[AMDGPU][GlobalISel] Legalization of G_ROTL and G_ROTR

[matejam]

Summary

Add implementation for the legalization of G_ROTL and G_ROTR machine instructions.
They are very similar to funnel shift instructions, the only difference is funnel shifts have 3 operands,
whereas rotate instructions have two operands, the first being the register that is being rotated and
the second being the number of shifts. The legalization of G_ROTL/G_ROTR is just lowering them into
funnel shift instructions if they are legal.

[arsenm]

llvm/lib/Target/AMDGPU/VOP3Instructions.td

346 ↗

(On Diff #356157)

Isn't there a lowering to turn this into fshr? Why do we need to directly select this?

llvm/test/CodeGen/AMDGPU/GlobalISel/rotate_pre_instselect.mir

6 ↗

(On Diff #356157)

Don't need the IR section, registers section or most of the MIR extras here

llvm/test/CodeGen/AMDGPU/GlobalISel/rotate_pre_regbankselect.mir

713 ↗

(On Diff #356157)

You can simplify out stuff like this, for the purpose of this test it doesn't need the readlanes

[matejam]

Changes in tests.

[foad]

llvm/lib/Target/AMDGPU/VOP3Instructions.td

346 ↗

(On Diff #356157)

I have previously discussed this with Mateja. There's a combine that turns funnel shifts into rotates, so I wasn't sure if that meant that the rotates were canonical and we had to handle them.

Thinking about it now, I guess we can say that rotates are illegal, and the combine should bail out in that case?

[matejam]

llvm/lib/Target/AMDGPU/VOP3Instructions.td

346 ↗

(On Diff #356157)

If we disabled the "funnel shift to rotate" combiner with a check if the rotate is legal or not, we would also be disabling that combiner for AArch64, specifically for aarch64-prelegalizer-combiner pass.

[foad]

llvm/lib/Target/AMDGPU/VOP3Instructions.td

346 ↗

(On Diff #356157)

That combine already checks "is legal or before legalizer", so it should be OK. (It is a bit inefficient, because the pre-legalizer combiner will convert funnel shift -> rotate and then the legalizer will convert it back again, but I think that is OK.)

[foad]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

2762

Can we guarantee that DstTy is a power of two size here? Otherwise just negating the rotate amount does not work.

2764

You could implement B.buildNeg for this, similar to the existing B.buildNot.

[matejam]

When rotating left, use the right negated rotation (negate the number of shifts) only in cases when the destination type size is a power of 2.

[matejam]

Add tests for non power of 2 types (i15, i31...).

[foad]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

2763

buildFNeg is only for floating point. You need buildNeg, which does not exist yet, but you could implement it.

[matejam]

Implement integer negation buildNeg() and use it instead of buildFNeg().

[matejam]

Formatting and refactoring.

[matejam]

Formatting and refactoring.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

2752

There's nothing target specific here, the generic LegalizerHelper can handle this

llvm/lib/Target/AMDGPU/VOP3Instructions.td

346 ↗

(On Diff #356157)

Saying anything is canonical in the MIR is a bit fuzzy. I do think the is legal or before legalizer checks aren't ideal now. I think the combiner needs to distinguish between being legal and not having legalization info

llvm/test/CodeGen/AMDGPU/GlobalISel/rotate_pre_regbankselect.mir

516–613 ↗

(On Diff #356157)

Don't need all this meta stuff

713 ↗

(On Diff #356157)

You should still remove these intrinsic copies

[matejam]

Move the legalization of G_ROTL and G_ROTR from AMDGPULegalizerInfo to LegalizerHelper.

[matejam]

Added a few comments and few minor updates.

[matejam]

Formatting.

[foad]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

5973–5974

This doesn't do what you think it does :) It sets IsFShLegal to the whole expression LI.isLegalOrCustom({FShOpc, {DstTy, AmtTy}}) || LI.isLegalOrCustom({RevFsh, {DstTy, AmtTy}})

5976

... so IsFShLegal is always true here.

[matejam]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

5976

Thanks! :)

[matejam]

Minor bug fix.

[foad]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

5978

What if it's not a power of two? In that case, should we fall through to the G_SHL/G_LSHR expansion below?

[matejam]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

5978

Yes, we should... we go into the "if (!IsFShLegal) " statement only if the instruction with the opcode FShOpc isn't legal and the other is. The problem is if the RevFsh instruction is legal and the FShOpc instruction isn't and the type isn't a power of 2, this code will try to build an instruction which ins't legal, instead of falling through to the G_SHL/G_LSHR part.
I'll fix that right away.

[matejam]

In case the funnel shift instruction is not legal and the type is not a power of 2, use the old way of lowering (G_ROTL/G_ROTR -> G_SHL/G_LSHR).

[foad]

LGTM, thanks!

