llvm-project/llvm/lib/Target/VE/VEInstrInfo.td

1502 lines
54 KiB
TableGen

//===-- VEInstrInfo.td - Target Description for VE Target -----------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// This file describes the VE instructions in TableGen format.
//
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// Instruction format superclass
//===----------------------------------------------------------------------===//
include "VEInstrFormats.td"
//===----------------------------------------------------------------------===//
// Feature predicates.
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// Instruction Pattern Stuff
//===----------------------------------------------------------------------===//
def simm7 : PatLeaf<(imm), [{ return isInt<7>(N->getSExtValue()); }]>;
def simm32 : PatLeaf<(imm), [{ return isInt<32>(N->getSExtValue()); }]>;
def uimm32 : PatLeaf<(imm), [{ return isUInt<32>(N->getZExtValue()); }]>;
def uimm6 : PatLeaf<(imm), [{ return isUInt<6>(N->getZExtValue()); }]>;
def lomsbzero : PatLeaf<(imm), [{ return (N->getZExtValue() & 0x80000000)
== 0; }]>;
def lozero : PatLeaf<(imm), [{ return (N->getZExtValue() & 0xffffffff)
== 0; }]>;
def fplomsbzero : PatLeaf<(fpimm), [{ return (N->getValueAPF().bitcastToAPInt()
.getZExtValue() & 0x80000000) == 0; }]>;
def fplozero : PatLeaf<(fpimm), [{ return (N->getValueAPF().bitcastToAPInt()
.getZExtValue() & 0xffffffff) == 0; }]>;
def CCSIOp : PatLeaf<(cond), [{
switch (N->get()) {
default: return true;
case ISD::SETULT:
case ISD::SETULE:
case ISD::SETUGT:
case ISD::SETUGE: return false;
}
}]>;
def CCUIOp : PatLeaf<(cond), [{
switch (N->get()) {
default: return true;
case ISD::SETLT:
case ISD::SETLE:
case ISD::SETGT:
case ISD::SETGE: return false;
}
}]>;
def LOFP32 : SDNodeXForm<fpimm, [{
// Get a integer immediate from fpimm
const APInt& imm = N->getValueAPF().bitcastToAPInt();
return CurDAG->getTargetConstant(Lo_32(imm.getZExtValue() & 0xffffffff),
SDLoc(N), MVT::i64);
}]>;
def HIFP32 : SDNodeXForm<fpimm, [{
// Get a integer immediate from fpimm
const APInt& imm = N->getValueAPF().bitcastToAPInt();
return CurDAG->getTargetConstant(Hi_32(imm.getZExtValue()),
SDLoc(N), MVT::i64);
}]>;
def LO32 : SDNodeXForm<imm, [{
return CurDAG->getTargetConstant(Lo_32(N->getZExtValue()),
SDLoc(N), MVT::i64);
}]>;
def HI32 : SDNodeXForm<imm, [{
// Transformation function: shift the immediate value down into the low bits.
return CurDAG->getTargetConstant(Hi_32(N->getZExtValue()),
SDLoc(N), MVT::i32);
}]>;
def icond2cc : SDNodeXForm<cond, [{
VECC::CondCodes cc;
switch (N->get()) {
default: llvm_unreachable("Unknown integer condition code!");
case ISD::SETEQ: cc = VECC::CC_IEQ; break;
case ISD::SETNE: cc = VECC::CC_INE; break;
case ISD::SETLT: cc = VECC::CC_IL; break;
case ISD::SETGT: cc = VECC::CC_IG; break;
case ISD::SETLE: cc = VECC::CC_ILE; break;
case ISD::SETGE: cc = VECC::CC_IGE; break;
case ISD::SETULT: cc = VECC::CC_IL; break;
case ISD::SETULE: cc = VECC::CC_ILE; break;
case ISD::SETUGT: cc = VECC::CC_IG; break;
case ISD::SETUGE: cc = VECC::CC_IGE; break;
}
return CurDAG->getTargetConstant(cc, SDLoc(N), MVT::i32);
}]>;
def fcond2cc : SDNodeXForm<cond, [{
VECC::CondCodes cc;
switch (N->get()) {
default: llvm_unreachable("Unknown float condition code!");
case ISD::SETFALSE: cc = VECC::CC_AF; break;
case ISD::SETEQ:
case ISD::SETOEQ: cc = VECC::CC_EQ; break;
case ISD::SETNE:
case ISD::SETONE: cc = VECC::CC_NE; break;
case ISD::SETLT:
case ISD::SETOLT: cc = VECC::CC_L; break;
case ISD::SETGT:
case ISD::SETOGT: cc = VECC::CC_G; break;
case ISD::SETLE:
case ISD::SETOLE: cc = VECC::CC_LE; break;
case ISD::SETGE:
case ISD::SETOGE: cc = VECC::CC_GE; break;
case ISD::SETO: cc = VECC::CC_NUM; break;
case ISD::SETUO: cc = VECC::CC_NAN; break;
case ISD::SETUEQ: cc = VECC::CC_EQNAN; break;
case ISD::SETUNE: cc = VECC::CC_NENAN; break;
case ISD::SETULT: cc = VECC::CC_LNAN; break;
case ISD::SETUGT: cc = VECC::CC_GNAN; break;
case ISD::SETULE: cc = VECC::CC_LENAN; break;
case ISD::SETUGE: cc = VECC::CC_GENAN; break;
case ISD::SETTRUE: cc = VECC::CC_AT; break;
}
return CurDAG->getTargetConstant(cc, SDLoc(N), MVT::i32);
}]>;
// Addressing modes.
def ADDRrr : ComplexPattern<iPTR, 2, "SelectADDRrr", [], []>;
def ADDRri : ComplexPattern<iPTR, 2, "SelectADDRri", [frameindex], []>;
// ASX format of memory address
def MEMrr : Operand<iPTR> {
let PrintMethod = "printMemASXOperand";
let MIOperandInfo = (ops ptr_rc, ptr_rc);
}
def MEMri : Operand<iPTR> {
let PrintMethod = "printMemASXOperand";
let MIOperandInfo = (ops ptr_rc, i64imm);
}
// AS format of memory address
def MEMASri : Operand<iPTR> {
let PrintMethod = "printMemASOperand";
let MIOperandInfo = (ops ptr_rc, i64imm);
}
// Branch targets have OtherVT type.
def brtarget32 : Operand<OtherVT> {
let EncoderMethod = "getBranchTarget32OpValue";
}
def calltarget : Operand<i64> {
let EncoderMethod = "getCallTargetOpValue";
let DecoderMethod = "DecodeCall";
}
def simm7Op32 : Operand<i32> {
let DecoderMethod = "DecodeSIMM7";
}
def simm7Op64 : Operand<i64> {
let DecoderMethod = "DecodeSIMM7";
}
def simm32Op32 : Operand<i32> {
let DecoderMethod = "DecodeSIMM32";
}
def simm32Op64 : Operand<i64> {
let DecoderMethod = "DecodeSIMM32";
}
def uimm6Op32 : Operand<i32> {
let DecoderMethod = "DecodeUIMM6";
}
def uimm6Op64 : Operand<i64> {
let DecoderMethod = "DecodeUIMM6";
}
// Operand for printing out a condition code.
let PrintMethod = "printCCOperand" in
def CCOp : Operand<i32>;
def VEhi : SDNode<"VEISD::Hi", SDTIntUnaryOp>;
def VElo : SDNode<"VEISD::Lo", SDTIntUnaryOp>;
// These are target-independent nodes, but have target-specific formats.
def SDT_SPCallSeqStart : SDCallSeqStart<[ SDTCisVT<0, i64>,
SDTCisVT<1, i64> ]>;
def SDT_SPCallSeqEnd : SDCallSeqEnd<[ SDTCisVT<0, i64>,
SDTCisVT<1, i64> ]>;
def callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_SPCallSeqStart,
[SDNPHasChain, SDNPOutGlue]>;
def callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_SPCallSeqEnd,
[SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
def SDT_SPCall : SDTypeProfile<0, -1, [SDTCisVT<0, i64>]>;
def call : SDNode<"VEISD::CALL", SDT_SPCall,
[SDNPHasChain, SDNPOptInGlue, SDNPOutGlue,
SDNPVariadic]>;
def retflag : SDNode<"VEISD::RET_FLAG", SDTNone,
[SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
def getGOT : Operand<iPTR>;
// GETFUNPLT for PIC
def GetFunPLT : SDNode<"VEISD::GETFUNPLT", SDTIntUnaryOp>;
// GETTLSADDR for TLS
def GetTLSAddr : SDNode<"VEISD::GETTLSADDR", SDT_SPCall,
[SDNPHasChain, SDNPOptInGlue, SDNPOutGlue,
SDNPVariadic]>;
//===----------------------------------------------------------------------===//
// VE Flag Conditions
//===----------------------------------------------------------------------===//
// Note that these values must be kept in sync with the CCOp::CondCode enum
// values.
