1 //===- HexagonPatterns.td - Selection Patterns for Hexagon -*- tablegen -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
13 // (3) Extend/truncate/saturate
19 // (9) Arithmetic/bitwise
29 // Guidelines (in no particular order):
30 // 1. Avoid relying on pattern ordering to give preference to one pattern
31 // over another, prefer using AddedComplexity instead. The reason for
32 // this is to avoid unintended conseqeuences (caused by altering the
33 // order) when making changes. The current order of patterns in this
34 // file obviously does play some role, but none of the ordering was
35 // deliberately chosen (other than to create a logical structure of
36 // this file). When making changes, adding AddedComplexity to existing
37 // patterns may be needed.
38 // 2. Maintain the logical structure of the file, try to put new patterns
39 // in designated sections.
40 // 3. Do not use A2_combinew instruction directly, use Combinew fragment
41 // instead. It uses REG_SEQUENCE, which is more amenable to optimizations.
42 // 4. Most selection macros are based on PatFrags. For DAGs that involve
43 // SDNodes, use pf1/pf2 to convert them to PatFrags. Use common frags
44 // whenever possible (see the Definitions section). When adding new
45 // macro, try to make is general to enable reuse across sections.
46 // 5. Compound instructions (e.g. Rx+Rs*Rt) are generated under the condition
47 // that the nested operation has only one use. Having it separated in case
48 // of multiple uses avoids duplication of (processor) work.
49 // 6. The v4 vector instructions (64-bit) are treated as core instructions,
50 // for example, A2_vaddh is in the "arithmetic" section with A2_add.
51 // 7. When adding a pattern for an instruction with a constant-extendable
52 // operand, allow all possible kinds of inputs for the immediate value
53 // (see AnyImm/anyimm and their variants in the Definitions section).
56 // --(0) Definitions -----------------------------------------------------
59 // This complex pattern exists only to create a machine instruction operand
60 // of type "frame index". There doesn't seem to be a way to do that directly
62 def AddrFI: ComplexPattern<i32, 1, "SelectAddrFI", [frameindex], []>;
64 // These complex patterns are not strictly necessary, since global address
65 // folding will happen during DAG combining. For distinguishing between GA
66 // and GP, pat frags with HexagonCONST32 and HexagonCONST32_GP can be used.
67 def AddrGA: ComplexPattern<i32, 1, "SelectAddrGA", [], []>;
68 def AddrGP: ComplexPattern<i32, 1, "SelectAddrGP", [], []>;
69 def AnyImm: ComplexPattern<i32, 1, "SelectAnyImm", [], []>;
70 def AnyInt: ComplexPattern<i32, 1, "SelectAnyInt", [], []>;
72 // Global address or a constant being a multiple of 2^n.
73 def AnyImm0: ComplexPattern<i32, 1, "SelectAnyImm0", [], []>;
74 def AnyImm1: ComplexPattern<i32, 1, "SelectAnyImm1", [], []>;
75 def AnyImm2: ComplexPattern<i32, 1, "SelectAnyImm2", [], []>;
76 def AnyImm3: ComplexPattern<i32, 1, "SelectAnyImm3", [], []>;
80 def V2I1: PatLeaf<(v2i1 PredRegs:$R)>;
81 def V4I1: PatLeaf<(v4i1 PredRegs:$R)>;
82 def V8I1: PatLeaf<(v8i1 PredRegs:$R)>;
83 def V4I8: PatLeaf<(v4i8 IntRegs:$R)>;
84 def V2I16: PatLeaf<(v2i16 IntRegs:$R)>;
86 def V8I8: PatLeaf<(v8i8 DoubleRegs:$R)>;
87 def V4I16: PatLeaf<(v4i16 DoubleRegs:$R)>;
88 def V2I32: PatLeaf<(v2i32 DoubleRegs:$R)>;
91 SDTypeProfile<1, 0, [SDTCisVec<0>]>;
93 SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisVec<1>, SDTCisSameAs<1,2>,
96 def HexagonPTRUE: SDNode<"HexagonISD::PTRUE", SDTVecLeaf>;
97 def HexagonPFALSE: SDNode<"HexagonISD::PFALSE", SDTVecLeaf>;
98 def HexagonVALIGN: SDNode<"HexagonISD::VALIGN", SDTVecVecIntOp>;
99 def HexagonVALIGNADDR: SDNode<"HexagonISD::VALIGNADDR", SDTIntUnaryOp>;
100 def HexagonMULHUS: SDNode<"HexagonISD::MULHUS", SDTIntBinOp>;
103 SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisInt<1>, SDTCisVT<2, OtherVT>]>;
104 def HexagonSSAT: SDNode<"HexagonISD::SSAT", SDTSaturate>;
105 def HexagonUSAT: SDNode<"HexagonISD::USAT", SDTSaturate>;
107 def ptrue: PatFrag<(ops), (HexagonPTRUE)>;
108 def pfalse: PatFrag<(ops), (HexagonPFALSE)>;
109 def pnot: PatFrag<(ops node:$Pu), (xor node:$Pu, ptrue)>;
111 def valign: PatFrag<(ops node:$Vt, node:$Vs, node:$Ru),
112 (HexagonVALIGN node:$Vt, node:$Vs, node:$Ru)>;
113 def valignaddr: PatFrag<(ops node:$Addr), (HexagonVALIGNADDR node:$Addr)>;
115 def ssat: PatFrag<(ops node:$V, node:$Ty), (HexagonSSAT node:$V, node:$Ty)>;
116 def usat: PatFrag<(ops node:$V, node:$Ty), (HexagonUSAT node:$V, node:$Ty)>;
118 // Pattern fragments to extract the low and high subregisters from a
120 def LoReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG $Rs, isub_lo)>;
121 def HiReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG $Rs, isub_hi)>;
123 def IsOrAdd: PatFrag<(ops node:$A, node:$B), (or node:$A, node:$B), [{
124 return isOrEquivalentToAdd(N);
127 def IsPow2_32: PatLeaf<(i32 imm), [{
128 uint32_t V = N->getZExtValue();
129 return isPowerOf2_32(V);
132 def IsPow2_64: PatLeaf<(i64 imm), [{
133 uint64_t V = N->getZExtValue();
134 return isPowerOf2_64(V);
137 def IsNPow2_32: PatLeaf<(i32 imm), [{
138 uint32_t NV = ~N->getZExtValue();
139 return isPowerOf2_32(NV);
142 def IsPow2_64L: PatLeaf<(i64 imm), [{
143 uint64_t V = N->getZExtValue();
144 return isPowerOf2_64(V) && Log2_64(V) < 32;
147 def IsPow2_64H: PatLeaf<(i64 imm), [{
148 uint64_t V = N->getZExtValue();
149 return isPowerOf2_64(V) && Log2_64(V) >= 32;
152 def IsNPow2_64L: PatLeaf<(i64 imm), [{
153 uint64_t NV = ~N->getZExtValue();
154 return isPowerOf2_64(NV) && Log2_64(NV) < 32;
157 def IsNPow2_64H: PatLeaf<(i64 imm), [{
158 uint64_t NV = ~N->getZExtValue();
159 return isPowerOf2_64(NV) && Log2_64(NV) >= 32;
162 class IsULE<int Width, int Arg>: PatLeaf<(i32 imm),
163 "uint64_t V = N->getZExtValue();" #
164 "return isUInt<" # Width # ">(V) && V <= " # Arg # ";"
167 class IsUGT<int Width, int Arg>: PatLeaf<(i32 imm),
168 "uint64_t V = N->getZExtValue();" #
169 "return isUInt<" # Width # ">(V) && V > " # Arg # ";"
172 def SDEC1: SDNodeXForm<imm, [{
173 int32_t V = N->getSExtValue();
174 return CurDAG->getSignedTargetConstant(V-1, SDLoc(N), MVT::i32);
177 def UDEC1: SDNodeXForm<imm, [{
178 uint32_t V = N->getZExtValue();
180 return CurDAG->getTargetConstant(V-1, SDLoc(N), MVT::i32);
183 def UDEC32: SDNodeXForm<imm, [{
184 uint32_t V = N->getZExtValue();
186 return CurDAG->getTargetConstant(V-32, SDLoc(N), MVT::i32);
189 class Subi<int From>: SDNodeXForm<imm,
190 "int32_t V = " # From # " - N->getSExtValue();" #
191 "return CurDAG->getTargetConstant(V, SDLoc(N), MVT::i32);"
194 def Log2_32: SDNodeXForm<imm, [{
195 uint32_t V = N->getZExtValue();
196 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
199 def Log2_64: SDNodeXForm<imm, [{
200 uint64_t V = N->getZExtValue();
201 return CurDAG->getTargetConstant(Log2_64(V), SDLoc(N), MVT::i32);
204 def LogN2_32: SDNodeXForm<imm, [{
205 uint32_t NV = ~N->getZExtValue();
206 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
209 def LogN2_64: SDNodeXForm<imm, [{
210 uint64_t NV = ~N->getZExtValue();
211 return CurDAG->getTargetConstant(Log2_64(NV), SDLoc(N), MVT::i32);
214 def NegImm8: SDNodeXForm<imm, [{
215 int8_t NV = -N->getSExtValue();
216 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
219 def NegImm16: SDNodeXForm<imm, [{
220 int16_t NV = -N->getSExtValue();
221 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
224 def NegImm32: SDNodeXForm<imm, [{
225 int32_t NV = -N->getSExtValue();
226 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
229 def SplatB: SDNodeXForm<imm, [{
230 uint32_t V = N->getZExtValue();
231 assert(isUInt<8>(V) || V >> 8 == 0xFFFFFF);
233 uint32_t S = V << 24 | V << 16 | V << 8 | V;
234 return CurDAG->getTargetConstant(S, SDLoc(N), MVT::i32);
237 def SplatH: SDNodeXForm<imm, [{
238 uint32_t V = N->getZExtValue();
239 assert(isUInt<16>(V) || V >> 16 == 0xFFFF);
241 return CurDAG->getTargetConstant(V << 16 | V, SDLoc(N), MVT::i32);
245 // Helpers for type promotions/contractions.
246 def I1toI32: OutPatFrag<(ops node:$Rs), (C2_muxii (i1 $Rs), 1, 0)>;
247 def I32toI1: OutPatFrag<(ops node:$Rs), (i1 (C2_cmpgtui (i32 $Rs), (i32 0)))>;
248 def ToZext64: OutPatFrag<(ops node:$Rs), (i64 (A4_combineir 0, (i32 $Rs)))>;
249 def ToSext64: OutPatFrag<(ops node:$Rs), (i64 (A2_sxtw (i32 $Rs)))>;
250 def ToAext64: OutPatFrag<(ops node:$Rs),
251 (REG_SEQUENCE DoubleRegs, (i32 (IMPLICIT_DEF)), isub_hi, (i32 $Rs), isub_lo)>;
253 def Combinew: OutPatFrag<(ops node:$Rs, node:$Rt),
254 (REG_SEQUENCE DoubleRegs, $Rs, isub_hi, $Rt, isub_lo)>;
256 def addrga: PatLeaf<(i32 AddrGA:$Addr)>;
257 def addrgp: PatLeaf<(i32 AddrGP:$Addr)>;
258 def anyimm: PatLeaf<(i32 AnyImm:$Imm)>;
259 def anyint: PatLeaf<(i32 AnyInt:$Imm)>;
261 // Global address or an aligned constant.
262 def anyimm0: PatLeaf<(i32 AnyImm0:$Addr)>;
263 def anyimm1: PatLeaf<(i32 AnyImm1:$Addr)>;
264 def anyimm2: PatLeaf<(i32 AnyImm2:$Addr)>;
265 def anyimm3: PatLeaf<(i32 AnyImm3:$Addr)>;
267 def f32ImmPred : PatLeaf<(f32 fpimm:$F)>;
268 def f64ImmPred : PatLeaf<(f64 fpimm:$F)>;
269 def f32zero: PatLeaf<(f32 fpimm:$F), [{
270 return N->isExactlyValue(APFloat::getZero(APFloat::IEEEsingle(), false));
273 // This complex pattern is really only to detect various forms of
274 // sign-extension i32->i64. The selected value will be of type i64
275 // whose low word is the value being extended. The high word is
277 def Usxtw: ComplexPattern<i64, 1, "DetectUseSxtw", [], []>;
279 def Aext64: PatFrag<(ops node:$Rs), (i64 (anyext node:$Rs))>;
280 def Zext64: PatFrag<(ops node:$Rs), (i64 (zext node:$Rs))>;
281 def Sext64: PatLeaf<(i64 Usxtw:$Rs)>;
283 def azext: PatFrags<(ops node:$Rs), [(zext node:$Rs), (anyext node:$Rs)]>;
284 def asext: PatFrags<(ops node:$Rs), [(sext node:$Rs), (anyext node:$Rs)]>;
286 def: Pat<(IsOrAdd (i32 AddrFI:$Rs), s32_0ImmPred:$off),
287 (PS_fi (i32 AddrFI:$Rs), imm:$off)>;
290 // Converters from unary/binary SDNode to PatFrag.
291 class pf1<SDNode Op> : PatFrag<(ops node:$a), (Op node:$a)>;
292 class pf2<SDNode Op> : PatFrag<(ops node:$a, node:$b), (Op node:$a, node:$b)>;
294 class Not2<PatFrag P>
295 : PatFrag<(ops node:$A, node:$B), (P node:$A, (not node:$B))>;
296 class VNot2<PatFrag P, PatFrag Not>
297 : PatFrag<(ops node:$A, node:$B), (P node:$A, (Not node:$B))>;
299 // If there is a constant operand that feeds the and/or instruction,
300 // do not generate the compound instructions.
301 // It is not always profitable, as some times we end up with a transfer.
302 // Check the below example.
303 // ra = #65820; rb = lsr(rb, #8); rc ^= and (rb, ra)
304 // Instead this is preferable.
305 // ra = and (#65820, lsr(ra, #8)); rb = xor(rb, ra)
306 class Su_ni1<PatFrag Op>
307 : PatFrag<Op.Operands, !head(Op.Fragments), [{
309 // Check if Op1 is an immediate operand.
310 SDValue Op1 = N->getOperand(1);
311 return !isa<ConstantSDNode>(Op1);
314 Op.OperandTransform>;
317 : PatFrag<Op.Operands, !head(Op.Fragments), [{ return hasOneUse(N); }],
318 Op.OperandTransform>;
320 // Main selection macros.
322 class OpR_R_pat<InstHexagon MI, PatFrag Op, ValueType ResVT, PatFrag RegPred>
323 : Pat<(ResVT (Op RegPred:$Rs)), (MI RegPred:$Rs)>;
325 class OpR_RI_pat<InstHexagon MI, PatFrag Op, ValueType ResType,
326 PatFrag RegPred, PatFrag ImmPred>
327 : Pat<(ResType (Op RegPred:$Rs, ImmPred:$I)),
328 (MI RegPred:$Rs, imm:$I)>;
330 class OpR_RR_pat<InstHexagon MI, PatFrag Op, ValueType ResType,
331 PatFrag RsPred, PatFrag RtPred = RsPred>
332 : Pat<(ResType (Op RsPred:$Rs, RtPred:$Rt)),
333 (MI RsPred:$Rs, RtPred:$Rt)>;
335 class AccRRI_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op,
336 PatFrag RegPred, PatFrag ImmPred>
337 : Pat<(AccOp RegPred:$Rx, (Op RegPred:$Rs, ImmPred:$I)),
338 (MI RegPred:$Rx, RegPred:$Rs, imm:$I)>;
340 class AccRRR_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op,
341 PatFrag RxPred, PatFrag RsPred, PatFrag RtPred>
342 : Pat<(AccOp RxPred:$Rx, (Op RsPred:$Rs, RtPred:$Rt)),
343 (MI RxPred:$Rx, RsPred:$Rs, RtPred:$Rt)>;
345 multiclass SelMinMax_pats<PatFrag CmpOp, PatFrag Val,
346 InstHexagon InstA, InstHexagon InstB> {
347 def: Pat<(select (i1 (CmpOp Val:$A, Val:$B)), Val:$A, Val:$B),
348 (InstA Val:$A, Val:$B)>;
349 def: Pat<(select (i1 (CmpOp Val:$A, Val:$B)), Val:$B, Val:$A),
350 (InstB Val:$A, Val:$B)>;
353 multiclass MinMax_pats<InstHexagon PickT, InstHexagon PickS,
354 SDPatternOperator Sel, SDPatternOperator CmpOp,
355 ValueType CmpType, PatFrag CmpPred> {
356 def: Pat<(Sel (CmpType (CmpOp CmpPred:$Vs, CmpPred:$Vt)),
357 CmpPred:$Vt, CmpPred:$Vs),
358 (PickT CmpPred:$Vs, CmpPred:$Vt)>;
359 def: Pat<(Sel (CmpType (CmpOp CmpPred:$Vs, CmpPred:$Vt)),
360 CmpPred:$Vs, CmpPred:$Vt),
361 (PickS CmpPred:$Vs, CmpPred:$Vt)>;
364 // Bitcasts between same-size vector types are no-ops, except for the
365 // actual type change.
366 multiclass NopCast_pat<ValueType Ty1, ValueType Ty2, RegisterClass RC> {
367 def: Pat<(Ty1 (bitconvert (Ty2 RC:$Val))), (Ty1 RC:$Val)>;
368 def: Pat<(Ty2 (bitconvert (Ty1 RC:$Val))), (Ty2 RC:$Val)>;
371 // Frags for commonly used SDNodes.
372 def Add: pf2<add>; def And: pf2<and>; def Sra: pf2<sra>;
373 def Sub: pf2<sub>; def Or: pf2<or>; def Srl: pf2<srl>;
374 def Mul: pf2<mul>; def Xor: pf2<xor>; def Shl: pf2<shl>;
376 def Smin: pf2<smin>; def Smax: pf2<smax>;
377 def Umin: pf2<umin>; def Umax: pf2<umax>;
381 def Fptosi: pf1<fp_to_sint>;
382 def Fptoui: pf1<fp_to_uint>;
383 def Sitofp: pf1<sint_to_fp>;
384 def Uitofp: pf1<uint_to_fp>;
387 // --(1) Immediate -------------------------------------------------------
390 def Imm64Lo: SDNodeXForm<imm, [{
391 return CurDAG->getSignedTargetConstant(int32_t(N->getSExtValue()),
394 def Imm64Hi: SDNodeXForm<imm, [{
395 return CurDAG->getSignedTargetConstant(int32_t(N->getSExtValue()>>32),
400 def SDTHexagonCONST32
401 : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisPtrTy<0>]>;
403 def HexagonJT: SDNode<"HexagonISD::JT", SDTIntUnaryOp>;
404 def HexagonCP: SDNode<"HexagonISD::CP", SDTIntUnaryOp>;
405 def HexagonCONST32: SDNode<"HexagonISD::CONST32", SDTHexagonCONST32>;
406 def HexagonCONST32_GP: SDNode<"HexagonISD::CONST32_GP", SDTHexagonCONST32>;
408 def TruncI64ToI32: SDNodeXForm<imm, [{
409 return CurDAG->getSignedTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i32);
412 def: Pat<(s32_0ImmPred:$s16), (A2_tfrsi imm:$s16)>;
413 def: Pat<(s8_0Imm64Pred:$s8), (A2_tfrpi (TruncI64ToI32 $s8))>;
415 def: Pat<(HexagonCONST32 tglobaltlsaddr:$A), (A2_tfrsi imm:$A)>;
416 def: Pat<(HexagonCONST32 bbl:$A), (A2_tfrsi imm:$A)>;
417 def: Pat<(HexagonCONST32 tglobaladdr:$A), (A2_tfrsi imm:$A)>;
418 def: Pat<(HexagonCONST32_GP tblockaddress:$A), (A2_tfrsi imm:$A)>;
419 def: Pat<(HexagonCONST32_GP tglobaladdr:$A), (A2_tfrsi imm:$A)>;
420 def: Pat<(HexagonJT tjumptable:$A), (A2_tfrsi imm:$A)>;
421 def: Pat<(HexagonCP tconstpool:$A), (A2_tfrsi imm:$A)>;
422 // The HVX load patterns also match CP directly. Make sure that if
423 // the selection of this opcode changes, it's updated in all places.
425 def: Pat<(i1 0), (PS_false)>;
426 def: Pat<(i1 1), (PS_true)>;
427 def: Pat<(i64 imm:$v), (CONST64 imm:$v)>,
428 Requires<[UseSmallData,NotOptTinyCore]>;
429 def: Pat<(i64 imm:$v),
430 (Combinew (A2_tfrsi (Imm64Hi $v)), (A2_tfrsi (Imm64Lo $v)))>;
432 def ftoi : SDNodeXForm<fpimm, [{
433 APInt I = N->getValueAPF().bitcastToAPInt();
434 return CurDAG->getTargetConstant(I.getZExtValue(), SDLoc(N),
435 MVT::getIntegerVT(I.getBitWidth()));
438 def: Pat<(f32ImmPred:$f), (A2_tfrsi (ftoi $f))>;
439 def: Pat<(f64ImmPred:$f), (CONST64 (ftoi $f))>;
441 def ToI32: OutPatFrag<(ops node:$V), (A2_tfrsi $V)>;
443 // --(2) Type cast -------------------------------------------------------
446 def: OpR_R_pat<F2_conv_sf2df, pf1<fpextend>, f64, F32>;
447 def: OpR_R_pat<F2_conv_df2sf, pf1<fpround>, f32, F64>;
449 def: OpR_R_pat<F2_conv_w2sf, pf1<sint_to_fp>, f32, I32>;
450 def: OpR_R_pat<F2_conv_d2sf, pf1<sint_to_fp>, f32, I64>;
451 def: OpR_R_pat<F2_conv_w2df, pf1<sint_to_fp>, f64, I32>;
452 def: OpR_R_pat<F2_conv_d2df, pf1<sint_to_fp>, f64, I64>;
454 def: OpR_R_pat<F2_conv_uw2sf, pf1<uint_to_fp>, f32, I32>;
455 def: OpR_R_pat<F2_conv_ud2sf, pf1<uint_to_fp>, f32, I64>;
456 def: OpR_R_pat<F2_conv_uw2df, pf1<uint_to_fp>, f64, I32>;
457 def: OpR_R_pat<F2_conv_ud2df, pf1<uint_to_fp>, f64, I64>;
459 def: OpR_R_pat<F2_conv_sf2w_chop, pf1<fp_to_sint>, i32, F32>;
460 def: OpR_R_pat<F2_conv_df2w_chop, pf1<fp_to_sint>, i32, F64>;
461 def: OpR_R_pat<F2_conv_sf2d_chop, pf1<fp_to_sint>, i64, F32>;
462 def: OpR_R_pat<F2_conv_df2d_chop, pf1<fp_to_sint>, i64, F64>;
464 def: OpR_R_pat<F2_conv_sf2uw_chop, pf1<fp_to_uint>, i32, F32>;
465 def: OpR_R_pat<F2_conv_df2uw_chop, pf1<fp_to_uint>, i32, F64>;
466 def: OpR_R_pat<F2_conv_sf2ud_chop, pf1<fp_to_uint>, i64, F32>;
467 def: OpR_R_pat<F2_conv_df2ud_chop, pf1<fp_to_uint>, i64, F64>;
469 // Bitcast is different than [fp|sint|uint]_to_[sint|uint|fp].
470 def: Pat<(i32 (bitconvert F32:$v)), (I32:$v)>;
471 def: Pat<(f32 (bitconvert I32:$v)), (F32:$v)>;
472 def: Pat<(i64 (bitconvert F64:$v)), (I64:$v)>;
473 def: Pat<(f64 (bitconvert I64:$v)), (F64:$v)>;
475 // Bit convert 32- and 64-bit types.
476 // All of these are bitcastable to one another: i32, v2i16, v4i8.
477 defm: NopCast_pat<i32, v2i16, IntRegs>;
478 defm: NopCast_pat<i32, v4i8, IntRegs>;
479 defm: NopCast_pat<v2i16, v4i8, IntRegs>;
480 // All of these are bitcastable to one another: i64, v2i32, v4i16, v8i8.
481 defm: NopCast_pat<i64, v2i32, DoubleRegs>;
482 defm: NopCast_pat<i64, v4i16, DoubleRegs>;
483 defm: NopCast_pat<i64, v8i8, DoubleRegs>;
484 defm: NopCast_pat<v2i32, v4i16, DoubleRegs>;
485 defm: NopCast_pat<v2i32, v8i8, DoubleRegs>;
486 defm: NopCast_pat<v4i16, v8i8, DoubleRegs>;
489 // --(3) Extend/truncate/saturate ----------------------------------------
492 def: Pat<(sext_inreg I32:$Rs, i8), (A2_sxtb I32:$Rs)>;
493 def: Pat<(sext_inreg I32:$Rs, i16), (A2_sxth I32:$Rs)>;
494 def: Pat<(sext_inreg I64:$Rs, i32), (A2_sxtw (LoReg $Rs))>;
495 def: Pat<(sext_inreg I64:$Rs, i16), (A2_sxtw (A2_sxth (LoReg $Rs)))>;
496 def: Pat<(sext_inreg I64:$Rs, i8), (A2_sxtw (A2_sxtb (LoReg $Rs)))>;
498 def: Pat<(i64 (sext I32:$Rs)), (A2_sxtw I32:$Rs)>;
499 def: Pat<(Zext64 I32:$Rs), (ToZext64 $Rs)>;
500 def: Pat<(Aext64 I32:$Rs), (ToZext64 $Rs)>;
502 def: Pat<(i32 (trunc I64:$Rs)), (LoReg $Rs)>;
503 def: Pat<(i1 (trunc I32:$Rs)), (S2_tstbit_i I32:$Rs, 0)>;
504 def: Pat<(i1 (trunc I64:$Rs)), (S2_tstbit_i (LoReg $Rs), 0)>;
506 let AddedComplexity = 20 in {
507 def: Pat<(and I32:$Rs, 255), (A2_zxtb I32:$Rs)>;
508 def: Pat<(and I32:$Rs, 65535), (A2_zxth I32:$Rs)>;
511 // Extensions from i1 or vectors of i1.
