1 //==- HexagonPatterns.td - Target Description 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
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)>;
90 def HQ8: PatLeaf<(VecQ8 HvxQR:$R)>;
91 def HQ16: PatLeaf<(VecQ16 HvxQR:$R)>;
92 def HQ32: PatLeaf<(VecQ32 HvxQR:$R)>;
94 def HVI8: PatLeaf<(VecI8 HvxVR:$R)>;
95 def HVI16: PatLeaf<(VecI16 HvxVR:$R)>;
96 def HVI32: PatLeaf<(VecI32 HvxVR:$R)>;
98 def HWI8: PatLeaf<(VecPI8 HvxWR:$R)>;
99 def HWI16: PatLeaf<(VecPI16 HvxWR:$R)>;
100 def HWI32: PatLeaf<(VecPI32 HvxWR:$R)>;
103 SDTypeProfile<1, 0, [SDTCisVec<0>]>;
105 SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisVec<1>, SDTCisSameAs<1,2>,
108 def HexagonPTRUE: SDNode<"HexagonISD::PTRUE", SDTVecLeaf>;
109 def HexagonPFALSE: SDNode<"HexagonISD::PFALSE", SDTVecLeaf>;
110 def HexagonVALIGN: SDNode<"HexagonISD::VALIGN", SDTVecVecIntOp>;
111 def HexagonVALIGNADDR: SDNode<"HexagonISD::VALIGNADDR", SDTIntUnaryOp>;
113 def ptrue: PatFrag<(ops), (HexagonPTRUE)>;
114 def pfalse: PatFrag<(ops), (HexagonPFALSE)>;
115 def pnot: PatFrag<(ops node:$Pu), (xor node:$Pu, ptrue)>;
117 def valign: PatFrag<(ops node:$Vt, node:$Vs, node:$Ru),
118 (HexagonVALIGN node:$Vt, node:$Vs, node:$Ru)>;
119 def valignaddr: PatFrag<(ops node:$Addr), (HexagonVALIGNADDR node:$Addr)>;
121 // Pattern fragments to extract the low and high subregisters from a
123 def LoReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG (i64 $Rs), isub_lo)>;
124 def HiReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG (i64 $Rs), isub_hi)>;
126 def IsOrAdd: PatFrag<(ops node:$A, node:$B), (or node:$A, node:$B), [{
127 return isOrEquivalentToAdd(N);
130 def IsPow2_32: PatLeaf<(i32 imm), [{
131 uint32_t V = N->getZExtValue();
132 return isPowerOf2_32(V);
135 def IsPow2_64: PatLeaf<(i64 imm), [{
136 uint64_t V = N->getZExtValue();
137 return isPowerOf2_64(V);
140 def IsNPow2_32: PatLeaf<(i32 imm), [{
141 uint32_t NV = ~N->getZExtValue();
142 return isPowerOf2_32(NV);
145 def IsPow2_64L: PatLeaf<(i64 imm), [{
146 uint64_t V = N->getZExtValue();
147 return isPowerOf2_64(V) && Log2_64(V) < 32;
150 def IsPow2_64H: PatLeaf<(i64 imm), [{
151 uint64_t V = N->getZExtValue();
152 return isPowerOf2_64(V) && Log2_64(V) >= 32;
155 def IsNPow2_64L: PatLeaf<(i64 imm), [{
156 uint64_t NV = ~N->getZExtValue();
157 return isPowerOf2_64(NV) && Log2_64(NV) < 32;
160 def IsNPow2_64H: PatLeaf<(i64 imm), [{
161 uint64_t NV = ~N->getZExtValue();
162 return isPowerOf2_64(NV) && Log2_64(NV) >= 32;
165 class IsULE<int Width, int Arg>: PatLeaf<(i32 imm),
166 "uint64_t V = N->getZExtValue();" #
167 "return isUInt<" # Width # ">(V) && V <= " # Arg # ";"
170 class IsUGT<int Width, int Arg>: PatLeaf<(i32 imm),
171 "uint64_t V = N->getZExtValue();" #
172 "return isUInt<" # Width # ">(V) && V > " # Arg # ";"
175 def SDEC1: SDNodeXForm<imm, [{
176 int32_t V = N->getSExtValue();
177 return CurDAG->getTargetConstant(V-1, SDLoc(N), MVT::i32);
180 def UDEC1: SDNodeXForm<imm, [{
181 uint32_t V = N->getZExtValue();
183 return CurDAG->getTargetConstant(V-1, SDLoc(N), MVT::i32);
186 def UDEC32: SDNodeXForm<imm, [{
187 uint32_t V = N->getZExtValue();
189 return CurDAG->getTargetConstant(V-32, SDLoc(N), MVT::i32);
192 class Subi<int From>: SDNodeXForm<imm,
193 "int32_t V = " # From # " - N->getSExtValue();" #
194 "return CurDAG->getTargetConstant(V, SDLoc(N), MVT::i32);"
197 def Log2_32: SDNodeXForm<imm, [{
198 uint32_t V = N->getZExtValue();
199 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
202 def Log2_64: SDNodeXForm<imm, [{
203 uint64_t V = N->getZExtValue();
204 return CurDAG->getTargetConstant(Log2_64(V), SDLoc(N), MVT::i32);
207 def LogN2_32: SDNodeXForm<imm, [{
208 uint32_t NV = ~N->getZExtValue();
209 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
212 def LogN2_64: SDNodeXForm<imm, [{
213 uint64_t NV = ~N->getZExtValue();
214 return CurDAG->getTargetConstant(Log2_64(NV), SDLoc(N), MVT::i32);
217 def NegImm8: SDNodeXForm<imm, [{
218 int8_t NV = -N->getSExtValue();
219 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
222 def NegImm16: SDNodeXForm<imm, [{
223 int16_t NV = -N->getSExtValue();
224 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
227 def NegImm32: SDNodeXForm<imm, [{
228 int32_t NV = -N->getSExtValue();
229 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
233 // Helpers for type promotions/contractions.
234 def I1toI32: OutPatFrag<(ops node:$Rs), (C2_muxii (i1 $Rs), 1, 0)>;
235 def I32toI1: OutPatFrag<(ops node:$Rs), (i1 (C2_cmpgtui (i32 $Rs), (i32 0)))>;
236 def ToZext64: OutPatFrag<(ops node:$Rs), (i64 (A4_combineir 0, (i32 $Rs)))>;
237 def ToSext64: OutPatFrag<(ops node:$Rs), (i64 (A2_sxtw (i32 $Rs)))>;
238 def ToAext64: OutPatFrag<(ops node:$Rs),
239 (REG_SEQUENCE DoubleRegs, (i32 (IMPLICIT_DEF)), isub_hi, (i32 $Rs), isub_lo)>;
241 def Combinew: OutPatFrag<(ops node:$Rs, node:$Rt),
242 (REG_SEQUENCE DoubleRegs, $Rs, isub_hi, $Rt, isub_lo)>;
244 def addrga: PatLeaf<(i32 AddrGA:$Addr)>;
245 def addrgp: PatLeaf<(i32 AddrGP:$Addr)>;
246 def anyimm: PatLeaf<(i32 AnyImm:$Imm)>;
247 def anyint: PatLeaf<(i32 AnyInt:$Imm)>;
249 // Global address or an aligned constant.
250 def anyimm0: PatLeaf<(i32 AnyImm0:$Addr)>;
251 def anyimm1: PatLeaf<(i32 AnyImm1:$Addr)>;
252 def anyimm2: PatLeaf<(i32 AnyImm2:$Addr)>;
253 def anyimm3: PatLeaf<(i32 AnyImm3:$Addr)>;
255 def f32ImmPred : PatLeaf<(f32 fpimm:$F)>;
256 def f64ImmPred : PatLeaf<(f64 fpimm:$F)>;
258 // This complex pattern is really only to detect various forms of
259 // sign-extension i32->i64. The selected value will be of type i64
260 // whose low word is the value being extended. The high word is
262 def Usxtw: ComplexPattern<i64, 1, "DetectUseSxtw", [], []>;
264 def Aext64: PatFrag<(ops node:$Rs), (i64 (anyext node:$Rs))>;
265 def Zext64: PatFrag<(ops node:$Rs), (i64 (zext node:$Rs))>;
266 def Sext64: PatLeaf<(i64 Usxtw:$Rs)>;
268 def azext: PatFrags<(ops node:$Rs), [(zext node:$Rs), (anyext node:$Rs)]>;
269 def asext: PatFrags<(ops node:$Rs), [(sext node:$Rs), (anyext node:$Rs)]>;
271 def: Pat<(IsOrAdd (i32 AddrFI:$Rs), s32_0ImmPred:$off),
272 (PS_fi (i32 AddrFI:$Rs), imm:$off)>;
275 // Converters from unary/binary SDNode to PatFrag.
276 class pf1<SDNode Op> : PatFrag<(ops node:$a), (Op node:$a)>;
277 class pf2<SDNode Op> : PatFrag<(ops node:$a, node:$b), (Op node:$a, node:$b)>;
279 class Not2<PatFrag P>
280 : PatFrag<(ops node:$A, node:$B), (P node:$A, (not node:$B))>;
282 // If there is a constant operand that feeds the and/or instruction,
283 // do not generate the compound instructions.
284 // It is not always profitable, as some times we end up with a transfer.
285 // Check the below example.
286 // ra = #65820; rb = lsr(rb, #8); rc ^= and (rb, ra)
287 // Instead this is preferable.
288 // ra = and (#65820, lsr(ra, #8)); rb = xor(rb, ra)
289 class Su_ni1<PatFrag Op>
290 : PatFrag<Op.Operands, !head(Op.Fragments), [{
292 // Check if Op1 is an immediate operand.
293 SDValue Op1 = N->getOperand(1);
294 return !isa<ConstantSDNode>(Op1);
297 Op.OperandTransform>;
300 : PatFrag<Op.Operands, !head(Op.Fragments), [{ return hasOneUse(N); }],
301 Op.OperandTransform>;
303 // Main selection macros.
305 class OpR_R_pat<InstHexagon MI, PatFrag Op, ValueType ResVT, PatFrag RegPred>
306 : Pat<(ResVT (Op RegPred:$Rs)), (MI RegPred:$Rs)>;
308 class OpR_RI_pat<InstHexagon MI, PatFrag Op, ValueType ResType,
309 PatFrag RegPred, PatFrag ImmPred>
310 : Pat<(ResType (Op RegPred:$Rs, ImmPred:$I)),
311 (MI RegPred:$Rs, imm:$I)>;
313 class OpR_RR_pat<InstHexagon MI, PatFrag Op, ValueType ResType,
314 PatFrag RsPred, PatFrag RtPred = RsPred>
315 : Pat<(ResType (Op RsPred:$Rs, RtPred:$Rt)),
316 (MI RsPred:$Rs, RtPred:$Rt)>;
318 class AccRRI_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op,
319 PatFrag RegPred, PatFrag ImmPred>
320 : Pat<(AccOp RegPred:$Rx, (Op RegPred:$Rs, ImmPred:$I)),
321 (MI RegPred:$Rx, RegPred:$Rs, imm:$I)>;
323 class AccRRR_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op,
324 PatFrag RxPred, PatFrag RsPred, PatFrag RtPred>
325 : Pat<(AccOp RxPred:$Rx, (Op RsPred:$Rs, RtPred:$Rt)),
326 (MI RxPred:$Rx, RsPred:$Rs, RtPred:$Rt)>;
328 multiclass SelMinMax_pats<PatFrag CmpOp, PatFrag Val,
329 InstHexagon InstA, InstHexagon InstB> {
330 def: Pat<(select (i1 (CmpOp Val:$A, Val:$B)), Val:$A, Val:$B),
331 (InstA Val:$A, Val:$B)>;
332 def: Pat<(select (i1 (CmpOp Val:$A, Val:$B)), Val:$B, Val:$A),
333 (InstB Val:$A, Val:$B)>;
336 multiclass MinMax_pats<InstHexagon PickT, InstHexagon PickS,
337 PatFrag Sel, PatFrag CmpOp,
338 ValueType CmpType, PatFrag CmpPred> {
339 def: Pat<(Sel (CmpType (CmpOp CmpPred:$Vs, CmpPred:$Vt)),
340 CmpPred:$Vt, CmpPred:$Vs),
341 (PickT CmpPred:$Vs, CmpPred:$Vt)>;
342 def: Pat<(Sel (CmpType (CmpOp CmpPred:$Vs, CmpPred:$Vt)),
343 CmpPred:$Vs, CmpPred:$Vt),
344 (PickS CmpPred:$Vs, CmpPred:$Vt)>;
348 // Frags for commonly used SDNodes.
349 def Add: pf2<add>; def And: pf2<and>; def Sra: pf2<sra>;
350 def Sub: pf2<sub>; def Or: pf2<or>; def Srl: pf2<srl>;
351 def Mul: pf2<mul>; def Xor: pf2<xor>; def Shl: pf2<shl>;
355 // --(1) Immediate -------------------------------------------------------
358 def SDTHexagonCONST32
359 : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisPtrTy<0>]>;
361 def HexagonJT: SDNode<"HexagonISD::JT", SDTIntUnaryOp>;
362 def HexagonCP: SDNode<"HexagonISD::CP", SDTIntUnaryOp>;
363 def HexagonCONST32: SDNode<"HexagonISD::CONST32", SDTHexagonCONST32>;
364 def HexagonCONST32_GP: SDNode<"HexagonISD::CONST32_GP", SDTHexagonCONST32>;
366 def TruncI64ToI32: SDNodeXForm<imm, [{
367 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i32);
370 def: Pat<(s32_0ImmPred:$s16), (A2_tfrsi imm:$s16)>;
371 def: Pat<(s8_0Imm64Pred:$s8), (A2_tfrpi (TruncI64ToI32 $s8))>;
373 def: Pat<(HexagonCONST32 tglobaltlsaddr:$A), (A2_tfrsi imm:$A)>;
374 def: Pat<(HexagonCONST32 bbl:$A), (A2_tfrsi imm:$A)>;
375 def: Pat<(HexagonCONST32 tglobaladdr:$A), (A2_tfrsi imm:$A)>;
376 def: Pat<(HexagonCONST32_GP tblockaddress:$A), (A2_tfrsi imm:$A)>;
377 def: Pat<(HexagonCONST32_GP tglobaladdr:$A), (A2_tfrsi imm:$A)>;
378 def: Pat<(HexagonJT tjumptable:$A), (A2_tfrsi imm:$A)>;
379 def: Pat<(HexagonCP tconstpool:$A), (A2_tfrsi imm:$A)>;
380 // The HVX load patterns also match CP directly. Make sure that if
381 // the selection of this opcode changes, it's updated in all places.