class CC_VAL<int N> : PatLeaf<(i32 N)>;
def CC_IG : CC_VAL< 0>; // Greater
def CC_IL : CC_VAL< 1>; // Less
def CC_INE : CC_VAL< 2>; // Not Equal
def CC_IEQ : CC_VAL< 3>; // Equal
def CC_IGE : CC_VAL< 4>; // Greater or Equal
def CC_ILE : CC_VAL< 5>; // Less or Equal
def CC_AF : CC_VAL< 6>; // Always false
def CC_G : CC_VAL< 7>; // Greater
def CC_L : CC_VAL< 8>; // Less
def CC_NE : CC_VAL< 9>; // Not Equal
def CC_EQ : CC_VAL<10>; // Equal
def CC_GE : CC_VAL<11>; // Greater or Equal
def CC_LE : CC_VAL<12>; // Less or Equal
def CC_NUM : CC_VAL<13>; // Number
def CC_NAN : CC_VAL<14>; // NaN
def CC_GNAN : CC_VAL<15>; // Greater or NaN
def CC_LNAN : CC_VAL<16>; // Less or NaN
def CC_NENAN : CC_VAL<17>; // Not Equal or NaN
def CC_EQNAN : CC_VAL<18>; // Equal or NaN
def CC_GENAN : CC_VAL<19>; // Greater or Equal or NaN
def CC_LENAN : CC_VAL<20>; // Less or Equal or NaN
def CC_AT : CC_VAL<21>; // Always true
//===----------------------------------------------------------------------===//
// VE Multiclasses for common instruction formats
//===----------------------------------------------------------------------===//
multiclass RMm<string opcStr, bits<8>opc,
RegisterClass RC, ValueType Ty,
Operand immOp, Operand immOp2,
SDPatternOperator OpNode=null_frag> {
def rri : RM<
opc, (outs RC:$sx), (ins RC:$sy, RC:$sz, immOp2:$imm32),
!strconcat(opcStr, " $sx, ${imm32}($sy, ${sz})")> {
let cy = 1;
let cz = 1;
let hasSideEffects = 0;
}
def rzi : RM<
opc, (outs RC:$sx), (ins RC:$sz, immOp2:$imm32),
!strconcat(opcStr, " $sx, ${imm32}(${sz})"),
[(set Ty:$sx, (OpNode Ty:$sz, (Ty simm32:$imm32)))]> {
let cy = 0;
let sy = 0;
let cz = 1;
let hasSideEffects = 0;
}
def zii : RM<
opc, (outs RC:$sx), (ins immOp:$sy, immOp2:$imm32),
!strconcat(opcStr, " $sx, ${imm32}(${sy})"),
[/* Not define DAG pattern here to avoid llvm uses LEAzii for all add
instructions.
(set Ty:$sx, (OpNode (Ty simm7:$sy), (Ty simm32:$imm32))) */]> {
let cy = 0;
let cz = 0;
let sz = 0;
let hasSideEffects = 0;
}
def zzi : RM<
opc, (outs RC:$sx), (ins immOp2:$imm32),
!strconcat(opcStr, " $sx, $imm32")> {
let cy = 0;
let sy = 0;
let cz = 0;
let sz = 0;
let hasSideEffects = 0;
}
}
multiclass RRmrr<string opcStr, bits<8>opc,
RegisterClass RCo, ValueType Tyo,
RegisterClass RCi, ValueType Tyi,
SDPatternOperator OpNode=null_frag> {
def rr : RR<opc, (outs RCo:$sx), (ins RCi:$sy, RCi:$sz),
!strconcat(opcStr, " $sx, $sy, $sz"),
[(set Tyo:$sx, (OpNode Tyi:$sy, Tyi:$sz))]>
{ let cy = 1; let cz = 1; let hasSideEffects = 0; }
}
multiclass RRmri<string opcStr, bits<8>opc,
RegisterClass RCo, ValueType Tyo,
RegisterClass RCi, ValueType Tyi, Operand immOp,
SDPatternOperator OpNode=null_frag> {
// VE calculates (OpNode $sy, $sz), but llvm requires to have immediate
// in RHS, so we use following definition.
def ri : RR<opc, (outs RCo:$sx), (ins RCi:$sz, immOp:$sy),
!strconcat(opcStr, " $sx, $sy, $sz"),
[(set Tyo:$sx, (OpNode Tyi:$sz, (Tyi simm7:$sy)))]>
{ let cy = 0; let cz = 1; let hasSideEffects = 0; }
}
multiclass RRmir<string opcStr, bits<8>opc,
RegisterClass RCo, ValueType Tyo,
RegisterClass RCi, ValueType Tyi, Operand immOp,
SDPatternOperator OpNode=null_frag> {
def ri : RR<opc, (outs RCo:$sx), (ins immOp:$sy, RCi:$sz),
!strconcat(opcStr, " $sx, $sy, $sz"),
[(set Tyo:$sx, (OpNode (Tyi simm7:$sy), Tyi:$sz))]>
{ let cy = 0; let cz = 1; let hasSideEffects = 0; }
}
multiclass RRmiz<string opcStr, bits<8>opc,
RegisterClass RCo, ValueType Tyo,
RegisterClass RCi, ValueType Tyi, Operand immOp,
SDPatternOperator OpNode=null_frag> {
def zi : RR<opc, (outs RCo:$sx), (ins immOp:$sy),
!strconcat(opcStr, " $sx, $sy"),
[(set Tyo:$sx, (OpNode (Tyi simm7:$sy), 0))]>
{ let cy = 0; let cz = 0; let sz = 0; let hasSideEffects = 0; }
}
multiclass RRNDmrm<string opcStr, bits<8>opc,
RegisterClass RCo, ValueType Tyo,
RegisterClass RCi, ValueType Tyi, Operand immOp2> {
def rm0 : RR<opc, (outs RCo:$sx), (ins RCi:$sy, immOp2:$sz),
!strconcat(opcStr, " $sx, $sy, (${sz})0")> {
let cy = 1;
let cz = 0;
let sz{6} = 1;
// (guess) tblgen conservatively assumes hasSideEffects when
// it fails to infer from a pattern.
let hasSideEffects = 0;
}
def rm1 : RR<opc, (outs RCo:$sx), (ins RCi:$sy, immOp2:$sz),
!strconcat(opcStr, " $sx, $sy, (${sz})1")> {
let cy = 1;
let cz = 0;
let hasSideEffects = 0;
}
}
multiclass RRNDmim<string opcStr, bits<8>opc,
RegisterClass RCo, ValueType Tyo,
RegisterClass RCi, ValueType Tyi,
Operand immOp, Operand immOp2> {
def im1 : RR<opc, (outs RCo:$sx), (ins immOp:$sy, immOp2:$sz),
!strconcat(opcStr, " $sx, $sy, (${sz})1")> {
let cy = 0;
let cz = 0;
let hasSideEffects = 0;
}
}
// Used by add, mul, div, and similar commutative instructions
// The order of operands are "$sx, $sy, $sz"
multiclass RRm<string opcStr, bits<8>opc,
RegisterClass RC, ValueType Ty,
Operand immOp, Operand immOp2,
SDPatternOperator OpNode=null_frag> :
RRmrr<opcStr, opc, RC, Ty, RC, Ty, OpNode>,
RRmri<opcStr, opc, RC, Ty, RC, Ty, immOp, OpNode>,
RRmiz<opcStr, opc, RC, Ty, RC, Ty, immOp, OpNode>,
RRNDmrm<opcStr, opc, RC, Ty, RC, Ty, immOp2>,
RRNDmim<opcStr, opc, RC, Ty, RC, Ty, immOp, immOp2>;
// Used by sub, and similar not commutative instructions
// The order of operands are "$sx, $sy, $sz"
multiclass RRNCm<string opcStr, bits<8>opc,
RegisterClass RC, ValueType Ty, Operand immOp, Operand immOp2,