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

5982

Nit: this would be slightly neater:

if (IsFshLegal) {
  return ...
} else if (isPowerOf2_32(EltSizeInBits)) {
  return ...
}

[matejam]

Refactoring.

llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h

   /// Build and insert a bitwise not,
   /// \p NegOne = G_CONSTANT -1
   /// \p Res = G_OR \p Op0, NegOne
   MachineInstrBuilder buildNot(const DstOp &Dst, const SrcOp &Src0) {
     auto NegOne = buildConstant(Dst.getLLTTy(*getMRI()), -1);
     return buildInstr(TargetOpcode::G_XOR, {Dst}, {Src0, NegOne});
   }
 
+  /// Build and insert integer negation
+  /// \p Zero = G_CONSTANT 0
+  /// \p Res = G_SUB Zero, \p Op0
+  MachineInstrBuilder buildNeg(const DstOp &Dst, const SrcOp &Src0) {
+    auto Zero = buildConstant(Dst.getLLTTy(*getMRI()), 0);
+    return buildInstr(TargetOpcode::G_SUB, {Dst}, {Zero, Src0});
+  }
+
   /// Build and insert \p Res = G_CTPOP \p Op0, \p Src0
   MachineInstrBuilder buildCTPOP(const DstOp &Dst, const SrcOp &Src0) {
     return buildInstr(TargetOpcode::G_CTPOP, {Dst}, {Src0});
   }
 
   /// Build and insert \p Res = G_CTLZ \p Op0, \p Src0
   MachineInstrBuilder buildCTLZ(const DstOp &Dst, const SrcOp &Src0) {
     return buildInstr(TargetOpcode::G_CTLZ, {Dst}, {Src0});

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

LegalizerHelper::fewerElementsVector(MachineInstr &MI, unsigned TypeIdx,
   case G_FMINNUM:
   case G_FMAXNUM:
   case G_FMINNUM_IEEE:
   case G_FMAXNUM_IEEE:
   case G_FMINIMUM:
   case G_FMAXIMUM:
   case G_FSHL:
   case G_FSHR:
+  case G_ROTL:
+  case G_ROTR:
   case G_FREEZE:
   case G_SADDSAT:
   case G_SSUBSAT:
   case G_UADDSAT:
   case G_USUBSAT:
     return reduceOperationWidth(MI, TypeIdx, NarrowTy);
   case G_UMULO:
   case G_SMULO:
LegalizerHelper::LegalizeResult LegalizerHelper::lowerRotate(MachineInstr &MI) {
   MIRBuilder.setInstrAndDebugLoc(MI);
 
   // If a rotate in the other direction is supported, use it.
   unsigned RevRot = IsLeft ? TargetOpcode::G_ROTR : TargetOpcode::G_ROTL;
   if (LI.isLegalOrCustom({RevRot, {DstTy, SrcTy}}) &&
       isPowerOf2_32(EltSizeInBits))
     return lowerRotateWithReverseRotate(MI);
 
+  // If a funnel shift is supported, use it.
+  unsigned FShOpc = IsLeft ? TargetOpcode::G_FSHL : TargetOpcode::G_FSHR;
+  unsigned RevFsh = !IsLeft ? TargetOpcode::G_FSHL : TargetOpcode::G_FSHR;
+  if (bool IsFShLegal = LI.isLegalOrCustom({FShOpc, {DstTy, AmtTy}}) ||
+      LI.isLegalOrCustom({RevFsh, {DstTy, AmtTy}})) {
+    // If a funnel shift in the other direction is supported, use it.
+    if (!IsFShLegal) {
+      if (isPowerOf2_32(EltSizeInBits)) {
+        FShOpc = RevFsh;
+        Amt = MIRBuilder.buildNeg(DstTy, Amt).getReg(0);
+      }
+    }
+    MIRBuilder.buildInstr(FShOpc, {Dst}, {Src, Src, Amt});
+    MI.eraseFromParent();
+    return Legalized;
+  }
+
   auto Zero = MIRBuilder.buildConstant(AmtTy, 0);
   unsigned ShOpc = IsLeft ? TargetOpcode::G_SHL : TargetOpcode::G_LSHR;
   unsigned RevShiftOpc = IsLeft ? TargetOpcode::G_LSHR : TargetOpcode::G_SHL;
   auto BitWidthMinusOneC = MIRBuilder.buildConstant(AmtTy, EltSizeInBits - 1);
   Register ShVal;
   Register RevShiftVal;
   if (isPowerOf2_32(EltSizeInBits)) {
     // (rotl x, c) -> x << (c & (w - 1)) | x >> (-c & (w - 1))
LegalizerHelper::lowerU64ToF32BitOps(MachineInstr &MI) {
 
   assert(MRI.getType(Src) == S64 && MRI.getType(Dst) == S32);
 
   // unsigned cul2f(ulong u) {
   //   uint lz = clz(u);
   //   uint e = (u != 0) ? 127U + 63U - lz : 0;
   //   u = (u << lz) & 0x7fffffffffffffffUL;
   //   ulong t = u & 0xffffffffffUL;
   //   uint v = (e << 23) | (uint)(u >> 40);
   //   uint r = t > 0x8000000000UL ? 1U : (t == 0x8000000000UL ? v & 1U : 0U);
   //   return as_float(v + r);
   // }
 
   auto Zero32 = MIRBuilder.buildConstant(S32, 0);
   auto Zero64 = MIRBuilder.buildConstant(S64, 0);
 
   auto LZ = MIRBuilder.buildCTLZ_ZERO_UNDEF(S32, Src);

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

if (ST.hasVOP3PInsts()) {
     // Prefer to promote to s32 before lowering if we don't have 16-bit
     // shifts. This avoid a lot of intermediate truncate and extend operations.
     SextInReg.lowerFor({{S32}, {S64}});
   }
 
   SextInReg
     .scalarize(0)
     .clampScalar(0, S32, S64)
     .lower();
 
