[RISCV] Optimize multiplication by constant

[benshi001]

Summary

... to shift/add or shift/sub.

[benshi001]

This patch can not cover all cases (especially "call __mulsi3" on rv32 without M extension), but at least it works well for most cases.

Maybe a better solution is make ISD::MUL as Custom, which I will try later. You are appreciated to review and accept such a partial optimization.

[lenary]

Thanks for the patch!

This optimisation is done by DAGCombine if you instead implement decomposeMulByConstant in RISCVTargetLowering. Read the comment on the TargetLoweringBase class to understand how to use it. This is preferrable, because we don't want to maintain a target-specific copy of this optimisation if we can avoid it.

It would be sensible to base your implementation on the one in the x86 backend: X86TargetLowering::decomposeMulByConstant, which deals with some phase ordering issues around legalisation.

[benshi001]

Thanks. I have uploaded a new version according to your suggestion!

In D82660#2117161, @lenary wrote:

Thanks for the patch!

This optimisation is done by DAGCombine if you instead implement decomposeMulByConstant in RISCVTargetLowering. Read the comment on the TargetLoweringBase class to understand how to use it. This is preferrable, because we don't want to maintain a target-specific copy of this optimisation if we can avoid it.

It would be sensible to base your implementation on the one in the x86 backend: X86TargetLowering::decomposeMulByConstant, which deals with some phase ordering issues around legalisation.

[lenary]

This is looking good.

I'm going to pre-commit the test additions today - if you could rebase your changes on top, that will allow us to see how this change affects the new testcases you added. I'll keep you as the author and let you know the sha so you can rebase on top of the commit.

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

2995

This TODO should not apply to RISC-V, yet.

[lenary]

In D82660#2126537, @lenary wrote:

This is looking good.

I'm going to pre-commit the test additions today - if you could rebase your changes on top, that will allow us to see how this change affects the new testcases you added. I'll keep you as the author and let you know the sha so you can rebase on top of the commit.

Done in rG003a086ffc0.

[benshi001]

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

2995

Thanks. I have rebased and fixed according to what you suggested.

[lenary]

I'm happy with this optimisation where this patch removes multiply libcalls.

Where the target has a the m extension, and especially for 64-bit multiplies on rv32im, I'm not sure this is an optimisation.

I think that, for the moment, we should add a guard to the hook to avoid this transformation where we do have mul instructions:

if (Subtarget.hasStdExtM())
  return false;

What do you think?

llvm/test/CodeGen/RISCV/mul.ll

546–547

I think this is a pessimisation, though I realise that depends on how slow the 32-bit multiplier is compared to add/shift.

[benshi001]

In D82660#2127626, @lenary wrote:

I'm happy with this optimisation where this patch removes multiply libcalls.

Where the target has a the m extension, and especially for 64-bit multiplies on rv32im, I'm not sure this is an optimisation.

I think that, for the moment, we should add a guard to the hook to avoid this transformation where we do have mul instructions:

if (Subtarget.hasStdExtM())
  return false;

What do you think?

Shall we loose the guard condition to that ?

if (!Subtarget.is64Bit && Subtarget.hasStdExtM())
   return false;

This will prevent the optimization for RV32IM, but still work for RV64IM。

I think a mul-instruction's latency is sure to be >=2, so all existing test cases will not have regresion.

[luismarques]

LGTM.
I'm not overly concerned about the occasional code size increases from doing the optimization for RV32IM, so the loosening of the condition is OK IMO.
Everything else seems to be in order now.
Maybe wait a couple of days more for @lenary's OK.

[lenary]

One issue, then I'm happy.

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

2990

getSExtValue will assert if the value does not fit into 64 bits - you need to do a check before you get there.