SDPatternOperator OpNode=null_frag> :
RRmrr<opcStr, opc, RC, Ty, RC, Ty, OpNode>,
RRmir<opcStr, opc, RC, Ty, RC, Ty, immOp, OpNode>,
RRmiz<opcStr, opc, RC, Ty, RC, Ty, immOp, OpNode>,
RRNDmrm<opcStr, opc, RC, Ty, RC, Ty, immOp2>,
RRNDmim<opcStr, opc, RC, Ty, RC, Ty, immOp, immOp2>;
// Used by fadd, fsub, and similar floating point instructions
// The order of operands are "$sx, $sy, $sz"
multiclass RRFm<string opcStr, bits<8>opc,
RegisterClass RC, ValueType Ty, Operand immOp, Operand immOp2,
SDPatternOperator OpNode=null_frag> :
RRmrr<opcStr, opc, RC, Ty, RC, Ty, OpNode>,
RRmir<opcStr, opc, RC, Ty, RC, Ty, immOp, null_frag>,
RRmiz<opcStr, opc, RC, Ty, RC, Ty, immOp, null_frag>,
RRNDmrm<opcStr, opc, RC, Ty, RC, Ty, immOp2>,
RRNDmim<opcStr, opc, RC, Ty, RC, Ty, immOp, immOp2>;
// Multiclass for RR type instructions
// Used by sra, sla, sll, and similar instructions
// The order of operands are "$sx, $sz, $sy"
multiclass RRIm<string opcStr, bits<8>opc,
RegisterClass RC, ValueType Ty,
Operand immOp, Operand immOp2,
SDPatternOperator OpNode=null_frag> {
def rr : RR<
opc, (outs RC:$sx), (ins RC:$sz, I32:$sy),
!strconcat(opcStr, " $sx, $sz, $sy"),
[(set Ty:$sx, (OpNode Ty:$sz, i32:$sy))]> {
let cy = 1;
let cz = 1;
let hasSideEffects = 0;
}
def ri : RR<
opc, (outs RC:$sx), (ins RC:$sz, immOp:$sy),
!strconcat(opcStr, " $sx, $sz, $sy"),
[(set Ty:$sx, (OpNode Ty:$sz, (i32 simm7:$sy)))]> {
let cy = 0;
let cz = 1;
let hasSideEffects = 0;
}
def rm0 : RR<
opc, (outs RC:$sx), (ins immOp2:$sz, I32:$sy),
!strconcat(opcStr, " $sx, (${sz})0, $sy")> {
let cy = 1;
let cz = 0;
let sz{6} = 1;
let hasSideEffects = 0;
}
def rm1 : RR<
opc, (outs RC:$sx), (ins immOp2:$sz, I32:$sy),
!strconcat(opcStr, " $sx, (${sz})1, $sy")> {
let cy = 1;
let cz = 0;
let hasSideEffects = 0;
}
def im0 : RR<
opc, (outs RC:$sx), (ins immOp2:$sz, immOp:$sy),
!strconcat(opcStr, " $sx, (${sz})0, $sy")> {
let cy = 0;
let cz = 0;
let sz{6} = 1;
let hasSideEffects = 0;
}
def im1 : RR<
opc, (outs RC:$sx), (ins immOp2:$sz, immOp:$sy),
!strconcat(opcStr, " $sx, (${sz})1, $sy")> {
let cy = 0;
let cz = 0;
let hasSideEffects = 0;
}
def zi : RR<
opc, (outs RC:$sx), (ins immOp:$sy),
!strconcat(opcStr, " $sx, $sy"),
[(set Ty:$sx, (OpNode 0, (i32 simm7:$sy)))]> {
let cy = 0;
let cz = 0;
let sz = 0;
let hasSideEffects = 0;
}
}
// Multiclass for RR type instructions
// Used by cmov instruction
let Constraints = "$sx = $sd", DisableEncoding = "$sd" in
multiclass RRCMOVm<string opcStr, bits<8>opc,
RegisterClass RC, ValueType Ty, Operand immOp, Operand immOp2> {
def rr : RR<
opc, (outs I64:$sx), (ins CCOp:$cf, RC:$sy, I64:$sz, I64:$sd),
!strconcat(opcStr, " $sx, $sz, $sy")> {
let cy = 1;
let cz = 1;
let hasSideEffects = 0;
}
def rm0 : RR<
opc, (outs I64:$sx), (ins CCOp:$cf, RC:$sy, immOp2:$sz, I64:$sd),
!strconcat(opcStr, " $sx, (${sz})0, $sy")> {
let cy = 1;
let cz = 0;
let sz{6} = 1;
let hasSideEffects = 0;
}
}
// Multiclass for RR type instructions with only 2 operands
// Used by pcnt, brv
let hasSideEffects = 0 in
multiclass RRI2m<string opcStr, bits<8>opc, RegisterClass RC, ValueType Ty,
Operand immOp2, SDPatternOperator OpNode=null_frag> {
def r : RR<
opc, (outs RC:$sx), (ins RC:$sz),
!strconcat(opcStr, " $sx, $sz"),
[(set Ty:$sx, (OpNode Ty:$sz))]> {
let cy = 1;
let cz = 1;
}
def i : RR<
opc, (outs RC:$sx), (ins RC:$sz),
!strconcat(opcStr, " $sx, $sz"),
[(set Ty:$sx, (OpNode Ty:$sz))]> {
let cy = 0;
let cz = 1;
}
def m0 : RR<
opc, (outs RC:$sx), (ins immOp2:$sz),
!strconcat(opcStr, " $sx, (${sz})0")> {
let cy = 1;
let cz = 0;
let sz{6} = 1;
}
def m1 : RR<
opc, (outs RC:$sx), (ins immOp2:$sz),
!strconcat(opcStr, " $sx, (${sz})1")> {
let cy = 1;
let cz = 0;
}
}
// Branch multiclass
let isBranch = 1, isTerminator = 1, hasDelaySlot = 1 in
multiclass BCRm<string opcStr, string opcStrAt, bits<8> opc,
RegisterClass RC, ValueType Ty, Operand immOp, Operand immOp2> {
def rr : CF<
opc, (outs),
(ins CCOp:$cf, RC:$sy, RC:$sz, brtarget32:$imm32),
!strconcat(opcStr, " $sy, $sz, $imm32")> {
let cy = 1;
let cz = 1;
let hasSideEffects = 0;
}
def ir : CF<
opc, (outs),
(ins CCOp:$cf, immOp:$sy, RC:$sz, brtarget32:$imm32),
!strconcat(opcStr, " $sy, $sz, $imm32")> {
let cy = 0;
let cz = 1;
let hasSideEffects = 0;
}
def rm0 : CF<
opc, (outs), (ins CCOp:$cf, RC:$sy, immOp2:$sz, brtarget32:$imm32),
!strconcat(opcStr, " $sy, (${sz})0, $imm32"), []> {
let cy = 1;
let cz = 0;
let sz{6} = 1;
let hasSideEffects = 0;
}
def rm1 : CF<
opc, (outs), (ins CCOp:$cf, RC:$sy, immOp2:$sz, brtarget32:$imm32),
!strconcat(opcStr, " $sy, (${sz})1, $imm32"), []> {
let cy = 1;
let cz = 0;
let hasSideEffects = 0;
}
def im0 : CF<
opc, (outs), (ins CCOp:$cf, immOp:$sy, immOp2:$sz, brtarget32:$imm32),
!strconcat(opcStr, " $sy, (${sz})0, $imm32"), []> {
let cy = 0;
let cz = 0;
let sz{6} = 1;
let hasSideEffects = 0;
}
def im1 : CF<
opc, (outs), (ins CCOp:$cf, immOp:$sy, immOp2:$sz, brtarget32:$imm32),
!strconcat(opcStr, " $sy, (${sz})1, $imm32"), []> {
let cy = 0;
let cz = 0;
let hasSideEffects = 0;
}
def a : CF<
opc, (outs), (ins brtarget32:$imm32),
!strconcat(opcStrAt, " $imm32"), []> {
let cy = 0;
let sy = 0;
let cz = 0;
let sz = 0;
let cf = 15; /* AT */
let isBarrier = 1;
let hasSideEffects = 0;
}
}
// Multiclass for floating point conversion instructions.
