[RISCV] Add an MIR pass to replace redundant sext.w instructions with copies.

[craig.topper]

Summary

Function calls and compare instructions tend to cause sext.w
instructions to be inserted. If we make good use of W instructions,
these operations can often end up being redundant. We don't always
detect these during SelectionDAG due to things like phis. There also
some cases caused by failure to turn extload into sextload in
SelectionDAG. extload selects to LW allowing later sext.ws to become
redundant.

This patch adds a pass that examines the input of sext.w instructions trying
to determine if it is already sign extended. Either by finding a
W instruction, other instructions that produce a sign extended result,
or looking through instructions that propagate sign bits. It uses
a worklist and visited set to search as far back as necessary.

[asb]

This is a nice improvement. I'm seeing a few cases where additional moves are generated when this pass is enabled. Presumably this is exposing a problem elsewhere (or just a case where the register allocator is slightly less good) - but I'm wondering if you have looked at this at all?

e.g. 20020529-1.c from the GCC torture suite:

--- a/output_rv64imafdc_lp64_O3/20020529-1.s
+++ b/output_rv64imafdc_lp64_O3/20020529-1.s
@@ -39,9 +39,10 @@ foo:                                    # @foo
        bnez    a0, .LBB0_5
 # %bb.9:                                # %if.end8.us.us
                                         #   in Loop: Header=BB0_8 Depth=1
-       sext.w  a2, a4
-       addiw   a4, a4, 1
-       bne     a2, a1, .LBB0_8
+       addiw   a2, a4, 1
+       mv      a3, a4
+       mv      a4, a2
+       bne     a3, a1, .LBB0_8
 .LBB0_10:                               # %if.then
        lui     a0, %hi(f1.beenhere)
        li      a1, 2

llvm/lib/Target/RISCV/RISCVSExtWRemoval.cpp

62

Is it just the AMO operations missing?

Are the missing instructions not listed due to further work being needed, or just lack of test coverage?

llvm/lib/Target/RISCV/RISCVTargetMachine.cpp

202

Only add the pass if targeting RV64?

[craig.topper]

llvm/lib/Target/RISCV/RISCVSExtWRemoval.cpp

62

The bitmanip instructions that weren’t up for ratification are also missing. Partly because I wasn’t sure if I we’d just have to delete it someday.

[craig.topper]

Only add the pass on RV64 target.

[craig.topper]

In D116397#3213615, @asb wrote:

This is a nice improvement. I'm seeing a few cases where additional moves are generated when this pass is enabled. Presumably this is exposing a problem elsewhere (or just a case where the register allocator is slightly less good) - but I'm wondering if you have looked at this at all?

e.g. 20020529-1.c from the GCC torture suite:

--- a/output_rv64imafdc_lp64_O3/20020529-1.s
+++ b/output_rv64imafdc_lp64_O3/20020529-1.s
@@ -39,9 +39,10 @@ foo:                                    # @foo
        bnez    a0, .LBB0_5
 # %bb.9:                                # %if.end8.us.us
                                         #   in Loop: Header=BB0_8 Depth=1
-       sext.w  a2, a4
-       addiw   a4, a4, 1
-       bne     a2, a1, .LBB0_8
+       addiw   a2, a4, 1
+       mv      a3, a4
+       mv      a4, a2
+       bne     a3, a1, .LBB0_8
 .LBB0_10:                               # %if.then
        lui     a0, %hi(f1.beenhere)
        li      a1, 2

This looks like a bad scheduling decision. The NoSchedModel gave a latency of 1 for the sext.w, but a latency of 0 for the COPY that replaced it. With the latency of 1 the sext.w got scheduled before the addiw, with the latency of 0 the copy got scheduled after.

With -mcpu=rocket-rv64 the COPY gets scheduled before.

[craig.topper]

In D116397#3214150, @craig.topper wrote:

In D116397#3213615, @asb wrote:

This is a nice improvement. I'm seeing a few cases where additional moves are generated when this pass is enabled. Presumably this is exposing a problem elsewhere (or just a case where the register allocator is slightly less good) - but I'm wondering if you have looked at this at all?

e.g. 20020529-1.c from the GCC torture suite:

--- a/output_rv64imafdc_lp64_O3/20020529-1.s
+++ b/output_rv64imafdc_lp64_O3/20020529-1.s
@@ -39,9 +39,10 @@ foo:                                    # @foo
        bnez    a0, .LBB0_5
 # %bb.9:                                # %if.end8.us.us
                                         #   in Loop: Header=BB0_8 Depth=1
-       sext.w  a2, a4
-       addiw   a4, a4, 1
-       bne     a2, a1, .LBB0_8
+       addiw   a2, a4, 1
+       mv      a3, a4
+       mv      a4, a2
+       bne     a3, a1, .LBB0_8
 .LBB0_10:                               # %if.then
        lui     a0, %hi(f1.beenhere)
        li      a1, 2

This looks like a bad scheduling decision. The NoSchedModel gave a latency of 1 for the sext.w, but a latency of 0 for the COPY that replaced it. With the latency of 1 the sext.w got scheduled before the addiw, with the latency of 0 the copy got scheduled after.

With -mcpu=rocket-rv64 the COPY gets scheduled before.

This is also caused by loop strength reduce not converting the loop exit condition to use the post inc value. This may be related to i32 not being a native type in datalayout. See also https://lists.llvm.org/pipermail/llvm-dev/2021-December/154474.html

[kito-cheng]

In D116397#3214159, @craig.topper wrote:Does it possible to just propagate the register IFF there is only one def (no phi)?

In D116397#3214150, @craig.topper wrote:

In D116397#3213615, @asb wrote:

This is a nice improvement. I'm seeing a few cases where additional moves are generated when this pass is enabled. Presumably this is exposing a problem elsewhere (or just a case where the register allocator is slightly less good) - but I'm wondering if you have looked at this at all?

e.g. 20020529-1.c from the GCC torture suite:

--- a/output_rv64imafdc_lp64_O3/20020529-1.s
+++ b/output_rv64imafdc_lp64_O3/20020529-1.s
@@ -39,9 +39,10 @@ foo:                                    # @foo
        bnez    a0, .LBB0_5
 # %bb.9:                                # %if.end8.us.us
                                         #   in Loop: Header=BB0_8 Depth=1
-       sext.w  a2, a4
-       addiw   a4, a4, 1
-       bne     a2, a1, .LBB0_8
+       addiw   a2, a4, 1
+       mv      a3, a4
+       mv      a4, a2
+       bne     a3, a1, .LBB0_8
 .LBB0_10:                               # %if.then
        lui     a0, %hi(f1.beenhere)
        li      a1, 2

This looks like a bad scheduling decision. The NoSchedModel gave a latency of 1 for the sext.w, but a latency of 0 for the COPY that replaced it. With the latency of 1 the sext.w got scheduled before the addiw, with the latency of 0 the copy got scheduled after.

With -mcpu=rocket-rv64 the COPY gets scheduled before.

