llvm-6502/lib/Target/Hexagon/HexagonInstrInfo.td
2014-12-11 17:08:21 +00:00

3466 lines
130 KiB
TableGen

//==- HexagonInstrInfo.td - Target Description for Hexagon -*- tablegen -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file describes the Hexagon instructions in TableGen format.
//
//===----------------------------------------------------------------------===//
include "HexagonInstrFormats.td"
include "HexagonOperands.td"
// Pattern fragment that combines the value type and the register class
// into a single parameter.
// The pat frags in the definitions below need to have a named register,
// otherwise i32 will be assumed regardless of the register class. The
// name of the register does not matter.
def I1 : PatLeaf<(i1 PredRegs:$R)>;
def I32 : PatLeaf<(i32 IntRegs:$R)>;
def I64 : PatLeaf<(i64 DoubleRegs:$R)>;
def F32 : PatLeaf<(f32 IntRegs:$R)>;
def F64 : PatLeaf<(f64 DoubleRegs:$R)>;
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// Compare
//===----------------------------------------------------------------------===//
let hasSideEffects = 0, isCompare = 1, InputType = "imm", isExtendable = 1,
opExtendable = 2 in
class T_CMP <string mnemonic, bits<2> MajOp, bit isNot, Operand ImmOp>
: ALU32Inst <(outs PredRegs:$dst),
(ins IntRegs:$src1, ImmOp:$src2),
"$dst = "#!if(isNot, "!","")#mnemonic#"($src1, #$src2)",
[], "",ALU32_2op_tc_2early_SLOT0123 >, ImmRegRel {
bits<2> dst;
bits<5> src1;
bits<10> src2;
let CextOpcode = mnemonic;
let opExtentBits = !if(!eq(mnemonic, "cmp.gtu"), 9, 10);
let isExtentSigned = !if(!eq(mnemonic, "cmp.gtu"), 0, 1);
let IClass = 0b0111;
let Inst{27-24} = 0b0101;
let Inst{23-22} = MajOp;
let Inst{21} = !if(!eq(mnemonic, "cmp.gtu"), 0, src2{9});
let Inst{20-16} = src1;
let Inst{13-5} = src2{8-0};
let Inst{4} = isNot;
let Inst{3-2} = 0b00;
let Inst{1-0} = dst;
}
def C2_cmpeqi : T_CMP <"cmp.eq", 0b00, 0, s10Ext>;
def C2_cmpgti : T_CMP <"cmp.gt", 0b01, 0, s10Ext>;
def C2_cmpgtui : T_CMP <"cmp.gtu", 0b10, 0, u9Ext>;
class T_CMP_pat <InstHexagon MI, PatFrag OpNode, PatLeaf ImmPred>
: Pat<(i1 (OpNode (i32 IntRegs:$src1), ImmPred:$src2)),
(MI IntRegs:$src1, ImmPred:$src2)>;
def : T_CMP_pat <C2_cmpeqi, seteq, s10ImmPred>;
def : T_CMP_pat <C2_cmpgti, setgt, s10ImmPred>;
def : T_CMP_pat <C2_cmpgtui, setugt, u9ImmPred>;
//===----------------------------------------------------------------------===//
// ALU32/ALU +
//===----------------------------------------------------------------------===//
def SDTHexagonI64I32I32 : SDTypeProfile<1, 2,
[SDTCisVT<0, i64>, SDTCisVT<1, i32>, SDTCisSameAs<1, 2>]>;
def HexagonCOMBINE : SDNode<"HexagonISD::COMBINE", SDTHexagonI64I32I32>;
let hasSideEffects = 0, hasNewValue = 1, InputType = "reg" in
class T_ALU32_3op<string mnemonic, bits<3> MajOp, bits<3> MinOp, bit OpsRev,
bit IsComm>
: ALU32_rr<(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt),
"$Rd = "#mnemonic#"($Rs, $Rt)",
[], "", ALU32_3op_tc_1_SLOT0123>, ImmRegRel, PredRel {
let isCommutable = IsComm;
let BaseOpcode = mnemonic#_rr;
let CextOpcode = mnemonic;
bits<5> Rs;
bits<5> Rt;
bits<5> Rd;
let IClass = 0b1111;
let Inst{27} = 0b0;
let Inst{26-24} = MajOp;
let Inst{23-21} = MinOp;
let Inst{20-16} = !if(OpsRev,Rt,Rs);
let Inst{12-8} = !if(OpsRev,Rs,Rt);
let Inst{4-0} = Rd;
}
let hasSideEffects = 0, hasNewValue = 1 in
class T_ALU32_3op_pred<string mnemonic, bits<3> MajOp, bits<3> MinOp,
bit OpsRev, bit PredNot, bit PredNew>
: ALU32_rr<(outs IntRegs:$Rd), (ins PredRegs:$Pu, IntRegs:$Rs, IntRegs:$Rt),
"if ("#!if(PredNot,"!","")#"$Pu"#!if(PredNew,".new","")#") "#
"$Rd = "#mnemonic#"($Rs, $Rt)",
[], "", ALU32_3op_tc_1_SLOT0123>, ImmRegRel, PredNewRel {
let isPredicated = 1;
let isPredicatedFalse = PredNot;
let isPredicatedNew = PredNew;
let BaseOpcode = mnemonic#_rr;
let CextOpcode = mnemonic;
bits<2> Pu;
bits<5> Rs;
bits<5> Rt;
bits<5> Rd;
let IClass = 0b1111;
let Inst{27} = 0b1;
let Inst{26-24} = MajOp;
let Inst{23-21} = MinOp;
let Inst{20-16} = !if(OpsRev,Rt,Rs);
let Inst{13} = PredNew;
let Inst{12-8} = !if(OpsRev,Rs,Rt);
let Inst{7} = PredNot;
let Inst{6-5} = Pu;
let Inst{4-0} = Rd;
}
class T_ALU32_combineh<string Op1, string Op2, bits<3> MajOp, bits<3> MinOp,
bit OpsRev>
: T_ALU32_3op<"", MajOp, MinOp, OpsRev, 0> {
let AsmString = "$Rd = combine($Rs"#Op1#", $Rt"#Op2#")";
}
let isCodeGenOnly = 0 in {
def A2_combine_hh : T_ALU32_combineh<".h", ".h", 0b011, 0b100, 1>;
def A2_combine_hl : T_ALU32_combineh<".h", ".l", 0b011, 0b101, 1>;
def A2_combine_lh : T_ALU32_combineh<".l", ".h", 0b011, 0b110, 1>;
def A2_combine_ll : T_ALU32_combineh<".l", ".l", 0b011, 0b111, 1>;
}
class T_ALU32_3op_sfx<string mnemonic, string suffix, bits<3> MajOp,
bits<3> MinOp, bit OpsRev, bit IsComm>
: T_ALU32_3op<"", MajOp, MinOp, OpsRev, IsComm> {
let AsmString = "$Rd = "#mnemonic#"($Rs, $Rt)"#suffix;
}
let Defs = [USR_OVF], Itinerary = ALU32_3op_tc_2_SLOT0123,
isCodeGenOnly = 0 in {
def A2_addsat : T_ALU32_3op_sfx<"add", ":sat", 0b110, 0b010, 0, 1>;
def A2_subsat : T_ALU32_3op_sfx<"sub", ":sat", 0b110, 0b110, 1, 0>;
}
multiclass T_ALU32_3op_p<string mnemonic, bits<3> MajOp, bits<3> MinOp,
bit OpsRev> {
def t : T_ALU32_3op_pred<mnemonic, MajOp, MinOp, OpsRev, 0, 0>;
def f : T_ALU32_3op_pred<mnemonic, MajOp, MinOp, OpsRev, 1, 0>;
def tnew : T_ALU32_3op_pred<mnemonic, MajOp, MinOp, OpsRev, 0, 1>;
def fnew : T_ALU32_3op_pred<mnemonic, MajOp, MinOp, OpsRev, 1, 1>;
}
multiclass T_ALU32_3op_A2<string mnemonic, bits<3> MajOp, bits<3> MinOp,
bit OpsRev, bit IsComm> {
let isPredicable = 1 in
def A2_#NAME : T_ALU32_3op <mnemonic, MajOp, MinOp, OpsRev, IsComm>;
defm A2_p#NAME : T_ALU32_3op_p<mnemonic, MajOp, MinOp, OpsRev>;
}
let isCodeGenOnly = 0 in {
defm add : T_ALU32_3op_A2<"add", 0b011, 0b000, 0, 1>;
defm and : T_ALU32_3op_A2<"and", 0b001, 0b000, 0, 1>;
defm or : T_ALU32_3op_A2<"or", 0b001, 0b001, 0, 1>;
defm sub : T_ALU32_3op_A2<"sub", 0b011, 0b001, 1, 0>;
defm xor : T_ALU32_3op_A2<"xor", 0b001, 0b011, 0, 1>;
}
// Pats for instruction selection.
class BinOp32_pat<SDNode Op, InstHexagon MI, ValueType ResT>
: Pat<(ResT (Op (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))),
(ResT (MI IntRegs:$Rs, IntRegs:$Rt))>;
def: BinOp32_pat<add, A2_add, i32>;
def: BinOp32_pat<and, A2_and, i32>;
def: BinOp32_pat<or, A2_or, i32>;
def: BinOp32_pat<sub, A2_sub, i32>;
def: BinOp32_pat<xor, A2_xor, i32>;
// A few special cases producing register pairs:
let OutOperandList = (outs DoubleRegs:$Rd), hasNewValue = 0,
isCodeGenOnly = 0 in {
def S2_packhl : T_ALU32_3op <"packhl", 0b101, 0b100, 0, 0>;
let isPredicable = 1 in
def A2_combinew : T_ALU32_3op <"combine", 0b101, 0b000, 0, 0>;
// Conditional combinew uses "newt/f" instead of "t/fnew".
def C2_ccombinewt : T_ALU32_3op_pred<"combine", 0b101, 0b000, 0, 0, 0>;
def C2_ccombinewf : T_ALU32_3op_pred<"combine", 0b101, 0b000, 0, 1, 0>;
def C2_ccombinewnewt : T_ALU32_3op_pred<"combine", 0b101, 0b000, 0, 0, 1>;
def C2_ccombinewnewf : T_ALU32_3op_pred<"combine", 0b101, 0b000, 0, 1, 1>;
}
let hasSideEffects = 0, hasNewValue = 1, isCompare = 1, InputType = "reg" in
class T_ALU32_3op_cmp<string mnemonic, bits<2> MinOp, bit IsNeg, bit IsComm>
: ALU32_rr<(outs PredRegs:$Pd), (ins IntRegs:$Rs, IntRegs:$Rt),
"$Pd = "#mnemonic#"($Rs, $Rt)",
[], "", ALU32_3op_tc_1_SLOT0123>, ImmRegRel {
let CextOpcode = mnemonic;
let isCommutable = IsComm;
bits<5> Rs;
bits<5> Rt;
bits<2> Pd;
let IClass = 0b1111;
let Inst{27-24} = 0b0010;
let Inst{22-21} = MinOp;
let Inst{20-16} = Rs;
let Inst{12-8} = Rt;
let Inst{4} = IsNeg;
let Inst{3-2} = 0b00;
let Inst{1-0} = Pd;
}
let Itinerary = ALU32_3op_tc_2early_SLOT0123, isCodeGenOnly = 0 in {
def C2_cmpeq : T_ALU32_3op_cmp< "cmp.eq", 0b00, 0, 1>;
def C2_cmpgt : T_ALU32_3op_cmp< "cmp.gt", 0b10, 0, 0>;
def C2_cmpgtu : T_ALU32_3op_cmp< "cmp.gtu", 0b11, 0, 0>;
}
// Patfrag to convert the usual comparison patfrags (e.g. setlt) to ones
// that reverse the order of the operands.
class RevCmp<PatFrag F> : PatFrag<(ops node:$rhs, node:$lhs), F.Fragment>;
// Pats for compares. They use PatFrags as operands, not SDNodes,
// since seteq/setgt/etc. are defined as ParFrags.
class T_cmp32_rr_pat<InstHexagon MI, PatFrag Op, ValueType VT>
: Pat<(VT (Op (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))),
(VT (MI IntRegs:$Rs, IntRegs:$Rt))>;
def: T_cmp32_rr_pat<C2_cmpeq, seteq, i1>;
def: T_cmp32_rr_pat<C2_cmpgt, setgt, i1>;
def: T_cmp32_rr_pat<C2_cmpgtu, setugt, i1>;
def: T_cmp32_rr_pat<C2_cmpgt, RevCmp<setlt>, i1>;
def: T_cmp32_rr_pat<C2_cmpgtu, RevCmp<setult>, i1>;
let CextOpcode = "MUX", InputType = "reg", hasNewValue = 1,
isCodeGenOnly = 0 in
def C2_mux: ALU32_rr<(outs IntRegs:$Rd),
(ins PredRegs:$Pu, IntRegs:$Rs, IntRegs:$Rt),
"$Rd = mux($Pu, $Rs, $Rt)", [], "", ALU32_3op_tc_1_SLOT0123>, ImmRegRel {
bits<5> Rd;
bits<2> Pu;
bits<5> Rs;
bits<5> Rt;
let CextOpcode = "mux";
let InputType = "reg";
let hasSideEffects = 0;
let IClass = 0b1111;
let Inst{27-24} = 0b0100;
let Inst{20-16} = Rs;
let Inst{12-8} = Rt;
let Inst{6-5} = Pu;
let Inst{4-0} = Rd;
}
def: Pat<(i32 (select (i1 PredRegs:$Pu), (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))),
(C2_mux PredRegs:$Pu, IntRegs:$Rs, IntRegs:$Rt)>;
// Combines the two immediates into a double register.
// Increase complexity to make it greater than any complexity of a combine
// that involves a register.
let isReMaterializable = 1, isMoveImm = 1, isAsCheapAsAMove = 1,
isExtentSigned = 1, isExtendable = 1, opExtentBits = 8, opExtendable = 1,
AddedComplexity = 75, isCodeGenOnly = 0 in
def A2_combineii: ALU32Inst <(outs DoubleRegs:$Rdd), (ins s8Ext:$s8, s8Imm:$S8),
"$Rdd = combine(#$s8, #$S8)",
[(set (i64 DoubleRegs:$Rdd),
(i64 (HexagonCOMBINE(i32 s8ExtPred:$s8), (i32 s8ImmPred:$S8))))]> {
bits<5> Rdd;
bits<8> s8;
bits<8> S8;
let IClass = 0b0111;
let Inst{27-23} = 0b11000;
let Inst{22-16} = S8{7-1};
let Inst{13} = S8{0};
let Inst{12-5} = s8;
let Inst{4-0} = Rdd;
}
//===----------------------------------------------------------------------===//
// Template class for predicated ADD of a reg and an Immediate value.
//===----------------------------------------------------------------------===//
let hasNewValue = 1 in
class T_Addri_Pred <bit PredNot, bit PredNew>
: ALU32_ri <(outs IntRegs:$Rd),
(ins PredRegs:$Pu, IntRegs:$Rs, s8Ext:$s8),
!if(PredNot, "if (!$Pu", "if ($Pu")#!if(PredNew,".new) $Rd = ",
") $Rd = ")#"add($Rs, #$s8)"> {
bits<5> Rd;
bits<2> Pu;
bits<5> Rs;
bits<8> s8;
let isPredicatedNew = PredNew;
let IClass = 0b0111;
let Inst{27-24} = 0b0100;
let Inst{23} = PredNot;
let Inst{22-21} = Pu;
let Inst{20-16} = Rs;
let Inst{13} = PredNew;
let Inst{12-5} = s8;
let Inst{4-0} = Rd;
}
//===----------------------------------------------------------------------===//
// A2_addi: Add a signed immediate to a register.
//===----------------------------------------------------------------------===//
let hasNewValue = 1 in
class T_Addri <Operand immOp, list<dag> pattern = [] >
: ALU32_ri <(outs IntRegs:$Rd),
(ins IntRegs:$Rs, immOp:$s16),
"$Rd = add($Rs, #$s16)", pattern,
//[(set (i32 IntRegs:$Rd), (add (i32 IntRegs:$Rs), (s16ExtPred:$s16)))],
"", ALU32_ADDI_tc_1_SLOT0123> {
bits<5> Rd;
bits<5> Rs;
bits<16> s16;
let IClass = 0b1011;
let Inst{27-21} = s16{15-9};
let Inst{20-16} = Rs;
let Inst{13-5} = s16{8-0};
let Inst{4-0} = Rd;
}
//===----------------------------------------------------------------------===//
// Multiclass for ADD of a register and an immediate value.
//===----------------------------------------------------------------------===//
multiclass Addri_Pred<string mnemonic, bit PredNot> {
let isPredicatedFalse = PredNot in {
def _c#NAME : T_Addri_Pred<PredNot, 0>;
// Predicate new
def _cdn#NAME : T_Addri_Pred<PredNot, 1>;
}
}
let isExtendable = 1, InputType = "imm" in
multiclass Addri_base<string mnemonic, SDNode OpNode> {
let CextOpcode = mnemonic, BaseOpcode = mnemonic#_ri in {
let opExtendable = 2, isExtentSigned = 1, opExtentBits = 16,
isPredicable = 1 in
def NAME : T_Addri< s16Ext, // Rd=add(Rs,#s16)
[(set (i32 IntRegs:$Rd),
(add IntRegs:$Rs, s16ExtPred:$s16))]>;
let opExtendable = 3, isExtentSigned = 1, opExtentBits = 8,
hasSideEffects = 0, isPredicated = 1 in {
defm Pt : Addri_Pred<mnemonic, 0>;
defm NotPt : Addri_Pred<mnemonic, 1>;
}
}
}
let isCodeGenOnly = 0 in
defm ADD_ri : Addri_base<"add", add>, ImmRegRel, PredNewRel;
//===----------------------------------------------------------------------===//
// Template class used for the following ALU32 instructions.
