[RISCV] Custom lower SHL_PARTS, SRA_PARTS, SRL_PARTS

[luismarques]

Summary

When not optimizing for minimum size (-Oz) we custom lower wide shifts
(SHL_PARTS, SRA_PARTS, SRL_PARTS) instead of expanding to a libcall. Although
conceptually independent from it, this patch builds upon the D59355 optimization
of SELECT sequences to obtain the optimized code exhibited in the updated tests.

[asb]

Thanks for this Luis. This needs coverage for the optForMinSize checks. I'd recommend extending shifts.ll with minsize versions of the tests, e.g.

define i64 @lshr64_minsize(i64 %a, i64 %b) minsize nounwind {                             
; RV32I-LABEL: lshr64_minsize:                                                            
; RV32I:       # %bb.0:                                                                   
; RV32I-NEXT:    addi sp, sp, -16                                                         
; RV32I-NEXT:    sw ra, 12(sp)                                                            
; RV32I-NEXT:    call __lshrdi3                                                           
; RV32I-NEXT:    lw ra, 12(sp)                                                            
; RV32I-NEXT:    addi sp, sp, 16                                                          
; RV32I-NEXT:    ret                                                                      
  %1 = lshr i64 %a, %b                                                                    
  ret i64 %1                                                                              
}

[luismarques]

Updates the patch to:

• Add a test covering the optForMinSize option
• Add a test that shifts a 128-bit value
• Add checks for RV64 (relevant for the 128-bit test)
• Fix the materialization of the -XLEN constant in RV64

[asb]

I think it would be worth rounding out the tests a bit, especially as it should just require a trivial amount of copy and paste and editing to do so:

• ashr64_minsize
• shl64_minsize
• lshr128
• ashr128

[lewis-revill]

lib/Target/RISCV/RISCVISelLowering.cpp

587

I'm sure there's an off-by-one error in my reasoning here, but aren't we already safe to assume that XLEN-Shamt can fit in the bits available for srl? Since we checked for if Shamt-XLEN < 0, the only case where XLEN-Shamt couldn't fit would be where XLEN-Shamt == XLEN IE when Shamt-XLEN == 0, which is handled by the basic case below.

I'm assuming this was the reasoning for the XLEN-1 stuff and the (Lo >>u 1) here, if not can someone explain it to me?

590

Nitpick: shouldn't this read Hi = Lo << (Shamt-XLEN)

[lewis-revill]

lib/Target/RISCV/RISCVISelLowering.cpp

587

I see at least part of my error. It should read 'IE when Shamt == 0`. Either way I'm not seeing how it should come up?

[lewis-revill]

lib/Target/RISCV/RISCVISelLowering.cpp

587

Okay sorry for the noise, I see why this is an issue now. Shamt == 0 is indeed handled by this case. Could we have some comment to that effect in this function? And likewise for lowerShiftRightParts?

[luismarques]

lib/Target/RISCV/RISCVISelLowering.cpp

587

I think the comment "// Shamt < XLEN" documents reasonably well that this branch also covers Shamt == 0, since "0 < XLEN". If I misunderstood your comment please clarify.

590

The (Shamt-XLEN) should have been in the "else" branch, not the "then" branch. These kind of mistakes were easy to make because the assembly has the "basic blocks" in the reverse order (I can't fix that in the pseudo-code without inverting the condition, which would be even more confusing). I had caught most of them, but that one slipped through. Good catch! I'll fix that together with the remaining issues.

[luismarques]

• Fixed a bug in the pseudo-code for lowerShiftLeftParts;
• Added more tests, which now cover all the 64-bit shift variants, with and without the minsize attribute, plus all of the 128-bit shift variants.

[asb]

This looks good to me. The GCC torture suite is happy with this code too, which serves as an extra (but non-exhaustive) sanity check. It's not immediately obvious to me that checking for OptForMinSize is better than OptForSize, but the precise criteria for when something should be done at Oz vs Os don't seem to be well specified and checking OptForMinSize is consistent with other backends, so the right choice for now.

lib/Target/RISCV/RISCVISelLowering.h

Instruction *emitLeadingFence(IRBuilder<> &Builder, Instruction *Inst,
                                 AtomicOrdering Ord) const override;
   Instruction *emitTrailingFence(IRBuilder<> &Builder, Instruction *Inst,
                                  AtomicOrdering Ord) const override;
 
