[RISCV] Add a pass to merge moving parameter registers instructions for Zcmp

[VincentWu]

Summary

This patch adds a pass to generate cm.mvsa01 & cm.mva01s.

RISCVMoveOptimizer.cpp which combines two mv inst into one cm.mva01s or cm.mva01s.

[kito-cheng]

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

1296–1305 ↗

(On Diff #521517)

This change seems right, but that should be a separated patch :)

[craig.topper]

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

1296–1305 ↗

(On Diff #521517)

Do we ever create such instructions?

[VincentWu]

address comments

[craig.topper]

llvm/lib/Target/RISCV/RISCVMoveOptimizer.cpp

62 ↗

(On Diff #521878)

const DestSourcePair &RegPair

67 ↗

(On Diff #521878)

Join the two ifs with &&

68 ↗

(On Diff #521878)

SR07RegClass.contains(Source) should work I think

74 ↗

(On Diff #521878)

const DestSourcePair &RegPair

80 ↗

(On Diff #521878)

SR07RegClass.contains(Dest)

80 ↗

(On Diff #521878)

Join the two ifs with &&

134 ↗

(On Diff #521878)

Can you pass the DestSourcePair from the caller?

173 ↗

(On Diff #521878)

Why are the available checks different between the two cases? Aren't we always moving the paired copy earlier in the basic block to the location of the first copy. So shouldn't the register checks be the same?

192 ↗

(On Diff #521878)

There is no reason to shorten Move to Mov in this function name. And function names should typically start with a verb. optimizeBlock is probably a better name

229 ↗

(On Diff #521878)

Use Fn.getSubtarget<RISCVSubtarget>()

230 ↗

(On Diff #521878)

Drop curly braces

llvm/lib/Target/RISCV/RISCVTargetMachine.cpp

352

Should we disable for -O0?

[craig.topper]

llvm/test/CodeGen/RISCV/cm_mvas_mvsa.ll

14

Shouldn't we have more tests to cover cases where the copies aren't next to each other and cases where registers are modified/used in between?

[VincentWu]

address comments

[craig.topper]

llvm/lib/Target/RISCV/RISCVMoveOptimizer.cpp

31 ↗

(On Diff #526402)

Subtarget doesn't need to be a member of the class. It's only used in runOnMachineFunction

110 ↗

(On Diff #526402)

Put ARegInFirstPair == RISCV::X10 in a variable instead of repeating it 4 times.

162 ↗

(On Diff #526402)

Why don't we need to check UsedRegUnits.available(DestReg) like we do for CM_MVSA01?

168 ↗

(On Diff #526402)

Why don't we need to we check ModifiedRegUnits.available(DestReg) like we do for RISCV::CM_MVA01S?

187 ↗

(On Diff #526402)

This function name is just a shortened form of the pass name. It should be a verb phrase as documented here https://llvm.org/docs/CodingStandards.html#name-types-functions-variables-and-enumerators-properly

[VincentWu]

address comments

[craig.topper]

llvm/lib/Target/RISCV/RISCVMoveOptimizer.cpp

181 ↗

(On Diff #528303)

Function names should start with a lower case letter

203 ↗

(On Diff #528303)

could -> can

231 ↗

(On Diff #528303)

Drop the curly braces

[VincentWu]

address comment

[craig.topper]

llvm/lib/Target/RISCV/RISCVMoveOptimizer.cpp

9 ↗

(On Diff #530424)

This description should mention Zcmp

223 ↗

(On Diff #530424)

This static_cast isn't necessary. RISCVSubtarget has an implementation of getInstrInfo that returns RISCVInstrInfo *

[craig.topper]

llvm/lib/Target/RISCV/RISCVMoveOptimizer.cpp

19 ↗

(On Diff #530424)

We're inconsistently using "move optimizer" and "move merging". Can we pick one?