I think this hook can be called before legalisation, so you may not get only legal types in this call.

llvm/lib/Target/RISCV/RISCVISelLowering.h

public:
 
   bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
                                          Type *Ty) const override {
     return true;
   }
   bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
   bool shouldConsiderGEPOffsetSplit() const override { return true; }
 
+  bool decomposeMulByConstant(LLVMContext &Context, EVT VT,
+                              SDValue C) const override;
+
   TargetLowering::AtomicExpansionKind
   shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
   Value *emitMaskedAtomicRMWIntrinsic(IRBuilder<> &Builder, AtomicRMWInst *AI,
                                       Value *AlignedAddr, Value *Incr,
                                       Value *Mask, Value *ShiftAmt,
                                       AtomicOrdering Ord) const override;
   TargetLowering::AtomicExpansionKind
   shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *CI) const override;

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

 #include "llvm/CodeGen/SelectionDAGISel.h"
 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
 #include "llvm/CodeGen/ValueTypes.h"
 #include "llvm/IR/DiagnosticInfo.h"
 #include "llvm/IR/DiagnosticPrinter.h"
 #include "llvm/IR/IntrinsicsRISCV.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/MathExtras.h"
 #include "llvm/Support/raw_ostream.h"
 
 using namespace llvm;
 
 #define DEBUG_TYPE "riscv-lower"
 
 STATISTIC(NumTailCalls, "Number of tail calls");
 
bool RISCVTargetLowering::shouldExtendTypeInLibCall(EVT Type) const {
   // arguments or return value is f32 type for LP64 ABI.
   RISCVABI::ABI ABI = Subtarget.getTargetABI();
   if (ABI == RISCVABI::ABI_LP64 && (Type == MVT::f32))
     return false;
 
   return true;
 }
 
+bool RISCVTargetLowering::decomposeMulByConstant(LLVMContext &Context, EVT VT,
+                                                 SDValue C) const {
+  // Check integral scalar types.
+  if (VT.isScalarInteger()) {
+    // Do not perform the transformation on riscv32 with the M extension.
+    if (!Subtarget.is64Bit() && Subtarget.hasStdExtM())
+      return false;
+    if (auto *ConstNode = dyn_cast<ConstantSDNode>(C.getNode())) {
+      if (ConstNode->getAPIntValue().getBitWidth() > 8 * sizeof(int64_t))
+        return false;
+      int64_t Imm = ConstNode->getSExtValue();
+      if (isPowerOf2_64(Imm + 1) || isPowerOf2_64(Imm - 1) ||
+          isPowerOf2_64(1 - Imm) || isPowerOf2_64(-1 - Imm))
+        return true;
+    }
+  }
+
+  return false;
+}
+
 #define GET_REGISTER_MATCHER
 #include "RISCVGenAsmMatcher.inc"
 
 Register
 RISCVTargetLowering::getRegisterByName(const char *RegName, LLT VT,
                                        const MachineFunction &MF) const {
   Register Reg = MatchRegisterAltName(RegName);
   if (Reg == RISCV::NoRegister)

llvm/test/CodeGen/RISCV/mul.ll

 ; RV64IM-NEXT:    ret
   %1 = mul i32 %a, %b
   ret i32 %1
 }
 
 define signext i32 @mul_constant(i32 %a) nounwind {
 ; RV32I-LABEL: mul_constant:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    addi a1, zero, 5
-; RV32I-NEXT:    call __mulsi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a1, a0, 2
+; RV32I-NEXT:    add a0, a1, a0
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: mul_constant:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    addi a1, zero, 5
 ; RV32IM-NEXT:    mul a0, a0, a1
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: mul_constant:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp)
-; RV64I-NEXT:    addi a1, zero, 5
-; RV64I-NEXT:    call __muldi3
-; RV64I-NEXT:    sext.w a0, a0
-; RV64I-NEXT:    ld ra, 8(sp)
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 2
+; RV64I-NEXT:    addw a0, a1, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: mul_constant:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    addi a1, zero, 5
-; RV64IM-NEXT:    mulw a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 2
+; RV64IM-NEXT:    addw a0, a1, a0
 ; RV64IM-NEXT:    ret
   %1 = mul i32 %a, 5
   ret i32 %1
 }
 
 define i32 @mul_pow2(i32 %a) nounwind {
 ; RV32I-LABEL: mul_pow2:
 ; RV32I:       # %bb.0:

 ; RV64IM-NEXT:    ret
   %1 = mul i64 %a, %b
   ret i64 %1
 }
 
 define i64 @mul64_constant(i64 %a) nounwind {
 ; RV32I-LABEL: mul64_constant:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    addi a2, zero, 5
-; RV32I-NEXT:    mv a3, zero
-; RV32I-NEXT:    call __muldi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a3, a0, 2
+; RV32I-NEXT:    add a2, a3, a0
+; RV32I-NEXT:    sltu a3, a2, a3
+; RV32I-NEXT:    srli a0, a0, 30
+; RV32I-NEXT:    slli a4, a1, 2
+; RV32I-NEXT:    or a0, a4, a0
+; RV32I-NEXT:    add a0, a0, a1
+; RV32I-NEXT:    add a1, a0, a3
+; RV32I-NEXT:    mv a0, a2
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: mul64_constant:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    addi a2, zero, 5
 ; RV32IM-NEXT:    mul a1, a1, a2
 ; RV32IM-NEXT:    mulhu a3, a0, a2
 ; RV32IM-NEXT:    add a1, a3, a1
 ; RV32IM-NEXT:    mul a0, a0, a2
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: mul64_constant:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp)
-; RV64I-NEXT:    addi a1, zero, 5
-; RV64I-NEXT:    call __muldi3
-; RV64I-NEXT:    ld ra, 8(sp)
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 2
+; RV64I-NEXT:    add a0, a1, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: mul64_constant:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    addi a1, zero, 5
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 2
+; RV64IM-NEXT:    add a0, a1, a0
 ; RV64IM-NEXT:    ret
   %1 = mul i64 %a, 5
   ret i64 %1
 }
 
 define i32 @mulhs(i32 %a, i32 %b) nounwind {
 ; RV32I-LABEL: mulhs:
 ; RV32I:       # %bb.0:
; RV64IM-NEXT:    ret
   %4 = lshr i64 %3, 32
   %5 = trunc i64 %4 to i32
   ret i32 %5
 }
 
 define i32 @muli32_p65(i32 %a) nounwind {
 ; RV32I-LABEL: muli32_p65:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    addi a1, zero, 65
-; RV32I-NEXT:    call __mulsi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a1, a0, 6
+; RV32I-NEXT:    add a0, a1, a0
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli32_p65:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    addi a1, zero, 65
 ; RV32IM-NEXT:    mul a0, a0, a1
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli32_p65:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp)
-; RV64I-NEXT:    addi a1, zero, 65
-; RV64I-NEXT:    call __muldi3
-; RV64I-NEXT:    ld ra, 8(sp)
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 6
+; RV64I-NEXT:    addw a0, a1, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli32_p65:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    addi a1, zero, 65
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 6
+; RV64IM-NEXT:    addw a0, a1, a0
 ; RV64IM-NEXT:    ret
   %1 = mul i32 %a, 65
   ret i32 %1
 }
 
 define i32 @muli32_p63(i32 %a) nounwind {
 ; RV32I-LABEL: muli32_p63:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    addi a1, zero, 63
-; RV32I-NEXT:    call __mulsi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a1, a0, 6
+; RV32I-NEXT:    sub a0, a1, a0
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli32_p63:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    addi a1, zero, 63
 ; RV32IM-NEXT:    mul a0, a0, a1
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli32_p63:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp)
-; RV64I-NEXT:    addi a1, zero, 63
-; RV64I-NEXT:    call __muldi3
-; RV64I-NEXT:    ld ra, 8(sp)
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 6
+; RV64I-NEXT:    subw a0, a1, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli32_p63:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    addi a1, zero, 63
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 6
+; RV64IM-NEXT:    subw a0, a1, a0
 ; RV64IM-NEXT:    ret
   %1 = mul i32 %a, 63
   ret i32 %1
 }
 