// Used by CVS/CVD/FLT and others
multiclass CVTm<string opcStr, bits<8> opc,
RegisterClass RCo, ValueType Tyo,
RegisterClass RCi, ValueType Tyi, Operand immOp,
SDPatternOperator OpNode=null_frag> {
def r : RR<opc, (outs RCo:$sx), (ins RCi:$sy),
!strconcat(opcStr, " $sx, $sy"),
[(set Tyo:$sx, (OpNode Tyi:$sy))]> {
let cy = 1;
let hasSideEffects = 0;
}
def i : RR<opc, (outs RCo:$sx), (ins immOp:$sy),
!strconcat(opcStr, " $sx, $sy"),
[/* (set Tyo:$sx, (OpNode (Tyi simm7:$sy))) */]> {
let cy = 0;
let hasSideEffects = 0;
}
}
//===----------------------------------------------------------------------===//
// Instructions
//===----------------------------------------------------------------------===//
// CMOV instructions
let cx = 0, cw = 0, cw2 = 0 in
defm CMOVL : RRCMOVm<"cmov.l.${cf}", 0x3B, I64, i64, simm7Op64, uimm6Op64>;
let cx = 0, cw = 1, cw2 = 0 in
defm CMOVW : RRCMOVm<"cmov.w.${cf}", 0x3B, I32, i32, simm7Op64, uimm6Op32>;
let cx = 0, cw = 0, cw2 = 1 in
defm CMOVD : RRCMOVm<"cmov.d.${cf}", 0x3B, I64, f64, simm7Op64, uimm6Op64>;
let cx = 0, cw = 1, cw2 = 1 in
defm CMOVS : RRCMOVm<"cmov.s.${cf}", 0x3B, F32, f32, simm7Op64, uimm6Op32>;
// LEA and LEASL instruction (load 32 bit imm to low or high part)
let cx = 0 in
defm LEA : RMm<"lea", 0x06, I64, i64, simm7Op64, simm32Op64, add>;
let cx = 1 in
defm LEASL : RMm<"lea.sl", 0x06, I64, i64, simm7Op64, simm32Op64>;
let isCodeGenOnly = 1 in {
let cx = 0 in
defm LEA32 : RMm<"lea", 0x06, I32, i32, simm7Op32, simm32Op32, add>;
}
let cx = 0, cy = 1, cz = 0, sz = 0, hasSideEffects = 0 in {
def LEAasx : RM<
0x06, (outs I64:$sx), (ins MEMri:$addr),
"lea $sx,$addr", [(set iPTR:$sx, ADDRri:$addr)]>;
}
// 5.3.2.2. Fixed-Point Arithmetic Operation Instructions
// ADD instruction
let cx = 0 in
defm ADD : RRm<"addu.l", 0x48, I64, i64, simm7Op64, uimm6Op64>;
let cx = 1 in
defm ADDUW : RRm<"addu.w", 0x48, I32, i32, simm7Op32, uimm6Op32>;
// ADS instruction
let cx = 0 in
defm ADS : RRm<"adds.w.sx", 0x4A, I32, i32, simm7Op32, uimm6Op32, add>;
let cx = 1 in
defm ADSU : RRm<"adds.w.zx", 0x4A, I32, i32, simm7Op32, uimm6Op32>;
// ADX instruction
let cx = 0 in
defm ADX : RRm<"adds.l", 0x59, I64, i64, simm7Op64, uimm6Op64, add>;
// SUB instruction
let cx = 0 in
defm SUB : RRm<"subu.l", 0x58, I64, i64, simm7Op64, uimm6Op64>;
let cx = 1 in
defm SUBUW : RRm<"subu.w", 0x58, I32, i32, simm7Op32, uimm6Op32>;
// SBS instruction
let cx = 0 in
defm SBS : RRNCm<"subs.w.sx", 0x5A, I32, i32, simm7Op32, uimm6Op32, sub>;
let cx = 1 in
defm SBSU : RRm<"subs.w.zx", 0x5A, I32, i32, simm7Op32, uimm6Op32>;
// SBX instruction
let cx = 0 in
defm SBX : RRNCm<"subs.l", 0x5B, I64, i64, simm7Op64, uimm6Op64, sub>;
// MPY instruction
let cx = 0 in
defm MPY : RRm<"mulu.l", 0x49, I64, i64, simm7Op64, uimm6Op64>;
let cx = 1 in
defm MPYUW : RRm<"mulu.w", 0x49, I32, i32, simm7Op32, uimm6Op32>;
// MPS instruction
let cx = 0 in
defm MPS : RRm<"muls.w.sx", 0x4B, I32, i32, simm7Op32, uimm6Op32, mul>;
let cx = 1 in
defm MPSU : RRm<"muls.w.zx", 0x4B, I32, i32, simm7Op32, uimm6Op32>;
// MPX instruction
let cx = 0 in
defm MPX : RRm<"muls.l", 0x6E, I64, i64, simm7Op64, uimm6Op64, mul>;
// DIV instruction
let cx = 0 in
defm DIV : RRNCm<"divu.l", 0x6F, I64, i64, simm7Op64, uimm6Op64, udiv>;
let cx = 1 in
defm DIVUW : RRNCm<"divu.w", 0x6F, I32, i32, simm7Op32, uimm6Op32, udiv>;
// DVS instruction
let cx = 0 in
defm DVS : RRNCm<"divs.w.sx", 0x7B, I32, i32, simm7Op32, uimm6Op32, sdiv>;
let cx = 1 in
defm DVSU : RRm<"divs.w.zx", 0x7B, I32, i32, simm7Op32, uimm6Op32>;
// DVX instruction
let cx = 0 in
defm DVX : RRNCm<"divs.l", 0x7F, I64, i64, simm7Op64, uimm6Op64, sdiv>;
// CMP instruction
let cx = 0 in
defm CMP : RRm<"cmpu.l", 0x55, I64, i64, simm7Op64, uimm6Op64>;
let cx = 1 in
defm CMPUW : RRm<"cmpu.w", 0x55, I32, i32, simm7Op32, uimm6Op32>;
// CPS instruction
let cx = 0 in
defm CPS : RRm<"cmps.w.sx", 0x7A, I32, i32, simm7Op32, uimm6Op32>;
let cx = 1 in
defm CPSU : RRm<"cmps.w.zx", 0x7A, I32, i32, simm7Op32, uimm6Op32>;
// CPX instruction
let cx = 0 in
defm CPX : RRm<"cmps.l", 0x6A, I64, i64, simm7Op64, uimm6Op64>;
// cx: sx/zx, cw: max/min
let cw = 0 in defm CMXa :
RRm<"maxs.l", 0x68, I64, i64, simm7Op64, uimm6Op64>;
let cx = 0, cw = 0 in defm CMSa :
RRm<"maxs.w.zx", 0x78, I32, i32, simm7Op32, uimm6Op32>;
let cw = 1 in defm CMXi :
RRm<"mins.l", 0x68, I64, i64, simm7Op64, uimm6Op64>;
let cx = 1, cw = 0 in defm CMSi :
RRm<"mins.w.zx", 0x78, I32, i32, simm7Op32, uimm6Op32>;
// 5.3.2.3. Logical Arithmetic Operation Instructions
let cx = 0 in {
defm AND : RRm<"and", 0x44, I64, i64, simm7Op64, uimm6Op64, and>;
defm OR : RRm<"or", 0x45, I64, i64, simm7Op64, uimm6Op64, or>;
defm XOR : RRm<"xor", 0x46, I64, i64, simm7Op64, uimm6Op64, xor>;
let isCodeGenOnly = 1 in {
defm AND32 : RRm<"and", 0x44, I32, i32, simm7Op32, uimm6Op32, and>;
defm OR32 : RRm<"or", 0x45, I32, i32, simm7Op32, uimm6Op32, or>;
defm XOR32 : RRm<"xor", 0x46, I32, i32, simm7Op32, uimm6Op32, xor>;
}
}
// Bits operations
let cx = 0 in {
defm PCNT : RRI2m<"pcnt", 0x38, I64, i64, uimm6Op64, ctpop>;
defm BRV : RRI2m<"brv", 0x39, I64, i64, uimm6Op64, bitreverse>;
defm LDZ : RRI2m<"ldz", 0x67, I64, i64, uimm6Op64, ctlz>;
defm BSWP : RRIm<"bswp", 0x2B, I64, i64, simm7Op64, uimm6Op64>;
}
// 5.3.2.4 Shift Instructions
let cx = 0 in
defm SRAX : RRIm<"sra.l", 0x77, I64, i64, simm7Op32, uimm6Op64, sra>;
let cx = 0 in
defm SRA : RRIm<"sra.w.sx", 0x76, I32, i32, simm7Op32, uimm6Op32, sra>;
let cx = 1 in
defm SRAU : RRIm<"sra.w.zx", 0x76, I32, i32, simm7Op32, uimm6Op32>;
let cx = 0 in
defm SLL : RRIm<"sll", 0x65, I64, i64, simm7Op32, uimm6Op64, shl>;
let cx = 0 in
defm SLA : RRIm<"sla.w.sx", 0x66, I32, i32, simm7Op32, uimm6Op32, shl>;
let cx = 1 in
defm SLAU : RRIm<"sla.w.zx", 0x66, I32, i32, simm7Op32, uimm6Op32>;
let cx = 0 in
defm SRL : RRIm<"srl", 0x75, I64, i64, simm7Op32, uimm6Op64, srl>;
def : Pat<(i32 (srl i32:$src, (i32 simm7:$val))),
(EXTRACT_SUBREG (SRLri (ANDrm0 (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
$src, sub_i32), 32), imm:$val), sub_i32)>;
def : Pat<(i32 (srl i32:$src, i32:$val)),
(EXTRACT_SUBREG (SRLrr (ANDrm0 (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
$src, sub_i32), 32), $val), sub_i32)>;
// 5.3.2.5. Floating-point Arithmetic Operation Instructions
let cx = 0 in
defm FAD : RRFm<"fadd.d", 0x4C, I64, f64, simm7Op64, uimm6Op64, fadd>;
let cx = 1 in
defm FADS : RRFm<"fadd.s", 0x4C, F32, f32, simm7Op32, uimm6Op32, fadd>;
let cx = 0 in
defm FSB : RRFm<"fsub.d", 0x5C, I64, f64, simm7Op64, uimm6Op64, fsub>;
let cx = 1 in
defm FSBS : RRFm<"fsub.s", 0x5C, F32, f32, simm7Op32, uimm6Op32, fsub>;
let cx = 0 in
defm FMP : RRFm<"fmul.d", 0x4D, I64, f64, simm7Op64, uimm6Op64, fmul>;
let cx = 1 in
defm FMPS : RRFm<"fmul.s", 0x4D, F32, f32, simm7Op32, uimm6Op32, fmul>;
let cx = 0 in
defm FDV : RRFm<"fdiv.d", 0x5D, I64, f64, simm7Op64, uimm6Op64, fdiv>;
let cx = 1 in
defm FDVS : RRFm<"fdiv.s", 0x5D, F32, f32, simm7Op32, uimm6Op32, fdiv>;
