoongArch_DEAD_REG_DEF_NAME; }
+};
+} // end anonymous namespace
+
+char LoongArchDeadRegisterDefinitions::ID = 0;
+INITIALIZE_PASS(LoongArchDeadRegisterDefinitions, DEBUG_TYPE,
+                LoongArch_DEAD_REG_DEF_NAME, false, false)
+
+FunctionPass *llvm::createLoongArchDeadRegisterDefinitionsPass() {
+  return new LoongArchDeadRegisterDefinitions();
+}
+
+bool LoongArchDeadRegisterDefinitions::runOnMachineFunction(
+    MachineFunction &MF) {
+  if (skipFunction(MF.getFunction()))
+    return false;
+
+  const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
+  const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
+  LiveIntervals &LIS = getAnalysis<LiveIntervals>();
+  LLVM_DEBUG(dbgs() << "***** LoongArchDeadRegisterDefinitions *****\n");
+
+  bool MadeChange = false;
+  for (MachineBasicBlock &MBB : MF) {
+    for (MachineInstr &MI : MBB) {
+      // We only handle non-computational instructions.
+      const MCInstrDesc &Desc = MI.getDesc();
+      if (!Desc.mayLoad() && !Desc.mayStore() &&
+          !Desc.hasUnmodeledSideEffects())
+        continue;
+      for (int I = 0, E = Desc.getNumDefs(); I != E; ++I) {
+        MachineOperand &MO = MI.getOperand(I);
+        if (!MO.isReg() || !MO.isDef() || MO.isEarlyClobber())
+          continue;
+        // Be careful not to change the register if it's a tied operand.
+        if (MI.isRegTiedToUseOperand(I)) {
+          LLVM_DEBUG(dbgs() << "    Ignoring, def is tied operand.\n");
+          continue;
+        }
+        Register Reg = MO.getReg();
+        if (!Reg.isVirtual() || !MO.isDead())
+          continue;
+        LLVM_DEBUG(dbgs() << "    Dead def operand #" << I << " in:\n      ";
+                   MI.print(dbgs()));
+        const TargetRegisterClass *RC = TII->getRegClass(Desc, I, TRI, MF);
+        if (!(RC && RC->contains(LoongArch::R0))) {
+          LLVM_DEBUG(dbgs() << "    Ignoring, register is not a GPR.\n");
+          continue;
+        }
+        assert(LIS.hasInterval(Reg));
+        LIS.removeInterval(Reg);
+        MO.setReg(LoongArch::R0);
+        LLVM_DEBUG(dbgs() << "    Replacing with zero register. New:\n      ";
+                   MI.print(dbgs()));
+        ++NumDeadDefsReplaced;
+        MadeChange = true;
+      }
+    }
+  }
+
+  return MadeChange;
+}

llvm/lib/Target/LoongArch/LoongArchTargetMachine.cpp

@@ -34,11 +34,19 @@ extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeLoongArchTarget() {
   RegisterTargetMachine<LoongArchTargetMachine> X(getTheLoongArch32Target());
   RegisterTargetMachine<LoongArchTargetMachine> Y(getTheLoongArch64Target());
   auto *PR = PassRegistry::getPassRegistry();
+  initializeLoongArchDeadRegisterDefinitionsPass(*PR);
   initializeLoongArchOptWInstrsPass(*PR);
   initializeLoongArchPreRAExpandPseudoPass(*PR);
   initializeLoongArchDAGToDAGISelLegacyPass(*PR);
 }
 
+static cl::opt<bool> EnableLoongArchDeadRegisterElimination(
+    "loongarch-enable-dead-defs", cl::Hidden,
+    cl::desc("Enable the pass that removes dead"
+             " definitons and replaces stores to"
+             " them with stores to r0"),
+    cl::init(true));
+
 static cl::opt<bool>
     EnableLoopDataPrefetch("loongarch-enable-loop-data-prefetch", cl::Hidden,
                            cl::desc("Enable the loop data prefetch pass"),
@@ -148,6 +156,8 @@ class LoongArchPassConfig : public TargetPassConfig {
   void addPreEmitPass2() override;
   void addMachineSSAOptimization() override;
   void addPreRegAlloc() override;
+  bool addRegAssignAndRewriteFast() override;
+  bool addRegAssignAndRewriteOptimized() override;
 };
 } // end namespace
 
@@ -200,3 +210,17 @@ void LoongArchPassConfig::addMachineSSAOptimization() {
 void LoongArchPassConfig::addPreRegAlloc() {
   addPass(createLoongArchPreRAExpandPseudoPass());
 }
+
+bool LoongArchPassConfig::addRegAssignAndRewriteFast() {
+  if (TM->getOptLevel() != CodeGenOptLevel::None &&
+      EnableLoongArchDeadRegisterElimination)
+    addPass(createLoongArchDeadRegisterDefinitionsPass());
+  return TargetPassConfig::addRegAssignAndRewriteFast();
+}
+
+bool LoongArchPassConfig::addRegAssignAndRewriteOptimized() {
+  if (TM->getOptLevel() != CodeGenOptLevel::None &&
+      EnableLoongArchDeadRegisterElimination)
+    addPass(createLoongArchDeadRegisterDefinitionsPass());
+  return TargetPassConfig::addRegAssignAndRewriteOptimized();
+}