383 def: Pat<(i1 0), (PS_false)>;
384 def: Pat<(i1 1), (PS_true)>;
385 def: Pat<(i64 imm:$v), (CONST64 imm:$v)>;
387 def ftoi : SDNodeXForm<fpimm, [{
388 APInt I = N->getValueAPF().bitcastToAPInt();
389 return CurDAG->getTargetConstant(I.getZExtValue(), SDLoc(N),
390 MVT::getIntegerVT(I.getBitWidth()));
393 def: Pat<(f32ImmPred:$f), (A2_tfrsi (ftoi $f))>;
394 def: Pat<(f64ImmPred:$f), (CONST64 (ftoi $f))>;
396 def ToI32: OutPatFrag<(ops node:$V), (A2_tfrsi $V)>;
398 // --(2) Type cast -------------------------------------------------------
401 def: OpR_R_pat<F2_conv_sf2df, pf1<fpextend>, f64, F32>;
402 def: OpR_R_pat<F2_conv_df2sf, pf1<fpround>, f32, F64>;
404 def: OpR_R_pat<F2_conv_w2sf, pf1<sint_to_fp>, f32, I32>;
405 def: OpR_R_pat<F2_conv_d2sf, pf1<sint_to_fp>, f32, I64>;
406 def: OpR_R_pat<F2_conv_w2df, pf1<sint_to_fp>, f64, I32>;
407 def: OpR_R_pat<F2_conv_d2df, pf1<sint_to_fp>, f64, I64>;
409 def: OpR_R_pat<F2_conv_uw2sf, pf1<uint_to_fp>, f32, I32>;
410 def: OpR_R_pat<F2_conv_ud2sf, pf1<uint_to_fp>, f32, I64>;
411 def: OpR_R_pat<F2_conv_uw2df, pf1<uint_to_fp>, f64, I32>;
412 def: OpR_R_pat<F2_conv_ud2df, pf1<uint_to_fp>, f64, I64>;
414 def: OpR_R_pat<F2_conv_sf2w_chop, pf1<fp_to_sint>, i32, F32>;
415 def: OpR_R_pat<F2_conv_df2w_chop, pf1<fp_to_sint>, i32, F64>;
416 def: OpR_R_pat<F2_conv_sf2d_chop, pf1<fp_to_sint>, i64, F32>;
417 def: OpR_R_pat<F2_conv_df2d_chop, pf1<fp_to_sint>, i64, F64>;
419 def: OpR_R_pat<F2_conv_sf2uw_chop, pf1<fp_to_uint>, i32, F32>;
420 def: OpR_R_pat<F2_conv_df2uw_chop, pf1<fp_to_uint>, i32, F64>;
421 def: OpR_R_pat<F2_conv_sf2ud_chop, pf1<fp_to_uint>, i64, F32>;
422 def: OpR_R_pat<F2_conv_df2ud_chop, pf1<fp_to_uint>, i64, F64>;
424 // Bitcast is different than [fp|sint|uint]_to_[sint|uint|fp].
425 def: Pat<(i32 (bitconvert F32:$v)), (I32:$v)>;
426 def: Pat<(f32 (bitconvert I32:$v)), (F32:$v)>;
427 def: Pat<(i64 (bitconvert F64:$v)), (I64:$v)>;
428 def: Pat<(f64 (bitconvert I64:$v)), (F64:$v)>;
430 multiclass Cast_pat<ValueType Ta, ValueType Tb, RegisterClass RC> {
431 def: Pat<(Tb (bitconvert (Ta RC:$Rs))), (Tb RC:$Rs)>;
432 def: Pat<(Ta (bitconvert (Tb RC:$Rs))), (Ta RC:$Rs)>;
435 // Bit convert vector types to integers.
436 defm: Cast_pat<v4i8, i32, IntRegs>;
437 defm: Cast_pat<v2i16, i32, IntRegs>;
438 defm: Cast_pat<v8i8, i64, DoubleRegs>;
439 defm: Cast_pat<v4i16, i64, DoubleRegs>;
440 defm: Cast_pat<v2i32, i64, DoubleRegs>;
443 // --(3) Extend/truncate -------------------------------------------------
446 def: Pat<(sext_inreg I32:$Rs, i8), (A2_sxtb I32:$Rs)>;
447 def: Pat<(sext_inreg I32:$Rs, i16), (A2_sxth I32:$Rs)>;
448 def: Pat<(sext_inreg I64:$Rs, i32), (A2_sxtw (LoReg $Rs))>;
449 def: Pat<(sext_inreg I64:$Rs, i16), (A2_sxtw (A2_sxth (LoReg $Rs)))>;
450 def: Pat<(sext_inreg I64:$Rs, i8), (A2_sxtw (A2_sxtb (LoReg $Rs)))>;
452 def: Pat<(i64 (sext I32:$Rs)), (A2_sxtw I32:$Rs)>;
453 def: Pat<(Zext64 I32:$Rs), (ToZext64 $Rs)>;
454 def: Pat<(Aext64 I32:$Rs), (ToZext64 $Rs)>;
456 def: Pat<(i32 (trunc I64:$Rs)), (LoReg $Rs)>;
457 def: Pat<(i1 (trunc I32:$Rs)), (S2_tstbit_i I32:$Rs, 0)>;
458 def: Pat<(i1 (trunc I64:$Rs)), (S2_tstbit_i (LoReg $Rs), 0)>;
460 let AddedComplexity = 20 in {
461 def: Pat<(and I32:$Rs, 255), (A2_zxtb I32:$Rs)>;
462 def: Pat<(and I32:$Rs, 65535), (A2_zxth I32:$Rs)>;
465 // Extensions from i1 or vectors of i1.
466 def: Pat<(i32 (azext I1:$Pu)), (C2_muxii I1:$Pu, 1, 0)>;
467 def: Pat<(i64 (azext I1:$Pu)), (ToZext64 (C2_muxii I1:$Pu, 1, 0))>;
468 def: Pat<(i32 (sext I1:$Pu)), (C2_muxii I1:$Pu, -1, 0)>;
469 def: Pat<(i64 (sext I1:$Pu)), (Combinew (C2_muxii PredRegs:$Pu, -1, 0),
470 (C2_muxii PredRegs:$Pu, -1, 0))>;
472 def: Pat<(v2i16 (sext V2I1:$Pu)), (S2_vtrunehb (C2_mask V2I1:$Pu))>;
473 def: Pat<(v2i32 (sext V2I1:$Pu)), (C2_mask V2I1:$Pu)>;
474 def: Pat<(v4i8 (sext V4I1:$Pu)), (S2_vtrunehb (C2_mask V4I1:$Pu))>;
475 def: Pat<(v4i16 (sext V4I1:$Pu)), (C2_mask V4I1:$Pu)>;
476 def: Pat<(v8i8 (sext V8I1:$Pu)), (C2_mask V8I1:$Pu)>;
478 def Vsplatpi: OutPatFrag<(ops node:$V),
479 (Combinew (A2_tfrsi $V), (A2_tfrsi $V))>;
481 def: Pat<(v2i16 (azext V2I1:$Pu)),
482 (A2_andir (LoReg (C2_mask V2I1:$Pu)), (i32 0x00010001))>;
483 def: Pat<(v2i32 (azext V2I1:$Pu)),
484 (A2_andp (C2_mask V2I1:$Pu), (A2_combineii (i32 1), (i32 1)))>;
485 def: Pat<(v4i8 (azext V4I1:$Pu)),
486 (A2_andir (LoReg (C2_mask V4I1:$Pu)), (i32 0x01010101))>;
487 def: Pat<(v4i16 (azext V4I1:$Pu)),
488 (A2_andp (C2_mask V4I1:$Pu), (Vsplatpi (i32 0x00010001)))>;
489 def: Pat<(v8i8 (azext V8I1:$Pu)),
490 (A2_andp (C2_mask V8I1:$Pu), (Vsplatpi (i32 0x01010101)))>;
492 def: Pat<(v4i16 (azext V4I8:$Rs)), (S2_vzxtbh V4I8:$Rs)>;
493 def: Pat<(v2i32 (azext V2I16:$Rs)), (S2_vzxthw V2I16:$Rs)>;
494 def: Pat<(v4i16 (sext V4I8:$Rs)), (S2_vsxtbh V4I8:$Rs)>;
495 def: Pat<(v2i32 (sext V2I16:$Rs)), (S2_vsxthw V2I16:$Rs)>;
497 def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i8)),
498 (Combinew (A2_sxtb (HiReg $Rs)), (A2_sxtb (LoReg $Rs)))>;
500 def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i16)),
501 (Combinew (A2_sxth (HiReg $Rs)), (A2_sxth (LoReg $Rs)))>;
503 // Truncate: from vector B copy all 'E'ven 'B'yte elements:
504 // A[0] = B[0]; A[1] = B[2]; A[2] = B[4]; A[3] = B[6];
505 def: Pat<(v4i8 (trunc V4I16:$Rs)),
506 (S2_vtrunehb V4I16:$Rs)>;
508 // Truncate: from vector B copy all 'O'dd 'B'yte elements:
509 // A[0] = B[1]; A[1] = B[3]; A[2] = B[5]; A[3] = B[7];
512 // Truncate: from vectors B and C copy all 'E'ven 'H'alf-word elements:
513 // A[0] = B[0]; A[1] = B[2]; A[2] = C[0]; A[3] = C[2];
516 def: Pat<(v2i16 (trunc V2I32:$Rs)),
517 (A2_combine_ll (HiReg $Rs), (LoReg $Rs))>;
520 // --(4) Logical ---------------------------------------------------------
523 def: Pat<(not I1:$Ps), (C2_not I1:$Ps)>;
524 def: Pat<(pnot V2I1:$Ps), (C2_not V2I1:$Ps)>;
525 def: Pat<(pnot V4I1:$Ps), (C2_not V4I1:$Ps)>;
526 def: Pat<(pnot V8I1:$Ps), (C2_not V8I1:$Ps)>;
527 def: Pat<(add I1:$Ps, -1), (C2_not I1:$Ps)>;
529 multiclass BoolOpR_RR_pat<InstHexagon MI, PatFrag Op> {
530 def: OpR_RR_pat<MI, Op, i1, I1>;
531 def: OpR_RR_pat<MI, Op, v2i1, V2I1>;
532 def: OpR_RR_pat<MI, Op, v4i1, V4I1>;
533 def: OpR_RR_pat<MI, Op, v8i1, V8I1>;
536 multiclass BoolAccRRR_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op> {
537 def: AccRRR_pat<MI, AccOp, Op, I1, I1, I1>;
538 def: AccRRR_pat<MI, AccOp, Op, V2I1, V2I1, V2I1>;
539 def: AccRRR_pat<MI, AccOp, Op, V4I1, V4I1, V4I1>;
540 def: AccRRR_pat<MI, AccOp, Op, V8I1, V8I1, V8I1>;
543 defm: BoolOpR_RR_pat<C2_and, And>;
544 defm: BoolOpR_RR_pat<C2_or, Or>;
545 defm: BoolOpR_RR_pat<C2_xor, Xor>;
546 defm: BoolOpR_RR_pat<C2_andn, Not2<And>>;
547 defm: BoolOpR_RR_pat<C2_orn, Not2<Or>>;
549 // op(Ps, op(Pt, Pu))
550 defm: BoolAccRRR_pat<C4_and_and, And, Su<And>>;
551 defm: BoolAccRRR_pat<C4_and_or, And, Su<Or>>;
552 defm: BoolAccRRR_pat<C4_or_and, Or, Su<And>>;
553 defm: BoolAccRRR_pat<C4_or_or, Or, Su<Or>>;
555 // op(Ps, op(Pt, ~Pu))
556 defm: BoolAccRRR_pat<C4_and_andn, And, Su<Not2<And>>>;
557 defm: BoolAccRRR_pat<C4_and_orn, And, Su<Not2<Or>>>;
558 defm: BoolAccRRR_pat<C4_or_andn, Or, Su<Not2<And>>>;
559 defm: BoolAccRRR_pat<C4_or_orn, Or, Su<Not2<Or>>>;
562 // --(5) Compare ---------------------------------------------------------
565 // Avoid negated comparisons, i.e. those of form "Pd = !cmp(...)".
566 // These cannot form compounds (e.g. J4_cmpeqi_tp0_jump_nt).
568 def: OpR_RI_pat<C2_cmpeqi, seteq, i1, I32, anyimm>;
569 def: OpR_RI_pat<C2_cmpgti, setgt, i1, I32, anyimm>;
570 def: OpR_RI_pat<C2_cmpgtui, setugt, i1, I32, anyimm>;
572 def: Pat<(i1 (setge I32:$Rs, s32_0ImmPred:$s10)),
573 (C2_cmpgti I32:$Rs, (SDEC1 imm:$s10))>;
574 def: Pat<(i1 (setuge I32:$Rs, u32_0ImmPred:$u9)),
575 (C2_cmpgtui I32:$Rs, (UDEC1 imm:$u9))>;
577 def: Pat<(i1 (setlt I32:$Rs, s32_0ImmPred:$s10)),
578 (C2_not (C2_cmpgti I32:$Rs, (SDEC1 imm:$s10)))>;
579 def: Pat<(i1 (setult I32:$Rs, u32_0ImmPred:$u9)),
580 (C2_not (C2_cmpgtui I32:$Rs, (UDEC1 imm:$u9)))>;
582 // Patfrag to convert the usual comparison patfrags (e.g. setlt) to ones
583 // that reverse the order of the operands.