[VincentWu]

address comments

[craig.topper]

LGTM

[VincentWu]

rebase

llvm/lib/Target/RISCV/CMakeLists.txt

add_llvm_target(RISCVCodeGen
   RISCVInstrInfo.cpp
   RISCVISelDAGToDAG.cpp
   RISCVISelLowering.cpp
   RISCVMachineFunctionInfo.cpp
   RISCVMacroFusion.cpp
   RISCVMergeBaseOffset.cpp
   RISCVOptWInstrs.cpp
   RISCVRedundantCopyElimination.cpp
+  RISCVMoveMerger.cpp
   RISCVRegisterInfo.cpp
   RISCVRVVInitUndef.cpp
   RISCVSubtarget.cpp
   RISCVTargetMachine.cpp
   RISCVTargetObjectFile.cpp
   RISCVTargetTransformInfo.cpp
   GISel/RISCVCallLowering.cpp
   GISel/RISCVInstructionSelector.cpp

llvm/lib/Target/RISCV/RISCV.h

 
 FunctionPass *createRISCVRedundantCopyEliminationPass();
 void initializeRISCVRedundantCopyEliminationPass(PassRegistry &);
 
 FunctionPass *createRISCVInitUndefPass();
 void initializeRISCVInitUndefPass(PassRegistry &);
 extern char &RISCVInitUndefID;
 
+FunctionPass *createRISCVMoveMergePass();
+void initializeRISCVMoveMergePass(PassRegistry &);
+
 InstructionSelector *createRISCVInstructionSelector(const RISCVTargetMachine &,
                                                     RISCVSubtarget &,
                                                     RISCVRegisterBankInfo &);
 void initializeRISCVDAGToDAGISelPass(PassRegistry &);
 } // namespace llvm
 
