[RISCV] Teach RISCVDAGToDAGISel::selectShiftMask to replace sub from constant with neg.

[craig.topper]

Summary

If the shift amount is (sub C, X) where C is 0 modulo the size of
the shift, we can replace it with neg or negw.

Similar is is done for AArch64 and X86.

[craig.topper]

There are a few small regressions in this as noted. Not sure exactly how to fix them yet.

llvm/test/CodeGen/RISCV/rotl-rotr.ll

259

This sub could be a t0, a2, -32 if we changed the bltz that uses t0 to check bgtz.

271

There appears to be a missed optimization here. We have (31 - (64 - a2)). Using two subs. This could be simplified to (a2 + (-33)) using an addi.

llvm/test/CodeGen/RISCV/shifts.ll

158

After we form the neg, this sub is only used by two branches. The sub could be reversed to an addi if the branch conditions were inverted.

This is also i128 shift expansion on rv32. Clang doesn't support __int128 on rv32 so this codegen doesn't matter a lot.

[khchen]

Cool, LGTM!
It seems it worth to have sub immediate + branch optimization.

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

 //===-- RISCVISelDAGToDAG.cpp - A dag to dag inst selector for RISCV ------===//
 //
 // 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
 //
 //===----------------------------------------------------------------------===//
 //
if (N.getOpcode() == ISD::AND && isa<ConstantSDNode>(N.getOperand(1))) {
 
     // SimplifyDemandedBits may have optimized the mask so try restoring any
     // bits that are known zero.
     KnownBits Known = CurDAG->computeKnownBits(N->getOperand(0));
     if (ShMask.isSubsetOf(AndMask | Known.Zero)) {
       ShAmt = N.getOperand(0);
       return true;
     }
+  } else if (N.getOpcode() == ISD::SUB &&
+             isa<ConstantSDNode>(N.getOperand(0))) {
+    uint64_t Imm = N.getConstantOperandVal(0);
+    // If we are shifting by N-X where N == 0 mod Size, then just shift by -X to
+    // generate a NEG instead of a SUB of a constant.
+    if (Imm != 0 && Imm % ShiftWidth == 0) {
+      SDLoc DL(N);
+      EVT VT = N.getValueType();
+      SDValue Zero =
+          CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL, RISCV::X0, VT);
+      unsigned NegOpc = VT == MVT::i64 ? RISCV::SUBW : RISCV::SUB;
+      MachineSDNode *Neg = CurDAG->getMachineNode(NegOpc, DL, VT, Zero,
+                                                  N.getOperand(1));
+      ShAmt = SDValue(Neg, 0);
+      return true;
+    }
   }
 
   ShAmt = N;
   return true;
 }
 
 bool RISCVDAGToDAGISel::selectSExti32(SDValue N, SDValue &Val) {
   if (N.getOpcode() == ISD::SIGN_EXTEND_INREG &&

llvm/test/CodeGen/RISCV/rotl-rotr.ll

 ; supported, they will be turned into ISD::ROTL or ISD::ROTR.
 
 ; FIXME: We don't match 32-bit rotates with Zbb even though we have ROLW/RORW
 ; instructions.
 
 define i32 @rotl_32(i32 %x, i32 %y) nounwind {
 ; RV32I-LABEL: rotl_32:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    li a2, 32
-; RV32I-NEXT:    sub a2, a2, a1
+; RV32I-NEXT:    neg a2, a1
 ; RV32I-NEXT:    sll a1, a0, a1
 ; RV32I-NEXT:    srl a0, a0, a2
 ; RV32I-NEXT:    or a0, a1, a0
 ; RV32I-NEXT:    ret
 ;
 ; RV64I-LABEL: rotl_32:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    li a2, 32
-; RV64I-NEXT:    subw a2, a2, a1
+; RV64I-NEXT:    negw a2, a1
 ; RV64I-NEXT:    sllw a1, a0, a1
 ; RV64I-NEXT:    srlw a0, a0, a2
 ; RV64I-NEXT:    or a0, a1, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV32ZBB-LABEL: rotl_32:
 ; RV32ZBB:       # %bb.0:
 ; RV32ZBB-NEXT:    rol a0, a0, a1
 ; RV32ZBB-NEXT:    ret
 ;
 ; RV64ZBB-LABEL: rotl_32:
 ; RV64ZBB:       # %bb.0:
 ; RV64ZBB-NEXT:    rolw a0, a0, a1
 ; RV64ZBB-NEXT:    ret
   %z = sub i32 32, %y
   %b = shl i32 %x, %y
   %c = lshr i32 %x, %z
   %d = or i32 %b, %c
   ret i32 %d
 }
 