// Rd=and(Rs,#s10)
// Rd=or(Rs,#s10)
//===----------------------------------------------------------------------===//
let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 10,
InputType = "imm", hasNewValue = 1 in
class T_ALU32ri_logical <string mnemonic, SDNode OpNode, bits<2> MinOp>
: ALU32_ri <(outs IntRegs:$Rd),
(ins IntRegs:$Rs, s10Ext:$s10),
"$Rd = "#mnemonic#"($Rs, #$s10)" ,
[(set (i32 IntRegs:$Rd), (OpNode (i32 IntRegs:$Rs), s10ExtPred:$s10))]> {
bits<5> Rd;
bits<5> Rs;
bits<10> s10;
let CextOpcode = mnemonic;
let IClass = 0b0111;
let Inst{27-24} = 0b0110;
let Inst{23-22} = MinOp;
let Inst{21} = s10{9};
let Inst{20-16} = Rs;
let Inst{13-5} = s10{8-0};
let Inst{4-0} = Rd;
}
let isCodeGenOnly = 0 in {
def OR_ri : T_ALU32ri_logical<"or", or, 0b10>, ImmRegRel;
def AND_ri : T_ALU32ri_logical<"and", and, 0b00>, ImmRegRel;
}
// Subtract register from immediate
// Rd32=sub(#s10,Rs32)
let isExtendable = 1, opExtendable = 1, isExtentSigned = 1, opExtentBits = 10,
CextOpcode = "sub", InputType = "imm", hasNewValue = 1, isCodeGenOnly = 0 in
def SUB_ri: ALU32_ri <(outs IntRegs:$Rd), (ins s10Ext:$s10, IntRegs:$Rs),
"$Rd = sub(#$s10, $Rs)" ,
[(set IntRegs:$Rd, (sub s10ExtPred:$s10, IntRegs:$Rs))] > ,
ImmRegRel {
bits<5> Rd;
bits<10> s10;
bits<5> Rs;
let IClass = 0b0111;
let Inst{27-22} = 0b011001;
let Inst{21} = s10{9};
let Inst{20-16} = Rs;
let Inst{13-5} = s10{8-0};
let Inst{4-0} = Rd;
}
// Nop.
let hasSideEffects = 0, isCodeGenOnly = 0 in
def A2_nop: ALU32Inst <(outs), (ins), "nop" > {
let IClass = 0b0111;
let Inst{27-24} = 0b1111;
}
// Rd = not(Rs) gets mapped to Rd=sub(#-1, Rs).
def : Pat<(not (i32 IntRegs:$src1)),
(SUB_ri -1, (i32 IntRegs:$src1))>;
let hasSideEffects = 0, hasNewValue = 1 in
class T_tfr16<bit isHi>
: ALU32Inst <(outs IntRegs:$Rx), (ins IntRegs:$src1, u16Imm:$u16),
"$Rx"#!if(isHi, ".h", ".l")#" = #$u16",
[], "$src1 = $Rx" > {
bits<5> Rx;
bits<16> u16;
let IClass = 0b0111;
let Inst{27-26} = 0b00;
let Inst{25-24} = !if(isHi, 0b10, 0b01);
let Inst{23-22} = u16{15-14};
let Inst{21} = 0b1;
let Inst{20-16} = Rx;
let Inst{13-0} = u16{13-0};
}
let isCodeGenOnly = 0 in {
def A2_tfril: T_tfr16<0>;
def A2_tfrih: T_tfr16<1>;
}
// Conditional transfer is an alias to conditional "Rd = add(Rs, #0)".
let isPredicated = 1, hasNewValue = 1, opNewValue = 0 in
class T_tfr_pred<bit isPredNot, bit isPredNew>
: ALU32Inst<(outs IntRegs:$dst),
(ins PredRegs:$src1, IntRegs:$src2),
"if ("#!if(isPredNot, "!", "")#
"$src1"#!if(isPredNew, ".new", "")#
") $dst = $src2"> {
bits<5> dst;
bits<2> src1;
bits<5> src2;
let isPredicatedFalse = isPredNot;
let isPredicatedNew = isPredNew;
let IClass = 0b0111;
let Inst{27-24} = 0b0100;
let Inst{23} = isPredNot;
let Inst{13} = isPredNew;
let Inst{12-5} = 0;
let Inst{4-0} = dst;
let Inst{22-21} = src1;
let Inst{20-16} = src2;
}
let isPredicable = 1 in
class T_tfr : ALU32Inst<(outs IntRegs:$dst), (ins IntRegs:$src),
"$dst = $src"> {
bits<5> dst;
bits<5> src;
let IClass = 0b0111;
let Inst{27-21} = 0b0000011;
let Inst{20-16} = src;
let Inst{13} = 0b0;
let Inst{4-0} = dst;
}
let InputType = "reg", hasNewValue = 1, hasSideEffects = 0 in
multiclass tfr_base<string CextOp> {
let CextOpcode = CextOp, BaseOpcode = CextOp in {
def NAME : T_tfr;
// Predicate
def t : T_tfr_pred<0, 0>;
def f : T_tfr_pred<1, 0>;
// Predicate new
def tnew : T_tfr_pred<0, 1>;
def fnew : T_tfr_pred<1, 1>;
}
}
// Assembler mapped to C2_ccombinew[t|f|newt|newf].
// Please don't add bits to this instruction as it'll be converted into
// 'combine' before object code emission.
let isPredicated = 1 in
class T_tfrp_pred<bit PredNot, bit PredNew>
: ALU32_rr <(outs DoubleRegs:$dst),
(ins PredRegs:$src1, DoubleRegs:$src2),
"if ("#!if(PredNot, "!", "")#"$src1"
#!if(PredNew, ".new", "")#") $dst = $src2" > {
let isPredicatedFalse = PredNot;
let isPredicatedNew = PredNew;
}
// Assembler mapped to A2_combinew.
// Please don't add bits to this instruction as it'll be converted into
// 'combine' before object code emission.
class T_tfrp : ALU32Inst <(outs DoubleRegs:$dst),
(ins DoubleRegs:$src),
"$dst = $src">;
let hasSideEffects = 0 in
multiclass TFR64_base<string BaseName> {
let BaseOpcode = BaseName in {
let isPredicable = 1 in
def NAME : T_tfrp;
// Predicate
def t : T_tfrp_pred <0, 0>;
def f : T_tfrp_pred <1, 0>;
// Predicate new
def tnew : T_tfrp_pred <0, 1>;
def fnew : T_tfrp_pred <1, 1>;
}
}
let InputType = "imm", isExtendable = 1, isExtentSigned = 1, opExtentBits = 12,
isMoveImm = 1, opExtendable = 2, BaseOpcode = "TFRI", CextOpcode = "TFR",
hasSideEffects = 0, isPredicated = 1, hasNewValue = 1 in
class T_TFRI_Pred<bit PredNot, bit PredNew>
: ALU32_ri<(outs IntRegs:$Rd), (ins PredRegs:$Pu, s12Ext:$s12),
"if ("#!if(PredNot,"!","")#"$Pu"#!if(PredNew,".new","")#") $Rd = #$s12",
[], "", ALU32_2op_tc_1_SLOT0123>, ImmRegRel, PredNewRel {
let isPredicatedFalse = PredNot;
let isPredicatedNew = PredNew;
bits<5> Rd;
bits<2> Pu;
bits<12> s12;
let IClass = 0b0111;
let Inst{27-24} = 0b1110;
let Inst{23} = PredNot;
let Inst{22-21} = Pu;
let Inst{20} = 0b0;
let Inst{19-16,12-5} = s12;
let Inst{13} = PredNew;
let Inst{4-0} = Rd;
}
let isCodeGenOnly = 0 in {
def C2_cmoveit : T_TFRI_Pred<0, 0>;
def C2_cmoveif : T_TFRI_Pred<1, 0>;
def C2_cmovenewit : T_TFRI_Pred<0, 1>;
def C2_cmovenewif : T_TFRI_Pred<1, 1>;
}
let InputType = "imm", isExtendable = 1, isExtentSigned = 1,
CextOpcode = "TFR", BaseOpcode = "TFRI", hasNewValue = 1, opNewValue = 0,
isAsCheapAsAMove = 1 , opExtendable = 1, opExtentBits = 16, isMoveImm = 1,
isPredicated = 0, isPredicable = 1, isReMaterializable = 1,
isCodeGenOnly = 0 in
def A2_tfrsi : ALU32Inst<(outs IntRegs:$Rd), (ins s16Ext:$s16), "$Rd = #$s16",
[(set (i32 IntRegs:$Rd), s16ExtPred:$s16)], "", ALU32_2op_tc_1_SLOT0123>,
ImmRegRel, PredRel {
bits<5> Rd;
bits<16> s16;
let IClass = 0b0111;
let Inst{27-24} = 0b1000;
let Inst{23-22,20-16,13-5} = s16;
let Inst{4-0} = Rd;
}
let isCodeGenOnly = 0 in
defm A2_tfr : tfr_base<"TFR">, ImmRegRel, PredNewRel;
defm A2_tfrp : TFR64_base<"TFR64">, PredNewRel;
// Assembler mapped
let isReMaterializable = 1, isMoveImm = 1, isAsCheapAsAMove = 1 in
def A2_tfrpi : ALU64_rr<(outs DoubleRegs:$dst), (ins s8Imm64:$src1),
"$dst = #$src1",
[(set (i64 DoubleRegs:$dst), s8Imm64Pred:$src1)]>;
// TODO: see if this instruction can be deleted..
let isExtendable = 1, opExtendable = 1, opExtentBits = 6 in
def TFRI64_V4 : ALU64_rr<(outs DoubleRegs:$dst), (ins u6Ext:$src1),
"$dst = #$src1">;
// Transfer control register.
let hasSideEffects = 0 in
def TFCR : CRInst<(outs CRRegs:$dst), (ins IntRegs:$src1),
"$dst = $src1",
[]>;
//===----------------------------------------------------------------------===//
// ALU32/ALU -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// ALU32/PERM +
//===----------------------------------------------------------------------===//
// Scalar mux register immediate.
let hasSideEffects = 0, isExtentSigned = 1, CextOpcode = "MUX",
InputType = "imm", hasNewValue = 1, isExtendable = 1, opExtentBits = 8 in
class T_MUX1 <bit MajOp, dag ins, string AsmStr>
: ALU32Inst <(outs IntRegs:$Rd), ins, AsmStr>, ImmRegRel {
bits<5> Rd;
bits<2> Pu;
bits<8> s8;
bits<5> Rs;
let IClass = 0b0111;
let Inst{27-24} = 0b0011;
let Inst{23} = MajOp;
let Inst{22-21} = Pu;
let Inst{20-16} = Rs;
let Inst{13} = 0b0;
let Inst{12-5} = s8;
let Inst{4-0} = Rd;
}
let opExtendable = 2, isCodeGenOnly = 0 in
def C2_muxri : T_MUX1<0b1, (ins PredRegs:$Pu, s8Ext:$s8, IntRegs:$Rs),
"$Rd = mux($Pu, #$s8, $Rs)">;
let opExtendable = 3, isCodeGenOnly = 0 in
def C2_muxir : T_MUX1<0b0, (ins PredRegs:$Pu, IntRegs:$Rs, s8Ext:$s8),
"$Rd = mux($Pu, $Rs, #$s8)">;
def : Pat<(i32 (select I1:$Pu, s8ExtPred:$s8, I32:$Rs)),
(C2_muxri I1:$Pu, s8ExtPred:$s8, I32:$Rs)>;
def : Pat<(i32 (select I1:$Pu, I32:$Rs, s8ExtPred:$s8)),
(C2_muxir I1:$Pu, I32:$Rs, s8ExtPred:$s8)>;
// C2_muxii: Scalar mux immediates.
let isExtentSigned = 1, hasNewValue = 1, isExtendable = 1,
opExtentBits = 8, opExtendable = 2, isCodeGenOnly = 0 in
def C2_muxii: ALU32Inst <(outs IntRegs:$Rd),
(ins PredRegs:$Pu, s8Ext:$s8, s8Imm:$S8),
"$Rd = mux($Pu, #$s8, #$S8)" ,
[(set (i32 IntRegs:$Rd),
(i32 (select I1:$Pu, s8ExtPred:$s8, s8ImmPred:$S8)))] > {
bits<5> Rd;
bits<2> Pu;
bits<8> s8;
bits<8> S8;
let IClass = 0b0111;
let Inst{27-25} = 0b101;
let Inst{24-23} = Pu;
let Inst{22-16} = S8{7-1};
let Inst{13} = S8{0};
let Inst{12-5} = s8;
let Inst{4-0} = Rd;
}
//===----------------------------------------------------------------------===//
// template class for non-predicated alu32_2op instructions
// - aslh, asrh, sxtb, sxth, zxth
//===----------------------------------------------------------------------===//
let hasNewValue = 1, opNewValue = 0 in
class T_ALU32_2op <string mnemonic, bits<3> minOp> :
ALU32Inst < (outs IntRegs:$Rd), (ins IntRegs:$Rs),
"$Rd = "#mnemonic#"($Rs)", [] > {
bits<5> Rd;
bits<5> Rs;
let IClass = 0b0111;
let Inst{27-24} = 0b0000;
let Inst{23-21} = minOp;
let Inst{13} = 0b0;
let Inst{4-0} = Rd;
let Inst{20-16} = Rs;
}
//===----------------------------------------------------------------------===//
// template class for predicated alu32_2op instructions
// - aslh, asrh, sxtb, sxth, zxtb, zxth
//===----------------------------------------------------------------------===//
let hasSideEffects = 0, validSubTargets = HasV4SubT,
hasNewValue = 1, opNewValue = 0 in
class T_ALU32_2op_Pred <string mnemonic, bits<3> minOp, bit isPredNot,
bit isPredNew > :
ALU32Inst <(outs IntRegs:$Rd), (ins PredRegs:$Pu, IntRegs:$Rs),
!if(isPredNot, "if (!$Pu", "if ($Pu")
#!if(isPredNew, ".new) ",") ")#"$Rd = "#mnemonic#"($Rs)"> {
bits<5> Rd;
bits<2> Pu;
bits<5> Rs;
let IClass = 0b0111;
let Inst{27-24} = 0b0000;
let Inst{23-21} = minOp;
let Inst{13} = 0b1;
let Inst{11} = isPredNot;
let Inst{10} = isPredNew;
let Inst{4-0} = Rd;
let Inst{9-8} = Pu;
let Inst{20-16} = Rs;
}
multiclass ALU32_2op_Pred<string mnemonic, bits<3> minOp, bit PredNot> {
let isPredicatedFalse = PredNot in {
def NAME : T_ALU32_2op_Pred<mnemonic, minOp, PredNot, 0>;
// Predicate new
let isPredicatedNew = 1 in
def NAME#new : T_ALU32_2op_Pred<mnemonic, minOp, PredNot, 1>;
}
}
multiclass ALU32_2op_base<string mnemonic, bits<3> minOp> {
let BaseOpcode = mnemonic in {
let isPredicable = 1, hasSideEffects = 0 in
def A2_#NAME : T_ALU32_2op<mnemonic, minOp>;
let validSubTargets = HasV4SubT, isPredicated = 1, hasSideEffects = 0 in {
defm A4_p#NAME#t : ALU32_2op_Pred<mnemonic, minOp, 0>;
defm A4_p#NAME#f : ALU32_2op_Pred<mnemonic, minOp, 1>;
}
}
}
let isCodeGenOnly = 0 in {
defm aslh : ALU32_2op_base<"aslh", 0b000>, PredNewRel;
defm asrh : ALU32_2op_base<"asrh", 0b001>, PredNewRel;
defm sxtb : ALU32_2op_base<"sxtb", 0b101>, PredNewRel;
defm sxth : ALU32_2op_base<"sxth", 0b111>, PredNewRel;
defm zxth : ALU32_2op_base<"zxth", 0b110>, PredNewRel;
}
// Rd=zxtb(Rs): assembler mapped to Rd=and(Rs,#255).
// Compiler would want to generate 'zxtb' instead of 'and' becuase 'zxtb' has
// predicated forms while 'and' doesn't. Since integrated assembler can't
// handle 'mapped' instructions, we need to encode 'zxtb' same as 'and' where
// immediate operand is set to '255'.
let hasNewValue = 1, opNewValue = 0 in
class T_ZXTB: ALU32Inst < (outs IntRegs:$Rd), (ins IntRegs:$Rs),
"$Rd = zxtb($Rs)", [] > { // Rd = and(Rs,255)
bits<5> Rd;
bits<5> Rs;
bits<10> s10 = 255;
let IClass = 0b0111;
let Inst{27-22} = 0b011000;
let Inst{4-0} = Rd;
let Inst{20-16} = Rs;
let Inst{21} = s10{9};
let Inst{13-5} = s10{8-0};
}
//Rd=zxtb(Rs): assembler mapped to "Rd=and(Rs,#255)
multiclass ZXTB_base <string mnemonic, bits<3> minOp> {
let BaseOpcode = mnemonic in {
let isPredicable = 1, hasSideEffects = 0 in
def A2_#NAME : T_ZXTB;
let validSubTargets = HasV4SubT, isPredicated = 1, hasSideEffects = 0 in {
defm A4_p#NAME#t : ALU32_2op_Pred<mnemonic, minOp, 0>;
defm A4_p#NAME#f : ALU32_2op_Pred<mnemonic, minOp, 1>;
}
}
}
let isCodeGenOnly=0 in
defm zxtb : ZXTB_base<"zxtb",0b100>, PredNewRel;
def: Pat<(shl I32:$src1, (i32 16)), (A2_aslh I32:$src1)>;
def: Pat<(sra I32:$src1, (i32 16)), (A2_asrh I32:$src1)>;
def: Pat<(sext_inreg I32:$src1, i8), (A2_sxtb I32:$src1)>;
def: Pat<(sext_inreg I32:$src1, i16), (A2_sxth I32:$src1)>;
// Mux.
def VMUX_prr64 : ALU64_rr<(outs DoubleRegs:$dst), (ins PredRegs:$src1,
DoubleRegs:$src2,
DoubleRegs:$src3),
"$dst = vmux($src1, $src2, $src3)",
[]>;
//===----------------------------------------------------------------------===//
// ALU32/PERM -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// ALU32/PRED +
//===----------------------------------------------------------------------===//
// SDNode for converting immediate C to C-1.
def DEC_CONST_SIGNED : SDNodeXForm<imm, [{
// Return the byte immediate const-1 as an SDNode.
int32_t imm = N->getSExtValue();
return XformSToSM1Imm(imm);
}]>;
// SDNode for converting immediate C to C-1.
def DEC_CONST_UNSIGNED : SDNodeXForm<imm, [{
// Return the byte immediate const-1 as an SDNode.
uint32_t imm = N->getZExtValue();
return XformUToUM1Imm(imm);
}]>;
def CTLZ_rr : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1),
"$dst = cl0($src1)",
[(set (i32 IntRegs:$dst), (ctlz (i32 IntRegs:$src1)))]>;
def CTTZ_rr : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1),
"$dst = ct0($src1)",
[(set (i32 IntRegs:$dst), (cttz (i32 IntRegs:$src1)))]>;
def CTLZ64_rr : SInst<(outs IntRegs:$dst), (ins DoubleRegs:$src1),
"$dst = cl0($src1)",
[(set (i32 IntRegs:$dst), (i32 (trunc (ctlz (i64 DoubleRegs:$src1)))))]>;
def CTTZ64_rr : SInst<(outs IntRegs:$dst), (ins DoubleRegs:$src1),
"$dst = ct0($src1)",
[(set (i32 IntRegs:$dst), (i32 (trunc (cttz (i64 DoubleRegs:$src1)))))]>;
def TSTBIT_rr : SInst<(outs PredRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
"$dst = tstbit($src1, $src2)",
[(set (i1 PredRegs:$dst),
(setne (and (shl 1, (i32 IntRegs:$src2)), (i32 IntRegs:$src1)), 0))]>;
//===----------------------------------------------------------------------===//
// ALU32/PRED -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// ALU64/ALU +
//===----------------------------------------------------------------------===//// Add.
//===----------------------------------------------------------------------===//
// Template Class
// Add/Subtract halfword
// Rd=add(Rt.L,Rs.[HL])[:sat]
// Rd=sub(Rt.L,Rs.[HL])[:sat]
// Rd=add(Rt.[LH],Rs.[HL])[:sat][:<16]
// Rd=sub(Rt.[LH],Rs.[HL])[:sat][:<16]
//===----------------------------------------------------------------------===//
let hasNewValue = 1, opNewValue = 0 in
class T_XTYPE_ADD_SUB <bits<2> LHbits, bit isSat, bit hasShift, bit isSub>
: ALU64Inst <(outs IntRegs:$Rd), (ins IntRegs:$Rt, IntRegs:$Rs),
"$Rd = "#!if(isSub,"sub","add")#"($Rt."
#!if(hasShift, !if(LHbits{1},"h","l"),"l") #", $Rs."