 #endif

llvm/lib/Target/RISCV/RISCVMoveMerger.cpp

+//===---------- RISCVMoveMerge.cpp - RISCV move merge pass -------------===//
+//
+// 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 file contains a pass that performs move related peephole optimizations
+// as Zcmp has specified. This pass should be run after register allocation.
+//
+//===----------------------------------------------------------------------===//
+
+#include "RISCVInstrInfo.h"
+#include "RISCVMachineFunctionInfo.h"
+
+using namespace llvm;
+
+#define RISCV_MOVE_MERGE_NAME "RISC-V Zcmp move merging pass"
+
+namespace {
+struct RISCVMoveMerge : public MachineFunctionPass {
+  static char ID;
+
+  RISCVMoveMerge() : MachineFunctionPass(ID) {
+    initializeRISCVMoveMergePass(*PassRegistry::getPassRegistry());
+  }
+
+  const RISCVInstrInfo *TII;
+  const TargetRegisterInfo *TRI;
+
+  // Track which register units have been modified and used.
+  LiveRegUnits ModifiedRegUnits, UsedRegUnits;
+
+  bool isCandidateToMergeMVA01S(const DestSourcePair &RegPair);
+  bool isCandidateToMergeMVSA01(const DestSourcePair &RegPair);
+  // Merge the two instructions indicated into a single pair instruction.
+  MachineBasicBlock::iterator
+  mergePairedInsns(MachineBasicBlock::iterator I,
+                   MachineBasicBlock::iterator Paired, unsigned Opcode);
+
+  // Look for C.MV instruction that can be combined with
+  // the given instruction into CM.MVA01S or CM.MVSA01. Return the matching
+  // instruction if one exists.
+  MachineBasicBlock::iterator
+  findMatchingInst(MachineBasicBlock::iterator &MBBI, unsigned InstOpcode,
+                   const DestSourcePair &RegPair);
+  bool mergeMoveSARegPair(MachineBasicBlock &MBB);
+  bool runOnMachineFunction(MachineFunction &Fn) override;
+
+  StringRef getPassName() const override { return RISCV_MOVE_MERGE_NAME; }
+};
+
+char RISCVMoveMerge::ID = 0;
+
+} // end of anonymous namespace
+
+INITIALIZE_PASS(RISCVMoveMerge, "riscv-move-merge", RISCV_MOVE_MERGE_NAME,
+                false, false)
+
+// Check if registers meet CM.MVA01S constraints.
+bool RISCVMoveMerge::isCandidateToMergeMVA01S(const DestSourcePair &RegPair) {
+  Register Destination = RegPair.Destination->getReg();
+  Register Source = RegPair.Source->getReg();
+  // If destination is not a0 or a1.
+  if ((Destination == RISCV::X10 || Destination == RISCV::X11) &&
+      RISCV::SR07RegClass.contains(Source))
+    return true;
+  return false;
+}
+
+// Check if registers meet CM.MVSA01 constraints.
+bool RISCVMoveMerge::isCandidateToMergeMVSA01(const DestSourcePair &RegPair) {
+  Register Destination = RegPair.Destination->getReg();
+  Register Source = RegPair.Source->getReg();
+  // If Source is s0 - s7.
+  if ((Source == RISCV::X10 || Source == RISCV::X11) &&
+      RISCV::SR07RegClass.contains(Destination))
+    return true;
+  return false;
+}
+
+MachineBasicBlock::iterator
+RISCVMoveMerge::mergePairedInsns(MachineBasicBlock::iterator I,
+                                 MachineBasicBlock::iterator Paired,
+                                 unsigned Opcode) {
+  const MachineOperand *Sreg1, *Sreg2;
+  MachineBasicBlock::iterator E = I->getParent()->end();
+  MachineBasicBlock::iterator NextI = next_nodbg(I, E);
+  DestSourcePair FirstPair = TII->isCopyInstrImpl(*I).value();
+  DestSourcePair PairedRegs = TII->isCopyInstrImpl(*Paired).value();
+  Register ARegInFirstPair = Opcode == RISCV::CM_MVA01S
+                                 ? FirstPair.Destination->getReg()
+                                 : FirstPair.Source->getReg();
+
+  if (NextI == Paired)
+    NextI = next_nodbg(NextI, E);
+  DebugLoc DL = I->getDebugLoc();
+
+  // The order of S-reg depends on which instruction holds A0, instead of
+  // the order of register pair.
+  // e,g.
+  //   mv a1, s1
+  //   mv a0, s2    =>  cm.mva01s s2,s1
+  //
+  //   mv a0, s2
+  //   mv a1, s1    =>  cm.mva01s s2,s1
+  bool StartWithX10 = ARegInFirstPair == RISCV::X10;
+  if (Opcode == RISCV::CM_MVA01S) {
+    Sreg1 = StartWithX10 ? FirstPair.Source : PairedRegs.Source;
+    Sreg2 = StartWithX10 ? PairedRegs.Source : FirstPair.Source;
+  } else {
+    Sreg1 = StartWithX10 ? FirstPair.Destination : PairedRegs.Destination;
