[LoongArch] Add codegen support for ISD::ROTL and ISD::ROTR

[xen0n]

llvm/lib/Target/LoongArch/LoongArchISelLowering.h

38–39

Are you placing ROTR before ROTL because it's the native instruction? Because alphabetically it would be the opposite.

Not a big concern though.

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

6

Isn't N-bit left-rotates equivalent to GRLen - N-bit right-rotates? I think you can just transform ROTL to equivalent ROTR then get optimized for free. In other words, at instruction selection time there shouldn't be any ROTLs left around.

[gonglingqin]

llvm/lib/Target/LoongArch/LoongArchISelLowering.h

38–39

Yes, thank you for your reminder. I'll sort it alphabetically

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

6

Yes, DAGCombiner has done this optimization. For example, when rotl_32 function is compiled with the --mtriple=loongarch32 option, DAGCombiner uses ISD::rotr, and the same optimization happens in other functions, such as rotl_i32_fshl. However, in some cases, such as when compiling rotl_32 functions with the --mtriple=loongarch64 option, the backend needs to add the optimizations you mentioned. In order to avoid this patch being too large to increase the difficulty of review, I will implement this optimization in another patch, thanks:)

[gonglingqin]

Address @xen0n's comments.

[xen0n]

I assume the remaining unoptimized LA64 cases are to be tackled in a later patch. For now this looks good. Thanks!

[RKSimon]

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

375

@gonglingqin You might want to review theses tests as the out of bounds shifts means they generate poison (which is exploited by rGae60706da07a128) - maybe use fshl/fshr intrinsics?

llvm/lib/Target/LoongArch/LoongArchISelLowering.h

enum NodeType : unsigned {
   CALL,
   RET,
   // 32-bit shifts, directly matching the semantics of the named LoongArch
   // instructions.
   SLL_W,
   SRA_W,
   SRL_W,
 
+  ROTL_W,
+  ROTR_W,
+
   // FPR<->GPR transfer operations
   MOVGR2FR_W_LA64,
   MOVFR2GR_S_LA64,
 
   FTINT,
 
   BSTRINS,
   BSTRPICK,

llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp

LoongArchTargetLowering::LoongArchTargetLowering(const TargetMachine &TM,
   setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, GRLenVT,
                    MVT::i1, Promote);
 
   // TODO: add necessary setOperationAction calls later.
   setOperationAction(ISD::SHL_PARTS, GRLenVT, Custom);
   setOperationAction(ISD::SRA_PARTS, GRLenVT, Custom);
   setOperationAction(ISD::SRL_PARTS, GRLenVT, Custom);
   setOperationAction(ISD::FP_TO_SINT, GRLenVT, Custom);
+  setOperationAction(ISD::ROTL, GRLenVT, Expand);
 
   setOperationAction({ISD::GlobalAddress, ISD::ConstantPool}, GRLenVT, Custom);
 
   if (Subtarget.is64Bit()) {
     setOperationAction(ISD::SHL, MVT::i32, Custom);
     setOperationAction(ISD::SRA, MVT::i32, Custom);
     setOperationAction(ISD::SRL, MVT::i32, Custom);
     setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
     setOperationAction(ISD::BITCAST, MVT::i32, Custom);
+    setOperationAction(ISD::ROTR, MVT::i32, Custom);
+    setOperationAction(ISD::ROTL, MVT::i32, Custom);
     if (Subtarget.hasBasicF() && !Subtarget.hasBasicD())
       setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
   }
 
   static const ISD::CondCode FPCCToExpand[] = {ISD::SETOGT, ISD::SETOGE,
                                                ISD::SETUGT, ISD::SETUGE};
 
   if (Subtarget.hasBasicF()) {
static LoongArchISD::NodeType getLoongArchWOpcode(unsigned Opcode) {
   default:
     llvm_unreachable("Unexpected opcode");
   case ISD::SHL:
     return LoongArchISD::SLL_W;
   case ISD::SRA:
     return LoongArchISD::SRA_W;
   case ISD::SRL:
     return LoongArchISD::SRL_W;
+  case ISD::ROTR:
+    return LoongArchISD::ROTR_W;
+  case ISD::ROTL:
+    return LoongArchISD::ROTL_W;
   }
 }
 