512 def: Pat<(i32 (azext I1:$Pu)), (C2_muxii I1:$Pu, 1, 0)>;
513 def: Pat<(i64 (azext I1:$Pu)), (ToZext64 (C2_muxii I1:$Pu, 1, 0))>;
514 def: Pat<(i32 (sext I1:$Pu)), (C2_muxii I1:$Pu, -1, 0)>;
515 def: Pat<(i64 (sext I1:$Pu)), (Combinew (C2_muxii PredRegs:$Pu, -1, 0),
516 (C2_muxii PredRegs:$Pu, -1, 0))>;
518 def: Pat<(v2i16 (sext V2I1:$Pu)), (S2_vtrunehb (C2_mask V2I1:$Pu))>;
519 def: Pat<(v2i32 (sext V2I1:$Pu)), (C2_mask V2I1:$Pu)>;
520 def: Pat<(v4i8 (sext V4I1:$Pu)), (S2_vtrunehb (C2_mask V4I1:$Pu))>;
521 def: Pat<(v4i16 (sext V4I1:$Pu)), (C2_mask V4I1:$Pu)>;
522 def: Pat<(v8i8 (sext V8I1:$Pu)), (C2_mask V8I1:$Pu)>;
524 def Vsplatpi: OutPatFrag<(ops node:$V),
525 (Combinew (A2_tfrsi $V), (A2_tfrsi $V))>;
527 def: Pat<(v2i16 (azext V2I1:$Pu)),
528 (A2_andir (S2_vtrunehb (C2_mask V2I1:$Pu)), (i32 0x00010001))>;
529 def: Pat<(v2i32 (azext V2I1:$Pu)),
530 (A2_andp (C2_mask V2I1:$Pu), (A2_combineii (i32 1), (i32 1)))>;
531 def: Pat<(v4i8 (azext V4I1:$Pu)),
532 (A2_andir (S2_vtrunehb (C2_mask V4I1:$Pu)), (i32 0x01010101))>;
533 def: Pat<(v4i16 (azext V4I1:$Pu)),
534 (A2_andp (C2_mask V4I1:$Pu), (Vsplatpi (i32 0x00010001)))>;
535 def: Pat<(v8i8 (azext V8I1:$Pu)),
536 (A2_andp (C2_mask V8I1:$Pu), (Vsplatpi (i32 0x01010101)))>;
538 def: Pat<(v4i16 (azext V4I8:$Rs)), (S2_vzxtbh V4I8:$Rs)>;
539 def: Pat<(v2i32 (azext V2I16:$Rs)), (S2_vzxthw V2I16:$Rs)>;
540 def: Pat<(v4i16 (sext V4I8:$Rs)), (S2_vsxtbh V4I8:$Rs)>;
541 def: Pat<(v2i32 (sext V2I16:$Rs)), (S2_vsxthw V2I16:$Rs)>;
543 def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i8)),
544 (Combinew (A2_sxtb (HiReg $Rs)), (A2_sxtb (LoReg $Rs)))>;
546 def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i16)),
547 (Combinew (A2_sxth (HiReg $Rs)), (A2_sxth (LoReg $Rs)))>;
549 // Truncate: from vector B copy all 'E'ven 'B'yte elements:
550 // A[0] = B[0]; A[1] = B[2]; A[2] = B[4]; A[3] = B[6];
551 def: Pat<(v4i8 (trunc V4I16:$Rs)),
552 (S2_vtrunehb V4I16:$Rs)>;
554 // Truncate: from vector B copy all 'O'dd 'B'yte elements:
555 // A[0] = B[1]; A[1] = B[3]; A[2] = B[5]; A[3] = B[7];
558 // Truncate: from vectors B and C copy all 'E'ven 'H'alf-word elements:
559 // A[0] = B[0]; A[1] = B[2]; A[2] = C[0]; A[3] = C[2];
562 def: Pat<(v2i16 (trunc V2I32:$Rs)),
563 (A2_combine_ll (HiReg $Rs), (LoReg $Rs))>;
566 def: Pat<(v2i1 (trunc V2I32:$Rs)),
567 (A4_vcmpweqi (A2_andp V2I32:$Rs, (A2_combineii (i32 1), (i32 1))),
569 def: Pat<(v4i1 (trunc V4I16:$Rs)),
570 (A4_vcmpheqi (Combinew (A2_andir (HiReg $Rs), (i32 0x00010001)),
571 (A2_andir (LoReg $Rs), (i32 0x00010001))),
573 def: Pat<(v8i1 (trunc V8I8:$Rs)),
574 (A4_vcmpbeqi (Combinew (A2_andir (HiReg $Rs), (i32 0x01010101)),
575 (A2_andir (LoReg $Rs), (i32 0x01010101))),
580 // Note: saturation assumes the same signed-ness for the input and the
582 def: Pat<(i32 (ssat I32:$Rs, i8)), (A2_satb I32:$Rs)>;
583 def: Pat<(i32 (ssat I32:$Rs, i16)), (A2_sath I32:$Rs)>;
584 def: Pat<(i32 (ssat I64:$Rs, i32)), (A2_sat I64:$Rs)>;
585 def: Pat<(i32 (usat I32:$Rs, i8)), (A2_satub I32:$Rs)>;
586 def: Pat<(i32 (usat I32:$Rs, i16)), (A2_satuh I32:$Rs)>;
587 def: Pat<(i32 (usat I64:$Rs, i32)),
588 (C2_mux (C2_cmpeqi (HiReg $Rs), (i32 0)), (LoReg $Rs), (i32 -1))>;
590 def: Pat<(v4i8 (ssat V4I16:$Rs, v4i8)), (S2_vsathb V4I16:$Rs)>;
591 def: Pat<(v2i16 (ssat V2I32:$Rs, v2i16)), (S2_vsatwh V2I32:$Rs)>;
592 def: Pat<(v4i8 (usat V4I16:$Rs, v4i8)), (S2_vsathub V4I16:$Rs)>;
593 def: Pat<(v2i16 (usat V2I32:$Rs, v2i16)), (S2_vsatwuh V2I32:$Rs)>;
596 // --(4) Logical ---------------------------------------------------------
599 def: Pat<(not I1:$Ps), (C2_not I1:$Ps)>;
600 def: Pat<(pnot V2I1:$Ps), (C2_not V2I1:$Ps)>;
601 def: Pat<(pnot V4I1:$Ps), (C2_not V4I1:$Ps)>;
602 def: Pat<(pnot V8I1:$Ps), (C2_not V8I1:$Ps)>;
603 def: Pat<(add I1:$Ps, -1), (C2_not I1:$Ps)>;
605 def: OpR_RR_pat<C2_and, And, i1, I1>;
606 def: OpR_RR_pat<C2_or, Or, i1, I1>;
607 def: OpR_RR_pat<C2_xor, Xor, i1, I1>;
608 def: OpR_RR_pat<C2_andn, Not2<And>, i1, I1>;
609 def: OpR_RR_pat<C2_orn, Not2<Or>, i1, I1>;
611 def: AccRRR_pat<C4_and_and, And, Su<And>, I1, I1, I1>;
612 def: AccRRR_pat<C4_and_or, And, Su< Or>, I1, I1, I1>;
613 def: AccRRR_pat<C4_or_and, Or, Su<And>, I1, I1, I1>;
614 def: AccRRR_pat<C4_or_or, Or, Su< Or>, I1, I1, I1>;
615 def: AccRRR_pat<C4_and_andn, And, Su<Not2<And>>, I1, I1, I1>;
616 def: AccRRR_pat<C4_and_orn, And, Su<Not2< Or>>, I1, I1, I1>;
617 def: AccRRR_pat<C4_or_andn, Or, Su<Not2<And>>, I1, I1, I1>;
618 def: AccRRR_pat<C4_or_orn, Or, Su<Not2< Or>>, I1, I1, I1>;
620 multiclass BoolvOpR_RR_pat<InstHexagon MI, PatFrag VOp> {
621 def: OpR_RR_pat<MI, VOp, v2i1, V2I1>;
622 def: OpR_RR_pat<MI, VOp, v4i1, V4I1>;
623 def: OpR_RR_pat<MI, VOp, v8i1, V8I1>;
626 multiclass BoolvAccRRR_pat<InstHexagon MI, PatFrag AccOp, PatFrag VOp> {
627 def: AccRRR_pat<MI, AccOp, VOp, V2I1, V2I1, V2I1>;
628 def: AccRRR_pat<MI, AccOp, VOp, V4I1, V4I1, V4I1>;
629 def: AccRRR_pat<MI, AccOp, VOp, V8I1, V8I1, V8I1>;
632 defm: BoolvOpR_RR_pat<C2_and, And>;
633 defm: BoolvOpR_RR_pat<C2_or, Or>;
634 defm: BoolvOpR_RR_pat<C2_xor, Xor>;
635 defm: BoolvOpR_RR_pat<C2_andn, VNot2<And, pnot>>;
636 defm: BoolvOpR_RR_pat<C2_orn, VNot2< Or, pnot>>;
638 // op(Ps, op(Pt, Pu))
639 defm: BoolvAccRRR_pat<C4_and_and, And, Su<And>>;
640 defm: BoolvAccRRR_pat<C4_and_or, And, Su<Or>>;
641 defm: BoolvAccRRR_pat<C4_or_and, Or, Su<And>>;
642 defm: BoolvAccRRR_pat<C4_or_or, Or, Su<Or>>;
644 // op(Ps, op(Pt, !Pu))
645 defm: BoolvAccRRR_pat<C4_and_andn, And, Su<VNot2<And, pnot>>>;
646 defm: BoolvAccRRR_pat<C4_and_orn, And, Su<VNot2< Or, pnot>>>;
647 defm: BoolvAccRRR_pat<C4_or_andn, Or, Su<VNot2<And, pnot>>>;
648 defm: BoolvAccRRR_pat<C4_or_orn, Or, Su<VNot2< Or, pnot>>>;
651 // --(5) Compare ---------------------------------------------------------
654 // Avoid negated comparisons, i.e. those of form "Pd = !cmp(...)".
655 // These cannot form compounds (e.g. J4_cmpeqi_tp0_jump_nt).
657 def: OpR_RI_pat<C2_cmpeqi, seteq, i1, I32, anyimm>;
658 def: OpR_RI_pat<C2_cmpgti, setgt, i1, I32, anyimm>;
659 def: OpR_RI_pat<C2_cmpgtui, setugt, i1, I32, anyimm>;
661 def: Pat<(i1 (setge I32:$Rs, s32_0ImmPred:$s10)),
662 (C2_cmpgti I32:$Rs, (SDEC1 imm:$s10))>;
663 def: Pat<(i1 (setuge I32:$Rs, u32_0ImmPred:$u9)),
664 (C2_cmpgtui I32:$Rs, (UDEC1 imm:$u9))>;
666 def: Pat<(i1 (setlt I32:$Rs, s32_0ImmPred:$s10)),
667 (C2_not (C2_cmpgti I32:$Rs, (SDEC1 imm:$s10)))>;
668 def: Pat<(i1 (setult I32:$Rs, u32_0ImmPred:$u9)),
669 (C2_not (C2_cmpgtui I32:$Rs, (UDEC1 imm:$u9)))>;
671 // Patfrag to convert the usual comparison patfrags (e.g. setlt) to ones
672 // that reverse the order of the operands.
673 class RevCmp<PatFrag F>
674 : PatFrag<(ops node:$rhs, node:$lhs), !head(F.Fragments), F.PredicateCode,
677 def: OpR_RR_pat<C2_cmpeq, seteq, i1, I32>;
678 def: OpR_RR_pat<C2_cmpgt, setgt, i1, I32>;
679 def: OpR_RR_pat<C2_cmpgtu, setugt, i1, I32>;
680 def: OpR_RR_pat<C2_cmpgt, RevCmp<setlt>, i1, I32>;
681 def: OpR_RR_pat<C2_cmpgtu, RevCmp<setult>, i1, I32>;
682 def: OpR_RR_pat<C2_cmpeqp, seteq, i1, I64>;
683 def: OpR_RR_pat<C2_cmpgtp, setgt, i1, I64>;
684 def: OpR_RR_pat<C2_cmpgtup, setugt, i1, I64>;
685 def: OpR_RR_pat<C2_cmpgtp, RevCmp<setlt>, i1, I64>;
686 def: OpR_RR_pat<C2_cmpgtup, RevCmp<setult>, i1, I64>;
687 def: OpR_RR_pat<A2_vcmpbeq, seteq, i1, V8I8>;
688 def: OpR_RR_pat<A2_vcmpbeq, seteq, v8i1, V8I8>;
689 def: OpR_RR_pat<A4_vcmpbgt, RevCmp<setlt>, i1, V8I8>;
690 def: OpR_RR_pat<A4_vcmpbgt, RevCmp<setlt>, v8i1, V8I8>;
691 def: OpR_RR_pat<A4_vcmpbgt, setgt, i1, V8I8>;
692 def: OpR_RR_pat<A4_vcmpbgt, setgt, v8i1, V8I8>;
693 def: OpR_RR_pat<A2_vcmpbgtu, RevCmp<setult>, i1, V8I8>;
694 def: OpR_RR_pat<A2_vcmpbgtu, RevCmp<setult>, v8i1, V8I8>;
695 def: OpR_RR_pat<A2_vcmpbgtu, setugt, i1, V8I8>;
696 def: OpR_RR_pat<A2_vcmpbgtu, setugt, v8i1, V8I8>;
697 def: OpR_RR_pat<A2_vcmpheq, seteq, i1, V4I16>;
698 def: OpR_RR_pat<A2_vcmpheq, seteq, v4i1, V4I16>;
699 def: OpR_RR_pat<A2_vcmphgt, RevCmp<setlt>, i1, V4I16>;
700 def: OpR_RR_pat<A2_vcmphgt, RevCmp<setlt>, v4i1, V4I16>;
701 def: OpR_RR_pat<A2_vcmphgt, setgt, i1, V4I16>;
702 def: OpR_RR_pat<A2_vcmphgt, setgt, v4i1, V4I16>;
703 def: OpR_RR_pat<A2_vcmphgtu, RevCmp<setult>, i1, V4I16>;
704 def: OpR_RR_pat<A2_vcmphgtu, RevCmp<setult>, v4i1, V4I16>;
705 def: OpR_RR_pat<A2_vcmphgtu, setugt, i1, V4I16>;
706 def: OpR_RR_pat<A2_vcmphgtu, setugt, v4i1, V4I16>;
707 def: OpR_RR_pat<A2_vcmpweq, seteq, i1, V2I32>;
708 def: OpR_RR_pat<A2_vcmpweq, seteq, v2i1, V2I32>;
709 def: OpR_RR_pat<A2_vcmpwgt, RevCmp<setlt>, i1, V2I32>;
710 def: OpR_RR_pat<A2_vcmpwgt, RevCmp<setlt>, v2i1, V2I32>;
711 def: OpR_RR_pat<A2_vcmpwgt, setgt, i1, V2I32>;
712 def: OpR_RR_pat<A2_vcmpwgt, setgt, v2i1, V2I32>;
713 def: OpR_RR_pat<A2_vcmpwgtu, RevCmp<setult>, i1, V2I32>;
714 def: OpR_RR_pat<A2_vcmpwgtu, RevCmp<setult>, v2i1, V2I32>;
715 def: OpR_RR_pat<A2_vcmpwgtu, setugt, i1, V2I32>;
716 def: OpR_RR_pat<A2_vcmpwgtu, setugt, v2i1, V2I32>;
718 def: OpR_RR_pat<F2_sfcmpeq, seteq, i1, F32>;
719 def: OpR_RR_pat<F2_sfcmpgt, setgt, i1, F32>;
720 def: OpR_RR_pat<F2_sfcmpge, setge, i1, F32>;
721 def: OpR_RR_pat<F2_sfcmpeq, setoeq, i1, F32>;
722 def: OpR_RR_pat<F2_sfcmpgt, setogt, i1, F32>;
723 def: OpR_RR_pat<F2_sfcmpge, setoge, i1, F32>;
724 def: OpR_RR_pat<F2_sfcmpgt, RevCmp<setolt>, i1, F32>;
725 def: OpR_RR_pat<F2_sfcmpge, RevCmp<setole>, i1, F32>;
726 def: OpR_RR_pat<F2_sfcmpgt, RevCmp<setlt>, i1, F32>;
727 def: OpR_RR_pat<F2_sfcmpge, RevCmp<setle>, i1, F32>;
728 def: OpR_RR_pat<F2_sfcmpuo, setuo, i1, F32>;
730 def: OpR_RR_pat<F2_dfcmpeq, seteq, i1, F64>;
731 def: OpR_RR_pat<F2_dfcmpgt, setgt, i1, F64>;
732 def: OpR_RR_pat<F2_dfcmpge, setge, i1, F64>;
733 def: OpR_RR_pat<F2_dfcmpeq, setoeq, i1, F64>;
734 def: OpR_RR_pat<F2_dfcmpgt, setogt, i1, F64>;
735 def: OpR_RR_pat<F2_dfcmpge, setoge, i1, F64>;
736 def: OpR_RR_pat<F2_dfcmpgt, RevCmp<setolt>, i1, F64>;
737 def: OpR_RR_pat<F2_dfcmpge, RevCmp<setole>, i1, F64>;
738 def: OpR_RR_pat<F2_dfcmpgt, RevCmp<setlt>, i1, F64>;
739 def: OpR_RR_pat<F2_dfcmpge, RevCmp<setle>, i1, F64>;
740 def: OpR_RR_pat<F2_dfcmpuo, setuo, i1, F64>;
742 // Avoid C4_cmpneqi, C4_cmpltei, C4_cmplteui, since they cannot form compounds.
744 def: Pat<(i1 (setne I32:$Rs, anyimm:$u5)),
745 (C2_not (C2_cmpeqi I32:$Rs, imm:$u5))>;
746 def: Pat<(i1 (setle I32:$Rs, anyimm:$u5)),
747 (C2_not (C2_cmpgti I32:$Rs, imm:$u5))>;
748 def: Pat<(i1 (setule I32:$Rs, anyimm:$u5)),
749 (C2_not (C2_cmpgtui I32:$Rs, imm:$u5))>;
751 class OpmR_RR_pat<PatFrag Output, PatFrag Op, ValueType ResType,
752 PatFrag RsPred, PatFrag RtPred = RsPred>
753 : Pat<(ResType (Op RsPred:$Rs, RtPred:$Rt)),
754 (Output RsPred:$Rs, RtPred:$Rt)>;
756 class Outn<InstHexagon MI>
757 : OutPatFrag<(ops node:$Rs, node:$Rt),
758 (C2_not (MI $Rs, $Rt))>;
760 def: OpmR_RR_pat<Outn<C2_cmpeq>, setne, i1, I32>;
761 def: OpmR_RR_pat<Outn<C2_cmpgt>, setle, i1, I32>;
762 def: OpmR_RR_pat<Outn<C2_cmpgtu>, setule, i1, I32>;
763 def: OpmR_RR_pat<Outn<C2_cmpgt>, RevCmp<setge>, i1, I32>;
764 def: OpmR_RR_pat<Outn<C2_cmpgtu>, RevCmp<setuge>, i1, I32>;
765 def: OpmR_RR_pat<Outn<C2_cmpeqp>, setne, i1, I64>;
766 def: OpmR_RR_pat<Outn<C2_cmpgtp>, setle, i1, I64>;
767 def: OpmR_RR_pat<Outn<C2_cmpgtup>, setule, i1, I64>;
768 def: OpmR_RR_pat<Outn<C2_cmpgtp>, RevCmp<setge>, i1, I64>;
769 def: OpmR_RR_pat<Outn<C2_cmpgtup>, RevCmp<setuge>, i1, I64>;
770 def: OpmR_RR_pat<Outn<A2_vcmpbeq>, setne, v8i1, V8I8>;
771 def: OpmR_RR_pat<Outn<A4_vcmpbgt>, setle, v8i1, V8I8>;
772 def: OpmR_RR_pat<Outn<A2_vcmpbgtu>, setule, v8i1, V8I8>;
773 def: OpmR_RR_pat<Outn<A4_vcmpbgt>, RevCmp<setge>, v8i1, V8I8>;
774 def: OpmR_RR_pat<Outn<A2_vcmpbgtu>, RevCmp<setuge>, v8i1, V8I8>;
775 def: OpmR_RR_pat<Outn<A2_vcmpheq>, setne, v4i1, V4I16>;
776 def: OpmR_RR_pat<Outn<A2_vcmphgt>, setle, v4i1, V4I16>;
777 def: OpmR_RR_pat<Outn<A2_vcmphgtu>, setule, v4i1, V4I16>;
778 def: OpmR_RR_pat<Outn<A2_vcmphgt>, RevCmp<setge>, v4i1, V4I16>;
779 def: OpmR_RR_pat<Outn<A2_vcmphgtu>, RevCmp<setuge>, v4i1, V4I16>;
780 def: OpmR_RR_pat<Outn<A2_vcmpweq>, setne, v2i1, V2I32>;
781 def: OpmR_RR_pat<Outn<A2_vcmpwgt>, setle, v2i1, V2I32>;
782 def: OpmR_RR_pat<Outn<A2_vcmpwgtu>, setule, v2i1, V2I32>;
783 def: OpmR_RR_pat<Outn<A2_vcmpwgt>, RevCmp<setge>, v2i1, V2I32>;
784 def: OpmR_RR_pat<Outn<A2_vcmpwgtu>, RevCmp<setuge>, v2i1, V2I32>;
786 let AddedComplexity = 100 in {
787 def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt), 255), 0)),
788 (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt)>;
789 def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt), 255), 0)),
790 (C2_not (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt))>;
791 def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt), 65535), 0)),
792 (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt)>;
793 def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt), 65535), 0)),
794 (C2_not (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt))>;
797 // PatFrag for AsserZext which takes the original type as a parameter.
798 def SDTAssertZext: SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisSameAs<0,1>]>;
799 def AssertZextSD: SDNode<"ISD::AssertZext", SDTAssertZext>;
800 class AssertZext<ValueType T>: PatFrag<(ops node:$A), (AssertZextSD $A, T)>;
802 multiclass Cmpb_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
803 PatLeaf ImmPred, int Mask> {
804 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
805 (MI I32:$Rs, imm:$I)>;
806 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
807 (MI I32:$Rs, imm:$I)>;
810 multiclass CmpbN_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
811 PatLeaf ImmPred, int Mask> {
812 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
813 (C2_not (MI I32:$Rs, imm:$I))>;
814 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
815 (C2_not (MI I32:$Rs, imm:$I))>;
818 multiclass CmpbND_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
819 PatLeaf ImmPred, int Mask> {
820 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
821 (C2_not (MI I32:$Rs, (UDEC1 imm:$I)))>;
822 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
823 (C2_not (MI I32:$Rs, (UDEC1 imm:$I)))>;
826 let AddedComplexity = 200 in {
827 defm: Cmpb_pat <A4_cmpbeqi, seteq, AssertZext<i8>, IsUGT<8,31>, 255>;
828 defm: CmpbN_pat <A4_cmpbeqi, setne, AssertZext<i8>, IsUGT<8,31>, 255>;
829 defm: Cmpb_pat <A4_cmpbgtui, setugt, AssertZext<i8>, IsUGT<32,31>, 255>;
830 defm: CmpbN_pat <A4_cmpbgtui, setule, AssertZext<i8>, IsUGT<32,31>, 255>;
831 defm: Cmpb_pat <A4_cmphgtui, setugt, AssertZext<i16>, IsUGT<32,31>, 65535>;
832 defm: CmpbN_pat <A4_cmphgtui, setule, AssertZext<i16>, IsUGT<32,31>, 65535>;
833 defm: CmpbND_pat<A4_cmpbgtui, setult, AssertZext<i8>, IsUGT<32,32>, 255>;
834 defm: CmpbND_pat<A4_cmphgtui, setult, AssertZext<i16>, IsUGT<32,32>, 65535>;
837 def: Pat<(i32 (zext (i1 (seteq I32:$Rs, I32:$Rt)))),
838 (A4_rcmpeq I32:$Rs, I32:$Rt)>;
839 def: Pat<(i32 (zext (i1 (setne I32:$Rs, I32:$Rt)))),
840 (A4_rcmpneq I32:$Rs, I32:$Rt)>;
841 def: Pat<(i32 (zext (i1 (seteq I32:$Rs, anyimm:$s8)))),
842 (A4_rcmpeqi I32:$Rs, imm:$s8)>;
843 def: Pat<(i32 (zext (i1 (setne I32:$Rs, anyimm:$s8)))),
844 (A4_rcmpneqi I32:$Rs, imm:$s8)>;
846 def: Pat<(i1 (seteq I1:$Ps, (i1 -1))), (I1:$Ps)>;
847 def: Pat<(i1 (setne I1:$Ps, (i1 -1))), (C2_not I1:$Ps)>;
848 def: Pat<(i1 (seteq I1:$Ps, I1:$Pt)), (C2_not (C2_xor I1:$Ps, I1:$Pt))>;
849 def: Pat<(i1 (setne I1:$Ps, I1:$Pt)), (C2_xor I1:$Ps, I1:$Pt)>;
851 multiclass BoolE_pat<PatFrag OpPred, ValueType ResTy> {
852 def: Pat<(ResTy (seteq OpPred:$Ps, OpPred:$Pt)), (C2_not (C2_xor $Ps, $Pt))>;
853 def: Pat<(ResTy (setne OpPred:$Ps, OpPred:$Pt)), (C2_xor $Ps, $Pt)>;
856 defm: BoolE_pat<I1, i1>;
857 defm: BoolE_pat<V2I1, v2i1>;
858 defm: BoolE_pat<V4I1, v4i1>;
859 defm: BoolE_pat<V8I1, v8i1>;
861 multiclass BoolL_pat<PatFrag OpPred, ValueType ResTy> {
862 // Signed "true" == -1
863 def: Pat<(ResTy (setlt OpPred:$Ps, OpPred:$Pt)), (C2_andn $Ps, $Pt)>;
864 def: Pat<(ResTy (setle OpPred:$Ps, OpPred:$Pt)), (C2_orn $Ps, $Pt)>;
865 def: Pat<(ResTy (setult OpPred:$Ps, OpPred:$Pt)), (C2_andn $Pt, $Ps)>;
866 def: Pat<(ResTy (setule OpPred:$Ps, OpPred:$Pt)), (C2_orn $Pt, $Ps)>;
869 defm: BoolL_pat<I1, i1>;
870 defm: BoolL_pat<V2I1, v2i1>;
871 defm: BoolL_pat<V4I1, v4i1>;
872 defm: BoolL_pat<V8I1, v8i1>;
874 // Floating-point comparisons with checks for ordered/unordered status.