584 class RevCmp<PatFrag F>
585 : PatFrag<(ops node:$rhs, node:$lhs), !head(F.Fragments), F.PredicateCode,
588 def: OpR_RR_pat<C2_cmpeq, seteq, i1, I32>;
589 def: OpR_RR_pat<C2_cmpgt, setgt, i1, I32>;
590 def: OpR_RR_pat<C2_cmpgtu, setugt, i1, I32>;
591 def: OpR_RR_pat<C2_cmpgt, RevCmp<setlt>, i1, I32>;
592 def: OpR_RR_pat<C2_cmpgtu, RevCmp<setult>, i1, I32>;
593 def: OpR_RR_pat<C2_cmpeqp, seteq, i1, I64>;
594 def: OpR_RR_pat<C2_cmpgtp, setgt, i1, I64>;
595 def: OpR_RR_pat<C2_cmpgtup, setugt, i1, I64>;
596 def: OpR_RR_pat<C2_cmpgtp, RevCmp<setlt>, i1, I64>;
597 def: OpR_RR_pat<C2_cmpgtup, RevCmp<setult>, i1, I64>;
598 def: OpR_RR_pat<A2_vcmpbeq, seteq, i1, V8I8>;
599 def: OpR_RR_pat<A2_vcmpbeq, seteq, v8i1, V8I8>;
600 def: OpR_RR_pat<A4_vcmpbgt, RevCmp<setlt>, i1, V8I8>;
601 def: OpR_RR_pat<A4_vcmpbgt, RevCmp<setlt>, v8i1, V8I8>;
602 def: OpR_RR_pat<A4_vcmpbgt, setgt, i1, V8I8>;
603 def: OpR_RR_pat<A4_vcmpbgt, setgt, v8i1, V8I8>;
604 def: OpR_RR_pat<A2_vcmpbgtu, RevCmp<setult>, i1, V8I8>;
605 def: OpR_RR_pat<A2_vcmpbgtu, RevCmp<setult>, v8i1, V8I8>;
606 def: OpR_RR_pat<A2_vcmpbgtu, setugt, i1, V8I8>;
607 def: OpR_RR_pat<A2_vcmpbgtu, setugt, v8i1, V8I8>;
608 def: OpR_RR_pat<A2_vcmpheq, seteq, i1, V4I16>;
609 def: OpR_RR_pat<A2_vcmpheq, seteq, v4i1, V4I16>;
610 def: OpR_RR_pat<A2_vcmphgt, RevCmp<setlt>, i1, V4I16>;
611 def: OpR_RR_pat<A2_vcmphgt, RevCmp<setlt>, v4i1, V4I16>;
612 def: OpR_RR_pat<A2_vcmphgt, setgt, i1, V4I16>;
613 def: OpR_RR_pat<A2_vcmphgt, setgt, v4i1, V4I16>;
614 def: OpR_RR_pat<A2_vcmphgtu, RevCmp<setult>, i1, V4I16>;
615 def: OpR_RR_pat<A2_vcmphgtu, RevCmp<setult>, v4i1, V4I16>;
616 def: OpR_RR_pat<A2_vcmphgtu, setugt, i1, V4I16>;
617 def: OpR_RR_pat<A2_vcmphgtu, setugt, v4i1, V4I16>;
618 def: OpR_RR_pat<A2_vcmpweq, seteq, i1, V2I32>;
619 def: OpR_RR_pat<A2_vcmpweq, seteq, v2i1, V2I32>;
620 def: OpR_RR_pat<A2_vcmpwgt, RevCmp<setlt>, i1, V2I32>;
621 def: OpR_RR_pat<A2_vcmpwgt, RevCmp<setlt>, v2i1, V2I32>;
622 def: OpR_RR_pat<A2_vcmpwgt, setgt, i1, V2I32>;
623 def: OpR_RR_pat<A2_vcmpwgt, setgt, v2i1, V2I32>;
624 def: OpR_RR_pat<A2_vcmpwgtu, RevCmp<setult>, i1, V2I32>;
625 def: OpR_RR_pat<A2_vcmpwgtu, RevCmp<setult>, v2i1, V2I32>;
626 def: OpR_RR_pat<A2_vcmpwgtu, setugt, i1, V2I32>;
627 def: OpR_RR_pat<A2_vcmpwgtu, setugt, v2i1, V2I32>;
629 def: OpR_RR_pat<F2_sfcmpeq, seteq, i1, F32>;
630 def: OpR_RR_pat<F2_sfcmpgt, setgt, i1, F32>;
631 def: OpR_RR_pat<F2_sfcmpge, setge, i1, F32>;
632 def: OpR_RR_pat<F2_sfcmpeq, setoeq, i1, F32>;
633 def: OpR_RR_pat<F2_sfcmpgt, setogt, i1, F32>;
634 def: OpR_RR_pat<F2_sfcmpge, setoge, i1, F32>;
635 def: OpR_RR_pat<F2_sfcmpgt, RevCmp<setolt>, i1, F32>;
636 def: OpR_RR_pat<F2_sfcmpge, RevCmp<setole>, i1, F32>;
637 def: OpR_RR_pat<F2_sfcmpgt, RevCmp<setlt>, i1, F32>;
638 def: OpR_RR_pat<F2_sfcmpge, RevCmp<setle>, i1, F32>;
639 def: OpR_RR_pat<F2_sfcmpuo, setuo, i1, F32>;
641 def: OpR_RR_pat<F2_dfcmpeq, seteq, i1, F64>;
642 def: OpR_RR_pat<F2_dfcmpgt, setgt, i1, F64>;
643 def: OpR_RR_pat<F2_dfcmpge, setge, i1, F64>;
644 def: OpR_RR_pat<F2_dfcmpeq, setoeq, i1, F64>;
645 def: OpR_RR_pat<F2_dfcmpgt, setogt, i1, F64>;
646 def: OpR_RR_pat<F2_dfcmpge, setoge, i1, F64>;
647 def: OpR_RR_pat<F2_dfcmpgt, RevCmp<setolt>, i1, F64>;
648 def: OpR_RR_pat<F2_dfcmpge, RevCmp<setole>, i1, F64>;
649 def: OpR_RR_pat<F2_dfcmpgt, RevCmp<setlt>, i1, F64>;
650 def: OpR_RR_pat<F2_dfcmpge, RevCmp<setle>, i1, F64>;
651 def: OpR_RR_pat<F2_dfcmpuo, setuo, i1, F64>;
653 // Avoid C4_cmpneqi, C4_cmpltei, C4_cmplteui, since they cannot form compounds.
655 def: Pat<(i1 (setne I32:$Rs, anyimm:$u5)),
656 (C2_not (C2_cmpeqi I32:$Rs, imm:$u5))>;
657 def: Pat<(i1 (setle I32:$Rs, anyimm:$u5)),
658 (C2_not (C2_cmpgti I32:$Rs, imm:$u5))>;
659 def: Pat<(i1 (setule I32:$Rs, anyimm:$u5)),
660 (C2_not (C2_cmpgtui I32:$Rs, imm:$u5))>;
662 class OpmR_RR_pat<PatFrag Output, PatFrag Op, ValueType ResType,
663 PatFrag RsPred, PatFrag RtPred = RsPred>
664 : Pat<(ResType (Op RsPred:$Rs, RtPred:$Rt)),
665 (Output RsPred:$Rs, RtPred:$Rt)>;
667 class Outn<InstHexagon MI>
668 : OutPatFrag<(ops node:$Rs, node:$Rt),
669 (C2_not (MI $Rs, $Rt))>;
671 def: OpmR_RR_pat<Outn<C2_cmpeq>, setne, i1, I32>;
672 def: OpmR_RR_pat<Outn<C2_cmpgt>, setle, i1, I32>;
673 def: OpmR_RR_pat<Outn<C2_cmpgtu>, setule, i1, I32>;
674 def: OpmR_RR_pat<Outn<C2_cmpgt>, RevCmp<setge>, i1, I32>;
675 def: OpmR_RR_pat<Outn<C2_cmpgtu>, RevCmp<setuge>, i1, I32>;
676 def: OpmR_RR_pat<Outn<C2_cmpeqp>, setne, i1, I64>;
677 def: OpmR_RR_pat<Outn<C2_cmpgtp>, setle, i1, I64>;
678 def: OpmR_RR_pat<Outn<C2_cmpgtup>, setule, i1, I64>;
679 def: OpmR_RR_pat<Outn<C2_cmpgtp>, RevCmp<setge>, i1, I64>;
680 def: OpmR_RR_pat<Outn<C2_cmpgtup>, RevCmp<setuge>, i1, I64>;
681 def: OpmR_RR_pat<Outn<A2_vcmpbeq>, setne, v8i1, V8I8>;
682 def: OpmR_RR_pat<Outn<A4_vcmpbgt>, setle, v8i1, V8I8>;
683 def: OpmR_RR_pat<Outn<A2_vcmpbgtu>, setule, v8i1, V8I8>;
684 def: OpmR_RR_pat<Outn<A4_vcmpbgt>, RevCmp<setge>, v8i1, V8I8>;
685 def: OpmR_RR_pat<Outn<A2_vcmpbgtu>, RevCmp<setuge>, v8i1, V8I8>;
686 def: OpmR_RR_pat<Outn<A2_vcmpheq>, setne, v4i1, V4I16>;
687 def: OpmR_RR_pat<Outn<A2_vcmphgt>, setle, v4i1, V4I16>;
688 def: OpmR_RR_pat<Outn<A2_vcmphgtu>, setule, v4i1, V4I16>;
689 def: OpmR_RR_pat<Outn<A2_vcmphgt>, RevCmp<setge>, v4i1, V4I16>;
690 def: OpmR_RR_pat<Outn<A2_vcmphgtu>, RevCmp<setuge>, v4i1, V4I16>;
691 def: OpmR_RR_pat<Outn<A2_vcmpweq>, setne, v2i1, V2I32>;
692 def: OpmR_RR_pat<Outn<A2_vcmpwgt>, setle, v2i1, V2I32>;
693 def: OpmR_RR_pat<Outn<A2_vcmpwgtu>, setule, v2i1, V2I32>;
694 def: OpmR_RR_pat<Outn<A2_vcmpwgt>, RevCmp<setge>, v2i1, V2I32>;
695 def: OpmR_RR_pat<Outn<A2_vcmpwgtu>, RevCmp<setuge>, v2i1, V2I32>;
697 let AddedComplexity = 100 in {
698 def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt), 255), 0)),
699 (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt)>;
700 def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt), 255), 0)),
701 (C2_not (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt))>;
702 def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt), 65535), 0)),
703 (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt)>;
704 def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt), 65535), 0)),
705 (C2_not (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt))>;
708 // PatFrag for AsserZext which takes the original type as a parameter.
709 def SDTAssertZext: SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisSameAs<0,1>]>;
710 def AssertZextSD: SDNode<"ISD::AssertZext", SDTAssertZext>;
711 class AssertZext<ValueType T>: PatFrag<(ops node:$A), (AssertZextSD $A, T)>;
713 multiclass Cmpb_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
714 PatLeaf ImmPred, int Mask> {
715 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
716 (MI I32:$Rs, imm:$I)>;
717 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
718 (MI I32:$Rs, imm:$I)>;
721 multiclass CmpbN_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
722 PatLeaf ImmPred, int Mask> {
723 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
724 (C2_not (MI I32:$Rs, imm:$I))>;
725 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
726 (C2_not (MI I32:$Rs, imm:$I))>;
729 multiclass CmpbND_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
730 PatLeaf ImmPred, int Mask> {
731 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
732 (C2_not (MI I32:$Rs, (UDEC1 imm:$I)))>;
733 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
734 (C2_not (MI I32:$Rs, (UDEC1 imm:$I)))>;
737 let AddedComplexity = 200 in {
738 defm: Cmpb_pat <A4_cmpbeqi, seteq, AssertZext<i8>, IsUGT<8,31>, 255>;
739 defm: CmpbN_pat <A4_cmpbeqi, setne, AssertZext<i8>, IsUGT<8,31>, 255>;
740 defm: Cmpb_pat <A4_cmpbgtui, setugt, AssertZext<i8>, IsUGT<32,31>, 255>;
741 defm: CmpbN_pat <A4_cmpbgtui, setule, AssertZext<i8>, IsUGT<32,31>, 255>;
742 defm: Cmpb_pat <A4_cmphgtui, setugt, AssertZext<i16>, IsUGT<32,31>, 65535>;
743 defm: CmpbN_pat <A4_cmphgtui, setule, AssertZext<i16>, IsUGT<32,31>, 65535>;
744 defm: CmpbND_pat<A4_cmpbgtui, setult, AssertZext<i8>, IsUGT<32,32>, 255>;
745 defm: CmpbND_pat<A4_cmphgtui, setult, AssertZext<i16>, IsUGT<32,32>, 65535>;
748 def: Pat<(i32 (zext (i1 (seteq I32:$Rs, I32:$Rt)))),
749 (A4_rcmpeq I32:$Rs, I32:$Rt)>;
750 def: Pat<(i32 (zext (i1 (setne I32:$Rs, I32:$Rt)))),
751 (A4_rcmpneq I32:$Rs, I32:$Rt)>;
752 def: Pat<(i32 (zext (i1 (seteq I32:$Rs, anyimm:$s8)))),
753 (A4_rcmpeqi I32:$Rs, imm:$s8)>;
754 def: Pat<(i32 (zext (i1 (setne I32:$Rs, anyimm:$s8)))),
755 (A4_rcmpneqi I32:$Rs, imm:$s8)>;
757 def: Pat<(i1 (seteq I1:$Ps, (i1 -1))), (I1:$Ps)>;
758 def: Pat<(i1 (setne I1:$Ps, (i1 -1))), (C2_not I1:$Ps)>;
759 def: Pat<(i1 (seteq I1:$Ps, I1:$Pt)), (C2_xor I1:$Ps, (C2_not I1:$Pt))>;
760 def: Pat<(i1 (setne I1:$Ps, I1:$Pt)), (C2_xor I1:$Ps, I1:$Pt)>;
762 // Floating-point comparisons with checks for ordered/unordered status.