 // Converts the given i8/i16/i32 operation to a target-specific SelectionDAG
 // node. Because i8/i16/i32 isn't a legal type for LA64, these operations would
 // otherwise be promoted to i64, making it difficult to select the
 // SLL_W/.../*W later one because the fact the operation was originally of
 // type i8/i16/i32 is lost.
void LoongArchTargetLowering::ReplaceNodeResults(
     SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
   SDLoc DL(N);
   switch (N->getOpcode()) {
   default:
     llvm_unreachable("Don't know how to legalize this operation");
   case ISD::SHL:
   case ISD::SRA:
   case ISD::SRL:
+  case ISD::ROTR:
     assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
            "Unexpected custom legalisation");
     if (N->getOperand(1).getOpcode() != ISD::Constant) {
       Results.push_back(customLegalizeToWOp(N, DAG));
       break;
     }
     break;
+  case ISD::ROTL:
+    ConstantSDNode *CN;
+    if ((CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))) {
+      Results.push_back(customLegalizeToWOp(N, DAG));
+      break;
+    }
+    break;
   case ISD::FP_TO_SINT: {
     assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
            "Unexpected custom legalisation");
     SDValue Src = N->getOperand(0);
     EVT VT = EVT::getFloatingPointVT(N->getValueSizeInBits(0));
     SDValue Dst = DAG.getNode(LoongArchISD::FTINT, DL, VT, Src);
     Results.push_back(DAG.getNode(ISD::BITCAST, DL, N->getValueType(0), Dst));
     break;
case LoongArchISD::node:                                                     \
     NODE_NAME_CASE(SLL_W)
     NODE_NAME_CASE(SRA_W)
     NODE_NAME_CASE(SRL_W)
     NODE_NAME_CASE(BSTRINS)
     NODE_NAME_CASE(BSTRPICK)
     NODE_NAME_CASE(MOVGR2FR_W_LA64)
     NODE_NAME_CASE(MOVFR2GR_S_LA64)
     NODE_NAME_CASE(FTINT)
+    NODE_NAME_CASE(ROTR_W)
+    NODE_NAME_CASE(ROTL_W)
   }
 #undef NODE_NAME_CASE
   return nullptr;
 }
 
 //===----------------------------------------------------------------------===//
 //                     Calling Convention Implementation
 //===----------------------------------------------------------------------===//

llvm/lib/Target/LoongArch/LoongArchInstrInfo.td

 def loongarch_call : SDNode<"LoongArchISD::CALL", SDT_LoongArchCall,
                             [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue,
                              SDNPVariadic]>;
 def loongarch_ret : SDNode<"LoongArchISD::RET", SDTNone,
                            [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
 def loongarch_sll_w : SDNode<"LoongArchISD::SLL_W", SDT_LoongArchIntBinOpW>;
 def loongarch_sra_w : SDNode<"LoongArchISD::SRA_W", SDT_LoongArchIntBinOpW>;
 def loongarch_srl_w : SDNode<"LoongArchISD::SRL_W", SDT_LoongArchIntBinOpW>;
+def loongarch_rotr_w : SDNode<"LoongArchISD::ROTR_W", SDT_LoongArchIntBinOpW>;
+def loongarch_rotl_w : SDNode<"LoongArchISD::ROTL_W", SDT_LoongArchIntBinOpW>;
 def loongarch_bstrins
     : SDNode<"LoongArchISD::BSTRINS", SDT_LoongArchBStrIns>;
 def loongarch_bstrpick
     : SDNode<"LoongArchISD::BSTRPICK", SDT_LoongArchBStrPick>;
 
 //===----------------------------------------------------------------------===//
 // Operand and SDNode transformation definitions.
 //===----------------------------------------------------------------------===//
return CurDAG->getTargetConstant(-N->getSExtValue(), SDLoc(N),
                                    N->getValueType(0));
 }]>;
 
 // FP immediate patterns.
 def fpimm0    : PatLeaf<(fpimm), [{return N->isExactlyValue(+0.0);}]>;
 def fpimm0neg : PatLeaf<(fpimm), [{return N->isExactlyValue(-0.0);}]>;
 def fpimm1    : PatLeaf<(fpimm), [{return N->isExactlyValue(+1.0);}]>;
 
+// Return an immediate subtracted from 32.
+def ImmSubFrom32 : SDNodeXForm<imm, [{
+  return CurDAG->getTargetConstant(32 - N->getZExtValue(), SDLoc(N),
+                                   N->getValueType(0));
+}]>;
+
 def CallSymbol: AsmOperandClass {
   let Name = "CallSymbol";
   let RenderMethod = "addImmOperands";
   let PredicateMethod = "isImm";
 }
 
 // A bare symbol used in call only.
 def call_symbol : Operand<iPTR> {

 def : PatGprGpr<sub, SUB_W>;
 def : PatGprGpr<sdiv, DIV_W>;
 def : PatGprGpr<udiv, DIV_WU>;
 def : PatGprGpr<srem, MOD_W>;
 def : PatGprGpr<urem, MOD_WU>;
 def : PatGprGpr<mul, MUL_W>;
 def : PatGprGpr<mulhs, MULH_W>;
 def : PatGprGpr<mulhu, MULH_WU>;
+def : PatGprGpr<rotr, ROTR_W>;
+def : PatGprImm<rotr, ROTRI_W, uimm5>;
 } // Predicates = [IsLA32]
 