876 class T3<InstHexagon MI1, InstHexagon MI2, InstHexagon MI3>
877 : OutPatFrag<(ops node:$Rs, node:$Rt),
878 (MI1 (MI2 $Rs, $Rt), (MI3 $Rs, $Rt))>;
880 class Cmpuf<InstHexagon MI>: T3<C2_or, F2_sfcmpuo, MI>;
881 class Cmpud<InstHexagon MI>: T3<C2_or, F2_dfcmpuo, MI>;
883 class Cmpufn<InstHexagon MI>: T3<C2_orn, F2_sfcmpuo, MI>;
884 class Cmpudn<InstHexagon MI>: T3<C2_orn, F2_dfcmpuo, MI>;
886 def: OpmR_RR_pat<Cmpuf<F2_sfcmpeq>, setueq, i1, F32>;
887 def: OpmR_RR_pat<Cmpuf<F2_sfcmpge>, setuge, i1, F32>;
888 def: OpmR_RR_pat<Cmpuf<F2_sfcmpgt>, setugt, i1, F32>;
889 def: OpmR_RR_pat<Cmpuf<F2_sfcmpge>, RevCmp<setule>, i1, F32>;
890 def: OpmR_RR_pat<Cmpuf<F2_sfcmpgt>, RevCmp<setult>, i1, F32>;
891 def: OpmR_RR_pat<Cmpufn<F2_sfcmpeq>, setune, i1, F32>;
893 def: OpmR_RR_pat<Cmpud<F2_dfcmpeq>, setueq, i1, F64>;
894 def: OpmR_RR_pat<Cmpud<F2_dfcmpge>, setuge, i1, F64>;
895 def: OpmR_RR_pat<Cmpud<F2_dfcmpgt>, setugt, i1, F64>;
896 def: OpmR_RR_pat<Cmpud<F2_dfcmpge>, RevCmp<setule>, i1, F64>;
897 def: OpmR_RR_pat<Cmpud<F2_dfcmpgt>, RevCmp<setult>, i1, F64>;
898 def: OpmR_RR_pat<Cmpudn<F2_dfcmpeq>, setune, i1, F64>;
900 def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setone, i1, F32>;
901 def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setne, i1, F32>;
903 def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setone, i1, F64>;
904 def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setne, i1, F64>;
906 def: OpmR_RR_pat<Outn<F2_sfcmpuo>, seto, i1, F32>;
907 def: OpmR_RR_pat<Outn<F2_dfcmpuo>, seto, i1, F64>;
910 // --(6) Select ----------------------------------------------------------
913 def: Pat<(select I1:$Pu, I32:$Rs, I32:$Rt),
914 (C2_mux I1:$Pu, I32:$Rs, I32:$Rt)>;
915 def: Pat<(select I1:$Pu, v4i8:$Rs, v4i8:$Rt),
916 (C2_mux I1:$Pu, v4i8:$Rs, v4i8:$Rt)>;
917 def: Pat<(select I1:$Pu, v2i16:$Rs, v2i16:$Rt),
918 (C2_mux I1:$Pu, v2i16:$Rs, v2i16:$Rt)>;
919 def: Pat<(select I1:$Pu, anyimm:$s8, I32:$Rs),
920 (C2_muxri I1:$Pu, imm:$s8, I32:$Rs)>;
921 def: Pat<(select I1:$Pu, I32:$Rs, anyimm:$s8),
922 (C2_muxir I1:$Pu, I32:$Rs, imm:$s8)>;
923 def: Pat<(select I1:$Pu, anyimm:$s8, s8_0ImmPred:$S8),
924 (C2_muxii I1:$Pu, imm:$s8, imm:$S8)>;
926 def: Pat<(select (not I1:$Pu), I32:$Rs, I32:$Rt),
927 (C2_mux I1:$Pu, I32:$Rt, I32:$Rs)>;
928 def: Pat<(select (not I1:$Pu), s8_0ImmPred:$S8, anyimm:$s8),
929 (C2_muxii I1:$Pu, imm:$s8, imm:$S8)>;
930 def: Pat<(select (not I1:$Pu), anyimm:$s8, I32:$Rs),
931 (C2_muxir I1:$Pu, I32:$Rs, imm:$s8)>;
932 def: Pat<(select (not I1:$Pu), I32:$Rs, anyimm:$s8),
933 (C2_muxri I1:$Pu, imm:$s8, I32:$Rs)>;
935 // Map from a 64-bit select to an emulated 64-bit mux.
936 // Hexagon does not support 64-bit MUXes; so emulate with combines.
937 def: Pat<(select I1:$Pu, I64:$Rs, I64:$Rt),
938 (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
939 (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
941 def: Pat<(select I1:$Pu, v2i32:$Rs, v2i32:$Rt),
942 (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
943 (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
945 def: Pat<(select I1:$Pu, F32:$Rs, f32ImmPred:$I),
946 (C2_muxir I1:$Pu, F32:$Rs, (ftoi $I))>;
947 def: Pat<(select I1:$Pu, f32ImmPred:$I, F32:$Rt),
948 (C2_muxri I1:$Pu, (ftoi $I), F32:$Rt)>;
949 def: Pat<(select I1:$Pu, F32:$Rs, F32:$Rt),
950 (C2_mux I1:$Pu, F32:$Rs, F32:$Rt)>;
951 def: Pat<(select I1:$Pu, F64:$Rs, F64:$Rt),
952 (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
953 (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
955 def: Pat<(select (i1 (setult F32:$Ra, F32:$Rb)), F32:$Rs, F32:$Rt),
956 (C2_mux (F2_sfcmpgt F32:$Rb, F32:$Ra), F32:$Rs, F32:$Rt)>;
957 def: Pat<(select (i1 (setult F64:$Ra, F64:$Rb)), F64:$Rs, F64:$Rt),
958 (C2_vmux (F2_dfcmpgt F64:$Rb, F64:$Ra), F64:$Rs, F64:$Rt)>;
960 def: Pat<(select (not I1:$Pu), f32ImmPred:$I, F32:$Rs),
961 (C2_muxir I1:$Pu, F32:$Rs, (ftoi $I))>;
962 def: Pat<(select (not I1:$Pu), F32:$Rt, f32ImmPred:$I),
963 (C2_muxri I1:$Pu, (ftoi $I), F32:$Rt)>;
965 def: Pat<(vselect V8I1:$Pu, V8I8:$Rs, V8I8:$Rt),
966 (C2_vmux V8I1:$Pu, V8I8:$Rs, V8I8:$Rt)>;
967 def: Pat<(vselect V4I1:$Pu, V4I16:$Rs, V4I16:$Rt),
968 (C2_vmux V4I1:$Pu, V4I16:$Rs, V4I16:$Rt)>;
969 def: Pat<(vselect V2I1:$Pu, V2I32:$Rs, V2I32:$Rt),
970 (C2_vmux V2I1:$Pu, V2I32:$Rs, V2I32:$Rt)>;
972 def: Pat<(vselect (pnot V8I1:$Pu), V8I8:$Rs, V8I8:$Rt),
973 (C2_vmux V8I1:$Pu, V8I8:$Rt, V8I8:$Rs)>;
974 def: Pat<(vselect (pnot V4I1:$Pu), V4I16:$Rs, V4I16:$Rt),
975 (C2_vmux V4I1:$Pu, V4I16:$Rt, V4I16:$Rs)>;
976 def: Pat<(vselect (pnot V2I1:$Pu), V2I32:$Rs, V2I32:$Rt),
977 (C2_vmux V2I1:$Pu, V2I32:$Rt, V2I32:$Rs)>;
980 // From LegalizeDAG.cpp: (Pu ? Pv : Pw) <=> (Pu & Pv) | (!Pu & Pw).
981 def: Pat<(select I1:$Pu, I1:$Ps, I1:$Pt),
982 (C4_or_andn (C2_and $Ps, $Pu), $Pt, $Pu)>;
984 def: Pat<(vselect V2I1:$Pu, V2I1:$Ps, V2I1:$Pt),
985 (C4_or_andn (C2_and $Ps, $Pu), $Pt, $Pu)>;
986 def: Pat<(vselect V4I1:$Pu, V4I1:$Ps, V4I1:$Pt),
987 (C4_or_andn (C2_and $Ps, $Pu), $Pt, $Pu)>;
988 def: Pat<(vselect V8I1:$Pu, V8I1:$Ps, V8I1:$Pt),
989 (C4_or_andn (C2_and $Ps, $Pu), $Pt, $Pu)>;
991 def: Pat<(select I1:$Pu, V2I1:$Ps, V2I1:$Pt),
992 (C2_tfrrp (C2_mux $Pu, (C2_tfrpr $Ps), (C2_tfrpr $Pt)))>;
993 def: Pat<(select I1:$Pu, V4I1:$Ps, V4I1:$Pt),
994 (C2_tfrrp (C2_mux $Pu, (C2_tfrpr $Ps), (C2_tfrpr $Pt)))>;
995 def: Pat<(select I1:$Pu, V8I1:$Ps, V8I1:$Pt),
996 (C2_tfrrp (C2_mux $Pu, (C2_tfrpr $Ps), (C2_tfrpr $Pt)))>;
998 def IsPosHalf : PatLeaf<(i32 IntRegs:$a), [{
999 return isPositiveHalfWord(N);
1002 multiclass SelMinMax16_pats<PatFrag CmpOp, InstHexagon InstA,
1003 InstHexagon InstB> {
1004 def: Pat<(sext_inreg (select (i1 (CmpOp IsPosHalf:$Rs, IsPosHalf:$Rt)),
1005 IsPosHalf:$Rs, IsPosHalf:$Rt), i16),
1006 (InstA IntRegs:$Rs, IntRegs:$Rt)>;
1007 def: Pat<(sext_inreg (select (i1 (CmpOp IsPosHalf:$Rs, IsPosHalf:$Rt)),
1008 IsPosHalf:$Rt, IsPosHalf:$Rs), i16),
1009 (InstB IntRegs:$Rs, IntRegs:$Rt)>;
1012 let AddedComplexity = 200 in {
1013 defm: SelMinMax16_pats<setge, A2_max, A2_min>;
1014 defm: SelMinMax16_pats<setgt, A2_max, A2_min>;
1015 defm: SelMinMax16_pats<setle, A2_min, A2_max>;
1016 defm: SelMinMax16_pats<setlt, A2_min, A2_max>;
1017 defm: SelMinMax16_pats<setuge, A2_maxu, A2_minu>;
1018 defm: SelMinMax16_pats<setugt, A2_maxu, A2_minu>;
1019 defm: SelMinMax16_pats<setule, A2_minu, A2_maxu>;
1020 defm: SelMinMax16_pats<setult, A2_minu, A2_maxu>;
1023 def: OpR_RR_pat<A2_min, Smin, i32, I32, I32>;
1024 def: OpR_RR_pat<A2_max, Smax, i32, I32, I32>;
1025 def: OpR_RR_pat<A2_minu, Umin, i32, I32, I32>;
1026 def: OpR_RR_pat<A2_maxu, Umax, i32, I32, I32>;
1027 def: OpR_RR_pat<A2_minp, Smin, i64, I64, I64>;
1028 def: OpR_RR_pat<A2_maxp, Smax, i64, I64, I64>;
1029 def: OpR_RR_pat<A2_minup, Umin, i64, I64, I64>;
1030 def: OpR_RR_pat<A2_maxup, Umax, i64, I64, I64>;
1032 let AddedComplexity = 100 in {
1033 defm: MinMax_pats<F2_sfmin, F2_sfmax, select, setogt, i1, F32>;
1034 defm: MinMax_pats<F2_sfmin, F2_sfmax, select, setoge, i1, F32>;
1035 defm: MinMax_pats<F2_sfmax, F2_sfmin, select, setolt, i1, F32>;
1036 defm: MinMax_pats<F2_sfmax, F2_sfmin, select, setole, i1, F32>;
1039 let AddedComplexity = 100, Predicates = [HasV67] in {
1040 defm: MinMax_pats<F2_dfmin, F2_dfmax, select, setogt, i1, F64>;
1041 defm: MinMax_pats<F2_dfmin, F2_dfmax, select, setoge, i1, F64>;
1042 defm: MinMax_pats<F2_dfmax, F2_dfmin, select, setolt, i1, F64>;
1043 defm: MinMax_pats<F2_dfmax, F2_dfmin, select, setole, i1, F64>;
1046 def: OpR_RR_pat<A2_vminb, Smin, v8i8, V8I8>;
1047 def: OpR_RR_pat<A2_vmaxb, Smax, v8i8, V8I8>;
1048 def: OpR_RR_pat<A2_vminub, Umin, v8i8, V8I8>;
1049 def: OpR_RR_pat<A2_vmaxub, Umax, v8i8, V8I8>;
1051 def: OpR_RR_pat<A2_vminh, Smin, v4i16, V4I16>;
1052 def: OpR_RR_pat<A2_vmaxh, Smax, v4i16, V4I16>;
1053 def: OpR_RR_pat<A2_vminuh, Umin, v4i16, V4I16>;
1054 def: OpR_RR_pat<A2_vmaxuh, Umax, v4i16, V4I16>;
1056 def: OpR_RR_pat<A2_vminw, Smin, v2i32, V2I32>;
1057 def: OpR_RR_pat<A2_vmaxw, Smax, v2i32, V2I32>;
1058 def: OpR_RR_pat<A2_vminuw, Umin, v2i32, V2I32>;
1059 def: OpR_RR_pat<A2_vmaxuw, Umax, v2i32, V2I32>;
1061 // --(7) Insert/extract --------------------------------------------------
1064 def SDTHexagonINSERT:
1065 SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>,
1066 SDTCisInt<0>, SDTCisVT<3, i32>, SDTCisVT<4, i32>]>;
1067 def HexagonINSERT: SDNode<"HexagonISD::INSERT", SDTHexagonINSERT>;
1069 let AddedComplexity = 10 in {
1070 def: Pat<(HexagonINSERT I32:$Rs, I32:$Rt, u5_0ImmPred:$u1, u5_0ImmPred:$u2),
1071 (S2_insert I32:$Rs, I32:$Rt, imm:$u1, imm:$u2)>;
1072 def: Pat<(HexagonINSERT I64:$Rs, I64:$Rt, u6_0ImmPred:$u1, u6_0ImmPred:$u2),
1073 (S2_insertp I64:$Rs, I64:$Rt, imm:$u1, imm:$u2)>;
1075 def: Pat<(HexagonINSERT I32:$Rs, I32:$Rt, I32:$Width, I32:$Off),
1076 (S2_insert_rp I32:$Rs, I32:$Rt, (Combinew $Width, $Off))>;
1077 def: Pat<(HexagonINSERT I64:$Rs, I64:$Rt, I32:$Width, I32:$Off),
1078 (S2_insertp_rp I64:$Rs, I64:$Rt, (Combinew $Width, $Off))>;
1080 def SDTHexagonEXTRACTU
1081 : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisInt<0>, SDTCisInt<1>,
1082 SDTCisVT<2, i32>, SDTCisVT<3, i32>]>;
1083 def HexagonEXTRACTU: SDNode<"HexagonISD::EXTRACTU", SDTHexagonEXTRACTU>;
1085 let AddedComplexity = 10 in {
1086 def: Pat<(HexagonEXTRACTU I32:$Rs, u5_0ImmPred:$u5, u5_0ImmPred:$U5),
1087 (S2_extractu I32:$Rs, imm:$u5, imm:$U5)>;
1088 def: Pat<(HexagonEXTRACTU I64:$Rs, u6_0ImmPred:$u6, u6_0ImmPred:$U6),
1089 (S2_extractup I64:$Rs, imm:$u6, imm:$U6)>;
1091 def: Pat<(HexagonEXTRACTU I32:$Rs, I32:$Width, I32:$Off),
1092 (S2_extractu_rp I32:$Rs, (Combinew $Width, $Off))>;
1093 def: Pat<(HexagonEXTRACTU I64:$Rs, I32:$Width, I32:$Off),
1094 (S2_extractup_rp I64:$Rs, (Combinew $Width, $Off))>;
1096 def: Pat<(v4i8 (splat_vector anyint:$V)), (ToI32 (SplatB $V))>;
1097 def: Pat<(v2i16 (splat_vector anyint:$V)), (ToI32 (SplatH $V))>;
1098 def: Pat<(v8i8 (splat_vector anyint:$V)),
1099 (Combinew (ToI32 (SplatB $V)), (ToI32 (SplatB $V)))>;
1100 def: Pat<(v4i16 (splat_vector anyint:$V)),
1101 (Combinew (ToI32 (SplatH $V)), (ToI32 (SplatH $V)))>;
1102 let AddedComplexity = 10 in
1103 def: Pat<(v2i32 (splat_vector s8_0ImmPred:$s8)),
1104 (A2_combineii imm:$s8, imm:$s8)>;
1105 def: Pat<(v2i32 (splat_vector anyimm:$V)), (Combinew (ToI32 $V), (ToI32 $V))>;
1107 def: Pat<(v4i8 (splat_vector I32:$Rs)), (S2_vsplatrb I32:$Rs)>;
1108 def: Pat<(v2i16 (splat_vector I32:$Rs)), (LoReg (S2_vsplatrh I32:$Rs))>;
1109 def: Pat<(v4i16 (splat_vector I32:$Rs)), (S2_vsplatrh I32:$Rs)>;
1110 def: Pat<(v2i32 (splat_vector I32:$Rs)), (Combinew I32:$Rs, I32:$Rs)>;
1112 let AddedComplexity = 10 in
1113 def: Pat<(v8i8 (splat_vector I32:$Rs)), (S6_vsplatrbp I32:$Rs)>,
1115 def: Pat<(v8i8 (splat_vector I32:$Rs)),
1116 (Combinew (S2_vsplatrb I32:$Rs), (S2_vsplatrb I32:$Rs))>;
1118 let AddedComplexity = 10 in {
1119 def: Pat<(sext_inreg (HexagonEXTRACTU I32:$Rs, 8, u5_0ImmPred:$U5), i8),
1120 (S4_extract I32:$Rs, 8, imm:$U5)>;
1121 def: Pat<(sext_inreg (HexagonEXTRACTU I32:$Rs, 16, u5_0ImmPred:$U5), i16),
1122 (S4_extract I32:$Rs, 16, imm:$U5)>;
1123 def: Pat<(sext_inreg (HexagonEXTRACTU I64:$Rs, 8, u6_0ImmPred:$U6), i8),
1124 (S4_extractp I64:$Rs, 8, imm:$U6)>;
1125 def: Pat<(sext_inreg (HexagonEXTRACTU I64:$Rs, 16, u6_0ImmPred:$U6), i16),
1126 (S4_extractp I64:$Rs, 16, imm:$U6)>;
1127 def: Pat<(sext_inreg (HexagonEXTRACTU I64:$Rs, 32, u6_0ImmPred:$U6), i32),
1128 (S4_extractp I64:$Rs, 32, imm:$U6)>;
1131 def: Pat<(sext_inreg (HexagonEXTRACTU I32:$Rs, 8, I32:$Off), i8),
1132 (S4_extract_rp I32:$Rs, (Combinew (ToI32 8), I32:$Off))>;
1133 def: Pat<(sext_inreg (HexagonEXTRACTU I32:$Rs, 16, I32:$Off), i16),
1134 (S4_extract_rp I32:$Rs, (Combinew (ToI32 16), I32:$Off))>;
1135 def: Pat<(sext_inreg (HexagonEXTRACTU I64:$Rs, 8, I32:$Off), i8),
1136 (S4_extractp_rp I64:$Rs, (Combinew (ToI32 8), I32:$Off))>;
1137 def: Pat<(sext_inreg (HexagonEXTRACTU I64:$Rs, 16, I32:$Off), i16),
1138 (S4_extractp_rp I64:$Rs, (Combinew (ToI32 16), I32:$Off))>;
1139 def: Pat<(sext_inreg (HexagonEXTRACTU I64:$Rs, 32, I32:$Off), i32),
1140 (S4_extractp_rp I64:$Rs, (Combinew (ToI32 32), I32:$Off))>;
1143 // --(8) Shift/permute ---------------------------------------------------
1146 def SDTHexagonI64I32I32: SDTypeProfile<1, 2,
1147 [SDTCisVT<0, i64>, SDTCisVT<1, i32>, SDTCisSameAs<1, 2>]>;
1149 def HexagonCOMBINE: SDNode<"HexagonISD::COMBINE", SDTHexagonI64I32I32>;
1151 def: Pat<(HexagonCOMBINE I32:$Rs, I32:$Rt), (Combinew $Rs, $Rt)>;
1153 // The complexity of the combines involving immediates should be greater
1154 // than the complexity of the combine with two registers.
1155 let AddedComplexity = 50 in {
1156 def: Pat<(HexagonCOMBINE I32:$Rs, anyimm:$s8),
1157 (A4_combineri IntRegs:$Rs, imm:$s8)>;
1158 def: Pat<(HexagonCOMBINE anyimm:$s8, I32:$Rs),
1159 (A4_combineir imm:$s8, IntRegs:$Rs)>;
1162 // The complexity of the combine with two immediates should be greater than
1163 // the complexity of a combine involving a register.
1164 let AddedComplexity = 75 in {
1165 def: Pat<(HexagonCOMBINE s8_0ImmPred:$s8, anyimm:$u6),
1166 (A4_combineii imm:$s8, imm:$u6)>;
1167 def: Pat<(HexagonCOMBINE anyimm:$s8, s8_0ImmPred:$S8),
1168 (A2_combineii imm:$s8, imm:$S8)>;
1171 def: Pat<(bswap I32:$Rs), (A2_swiz I32:$Rs)>;
1172 def: Pat<(bswap I64:$Rss), (Combinew (A2_swiz (LoReg $Rss)),
1173 (A2_swiz (HiReg $Rss)))>;
1175 def: Pat<(bswap V2I16:$Rs), (A2_combine_lh (A2_swiz $Rs), (A2_swiz $Rs))>;
1176 def: Pat<(bswap V2I32:$Rs), (Combinew (A2_swiz (HiReg $Rs)),
1177 (A2_swiz (LoReg $Rs)))>;
1178 def: Pat<(bswap V4I16:$Rs), (A2_orp (S2_lsr_i_vh $Rs, 8),
1179 (S2_asl_i_vh $Rs, 8))>;
1181 def: Pat<(shl s6_0ImmPred:$s6, I32:$Rt), (S4_lsli imm:$s6, I32:$Rt)>;
1182 def: Pat<(shl I32:$Rs, (i32 16)), (A2_aslh I32:$Rs)>;
1183 def: Pat<(sra I32:$Rs, (i32 16)), (A2_asrh I32:$Rs)>;
1185 def: OpR_RI_pat<S2_asr_i_r, Sra, i32, I32, u5_0ImmPred>;
1186 def: OpR_RI_pat<S2_lsr_i_r, Srl, i32, I32, u5_0ImmPred>;
1187 def: OpR_RI_pat<S2_asl_i_r, Shl, i32, I32, u5_0ImmPred>;
1188 def: OpR_RI_pat<S2_asr_i_p, Sra, i64, I64, u6_0ImmPred>;
1189 def: OpR_RI_pat<S2_lsr_i_p, Srl, i64, I64, u6_0ImmPred>;
1190 def: OpR_RI_pat<S2_asl_i_p, Shl, i64, I64, u6_0ImmPred>;
1191 def: OpR_RI_pat<S2_asr_i_vh, Sra, v4i16, V4I16, u4_0ImmPred>;
1192 def: OpR_RI_pat<S2_lsr_i_vh, Srl, v4i16, V4I16, u4_0ImmPred>;
1193 def: OpR_RI_pat<S2_asl_i_vh, Shl, v4i16, V4I16, u4_0ImmPred>;
1194 def: OpR_RI_pat<S2_asr_i_vh, Sra, v2i32, V2I32, u5_0ImmPred>;
1195 def: OpR_RI_pat<S2_lsr_i_vh, Srl, v2i32, V2I32, u5_0ImmPred>;
1196 def: OpR_RI_pat<S2_asl_i_vh, Shl, v2i32, V2I32, u5_0ImmPred>;
1198 def: OpR_RR_pat<S2_asr_r_r, Sra, i32, I32, I32>;
1199 def: OpR_RR_pat<S2_lsr_r_r, Srl, i32, I32, I32>;
1200 def: OpR_RR_pat<S2_asl_r_r, Shl, i32, I32, I32>;
1201 def: OpR_RR_pat<S2_asr_r_p, Sra, i64, I64, I32>;
1202 def: OpR_RR_pat<S2_lsr_r_p, Srl, i64, I64, I32>;
1203 def: OpR_RR_pat<S2_asl_r_p, Shl, i64, I64, I32>;
1206 def IsMul8_U3: PatLeaf<(i32 imm), [{
1207 uint64_t V = N->getZExtValue();
1208 return V % 8 == 0 && isUInt<3>(V / 8);
1211 def Divu8: SDNodeXForm<imm, [{
1212 return CurDAG->getTargetConstant(N->getZExtValue() / 8, SDLoc(N), MVT::i32);
1215 // Funnel shift-left.
1216 def FShl32i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1217 (HiReg (S2_asl_i_p (Combinew $Rs, $Rt), $S))>;
1218 def FShl32r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1219 (HiReg (S2_asl_r_p (Combinew $Rs, $Rt), $Ru))>;
1221 def FShl64i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1222 (S2_lsr_i_p_or (S2_asl_i_p $Rs, $S), $Rt, (Subi<64> $S))>;
1223 def FShl64r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1224 (S2_lsr_r_p_or (S2_asl_r_p $Rs, $Ru), $Rt, (A2_subri 64, $Ru))>;
1226 // Combined SDNodeXForm: (Divu8 (Subi<64> $S))
1227 def Divu64_8: SDNodeXForm<imm, [{
1228 return CurDAG->getTargetConstant((64 - N->getSExtValue()) / 8,
1229 SDLoc(N), MVT::i32);
1233 let AddedComplexity = 100 in {
1234 def: Pat<(fshl I32:$Rs, I32:$Rt, (i32 16)),
1235 (A2_combine_lh I32:$Rs, I32:$Rt)>;
1236 def: Pat<(fshl I64:$Rs, I64:$Rt, IsMul8_U3:$S),
1237 (S2_valignib I64:$Rs, I64:$Rt, (Divu64_8 $S))>;
1240 let Predicates = [HasV60], AddedComplexity = 50 in {
1241 def: OpR_RI_pat<S6_rol_i_r, Rol, i32, I32, u5_0ImmPred>;
1242 def: OpR_RI_pat<S6_rol_i_p, Rol, i64, I64, u6_0ImmPred>;
1244 let AddedComplexity = 30 in {
1245 def: Pat<(rotl I32:$Rs, u5_0ImmPred:$S), (FShl32i $Rs, $Rs, imm:$S)>;
1246 def: Pat<(rotl I64:$Rs, u6_0ImmPred:$S), (FShl64i $Rs, $Rs, imm:$S)>;
1247 def: Pat<(fshl I32:$Rs, I32:$Rt, u5_0ImmPred:$S), (FShl32i $Rs, $Rt, imm:$S)>;
1248 def: Pat<(fshl I64:$Rs, I64:$Rt, u6_0ImmPred:$S), (FShl64i $Rs, $Rt, imm:$S)>;
1250 def: Pat<(rotl I32:$Rs, I32:$Rt), (FShl32r $Rs, $Rs, $Rt)>;
1251 def: Pat<(rotl I64:$Rs, I32:$Rt), (FShl64r $Rs, $Rs, $Rt)>;
1252 def: Pat<(fshl I32:$Rs, I32:$Rt, I32:$Ru), (FShl32r $Rs, $Rt, $Ru)>;
1253 def: Pat<(fshl I64:$Rs, I64:$Rt, I32:$Ru), (FShl64r $Rs, $Rt, $Ru)>;
1255 // Funnel shift-right.