+    Sreg2 = StartWithX10 ? PairedRegs.Destination : FirstPair.Destination;
+  }
+
+  BuildMI(*I->getParent(), I, DL, TII->get(Opcode)).add(*Sreg1).add(*Sreg2);
+
+  I->eraseFromParent();
+  Paired->eraseFromParent();
+  return NextI;
+}
+
+MachineBasicBlock::iterator
+RISCVMoveMerge::findMatchingInst(MachineBasicBlock::iterator &MBBI,
+                                 unsigned InstOpcode,
+                                 const DestSourcePair &RegPair) {
+  MachineBasicBlock::iterator E = MBBI->getParent()->end();
+
+  // Track which register units have been modified and used between the first
+  // insn and the second insn.
+  ModifiedRegUnits.clear();
+  UsedRegUnits.clear();
+
+  for (MachineBasicBlock::iterator I = next_nodbg(MBBI, E); I != E;
+       I = next_nodbg(I, E)) {
+
+    MachineInstr &MI = *I;
+
+    if (auto SecondPair = TII->isCopyInstrImpl(MI)) {
+      Register SourceReg = SecondPair->Source->getReg();
+      Register DestReg = SecondPair->Destination->getReg();
+
+      if (InstOpcode == RISCV::CM_MVA01S &&
+          isCandidateToMergeMVA01S(*SecondPair)) {
+        // If register pair is valid and destination registers are different.
+        if ((RegPair.Destination->getReg() == DestReg))
+          return E;
+
+        //  If paired destination register was modified or used, the source reg
+        //  was modified, there is no possibility of finding matching
+        //  instruction so exit early.
+        if (!ModifiedRegUnits.available(DestReg) ||
+            !UsedRegUnits.available(DestReg) ||
+            !ModifiedRegUnits.available(SourceReg))
+          return E;
+
+        return I;
+      } else if (InstOpcode == RISCV::CM_MVSA01 &&
+                 isCandidateToMergeMVSA01(*SecondPair)) {
+        if ((RegPair.Source->getReg() == SourceReg) ||
+            (RegPair.Destination->getReg() == DestReg))
+          return E;
+
+        if (!ModifiedRegUnits.available(DestReg) ||
+            !UsedRegUnits.available(DestReg) ||
+            !ModifiedRegUnits.available(SourceReg))
+          return E;
+
+        return I;
+      }
+    }
+    // Update modified / used register units.
+    LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI);
+  }
+  return E;
+}
+
+// Finds instructions, which could be represented as C.MV instructions and
+// merged into CM.MVA01S or CM.MVSA01.
+bool RISCVMoveMerge::mergeMoveSARegPair(MachineBasicBlock &MBB) {
+  bool Modified = false;
+
+  for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
+       MBBI != E;) {
+    // Check if the instruction can be compressed to C.MV instruction. If it
+    // can, return Dest/Src register pair.
+    auto RegPair = TII->isCopyInstrImpl(*MBBI);
+    if (RegPair.has_value()) {
+      unsigned Opcode = 0;
+
+      if (isCandidateToMergeMVA01S(*RegPair))
+        Opcode = RISCV::CM_MVA01S;
+      else if (isCandidateToMergeMVSA01(*RegPair))
+        Opcode = RISCV::CM_MVSA01;
+      else {
+        ++MBBI;
+        continue;
+      }
+
+      MachineBasicBlock::iterator Paired =
+          findMatchingInst(MBBI, Opcode, RegPair.value());
+      // If matching instruction can be found merge them.
+      if (Paired != E) {
+        MBBI = mergePairedInsns(MBBI, Paired, Opcode);
+        Modified = true;
+        continue;
+      }
+    }
+    ++MBBI;
+  }
+  return Modified;
+}
+
+bool RISCVMoveMerge::runOnMachineFunction(MachineFunction &Fn) {
+  if (skipFunction(Fn.getFunction()))
+    return false;
+
+  const RISCVSubtarget *Subtarget = &Fn.getSubtarget<RISCVSubtarget>();
+  if (!Subtarget->hasStdExtZcmp())
+    return false;
+
+  TII = Subtarget->getInstrInfo();
+  TRI = Subtarget->getRegisterInfo();
+  // Resize the modified and used register unit trackers.  We do this once
+  // per function and then clear the register units each time we optimize a
+  // move.
+  ModifiedRegUnits.init(*TRI);
+  UsedRegUnits.init(*TRI);
+  bool Modified = false;
+  for (auto &MBB : Fn)
+    Modified |= mergeMoveSARegPair(MBB);
+  return Modified;
+}
+
+/// createRISCVMoveMergePass - returns an instance of the
+/// move merge pass.
+FunctionPass *llvm::createRISCVMoveMergePass() { return new RISCVMoveMerge(); }