 let Predicates = [IsLA64] in {
 def : PatGprGpr<add, ADD_D>;
 def : PatGprGpr_32<add, ADD_W>;
 def : PatGprImm<add, ADDI_D, simm12>;
 def : PatGprImm_32<add, ADDI_W, simm12>;
 def : PatGprGpr<sub, SUB_D>;
 def : PatGprGpr_32<sub, SUB_W>;
 def : PatGprGpr<sdiv, DIV_D>;
 def : PatGprGpr<udiv, DIV_DU>;
 def : PatGprGpr<srem, MOD_D>;
 def : PatGprGpr<urem, MOD_DU>;
+def : PatGprGpr<rotr, ROTR_D>;
+def : PatGprGpr<loongarch_rotr_w, ROTR_W>;
+def : PatGprImm<rotr, ROTRI_D, uimm6>;
+def : PatGprImm_32<rotr, ROTRI_W, uimm5>;
+def : Pat<(loongarch_rotl_w GPR:$rj, uimm5:$imm),
+          (ROTRI_W GPR:$rj, (ImmSubFrom32 uimm5:$imm))>;
+def : Pat<(sext_inreg (loongarch_rotl_w GPR:$rj, uimm5:$imm), i32),
+          (ROTRI_W GPR:$rj, (ImmSubFrom32 uimm5:$imm))>;
 // TODO: Select "_W[U]" instructions for i32xi32 if only lower 32 bits of the
 // product are used.
 def : PatGprGpr<mul, MUL_D>;
 def : PatGprGpr<mulhs, MULH_D>;
 def : PatGprGpr<mulhu, MULH_DU>;
 // Select MULW_D_W for calculating the full 64 bits product of i32xi32 signed
 // multiplication.
 def : Pat<(i64 (mul (sext_inreg GPR:$rj, i32), (sext_inreg GPR:$rk, i32))),