+  getActionDefinitionsBuilder({G_ROTR, G_ROTL})
+    .scalarize(0)
+    .lower();
+
   // TODO: Only Try to form v2s16 with legal packed instructions.
   getActionDefinitionsBuilder(G_FSHR)
     .legalFor({{S32, S32}})
     .lowerFor({{V2S16, V2S16}})
     .fewerElementsIf(elementTypeIs(0, S16), changeTo(0, V2S16))
     .scalarize(0)
     .lower();
 
bool AMDGPULegalizerInfo::legalizeBuildVector(
 
   auto Merge = B.buildMerge(S32, {Src0, Src1});
   B.buildBitcast(Dst, Merge);
 
   MI.eraseFromParent();
   return true;
 }
 
 // Check that this is a G_XOR x, -1
 static bool isNot(const MachineRegisterInfo &MRI, const MachineInstr &MI) {
   if (MI.getOpcode() != TargetOpcode::G_XOR)
     return false;
   auto ConstVal = getConstantVRegSExtVal(MI.getOperand(2).getReg(), MRI);
   return ConstVal && *ConstVal == -1;
 }
 
 // Return the use branch instruction, otherwise null if the usage is invalid.
 static MachineInstr *
 verifyCFIntrinsic(MachineInstr &MI, MachineRegisterInfo &MRI, MachineInstr *&Br,
                   MachineBasicBlock *&UncondBrTarget, bool &Negated) {
   Register CondDef = MI.getOperand(0).getReg();
   if (!MRI.hasOneNonDBGUse(CondDef))
     return nullptr;
 
   MachineBasicBlock *Parent = MI.getParent();
   MachineInstr *UseMI = &*MRI.use_instr_nodbg_begin(CondDef);
 
   if (isNot(MRI, *UseMI)) {
     Register NegatedCond = UseMI->getOperand(0).getReg();