This is also caused by loop strength reduce not converting the loop exit condition to use the post inc value. This may be related to i32 not being a native type in datalayout. See also https://lists.llvm.org/pipermail/llvm-dev/2021-December/154474.html

Does it possible to just propagate the register IFF there is only one def (no phi) rather than insert COPY? so that the extra mv can gone?

[kito-cheng]

Oh I take a closer look, one def (no phi) or not should not matter, I tried to add MachineCopyPropagation but not work since it require non-SSA, in theory we could remove that by a simple copy propagation after this pass.

Before:

bb.5.if.end.us.us:
; predecessors: %bb.4, %bb.6
  successors: %bb.6(0x7ffff800), %bb.11(0x00000800); %bb.6(100.00%), %bb.11(0.00%)

  %3:gpr = PHI %1:gpr, %bb.4, %4:gpr, %bb.6
  %23:gpr = COPY $x0
  BNE %22:gpr, %23:gpr, %bb.11
  PseudoBR %bb.6

bb.6.if.end8.us.us:
; predecessors: %bb.5
  successors: %bb.10(0x04000000), %bb.5(0x7c000000); %bb.10(3.12%), %bb.5(96.88%)

  %4:gpr = ADDIW %3:gpr, 1
  %25:gpr = ADDIW %3:gpr, 0
  BEQ killed %25:gpr, %21:gpr, %bb.10
  PseudoBR %bb.5

After:

bb.5.if.end.us.us:
; predecessors: %bb.4, %bb.6 
  successors: %bb.6(0x7ffff800), %bb.11(0x00000800); %bb.6(100.00%), %bb.11(0.00%)

  %3:gpr = PHI %1:gpr, %bb.4, %4:gpr, %bb.6 
  %23:gpr = COPY $x0
  BNE %22:gpr, %23:gpr, %bb.11
  PseudoBR %bb.6 

bb.6.if.end8.us.us:
; predecessors: %bb.5 
  successors: %bb.10(0x04000000), %bb.5(0x7c000000); %bb.10(3.12%), %bb.5(96.88%)

  %4:gpr = ADDIW %3:gpr, 1
  %25:gpr = COPY %3:gpr
  BEQ killed %25:gpr, %21:gpr, %bb.10
  PseudoBR %bb.5

[craig.topper]

Merge the source and destination register instead of inserting a copy.

[craig.topper]

Constrain the reg class before merging the registers. If we can't constrain, don't replace the sext.w.

[kito-cheng]

Quick patch for that, but seems also change the BB layout of 20020529-1...

diff --git a/llvm/lib/Target/RISCV/RISCVSExtWRemoval.cpp b/llvm/lib/Target/RISCV/RISCVSExtWRemoval.cpp
index acd79b4ab742..e8774a033ca3 100644
--- a/llvm/lib/Target/RISCV/RISCVSExtWRemoval.cpp
+++ b/llvm/lib/Target/RISCV/RISCVSExtWRemoval.cpp
@@ -251,9 +251,19 @@ bool RISCVSExtWRemoval::runOnMachineFunction(MachineFunction &MF) {
         continue;
 
       LLVM_DEBUG(dbgs() << "Removing redundant sign-extension\n");
-      BuildMI(MBB, MI, MI->getDebugLoc(), TII->get(RISCV::COPY),
-              MI->getOperand(0).getReg())
-          .add(MI->getOperand(1));
+      Register DstReg = MI->getOperand(0).getReg();
+      LLVM_DEBUG(dbgs() << "Replace " << DstReg << " to " << SrcReg << "\n");
+      // Rewrite all uses.
+      // const TargetRegisterInfo &TRI = ST.getRegisterInfo();
+      const TargetRegisterInfo *TRI =
+          MF.getSubtarget<RISCVSubtarget>().getRegisterInfo();
+      for (MachineBasicBlock &MBB : MF) {
+        for (auto I = MBB.begin(), IE = MBB.end(); I != IE;) {
+          MachineInstr *MI = &*I++;
+          if (auto MO = MI->findRegisterUseOperand(DstReg))
+            MO->substVirtReg(SrcReg, 0, *TRI);
+        }
+      }
       MI->eraseFromParent();
       ++NumRemovedSExtW;
       MadeChange = true;

[kito-cheng]

LGTM without copy prorogation improvement, copy prorogation might cause extra code gen issue, so I think that part could improve later with more investigation.

[kito-cheng]

Oh, my mistake I missed the updated version, LGTM.

[asb]

This LGTM, thanks!

llvm/lib/Target/RISCV/CMakeLists.txt

add_llvm_target(RISCVCodeGen
   RISCVInstructionSelector.cpp
   RISCVISelDAGToDAG.cpp
   RISCVISelLowering.cpp
   RISCVLegalizerInfo.cpp
   RISCVMCInstLower.cpp
   RISCVMergeBaseOffset.cpp
   RISCVRegisterBankInfo.cpp
   RISCVRegisterInfo.cpp
+  RISCVSExtWRemoval.cpp
   RISCVSubtarget.cpp
   RISCVTargetMachine.cpp
   RISCVTargetObjectFile.cpp
   RISCVTargetTransformInfo.cpp
 
   LINK_COMPONENTS
   Analysis
   AsmPrinter

llvm/lib/Target/RISCV/RISCV.h

 bool LowerRISCVMachineOperandToMCOperand(const MachineOperand &MO,
                                          MCOperand &MCOp, const AsmPrinter &AP);
 
 FunctionPass *createRISCVISelDag(RISCVTargetMachine &TM);
 
 FunctionPass *createRISCVGatherScatterLoweringPass();
 void initializeRISCVGatherScatterLoweringPass(PassRegistry &);
 
+FunctionPass *createRISCVSExtWRemovalPass();
+void initializeRISCVSExtWRemovalPass(PassRegistry &);
+
 FunctionPass *createRISCVMergeBaseOffsetOptPass();
 void initializeRISCVMergeBaseOffsetOptPass(PassRegistry &);
 
 FunctionPass *createRISCVExpandPseudoPass();
 void initializeRISCVExpandPseudoPass(PassRegistry &);
 
 FunctionPass *createRISCVExpandAtomicPseudoPass();
 void initializeRISCVExpandAtomicPseudoPass(PassRegistry &);