764 class T3<InstHexagon MI1, InstHexagon MI2, InstHexagon MI3>
765 : OutPatFrag<(ops node:$Rs, node:$Rt),
766 (MI1 (MI2 $Rs, $Rt), (MI3 $Rs, $Rt))>;
768 class Cmpuf<InstHexagon MI>: T3<C2_or, F2_sfcmpuo, MI>;
769 class Cmpud<InstHexagon MI>: T3<C2_or, F2_dfcmpuo, MI>;
771 class Cmpufn<InstHexagon MI>: T3<C2_orn, F2_sfcmpuo, MI>;
772 class Cmpudn<InstHexagon MI>: T3<C2_orn, F2_dfcmpuo, MI>;
774 def: OpmR_RR_pat<Cmpuf<F2_sfcmpeq>, setueq, i1, F32>;
775 def: OpmR_RR_pat<Cmpuf<F2_sfcmpge>, setuge, i1, F32>;
776 def: OpmR_RR_pat<Cmpuf<F2_sfcmpgt>, setugt, i1, F32>;
777 def: OpmR_RR_pat<Cmpuf<F2_sfcmpge>, RevCmp<setule>, i1, F32>;
778 def: OpmR_RR_pat<Cmpuf<F2_sfcmpgt>, RevCmp<setult>, i1, F32>;
779 def: OpmR_RR_pat<Cmpufn<F2_sfcmpeq>, setune, i1, F32>;
781 def: OpmR_RR_pat<Cmpud<F2_dfcmpeq>, setueq, i1, F64>;
782 def: OpmR_RR_pat<Cmpud<F2_dfcmpge>, setuge, i1, F64>;
783 def: OpmR_RR_pat<Cmpud<F2_dfcmpgt>, setugt, i1, F64>;
784 def: OpmR_RR_pat<Cmpud<F2_dfcmpge>, RevCmp<setule>, i1, F64>;
785 def: OpmR_RR_pat<Cmpud<F2_dfcmpgt>, RevCmp<setult>, i1, F64>;
786 def: OpmR_RR_pat<Cmpudn<F2_dfcmpeq>, setune, i1, F64>;
788 def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setone, i1, F32>;
789 def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setne, i1, F32>;
791 def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setone, i1, F64>;
792 def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setne, i1, F64>;
794 def: OpmR_RR_pat<Outn<F2_sfcmpuo>, seto, i1, F32>;
795 def: OpmR_RR_pat<Outn<F2_dfcmpuo>, seto, i1, F64>;
798 // --(6) Select ----------------------------------------------------------
801 def: Pat<(select I1:$Pu, I32:$Rs, I32:$Rt),
802 (C2_mux I1:$Pu, I32:$Rs, I32:$Rt)>;
803 def: Pat<(select I1:$Pu, anyimm:$s8, I32:$Rs),
804 (C2_muxri I1:$Pu, imm:$s8, I32:$Rs)>;
805 def: Pat<(select I1:$Pu, I32:$Rs, anyimm:$s8),
806 (C2_muxir I1:$Pu, I32:$Rs, imm:$s8)>;
807 def: Pat<(select I1:$Pu, anyimm:$s8, s8_0ImmPred:$S8),
808 (C2_muxii I1:$Pu, imm:$s8, imm:$S8)>;
810 def: Pat<(select (not I1:$Pu), I32:$Rs, I32:$Rt),
811 (C2_mux I1:$Pu, I32:$Rt, I32:$Rs)>;
812 def: Pat<(select (not I1:$Pu), s8_0ImmPred:$S8, anyimm:$s8),
813 (C2_muxii I1:$Pu, imm:$s8, imm:$S8)>;
814 def: Pat<(select (not I1:$Pu), anyimm:$s8, I32:$Rs),
815 (C2_muxir I1:$Pu, I32:$Rs, imm:$s8)>;
816 def: Pat<(select (not I1:$Pu), I32:$Rs, anyimm:$s8),
817 (C2_muxri I1:$Pu, imm:$s8, I32:$Rs)>;
819 // Map from a 64-bit select to an emulated 64-bit mux.
820 // Hexagon does not support 64-bit MUXes; so emulate with combines.
821 def: Pat<(select I1:$Pu, I64:$Rs, I64:$Rt),
822 (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
823 (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
825 def: Pat<(select I1:$Pu, F32:$Rs, f32ImmPred:$I),
826 (C2_muxir I1:$Pu, F32:$Rs, (ftoi $I))>;
827 def: Pat<(select I1:$Pu, f32ImmPred:$I, F32:$Rt),
828 (C2_muxri I1:$Pu, (ftoi $I), F32:$Rt)>;
829 def: Pat<(select I1:$Pu, F32:$Rs, F32:$Rt),
830 (C2_mux I1:$Pu, F32:$Rs, F32:$Rt)>;
831 def: Pat<(select I1:$Pu, F64:$Rs, F64:$Rt),
832 (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
833 (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
835 def: Pat<(select (i1 (setult F32:$Ra, F32:$Rb)), F32:$Rs, F32:$Rt),
836 (C2_mux (F2_sfcmpgt F32:$Rb, F32:$Ra), F32:$Rs, F32:$Rt)>;
837 def: Pat<(select (i1 (setult F64:$Ra, F64:$Rb)), F64:$Rs, F64:$Rt),
838 (C2_vmux (F2_dfcmpgt F64:$Rb, F64:$Ra), F64:$Rs, F64:$Rt)>;
840 def: Pat<(select (not I1:$Pu), f32ImmPred:$I, F32:$Rs),
841 (C2_muxir I1:$Pu, F32:$Rs, (ftoi $I))>;
842 def: Pat<(select (not I1:$Pu), F32:$Rt, f32ImmPred:$I),
843 (C2_muxri I1:$Pu, (ftoi $I), F32:$Rt)>;
845 def: Pat<(vselect V8I1:$Pu, V8I8:$Rs, V8I8:$Rt),
846 (C2_vmux V8I1:$Pu, V8I8:$Rs, V8I8:$Rt)>;
847 def: Pat<(vselect V4I1:$Pu, V4I16:$Rs, V4I16:$Rt),
848 (C2_vmux V4I1:$Pu, V4I16:$Rs, V4I16:$Rt)>;
849 def: Pat<(vselect V2I1:$Pu, V2I32:$Rs, V2I32:$Rt),
850 (C2_vmux V2I1:$Pu, V2I32:$Rs, V2I32:$Rt)>;
852 def: Pat<(vselect (pnot V8I1:$Pu), V8I8:$Rs, V8I8:$Rt),
853 (C2_vmux V8I1:$Pu, V8I8:$Rt, V8I8:$Rs)>;
854 def: Pat<(vselect (pnot V4I1:$Pu), V4I16:$Rs, V4I16:$Rt),
855 (C2_vmux V4I1:$Pu, V4I16:$Rt, V4I16:$Rs)>;
856 def: Pat<(vselect (pnot V2I1:$Pu), V2I32:$Rs, V2I32:$Rt),
857 (C2_vmux V2I1:$Pu, V2I32:$Rt, V2I32:$Rs)>;
860 // From LegalizeDAG.cpp: (Pu ? Pv : Pw) <=> (Pu & Pv) | (!Pu & Pw).
861 def: Pat<(select I1:$Pu, I1:$Pv, I1:$Pw),
862 (C2_or (C2_and I1:$Pu, I1:$Pv),
863 (C2_andn I1:$Pw, I1:$Pu))>;
866 def IsPosHalf : PatLeaf<(i32 IntRegs:$a), [{
867 return isPositiveHalfWord(N);
870 multiclass SelMinMax16_pats<PatFrag CmpOp, InstHexagon InstA,
872 def: Pat<(sext_inreg (select (i1 (CmpOp IsPosHalf:$Rs, IsPosHalf:$Rt)),
873 IsPosHalf:$Rs, IsPosHalf:$Rt), i16),
874 (InstA IntRegs:$Rs, IntRegs:$Rt)>;
875 def: Pat<(sext_inreg (select (i1 (CmpOp IsPosHalf:$Rs, IsPosHalf:$Rt)),
876 IsPosHalf:$Rt, IsPosHalf:$Rs), i16),
877 (InstB IntRegs:$Rs, IntRegs:$Rt)>;
880 let AddedComplexity = 200 in {
881 defm: SelMinMax16_pats<setge, A2_max, A2_min>;
882 defm: SelMinMax16_pats<setgt, A2_max, A2_min>;
883 defm: SelMinMax16_pats<setle, A2_min, A2_max>;
884 defm: SelMinMax16_pats<setlt, A2_min, A2_max>;
885 defm: SelMinMax16_pats<setuge, A2_maxu, A2_minu>;
886 defm: SelMinMax16_pats<setugt, A2_maxu, A2_minu>;
887 defm: SelMinMax16_pats<setule, A2_minu, A2_maxu>;
888 defm: SelMinMax16_pats<setult, A2_minu, A2_maxu>;
891 let AddedComplexity = 200 in {
892 defm: MinMax_pats<A2_min, A2_max, select, setgt, i1, I32>;
893 defm: MinMax_pats<A2_min, A2_max, select, setge, i1, I32>;
894 defm: MinMax_pats<A2_max, A2_min, select, setlt, i1, I32>;
895 defm: MinMax_pats<A2_max, A2_min, select, setle, i1, I32>;
896 defm: MinMax_pats<A2_minu, A2_maxu, select, setugt, i1, I32>;
897 defm: MinMax_pats<A2_minu, A2_maxu, select, setuge, i1, I32>;
898 defm: MinMax_pats<A2_maxu, A2_minu, select, setult, i1, I32>;
899 defm: MinMax_pats<A2_maxu, A2_minu, select, setule, i1, I32>;
901 defm: MinMax_pats<A2_minp, A2_maxp, select, setgt, i1, I64>;
902 defm: MinMax_pats<A2_minp, A2_maxp, select, setge, i1, I64>;
903 defm: MinMax_pats<A2_maxp, A2_minp, select, setlt, i1, I64>;
904 defm: MinMax_pats<A2_maxp, A2_minp, select, setle, i1, I64>;
905 defm: MinMax_pats<A2_minup, A2_maxup, select, setugt, i1, I64>;
906 defm: MinMax_pats<A2_minup, A2_maxup, select, setuge, i1, I64>;
907 defm: MinMax_pats<A2_maxup, A2_minup, select, setult, i1, I64>;
908 defm: MinMax_pats<A2_maxup, A2_minup, select, setule, i1, I64>;
911 let AddedComplexity = 100 in {
912 defm: MinMax_pats<F2_sfmin, F2_sfmax, select, setogt, i1, F32>;
913 defm: MinMax_pats<F2_sfmin, F2_sfmax, select, setoge, i1, F32>;
914 defm: MinMax_pats<F2_sfmax, F2_sfmin, select, setolt, i1, F32>;
915 defm: MinMax_pats<F2_sfmax, F2_sfmin, select, setole, i1, F32>;
918 defm: MinMax_pats<A2_vminb, A2_vmaxb, vselect, setgt, v8i1, V8I8>;
919 defm: MinMax_pats<A2_vminb, A2_vmaxb, vselect, setge, v8i1, V8I8>;
920 defm: MinMax_pats<A2_vminh, A2_vmaxh, vselect, setgt, v4i1, V4I16>;
921 defm: MinMax_pats<A2_vminh, A2_vmaxh, vselect, setge, v4i1, V4I16>;
922 defm: MinMax_pats<A2_vminw, A2_vmaxw, vselect, setgt, v2i1, V2I32>;
923 defm: MinMax_pats<A2_vminw, A2_vmaxw, vselect, setge, v2i1, V2I32>;
924 defm: MinMax_pats<A2_vminub, A2_vmaxub, vselect, setugt, v8i1, V8I8>;
925 defm: MinMax_pats<A2_vminub, A2_vmaxub, vselect, setuge, v8i1, V8I8>;
926 defm: MinMax_pats<A2_vminuh, A2_vmaxuh, vselect, setugt, v4i1, V4I16>;
927 defm: MinMax_pats<A2_vminuh, A2_vmaxuh, vselect, setuge, v4i1, V4I16>;
928 defm: MinMax_pats<A2_vminuw, A2_vmaxuw, vselect, setugt, v2i1, V2I32>;
929 defm: MinMax_pats<A2_vminuw, A2_vmaxuw, vselect, setuge, v2i1, V2I32>;
931 // --(7) Insert/extract --------------------------------------------------
934 def SDTHexagonINSERT:
935 SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>,
936 SDTCisInt<0>, SDTCisVT<3, i32>, SDTCisVT<4, i32>]>;
937 def HexagonINSERT: SDNode<"HexagonISD::INSERT", SDTHexagonINSERT>;
939 let AddedComplexity = 10 in {
940 def: Pat<(HexagonINSERT I32:$Rs, I32:$Rt, u5_0ImmPred:$u1, u5_0ImmPred:$u2),
941 (S2_insert I32:$Rs, I32:$Rt, imm:$u1, imm:$u2)>;
942 def: Pat<(HexagonINSERT I64:$Rs, I64:$Rt, u6_0ImmPred:$u1, u6_0ImmPred:$u2),
943 (S2_insertp I64:$Rs, I64:$Rt, imm:$u1, imm:$u2)>;
945 def: Pat<(HexagonINSERT I32:$Rs, I32:$Rt, I32:$Width, I32:$Off),
946 (S2_insert_rp I32:$Rs, I32:$Rt, (Combinew $Width, $Off))>;
947 def: Pat<(HexagonINSERT I64:$Rs, I64:$Rt, I32:$Width, I32:$Off),
948 (S2_insertp_rp I64:$Rs, I64:$Rt, (Combinew $Width, $Off))>;
950 def SDTHexagonEXTRACTU
951 : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisInt<0>, SDTCisInt<1>,
952 SDTCisVT<2, i32>, SDTCisVT<3, i32>]>;
953 def HexagonEXTRACTU: SDNode<"HexagonISD::EXTRACTU", SDTHexagonEXTRACTU>;
955 let AddedComplexity = 10 in {
956 def: Pat<(HexagonEXTRACTU I32:$Rs, u5_0ImmPred:$u5, u5_0ImmPred:$U5),
957 (S2_extractu I32:$Rs, imm:$u5, imm:$U5)>;
958 def: Pat<(HexagonEXTRACTU I64:$Rs, u6_0ImmPred:$u6, u6_0ImmPred:$U6),
959 (S2_extractup I64:$Rs, imm:$u6, imm:$U6)>;
961 def: Pat<(HexagonEXTRACTU I32:$Rs, I32:$Width, I32:$Off),
962 (S2_extractu_rp I32:$Rs, (Combinew $Width, $Off))>;
963 def: Pat<(HexagonEXTRACTU I64:$Rs, I32:$Width, I32:$Off),
964 (S2_extractup_rp I64:$Rs, (Combinew $Width, $Off))>;
966 def SDTHexagonVSPLAT:
967 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVT<1, i32>]>;
969 def HexagonVSPLAT: SDNode<"HexagonISD::VSPLAT", SDTHexagonVSPLAT>;
971 def: Pat<(v4i8 (HexagonVSPLAT I32:$Rs)), (S2_vsplatrb I32:$Rs)>;
972 def: Pat<(v4i16 (HexagonVSPLAT I32:$Rs)), (S2_vsplatrh I32:$Rs)>;
973 def: Pat<(v2i32 (HexagonVSPLAT s8_0ImmPred:$s8)),
974 (A2_combineii imm:$s8, imm:$s8)>;
975 def: Pat<(v2i32 (HexagonVSPLAT I32:$Rs)), (Combinew I32:$Rs, I32:$Rs)>;
977 let AddedComplexity = 10 in
978 def: Pat<(v8i8 (HexagonVSPLAT I32:$Rs)), (S6_vsplatrbp I32:$Rs)>,
980 def: Pat<(v8i8 (HexagonVSPLAT I32:$Rs)),
981 (Combinew (S2_vsplatrb I32:$Rs), (S2_vsplatrb I32:$Rs))>;
984 // --(8) Shift/permute ---------------------------------------------------
987 def SDTHexagonI64I32I32: SDTypeProfile<1, 2,
988 [SDTCisVT<0, i64>, SDTCisVT<1, i32>, SDTCisSameAs<1, 2>]>;
990 def HexagonCOMBINE: SDNode<"HexagonISD::COMBINE", SDTHexagonI64I32I32>;
992 def: Pat<(HexagonCOMBINE I32:$Rs, I32:$Rt), (Combinew $Rs, $Rt)>;
994 // The complexity of the combines involving immediates should be greater
995 // than the complexity of the combine with two registers.