 define i32 @rotr_32(i32 %x, i32 %y) nounwind {
 ; RV32I-LABEL: rotr_32:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    li a2, 32
-; RV32I-NEXT:    sub a2, a2, a1
+; RV32I-NEXT:    neg a2, a1
 ; RV32I-NEXT:    srl a1, a0, a1
 ; RV32I-NEXT:    sll a0, a0, a2
 ; RV32I-NEXT:    or a0, a1, a0
 ; RV32I-NEXT:    ret
 ;
 ; RV64I-LABEL: rotr_32:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    li a2, 32
-; RV64I-NEXT:    subw a2, a2, a1
+; RV64I-NEXT:    negw a2, a1
 ; RV64I-NEXT:    srlw a1, a0, a1
 ; RV64I-NEXT:    sllw a0, a0, a2
 ; RV64I-NEXT:    or a0, a1, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV32ZBB-LABEL: rotr_32:
 ; RV32ZBB:       # %bb.0:
 ; RV32ZBB-NEXT:    ror a0, a0, a1
; RV64ZBB-NEXT:    ret
   ret i32 %d
 }
 
 define i64 @rotl_64(i64 %x, i64 %y) nounwind {
 ; RV32I-LABEL: rotl_64:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    mv a3, a1
 ; RV32I-NEXT:    addi a5, a2, -32
-; RV32I-NEXT:    li a4, 31
+; RV32I-NEXT:    li a6, 31
 ; RV32I-NEXT:    bltz a5, .LBB2_2
 ; RV32I-NEXT:  # %bb.1:
 ; RV32I-NEXT:    sll a1, a0, a5
 ; RV32I-NEXT:    j .LBB2_3
 ; RV32I-NEXT:  .LBB2_2:
 ; RV32I-NEXT:    sll a1, a3, a2
-; RV32I-NEXT:    sub a6, a4, a2
+; RV32I-NEXT:    sub a4, a6, a2
 ; RV32I-NEXT:    srli a7, a0, 1
-; RV32I-NEXT:    srl a6, a7, a6
-; RV32I-NEXT:    or a1, a1, a6
+; RV32I-NEXT:    srl a4, a7, a4
+; RV32I-NEXT:    or a1, a1, a4
 ; RV32I-NEXT:  .LBB2_3:
-; RV32I-NEXT:    li a6, 32
-; RV32I-NEXT:    sub a6, a6, a2
-; RV32I-NEXT:    bltz a6, .LBB2_5
+; RV32I-NEXT:    neg a7, a2
+; RV32I-NEXT:    li a4, 32
+; RV32I-NEXT:    sub t0, a4, a2
+; RV32I-NEXT:    srl a4, a3, a7
+; RV32I-NEXT:    bltz t0, .LBB2_6
 ; RV32I-NEXT:  # %bb.4:
-; RV32I-NEXT:    srl a4, a3, a6
-; RV32I-NEXT:    bltz a5, .LBB2_6
-; RV32I-NEXT:    j .LBB2_7
+; RV32I-NEXT:    bltz a5, .LBB2_7
 ; RV32I-NEXT:  .LBB2_5:
-; RV32I-NEXT:    li a6, 64
-; RV32I-NEXT:    sub a6, a6, a2
-; RV32I-NEXT:    srl a7, a0, a6
-; RV32I-NEXT:    sub a4, a4, a6
-; RV32I-NEXT:    slli t0, a3, 1
-; RV32I-NEXT:    sll a4, t0, a4
-; RV32I-NEXT:    or a4, a7, a4
-; RV32I-NEXT:    srl a3, a3, a6
-; RV32I-NEXT:    or a1, a1, a3
-; RV32I-NEXT:    bgez a5, .LBB2_7
+; RV32I-NEXT:    mv a0, a4
+; RV32I-NEXT:    ret
 ; RV32I-NEXT:  .LBB2_6:
+; RV32I-NEXT:    srl a7, a0, a7
+; RV32I-NEXT:    li t0, 64
+; RV32I-NEXT:    sub t0, t0, a2
+; RV32I-NEXT:    sub a6, a6, t0
+; RV32I-NEXT:    slli a3, a3, 1
+; RV32I-NEXT:    sll a3, a3, a6
+; RV32I-NEXT:    or a3, a7, a3
+; RV32I-NEXT:    or a1, a1, a4
+; RV32I-NEXT:    mv a4, a3
+; RV32I-NEXT:    bgez a5, .LBB2_5
+; RV32I-NEXT:  .LBB2_7:
 ; RV32I-NEXT:    sll a0, a0, a2
 ; RV32I-NEXT:    or a4, a4, a0