1256 def FShr32i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1257 (LoReg (S2_lsr_i_p (Combinew $Rs, $Rt), $S))>;
1258 def FShr32r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1259 (LoReg (S2_lsr_r_p (Combinew $Rs, $Rt), $Ru))>;
1261 def FShr64i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1262 (S2_asl_i_p_or (S2_lsr_i_p $Rt, $S), $Rs, (Subi<64> $S))>;
1263 def FShr64r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1264 (S2_asl_r_p_or (S2_lsr_r_p $Rt, $Ru), $Rs, (A2_subri 64, $Ru))>;
1267 let AddedComplexity = 100 in {
1268 def: Pat<(fshr I32:$Rs, I32:$Rt, (i32 16)),
1269 (A2_combine_lh I32:$Rs, I32:$Rt)>;
1270 def: Pat<(fshr I64:$Rs, I64:$Rt, IsMul8_U3:$S),
1271 (S2_valignib I64:$Rs, I64:$Rt, (Divu8 $S))>;
1274 let Predicates = [HasV60], AddedComplexity = 50 in {
1275 def: Pat<(rotr I32:$Rs, u5_0ImmPred:$S), (S6_rol_i_r I32:$Rs, (Subi<32> $S))>;
1276 def: Pat<(rotr I64:$Rs, u6_0ImmPred:$S), (S6_rol_i_p I64:$Rs, (Subi<64> $S))>;
1278 let AddedComplexity = 30 in {
1279 def: Pat<(rotr I32:$Rs, u5_0ImmPred:$S), (FShr32i $Rs, $Rs, imm:$S)>;
1280 def: Pat<(rotr I64:$Rs, u6_0ImmPred:$S), (FShr64i $Rs, $Rs, imm:$S)>;
1281 def: Pat<(fshr I32:$Rs, I32:$Rt, u5_0ImmPred:$S), (FShr32i $Rs, $Rt, imm:$S)>;
1282 def: Pat<(fshr I64:$Rs, I64:$Rt, u6_0ImmPred:$S), (FShr64i $Rs, $Rt, imm:$S)>;
1284 def: Pat<(rotr I32:$Rs, I32:$Rt), (FShr32r $Rs, $Rs, $Rt)>;
1285 def: Pat<(rotr I64:$Rs, I32:$Rt), (FShr64r $Rs, $Rs, $Rt)>;
1286 def: Pat<(fshr I32:$Rs, I32:$Rt, I32:$Ru), (FShr32r $Rs, $Rt, $Ru)>;
1287 def: Pat<(fshr I64:$Rs, I64:$Rt, I32:$Ru), (FShr64r $Rs, $Rt, $Ru)>;
1290 def: Pat<(sra (add (sra I32:$Rs, u5_0ImmPred:$u5), 1), (i32 1)),
1291 (S2_asr_i_r_rnd I32:$Rs, imm:$u5)>;
1292 def: Pat<(sra (add (sra I64:$Rs, u6_0ImmPred:$u6), 1), (i32 1)),
1293 (S2_asr_i_p_rnd I64:$Rs, imm:$u6)>;
1295 // Prefer S2_addasl_rrri over S2_asl_i_r_acc.
1296 let AddedComplexity = 120 in
1297 def: Pat<(add I32:$Rt, (shl I32:$Rs, u3_0ImmPred:$u3)),
1298 (S2_addasl_rrri IntRegs:$Rt, IntRegs:$Rs, imm:$u3)>;
1300 let AddedComplexity = 100 in {
1301 def: AccRRI_pat<S2_asr_i_r_acc, Add, Su<Sra>, I32, u5_0ImmPred>;
1302 def: AccRRI_pat<S2_asr_i_r_nac, Sub, Su<Sra>, I32, u5_0ImmPred>;
1303 def: AccRRI_pat<S2_asr_i_r_and, And, Su<Sra>, I32, u5_0ImmPred>;
1304 def: AccRRI_pat<S2_asr_i_r_or, Or, Su<Sra>, I32, u5_0ImmPred>;
1306 def: AccRRI_pat<S2_asr_i_p_acc, Add, Su<Sra>, I64, u6_0ImmPred>;
1307 def: AccRRI_pat<S2_asr_i_p_nac, Sub, Su<Sra>, I64, u6_0ImmPred>;
1308 def: AccRRI_pat<S2_asr_i_p_and, And, Su<Sra>, I64, u6_0ImmPred>;
1309 def: AccRRI_pat<S2_asr_i_p_or, Or, Su<Sra>, I64, u6_0ImmPred>;
1311 def: AccRRI_pat<S2_lsr_i_r_acc, Add, Su<Srl>, I32, u5_0ImmPred>;
1312 def: AccRRI_pat<S2_lsr_i_r_nac, Sub, Su<Srl>, I32, u5_0ImmPred>;
1313 def: AccRRI_pat<S2_lsr_i_r_and, And, Su<Srl>, I32, u5_0ImmPred>;
1314 def: AccRRI_pat<S2_lsr_i_r_or, Or, Su<Srl>, I32, u5_0ImmPred>;
1315 def: AccRRI_pat<S2_lsr_i_r_xacc, Xor, Su<Srl>, I32, u5_0ImmPred>;
1317 def: AccRRI_pat<S2_lsr_i_p_acc, Add, Su<Srl>, I64, u6_0ImmPred>;
1318 def: AccRRI_pat<S2_lsr_i_p_nac, Sub, Su<Srl>, I64, u6_0ImmPred>;
1319 def: AccRRI_pat<S2_lsr_i_p_and, And, Su<Srl>, I64, u6_0ImmPred>;
1320 def: AccRRI_pat<S2_lsr_i_p_or, Or, Su<Srl>, I64, u6_0ImmPred>;
1321 def: AccRRI_pat<S2_lsr_i_p_xacc, Xor, Su<Srl>, I64, u6_0ImmPred>;
1323 def: AccRRI_pat<S2_asl_i_r_acc, Add, Su<Shl>, I32, u5_0ImmPred>;
1324 def: AccRRI_pat<S2_asl_i_r_nac, Sub, Su<Shl>, I32, u5_0ImmPred>;
1325 def: AccRRI_pat<S2_asl_i_r_and, And, Su<Shl>, I32, u5_0ImmPred>;
1326 def: AccRRI_pat<S2_asl_i_r_or, Or, Su<Shl>, I32, u5_0ImmPred>;
1327 def: AccRRI_pat<S2_asl_i_r_xacc, Xor, Su<Shl>, I32, u5_0ImmPred>;
1329 def: AccRRI_pat<S2_asl_i_p_acc, Add, Su<Shl>, I64, u6_0ImmPred>;
1330 def: AccRRI_pat<S2_asl_i_p_nac, Sub, Su<Shl>, I64, u6_0ImmPred>;
1331 def: AccRRI_pat<S2_asl_i_p_and, And, Su<Shl>, I64, u6_0ImmPred>;
1332 def: AccRRI_pat<S2_asl_i_p_or, Or, Su<Shl>, I64, u6_0ImmPred>;
1333 def: AccRRI_pat<S2_asl_i_p_xacc, Xor, Su<Shl>, I64, u6_0ImmPred>;
1335 let Predicates = [HasV60] in {
1336 def: AccRRI_pat<S6_rol_i_r_acc, Add, Su<Rol>, I32, u5_0ImmPred>;
1337 def: AccRRI_pat<S6_rol_i_r_nac, Sub, Su<Rol>, I32, u5_0ImmPred>;
1338 def: AccRRI_pat<S6_rol_i_r_and, And, Su<Rol>, I32, u5_0ImmPred>;
1339 def: AccRRI_pat<S6_rol_i_r_or, Or, Su<Rol>, I32, u5_0ImmPred>;
1340 def: AccRRI_pat<S6_rol_i_r_xacc, Xor, Su<Rol>, I32, u5_0ImmPred>;
1342 def: AccRRI_pat<S6_rol_i_p_acc, Add, Su<Rol>, I64, u6_0ImmPred>;
1343 def: AccRRI_pat<S6_rol_i_p_nac, Sub, Su<Rol>, I64, u6_0ImmPred>;
1344 def: AccRRI_pat<S6_rol_i_p_and, And, Su<Rol>, I64, u6_0ImmPred>;
1345 def: AccRRI_pat<S6_rol_i_p_or, Or, Su<Rol>, I64, u6_0ImmPred>;
1346 def: AccRRI_pat<S6_rol_i_p_xacc, Xor, Su<Rol>, I64, u6_0ImmPred>;
1350 let AddedComplexity = 100 in {
1351 def: AccRRR_pat<S2_asr_r_r_acc, Add, Su<Sra>, I32, I32, I32>;
1352 def: AccRRR_pat<S2_asr_r_r_nac, Sub, Su<Sra>, I32, I32, I32>;
1353 def: AccRRR_pat<S2_asr_r_r_and, And, Su<Sra>, I32, I32, I32>;
1354 def: AccRRR_pat<S2_asr_r_r_or, Or, Su<Sra>, I32, I32, I32>;
1356 def: AccRRR_pat<S2_asr_r_p_acc, Add, Su<Sra>, I64, I64, I32>;
1357 def: AccRRR_pat<S2_asr_r_p_nac, Sub, Su<Sra>, I64, I64, I32>;
1358 def: AccRRR_pat<S2_asr_r_p_and, And, Su<Sra>, I64, I64, I32>;
1359 def: AccRRR_pat<S2_asr_r_p_or, Or, Su<Sra>, I64, I64, I32>;
1360 def: AccRRR_pat<S2_asr_r_p_xor, Xor, Su<Sra>, I64, I64, I32>;
1362 def: AccRRR_pat<S2_lsr_r_r_acc, Add, Su<Srl>, I32, I32, I32>;
1363 def: AccRRR_pat<S2_lsr_r_r_nac, Sub, Su<Srl>, I32, I32, I32>;
1364 def: AccRRR_pat<S2_lsr_r_r_and, And, Su<Srl>, I32, I32, I32>;
1365 def: AccRRR_pat<S2_lsr_r_r_or, Or, Su<Srl>, I32, I32, I32>;
1367 def: AccRRR_pat<S2_lsr_r_p_acc, Add, Su<Srl>, I64, I64, I32>;
1368 def: AccRRR_pat<S2_lsr_r_p_nac, Sub, Su<Srl>, I64, I64, I32>;
1369 def: AccRRR_pat<S2_lsr_r_p_and, And, Su<Srl>, I64, I64, I32>;
1370 def: AccRRR_pat<S2_lsr_r_p_or, Or, Su<Srl>, I64, I64, I32>;
1371 def: AccRRR_pat<S2_lsr_r_p_xor, Xor, Su<Srl>, I64, I64, I32>;
1373 def: AccRRR_pat<S2_asl_r_r_acc, Add, Su<Shl>, I32, I32, I32>;
1374 def: AccRRR_pat<S2_asl_r_r_nac, Sub, Su<Shl>, I32, I32, I32>;
1375 def: AccRRR_pat<S2_asl_r_r_and, And, Su<Shl>, I32, I32, I32>;
1376 def: AccRRR_pat<S2_asl_r_r_or, Or, Su<Shl>, I32, I32, I32>;
1378 def: AccRRR_pat<S2_asl_r_p_acc, Add, Su<Shl>, I64, I64, I32>;
1379 def: AccRRR_pat<S2_asl_r_p_nac, Sub, Su<Shl>, I64, I64, I32>;
1380 def: AccRRR_pat<S2_asl_r_p_and, And, Su<Shl>, I64, I64, I32>;
1381 def: AccRRR_pat<S2_asl_r_p_or, Or, Su<Shl>, I64, I64, I32>;
1382 def: AccRRR_pat<S2_asl_r_p_xor, Xor, Su<Shl>, I64, I64, I32>;
1386 class OpshIRI_pat<InstHexagon MI, PatFrag Op, PatFrag ShOp,
1387 PatFrag RegPred, PatFrag ImmPred>
1388 : Pat<(Op anyimm:$u8, (ShOp RegPred:$Rs, ImmPred:$U5)),
1389 (MI anyimm:$u8, RegPred:$Rs, imm:$U5)>;
1391 let AddedComplexity = 200, Predicates = [UseCompound] in {
1392 def: OpshIRI_pat<S4_addi_asl_ri, Add, Su<Shl>, I32, u5_0ImmPred>;
1393 def: OpshIRI_pat<S4_addi_lsr_ri, Add, Su<Srl>, I32, u5_0ImmPred>;
1394 def: OpshIRI_pat<S4_subi_asl_ri, Sub, Su<Shl>, I32, u5_0ImmPred>;
1395 def: OpshIRI_pat<S4_subi_lsr_ri, Sub, Su<Srl>, I32, u5_0ImmPred>;
1396 def: OpshIRI_pat<S4_andi_asl_ri, And, Su<Shl>, I32, u5_0ImmPred>;
1397 def: OpshIRI_pat<S4_andi_lsr_ri, And, Su<Srl>, I32, u5_0ImmPred>;
1398 def: OpshIRI_pat<S4_ori_asl_ri, Or, Su<Shl>, I32, u5_0ImmPred>;
1399 def: OpshIRI_pat<S4_ori_lsr_ri, Or, Su<Srl>, I32, u5_0ImmPred>;
1402 // Prefer this pattern to S2_asl_i_p_or for the special case of joining
1403 // two 32-bit words into a 64-bit word.
1404 let AddedComplexity = 200 in
1405 def: Pat<(or (shl (Aext64 I32:$a), (i32 32)), (Zext64 I32:$b)),
1406 (Combinew I32:$a, I32:$b)>;
1408 def: Pat<(or (or (or (shl (Zext64 (and I32:$b, (i32 65535))), (i32 16)),
1409 (Zext64 (and I32:$a, (i32 65535)))),
1410 (shl (Aext64 (and I32:$c, (i32 65535))), (i32 32))),
1411 (shl (Aext64 I32:$d), (i32 48))),
1412 (Combinew (A2_combine_ll I32:$d, I32:$c),
1413 (A2_combine_ll I32:$b, I32:$a))>;
1415 let AddedComplexity = 200 in {
1416 def: Pat<(or (shl I32:$Rt, (i32 16)), (and I32:$Rs, (i32 65535))),
1417 (A2_combine_ll I32:$Rt, I32:$Rs)>;
1418 def: Pat<(or (shl I32:$Rt, (i32 16)), (srl I32:$Rs, (i32 16))),
1419 (A2_combine_lh I32:$Rt, I32:$Rs)>;
1420 def: Pat<(or (and I32:$Rt, (i32 268431360)), (and I32:$Rs, (i32 65535))),
1421 (A2_combine_hl I32:$Rt, I32:$Rs)>;
1422 def: Pat<(or (and I32:$Rt, (i32 268431360)), (srl I32:$Rs, (i32 16))),
1423 (A2_combine_hh I32:$Rt, I32:$Rs)>;
1426 def SDTHexagonVShift
1427 : SDTypeProfile<1, 2, [SDTCisSameAs<0, 1>, SDTCisVec<0>, SDTCisVT<2, i32>]>;
1429 def HexagonVASL: SDNode<"HexagonISD::VASL", SDTHexagonVShift>;
1430 def HexagonVASR: SDNode<"HexagonISD::VASR", SDTHexagonVShift>;
1431 def HexagonVLSR: SDNode<"HexagonISD::VLSR", SDTHexagonVShift>;
1433 // Funnel shifts with the shift amount module element bit width.
1434 def HexagonMFSHL: SDNode<"HexagonISD::MFSHL", SDTIntShiftDOp>;
1435 def HexagonMFSHR: SDNode<"HexagonISD::MFSHR", SDTIntShiftDOp>;
1437 def: OpR_RI_pat<S2_asl_i_vw, pf2<HexagonVASL>, v2i32, V2I32, u5_0ImmPred>;
1438 def: OpR_RI_pat<S2_asl_i_vh, pf2<HexagonVASL>, v4i16, V4I16, u4_0ImmPred>;
1439 def: OpR_RI_pat<S2_asr_i_vw, pf2<HexagonVASR>, v2i32, V2I32, u5_0ImmPred>;
1440 def: OpR_RI_pat<S2_asr_i_vh, pf2<HexagonVASR>, v4i16, V4I16, u4_0ImmPred>;
1441 def: OpR_RI_pat<S2_lsr_i_vw, pf2<HexagonVLSR>, v2i32, V2I32, u5_0ImmPred>;
1442 def: OpR_RI_pat<S2_lsr_i_vh, pf2<HexagonVLSR>, v4i16, V4I16, u4_0ImmPred>;
1444 def: OpR_RR_pat<S2_asl_r_vw, pf2<HexagonVASL>, v2i32, V2I32, I32>;
1445 def: OpR_RR_pat<S2_asl_r_vh, pf2<HexagonVASL>, v4i16, V4I16, I32>;
1446 def: OpR_RR_pat<S2_asr_r_vw, pf2<HexagonVASR>, v2i32, V2I32, I32>;
1447 def: OpR_RR_pat<S2_asr_r_vh, pf2<HexagonVASR>, v4i16, V4I16, I32>;
1448 def: OpR_RR_pat<S2_lsr_r_vw, pf2<HexagonVLSR>, v2i32, V2I32, I32>;
1449 def: OpR_RR_pat<S2_lsr_r_vh, pf2<HexagonVLSR>, v4i16, V4I16, I32>;
1451 def: Pat<(sra V2I32:$b, (v2i32 (splat_vector u5_0ImmPred:$c))),
1452 (S2_asr_i_vw V2I32:$b, imm:$c)>;
1453 def: Pat<(srl V2I32:$b, (v2i32 (splat_vector u5_0ImmPred:$c))),
1454 (S2_lsr_i_vw V2I32:$b, imm:$c)>;
1455 def: Pat<(shl V2I32:$b, (v2i32 (splat_vector u5_0ImmPred:$c))),
1456 (S2_asl_i_vw V2I32:$b, imm:$c)>;
1457 def: Pat<(sra V4I16:$b, (v4i16 (splat_vector u4_0ImmPred:$c))),
1458 (S2_asr_i_vh V4I16:$b, imm:$c)>;
1459 def: Pat<(srl V4I16:$b, (v4i16 (splat_vector u4_0ImmPred:$c))),
1460 (S2_lsr_i_vh V4I16:$b, imm:$c)>;
1461 def: Pat<(shl V4I16:$b, (v4i16 (splat_vector u4_0ImmPred:$c))),
1462 (S2_asl_i_vh V4I16:$b, imm:$c)>;
1464 def: Pat<(HexagonVASR V2I16:$Rs, u4_0ImmPred:$S),
1465 (LoReg (S2_asr_i_vh (ToAext64 $Rs), imm:$S))>;
1466 def: Pat<(HexagonVASL V2I16:$Rs, u4_0ImmPred:$S),
1467 (LoReg (S2_asl_i_vh (ToAext64 $Rs), imm:$S))>;
1468 def: Pat<(HexagonVLSR V2I16:$Rs, u4_0ImmPred:$S),
1469 (LoReg (S2_lsr_i_vh (ToAext64 $Rs), imm:$S))>;
1470 def: Pat<(HexagonVASR V2I16:$Rs, I32:$Rt),
1471 (LoReg (S2_asr_i_vh (ToAext64 $Rs), I32:$Rt))>;
1472 def: Pat<(HexagonVASL V2I16:$Rs, I32:$Rt),
1473 (LoReg (S2_asl_i_vh (ToAext64 $Rs), I32:$Rt))>;
1474 def: Pat<(HexagonVLSR V2I16:$Rs, I32:$Rt),
1475 (LoReg (S2_lsr_i_vh (ToAext64 $Rs), I32:$Rt))>;
1478 // --(9) Arithmetic/bitwise ----------------------------------------------
1481 def: Pat<(abs I32:$Rs), (A2_abs I32:$Rs)>;
1482 def: Pat<(abs I64:$Rs), (A2_absp I64:$Rs)>;
1483 def: Pat<(not I32:$Rs), (A2_subri -1, I32:$Rs)>;
1484 def: Pat<(not I64:$Rs), (A2_notp I64:$Rs)>;
1485 def: Pat<(ineg I64:$Rs), (A2_negp I64:$Rs)>;
1487 def: Pat<(fabs F32:$Rs), (S2_clrbit_i F32:$Rs, 31)>;
1488 def: Pat<(fneg F32:$Rs), (S2_togglebit_i F32:$Rs, 31)>;
1490 def: Pat<(fabs F64:$Rs),
1491 (Combinew (S2_clrbit_i (HiReg $Rs), 31),
1492 (i32 (LoReg $Rs)))>;
1493 def: Pat<(fneg F64:$Rs),
1494 (Combinew (S2_togglebit_i (HiReg $Rs), 31),
1495 (i32 (LoReg $Rs)))>;
1497 def: Pat<(add I32:$Rs, anyimm:$s16), (A2_addi I32:$Rs, imm:$s16)>;
1498 def: Pat<(or I32:$Rs, anyimm:$s10), (A2_orir I32:$Rs, imm:$s10)>;
1499 def: Pat<(and I32:$Rs, anyimm:$s10), (A2_andir I32:$Rs, imm:$s10)>;
1500 def: Pat<(sub anyimm:$s10, I32:$Rs), (A2_subri imm:$s10, I32:$Rs)>;
1502 def: OpR_RR_pat<A2_add, Add, i32, I32>;
1503 def: OpR_RR_pat<A2_sub, Sub, i32, I32>;
1504 def: OpR_RR_pat<A2_and, And, i32, I32>;
1505 def: OpR_RR_pat<A2_or, Or, i32, I32>;
1506 def: OpR_RR_pat<A2_xor, Xor, i32, I32>;
1507 def: OpR_RR_pat<A2_addp, Add, i64, I64>;
1508 def: OpR_RR_pat<A2_subp, Sub, i64, I64>;
1509 def: OpR_RR_pat<A2_andp, And, i64, I64>;
1510 def: OpR_RR_pat<A2_orp, Or, i64, I64>;
1511 def: OpR_RR_pat<A2_xorp, Xor, i64, I64>;
1512 def: OpR_RR_pat<A4_andnp, Not2<And>, i64, I64>;
1513 def: OpR_RR_pat<A4_ornp, Not2<Or>, i64, I64>;
1515 def: OpR_RR_pat<A2_svaddh, Add, v2i16, V2I16>;
1516 def: OpR_RR_pat<A2_svsubh, Sub, v2i16, V2I16>;
1518 def: OpR_RR_pat<A2_vaddub, Add, v8i8, V8I8>;
1519 def: OpR_RR_pat<A2_vaddh, Add, v4i16, V4I16>;
1520 def: OpR_RR_pat<A2_vaddw, Add, v2i32, V2I32>;
1521 def: OpR_RR_pat<A2_vsubub, Sub, v8i8, V8I8>;
1522 def: OpR_RR_pat<A2_vsubh, Sub, v4i16, V4I16>;
1523 def: OpR_RR_pat<A2_vsubw, Sub, v2i32, V2I32>;
1525 def: OpR_RR_pat<A2_and, And, v4i8, V4I8>;
1526 def: OpR_RR_pat<A2_xor, Xor, v4i8, V4I8>;
1527 def: OpR_RR_pat<A2_or, Or, v4i8, V4I8>;
1528 def: OpR_RR_pat<A2_and, And, v2i16, V2I16>;
1529 def: OpR_RR_pat<A2_xor, Xor, v2i16, V2I16>;
1530 def: OpR_RR_pat<A2_or, Or, v2i16, V2I16>;
1531 def: OpR_RR_pat<A2_andp, And, v8i8, V8I8>;
1532 def: OpR_RR_pat<A2_orp, Or, v8i8, V8I8>;
1533 def: OpR_RR_pat<A2_xorp, Xor, v8i8, V8I8>;
1534 def: OpR_RR_pat<A2_andp, And, v4i16, V4I16>;
1535 def: OpR_RR_pat<A2_orp, Or, v4i16, V4I16>;
1536 def: OpR_RR_pat<A2_xorp, Xor, v4i16, V4I16>;
1537 def: OpR_RR_pat<A2_andp, And, v2i32, V2I32>;
1538 def: OpR_RR_pat<A2_orp, Or, v2i32, V2I32>;
1539 def: OpR_RR_pat<A2_xorp, Xor, v2i32, V2I32>;
1541 def: OpR_RR_pat<M2_mpyi, Mul, i32, I32>;
1542 def: OpR_RR_pat<M2_mpy_up, pf2<mulhs>, i32, I32>;
1543 def: OpR_RR_pat<M2_mpyu_up, pf2<mulhu>, i32, I32>;
1544 def: OpR_RI_pat<M2_mpysip, Mul, i32, I32, u32_0ImmPred>;
1545 def: OpR_RI_pat<M2_mpysmi, Mul, i32, I32, s32_0ImmPred>;
1547 // Arithmetic on predicates.
1548 def: OpR_RR_pat<C2_xor, Add, i1, I1>;
1549 def: OpR_RR_pat<C2_xor, Add, v2i1, V2I1>;
1550 def: OpR_RR_pat<C2_xor, Add, v4i1, V4I1>;
1551 def: OpR_RR_pat<C2_xor, Add, v8i1, V8I1>;
1552 def: OpR_RR_pat<C2_xor, Sub, i1, I1>;
1553 def: OpR_RR_pat<C2_xor, Sub, v2i1, V2I1>;
1554 def: OpR_RR_pat<C2_xor, Sub, v4i1, V4I1>;
1555 def: OpR_RR_pat<C2_xor, Sub, v8i1, V8I1>;
1556 def: OpR_RR_pat<C2_and, Mul, i1, I1>;
1557 def: OpR_RR_pat<C2_and, Mul, v2i1, V2I1>;
1558 def: OpR_RR_pat<C2_and, Mul, v4i1, V4I1>;
1559 def: OpR_RR_pat<C2_and, Mul, v8i1, V8I1>;
1561 def: OpR_RR_pat<F2_sfadd, pf2<fadd>, f32, F32>;
1562 def: OpR_RR_pat<F2_sfsub, pf2<fsub>, f32, F32>;
1563 def: OpR_RR_pat<F2_sfmpy, pf2<fmul>, f32, F32>;
1564 def: OpR_RR_pat<F2_sfmin, pf2<fminnum>, f32, F32>;
1565 def: OpR_RR_pat<F2_sfmax, pf2<fmaxnum>, f32, F32>;
1567 let Predicates = [HasV66] in {
1568 def: OpR_RR_pat<F2_dfadd, pf2<fadd>, f64, F64>;
1569 def: OpR_RR_pat<F2_dfsub, pf2<fsub>, f64, F64>;
1572 def DfMpy: OutPatFrag<(ops node:$Rs, node:$Rt),
1576 (F2_dfmpyll $Rs, $Rt),
1581 let Predicates = [HasV67,UseUnsafeMath], AddedComplexity = 50 in {
1582 def: Pat<(fmul F64:$Rs, F64:$Rt), (DfMpy $Rs, $Rt)>;
1584 let Predicates = [HasV67] in {
1585 def: OpR_RR_pat<F2_dfmin, pf2<fminnum>, f64, F64>;
1586 def: OpR_RR_pat<F2_dfmax, pf2<fmaxnum>, f64, F64>;
1588 def: Pat<(fmul F64:$Rs, F64:$Rt), (DfMpy (F2_dfmpyfix $Rs, $Rt),
1589 (F2_dfmpyfix $Rt, $Rs))>;
1592 // In expressions like a0*b0 + a1*b1 + ..., prefer to generate multiply-add,
1593 // over add-add with individual multiplies as inputs.