   ISD::NodeType getExtendForAtomicOps() const override {
     return ISD::SIGN_EXTEND;
   }
 
+  bool shouldExpandShift(SelectionDAG &DAG, SDNode *N) const override {
+    if (DAG.getMachineFunction().getFunction().hasMinSize())
+      return false;
+    return true;
+  }
+
 private:
   void analyzeInputArgs(MachineFunction &MF, CCState &CCInfo,
                         const SmallVectorImpl<ISD::InputArg> &Ins,
                         bool IsRet) const;
   void analyzeOutputArgs(MachineFunction &MF, CCState &CCInfo,
                          const SmallVectorImpl<ISD::OutputArg> &Outs,
                          bool IsRet, CallLoweringInfo *CLI) const;
   // Lower incoming arguments, copy physregs into vregs
private:
 
   SDValue lowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerSELECT(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerVASTART(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
+  SDValue lowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
+  SDValue lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, bool IsSRA) const;
 
   bool isEligibleForTailCallOptimization(
       CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF,
       const SmallVector<CCValAssign, 16> &ArgLocs) const;
 
   TargetLowering::AtomicExpansionKind
   shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override;
   virtual Value *emitMaskedAtomicRMWIntrinsic(

lib/Target/RISCV/RISCVISelLowering.cpp

if (Subtarget.is64Bit() && Subtarget.hasStdExtM()) {
     setOperationAction(ISD::UREM, MVT::i32, Custom);
   }
 
   setOperationAction(ISD::SDIVREM, XLenVT, Expand);
   setOperationAction(ISD::UDIVREM, XLenVT, Expand);
   setOperationAction(ISD::SMUL_LOHI, XLenVT, Expand);
   setOperationAction(ISD::UMUL_LOHI, XLenVT, Expand);
 
-  setOperationAction(ISD::SHL_PARTS, XLenVT, Expand);
-  setOperationAction(ISD::SRL_PARTS, XLenVT, Expand);
-  setOperationAction(ISD::SRA_PARTS, XLenVT, Expand);
+  setOperationAction(ISD::SHL_PARTS, XLenVT, Custom);
+  setOperationAction(ISD::SRL_PARTS, XLenVT, Custom);
+  setOperationAction(ISD::SRA_PARTS, XLenVT, Custom);
 
   setOperationAction(ISD::ROTL, XLenVT, Expand);
   setOperationAction(ISD::ROTR, XLenVT, Expand);
   setOperationAction(ISD::BSWAP, XLenVT, Expand);
   setOperationAction(ISD::CTTZ, XLenVT, Expand);
   setOperationAction(ISD::CTLZ, XLenVT, Expand);
   setOperationAction(ISD::CTPOP, XLenVT, Expand);
 
SDValue RISCVTargetLowering::LowerOperation(SDValue Op,
   case ISD::SELECT:
     return lowerSELECT(Op, DAG);
   case ISD::VASTART:
     return lowerVASTART(Op, DAG);
   case ISD::FRAMEADDR:
     return lowerFRAMEADDR(Op, DAG);
   case ISD::RETURNADDR:
     return lowerRETURNADDR(Op, DAG);
+  case ISD::SHL_PARTS:
+    return lowerShiftLeftParts(Op, DAG);
+  case ISD::SRA_PARTS:
+    return lowerShiftRightParts(Op, DAG, true);
+  case ISD::SRL_PARTS:
+    return lowerShiftRightParts(Op, DAG, false);
   case ISD::BITCAST: {
     assert(Subtarget.is64Bit() && Subtarget.hasStdExtF() &&
            "Unexpected custom legalisation");
     SDLoc DL(Op);
     SDValue Op0 = Op.getOperand(0);
     if (Op.getValueType() != MVT::f32 || Op0.getValueType() != MVT::i32)
       return SDValue();
     SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0);
SDValue RISCVTargetLowering::lowerRETURNADDR(SDValue Op,
   }
 
   // Return the value of the return address register, marking it an implicit
   // live-in.
   unsigned Reg = MF.addLiveIn(RI.getRARegister(), getRegClassFor(XLenVT));
   return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, XLenVT);
 }
 