-; RV32I-NEXT:  .LBB2_7:
 ; RV32I-NEXT:    mv a0, a4
 ; RV32I-NEXT:    ret
 ;
 ; RV64I-LABEL: rotl_64:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    li a2, 64
-; RV64I-NEXT:    sub a2, a2, a1
+; RV64I-NEXT:    negw a2, a1
 ; RV64I-NEXT:    sll a1, a0, a1
 ; RV64I-NEXT:    srl a0, a0, a2
 ; RV64I-NEXT:    or a0, a1, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV32ZBB-LABEL: rotl_64:
 ; RV32ZBB:       # %bb.0:
 ; RV32ZBB-NEXT:    mv a3, a1
 ; RV32ZBB-NEXT:    addi a5, a2, -32
-; RV32ZBB-NEXT:    li a4, 31
+; RV32ZBB-NEXT:    li a6, 31
 ; RV32ZBB-NEXT:    bltz a5, .LBB2_2
 ; RV32ZBB-NEXT:  # %bb.1:
 ; RV32ZBB-NEXT:    sll a1, a0, a5
 ; RV32ZBB-NEXT:    j .LBB2_3
 ; RV32ZBB-NEXT:  .LBB2_2:
 ; RV32ZBB-NEXT:    sll a1, a3, a2
-; RV32ZBB-NEXT:    sub a6, a4, a2
+; RV32ZBB-NEXT:    sub a4, a6, a2
 ; RV32ZBB-NEXT:    srli a7, a0, 1
-; RV32ZBB-NEXT:    srl a6, a7, a6
-; RV32ZBB-NEXT:    or a1, a1, a6
+; RV32ZBB-NEXT:    srl a4, a7, a4
+; RV32ZBB-NEXT:    or a1, a1, a4
 ; RV32ZBB-NEXT:  .LBB2_3:
-; RV32ZBB-NEXT:    li a6, 32
-; RV32ZBB-NEXT:    sub a6, a6, a2
-; RV32ZBB-NEXT:    bltz a6, .LBB2_5
+; RV32ZBB-NEXT:    neg a7, a2
+; RV32ZBB-NEXT:    li a4, 32
+; RV32ZBB-NEXT:    sub t0, a4, a2
+; RV32ZBB-NEXT:    srl a4, a3, a7
+; RV32ZBB-NEXT:    bltz t0, .LBB2_6
 ; RV32ZBB-NEXT:  # %bb.4:
-; RV32ZBB-NEXT:    srl a4, a3, a6
-; RV32ZBB-NEXT:    bltz a5, .LBB2_6
-; RV32ZBB-NEXT:    j .LBB2_7
+; RV32ZBB-NEXT:    bltz a5, .LBB2_7
 ; RV32ZBB-NEXT:  .LBB2_5:
-; RV32ZBB-NEXT:    li a6, 64
-; RV32ZBB-NEXT:    sub a6, a6, a2
-; RV32ZBB-NEXT:    srl a7, a0, a6
-; RV32ZBB-NEXT:    sub a4, a4, a6
-; RV32ZBB-NEXT:    slli t0, a3, 1
-; RV32ZBB-NEXT:    sll a4, t0, a4
-; RV32ZBB-NEXT:    or a4, a7, a4
-; RV32ZBB-NEXT:    srl a3, a3, a6
-; RV32ZBB-NEXT:    or a1, a1, a3
-; RV32ZBB-NEXT:    bgez a5, .LBB2_7
+; RV32ZBB-NEXT:    mv a0, a4
+; RV32ZBB-NEXT:    ret
 ; RV32ZBB-NEXT:  .LBB2_6:
+; RV32ZBB-NEXT:    srl a7, a0, a7
+; RV32ZBB-NEXT:    li t0, 64
+; RV32ZBB-NEXT:    sub t0, t0, a2
+; RV32ZBB-NEXT:    sub a6, a6, t0
+; RV32ZBB-NEXT:    slli a3, a3, 1
+; RV32ZBB-NEXT:    sll a3, a3, a6
+; RV32ZBB-NEXT:    or a3, a7, a3
+; RV32ZBB-NEXT:    or a1, a1, a4
+; RV32ZBB-NEXT:    mv a4, a3
+; RV32ZBB-NEXT:    bgez a5, .LBB2_5
+; RV32ZBB-NEXT:  .LBB2_7:
 ; RV32ZBB-NEXT:    sll a0, a0, a2
 ; RV32ZBB-NEXT:    or a4, a4, a0
-; RV32ZBB-NEXT:  .LBB2_7:
 ; RV32ZBB-NEXT:    mv a0, a4
 ; RV32ZBB-NEXT:    ret
 ;
 ; RV64ZBB-LABEL: rotl_64:
 ; RV64ZBB:       # %bb.0:
 ; RV64ZBB-NEXT:    rol a0, a0, a1
 ; RV64ZBB-NEXT:    ret
   %z = sub i64 64, %y
   %b = shl i64 %x, %y
   %c = lshr i64 %x, %z
   %d = or i64 %b, %c
   ret i64 %d
 }
 