#!if(hasShift, !if(LHbits{0},"h)","l)"), !if(LHbits{1},"h)","l)"))
#!if(isSat,":sat","")
#!if(hasShift,":<<16",""), [], "", ALU64_tc_1_SLOT23> {
bits<5> Rd;
bits<5> Rt;
bits<5> Rs;
let IClass = 0b1101;
let Inst{27-23} = 0b01010;
let Inst{22} = hasShift;
let Inst{21} = isSub;
let Inst{7} = isSat;
let Inst{6-5} = LHbits;
let Inst{4-0} = Rd;
let Inst{12-8} = Rt;
let Inst{20-16} = Rs;
}
//Rd=sub(Rt.L,Rs.[LH])
let isCodeGenOnly = 0 in {
def A2_subh_l16_ll : T_XTYPE_ADD_SUB <0b00, 0, 0, 1>;
def A2_subh_l16_hl : T_XTYPE_ADD_SUB <0b10, 0, 0, 1>;
}
let isCodeGenOnly = 0 in {
//Rd=add(Rt.L,Rs.[LH])
def A2_addh_l16_ll : T_XTYPE_ADD_SUB <0b00, 0, 0, 0>;
def A2_addh_l16_hl : T_XTYPE_ADD_SUB <0b10, 0, 0, 0>;
}
let Itinerary = ALU64_tc_2_SLOT23, Defs = [USR_OVF], isCodeGenOnly = 0 in {
//Rd=sub(Rt.L,Rs.[LH]):sat
def A2_subh_l16_sat_ll : T_XTYPE_ADD_SUB <0b00, 1, 0, 1>;
def A2_subh_l16_sat_hl : T_XTYPE_ADD_SUB <0b10, 1, 0, 1>;
//Rd=add(Rt.L,Rs.[LH]):sat
def A2_addh_l16_sat_ll : T_XTYPE_ADD_SUB <0b00, 1, 0, 0>;
def A2_addh_l16_sat_hl : T_XTYPE_ADD_SUB <0b10, 1, 0, 0>;
}
//Rd=sub(Rt.[LH],Rs.[LH]):<<16
let isCodeGenOnly = 0 in {
def A2_subh_h16_ll : T_XTYPE_ADD_SUB <0b00, 0, 1, 1>;
def A2_subh_h16_lh : T_XTYPE_ADD_SUB <0b01, 0, 1, 1>;
def A2_subh_h16_hl : T_XTYPE_ADD_SUB <0b10, 0, 1, 1>;
def A2_subh_h16_hh : T_XTYPE_ADD_SUB <0b11, 0, 1, 1>;
}
//Rd=add(Rt.[LH],Rs.[LH]):<<16
let isCodeGenOnly = 0 in {
def A2_addh_h16_ll : T_XTYPE_ADD_SUB <0b00, 0, 1, 0>;
def A2_addh_h16_lh : T_XTYPE_ADD_SUB <0b01, 0, 1, 0>;
def A2_addh_h16_hl : T_XTYPE_ADD_SUB <0b10, 0, 1, 0>;
def A2_addh_h16_hh : T_XTYPE_ADD_SUB <0b11, 0, 1, 0>;
}
let Itinerary = ALU64_tc_2_SLOT23, Defs = [USR_OVF], isCodeGenOnly = 0 in {
//Rd=sub(Rt.[LH],Rs.[LH]):sat:<<16
def A2_subh_h16_sat_ll : T_XTYPE_ADD_SUB <0b00, 1, 1, 1>;
def A2_subh_h16_sat_lh : T_XTYPE_ADD_SUB <0b01, 1, 1, 1>;
def A2_subh_h16_sat_hl : T_XTYPE_ADD_SUB <0b10, 1, 1, 1>;
def A2_subh_h16_sat_hh : T_XTYPE_ADD_SUB <0b11, 1, 1, 1>;
//Rd=add(Rt.[LH],Rs.[LH]):sat:<<16
def A2_addh_h16_sat_ll : T_XTYPE_ADD_SUB <0b00, 1, 1, 0>;
def A2_addh_h16_sat_lh : T_XTYPE_ADD_SUB <0b01, 1, 1, 0>;
def A2_addh_h16_sat_hl : T_XTYPE_ADD_SUB <0b10, 1, 1, 0>;
def A2_addh_h16_sat_hh : T_XTYPE_ADD_SUB <0b11, 1, 1, 0>;
}
// Add halfword.
def: Pat<(sext_inreg (add I32:$src1, I32:$src2), i16),
(A2_addh_l16_ll I32:$src1, I32:$src2)>;
def: Pat<(sra (add (shl I32:$src1, (i32 16)), I32:$src2), (i32 16)),
(A2_addh_l16_hl I32:$src1, I32:$src2)>;
def: Pat<(shl (add I32:$src1, I32:$src2), (i32 16)),
(A2_addh_h16_ll I32:$src1, I32:$src2)>;
// Subtract halfword.
def: Pat<(sext_inreg (sub I32:$src1, I32:$src2), i16),
(A2_subh_l16_ll I32:$src1, I32:$src2)>;
def: Pat<(shl (sub I32:$src1, I32:$src2), (i32 16)),
(A2_subh_h16_ll I32:$src1, I32:$src2)>;
let hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in
def S2_parityp: ALU64Inst<(outs IntRegs:$Rd),
(ins DoubleRegs:$Rs, DoubleRegs:$Rt),
"$Rd = parity($Rs, $Rt)", [], "", ALU64_tc_2_SLOT23> {
bits<5> Rd;
bits<5> Rs;
bits<5> Rt;
let IClass = 0b1101;
let Inst{27-24} = 0b0000;
let Inst{20-16} = Rs;
let Inst{12-8} = Rt;
let Inst{4-0} = Rd;
}
let hasNewValue = 1, opNewValue = 0, hasSideEffects = 0 in
class T_XTYPE_MIN_MAX < bit isMax, bit isUnsigned >
: ALU64Inst < (outs IntRegs:$Rd), (ins IntRegs:$Rt, IntRegs:$Rs),
"$Rd = "#!if(isMax,"max","min")#!if(isUnsigned,"u","")
#"($Rt, $Rs)", [], "", ALU64_tc_2_SLOT23> {
bits<5> Rd;
bits<5> Rt;
bits<5> Rs;
let IClass = 0b1101;
let Inst{27-23} = 0b01011;
let Inst{22-21} = !if(isMax, 0b10, 0b01);
let Inst{7} = isUnsigned;
let Inst{4-0} = Rd;
let Inst{12-8} = !if(isMax, Rs, Rt);
let Inst{20-16} = !if(isMax, Rt, Rs);
}
let isCodeGenOnly = 0 in {
def A2_min : T_XTYPE_MIN_MAX < 0, 0 >;
def A2_minu : T_XTYPE_MIN_MAX < 0, 1 >;
def A2_max : T_XTYPE_MIN_MAX < 1, 0 >;
def A2_maxu : T_XTYPE_MIN_MAX < 1, 1 >;
}
// Here, depending on the operand being selected, we'll either generate a
// min or max instruction.
// Ex:
// (a>b)?a:b --> max(a,b) => Here check performed is '>' and the value selected
// is the larger of two. So, the corresponding HexagonInst is passed in 'Inst'.
// (a>b)?b:a --> min(a,b) => Here check performed is '>' but the smaller value
// is selected and the corresponding HexagonInst is passed in 'SwapInst'.
multiclass T_MinMax_pats <PatFrag Op, RegisterClass RC, ValueType VT,
InstHexagon Inst, InstHexagon SwapInst> {
def: Pat<(select (i1 (Op (VT RC:$src1), (VT RC:$src2))),
(VT RC:$src1), (VT RC:$src2)),
(Inst RC:$src1, RC:$src2)>;
def: Pat<(select (i1 (Op (VT RC:$src1), (VT RC:$src2))),
(VT RC:$src2), (VT RC:$src1)),
(SwapInst RC:$src1, RC:$src2)>;
}
multiclass MinMax_pats <PatFrag Op, InstHexagon Inst, InstHexagon SwapInst> {
defm: T_MinMax_pats<Op, IntRegs, i32, Inst, SwapInst>;
def: Pat<(sext_inreg (i32 (select (i1 (Op (i32 PositiveHalfWord:$src1),
(i32 PositiveHalfWord:$src2))),
(i32 PositiveHalfWord:$src1),
(i32 PositiveHalfWord:$src2))), i16),
(Inst IntRegs:$src1, IntRegs:$src2)>;
def: Pat<(sext_inreg (i32 (select (i1 (Op (i32 PositiveHalfWord:$src1),
(i32 PositiveHalfWord:$src2))),
(i32 PositiveHalfWord:$src2),
(i32 PositiveHalfWord:$src1))), i16),
(SwapInst IntRegs:$src1, IntRegs:$src2)>;
}
let AddedComplexity = 200 in {
defm: MinMax_pats<setge, A2_max, A2_min>;
defm: MinMax_pats<setgt, A2_max, A2_min>;
defm: MinMax_pats<setle, A2_min, A2_max>;
defm: MinMax_pats<setlt, A2_min, A2_max>;
defm: MinMax_pats<setuge, A2_maxu, A2_minu>;
defm: MinMax_pats<setugt, A2_maxu, A2_minu>;
defm: MinMax_pats<setule, A2_minu, A2_maxu>;
defm: MinMax_pats<setult, A2_minu, A2_maxu>;
}
class T_cmp64_rr<string mnemonic, bits<3> MinOp, bit IsComm>
: ALU64_rr<(outs PredRegs:$Pd), (ins DoubleRegs:$Rs, DoubleRegs:$Rt),
"$Pd = "#mnemonic#"($Rs, $Rt)", [], "", ALU64_tc_2early_SLOT23> {
let isCompare = 1;
let isCommutable = IsComm;
let hasSideEffects = 0;
bits<2> Pd;
bits<5> Rs;
bits<5> Rt;
let IClass = 0b1101;
let Inst{27-21} = 0b0010100;
let Inst{20-16} = Rs;
let Inst{12-8} = Rt;
let Inst{7-5} = MinOp;
let Inst{1-0} = Pd;
}
let isCodeGenOnly = 0 in {
def C2_cmpeqp : T_cmp64_rr<"cmp.eq", 0b000, 1>;
def C2_cmpgtp : T_cmp64_rr<"cmp.gt", 0b010, 0>;
def C2_cmpgtup : T_cmp64_rr<"cmp.gtu", 0b100, 0>;
}
class T_cmp64_rr_pat<InstHexagon MI, PatFrag CmpOp>
: Pat<(i1 (CmpOp (i64 DoubleRegs:$Rs), (i64 DoubleRegs:$Rt))),
(i1 (MI DoubleRegs:$Rs, DoubleRegs:$Rt))>;
def: T_cmp64_rr_pat<C2_cmpeqp, seteq>;
def: T_cmp64_rr_pat<C2_cmpgtp, setgt>;
def: T_cmp64_rr_pat<C2_cmpgtup, setugt>;
def: T_cmp64_rr_pat<C2_cmpgtp, RevCmp<setlt>>;
def: T_cmp64_rr_pat<C2_cmpgtup, RevCmp<setult>>;
class T_ALU64_rr<string mnemonic, string suffix, bits<4> RegType,
bits<3> MajOp, bits<3> MinOp, bit OpsRev, bit IsComm,
string Op2Pfx>
: ALU64_rr<(outs DoubleRegs:$Rd), (ins DoubleRegs:$Rs, DoubleRegs:$Rt),
"$Rd = " #mnemonic# "($Rs, " #Op2Pfx# "$Rt)" #suffix, [],
"", ALU64_tc_1_SLOT23> {
let hasSideEffects = 0;
let isCommutable = IsComm;
bits<5> Rs;
bits<5> Rt;
bits<5> Rd;
let IClass = 0b1101;
let Inst{27-24} = RegType;
let Inst{23-21} = MajOp;
let Inst{20-16} = !if (OpsRev,Rt,Rs);
let Inst{12-8} = !if (OpsRev,Rs,Rt);
let Inst{7-5} = MinOp;
let Inst{4-0} = Rd;
}
class T_ALU64_arith<string mnemonic, bits<3> MajOp, bits<3> MinOp, bit IsSat,
bit OpsRev, bit IsComm>
: T_ALU64_rr<mnemonic, !if(IsSat,":sat",""), 0b0011, MajOp, MinOp, OpsRev,
IsComm, "">;
let isCodeGenOnly = 0 in {
def A2_addp : T_ALU64_arith<"add", 0b000, 0b111, 0, 0, 1>;
def A2_subp : T_ALU64_arith<"sub", 0b001, 0b111, 0, 1, 0>;
}
def: Pat<(i64 (add I64:$Rs, I64:$Rt)), (A2_addp I64:$Rs, I64:$Rt)>;
def: Pat<(i64 (sub I64:$Rs, I64:$Rt)), (A2_subp I64:$Rs, I64:$Rt)>;
class T_ALU64_logical<string mnemonic, bits<3> MinOp, bit OpsRev, bit IsComm,
bit IsNeg>
: T_ALU64_rr<mnemonic, "", 0b0011, 0b111, MinOp, OpsRev, IsComm,
!if(IsNeg,"~","")>;
let isCodeGenOnly = 0 in {
def A2_andp : T_ALU64_logical<"and", 0b000, 0, 1, 0>;
def A2_orp : T_ALU64_logical<"or", 0b010, 0, 1, 0>;
def A2_xorp : T_ALU64_logical<"xor", 0b100, 0, 1, 0>;
}
def: Pat<(i64 (and I64:$Rs, I64:$Rt)), (A2_andp I64:$Rs, I64:$Rt)>;
def: Pat<(i64 (or I64:$Rs, I64:$Rt)), (A2_orp I64:$Rs, I64:$Rt)>;
def: Pat<(i64 (xor I64:$Rs, I64:$Rt)), (A2_xorp I64:$Rs, I64:$Rt)>;
//===----------------------------------------------------------------------===//
// ALU64/ALU -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// ALU64/BIT +
//===----------------------------------------------------------------------===//
//
//===----------------------------------------------------------------------===//
// ALU64/BIT -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// ALU64/PERM +
//===----------------------------------------------------------------------===//
//
//===----------------------------------------------------------------------===//
// ALU64/PERM -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// CR +
//===----------------------------------------------------------------------===//
// Logical reductions on predicates.
// Looping instructions.
// Pipelined looping instructions.
// Logical operations on predicates.
let hasSideEffects = 0 in
class T_LOGICAL_1OP<string MnOp, bits<2> OpBits>
: CRInst<(outs PredRegs:$Pd), (ins PredRegs:$Ps),
"$Pd = " # MnOp # "($Ps)", [], "", CR_tc_2early_SLOT23> {
bits<2> Pd;
bits<2> Ps;
let IClass = 0b0110;
let Inst{27-23} = 0b10111;
let Inst{22-21} = OpBits;
let Inst{20} = 0b0;
let Inst{17-16} = Ps;
let Inst{13} = 0b0;
let Inst{1-0} = Pd;
}
let isCodeGenOnly = 0 in {
def C2_any8 : T_LOGICAL_1OP<"any8", 0b00>;
def C2_all8 : T_LOGICAL_1OP<"all8", 0b01>;
def C2_not : T_LOGICAL_1OP<"not", 0b10>;
}
def: Pat<(i1 (not (i1 PredRegs:$Ps))),
(C2_not PredRegs:$Ps)>;
let hasSideEffects = 0 in
class T_LOGICAL_2OP<string MnOp, bits<3> OpBits, bit IsNeg, bit Rev>
: CRInst<(outs PredRegs:$Pd), (ins PredRegs:$Ps, PredRegs:$Pt),
"$Pd = " # MnOp # "($Ps, " # !if (IsNeg,"!","") # "$Pt)",
[], "", CR_tc_2early_SLOT23> {
bits<2> Pd;
bits<2> Ps;
bits<2> Pt;
let IClass = 0b0110;
let Inst{27-24} = 0b1011;
let Inst{23-21} = OpBits;
let Inst{20} = 0b0;
let Inst{17-16} = !if(Rev,Pt,Ps); // Rs and Rt are reversed for some
let Inst{13} = 0b0; // instructions.
let Inst{9-8} = !if(Rev,Ps,Pt);
let Inst{1-0} = Pd;
}
let isCodeGenOnly = 0 in {
def C2_and : T_LOGICAL_2OP<"and", 0b000, 0, 1>;
def C2_or : T_LOGICAL_2OP<"or", 0b001, 0, 1>;
def C2_xor : T_LOGICAL_2OP<"xor", 0b010, 0, 0>;
def C2_andn : T_LOGICAL_2OP<"and", 0b011, 1, 1>;
def C2_orn : T_LOGICAL_2OP<"or", 0b111, 1, 1>;
}
def: Pat<(i1 (and I1:$Ps, I1:$Pt)), (C2_and I1:$Ps, I1:$Pt)>;
def: Pat<(i1 (or I1:$Ps, I1:$Pt)), (C2_or I1:$Ps, I1:$Pt)>;
def: Pat<(i1 (xor I1:$Ps, I1:$Pt)), (C2_xor I1:$Ps, I1:$Pt)>;
def: Pat<(i1 (and I1:$Ps, (not I1:$Pt))), (C2_andn I1:$Ps, I1:$Pt)>;
def: Pat<(i1 (or I1:$Ps, (not I1:$Pt))), (C2_orn I1:$Ps, I1:$Pt)>;
let hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in
def C2_vitpack : SInst<(outs IntRegs:$Rd), (ins PredRegs:$Ps, PredRegs:$Pt),
"$Rd = vitpack($Ps, $Pt)", [], "", S_2op_tc_1_SLOT23> {
bits<5> Rd;
bits<2> Ps;
bits<2> Pt;
let IClass = 0b1000;
let Inst{27-24} = 0b1001;
let Inst{22-21} = 0b00;
let Inst{17-16} = Ps;
let Inst{9-8} = Pt;
let Inst{4-0} = Rd;
}
let hasSideEffects = 0, isCodeGenOnly = 0 in
def C2_mask : SInst<(outs DoubleRegs:$Rd), (ins PredRegs:$Pt),
"$Rd = mask($Pt)", [], "", S_2op_tc_1_SLOT23> {
bits<5> Rd;
bits<2> Pt;
let IClass = 0b1000;
let Inst{27-24} = 0b0110;
let Inst{9-8} = Pt;
let Inst{4-0} = Rd;
}
def VALIGN_rrp : SInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1,
DoubleRegs:$src2,
PredRegs:$src3),
"$dst = valignb($src1, $src2, $src3)",
[]>;
def VSPLICE_rrp : SInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1,
DoubleRegs:$src2,
PredRegs:$src3),
"$dst = vspliceb($src1, $src2, $src3)",
[]>;
// User control register transfer.