+SDValue RISCVTargetLowering::lowerShiftLeftParts(SDValue Op,
+                                                 SelectionDAG &DAG) const {
+  SDLoc DL(Op);
+  SDValue Lo = Op.getOperand(0);
+  SDValue Hi = Op.getOperand(1);
+  SDValue Shamt = Op.getOperand(2);
+  EVT VT = Lo.getValueType();
+
+  // if Shamt-XLEN < 0: // Shamt < XLEN
+  //   Lo = Lo << Shamt
+  //   Hi = (Hi << Shamt) | ((Lo >>u 1) >>u (XLEN-1 - Shamt))
+  // else:
+  //   Lo = 0
+  //   Hi = Lo << (Shamt-XLEN)
+
+  SDValue Zero = DAG.getConstant(0, DL, VT);
+  SDValue One = DAG.getConstant(1, DL, VT);
+  SDValue MinusXLen = DAG.getConstant(-(int)Subtarget.getXLen(), DL, VT);
+  SDValue XLenMinus1 = DAG.getConstant(Subtarget.getXLen() - 1, DL, VT);
+  SDValue ShamtMinusXLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusXLen);
+  SDValue XLenMinus1Shamt = DAG.getNode(ISD::SUB, DL, VT, XLenMinus1, Shamt);
+
+  SDValue LoTrue = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt);
+  SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo, One);
+  SDValue ShiftRightLo =
+      DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, XLenMinus1Shamt);
+  SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt);
+  SDValue HiTrue = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
+  SDValue HiFalse = DAG.getNode(ISD::SHL, DL, VT, Lo, ShamtMinusXLen);
+
+  SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusXLen, Zero, ISD::SETLT);
+
+  Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, Zero);
+  Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse);
+
+  SDValue Parts[2] = {Lo, Hi};
+  return DAG.getMergeValues(Parts, DL);
+}
+
+SDValue RISCVTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
+                                                  bool IsSRA) const {
+  SDLoc DL(Op);
+  SDValue Lo = Op.getOperand(0);
+  SDValue Hi = Op.getOperand(1);
+  SDValue Shamt = Op.getOperand(2);
+  EVT VT = Lo.getValueType();
+
+  // SRA expansion:
+  //   if Shamt-XLEN < 0: // Shamt < XLEN
+  //     Lo = (Lo >>u Shamt) | ((Hi << 1) << (XLEN-1 - Shamt))
+  //     Hi = Hi >>s Shamt
+  //   else:
+  //     Lo = Hi >>s (Shamt-XLEN);
+  //     Hi = Hi >>s (XLEN-1)
+  //
+  // SRL expansion:
+  //   if Shamt-XLEN < 0: // Shamt < XLEN
+  //     Lo = (Lo >>u Shamt) | ((Hi << 1) << (XLEN-1 - Shamt))
+  //     Hi = Hi >>u Shamt
+  //   else:
+  //     Lo = Hi >>u (Shamt-XLEN);
+  //     Hi = 0;
+
+  unsigned ShiftRightOp = IsSRA ? ISD::SRA : ISD::SRL;
+
+  SDValue Zero = DAG.getConstant(0, DL, VT);
+  SDValue One = DAG.getConstant(1, DL, VT);
+  SDValue MinusXLen = DAG.getConstant(-(int)Subtarget.getXLen(), DL, VT);
+  SDValue XLenMinus1 = DAG.getConstant(Subtarget.getXLen() - 1, DL, VT);
+  SDValue ShamtMinusXLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusXLen);
+  SDValue XLenMinus1Shamt = DAG.getNode(ISD::SUB, DL, VT, XLenMinus1, Shamt);
+
+  SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt);
+  SDValue ShiftLeftHi1 = DAG.getNode(ISD::SHL, DL, VT, Hi, One);
+  SDValue ShiftLeftHi =
+      DAG.getNode(ISD::SHL, DL, VT, ShiftLeftHi1, XLenMinus1Shamt);
+  SDValue LoTrue = DAG.getNode(ISD::OR, DL, VT, ShiftRightLo, ShiftLeftHi);
+  SDValue HiTrue = DAG.getNode(ShiftRightOp, DL, VT, Hi, Shamt);
+  SDValue LoFalse = DAG.getNode(ShiftRightOp, DL, VT, Hi, ShamtMinusXLen);
+  SDValue HiFalse =
+      IsSRA ? DAG.getNode(ISD::SRA, DL, VT, Hi, XLenMinus1) : Zero;
+
+  SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusXLen, Zero, ISD::SETLT);
+
+  Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, LoFalse);
+  Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse);
+
+  SDValue Parts[2] = {Lo, Hi};
+  return DAG.getMergeValues(Parts, DL);
+}
+
 // Returns the opcode of the target-specific SDNode that implements the 32-bit
 // form of the given Opcode.
 static RISCVISD::NodeType getRISCVWOpcode(unsigned Opcode) {
   switch (Opcode) {
   default:
     llvm_unreachable("Unexpected opcode");
   case ISD::SHL:
     return RISCVISD::SLLW;