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

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc --mtriple=loongarch32 < %s | FileCheck %s --check-prefix=LA32
+; RUN: llc --mtriple=loongarch64 < %s | FileCheck %s --check-prefix=LA64
+
+;; TODO: Add optimization to ISD::ROTL
+
+define i32 @rotl_32(i32 %x, i32 %y) nounwind {
+; LA32-LABEL: rotl_32:
+; LA32:       # %bb.0:
+; LA32-NEXT:    ori $a2, $zero, 32
+; LA32-NEXT:    sub.w $a1, $a2, $a1
+; LA32-NEXT:    rotr.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotl_32:
+; LA64:       # %bb.0:
+; LA64-NEXT:    sub.d $a2, $zero, $a1
+; LA64-NEXT:    sll.w $a1, $a0, $a1
+; LA64-NEXT:    srl.w $a0, $a0, $a2
+; LA64-NEXT:    or $a0, $a1, $a0
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %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 {
+; LA32-LABEL: rotr_32:
+; LA32:       # %bb.0:
+; LA32-NEXT:    rotr.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotr_32:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotr.w $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %z = sub i32 32, %y
+  %b = lshr i32 %x, %y
+  %c = shl i32 %x, %z
+  %d = or i32 %b, %c
+  ret i32 %d
+}
+
+define i64 @rotl_64(i64 %x, i64 %y) nounwind {
+; LA32-LABEL: rotl_64:
+; LA32:       # %bb.0:
+; LA32-NEXT:    xori $a3, $a2, 31
+; LA32-NEXT:    srli.w $a4, $a0, 1
+; LA32-NEXT:    srl.w $a3, $a4, $a3
+; LA32-NEXT:    sll.w $a4, $a1, $a2
+; LA32-NEXT:    or $a3, $a4, $a3
+; LA32-NEXT:    addi.w $a4, $a2, -32
+; LA32-NEXT:    slti $a5, $a4, 0
+; LA32-NEXT:    maskeqz $a3, $a3, $a5
+; LA32-NEXT:    sll.w $a6, $a0, $a4
+; LA32-NEXT:    masknez $a5, $a6, $a5
+; LA32-NEXT:    or $a3, $a3, $a5
+; LA32-NEXT:    ori $a5, $zero, 64
+; LA32-NEXT:    sub.w $a5, $a5, $a2
+; LA32-NEXT:    xori $a5, $a5, 31
+; LA32-NEXT:    slli.w $a6, $a1, 1
+; LA32-NEXT:    sll.w $a5, $a6, $a5
+; LA32-NEXT:    sub.w $a6, $zero, $a2
+; LA32-NEXT:    srl.w $a7, $a1, $a6
+; LA32-NEXT:    ori $a1, $zero, 32
+; LA32-NEXT:    sub.w $t0, $a1, $a2
+; LA32-NEXT:    srai.w $a1, $t0, 31
+; LA32-NEXT:    and $a1, $a1, $a7
+; LA32-NEXT:    or $a1, $a3, $a1
+; LA32-NEXT:    srl.w $a3, $a0, $a6
+; LA32-NEXT:    or $a3, $a3, $a5
+; LA32-NEXT:    slti $a5, $t0, 0
+; LA32-NEXT:    masknez $a6, $a7, $a5
+; LA32-NEXT:    maskeqz $a3, $a3, $a5
+; LA32-NEXT:    or $a3, $a3, $a6
+; LA32-NEXT:    sll.w $a0, $a0, $a2
+; LA32-NEXT:    srai.w $a2, $a4, 31
+; LA32-NEXT:    and $a0, $a2, $a0
+; LA32-NEXT:    or $a0, $a0, $a3
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotl_64:
+; LA64:       # %bb.0:
+; LA64-NEXT:    ori $a2, $zero, 64
+; LA64-NEXT:    sub.d $a1, $a2, $a1
+; LA64-NEXT:    rotr.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %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 {
+; LA32-LABEL: rotr_64:
+; LA32:       # %bb.0:
+; LA32-NEXT:    xori $a3, $a2, 31
+; LA32-NEXT:    slli.w $a4, $a1, 1
+; LA32-NEXT:    sll.w $a3, $a4, $a3
+; LA32-NEXT:    srl.w $a4, $a0, $a2
+; LA32-NEXT:    or $a3, $a4, $a3
+; LA32-NEXT:    addi.w $a4, $a2, -32
+; LA32-NEXT:    slti $a5, $a4, 0
+; LA32-NEXT:    maskeqz $a3, $a3, $a5
+; LA32-NEXT:    srl.w $a6, $a1, $a4
+; LA32-NEXT:    masknez $a5, $a6, $a5
+; LA32-NEXT:    or $a3, $a3, $a5
+; LA32-NEXT:    ori $a5, $zero, 64
+; LA32-NEXT:    sub.w $a5, $a5, $a2
+; LA32-NEXT:    xori $a5, $a5, 31
+; LA32-NEXT:    srli.w $a6, $a0, 1
+; LA32-NEXT:    srl.w $a5, $a6, $a5
+; LA32-NEXT:    sub.w $a6, $zero, $a2
+; LA32-NEXT:    sll.w $a7, $a0, $a6
+; LA32-NEXT:    ori $a0, $zero, 32
+; LA32-NEXT:    sub.w $t0, $a0, $a2
+; LA32-NEXT:    srai.w $a0, $t0, 31
+; LA32-NEXT:    and $a0, $a0, $a7
+; LA32-NEXT:    or $a0, $a3, $a0
+; LA32-NEXT:    sll.w $a3, $a1, $a6
+; LA32-NEXT:    or $a3, $a3, $a5
+; LA32-NEXT:    slti $a5, $t0, 0