llvm/lib/Target/RISCV/RISCVSExtWRemoval.cpp

+//===-------------- RISCVSExtWRemoval.cpp - MI sext.w Removal -------------===//
+//
+// 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 removes unneeded sext.w instructions at the MI level.
+//
+//===---------------------------------------------------------------------===//
+
+#include "RISCV.h"
+#include "RISCVSubtarget.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/TargetInstrInfo.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "riscv-sextw-removal"
+
+STATISTIC(NumRemovedSExtW, "Number of removed sign-extensions");
+
+static cl::opt<bool> DisableSExtWRemoval("riscv-disable-sextw-removal",
+                                         cl::desc("Disable removal of sext.w"),
+                                         cl::init(false), cl::Hidden);
+namespace {
+
+class RISCVSExtWRemoval : public MachineFunctionPass {
+public:
+  static char ID;
+
+  RISCVSExtWRemoval() : MachineFunctionPass(ID) {
+    initializeRISCVSExtWRemovalPass(*PassRegistry::getPassRegistry());
+  }
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.setPreservesCFG();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+  StringRef getPassName() const override { return "RISCV sext.w Removal"; }
+};
+
+} // end anonymous namespace
+
+char RISCVSExtWRemoval::ID = 0;
+INITIALIZE_PASS(RISCVSExtWRemoval, DEBUG_TYPE, "RISCV sext.w Removal", false,
+                false)
+
+FunctionPass *llvm::createRISCVSExtWRemovalPass() {
+  return new RISCVSExtWRemoval();
+}
+
+// This function returns true if the machine instruction always outputs a value
+// where bits 63:32 match bit 31.
+// TODO: Allocate a bit in TSFlags for the W instructions?
+// TODO: Add other W instructions.
+static bool isSignExtendingOpW(const MachineInstr &MI) {
+  switch (MI.getOpcode()) {
+  case RISCV::LUI:
+  case RISCV::LW:
+  case RISCV::ADDW:
+  case RISCV::ADDIW:
+  case RISCV::SUBW:
+  case RISCV::MULW:
+  case RISCV::SLLW:
+  case RISCV::SLLIW:
+  case RISCV::SRAW:
+  case RISCV::SRAIW:
+  case RISCV::SRLW:
+  case RISCV::SRLIW:
+  case RISCV::DIVW:
+  case RISCV::DIVUW:
+  case RISCV::REMW:
+  case RISCV::REMUW:
+  case RISCV::ROLW:
+  case RISCV::RORW:
+  case RISCV::RORIW:
+  case RISCV::CLZW:
+  case RISCV::CTZW:
+  case RISCV::CPOPW:
+  case RISCV::FCVT_W_H:
+  case RISCV::FCVT_WU_H:
+  case RISCV::FCVT_W_S:
+  case RISCV::FCVT_WU_S:
+  case RISCV::FCVT_W_D:
+  case RISCV::FCVT_WU_D:
+  // The following aren't W instructions, but are either sign extended from a
+  // smaller size or put zeros in bits 63:31.
+  case RISCV::LBU:
+  case RISCV::LHU:
+  case RISCV::LB:
+  case RISCV::LH:
+  case RISCV::SEXTB:
+  case RISCV::SEXTH:
+  case RISCV::ZEXTH_RV64:
+    return true;
+  }
+
+  // The LI pattern ADDI rd, X0, imm is sign extended.
+  if (MI.getOpcode() == RISCV::ADDI && MI.getOperand(1).isReg() &&
+      MI.getOperand(1).getReg() == RISCV::X0)
+    return true;
+
+  // An ANDI with an 11 bit immediate will zero bits 63:11.
+  if (MI.getOpcode() == RISCV::ANDI && isUInt<11>(MI.getOperand(2).getImm()))
+    return true;
+
+  // Copying from X0 produces zero.
+  if (MI.getOpcode() == RISCV::COPY && MI.getOperand(1).getReg() == RISCV::X0)
+    return true;
+
+  return false;
+}
+
+static bool isSignExtendedW(const MachineInstr &OrigMI,
+                            MachineRegisterInfo &MRI) {
+
+  SmallPtrSet<const MachineInstr *, 4> Visited;
+  SmallVector<const MachineInstr *, 4> Worklist;
+
+  Worklist.push_back(&OrigMI);
+
+  while (!Worklist.empty()) {
+    const MachineInstr *MI = Worklist.pop_back_val();
+
+    // If we already visited this instruction, we don't need to check it again.
+    if (!Visited.insert(MI).second)
+      continue;
+
+    // If this is a sign extending operation we don't need to look any further.
+    if (isSignExtendingOpW(*MI))
+      continue;
+
+    // Is this an instruction that propagates sign extend.
+    switch (MI->getOpcode()) {
+    default:
+      // Unknown opcode, give up.
+      return false;
+    case RISCV::COPY: {
+      Register SrcReg = MI->getOperand(1).getReg();
+
+      // TODO: Handle arguments and returns from calls?
+
+      // If this is a copy from another register, check its source instruction.
+      if (!SrcReg.isVirtual())
+        return false;
+      const MachineInstr *SrcMI = MRI.getVRegDef(SrcReg);
+      if (!SrcMI)
+        return false;
+
+      // Add SrcMI to the worklist.
+      Worklist.push_back(SrcMI);
+      break;
+    }
+    case RISCV::ANDI:
+    case RISCV::ORI:
+    case RISCV::XORI: {
+      // Logical operations use a sign extended 12-bit immediate. We just need
+      // to check if the other operand is sign extended.
+      Register SrcReg = MI->getOperand(1).getReg();
+      if (!SrcReg.isVirtual())
+        return false;
+      const MachineInstr *SrcMI = MRI.getVRegDef(SrcReg);
+      if (!SrcMI)
+        return false;
+
+      // Add SrcMI to the worklist.
+      Worklist.push_back(SrcMI);
+      break;
+    }
+    case RISCV::AND:
+    case RISCV::OR:
+    case RISCV::XOR:
+    case RISCV::ANDN:
+    case RISCV::ORN:
+    case RISCV::XNOR:
+    case RISCV::MAX:
+    case RISCV::MAXU:
+    case RISCV::MIN:
+    case RISCV::MINU:
+    case RISCV::PHI: {
+      // If all incoming values are sign-extended, the output of AND, OR, XOR,
+      // MIN, MAX, or PHI is also sign-extended.
+
+      // The input registers for PHI are operand 1, 3, ...
+      // The input registers for others are operand 1 and 2.
+      unsigned E = 3, D = 1;
+      if (MI->getOpcode() == RISCV::PHI) {
+        E = MI->getNumOperands();
+        D = 2;
+      }
+
+      for (unsigned I = 1; I != E; I += D) {
+        if (!MI->getOperand(I).isReg())
+          return false;
+
+        Register SrcReg = MI->getOperand(I).getReg();
+        if (!SrcReg.isVirtual())
+          return false;
+        const MachineInstr *SrcMI = MRI.getVRegDef(SrcReg);
+        if (!SrcMI)
+          return false;
+
+        // Add SrcMI to the worklist.
+        Worklist.push_back(SrcMI);
+      }
+
+      break;
+    }
+    }
+  }
+
+  // If we get here, then every node we visited produces a sign extended value
+  // or propagated sign extended values. So the result must be sign extended.
+  return true;
+}
+
+bool RISCVSExtWRemoval::runOnMachineFunction(MachineFunction &MF) {
+  if (skipFunction(MF.getFunction()) || DisableSExtWRemoval)
+    return false;
+
+  MachineRegisterInfo &MRI = MF.getRegInfo();
+  const RISCVSubtarget &ST = MF.getSubtarget<RISCVSubtarget>();
+
+  if (!ST.is64Bit())
+    return false;
+
+  bool MadeChange = false;
+  for (MachineBasicBlock &MBB : MF) {
+    for (auto I = MBB.begin(), IE = MBB.end(); I != IE;) {
+      MachineInstr *MI = &*I++;
+
+      // We're looking for the sext.w pattern ADDIW rd, rs1, 0.
+      if (MI->getOpcode() != RISCV::ADDIW || !MI->getOperand(2).isImm() ||
+          MI->getOperand(2).getImm() != 0 || !MI->getOperand(1).isReg())
+        continue;
+
+      // Input should be a virtual register.
+      Register SrcReg = MI->getOperand(1).getReg();
+      if (!SrcReg.isVirtual())
+        continue;
+
+      const MachineInstr &SrcMI = *MRI.getVRegDef(SrcReg);
+      if (!isSignExtendedW(SrcMI, MRI))
+        continue;
+
+      Register DstReg = MI->getOperand(0).getReg();
+      if (!MRI.constrainRegClass(SrcReg, MRI.getRegClass(DstReg)))
+        continue;
+
+      LLVM_DEBUG(dbgs() << "Removing redundant sign-extension\n");
+      MRI.replaceRegWith(DstReg, SrcReg);
+      MRI.clearKillFlags(SrcReg);
+      MI->eraseFromParent();
+      ++NumRemovedSExtW;
+      MadeChange = true;
+    }
+  }
+
+  return MadeChange;
+}