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-rotl-rotr.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -global-isel -march=amdgcn -mcpu=tahiti -verify-machineinstrs -run-pass=legalizer %s -o - | FileCheck %s --check-prefixes=GFX,GFX6
+# RUN: llc -global-isel -march=amdgcn -mcpu=fiji -verify-machineinstrs -run-pass=legalizer %s -o - | FileCheck %s --check-prefixes=GFX,GFX8
+
+---
+name:            rotl_i15
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0, $sgpr1
+
+    ; GFX6-LABEL: name: rotl_i15
+    ; GFX6: liveins: $sgpr0, $sgpr1
+    ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $sgpr0
+    ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $sgpr1
+    ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 14
+    ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 15
+    ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 32767
+    ; GFX6: [[COPY2:%[0-9]+]]:_(s32) = COPY [[COPY1]](s32)
+    ; GFX6: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C2]]
+    ; GFX6: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C1]](s32)
+    ; GFX6: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C2]]
+    ; GFX6: [[UITOFP:%[0-9]+]]:_(s32) = G_UITOFP [[AND1]](s32)
+    ; GFX6: [[AMDGPU_RCP_IFLAG:%[0-9]+]]:_(s32) = G_AMDGPU_RCP_IFLAG [[UITOFP]](s32)
+    ; GFX6: [[C3:%[0-9]+]]:_(s32) = G_FCONSTANT float 0x41EFFFFFC0000000
+    ; GFX6: [[FMUL:%[0-9]+]]:_(s32) = G_FMUL [[AMDGPU_RCP_IFLAG]], [[C3]]
+    ; GFX6: [[FPTOUI:%[0-9]+]]:_(s32) = G_FPTOUI [[FMUL]](s32)
+    ; GFX6: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; GFX6: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C4]], [[AND1]]
+    ; GFX6: [[MUL:%[0-9]+]]:_(s32) = G_MUL [[SUB]], [[FPTOUI]]
+    ; GFX6: [[UMULH:%[0-9]+]]:_(s32) = G_UMULH [[FPTOUI]], [[MUL]]
+    ; GFX6: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[FPTOUI]], [[UMULH]]
+    ; GFX6: [[UMULH1:%[0-9]+]]:_(s32) = G_UMULH [[AND]], [[ADD]]
+    ; GFX6: [[MUL1:%[0-9]+]]:_(s32) = G_MUL [[UMULH1]], [[AND1]]
+    ; GFX6: [[SUB1:%[0-9]+]]:_(s32) = G_SUB [[AND]], [[MUL1]]
+    ; GFX6: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX6: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(uge), [[SUB1]](s32), [[AND1]]
+    ; GFX6: [[SUB2:%[0-9]+]]:_(s32) = G_SUB [[SUB1]], [[AND1]]
+    ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[SUB2]], [[SUB1]]
+    ; GFX6: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(uge), [[SELECT]](s32), [[AND1]]
+    ; GFX6: [[SUB3:%[0-9]+]]:_(s32) = G_SUB [[SELECT]], [[AND1]]
+    ; GFX6: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[SUB3]], [[SELECT]]
+    ; GFX6: [[COPY4:%[0-9]+]]:_(s32) = COPY [[SELECT1]](s32)
+    ; GFX6: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C2]]
+    ; GFX6: [[COPY5:%[0-9]+]]:_(s32) = COPY [[COPY]](s32)
+    ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[COPY5]], [[AND2]](s32)
+    ; GFX6: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C]](s32)
+    ; GFX6: [[COPY7:%[0-9]+]]:_(s32) = COPY [[SELECT1]](s32)
+    ; GFX6: [[SUB4:%[0-9]+]]:_(s32) = G_SUB [[COPY6]], [[COPY7]]
+    ; GFX6: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C5]](s32)
+    ; GFX6: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C2]]
+    ; GFX6: [[COPY9:%[0-9]+]]:_(s32) = COPY [[COPY]](s32)
+    ; GFX6: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C2]]
+    ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[AND4]], [[AND3]](s32)
+    ; GFX6: [[COPY10:%[0-9]+]]:_(s32) = COPY [[SUB4]](s32)
+    ; GFX6: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C2]]
+    ; GFX6: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
+    ; GFX6: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C2]]
+    ; GFX6: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[AND6]], [[AND5]](s32)
+    ; GFX6: [[COPY12:%[0-9]+]]:_(s32) = COPY [[SHL]](s32)
+    ; GFX6: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
+    ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[COPY12]], [[COPY13]]
+    ; GFX6: [[COPY14:%[0-9]+]]:_(s32) = COPY [[OR]](s32)
+    ; GFX6: $sgpr0 = COPY [[COPY14]](s32)
+    ; GFX8-LABEL: name: rotl_i15
+    ; GFX8: liveins: $sgpr0, $sgpr1
+    ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $sgpr0
+    ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $sgpr1
+    ; GFX8: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 15
+    ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 32767
+    ; GFX8: [[COPY2:%[0-9]+]]:_(s32) = COPY [[COPY1]](s32)
+    ; GFX8: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
+    ; GFX8: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C]](s32)
+    ; GFX8: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
+    ; GFX8: [[UITOFP:%[0-9]+]]:_(s32) = G_UITOFP [[AND1]](s32)
+    ; GFX8: [[AMDGPU_RCP_IFLAG:%[0-9]+]]:_(s32) = G_AMDGPU_RCP_IFLAG [[UITOFP]](s32)
+    ; GFX8: [[C2:%[0-9]+]]:_(s32) = G_FCONSTANT float 0x41EFFFFFC0000000
+    ; GFX8: [[FMUL:%[0-9]+]]:_(s32) = G_FMUL [[AMDGPU_RCP_IFLAG]], [[C2]]
+    ; GFX8: [[FPTOUI:%[0-9]+]]:_(s32) = G_FPTOUI [[FMUL]](s32)