 define i64 @rotr_64(i64 %x, i64 %y) nounwind {
 ; RV32I-LABEL: rotr_64:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    mv a3, a0
+; RV32I-NEXT:    mv a4, a0
 ; RV32I-NEXT:    addi a5, a2, -32
-; RV32I-NEXT:    li a4, 31
+; RV32I-NEXT:    li a6, 31
 ; RV32I-NEXT:    bltz a5, .LBB3_2
 ; RV32I-NEXT:  # %bb.1:
 ; RV32I-NEXT:    srl a0, a1, a5
 ; RV32I-NEXT:    j .LBB3_3
 ; RV32I-NEXT:  .LBB3_2:
-; RV32I-NEXT:    srl a0, a3, a2
-; RV32I-NEXT:    sub a6, a4, a2
+; RV32I-NEXT:    srl a0, a4, a2
+; RV32I-NEXT:    sub a3, a6, a2
 ; RV32I-NEXT:    slli a7, a1, 1
-; RV32I-NEXT:    sll a6, a7, a6
-; RV32I-NEXT:    or a0, a0, a6
+; RV32I-NEXT:    sll a3, a7, a3
+; RV32I-NEXT:    or a0, a0, a3
 ; RV32I-NEXT:  .LBB3_3:
-; RV32I-NEXT:    li a6, 32
-; RV32I-NEXT:    sub a6, a6, a2
-; RV32I-NEXT:    bltz a6, .LBB3_5
+; RV32I-NEXT:    neg a7, a2
+; RV32I-NEXT:    li a3, 32
+; RV32I-NEXT:    sub t0, a3, a2
+; RV32I-NEXT:    sll a3, a4, a7
+; RV32I-NEXT:    bltz t0, .LBB3_6
 ; RV32I-NEXT:  # %bb.4:
-; RV32I-NEXT:    sll a4, a3, a6
-; RV32I-NEXT:    bltz a5, .LBB3_6
-; RV32I-NEXT:    j .LBB3_7
+; RV32I-NEXT:    bltz a5, .LBB3_7
 ; RV32I-NEXT:  .LBB3_5:
-; RV32I-NEXT:    li a6, 64
-; RV32I-NEXT:    sub a6, a6, a2
-; RV32I-NEXT:    sll a7, a1, a6
-; RV32I-NEXT:    sub a4, a4, a6
-; RV32I-NEXT:    srli t0, a3, 1
-; RV32I-NEXT:    srl a4, t0, a4
+; RV32I-NEXT:    mv a1, a3
+; RV32I-NEXT:    ret
+; RV32I-NEXT:  .LBB3_6:
+; RV32I-NEXT:    sll a7, a1, a7
+; RV32I-NEXT:    li t0, 64
+; RV32I-NEXT:    sub t0, t0, a2
+; RV32I-NEXT:    sub a6, a6, t0
+; RV32I-NEXT:    srli a4, a4, 1
+; RV32I-NEXT:    srl a4, a4, a6
 ; RV32I-NEXT:    or a4, a7, a4
-; RV32I-NEXT:    sll a3, a3, a6
 ; RV32I-NEXT:    or a0, a0, a3
-; RV32I-NEXT:    bgez a5, .LBB3_7
-; RV32I-NEXT:  .LBB3_6:
-; RV32I-NEXT:    srl a1, a1, a2
-; RV32I-NEXT:    or a4, a4, a1
+; RV32I-NEXT:    mv a3, a4
+; RV32I-NEXT:    bgez a5, .LBB3_5
 ; RV32I-NEXT:  .LBB3_7:
-; RV32I-NEXT:    mv a1, a4
+; RV32I-NEXT:    srl a1, a1, a2
+; RV32I-NEXT:    or a3, a3, a1
+; RV32I-NEXT:    mv a1, a3
 ; RV32I-NEXT:    ret
 ;
 ; RV64I-LABEL: rotr_64:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    li a2, 64
-; RV64I-NEXT:    sub a2, a2, a1
+; RV64I-NEXT:    negw a2, a1
 ; RV64I-NEXT:    srl a1, a0, a1
 ; RV64I-NEXT:    sll a0, a0, a2