//===----------------------------------------------------------------------===//
// CR -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// JR +
//===----------------------------------------------------------------------===//
def retflag : SDNode<"HexagonISD::RET_FLAG", SDTNone,
[SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
def eh_return: SDNode<"HexagonISD::EH_RETURN", SDTNone, [SDNPHasChain]>;
def SDHexagonBR_JT: SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
def HexagonBR_JT: SDNode<"HexagonISD::BR_JT", SDHexagonBR_JT, [SDNPHasChain]>;
class CondStr<string CReg, bit True, bit New> {
string S = "if (" # !if(True,"","!") # CReg # !if(New,".new","") # ") ";
}
class JumpOpcStr<string Mnemonic, bit New, bit Taken> {
string S = Mnemonic # !if(New, !if(Taken,":t",":nt"), "");
}
let isBranch = 1, isBarrier = 1, Defs = [PC], hasSideEffects = 0,
isPredicable = 1,
isExtendable = 1, opExtendable = 0, isExtentSigned = 1,
opExtentBits = 24, opExtentAlign = 2, InputType = "imm" in
class T_JMP<string ExtStr>
: JInst<(outs), (ins brtarget:$dst),
"jump " # ExtStr # "$dst",
[], "", J_tc_2early_SLOT23> {
bits<24> dst;
let IClass = 0b0101;
let Inst{27-25} = 0b100;
let Inst{24-16} = dst{23-15};
let Inst{13-1} = dst{14-2};
}
let isBranch = 1, Defs = [PC], hasSideEffects = 0, isPredicated = 1,
isExtendable = 1, opExtendable = 1, isExtentSigned = 1,
opExtentBits = 17, opExtentAlign = 2, InputType = "imm" in
class T_JMP_c<bit PredNot, bit isPredNew, bit isTak, string ExtStr>
: JInst<(outs), (ins PredRegs:$src, brtarget:$dst),
CondStr<"$src", !if(PredNot,0,1), isPredNew>.S #
JumpOpcStr<"jump", isPredNew, isTak>.S # " " #
ExtStr # "$dst",
[], "", J_tc_2early_SLOT23>, ImmRegRel {
let isTaken = isTak;
let isPredicatedFalse = PredNot;
let isPredicatedNew = isPredNew;
bits<2> src;
bits<17> dst;
let IClass = 0b0101;
let Inst{27-24} = 0b1100;
let Inst{21} = PredNot;
let Inst{12} = !if(isPredNew, isTak, zero);
let Inst{11} = isPredNew;
let Inst{9-8} = src;
let Inst{23-22} = dst{16-15};
let Inst{20-16} = dst{14-10};
let Inst{13} = dst{9};
let Inst{7-1} = dst{8-2};
}
multiclass JMP_Pred<bit PredNot, string ExtStr> {
def NAME : T_JMP_c<PredNot, 0, 0, ExtStr>;
// Predicate new
def NAME#newpt : T_JMP_c<PredNot, 1, 1, ExtStr>; // taken
def NAME#new : T_JMP_c<PredNot, 1, 0, ExtStr>; // not taken
}
multiclass JMP_base<string BaseOp, string ExtStr> {
let BaseOpcode = BaseOp in {
def NAME : T_JMP<ExtStr>;
defm t : JMP_Pred<0, ExtStr>;
defm f : JMP_Pred<1, ExtStr>;
}
}
// Jumps to address stored in a register, JUMPR_MISC
// if ([[!]P[.new]]) jumpr[:t/nt] Rs
let isBranch = 1, isIndirectBranch = 1, isBarrier = 1, Defs = [PC],
isPredicable = 1, hasSideEffects = 0, InputType = "reg" in
class T_JMPr
: JRInst<(outs), (ins IntRegs:$dst),
"jumpr $dst", [], "", J_tc_2early_SLOT2> {
bits<5> dst;
let IClass = 0b0101;
let Inst{27-21} = 0b0010100;
let Inst{20-16} = dst;
}
let isBranch = 1, isIndirectBranch = 1, Defs = [PC], isPredicated = 1,
hasSideEffects = 0, InputType = "reg" in
class T_JMPr_c <bit PredNot, bit isPredNew, bit isTak>
: JRInst <(outs), (ins PredRegs:$src, IntRegs:$dst),
CondStr<"$src", !if(PredNot,0,1), isPredNew>.S #
JumpOpcStr<"jumpr", isPredNew, isTak>.S # " $dst", [],
"", J_tc_2early_SLOT2> {
let isTaken = isTak;
let isPredicatedFalse = PredNot;
let isPredicatedNew = isPredNew;
bits<2> src;
bits<5> dst;
let IClass = 0b0101;
let Inst{27-22} = 0b001101;
let Inst{21} = PredNot;
let Inst{20-16} = dst;
let Inst{12} = !if(isPredNew, isTak, zero);
let Inst{11} = isPredNew;
let Inst{9-8} = src;
}
multiclass JMPR_Pred<bit PredNot> {
def NAME: T_JMPr_c<PredNot, 0, 0>;
// Predicate new
def NAME#newpt : T_JMPr_c<PredNot, 1, 1>; // taken
def NAME#new : T_JMPr_c<PredNot, 1, 0>; // not taken
}
multiclass JMPR_base<string BaseOp> {
let BaseOpcode = BaseOp in {
def NAME : T_JMPr;
defm t : JMPR_Pred<0>;
defm f : JMPR_Pred<1>;
}
}
let isCall = 1, hasSideEffects = 1 in
class JUMPR_MISC_CALLR<bit isPred, bit isPredNot,
dag InputDag = (ins IntRegs:$Rs)>
: JRInst<(outs), InputDag,
!if(isPred, !if(isPredNot, "if (!$Pu) callr $Rs",
"if ($Pu) callr $Rs"),
"callr $Rs"),
[], "", J_tc_2early_SLOT2> {
bits<5> Rs;
bits<2> Pu;
let isPredicated = isPred;
let isPredicatedFalse = isPredNot;
let IClass = 0b0101;
let Inst{27-25} = 0b000;
let Inst{24-23} = !if (isPred, 0b10, 0b01);
let Inst{22} = 0;
let Inst{21} = isPredNot;
let Inst{9-8} = !if (isPred, Pu, 0b00);
let Inst{20-16} = Rs;
}
let Defs = VolatileV3.Regs, isCodeGenOnly = 0 in {
def J2_callrt : JUMPR_MISC_CALLR<1, 0, (ins PredRegs:$Pu, IntRegs:$Rs)>;
def J2_callrf : JUMPR_MISC_CALLR<1, 1, (ins PredRegs:$Pu, IntRegs:$Rs)>;
}
let isTerminator = 1, hasSideEffects = 0, isCodeGenOnly = 0 in {
defm J2_jump : JMP_base<"JMP", "">, PredNewRel;
// Deal with explicit assembly
// - never extened a jump #, always extend a jump ##
let isAsmParserOnly = 1 in {
defm J2_jump_ext : JMP_base<"JMP", "##">;
defm J2_jump_noext : JMP_base<"JMP", "#">;
}
defm J2_jumpr : JMPR_base<"JMPr">, PredNewRel;
let isReturn = 1, isCodeGenOnly = 1 in
defm JMPret : JMPR_base<"JMPret">, PredNewRel;
}
def: Pat<(br bb:$dst),
(J2_jump brtarget:$dst)>;
def: Pat<(retflag),
(JMPret (i32 R31))>;
def: Pat<(brcond (i1 PredRegs:$src1), bb:$offset),
(J2_jumpt PredRegs:$src1, bb:$offset)>;
// A return through builtin_eh_return.
let isReturn = 1, isTerminator = 1, isBarrier = 1, hasSideEffects = 0,
isCodeGenOnly = 1, Defs = [PC], Uses = [R28], isPredicable = 0 in
def EH_RETURN_JMPR : T_JMPr;
def: Pat<(eh_return),
(EH_RETURN_JMPR (i32 R31))>;
def: Pat<(HexagonBR_JT (i32 IntRegs:$dst)),
(J2_jumpr IntRegs:$dst)>;
def: Pat<(brind (i32 IntRegs:$dst)),
(J2_jumpr IntRegs:$dst)>;
//===----------------------------------------------------------------------===//
// JR -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// LD +
//===----------------------------------------------------------------------===//
///
// Load -- MEMri operand
multiclass LD_MEMri_Pbase<string mnemonic, RegisterClass RC,
bit isNot, bit isPredNew> {
let isPredicatedNew = isPredNew in
def NAME : LDInst2<(outs RC:$dst),
(ins PredRegs:$src1, MEMri:$addr),
!if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
") ")#"$dst = "#mnemonic#"($addr)",
[]>;
}
multiclass LD_MEMri_Pred<string mnemonic, RegisterClass RC, bit PredNot> {
let isPredicatedFalse = PredNot in {
defm _c#NAME : LD_MEMri_Pbase<mnemonic, RC, PredNot, 0>;
// Predicate new
defm _cdn#NAME : LD_MEMri_Pbase<mnemonic, RC, PredNot, 1>;
}
}
let isExtendable = 1, hasSideEffects = 0 in
multiclass LD_MEMri<string mnemonic, string CextOp, RegisterClass RC,
bits<5> ImmBits, bits<5> PredImmBits> {
let CextOpcode = CextOp, BaseOpcode = CextOp in {
let opExtendable = 2, isExtentSigned = 1, opExtentBits = ImmBits,
isPredicable = 1 in
def NAME : LDInst2<(outs RC:$dst), (ins MEMri:$addr),
"$dst = "#mnemonic#"($addr)",
[]>;
let opExtendable = 3, isExtentSigned = 0, opExtentBits = PredImmBits,
isPredicated = 1 in {
defm Pt : LD_MEMri_Pred<mnemonic, RC, 0 >;
defm NotPt : LD_MEMri_Pred<mnemonic, RC, 1 >;
}
}
}
let addrMode = BaseImmOffset, isMEMri = "true" in {
let accessSize = ByteAccess in {
defm LDrib: LD_MEMri < "memb", "LDrib", IntRegs, 11, 6>, AddrModeRel;
defm LDriub: LD_MEMri < "memub" , "LDriub", IntRegs, 11, 6>, AddrModeRel;
}
let accessSize = HalfWordAccess in {
defm LDrih: LD_MEMri < "memh", "LDrih", IntRegs, 12, 7>, AddrModeRel;
defm LDriuh: LD_MEMri < "memuh", "LDriuh", IntRegs, 12, 7>, AddrModeRel;
}
let accessSize = WordAccess in
defm LDriw: LD_MEMri < "memw", "LDriw", IntRegs, 13, 8>, AddrModeRel;
let accessSize = DoubleWordAccess in
defm LDrid: LD_MEMri < "memd", "LDrid", DoubleRegs, 14, 9>, AddrModeRel;
}
def : Pat < (i32 (sextloadi8 ADDRriS11_0:$addr)),
(LDrib ADDRriS11_0:$addr) >;
def : Pat < (i32 (zextloadi8 ADDRriS11_0:$addr)),
(LDriub ADDRriS11_0:$addr) >;
def : Pat < (i32 (sextloadi16 ADDRriS11_1:$addr)),
(LDrih ADDRriS11_1:$addr) >;
def : Pat < (i32 (zextloadi16 ADDRriS11_1:$addr)),
(LDriuh ADDRriS11_1:$addr) >;
def : Pat < (i32 (load ADDRriS11_2:$addr)),
(LDriw ADDRriS11_2:$addr) >;
def : Pat < (i64 (load ADDRriS11_3:$addr)),
(LDrid ADDRriS11_3:$addr) >;
// Load - Base with Immediate offset addressing mode
multiclass LD_Idxd_Pbase<string mnemonic, RegisterClass RC, Operand predImmOp,
bit isNot, bit isPredNew> {
let isPredicatedNew = isPredNew in
def NAME : LDInst2<(outs RC:$dst),
(ins PredRegs:$src1, IntRegs:$src2, predImmOp:$src3),
!if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
") ")#"$dst = "#mnemonic#"($src2+#$src3)",
[]>;
}
multiclass LD_Idxd_Pred<string mnemonic, RegisterClass RC, Operand predImmOp,
bit PredNot> {
let isPredicatedFalse = PredNot in {
defm _c#NAME : LD_Idxd_Pbase<mnemonic, RC, predImmOp, PredNot, 0>;
// Predicate new
defm _cdn#NAME : LD_Idxd_Pbase<mnemonic, RC, predImmOp, PredNot, 1>;
}
}
let isExtendable = 1, hasSideEffects = 0 in
multiclass LD_Idxd<string mnemonic, string CextOp, RegisterClass RC,
Operand ImmOp, Operand predImmOp, bits<5> ImmBits,
bits<5> PredImmBits> {
let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed in {
let opExtendable = 2, isExtentSigned = 1, opExtentBits = ImmBits,
isPredicable = 1, AddedComplexity = 20 in
def NAME : LDInst2<(outs RC:$dst), (ins IntRegs:$src1, ImmOp:$offset),
"$dst = "#mnemonic#"($src1+#$offset)",
[]>;
let opExtendable = 3, isExtentSigned = 0, opExtentBits = PredImmBits,
isPredicated = 1 in {
defm Pt : LD_Idxd_Pred<mnemonic, RC, predImmOp, 0 >;
defm NotPt : LD_Idxd_Pred<mnemonic, RC, predImmOp, 1 >;
}
}
}
let addrMode = BaseImmOffset in {
let accessSize = ByteAccess in {
defm LDrib_indexed: LD_Idxd <"memb", "LDrib", IntRegs, s11_0Ext, u6_0Ext,
11, 6>, AddrModeRel;
defm LDriub_indexed: LD_Idxd <"memub" , "LDriub", IntRegs, s11_0Ext, u6_0Ext,
11, 6>, AddrModeRel;
}
let accessSize = HalfWordAccess in {
defm LDrih_indexed: LD_Idxd <"memh", "LDrih", IntRegs, s11_1Ext, u6_1Ext,
12, 7>, AddrModeRel;
defm LDriuh_indexed: LD_Idxd <"memuh", "LDriuh", IntRegs, s11_1Ext, u6_1Ext,
12, 7>, AddrModeRel;
}
let accessSize = WordAccess in
defm LDriw_indexed: LD_Idxd <"memw", "LDriw", IntRegs, s11_2Ext, u6_2Ext,
13, 8>, AddrModeRel;
let accessSize = DoubleWordAccess in
defm LDrid_indexed: LD_Idxd <"memd", "LDrid", DoubleRegs, s11_3Ext, u6_3Ext,
14, 9>, AddrModeRel;
}
let AddedComplexity = 20 in {
def : Pat < (i32 (sextloadi8 (add IntRegs:$src1, s11_0ExtPred:$offset))),
(LDrib_indexed IntRegs:$src1, s11_0ExtPred:$offset) >;
def : Pat < (i32 (zextloadi8 (add IntRegs:$src1, s11_0ExtPred:$offset))),
(LDriub_indexed IntRegs:$src1, s11_0ExtPred:$offset) >;
def : Pat < (i32 (sextloadi16 (add IntRegs:$src1, s11_1ExtPred:$offset))),
(LDrih_indexed IntRegs:$src1, s11_1ExtPred:$offset) >;
def : Pat < (i32 (zextloadi16 (add IntRegs:$src1, s11_1ExtPred:$offset))),
(LDriuh_indexed IntRegs:$src1, s11_1ExtPred:$offset) >;
def : Pat < (i32 (load (add IntRegs:$src1, s11_2ExtPred:$offset))),
(LDriw_indexed IntRegs:$src1, s11_2ExtPred:$offset) >;
def : Pat < (i64 (load (add IntRegs:$src1, s11_3ExtPred:$offset))),
(LDrid_indexed IntRegs:$src1, s11_3ExtPred:$offset) >;
}
//===----------------------------------------------------------------------===//
// Post increment load
//===----------------------------------------------------------------------===//
multiclass LD_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp,
bit isNot, bit isPredNew> {
let isPredicatedNew = isPredNew in
def NAME : LDInst2PI<(outs RC:$dst, IntRegs:$dst2),
(ins PredRegs:$src1, IntRegs:$src2, ImmOp:$offset),
!if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
") ")#"$dst = "#mnemonic#"($src2++#$offset)",
[],
"$src2 = $dst2">;
}
multiclass LD_PostInc_Pred<string mnemonic, RegisterClass RC,
Operand ImmOp, bit PredNot> {
let isPredicatedFalse = PredNot in {
defm _c#NAME : LD_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 0>;
// Predicate new
let Predicates = [HasV4T], validSubTargets = HasV4SubT in
defm _cdn#NAME#_V4 : LD_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 1>;
}
}
multiclass LD_PostInc<string mnemonic, string BaseOp, RegisterClass RC,
Operand ImmOp> {
let BaseOpcode = "POST_"#BaseOp in {
let isPredicable = 1 in
def NAME : LDInst2PI<(outs RC:$dst, IntRegs:$dst2),
(ins IntRegs:$src1, ImmOp:$offset),
"$dst = "#mnemonic#"($src1++#$offset)",
[],
"$src1 = $dst2">;
let isPredicated = 1 in {
defm Pt : LD_PostInc_Pred<mnemonic, RC, ImmOp, 0 >;
defm NotPt : LD_PostInc_Pred<mnemonic, RC, ImmOp, 1 >;
}
}
}
let hasCtrlDep = 1, hasSideEffects = 0, addrMode = PostInc in {
defm POST_LDrib : LD_PostInc<"memb", "LDrib", IntRegs, s4_0Imm>,
PredNewRel;
defm POST_LDriub : LD_PostInc<"memub", "LDriub", IntRegs, s4_0Imm>,
PredNewRel;
defm POST_LDrih : LD_PostInc<"memh", "LDrih", IntRegs, s4_1Imm>,
PredNewRel;
defm POST_LDriuh : LD_PostInc<"memuh", "LDriuh", IntRegs, s4_1Imm>,
PredNewRel;
defm POST_LDriw : LD_PostInc<"memw", "LDriw", IntRegs, s4_2Imm>,
PredNewRel;
defm POST_LDrid : LD_PostInc<"memd", "LDrid", DoubleRegs, s4_3Imm>,
PredNewRel;
}
def : Pat< (i32 (extloadi1 ADDRriS11_0:$addr)),
(i32 (LDrib ADDRriS11_0:$addr)) >;
// Load byte any-extend.
def : Pat < (i32 (extloadi8 ADDRriS11_0:$addr)),
(i32 (LDrib ADDRriS11_0:$addr)) >;
// Indexed load byte any-extend.
let AddedComplexity = 20 in
def : Pat < (i32 (extloadi8 (add IntRegs:$src1, s11_0ImmPred:$offset))),
(i32 (LDrib_indexed IntRegs:$src1, s11_0ImmPred:$offset)) >;
def : Pat < (i32 (extloadi16 ADDRriS11_1:$addr)),
(i32 (LDrih ADDRriS11_1:$addr))>;
let AddedComplexity = 20 in
def : Pat < (i32 (extloadi16 (add IntRegs:$src1, s11_1ImmPred:$offset))),
(i32 (LDrih_indexed IntRegs:$src1, s11_1ImmPred:$offset)) >;
let AddedComplexity = 10 in
def : Pat < (i32 (zextloadi1 ADDRriS11_0:$addr)),
(i32 (LDriub ADDRriS11_0:$addr))>;
let AddedComplexity = 20 in
def : Pat < (i32 (zextloadi1 (add IntRegs:$src1, s11_0ImmPred:$offset))),
(i32 (LDriub_indexed IntRegs:$src1, s11_0ImmPred:$offset))>;
// Load predicate.
let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 13,
isPseudo = 1, Defs = [R10,R11,D5], hasSideEffects = 0 in
def LDriw_pred : LDInst2<(outs PredRegs:$dst),
(ins MEMri:$addr),
"Error; should not emit",
[]>;
// Deallocate stack frame.
let Defs = [R29, R30, R31], Uses = [R29], hasSideEffects = 0 in {
def DEALLOCFRAME : LDInst2<(outs), (ins),
"deallocframe",
[]>;
}
// Load and unpack bytes to halfwords.
//===----------------------------------------------------------------------===//
// LD -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// MTYPE/ALU +
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// MTYPE/ALU -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// MTYPE/COMPLEX +
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// MTYPE/COMPLEX -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// MTYPE/MPYH +
//===----------------------------------------------------------------------===//
// Multiply and use lower result.
// Rd=+mpyi(Rs,#u8)
let isExtendable = 1, opExtendable = 2, isExtentSigned = 0, opExtentBits = 8 in
def MPYI_riu : MInst<(outs IntRegs:$dst), (ins IntRegs:$src1, u8Ext:$src2),
"$dst =+ mpyi($src1, #$src2)",
[(set (i32 IntRegs:$dst), (mul (i32 IntRegs:$src1),
u8ExtPred:$src2))]>;
// Rd=-mpyi(Rs,#u8)
def MPYI_rin : MInst<(outs IntRegs:$dst), (ins IntRegs:$src1, u8Imm:$src2),
"$dst =- mpyi($src1, #$src2)",
[(set (i32 IntRegs:$dst), (ineg (mul (i32 IntRegs:$src1),
u8ImmPred:$src2)))]>;
// Rd=mpyi(Rs,#m9)
// s9 is NOT the same as m9 - but it works.. so far.