+    ; GFX8: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; GFX8: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C3]], [[AND1]]
+    ; GFX8: [[MUL:%[0-9]+]]:_(s32) = G_MUL [[SUB]], [[FPTOUI]]
+    ; GFX8: [[UMULH:%[0-9]+]]:_(s32) = G_UMULH [[FPTOUI]], [[MUL]]
+    ; GFX8: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[FPTOUI]], [[UMULH]]
+    ; GFX8: [[UMULH1:%[0-9]+]]:_(s32) = G_UMULH [[AND]], [[ADD]]
+    ; GFX8: [[MUL1:%[0-9]+]]:_(s32) = G_MUL [[UMULH1]], [[AND1]]
+    ; GFX8: [[SUB1:%[0-9]+]]:_(s32) = G_SUB [[AND]], [[MUL1]]
+    ; GFX8: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX8: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(uge), [[SUB1]](s32), [[AND1]]
+    ; GFX8: [[SUB2:%[0-9]+]]:_(s32) = G_SUB [[SUB1]], [[AND1]]
+    ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[SUB2]], [[SUB1]]
+    ; GFX8: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(uge), [[SELECT]](s32), [[AND1]]
+    ; GFX8: [[SUB3:%[0-9]+]]:_(s32) = G_SUB [[SELECT]], [[AND1]]
+    ; GFX8: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[SUB3]], [[SELECT]]
+    ; GFX8: [[COPY4:%[0-9]+]]:_(s32) = COPY [[SELECT1]](s32)
+    ; GFX8: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
+    ; GFX8: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; GFX8: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[AND2]](s32)
+    ; GFX8: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[TRUNC]], [[TRUNC1]](s16)
+    ; GFX8: [[C5:%[0-9]+]]:_(s16) = G_CONSTANT i16 14
+    ; GFX8: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[SELECT1]](s32)
+    ; GFX8: [[SUB4:%[0-9]+]]:_(s16) = G_SUB [[C5]], [[TRUNC2]]
+    ; GFX8: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C4]](s32)
+    ; GFX8: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
+    ; GFX8: [[C6:%[0-9]+]]:_(s16) = G_CONSTANT i16 32767
+    ; GFX8: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C6]]
+    ; GFX8: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[AND3]](s32)
+    ; GFX8: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[AND4]], [[TRUNC3]](s16)
+    ; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SUB4]](s16)
+    ; GFX8: [[AND5:%[0-9]+]]:_(s32) = G_AND [[ANYEXT]], [[C1]]
+    ; GFX8: [[COPY6:%[0-9]+]]:_(s16) = COPY [[LSHR]](s16)
+    ; GFX8: [[AND6:%[0-9]+]]:_(s16) = G_AND [[COPY6]], [[C6]]
+    ; GFX8: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[AND5]](s32)
+    ; GFX8: [[LSHR1:%[0-9]+]]:_(s16) = G_LSHR [[AND6]], [[TRUNC4]](s16)
+    ; GFX8: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[SHL]](s16)
+    ; GFX8: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[LSHR1]](s16)
+    ; GFX8: [[OR:%[0-9]+]]:_(s32) = G_OR [[ANYEXT1]], [[ANYEXT2]]
+    ; GFX8: [[COPY7:%[0-9]+]]:_(s32) = COPY [[OR]](s32)
+    ; GFX8: $sgpr0 = COPY [[COPY7]](s32)
+    %2:_(s32) = COPY $sgpr0
+    %0:_(s15) = G_TRUNC %2(s32)
+    %3:_(s32) = COPY $sgpr1
+    %1:_(s15) = G_TRUNC %3(s32)
+    %5:_(s15) = G_ROTL %0, %1(s15)
+    %4:_(s32) = G_ANYEXT %5(s15)
+    $sgpr0 = COPY %4
+
+...
+---
+name:            rotl_i16
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0, $sgpr1
+
+    ; GFX6-LABEL: name: rotl_i16
+    ; GFX6: liveins: $sgpr0, $sgpr1
+    ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $sgpr0
+    ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $sgpr1
+    ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
+    ; GFX6: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 15
+    ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; GFX6: [[COPY2:%[0-9]+]]:_(s32) = COPY [[COPY1]](s32)
+    ; GFX6: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C1]], [[COPY2]]
+    ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SUB]](s32)
+    ; GFX6: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C]]
+    ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[AND]](s16)
+    ; GFX6: [[COPY3:%[0-9]+]]:_(s32) = COPY [[COPY]](s32)
+    ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[COPY3]], [[ZEXT]](s32)
+    ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; GFX6: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C]]
+    ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[AND1]](s16)
+    ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
+    ; GFX6: [[COPY4:%[0-9]+]]:_(s32) = COPY [[COPY]](s32)
+    ; GFX6: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C2]]
+    ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[AND2]], [[ZEXT1]](s32)
+    ; GFX6: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX6: [[OR:%[0-9]+]]:_(s16) = G_OR [[TRUNC2]], [[TRUNC3]]
+    ; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; GFX6: $sgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX8-LABEL: name: rotl_i16
+    ; GFX8: liveins: $sgpr0, $sgpr1
+    ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $sgpr0
+    ; GFX8: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $sgpr1
+    ; GFX8: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
+    ; GFX8: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 0
+    ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 15