1594 let AddedComplexity = 10 in {
1595 def: AccRRI_pat<M2_macsip, Add, Su<Mul>, I32, u32_0ImmPred>;
1596 def: AccRRI_pat<M2_macsin, Sub, Su<Mul>, I32, u32_0ImmPred>;
1597 def: AccRRR_pat<M2_maci, Add, Su<Mul>, I32, I32, I32>;
1598 let Predicates = [HasV66] in
1599 def: AccRRR_pat<M2_mnaci, Sub, Su<Mul>, I32, I32, I32>;
1602 def: AccRRI_pat<M2_naccii, Sub, Su<Add>, I32, s32_0ImmPred>;
1603 def: AccRRI_pat<M2_accii, Add, Su<Add>, I32, s32_0ImmPred>;
1604 def: AccRRR_pat<M2_acci, Add, Su<Add>, I32, I32, I32>;
1608 def: Pat<(v2i32 (mulhu V2I32:$Rss, V2I32:$Rtt)),
1609 (Combinew (M2_mpyu_up (HiReg $Rss), (HiReg $Rtt)),
1610 (M2_mpyu_up (LoReg $Rss), (LoReg $Rtt)))>;
1612 def: Pat<(v2i32 (mulhs V2I32:$Rss, V2I32:$Rtt)),
1613 (Combinew (M2_mpy_up (HiReg $Rss), (HiReg $Rtt)),
1614 (M2_mpy_up (LoReg $Rss), (LoReg $Rtt)))>;
1617 OutPatFrag<(ops node:$Rs, node:$Rt), (S2_vtrunohb (M5_vmpybuu $Rs, $Rt))>;
1619 OutPatFrag<(ops node:$Rss, node:$Rtt),
1620 (Combinew (Mulhub4 (HiReg $Rss), (HiReg $Rtt)),
1621 (Mulhub4 (LoReg $Rss), (LoReg $Rtt)))>;
1623 // (mux (x >= 0), 0, y)
1625 OutPatFrag<(ops node:$Rss, node:$Rtt),
1626 (C2_vmux (A4_vcmpbgti $Rss, -1), (A2_tfrpi 0), $Rtt)>;
1628 def: Pat<(v4i8 (mulhu V4I8:$Rs, V4I8:$Rt)), (Mulhub4 $Rs, $Rt)>;
1629 def: Pat<(v8i8 (mulhu V8I8:$Rss, V8I8:$Rtt)), (Mulhub8 $Rss, $Rtt)>;
1631 // (Mulhs x, y) = (Mulhu x, y) - (x < 0 ? y : 0) - (y < 0 ? x : 0)
1633 OutPatFrag<(ops node:$Rss, node:$Rtt),
1634 (A2_vsubub (Mulhub8 $Rss, $Rtt),
1635 (A2_vaddub (Negbytes8 $Rss, $Rtt),
1636 (Negbytes8 $Rtt, $Rss)))>;
1638 def: Pat<(v4i8 (mulhs V4I8:$Rs, V4I8:$Rt)),
1639 (LoReg (Mulhsb8 (v8i8 (ToAext64 $Rs)), (v8i8 (ToAext64 $Rt))))>;
1640 def: Pat<(v8i8 (mulhs V8I8:$Rss, V8I8:$Rtt)), (Mulhsb8 $Rss, $Rtt)>;
1642 // v2i16 *s v2i16 -> v2i32
1644 OutPatFrag<(ops node:$Rs, node:$Rt), (M2_vmpy2s_s0 $Rs, $Rt)>;
1647 OutPatFrag<(ops node:$Rs, node:$Rt),
1648 (A2_combine_hh (HiReg (Muli16 $Rs, $Rt)),
1649 (LoReg (Muli16 $Rs, $Rt)))>;
1651 OutPatFrag<(ops node:$Rss, node:$Rtt),
1652 (Combinew (Mulhsh2 (HiReg $Rss), (HiReg $Rtt)),
1653 (Mulhsh2 (LoReg $Rss), (LoReg $Rtt)))>;
1655 def: Pat<(v2i16 (mulhs V2I16:$Rs, V2I16:$Rt)), (Mulhsh2 $Rs, $Rt)>;
1656 def: Pat<(v4i16 (mulhs V4I16:$Rss, V4I16:$Rtt)), (Mulhsh4 $Rss, $Rtt)>;
1658 def: Pat<(v2i16 (mulhu V2I16:$Rs, V2I16:$Rt)),
1661 (A2_svaddh (LoReg (A2_andp (Combinew $Rt, $Rs),
1662 (S2_asr_i_vh (Combinew $Rs, $Rt), 15))),
1663 (HiReg (A2_andp (Combinew $Rt, $Rs),
1664 (S2_asr_i_vh (Combinew $Rs, $Rt), 15)))))>;
1666 def: Pat<(v4i16 (mulhu V4I16:$Rss, V4I16:$Rtt)),
1668 (Mulhsh4 $Rss, $Rtt),
1669 (A2_vaddh (A2_andp V4I16:$Rss, (S2_asr_i_vh $Rtt, 15)),
1670 (A2_andp V4I16:$Rtt, (S2_asr_i_vh $Rss, 15))))>;
1673 def: Pat<(ineg (mul I32:$Rs, u8_0ImmPred:$u8)),
1674 (M2_mpysin IntRegs:$Rs, imm:$u8)>;
1676 def n8_0ImmPred: PatLeaf<(i32 imm), [{
1677 int64_t V = N->getSExtValue();
1678 return -255 <= V && V <= 0;
1681 // Change the sign of the immediate for Rd=-mpyi(Rs,#u8)
1682 def: Pat<(mul I32:$Rs, n8_0ImmPred:$n8),
1683 (M2_mpysin I32:$Rs, (NegImm8 imm:$n8))>;
1685 def: Pat<(add Sext64:$Rs, I64:$Rt),
1686 (A2_addsp (LoReg Sext64:$Rs), I64:$Rt)>;
1688 def: AccRRR_pat<M4_and_and, And, Su_ni1<And>, I32, I32, I32>;
1689 def: AccRRR_pat<M4_and_or, And, Su_ni1<Or>, I32, I32, I32>;
1690 def: AccRRR_pat<M4_and_xor, And, Su<Xor>, I32, I32, I32>;
1691 def: AccRRR_pat<M4_or_and, Or, Su_ni1<And>, I32, I32, I32>;
1692 def: AccRRR_pat<M4_or_or, Or, Su_ni1<Or>, I32, I32, I32>;
1693 def: AccRRR_pat<M4_or_xor, Or, Su<Xor>, I32, I32, I32>;
1694 def: AccRRR_pat<M4_xor_and, Xor, Su_ni1<And>, I32, I32, I32>;
1695 def: AccRRR_pat<M4_xor_or, Xor, Su_ni1<Or>, I32, I32, I32>;
1696 def: AccRRR_pat<M2_xor_xacc, Xor, Su<Xor>, I32, I32, I32>;
1697 def: AccRRR_pat<M4_xor_xacc, Xor, Su<Xor>, I64, I64, I64>;
1699 // For dags like (or (and (not _), _), (shl _, _)) where the "or" with
1700 // one argument matches the patterns below, and with the other argument
1701 // matches S2_asl_r_r_or, etc, prefer the patterns below.
1702 let AddedComplexity = 110 in { // greater than S2_asl_r_r_and/or/xor.
1703 def: AccRRR_pat<M4_and_andn, And, Su<Not2<And>>, I32, I32, I32>;
1704 def: AccRRR_pat<M4_or_andn, Or, Su<Not2<And>>, I32, I32, I32>;
1705 def: AccRRR_pat<M4_xor_andn, Xor, Su<Not2<And>>, I32, I32, I32>;
1708 // S4_addaddi and S4_subaddi don't have tied operands, so give them
1709 // a bit of preference.
1710 let AddedComplexity = 30, Predicates = [UseCompound] in {
1711 def: Pat<(add I32:$Rs, (Su<Add> I32:$Ru, anyimm:$s6)),
1712 (S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
1713 def: Pat<(add anyimm:$s6, (Su<Add> I32:$Rs, I32:$Ru)),
1714 (S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
1715 def: Pat<(add I32:$Rs, (Su<Sub> anyimm:$s6, I32:$Ru)),
1716 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1717 def: Pat<(sub (Su<Add> I32:$Rs, anyimm:$s6), I32:$Ru),
1718 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1719 def: Pat<(add (Su<Sub> I32:$Rs, I32:$Ru), anyimm:$s6),
1720 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1723 let Predicates = [UseCompound] in
1724 def: Pat<(or I32:$Ru, (Su<And> I32:$Rx, anyimm:$s10)),
1725 (S4_or_andix IntRegs:$Ru, IntRegs:$Rx, imm:$s10)>;
1727 def: Pat<(or I32:$Rx, (Su<And> I32:$Rs, anyimm:$s10)),
1728 (S4_or_andi IntRegs:$Rx, IntRegs:$Rs, imm:$s10)>;
1729 def: Pat<(or I32:$Rx, (Su<Or> I32:$Rs, anyimm:$s10)),
1730 (S4_or_ori IntRegs:$Rx, IntRegs:$Rs, imm:$s10)>;
1733 def: Pat<(i32 (trunc (sra (Su<Mul> Sext64:$Rs, Sext64:$Rt), (i32 32)))),
1734 (M2_mpy_up (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1735 def: Pat<(i32 (trunc (srl (Su<Mul> Sext64:$Rs, Sext64:$Rt), (i32 32)))),
1736 (M2_mpy_up (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1738 def: Pat<(mul (Zext64 I32:$Rs), (Zext64 I32:$Rt)),
1739 (M2_dpmpyuu_s0 I32:$Rs, I32:$Rt)>;
1740 def: Pat<(mul (Aext64 I32:$Rs), (Aext64 I32:$Rt)),
1741 (M2_dpmpyuu_s0 I32:$Rs, I32:$Rt)>;
1742 def: Pat<(mul Sext64:$Rs, Sext64:$Rt),
1743 (M2_dpmpyss_s0 (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1745 def: Pat<(add I64:$Rx, (Su<Mul> Sext64:$Rs, Sext64:$Rt)),
1746 (M2_dpmpyss_acc_s0 I64:$Rx, (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1747 def: Pat<(sub I64:$Rx, (Su<Mul> Sext64:$Rs, Sext64:$Rt)),
1748 (M2_dpmpyss_nac_s0 I64:$Rx, (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1749 def: Pat<(add I64:$Rx, (Su<Mul> (Aext64 I32:$Rs), (Aext64 I32:$Rt))),
1750 (M2_dpmpyuu_acc_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1751 def: Pat<(add I64:$Rx, (Su<Mul> (Zext64 I32:$Rs), (Zext64 I32:$Rt))),
1752 (M2_dpmpyuu_acc_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1753 def: Pat<(sub I64:$Rx, (Su<Mul> (Aext64 I32:$Rs), (Aext64 I32:$Rt))),
1754 (M2_dpmpyuu_nac_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1755 def: Pat<(sub I64:$Rx, (Su<Mul> (Zext64 I32:$Rs), (Zext64 I32:$Rt))),
1756 (M2_dpmpyuu_nac_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1759 def: Pat<(sext_inreg (add I32:$Rt, I32:$Rs), i16),
1760 (A2_addh_l16_ll I32:$Rt, I32:$Rs)>;
1761 def: Pat<(sra (add (shl I32:$Rt, (i32 16)), I32:$Rs), (i32 16)),
1762 (A2_addh_l16_hl I32:$Rt, I32:$Rs)>;
1763 def: Pat<(shl (add I32:$Rt, I32:$Rs), (i32 16)),
1764 (A2_addh_h16_ll I32:$Rt, I32:$Rs)>;
1766 // Subtract halfword.
1767 def: Pat<(sext_inreg (sub I32:$Rt, I32:$Rs), i16),
1768 (A2_subh_l16_ll I32:$Rt, I32:$Rs)>;
1769 def: Pat<(sra (add (shl I32:$Rt, (i32 16)), I32:$Rs), (i32 16)),
1770 (A2_addh_l16_hl I32:$Rt, I32:$Rs)>;
1771 def: Pat<(shl (sub I32:$Rt, I32:$Rs), (i32 16)),
1772 (A2_subh_h16_ll I32:$Rt, I32:$Rs)>;
1774 def: Pat<(mul I64:$Rss, I64:$Rtt),
1776 (M2_maci (M2_maci (HiReg (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt))),
1781 (i32 (LoReg (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt)))))>;
1783 def MulHU : OutPatFrag<(ops node:$Rss, node:$Rtt),
1789 (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt)), 32),
1792 (A4_combineir 0, (LoReg (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt))))),
1796 (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt)), 32))>;
1798 // Multiply 64-bit unsigned and use upper result.
1799 def : Pat <(mulhu I64:$Rss, I64:$Rtt), (MulHU $Rss, $Rtt)>;
1801 // Multiply 64-bit signed and use upper result.
1803 // For two signed 64-bit integers A and B, let A' and B' denote A and B
1804 // with the sign bit cleared. Then A = -2^63*s(A) + A', where s(A) is the
1805 // sign bit of A (and identically for B). With this notation, the signed
1806 // product A*B can be written as:
1807 // AB = (-2^63 s(A) + A') * (-2^63 s(B) + B')
1808 // = 2^126 s(A)s(B) - 2^63 [s(A)B'+s(B)A'] + A'B'
1809 // = 2^126 s(A)s(B) + 2^63 [s(A)B'+s(B)A'] + A'B' - 2*2^63 [s(A)B'+s(B)A']
1810 // = (unsigned product AB) - 2^64 [s(A)B'+s(B)A']
1812 // Clear the sign bit in a 64-bit register.
1813 def ClearSign : OutPatFrag<(ops node:$Rss),
1814 (Combinew (S2_clrbit_i (HiReg $Rss), 31), (i32 (LoReg $Rss)))>;
1816 def : Pat <(mulhs I64:$Rss, I64:$Rtt),
1820 (A2_andp (S2_asr_i_p $Rss, 63), (ClearSign $Rtt)),
1821 (A2_andp (S2_asr_i_p $Rtt, 63), (ClearSign $Rss))))>;
1823 // Prefer these instructions over M2_macsip/M2_macsin: the macsi* instructions
1824 // will put the immediate addend into a register, while these instructions will
1825 // use it directly. Such a construct does not appear in the middle of a gep,
1826 // where M2_macsip would be preferable.
1827 let AddedComplexity = 20, Predicates = [UseCompound] in {
1828 def: Pat<(add (Su<Mul> I32:$Rs, u6_0ImmPred:$U6), anyimm:$u6),
1829 (M4_mpyri_addi imm:$u6, IntRegs:$Rs, imm:$U6)>;
1830 def: Pat<(add (Su<Mul> I32:$Rs, I32:$Rt), anyimm:$u6),
1831 (M4_mpyrr_addi imm:$u6, IntRegs:$Rs, IntRegs:$Rt)>;
1834 // Keep these instructions less preferable to M2_macsip/M2_macsin.
1835 let Predicates = [UseCompound] in {
1836 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, u6_2ImmPred:$u6_2)),
1837 (M4_mpyri_addr_u2 IntRegs:$Ru, imm:$u6_2, IntRegs:$Rs)>;
1838 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, anyimm:$u6)),
1839 (M4_mpyri_addr IntRegs:$Ru, IntRegs:$Rs, imm:$u6)>;
1840 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Ry, I32:$Rs)),
1841 (M4_mpyrr_addr IntRegs:$Ru, IntRegs:$Ry, IntRegs:$Rs)>;
1844 def: Pat<(fma F32:$Rs, F32:$Rt, F32:$Rx),
1845 (F2_sffma F32:$Rx, F32:$Rs, F32:$Rt)>;
1846 def: Pat<(fma (fneg F32:$Rs), F32:$Rt, F32:$Rx),
1847 (F2_sffms F32:$Rx, F32:$Rs, F32:$Rt)>;
1849 def: Pat<(mul V2I32:$Rs, V2I32:$Rt),
1850 (PS_vmulw V2I32:$Rs, V2I32:$Rt)>;
1851 def: Pat<(add V2I32:$Rx, (mul V2I32:$Rs, V2I32:$Rt)),
1852 (PS_vmulw_acc V2I32:$Rx, V2I32:$Rs, V2I32:$Rt)>;
1854 // Add/subtract two v4i8: Hexagon does not have an insn for this one, so
1855 // we use the double add v8i8, and use only the low part of the result.
1856 def: Pat<(add V4I8:$Rs, V4I8:$Rt),
1857 (LoReg (A2_vaddub (ToAext64 $Rs), (ToAext64 $Rt)))>;
1858 def: Pat<(sub V4I8:$Rs, V4I8:$Rt),
1859 (LoReg (A2_vsubub (ToAext64 $Rs), (ToAext64 $Rt)))>;
1861 // Use M2_vmpy2s_s0 for half-word vector multiply. It multiplies two
1862 // half-words, and saturates the result to a 32-bit value, except the
1863 // saturation never happens (it can only occur with scaling).
1864 def: Pat<(v2i16 (mul V2I16:$Rs, V2I16:$Rt)),
1865 (LoReg (S2_vtrunewh (IMPLICIT_DEF),
1866 (M2_vmpy2s_s0 V2I16:$Rs, V2I16:$Rt)))>;
1867 def: Pat<(v4i16 (mul V4I16:$Rs, V4I16:$Rt)),
1868 (S2_vtrunewh (M2_vmpy2s_s0 (HiReg $Rs), (HiReg $Rt)),
1869 (M2_vmpy2s_s0 (LoReg $Rs), (LoReg $Rt)))>;
1871 // Multiplies two v4i8 vectors.
1872 def: Pat<(v4i8 (mul V4I8:$Rs, V4I8:$Rt)),
1873 (S2_vtrunehb (M5_vmpybuu V4I8:$Rs, V4I8:$Rt))>;
1875 // Multiplies two v8i8 vectors.
1876 def: Pat<(v8i8 (mul V8I8:$Rs, V8I8:$Rt)),
1877 (Combinew (S2_vtrunehb (M5_vmpybuu (HiReg $Rs), (HiReg $Rt))),
1878 (S2_vtrunehb (M5_vmpybuu (LoReg $Rs), (LoReg $Rt))))>;
1881 // --(10) Bit ------------------------------------------------------------
1884 // Count leading zeros.
1885 def: Pat<(i32 (ctlz I32:$Rs)), (S2_cl0 I32:$Rs)>;
1886 def: Pat<(i32 (trunc (ctlz I64:$Rss))), (S2_cl0p I64:$Rss)>;
1888 // Count trailing zeros.
1889 def: Pat<(i32 (cttz I32:$Rs)), (S2_ct0 I32:$Rs)>;
1890 def: Pat<(i32 (trunc (cttz I64:$Rss))), (S2_ct0p I64:$Rss)>;
1892 // Count leading ones.
1893 def: Pat<(i32 (ctlz (not I32:$Rs))), (S2_cl1 I32:$Rs)>;
1894 def: Pat<(i32 (trunc (ctlz (not I64:$Rss)))), (S2_cl1p I64:$Rss)>;
1896 // Count trailing ones.
1897 def: Pat<(i32 (cttz (not I32:$Rs))), (S2_ct1 I32:$Rs)>;
1898 def: Pat<(i32 (trunc (cttz (not I64:$Rss)))), (S2_ct1p I64:$Rss)>;
1900 // Define leading/trailing patterns that require zero-extensions to 64 bits.
1901 def: Pat<(i64 (ctlz I64:$Rss)), (ToZext64 (S2_cl0p I64:$Rss))>;
1902 def: Pat<(i64 (cttz I64:$Rss)), (ToZext64 (S2_ct0p I64:$Rss))>;
1903 def: Pat<(i64 (ctlz (not I64:$Rss))), (ToZext64 (S2_cl1p I64:$Rss))>;
1904 def: Pat<(i64 (cttz (not I64:$Rss))), (ToZext64 (S2_ct1p I64:$Rss))>;
1906 def: Pat<(i64 (ctpop I64:$Rss)), (ToZext64 (S5_popcountp I64:$Rss))>;
1907 def: Pat<(i32 (ctpop I32:$Rs)), (S5_popcountp (A4_combineir 0, I32:$Rs))>;
1909 def: Pat<(bitreverse I32:$Rs), (S2_brev I32:$Rs)>;
1910 def: Pat<(bitreverse I64:$Rss), (S2_brevp I64:$Rss)>;
1912 def: Pat<(bitreverse V4I8:$Rs), (A2_swiz (S2_brev $Rs))>;
1913 def: Pat<(bitreverse V8I8:$Rs), (Combinew (A2_swiz (LoReg (S2_brevp $Rs))),
1914 (A2_swiz (HiReg (S2_brevp $Rs))))>;
1915 def: Pat<(bitreverse V2I16:$Rs), (A2_combine_lh (S2_brev $Rs),
1917 def: Pat<(bitreverse V4I16:$Rs),
1918 (Combinew (A2_combine_lh (LoReg (S2_brevp $Rs)),
1919 (LoReg (S2_brevp $Rs))),
1920 (A2_combine_lh (HiReg (S2_brevp $Rs)),
1921 (HiReg (S2_brevp $Rs))))>;
1922 def: Pat<(bitreverse V2I32:$Rs),
1923 (Combinew (i32 (LoReg (S2_brevp $Rs))),
1924 (i32 (HiReg (S2_brevp $Rs))))>;
1926 let AddedComplexity = 20 in { // Complexity greater than and/or/xor
1927 def: Pat<(and I32:$Rs, IsNPow2_32:$V),
1928 (S2_clrbit_i IntRegs:$Rs, (LogN2_32 $V))>;
1929 def: Pat<(or I32:$Rs, IsPow2_32:$V),
1930 (S2_setbit_i IntRegs:$Rs, (Log2_32 $V))>;
1931 def: Pat<(xor I32:$Rs, IsPow2_32:$V),
1932 (S2_togglebit_i IntRegs:$Rs, (Log2_32 $V))>;
1934 def: Pat<(and I32:$Rs, (not (shl 1, I32:$Rt))),
1935 (S2_clrbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1936 def: Pat<(or I32:$Rs, (shl 1, I32:$Rt)),
1937 (S2_setbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1938 def: Pat<(xor I32:$Rs, (shl 1, I32:$Rt)),
1939 (S2_togglebit_r IntRegs:$Rs, IntRegs:$Rt)>;
1942 // Clr/set/toggle bit for 64-bit values with immediate bit index.
1943 let AddedComplexity = 20 in { // Complexity greater than and/or/xor
1944 def: Pat<(and I64:$Rss, IsNPow2_64L:$V),
1945 (Combinew (i32 (HiReg $Rss)),
1946 (S2_clrbit_i (LoReg $Rss), (LogN2_64 $V)))>;
1947 def: Pat<(and I64:$Rss, IsNPow2_64H:$V),
1948 (Combinew (S2_clrbit_i (HiReg $Rss), (UDEC32 (i32 (LogN2_64 $V)))),
1949 (i32 (LoReg $Rss)))>;
1951 def: Pat<(or I64:$Rss, IsPow2_64L:$V),
1952 (Combinew (i32 (HiReg $Rss)),
1953 (S2_setbit_i (LoReg $Rss), (Log2_64 $V)))>;
1954 def: Pat<(or I64:$Rss, IsPow2_64H:$V),
1955 (Combinew (S2_setbit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
1956 (i32 (LoReg $Rss)))>;
1958 def: Pat<(xor I64:$Rss, IsPow2_64L:$V),
1959 (Combinew (i32 (HiReg $Rss)),
1960 (S2_togglebit_i (LoReg $Rss), (Log2_64 $V)))>;
1961 def: Pat<(xor I64:$Rss, IsPow2_64H:$V),
1962 (Combinew (S2_togglebit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
1963 (i32 (LoReg $Rss)))>;
1967 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
1968 def: Pat<(i1 (setne (and (shl 1, u5_0ImmPred:$u5), I32:$Rs), 0)),
1969 (S2_tstbit_i IntRegs:$Rs, imm:$u5)>;
1970 def: Pat<(i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1971 (S2_tstbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1972 def: Pat<(i1 (trunc I32:$Rs)),
1973 (S2_tstbit_i IntRegs:$Rs, 0)>;
1974 def: Pat<(i1 (trunc I64:$Rs)),
1975 (S2_tstbit_i (LoReg DoubleRegs:$Rs), 0)>;
1978 def: Pat<(and (srl I32:$Rs, u5_0ImmPred:$u5), 1),
1979 (I1toI32 (S2_tstbit_i I32:$Rs, imm:$u5))>;
1980 def: Pat<(and (srl I64:$Rss, IsULE<32,31>:$u6), 1),
1981 (ToZext64 (I1toI32 (S2_tstbit_i (LoReg $Rss), imm:$u6)))>;
1982 def: Pat<(and (srl I64:$Rss, IsUGT<32,31>:$u6), 1),
1983 (ToZext64 (I1toI32 (S2_tstbit_i (HiReg $Rss), (UDEC32 $u6))))>;
1985 def: Pat<(and (not (srl I32:$Rs, u5_0ImmPred:$u5)), 1),
1986 (I1toI32 (S4_ntstbit_i I32:$Rs, imm:$u5))>;
1987 def: Pat<(and (not (srl I64:$Rss, IsULE<32,31>:$u6)), 1),
1988 (ToZext64 (I1toI32 (S4_ntstbit_i (LoReg $Rss), imm:$u6)))>;
1989 def: Pat<(and (not (srl I64:$Rss, IsUGT<32,31>:$u6)), 1),
1990 (ToZext64 (I1toI32 (S4_ntstbit_i (HiReg $Rss), (UDEC32 $u6))))>;
1992 let AddedComplexity = 20 in { // Complexity greater than compare reg-imm.
1993 def: Pat<(i1 (seteq (and I32:$Rs, u6_0ImmPred:$u6), 0)),
1994 (C2_bitsclri IntRegs:$Rs, imm:$u6)>;
1995 def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), 0)),
1996 (C2_bitsclr IntRegs:$Rs, IntRegs:$Rt)>;
1999 let AddedComplexity = 10 in // Complexity greater than compare reg-reg.
2000 def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), IntRegs:$Rt)),
2001 (C2_bitsset IntRegs:$Rs, IntRegs:$Rt)>;
2004 SDTypeProfile<1, 2, [SDTCisVT<0, i1>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
2005 def HexagonTSTBIT: SDNode<"HexagonISD::TSTBIT", SDTTestBit>;
2007 def: Pat<(HexagonTSTBIT I32:$Rs, u5_0ImmPred:$u5),
2008 (S2_tstbit_i I32:$Rs, imm:$u5)>;
2009 def: Pat<(HexagonTSTBIT I32:$Rs, I32:$Rt),
2010 (S2_tstbit_r I32:$Rs, I32:$Rt)>;
2012 // Add extra complexity to prefer these instructions over bitsset/bitsclr.
2013 // The reason is that tstbit/ntstbit can be folded into a compound instruction:
2014 // if ([!]tstbit(...)) jump ...
2015 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
2016 def: Pat<(i1 (seteq (and I32:$Rs, IsPow2_32:$u5), 0)),
2017 (S4_ntstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
2018 def: Pat<(i1 (setne (and I32:$Rs, IsPow2_32:$u5), 0)),
2019 (S2_tstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
2020 def: Pat<(i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)),
2021 (S4_ntstbit_r I32:$Rs, I32:$Rt)>;
2022 def: Pat<(i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)),
2023 (S2_tstbit_r I32:$Rs, I32:$Rt)>;
2026 def: Pat<(i1 (seteq (and I64:$Rs, IsPow2_64L:$u6), 0)),
2027 (S4_ntstbit_i (LoReg $Rs), (Log2_64 $u6))>;
2028 def: Pat<(i1 (seteq (and I64:$Rs, IsPow2_64H:$u6), 0)),
2029 (S4_ntstbit_i (HiReg $Rs), (UDEC32 (i32 (Log2_64 $u6))))>;
2030 def: Pat<(i1 (setne (and I64:$Rs, IsPow2_64L:$u6), 0)),
2031 (S2_tstbit_i (LoReg $Rs), (Log2_64 imm:$u6))>;
2032 def: Pat<(i1 (setne (and I64:$Rs, IsPow2_64H:$u6), 0)),
2033 (S2_tstbit_i (HiReg $Rs), (UDEC32 (i32 (Log2_64 imm:$u6))))>;
2035 // Do not increase complexity of these patterns. In the DAG, "cmp i8" may be
2036 // represented as a compare against "value & 0xFF", which is an exact match
2037 // for cmpb (same for cmph). The patterns below do not contain any additional
2038 // complexity that would make them preferable, and if they were actually used
2039 // instead of cmpb/cmph, they would result in a compare against register that
2040 // is loaded with the byte/half mask (i.e. 0xFF or 0xFFFF).