// Assembler maps to either Rd=+mpyi(Rs,#u8 or Rd=-mpyi(Rs,#u8)
// depending on the value of m9. See Arch Spec.
let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 9,
CextOpcode = "MPYI", InputType = "imm" in
def MPYI_ri : MInst<(outs IntRegs:$dst), (ins IntRegs:$src1, s9Ext:$src2),
"$dst = mpyi($src1, #$src2)",
[(set (i32 IntRegs:$dst), (mul (i32 IntRegs:$src1),
s9ExtPred:$src2))]>, ImmRegRel;
// Rd=mpyi(Rs,Rt)
let CextOpcode = "MPYI", InputType = "reg" in
def MPYI : MInst<(outs IntRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
"$dst = mpyi($src1, $src2)",
[(set (i32 IntRegs:$dst), (mul (i32 IntRegs:$src1),
(i32 IntRegs:$src2)))]>, ImmRegRel;
// Rx+=mpyi(Rs,#u8)
let isExtendable = 1, opExtendable = 3, isExtentSigned = 0, opExtentBits = 8,
CextOpcode = "MPYI_acc", InputType = "imm" in
def MPYI_acc_ri : MInst_acc<(outs IntRegs:$dst),
(ins IntRegs:$src1, IntRegs:$src2, u8Ext:$src3),
"$dst += mpyi($src2, #$src3)",
[(set (i32 IntRegs:$dst),
(add (mul (i32 IntRegs:$src2), u8ExtPred:$src3),
(i32 IntRegs:$src1)))],
"$src1 = $dst">, ImmRegRel;
// Rx+=mpyi(Rs,Rt)
let CextOpcode = "MPYI_acc", InputType = "reg" in
def MPYI_acc_rr : MInst_acc<(outs IntRegs:$dst),
(ins IntRegs:$src1, IntRegs:$src2, IntRegs:$src3),
"$dst += mpyi($src2, $src3)",
[(set (i32 IntRegs:$dst),
(add (mul (i32 IntRegs:$src2), (i32 IntRegs:$src3)),
(i32 IntRegs:$src1)))],
"$src1 = $dst">, ImmRegRel;
// Rx-=mpyi(Rs,#u8)
let isExtendable = 1, opExtendable = 3, isExtentSigned = 0, opExtentBits = 8 in
def MPYI_sub_ri : MInst_acc<(outs IntRegs:$dst),
(ins IntRegs:$src1, IntRegs:$src2, u8Ext:$src3),
"$dst -= mpyi($src2, #$src3)",
[(set (i32 IntRegs:$dst),
(sub (i32 IntRegs:$src1), (mul (i32 IntRegs:$src2),
u8ExtPred:$src3)))],
"$src1 = $dst">;
// Multiply and use upper result.
// Rd=mpy(Rs,Rt.H):<<1:rnd:sat
// Rd=mpy(Rs,Rt.L):<<1:rnd:sat
// Rd=mpy(Rs,Rt)
def MPY : MInst<(outs IntRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
"$dst = mpy($src1, $src2)",
[(set (i32 IntRegs:$dst), (mulhs (i32 IntRegs:$src1),
(i32 IntRegs:$src2)))]>;
// Rd=mpy(Rs,Rt):rnd
// Rd=mpyu(Rs,Rt)
def MPYU : MInst<(outs IntRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
"$dst = mpyu($src1, $src2)",
[(set (i32 IntRegs:$dst), (mulhu (i32 IntRegs:$src1),
(i32 IntRegs:$src2)))]>;
// Multiply and use full result.
// Rdd=mpyu(Rs,Rt)
def MPYU64 : MInst<(outs DoubleRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
"$dst = mpyu($src1, $src2)",
[(set (i64 DoubleRegs:$dst),
(mul (i64 (anyext (i32 IntRegs:$src1))),
(i64 (anyext (i32 IntRegs:$src2)))))]>;
// Rdd=mpy(Rs,Rt)
def MPY64 : MInst<(outs DoubleRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
"$dst = mpy($src1, $src2)",
[(set (i64 DoubleRegs:$dst),
(mul (i64 (sext (i32 IntRegs:$src1))),
(i64 (sext (i32 IntRegs:$src2)))))]>;
// Multiply and accumulate, use full result.
// Rxx[+-]=mpy(Rs,Rt)
// Rxx+=mpy(Rs,Rt)
def MPY64_acc : MInst_acc<(outs DoubleRegs:$dst),
(ins DoubleRegs:$src1, IntRegs:$src2, IntRegs:$src3),
"$dst += mpy($src2, $src3)",
[(set (i64 DoubleRegs:$dst),
(add (mul (i64 (sext (i32 IntRegs:$src2))),
(i64 (sext (i32 IntRegs:$src3)))),
(i64 DoubleRegs:$src1)))],
"$src1 = $dst">;
// Rxx-=mpy(Rs,Rt)
def MPY64_sub : MInst_acc<(outs DoubleRegs:$dst),
(ins DoubleRegs:$src1, IntRegs:$src2, IntRegs:$src3),
"$dst -= mpy($src2, $src3)",
[(set (i64 DoubleRegs:$dst),
(sub (i64 DoubleRegs:$src1),
(mul (i64 (sext (i32 IntRegs:$src2))),
(i64 (sext (i32 IntRegs:$src3))))))],
"$src1 = $dst">;
// Rxx[+-]=mpyu(Rs,Rt)
// Rxx+=mpyu(Rs,Rt)
def MPYU64_acc : MInst_acc<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1,
IntRegs:$src2, IntRegs:$src3),
"$dst += mpyu($src2, $src3)",
[(set (i64 DoubleRegs:$dst),
(add (mul (i64 (anyext (i32 IntRegs:$src2))),
(i64 (anyext (i32 IntRegs:$src3)))),
(i64 DoubleRegs:$src1)))], "$src1 = $dst">;
// Rxx-=mpyu(Rs,Rt)
def MPYU64_sub : MInst_acc<(outs DoubleRegs:$dst),
(ins DoubleRegs:$src1, IntRegs:$src2, IntRegs:$src3),
"$dst -= mpyu($src2, $src3)",
[(set (i64 DoubleRegs:$dst),
(sub (i64 DoubleRegs:$src1),
(mul (i64 (anyext (i32 IntRegs:$src2))),
(i64 (anyext (i32 IntRegs:$src3))))))],
"$src1 = $dst">;
let InputType = "reg", CextOpcode = "ADD_acc" in
def ADDrr_acc : MInst_acc<(outs IntRegs: $dst), (ins IntRegs:$src1,
IntRegs:$src2, IntRegs:$src3),
"$dst += add($src2, $src3)",
[(set (i32 IntRegs:$dst), (add (add (i32 IntRegs:$src2),
(i32 IntRegs:$src3)),
(i32 IntRegs:$src1)))],
"$src1 = $dst">, ImmRegRel;
let isExtendable = 1, opExtendable = 3, isExtentSigned = 1, opExtentBits = 8,
InputType = "imm", CextOpcode = "ADD_acc" in
def ADDri_acc : MInst_acc<(outs IntRegs: $dst), (ins IntRegs:$src1,
IntRegs:$src2, s8Ext:$src3),
"$dst += add($src2, #$src3)",
[(set (i32 IntRegs:$dst), (add (add (i32 IntRegs:$src2),
s8_16ExtPred:$src3),
(i32 IntRegs:$src1)))],
"$src1 = $dst">, ImmRegRel;
let CextOpcode = "SUB_acc", InputType = "reg" in
def SUBrr_acc : MInst_acc<(outs IntRegs: $dst), (ins IntRegs:$src1,
IntRegs:$src2, IntRegs:$src3),
"$dst -= add($src2, $src3)",
[(set (i32 IntRegs:$dst),
(sub (i32 IntRegs:$src1), (add (i32 IntRegs:$src2),
(i32 IntRegs:$src3))))],
"$src1 = $dst">, ImmRegRel;
let isExtendable = 1, opExtendable = 3, isExtentSigned = 1, opExtentBits = 8,
CextOpcode = "SUB_acc", InputType = "imm" in
def SUBri_acc : MInst_acc<(outs IntRegs: $dst), (ins IntRegs:$src1,
IntRegs:$src2, s8Ext:$src3),
"$dst -= add($src2, #$src3)",
[(set (i32 IntRegs:$dst), (sub (i32 IntRegs:$src1),
(add (i32 IntRegs:$src2),
s8_16ExtPred:$src3)))],
"$src1 = $dst">, ImmRegRel;
//===----------------------------------------------------------------------===//
// MTYPE/MPYH -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// MTYPE/MPYS +
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// MTYPE/MPYS -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// MTYPE/VB +
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// MTYPE/VB -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// MTYPE/VH +
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// MTYPE/VH -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// ST +
//===----------------------------------------------------------------------===//
///
// Store doubleword.
//===----------------------------------------------------------------------===//
// Post increment store
//===----------------------------------------------------------------------===//
multiclass ST_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp,
bit isNot, bit isPredNew> {
let isPredicatedNew = isPredNew in
def NAME : STInst2PI<(outs IntRegs:$dst),
(ins PredRegs:$src1, IntRegs:$src2, ImmOp:$offset, RC:$src3),
!if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
") ")#mnemonic#"($src2++#$offset) = $src3",
[],
"$src2 = $dst">;
}
multiclass ST_PostInc_Pred<string mnemonic, RegisterClass RC,
Operand ImmOp, bit PredNot> {
let isPredicatedFalse = PredNot in {
defm _c#NAME : ST_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 0>;
// Predicate new
let Predicates = [HasV4T], validSubTargets = HasV4SubT in
defm _cdn#NAME#_V4 : ST_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 1>;
}
}
let hasCtrlDep = 1, isNVStorable = 1, hasSideEffects = 0 in
multiclass ST_PostInc<string mnemonic, string BaseOp, RegisterClass RC,
Operand ImmOp> {
let hasCtrlDep = 1, BaseOpcode = "POST_"#BaseOp in {
let isPredicable = 1 in
def NAME : STInst2PI<(outs IntRegs:$dst),
(ins IntRegs:$src1, ImmOp:$offset, RC:$src2),
mnemonic#"($src1++#$offset) = $src2",
[],
"$src1 = $dst">;
let isPredicated = 1 in {
defm Pt : ST_PostInc_Pred<mnemonic, RC, ImmOp, 0 >;
defm NotPt : ST_PostInc_Pred<mnemonic, RC, ImmOp, 1 >;
}
}
}
defm POST_STbri: ST_PostInc <"memb", "STrib", IntRegs, s4_0Imm>, AddrModeRel;
defm POST_SThri: ST_PostInc <"memh", "STrih", IntRegs, s4_1Imm>, AddrModeRel;
defm POST_STwri: ST_PostInc <"memw", "STriw", IntRegs, s4_2Imm>, AddrModeRel;
let isNVStorable = 0 in
defm POST_STdri: ST_PostInc <"memd", "STrid", DoubleRegs, s4_3Imm>, AddrModeRel;
def : Pat<(post_truncsti8 (i32 IntRegs:$src1), IntRegs:$src2,
s4_3ImmPred:$offset),
(POST_STbri IntRegs:$src2, s4_0ImmPred:$offset, IntRegs:$src1)>;
def : Pat<(post_truncsti16 (i32 IntRegs:$src1), IntRegs:$src2,
s4_3ImmPred:$offset),
(POST_SThri IntRegs:$src2, s4_1ImmPred:$offset, IntRegs:$src1)>;
def : Pat<(post_store (i32 IntRegs:$src1), IntRegs:$src2, s4_2ImmPred:$offset),
(POST_STwri IntRegs:$src2, s4_1ImmPred:$offset, IntRegs:$src1)>;
def : Pat<(post_store (i64 DoubleRegs:$src1), IntRegs:$src2,
s4_3ImmPred:$offset),
(POST_STdri IntRegs:$src2, s4_3ImmPred:$offset, DoubleRegs:$src1)>;
//===----------------------------------------------------------------------===//
// multiclass for the store instructions with MEMri operand.
//===----------------------------------------------------------------------===//
multiclass ST_MEMri_Pbase<string mnemonic, RegisterClass RC, bit isNot,
bit isPredNew> {
let isPredicatedNew = isPredNew in
def NAME : STInst2<(outs),
(ins PredRegs:$src1, MEMri:$addr, RC: $src2),
!if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
") ")#mnemonic#"($addr) = $src2",
[]>;
}
multiclass ST_MEMri_Pred<string mnemonic, RegisterClass RC, bit PredNot> {
let isPredicatedFalse = PredNot in {
defm _c#NAME : ST_MEMri_Pbase<mnemonic, RC, PredNot, 0>;
// Predicate new
let validSubTargets = HasV4SubT, Predicates = [HasV4T] in
defm _cdn#NAME#_V4 : ST_MEMri_Pbase<mnemonic, RC, PredNot, 1>;
}
}
let isExtendable = 1, isNVStorable = 1, hasSideEffects = 0 in
multiclass ST_MEMri<string mnemonic, string CextOp, RegisterClass RC,
bits<5> ImmBits, bits<5> PredImmBits> {
let CextOpcode = CextOp, BaseOpcode = CextOp in {
let opExtendable = 1, isExtentSigned = 1, opExtentBits = ImmBits,
isPredicable = 1 in
def NAME : STInst2<(outs),
(ins MEMri:$addr, RC:$src),
mnemonic#"($addr) = $src",
[]>;
let opExtendable = 2, isExtentSigned = 0, opExtentBits = PredImmBits,
isPredicated = 1 in {
defm Pt : ST_MEMri_Pred<mnemonic, RC, 0>;
defm NotPt : ST_MEMri_Pred<mnemonic, RC, 1>;
}
}
}
let addrMode = BaseImmOffset, isMEMri = "true" in {
let accessSize = ByteAccess in
defm STrib: ST_MEMri < "memb", "STrib", IntRegs, 11, 6>, AddrModeRel;
let accessSize = HalfWordAccess in
defm STrih: ST_MEMri < "memh", "STrih", IntRegs, 12, 7>, AddrModeRel;
let accessSize = WordAccess in
defm STriw: ST_MEMri < "memw", "STriw", IntRegs, 13, 8>, AddrModeRel;
let accessSize = DoubleWordAccess, isNVStorable = 0 in
defm STrid: ST_MEMri < "memd", "STrid", DoubleRegs, 14, 9>, AddrModeRel;
}
def : Pat<(truncstorei8 (i32 IntRegs:$src1), ADDRriS11_0:$addr),
(STrib ADDRriS11_0:$addr, (i32 IntRegs:$src1))>;
def : Pat<(truncstorei16 (i32 IntRegs:$src1), ADDRriS11_1:$addr),
(STrih ADDRriS11_1:$addr, (i32 IntRegs:$src1))>;
def : Pat<(store (i32 IntRegs:$src1), ADDRriS11_2:$addr),
(STriw ADDRriS11_2:$addr, (i32 IntRegs:$src1))>;
def : Pat<(store (i64 DoubleRegs:$src1), ADDRriS11_3:$addr),
(STrid ADDRriS11_3:$addr, (i64 DoubleRegs:$src1))>;
//===----------------------------------------------------------------------===//
// multiclass for the store instructions with base+immediate offset
// addressing mode
//===----------------------------------------------------------------------===//
multiclass ST_Idxd_Pbase<string mnemonic, RegisterClass RC, Operand predImmOp,
bit isNot, bit isPredNew> {
let isPredicatedNew = isPredNew in
def NAME : STInst2<(outs),
(ins PredRegs:$src1, IntRegs:$src2, predImmOp:$src3, RC: $src4),
!if(isNot, "if (!$src1", "if ($src1")#!if(isPredNew, ".new) ",
") ")#mnemonic#"($src2+#$src3) = $src4",
[]>;
}
multiclass ST_Idxd_Pred<string mnemonic, RegisterClass RC, Operand predImmOp,
bit PredNot> {
let isPredicatedFalse = PredNot, isPredicated = 1 in {
defm _c#NAME : ST_Idxd_Pbase<mnemonic, RC, predImmOp, PredNot, 0>;
// Predicate new
let validSubTargets = HasV4SubT, Predicates = [HasV4T] in
defm _cdn#NAME#_V4 : ST_Idxd_Pbase<mnemonic, RC, predImmOp, PredNot, 1>;
}
}
let isExtendable = 1, isNVStorable = 1, hasSideEffects = 0 in
multiclass ST_Idxd<string mnemonic, string CextOp, RegisterClass RC,
Operand ImmOp, Operand predImmOp, bits<5> ImmBits,
bits<5> PredImmBits> {
let CextOpcode = CextOp, BaseOpcode = CextOp#_indexed in {
let opExtendable = 1, isExtentSigned = 1, opExtentBits = ImmBits,
isPredicable = 1 in
def NAME : STInst2<(outs),
(ins IntRegs:$src1, ImmOp:$src2, RC:$src3),
mnemonic#"($src1+#$src2) = $src3",
[]>;
let opExtendable = 2, isExtentSigned = 0, opExtentBits = PredImmBits in {
defm Pt : ST_Idxd_Pred<mnemonic, RC, predImmOp, 0>;
defm NotPt : ST_Idxd_Pred<mnemonic, RC, predImmOp, 1>;
}
}
}
let addrMode = BaseImmOffset, InputType = "reg" in {
let accessSize = ByteAccess in
defm STrib_indexed: ST_Idxd < "memb", "STrib", IntRegs, s11_0Ext,
u6_0Ext, 11, 6>, AddrModeRel, ImmRegRel;
let accessSize = HalfWordAccess in
defm STrih_indexed: ST_Idxd < "memh", "STrih", IntRegs, s11_1Ext,
u6_1Ext, 12, 7>, AddrModeRel, ImmRegRel;
let accessSize = WordAccess in
defm STriw_indexed: ST_Idxd < "memw", "STriw", IntRegs, s11_2Ext,
u6_2Ext, 13, 8>, AddrModeRel, ImmRegRel;
let accessSize = DoubleWordAccess, isNVStorable = 0 in
defm STrid_indexed: ST_Idxd < "memd", "STrid", DoubleRegs, s11_3Ext,
u6_3Ext, 14, 9>, AddrModeRel;
}
let AddedComplexity = 10 in {
def : Pat<(truncstorei8 (i32 IntRegs:$src1), (add IntRegs:$src2,
s11_0ExtPred:$offset)),
(STrib_indexed IntRegs:$src2, s11_0ImmPred:$offset,
(i32 IntRegs:$src1))>;
def : Pat<(truncstorei16 (i32 IntRegs:$src1), (add IntRegs:$src2,
s11_1ExtPred:$offset)),
(STrih_indexed IntRegs:$src2, s11_1ImmPred:$offset,
(i32 IntRegs:$src1))>;
def : Pat<(store (i32 IntRegs:$src1), (add IntRegs:$src2,
s11_2ExtPred:$offset)),
(STriw_indexed IntRegs:$src2, s11_2ImmPred:$offset,
(i32 IntRegs:$src1))>;
def : Pat<(store (i64 DoubleRegs:$src1), (add IntRegs:$src2,
s11_3ExtPred:$offset)),
(STrid_indexed IntRegs:$src2, s11_3ImmPred:$offset,
(i64 DoubleRegs:$src1))>;
}
// memh(Rx++#s4:1)=Rt.H
// Store word.