 ; RV64I-NEXT:    or a0, a1, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV32ZBB-LABEL: rotr_64:
 ; RV32ZBB:       # %bb.0:
-; RV32ZBB-NEXT:    mv a3, a0
+; RV32ZBB-NEXT:    mv a4, a0
 ; RV32ZBB-NEXT:    addi a5, a2, -32
-; RV32ZBB-NEXT:    li a4, 31
+; RV32ZBB-NEXT:    li a6, 31
 ; RV32ZBB-NEXT:    bltz a5, .LBB3_2
 ; RV32ZBB-NEXT:  # %bb.1:
 ; RV32ZBB-NEXT:    srl a0, a1, a5
 ; RV32ZBB-NEXT:    j .LBB3_3
 ; RV32ZBB-NEXT:  .LBB3_2:
-; RV32ZBB-NEXT:    srl a0, a3, a2
-; RV32ZBB-NEXT:    sub a6, a4, a2
+; RV32ZBB-NEXT:    srl a0, a4, a2
+; RV32ZBB-NEXT:    sub a3, a6, a2
 ; RV32ZBB-NEXT:    slli a7, a1, 1
-; RV32ZBB-NEXT:    sll a6, a7, a6
-; RV32ZBB-NEXT:    or a0, a0, a6
+; RV32ZBB-NEXT:    sll a3, a7, a3
+; RV32ZBB-NEXT:    or a0, a0, a3
 ; RV32ZBB-NEXT:  .LBB3_3:
-; RV32ZBB-NEXT:    li a6, 32
-; RV32ZBB-NEXT:    sub a6, a6, a2
-; RV32ZBB-NEXT:    bltz a6, .LBB3_5
+; RV32ZBB-NEXT:    neg a7, a2
+; RV32ZBB-NEXT:    li a3, 32
+; RV32ZBB-NEXT:    sub t0, a3, a2
+; RV32ZBB-NEXT:    sll a3, a4, a7
+; RV32ZBB-NEXT:    bltz t0, .LBB3_6
 ; RV32ZBB-NEXT:  # %bb.4:
-; RV32ZBB-NEXT:    sll a4, a3, a6
-; RV32ZBB-NEXT:    bltz a5, .LBB3_6
-; RV32ZBB-NEXT:    j .LBB3_7
+; RV32ZBB-NEXT:    bltz a5, .LBB3_7
 ; RV32ZBB-NEXT:  .LBB3_5:
-; RV32ZBB-NEXT:    li a6, 64
-; RV32ZBB-NEXT:    sub a6, a6, a2
-; RV32ZBB-NEXT:    sll a7, a1, a6
-; RV32ZBB-NEXT:    sub a4, a4, a6
-; RV32ZBB-NEXT:    srli t0, a3, 1
-; RV32ZBB-NEXT:    srl a4, t0, a4
+; RV32ZBB-NEXT:    mv a1, a3
+; RV32ZBB-NEXT:    ret
+; RV32ZBB-NEXT:  .LBB3_6:
+; RV32ZBB-NEXT:    sll a7, a1, a7
+; RV32ZBB-NEXT:    li t0, 64
+; RV32ZBB-NEXT:    sub t0, t0, a2
+; RV32ZBB-NEXT:    sub a6, a6, t0
+; RV32ZBB-NEXT:    srli a4, a4, 1
+; RV32ZBB-NEXT:    srl a4, a4, a6
 ; RV32ZBB-NEXT:    or a4, a7, a4
-; RV32ZBB-NEXT:    sll a3, a3, a6
 ; RV32ZBB-NEXT:    or a0, a0, a3
-; RV32ZBB-NEXT:    bgez a5, .LBB3_7
-; RV32ZBB-NEXT:  .LBB3_6:
-; RV32ZBB-NEXT:    srl a1, a1, a2
-; RV32ZBB-NEXT:    or a4, a4, a1
+; RV32ZBB-NEXT:    mv a3, a4
+; RV32ZBB-NEXT:    bgez a5, .LBB3_5
 ; RV32ZBB-NEXT:  .LBB3_7:
-; RV32ZBB-NEXT:    mv a1, a4
+; RV32ZBB-NEXT:    srl a1, a1, a2
+; RV32ZBB-NEXT:    or a3, a3, a1
+; RV32ZBB-NEXT:    mv a1, a3
 ; RV32ZBB-NEXT:    ret
 ;
 ; RV64ZBB-LABEL: rotr_64:
 ; RV64ZBB:       # %bb.0:
 ; RV64ZBB-NEXT:    ror a0, a0, a1
 ; RV64ZBB-NEXT:    ret
   %z = sub i64 64, %y
   %b = lshr i64 %x, %y