996 let AddedComplexity = 50 in {
997 def: Pat<(HexagonCOMBINE I32:$Rs, anyimm:$s8),
998 (A4_combineri IntRegs:$Rs, imm:$s8)>;
999 def: Pat<(HexagonCOMBINE anyimm:$s8, I32:$Rs),
1000 (A4_combineir imm:$s8, IntRegs:$Rs)>;
1003 // The complexity of the combine with two immediates should be greater than
1004 // the complexity of a combine involving a register.
1005 let AddedComplexity = 75 in {
1006 def: Pat<(HexagonCOMBINE s8_0ImmPred:$s8, anyimm:$u6),
1007 (A4_combineii imm:$s8, imm:$u6)>;
1008 def: Pat<(HexagonCOMBINE anyimm:$s8, s8_0ImmPred:$S8),
1009 (A2_combineii imm:$s8, imm:$S8)>;
1012 def: Pat<(bswap I32:$Rs), (A2_swiz I32:$Rs)>;
1013 def: Pat<(bswap I64:$Rss), (Combinew (A2_swiz (LoReg $Rss)),
1014 (A2_swiz (HiReg $Rss)))>;
1016 def: Pat<(shl s6_0ImmPred:$s6, I32:$Rt), (S4_lsli imm:$s6, I32:$Rt)>;
1017 def: Pat<(shl I32:$Rs, (i32 16)), (A2_aslh I32:$Rs)>;
1018 def: Pat<(sra I32:$Rs, (i32 16)), (A2_asrh I32:$Rs)>;
1020 def: OpR_RI_pat<S2_asr_i_r, Sra, i32, I32, u5_0ImmPred>;
1021 def: OpR_RI_pat<S2_lsr_i_r, Srl, i32, I32, u5_0ImmPred>;
1022 def: OpR_RI_pat<S2_asl_i_r, Shl, i32, I32, u5_0ImmPred>;
1023 def: OpR_RI_pat<S2_asr_i_p, Sra, i64, I64, u6_0ImmPred>;
1024 def: OpR_RI_pat<S2_lsr_i_p, Srl, i64, I64, u6_0ImmPred>;
1025 def: OpR_RI_pat<S2_asl_i_p, Shl, i64, I64, u6_0ImmPred>;
1026 def: OpR_RI_pat<S2_asr_i_vh, Sra, v4i16, V4I16, u4_0ImmPred>;
1027 def: OpR_RI_pat<S2_lsr_i_vh, Srl, v4i16, V4I16, u4_0ImmPred>;
1028 def: OpR_RI_pat<S2_asl_i_vh, Shl, v4i16, V4I16, u4_0ImmPred>;
1029 def: OpR_RI_pat<S2_asr_i_vh, Sra, v2i32, V2I32, u5_0ImmPred>;
1030 def: OpR_RI_pat<S2_lsr_i_vh, Srl, v2i32, V2I32, u5_0ImmPred>;
1031 def: OpR_RI_pat<S2_asl_i_vh, Shl, v2i32, V2I32, u5_0ImmPred>;
1033 def: OpR_RR_pat<S2_asr_r_r, Sra, i32, I32, I32>;
1034 def: OpR_RR_pat<S2_lsr_r_r, Srl, i32, I32, I32>;
1035 def: OpR_RR_pat<S2_asl_r_r, Shl, i32, I32, I32>;
1036 def: OpR_RR_pat<S2_asr_r_p, Sra, i64, I64, I32>;
1037 def: OpR_RR_pat<S2_lsr_r_p, Srl, i64, I64, I32>;
1038 def: OpR_RR_pat<S2_asl_r_p, Shl, i64, I64, I32>;
1041 def IsMul8_U3: PatLeaf<(i32 imm), [{
1042 uint64_t V = N->getZExtValue();
1043 return V % 8 == 0 && isUInt<3>(V / 8);
1046 def Divu8: SDNodeXForm<imm, [{
1047 return CurDAG->getTargetConstant(N->getZExtValue() / 8, SDLoc(N), MVT::i32);
1050 // Funnel shift-left.
1051 def FShl32i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1052 (HiReg (S2_asl_i_p (Combinew $Rs, $Rt), $S))>;
1053 def FShl32r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1054 (HiReg (S2_asl_r_p (Combinew $Rs, $Rt), $Ru))>;
1056 def FShl64i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1057 (S2_lsr_i_p_or (S2_asl_i_p $Rt, $S), $Rs, (Subi<64> $S))>;
1058 def FShl64r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1059 (S2_lsr_r_p_or (S2_asl_r_p $Rt, $Ru), $Rs, (A2_subri 64, $Ru))>;
1061 // Combined SDNodeXForm: (Divu8 (Subi<64> $S))
1062 def Divu64_8: SDNodeXForm<imm, [{
1063 return CurDAG->getTargetConstant((64 - N->getSExtValue()) / 8,
1064 SDLoc(N), MVT::i32);
1068 let AddedComplexity = 100 in {
1069 def: Pat<(fshl I32:$Rs, I32:$Rt, (i32 16)),
1070 (A2_combine_hl I32:$Rs, I32:$Rt)>;
1071 def: Pat<(fshl I64:$Rs, I64:$Rt, IsMul8_U3:$S),
1072 (S2_valignib I64:$Rs, I64:$Rt, (Divu64_8 $S))>;
1075 let Predicates = [HasV60], AddedComplexity = 50 in {
1076 def: OpR_RI_pat<S6_rol_i_r, Rol, i32, I32, u5_0ImmPred>;
1077 def: OpR_RI_pat<S6_rol_i_p, Rol, i64, I64, u6_0ImmPred>;
1079 let AddedComplexity = 30 in {
1080 def: Pat<(rotl I32:$Rs, u5_0ImmPred:$S), (FShl32i $Rs, $Rs, imm:$S)>;
1081 def: Pat<(rotl I64:$Rs, u6_0ImmPred:$S), (FShl64i $Rs, $Rs, imm:$S)>;
1082 def: Pat<(fshl I32:$Rs, I32:$Rt, u5_0ImmPred:$S), (FShl32i $Rs, $Rt, imm:$S)>;
1083 def: Pat<(fshl I64:$Rs, I64:$Rt, u6_0ImmPred:$S), (FShl64i $Rs, $Rt, imm:$S)>;
1085 def: Pat<(rotl I32:$Rs, I32:$Rt), (FShl32r $Rs, $Rs, $Rt)>;
1086 def: Pat<(rotl I64:$Rs, I32:$Rt), (FShl64r $Rs, $Rs, $Rt)>;
1087 def: Pat<(fshl I32:$Rs, I32:$Rt, I32:$Ru), (FShl32r $Rs, $Rt, $Ru)>;
1088 def: Pat<(fshl I64:$Rs, I64:$Rt, I32:$Ru), (FShl64r $Rs, $Rt, $Ru)>;
1090 // Funnel shift-right.
1091 def FShr32i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1092 (LoReg (S2_lsr_i_p (Combinew $Rs, $Rt), $S))>;
1093 def FShr32r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1094 (LoReg (S2_lsr_r_p (Combinew $Rs, $Rt), $Ru))>;
1096 def FShr64i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1097 (S2_asl_i_p_or (S2_lsr_i_p $Rt, $S), $Rs, (Subi<64> $S))>;
1098 def FShr64r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1099 (S2_asl_r_p_or (S2_lsr_r_p $Rt, $Ru), $Rs, (A2_subri 64, $Ru))>;
1102 let AddedComplexity = 100 in {
1103 def: Pat<(fshr I32:$Rs, I32:$Rt, (i32 16)),
1104 (A2_combine_hl I32:$Rs, I32:$Rt)>;
1105 def: Pat<(fshr I64:$Rs, I64:$Rt, IsMul8_U3:$S),
1106 (S2_valignib I64:$Rs, I64:$Rt, (Divu8 $S))>;
1109 let Predicates = [HasV60], AddedComplexity = 50 in {
1110 def: Pat<(rotr I32:$Rs, u5_0ImmPred:$S), (S6_rol_i_r I32:$Rs, (Subi<32> $S))>;
1111 def: Pat<(rotr I64:$Rs, u6_0ImmPred:$S), (S6_rol_i_p I64:$Rs, (Subi<64> $S))>;
1113 let AddedComplexity = 30 in {
1114 def: Pat<(rotr I32:$Rs, u5_0ImmPred:$S), (FShr32i $Rs, $Rs, imm:$S)>;
1115 def: Pat<(rotr I64:$Rs, u6_0ImmPred:$S), (FShr64i $Rs, $Rs, imm:$S)>;
1116 def: Pat<(fshr I32:$Rs, I32:$Rt, u5_0ImmPred:$S), (FShr32i $Rs, $Rt, imm:$S)>;
1117 def: Pat<(fshr I64:$Rs, I64:$Rt, u6_0ImmPred:$S), (FShr64i $Rs, $Rt, imm:$S)>;
1119 def: Pat<(rotr I32:$Rs, I32:$Rt), (FShr32r $Rs, $Rs, $Rt)>;
1120 def: Pat<(rotr I64:$Rs, I32:$Rt), (FShr64r $Rs, $Rs, $Rt)>;
1121 def: Pat<(fshr I32:$Rs, I32:$Rt, I32:$Ru), (FShr32r $Rs, $Rt, $Ru)>;
1122 def: Pat<(fshr I64:$Rs, I64:$Rt, I32:$Ru), (FShr64r $Rs, $Rt, $Ru)>;
1125 def: Pat<(sra (add (sra I32:$Rs, u5_0ImmPred:$u5), 1), (i32 1)),
1126 (S2_asr_i_r_rnd I32:$Rs, imm:$u5)>;
1127 def: Pat<(sra (add (sra I64:$Rs, u6_0ImmPred:$u6), 1), (i32 1)),
1128 (S2_asr_i_p_rnd I64:$Rs, imm:$u6)>;
1130 // Prefer S2_addasl_rrri over S2_asl_i_r_acc.