llvm/lib/Target/RISCV/RISCVTargetMachine.cpp

 
 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeRISCVTarget() {
   RegisterTargetMachine<RISCVTargetMachine> X(getTheRISCV32Target());
   RegisterTargetMachine<RISCVTargetMachine> Y(getTheRISCV64Target());
   auto *PR = PassRegistry::getPassRegistry();
   initializeGlobalISel(*PR);
   initializeRISCVGatherScatterLoweringPass(*PR);
   initializeRISCVMergeBaseOffsetOptPass(*PR);
+  initializeRISCVSExtWRemovalPass(*PR);
   initializeRISCVExpandPseudoPass(*PR);
   initializeRISCVInsertVSETVLIPass(*PR);
 }
 
 static StringRef computeDataLayout(const Triple &TT) {
   if (TT.isArch64Bit())
     return "e-m:e-p:64:64-i64:64-i128:128-n64-S128";
   assert(TT.isArch32Bit() && "only RV32 and RV64 are currently supported");
public:
   bool addInstSelector() override;
   bool addIRTranslator() override;
   bool addLegalizeMachineIR() override;
   bool addRegBankSelect() override;
   bool addGlobalInstructionSelect() override;
   void addPreEmitPass() override;
   void addPreEmitPass2() override;
   void addPreSched2() override;
+  void addMachineSSAOptimization() override;
   void addPreRegAlloc() override;
 };
 } // namespace
 
 TargetPassConfig *RISCVTargetMachine::createPassConfig(PassManagerBase &PM) {
   return new RISCVPassConfig(*this, PM);
 }
 

 void RISCVPassConfig::addPreEmitPass2() {
   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());
 }
 
+void RISCVPassConfig::addMachineSSAOptimization() {
+  TargetPassConfig::addMachineSSAOptimization();
+
+  if (TM->getTargetTriple().getArch() == Triple::riscv64)
+    addPass(createRISCVSExtWRemovalPass());
+}
+
 void RISCVPassConfig::addPreRegAlloc() {
   if (TM->getOptLevel() != CodeGenOpt::None)
     addPass(createRISCVMergeBaseOffsetOptPass());
   addPass(createRISCVInsertVSETVLIPass());
 }

llvm/test/CodeGen/RISCV/atomic-rmw.ll

 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a3, 0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB145_4
 ; RV64I-NEXT:  .LBB145_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    blt s1, a0, .LBB145_1
+; RV64I-NEXT:    blt s1, a3, .LBB145_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB145_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB145_1
 ; RV64I-NEXT:  .LBB145_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 2
 ; RV64I-NEXT:    li a4, 2
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB146_4
 ; RV64I-NEXT:  .LBB146_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    blt s1, a0, .LBB146_1
+; RV64I-NEXT:    blt s1, a3, .LBB146_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB146_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB146_1
 ; RV64I-NEXT:  .LBB146_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 3
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB147_4
 ; RV64I-NEXT:  .LBB147_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    blt s1, a0, .LBB147_1
+; RV64I-NEXT:    blt s1, a3, .LBB147_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB147_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB147_1
 ; RV64I-NEXT:  .LBB147_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 4
 ; RV64I-NEXT:    li a4, 2
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB148_4
 ; RV64I-NEXT:  .LBB148_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    blt s1, a0, .LBB148_1
+; RV64I-NEXT:    blt s1, a3, .LBB148_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB148_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB148_1
 ; RV64I-NEXT:  .LBB148_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 5
 ; RV64I-NEXT:    li a4, 5
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB149_4
 ; RV64I-NEXT:  .LBB149_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    blt s1, a0, .LBB149_1
+; RV64I-NEXT:    blt s1, a3, .LBB149_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB149_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB149_1
 ; RV64I-NEXT:  .LBB149_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a3, 0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB150_4
 ; RV64I-NEXT:  .LBB150_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bge s1, a0, .LBB150_1
+; RV64I-NEXT:    bge s1, a3, .LBB150_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB150_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB150_1
 ; RV64I-NEXT:  .LBB150_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 2
 ; RV64I-NEXT:    li a4, 2
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB151_4
 ; RV64I-NEXT:  .LBB151_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bge s1, a0, .LBB151_1
+; RV64I-NEXT:    bge s1, a3, .LBB151_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB151_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB151_1
 ; RV64I-NEXT:  .LBB151_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 3
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB152_4
 ; RV64I-NEXT:  .LBB152_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bge s1, a0, .LBB152_1
+; RV64I-NEXT:    bge s1, a3, .LBB152_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB152_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB152_1
 ; RV64I-NEXT:  .LBB152_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 4
 ; RV64I-NEXT:    li a4, 2
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB153_4
 ; RV64I-NEXT:  .LBB153_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bge s1, a0, .LBB153_1
+; RV64I-NEXT:    bge s1, a3, .LBB153_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB153_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB153_1
 ; RV64I-NEXT:  .LBB153_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 5
 ; RV64I-NEXT:    li a4, 5
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB154_4
 ; RV64I-NEXT:  .LBB154_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bge s1, a0, .LBB154_1
+; RV64I-NEXT:    bge s1, a3, .LBB154_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB154_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB154_1
 ; RV64I-NEXT:  .LBB154_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a3, 0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB155_4
 ; RV64I-NEXT:  .LBB155_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bltu s1, a0, .LBB155_1
+; RV64I-NEXT:    bltu s1, a3, .LBB155_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB155_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB155_1
 ; RV64I-NEXT:  .LBB155_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 2
 ; RV64I-NEXT:    li a4, 2
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB156_4
 ; RV64I-NEXT:  .LBB156_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bltu s1, a0, .LBB156_1
+; RV64I-NEXT:    bltu s1, a3, .LBB156_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB156_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB156_1
 ; RV64I-NEXT:  .LBB156_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 3
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB157_4
 ; RV64I-NEXT:  .LBB157_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bltu s1, a0, .LBB157_1
+; RV64I-NEXT:    bltu s1, a3, .LBB157_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB157_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB157_1
 ; RV64I-NEXT:  .LBB157_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 4
 ; RV64I-NEXT:    li a4, 2
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB158_4
 ; RV64I-NEXT:  .LBB158_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bltu s1, a0, .LBB158_1
+; RV64I-NEXT:    bltu s1, a3, .LBB158_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB158_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB158_1
 ; RV64I-NEXT:  .LBB158_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 5
 ; RV64I-NEXT:    li a4, 5
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB159_4
 ; RV64I-NEXT:  .LBB159_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bltu s1, a0, .LBB159_1
+; RV64I-NEXT:    bltu s1, a3, .LBB159_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB159_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB159_1
 ; RV64I-NEXT:  .LBB159_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a3, 0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB160_4
 ; RV64I-NEXT:  .LBB160_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bgeu s1, a0, .LBB160_1
+; RV64I-NEXT:    bgeu s1, a3, .LBB160_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB160_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB160_1
 ; RV64I-NEXT:  .LBB160_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 2
 ; RV64I-NEXT:    li a4, 2
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB161_4
 ; RV64I-NEXT:  .LBB161_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bgeu s1, a0, .LBB161_1
+; RV64I-NEXT:    bgeu s1, a3, .LBB161_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB161_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB161_1
 ; RV64I-NEXT:  .LBB161_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 3
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB162_4
 ; RV64I-NEXT:  .LBB162_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bgeu s1, a0, .LBB162_1
+; RV64I-NEXT:    bgeu s1, a3, .LBB162_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB162_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB162_1
 ; RV64I-NEXT:  .LBB162_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 4
 ; RV64I-NEXT:    li a4, 2
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB163_4
 ; RV64I-NEXT:  .LBB163_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bgeu s1, a0, .LBB163_1
+; RV64I-NEXT:    bgeu s1, a3, .LBB163_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB163_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB163_1
 ; RV64I-NEXT:  .LBB163_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