llvm/test/CodeGen/LoongArch/global-address.ll

@@ -14,40 +14,40 @@ define void @foo() nounwind {
 ; LA32NOPIC:       # %bb.0:
 ; LA32NOPIC-NEXT:    pcalau12i $a0, %got_pc_hi20(G)
 ; LA32NOPIC-NEXT:    ld.w $a0, $a0, %got_pc_lo12(G)
-; LA32NOPIC-NEXT:    ld.w $a0, $a0, 0
+; LA32NOPIC-NEXT:    ld.w $zero, $a0, 0
 ; LA32NOPIC-NEXT:    pcalau12i $a0, %pc_hi20(g)
 ; LA32NOPIC-NEXT:    addi.w $a0, $a0, %pc_lo12(g)
-; LA32NOPIC-NEXT:    ld.w $a0, $a0, 0
+; LA32NOPIC-NEXT:    ld.w $zero, $a0, 0
 ; LA32NOPIC-NEXT:    ret
 ;
 ; LA32PIC-LABEL: foo:
 ; LA32PIC:       # %bb.0:
 ; LA32PIC-NEXT:    pcalau12i $a0, %got_pc_hi20(G)
 ; LA32PIC-NEXT:    ld.w $a0, $a0, %got_pc_lo12(G)
-; LA32PIC-NEXT:    ld.w $a0, $a0, 0
+; LA32PIC-NEXT:    ld.w $zero, $a0, 0
 ; LA32PIC-NEXT:    pcalau12i $a0, %pc_hi20(.Lg$local)
 ; LA32PIC-NEXT:    addi.w $a0, $a0, %pc_lo12(.Lg$local)
-; LA32PIC-NEXT:    ld.w $a0, $a0, 0
+; LA32PIC-NEXT:    ld.w $zero, $a0, 0
 ; LA32PIC-NEXT:    ret
 ;
 ; LA64NOPIC-LABEL: foo:
 ; LA64NOPIC:       # %bb.0:
 ; LA64NOPIC-NEXT:    pcalau12i $a0, %got_pc_hi20(G)
 ; LA64NOPIC-NEXT:    ld.d $a0, $a0, %got_pc_lo12(G)
-; LA64NOPIC-NEXT:    ld.w $a0, $a0, 0
+; LA64NOPIC-NEXT:    ld.w $zero, $a0, 0
 ; LA64NOPIC-NEXT:    pcalau12i $a0, %pc_hi20(g)
 ; LA64NOPIC-NEXT:    addi.d $a0, $a0, %pc_lo12(g)
-; LA64NOPIC-NEXT:    ld.w $a0, $a0, 0
+; LA64NOPIC-NEXT:    ld.w $zero, $a0, 0
 ; LA64NOPIC-NEXT:    ret
 ;
 ; LA64PIC-LABEL: foo:
 ; LA64PIC:       # %bb.0:
 ; LA64PIC-NEXT:    pcalau12i $a0, %got_pc_hi20(G)
 ; LA64PIC-NEXT:    ld.d $a0, $a0, %got_pc_lo12(G)
-; LA64PIC-NEXT:    ld.w $a0, $a0, 0
+; LA64PIC-NEXT:    ld.w $zero, $a0, 0
 ; LA64PIC-NEXT:    pcalau12i $a0, %pc_hi20(.Lg$local)
 ; LA64PIC-NEXT:    addi.d $a0, $a0, %pc_lo12(.Lg$local)
-; LA64PIC-NEXT:    ld.w $a0, $a0, 0
+; LA64PIC-NEXT:    ld.w $zero, $a0, 0
 ; LA64PIC-NEXT:    ret
 ;
 ; LA64LARGENOPIC-LABEL: foo:
@@ -57,13 +57,13 @@ define void @foo() nounwind {
 ; LA64LARGENOPIC-NEXT:    lu32i.d $t8, %got64_pc_lo20(G)
 ; LA64LARGENOPIC-NEXT:    lu52i.d $t8, $t8, %got64_pc_hi12(G)
 ; LA64LARGENOPIC-NEXT:    ldx.d $a0, $t8, $a0
-; LA64LARGENOPIC-NEXT:    ld.w $a0, $a0, 0
+; LA64LARGENOPIC-NEXT:    ld.w $zero, $a0, 0
 ; LA64LARGENOPIC-NEXT:    pcalau12i $a0, %pc_hi20(g)
 ; LA64LARGENOPIC-NEXT:    addi.d $t8, $zero, %pc_lo12(g)
 ; LA64LARGENOPIC-NEXT:    lu32i.d $t8, %pc64_lo20(g)
 ; LA64LARGENOPIC-NEXT:    lu52i.d $t8, $t8, %pc64_hi12(g)
 ; LA64LARGENOPIC-NEXT:    add.d $a0, $t8, $a0
-; LA64LARGENOPIC-NEXT:    ld.w $a0, $a0, 0
+; LA64LARGENOPIC-NEXT:    ld.w $zero, $a0, 0
 ; LA64LARGENOPIC-NEXT:    ret
 ;
 ; LA64LARGEPIC-LABEL: foo:
@@ -73,13 +73,13 @@ define void @foo() nounwind {
 ; LA64LARGEPIC-NEXT:    lu32i.d $t8, %got64_pc_lo20(G)
 ; LA64LARGEPIC-NEXT:    lu52i.d $t8, $t8, %got64_pc_hi12(G)
 ; LA64LARGEPIC-NEXT:    ldx.d $a0, $t8, $a0
-; LA64LARGEPIC-NEXT:    ld.w $a0, $a0, 0
+; LA64LARGEPIC-NEXT:    ld.w $zero, $a0, 0
 ; LA64LARGEPIC-NEXT:    pcalau12i $a0, %pc_hi20(.Lg$local)
 ; LA64LARGEPIC-NEXT:    addi.d $t8, $zero, %pc_lo12(.Lg$local)
 ; LA64LARGEPIC-NEXT:    lu32i.d $t8, %pc64_lo20(.Lg$local)
 ; LA64LARGEPIC-NEXT:    lu52i.d $t8, $t8, %pc64_hi12(.Lg$local)
 ; LA64LARGEPIC-NEXT:    add.d $a0, $t8, $a0
-; LA64LARGEPIC-NEXT:    ld.w $a0, $a0, 0
+; LA64LARGEPIC-NEXT:    ld.w $zero, $a0, 0
 ; LA64LARGEPIC-NEXT:    ret
   %V = load volatile i32, ptr @G
   %v = load volatile i32, ptr @g