1131 let AddedComplexity = 120 in
1132 def: Pat<(add I32:$Rt, (shl I32:$Rs, u3_0ImmPred:$u3)),
1133 (S2_addasl_rrri IntRegs:$Rt, IntRegs:$Rs, imm:$u3)>;
1135 let AddedComplexity = 100 in {
1136 def: AccRRI_pat<S2_asr_i_r_acc, Add, Su<Sra>, I32, u5_0ImmPred>;
1137 def: AccRRI_pat<S2_asr_i_r_nac, Sub, Su<Sra>, I32, u5_0ImmPred>;
1138 def: AccRRI_pat<S2_asr_i_r_and, And, Su<Sra>, I32, u5_0ImmPred>;
1139 def: AccRRI_pat<S2_asr_i_r_or, Or, Su<Sra>, I32, u5_0ImmPred>;
1141 def: AccRRI_pat<S2_asr_i_p_acc, Add, Su<Sra>, I64, u6_0ImmPred>;
1142 def: AccRRI_pat<S2_asr_i_p_nac, Sub, Su<Sra>, I64, u6_0ImmPred>;
1143 def: AccRRI_pat<S2_asr_i_p_and, And, Su<Sra>, I64, u6_0ImmPred>;
1144 def: AccRRI_pat<S2_asr_i_p_or, Or, Su<Sra>, I64, u6_0ImmPred>;
1146 def: AccRRI_pat<S2_lsr_i_r_acc, Add, Su<Srl>, I32, u5_0ImmPred>;
1147 def: AccRRI_pat<S2_lsr_i_r_nac, Sub, Su<Srl>, I32, u5_0ImmPred>;
1148 def: AccRRI_pat<S2_lsr_i_r_and, And, Su<Srl>, I32, u5_0ImmPred>;
1149 def: AccRRI_pat<S2_lsr_i_r_or, Or, Su<Srl>, I32, u5_0ImmPred>;
1150 def: AccRRI_pat<S2_lsr_i_r_xacc, Xor, Su<Srl>, I32, u5_0ImmPred>;
1152 def: AccRRI_pat<S2_lsr_i_p_acc, Add, Su<Srl>, I64, u6_0ImmPred>;
1153 def: AccRRI_pat<S2_lsr_i_p_nac, Sub, Su<Srl>, I64, u6_0ImmPred>;
1154 def: AccRRI_pat<S2_lsr_i_p_and, And, Su<Srl>, I64, u6_0ImmPred>;
1155 def: AccRRI_pat<S2_lsr_i_p_or, Or, Su<Srl>, I64, u6_0ImmPred>;
1156 def: AccRRI_pat<S2_lsr_i_p_xacc, Xor, Su<Srl>, I64, u6_0ImmPred>;
1158 def: AccRRI_pat<S2_asl_i_r_acc, Add, Su<Shl>, I32, u5_0ImmPred>;
1159 def: AccRRI_pat<S2_asl_i_r_nac, Sub, Su<Shl>, I32, u5_0ImmPred>;
1160 def: AccRRI_pat<S2_asl_i_r_and, And, Su<Shl>, I32, u5_0ImmPred>;
1161 def: AccRRI_pat<S2_asl_i_r_or, Or, Su<Shl>, I32, u5_0ImmPred>;
1162 def: AccRRI_pat<S2_asl_i_r_xacc, Xor, Su<Shl>, I32, u5_0ImmPred>;
1164 def: AccRRI_pat<S2_asl_i_p_acc, Add, Su<Shl>, I64, u6_0ImmPred>;
1165 def: AccRRI_pat<S2_asl_i_p_nac, Sub, Su<Shl>, I64, u6_0ImmPred>;
1166 def: AccRRI_pat<S2_asl_i_p_and, And, Su<Shl>, I64, u6_0ImmPred>;
1167 def: AccRRI_pat<S2_asl_i_p_or, Or, Su<Shl>, I64, u6_0ImmPred>;
1168 def: AccRRI_pat<S2_asl_i_p_xacc, Xor, Su<Shl>, I64, u6_0ImmPred>;
1170 let Predicates = [HasV60] in {
1171 def: AccRRI_pat<S6_rol_i_r_acc, Add, Su<Rol>, I32, u5_0ImmPred>;
1172 def: AccRRI_pat<S6_rol_i_r_nac, Sub, Su<Rol>, I32, u5_0ImmPred>;
1173 def: AccRRI_pat<S6_rol_i_r_and, And, Su<Rol>, I32, u5_0ImmPred>;
1174 def: AccRRI_pat<S6_rol_i_r_or, Or, Su<Rol>, I32, u5_0ImmPred>;
1175 def: AccRRI_pat<S6_rol_i_r_xacc, Xor, Su<Rol>, I32, u5_0ImmPred>;
1177 def: AccRRI_pat<S6_rol_i_p_acc, Add, Su<Rol>, I64, u6_0ImmPred>;
1178 def: AccRRI_pat<S6_rol_i_p_nac, Sub, Su<Rol>, I64, u6_0ImmPred>;
1179 def: AccRRI_pat<S6_rol_i_p_and, And, Su<Rol>, I64, u6_0ImmPred>;
1180 def: AccRRI_pat<S6_rol_i_p_or, Or, Su<Rol>, I64, u6_0ImmPred>;
1181 def: AccRRI_pat<S6_rol_i_p_xacc, Xor, Su<Rol>, I64, u6_0ImmPred>;
1185 let AddedComplexity = 100 in {
1186 def: AccRRR_pat<S2_asr_r_r_acc, Add, Su<Sra>, I32, I32, I32>;
1187 def: AccRRR_pat<S2_asr_r_r_nac, Sub, Su<Sra>, I32, I32, I32>;
1188 def: AccRRR_pat<S2_asr_r_r_and, And, Su<Sra>, I32, I32, I32>;
1189 def: AccRRR_pat<S2_asr_r_r_or, Or, Su<Sra>, I32, I32, I32>;
1191 def: AccRRR_pat<S2_asr_r_p_acc, Add, Su<Sra>, I64, I64, I32>;
1192 def: AccRRR_pat<S2_asr_r_p_nac, Sub, Su<Sra>, I64, I64, I32>;
1193 def: AccRRR_pat<S2_asr_r_p_and, And, Su<Sra>, I64, I64, I32>;
1194 def: AccRRR_pat<S2_asr_r_p_or, Or, Su<Sra>, I64, I64, I32>;
1195 def: AccRRR_pat<S2_asr_r_p_xor, Xor, Su<Sra>, I64, I64, I32>;
1197 def: AccRRR_pat<S2_lsr_r_r_acc, Add, Su<Srl>, I32, I32, I32>;
1198 def: AccRRR_pat<S2_lsr_r_r_nac, Sub, Su<Srl>, I32, I32, I32>;
1199 def: AccRRR_pat<S2_lsr_r_r_and, And, Su<Srl>, I32, I32, I32>;
1200 def: AccRRR_pat<S2_lsr_r_r_or, Or, Su<Srl>, I32, I32, I32>;
1202 def: AccRRR_pat<S2_lsr_r_p_acc, Add, Su<Srl>, I64, I64, I32>;
1203 def: AccRRR_pat<S2_lsr_r_p_nac, Sub, Su<Srl>, I64, I64, I32>;
1204 def: AccRRR_pat<S2_lsr_r_p_and, And, Su<Srl>, I64, I64, I32>;
1205 def: AccRRR_pat<S2_lsr_r_p_or, Or, Su<Srl>, I64, I64, I32>;
1206 def: AccRRR_pat<S2_lsr_r_p_xor, Xor, Su<Srl>, I64, I64, I32>;
1208 def: AccRRR_pat<S2_asl_r_r_acc, Add, Su<Shl>, I32, I32, I32>;
1209 def: AccRRR_pat<S2_asl_r_r_nac, Sub, Su<Shl>, I32, I32, I32>;
1210 def: AccRRR_pat<S2_asl_r_r_and, And, Su<Shl>, I32, I32, I32>;
1211 def: AccRRR_pat<S2_asl_r_r_or, Or, Su<Shl>, I32, I32, I32>;
1213 def: AccRRR_pat<S2_asl_r_p_acc, Add, Su<Shl>, I64, I64, I32>;
1214 def: AccRRR_pat<S2_asl_r_p_nac, Sub, Su<Shl>, I64, I64, I32>;
1215 def: AccRRR_pat<S2_asl_r_p_and, And, Su<Shl>, I64, I64, I32>;
1216 def: AccRRR_pat<S2_asl_r_p_or, Or, Su<Shl>, I64, I64, I32>;
1217 def: AccRRR_pat<S2_asl_r_p_xor, Xor, Su<Shl>, I64, I64, I32>;
1221 class OpshIRI_pat<InstHexagon MI, PatFrag Op, PatFrag ShOp,
1222 PatFrag RegPred, PatFrag ImmPred>
1223 : Pat<(Op anyimm:$u8, (ShOp RegPred:$Rs, ImmPred:$U5)),
1224 (MI anyimm:$u8, RegPred:$Rs, imm:$U5)>;
1226 let AddedComplexity = 200 in {
1227 def: OpshIRI_pat<S4_addi_asl_ri, Add, Su<Shl>, I32, u5_0ImmPred>;
1228 def: OpshIRI_pat<S4_addi_lsr_ri, Add, Su<Srl>, I32, u5_0ImmPred>;
1229 def: OpshIRI_pat<S4_subi_asl_ri, Sub, Su<Shl>, I32, u5_0ImmPred>;
1230 def: OpshIRI_pat<S4_subi_lsr_ri, Sub, Su<Srl>, I32, u5_0ImmPred>;
1231 def: OpshIRI_pat<S4_andi_asl_ri, And, Su<Shl>, I32, u5_0ImmPred>;
1232 def: OpshIRI_pat<S4_andi_lsr_ri, And, Su<Srl>, I32, u5_0ImmPred>;
1233 def: OpshIRI_pat<S4_ori_asl_ri, Or, Su<Shl>, I32, u5_0ImmPred>;
1234 def: OpshIRI_pat<S4_ori_lsr_ri, Or, Su<Srl>, I32, u5_0ImmPred>;
1237 // Prefer this pattern to S2_asl_i_p_or for the special case of joining
1238 // two 32-bit words into a 64-bit word.
1239 let AddedComplexity = 200 in
1240 def: Pat<(or (shl (Aext64 I32:$a), (i32 32)), (Zext64 I32:$b)),
1241 (Combinew I32:$a, I32:$b)>;
1243 def: Pat<(or (or (or (shl (Zext64 (and I32:$b, (i32 65535))), (i32 16)),
1244 (Zext64 (and I32:$a, (i32 65535)))),
1245 (shl (Aext64 (and I32:$c, (i32 65535))), (i32 32))),
1246 (shl (Aext64 I32:$d), (i32 48))),
1247 (Combinew (A2_combine_ll I32:$d, I32:$c),
1248 (A2_combine_ll I32:$b, I32:$a))>;
1250 let AddedComplexity = 200 in {
1251 def: Pat<(or (shl I32:$Rt, (i32 16)), (and I32:$Rs, (i32 65535))),
1252 (A2_combine_ll I32:$Rt, I32:$Rs)>;
1253 def: Pat<(or (shl I32:$Rt, (i32 16)), (srl I32:$Rs, (i32 16))),
1254 (A2_combine_lh I32:$Rt, I32:$Rs)>;
1255 def: Pat<(or (and I32:$Rt, (i32 268431360)), (and I32:$Rs, (i32 65535))),
1256 (A2_combine_hl I32:$Rt, I32:$Rs)>;
1257 def: Pat<(or (and I32:$Rt, (i32 268431360)), (srl I32:$Rs, (i32 16))),
1258 (A2_combine_hh I32:$Rt, I32:$Rs)>;
1261 def SDTHexagonVShift
1262 : SDTypeProfile<1, 2, [SDTCisSameAs<0, 1>, SDTCisVec<0>, SDTCisVT<2, i32>]>;
1264 def HexagonVASL: SDNode<"HexagonISD::VASL", SDTHexagonVShift>;
1265 def HexagonVASR: SDNode<"HexagonISD::VASR", SDTHexagonVShift>;
1266 def HexagonVLSR: SDNode<"HexagonISD::VLSR", SDTHexagonVShift>;
1268 def: OpR_RI_pat<S2_asl_i_vw, pf2<HexagonVASL>, v2i32, V2I32, u5_0ImmPred>;
1269 def: OpR_RI_pat<S2_asl_i_vh, pf2<HexagonVASL>, v4i16, V4I16, u4_0ImmPred>;
1270 def: OpR_RI_pat<S2_asr_i_vw, pf2<HexagonVASR>, v2i32, V2I32, u5_0ImmPred>;
1271 def: OpR_RI_pat<S2_asr_i_vh, pf2<HexagonVASR>, v4i16, V4I16, u4_0ImmPred>;
1272 def: OpR_RI_pat<S2_lsr_i_vw, pf2<HexagonVLSR>, v2i32, V2I32, u5_0ImmPred>;
1273 def: OpR_RI_pat<S2_lsr_i_vh, pf2<HexagonVLSR>, v4i16, V4I16, u4_0ImmPred>;
1275 def: OpR_RR_pat<S2_asl_r_vw, pf2<HexagonVASL>, v2i32, V2I32, I32>;
1276 def: OpR_RR_pat<S2_asl_r_vh, pf2<HexagonVASL>, v4i16, V4I16, I32>;
1277 def: OpR_RR_pat<S2_asr_r_vw, pf2<HexagonVASR>, v2i32, V2I32, I32>;
1278 def: OpR_RR_pat<S2_asr_r_vh, pf2<HexagonVASR>, v4i16, V4I16, I32>;
1279 def: OpR_RR_pat<S2_lsr_r_vw, pf2<HexagonVLSR>, v2i32, V2I32, I32>;
1280 def: OpR_RR_pat<S2_lsr_r_vh, pf2<HexagonVLSR>, v4i16, V4I16, I32>;
1282 def: Pat<(sra V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c))),
1283 (S2_asr_i_vw V2I32:$b, imm:$c)>;
1284 def: Pat<(srl V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c))),
1285 (S2_lsr_i_vw V2I32:$b, imm:$c)>;
1286 def: Pat<(shl V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c))),
1287 (S2_asl_i_vw V2I32:$b, imm:$c)>;
1288 def: Pat<(sra V4I16:$b, (v4i16 (HexagonVSPLAT u4_0ImmPred:$c))),
1289 (S2_asr_i_vh V4I16:$b, imm:$c)>;
1290 def: Pat<(srl V4I16:$b, (v4i16 (HexagonVSPLAT u4_0ImmPred:$c))),
1291 (S2_lsr_i_vh V4I16:$b, imm:$c)>;
1292 def: Pat<(shl V4I16:$b, (v4i16 (HexagonVSPLAT u4_0ImmPred:$c))),
1293 (S2_asl_i_vh V4I16:$b, imm:$c)>;
1295 def: Pat<(HexagonVASR V2I16:$Rs, u4_0ImmPred:$S),
1296 (LoReg (S2_asr_i_vh (ToAext64 $Rs), imm:$S))>;
1297 def: Pat<(HexagonVASL V2I16:$Rs, u4_0ImmPred:$S),
1298 (LoReg (S2_asl_i_vh (ToAext64 $Rs), imm:$S))>;
1299 def: Pat<(HexagonVLSR V2I16:$Rs, u4_0ImmPred:$S),
1300 (LoReg (S2_lsr_i_vh (ToAext64 $Rs), imm:$S))>;
1301 def: Pat<(HexagonVASR V2I16:$Rs, I32:$Rt),
1302 (LoReg (S2_asr_i_vh (ToAext64 $Rs), I32:$Rt))>;
1303 def: Pat<(HexagonVASL V2I16:$Rs, I32:$Rt),
1304 (LoReg (S2_asl_i_vh (ToAext64 $Rs), I32:$Rt))>;
1305 def: Pat<(HexagonVLSR V2I16:$Rs, I32:$Rt),
1306 (LoReg (S2_lsr_i_vh (ToAext64 $Rs), I32:$Rt))>;
1309 // --(9) Arithmetic/bitwise ----------------------------------------------
1312 def: Pat<(abs I32:$Rs), (A2_abs I32:$Rs)>;
1313 def: Pat<(abs I64:$Rs), (A2_absp I64:$Rs)>;
1314 def: Pat<(not I32:$Rs), (A2_subri -1, I32:$Rs)>;
1315 def: Pat<(not I64:$Rs), (A2_notp I64:$Rs)>;
1316 def: Pat<(ineg I64:$Rs), (A2_negp I64:$Rs)>;
1318 def: Pat<(fabs F32:$Rs), (S2_clrbit_i F32:$Rs, 31)>;
1319 def: Pat<(fneg F32:$Rs), (S2_togglebit_i F32:$Rs, 31)>;
1321 def: Pat<(fabs F64:$Rs),
1322 (Combinew (S2_clrbit_i (HiReg $Rs), 31),
1323 (i32 (LoReg $Rs)))>;
1324 def: Pat<(fneg F64:$Rs),
1325 (Combinew (S2_togglebit_i (HiReg $Rs), 31),
1326 (i32 (LoReg $Rs)))>;
1328 def: Pat<(add I32:$Rs, anyimm:$s16), (A2_addi I32:$Rs, imm:$s16)>;
1329 def: Pat<(or I32:$Rs, anyimm:$s10), (A2_orir I32:$Rs, imm:$s10)>;
1330 def: Pat<(and I32:$Rs, anyimm:$s10), (A2_andir I32:$Rs, imm:$s10)>;
1331 def: Pat<(sub anyimm:$s10, I32:$Rs), (A2_subri imm:$s10, I32:$Rs)>;
1333 def: OpR_RR_pat<A2_add, Add, i32, I32>;
1334 def: OpR_RR_pat<A2_sub, Sub, i32, I32>;
1335 def: OpR_RR_pat<A2_and, And, i32, I32>;
1336 def: OpR_RR_pat<A2_or, Or, i32, I32>;
1337 def: OpR_RR_pat<A2_xor, Xor, i32, I32>;
1338 def: OpR_RR_pat<A2_addp, Add, i64, I64>;
1339 def: OpR_RR_pat<A2_subp, Sub, i64, I64>;
1340 def: OpR_RR_pat<A2_andp, And, i64, I64>;
1341 def: OpR_RR_pat<A2_orp, Or, i64, I64>;
1342 def: OpR_RR_pat<A2_xorp, Xor, i64, I64>;
1343 def: OpR_RR_pat<A4_andnp, Not2<And>, i64, I64>;
1344 def: OpR_RR_pat<A4_ornp, Not2<Or>, i64, I64>;
1346 def: OpR_RR_pat<A2_svaddh, Add, v2i16, V2I16>;
1347 def: OpR_RR_pat<A2_svsubh, Sub, v2i16, V2I16>;
1349 def: OpR_RR_pat<A2_vaddub, Add, v8i8, V8I8>;
1350 def: OpR_RR_pat<A2_vaddh, Add, v4i16, V4I16>;
1351 def: OpR_RR_pat<A2_vaddw, Add, v2i32, V2I32>;
1352 def: OpR_RR_pat<A2_vsubub, Sub, v8i8, V8I8>;
1353 def: OpR_RR_pat<A2_vsubh, Sub, v4i16, V4I16>;
1354 def: OpR_RR_pat<A2_vsubw, Sub, v2i32, V2I32>;
1356 def: OpR_RR_pat<A2_and, And, v4i8, V4I8>;
1357 def: OpR_RR_pat<A2_xor, Xor, v4i8, V4I8>;
1358 def: OpR_RR_pat<A2_or, Or, v4i8, V4I8>;
1359 def: OpR_RR_pat<A2_and, And, v2i16, V2I16>;
1360 def: OpR_RR_pat<A2_xor, Xor, v2i16, V2I16>;
1361 def: OpR_RR_pat<A2_or, Or, v2i16, V2I16>;
1362 def: OpR_RR_pat<A2_andp, And, v8i8, V8I8>;
1363 def: OpR_RR_pat<A2_orp, Or, v8i8, V8I8>;
1364 def: OpR_RR_pat<A2_xorp, Xor, v8i8, V8I8>;
1365 def: OpR_RR_pat<A2_andp, And, v4i16, V4I16>;
1366 def: OpR_RR_pat<A2_orp, Or, v4i16, V4I16>;
1367 def: OpR_RR_pat<A2_xorp, Xor, v4i16, V4I16>;
1368 def: OpR_RR_pat<A2_andp, And, v2i32, V2I32>;
1369 def: OpR_RR_pat<A2_orp, Or, v2i32, V2I32>;
1370 def: OpR_RR_pat<A2_xorp, Xor, v2i32, V2I32>;
1372 def: OpR_RR_pat<M2_mpyi, Mul, i32, I32>;
1373 def: OpR_RR_pat<M2_mpy_up, pf2<mulhs>, i32, I32>;
1374 def: OpR_RR_pat<M2_mpyu_up, pf2<mulhu>, i32, I32>;
1375 def: OpR_RI_pat<M2_mpysip, Mul, i32, I32, u32_0ImmPred>;
1376 def: OpR_RI_pat<M2_mpysmi, Mul, i32, I32, s32_0ImmPred>;
1378 // Arithmetic on predicates.