 ; RV64I-NEXT:    li a3, 5
 ; RV64I-NEXT:    li a4, 5
 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB164_4
 ; RV64I-NEXT:  .LBB164_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bgeu s1, a0, .LBB164_1
+; RV64I-NEXT:    bgeu s1, a3, .LBB164_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB164_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB164_1
 ; RV64I-NEXT:  .LBB164_4: # %atomicrmw.end
 ; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload

llvm/test/CodeGen/RISCV/atomic-signext.ll

 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a3, 0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB32_4
 ; RV64I-NEXT:  .LBB32_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    blt s1, a0, .LBB32_1
+; RV64I-NEXT:    blt s1, a3, .LBB32_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB32_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB32_1
 ; RV64I-NEXT:  .LBB32_4: # %atomicrmw.end
-; RV64I-NEXT:    sext.w a0, a3
+; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s1, 24(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s2, 16(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    addi sp, sp, 48
 ; RV64I-NEXT:    ret
 ;
 ; RV64IA-LABEL: atomicrmw_max_i32_monotonic:

 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a3, 0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB33_4
 ; RV64I-NEXT:  .LBB33_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bge s1, a0, .LBB33_1
+; RV64I-NEXT:    bge s1, a3, .LBB33_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB33_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB33_1
 ; RV64I-NEXT:  .LBB33_4: # %atomicrmw.end
-; RV64I-NEXT:    sext.w a0, a3
+; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s1, 24(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s2, 16(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    addi sp, sp, 48
 ; RV64I-NEXT:    ret
 ;
 ; RV64IA-LABEL: atomicrmw_min_i32_monotonic:

 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a3, 0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB34_4
 ; RV64I-NEXT:  .LBB34_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bltu s1, a0, .LBB34_1
+; RV64I-NEXT:    bltu s1, a3, .LBB34_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB34_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB34_1
 ; RV64I-NEXT:  .LBB34_4: # %atomicrmw.end
-; RV64I-NEXT:    sext.w a0, a3
+; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s1, 24(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s2, 16(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    addi sp, sp, 48
 ; RV64I-NEXT:    ret
 ;
 ; RV64IA-LABEL: atomicrmw_umax_i32_monotonic:

 ; RV64I-NEXT:    mv a0, s0
 ; RV64I-NEXT:    li a3, 0
 ; RV64I-NEXT:    li a4, 0
 ; RV64I-NEXT:    call __atomic_compare_exchange_4@plt
 ; RV64I-NEXT:    lw a3, 12(sp)
 ; RV64I-NEXT:    bnez a0, .LBB35_4
 ; RV64I-NEXT:  .LBB35_2: # %atomicrmw.start
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
-; RV64I-NEXT:    sext.w a0, a3
 ; RV64I-NEXT:    mv a2, a3
-; RV64I-NEXT:    bgeu s1, a0, .LBB35_1
+; RV64I-NEXT:    bgeu s1, a3, .LBB35_1
 ; RV64I-NEXT:  # %bb.3: # %atomicrmw.start
 ; RV64I-NEXT:    # in Loop: Header=BB35_2 Depth=1
 ; RV64I-NEXT:    mv a2, s2
 ; RV64I-NEXT:    j .LBB35_1
 ; RV64I-NEXT:  .LBB35_4: # %atomicrmw.end
-; RV64I-NEXT:    sext.w a0, a3
+; RV64I-NEXT:    mv a0, a3
 ; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s1, 24(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    ld s2, 16(sp) # 8-byte Folded Reload
 ; RV64I-NEXT:    addi sp, sp, 48
 ; RV64I-NEXT:    ret
 ;
 ; RV64IA-LABEL: atomicrmw_umin_i32_monotonic:

llvm/test/CodeGen/RISCV/rv64zbs.ll

 define signext i32 @bclr_i32_load(i32* %p, i32 signext %b) nounwind {
 ; RV64I-LABEL: bclr_i32_load:
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    lw a0, 0(a0)
 ; RV64I-NEXT:    li a2, 1
 ; RV64I-NEXT:    sllw a1, a2, a1
 ; RV64I-NEXT:    not a1, a1
 ; RV64I-NEXT:    and a0, a1, a0
-; RV64I-NEXT:    sext.w a0, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64ZBS-LABEL: bclr_i32_load:
 ; RV64ZBS:       # %bb.0:
 ; RV64ZBS-NEXT:    lw a0, 0(a0)
 ; RV64ZBS-NEXT:    li a2, 1
 ; RV64ZBS-NEXT:    sllw a1, a2, a1
 ; RV64ZBS-NEXT:    not a1, a1
 ; RV64ZBS-NEXT:    and a0, a1, a0
-; RV64ZBS-NEXT:    sext.w a0, a0
 ; RV64ZBS-NEXT:    ret
   %a = load i32, i32* %p
   %shl = shl i32 1, %b
   %neg = xor i32 %shl, -1
   %and1 = and i32 %neg, %a
   ret i32 %and1
 }
 

 
 define signext i32 @bset_i32_load(i32* %p, i32 signext %b) nounwind {
 ; RV64I-LABEL: bset_i32_load:
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    lw a0, 0(a0)
 ; RV64I-NEXT:    li a2, 1
 ; RV64I-NEXT:    sllw a1, a2, a1
 ; RV64I-NEXT:    or a0, a1, a0
-; RV64I-NEXT:    sext.w a0, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64ZBS-LABEL: bset_i32_load:
 ; RV64ZBS:       # %bb.0:
 ; RV64ZBS-NEXT:    lw a0, 0(a0)
 ; RV64ZBS-NEXT:    li a2, 1
 ; RV64ZBS-NEXT:    sllw a1, a2, a1
 ; RV64ZBS-NEXT:    or a0, a1, a0
-; RV64ZBS-NEXT:    sext.w a0, a0
 ; RV64ZBS-NEXT:    ret
   %a = load i32, i32* %p
   %shl = shl i32 1, %b
   %or = or i32 %shl, %a
   ret i32 %or
 }
 