+; LA32-NEXT:    masknez $a6, $a7, $a5
+; LA32-NEXT:    maskeqz $a3, $a3, $a5
+; LA32-NEXT:    or $a3, $a3, $a6
+; LA32-NEXT:    srl.w $a1, $a1, $a2
+; LA32-NEXT:    srai.w $a2, $a4, 31
+; LA32-NEXT:    and $a1, $a2, $a1
+; LA32-NEXT:    or $a1, $a1, $a3
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotr_64:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotr.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %z = sub i64 64, %y
+  %b = lshr i64 %x, %y
+  %c = shl i64 %x, %z
+  %d = or i64 %b, %c
+  ret i64 %d
+}
+
+define i32 @rotl_32_mask(i32 %x, i32 %y) nounwind {
+; LA32-LABEL: rotl_32_mask:
+; LA32:       # %bb.0:
+; LA32-NEXT:    sub.w $a1, $zero, $a1
+; LA32-NEXT:    rotr.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotl_32_mask:
+; LA64:       # %bb.0:
+; LA64-NEXT:    sub.d $a2, $zero, $a1
+; LA64-NEXT:    sll.w $a1, $a0, $a1
+; LA64-NEXT:    srl.w $a0, $a0, $a2
+; LA64-NEXT:    or $a0, $a1, $a0
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %z = sub i32 0, %y
+  %and = and i32 %z, 31
+  %b = shl i32 %x, %y
+  %c = lshr i32 %x, %and
+  %d = or i32 %b, %c
+  ret i32 %d
+}
+
+define i32 @rotl_32_mask_and_63_and_31(i32 %x, i32 %y) nounwind {
+; LA32-LABEL: rotl_32_mask_and_63_and_31:
+; LA32:       # %bb.0:
+; LA32-NEXT:    sub.w $a1, $zero, $a1
+; LA32-NEXT:    rotr.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotl_32_mask_and_63_and_31:
+; LA64:       # %bb.0:
+; LA64-NEXT:    sub.d $a2, $zero, $a1
+; LA64-NEXT:    sll.w $a1, $a0, $a1
+; LA64-NEXT:    srl.w $a0, $a0, $a2
+; LA64-NEXT:    or $a0, $a1, $a0
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %a = and i32 %y, 63
+  %b = shl i32 %x, %a
+  %c = sub i32 0, %y
+  %d = and i32 %c, 31
+  %e = lshr i32 %x, %d
+  %f = or i32 %b, %e
+  ret i32 %f
+}
+
+define i32 @rotl_32_mask_or_64_or_32(i32 %x, i32 %y) nounwind {
+; LA32-LABEL: rotl_32_mask_or_64_or_32:
+; LA32:       # %bb.0:
+; LA32-NEXT:    sub.w $a1, $zero, $a1
+; LA32-NEXT:    rotr.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotl_32_mask_or_64_or_32:
+; LA64:       # %bb.0:
+; LA64-NEXT:    sub.d $a2, $zero, $a1
+; LA64-NEXT:    sll.w $a1, $a0, $a1
+; LA64-NEXT:    srl.w $a0, $a0, $a2
+; LA64-NEXT:    or $a0, $a1, $a0
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %a = or i32 %y, 64
+  %b = shl i32 %x, %a
+  %c = sub i32 0, %y
+  %d = or i32 %c, 32
+  %e = lshr i32 %x, %d
+  %f = or i32 %b, %e
+  ret i32 %f
+}
+
+define i32 @rotr_32_mask(i32 %x, i32 %y) nounwind {
+; LA32-LABEL: rotr_32_mask:
+; LA32:       # %bb.0:
+; LA32-NEXT:    rotr.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotr_32_mask:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotr.w $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %z = sub i32 0, %y
+  %and = and i32 %z, 31
+  %b = lshr i32 %x, %y
+  %c = shl i32 %x, %and
+  %d = or i32 %b, %c
+  ret i32 %d
+}
+
+define i32 @rotr_32_mask_and_63_and_31(i32 %x, i32 %y) nounwind {
+; LA32-LABEL: rotr_32_mask_and_63_and_31:
+; LA32:       # %bb.0:
+; LA32-NEXT:    rotr.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotr_32_mask_and_63_and_31:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotr.w $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %a = and i32 %y, 63
+  %b = lshr i32 %x, %a
+  %c = sub i32 0, %y
+  %d = and i32 %c, 31
+  %e = shl i32 %x, %d
+  %f = or i32 %b, %e
+  ret i32 %f
+}
+
+define i32 @rotr_32_mask_or_64_or_32(i32 %x, i32 %y) nounwind {
+; LA32-LABEL: rotr_32_mask_or_64_or_32:
+; LA32:       # %bb.0:
+; LA32-NEXT:    rotr.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotr_32_mask_or_64_or_32:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotr.w $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %a = or i32 %y, 64
+  %b = lshr i32 %x, %a
+  %c = sub i32 0, %y