2041 def: Pat<(i1 (setne (and I32:$Rs, u6_0ImmPred:$u6), 0)),
2042 (C4_nbitsclri I32:$Rs, imm:$u6)>;
2043 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), 0)),
2044 (C4_nbitsclr I32:$Rs, I32:$Rt)>;
2045 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), I32:$Rt)),
2046 (C4_nbitsset I32:$Rs, I32:$Rt)>;
2048 // Special patterns to address certain cases where the "top-down" matching
2049 // algorithm would cause suboptimal selection.
2051 let AddedComplexity = 100 in {
2052 // Avoid A4_rcmp[n]eqi in these cases:
2053 def: Pat<(i32 (zext (i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
2054 (I1toI32 (S4_ntstbit_r IntRegs:$Rs, IntRegs:$Rt))>;
2055 def: Pat<(i32 (zext (i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
2056 (I1toI32 (S2_tstbit_r IntRegs:$Rs, IntRegs:$Rt))>;
2057 def: Pat<(i32 (zext (i1 (seteq (and I32:$Rs, IsPow2_32:$u5), 0)))),
2058 (I1toI32 (S4_ntstbit_i I32:$Rs, (Log2_32 imm:$u5)))>;
2059 def: Pat<(i32 (zext (i1 (setne (and I32:$Rs, IsPow2_32:$u5), 0)))),
2060 (I1toI32 (S2_tstbit_i I32:$Rs, (Log2_32 imm:$u5)))>;
2061 def: Pat<(i32 (zext (i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
2062 (I1toI32 (S4_ntstbit_r I32:$Rs, I32:$Rt))>;
2063 def: Pat<(i32 (zext (i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
2064 (I1toI32 (S2_tstbit_r I32:$Rs, I32:$Rt))>;
2067 // --(11) PIC ------------------------------------------------------------
2070 def SDT_HexagonAtGot
2071 : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
2072 def SDT_HexagonAtPcrel
2073 : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
2075 // AT_GOT address-of-GOT, address-of-global, offset-in-global
2076 def HexagonAtGot : SDNode<"HexagonISD::AT_GOT", SDT_HexagonAtGot>;
2077 // AT_PCREL address-of-global
2078 def HexagonAtPcrel : SDNode<"HexagonISD::AT_PCREL", SDT_HexagonAtPcrel>;
2080 def: Pat<(HexagonAtGot I32:$got, I32:$addr, (i32 0)),
2081 (L2_loadri_io I32:$got, imm:$addr)>;
2082 def: Pat<(HexagonAtGot I32:$got, I32:$addr, s30_2ImmPred:$off),
2083 (A2_addi (L2_loadri_io I32:$got, imm:$addr), imm:$off)>;
2084 def: Pat<(HexagonAtPcrel I32:$addr),
2085 (C4_addipc imm:$addr)>;
2087 // The HVX load patterns also match AT_PCREL directly. Make sure that
2088 // if the selection of this opcode changes, it's updated in all places.
2091 // --(12) Load -----------------------------------------------------------
2094 def L1toI32: OutPatFrag<(ops node:$Rs), (A2_subri 0, (i32 $Rs))>;
2095 def L1toI64: OutPatFrag<(ops node:$Rs), (ToSext64 (L1toI32 $Rs))>;
2097 def extloadv2i8: PatFrag<(ops node:$ptr), (extload node:$ptr), [{
2098 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
2100 def extloadv4i8: PatFrag<(ops node:$ptr), (extload node:$ptr), [{
2101 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
2104 def zextloadv2i8: PatFrag<(ops node:$ptr), (zextload node:$ptr), [{
2105 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
2107 def zextloadv4i8: PatFrag<(ops node:$ptr), (zextload node:$ptr), [{
2108 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
2111 def sextloadv2i8: PatFrag<(ops node:$ptr), (sextload node:$ptr), [{
2112 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
2114 def sextloadv4i8: PatFrag<(ops node:$ptr), (sextload node:$ptr), [{
2115 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
2118 // Patterns to select load-indexed: Rs + Off.
2119 // - frameindex [+ imm],
2120 multiclass Loadxfi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
2122 def: Pat<(VT (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
2123 (VT (MI AddrFI:$fi, imm:$Off))>;
2124 def: Pat<(VT (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
2125 (VT (MI AddrFI:$fi, imm:$Off))>;
2126 def: Pat<(VT (Load AddrFI:$fi)), (VT (MI AddrFI:$fi, 0))>;
2129 // Patterns to select load-indexed: Rs + Off.
2130 // - base reg [+ imm]
2131 multiclass Loadxgi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
2133 def: Pat<(VT (Load (add I32:$Rs, ImmPred:$Off))),
2134 (VT (MI IntRegs:$Rs, imm:$Off))>;
2135 def: Pat<(VT (Load (IsOrAdd I32:$Rs, ImmPred:$Off))),
2136 (VT (MI IntRegs:$Rs, imm:$Off))>;
2137 def: Pat<(VT (Load I32:$Rs)), (VT (MI IntRegs:$Rs, 0))>;
2140 // Patterns to select load-indexed: Rs + Off. Combines Loadxfi + Loadxgi.
2141 multiclass Loadxi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
2143 defm: Loadxfi_pat<Load, VT, ImmPred, MI>;
2144 defm: Loadxgi_pat<Load, VT, ImmPred, MI>;
2147 // Patterns to select load reg indexed: Rs + Off with a value modifier.
2148 // - frameindex [+ imm]
2149 multiclass Loadxfim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2150 PatLeaf ImmPred, InstHexagon MI> {
2151 def: Pat<(VT (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
2152 (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>;
2153 def: Pat<(VT (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
2154 (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>;
2155 def: Pat<(VT (Load AddrFI:$fi)), (VT (ValueMod (MI AddrFI:$fi, 0)))>;
2158 // Patterns to select load reg indexed: Rs + Off with a value modifier.
2159 // - base reg [+ imm]
2160 multiclass Loadxgim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2161 PatLeaf ImmPred, InstHexagon MI> {
2162 def: Pat<(VT (Load (add I32:$Rs, ImmPred:$Off))),
2163 (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
2164 def: Pat<(VT (Load (IsOrAdd I32:$Rs, ImmPred:$Off))),
2165 (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
2166 def: Pat<(VT (Load I32:$Rs)), (VT (ValueMod (MI IntRegs:$Rs, 0)))>;
2169 // Patterns to select load reg indexed: Rs + Off with a value modifier.
2170 // Combines Loadxfim + Loadxgim.
2171 multiclass Loadxim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2172 PatLeaf ImmPred, InstHexagon MI> {
2173 defm: Loadxfim_pat<Load, VT, ValueMod, ImmPred, MI>;
2174 defm: Loadxgim_pat<Load, VT, ValueMod, ImmPred, MI>;
2177 // Pattern to select load reg reg-indexed: Rs + Rt<<u2.
2178 class Loadxr_shl_pat<PatFrag Load, ValueType VT, InstHexagon MI>
2179 : Pat<(VT (Load (add I32:$Rs, (i32 (shl I32:$Rt, u2_0ImmPred:$u2))))),
2180 (VT (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2))>;
2182 // Pattern to select load reg reg-indexed: Rs + Rt<<0.
2183 class Loadxr_add_pat<PatFrag Load, ValueType VT, InstHexagon MI>
2184 : Pat<(VT (Load (add I32:$Rs, I32:$Rt))),
2185 (VT (MI IntRegs:$Rs, IntRegs:$Rt, 0))>;
2187 // Pattern to select load reg reg-indexed: Rs + Rt<<u2 with value modifier.
2188 class Loadxrm_shl_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2190 : Pat<(VT (Load (add I32:$Rs, (i32 (shl I32:$Rt, u2_0ImmPred:$u2))))),
2191 (VT (ValueMod (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2)))>;
2193 // Pattern to select load reg reg-indexed: Rs + Rt<<0 with value modifier.
2194 class Loadxrm_add_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2196 : Pat<(VT (Load (add I32:$Rs, I32:$Rt))),
2197 (VT (ValueMod (MI IntRegs:$Rs, IntRegs:$Rt, 0)))>;
2199 // Pattern to select load long-offset reg-indexed: Addr + Rt<<u2.
2200 // Don't match for u2==0, instead use reg+imm for those cases.
2201 class Loadxu_pat<PatFrag Load, ValueType VT, PatFrag ImmPred, InstHexagon MI>
2202 : Pat<(VT (Load (add (shl IntRegs:$Rt, u2_0ImmPred:$u2), ImmPred:$Addr))),
2203 (VT (MI IntRegs:$Rt, imm:$u2, ImmPred:$Addr))>;
2205 class Loadxum_pat<PatFrag Load, ValueType VT, PatFrag ImmPred, PatFrag ValueMod,
2207 : Pat<(VT (Load (add (shl IntRegs:$Rt, u2_0ImmPred:$u2), ImmPred:$Addr))),
2208 (VT (ValueMod (MI IntRegs:$Rt, imm:$u2, ImmPred:$Addr)))>;
2210 // Pattern to select load absolute.
2211 class Loada_pat<PatFrag Load, ValueType VT, PatFrag Addr, InstHexagon MI>
2212 : Pat<(VT (Load Addr:$addr)), (MI Addr:$addr)>;
2214 // Pattern to select load absolute with value modifier.
2215 class Loadam_pat<PatFrag Load, ValueType VT, PatFrag Addr, PatFrag ValueMod,
2217 : Pat<(VT (Load Addr:$addr)), (ValueMod (MI Addr:$addr))>;
2220 let AddedComplexity = 20 in {
2221 defm: Loadxi_pat<extloadi1, i32, anyimm0, L2_loadrub_io>;
2222 defm: Loadxi_pat<extloadi8, i32, anyimm0, L2_loadrub_io>;
2223 defm: Loadxi_pat<extloadi16, i32, anyimm1, L2_loadruh_io>;
2224 defm: Loadxi_pat<extloadv2i8, v2i16, anyimm1, L2_loadbzw2_io>;
2225 defm: Loadxi_pat<extloadv4i8, v4i16, anyimm2, L2_loadbzw4_io>;
2226 defm: Loadxi_pat<sextloadi8, i32, anyimm0, L2_loadrb_io>;
2227 defm: Loadxi_pat<sextloadi16, i32, anyimm1, L2_loadrh_io>;
2228 defm: Loadxi_pat<sextloadv2i8, v2i16, anyimm1, L2_loadbsw2_io>;
2229 defm: Loadxi_pat<sextloadv4i8, v4i16, anyimm2, L2_loadbsw4_io>;
2230 defm: Loadxi_pat<zextloadi1, i32, anyimm0, L2_loadrub_io>;
2231 defm: Loadxi_pat<zextloadi8, i32, anyimm0, L2_loadrub_io>;
2232 defm: Loadxi_pat<zextloadi16, i32, anyimm1, L2_loadruh_io>;
2233 defm: Loadxi_pat<zextloadv2i8, v2i16, anyimm1, L2_loadbzw2_io>;
2234 defm: Loadxi_pat<zextloadv4i8, v4i16, anyimm2, L2_loadbzw4_io>;
2235 defm: Loadxi_pat<load, i32, anyimm2, L2_loadri_io>;
2236 defm: Loadxi_pat<load, v2i16, anyimm2, L2_loadri_io>;
2237 defm: Loadxi_pat<load, v4i8, anyimm2, L2_loadri_io>;
2238 defm: Loadxi_pat<load, i64, anyimm3, L2_loadrd_io>;
2239 defm: Loadxi_pat<load, v2i32, anyimm3, L2_loadrd_io>;
2240 defm: Loadxi_pat<load, v4i16, anyimm3, L2_loadrd_io>;
2241 defm: Loadxi_pat<load, v8i8, anyimm3, L2_loadrd_io>;
2242 defm: Loadxi_pat<load, f32, anyimm2, L2_loadri_io>;
2243 defm: Loadxi_pat<load, f64, anyimm3, L2_loadrd_io>;
2246 defm: Loadxi_pat<atomic_load_8 , i32, anyimm0, L2_loadrub_io>;
2247 defm: Loadxi_pat<atomic_load_16, i32, anyimm1, L2_loadruh_io>;
2248 defm: Loadxi_pat<atomic_load_32, i32, anyimm2, L2_loadri_io>;
2249 defm: Loadxi_pat<atomic_load_64, i64, anyimm3, L2_loadrd_io>;
2252 let AddedComplexity = 30 in {
2253 // Loads of i1 are loading a byte, and the byte should be either 0 or 1.
2254 // It doesn't matter if it's sign- or zero-extended, so use zero-extension
2256 defm: Loadxim_pat<sextloadi1, i32, L1toI32, anyimm0, L2_loadrub_io>;
2257 defm: Loadxim_pat<extloadi1, i64, ToAext64, anyimm0, L2_loadrub_io>;
2258 defm: Loadxim_pat<sextloadi1, i64, L1toI64, anyimm0, L2_loadrub_io>;
2259 defm: Loadxim_pat<zextloadi1, i64, ToZext64, anyimm0, L2_loadrub_io>;
2261 defm: Loadxim_pat<extloadi8, i64, ToAext64, anyimm0, L2_loadrub_io>;
2262 defm: Loadxim_pat<extloadi16, i64, ToAext64, anyimm1, L2_loadruh_io>;
2263 defm: Loadxim_pat<extloadi32, i64, ToAext64, anyimm2, L2_loadri_io>;
2264 defm: Loadxim_pat<zextloadi8, i64, ToZext64, anyimm0, L2_loadrub_io>;
2265 defm: Loadxim_pat<zextloadi16, i64, ToZext64, anyimm1, L2_loadruh_io>;
2266 defm: Loadxim_pat<zextloadi32, i64, ToZext64, anyimm2, L2_loadri_io>;
2267 defm: Loadxim_pat<sextloadi8, i64, ToSext64, anyimm0, L2_loadrb_io>;
2268 defm: Loadxim_pat<sextloadi16, i64, ToSext64, anyimm1, L2_loadrh_io>;
2269 defm: Loadxim_pat<sextloadi32, i64, ToSext64, anyimm2, L2_loadri_io>;
2272 let AddedComplexity = 60 in {
2273 def: Loadxu_pat<extloadi1, i32, anyimm0, L4_loadrub_ur>;
2274 def: Loadxu_pat<extloadi8, i32, anyimm0, L4_loadrub_ur>;
2275 def: Loadxu_pat<extloadi16, i32, anyimm1, L4_loadruh_ur>;
2276 def: Loadxu_pat<extloadv2i8, v2i16, anyimm1, L4_loadbzw2_ur>;
2277 def: Loadxu_pat<extloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
2278 def: Loadxu_pat<sextloadi8, i32, anyimm0, L4_loadrb_ur>;
2279 def: Loadxu_pat<sextloadi16, i32, anyimm1, L4_loadrh_ur>;
2280 def: Loadxu_pat<sextloadv2i8, v2i16, anyimm1, L4_loadbsw2_ur>;
2281 def: Loadxu_pat<sextloadv4i8, v4i16, anyimm2, L4_loadbsw4_ur>;
2282 def: Loadxu_pat<zextloadi1, i32, anyimm0, L4_loadrub_ur>;
2283 def: Loadxu_pat<zextloadi8, i32, anyimm0, L4_loadrub_ur>;
2284 def: Loadxu_pat<zextloadi16, i32, anyimm1, L4_loadruh_ur>;
2285 def: Loadxu_pat<zextloadv2i8, v2i16, anyimm1, L4_loadbzw2_ur>;
2286 def: Loadxu_pat<zextloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
2287 def: Loadxu_pat<load, i32, anyimm2, L4_loadri_ur>;
2288 def: Loadxu_pat<load, v2i16, anyimm2, L4_loadri_ur>;
2289 def: Loadxu_pat<load, v4i8, anyimm2, L4_loadri_ur>;
2290 def: Loadxu_pat<load, i64, anyimm3, L4_loadrd_ur>;
2291 def: Loadxu_pat<load, v2i32, anyimm3, L4_loadrd_ur>;
2292 def: Loadxu_pat<load, v4i16, anyimm3, L4_loadrd_ur>;
2293 def: Loadxu_pat<load, v8i8, anyimm3, L4_loadrd_ur>;
2294 def: Loadxu_pat<load, f32, anyimm2, L4_loadri_ur>;
2295 def: Loadxu_pat<load, f64, anyimm3, L4_loadrd_ur>;
2297 def: Loadxum_pat<sextloadi1, i32, anyimm0, L1toI32, L4_loadrub_ur>;
2298 def: Loadxum_pat<extloadi1, i64, anyimm0, ToAext64, L4_loadrub_ur>;
2299 def: Loadxum_pat<sextloadi1, i64, anyimm0, L1toI64, L4_loadrub_ur>;
2300 def: Loadxum_pat<zextloadi1, i64, anyimm0, ToZext64, L4_loadrub_ur>;
2302 def: Loadxum_pat<sextloadi8, i64, anyimm0, ToSext64, L4_loadrb_ur>;
2303 def: Loadxum_pat<zextloadi8, i64, anyimm0, ToZext64, L4_loadrub_ur>;
2304 def: Loadxum_pat<extloadi8, i64, anyimm0, ToAext64, L4_loadrub_ur>;
2305 def: Loadxum_pat<sextloadi16, i64, anyimm1, ToSext64, L4_loadrh_ur>;
2306 def: Loadxum_pat<zextloadi16, i64, anyimm1, ToZext64, L4_loadruh_ur>;
2307 def: Loadxum_pat<extloadi16, i64, anyimm1, ToAext64, L4_loadruh_ur>;
2308 def: Loadxum_pat<sextloadi32, i64, anyimm2, ToSext64, L4_loadri_ur>;
2309 def: Loadxum_pat<zextloadi32, i64, anyimm2, ToZext64, L4_loadri_ur>;
2310 def: Loadxum_pat<extloadi32, i64, anyimm2, ToAext64, L4_loadri_ur>;
2313 let AddedComplexity = 40 in {
2314 def: Loadxr_shl_pat<extloadi1, i32, L4_loadrub_rr>;
2315 def: Loadxr_shl_pat<extloadi8, i32, L4_loadrub_rr>;
2316 def: Loadxr_shl_pat<zextloadi1, i32, L4_loadrub_rr>;
2317 def: Loadxr_shl_pat<zextloadi8, i32, L4_loadrub_rr>;
2318 def: Loadxr_shl_pat<sextloadi8, i32, L4_loadrb_rr>;
2319 def: Loadxr_shl_pat<extloadi16, i32, L4_loadruh_rr>;
2320 def: Loadxr_shl_pat<zextloadi16, i32, L4_loadruh_rr>;
2321 def: Loadxr_shl_pat<sextloadi16, i32, L4_loadrh_rr>;
2322 def: Loadxr_shl_pat<load, i32, L4_loadri_rr>;
2323 def: Loadxr_shl_pat<load, v2i16, L4_loadri_rr>;
2324 def: Loadxr_shl_pat<load, v4i8, L4_loadri_rr>;
2325 def: Loadxr_shl_pat<load, i64, L4_loadrd_rr>;
2326 def: Loadxr_shl_pat<load, v2i32, L4_loadrd_rr>;
2327 def: Loadxr_shl_pat<load, v4i16, L4_loadrd_rr>;
2328 def: Loadxr_shl_pat<load, v8i8, L4_loadrd_rr>;
2329 def: Loadxr_shl_pat<load, f32, L4_loadri_rr>;
2330 def: Loadxr_shl_pat<load, f64, L4_loadrd_rr>;
2333 let AddedComplexity = 20 in {
2334 def: Loadxr_add_pat<extloadi1, i32, L4_loadrub_rr>;
2335 def: Loadxr_add_pat<extloadi8, i32, L4_loadrub_rr>;
2336 def: Loadxr_add_pat<zextloadi8, i32, L4_loadrub_rr>;
2337 def: Loadxr_add_pat<zextloadi1, i32, L4_loadrub_rr>;
2338 def: Loadxr_add_pat<sextloadi8, i32, L4_loadrb_rr>;
2339 def: Loadxr_add_pat<extloadi16, i32, L4_loadruh_rr>;
2340 def: Loadxr_add_pat<zextloadi16, i32, L4_loadruh_rr>;
2341 def: Loadxr_add_pat<sextloadi16, i32, L4_loadrh_rr>;
2342 def: Loadxr_add_pat<load, i32, L4_loadri_rr>;
2343 def: Loadxr_add_pat<load, v2i16, L4_loadri_rr>;
2344 def: Loadxr_add_pat<load, v4i8, L4_loadri_rr>;
2345 def: Loadxr_add_pat<load, i64, L4_loadrd_rr>;
2346 def: Loadxr_add_pat<load, v2i32, L4_loadrd_rr>;
2347 def: Loadxr_add_pat<load, v4i16, L4_loadrd_rr>;
2348 def: Loadxr_add_pat<load, v8i8, L4_loadrd_rr>;
2349 def: Loadxr_add_pat<load, f32, L4_loadri_rr>;
2350 def: Loadxr_add_pat<load, f64, L4_loadrd_rr>;
2353 let AddedComplexity = 40 in {
2354 def: Loadxrm_shl_pat<sextloadi1, i32, L1toI32, L4_loadrub_rr>;
2355 def: Loadxrm_shl_pat<extloadi1, i64, ToAext64, L4_loadrub_rr>;
2356 def: Loadxrm_shl_pat<sextloadi1, i64, L1toI64, L4_loadrub_rr>;
2357 def: Loadxrm_shl_pat<zextloadi1, i64, ToZext64, L4_loadrub_rr>;
2359 def: Loadxrm_shl_pat<extloadi8, i64, ToAext64, L4_loadrub_rr>;
2360 def: Loadxrm_shl_pat<zextloadi8, i64, ToZext64, L4_loadrub_rr>;
2361 def: Loadxrm_shl_pat<sextloadi8, i64, ToSext64, L4_loadrb_rr>;
2362 def: Loadxrm_shl_pat<extloadi16, i64, ToAext64, L4_loadruh_rr>;
2363 def: Loadxrm_shl_pat<zextloadi16, i64, ToZext64, L4_loadruh_rr>;
2364 def: Loadxrm_shl_pat<sextloadi16, i64, ToSext64, L4_loadrh_rr>;
2365 def: Loadxrm_shl_pat<extloadi32, i64, ToAext64, L4_loadri_rr>;
2366 def: Loadxrm_shl_pat<zextloadi32, i64, ToZext64, L4_loadri_rr>;
2367 def: Loadxrm_shl_pat<sextloadi32, i64, ToSext64, L4_loadri_rr>;
2370 let AddedComplexity = 30 in {
2371 def: Loadxrm_add_pat<sextloadi1, i32, L1toI32, L4_loadrub_rr>;
2372 def: Loadxrm_add_pat<extloadi1, i64, ToAext64, L4_loadrub_rr>;
2373 def: Loadxrm_add_pat<sextloadi1, i64, L1toI64, L4_loadrub_rr>;
2374 def: Loadxrm_add_pat<zextloadi1, i64, ToZext64, L4_loadrub_rr>;
2376 def: Loadxrm_add_pat<extloadi8, i64, ToAext64, L4_loadrub_rr>;
2377 def: Loadxrm_add_pat<zextloadi8, i64, ToZext64, L4_loadrub_rr>;
2378 def: Loadxrm_add_pat<sextloadi8, i64, ToSext64, L4_loadrb_rr>;
2379 def: Loadxrm_add_pat<extloadi16, i64, ToAext64, L4_loadruh_rr>;
2380 def: Loadxrm_add_pat<zextloadi16, i64, ToZext64, L4_loadruh_rr>;
2381 def: Loadxrm_add_pat<sextloadi16, i64, ToSext64, L4_loadrh_rr>;
2382 def: Loadxrm_add_pat<extloadi32, i64, ToAext64, L4_loadri_rr>;
2383 def: Loadxrm_add_pat<zextloadi32, i64, ToZext64, L4_loadri_rr>;
2384 def: Loadxrm_add_pat<sextloadi32, i64, ToSext64, L4_loadri_rr>;
2389 let AddedComplexity = 60 in {
2390 def: Loada_pat<extloadi1, i32, anyimm0, PS_loadrubabs>;
2391 def: Loada_pat<zextloadi1, i32, anyimm0, PS_loadrubabs>;
2392 def: Loada_pat<extloadi8, i32, anyimm0, PS_loadrubabs>;
2393 def: Loada_pat<sextloadi8, i32, anyimm0, PS_loadrbabs>;
2394 def: Loada_pat<zextloadi8, i32, anyimm0, PS_loadrubabs>;
2395 def: Loada_pat<extloadi16, i32, anyimm1, PS_loadruhabs>;
2396 def: Loada_pat<sextloadi16, i32, anyimm1, PS_loadrhabs>;
2397 def: Loada_pat<zextloadi16, i32, anyimm1, PS_loadruhabs>;
2398 def: Loada_pat<load, i32, anyimm2, PS_loadriabs>;
2399 def: Loada_pat<load, v2i16, anyimm2, PS_loadriabs>;
2400 def: Loada_pat<load, v4i8, anyimm2, PS_loadriabs>;
2401 def: Loada_pat<load, i64, anyimm3, PS_loadrdabs>;
2402 def: Loada_pat<load, v2i32, anyimm3, PS_loadrdabs>;
2403 def: Loada_pat<load, v4i16, anyimm3, PS_loadrdabs>;
2404 def: Loada_pat<load, v8i8, anyimm3, PS_loadrdabs>;
2405 def: Loada_pat<load, f32, anyimm2, PS_loadriabs>;
2406 def: Loada_pat<load, f64, anyimm3, PS_loadrdabs>;
2408 def: Loada_pat<atomic_load_8, i32, anyimm0, PS_loadrubabs>;
2409 def: Loada_pat<atomic_load_16, i32, anyimm1, PS_loadruhabs>;
2410 def: Loada_pat<atomic_load_32, i32, anyimm2, PS_loadriabs>;
2411 def: Loada_pat<atomic_load_64, i64, anyimm3, PS_loadrdabs>;
2414 let AddedComplexity = 30 in {
2415 def: Loadam_pat<load, i1, anyimm0, I32toI1, PS_loadrubabs>;
2416 def: Loadam_pat<sextloadi1, i32, anyimm0, L1toI32, PS_loadrubabs>;
2417 def: Loadam_pat<extloadi1, i64, anyimm0, ToZext64, PS_loadrubabs>;
2418 def: Loadam_pat<sextloadi1, i64, anyimm0, L1toI64, PS_loadrubabs>;
2419 def: Loadam_pat<zextloadi1, i64, anyimm0, ToZext64, PS_loadrubabs>;
2421 def: Loadam_pat<extloadi8, i64, anyimm0, ToAext64, PS_loadrubabs>;
2422 def: Loadam_pat<sextloadi8, i64, anyimm0, ToSext64, PS_loadrbabs>;
2423 def: Loadam_pat<zextloadi8, i64, anyimm0, ToZext64, PS_loadrubabs>;
2424 def: Loadam_pat<extloadi16, i64, anyimm1, ToAext64, PS_loadruhabs>;
2425 def: Loadam_pat<sextloadi16, i64, anyimm1, ToSext64, PS_loadrhabs>;
2426 def: Loadam_pat<zextloadi16, i64, anyimm1, ToZext64, PS_loadruhabs>;
2427 def: Loadam_pat<extloadi32, i64, anyimm2, ToAext64, PS_loadriabs>;
2428 def: Loadam_pat<sextloadi32, i64, anyimm2, ToSext64, PS_loadriabs>;
2429 def: Loadam_pat<zextloadi32, i64, anyimm2, ToZext64, PS_loadriabs>;
2432 // GP-relative address
2434 let AddedComplexity = 100 in {
2435 def: Loada_pat<extloadi1, i32, addrgp, L2_loadrubgp>;
2436 def: Loada_pat<zextloadi1, i32, addrgp, L2_loadrubgp>;
2437 def: Loada_pat<extloadi8, i32, addrgp, L2_loadrubgp>;
2438 def: Loada_pat<sextloadi8, i32, addrgp, L2_loadrbgp>;
2439 def: Loada_pat<zextloadi8, i32, addrgp, L2_loadrubgp>;
2440 def: Loada_pat<extloadi16, i32, addrgp, L2_loadruhgp>;
2441 def: Loada_pat<sextloadi16, i32, addrgp, L2_loadrhgp>;
2442 def: Loada_pat<zextloadi16, i32, addrgp, L2_loadruhgp>;
2443 def: Loada_pat<load, i32, addrgp, L2_loadrigp>;
2444 def: Loada_pat<load, v2i16, addrgp, L2_loadrigp>;
2445 def: Loada_pat<load, v4i8, addrgp, L2_loadrigp>;
2446 def: Loada_pat<load, i64, addrgp, L2_loadrdgp>;
2447 def: Loada_pat<load, v2i32, addrgp, L2_loadrdgp>;
2448 def: Loada_pat<load, v4i16, addrgp, L2_loadrdgp>;
2449 def: Loada_pat<load, v8i8, addrgp, L2_loadrdgp>;
2450 def: Loada_pat<load, f32, addrgp, L2_loadrigp>;
2451 def: Loada_pat<load, f64, addrgp, L2_loadrdgp>;
2453 def: Loada_pat<atomic_load_8, i32, addrgp, L2_loadrubgp>;
2454 def: Loada_pat<atomic_load_16, i32, addrgp, L2_loadruhgp>;
2455 def: Loada_pat<atomic_load_32, i32, addrgp, L2_loadrigp>;
2456 def: Loada_pat<atomic_load_64, i64, addrgp, L2_loadrdgp>;
2459 let AddedComplexity = 70 in {
2460 def: Loadam_pat<sextloadi1, i32, addrgp, L1toI32, L2_loadrubgp>;
2461 def: Loadam_pat<extloadi1, i64, addrgp, ToAext64, L2_loadrubgp>;
2462 def: Loadam_pat<sextloadi1, i64, addrgp, L1toI64, L2_loadrubgp>;
2463 def: Loadam_pat<zextloadi1, i64, addrgp, ToZext64, L2_loadrubgp>;
2465 def: Loadam_pat<extloadi8, i64, addrgp, ToAext64, L2_loadrubgp>;
2466 def: Loadam_pat<sextloadi8, i64, addrgp, ToSext64, L2_loadrbgp>;
2467 def: Loadam_pat<zextloadi8, i64, addrgp, ToZext64, L2_loadrubgp>;
2468 def: Loadam_pat<extloadi16, i64, addrgp, ToAext64, L2_loadruhgp>;
2469 def: Loadam_pat<sextloadi16, i64, addrgp, ToSext64, L2_loadrhgp>;
2470 def: Loadam_pat<zextloadi16, i64, addrgp, ToZext64, L2_loadruhgp>;
2471 def: Loadam_pat<extloadi32, i64, addrgp, ToAext64, L2_loadrigp>;
2472 def: Loadam_pat<sextloadi32, i64, addrgp, ToSext64, L2_loadrigp>;
2473 def: Loadam_pat<zextloadi32, i64, addrgp, ToZext64, L2_loadrigp>;
2475 def: Loadam_pat<load, i1, addrgp, I32toI1, L2_loadrubgp>;
2478 // Patterns for loads of i1:
2479 def: Pat<(i1 (load AddrFI:$fi)),
2480 (C2_tfrrp (L2_loadrub_io AddrFI:$fi, 0))>;
2481 def: Pat<(i1 (load (add I32:$Rs, anyimm0:$Off))),
2482 (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, imm:$Off))>;
2483 def: Pat<(i1 (load I32:$Rs)),
2484 (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, 0))>;
2487 // --(13) Store ----------------------------------------------------------
2490 class Storepi_pat<PatFrag Store, PatFrag Value, PatFrag Offset, InstHexagon MI>
2491 : Pat<(Store Value:$Rt, I32:$Rx, Offset:$s4),
2492 (MI I32:$Rx, imm:$s4, Value:$Rt)>;
2494 def: Storepi_pat<post_truncsti8, I32, s4_0ImmPred, S2_storerb_pi>;
2495 def: Storepi_pat<post_truncsti16, I32, s4_1ImmPred, S2_storerh_pi>;
2496 def: Storepi_pat<post_store, I32, s4_2ImmPred, S2_storeri_pi>;
2497 def: Storepi_pat<post_store, I64, s4_3ImmPred, S2_storerd_pi>;
2499 // Patterns for generating stores, where the address takes different forms:
2501 // - frameindex + offset,
2503 // - simple (base address without offset).