llvm/lib/Target/RISCV/RISCVTargetMachine.cpp

extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeRISCVTarget() {
   initializeRISCVMergeBaseOffsetOptPass(*PR);
   initializeRISCVOptWInstrsPass(*PR);
   initializeRISCVPreRAExpandPseudoPass(*PR);
   initializeRISCVExpandPseudoPass(*PR);
   initializeRISCVInsertVSETVLIPass(*PR);
   initializeRISCVInsertReadWriteCSRPass(*PR);
   initializeRISCVDAGToDAGISelPass(*PR);
   initializeRISCVInitUndefPass(*PR);
+  initializeRISCVMoveMergePass(*PR);
 }
 
 static StringRef computeDataLayout(const Triple &TT) {
   if (TT.isArch64Bit())
     return "e-m:e-p:64:64-i64:64-i128:128-n32:64-S128";
   assert(TT.isArch32Bit() && "only RV32 and RV64 are currently supported");
   return "e-m:e-p:32:32-i64:64-n32-S128";
 }
void RISCVPassConfig::addPreEmitPass() {
   // propagation after the machine outliner (which runs after addPreEmitPass)
   // currently leads to incorrect code-gen, where copies to registers within
   // outlined functions are removed erroneously.
   if (TM->getOptLevel() >= CodeGenOpt::Default && EnableRISCVCopyPropagation)
     addPass(createMachineCopyPropagationPass(true));
 }
 
 void RISCVPassConfig::addPreEmitPass2() {
+  if (TM->getOptLevel() != CodeGenOpt::None)
+    addPass(createRISCVMoveMergePass());
   addPass(createRISCVExpandPseudoPass());
 
   // Schedule the expansion of AMOs at the last possible moment, avoiding the
   // possibility for other passes to break the requirements for forward
   // progress in the LR/SC block.
   addPass(createRISCVExpandAtomicPseudoPass());
 }
 

llvm/test/CodeGen/RISCV/O3-pipeline.ll

 ; CHECK-NEXT:       StackMap Liveness Analysis
 ; CHECK-NEXT:       Live DEBUG_VALUE analysis
 ; CHECK-NEXT:       Machine Sanitizer Binary Metadata
 ; CHECK-NEXT:     Machine Outliner
 ; CHECK-NEXT:     FunctionPass Manager
 ; CHECK-NEXT:       Lazy Machine Block Frequency Analysis
 ; CHECK-NEXT:       Machine Optimization Remark Emitter
 ; CHECK-NEXT:       Stack Frame Layout Analysis
+; CHECK-NEXT:       RISC-V Zcmp move merging pass
 ; CHECK-NEXT:       RISC-V pseudo instruction expansion pass
 ; CHECK-NEXT:       RISC-V atomic pseudo instruction expansion pass
 ; CHECK-NEXT:       Lazy Machine Block Frequency Analysis
 ; CHECK-NEXT:       Machine Optimization Remark Emitter
 ; CHECK-NEXT:       RISC-V Assembly Printer
 ; CHECK-NEXT:       Free MachineFunction