+    ; GFX8: [[SUB:%[0-9]+]]:_(s16) = G_SUB [[C]], [[TRUNC1]]
+    ; GFX8: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; GFX8: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[TRUNC]], [[AND]](s16)
+    ; GFX8: [[AND1:%[0-9]+]]:_(s16) = G_AND [[SUB]], [[C1]]
+    ; GFX8: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC]], [[AND1]](s16)
+    ; GFX8: [[OR:%[0-9]+]]:_(s16) = G_OR [[SHL]], [[LSHR]]
+    ; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; GFX8: $sgpr0 = COPY [[ANYEXT]](s32)
+    %2:_(s32) = COPY $sgpr0
+    %0:_(s16) = G_TRUNC %2(s32)
+    %3:_(s32) = COPY $sgpr1
+    %1:_(s16) = G_TRUNC %3(s32)
+    %5:_(s16) = G_ROTL %0, %1(s16)
+    %4:_(s32) = G_ANYEXT %5(s16)
+    $sgpr0 = COPY %4
+
+...
+---
+name:            rotl_i32
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0, $sgpr1
+
+    ; GFX-LABEL: name: rotl_i32
+    ; GFX: liveins: $sgpr0, $sgpr1
+    ; GFX: [[COPY:%[0-9]+]]:_(s32) = COPY $sgpr0
+    ; GFX: [[COPY1:%[0-9]+]]:_(s32) = COPY $sgpr1
+    ; GFX: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX: [[FSHR:%[0-9]+]]:_(s32) = G_FSHR [[COPY]], [[COPY]], [[C]](s32)
+    ; GFX: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[COPY]], [[C]](s32)
+    ; GFX: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 -1
+    ; GFX: [[XOR:%[0-9]+]]:_(s32) = G_XOR [[COPY1]], [[C1]]
+    ; GFX: [[FSHR1:%[0-9]+]]:_(s32) = G_FSHR [[LSHR]], [[FSHR]], [[XOR]](s32)
+    ; GFX: $sgpr0 = COPY [[FSHR1]](s32)
+    %0:_(s32) = COPY $sgpr0
+    %1:_(s32) = COPY $sgpr1
+    %2:_(s32) = G_ROTL %0, %1(s32)
+    $sgpr0 = COPY %2
+
+...
+---
+name:            rotl_i31
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0, $sgpr1
+
+    ; GFX-LABEL: name: rotl_i31
+    ; GFX: liveins: $sgpr0, $sgpr1
+    ; GFX: [[COPY:%[0-9]+]]:_(s32) = COPY $sgpr0
+    ; GFX: [[COPY1:%[0-9]+]]:_(s32) = COPY $sgpr1
+    ; GFX: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 30
+    ; GFX: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
+    ; GFX: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 2147483647
+    ; GFX: [[COPY2:%[0-9]+]]:_(s32) = COPY [[COPY1]](s32)
+    ; GFX: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C2]]
+    ; GFX: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C1]](s32)
+    ; GFX: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C2]]
+    ; GFX: [[UITOFP:%[0-9]+]]:_(s32) = G_UITOFP [[AND1]](s32)
+    ; GFX: [[AMDGPU_RCP_IFLAG:%[0-9]+]]:_(s32) = G_AMDGPU_RCP_IFLAG [[UITOFP]](s32)
+    ; GFX: [[C3:%[0-9]+]]:_(s32) = G_FCONSTANT float 0x41EFFFFFC0000000
+    ; GFX: [[FMUL:%[0-9]+]]:_(s32) = G_FMUL [[AMDGPU_RCP_IFLAG]], [[C3]]
+    ; GFX: [[FPTOUI:%[0-9]+]]:_(s32) = G_FPTOUI [[FMUL]](s32)
+    ; GFX: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; GFX: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C4]], [[AND1]]
+    ; GFX: [[MUL:%[0-9]+]]:_(s32) = G_MUL [[SUB]], [[FPTOUI]]
+    ; GFX: [[UMULH:%[0-9]+]]:_(s32) = G_UMULH [[FPTOUI]], [[MUL]]
+    ; GFX: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[FPTOUI]], [[UMULH]]
+    ; GFX: [[UMULH1:%[0-9]+]]:_(s32) = G_UMULH [[AND]], [[ADD]]
+    ; GFX: [[MUL1:%[0-9]+]]:_(s32) = G_MUL [[UMULH1]], [[AND1]]
+    ; GFX: [[SUB1:%[0-9]+]]:_(s32) = G_SUB [[AND]], [[MUL1]]
+    ; GFX: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(uge), [[SUB1]](s32), [[AND1]]
+    ; GFX: [[SUB2:%[0-9]+]]:_(s32) = G_SUB [[SUB1]], [[AND1]]
+    ; GFX: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[SUB2]], [[SUB1]]
+    ; GFX: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(uge), [[SELECT]](s32), [[AND1]]
+    ; GFX: [[SUB3:%[0-9]+]]:_(s32) = G_SUB [[SELECT]], [[AND1]]
+    ; GFX: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[SUB3]], [[SELECT]]
+    ; GFX: [[COPY4:%[0-9]+]]:_(s32) = COPY [[SELECT1]](s32)
+    ; GFX: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C2]]
+    ; GFX: [[COPY5:%[0-9]+]]:_(s32) = COPY [[COPY]](s32)
+    ; GFX: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[COPY5]], [[AND2]](s32)
+    ; GFX: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C]](s32)
+    ; GFX: [[COPY7:%[0-9]+]]:_(s32) = COPY [[SELECT1]](s32)
+    ; GFX: [[SUB4:%[0-9]+]]:_(s32) = G_SUB [[COPY6]], [[COPY7]]
+    ; GFX: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C5]](s32)
+    ; GFX: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C2]]
+    ; GFX: [[COPY9:%[0-9]+]]:_(s32) = COPY [[COPY]](s32)
+    ; GFX: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C2]]
+    ; GFX: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[AND4]], [[AND3]](s32)
+    ; GFX: [[COPY10:%[0-9]+]]:_(s32) = COPY [[SUB4]](s32)
+    ; GFX: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C2]]
+    ; GFX: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
+    ; GFX: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C2]]
+    ; GFX: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[AND6]], [[AND5]](s32)
+    ; GFX: [[COPY12:%[0-9]+]]:_(s32) = COPY [[SHL]](s32)
+    ; GFX: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
+    ; GFX: [[OR:%[0-9]+]]:_(s32) = G_OR [[COPY12]], [[COPY13]]