1379 def: OpR_RR_pat<C2_xor, Add, i1, I1>;
1380 def: OpR_RR_pat<C2_xor, Add, v2i1, V2I1>;
1381 def: OpR_RR_pat<C2_xor, Add, v4i1, V4I1>;
1382 def: OpR_RR_pat<C2_xor, Add, v8i1, V8I1>;
1383 def: OpR_RR_pat<C2_xor, Sub, i1, I1>;
1384 def: OpR_RR_pat<C2_xor, Sub, v2i1, V2I1>;
1385 def: OpR_RR_pat<C2_xor, Sub, v4i1, V4I1>;
1386 def: OpR_RR_pat<C2_xor, Sub, v8i1, V8I1>;
1387 def: OpR_RR_pat<C2_and, Mul, i1, I1>;
1388 def: OpR_RR_pat<C2_and, Mul, v2i1, V2I1>;
1389 def: OpR_RR_pat<C2_and, Mul, v4i1, V4I1>;
1390 def: OpR_RR_pat<C2_and, Mul, v8i1, V8I1>;
1392 def: OpR_RR_pat<F2_sfadd, pf2<fadd>, f32, F32>;
1393 def: OpR_RR_pat<F2_sfsub, pf2<fsub>, f32, F32>;
1394 def: OpR_RR_pat<F2_sfmpy, pf2<fmul>, f32, F32>;
1395 def: OpR_RR_pat<F2_sfmin, pf2<fminnum>, f32, F32>;
1396 def: OpR_RR_pat<F2_sfmax, pf2<fmaxnum>, f32, F32>;
1398 let Predicates = [HasV66] in {
1399 def: OpR_RR_pat<F2_dfadd, pf2<fadd>, f64, F64>;
1400 def: OpR_RR_pat<F2_dfsub, pf2<fsub>, f64, F64>;
1403 // In expressions like a0*b0 + a1*b1 + ..., prefer to generate multiply-add,
1404 // over add-add with individual multiplies as inputs.
1405 let AddedComplexity = 10 in {
1406 def: AccRRI_pat<M2_macsip, Add, Su<Mul>, I32, u32_0ImmPred>;
1407 def: AccRRI_pat<M2_macsin, Sub, Su<Mul>, I32, u32_0ImmPred>;
1408 def: AccRRR_pat<M2_maci, Add, Su<Mul>, I32, I32, I32>;
1409 let Predicates = [HasV66] in
1410 def: AccRRR_pat<M2_mnaci, Sub, Su<Mul>, I32, I32, I32>;
1413 def: AccRRI_pat<M2_naccii, Sub, Su<Add>, I32, s32_0ImmPred>;
1414 def: AccRRI_pat<M2_accii, Add, Su<Add>, I32, s32_0ImmPred>;
1415 def: AccRRR_pat<M2_acci, Add, Su<Add>, I32, I32, I32>;
1419 def: Pat<(v2i32 (mulhu V2I32:$Rss, V2I32:$Rtt)),
1420 (Combinew (M2_mpyu_up (HiReg $Rss), (HiReg $Rtt)),
1421 (M2_mpyu_up (LoReg $Rss), (LoReg $Rtt)))>;
1423 def: Pat<(v2i32 (mulhs V2I32:$Rs, V2I32:$Rt)),
1424 (Combinew (M2_mpy_up (HiReg $Rs), (HiReg $Rt)),
1425 (M2_mpy_up (LoReg $Rt), (LoReg $Rt)))>;
1428 OutPatFrag<(ops node:$Rss, node:$Rtt),
1429 (Combinew (S2_vtrunohb (M5_vmpybuu (HiReg $Rss), (HiReg $Rtt))),
1430 (S2_vtrunohb (M5_vmpybuu (LoReg $Rss), (LoReg $Rtt))))>;
1432 // Equivalent of byte-wise arithmetic shift right by 7 in v8i8.
1434 OutPatFrag<(ops node:$Rss), (C2_mask (C2_not (A4_vcmpbgti $Rss, 0)))>;
1436 def: Pat<(v8i8 (mulhu V8I8:$Rss, V8I8:$Rtt)),
1437 (Mulhub $Rss, $Rtt)>;
1439 def: Pat<(v8i8 (mulhs V8I8:$Rss, V8I8:$Rtt)),
1441 (Mulhub $Rss, $Rtt),
1442 (A2_vaddub (A2_andp V8I8:$Rss, (Asr7 $Rtt)),
1443 (A2_andp V8I8:$Rtt, (Asr7 $Rss))))>;
1446 OutPatFrag<(ops node:$Rs, node:$Rt), (M2_vmpy2s_s0 $Rs, $Rt)>;
1448 OutPatFrag<(ops node:$Rss, node:$Rtt), (Mpysh (HiReg $Rss), (HiReg $Rtt))>;
1450 OutPatFrag<(ops node:$Rss, node:$Rtt), (Mpysh (LoReg $Rss), (LoReg $Rtt))>;
1453 OutPatFrag<(ops node:$Rss, node:$Rtt),
1454 (Combinew (A2_combine_hh (HiReg (Mpyshh $Rss, $Rtt)),
1455 (LoReg (Mpyshh $Rss, $Rtt))),
1456 (A2_combine_hh (HiReg (Mpyshl $Rss, $Rtt)),
1457 (LoReg (Mpyshl $Rss, $Rtt))))>;
1459 def: Pat<(v4i16 (mulhs V4I16:$Rss, V4I16:$Rtt)), (Mulhsh $Rss, $Rtt)>;
1461 def: Pat<(v4i16 (mulhu V4I16:$Rss, V4I16:$Rtt)),
1463 (Mulhsh $Rss, $Rtt),
1464 (A2_vaddh (A2_andp V4I16:$Rss, (S2_asr_i_vh $Rtt, 15)),
1465 (A2_andp V4I16:$Rtt, (S2_asr_i_vh $Rss, 15))))>;
1468 def: Pat<(ineg (mul I32:$Rs, u8_0ImmPred:$u8)),
1469 (M2_mpysin IntRegs:$Rs, imm:$u8)>;
1471 def n8_0ImmPred: PatLeaf<(i32 imm), [{
1472 int64_t V = N->getSExtValue();
1473 return -255 <= V && V <= 0;
1476 // Change the sign of the immediate for Rd=-mpyi(Rs,#u8)
1477 def: Pat<(mul I32:$Rs, n8_0ImmPred:$n8),
1478 (M2_mpysin I32:$Rs, (NegImm8 imm:$n8))>;
1480 def: Pat<(add Sext64:$Rs, I64:$Rt),
1481 (A2_addsp (LoReg Sext64:$Rs), I64:$Rt)>;
1483 def: AccRRR_pat<M4_and_and, And, Su_ni1<And>, I32, I32, I32>;
1484 def: AccRRR_pat<M4_and_or, And, Su_ni1<Or>, I32, I32, I32>;
1485 def: AccRRR_pat<M4_and_xor, And, Su<Xor>, I32, I32, I32>;
1486 def: AccRRR_pat<M4_or_and, Or, Su_ni1<And>, I32, I32, I32>;
1487 def: AccRRR_pat<M4_or_or, Or, Su_ni1<Or>, I32, I32, I32>;
1488 def: AccRRR_pat<M4_or_xor, Or, Su<Xor>, I32, I32, I32>;
1489 def: AccRRR_pat<M4_xor_and, Xor, Su_ni1<And>, I32, I32, I32>;
1490 def: AccRRR_pat<M4_xor_or, Xor, Su_ni1<Or>, I32, I32, I32>;
1491 def: AccRRR_pat<M2_xor_xacc, Xor, Su<Xor>, I32, I32, I32>;
1492 def: AccRRR_pat<M4_xor_xacc, Xor, Su<Xor>, I64, I64, I64>;
1494 // For dags like (or (and (not _), _), (shl _, _)) where the "or" with
1495 // one argument matches the patterns below, and with the other argument
1496 // matches S2_asl_r_r_or, etc, prefer the patterns below.
1497 let AddedComplexity = 110 in { // greater than S2_asl_r_r_and/or/xor.
1498 def: AccRRR_pat<M4_and_andn, And, Su<Not2<And>>, I32, I32, I32>;
1499 def: AccRRR_pat<M4_or_andn, Or, Su<Not2<And>>, I32, I32, I32>;
1500 def: AccRRR_pat<M4_xor_andn, Xor, Su<Not2<And>>, I32, I32, I32>;
1503 // S4_addaddi and S4_subaddi don't have tied operands, so give them
1504 // a bit of preference.
1505 let AddedComplexity = 30 in {
1506 def: Pat<(add I32:$Rs, (Su<Add> I32:$Ru, anyimm:$s6)),
1507 (S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
1508 def: Pat<(add anyimm:$s6, (Su<Add> I32:$Rs, I32:$Ru)),
1509 (S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
1510 def: Pat<(add I32:$Rs, (Su<Sub> anyimm:$s6, I32:$Ru)),
1511 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1512 def: Pat<(sub (Su<Add> I32:$Rs, anyimm:$s6), I32:$Ru),
1513 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1514 def: Pat<(add (Su<Sub> I32:$Rs, I32:$Ru), anyimm:$s6),
1515 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1518 def: Pat<(or I32:$Ru, (Su<And> I32:$Rx, anyimm:$s10)),
1519 (S4_or_andix IntRegs:$Ru, IntRegs:$Rx, imm:$s10)>;
1520 def: Pat<(or I32:$Rx, (Su<And> I32:$Rs, anyimm:$s10)),
1521 (S4_or_andi IntRegs:$Rx, IntRegs:$Rs, imm:$s10)>;
1522 def: Pat<(or I32:$Rx, (Su<Or> I32:$Rs, anyimm:$s10)),
1523 (S4_or_ori IntRegs:$Rx, IntRegs:$Rs, imm:$s10)>;
1526 def: Pat<(i32 (trunc (sra (Su<Mul> Sext64:$Rs, Sext64:$Rt), (i32 32)))),
1527 (M2_mpy_up (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1528 def: Pat<(i32 (trunc (srl (Su<Mul> Sext64:$Rs, Sext64:$Rt), (i32 32)))),
1529 (M2_mpy_up (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1531 def: Pat<(mul (Zext64 I32:$Rs), (Zext64 I32:$Rt)),
1532 (M2_dpmpyuu_s0 I32:$Rs, I32:$Rt)>;
1533 def: Pat<(mul (Aext64 I32:$Rs), (Aext64 I32:$Rt)),
1534 (M2_dpmpyuu_s0 I32:$Rs, I32:$Rt)>;
1535 def: Pat<(mul Sext64:$Rs, Sext64:$Rt),
1536 (M2_dpmpyss_s0 (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1538 def: Pat<(add I64:$Rx, (Su<Mul> Sext64:$Rs, Sext64:$Rt)),
1539 (M2_dpmpyss_acc_s0 I64:$Rx, (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1540 def: Pat<(sub I64:$Rx, (Su<Mul> Sext64:$Rs, Sext64:$Rt)),
1541 (M2_dpmpyss_nac_s0 I64:$Rx, (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1542 def: Pat<(add I64:$Rx, (Su<Mul> (Aext64 I32:$Rs), (Aext64 I32:$Rt))),
1543 (M2_dpmpyuu_acc_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1544 def: Pat<(add I64:$Rx, (Su<Mul> (Zext64 I32:$Rs), (Zext64 I32:$Rt))),
1545 (M2_dpmpyuu_acc_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1546 def: Pat<(sub I64:$Rx, (Su<Mul> (Aext64 I32:$Rs), (Aext64 I32:$Rt))),
1547 (M2_dpmpyuu_nac_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1548 def: Pat<(sub I64:$Rx, (Su<Mul> (Zext64 I32:$Rs), (Zext64 I32:$Rt))),
1549 (M2_dpmpyuu_nac_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1552 def: Pat<(sext_inreg (add I32:$Rt, I32:$Rs), i16),
1553 (A2_addh_l16_ll I32:$Rt, I32:$Rs)>;
1554 def: Pat<(sra (add (shl I32:$Rt, (i32 16)), I32:$Rs), (i32 16)),
1555 (A2_addh_l16_hl I32:$Rt, I32:$Rs)>;
1556 def: Pat<(shl (add I32:$Rt, I32:$Rs), (i32 16)),
1557 (A2_addh_h16_ll I32:$Rt, I32:$Rs)>;
1559 // Subtract halfword.