// FCP instruction
let cx = 0 in
defm FCP : RRm<"fcmp.d", 0x7E, I64, f64, simm7Op64, uimm6Op64>;
let cx = 1 in
defm FCPS : RRm<"fcmp.s", 0x7E, F32, f32, simm7Op32, uimm6Op32>;
// FCM
let cw = 0 in {
let cx = 0 in
defm FCMA : RRm<"fmax.d", 0x3E, I64, f64, simm7Op64, uimm6Op64>;
let cx = 1 in
defm FCMAS : RRm<"fmax.s", 0x3E, F32, f32, simm7Op32, uimm6Op32>;
}
let cw = 1 in {
let cx = 0 in
defm FCMI : RRm<"fmin.d", 0x3E, I64, f64, simm7Op64, uimm6Op64>;
let cx = 1 in
defm FCMIS : RRm<"fmin.s", 0x3E, F32, f32, simm7Op32, uimm6Op32>;
}
let cx = 0, cw = 0 /* sign extend */, cz = 1, sz = 0 /* round toward zero */ in
defm FIX : CVTm<"cvt.w.d.sx.rz", 0x4E, I32, i32, I64, f64, simm7Op32, fp_to_sint>;
let cx = 1, cw = 0 /* sign extend */, cz = 1, sz = 0 /* round toward zero */ in
defm FIXS : CVTm<"cvt.w.s.sx.rz", 0x4E, I32, i32, F32, f32, simm7Op32, fp_to_sint>;
let cx = 0, cz = 1, sz = 0 /* round toward zero */ in
defm FIXX : CVTm<"cvt.l.d.rz", 0x4F, I64, i64, I64, f64, simm7Op64, fp_to_sint>;
let cz = 0, sz = 0 in {
let cx = 0 in
defm FLT : CVTm<"cvt.d.w", 0x5E, I64, f64, I32, i32, simm7Op32, sint_to_fp>;
let cx = 1 in
defm FLTS : CVTm<"cvt.s.w", 0x5E, F32, f32, I32, i32, simm7Op32, sint_to_fp>;
let cx = 0 in
defm FLTX : CVTm<"cvt.d.l", 0x5F, I64, f64, I64, i64, simm7Op64, sint_to_fp>;
let cx = 0 in
defm CVS : CVTm<"cvt.s.d", 0x1F, F32, f32, I64, f64, simm7Op64, fpround>;
let cx = 0 in
defm CVD : CVTm<"cvt.d.s", 0x0F, I64, f64, F32, f32, simm7Op32, fpextend>;
}
// Load and Store instructions
// As 1st step, only uses sz and imm32 to represent $addr
let mayLoad = 1, hasSideEffects = 0 in {
let cy = 0, sy = 0, cz = 1 in {
let cx = 0 in
def LDSri : RM<
0x01, (outs I64:$sx), (ins MEMri:$addr),
"ld $sx, $addr",
[(set i64:$sx, (load ADDRri:$addr))]>;
let cx = 0 in
def LDUri : RM<
0x02, (outs F32:$sx), (ins MEMri:$addr),
"ldu $sx, $addr",
[(set f32:$sx, (load ADDRri:$addr))]>;
let cx = 0 in
def LDLri : RM<
0x03, (outs I32:$sx), (ins MEMri:$addr),
"ldl.sx $sx, $addr",
[(set i32:$sx, (load ADDRri:$addr))]>;
let cx = 1 in
def LDLUri : RM<
0x03, (outs I32:$sx), (ins MEMri:$addr),
"ldl.zx $sx, $addr",
[(set i32:$sx, (load ADDRri:$addr))]>;
let cx = 0 in
def LD2Bri : RM<
0x04, (outs I32:$sx), (ins MEMri:$addr),
"ld2b.sx $sx, $addr",
[(set i32:$sx, (sextloadi16 ADDRri:$addr))]>;
let cx = 1 in
def LD2BUri : RM<
0x04, (outs I32:$sx), (ins MEMri:$addr),
"ld2b.zx $sx, $addr",
[(set i32:$sx, (zextloadi16 ADDRri:$addr))]>;
let cx = 0 in
def LD1Bri : RM<
0x05, (outs I32:$sx), (ins MEMri:$addr),
"ld1b.sx $sx, $addr",
[(set i32:$sx, (sextloadi8 ADDRri:$addr))]>;
let cx = 1 in
def LD1BUri : RM<
0x05, (outs I32:$sx), (ins MEMri:$addr),
"ld1b.zx $sx, $addr",
[(set i32:$sx, (zextloadi8 ADDRri:$addr))]>;
}
}
let mayStore = 1, hasSideEffects = 0 in {
let cx = 0, cy = 0, sy = 0, cz = 1 in {
def STSri : RM<
0x11, (outs), (ins MEMri:$addr, I64:$sx),
"st $sx, $addr",
[(store i64:$sx, ADDRri:$addr)]>;
def STUri : RM<
0x12, (outs), (ins MEMri:$addr, F32:$sx),
"stu $sx, $addr",
[(store f32:$sx, ADDRri:$addr)]>;
def STLri : RM<
0x13, (outs), (ins MEMri:$addr, I32:$sx),
"stl $sx, $addr",
[(store i32:$sx, ADDRri:$addr)]>;
def ST2Bri : RM<
0x14, (outs), (ins MEMri:$addr, I32:$sx),
"st2b $sx, $addr",
[(truncstorei16 i32:$sx, ADDRri:$addr)]>;
def ST1Bri : RM<
0x15, (outs), (ins MEMri:$addr, I32:$sx),
"st1b $sx, $addr",
[(truncstorei8 i32:$sx, ADDRri:$addr)]>;
}
}
def : Pat<(f64 (load ADDRri:$addr)), (LDSri ADDRri:$addr)>;
def : Pat<(store f64:$sx, ADDRri:$addr), (STSri ADDRri:$addr, $sx)>;
// Control-flow
// Jump instruction
let cx = 0, cx2 = 0, bpf = 0 /* NONE */, cy = 1, cz = 1,
isBranch = 1, isTerminator = 1, hasDelaySlot = 1, hasSideEffects = 0 in
def BC : CF<
0x19, (outs), (ins CCOp:$cf, I64:$sy, brtarget32:$imm32),
"b.${cf}.l $sy, $imm32">;
// Jump always instruction is treated as a special case of jump in order
// to make finding unconditional jump easy.
let cx = 0, cx2 = 0, bpf = 0 /* NONE */, cf = 15 /* AT */, cy = 0, sy = 0,
cz = 1,
isBranch = 1, isTerminator = 1, isBarrier = 1, isIndirectBranch = 1,
hasDelaySlot = 1, isCodeGenOnly = 1, hasSideEffects = 0 in {
def BArr : CF<
0x19, (outs), (ins MEMrr:$addr),
"b.l $addr",
[(brind ADDRrr:$addr)]>;
def BAri : CF<
0x19, (outs), (ins MEMri:$addr),
"b.l $addr",
[(brind ADDRri:$addr)]>;
}
// Jump never instruction is also a special case of jump.
let cx = 0, cx2 = 0, bpf = 0 /* NONE */, cf = 0 /* AF */, cy = 1, sy = 0,
cz = 1,
isBranch = 1, isTerminator = 1, hasDelaySlot = 1, hasSideEffects = 0 in
def BN : CF<
0x19, (outs), (ins brtarget32:$imm32),
"b.af.l $imm32">;
// Return instruction is also a special case of jump.
let cx = 0, cx2 = 0, bpf = 0 /* NONE */, cf = 15 /* AT */, cy = 0, sy = 0,
cz = 1, sz = 0x10 /* SX10 */, imm32 = 0, Uses = [SX10],
isReturn = 1, isTerminator = 1, hasDelaySlot = 1, isBarrier = 1,
isCodeGenOnly = 1, hasSideEffects = 0 in
def RET : CF<
0x19, (outs), (ins),
"b.l (,%lr)",
[(retflag)]>;
// Branch and Save IC
let cx = 0, cy = 0, cy = 0, cz = 1, hasSideEffects = 0 /* , Uses = [IC] */ in
def BSIC : RM<0x08, (outs), (ins I64:$sx, I64:$sz), "bsic $sx, (, ${sz})">;
// Branch instruction
let cx = 0, cx2 = 0, bpf = 0 /* NONE */ in
defm BCRL : BCRm<"br${cf}.l", "br.l", 0x18, I64, i64, simm7Op64, uimm6Op64>;
let cx = 1, cx2 = 0, bpf = 0 /* NONE */ in
defm BCRW : BCRm<"br${cf}.w", "br.w", 0x18, I32, i32, simm7Op32, uimm6Op32>;
let cx = 0, cx2 = 1, bpf = 0 /* NONE */ in
defm BCRD : BCRm<"br${cf}.d", "br.d", 0x18, I64, f64, simm7Op64, uimm6Op64>;
let cx = 1, cx2 = 1, bpf = 0 /* NONE */ in
defm BCRS : BCRm<"br${cf}.s", "br.s", 0x18, F32, f32, simm7Op32, uimm6Op32>;
let cx = 0, cy = 0, cz = 1, hasSideEffects = 0 in {
let sy = 3 in
def SHMri : RM<
0x31, (outs), (ins MEMASri:$addr, I64:$sx),
"shm.l $sx, $addr">;
}
let cx = 0, sx = 0, cy = 0, sy = 0, cz = 0, sz = 0, hasSideEffects = 0 in
def MONC : RR<
0x3F, (outs), (ins),
"monc">;
// Save Instruction Counter
let cx = 0, cy = 0, sy = 0, cz = 0, sz = 0, hasSideEffects = 0 /* , Uses = [IC] */ in
def SIC : RR<0x28, (outs I32:$sx), (ins), "sic $sx">;
//===----------------------------------------------------------------------===//
// Instructions for CodeGenOnly
//===----------------------------------------------------------------------===//
let isCodeGenOnly = 1 in {
// Call instruction
let Defs = [SX10], Uses = [SX11], hasDelaySlot = 1, isCall = 1, hasSideEffects = 0 in {
let cx = 0, sx = 10, cy = 0, sy = 0, cz = 0, sz = 0 in
def CALL : RM<
0x08, (outs), (ins calltarget:$imm32, variable_ops),
"bsic %lr, $imm32">;
// use sz to represent a register
let cx = 0, sx = 10, cy = 0, sy = 0, cz = 1, imm32 = 0 in
def CALLr : RM<
0x08, (outs), (ins I64:$sz, variable_ops),
"bsic %lr, (,$sz)">;
}
}
//===----------------------------------------------------------------------===//
// Pattern Matchings
//===----------------------------------------------------------------------===//
// Small immediates.