+    ; GFX: [[COPY14:%[0-9]+]]:_(s32) = COPY [[OR]](s32)
+    ; GFX: $sgpr0 = COPY [[COPY14]](s32)
+    %0:_(s32) = COPY $sgpr0
+    %1:_(s32) = COPY $sgpr1
+    %2:_(s31) = G_TRUNC %0(s32)
+    %3:_(s31) = G_TRUNC %1(s32)
+    %4:_(s31) = G_ROTL %2, %3(s31)
+    %5:_(s32) = G_ANYEXT %4(s31)
+    $sgpr0 = COPY %5
+
+...
+---
+name:            rotl_i64
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1, $sgpr2_sgpr3
+
+    ; GFX-LABEL: name: rotl_i64
+    ; GFX: liveins: $sgpr0_sgpr1, $sgpr2_sgpr3
+    ; GFX: [[COPY:%[0-9]+]]:_(s64) = COPY $sgpr0_sgpr1
+    ; GFX: [[COPY1:%[0-9]+]]:_(s64) = COPY $sgpr2_sgpr3
+    ; GFX: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+    ; GFX: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 63
+    ; GFX: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[C]](s64)
+    ; GFX: [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY1]](s64)
+    ; GFX: [[USUBO:%[0-9]+]]:_(s32), [[USUBO1:%[0-9]+]]:_(s1) = G_USUBO [[UV]], [[UV2]]
+    ; GFX: [[USUBE:%[0-9]+]]:_(s32), [[USUBE1:%[0-9]+]]:_(s1) = G_USUBE [[UV1]], [[UV3]], [[USUBO1]]
+    ; GFX: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[USUBO]](s32), [[USUBE]](s32)
+    ; GFX: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C1]]
+    ; GFX: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[AND]](s64)
+    ; GFX: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[COPY]], [[TRUNC]](s32)
+    ; GFX: [[AND1:%[0-9]+]]:_(s64) = G_AND [[MV]], [[C1]]
+    ; GFX: [[TRUNC1:%[0-9]+]]:_(s32) = G_TRUNC [[AND1]](s64)
+    ; GFX: [[LSHR:%[0-9]+]]:_(s64) = G_LSHR [[COPY]], [[TRUNC1]](s32)
+    ; GFX: [[OR:%[0-9]+]]:_(s64) = G_OR [[SHL]], [[LSHR]]
+    ; GFX: $sgpr0_sgpr1 = COPY [[OR]](s64)
+    %0:_(s64) = COPY $sgpr0_sgpr1
+    %1:_(s64) = COPY $sgpr2_sgpr3
+    %2:_(s64) = G_ROTL %0, %1(s64)
+    $sgpr0_sgpr1 = COPY %2
+
+...
+---
+name:            rotl_v4i32
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4_sgpr5_sgpr6_sgpr7
+
+    ; GFX-LABEL: name: rotl_v4i32
+    ; GFX: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4_sgpr5_sgpr6_sgpr7
+    ; GFX: [[COPY:%[0-9]+]]:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    ; GFX: [[COPY1:%[0-9]+]]:_(<4 x s32>) = COPY $sgpr4_sgpr5_sgpr6_sgpr7
+    ; GFX: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<4 x s32>)
+    ; GFX: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY1]](<4 x s32>)
+    ; GFX: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
+    ; GFX: [[FSHR:%[0-9]+]]:_(s32) = G_FSHR [[UV]], [[UV]], [[C]](s32)
+    ; GFX: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
+    ; GFX: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 -1
+    ; GFX: [[XOR:%[0-9]+]]:_(s32) = G_XOR [[UV4]], [[C1]]
+    ; GFX: [[FSHR1:%[0-9]+]]:_(s32) = G_FSHR [[LSHR]], [[FSHR]], [[XOR]](s32)
+    ; GFX: [[FSHR2:%[0-9]+]]:_(s32) = G_FSHR [[UV1]], [[UV1]], [[C]](s32)
+    ; GFX: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
+    ; GFX: [[XOR1:%[0-9]+]]:_(s32) = G_XOR [[UV5]], [[C1]]
+    ; GFX: [[FSHR3:%[0-9]+]]:_(s32) = G_FSHR [[LSHR1]], [[FSHR2]], [[XOR1]](s32)
+    ; GFX: [[FSHR4:%[0-9]+]]:_(s32) = G_FSHR [[UV2]], [[UV2]], [[C]](s32)
+    ; GFX: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
+    ; GFX: [[XOR2:%[0-9]+]]:_(s32) = G_XOR [[UV6]], [[C1]]
+    ; GFX: [[FSHR5:%[0-9]+]]:_(s32) = G_FSHR [[LSHR2]], [[FSHR4]], [[XOR2]](s32)
+    ; GFX: [[FSHR6:%[0-9]+]]:_(s32) = G_FSHR [[UV3]], [[UV3]], [[C]](s32)
+    ; GFX: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
+    ; GFX: [[XOR3:%[0-9]+]]:_(s32) = G_XOR [[UV7]], [[C1]]
+    ; GFX: [[FSHR7:%[0-9]+]]:_(s32) = G_FSHR [[LSHR3]], [[FSHR6]], [[XOR3]](s32)
+    ; GFX: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[FSHR1]](s32), [[FSHR3]](s32), [[FSHR5]](s32), [[FSHR7]](s32)
+    ; GFX: $sgpr0_sgpr1_sgpr2_sgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    %0:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:_(<4 x s32>) = COPY $sgpr4_sgpr5_sgpr6_sgpr7
+    %2:_(<4 x s32>) = G_ROTL %0, %1(<4 x s32>)
+    $sgpr0_sgpr1_sgpr2_sgpr3 = COPY %2
+
+...
+---
+name:            rotr_i16
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0, $sgpr1
+
+    ; GFX6-LABEL: name: rotr_i16
+    ; GFX6: liveins: $sgpr0, $sgpr1
+    ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $sgpr0
+    ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $sgpr1
+    ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
+    ; GFX6: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 15
+    ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; GFX6: [[COPY2:%[0-9]+]]:_(s32) = COPY [[COPY1]](s32)
+    ; GFX6: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C1]], [[COPY2]]
+    ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SUB]](s32)
+    ; GFX6: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C]]
+    ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[AND]](s16)
+    ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
+    ; GFX6: [[COPY3:%[0-9]+]]:_(s32) = COPY [[COPY]](s32)
+    ; GFX6: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C2]]