+  %d = or i32 %c, 32
+  %e = shl i32 %x, %d
+  %f = or i32 %b, %e
+  ret i32 %f
+}
+
+define i64 @rotl_64_mask(i64 %x, i64 %y) nounwind {
+; LA32-LABEL: rotl_64_mask:
+; LA32:       # %bb.0:
+; LA32-NEXT:    xori $a3, $a2, 31
+; LA32-NEXT:    srli.w $a4, $a0, 1
+; LA32-NEXT:    srl.w $a3, $a4, $a3
+; LA32-NEXT:    sll.w $a4, $a1, $a2
+; LA32-NEXT:    or $a3, $a4, $a3
+; LA32-NEXT:    sub.w $a4, $zero, $a2
+; LA32-NEXT:    srl.w $a5, $a1, $a4
+; LA32-NEXT:    andi $a6, $a4, 63
+; LA32-NEXT:    addi.w $a7, $a6, -32
+; LA32-NEXT:    srai.w $t0, $a7, 31
+; LA32-NEXT:    and $a5, $t0, $a5
+; LA32-NEXT:    addi.w $t0, $a2, -32
+; LA32-NEXT:    slti $t1, $t0, 0
+; LA32-NEXT:    maskeqz $a3, $a3, $t1
+; LA32-NEXT:    sll.w $t2, $a0, $t0
+; LA32-NEXT:    masknez $t1, $t2, $t1
+; LA32-NEXT:    or $a3, $a3, $t1
+; LA32-NEXT:    xori $a6, $a6, 31
+; LA32-NEXT:    slli.w $t1, $a1, 1
+; LA32-NEXT:    sll.w $a6, $t1, $a6
+; LA32-NEXT:    or $a3, $a3, $a5
+; LA32-NEXT:    srl.w $a4, $a0, $a4
+; LA32-NEXT:    or $a4, $a4, $a6
+; LA32-NEXT:    srl.w $a1, $a1, $a7
+; LA32-NEXT:    slti $a5, $a7, 0
+; LA32-NEXT:    masknez $a1, $a1, $a5
+; LA32-NEXT:    maskeqz $a4, $a4, $a5
+; LA32-NEXT:    or $a1, $a4, $a1
+; LA32-NEXT:    sll.w $a0, $a0, $a2
+; LA32-NEXT:    srai.w $a2, $t0, 31
+; LA32-NEXT:    and $a0, $a2, $a0
+; LA32-NEXT:    or $a0, $a0, $a1
+; LA32-NEXT:    move $a1, $a3
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotl_64_mask:
+; LA64:       # %bb.0:
+; LA64-NEXT:    sub.d $a1, $zero, $a1
+; LA64-NEXT:    rotr.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %z = sub i64 0, %y
+  %and = and i64 %z, 63
+  %b = shl i64 %x, %y
+  %c = lshr i64 %x, %and
+  %d = or i64 %b, %c
+  ret i64 %d
+}
+
+define i64 @rotl_64_mask_and_127_and_63(i64 %x, i64 %y) nounwind {
+; LA32-LABEL: rotl_64_mask_and_127_and_63:
+; LA32:       # %bb.0:
+; LA32-NEXT:    srli.w $a3, $a0, 1
+; LA32-NEXT:    andi $a4, $a2, 127
+; LA32-NEXT:    xori $a5, $a4, 31
+; LA32-NEXT:    srl.w $a3, $a3, $a5
+; LA32-NEXT:    sll.w $a5, $a1, $a2
+; LA32-NEXT:    or $a3, $a5, $a3
+; LA32-NEXT:    sub.w $a5, $zero, $a2
+; LA32-NEXT:    srl.w $a6, $a1, $a5
+; LA32-NEXT:    andi $a7, $a5, 63
+; LA32-NEXT:    addi.w $t0, $a7, -32
+; LA32-NEXT:    srai.w $t1, $t0, 31
+; LA32-NEXT:    and $a6, $t1, $a6
+; LA32-NEXT:    addi.w $a4, $a4, -32
+; LA32-NEXT:    slti $t1, $a4, 0
+; LA32-NEXT:    maskeqz $a3, $a3, $t1
+; LA32-NEXT:    sll.w $t2, $a0, $a4
+; LA32-NEXT:    masknez $t1, $t2, $t1
+; LA32-NEXT:    or $a3, $a3, $t1
+; LA32-NEXT:    xori $a7, $a7, 31
+; LA32-NEXT:    slli.w $t1, $a1, 1
+; LA32-NEXT:    sll.w $a7, $t1, $a7
+; LA32-NEXT:    or $a3, $a3, $a6
+; LA32-NEXT:    srl.w $a5, $a0, $a5
+; LA32-NEXT:    or $a5, $a5, $a7
+; LA32-NEXT:    srl.w $a1, $a1, $t0
+; LA32-NEXT:    slti $a6, $t0, 0
+; LA32-NEXT:    masknez $a1, $a1, $a6
+; LA32-NEXT:    maskeqz $a5, $a5, $a6
+; LA32-NEXT:    or $a1, $a5, $a1
+; LA32-NEXT:    sll.w $a0, $a0, $a2
+; LA32-NEXT:    srai.w $a2, $a4, 31
+; LA32-NEXT:    and $a0, $a2, $a0
+; LA32-NEXT:    or $a0, $a0, $a1
+; LA32-NEXT:    move $a1, $a3
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotl_64_mask_and_127_and_63:
+; LA64:       # %bb.0:
+; LA64-NEXT:    sub.d $a1, $zero, $a1
+; LA64-NEXT:    rotr.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %a = and i64 %y, 127
+  %b = shl i64 %x, %a
+  %c = sub i64 0, %y
+  %d = and i64 %c, 63
+  %e = lshr i64 %x, %d
+  %f = or i64 %b, %e
+  ret i64 %f
+}
+
+define i64 @rotl_64_mask_or_128_or_64(i64 %x, i64 %y) nounwind {
+; LA32-LABEL: rotl_64_mask_or_128_or_64:
+; LA32:       # %bb.0:
+; LA32-NEXT:    sll.w $a3, $a0, $a2
+; LA32-NEXT:    sub.w $a0, $zero, $a2
+; LA32-NEXT:    srl.w $a0, $a1, $a0
+; LA32-NEXT:    move $a1, $a3