// Store predicate.
let Defs = [R10,R11,D5], hasSideEffects = 0 in
def STriw_pred : STInst2<(outs),
(ins MEMri:$addr, PredRegs:$src1),
"Error; should not emit",
[]>;
// Allocate stack frame.
let Defs = [R29, R30], Uses = [R31, R30], hasSideEffects = 0 in {
def ALLOCFRAME : STInst2<(outs),
(ins i32imm:$amt),
"allocframe(#$amt)",
[]>;
}
//===----------------------------------------------------------------------===//
// ST -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// STYPE/ALU +
//===----------------------------------------------------------------------===//
// Logical NOT.
def NOT_rr64 : ALU64_rr<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1),
"$dst = not($src1)",
[(set (i64 DoubleRegs:$dst), (not (i64 DoubleRegs:$src1)))]>;
let hasSideEffects = 0 in
class T_S2op_1 <string mnemonic, bits<4> RegTyBits, RegisterClass RCOut,
RegisterClass RCIn, bits<2> MajOp, bits<3> MinOp, bit isSat>
: SInst <(outs RCOut:$dst), (ins RCIn:$src),
"$dst = "#mnemonic#"($src)"#!if(isSat, ":sat", ""),
[], "", S_2op_tc_1_SLOT23 > {
bits<5> dst;
bits<5> src;
let IClass = 0b1000;
let Inst{27-24} = RegTyBits;
let Inst{23-22} = MajOp;
let Inst{21} = 0b0;
let Inst{20-16} = src;
let Inst{7-5} = MinOp;
let Inst{4-0} = dst;
}
class T_S2op_1_di <string mnemonic, bits<2> MajOp, bits<3> MinOp>
: T_S2op_1 <mnemonic, 0b0100, DoubleRegs, IntRegs, MajOp, MinOp, 0>;
// Sign extend word to doubleword
let isCodeGenOnly = 0 in
def A2_sxtw : T_S2op_1_di <"sxtw", 0b01, 0b000>;
def: Pat <(i64 (sext I32:$src)), (A2_sxtw I32:$src)>;
//===----------------------------------------------------------------------===//
// STYPE/ALU -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// STYPE/BIT +
//===----------------------------------------------------------------------===//
// clrbit.
def CLRBIT : ALU64_rr<(outs IntRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2),
"$dst = clrbit($src1, #$src2)",
[(set (i32 IntRegs:$dst), (and (i32 IntRegs:$src1),
(not
(shl 1, u5ImmPred:$src2))))]>;
def CLRBIT_31 : ALU64_rr<(outs IntRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2),
"$dst = clrbit($src1, #$src2)",
[]>;
// Map from r0 = and(r1, 2147483647) to r0 = clrbit(r1, #31).
def : Pat <(and (i32 IntRegs:$src1), 2147483647),
(CLRBIT_31 (i32 IntRegs:$src1), 31)>;
// setbit.
def SETBIT : ALU64_rr<(outs IntRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2),
"$dst = setbit($src1, #$src2)",
[(set (i32 IntRegs:$dst), (or (i32 IntRegs:$src1),
(shl 1, u5ImmPred:$src2)))]>;
// Map from r0 = or(r1, -2147483648) to r0 = setbit(r1, #31).
def SETBIT_31 : ALU64_rr<(outs IntRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2),
"$dst = setbit($src1, #$src2)",
[]>;
def : Pat <(or (i32 IntRegs:$src1), -2147483648),
(SETBIT_31 (i32 IntRegs:$src1), 31)>;
// togglebit.
def TOGBIT : ALU64_rr<(outs IntRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2),
"$dst = setbit($src1, #$src2)",
[(set (i32 IntRegs:$dst), (xor (i32 IntRegs:$src1),
(shl 1, u5ImmPred:$src2)))]>;
// Map from r0 = xor(r1, -2147483648) to r0 = togglebit(r1, #31).
def TOGBIT_31 : ALU64_rr<(outs IntRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2),
"$dst = togglebit($src1, #$src2)",
[]>;
def : Pat <(xor (i32 IntRegs:$src1), -2147483648),
(TOGBIT_31 (i32 IntRegs:$src1), 31)>;
//===----------------------------------------------------------------------===//
// STYPE/BIT -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// STYPE/PRED +
//===----------------------------------------------------------------------===//
// Predicate transfer.
let hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in
def C2_tfrpr : SInst<(outs IntRegs:$Rd), (ins PredRegs:$Ps),
"$Rd = $Ps", [], "", S_2op_tc_1_SLOT23> {
bits<5> Rd;
bits<2> Ps;
let IClass = 0b1000;
let Inst{27-24} = 0b1001;
let Inst{22} = 0b1;
let Inst{17-16} = Ps;
let Inst{4-0} = Rd;
}
// Transfer general register to predicate.
let hasSideEffects = 0, isCodeGenOnly = 0 in
def C2_tfrrp: SInst<(outs PredRegs:$Pd), (ins IntRegs:$Rs),
"$Pd = $Rs", [], "", S_2op_tc_2early_SLOT23> {
bits<2> Pd;
bits<5> Rs;
let IClass = 0b1000;
let Inst{27-21} = 0b0101010;
let Inst{20-16} = Rs;
let Inst{1-0} = Pd;
}
let hasSideEffects = 0 in
class T_TEST_BIT_IMM<string MnOp, bits<3> MajOp>
: SInst<(outs PredRegs:$Pd), (ins IntRegs:$Rs, u5Imm:$u5),
"$Pd = "#MnOp#"($Rs, #$u5)",
[], "", S_2op_tc_2early_SLOT23> {
bits<2> Pd;
bits<5> Rs;
bits<5> u5;
let IClass = 0b1000;
let Inst{27-24} = 0b0101;
let Inst{23-21} = MajOp;
let Inst{20-16} = Rs;
let Inst{13} = 0;
let Inst{12-8} = u5;
let Inst{1-0} = Pd;
}
def S2_tstbit_i : T_TEST_BIT_IMM<"tstbit", 0b000>;
let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
def: Pat<(i1 (trunc (i32 IntRegs:$Rs))),
(S2_tstbit_i IntRegs:$Rs, 0)>;
}
//===----------------------------------------------------------------------===//
// STYPE/PRED -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// STYPE/SHIFT +
//===----------------------------------------------------------------------===//
// Shift by immediate.
def ASR_ri : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2),
"$dst = asr($src1, #$src2)",
[(set (i32 IntRegs:$dst), (sra (i32 IntRegs:$src1),
u5ImmPred:$src2))]>;
def ASRd_ri : SInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1, u6Imm:$src2),
"$dst = asr($src1, #$src2)",
[(set (i64 DoubleRegs:$dst), (sra (i64 DoubleRegs:$src1),
u6ImmPred:$src2))]>;
def ASL : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2),
"$dst = asl($src1, #$src2)",
[(set (i32 IntRegs:$dst), (shl (i32 IntRegs:$src1),
u5ImmPred:$src2))]>;
def ASLd_ri : SInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1, u6Imm:$src2),
"$dst = asl($src1, #$src2)",
[(set (i64 DoubleRegs:$dst), (shl (i64 DoubleRegs:$src1),
u6ImmPred:$src2))]>;
def LSR_ri : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2),
"$dst = lsr($src1, #$src2)",
[(set (i32 IntRegs:$dst), (srl (i32 IntRegs:$src1),
u5ImmPred:$src2))]>;
def LSRd_ri : SInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1, u6Imm:$src2),
"$dst = lsr($src1, #$src2)",
[(set (i64 DoubleRegs:$dst), (srl (i64 DoubleRegs:$src1),
u6ImmPred:$src2))]>;
// Shift by immediate and add.
let AddedComplexity = 100 in
def ADDASL : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2,
u3Imm:$src3),
"$dst = addasl($src1, $src2, #$src3)",
[(set (i32 IntRegs:$dst), (add (i32 IntRegs:$src1),
(shl (i32 IntRegs:$src2),
u3ImmPred:$src3)))]>;
// Shift by register.
def ASL_rr : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
"$dst = asl($src1, $src2)",
[(set (i32 IntRegs:$dst), (shl (i32 IntRegs:$src1),
(i32 IntRegs:$src2)))]>;
def ASR_rr : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
"$dst = asr($src1, $src2)",
[(set (i32 IntRegs:$dst), (sra (i32 IntRegs:$src1),
(i32 IntRegs:$src2)))]>;
def LSL_rr : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
"$dst = lsl($src1, $src2)",
[(set (i32 IntRegs:$dst), (shl (i32 IntRegs:$src1),
(i32 IntRegs:$src2)))]>;
def LSR_rr : SInst<(outs IntRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
"$dst = lsr($src1, $src2)",
[(set (i32 IntRegs:$dst), (srl (i32 IntRegs:$src1),
(i32 IntRegs:$src2)))]>;
def ASLd : SInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1, IntRegs:$src2),
"$dst = asl($src1, $src2)",
[(set (i64 DoubleRegs:$dst), (shl (i64 DoubleRegs:$src1),
(i32 IntRegs:$src2)))]>;
def LSLd : SInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1, IntRegs:$src2),
"$dst = lsl($src1, $src2)",
[(set (i64 DoubleRegs:$dst), (shl (i64 DoubleRegs:$src1),
(i32 IntRegs:$src2)))]>;
def ASRd_rr : SInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1,
IntRegs:$src2),
"$dst = asr($src1, $src2)",
[(set (i64 DoubleRegs:$dst), (sra (i64 DoubleRegs:$src1),
(i32 IntRegs:$src2)))]>;
def LSRd_rr : SInst<(outs DoubleRegs:$dst), (ins DoubleRegs:$src1,
IntRegs:$src2),
"$dst = lsr($src1, $src2)",
[(set (i64 DoubleRegs:$dst), (srl (i64 DoubleRegs:$src1),
(i32 IntRegs:$src2)))]>;
//===----------------------------------------------------------------------===//
// STYPE/SHIFT -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// STYPE/VH +
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// STYPE/VH -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// STYPE/VW +
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// STYPE/VW -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// SYSTEM/SUPER +
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// SYSTEM/USER +
//===----------------------------------------------------------------------===//
def SDHexagonBARRIER: SDTypeProfile<0, 0, []>;
def HexagonBARRIER: SDNode<"HexagonISD::BARRIER", SDHexagonBARRIER,
[SDNPHasChain]>;
let hasSideEffects = 1, isSolo = 1 in
def BARRIER : SYSInst<(outs), (ins),
"barrier",
[(HexagonBARRIER)]>;
//===----------------------------------------------------------------------===//
// SYSTEM/SUPER -
//===----------------------------------------------------------------------===//
// TFRI64 - assembly mapped.
let isReMaterializable = 1 in
def TFRI64 : ALU64_rr<(outs DoubleRegs:$dst), (ins s8Imm64:$src1),
"$dst = #$src1",
[(set (i64 DoubleRegs:$dst), s8Imm64Pred:$src1)]>;
let AddedComplexity = 100, isPredicated = 1 in
def TFR_condset_ri : ALU32_rr<(outs IntRegs:$dst),
(ins PredRegs:$src1, IntRegs:$src2, s12Imm:$src3),
"Error; should not emit",
[(set (i32 IntRegs:$dst),
(i32 (select (i1 PredRegs:$src1), (i32 IntRegs:$src2),
s12ImmPred:$src3)))]>;
let AddedComplexity = 100, isPredicated = 1 in
def TFR_condset_ir : ALU32_rr<(outs IntRegs:$dst),
(ins PredRegs:$src1, s12Imm:$src2, IntRegs:$src3),
"Error; should not emit",
[(set (i32 IntRegs:$dst),
(i32 (select (i1 PredRegs:$src1), s12ImmPred:$src2,
(i32 IntRegs:$src3))))]>;
let AddedComplexity = 100, isPredicated = 1 in
def TFR_condset_ii : ALU32_rr<(outs IntRegs:$dst),
(ins PredRegs:$src1, s12Imm:$src2, s12Imm:$src3),
"Error; should not emit",
[(set (i32 IntRegs:$dst),
(i32 (select (i1 PredRegs:$src1), s12ImmPred:$src2,
s12ImmPred:$src3)))]>;
// Generate frameindex addresses.
let isReMaterializable = 1 in
def TFR_FI : ALU32_ri<(outs IntRegs:$dst), (ins FrameIndex:$src1),
"$dst = add($src1)",
[(set (i32 IntRegs:$dst), ADDRri:$src1)]>;
//
// CR - Type.
//
let hasSideEffects = 0, Defs = [SA0, LC0] in {
def LOOP0_i : CRInst<(outs), (ins brtarget:$offset, u10Imm:$src2),
"loop0($offset, #$src2)",
[]>;
}
let hasSideEffects = 0, Defs = [SA0, LC0] in {
def LOOP0_r : CRInst<(outs), (ins brtarget:$offset, IntRegs:$src2),
"loop0($offset, $src2)",
[]>;
}
let isBranch = 1, isTerminator = 1, hasSideEffects = 0,
Defs = [PC, LC0], Uses = [SA0, LC0] in {
def ENDLOOP0 : Endloop<(outs), (ins brtarget:$offset),
":endloop0",
[]>;
}
// Support for generating global address.
// Taken from X86InstrInfo.td.
def SDTHexagonCONST32 : SDTypeProfile<1, 1, [
SDTCisVT<0, i32>,
SDTCisVT<1, i32>,
SDTCisPtrTy<0>]>;
def HexagonCONST32 : SDNode<"HexagonISD::CONST32", SDTHexagonCONST32>;
def HexagonCONST32_GP : SDNode<"HexagonISD::CONST32_GP", SDTHexagonCONST32>;
// HI/LO Instructions
let isReMaterializable = 1, isMoveImm = 1, hasSideEffects = 0 in
def LO : ALU32_ri<(outs IntRegs:$dst), (ins globaladdress:$global),
"$dst.l = #LO($global)",
[]>;
let isReMaterializable = 1, isMoveImm = 1, hasSideEffects = 0 in
def HI : ALU32_ri<(outs IntRegs:$dst), (ins globaladdress:$global),
"$dst.h = #HI($global)",
[]>;
let isReMaterializable = 1, isMoveImm = 1, hasSideEffects = 0 in
def LOi : ALU32_ri<(outs IntRegs:$dst), (ins i32imm:$imm_value),
"$dst.l = #LO($imm_value)",
[]>;
let isReMaterializable = 1, isMoveImm = 1, hasSideEffects = 0 in
def HIi : ALU32_ri<(outs IntRegs:$dst), (ins i32imm:$imm_value),
"$dst.h = #HI($imm_value)",
[]>;
let isReMaterializable = 1, isMoveImm = 1, hasSideEffects = 0 in
def LO_jt : ALU32_ri<(outs IntRegs:$dst), (ins jumptablebase:$jt),
"$dst.l = #LO($jt)",
[]>;
let isReMaterializable = 1, isMoveImm = 1, hasSideEffects = 0 in
def HI_jt : ALU32_ri<(outs IntRegs:$dst), (ins jumptablebase:$jt),
"$dst.h = #HI($jt)",
[]>;
let isReMaterializable = 1, isMoveImm = 1, hasSideEffects = 0 in
def LO_label : ALU32_ri<(outs IntRegs:$dst), (ins bblabel:$label),
"$dst.l = #LO($label)",
[]>;
let isReMaterializable = 1, isMoveImm = 1 , hasSideEffects = 0 in
def HI_label : ALU32_ri<(outs IntRegs:$dst), (ins bblabel:$label),
"$dst.h = #HI($label)",
[]>;
// This pattern is incorrect. When we add small data, we should change
// this pattern to use memw(#foo).
// This is for sdata.
let isMoveImm = 1 in
def CONST32 : LDInst<(outs IntRegs:$dst), (ins globaladdress:$global),
"$dst = CONST32(#$global)",
[(set (i32 IntRegs:$dst),
(load (HexagonCONST32 tglobaltlsaddr:$global)))]>;
// This is for non-sdata.
let isReMaterializable = 1, isMoveImm = 1 in
def CONST32_set : LDInst2<(outs IntRegs:$dst), (ins globaladdress:$global),
"$dst = CONST32(#$global)",
[(set (i32 IntRegs:$dst),
(HexagonCONST32 tglobaladdr:$global))]>;
let isReMaterializable = 1, isMoveImm = 1 in
def CONST32_set_jt : LDInst2<(outs IntRegs:$dst), (ins jumptablebase:$jt),
"$dst = CONST32(#$jt)",
[(set (i32 IntRegs:$dst),
(HexagonCONST32 tjumptable:$jt))]>;
let isReMaterializable = 1, isMoveImm = 1 in
def CONST32GP_set : LDInst2<(outs IntRegs:$dst), (ins globaladdress:$global),
"$dst = CONST32(#$global)",
[(set (i32 IntRegs:$dst),
(HexagonCONST32_GP tglobaladdr:$global))]>;
let isReMaterializable = 1, isMoveImm = 1 in
def CONST32_Int_Real : LDInst2<(outs IntRegs:$dst), (ins i32imm:$global),
"$dst = CONST32(#$global)",
[(set (i32 IntRegs:$dst), imm:$global) ]>;
// Map BlockAddress lowering to CONST32_Int_Real
def : Pat<(HexagonCONST32_GP tblockaddress:$addr),
(CONST32_Int_Real tblockaddress:$addr)>;
let isReMaterializable = 1, isMoveImm = 1 in
def CONST32_Label : LDInst2<(outs IntRegs:$dst), (ins bblabel:$label),
"$dst = CONST32($label)",
[(set (i32 IntRegs:$dst), (HexagonCONST32 bbl:$label))]>;
let isReMaterializable = 1, isMoveImm = 1 in
def CONST64_Int_Real : LDInst2<(outs DoubleRegs:$dst), (ins i64imm:$global),
"$dst = CONST64(#$global)",
[(set (i64 DoubleRegs:$dst), imm:$global) ]>;
def TFR_PdFalse : SInst<(outs PredRegs:$dst), (ins),
"$dst = xor($dst, $dst)",
[(set (i1 PredRegs:$dst), 0)]>;
def MPY_trsext : MInst<(outs IntRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2),
"$dst = mpy($src1, $src2)",
[(set (i32 IntRegs:$dst),
(trunc (i64 (srl (i64 (mul (i64 (sext (i32 IntRegs:$src1))),
(i64 (sext (i32 IntRegs:$src2))))),
(i32 32)))))]>;
// Pseudo instructions.
def SDT_SPCallSeqStart : SDCallSeqStart<[ SDTCisVT<0, i32> ]>;
def SDT_SPCallSeqEnd : SDCallSeqEnd<[ SDTCisVT<0, i32>,
SDTCisVT<1, i32> ]>;
def callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_SPCallSeqEnd,
[SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
def callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_SPCallSeqStart,
[SDNPHasChain, SDNPOutGlue]>;
def SDT_SPCall : SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
def call : SDNode<"HexagonISD::CALL", SDT_SPCall,
[SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
// For tailcalls a HexagonTCRet SDNode has 3 SDNode Properties - a chain,
// Optional Flag and Variable Arguments.