 define i64 @muli64_p65(i64 %a) nounwind {
 ; RV32I-LABEL: muli64_p65:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    addi a2, zero, 65
-; RV32I-NEXT:    mv a3, zero
-; RV32I-NEXT:    call __muldi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a3, a0, 6
+; RV32I-NEXT:    add a2, a3, a0
+; RV32I-NEXT:    sltu a3, a2, a3
+; RV32I-NEXT:    srli a0, a0, 26
+; RV32I-NEXT:    slli a4, a1, 6
+; RV32I-NEXT:    or a0, a4, a0
+; RV32I-NEXT:    add a0, a0, a1
+; RV32I-NEXT:    add a1, a0, a3
+; RV32I-NEXT:    mv a0, a2
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli64_p65:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    addi a2, zero, 65
 ; RV32IM-NEXT:    mul a1, a1, a2
 ; RV32IM-NEXT:    mulhu a3, a0, a2
 ; RV32IM-NEXT:    add a1, a3, a1
 ; RV32IM-NEXT:    mul a0, a0, a2
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli64_p65:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp)
-; RV64I-NEXT:    addi a1, zero, 65
-; RV64I-NEXT:    call __muldi3
-; RV64I-NEXT:    ld ra, 8(sp)
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 6
+; RV64I-NEXT:    add a0, a1, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli64_p65:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    addi a1, zero, 65
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 6
+; RV64IM-NEXT:    add a0, a1, a0
 ; RV64IM-NEXT:    ret
   %1 = mul i64 %a, 65
   ret i64 %1
 }
 
 define i64 @muli64_p63(i64 %a) nounwind {
 ; RV32I-LABEL: muli64_p63:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    addi a2, zero, 63
-; RV32I-NEXT:    mv a3, zero
-; RV32I-NEXT:    call __muldi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a2, a0, 6
+; RV32I-NEXT:    sltu a3, a2, a0
+; RV32I-NEXT:    srli a4, a0, 26
+; RV32I-NEXT:    slli a5, a1, 6
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    sub a1, a4, a1
+; RV32I-NEXT:    sub a1, a1, a3
+; RV32I-NEXT:    sub a0, a2, a0
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli64_p63:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    addi a2, zero, 63
 ; RV32IM-NEXT:    mul a1, a1, a2
 ; RV32IM-NEXT:    mulhu a3, a0, a2
 ; RV32IM-NEXT:    add a1, a3, a1
 ; RV32IM-NEXT:    mul a0, a0, a2
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli64_p63:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp)
-; RV64I-NEXT:    addi a1, zero, 63
-; RV64I-NEXT:    call __muldi3
-; RV64I-NEXT:    ld ra, 8(sp)
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 6
+; RV64I-NEXT:    sub a0, a1, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli64_p63:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    addi a1, zero, 63
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 6
+; RV64IM-NEXT:    sub a0, a1, a0
 ; RV64IM-NEXT:    ret
   %1 = mul i64 %a, 63
   ret i64 %1
 }
 
 define i32 @muli32_m63(i32 %a) nounwind {
 ; RV32I-LABEL: muli32_m63:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    addi a1, zero, -63
-; RV32I-NEXT:    call __mulsi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a1, a0, 6
+; RV32I-NEXT:    sub a0, a0, a1
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli32_m63:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    addi a1, zero, -63
 ; RV32IM-NEXT:    mul a0, a0, a1
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli32_m63:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp)
-; RV64I-NEXT:    addi a1, zero, -63
-; RV64I-NEXT:    call __muldi3
-; RV64I-NEXT:    ld ra, 8(sp)
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 6
+; RV64I-NEXT:    subw a0, a0, a1
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli32_m63:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    addi a1, zero, -63
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 6
+; RV64IM-NEXT:    subw a0, a0, a1
 ; RV64IM-NEXT:    ret
   %1 = mul i32 %a, -63
   ret i32 %1
 }
 