2504 // These would usually be used together (via Storexi_pat defined below), but
2505 // in some cases one may want to apply different properties (such as
2506 // AddedComplexity) to the individual patterns.
2507 class Storexi_fi_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2508 : Pat<(Store Value:$Rs, AddrFI:$fi), (MI AddrFI:$fi, 0, Value:$Rs)>;
2510 multiclass Storexi_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2512 def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
2513 (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
2514 def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
2515 (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
2518 multiclass Storexi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2520 def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
2521 (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
2522 def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
2523 (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
2526 class Storexi_base_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2527 : Pat<(Store Value:$Rt, I32:$Rs),
2528 (MI IntRegs:$Rs, 0, Value:$Rt)>;
2530 // Patterns for generating stores, where the address takes different forms,
2531 // and where the value being stored is transformed through the value modifier
2532 // ValueMod. The address forms are same as above.
2533 class Storexim_fi_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2535 : Pat<(Store Value:$Rs, AddrFI:$fi),
2536 (MI AddrFI:$fi, 0, (ValueMod Value:$Rs))>;
2538 multiclass Storexim_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2539 PatFrag ValueMod, InstHexagon MI> {
2540 def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
2541 (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
2542 def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
2543 (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
2546 multiclass Storexim_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2547 PatFrag ValueMod, InstHexagon MI> {
2548 def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
2549 (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
2550 def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
2551 (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
2554 class Storexim_base_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2556 : Pat<(Store Value:$Rt, I32:$Rs),
2557 (MI IntRegs:$Rs, 0, (ValueMod Value:$Rt))>;
2559 multiclass Storexi_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
2561 defm: Storexi_fi_add_pat <Store, Value, ImmPred, MI>;
2562 def: Storexi_fi_pat <Store, Value, MI>;
2563 defm: Storexi_add_pat <Store, Value, ImmPred, MI>;
2566 multiclass Storexim_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
2567 PatFrag ValueMod, InstHexagon MI> {
2568 defm: Storexim_fi_add_pat <Store, Value, ImmPred, ValueMod, MI>;
2569 def: Storexim_fi_pat <Store, Value, ValueMod, MI>;
2570 defm: Storexim_add_pat <Store, Value, ImmPred, ValueMod, MI>;
2574 class Storexu_shl_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred, InstHexagon MI>
2575 : Pat<(Store Value:$Rt, (add (shl I32:$Ru, u2_0ImmPred:$u2), ImmPred:$A)),
2576 (MI IntRegs:$Ru, imm:$u2, ImmPred:$A, Value:$Rt)>;
2579 class Storexr_shl_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2580 : Pat<(Store Value:$Ru, (add I32:$Rs, (shl I32:$Rt, u2_0ImmPred:$u2))),
2581 (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2, Value:$Ru)>;
2584 class Storexr_add_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2585 : Pat<(Store Value:$Ru, (add I32:$Rs, I32:$Rt)),
2586 (MI IntRegs:$Rs, IntRegs:$Rt, 0, Value:$Ru)>;
2588 class Storea_pat<PatFrag Store, PatFrag Value, PatFrag Addr, InstHexagon MI>
2589 : Pat<(Store Value:$val, Addr:$addr), (MI Addr:$addr, Value:$val)>;
2591 class Stoream_pat<PatFrag Store, PatFrag Value, PatFrag Addr, PatFrag ValueMod,
2593 : Pat<(Store Value:$val, Addr:$addr),
2594 (MI Addr:$addr, (ValueMod Value:$val))>;
2596 def IMM_BYTE : SDNodeXForm<imm, [{
2597 // -1 can be represented as 255, etc.
2598 // assigning to a byte restores our desired signed value.
2599 int8_t imm = N->getSExtValue();
2600 return CurDAG->getSignedTargetConstant(imm, SDLoc(N), MVT::i32);
2603 def IMM_HALF : SDNodeXForm<imm, [{
2604 // -1 can be represented as 65535, etc.
2605 // assigning to a short restores our desired signed value.
2606 int16_t imm = N->getSExtValue();
2607 return CurDAG->getSignedTargetConstant(imm, SDLoc(N), MVT::i32);
2610 def IMM_WORD : SDNodeXForm<imm, [{
2611 // -1 can be represented as 4294967295, etc.
2612 // Currently, it's not doing this. But some optimization
2613 // might convert -1 to a large +ve number.
2614 // assigning to a word restores our desired signed value.
2615 int32_t imm = N->getSExtValue();
2616 return CurDAG->getSignedTargetConstant(imm, SDLoc(N), MVT::i32);
2619 def ToImmByte : OutPatFrag<(ops node:$R), (IMM_BYTE $R)>;
2620 def ToImmHalf : OutPatFrag<(ops node:$R), (IMM_HALF $R)>;
2621 def ToImmWord : OutPatFrag<(ops node:$R), (IMM_WORD $R)>;
2623 // Even though the offset is not extendable in the store-immediate, we
2624 // can still generate the fi# in the base address. If the final offset
2625 // is not valid for the instruction, we will replace it with a scratch
2627 class SmallStackStore<PatFrag Store>
2628 : PatFrag<(ops node:$Val, node:$Addr), (Store node:$Val, node:$Addr), [{
2629 return isSmallStackStore(cast<StoreSDNode>(N));
2632 // This is the complement of SmallStackStore.
2633 class LargeStackStore<PatFrag Store>
2634 : PatFrag<(ops node:$Val, node:$Addr), (Store node:$Val, node:$Addr), [{
2635 return !isSmallStackStore(cast<StoreSDNode>(N));
2638 // Preferred addressing modes for various combinations of stored value
2639 // and address computation.
2640 // For stores where the address and value are both immediates, prefer
2641 // store-immediate. The reason is that the constant-extender optimization
2642 // can replace store-immediate with a store-register, but there is nothing
2643 // to generate a store-immediate out of a store-register.
2645 // C R F F+C R+C R+R R<<S+C R<<S+R
2646 // --+-------+-----+-----+------+-----+-----+--------+--------
2647 // C | imm | imm | imm | imm | imm | rr | ur | rr
2648 // R | abs* | io | io | io | io | rr | ur | rr
2650 // (*) Absolute or GP-relative.
2652 // Note that any expression can be matched by Reg. In particular, an immediate
2653 // can always be placed in a register, so patterns checking for Imm should
2654 // have a higher priority than the ones involving Reg that could also match.
2655 // For example, *(p+4) could become r1=#4; memw(r0+r1<<#0) instead of the
2656 // preferred memw(r0+#4). Similarly Reg+Imm or Reg+Reg should be tried before
2659 // The order in which the different combinations are tried:
2661 // C F R F+C R+C R+R R<<S+C R<<S+R
2662 // --+-------+-----+-----+------+-----+-----+--------+--------
2663 // C | 1 | 6 | - | 5 | 9 | - | - | -
2664 // R | 2 | 8 | 12 | 7 | 10 | 11 | 3 | 4
2667 // First, match the unusual case of doubleword store into Reg+Imm4, i.e.
2668 // a store where the offset Imm4 is a multiple of 4, but not of 8. This
2669 // implies that Reg is also a proper multiple of 4. To still generate a
2670 // doubleword store, add 4 to Reg, and subtract 4 from the offset.
2672 def s30_2ProperPred : PatLeaf<(i32 imm), [{
2673 int64_t v = (int64_t)N->getSExtValue();
2674 return isShiftedInt<30,2>(v) && !isShiftedInt<29,3>(v);
2676 def RoundTo8 : SDNodeXForm<imm, [{
2677 int32_t Imm = N->getSExtValue();
2678 return CurDAG->getSignedTargetConstant(Imm & -8, SDLoc(N), MVT::i32);
2681 let AddedComplexity = 150 in
2682 def: Pat<(store I64:$Ru, (add I32:$Rs, s30_2ProperPred:$Off)),
2683 (S2_storerd_io (A2_addi I32:$Rs, 4), (RoundTo8 $Off), I64:$Ru)>;
2685 class Storexi_abs_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2686 : Pat<(Store Value:$val, anyimm:$addr),
2687 (MI (ToI32 $addr), 0, Value:$val)>;
2688 class Storexim_abs_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2690 : Pat<(Store Value:$val, anyimm:$addr),
2691 (MI (ToI32 $addr), 0, (ValueMod Value:$val))>;
2693 let AddedComplexity = 140 in {
2694 def: Storexim_abs_pat<truncstorei8, anyint, ToImmByte, S4_storeirb_io>;
2695 def: Storexim_abs_pat<truncstorei16, anyint, ToImmHalf, S4_storeirh_io>;
2696 def: Storexim_abs_pat<store, anyint, ToImmWord, S4_storeiri_io>;
2698 def: Storexi_abs_pat<truncstorei8, anyimm, S4_storeirb_io>;
2699 def: Storexi_abs_pat<truncstorei16, anyimm, S4_storeirh_io>;
2700 def: Storexi_abs_pat<store, anyimm, S4_storeiri_io>;
2703 // GP-relative address
2704 let AddedComplexity = 120 in {
2705 def: Storea_pat<truncstorei8, I32, addrgp, S2_storerbgp>;
2706 def: Storea_pat<truncstorei16, I32, addrgp, S2_storerhgp>;
2707 def: Storea_pat<store, I32, addrgp, S2_storerigp>;
2708 def: Storea_pat<store, V4I8, addrgp, S2_storerigp>;
2709 def: Storea_pat<store, V2I16, addrgp, S2_storerigp>;
2710 def: Storea_pat<store, I64, addrgp, S2_storerdgp>;
2711 def: Storea_pat<store, V8I8, addrgp, S2_storerdgp>;
2712 def: Storea_pat<store, V4I16, addrgp, S2_storerdgp>;
2713 def: Storea_pat<store, V2I32, addrgp, S2_storerdgp>;
2714 def: Storea_pat<store, F32, addrgp, S2_storerigp>;
2715 def: Storea_pat<store, F64, addrgp, S2_storerdgp>;
2716 def: Storea_pat<atomic_store_8, I32, addrgp, S2_storerbgp>;
2717 def: Storea_pat<atomic_store_16, I32, addrgp, S2_storerhgp>;
2718 def: Storea_pat<atomic_store_32, I32, addrgp, S2_storerigp>;
2719 def: Storea_pat<atomic_store_32, V4I8, addrgp, S2_storerigp>;
2720 def: Storea_pat<atomic_store_32, V2I16, addrgp, S2_storerigp>;
2721 def: Storea_pat<atomic_store_64, I64, addrgp, S2_storerdgp>;
2722 def: Storea_pat<atomic_store_64, V8I8, addrgp, S2_storerdgp>;
2723 def: Storea_pat<atomic_store_64, V4I16, addrgp, S2_storerdgp>;
2724 def: Storea_pat<atomic_store_64, V2I32, addrgp, S2_storerdgp>;
2726 def: Stoream_pat<truncstorei8, I64, addrgp, LoReg, S2_storerbgp>;
2727 def: Stoream_pat<truncstorei16, I64, addrgp, LoReg, S2_storerhgp>;
2728 def: Stoream_pat<truncstorei32, I64, addrgp, LoReg, S2_storerigp>;
2729 def: Stoream_pat<store, I1, addrgp, I1toI32, S2_storerbgp>;
2733 let AddedComplexity = 110 in {
2734 def: Storea_pat<truncstorei8, I32, anyimm0, PS_storerbabs>;
2735 def: Storea_pat<truncstorei16, I32, anyimm1, PS_storerhabs>;
2736 def: Storea_pat<store, I32, anyimm2, PS_storeriabs>;
2737 def: Storea_pat<store, V4I8, anyimm2, PS_storeriabs>;
2738 def: Storea_pat<store, V2I16, anyimm2, PS_storeriabs>;
2739 def: Storea_pat<store, I64, anyimm3, PS_storerdabs>;
2740 def: Storea_pat<store, V8I8, anyimm3, PS_storerdabs>;
2741 def: Storea_pat<store, V4I16, anyimm3, PS_storerdabs>;
2742 def: Storea_pat<store, V2I32, anyimm3, PS_storerdabs>;
2743 def: Storea_pat<store, F32, anyimm2, PS_storeriabs>;
2744 def: Storea_pat<store, F64, anyimm3, PS_storerdabs>;
2745 def: Storea_pat<atomic_store_8, I32, anyimm0, PS_storerbabs>;
2746 def: Storea_pat<atomic_store_16, I32, anyimm1, PS_storerhabs>;
2747 def: Storea_pat<atomic_store_32, I32, anyimm2, PS_storeriabs>;
2748 def: Storea_pat<atomic_store_32, V4I8, anyimm2, PS_storeriabs>;
2749 def: Storea_pat<atomic_store_32, V2I16, anyimm2, PS_storeriabs>;
2750 def: Storea_pat<atomic_store_64, I64, anyimm3, PS_storerdabs>;
2751 def: Storea_pat<atomic_store_64, V8I8, anyimm3, PS_storerdabs>;
2752 def: Storea_pat<atomic_store_64, V4I16, anyimm3, PS_storerdabs>;
2753 def: Storea_pat<atomic_store_64, V2I32, anyimm3, PS_storerdabs>;
2755 def: Stoream_pat<truncstorei8, I64, anyimm0, LoReg, PS_storerbabs>;
2756 def: Stoream_pat<truncstorei16, I64, anyimm1, LoReg, PS_storerhabs>;
2757 def: Stoream_pat<truncstorei32, I64, anyimm2, LoReg, PS_storeriabs>;
2758 def: Stoream_pat<store, I1, anyimm0, I1toI32, PS_storerbabs>;
2762 let AddedComplexity = 100 in {
2763 def: Storexu_shl_pat<truncstorei8, I32, anyimm0, S4_storerb_ur>;
2764 def: Storexu_shl_pat<truncstorei16, I32, anyimm1, S4_storerh_ur>;
2765 def: Storexu_shl_pat<store, I32, anyimm2, S4_storeri_ur>;
2766 def: Storexu_shl_pat<store, V4I8, anyimm2, S4_storeri_ur>;
2767 def: Storexu_shl_pat<store, V2I16, anyimm2, S4_storeri_ur>;
2768 def: Storexu_shl_pat<store, I64, anyimm3, S4_storerd_ur>;
2769 def: Storexu_shl_pat<store, V8I8, anyimm3, S4_storerd_ur>;
2770 def: Storexu_shl_pat<store, V4I16, anyimm3, S4_storerd_ur>;
2771 def: Storexu_shl_pat<store, V2I32, anyimm3, S4_storerd_ur>;
2772 def: Storexu_shl_pat<store, F32, anyimm2, S4_storeri_ur>;
2773 def: Storexu_shl_pat<store, F64, anyimm3, S4_storerd_ur>;
2775 def: Pat<(store I1:$Pu, (add (shl I32:$Rs, u2_0ImmPred:$u2), anyimm:$A)),
2776 (S4_storerb_ur IntRegs:$Rs, imm:$u2, imm:$A, (I1toI32 I1:$Pu))>;
2780 let AddedComplexity = 90 in {
2781 def: Storexr_shl_pat<truncstorei8, I32, S4_storerb_rr>;
2782 def: Storexr_shl_pat<truncstorei16, I32, S4_storerh_rr>;
2783 def: Storexr_shl_pat<store, I32, S4_storeri_rr>;
2784 def: Storexr_shl_pat<store, V4I8, S4_storeri_rr>;
2785 def: Storexr_shl_pat<store, V2I16, S4_storeri_rr>;
2786 def: Storexr_shl_pat<store, I64, S4_storerd_rr>;
2787 def: Storexr_shl_pat<store, V8I8, S4_storerd_rr>;
2788 def: Storexr_shl_pat<store, V4I16, S4_storerd_rr>;
2789 def: Storexr_shl_pat<store, V2I32, S4_storerd_rr>;
2790 def: Storexr_shl_pat<store, F32, S4_storeri_rr>;
2791 def: Storexr_shl_pat<store, F64, S4_storerd_rr>;
2793 def: Pat<(store I1:$Pu, (add (shl I32:$Rs, u2_0ImmPred:$u2), I32:$Rt)),
2794 (S4_storerb_rr IntRegs:$Rt, IntRegs:$Rs, imm:$u2, (I1toI32 I1:$Pu))>;
2797 class SS_<PatFrag F> : SmallStackStore<F>;
2798 class LS_<PatFrag F> : LargeStackStore<F>;
2800 multiclass IMFA_<PatFrag S, PatFrag V, PatFrag O, PatFrag M, InstHexagon I> {
2801 defm: Storexim_fi_add_pat<S, V, O, M, I>;
2803 multiclass IFA_<PatFrag S, PatFrag V, PatFrag O, InstHexagon I> {
2804 defm: Storexi_fi_add_pat<S, V, O, I>;
2807 // Fi+Imm, store-immediate
2808 let AddedComplexity = 80 in {
2809 defm: IMFA_<SS_<truncstorei8>, anyint, u6_0ImmPred, ToImmByte, S4_storeirb_io>;
2810 defm: IMFA_<SS_<truncstorei16>, anyint, u6_1ImmPred, ToImmHalf, S4_storeirh_io>;
2811 defm: IMFA_<SS_<store>, anyint, u6_2ImmPred, ToImmWord, S4_storeiri_io>;
2813 defm: IFA_<SS_<truncstorei8>, anyimm, u6_0ImmPred, S4_storeirb_io>;
2814 defm: IFA_<SS_<truncstorei16>, anyimm, u6_1ImmPred, S4_storeirh_io>;
2815 defm: IFA_<SS_<store>, anyimm, u6_2ImmPred, S4_storeiri_io>;
2817 // For large-stack stores, generate store-register (prefer explicit Fi
2819 defm: IMFA_<LS_<truncstorei8>, anyimm, u6_0ImmPred, ToI32, S2_storerb_io>;
2820 defm: IMFA_<LS_<truncstorei16>, anyimm, u6_1ImmPred, ToI32, S2_storerh_io>;
2821 defm: IMFA_<LS_<store>, anyimm, u6_2ImmPred, ToI32, S2_storeri_io>;
2824 // Fi, store-immediate
2825 let AddedComplexity = 70 in {
2826 def: Storexim_fi_pat<SS_<truncstorei8>, anyint, ToImmByte, S4_storeirb_io>;
2827 def: Storexim_fi_pat<SS_<truncstorei16>, anyint, ToImmHalf, S4_storeirh_io>;
2828 def: Storexim_fi_pat<SS_<store>, anyint, ToImmWord, S4_storeiri_io>;
2830 def: Storexi_fi_pat<SS_<truncstorei8>, anyimm, S4_storeirb_io>;
2831 def: Storexi_fi_pat<SS_<truncstorei16>, anyimm, S4_storeirh_io>;
2832 def: Storexi_fi_pat<SS_<store>, anyimm, S4_storeiri_io>;
2834 // For large-stack stores, generate store-register (prefer explicit Fi
2836 def: Storexim_fi_pat<LS_<truncstorei8>, anyimm, ToI32, S2_storerb_io>;
2837 def: Storexim_fi_pat<LS_<truncstorei16>, anyimm, ToI32, S2_storerh_io>;
2838 def: Storexim_fi_pat<LS_<store>, anyimm, ToI32, S2_storeri_io>;
2841 // Fi+Imm, Fi, store-register
2842 let AddedComplexity = 60 in {
2843 defm: Storexi_fi_add_pat<truncstorei8, I32, anyimm, S2_storerb_io>;
2844 defm: Storexi_fi_add_pat<truncstorei16, I32, anyimm, S2_storerh_io>;
2845 defm: Storexi_fi_add_pat<store, I32, anyimm, S2_storeri_io>;
2846 defm: Storexi_fi_add_pat<store, V4I8, anyimm, S2_storeri_io>;
2847 defm: Storexi_fi_add_pat<store, V2I16, anyimm, S2_storeri_io>;
2848 defm: Storexi_fi_add_pat<store, I64, anyimm, S2_storerd_io>;
2849 defm: Storexi_fi_add_pat<store, V8I8, anyimm, S2_storerd_io>;
2850 defm: Storexi_fi_add_pat<store, V4I16, anyimm, S2_storerd_io>;
2851 defm: Storexi_fi_add_pat<store, V2I32, anyimm, S2_storerd_io>;
2852 defm: Storexi_fi_add_pat<store, F32, anyimm, S2_storeri_io>;
2853 defm: Storexi_fi_add_pat<store, F64, anyimm, S2_storerd_io>;
2854 defm: Storexim_fi_add_pat<store, I1, anyimm, I1toI32, S2_storerb_io>;
2856 def: Storexi_fi_pat<truncstorei8, I32, S2_storerb_io>;
2857 def: Storexi_fi_pat<truncstorei16, I32, S2_storerh_io>;
2858 def: Storexi_fi_pat<store, I32, S2_storeri_io>;
2859 def: Storexi_fi_pat<store, V4I8, S2_storeri_io>;
2860 def: Storexi_fi_pat<store, V2I16, S2_storeri_io>;
2861 def: Storexi_fi_pat<store, I64, S2_storerd_io>;
2862 def: Storexi_fi_pat<store, V8I8, S2_storerd_io>;
2863 def: Storexi_fi_pat<store, V4I16, S2_storerd_io>;
2864 def: Storexi_fi_pat<store, V2I32, S2_storerd_io>;
2865 def: Storexi_fi_pat<store, F32, S2_storeri_io>;
2866 def: Storexi_fi_pat<store, F64, S2_storerd_io>;
2867 def: Storexim_fi_pat<store, I1, I1toI32, S2_storerb_io>;
2871 multiclass IMRA_<PatFrag S, PatFrag V, PatFrag O, PatFrag M, InstHexagon I> {
2872 defm: Storexim_add_pat<S, V, O, M, I>;
2874 multiclass IRA_<PatFrag S, PatFrag V, PatFrag O, InstHexagon I> {
2875 defm: Storexi_add_pat<S, V, O, I>;
2878 // Reg+Imm, store-immediate
2879 let AddedComplexity = 50 in {
2880 defm: IMRA_<truncstorei8, anyint, u6_0ImmPred, ToImmByte, S4_storeirb_io>;
2881 defm: IMRA_<truncstorei16, anyint, u6_1ImmPred, ToImmHalf, S4_storeirh_io>;
2882 defm: IMRA_<store, anyint, u6_2ImmPred, ToImmWord, S4_storeiri_io>;
2884 defm: IRA_<truncstorei8, anyimm, u6_0ImmPred, S4_storeirb_io>;
2885 defm: IRA_<truncstorei16, anyimm, u6_1ImmPred, S4_storeirh_io>;
2886 defm: IRA_<store, anyimm, u6_2ImmPred, S4_storeiri_io>;
2889 // Reg+Imm, store-register
2890 let AddedComplexity = 40 in {
2891 defm: Storexi_pat<truncstorei8, I32, anyimm0, S2_storerb_io>;
2892 defm: Storexi_pat<truncstorei16, I32, anyimm1, S2_storerh_io>;
2893 defm: Storexi_pat<store, I32, anyimm2, S2_storeri_io>;
2894 defm: Storexi_pat<store, V4I8, anyimm2, S2_storeri_io>;
2895 defm: Storexi_pat<store, V2I16, anyimm2, S2_storeri_io>;
2896 defm: Storexi_pat<store, I64, anyimm3, S2_storerd_io>;
2897 defm: Storexi_pat<store, V8I8, anyimm3, S2_storerd_io>;
2898 defm: Storexi_pat<store, V4I16, anyimm3, S2_storerd_io>;
2899 defm: Storexi_pat<store, V2I32, anyimm3, S2_storerd_io>;
2900 defm: Storexi_pat<store, F32, anyimm2, S2_storeri_io>;
2901 defm: Storexi_pat<store, F64, anyimm3, S2_storerd_io>;
2903 defm: Storexim_pat<truncstorei8, I64, anyimm0, LoReg, S2_storerb_io>;
2904 defm: Storexim_pat<truncstorei16, I64, anyimm1, LoReg, S2_storerh_io>;
2905 defm: Storexim_pat<truncstorei32, I64, anyimm2, LoReg, S2_storeri_io>;
2906 defm: Storexim_pat<store, I1, anyimm0, I1toI32, S2_storerb_io>;
2908 defm: Storexi_pat<atomic_store_8, I32, anyimm0, S2_storerb_io>;
2909 defm: Storexi_pat<atomic_store_16, I32, anyimm1, S2_storerh_io>;
2910 defm: Storexi_pat<atomic_store_32, I32, anyimm2, S2_storeri_io>;
2911 defm: Storexi_pat<atomic_store_32, V4I8, anyimm2, S2_storeri_io>;
2912 defm: Storexi_pat<atomic_store_32, V2I16, anyimm2, S2_storeri_io>;
2913 defm: Storexi_pat<atomic_store_64, I64, anyimm3, S2_storerd_io>;
2914 defm: Storexi_pat<atomic_store_64, V8I8, anyimm3, S2_storerd_io>;
2915 defm: Storexi_pat<atomic_store_64, V4I16, anyimm3, S2_storerd_io>;
2916 defm: Storexi_pat<atomic_store_64, V2I32, anyimm3, S2_storerd_io>;
2920 let AddedComplexity = 30 in {
2921 def: Storexr_add_pat<truncstorei8, I32, S4_storerb_rr>;
2922 def: Storexr_add_pat<truncstorei16, I32, S4_storerh_rr>;
2923 def: Storexr_add_pat<store, I32, S4_storeri_rr>;
2924 def: Storexr_add_pat<store, V4I8, S4_storeri_rr>;
2925 def: Storexr_add_pat<store, V2I16, S4_storeri_rr>;
2926 def: Storexr_add_pat<store, I64, S4_storerd_rr>;
2927 def: Storexr_add_pat<store, V8I8, S4_storerd_rr>;
2928 def: Storexr_add_pat<store, V4I16, S4_storerd_rr>;
2929 def: Storexr_add_pat<store, V2I32, S4_storerd_rr>;