test/CodeGen/RISCV/alu64.ll

 define i64 @sll(i64 %a, i64 %b) nounwind {
 ; RV64I-LABEL: sll:
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    sll a0, a0, a1
 ; RV64I-NEXT:    ret
 ;
 ; RV32I-LABEL: sll:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    call __ashldi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    addi a3, a2, -32
+; RV32I-NEXT:    bltz a3, .LBB11_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    sll a1, a0, a3
+; RV32I-NEXT:    mv a0, zero
+; RV32I-NEXT:    ret
+; RV32I-NEXT:  .LBB11_2:
+; RV32I-NEXT:    addi a3, zero, 31
+; RV32I-NEXT:    sub a3, a3, a2
+; RV32I-NEXT:    srli a4, a0, 1
+; RV32I-NEXT:    srl a3, a4, a3
+; RV32I-NEXT:    sll a1, a1, a2
+; RV32I-NEXT:    or a1, a1, a3
+; RV32I-NEXT:    sll a0, a0, a2
 ; RV32I-NEXT:    ret
   %1 = shl i64 %a, %b
   ret i64 %1
 }
 
 define i64 @slt(i64 %a, i64 %b) nounwind {
 ; RV64I-LABEL: slt:
 ; RV64I:       # %bb.0:

 define i64 @srl(i64 %a, i64 %b) nounwind {
 ; RV64I-LABEL: srl:
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    srl a0, a0, a1
 ; RV64I-NEXT:    ret
 ;
 ; RV32I-LABEL: srl:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    call __lshrdi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    addi a3, a2, -32
+; RV32I-NEXT:    bltz a3, .LBB15_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    srl a0, a1, a3
+; RV32I-NEXT:    mv a1, zero
+; RV32I-NEXT:    ret
+; RV32I-NEXT:  .LBB15_2:
+; RV32I-NEXT:    addi a3, zero, 31
+; RV32I-NEXT:    sub a3, a3, a2
+; RV32I-NEXT:    slli a4, a1, 1
+; RV32I-NEXT:    sll a3, a4, a3
+; RV32I-NEXT:    srl a0, a0, a2
+; RV32I-NEXT:    or a0, a0, a3
+; RV32I-NEXT:    srl a1, a1, a2
 ; RV32I-NEXT:    ret
   %1 = lshr i64 %a, %b
   ret i64 %1
 }
 
 define i64 @sra(i64 %a, i64 %b) nounwind {
 ; RV64I-LABEL: sra:
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    sra a0, a0, a1
 ; RV64I-NEXT:    ret
 ;
 ; RV32I-LABEL: sra:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp)
-; RV32I-NEXT:    call __ashrdi3
-; RV32I-NEXT:    lw ra, 12(sp)
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    addi a3, a2, -32
+; RV32I-NEXT:    bltz a3, .LBB16_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    sra a0, a1, a3
+; RV32I-NEXT:    srai a1, a1, 31
+; RV32I-NEXT:    ret
+; RV32I-NEXT:  .LBB16_2:
+; RV32I-NEXT:    addi a3, zero, 31
+; RV32I-NEXT:    sub a3, a3, a2
+; RV32I-NEXT:    slli a4, a1, 1
+; RV32I-NEXT:    sll a3, a4, a3
+; RV32I-NEXT:    srl a0, a0, a2
+; RV32I-NEXT:    or a0, a0, a3
+; RV32I-NEXT:    sra a1, a1, a2
 ; RV32I-NEXT:    ret
   %1 = ashr i64 %a, %b
   ret i64 %1
 }
 
 define i64 @or(i64 %a, i64 %b) nounwind {
 ; RV64I-LABEL: or:
 ; RV64I:       # %bb.0:

test/CodeGen/RISCV/shifts.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=riscv32 -verify-machineinstrs < %s \
 ; RUN:   | FileCheck %s -check-prefix=RV32I
+; RUN: llc -mtriple=riscv64 -verify-machineinstrs < %s \
+; RUN:   | FileCheck %s -check-prefix=RV64I
 