 define i32 @muli32_m65(i32 %a) nounwind {
 ; RV32I-LABEL: muli32_m65:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    addi a1, zero, -65
-; RV32I-NEXT:    call __mulsi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a1, a0, 6
+; RV32I-NEXT:    add a0, a1, a0
+; RV32I-NEXT:    neg a0, a0
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli32_m65:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    addi a1, zero, -65
 ; RV32IM-NEXT:    mul a0, a0, a1
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli32_m65:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp)
-; RV64I-NEXT:    addi a1, zero, -65
-; RV64I-NEXT:    call __muldi3
-; RV64I-NEXT:    ld ra, 8(sp)
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 6
+; RV64I-NEXT:    add a0, a1, a0
+; RV64I-NEXT:    negw a0, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli32_m65:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    addi a1, zero, -65
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 6
+; RV64IM-NEXT:    add a0, a1, a0
+; RV64IM-NEXT:    negw a0, a0
 ; RV64IM-NEXT:    ret
   %1 = mul i32 %a, -65
   ret i32 %1
 }
 
 define i64 @muli64_m63(i64 %a) nounwind {
 ; RV32I-LABEL: muli64_m63:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    addi a2, zero, -63
-; RV32I-NEXT:    addi a3, zero, -1
-; RV32I-NEXT:    call __muldi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a2, a0, 6
+; RV32I-NEXT:    sltu a3, a0, a2
+; RV32I-NEXT:    srli a4, a0, 26
+; RV32I-NEXT:    slli a5, a1, 6
+; RV32I-NEXT:    or a4, a5, a4
+; RV32I-NEXT:    sub a1, a1, a4
+; RV32I-NEXT:    sub a1, a1, a3
+; RV32I-NEXT:    sub a0, a0, a2
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli64_m63:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    addi a2, zero, -63
 ; RV32IM-NEXT:    mul a1, a1, a2
 ; RV32IM-NEXT:    mulhu a3, a0, a2
 ; RV32IM-NEXT:    sub a3, a3, a0
 ; RV32IM-NEXT:    add a1, a3, a1
 ; RV32IM-NEXT:    mul a0, a0, a2
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli64_m63:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp)
-; RV64I-NEXT:    addi a1, zero, -63
-; RV64I-NEXT:    call __muldi3
-; RV64I-NEXT:    ld ra, 8(sp)
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 6
+; RV64I-NEXT:    sub a0, a0, a1
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli64_m63:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    addi a1, zero, -63
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 6
+; RV64IM-NEXT:    sub a0, a0, a1
 ; RV64IM-NEXT:    ret
   %1 = mul i64 %a, -63
   ret i64 %1
 }
 
 define i64 @muli64_m65(i64 %a) nounwind {
 ; RV32I-LABEL: muli64_m65:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    addi a2, zero, -65
-; RV32I-NEXT:    addi a3, zero, -1
-; RV32I-NEXT:    call __muldi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a2, a0, 6
+; RV32I-NEXT:    add a3, a2, a0
+; RV32I-NEXT:    sltu a2, a3, a2
+; RV32I-NEXT:    srli a0, a0, 26
+; RV32I-NEXT:    slli a4, a1, 6
+; RV32I-NEXT:    or a0, a4, a0
+; RV32I-NEXT:    add a0, a0, a1
+; RV32I-NEXT:    add a0, a0, a2
+; RV32I-NEXT:    snez a1, a3
+; RV32I-NEXT:    add a0, a0, a1
+; RV32I-NEXT:    neg a1, a0
+; RV32I-NEXT:    neg a0, a3
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli64_m65:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    addi a2, zero, -65
 ; RV32IM-NEXT:    mul a1, a1, a2
 ; RV32IM-NEXT:    mulhu a3, a0, a2
 ; RV32IM-NEXT:    sub a3, a3, a0
 ; RV32IM-NEXT:    add a1, a3, a1
 ; RV32IM-NEXT:    mul a0, a0, a2
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli64_m65:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp)
-; RV64I-NEXT:    addi a1, zero, -65
-; RV64I-NEXT:    call __muldi3
-; RV64I-NEXT:    ld ra, 8(sp)
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 6
+; RV64I-NEXT:    add a0, a1, a0
+; RV64I-NEXT:    neg a0, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli64_m65:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    addi a1, zero, -65
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 6
+; RV64IM-NEXT:    add a0, a1, a0
+; RV64IM-NEXT:    neg a0, a0
 ; RV64IM-NEXT:    ret
   %1 = mul i64 %a, -65
   ret i64 %1
 }