llvm/test/CodeGen/LoongArch/intrinsic-la64.ll

@@ -178,7 +178,7 @@ entry:
 define void @csrrd_d_noret() {
 ; CHECK-LABEL: csrrd_d_noret:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    csrrd $a0, 1
+; CHECK-NEXT:    csrrd $zero, 1
 ; CHECK-NEXT:    ret
 entry:
   %0 = tail call i64 @llvm.loongarch.csrrd.d(i32 1)
@@ -240,7 +240,7 @@ entry:
 define void @iocsrrd_d_noret(i32 %a) {
 ; CHECK-LABEL: iocsrrd_d_noret:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    iocsrrd.d $a0, $a0
+; CHECK-NEXT:    iocsrrd.d $zero, $a0
 ; CHECK-NEXT:    ret
 entry:
   %0 = tail call i64 @llvm.loongarch.iocsrrd.d(i32 %a)
@@ -290,7 +290,7 @@ entry:
 define void @lddir_d_noret(i64 %a) {
 ; CHECK-LABEL: lddir_d_noret:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    lddir $a0, $a0, 1
+; CHECK-NEXT:    lddir $zero, $a0, 1
 ; CHECK-NEXT:    ret
 entry:
   %0 = tail call i64 @llvm.loongarch.lddir.d(i64 %a, i64 1)

llvm/test/CodeGen/LoongArch/intrinsic.ll

@@ -73,7 +73,7 @@ entry:
 define void @movfcsr2gr_noret() nounwind {
 ; CHECK-LABEL: movfcsr2gr_noret:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    movfcsr2gr $a0, $fcsr1
+; CHECK-NEXT:    movfcsr2gr $zero, $fcsr1
 ; CHECK-NEXT:    ret
 entry:
   %res = call i32 @llvm.loongarch.movfcsr2gr(i32 1)
@@ -103,7 +103,7 @@ entry:
 define void @csrrd_w_noret() {
 ; CHECK-LABEL: csrrd_w_noret:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    csrrd $a0, 1
+; CHECK-NEXT:    csrrd $zero, 1
 ; CHECK-NEXT:    ret
 entry:
   %0 = tail call i32 @llvm.loongarch.csrrd.w(i32 1)
@@ -185,7 +185,7 @@ entry:
 define void @iocsrrd_b_noret(i32 %a) {
 ; CHECK-LABEL: iocsrrd_b_noret:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    iocsrrd.b $a0, $a0
+; CHECK-NEXT:    iocsrrd.b $zero, $a0
 ; CHECK-NEXT:    ret
 entry:
   %0 = tail call i32 @llvm.loongarch.iocsrrd.b(i32 %a)
@@ -195,7 +195,7 @@ entry:
 define void @iocsrrd_h_noret(i32 %a) {
 ; CHECK-LABEL: iocsrrd_h_noret:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    iocsrrd.h $a0, $a0
+; CHECK-NEXT:    iocsrrd.h $zero, $a0
 ; CHECK-NEXT:    ret
 entry:
   %0 = tail call i32 @llvm.loongarch.iocsrrd.h(i32 %a)
@@ -205,7 +205,7 @@ entry:
 define void @iocsrrd_w_noret(i32 %a) {
 ; CHECK-LABEL: iocsrrd_w_noret:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    iocsrrd.w $a0, $a0
+; CHECK-NEXT:    iocsrrd.w $zero, $a0
 ; CHECK-NEXT:    ret
 entry:
   %0 = tail call i32 @llvm.loongarch.iocsrrd.w(i32 %a)