 ; We can use bsetw for 1 << x by setting the first source to zero.

 
 define signext i32 @binv_i32_load(i32* %p, i32 signext %b) nounwind {
 ; RV64I-LABEL: binv_i32_load:
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    lw a0, 0(a0)
 ; RV64I-NEXT:    li a2, 1
 ; RV64I-NEXT:    sllw a1, a2, a1
 ; RV64I-NEXT:    xor a0, a1, a0
-; RV64I-NEXT:    sext.w a0, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64ZBS-LABEL: binv_i32_load:
 ; RV64ZBS:       # %bb.0:
 ; RV64ZBS-NEXT:    lw a0, 0(a0)
 ; RV64ZBS-NEXT:    li a2, 1
 ; RV64ZBS-NEXT:    sllw a1, a2, a1
 ; RV64ZBS-NEXT:    xor a0, a1, a0
-; RV64ZBS-NEXT:    sext.w a0, a0
 ; RV64ZBS-NEXT:    ret
   %a = load i32, i32* %p
   %shl = shl i32 1, %b
   %xor = xor i32 %shl, %a
   ret i32 %xor
 }
 
 define i64 @binv_i64(i64 %a, i64 %b) nounwind {

llvm/test/CodeGen/RISCV/sextw-removal.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=riscv64 -mattr=+m,+f \
+; RUN:   | FileCheck %s --check-prefixes=CHECK,RV64I
+; RUN: llc < %s -mtriple=riscv64 -mattr=+m,+experimental-zbb,+f \
+; RUN:   | FileCheck %s --check-prefixes=CHECK,RV64ZBB
+; RUN: llc < %s -mtriple=riscv64 -mattr=+m,+experimental-zbb,+f \
+; RUN:   -riscv-disable-sextw-removal | FileCheck %s --check-prefix=NOREMOVAL
+
+define void @test1(i32 signext %arg, i32 signext %arg1) nounwind {
+; CHECK-LABEL: test1:
+; CHECK:       # %bb.0: # %bb
+; CHECK-NEXT:    addi sp, sp, -32
+; CHECK-NEXT:    sd ra, 24(sp) # 8-byte Folded Spill
+; CHECK-NEXT:    sd s0, 16(sp) # 8-byte Folded Spill
+; CHECK-NEXT:    sd s1, 8(sp) # 8-byte Folded Spill
+; CHECK-NEXT:    mv s0, a1
+; CHECK-NEXT:    sraw s1, a0, a1
+; CHECK-NEXT:  .LBB0_1: # %bb2
+; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    mv a0, s1
+; CHECK-NEXT:    call bar@plt
+; CHECK-NEXT:    sllw s1, s1, s0
+; CHECK-NEXT:    bnez a0, .LBB0_1
+; CHECK-NEXT:  # %bb.2: # %bb7
+; CHECK-NEXT:    ld ra, 24(sp) # 8-byte Folded Reload
+; CHECK-NEXT:    ld s0, 16(sp) # 8-byte Folded Reload
+; CHECK-NEXT:    ld s1, 8(sp) # 8-byte Folded Reload
+; CHECK-NEXT:    addi sp, sp, 32
+; CHECK-NEXT:    ret
+;
+; NOREMOVAL-LABEL: test1:
+; NOREMOVAL:       # %bb.0: # %bb
+; NOREMOVAL-NEXT:    addi sp, sp, -32
+; NOREMOVAL-NEXT:    sd ra, 24(sp) # 8-byte Folded Spill
+; NOREMOVAL-NEXT:    sd s0, 16(sp) # 8-byte Folded Spill
+; NOREMOVAL-NEXT:    sd s1, 8(sp) # 8-byte Folded Spill
+; NOREMOVAL-NEXT:    mv s0, a1
+; NOREMOVAL-NEXT:    sraw s1, a0, a1
+; NOREMOVAL-NEXT:  .LBB0_1: # %bb2
+; NOREMOVAL-NEXT:    # =>This Inner Loop Header: Depth=1
+; NOREMOVAL-NEXT:    sext.w a0, s1
+; NOREMOVAL-NEXT:    call bar@plt
+; NOREMOVAL-NEXT:    sllw s1, s1, s0
+; NOREMOVAL-NEXT:    bnez a0, .LBB0_1
+; NOREMOVAL-NEXT:  # %bb.2: # %bb7
+; NOREMOVAL-NEXT:    ld ra, 24(sp) # 8-byte Folded Reload
+; NOREMOVAL-NEXT:    ld s0, 16(sp) # 8-byte Folded Reload
+; NOREMOVAL-NEXT:    ld s1, 8(sp) # 8-byte Folded Reload
+; NOREMOVAL-NEXT:    addi sp, sp, 32
+; NOREMOVAL-NEXT:    ret
+bb:
+  %i = ashr i32 %arg, %arg1
+  br label %bb2
+
+bb2:                                              ; preds = %bb2, %bb
+  %i3 = phi i32 [ %i, %bb ], [ %i5, %bb2 ]
+  %i4 = tail call signext i32 @bar(i32 signext %i3)
+  %i5 = shl i32 %i3, %arg1
+  %i6 = icmp eq i32 %i4, 0
+  br i1 %i6, label %bb7, label %bb2
+
+bb7:                                              ; preds = %bb2
+  ret void
+}
+
+declare signext i32 @bar(i32 signext)
+
+; The load here will be an anyext load in isel and sext.w will be emitted for
+; the ret. Make sure we can look through logic ops to prove the sext.w is
+; unnecessary.
+define signext i32 @test2(i32* %p, i32 signext %b) nounwind {
+; RV64I-LABEL: test2:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    lw a0, 0(a0)
+; RV64I-NEXT:    li a2, 1
+; RV64I-NEXT:    sllw a1, a2, a1
+; RV64I-NEXT:    not a1, a1
+; RV64I-NEXT:    and a0, a1, a0
+; RV64I-NEXT:    ret
+;
+; RV64ZBB-LABEL: test2:
+; RV64ZBB:       # %bb.0:
+; RV64ZBB-NEXT:    lw a0, 0(a0)
+; RV64ZBB-NEXT:    li a2, 1
+; RV64ZBB-NEXT:    sllw a1, a2, a1
+; RV64ZBB-NEXT:    andn a0, a0, a1
+; RV64ZBB-NEXT:    ret
+;
+; NOREMOVAL-LABEL: test2:
+; NOREMOVAL:       # %bb.0:
+; NOREMOVAL-NEXT:    lw a0, 0(a0)
+; NOREMOVAL-NEXT:    li a2, 1
+; NOREMOVAL-NEXT:    sllw a1, a2, a1
+; NOREMOVAL-NEXT:    andn a0, a0, a1
+; NOREMOVAL-NEXT:    sext.w a0, a0
+; NOREMOVAL-NEXT:    ret
+  %a = load i32, i32* %p
+  %shl = shl i32 1, %b
+  %neg = xor i32 %shl, -1
+  %and1 = and i32 %neg, %a
+  ret i32 %and1
+}
+
+define signext i32 @test3(i32* %p, i32 signext %b) nounwind {
+; RV64I-LABEL: test3:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    lw a0, 0(a0)
+; RV64I-NEXT:    li a2, 1
+; RV64I-NEXT:    sllw a1, a2, a1
+; RV64I-NEXT:    not a1, a1
+; RV64I-NEXT:    or a0, a1, a0
+; RV64I-NEXT:    ret
+;
+; RV64ZBB-LABEL: test3:
+; RV64ZBB:       # %bb.0:
+; RV64ZBB-NEXT:    lw a0, 0(a0)
+; RV64ZBB-NEXT:    li a2, 1
+; RV64ZBB-NEXT:    sllw a1, a2, a1
+; RV64ZBB-NEXT:    orn a0, a0, a1
+; RV64ZBB-NEXT:    ret
+;
+; NOREMOVAL-LABEL: test3:
+; NOREMOVAL:       # %bb.0:
+; NOREMOVAL-NEXT:    lw a0, 0(a0)
+; NOREMOVAL-NEXT:    li a2, 1
+; NOREMOVAL-NEXT:    sllw a1, a2, a1
+; NOREMOVAL-NEXT:    orn a0, a0, a1
+; NOREMOVAL-NEXT:    sext.w a0, a0
+; NOREMOVAL-NEXT:    ret
+  %a = load i32, i32* %p
+  %shl = shl i32 1, %b
+  %neg = xor i32 %shl, -1
+  %and1 = or i32 %neg, %a
+  ret i32 %and1
+}
+
+define signext i32 @test4(i32* %p, i32 signext %b) nounwind {
+; RV64I-LABEL: test4:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    lw a0, 0(a0)
+; RV64I-NEXT:    li a2, 1
+; RV64I-NEXT:    sllw a1, a2, a1
+; RV64I-NEXT:    xor a0, a1, a0
+; RV64I-NEXT:    not a0, a0
+; RV64I-NEXT:    ret
+;
+; RV64ZBB-LABEL: test4:
+; RV64ZBB:       # %bb.0:
+; RV64ZBB-NEXT:    lw a0, 0(a0)
+; RV64ZBB-NEXT:    li a2, 1
+; RV64ZBB-NEXT:    sllw a1, a2, a1
+; RV64ZBB-NEXT:    xnor a0, a1, a0
+; RV64ZBB-NEXT:    ret
+;