+    ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[AND1]], [[ZEXT]](s32)
+    ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX6: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C]]
+    ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[AND2]](s16)
+    ; GFX6: [[COPY4:%[0-9]+]]:_(s32) = COPY [[COPY]](s32)
+    ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[COPY4]], [[ZEXT1]](s32)
+    ; GFX6: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; GFX6: [[OR:%[0-9]+]]:_(s16) = G_OR [[TRUNC2]], [[TRUNC3]]
+    ; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; GFX6: $sgpr0 = COPY [[ANYEXT]](s32)
+    ; GFX8-LABEL: name: rotr_i16
+    ; GFX8: liveins: $sgpr0, $sgpr1
+    ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $sgpr0
+    ; GFX8: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $sgpr1
+    ; GFX8: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
+    ; GFX8: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 0
+    ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 15
+    ; GFX8: [[SUB:%[0-9]+]]:_(s16) = G_SUB [[C]], [[TRUNC1]]
+    ; GFX8: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
+    ; GFX8: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC]], [[AND]](s16)
+    ; GFX8: [[AND1:%[0-9]+]]:_(s16) = G_AND [[SUB]], [[C1]]
+    ; GFX8: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[TRUNC]], [[AND1]](s16)
+    ; GFX8: [[OR:%[0-9]+]]:_(s16) = G_OR [[LSHR]], [[SHL]]
+    ; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
+    ; GFX8: $sgpr0 = COPY [[ANYEXT]](s32)
+    %2:_(s32) = COPY $sgpr0
+    %0:_(s16) = G_TRUNC %2(s32)
+    %3:_(s32) = COPY $sgpr1
+    %1:_(s16) = G_TRUNC %3(s32)
+    %5:_(s16) = G_ROTR %0, %1(s16)
+    %4:_(s32) = G_ANYEXT %5(s16)
+    $sgpr0 = COPY %4
+
+...
+---
+name:            rotr_i32
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0, $sgpr1
+
+    ; GFX-LABEL: name: rotr_i32
+    ; GFX: liveins: $sgpr0, $sgpr1
+    ; GFX: [[COPY:%[0-9]+]]:_(s32) = COPY $sgpr0
+    ; GFX: [[COPY1:%[0-9]+]]:_(s32) = COPY $sgpr1
+    ; GFX: [[FSHR:%[0-9]+]]:_(s32) = G_FSHR [[COPY]], [[COPY]], [[COPY1]](s32)
+    ; GFX: $sgpr0 = COPY [[FSHR]](s32)
+    %0:_(s32) = COPY $sgpr0
+    %1:_(s32) = COPY $sgpr1
+    %2:_(s32) = G_ROTR %0, %1(s32)
+    $sgpr0 = COPY %2
+
+...
+---
+name:            rotr_i64
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1, $sgpr2_sgpr3
+
+    ; GFX-LABEL: name: rotr_i64
+    ; GFX: liveins: $sgpr0_sgpr1, $sgpr2_sgpr3
+    ; GFX: [[COPY:%[0-9]+]]:_(s64) = COPY $sgpr0_sgpr1
+    ; GFX: [[COPY1:%[0-9]+]]:_(s64) = COPY $sgpr2_sgpr3
+    ; GFX: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+    ; GFX: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 63
+    ; GFX: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[C]](s64)
+    ; GFX: [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY1]](s64)
+    ; GFX: [[USUBO:%[0-9]+]]:_(s32), [[USUBO1:%[0-9]+]]:_(s1) = G_USUBO [[UV]], [[UV2]]
+    ; GFX: [[USUBE:%[0-9]+]]:_(s32), [[USUBE1:%[0-9]+]]:_(s1) = G_USUBE [[UV1]], [[UV3]], [[USUBO1]]
+    ; GFX: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[USUBO]](s32), [[USUBE]](s32)
+    ; GFX: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C1]]
+    ; GFX: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[AND]](s64)
+    ; GFX: [[LSHR:%[0-9]+]]:_(s64) = G_LSHR [[COPY]], [[TRUNC]](s32)
+    ; GFX: [[AND1:%[0-9]+]]:_(s64) = G_AND [[MV]], [[C1]]
+    ; GFX: [[TRUNC1:%[0-9]+]]:_(s32) = G_TRUNC [[AND1]](s64)
+    ; GFX: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[COPY]], [[TRUNC1]](s32)
+    ; GFX: [[OR:%[0-9]+]]:_(s64) = G_OR [[LSHR]], [[SHL]]
+    ; GFX: $sgpr0_sgpr1 = COPY [[OR]](s64)
+    %0:_(s64) = COPY $sgpr0_sgpr1
+    %1:_(s64) = COPY $sgpr2_sgpr3
+    %2:_(s64) = G_ROTR %0, %1(s64)
+    $sgpr0_sgpr1 = COPY %2
+
+...
+---
+name:            rotr_v4i32
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4_sgpr5_sgpr6_sgpr7
+
+    ; GFX-LABEL: name: rotr_v4i32
+    ; GFX: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4_sgpr5_sgpr6_sgpr7
+    ; GFX: [[COPY:%[0-9]+]]:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    ; GFX: [[COPY1:%[0-9]+]]:_(<4 x s32>) = COPY $sgpr4_sgpr5_sgpr6_sgpr7
+    ; GFX: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<4 x s32>)
+    ; GFX: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY1]](<4 x s32>)
+    ; GFX: [[FSHR:%[0-9]+]]:_(s32) = G_FSHR [[UV]], [[UV]], [[UV4]](s32)
+    ; GFX: [[FSHR1:%[0-9]+]]:_(s32) = G_FSHR [[UV1]], [[UV1]], [[UV5]](s32)
+    ; GFX: [[FSHR2:%[0-9]+]]:_(s32) = G_FSHR [[UV2]], [[UV2]], [[UV6]](s32)
+    ; GFX: [[FSHR3:%[0-9]+]]:_(s32) = G_FSHR [[UV3]], [[UV3]], [[UV7]](s32)
+    ; GFX: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[FSHR]](s32), [[FSHR1]](s32), [[FSHR2]](s32), [[FSHR3]](s32)
+    ; GFX: $sgpr0_sgpr1_sgpr2_sgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>)
+    %0:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:_(<4 x s32>) = COPY $sgpr4_sgpr5_sgpr6_sgpr7
+    %2:_(<4 x s32>) = G_ROTR %0, %1(<4 x s32>)
+    $sgpr0_sgpr1_sgpr2_sgpr3 = COPY %2
+
+...