 ; Basic shift support is tested as part of ALU.ll. This file ensures that
 ; shifts which may not be supported natively are lowered properly.
 
 define i64 @lshr64(i64 %a, i64 %b) nounwind {
 ; RV32I-LABEL: lshr64:
 ; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi a3, a2, -32
+; RV32I-NEXT:    bltz a3, .LBB0_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    srl a0, a1, a3
+; RV32I-NEXT:    mv a1, zero
+; RV32I-NEXT:    ret
+; RV32I-NEXT:  .LBB0_2:
+; RV32I-NEXT:    addi a3, zero, 31
+; RV32I-NEXT:    sub a3, a3, a2
+; RV32I-NEXT:    slli a4, a1, 1
+; RV32I-NEXT:    sll a3, a4, a3
+; RV32I-NEXT:    srl a0, a0, a2
+; RV32I-NEXT:    or a0, a0, a3
+; RV32I-NEXT:    srl a1, a1, a2
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: lshr64:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    srl a0, a0, a1
+; RV64I-NEXT:    ret
+  %1 = lshr i64 %a, %b
+  ret i64 %1
+}
+
+define i64 @lshr64_minsize(i64 %a, i64 %b) minsize nounwind {
+; RV32I-LABEL: lshr64_minsize:
+; RV32I:       # %bb.0:
 ; RV32I-NEXT:    addi sp, sp, -16
 ; RV32I-NEXT:    sw ra, 12(sp)
 ; RV32I-NEXT:    call __lshrdi3
 ; RV32I-NEXT:    lw ra, 12(sp)
 ; RV32I-NEXT:    addi sp, sp, 16
 ; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: lshr64_minsize:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    srl a0, a0, a1
+; RV64I-NEXT:    ret
   %1 = lshr i64 %a, %b
   ret i64 %1
 }
 
 define i64 @ashr64(i64 %a, i64 %b) nounwind {
 ; RV32I-LABEL: ashr64:
 ; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi a3, a2, -32
+; RV32I-NEXT:    bltz a3, .LBB2_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    sra a0, a1, a3
+; RV32I-NEXT:    srai a1, a1, 31
+; RV32I-NEXT:    ret
+; RV32I-NEXT:  .LBB2_2:
+; RV32I-NEXT:    addi a3, zero, 31
+; RV32I-NEXT:    sub a3, a3, a2
+; RV32I-NEXT:    slli a4, a1, 1
+; RV32I-NEXT:    sll a3, a4, a3
+; RV32I-NEXT:    srl a0, a0, a2
+; RV32I-NEXT:    or a0, a0, a3
+; RV32I-NEXT:    sra a1, a1, a2
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: ashr64:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    sra a0, a0, a1
+; RV64I-NEXT:    ret
+  %1 = ashr i64 %a, %b
+  ret i64 %1
+}
+
+define i64 @ashr64_minsize(i64 %a, i64 %b) minsize nounwind {
+; RV32I-LABEL: ashr64_minsize:
+; RV32I:       # %bb.0:
 ; RV32I-NEXT:    addi sp, sp, -16
 ; RV32I-NEXT:    sw ra, 12(sp)
 ; RV32I-NEXT:    call __ashrdi3
 ; RV32I-NEXT:    lw ra, 12(sp)
 ; RV32I-NEXT:    addi sp, sp, 16
 ; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: ashr64_minsize:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    sra a0, a0, a1
+; RV64I-NEXT:    ret
   %1 = ashr i64 %a, %b
   ret i64 %1
 }
 