// Its 1 Operand has pointer type.
def HexagonTCRet : SDNode<"HexagonISD::TC_RETURN", SDT_SPCall,
[SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
let Defs = [R29, R30], Uses = [R31, R30, R29] in {
def ADJCALLSTACKDOWN : Pseudo<(outs), (ins i32imm:$amt),
"Should never be emitted",
[(callseq_start timm:$amt)]>;
}
let Defs = [R29, R30, R31], Uses = [R29] in {
def ADJCALLSTACKUP : Pseudo<(outs), (ins i32imm:$amt1, i32imm:$amt2),
"Should never be emitted",
[(callseq_end timm:$amt1, timm:$amt2)]>;
}
// Call subroutine.
let isCall = 1, hasSideEffects = 0,
Defs = [D0, D1, D2, D3, D4, D5, D6, D7, D8, D9, D10,
R22, R23, R28, R31, P0, P1, P2, P3, LC0, LC1, SA0, SA1] in {
def CALL : JInst<(outs), (ins calltarget:$dst),
"call $dst", []>;
}
// Call subroutine indirectly.
let Defs = VolatileV3.Regs, isCodeGenOnly = 0 in
def J2_callr : JUMPR_MISC_CALLR<0, 1>;
// Indirect tail-call.
let isCodeGenOnly = 1, isCall = 1, isReturn = 1 in
def TCRETURNR : T_JMPr;
// Direct tail-calls.
let isCall = 1, isReturn = 1, isBarrier = 1, isPredicable = 0,
isTerminator = 1, isCodeGenOnly = 1 in {
def TCRETURNtg : JInst<(outs), (ins calltarget:$dst), "jump $dst",
[], "", J_tc_2early_SLOT23>;
def TCRETURNtext : JInst<(outs), (ins calltarget:$dst), "jump $dst",
[], "", J_tc_2early_SLOT23>;
}
// Map call instruction.
def : Pat<(call (i32 IntRegs:$dst)),
(J2_callr (i32 IntRegs:$dst))>, Requires<[HasV2TOnly]>;
def : Pat<(call tglobaladdr:$dst),
(CALL tglobaladdr:$dst)>, Requires<[HasV2TOnly]>;
def : Pat<(call texternalsym:$dst),
(CALL texternalsym:$dst)>, Requires<[HasV2TOnly]>;
//Tail calls.
def : Pat<(HexagonTCRet tglobaladdr:$dst),
(TCRETURNtg tglobaladdr:$dst)>;
def : Pat<(HexagonTCRet texternalsym:$dst),
(TCRETURNtext texternalsym:$dst)>;
def : Pat<(HexagonTCRet (i32 IntRegs:$dst)),
(TCRETURNR (i32 IntRegs:$dst))>;
// Atomic load and store support
// 8 bit atomic load
def : Pat<(atomic_load_8 ADDRriS11_0:$src1),
(i32 (LDriub ADDRriS11_0:$src1))>;
def : Pat<(atomic_load_8 (add (i32 IntRegs:$src1), s11_0ImmPred:$offset)),
(i32 (LDriub_indexed (i32 IntRegs:$src1), s11_0ImmPred:$offset))>;
// 16 bit atomic load
def : Pat<(atomic_load_16 ADDRriS11_1:$src1),
(i32 (LDriuh ADDRriS11_1:$src1))>;
def : Pat<(atomic_load_16 (add (i32 IntRegs:$src1), s11_1ImmPred:$offset)),
(i32 (LDriuh_indexed (i32 IntRegs:$src1), s11_1ImmPred:$offset))>;
def : Pat<(atomic_load_32 ADDRriS11_2:$src1),
(i32 (LDriw ADDRriS11_2:$src1))>;
def : Pat<(atomic_load_32 (add (i32 IntRegs:$src1), s11_2ImmPred:$offset)),
(i32 (LDriw_indexed (i32 IntRegs:$src1), s11_2ImmPred:$offset))>;
// 64 bit atomic load
def : Pat<(atomic_load_64 ADDRriS11_3:$src1),
(i64 (LDrid ADDRriS11_3:$src1))>;
def : Pat<(atomic_load_64 (add (i32 IntRegs:$src1), s11_3ImmPred:$offset)),
(i64 (LDrid_indexed (i32 IntRegs:$src1), s11_3ImmPred:$offset))>;
def : Pat<(atomic_store_8 ADDRriS11_0:$src2, (i32 IntRegs:$src1)),
(STrib ADDRriS11_0:$src2, (i32 IntRegs:$src1))>;
def : Pat<(atomic_store_8 (add (i32 IntRegs:$src2), s11_0ImmPred:$offset),
(i32 IntRegs:$src1)),
(STrib_indexed (i32 IntRegs:$src2), s11_0ImmPred:$offset,
(i32 IntRegs:$src1))>;
def : Pat<(atomic_store_16 ADDRriS11_1:$src2, (i32 IntRegs:$src1)),
(STrih ADDRriS11_1:$src2, (i32 IntRegs:$src1))>;
def : Pat<(atomic_store_16 (i32 IntRegs:$src1),
(add (i32 IntRegs:$src2), s11_1ImmPred:$offset)),
(STrih_indexed (i32 IntRegs:$src2), s11_1ImmPred:$offset,
(i32 IntRegs:$src1))>;
def : Pat<(atomic_store_32 ADDRriS11_2:$src2, (i32 IntRegs:$src1)),
(STriw ADDRriS11_2:$src2, (i32 IntRegs:$src1))>;
def : Pat<(atomic_store_32 (add (i32 IntRegs:$src2), s11_2ImmPred:$offset),
(i32 IntRegs:$src1)),
(STriw_indexed (i32 IntRegs:$src2), s11_2ImmPred:$offset,
(i32 IntRegs:$src1))>;
def : Pat<(atomic_store_64 ADDRriS11_3:$src2, (i64 DoubleRegs:$src1)),
(STrid ADDRriS11_3:$src2, (i64 DoubleRegs:$src1))>;
def : Pat<(atomic_store_64 (add (i32 IntRegs:$src2), s11_3ImmPred:$offset),
(i64 DoubleRegs:$src1)),
(STrid_indexed (i32 IntRegs:$src2), s11_3ImmPred:$offset,
(i64 DoubleRegs:$src1))>;
// Map from r0 = and(r1, 65535) to r0 = zxth(r1)
def : Pat <(and (i32 IntRegs:$src1), 65535),
(A2_zxth (i32 IntRegs:$src1))>;
// Map from r0 = and(r1, 255) to r0 = zxtb(r1).
def : Pat <(and (i32 IntRegs:$src1), 255),
(A2_zxtb (i32 IntRegs:$src1))>;
// Map Add(p1, true) to p1 = not(p1).
// Add(p1, false) should never be produced,
// if it does, it got to be mapped to NOOP.
def : Pat <(add (i1 PredRegs:$src1), -1),
(C2_not (i1 PredRegs:$src1))>;
// Map from p0 = pnot(p0); r0 = mux(p0, #i, #j) => r0 = mux(p0, #j, #i).
def : Pat <(select (not (i1 PredRegs:$src1)), s8ImmPred:$src2, s8ImmPred:$src3),
(i32 (TFR_condset_ii (i1 PredRegs:$src1), s8ImmPred:$src3,
s8ImmPred:$src2))>;
// Map from p0 = pnot(p0); r0 = select(p0, #i, r1)
// => r0 = TFR_condset_ri(p0, r1, #i)
def : Pat <(select (not (i1 PredRegs:$src1)), s12ImmPred:$src2,
(i32 IntRegs:$src3)),
(i32 (TFR_condset_ri (i1 PredRegs:$src1), (i32 IntRegs:$src3),
s12ImmPred:$src2))>;
// Map from p0 = pnot(p0); r0 = mux(p0, r1, #i)
// => r0 = TFR_condset_ir(p0, #i, r1)
def : Pat <(select (not (i1 PredRegs:$src1)), IntRegs:$src2, s12ImmPred:$src3),
(i32 (TFR_condset_ir (i1 PredRegs:$src1), s12ImmPred:$src3,
(i32 IntRegs:$src2)))>;
// Map from p0 = pnot(p0); if (p0) jump => if (!p0) jump.
def : Pat <(brcond (not (i1 PredRegs:$src1)), bb:$offset),
(J2_jumpf (i1 PredRegs:$src1), bb:$offset)>;
// Map from p2 = pnot(p2); p1 = and(p0, p2) => p1 = and(p0, !p2).
def : Pat <(and (i1 PredRegs:$src1), (not (i1 PredRegs:$src2))),
(i1 (C2_andn (i1 PredRegs:$src1), (i1 PredRegs:$src2)))>;
let AddedComplexity = 100 in
def : Pat <(i64 (zextloadi1 (HexagonCONST32 tglobaladdr:$global))),
(i64 (A2_combinew (A2_tfrsi 0),
(LDriub_indexed (CONST32_set tglobaladdr:$global), 0)))>,
Requires<[NoV4T]>;
// Map from i1 loads to 32 bits. This assumes that the i1* is byte aligned.
let AddedComplexity = 10 in
def : Pat <(i32 (zextloadi1 ADDRriS11_0:$addr)),
(i32 (A2_and (i32 (LDrib ADDRriS11_0:$addr)), (A2_tfrsi 0x1)))>;
// Map from Rdd = sign_extend_inreg(Rss, i32) -> Rdd = A2_sxtw(Rss.lo).
def : Pat <(i64 (sext_inreg (i64 DoubleRegs:$src1), i32)),
(i64 (A2_sxtw (i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src1), subreg_loreg))))>;
// Map from Rdd = sign_extend_inreg(Rss, i16) -> Rdd = A2_sxtw(SXTH(Rss.lo)).
def : Pat <(i64 (sext_inreg (i64 DoubleRegs:$src1), i16)),
(i64 (A2_sxtw (i32 (A2_sxth (i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src1),
subreg_loreg))))))>;
// Map from Rdd = sign_extend_inreg(Rss, i8) -> Rdd = A2_sxtw(SXTB(Rss.lo)).
def : Pat <(i64 (sext_inreg (i64 DoubleRegs:$src1), i8)),
(i64 (A2_sxtw (i32 (A2_sxtb (i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src1),
subreg_loreg))))))>;
// We want to prevent emitting pnot's as much as possible.
// Map brcond with an unsupported setcc to a J2_jumpf.
def : Pat <(brcond (i1 (setne (i32 IntRegs:$src1), (i32 IntRegs:$src2))),
bb:$offset),
(J2_jumpf (C2_cmpeq (i32 IntRegs:$src1), (i32 IntRegs:$src2)),
bb:$offset)>;
def : Pat <(brcond (i1 (setne (i32 IntRegs:$src1), s10ImmPred:$src2)),
bb:$offset),
(J2_jumpf (C2_cmpeqi (i32 IntRegs:$src1), s10ImmPred:$src2), bb:$offset)>;
def : Pat <(brcond (i1 (setne (i1 PredRegs:$src1), (i1 -1))), bb:$offset),
(J2_jumpf (i1 PredRegs:$src1), bb:$offset)>;
def : Pat <(brcond (i1 (setne (i1 PredRegs:$src1), (i1 0))), bb:$offset),
(J2_jumpt (i1 PredRegs:$src1), bb:$offset)>;
// cmp.lt(Rs, Imm) -> !cmp.ge(Rs, Imm) -> !cmp.gt(Rs, Imm-1)
def : Pat <(brcond (i1 (setlt (i32 IntRegs:$src1), s8ImmPred:$src2)),
bb:$offset),
(J2_jumpf (C2_cmpgti (i32 IntRegs:$src1),
(DEC_CONST_SIGNED s8ImmPred:$src2)), bb:$offset)>;
// cmp.lt(r0, r1) -> cmp.gt(r1, r0)
def : Pat <(brcond (i1 (setlt (i32 IntRegs:$src1), (i32 IntRegs:$src2))),
bb:$offset),
(J2_jumpt (C2_cmpgt (i32 IntRegs:$src2), (i32 IntRegs:$src1)), bb:$offset)>;
def : Pat <(brcond (i1 (setuge (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))),
bb:$offset),
(J2_jumpf (C2_cmpgtup (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1)),
bb:$offset)>;
def : Pat <(brcond (i1 (setule (i32 IntRegs:$src1), (i32 IntRegs:$src2))),
bb:$offset),
(J2_jumpf (C2_cmpgtu (i32 IntRegs:$src1), (i32 IntRegs:$src2)),
bb:$offset)>;
def : Pat <(brcond (i1 (setule (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))),
bb:$offset),
(J2_jumpf (C2_cmpgtup (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2)),
bb:$offset)>;
// Map from a 64-bit select to an emulated 64-bit mux.
// Hexagon does not support 64-bit MUXes; so emulate with combines.
def : Pat <(select (i1 PredRegs:$src1), (i64 DoubleRegs:$src2),
(i64 DoubleRegs:$src3)),
(i64 (A2_combinew (i32 (C2_mux (i1 PredRegs:$src1),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src2),
subreg_hireg)),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src3),
subreg_hireg)))),
(i32 (C2_mux (i1 PredRegs:$src1),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src2),
subreg_loreg)),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src3),
subreg_loreg))))))>;
// Map from a 1-bit select to logical ops.
// From LegalizeDAG.cpp: (B1 ? B2 : B3) <=> (B1 & B2)|(!B1&B3).
def : Pat <(select (i1 PredRegs:$src1), (i1 PredRegs:$src2),
(i1 PredRegs:$src3)),
(C2_or (C2_and (i1 PredRegs:$src1), (i1 PredRegs:$src2)),
(C2_and (C2_not (i1 PredRegs:$src1)), (i1 PredRegs:$src3)))>;
// Map Pd = load(addr) -> Rs = load(addr); Pd = Rs.
def : Pat<(i1 (load ADDRriS11_2:$addr)),
(i1 (C2_tfrrp (i32 (LDrib ADDRriS11_2:$addr))))>;
// Map for truncating from 64 immediates to 32 bit immediates.
def : Pat<(i32 (trunc (i64 DoubleRegs:$src))),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src), subreg_loreg))>;
// Map for truncating from i64 immediates to i1 bit immediates.
def : Pat<(i1 (trunc (i64 DoubleRegs:$src))),
(i1 (C2_tfrrp (i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src),
subreg_loreg))))>;
// Map memb(Rs) = Rdd -> memb(Rs) = Rt.
def : Pat<(truncstorei8 (i64 DoubleRegs:$src), ADDRriS11_0:$addr),
(STrib ADDRriS11_0:$addr, (i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src),
subreg_loreg)))>;
// Map memh(Rs) = Rdd -> memh(Rs) = Rt.
def : Pat<(truncstorei16 (i64 DoubleRegs:$src), ADDRriS11_0:$addr),
(STrih ADDRriS11_0:$addr, (i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src),
subreg_loreg)))>;
// Map memw(Rs) = Rdd -> memw(Rs) = Rt
def : Pat<(truncstorei32 (i64 DoubleRegs:$src), ADDRriS11_0:$addr),
(STriw ADDRriS11_0:$addr, (i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src),
subreg_loreg)))>;
// Map memw(Rs) = Rdd -> memw(Rs) = Rt.
def : Pat<(truncstorei32 (i64 DoubleRegs:$src), ADDRriS11_0:$addr),
(STriw ADDRriS11_0:$addr, (i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src),
subreg_loreg)))>;
// Map from i1 = constant<-1>; memw(addr) = i1 -> r0 = 1; memw(addr) = r0.
def : Pat<(store (i1 -1), ADDRriS11_2:$addr),
(STrib ADDRriS11_2:$addr, (A2_tfrsi 1))>;
// Map from i1 = constant<-1>; store i1 -> r0 = 1; store r0.
def : Pat<(store (i1 -1), ADDRriS11_2:$addr),
(STrib ADDRriS11_2:$addr, (A2_tfrsi 1))>;
// Map from memb(Rs) = Pd -> Rt = mux(Pd, #0, #1); store Rt.
def : Pat<(store (i1 PredRegs:$src1), ADDRriS11_2:$addr),
(STrib ADDRriS11_2:$addr, (i32 (C2_muxii (i1 PredRegs:$src1), 1, 0)) )>;
// Map Rdd = anyext(Rs) -> Rdd = A2_sxtw(Rs).
// Hexagon_TODO: We can probably use combine but that will cost 2 instructions.
// Better way to do this?
def : Pat<(i64 (anyext (i32 IntRegs:$src1))),
(i64 (A2_sxtw (i32 IntRegs:$src1)))>;
// Map cmple -> cmpgt.
// rs <= rt -> !(rs > rt).
def : Pat<(i1 (setle (i32 IntRegs:$src1), s10ExtPred:$src2)),
(i1 (C2_not (C2_cmpgti (i32 IntRegs:$src1), s10ExtPred:$src2)))>;
// rs <= rt -> !(rs > rt).
def : Pat<(i1 (setle (i32 IntRegs:$src1), (i32 IntRegs:$src2))),
(i1 (C2_not (C2_cmpgt (i32 IntRegs:$src1), (i32 IntRegs:$src2))))>;
// Rss <= Rtt -> !(Rss > Rtt).
def : Pat<(i1 (setle (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))),
(i1 (C2_not (C2_cmpgtp (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))))>;
// Map cmpne -> cmpeq.
// Hexagon_TODO: We should improve on this.
// rs != rt -> !(rs == rt).
def : Pat <(i1 (setne (i32 IntRegs:$src1), s10ExtPred:$src2)),
(i1 (C2_not(i1 (C2_cmpeqi (i32 IntRegs:$src1), s10ExtPred:$src2))))>;
// Map cmpne(Rs) -> !cmpeqe(Rs).
// rs != rt -> !(rs == rt).
def : Pat <(i1 (setne (i32 IntRegs:$src1), (i32 IntRegs:$src2))),
(i1 (C2_not (i1 (C2_cmpeq (i32 IntRegs:$src1), (i32 IntRegs:$src2)))))>;
// Convert setne back to xor for hexagon since we compute w/ pred registers.
def : Pat <(i1 (setne (i1 PredRegs:$src1), (i1 PredRegs:$src2))),
(i1 (C2_xor (i1 PredRegs:$src1), (i1 PredRegs:$src2)))>;
// Map cmpne(Rss) -> !cmpew(Rss).
// rs != rt -> !(rs == rt).
def : Pat <(i1 (setne (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))),
(i1 (C2_not (i1 (C2_cmpeqp (i64 DoubleRegs:$src1),
(i64 DoubleRegs:$src2)))))>;
// Map cmpge(Rs, Rt) -> !(cmpgt(Rs, Rt).
// rs >= rt -> !(rt > rs).
def : Pat <(i1 (setge (i32 IntRegs:$src1), (i32 IntRegs:$src2))),
(i1 (C2_not (i1 (C2_cmpgt (i32 IntRegs:$src2), (i32 IntRegs:$src1)))))>;
// cmpge(Rs, Imm) -> cmpgt(Rs, Imm-1)
def : Pat <(i1 (setge (i32 IntRegs:$src1), s8ExtPred:$src2)),
(i1 (C2_cmpgti (i32 IntRegs:$src1), (DEC_CONST_SIGNED s8ExtPred:$src2)))>;
// Map cmpge(Rss, Rtt) -> !cmpgt(Rtt, Rss).
// rss >= rtt -> !(rtt > rss).
def : Pat <(i1 (setge (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))),
(i1 (C2_not (i1 (C2_cmpgtp (i64 DoubleRegs:$src2),
(i64 DoubleRegs:$src1)))))>;
// Map cmplt(Rs, Imm) -> !cmpge(Rs, Imm).
// !cmpge(Rs, Imm) -> !cmpgt(Rs, Imm-1).
// rs < rt -> !(rs >= rt).
def : Pat <(i1 (setlt (i32 IntRegs:$src1), s8ExtPred:$src2)),
(i1 (C2_not (C2_cmpgti (i32 IntRegs:$src1), (DEC_CONST_SIGNED s8ExtPred:$src2))))>;
// Map cmplt(Rs, Rt) -> cmpgt(Rt, Rs).
// rs < rt -> rt > rs.
// We can let assembler map it, or we can do in the compiler itself.
def : Pat <(i1 (setlt (i32 IntRegs:$src1), (i32 IntRegs:$src2))),
(i1 (C2_cmpgt (i32 IntRegs:$src2), (i32 IntRegs:$src1)))>;
// Map cmplt(Rss, Rtt) -> cmpgt(Rtt, Rss).
// rss < rtt -> (rtt > rss).
def : Pat <(i1 (setlt (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))),
(i1 (C2_cmpgtp (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1)))>;
// Map from cmpltu(Rs, Rd) -> cmpgtu(Rd, Rs)
// rs < rt -> rt > rs.