llvm/test/CodeGen/RISCV/cm_mvas_mvsa.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
+; RUN: llc -mtriple=riscv32 -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefixes=CHECK32I %s
+; RUN: llc -mtriple=riscv32 -mattr=+experimental-zcmp -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefixes=CHECK32ZCMP %s
+; RUN: llc -mtriple=riscv64 -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefixes=CHECK64I %s
+; RUN: llc -mtriple=riscv64 -mattr=+experimental-zcmp -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefixes=CHECK64ZCMP %s
+
+declare i32 @foo(i32)
+declare i32 @func(i32,i32)
+
+define i32 @zcmp_mv(i32 %num, i32 %f) nounwind {
+; CHECK32I-LABEL: zcmp_mv:
+; CHECK32I:       # %bb.0:
+; CHECK32I-NEXT:    addi sp, sp, -16
+; CHECK32I-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
+; CHECK32I-NEXT:    sw s0, 8(sp) # 4-byte Folded Spill
+; CHECK32I-NEXT:    sw s1, 4(sp) # 4-byte Folded Spill
+; CHECK32I-NEXT:    sw s2, 0(sp) # 4-byte Folded Spill
+; CHECK32I-NEXT:    mv s0, a1
+; CHECK32I-NEXT:    mv s1, a0
+; CHECK32I-NEXT:    call func@plt
+; CHECK32I-NEXT:    mv s2, a0
+; CHECK32I-NEXT:    mv a0, s1
+; CHECK32I-NEXT:    mv a1, s0
+; CHECK32I-NEXT:    call func@plt
+; CHECK32I-NEXT:    add a0, s2, s0
+; CHECK32I-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
+; CHECK32I-NEXT:    lw s0, 8(sp) # 4-byte Folded Reload
+; CHECK32I-NEXT:    lw s1, 4(sp) # 4-byte Folded Reload
+; CHECK32I-NEXT:    lw s2, 0(sp) # 4-byte Folded Reload
+; CHECK32I-NEXT:    addi sp, sp, 16
+; CHECK32I-NEXT:    ret
+;
+; CHECK32ZCMP-LABEL: zcmp_mv:
+; CHECK32ZCMP:       # %bb.0:
+; CHECK32ZCMP-NEXT:    addi sp, sp, -16
+; CHECK32ZCMP-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
+; CHECK32ZCMP-NEXT:    sw s0, 8(sp) # 4-byte Folded Spill
+; CHECK32ZCMP-NEXT:    sw s1, 4(sp) # 4-byte Folded Spill
+; CHECK32ZCMP-NEXT:    sw s2, 0(sp) # 4-byte Folded Spill
+; CHECK32ZCMP-NEXT:    cm.mvsa01 s1, s0
+; CHECK32ZCMP-NEXT:    call func@plt
+; CHECK32ZCMP-NEXT:    mv s2, a0
+; CHECK32ZCMP-NEXT:    cm.mva01s s1, s0
+; CHECK32ZCMP-NEXT:    call func@plt
+; CHECK32ZCMP-NEXT:    add a0, s2, s0
+; CHECK32ZCMP-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
+; CHECK32ZCMP-NEXT:    lw s0, 8(sp) # 4-byte Folded Reload
+; CHECK32ZCMP-NEXT:    lw s1, 4(sp) # 4-byte Folded Reload
+; CHECK32ZCMP-NEXT:    lw s2, 0(sp) # 4-byte Folded Reload
+; CHECK32ZCMP-NEXT:    addi sp, sp, 16
+; CHECK32ZCMP-NEXT:    ret
+;
+; CHECK64I-LABEL: zcmp_mv:
+; CHECK64I:       # %bb.0:
+; CHECK64I-NEXT:    addi sp, sp, -32
+; CHECK64I-NEXT:    sd ra, 24(sp) # 8-byte Folded Spill
+; CHECK64I-NEXT:    sd s0, 16(sp) # 8-byte Folded Spill