llvm/test/CodeGen/RISCV/shifts.ll

 }
 
 define i128 @lshr128(i128 %a, i128 %b) nounwind {
 ; RV32I-LABEL: lshr128:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lw a2, 0(a2)
 ; RV32I-NEXT:    lw a5, 8(a1)
 ; RV32I-NEXT:    lw a4, 12(a1)
+; RV32I-NEXT:    neg a6, a2
 ; RV32I-NEXT:    li a3, 64
-; RV32I-NEXT:    sub t0, a3, a2
-; RV32I-NEXT:    li a6, 32
-; RV32I-NEXT:    sub t1, a6, a2
 ; RV32I-NEXT:    li t2, 31
+; RV32I-NEXT:    li a7, 32
+; RV32I-NEXT:    sub t1, a7, a2
+; RV32I-NEXT:    sll t0, a5, a6
 ; RV32I-NEXT:    bltz t1, .LBB6_2
 ; RV32I-NEXT:  # %bb.1:
-; RV32I-NEXT:    sll t6, a5, t1
+; RV32I-NEXT:    mv t6, t0
 ; RV32I-NEXT:    j .LBB6_3
 ; RV32I-NEXT:  .LBB6_2:
-; RV32I-NEXT:    sll a6, a4, t0
-; RV32I-NEXT:    sub a7, t2, t0
+; RV32I-NEXT:    sll a6, a4, a6
+; RV32I-NEXT:    sub a7, a3, a2
+; RV32I-NEXT:    sub a7, t2, a7
 ; RV32I-NEXT:    srli t3, a5, 1
 ; RV32I-NEXT:    srl a7, t3, a7
 ; RV32I-NEXT:    or t6, a6, a7
 ; RV32I-NEXT:  .LBB6_3:
 ; RV32I-NEXT:    lw t5, 4(a1)
 ; RV32I-NEXT:    addi a6, a2, -32
 ; RV32I-NEXT:    bgez a6, .LBB6_5
 ; RV32I-NEXT:  # %bb.4:

 ; RV32I-NEXT:    j .LBB6_15
 ; RV32I-NEXT:  .LBB6_13:
 ; RV32I-NEXT:    srl t6, a1, a2
 ; RV32I-NEXT:    slli t5, t5, 1
 ; RV32I-NEXT:    sll t5, t5, t2
 ; RV32I-NEXT:    or t5, t6, t5
 ; RV32I-NEXT:    bgez t1, .LBB6_15
 ; RV32I-NEXT:  .LBB6_14:
-; RV32I-NEXT:    sll t0, a5, t0
 ; RV32I-NEXT:    or t5, t5, t0
 ; RV32I-NEXT:  .LBB6_15:
 ; RV32I-NEXT:    slli t0, a4, 1
 ; RV32I-NEXT:    bltz t4, .LBB6_17
 ; RV32I-NEXT:  # %bb.16:
 ; RV32I-NEXT:    srl t1, a4, t4
 ; RV32I-NEXT:    bgeu a2, a3, .LBB6_18
 ; RV32I-NEXT:    j .LBB6_19