def : Pat<(i32 simm7:$val), (OR32im1 imm:$val, 0)>;
def : Pat<(i64 simm7:$val), (ORim1 imm:$val, 0)>;
// Medium immediates.
def : Pat<(i32 simm32:$val), (LEA32zzi imm:$val)>;
def : Pat<(i64 simm32:$val), (LEAzzi imm:$val)>;
def : Pat<(i64 uimm32:$val), (ANDrm0 (LEAzzi imm:$val), 32)>;
// Arbitrary immediates.
def : Pat<(i64 lozero:$val),
(LEASLzzi (HI32 imm:$val))>;
def : Pat<(i64 lomsbzero:$val),
(LEASLrzi (LEAzzi (LO32 imm:$val)), (HI32 imm:$val))>;
def : Pat<(i64 imm:$val),
(LEASLrzi (ANDrm0 (LEAzzi (LO32 imm:$val)), 32),
(HI32 imm:$val))>;
// floating point
def : Pat<(f32 fpimm:$val),
(COPY_TO_REGCLASS (LEASLzzi (LOFP32 $val)), F32)>;
def : Pat<(f64 fplozero:$val),
(LEASLzzi (HIFP32 $val))>;
def : Pat<(f64 fplomsbzero:$val),
(LEASLrzi (LEAzzi (LOFP32 $val)), (HIFP32 $val))>;
def : Pat<(f64 fpimm:$val),
(LEASLrzi (ANDrm0 (LEAzzi (LOFP32 $val)), 32),
(HIFP32 $val))>;
// The same integer registers are used for i32 and i64 values.
// When registers hold i32 values, the high bits are unused.
// TODO Use standard expansion for shift-based lowering of sext_inreg
// Cast to i1
def : Pat<(sext_inreg I32:$src, i1),
(SRAri (SLAri $src, 31), 31)>;
def : Pat<(sext_inreg I64:$src, i1),
(SRAXri (SLLri $src, 63), 63)>;
// Cast to i8
def : Pat<(sext_inreg I32:$src, i8),
(SRAri (SLAri $src, 24), 24)>;
def : Pat<(sext_inreg I64:$src, i8),
(SRAXri (SLLri $src, 56), 56)>;
def : Pat<(sext_inreg (i32 (trunc i64:$src)), i8),
(EXTRACT_SUBREG (SRAXri (SLLri $src, 56), 56), sub_i32)>;
def : Pat<(and (trunc i64:$src), 0xff),
(AND32rm0 (EXTRACT_SUBREG $src, sub_i32), 56)>;
// Cast to i16
def : Pat<(sext_inreg I32:$src, i16),
(SRAri (SLAri $src, 16), 16)>;
def : Pat<(sext_inreg I64:$src, i16),
(SRAXri (SLLri $src, 48), 48)>;
def : Pat<(sext_inreg (i32 (trunc i64:$src)), i16),
(EXTRACT_SUBREG (SRAXri (SLLri $src, 48), 48), sub_i32)>;
def : Pat<(and (trunc i64:$src), 0xffff),
(AND32rm0 (EXTRACT_SUBREG $src, sub_i32), 48)>;
// Cast to i32
def : Pat<(i32 (trunc i64:$src)),
(ADSrm1 (EXTRACT_SUBREG $src, sub_i32), 0)>;
// Cast to i64
def : Pat<(sext_inreg I64:$src, i32),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)),
(ADSrm1 (EXTRACT_SUBREG $src, sub_i32), 0), sub_i32)>;
def : Pat<(i64 (sext i32:$sy)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), (ADSrm1 $sy, 0), sub_i32)>;
def : Pat<(i64 (zext i32:$sy)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), (ADSUrm1 $sy, 0), sub_i32)>;
def : Pat<(i64 (fp_to_sint f32:$sy)), (FIXXr (CVDr $sy))>;
// Cast to f32
def : Pat<(f32 (sint_to_fp i64:$sy)), (CVSr (FLTXr i64:$sy))>;
def : Pat<(i64 (anyext i32:$sy)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $sy, sub_i32)>;
// extload, sextload and zextload stuff
def : Pat<(i64 (sextloadi8 ADDRri:$addr)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), (LD1Bri MEMri:$addr), sub_i32)>;
def : Pat<(i64 (zextloadi8 ADDRri:$addr)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), (LD1BUri MEMri:$addr), sub_i32)>;
def : Pat<(i64 (sextloadi16 ADDRri:$addr)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), (LD2Bri MEMri:$addr), sub_i32)>;
def : Pat<(i64 (zextloadi16 ADDRri:$addr)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), (LD2BUri MEMri:$addr), sub_i32)>;
def : Pat<(i64 (sextloadi32 ADDRri:$addr)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), (LDLri MEMri:$addr), sub_i32)>;
def : Pat<(i64 (zextloadi32 ADDRri:$addr)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), (LDLUri MEMri:$addr), sub_i32)>;
def : Pat<(i64 (extloadi8 ADDRri:$addr)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), (LD1BUri MEMri:$addr), sub_i32)>;
def : Pat<(i64 (extloadi16 ADDRri:$addr)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), (LD2BUri MEMri:$addr), sub_i32)>;
def : Pat<(i64 (extloadi32 ADDRri:$addr)),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), (LDLri MEMri:$addr), sub_i32)>;
// anyextload
def : Pat<(extloadi8 ADDRri:$addr), (LD1BUri MEMri:$addr)>;
def : Pat<(extloadi16 ADDRri:$addr), (LD2BUri MEMri:$addr)>;
// truncstore
def : Pat<(truncstorei8 i64:$src, ADDRri:$addr),
(ST1Bri MEMri:$addr, (EXTRACT_SUBREG $src, sub_i32))>;
def : Pat<(truncstorei16 i64:$src, ADDRri:$addr),
(ST2Bri MEMri:$addr, (EXTRACT_SUBREG $src, sub_i32))>;
def : Pat<(truncstorei32 i64:$src, ADDRri:$addr),
(STLri MEMri:$addr, (EXTRACT_SUBREG $src, sub_i32))>;
// Address calculation and its optimization
def : Pat<(VEhi tglobaladdr:$in), (LEASLzzi tglobaladdr:$in)>;
def : Pat<(VElo tglobaladdr:$in), (ANDrm0 (LEAzzi tglobaladdr:$in), 32)>;
def : Pat<(add (VEhi tglobaladdr:$in1), (VElo tglobaladdr:$in2)),
(LEASLrzi (ANDrm0 (LEAzzi tglobaladdr:$in2), 32),
(tglobaladdr:$in1))>;
// GlobalTLS address calculation and its optimization
def : Pat<(VEhi tglobaltlsaddr:$in), (LEASLzzi tglobaltlsaddr:$in)>;
def : Pat<(VElo tglobaltlsaddr:$in), (ANDrm0 (LEAzzi tglobaltlsaddr:$in), 32)>;
def : Pat<(add (VEhi tglobaltlsaddr:$in1), (VElo tglobaltlsaddr:$in2)),
(LEASLrzi (ANDrm0 (LEAzzi tglobaltlsaddr:$in2), 32),
(tglobaltlsaddr:$in1))>;
// Address calculation and its optimization
def : Pat<(VEhi texternalsym:$in), (LEASLzzi texternalsym:$in)>;
def : Pat<(VElo texternalsym:$in), (ANDrm0 (LEAzzi texternalsym:$in), 32)>;
def : Pat<(add (VEhi texternalsym:$in1), (VElo texternalsym:$in2)),
(LEASLrzi (ANDrm0 (LEAzzi texternalsym:$in2), 32),
(texternalsym:$in1))>;
// Calls
def : Pat<(call tglobaladdr:$dst),
(CALL tglobaladdr:$dst)>;
def : Pat<(call i64:$dst),
(CALLr i64:$dst)>;
// Branches
def : Pat<(br bb:$addr), (BCRLa bb:$addr)>;
// brcc
def : Pat<(brcc CCSIOp:$cond, i32:$l, i32:$r, bb:$addr),
(BCRWrr (icond2cc $cond), $l, $r, bb:$addr)>;
def : Pat<(brcc CCUIOp:$cond, i32:$l, i32:$r, bb:$addr),
(BCRWir (icond2cc $cond), 0, (CMPUWrr $r, $l), bb:$addr)>;
def : Pat<(brcc CCSIOp:$cond, i64:$l, i64:$r, bb:$addr),
(BCRLrr (icond2cc $cond), $l, $r, bb:$addr)>;