+; CHECK64I-NEXT:    sd s1, 8(sp) # 8-byte Folded Spill
+; CHECK64I-NEXT:    sd s2, 0(sp) # 8-byte Folded Spill
+; CHECK64I-NEXT:    mv s0, a1
+; CHECK64I-NEXT:    mv s1, a0
+; CHECK64I-NEXT:    call func@plt
+; CHECK64I-NEXT:    mv s2, a0
+; CHECK64I-NEXT:    mv a0, s1
+; CHECK64I-NEXT:    mv a1, s0
+; CHECK64I-NEXT:    call func@plt
+; CHECK64I-NEXT:    addw a0, s2, s0
+; CHECK64I-NEXT:    ld ra, 24(sp) # 8-byte Folded Reload
+; CHECK64I-NEXT:    ld s0, 16(sp) # 8-byte Folded Reload
+; CHECK64I-NEXT:    ld s1, 8(sp) # 8-byte Folded Reload
+; CHECK64I-NEXT:    ld s2, 0(sp) # 8-byte Folded Reload
+; CHECK64I-NEXT:    addi sp, sp, 32
+; CHECK64I-NEXT:    ret
+;
+; CHECK64ZCMP-LABEL: zcmp_mv:
+; CHECK64ZCMP:       # %bb.0:
+; CHECK64ZCMP-NEXT:    addi sp, sp, -32
+; CHECK64ZCMP-NEXT:    sd ra, 24(sp) # 8-byte Folded Spill
+; CHECK64ZCMP-NEXT:    sd s0, 16(sp) # 8-byte Folded Spill
+; CHECK64ZCMP-NEXT:    sd s1, 8(sp) # 8-byte Folded Spill
+; CHECK64ZCMP-NEXT:    sd s2, 0(sp) # 8-byte Folded Spill
+; CHECK64ZCMP-NEXT:    cm.mvsa01 s1, s0
+; CHECK64ZCMP-NEXT:    call func@plt
+; CHECK64ZCMP-NEXT:    mv s2, a0
+; CHECK64ZCMP-NEXT:    cm.mva01s s1, s0
+; CHECK64ZCMP-NEXT:    call func@plt
+; CHECK64ZCMP-NEXT:    addw a0, s2, s0
+; CHECK64ZCMP-NEXT:    ld ra, 24(sp) # 8-byte Folded Reload
+; CHECK64ZCMP-NEXT:    ld s0, 16(sp) # 8-byte Folded Reload
+; CHECK64ZCMP-NEXT:    ld s1, 8(sp) # 8-byte Folded Reload
+; CHECK64ZCMP-NEXT:    ld s2, 0(sp) # 8-byte Folded Reload
+; CHECK64ZCMP-NEXT:    addi sp, sp, 32
+; CHECK64ZCMP-NEXT:    ret
+  %call = call i32 @func(i32 %num, i32 %f)
+  %call1 = call i32 @func(i32 %num, i32 %f)
+  %res = add i32 %call, %f
+  ret i32 %res
+}
+
+define i32 @not_zcmp_mv(i32 %num, i32 %f) nounwind {
+; CHECK32I-LABEL: not_zcmp_mv:
+; CHECK32I:       # %bb.0:
+; CHECK32I-NEXT:    addi sp, sp, -16
+; CHECK32I-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
+; CHECK32I-NEXT:    sw s0, 8(sp) # 4-byte Folded Spill
+; CHECK32I-NEXT:    sw s1, 4(sp) # 4-byte Folded Spill
+; CHECK32I-NEXT:    mv s0, a1
+; CHECK32I-NEXT:    call foo@plt
+; CHECK32I-NEXT:    mv s1, a0
+; CHECK32I-NEXT:    mv a0, s0
+; CHECK32I-NEXT:    call foo@plt
+; CHECK32I-NEXT:    mv a0, s1
+; CHECK32I-NEXT:    call foo@plt
+; CHECK32I-NEXT:    li a0, 1
+; CHECK32I-NEXT:    mv a1, s0
+; CHECK32I-NEXT:    call func@plt
+; CHECK32I-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
+; CHECK32I-NEXT:    lw s0, 8(sp) # 4-byte Folded Reload
+; CHECK32I-NEXT:    lw s1, 4(sp) # 4-byte Folded Reload
+; CHECK32I-NEXT:    addi sp, sp, 16
+; CHECK32I-NEXT:    ret
+;
+; CHECK32ZCMP-LABEL: not_zcmp_mv:
+; CHECK32ZCMP:       # %bb.0:
+; CHECK32ZCMP-NEXT:    addi sp, sp, -16