+; NOREMOVAL-LABEL: test4:
+; NOREMOVAL:       # %bb.0:
+; NOREMOVAL-NEXT:    lw a0, 0(a0)
+; NOREMOVAL-NEXT:    li a2, 1
+; NOREMOVAL-NEXT:    sllw a1, a2, a1
+; NOREMOVAL-NEXT:    xnor a0, a1, a0
+; NOREMOVAL-NEXT:    sext.w a0, a0
+; NOREMOVAL-NEXT:    ret
+  %a = load i32, i32* %p
+  %shl = shl i32 1, %b
+  %neg = xor i32 %shl, -1
+  %and1 = xor i32 %neg, %a
+  ret i32 %and1
+}
+
+; Make sure we don't put a sext.w before bar when using cpopw.
+define void @test5(i32 signext %arg, i32 signext %arg1) nounwind {
+; RV64I-LABEL: test5:
+; RV64I:       # %bb.0: # %bb
+; RV64I-NEXT:    addi sp, sp, -48
+; RV64I-NEXT:    sd ra, 40(sp) # 8-byte Folded Spill
+; RV64I-NEXT:    sd s0, 32(sp) # 8-byte Folded Spill
+; RV64I-NEXT:    sd s1, 24(sp) # 8-byte Folded Spill
+; RV64I-NEXT:    sd s2, 16(sp) # 8-byte Folded Spill
+; RV64I-NEXT:    sd s3, 8(sp) # 8-byte Folded Spill
+; RV64I-NEXT:    sraw a0, a0, a1
+; RV64I-NEXT:    lui a1, 349525
+; RV64I-NEXT:    addiw s2, a1, 1365
+; RV64I-NEXT:    lui a1, 209715
+; RV64I-NEXT:    addiw s1, a1, 819
+; RV64I-NEXT:    lui a1, 61681
+; RV64I-NEXT:    addiw s3, a1, -241
+; RV64I-NEXT:    lui a1, 4112
+; RV64I-NEXT:    addiw s0, a1, 257
+; RV64I-NEXT:  .LBB4_1: # %bb2
+; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
+; RV64I-NEXT:    call bar@plt
+; RV64I-NEXT:    mv a1, a0
+; RV64I-NEXT:    srli a0, a0, 1
+; RV64I-NEXT:    and a0, a0, s2
+; RV64I-NEXT:    subw a0, a1, a0
+; RV64I-NEXT:    and a2, a0, s1
+; RV64I-NEXT:    srli a0, a0, 2
+; RV64I-NEXT:    and a0, a0, s1
+; RV64I-NEXT:    add a0, a2, a0
+; RV64I-NEXT:    srli a2, a0, 4
+; RV64I-NEXT:    add a0, a0, a2
+; RV64I-NEXT:    and a0, a0, s3
+; RV64I-NEXT:    mulw a0, a0, s0
+; RV64I-NEXT:    srliw a0, a0, 24
+; RV64I-NEXT:    bnez a1, .LBB4_1
+; RV64I-NEXT:  # %bb.2: # %bb7
+; RV64I-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
+; RV64I-NEXT:    ld s0, 32(sp) # 8-byte Folded Reload
+; RV64I-NEXT:    ld s1, 24(sp) # 8-byte Folded Reload
+; RV64I-NEXT:    ld s2, 16(sp) # 8-byte Folded Reload
+; RV64I-NEXT:    ld s3, 8(sp) # 8-byte Folded Reload
+; RV64I-NEXT:    addi sp, sp, 48
+; RV64I-NEXT:    ret
+;
+; RV64ZBB-LABEL: test5:
+; RV64ZBB:       # %bb.0: # %bb
+; RV64ZBB-NEXT:    addi sp, sp, -16
+; RV64ZBB-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
+; RV64ZBB-NEXT:    sraw a0, a0, a1
+; RV64ZBB-NEXT:  .LBB4_1: # %bb2
+; RV64ZBB-NEXT:    # =>This Inner Loop Header: Depth=1
+; RV64ZBB-NEXT:    call bar@plt
+; RV64ZBB-NEXT:    mv a1, a0
+; RV64ZBB-NEXT:    cpopw a0, a0
+; RV64ZBB-NEXT:    bnez a1, .LBB4_1
+; RV64ZBB-NEXT:  # %bb.2: # %bb7
+; RV64ZBB-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
+; RV64ZBB-NEXT:    addi sp, sp, 16
+; RV64ZBB-NEXT:    ret
+;
+; NOREMOVAL-LABEL: test5:
+; NOREMOVAL:       # %bb.0: # %bb
+; NOREMOVAL-NEXT:    addi sp, sp, -16
+; NOREMOVAL-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
+; NOREMOVAL-NEXT:    sraw a1, a0, a1
+; NOREMOVAL-NEXT:  .LBB4_1: # %bb2
+; NOREMOVAL-NEXT:    # =>This Inner Loop Header: Depth=1
+; NOREMOVAL-NEXT:    sext.w a0, a1
+; NOREMOVAL-NEXT:    call bar@plt
+; NOREMOVAL-NEXT:    cpopw a1, a0
+; NOREMOVAL-NEXT:    bnez a0, .LBB4_1
+; NOREMOVAL-NEXT:  # %bb.2: # %bb7
+; NOREMOVAL-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
+; NOREMOVAL-NEXT:    addi sp, sp, 16
+; NOREMOVAL-NEXT:    ret
+bb:
+  %i = ashr i32 %arg, %arg1
+  br label %bb2
+
+bb2:                                              ; preds = %bb2, %bb
+  %i3 = phi i32 [ %i, %bb ], [ %i5, %bb2 ]
+  %i4 = tail call signext i32 @bar(i32 signext %i3)
+  %i5 = tail call i32 @llvm.ctpop.i32(i32 %i4)
+  %i6 = icmp eq i32 %i4, 0
+  br i1 %i6, label %bb7, label %bb2
+
+bb7:                                              ; preds = %bb2
+  ret void
+}
+
+declare i32 @llvm.ctpop.i32(i32)
+
+define void @test6(i32 signext %arg, i32 signext %arg1) nounwind {
+; CHECK-LABEL: test6:
+; CHECK:       # %bb.0: # %bb
+; CHECK-NEXT:    addi sp, sp, -16
+; CHECK-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
+; CHECK-NEXT:    sraw a0, a0, a1
+; CHECK-NEXT:    fmv.w.x ft0, zero
+; CHECK-NEXT:    fsw ft0, 4(sp) # 4-byte Folded Spill
+; CHECK-NEXT:  .LBB5_1: # %bb2
+; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    call baz@plt
+; CHECK-NEXT:    fmv.w.x ft0, a0
+; CHECK-NEXT:    flw ft1, 4(sp) # 4-byte Folded Reload
+; CHECK-NEXT:    feq.s a1, ft0, ft1
+; CHECK-NEXT:    fcvt.w.s a0, ft0, rtz
+; CHECK-NEXT:    beqz a1, .LBB5_1
+; CHECK-NEXT:  # %bb.2: # %bb7
+; CHECK-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
+; CHECK-NEXT:    addi sp, sp, 16
+; CHECK-NEXT:    ret
+;
+; NOREMOVAL-LABEL: test6:
+; NOREMOVAL:       # %bb.0: # %bb
+; NOREMOVAL-NEXT:    addi sp, sp, -16
+; NOREMOVAL-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
+; NOREMOVAL-NEXT:    sraw a0, a0, a1
+; NOREMOVAL-NEXT:    fmv.w.x ft0, zero
+; NOREMOVAL-NEXT:    fsw ft0, 4(sp) # 4-byte Folded Spill
+; NOREMOVAL-NEXT:  .LBB5_1: # %bb2
+; NOREMOVAL-NEXT:    # =>This Inner Loop Header: Depth=1
+; NOREMOVAL-NEXT:    sext.w a0, a0
+; NOREMOVAL-NEXT:    call baz@plt
+; NOREMOVAL-NEXT:    fmv.w.x ft0, a0
+; NOREMOVAL-NEXT:    flw ft1, 4(sp) # 4-byte Folded Reload
+; NOREMOVAL-NEXT:    feq.s a1, ft0, ft1
+; NOREMOVAL-NEXT:    fcvt.w.s a0, ft0, rtz
+; NOREMOVAL-NEXT:    beqz a1, .LBB5_1
+; NOREMOVAL-NEXT:  # %bb.2: # %bb7
+; NOREMOVAL-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
+; NOREMOVAL-NEXT:    addi sp, sp, 16
+; NOREMOVAL-NEXT:    ret
+bb:
+  %i = ashr i32 %arg, %arg1
+  br label %bb2
+
+bb2:                                              ; preds = %bb2, %bb
+  %i3 = phi i32 [ %i, %bb ], [ %i5, %bb2 ]
+  %i4 = tail call float @baz(i32 signext %i3)
+  %i5 = fptosi float %i4 to i32
+  %i6 = fcmp oeq float %i4, zeroinitializer
+  br i1 %i6, label %bb7, label %bb2
+
+bb7:                                              ; preds = %bb2
+  ret void
+}
+declare float @baz(i32 signext %i3)