def : Pat<(brcc CCUIOp:$cond, i64:$l, i64:$r, bb:$addr),
(BCRLir (icond2cc $cond), 0, (CMPrr $r, $l), bb:$addr)>;
def : Pat<(brcc cond:$cond, f32:$l, f32:$r, bb:$addr),
(BCRSrr (fcond2cc $cond), $l, $r, bb:$addr)>;
def : Pat<(brcc cond:$cond, f64:$l, f64:$r, bb:$addr),
(BCRDrr (fcond2cc $cond), $l, $r, bb:$addr)>;
//===----------------------------------------------------------------------===//
// Pseudo Instructions
//===----------------------------------------------------------------------===//
// GETGOT for PIC
let Defs = [SX15 /* %got */, SX16 /* %plt */], hasSideEffects = 0 in {
def GETGOT : Pseudo<(outs getGOT:$getpcseq), (ins), "$getpcseq">;
}
// GETFUNPLT for PIC
let hasSideEffects = 0 in
def GETFUNPLT : Pseudo<(outs I64:$dst), (ins i64imm:$addr),
"$dst, $addr",
[(set iPTR:$dst, (GetFunPLT tglobaladdr:$addr))] >;
def : Pat<(GetFunPLT tglobaladdr:$dst),
(GETFUNPLT tglobaladdr:$dst)>;
def : Pat<(GetFunPLT texternalsym:$dst),
(GETFUNPLT texternalsym:$dst)>;
// GETTLSADDR for TLS
let Defs = [SX0, SX10, SX12], hasSideEffects = 0 in
def GETTLSADDR : Pseudo<(outs), (ins i64imm:$addr),
"# GETTLSADDR $addr",
[(GetTLSAddr tglobaltlsaddr:$addr)] >;
def : Pat<(GetTLSAddr tglobaltlsaddr:$dst),
(GETTLSADDR tglobaltlsaddr:$dst)>;
let Defs = [SX11], Uses = [SX11], hasSideEffects = 0 in {
def ADJCALLSTACKDOWN : Pseudo<(outs), (ins i64imm:$amt, i64imm:$amt2),
"# ADJCALLSTACKDOWN $amt, $amt2",
[(callseq_start timm:$amt, timm:$amt2)]>;
def ADJCALLSTACKUP : Pseudo<(outs), (ins i64imm:$amt1, i64imm:$amt2),
"# ADJCALLSTACKUP $amt1",
[(callseq_end timm:$amt1, timm:$amt2)]>;
}
let Defs = [SX8], Uses = [SX8, SX11], hasSideEffects = 0 in
def EXTEND_STACK : Pseudo<(outs), (ins),
"# EXTEND STACK",
[]>;
let hasSideEffects = 0 in
def EXTEND_STACK_GUARD : Pseudo<(outs), (ins),
"# EXTEND STACK GUARD",
[]>;
// SETCC pattern matches
//
// CMP %tmp, lhs, rhs ; compare lhs and rhs
// or %res, 0, (0)1 ; initialize by 0
// CMOV %res, (63)0, %tmp ; set 1 if %tmp is true
def : Pat<(i32 (setcc i64:$LHS, i64:$RHS, CCSIOp:$cond)),
(EXTRACT_SUBREG
(CMOVLrm0 (icond2cc $cond),
(CPXrr i64:$LHS, i64:$RHS),
63,
(ORim1 0, 0)), sub_i32)>;
def : Pat<(i32 (setcc i64:$LHS, i64:$RHS, CCUIOp:$cond)),
(EXTRACT_SUBREG
(CMOVLrm0 (icond2cc $cond),
(CMPrr i64:$LHS, i64:$RHS),
63,
(ORim1 0, 0)), sub_i32)>;
def : Pat<(i32 (setcc i32:$LHS, i32:$RHS, CCSIOp:$cond)),
(EXTRACT_SUBREG
(CMOVWrm0 (icond2cc $cond),
(CPSrr i32:$LHS, i32:$RHS),
63,
(ORim1 0, 0)), sub_i32)>;
def : Pat<(i32 (setcc i32:$LHS, i32:$RHS, CCUIOp:$cond)),
(EXTRACT_SUBREG
(CMOVWrm0 (icond2cc $cond),
(CMPUWrr i32:$LHS, i32:$RHS),
63,
(ORim1 0, 0)), sub_i32)>;
def : Pat<(i32 (setcc f64:$LHS, f64:$RHS, cond:$cond)),
(EXTRACT_SUBREG
(CMOVDrm0 (fcond2cc $cond),
(FCPrr f64:$LHS, f64:$RHS),
63,
(ORim1 0, 0)), sub_i32)>;
def : Pat<(i32 (setcc f32:$LHS, f32:$RHS, cond:$cond)),
(EXTRACT_SUBREG
(CMOVSrm0 (fcond2cc $cond),
(FCPSrr f32:$LHS, f32:$RHS),
63,
(ORim1 0, 0)), sub_i32)>;
// Special SELECTCC pattern matches
// Use min/max for better performance.
//
// MAX/MIN %res, %lhs, %rhs
def : Pat<(f64 (selectcc f64:$LHS, f64:$RHS, f64:$LHS, f64:$RHS, SETOGT)),
(FCMArr $LHS, $RHS)>;
def : Pat<(f32 (selectcc f32:$LHS, f32:$RHS, f32:$LHS, f32:$RHS, SETOGT)),
(FCMASrr $LHS, $RHS)>;
def : Pat<(i64 (selectcc i64:$LHS, i64:$RHS, i64:$LHS, i64:$RHS, SETGT)),
(CMXarr $LHS, $RHS)>;
def : Pat<(i32 (selectcc i32:$LHS, i32:$RHS, i32:$LHS, i32:$RHS, SETGT)),
(CMSarr $LHS, $RHS)>;
def : Pat<(f64 (selectcc f64:$LHS, f64:$RHS, f64:$LHS, f64:$RHS, SETOGE)),
(FCMArr $LHS, $RHS)>;
def : Pat<(f32 (selectcc f32:$LHS, f32:$RHS, f32:$LHS, f32:$RHS, SETOGE)),
(FCMASrr $LHS, $RHS)>;
def : Pat<(i64 (selectcc i64:$LHS, i64:$RHS, i64:$LHS, i64:$RHS, SETGE)),
(CMXarr $LHS, $RHS)>;
def : Pat<(i32 (selectcc i32:$LHS, i32:$RHS, i32:$LHS, i32:$RHS, SETGE)),
(CMSarr $LHS, $RHS)>;
def : Pat<(f64 (selectcc f64:$LHS, f64:$RHS, f64:$LHS, f64:$RHS, SETOLT)),
(FCMIrr $LHS, $RHS)>;
def : Pat<(f32 (selectcc f32:$LHS, f32:$RHS, f32:$LHS, f32:$RHS, SETOLT)),
(FCMISrr $LHS, $RHS)>;
def : Pat<(i64 (selectcc i64:$LHS, i64:$RHS, i64:$LHS, i64:$RHS, SETLT)),
(CMXirr $LHS, $RHS)>;
def : Pat<(i32 (selectcc i32:$LHS, i32:$RHS, i32:$LHS, i32:$RHS, SETLT)),
(CMSirr $LHS, $RHS)>;
def : Pat<(f64 (selectcc f64:$LHS, f64:$RHS, f64:$LHS, f64:$RHS, SETOLE)),
(FCMIrr $LHS, $RHS)>;
def : Pat<(f32 (selectcc f32:$LHS, f32:$RHS, f32:$LHS, f32:$RHS, SETOLE)),
(FCMISrr $LHS, $RHS)>;
def : Pat<(i64 (selectcc i64:$LHS, i64:$RHS, i64:$LHS, i64:$RHS, SETLE)),
(CMXirr $LHS, $RHS)>;
def : Pat<(i32 (selectcc i32:$LHS, i32:$RHS, i32:$LHS, i32:$RHS, SETLE)),
(CMSirr $LHS, $RHS)>;
// Generic SELECTCC pattern matches
//
// CMP %tmp, %l, %r ; compare %l and %r
// or %res, %f, (0)1 ; initialize by %f
// CMOV %res, %t, %tmp ; set %t if %tmp is true
// selectcc for i64 result
def : Pat<(i64 (selectcc i32:$l, i32:$r, i64:$t, i64:$f, CCSIOp:$cond)),
(CMOVWrr (icond2cc $cond), (CPSrr $l, $r), $t, $f)>;
def : Pat<(i64 (selectcc i32:$l, i32:$r, i64:$t, i64:$f, CCUIOp:$cond)),
(CMOVWrr (icond2cc $cond), (CMPUWrr $l, $r), $t, $f)>;
def : Pat<(i64 (selectcc i64:$l, i64:$r, i64:$t, i64:$f, CCSIOp:$cond)),
(CMOVLrr (icond2cc $cond), (CPXrr $l, $r), $t, $f)>;
def : Pat<(i64 (selectcc i64:$l, i64:$r, i64:$t, i64:$f, CCUIOp:$cond)),
(CMOVLrr (icond2cc $cond), (CMPrr $l, $r), $t, $f)>;
def : Pat<(i64 (selectcc f32:$l, f32:$r, i64:$t, i64:$f, cond:$cond)),
(CMOVSrr (fcond2cc $cond), (FCPSrr $l, $r), $t, $f)>;
def : Pat<(i64 (selectcc f64:$l, f64:$r, i64:$t, i64:$f, cond:$cond)),
(CMOVDrr (fcond2cc $cond), (FCPrr $l, $r), $t, $f)>;
// selectcc for i32 result
def : Pat<(i32 (selectcc i32:$l, i32:$r, i32:$t, i32:$f, CCSIOp:$cond)),
(EXTRACT_SUBREG
(CMOVWrr (icond2cc $cond),
(CPSrr $l, $r),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $t, sub_i32),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $f, sub_i32)),