+; CHECK32ZCMP-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
+; CHECK32ZCMP-NEXT:    sw s0, 8(sp) # 4-byte Folded Spill
+; CHECK32ZCMP-NEXT:    sw s1, 4(sp) # 4-byte Folded Spill
+; CHECK32ZCMP-NEXT:    mv s0, a1
+; CHECK32ZCMP-NEXT:    call foo@plt
+; CHECK32ZCMP-NEXT:    mv s1, a0
+; CHECK32ZCMP-NEXT:    mv a0, s0
+; CHECK32ZCMP-NEXT:    call foo@plt
+; CHECK32ZCMP-NEXT:    mv a0, s1
+; CHECK32ZCMP-NEXT:    call foo@plt
+; CHECK32ZCMP-NEXT:    li a0, 1
+; CHECK32ZCMP-NEXT:    mv a1, s0
+; CHECK32ZCMP-NEXT:    call func@plt
+; CHECK32ZCMP-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
+; CHECK32ZCMP-NEXT:    lw s0, 8(sp) # 4-byte Folded Reload
+; CHECK32ZCMP-NEXT:    lw s1, 4(sp) # 4-byte Folded Reload
+; CHECK32ZCMP-NEXT:    addi sp, sp, 16
+; CHECK32ZCMP-NEXT:    ret
+;
+; CHECK64I-LABEL: not_zcmp_mv:
+; CHECK64I:       # %bb.0:
+; CHECK64I-NEXT:    addi sp, sp, -32
+; CHECK64I-NEXT:    sd ra, 24(sp) # 8-byte Folded Spill
+; CHECK64I-NEXT:    sd s0, 16(sp) # 8-byte Folded Spill
+; CHECK64I-NEXT:    sd s1, 8(sp) # 8-byte Folded Spill
+; CHECK64I-NEXT:    mv s0, a1
+; CHECK64I-NEXT:    call foo@plt
+; CHECK64I-NEXT:    mv s1, a0
+; CHECK64I-NEXT:    mv a0, s0
+; CHECK64I-NEXT:    call foo@plt
+; CHECK64I-NEXT:    mv a0, s1
+; CHECK64I-NEXT:    call foo@plt
+; CHECK64I-NEXT:    li a0, 1
+; CHECK64I-NEXT:    mv a1, s0
+; CHECK64I-NEXT:    call func@plt
+; CHECK64I-NEXT:    ld ra, 24(sp) # 8-byte Folded Reload
+; CHECK64I-NEXT:    ld s0, 16(sp) # 8-byte Folded Reload
+; CHECK64I-NEXT:    ld s1, 8(sp) # 8-byte Folded Reload
+; CHECK64I-NEXT:    addi sp, sp, 32
+; CHECK64I-NEXT:    ret
+;
+; CHECK64ZCMP-LABEL: not_zcmp_mv:
+; CHECK64ZCMP:       # %bb.0:
+; CHECK64ZCMP-NEXT:    addi sp, sp, -32
+; CHECK64ZCMP-NEXT:    sd ra, 24(sp) # 8-byte Folded Spill
+; CHECK64ZCMP-NEXT:    sd s0, 16(sp) # 8-byte Folded Spill
+; CHECK64ZCMP-NEXT:    sd s1, 8(sp) # 8-byte Folded Spill
+; CHECK64ZCMP-NEXT:    mv s0, a1
+; CHECK64ZCMP-NEXT:    call foo@plt
+; CHECK64ZCMP-NEXT:    mv s1, a0
+; CHECK64ZCMP-NEXT:    mv a0, s0
+; CHECK64ZCMP-NEXT:    call foo@plt
+; CHECK64ZCMP-NEXT:    mv a0, s1
+; CHECK64ZCMP-NEXT:    call foo@plt
+; CHECK64ZCMP-NEXT:    li a0, 1
+; CHECK64ZCMP-NEXT:    mv a1, s0
+; CHECK64ZCMP-NEXT:    call func@plt
+; CHECK64ZCMP-NEXT:    ld ra, 24(sp) # 8-byte Folded Reload
+; CHECK64ZCMP-NEXT:    ld s0, 16(sp) # 8-byte Folded Reload
+; CHECK64ZCMP-NEXT:    ld s1, 8(sp) # 8-byte Folded Reload
+; CHECK64ZCMP-NEXT:    addi sp, sp, 32
+; CHECK64ZCMP-NEXT:    ret
+  %call = call i32 @foo(i32 %num)
+  %call1 = call i32 @foo(i32 %f)
+  %tmp = call i32 @foo(i32 %call)
+  %res = call i32 @func(i32 1, i32 %f)
+  ret i32 %res
+}