 define i128 @ashr128(i128 %a, i128 %b) nounwind {
 ; RV32I-LABEL: ashr128:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    addi sp, sp, -16
 ; RV32I-NEXT:    sw s0, 12(sp) # 4-byte Folded Spill
 ; RV32I-NEXT:    lw a2, 0(a2)
 ; RV32I-NEXT:    lw a5, 8(a1)
 ; RV32I-NEXT:    lw a4, 12(a1)
+; RV32I-NEXT:    neg a6, a2
 ; RV32I-NEXT:    li a3, 64
-; RV32I-NEXT:    sub t1, a3, a2
-; RV32I-NEXT:    li a6, 32
-; RV32I-NEXT:    sub t2, a6, a2
-; RV32I-NEXT:    li t4, 31
+; RV32I-NEXT:    li t3, 31
+; RV32I-NEXT:    li a7, 32
+; RV32I-NEXT:    sub t2, a7, a2
+; RV32I-NEXT:    sll t1, a5, a6
 ; RV32I-NEXT:    bltz t2, .LBB7_2
 ; RV32I-NEXT:  # %bb.1:
-; RV32I-NEXT:    sll s0, a5, t2
+; RV32I-NEXT:    mv s0, t1
 ; RV32I-NEXT:    j .LBB7_3
 ; RV32I-NEXT:  .LBB7_2:
-; RV32I-NEXT:    sll a6, a4, t1
-; RV32I-NEXT:    sub a7, t4, t1
+; RV32I-NEXT:    sll a6, a4, a6
+; RV32I-NEXT:    sub a7, a3, a2
+; RV32I-NEXT:    sub a7, t3, a7
 ; RV32I-NEXT:    srli t0, a5, 1
 ; RV32I-NEXT:    srl a7, t0, a7
 ; RV32I-NEXT:    or s0, a6, a7
 ; RV32I-NEXT:  .LBB7_3:
 ; RV32I-NEXT:    lw t6, 4(a1)
 ; RV32I-NEXT:    addi a6, a2, -32
 ; RV32I-NEXT:    bgez a6, .LBB7_5
 ; RV32I-NEXT:  # %bb.4:
 ; RV32I-NEXT:    srl a7, t6, a2
 ; RV32I-NEXT:    or s0, s0, a7
 ; RV32I-NEXT:  .LBB7_5:
-; RV32I-NEXT:    addi t3, a2, -64
+; RV32I-NEXT:    addi t4, a2, -64
 ; RV32I-NEXT:    addi t5, a2, -96
 ; RV32I-NEXT:    srai a7, a4, 31
 ; RV32I-NEXT:    bltz t5, .LBB7_7
 ; RV32I-NEXT:  # %bb.6:
 ; RV32I-NEXT:    mv t0, a7
 ; RV32I-NEXT:    bgeu a2, a3, .LBB7_8
 ; RV32I-NEXT:    j .LBB7_9
 ; RV32I-NEXT:  .LBB7_7:
-; RV32I-NEXT:    sra t0, a4, t3
+; RV32I-NEXT:    sra t0, a4, t4
 ; RV32I-NEXT:    bltu a2, a3, .LBB7_9
 ; RV32I-NEXT:  .LBB7_8:
 ; RV32I-NEXT:    mv s0, t0
 ; RV32I-NEXT:  .LBB7_9:
 ; RV32I-NEXT:    mv t0, t6
 ; RV32I-NEXT:    beqz a2, .LBB7_11
 ; RV32I-NEXT:  # %bb.10:
 ; RV32I-NEXT:    mv t0, s0
 ; RV32I-NEXT:  .LBB7_11:
 ; RV32I-NEXT:    lw a1, 0(a1)
-; RV32I-NEXT:    sub t4, t4, a2
+; RV32I-NEXT:    sub t3, t3, a2
 ; RV32I-NEXT:    bltz a6, .LBB7_13
 ; RV32I-NEXT:  # %bb.12:
 ; RV32I-NEXT:    srl t6, t6, a6
 ; RV32I-NEXT:    bltz t2, .LBB7_14
 ; RV32I-NEXT:    j .LBB7_15
 ; RV32I-NEXT:  .LBB7_13:
 ; RV32I-NEXT:    srl s0, a1, a2
 ; RV32I-NEXT:    slli t6, t6, 1
-; RV32I-NEXT:    sll t6, t6, t4
+; RV32I-NEXT:    sll t6, t6, t3
 ; RV32I-NEXT:    or t6, s0, t6
 ; RV32I-NEXT:    bgez t2, .LBB7_15
 ; RV32I-NEXT:  .LBB7_14:
-; RV32I-NEXT:    sll t1, a5, t1
 ; RV32I-NEXT:    or t6, t6, t1
 ; RV32I-NEXT:  .LBB7_15:
 ; RV32I-NEXT:    slli t1, a4, 1
 ; RV32I-NEXT:    bltz t5, .LBB7_17
 ; RV32I-NEXT:  # %bb.16:
 ; RV32I-NEXT:    sra t2, a4, t5
 ; RV32I-NEXT:    bgeu a2, a3, .LBB7_18
 ; RV32I-NEXT:    j .LBB7_19
 ; RV32I-NEXT:  .LBB7_17:
 ; RV32I-NEXT:    li t2, 95
 ; RV32I-NEXT:    sub t2, t2, a2
 ; RV32I-NEXT:    sll t2, t1, t2
-; RV32I-NEXT:    srl t3, a5, t3
-; RV32I-NEXT:    or t2, t3, t2
+; RV32I-NEXT:    srl t4, a5, t4
+; RV32I-NEXT:    or t2, t4, t2
 ; RV32I-NEXT:    bltu a2, a3, .LBB7_19
 ; RV32I-NEXT:  .LBB7_18:
 ; RV32I-NEXT:    mv t6, t2
 ; RV32I-NEXT:  .LBB7_19:
 ; RV32I-NEXT:    bnez a2, .LBB7_22
 ; RV32I-NEXT:  # %bb.20:
 ; RV32I-NEXT:    bltz a6, .LBB7_23
 ; RV32I-NEXT:  .LBB7_21:
 ; RV32I-NEXT:    sra a5, a4, a6
 ; RV32I-NEXT:    bgeu a2, a3, .LBB7_24
 ; RV32I-NEXT:    j .LBB7_25
 ; RV32I-NEXT:  .LBB7_22:
 ; RV32I-NEXT:    mv a1, t6
 ; RV32I-NEXT:    bgez a6, .LBB7_21
 ; RV32I-NEXT:  .LBB7_23:
 ; RV32I-NEXT:    srl a5, a5, a2
-; RV32I-NEXT:    sll t1, t1, t4
+; RV32I-NEXT:    sll t1, t1, t3
 ; RV32I-NEXT:    or a5, a5, t1
 ; RV32I-NEXT:    bltu a2, a3, .LBB7_25
 ; RV32I-NEXT:  .LBB7_24:
 ; RV32I-NEXT:    mv a5, a7
 ; RV32I-NEXT:  .LBB7_25:
 ; RV32I-NEXT:    bltz a6, .LBB7_27
 ; RV32I-NEXT:  # %bb.26:
 ; RV32I-NEXT:    mv a4, a7