// We can let assembler map it, or we can do in the compiler itself.
def : Pat <(i1 (setult (i32 IntRegs:$src1), (i32 IntRegs:$src2))),
(i1 (C2_cmpgtu (i32 IntRegs:$src2), (i32 IntRegs:$src1)))>;
// Map from cmpltu(Rss, Rdd) -> cmpgtu(Rdd, Rss).
// rs < rt -> rt > rs.
def : Pat <(i1 (setult (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))),
(i1 (C2_cmpgtup (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1)))>;
// Generate cmpgeu(Rs, #0) -> cmpeq(Rs, Rs)
def : Pat <(i1 (setuge (i32 IntRegs:$src1), 0)),
(i1 (C2_cmpeq (i32 IntRegs:$src1), (i32 IntRegs:$src1)))>;
// Generate cmpgeu(Rs, #u8) -> cmpgtu(Rs, #u8 -1)
def : Pat <(i1 (setuge (i32 IntRegs:$src1), u8ExtPred:$src2)),
(i1 (C2_cmpgtui (i32 IntRegs:$src1), (DEC_CONST_UNSIGNED u8ExtPred:$src2)))>;
// Generate cmpgtu(Rs, #u9)
def : Pat <(i1 (setugt (i32 IntRegs:$src1), u9ExtPred:$src2)),
(i1 (C2_cmpgtui (i32 IntRegs:$src1), u9ExtPred:$src2))>;
// Map from Rs >= Rt -> !(Rt > Rs).
// rs >= rt -> !(rt > rs).
def : Pat <(i1 (setuge (i32 IntRegs:$src1), (i32 IntRegs:$src2))),
(i1 (C2_not (C2_cmpgtu (i32 IntRegs:$src2), (i32 IntRegs:$src1))))>;
// Map from Rs >= Rt -> !(Rt > Rs).
// rs >= rt -> !(rt > rs).
def : Pat <(i1 (setuge (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))),
(i1 (C2_not (C2_cmpgtup (i64 DoubleRegs:$src2), (i64 DoubleRegs:$src1))))>;
// Map from cmpleu(Rs, Rt) -> !cmpgtu(Rs, Rt).
// Map from (Rs <= Rt) -> !(Rs > Rt).
def : Pat <(i1 (setule (i32 IntRegs:$src1), (i32 IntRegs:$src2))),
(i1 (C2_not (C2_cmpgtu (i32 IntRegs:$src1), (i32 IntRegs:$src2))))>;
// Map from cmpleu(Rss, Rtt) -> !cmpgtu(Rss, Rtt-1).
// Map from (Rs <= Rt) -> !(Rs > Rt).
def : Pat <(i1 (setule (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))),
(i1 (C2_not (C2_cmpgtup (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2))))>;
// Sign extends.
// i1 -> i32
def : Pat <(i32 (sext (i1 PredRegs:$src1))),
(i32 (C2_muxii (i1 PredRegs:$src1), -1, 0))>;
// i1 -> i64
def : Pat <(i64 (sext (i1 PredRegs:$src1))),
(i64 (A2_combinew (A2_tfrsi -1), (C2_muxii (i1 PredRegs:$src1), -1, 0)))>;
// Convert sign-extended load back to load and sign extend.
// i8 -> i64
def: Pat <(i64 (sextloadi8 ADDRriS11_0:$src1)),
(i64 (A2_sxtw (LDrib ADDRriS11_0:$src1)))>;
// Convert any-extended load back to load and sign extend.
// i8 -> i64
def: Pat <(i64 (extloadi8 ADDRriS11_0:$src1)),
(i64 (A2_sxtw (LDrib ADDRriS11_0:$src1)))>;
// Convert sign-extended load back to load and sign extend.
// i16 -> i64
def: Pat <(i64 (sextloadi16 ADDRriS11_1:$src1)),
(i64 (A2_sxtw (LDrih ADDRriS11_1:$src1)))>;
// Convert sign-extended load back to load and sign extend.
// i32 -> i64
def: Pat <(i64 (sextloadi32 ADDRriS11_2:$src1)),
(i64 (A2_sxtw (LDriw ADDRriS11_2:$src1)))>;
// Zero extends.
// i1 -> i32
def : Pat <(i32 (zext (i1 PredRegs:$src1))),
(i32 (C2_muxii (i1 PredRegs:$src1), 1, 0))>;
// i1 -> i64
def : Pat <(i64 (zext (i1 PredRegs:$src1))),
(i64 (A2_combinew (A2_tfrsi 0), (C2_muxii (i1 PredRegs:$src1), 1, 0)))>,
Requires<[NoV4T]>;
// i32 -> i64
def : Pat <(i64 (zext (i32 IntRegs:$src1))),
(i64 (A2_combinew (A2_tfrsi 0), (i32 IntRegs:$src1)))>,
Requires<[NoV4T]>;
// i8 -> i64
def: Pat <(i64 (zextloadi8 ADDRriS11_0:$src1)),
(i64 (A2_combinew (A2_tfrsi 0), (LDriub ADDRriS11_0:$src1)))>,
Requires<[NoV4T]>;
let AddedComplexity = 20 in
def: Pat <(i64 (zextloadi8 (add (i32 IntRegs:$src1),
s11_0ExtPred:$offset))),
(i64 (A2_combinew (A2_tfrsi 0), (LDriub_indexed IntRegs:$src1,
s11_0ExtPred:$offset)))>,
Requires<[NoV4T]>;
// i1 -> i64
def: Pat <(i64 (zextloadi1 ADDRriS11_0:$src1)),
(i64 (A2_combinew (A2_tfrsi 0), (LDriub ADDRriS11_0:$src1)))>,
Requires<[NoV4T]>;
let AddedComplexity = 20 in
def: Pat <(i64 (zextloadi1 (add (i32 IntRegs:$src1),
s11_0ExtPred:$offset))),
(i64 (A2_combinew (A2_tfrsi 0), (LDriub_indexed IntRegs:$src1,
s11_0ExtPred:$offset)))>,
Requires<[NoV4T]>;
// i16 -> i64
def: Pat <(i64 (zextloadi16 ADDRriS11_1:$src1)),
(i64 (A2_combinew (A2_tfrsi 0), (LDriuh ADDRriS11_1:$src1)))>,
Requires<[NoV4T]>;
let AddedComplexity = 20 in
def: Pat <(i64 (zextloadi16 (add (i32 IntRegs:$src1),
s11_1ExtPred:$offset))),
(i64 (A2_combinew (A2_tfrsi 0), (LDriuh_indexed IntRegs:$src1,
s11_1ExtPred:$offset)))>,
Requires<[NoV4T]>;
// i32 -> i64
def: Pat <(i64 (zextloadi32 ADDRriS11_2:$src1)),
(i64 (A2_combinew (A2_tfrsi 0), (LDriw ADDRriS11_2:$src1)))>,
Requires<[NoV4T]>;
let AddedComplexity = 100 in
def: Pat <(i64 (zextloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))),
(i64 (A2_combinew (A2_tfrsi 0), (LDriw_indexed IntRegs:$src1,
s11_2ExtPred:$offset)))>,
Requires<[NoV4T]>;
let AddedComplexity = 10 in
def: Pat <(i32 (zextloadi1 ADDRriS11_0:$src1)),
(i32 (LDriw ADDRriS11_0:$src1))>;
// Map from Rs = Pd to Pd = mux(Pd, #1, #0)
def : Pat <(i32 (zext (i1 PredRegs:$src1))),
(i32 (C2_muxii (i1 PredRegs:$src1), 1, 0))>;
// Map from Rs = Pd to Pd = mux(Pd, #1, #0)
def : Pat <(i32 (anyext (i1 PredRegs:$src1))),
(i32 (C2_muxii (i1 PredRegs:$src1), 1, 0))>;
// Map from Rss = Pd to Rdd = A2_sxtw (mux(Pd, #1, #0))
def : Pat <(i64 (anyext (i1 PredRegs:$src1))),
(i64 (A2_sxtw (i32 (C2_muxii (i1 PredRegs:$src1), 1, 0))))>;
let AddedComplexity = 100 in
def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh),
(i32 32))),
(i64 (zextloadi32 (i32 (add IntRegs:$src2,
s11_2ExtPred:$offset2)))))),
(i64 (A2_combinew (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg),
(LDriw_indexed IntRegs:$src2,
s11_2ExtPred:$offset2)))>;
def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh),
(i32 32))),
(i64 (zextloadi32 ADDRriS11_2:$srcLow)))),
(i64 (A2_combinew (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg),
(LDriw ADDRriS11_2:$srcLow)))>;
def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh),
(i32 32))),
(i64 (zext (i32 IntRegs:$srcLow))))),
(i64 (A2_combinew (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg),
IntRegs:$srcLow))>;
let AddedComplexity = 100 in
def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh),
(i32 32))),
(i64 (zextloadi32 (i32 (add IntRegs:$src2,
s11_2ExtPred:$offset2)))))),
(i64 (A2_combinew (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg),
(LDriw_indexed IntRegs:$src2,
s11_2ExtPred:$offset2)))>;
def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh),
(i32 32))),
(i64 (zextloadi32 ADDRriS11_2:$srcLow)))),
(i64 (A2_combinew (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg),
(LDriw ADDRriS11_2:$srcLow)))>;
def: Pat<(i64 (or (i64 (shl (i64 DoubleRegs:$srcHigh),
(i32 32))),
(i64 (zext (i32 IntRegs:$srcLow))))),
(i64 (A2_combinew (EXTRACT_SUBREG (i64 DoubleRegs:$srcHigh), subreg_loreg),
IntRegs:$srcLow))>;
// Any extended 64-bit load.
// anyext i32 -> i64
def: Pat <(i64 (extloadi32 ADDRriS11_2:$src1)),
(i64 (A2_combinew (A2_tfrsi 0), (LDriw ADDRriS11_2:$src1)))>,
Requires<[NoV4T]>;
// When there is an offset we should prefer the pattern below over the pattern above.
// The complexity of the above is 13 (gleaned from HexagonGenDAGIsel.inc)
// So this complexity below is comfortably higher to allow for choosing the below.
// If this is not done then we generate addresses such as
// ********************************************
// r1 = add (r0, #4)
// r1 = memw(r1 + #0)
// instead of
// r1 = memw(r0 + #4)
// ********************************************
let AddedComplexity = 100 in
def: Pat <(i64 (extloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))),
(i64 (A2_combinew (A2_tfrsi 0), (LDriw_indexed IntRegs:$src1,
s11_2ExtPred:$offset)))>,
Requires<[NoV4T]>;
// anyext i16 -> i64.
def: Pat <(i64 (extloadi16 ADDRriS11_2:$src1)),
(i64 (A2_combinew (A2_tfrsi 0), (LDrih ADDRriS11_2:$src1)))>,
Requires<[NoV4T]>;
let AddedComplexity = 20 in
def: Pat <(i64 (extloadi16 (add (i32 IntRegs:$src1),
s11_1ExtPred:$offset))),
(i64 (A2_combinew (A2_tfrsi 0), (LDrih_indexed IntRegs:$src1,
s11_1ExtPred:$offset)))>,
Requires<[NoV4T]>;
// Map from Rdd = zxtw(Rs) -> Rdd = combine(0, Rs).
def : Pat<(i64 (zext (i32 IntRegs:$src1))),
(i64 (A2_combinew (A2_tfrsi 0), (i32 IntRegs:$src1)))>,
Requires<[NoV4T]>;
// Multiply 64-bit unsigned and use upper result.
def : Pat <(mulhu (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2)),
(i64
(MPYU64_acc
(i64
(A2_combinew
(A2_tfrsi 0),
(i32
(EXTRACT_SUBREG
(i64
(LSRd_ri
(i64
(MPYU64_acc
(i64
(MPYU64_acc
(i64
(A2_combinew (A2_tfrsi 0),
(i32
(EXTRACT_SUBREG
(i64
(LSRd_ri
(i64
(MPYU64 (i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src1),
subreg_loreg)),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src2),
subreg_loreg)))), 32)),
subreg_loreg)))),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src1), subreg_hireg)),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src2), subreg_loreg)))),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src1), subreg_loreg)),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src2), subreg_hireg)))),
32)), subreg_loreg)))),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src1), subreg_hireg)),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src2), subreg_hireg))))>;
// Multiply 64-bit signed and use upper result.
def : Pat <(mulhs (i64 DoubleRegs:$src1), (i64 DoubleRegs:$src2)),
(i64
(MPY64_acc
(i64
(A2_combinew (A2_tfrsi 0),
(i32
(EXTRACT_SUBREG
(i64
(LSRd_ri
(i64
(MPY64_acc
(i64
(MPY64_acc
(i64
(A2_combinew (A2_tfrsi 0),
(i32
(EXTRACT_SUBREG
(i64
(LSRd_ri
(i64
(MPYU64 (i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src1),
subreg_loreg)),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src2),
subreg_loreg)))), 32)),
subreg_loreg)))),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src1), subreg_hireg)),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src2), subreg_loreg)))),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src1), subreg_loreg)),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src2), subreg_hireg)))),
32)), subreg_loreg)))),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src1), subreg_hireg)),
(i32 (EXTRACT_SUBREG (i64 DoubleRegs:$src2), subreg_hireg))))>;
// Hexagon specific ISD nodes.
//def SDTHexagonADJDYNALLOC : SDTypeProfile<1, 2, [SDTCisSameAs<0, 1>]>;
def SDTHexagonADJDYNALLOC : SDTypeProfile<1, 2,
[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
def Hexagon_ADJDYNALLOC : SDNode<"HexagonISD::ADJDYNALLOC",
SDTHexagonADJDYNALLOC>;
// Needed to tag these instructions for stack layout.
let usesCustomInserter = 1 in
def ADJDYNALLOC : ALU32_ri<(outs IntRegs:$dst), (ins IntRegs:$src1,
s16Imm:$src2),
"$dst = add($src1, #$src2)",
[(set (i32 IntRegs:$dst),
(Hexagon_ADJDYNALLOC (i32 IntRegs:$src1),
s16ImmPred:$src2))]>;
def SDTHexagonARGEXTEND : SDTypeProfile<1, 1, [SDTCisVT<0, i32>]>;
def Hexagon_ARGEXTEND : SDNode<"HexagonISD::ARGEXTEND", SDTHexagonARGEXTEND>;
def ARGEXTEND : ALU32_rr <(outs IntRegs:$dst), (ins IntRegs:$src1),
"$dst = $src1",
[(set (i32 IntRegs:$dst),
(Hexagon_ARGEXTEND (i32 IntRegs:$src1)))]>;
let AddedComplexity = 100 in
def : Pat<(i32 (sext_inreg (Hexagon_ARGEXTEND (i32 IntRegs:$src1)), i16)),
(COPY (i32 IntRegs:$src1))>;
def HexagonWrapperJT: SDNode<"HexagonISD::WrapperJT", SDTIntUnaryOp>;
def : Pat<(HexagonWrapperJT tjumptable:$dst),
(i32 (CONST32_set_jt tjumptable:$dst))>;
// XTYPE/SHIFT
// Multi-class for logical operators :
// Shift by immediate/register and accumulate/logical
multiclass xtype_imm<string OpcStr, SDNode OpNode1, SDNode OpNode2> {
def _ri : SInst_acc<(outs IntRegs:$dst),
(ins IntRegs:$src1, IntRegs:$src2, u5Imm:$src3),
!strconcat("$dst ", !strconcat(OpcStr, "($src2, #$src3)")),
[(set (i32 IntRegs:$dst),
(OpNode2 (i32 IntRegs:$src1),
(OpNode1 (i32 IntRegs:$src2),
u5ImmPred:$src3)))],
"$src1 = $dst">;
def d_ri : SInst_acc<(outs DoubleRegs:$dst),
(ins DoubleRegs:$src1, DoubleRegs:$src2, u6Imm:$src3),
!strconcat("$dst ", !strconcat(OpcStr, "($src2, #$src3)")),
[(set (i64 DoubleRegs:$dst), (OpNode2 (i64 DoubleRegs:$src1),
(OpNode1 (i64 DoubleRegs:$src2), u6ImmPred:$src3)))],
"$src1 = $dst">;
}
// Multi-class for logical operators :
// Shift by register and accumulate/logical (32/64 bits)
multiclass xtype_reg<string OpcStr, SDNode OpNode1, SDNode OpNode2> {
def _rr : SInst_acc<(outs IntRegs:$dst),
(ins IntRegs:$src1, IntRegs:$src2, IntRegs:$src3),
!strconcat("$dst ", !strconcat(OpcStr, "($src2, $src3)")),
[(set (i32 IntRegs:$dst),
(OpNode2 (i32 IntRegs:$src1),
(OpNode1 (i32 IntRegs:$src2),
(i32 IntRegs:$src3))))],
"$src1 = $dst">;
def d_rr : SInst_acc<(outs DoubleRegs:$dst),
(ins DoubleRegs:$src1, DoubleRegs:$src2, IntRegs:$src3),
!strconcat("$dst ", !strconcat(OpcStr, "($src2, $src3)")),
[(set (i64 DoubleRegs:$dst),
(OpNode2 (i64 DoubleRegs:$src1),
(OpNode1 (i64 DoubleRegs:$src2),
(i32 IntRegs:$src3))))],
"$src1 = $dst">;
}
multiclass basic_xtype_imm<string OpcStr, SDNode OpNode> {
let AddedComplexity = 100 in
defm _ADD : xtype_imm< !strconcat("+= ", OpcStr), OpNode, add>;
defm _SUB : xtype_imm< !strconcat("-= ", OpcStr), OpNode, sub>;
defm _AND : xtype_imm< !strconcat("&= ", OpcStr), OpNode, and>;
defm _OR : xtype_imm< !strconcat("|= ", OpcStr), OpNode, or>;
}
multiclass basic_xtype_reg<string OpcStr, SDNode OpNode> {
let AddedComplexity = 100 in
defm _ADD : xtype_reg< !strconcat("+= ", OpcStr), OpNode, add>;
defm _SUB : xtype_reg< !strconcat("-= ", OpcStr), OpNode, sub>;
defm _AND : xtype_reg< !strconcat("&= ", OpcStr), OpNode, and>;
defm _OR : xtype_reg< !strconcat("|= ", OpcStr), OpNode, or>;
}
multiclass xtype_xor_imm<string OpcStr, SDNode OpNode> {
let AddedComplexity = 100 in
defm _XOR : xtype_imm< !strconcat("^= ", OpcStr), OpNode, xor>;
}
defm ASL : basic_xtype_imm<"asl", shl>, basic_xtype_reg<"asl", shl>,
xtype_xor_imm<"asl", shl>;
defm LSR : basic_xtype_imm<"lsr", srl>, basic_xtype_reg<"lsr", srl>,
xtype_xor_imm<"lsr", srl>;
defm ASR : basic_xtype_imm<"asr", sra>, basic_xtype_reg<"asr", sra>;
defm LSL : basic_xtype_reg<"lsl", shl>;
// Change the sign of the immediate for Rd=-mpyi(Rs,#u8)
def : Pat <(mul (i32 IntRegs:$src1), (ineg n8ImmPred:$src2)),
(i32 (MPYI_rin (i32 IntRegs:$src1), u8ImmPred:$src2))>;
//===----------------------------------------------------------------------===//
// V3 Instructions +
//===----------------------------------------------------------------------===//
include "HexagonInstrInfoV3.td"
//===----------------------------------------------------------------------===//
// V3 Instructions -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// V4 Instructions +
//===----------------------------------------------------------------------===//
include "HexagonInstrInfoV4.td"
//===----------------------------------------------------------------------===//
// V4 Instructions -
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// V5 Instructions +
//===----------------------------------------------------------------------===//
include "HexagonInstrInfoV5.td"
//===----------------------------------------------------------------------===//
// V5 Instructions -
//===----------------------------------------------------------------------===//