2930 def: Storexr_add_pat<store, F32, S4_storeri_rr>;
2931 def: Storexr_add_pat<store, F64, S4_storerd_rr>;
2933 def: Pat<(store I1:$Pu, (add I32:$Rs, I32:$Rt)),
2934 (S4_storerb_rr IntRegs:$Rs, IntRegs:$Rt, 0, (I1toI32 I1:$Pu))>;
2937 // Reg, store-immediate
2938 let AddedComplexity = 20 in {
2939 def: Storexim_base_pat<truncstorei8, anyint, ToImmByte, S4_storeirb_io>;
2940 def: Storexim_base_pat<truncstorei16, anyint, ToImmHalf, S4_storeirh_io>;
2941 def: Storexim_base_pat<store, anyint, ToImmWord, S4_storeiri_io>;
2943 def: Storexi_base_pat<truncstorei8, anyimm, S4_storeirb_io>;
2944 def: Storexi_base_pat<truncstorei16, anyimm, S4_storeirh_io>;
2945 def: Storexi_base_pat<store, anyimm, S4_storeiri_io>;
2948 // Reg, store-register
2949 let AddedComplexity = 10 in {
2950 def: Storexi_base_pat<truncstorei8, I32, S2_storerb_io>;
2951 def: Storexi_base_pat<truncstorei16, I32, S2_storerh_io>;
2952 def: Storexi_base_pat<store, I32, S2_storeri_io>;
2953 def: Storexi_base_pat<store, V4I8, S2_storeri_io>;
2954 def: Storexi_base_pat<store, V2I16, S2_storeri_io>;
2955 def: Storexi_base_pat<store, I64, S2_storerd_io>;
2956 def: Storexi_base_pat<store, V8I8, S2_storerd_io>;
2957 def: Storexi_base_pat<store, V4I16, S2_storerd_io>;
2958 def: Storexi_base_pat<store, V2I32, S2_storerd_io>;
2959 def: Storexi_base_pat<store, F32, S2_storeri_io>;
2960 def: Storexi_base_pat<store, F64, S2_storerd_io>;
2962 def: Storexim_base_pat<truncstorei8, I64, LoReg, S2_storerb_io>;
2963 def: Storexim_base_pat<truncstorei16, I64, LoReg, S2_storerh_io>;
2964 def: Storexim_base_pat<truncstorei32, I64, LoReg, S2_storeri_io>;
2965 def: Storexim_base_pat<store, I1, I1toI32, S2_storerb_io>;
2967 def: Storexi_base_pat<atomic_store_8, I32, S2_storerb_io>;
2968 def: Storexi_base_pat<atomic_store_16, I32, S2_storerh_io>;
2969 def: Storexi_base_pat<atomic_store_32, I32, S2_storeri_io>;
2970 def: Storexi_base_pat<atomic_store_32, V4I8, S2_storeri_io>;
2971 def: Storexi_base_pat<atomic_store_32, V2I16, S2_storeri_io>;
2972 def: Storexi_base_pat<atomic_store_64, I64, S2_storerd_io>;
2973 def: Storexi_base_pat<atomic_store_64, V8I8, S2_storerd_io>;
2974 def: Storexi_base_pat<atomic_store_64, V4I16, S2_storerd_io>;
2975 def: Storexi_base_pat<atomic_store_64, V2I32, S2_storerd_io>;
2979 // --(14) Memop ----------------------------------------------------------
2982 def m5_0Imm8Pred : PatLeaf<(i32 imm), [{
2983 int8_t V = N->getSExtValue();
2984 return -32 < V && V <= -1;
2987 def m5_0Imm16Pred : PatLeaf<(i32 imm), [{
2988 int16_t V = N->getSExtValue();
2989 return -32 < V && V <= -1;
2992 def m5_0ImmPred : PatLeaf<(i32 imm), [{
2993 int64_t V = N->getSExtValue();
2994 return -31 <= V && V <= -1;
2997 def IsNPow2_8 : PatLeaf<(i32 imm), [{
2998 uint8_t NV = ~N->getZExtValue();
2999 return isPowerOf2_32(NV);
3002 def IsNPow2_16 : PatLeaf<(i32 imm), [{
3003 uint16_t NV = ~N->getZExtValue();
3004 return isPowerOf2_32(NV);
3007 def Log2_8 : SDNodeXForm<imm, [{
3008 uint8_t V = N->getZExtValue();
3009 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
3012 def Log2_16 : SDNodeXForm<imm, [{
3013 uint16_t V = N->getZExtValue();
3014 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
3017 def LogN2_8 : SDNodeXForm<imm, [{
3018 uint8_t NV = ~N->getZExtValue();
3019 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
3022 def LogN2_16 : SDNodeXForm<imm, [{
3023 uint16_t NV = ~N->getZExtValue();
3024 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
3027 def IdImm : SDNodeXForm<imm, [{ return SDValue(N, 0); }]>;
3029 multiclass Memopxr_base_pat<PatFrag Load, PatFrag Store, SDNode Oper,
3032 def: Pat<(Store (Oper (Load I32:$Rs), I32:$A), I32:$Rs),
3033 (MI I32:$Rs, 0, I32:$A)>;
3035 def: Pat<(Store (Oper (Load AddrFI:$Rs), I32:$A), AddrFI:$Rs),
3036 (MI AddrFI:$Rs, 0, I32:$A)>;
3039 multiclass Memopxr_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
3040 SDNode Oper, InstHexagon MI> {
3042 def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), I32:$A),
3043 (add I32:$Rs, ImmPred:$Off)),
3044 (MI I32:$Rs, imm:$Off, I32:$A)>;
3045 def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), I32:$A),
3046 (IsOrAdd I32:$Rs, ImmPred:$Off)),
3047 (MI I32:$Rs, imm:$Off, I32:$A)>;
3049 def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), I32:$A),
3050 (add AddrFI:$Rs, ImmPred:$Off)),
3051 (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
3052 def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), I32:$A),
3053 (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
3054 (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
3057 multiclass Memopxr_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
3058 SDNode Oper, InstHexagon MI> {
3059 let Predicates = [UseMEMOPS] in {
3060 defm: Memopxr_base_pat <Load, Store, Oper, MI>;
3061 defm: Memopxr_add_pat <Load, Store, ImmPred, Oper, MI>;
3065 let AddedComplexity = 200 in {
3067 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, add,
3068 /*anyext*/ L4_add_memopb_io>;
3069 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, add,
3070 /*sext*/ L4_add_memopb_io>;
3071 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, add,
3072 /*zext*/ L4_add_memopb_io>;
3073 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, add,
3074 /*anyext*/ L4_add_memoph_io>;
3075 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, add,
3076 /*sext*/ L4_add_memoph_io>;
3077 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, add,
3078 /*zext*/ L4_add_memoph_io>;
3079 defm: Memopxr_pat<load, store, u6_2ImmPred, add, L4_add_memopw_io>;
3082 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, sub,
3083 /*anyext*/ L4_sub_memopb_io>;
3084 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub,
3085 /*sext*/ L4_sub_memopb_io>;
3086 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub,
3087 /*zext*/ L4_sub_memopb_io>;
3088 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, sub,
3089 /*anyext*/ L4_sub_memoph_io>;
3090 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub,
3091 /*sext*/ L4_sub_memoph_io>;
3092 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub,
3093 /*zext*/ L4_sub_memoph_io>;
3094 defm: Memopxr_pat<load, store, u6_2ImmPred, sub, L4_sub_memopw_io>;
3097 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, and,
3098 /*anyext*/ L4_and_memopb_io>;
3099 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, and,
3100 /*sext*/ L4_and_memopb_io>;
3101 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, and,
3102 /*zext*/ L4_and_memopb_io>;
3103 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, and,
3104 /*anyext*/ L4_and_memoph_io>;
3105 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, and,
3106 /*sext*/ L4_and_memoph_io>;
3107 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, and,
3108 /*zext*/ L4_and_memoph_io>;
3109 defm: Memopxr_pat<load, store, u6_2ImmPred, and, L4_and_memopw_io>;
3112 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, or,
3113 /*anyext*/ L4_or_memopb_io>;
3114 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, or,
3115 /*sext*/ L4_or_memopb_io>;
3116 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, or,
3117 /*zext*/ L4_or_memopb_io>;
3118 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, or,
3119 /*anyext*/ L4_or_memoph_io>;
3120 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, or,
3121 /*sext*/ L4_or_memoph_io>;
3122 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, or,
3123 /*zext*/ L4_or_memoph_io>;
3124 defm: Memopxr_pat<load, store, u6_2ImmPred, or, L4_or_memopw_io>;
3128 multiclass Memopxi_base_pat<PatFrag Load, PatFrag Store, SDNode Oper,
3129 PatFrag Arg, SDNodeXForm ArgMod, InstHexagon MI> {
3131 def: Pat<(Store (Oper (Load I32:$Rs), Arg:$A), I32:$Rs),
3132 (MI I32:$Rs, 0, (ArgMod Arg:$A))>;
3134 def: Pat<(Store (Oper (Load AddrFI:$Rs), Arg:$A), AddrFI:$Rs),
3135 (MI AddrFI:$Rs, 0, (ArgMod Arg:$A))>;
3138 multiclass Memopxi_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
3139 SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
3142 def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), Arg:$A),
3143 (add I32:$Rs, ImmPred:$Off)),
3144 (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
3145 def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), Arg:$A),
3146 (IsOrAdd I32:$Rs, ImmPred:$Off)),
3147 (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
3149 def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
3150 (add AddrFI:$Rs, ImmPred:$Off)),
3151 (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
3152 def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
3153 (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
3154 (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
3157 multiclass Memopxi_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
3158 SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
3160 let Predicates = [UseMEMOPS] in {
3161 defm: Memopxi_base_pat <Load, Store, Oper, Arg, ArgMod, MI>;
3162 defm: Memopxi_add_pat <Load, Store, ImmPred, Oper, Arg, ArgMod, MI>;
3166 let AddedComplexity = 220 in {
3168 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
3169 /*anyext*/ IdImm, L4_iadd_memopb_io>;
3170 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
3171 /*sext*/ IdImm, L4_iadd_memopb_io>;
3172 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
3173 /*zext*/ IdImm, L4_iadd_memopb_io>;
3174 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
3175 /*anyext*/ IdImm, L4_iadd_memoph_io>;
3176 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
3177 /*sext*/ IdImm, L4_iadd_memoph_io>;
3178 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
3179 /*zext*/ IdImm, L4_iadd_memoph_io>;
3180 defm: Memopxi_pat<load, store, u6_2ImmPred, add, u5_0ImmPred, IdImm,
3182 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
3183 /*anyext*/ NegImm8, L4_iadd_memopb_io>;
3184 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
3185 /*sext*/ NegImm8, L4_iadd_memopb_io>;
3186 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
3187 /*zext*/ NegImm8, L4_iadd_memopb_io>;
3188 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
3189 /*anyext*/ NegImm16, L4_iadd_memoph_io>;
3190 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
3191 /*sext*/ NegImm16, L4_iadd_memoph_io>;
3192 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
3193 /*zext*/ NegImm16, L4_iadd_memoph_io>;
3194 defm: Memopxi_pat<load, store, u6_2ImmPred, sub, m5_0ImmPred, NegImm32,
3198 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
3199 /*anyext*/ IdImm, L4_isub_memopb_io>;
3200 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
3201 /*sext*/ IdImm, L4_isub_memopb_io>;
3202 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
3203 /*zext*/ IdImm, L4_isub_memopb_io>;
3204 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
3205 /*anyext*/ IdImm, L4_isub_memoph_io>;
3206 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
3207 /*sext*/ IdImm, L4_isub_memoph_io>;
3208 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
3209 /*zext*/ IdImm, L4_isub_memoph_io>;
3210 defm: Memopxi_pat<load, store, u6_2ImmPred, sub, u5_0ImmPred, IdImm,
3212 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
3213 /*anyext*/ NegImm8, L4_isub_memopb_io>;
3214 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
3215 /*sext*/ NegImm8, L4_isub_memopb_io>;
3216 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
3217 /*zext*/ NegImm8, L4_isub_memopb_io>;
3218 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
3219 /*anyext*/ NegImm16, L4_isub_memoph_io>;
3220 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
3221 /*sext*/ NegImm16, L4_isub_memoph_io>;
3222 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
3223 /*zext*/ NegImm16, L4_isub_memoph_io>;
3224 defm: Memopxi_pat<load, store, u6_2ImmPred, add, m5_0ImmPred, NegImm32,
3228 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
3229 /*anyext*/ LogN2_8, L4_iand_memopb_io>;
3230 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
3231 /*sext*/ LogN2_8, L4_iand_memopb_io>;
3232 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
3233 /*zext*/ LogN2_8, L4_iand_memopb_io>;
3234 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
3235 /*anyext*/ LogN2_16, L4_iand_memoph_io>;
3236 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
3237 /*sext*/ LogN2_16, L4_iand_memoph_io>;
3238 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
3239 /*zext*/ LogN2_16, L4_iand_memoph_io>;
3240 defm: Memopxi_pat<load, store, u6_2ImmPred, and, IsNPow2_32,
3241 LogN2_32, L4_iand_memopw_io>;
3244 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3245 /*anyext*/ Log2_8, L4_ior_memopb_io>;
3246 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3247 /*sext*/ Log2_8, L4_ior_memopb_io>;
3248 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3249 /*zext*/ Log2_8, L4_ior_memopb_io>;
3250 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3251 /*anyext*/ Log2_16, L4_ior_memoph_io>;
3252 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3253 /*sext*/ Log2_16, L4_ior_memoph_io>;
3254 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3255 /*zext*/ Log2_16, L4_ior_memoph_io>;
3256 defm: Memopxi_pat<load, store, u6_2ImmPred, or, IsPow2_32,
3257 Log2_32, L4_ior_memopw_io>;
3261 // --(15) Call -----------------------------------------------------------
3264 // Pseudo instructions.
3265 def SDT_SPCallSeqStart
3266 : SDCallSeqStart<[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3267 def SDT_SPCallSeqEnd
3268 : SDCallSeqEnd<[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3270 def callseq_start: SDNode<"ISD::CALLSEQ_START", SDT_SPCallSeqStart,
3271 [SDNPHasChain, SDNPOutGlue]>;
3272 def callseq_end: SDNode<"ISD::CALLSEQ_END", SDT_SPCallSeqEnd,
3273 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
3275 def SDT_SPCall: SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
3277 def HexagonTCRet: SDNode<"HexagonISD::TC_RETURN", SDT_SPCall,
3278 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
3279 def callv3: SDNode<"HexagonISD::CALL", SDT_SPCall,
3280 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
3281 def callv3nr: SDNode<"HexagonISD::CALLnr", SDT_SPCall,
3282 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
3284 def: Pat<(callseq_start timm:$amt, timm:$amt2),
3285 (ADJCALLSTACKDOWN imm:$amt, imm:$amt2)>;
3286 def: Pat<(callseq_end timm:$amt1, timm:$amt2),
3287 (ADJCALLSTACKUP imm:$amt1, imm:$amt2)>;
3289 def: Pat<(HexagonTCRet tglobaladdr:$dst), (PS_tailcall_i tglobaladdr:$dst)>;
3290 def: Pat<(HexagonTCRet texternalsym:$dst), (PS_tailcall_i texternalsym:$dst)>;
3291 def: Pat<(HexagonTCRet I32:$dst), (PS_tailcall_r I32:$dst)>;
3293 def: Pat<(callv3 I32:$dst), (J2_callr I32:$dst)>;
3294 def: Pat<(callv3 tglobaladdr:$dst), (J2_call tglobaladdr:$dst)>;
3295 def: Pat<(callv3 texternalsym:$dst), (J2_call texternalsym:$dst)>;
3296 def: Pat<(callv3 tglobaltlsaddr:$dst), (J2_call tglobaltlsaddr:$dst)>;
3298 def: Pat<(callv3nr I32:$dst), (PS_callr_nr I32:$dst)>;
3299 def: Pat<(callv3nr tglobaladdr:$dst), (PS_call_nr tglobaladdr:$dst)>;
3300 def: Pat<(callv3nr texternalsym:$dst), (PS_call_nr texternalsym:$dst)>;
3302 def retglue : SDNode<"HexagonISD::RET_GLUE", SDTNone,
3303 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
3304 def eh_return: SDNode<"HexagonISD::EH_RETURN", SDTNone, [SDNPHasChain]>;
3306 def: Pat<(retglue), (PS_jmpret (i32 R31))>;
3307 def: Pat<(eh_return), (EH_RETURN_JMPR (i32 R31))>;
3310 // --(16) Branch ---------------------------------------------------------
3313 def: Pat<(br bb:$dst), (J2_jump b30_2Imm:$dst)>;
3314 def: Pat<(brind I32:$dst), (J2_jumpr I32:$dst)>;
3316 def: Pat<(brcond I1:$Pu, bb:$dst),
3317 (J2_jumpt I1:$Pu, bb:$dst)>;
3318 def: Pat<(brcond (not I1:$Pu), bb:$dst),
3319 (J2_jumpf I1:$Pu, bb:$dst)>;
3320 def: Pat<(brcond (i1 (setne I1:$Pu, -1)), bb:$dst),
3321 (J2_jumpf I1:$Pu, bb:$dst)>;
3322 def: Pat<(brcond (i1 (seteq I1:$Pu, 0)), bb:$dst),
3323 (J2_jumpf I1:$Pu, bb:$dst)>;
3324 def: Pat<(brcond (i1 (setne I1:$Pu, 0)), bb:$dst),
3325 (J2_jumpt I1:$Pu, bb:$dst)>;
3328 // --(17) Misc -----------------------------------------------------------
3331 // Generate code of the form 'C2_muxii(cmpbgtui(Rdd, C-1),0,1)'
3332 // for C code of the form r = (c>='0' && c<='9') ? 1 : 0.
3333 // The isdigit transformation relies on two 'clever' aspects:
3334 // 1) The data type is unsigned which allows us to eliminate a zero test after
3335 // biasing the expression by 48. We are depending on the representation of
3336 // the unsigned types, and semantics.
3337 // 2) The front end has converted <= 9 into < 10 on entry to LLVM.
3340 // retval = (c >= '0' && c <= '9') ? 1 : 0;
3341 // The code is transformed upstream of llvm into
3342 // retval = (c-48) < 10 ? 1 : 0;
3344 def u7_0PosImmPred : ImmLeaf<i32, [{
3345 // True if the immediate fits in an 7-bit unsigned field and is positive.
3346 return Imm > 0 && isUInt<7>(Imm);
3349 let AddedComplexity = 139 in
3350 def: Pat<(i32 (zext (i1 (setult (and I32:$Rs, 255), u7_0PosImmPred:$u7)))),
3351 (C2_muxii (A4_cmpbgtui IntRegs:$Rs, (UDEC1 imm:$u7)), 0, 1)>;
3353 let AddedComplexity = 100 in
3354 def: Pat<(or (or (shl (HexagonINSERT (i32 (zextloadi8 (add I32:$b, 2))),
3355 (i32 (extloadi8 (add I32:$b, 3))),
3358 (shl (i32 (zextloadi8 (add I32:$b, 1))), (i32 8))),
3359 (zextloadi8 I32:$b)),
3360 (A2_swiz (L2_loadri_io I32:$b, 0))>;
3363 // We need custom lowering of ISD::PREFETCH into HexagonISD::DCFETCH
3364 // because the SDNode ISD::PREFETCH has properties MayLoad and MayStore.
3365 // We don't really want either one here.
3366 def SDTHexagonDCFETCH: SDTypeProfile<0, 2, [SDTCisPtrTy<0>,SDTCisInt<1>]>;
3367 def HexagonDCFETCH: SDNode<"HexagonISD::DCFETCH", SDTHexagonDCFETCH,
3370 def: Pat<(HexagonDCFETCH IntRegs:$Rs, u11_3ImmPred:$u11_3),
3371 (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
3372 def: Pat<(HexagonDCFETCH (i32 (add IntRegs:$Rs, u11_3ImmPred:$u11_3)), (i32 0)),
3373 (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
3375 def SDTHexagonALLOCA
3376 : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3378 : SDNode<"HexagonISD::ALLOCA", SDTHexagonALLOCA, [SDNPHasChain]>;
3380 def: Pat<(HexagonALLOCA I32:$Rs, (i32 imm:$A)),
3381 (PS_alloca IntRegs:$Rs, imm:$A)>;
3383 def HexagonBARRIER: SDNode<"HexagonISD::BARRIER", SDTNone, [SDNPHasChain]>;
3384 def: Pat<(HexagonBARRIER), (Y2_barrier)>;
3386 def: Pat<(trap), (PS_crash)>;
3387 def: Pat<(debugtrap), (Y2_break)>;
3389 // Read cycle counter.
3390 def SDTInt64Leaf: SDTypeProfile<1, 0, [SDTCisVT<0, i64>]>;
3391 def HexagonREADCYCLE: SDNode<"HexagonISD::READCYCLE", SDTInt64Leaf,
3394 def: Pat<(HexagonREADCYCLE), (A4_tfrcpp UPCYCLE)>;
3396 // Read time counter.
3397 def HexagonREADTIMER: SDNode<"HexagonISD::READTIMER", SDTInt64Leaf,
3400 def: Pat<(HexagonREADTIMER), (A4_tfrcpp UTIMER)>;
3402 // The declared return value of the store-locked intrinsics is i32, but
3403 // the instructions actually define i1. To avoid register copies from
3404 // IntRegs to PredRegs and back, fold the entire pattern checking the
3405 // result against true/false.
3406 let AddedComplexity = 100 in {
3407 def: Pat<(i1 (setne (int_hexagon_S2_storew_locked I32:$Rs, I32:$Rt), 0)),
3408 (S2_storew_locked I32:$Rs, I32:$Rt)>;
3409 def: Pat<(i1 (seteq (int_hexagon_S2_storew_locked I32:$Rs, I32:$Rt), 0)),
3410 (C2_not (S2_storew_locked I32:$Rs, I32:$Rt))>;
3411 def: Pat<(i1 (setne (int_hexagon_S4_stored_locked I32:$Rs, I64:$Rt), 0)),
3412 (S4_stored_locked I32:$Rs, I64:$Rt)>;
3413 def: Pat<(i1 (seteq (int_hexagon_S4_stored_locked I32:$Rs, I64:$Rt), 0)),
3414 (C2_not (S4_stored_locked I32:$Rs, I64:$Rt))>;
3417 def: Pat<(int_hexagon_instrprof_custom (HexagonAtPcrel tglobaladdr:$addr), u32_0ImmPred:$I),
3418 (PS_call_instrprof_custom tglobaladdr:$addr, imm:$I)>;
3420 def: Pat<(int_hexagon_instrprof_custom (HexagonCONST32 tglobaladdr:$addr), u32_0ImmPred:$I),
3421 (PS_call_instrprof_custom tglobaladdr:$addr, imm:$I)>;