 define i64 @shl64(i64 %a, i64 %b) nounwind {
 ; RV32I-LABEL: shl64:
 ; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi a3, a2, -32
+; RV32I-NEXT:    bltz a3, .LBB4_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    sll a1, a0, a3
+; RV32I-NEXT:    mv a0, zero
+; RV32I-NEXT:    ret
+; RV32I-NEXT:  .LBB4_2:
+; RV32I-NEXT:    addi a3, zero, 31
+; RV32I-NEXT:    sub a3, a3, a2
+; RV32I-NEXT:    srli a4, a0, 1
+; RV32I-NEXT:    srl a3, a4, a3
+; RV32I-NEXT:    sll a1, a1, a2
+; RV32I-NEXT:    or a1, a1, a3
+; RV32I-NEXT:    sll a0, a0, a2
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: shl64:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    sll a0, a0, a1
+; RV64I-NEXT:    ret
+  %1 = shl i64 %a, %b
+  ret i64 %1
+}
+
+define i64 @shl64_minsize(i64 %a, i64 %b) minsize nounwind {
+; RV32I-LABEL: shl64_minsize:
+; RV32I:       # %bb.0:
 ; RV32I-NEXT:    addi sp, sp, -16
 ; RV32I-NEXT:    sw ra, 12(sp)
 ; RV32I-NEXT:    call __ashldi3
 ; RV32I-NEXT:    lw ra, 12(sp)
 ; RV32I-NEXT:    addi sp, sp, 16
 ; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: shl64_minsize:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    sll a0, a0, a1
+; RV64I-NEXT:    ret
   %1 = shl i64 %a, %b
   ret i64 %1
 }
+
+define i128 @lshr128(i128 %a, i128 %b) nounwind {
+; RV32I-LABEL: lshr128:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 44(sp)
+; RV32I-NEXT:    sw s0, 40(sp)
+; RV32I-NEXT:    mv s0, a0
+; RV32I-NEXT:    lw a0, 12(a1)
+; RV32I-NEXT:    sw a0, 20(sp)
+; RV32I-NEXT:    lw a0, 8(a1)
+; RV32I-NEXT:    sw a0, 16(sp)
+; RV32I-NEXT:    lw a0, 4(a1)
+; RV32I-NEXT:    sw a0, 12(sp)
+; RV32I-NEXT:    lw a0, 0(a1)
+; RV32I-NEXT:    sw a0, 8(sp)
+; RV32I-NEXT:    lw a2, 0(a2)
+; RV32I-NEXT:    addi a0, sp, 24
+; RV32I-NEXT:    addi a1, sp, 8
+; RV32I-NEXT:    call __lshrti3
+; RV32I-NEXT:    lw a0, 36(sp)
+; RV32I-NEXT:    sw a0, 12(s0)
+; RV32I-NEXT:    lw a0, 32(sp)
+; RV32I-NEXT:    sw a0, 8(s0)
+; RV32I-NEXT:    lw a0, 28(sp)
+; RV32I-NEXT:    sw a0, 4(s0)
+; RV32I-NEXT:    lw a0, 24(sp)
+; RV32I-NEXT:    sw a0, 0(s0)
+; RV32I-NEXT:    lw s0, 40(sp)
+; RV32I-NEXT:    lw ra, 44(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: lshr128:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    addi a3, a2, -64
+; RV64I-NEXT:    bltz a3, .LBB6_2
+; RV64I-NEXT:  # %bb.1:
+; RV64I-NEXT:    srl a0, a1, a3
+; RV64I-NEXT:    mv a1, zero
+; RV64I-NEXT:    ret
+; RV64I-NEXT:  .LBB6_2:
+; RV64I-NEXT:    addi a3, zero, 63
+; RV64I-NEXT:    sub a3, a3, a2
+; RV64I-NEXT:    slli a4, a1, 1
+; RV64I-NEXT:    sll a3, a4, a3
+; RV64I-NEXT:    srl a0, a0, a2
+; RV64I-NEXT:    or a0, a0, a3
+; RV64I-NEXT:    srl a1, a1, a2
+; RV64I-NEXT:    ret
+  %1 = lshr i128 %a, %b
+  ret i128 %1
+}
+
+define i128 @ashr128(i128 %a, i128 %b) nounwind {
+; RV32I-LABEL: ashr128:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 44(sp)
+; RV32I-NEXT:    sw s0, 40(sp)