 }
 
 define i128 @shl128(i128 %a, i128 %b) nounwind {
 ; RV32I-LABEL: shl128:
 ; RV32I:       # %bb.0:
 ; RV32I-NEXT:    lw a2, 0(a2)
 ; RV32I-NEXT:    lw a5, 4(a1)
 ; RV32I-NEXT:    lw a4, 0(a1)
+; RV32I-NEXT:    neg a6, a2
 ; RV32I-NEXT:    li a3, 64
-; RV32I-NEXT:    sub t0, a3, a2
-; RV32I-NEXT:    li a6, 32
-; RV32I-NEXT:    sub t1, a6, a2
 ; RV32I-NEXT:    li t2, 31
+; RV32I-NEXT:    li a7, 32
+; RV32I-NEXT:    sub t1, a7, a2
+; RV32I-NEXT:    srl t0, a5, a6
 ; RV32I-NEXT:    bltz t1, .LBB8_2
 ; RV32I-NEXT:  # %bb.1:
-; RV32I-NEXT:    srl t6, a5, t1
+; RV32I-NEXT:    mv t6, t0
 ; RV32I-NEXT:    j .LBB8_3
 ; RV32I-NEXT:  .LBB8_2:
-; RV32I-NEXT:    srl a6, a4, t0
-; RV32I-NEXT:    sub a7, t2, t0
+; RV32I-NEXT:    srl a6, a4, a6
+; RV32I-NEXT:    sub a7, a3, a2
+; RV32I-NEXT:    sub a7, t2, a7
 ; RV32I-NEXT:    slli t3, a5, 1
 ; RV32I-NEXT:    sll a7, t3, a7
 ; RV32I-NEXT:    or t6, a6, a7
 ; RV32I-NEXT:  .LBB8_3:
 ; RV32I-NEXT:    lw t5, 8(a1)
 ; RV32I-NEXT:    addi a6, a2, -32
 ; RV32I-NEXT:    bgez a6, .LBB8_5
 ; RV32I-NEXT:  # %bb.4:

 ; RV32I-NEXT:    j .LBB8_15
 ; RV32I-NEXT:  .LBB8_13:
 ; RV32I-NEXT:    sll t6, a1, a2
 ; RV32I-NEXT:    srli t5, t5, 1
 ; RV32I-NEXT:    srl t5, t5, t2
 ; RV32I-NEXT:    or t5, t6, t5
 ; RV32I-NEXT:    bgez t1, .LBB8_15
 ; RV32I-NEXT:  .LBB8_14:
-; RV32I-NEXT:    srl t0, a5, t0
 ; RV32I-NEXT:    or t5, t5, t0
 ; RV32I-NEXT:  .LBB8_15:
 ; RV32I-NEXT:    srli t0, a4, 1
 ; RV32I-NEXT:    bltz t4, .LBB8_17
 ; RV32I-NEXT:  # %bb.16:
 ; RV32I-NEXT:    sll t1, a4, t4
 ; RV32I-NEXT:    bgeu a2, a3, .LBB8_18
 ; RV32I-NEXT:    j .LBB8_19