sub_i32)>;
def : Pat<(i32 (selectcc i32:$l, i32:$r, i32:$t, i32:$f, CCUIOp:$cond)),
(EXTRACT_SUBREG
(CMOVWrr (icond2cc $cond),
(CMPUWrr $l, $r),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $t, sub_i32),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $f, sub_i32)),
sub_i32)>;
def : Pat<(i32 (selectcc i64:$l, i64:$r, i32:$t, i32:$f, CCSIOp:$cond)),
(EXTRACT_SUBREG
(CMOVLrr (icond2cc $cond),
(CPXrr $l, $r),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $t, sub_i32),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $f, sub_i32)),
sub_i32)>;
def : Pat<(i32 (selectcc i64:$l, i64:$r, i32:$t, i32:$f, CCUIOp:$cond)),
(EXTRACT_SUBREG
(CMOVLrr (icond2cc $cond),
(CMPrr $l, $r),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $t, sub_i32),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $f, sub_i32)),
sub_i32)>;
def : Pat<(i32 (selectcc f32:$l, f32:$r, i32:$t, i32:$f, cond:$cond)),
(EXTRACT_SUBREG
(CMOVSrr (fcond2cc $cond),
(FCPSrr $l, $r),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $t, sub_i32),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $f, sub_i32)),
sub_i32)>;
def : Pat<(i32 (selectcc f64:$l, f64:$r, i32:$t, i32:$f, cond:$cond)),
(EXTRACT_SUBREG
(CMOVDrr (fcond2cc $cond),
(FCPrr $l, $r),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $t, sub_i32),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $f, sub_i32)),
sub_i32)>;
// selectcc for f64 result
def : Pat<(f64 (selectcc i32:$l, i32:$r, f64:$t, f64:$f, CCSIOp:$cond)),
(CMOVWrr (icond2cc $cond), (CPSrr $l, $r), $t, $f)>;
def : Pat<(f64 (selectcc i32:$l, i32:$r, f64:$t, f64:$f, CCUIOp:$cond)),
(CMOVWrr (icond2cc $cond), (CMPUWrr $l, $r), $t, $f)>;
def : Pat<(f64 (selectcc i64:$l, i64:$r, f64:$t, f64:$f, CCSIOp:$cond)),
(CMOVLrr (icond2cc $cond), (CPXrr $l, $r), $t, $f)>;
def : Pat<(f64 (selectcc i64:$l, i64:$r, f64:$t, f64:$f, CCUIOp:$cond)),
(CMOVLrr (icond2cc $cond), (CMPrr $l, $r), $t, $f)>;
def : Pat<(f64 (selectcc f32:$l, f32:$r, f64:$t, f64:$f, cond:$cond)),
(CMOVSrr (fcond2cc $cond), (FCPSrr $l, $r), $t, $f)>;
def : Pat<(f64 (selectcc f64:$l, f64:$r, f64:$t, f64:$f, cond:$cond)),
(CMOVDrr (fcond2cc $cond), (FCPrr $l, $r), $t, $f)>;
// selectcc for f32 result
def : Pat<(f32 (selectcc i32:$l, i32:$r, f32:$t, f32:$f, CCSIOp:$cond)),
(EXTRACT_SUBREG
(CMOVWrr (icond2cc $cond),
(CPSrr $l, $r),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $t, sub_f32),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $f, sub_f32)),
sub_f32)>;
def : Pat<(f32 (selectcc i32:$l, i32:$r, f32:$t, f32:$f, CCUIOp:$cond)),
(EXTRACT_SUBREG
(CMOVWrr (icond2cc $cond),
(CMPUWrr $l, $r),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $t, sub_f32),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $f, sub_f32)),
sub_f32)>;
def : Pat<(f32 (selectcc i64:$l, i64:$r, f32:$t, f32:$f, CCSIOp:$cond)),
(EXTRACT_SUBREG
(CMOVLrr (icond2cc $cond),
(CPXrr $l, $r),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $t, sub_f32),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $f, sub_f32)),
sub_f32)>;
def : Pat<(f32 (selectcc i64:$l, i64:$r, f32:$t, f32:$f, CCUIOp:$cond)),
(EXTRACT_SUBREG
(CMOVLrr (icond2cc $cond),
(CMPrr $l, $r),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $t, sub_f32),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $f, sub_f32)),
sub_f32)>;
def : Pat<(f32 (selectcc f32:$l, f32:$r, f32:$t, f32:$f, cond:$cond)),
(EXTRACT_SUBREG
(CMOVSrr (fcond2cc $cond),
(FCPSrr $l, $r),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $t, sub_f32),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $f, sub_f32)),
sub_f32)>;
def : Pat<(f32 (selectcc f64:$l, f64:$r, f32:$t, f32:$f, cond:$cond)),
(EXTRACT_SUBREG
(CMOVDrr (fcond2cc $cond),
(FCPrr $l, $r),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $t, sub_f32),
(INSERT_SUBREG (f64 (IMPLICIT_DEF)), $f, sub_f32)),
sub_f32)>;
// Generic SELECT pattern matches
// Use cmov.w for all cases since %pred holds i32.
//
// CMOV.w.ne %res, %tval, %tmp ; set tval if %tmp is true
def : Pat<(i64 (select i32:$pred, i64:$t, i64:$f)),
(CMOVWrr CC_INE, $pred, $t, $f)>;
def : Pat<(i32 (select i32:$pred, i32:$t, i32:$f)),
(EXTRACT_SUBREG
(CMOVWrr CC_INE, $pred,
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $t, sub_i32),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $f, sub_i32)),
sub_i32)>;
def : Pat<(f64 (select i32:$pred, f64:$t, f64:$f)),
(CMOVWrr CC_INE, $pred, $t, $f)>;
def : Pat<(f32 (select i32:$pred, f32:$t, f32:$f)),
(EXTRACT_SUBREG
(CMOVWrr CC_INE, $pred,
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $t, sub_f32),
(INSERT_SUBREG (i64 (IMPLICIT_DEF)), $f, sub_f32)),
sub_f32)>;
// bitconvert
def : Pat<(f64 (bitconvert i64:$src)), (COPY_TO_REGCLASS $src, I64)>;
def : Pat<(i64 (bitconvert f64:$src)), (COPY_TO_REGCLASS $src, I64)>;
def : Pat<(i32 (bitconvert f32:$op)),
(EXTRACT_SUBREG (SRAXri (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
$op, sub_f32), 32), sub_i32)>;
def : Pat<(f32 (bitconvert i32:$op)),
(EXTRACT_SUBREG (SLLri (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
$op, sub_i32), 32), sub_f32)>;
// Bits operations pattern matchings.
def : Pat<(i32 (ctpop i32:$src)),
(EXTRACT_SUBREG (PCNTr (ANDrm0 (INSERT_SUBREG
(i64 (IMPLICIT_DEF)), $src, sub_i32), 32)), sub_i32)>;
def : Pat<(i32 (ctlz i32:$src)),
(EXTRACT_SUBREG (LDZr (SLLri (INSERT_SUBREG
(i64 (IMPLICIT_DEF)), $src, sub_i32), 32)), sub_i32)>;
def : Pat<(i64 (bswap i64:$src)),
(BSWPri $src, 0)>;
def : Pat<(i32 (bswap i32:$src)),
(EXTRACT_SUBREG (BSWPri (INSERT_SUBREG
(i64 (IMPLICIT_DEF)), $src, sub_i32), 1), sub_i32)>;
// Several special pattern matches to optimize code
def : Pat<(i32 (and i32:$lhs, 0xff)),
(AND32rm0 $lhs, 56)>;
def : Pat<(i32 (and i32:$lhs, 0xffff)),
(AND32rm0 $lhs, 48)>;
def : Pat<(i32 (and i32:$lhs, 0xffffffff)),
(AND32rm0 $lhs, 32)>;