+; RV32I-NEXT:    mv s0, a0
+; RV32I-NEXT:    lw a0, 12(a1)
+; RV32I-NEXT:    sw a0, 20(sp)
+; RV32I-NEXT:    lw a0, 8(a1)
+; RV32I-NEXT:    sw a0, 16(sp)
+; RV32I-NEXT:    lw a0, 4(a1)
+; RV32I-NEXT:    sw a0, 12(sp)
+; RV32I-NEXT:    lw a0, 0(a1)
+; RV32I-NEXT:    sw a0, 8(sp)
+; RV32I-NEXT:    lw a2, 0(a2)
+; RV32I-NEXT:    addi a0, sp, 24
+; RV32I-NEXT:    addi a1, sp, 8
+; RV32I-NEXT:    call __ashrti3
+; RV32I-NEXT:    lw a0, 36(sp)
+; RV32I-NEXT:    sw a0, 12(s0)
+; RV32I-NEXT:    lw a0, 32(sp)
+; RV32I-NEXT:    sw a0, 8(s0)
+; RV32I-NEXT:    lw a0, 28(sp)
+; RV32I-NEXT:    sw a0, 4(s0)
+; RV32I-NEXT:    lw a0, 24(sp)
+; RV32I-NEXT:    sw a0, 0(s0)
+; RV32I-NEXT:    lw s0, 40(sp)
+; RV32I-NEXT:    lw ra, 44(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: ashr128:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    addi a3, a2, -64
+; RV64I-NEXT:    bltz a3, .LBB7_2
+; RV64I-NEXT:  # %bb.1:
+; RV64I-NEXT:    sra a0, a1, a3
+; RV64I-NEXT:    srai a1, a1, 63
+; RV64I-NEXT:    ret
+; RV64I-NEXT:  .LBB7_2:
+; RV64I-NEXT:    addi a3, zero, 63
+; RV64I-NEXT:    sub a3, a3, a2
+; RV64I-NEXT:    slli a4, a1, 1
+; RV64I-NEXT:    sll a3, a4, a3
+; RV64I-NEXT:    srl a0, a0, a2
+; RV64I-NEXT:    or a0, a0, a3
+; RV64I-NEXT:    sra a1, a1, a2
+; RV64I-NEXT:    ret
+  %1 = ashr i128 %a, %b
+  ret i128 %1
+}
+
+define i128 @shl128(i128 %a, i128 %b) nounwind {
+; RV32I-LABEL: shl128:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -48
+; RV32I-NEXT:    sw ra, 44(sp)
+; RV32I-NEXT:    sw s0, 40(sp)
+; RV32I-NEXT:    mv s0, a0
+; RV32I-NEXT:    lw a0, 12(a1)
+; RV32I-NEXT:    sw a0, 20(sp)
+; RV32I-NEXT:    lw a0, 8(a1)
+; RV32I-NEXT:    sw a0, 16(sp)
+; RV32I-NEXT:    lw a0, 4(a1)
+; RV32I-NEXT:    sw a0, 12(sp)
+; RV32I-NEXT:    lw a0, 0(a1)
+; RV32I-NEXT:    sw a0, 8(sp)
+; RV32I-NEXT:    lw a2, 0(a2)
+; RV32I-NEXT:    addi a0, sp, 24
+; RV32I-NEXT:    addi a1, sp, 8
+; RV32I-NEXT:    call __ashlti3
+; RV32I-NEXT:    lw a0, 36(sp)
+; RV32I-NEXT:    sw a0, 12(s0)
+; RV32I-NEXT:    lw a0, 32(sp)
+; RV32I-NEXT:    sw a0, 8(s0)
+; RV32I-NEXT:    lw a0, 28(sp)
+; RV32I-NEXT:    sw a0, 4(s0)
+; RV32I-NEXT:    lw a0, 24(sp)
+; RV32I-NEXT:    sw a0, 0(s0)
+; RV32I-NEXT:    lw s0, 40(sp)
+; RV32I-NEXT:    lw ra, 44(sp)
+; RV32I-NEXT:    addi sp, sp, 48
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: shl128:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    addi a3, a2, -64
+; RV64I-NEXT:    bltz a3, .LBB8_2
+; RV64I-NEXT:  # %bb.1:
+; RV64I-NEXT:    sll a1, a0, a3
+; RV64I-NEXT:    mv a0, zero
+; RV64I-NEXT:    ret
+; RV64I-NEXT:  .LBB8_2:
+; RV64I-NEXT:    addi a3, zero, 63
+; RV64I-NEXT:    sub a3, a3, a2
+; RV64I-NEXT:    srli a4, a0, 1
+; RV64I-NEXT:    srl a3, a4, a3
+; RV64I-NEXT:    sll a1, a1, a2
+; RV64I-NEXT:    or a1, a1, a3
+; RV64I-NEXT:    sll a0, a0, a2
+; RV64I-NEXT:    ret
+  %1 = shl i128 %a, %b
+  ret i128 %1
+}