llvm/test/CodeGen/RISCV/split-offsets.ll

 ; RV64I:       # %bb.0: # %entry
 ; RV64I-NEXT:    li a3, 0
 ; RV64I-NEXT:    ld a4, 0(a0)
 ; RV64I-NEXT:    lui a0, 20
 ; RV64I-NEXT:    addiw a5, a0, -1920
 ; RV64I-NEXT:    add a0, a1, a5
 ; RV64I-NEXT:    add a1, a4, a5
 ; RV64I-NEXT:    sext.w a2, a2
-; RV64I-NEXT:    sext.w a4, a3
-; RV64I-NEXT:    bge a4, a2, .LBB1_2
+; RV64I-NEXT:    bge a3, a2, .LBB1_2
 ; RV64I-NEXT:  .LBB1_1: # %while_body
 ; RV64I-NEXT:    # =>This Inner Loop Header: Depth=1
 ; RV64I-NEXT:    addiw a4, a3, 1
 ; RV64I-NEXT:    sw a4, 0(a1)
 ; RV64I-NEXT:    sw a3, 4(a1)
 ; RV64I-NEXT:    sw a4, 0(a0)
 ; RV64I-NEXT:    sw a3, 4(a0)
 ; RV64I-NEXT:    mv a3, a4
-; RV64I-NEXT:    sext.w a4, a3
-; RV64I-NEXT:    blt a4, a2, .LBB1_1
+; RV64I-NEXT:    blt a3, a2, .LBB1_1
 ; RV64I-NEXT:  .LBB1_2: # %while_end
 ; RV64I-NEXT:    ret
 entry:
   %s = load [65536 x i32]*, [65536 x i32]** %sp
   br label %while_cond
 while_cond:
   %phi = phi i32 [ 0, %entry ], [ %i, %while_body ]
   %gep0 = getelementptr [65536 x i32], [65536 x i32]* %s, i64 0, i32 20000