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotl_64_mask_or_128_or_64:
+; LA64:       # %bb.0:
+; LA64-NEXT:    sub.d $a1, $zero, $a1
+; LA64-NEXT:    rotr.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %a = or i64 %y, 128
+  %b = shl i64 %x, %a
+  %c = sub i64 0, %y
+  %d = or i64 %c, 64
+  %e = lshr i64 %x, %d
+  %f = or i64 %b, %e
+  ret i64 %f
+}
+
+define i64 @rotr_64_mask(i64 %x, i64 %y) nounwind {
+; LA32-LABEL: rotr_64_mask:
+; LA32:       # %bb.0:
+; LA32-NEXT:    xori $a3, $a2, 31
+; LA32-NEXT:    slli.w $a4, $a1, 1
+; LA32-NEXT:    sll.w $a3, $a4, $a3
+; LA32-NEXT:    srl.w $a4, $a0, $a2
+; LA32-NEXT:    or $a3, $a4, $a3
+; LA32-NEXT:    sub.w $a4, $zero, $a2
+; LA32-NEXT:    sll.w $a5, $a0, $a4
+; LA32-NEXT:    andi $a6, $a4, 63
+; LA32-NEXT:    addi.w $a7, $a6, -32
+; LA32-NEXT:    srai.w $t0, $a7, 31
+; LA32-NEXT:    and $a5, $t0, $a5
+; LA32-NEXT:    addi.w $t0, $a2, -32
+; LA32-NEXT:    slti $t1, $t0, 0
+; LA32-NEXT:    maskeqz $a3, $a3, $t1
+; LA32-NEXT:    srl.w $t2, $a1, $t0
+; LA32-NEXT:    masknez $t1, $t2, $t1
+; LA32-NEXT:    or $a3, $a3, $t1
+; LA32-NEXT:    xori $a6, $a6, 31
+; LA32-NEXT:    srli.w $t1, $a0, 1
+; LA32-NEXT:    srl.w $a6, $t1, $a6
+; LA32-NEXT:    or $a3, $a3, $a5
+; LA32-NEXT:    sll.w $a4, $a1, $a4
+; LA32-NEXT:    or $a4, $a4, $a6
+; LA32-NEXT:    sll.w $a0, $a0, $a7
+; LA32-NEXT:    slti $a5, $a7, 0
+; LA32-NEXT:    masknez $a0, $a0, $a5
+; LA32-NEXT:    maskeqz $a4, $a4, $a5
+; LA32-NEXT:    or $a0, $a4, $a0
+; LA32-NEXT:    srl.w $a1, $a1, $a2
+; LA32-NEXT:    srai.w $a2, $t0, 31
+; LA32-NEXT:    and $a1, $a2, $a1
+; LA32-NEXT:    or $a1, $a1, $a0
+; LA32-NEXT:    move $a0, $a3
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotr_64_mask:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotr.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %z = sub i64 0, %y
+  %and = and i64 %z, 63
+  %b = lshr i64 %x, %y
+  %c = shl i64 %x, %and
+  %d = or i64 %b, %c
+  ret i64 %d
+}
+
+define i64 @rotr_64_mask_and_127_and_63(i64 %x, i64 %y) nounwind {
+; LA32-LABEL: rotr_64_mask_and_127_and_63:
+; LA32:       # %bb.0:
+; LA32-NEXT:    slli.w $a3, $a1, 1
+; LA32-NEXT:    andi $a4, $a2, 127
+; LA32-NEXT:    xori $a5, $a4, 31
+; LA32-NEXT:    sll.w $a3, $a3, $a5
+; LA32-NEXT:    srl.w $a5, $a0, $a2
+; LA32-NEXT:    or $a3, $a5, $a3
+; LA32-NEXT:    sub.w $a5, $zero, $a2
+; LA32-NEXT:    sll.w $a6, $a0, $a5
+; LA32-NEXT:    andi $a7, $a5, 63
+; LA32-NEXT:    addi.w $t0, $a7, -32
+; LA32-NEXT:    srai.w $t1, $t0, 31
+; LA32-NEXT:    and $a6, $t1, $a6
+; LA32-NEXT:    addi.w $a4, $a4, -32
+; LA32-NEXT:    slti $t1, $a4, 0
+; LA32-NEXT:    maskeqz $a3, $a3, $t1
+; LA32-NEXT:    srl.w $t2, $a1, $a4
+; LA32-NEXT:    masknez $t1, $t2, $t1
+; LA32-NEXT:    or $a3, $a3, $t1
+; LA32-NEXT:    xori $a7, $a7, 31
+; LA32-NEXT:    srli.w $t1, $a0, 1
+; LA32-NEXT:    srl.w $a7, $t1, $a7
+; LA32-NEXT:    or $a3, $a3, $a6
+; LA32-NEXT:    sll.w $a5, $a1, $a5
+; LA32-NEXT:    or $a5, $a5, $a7
+; LA32-NEXT:    sll.w $a0, $a0, $t0
+; LA32-NEXT:    slti $a6, $t0, 0
+; LA32-NEXT:    masknez $a0, $a0, $a6
+; LA32-NEXT:    maskeqz $a5, $a5, $a6
+; LA32-NEXT:    or $a0, $a5, $a0
+; LA32-NEXT:    srl.w $a1, $a1, $a2
+; LA32-NEXT:    srai.w $a2, $a4, 31
+; LA32-NEXT:    and $a1, $a2, $a1
+; LA32-NEXT:    or $a1, $a1, $a0
+; LA32-NEXT:    move $a0, $a3
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotr_64_mask_and_127_and_63:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotr.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %a = and i64 %y, 127
+  %b = lshr i64 %x, %a
+  %c = sub i64 0, %y
+  %d = and i64 %c, 63
+  %e = shl i64 %x, %d
+  %f = or i64 %b, %e
+  ret i64 %f
+}
+
+define i64 @rotr_64_mask_or_128_or_64(i64 %x, i64 %y) nounwind {
+; LA32-LABEL: rotr_64_mask_or_128_or_64:
+; LA32:       # %bb.0:
+; LA32-NEXT:    srl.w $a3, $a1, $a2
+; LA32-NEXT:    sub.w $a1, $zero, $a2
+; LA32-NEXT:    sll.w $a1, $a0, $a1
+; LA32-NEXT:    move $a0, $a3
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotr_64_mask_or_128_or_64:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotr.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %a = or i64 %y, 128
+  %b = lshr i64 %x, %a
+  %c = sub i64 0, %y
+  %d = or i64 %c, 64
+  %e = shl i64 %x, %d
+  %f = or i64 %b, %e
+  ret i64 %f
+}
+
+define i32 @rotri_i32(i32 %a) nounwind {
+; LA32-LABEL: rotri_i32:
+; LA32:       # %bb.0:
+; LA32-NEXT:    rotri.w $a0, $a0, 16
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotri_i32:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotri.w $a0, $a0, 16
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %shl = shl i32 %a, 16
+  %shr = lshr i32 %a, 16
+  %or = or i32 %shl, %shr
+  ret i32 %or
+}
+
+define i64 @rotri_i64(i64 %a) nounwind {
+; LA32-LABEL: rotri_i64:
+; LA32:       # %bb.0:
+; LA32-NEXT:    move $a2, $a0
+; LA32-NEXT:    move $a0, $a1
+; LA32-NEXT:    move $a1, $a2
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotri_i64:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotri.d $a0, $a0, 32
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %shl = shl i64 %a, 32
+  %shr = lshr i64 %a, 32
+  %or = or i64 %shl, %shr
+  ret i64 %or
+}
+
+declare i32 @llvm.fshl.i32(i32, i32, i32)
+declare i64 @llvm.fshl.i64(i64, i64, i64)
+declare i32 @llvm.fshr.i32(i32, i32, i32)
+declare i64 @llvm.fshr.i64(i64, i64, i64)
+
+define signext i32 @rotl_i32_fshl(i32 signext %a) nounwind {
+; LA32-LABEL: rotl_i32_fshl:
+; LA32:       # %bb.0:
+; LA32-NEXT:    rotri.w $a0, $a0, 20
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotl_i32_fshl:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotri.w $a0, $a0, 20
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %or = tail call i32 @llvm.fshl.i32(i32 %a, i32 %a, i32 12)
+  ret i32 %or
+}
+
+define i64 @rotl_i64_fshl(i64 %a) nounwind {
+; LA32-LABEL: rotl_i64_fshl:
+; LA32:       # %bb.0:
+; LA32-NEXT:    srli.w $a2, $a1, 20
+; LA32-NEXT:    slli.w $a3, $a0, 12
+; LA32-NEXT:    or $a2, $a3, $a2
+; LA32-NEXT:    srli.w $a0, $a0, 20
+; LA32-NEXT:    slli.w $a1, $a1, 12
+; LA32-NEXT:    or $a1, $a1, $a0
+; LA32-NEXT:    move $a0, $a2
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotl_i64_fshl:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotri.d $a0, $a0, 52
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %or = tail call i64 @llvm.fshl.i64(i64 %a, i64 %a, i64 12)
+  ret i64 %or
+}
+
+define signext i32 @rotr_i32_fshr(i32 signext %a) nounwind {
+; LA32-LABEL: rotr_i32_fshr:
+; LA32:       # %bb.0:
+; LA32-NEXT:    rotri.w $a0, $a0, 12
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotr_i32_fshr:
+; LA64:       # %bb.0:
+; LA64-NEXT:    slli.d $a1, $a0, 20
+; LA64-NEXT:    bstrpick.d $a0, $a0, 31, 12
+; LA64-NEXT:    or $a0, $a0, $a1
+; LA64-NEXT:    addi.w $a0, $a0, 0
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %or = tail call i32 @llvm.fshr.i32(i32 %a, i32 %a, i32 12)
+  ret i32 %or
+}
+
+define i64 @rotr_i64_fshr(i64 %a) nounwind {
+; LA32-LABEL: rotr_i64_fshr:
+; LA32:       # %bb.0:
+; LA32-NEXT:    srli.w $a2, $a0, 12
+; LA32-NEXT:    slli.w $a3, $a1, 20
+; LA32-NEXT:    or $a2, $a3, $a2
+; LA32-NEXT:    srli.w $a1, $a1, 12
+; LA32-NEXT:    slli.w $a0, $a0, 20
+; LA32-NEXT:    or $a1, $a0, $a1
+; LA32-NEXT:    move $a0, $a2
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: rotr_i64_fshr:
+; LA64:       # %bb.0:
+; LA64-NEXT:    rotri.d $a0, $a0, 12
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %or = tail call i64 @llvm.fshr.i64(i64 %a, i64 %a, i64 12)
+  ret i64 %or
+}