1560 def: Pat<(sext_inreg (sub I32:$Rt, I32:$Rs), i16),
1561 (A2_subh_l16_ll I32:$Rt, I32:$Rs)>;
1562 def: Pat<(sra (add (shl I32:$Rt, (i32 16)), I32:$Rs), (i32 16)),
1563 (A2_addh_l16_hl I32:$Rt, I32:$Rs)>;
1564 def: Pat<(shl (sub I32:$Rt, I32:$Rs), (i32 16)),
1565 (A2_subh_h16_ll I32:$Rt, I32:$Rs)>;
1567 def: Pat<(mul I64:$Rss, I64:$Rtt),
1569 (M2_maci (M2_maci (HiReg (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt))),
1574 (i32 (LoReg (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt)))))>;
1576 def MulHU : OutPatFrag<(ops node:$Rss, node:$Rtt),
1582 (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt)), 32),
1585 (A4_combineir 0, (LoReg (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt))))),
1589 (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt)), 32))>;
1591 // Multiply 64-bit unsigned and use upper result.
1592 def : Pat <(mulhu I64:$Rss, I64:$Rtt), (MulHU $Rss, $Rtt)>;
1594 // Multiply 64-bit signed and use upper result.
1596 // For two signed 64-bit integers A and B, let A' and B' denote A and B
1597 // with the sign bit cleared. Then A = -2^63*s(A) + A', where s(A) is the
1598 // sign bit of A (and identically for B). With this notation, the signed
1599 // product A*B can be written as:
1600 // AB = (-2^63 s(A) + A') * (-2^63 s(B) + B')
1601 // = 2^126 s(A)s(B) - 2^63 [s(A)B'+s(B)A'] + A'B'
1602 // = 2^126 s(A)s(B) + 2^63 [s(A)B'+s(B)A'] + A'B' - 2*2^63 [s(A)B'+s(B)A']
1603 // = (unsigned product AB) - 2^64 [s(A)B'+s(B)A']
1605 // Clear the sign bit in a 64-bit register.
1606 def ClearSign : OutPatFrag<(ops node:$Rss),
1607 (Combinew (S2_clrbit_i (HiReg $Rss), 31), (i32 (LoReg $Rss)))>;
1609 def : Pat <(mulhs I64:$Rss, I64:$Rtt),
1613 (A2_andp (S2_asr_i_p $Rss, 63), (ClearSign $Rtt)),
1614 (A2_andp (S2_asr_i_p $Rtt, 63), (ClearSign $Rss))))>;
1616 // Prefer these instructions over M2_macsip/M2_macsin: the macsi* instructions
1617 // will put the immediate addend into a register, while these instructions will
1618 // use it directly. Such a construct does not appear in the middle of a gep,
1619 // where M2_macsip would be preferable.
1620 let AddedComplexity = 20 in {
1621 def: Pat<(add (Su<Mul> I32:$Rs, u6_0ImmPred:$U6), anyimm:$u6),
1622 (M4_mpyri_addi imm:$u6, IntRegs:$Rs, imm:$U6)>;
1623 def: Pat<(add (Su<Mul> I32:$Rs, I32:$Rt), anyimm:$u6),
1624 (M4_mpyrr_addi imm:$u6, IntRegs:$Rs, IntRegs:$Rt)>;
1627 // Keep these instructions less preferable to M2_macsip/M2_macsin.
1628 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, u6_2ImmPred:$u6_2)),
1629 (M4_mpyri_addr_u2 IntRegs:$Ru, imm:$u6_2, IntRegs:$Rs)>;
1630 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, anyimm:$u6)),
1631 (M4_mpyri_addr IntRegs:$Ru, IntRegs:$Rs, imm:$u6)>;
1632 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Ry, I32:$Rs)),
1633 (M4_mpyrr_addr IntRegs:$Ru, IntRegs:$Ry, IntRegs:$Rs)>;
1636 def: Pat<(fma F32:$Rs, F32:$Rt, F32:$Rx),
1637 (F2_sffma F32:$Rx, F32:$Rs, F32:$Rt)>;
1638 def: Pat<(fma (fneg F32:$Rs), F32:$Rt, F32:$Rx),
1639 (F2_sffms F32:$Rx, F32:$Rs, F32:$Rt)>;
1640 def: Pat<(fma F32:$Rs, (fneg F32:$Rt), F32:$Rx),
1641 (F2_sffms F32:$Rx, F32:$Rs, F32:$Rt)>;
1644 def: Pat<(mul V2I32:$Rs, V2I32:$Rt),
1645 (PS_vmulw V2I32:$Rs, V2I32:$Rt)>;
1646 def: Pat<(add V2I32:$Rx, (mul V2I32:$Rs, V2I32:$Rt)),
1647 (PS_vmulw_acc V2I32:$Rx, V2I32:$Rs, V2I32:$Rt)>;
1649 // Add/subtract two v4i8: Hexagon does not have an insn for this one, so
1650 // we use the double add v8i8, and use only the low part of the result.
1651 def: Pat<(add V4I8:$Rs, V4I8:$Rt),
1652 (LoReg (A2_vaddub (ToAext64 $Rs), (ToAext64 $Rt)))>;
1653 def: Pat<(sub V4I8:$Rs, V4I8:$Rt),
1654 (LoReg (A2_vsubub (ToAext64 $Rs), (ToAext64 $Rt)))>;
1656 // Use M2_vmpy2s_s0 for half-word vector multiply. It multiplies two
1657 // half-words, and saturates the result to a 32-bit value, except the
1658 // saturation never happens (it can only occur with scaling).
1659 def: Pat<(v2i16 (mul V2I16:$Rs, V2I16:$Rt)),
1660 (LoReg (S2_vtrunewh (A2_combineii 0, 0),
1661 (M2_vmpy2s_s0 V2I16:$Rs, V2I16:$Rt)))>;
1662 def: Pat<(v4i16 (mul V4I16:$Rs, V4I16:$Rt)),
1663 (S2_vtrunewh (M2_vmpy2s_s0 (HiReg $Rs), (HiReg $Rt)),
1664 (M2_vmpy2s_s0 (LoReg $Rs), (LoReg $Rt)))>;
1666 // Multiplies two v4i8 vectors.
1667 def: Pat<(v4i8 (mul V4I8:$Rs, V4I8:$Rt)),
1668 (S2_vtrunehb (M5_vmpybuu V4I8:$Rs, V4I8:$Rt))>;
1670 // Multiplies two v8i8 vectors.
1671 def: Pat<(v8i8 (mul V8I8:$Rs, V8I8:$Rt)),
1672 (Combinew (S2_vtrunehb (M5_vmpybuu (HiReg $Rs), (HiReg $Rt))),
1673 (S2_vtrunehb (M5_vmpybuu (LoReg $Rs), (LoReg $Rt))))>;
1676 // --(10) Bit ------------------------------------------------------------
1679 // Count leading zeros.
1680 def: Pat<(i32 (ctlz I32:$Rs)), (S2_cl0 I32:$Rs)>;
1681 def: Pat<(i32 (trunc (ctlz I64:$Rss))), (S2_cl0p I64:$Rss)>;
1683 // Count trailing zeros.
1684 def: Pat<(i32 (cttz I32:$Rs)), (S2_ct0 I32:$Rs)>;
1685 def: Pat<(i32 (trunc (cttz I64:$Rss))), (S2_ct0p I64:$Rss)>;
1687 // Count leading ones.
1688 def: Pat<(i32 (ctlz (not I32:$Rs))), (S2_cl1 I32:$Rs)>;
1689 def: Pat<(i32 (trunc (ctlz (not I64:$Rss)))), (S2_cl1p I64:$Rss)>;
1691 // Count trailing ones.
1692 def: Pat<(i32 (cttz (not I32:$Rs))), (S2_ct1 I32:$Rs)>;
1693 def: Pat<(i32 (trunc (cttz (not I64:$Rss)))), (S2_ct1p I64:$Rss)>;
1695 // Define leading/trailing patterns that require zero-extensions to 64 bits.
1696 def: Pat<(i64 (ctlz I64:$Rss)), (ToZext64 (S2_cl0p I64:$Rss))>;
1697 def: Pat<(i64 (cttz I64:$Rss)), (ToZext64 (S2_ct0p I64:$Rss))>;
1698 def: Pat<(i64 (ctlz (not I64:$Rss))), (ToZext64 (S2_cl1p I64:$Rss))>;
1699 def: Pat<(i64 (cttz (not I64:$Rss))), (ToZext64 (S2_ct1p I64:$Rss))>;
1701 def: Pat<(i64 (ctpop I64:$Rss)), (ToZext64 (S5_popcountp I64:$Rss))>;
1702 def: Pat<(i32 (ctpop I32:$Rs)), (S5_popcountp (A4_combineir 0, I32:$Rs))>;
1704 def: Pat<(bitreverse I32:$Rs), (S2_brev I32:$Rs)>;
1705 def: Pat<(bitreverse I64:$Rss), (S2_brevp I64:$Rss)>;
1707 let AddedComplexity = 20 in { // Complexity greater than and/or/xor
1708 def: Pat<(and I32:$Rs, IsNPow2_32:$V),
1709 (S2_clrbit_i IntRegs:$Rs, (LogN2_32 $V))>;
1710 def: Pat<(or I32:$Rs, IsPow2_32:$V),
1711 (S2_setbit_i IntRegs:$Rs, (Log2_32 $V))>;
1712 def: Pat<(xor I32:$Rs, IsPow2_32:$V),
1713 (S2_togglebit_i IntRegs:$Rs, (Log2_32 $V))>;
1715 def: Pat<(and I32:$Rs, (not (shl 1, I32:$Rt))),
1716 (S2_clrbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1717 def: Pat<(or I32:$Rs, (shl 1, I32:$Rt)),
1718 (S2_setbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1719 def: Pat<(xor I32:$Rs, (shl 1, I32:$Rt)),
1720 (S2_togglebit_r IntRegs:$Rs, IntRegs:$Rt)>;
1723 // Clr/set/toggle bit for 64-bit values with immediate bit index.
1724 let AddedComplexity = 20 in { // Complexity greater than and/or/xor
1725 def: Pat<(and I64:$Rss, IsNPow2_64L:$V),
1726 (Combinew (i32 (HiReg $Rss)),
1727 (S2_clrbit_i (LoReg $Rss), (LogN2_64 $V)))>;
1728 def: Pat<(and I64:$Rss, IsNPow2_64H:$V),
1729 (Combinew (S2_clrbit_i (HiReg $Rss), (UDEC32 (i32 (LogN2_64 $V)))),
1730 (i32 (LoReg $Rss)))>;
1732 def: Pat<(or I64:$Rss, IsPow2_64L:$V),
1733 (Combinew (i32 (HiReg $Rss)),
1734 (S2_setbit_i (LoReg $Rss), (Log2_64 $V)))>;
1735 def: Pat<(or I64:$Rss, IsPow2_64H:$V),
1736 (Combinew (S2_setbit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
1737 (i32 (LoReg $Rss)))>;
1739 def: Pat<(xor I64:$Rss, IsPow2_64L:$V),
1740 (Combinew (i32 (HiReg $Rss)),
1741 (S2_togglebit_i (LoReg $Rss), (Log2_64 $V)))>;
1742 def: Pat<(xor I64:$Rss, IsPow2_64H:$V),
1743 (Combinew (S2_togglebit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
1744 (i32 (LoReg $Rss)))>;
1748 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
1749 def: Pat<(i1 (setne (and (shl 1, u5_0ImmPred:$u5), I32:$Rs), 0)),
1750 (S2_tstbit_i IntRegs:$Rs, imm:$u5)>;
1751 def: Pat<(i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1752 (S2_tstbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1753 def: Pat<(i1 (trunc I32:$Rs)),
1754 (S2_tstbit_i IntRegs:$Rs, 0)>;
1755 def: Pat<(i1 (trunc I64:$Rs)),
1756 (S2_tstbit_i (LoReg DoubleRegs:$Rs), 0)>;
1759 def: Pat<(and (srl I32:$Rs, u5_0ImmPred:$u5), 1),
1760 (I1toI32 (S2_tstbit_i I32:$Rs, imm:$u5))>;
1761 def: Pat<(and (srl I64:$Rss, IsULE<32,31>:$u6), 1),
1762 (ToZext64 (I1toI32 (S2_tstbit_i (LoReg $Rss), imm:$u6)))>;
1763 def: Pat<(and (srl I64:$Rss, IsUGT<32,31>:$u6), 1),
1764 (ToZext64 (I1toI32 (S2_tstbit_i (HiReg $Rss), (UDEC32 $u6))))>;
1766 def: Pat<(and (not (srl I32:$Rs, u5_0ImmPred:$u5)), 1),
1767 (I1toI32 (S4_ntstbit_i I32:$Rs, imm:$u5))>;
1768 def: Pat<(and (not (srl I64:$Rss, IsULE<32,31>:$u6)), 1),
1769 (ToZext64 (I1toI32 (S4_ntstbit_i (LoReg $Rss), imm:$u6)))>;
1770 def: Pat<(and (not (srl I64:$Rss, IsUGT<32,31>:$u6)), 1),
1771 (ToZext64 (I1toI32 (S4_ntstbit_i (HiReg $Rss), (UDEC32 $u6))))>;
1773 let AddedComplexity = 20 in { // Complexity greater than compare reg-imm.
1774 def: Pat<(i1 (seteq (and I32:$Rs, u6_0ImmPred:$u6), 0)),
1775 (C2_bitsclri IntRegs:$Rs, imm:$u6)>;
1776 def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), 0)),
1777 (C2_bitsclr IntRegs:$Rs, IntRegs:$Rt)>;
1780 let AddedComplexity = 10 in // Complexity greater than compare reg-reg.
1781 def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), IntRegs:$Rt)),
1782 (C2_bitsset IntRegs:$Rs, IntRegs:$Rt)>;
1785 SDTypeProfile<1, 2, [SDTCisVT<0, i1>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
1786 def HexagonTSTBIT: SDNode<"HexagonISD::TSTBIT", SDTTestBit>;
1788 def: Pat<(HexagonTSTBIT I32:$Rs, u5_0ImmPred:$u5),
1789 (S2_tstbit_i I32:$Rs, imm:$u5)>;
1790 def: Pat<(HexagonTSTBIT I32:$Rs, I32:$Rt),
1791 (S2_tstbit_r I32:$Rs, I32:$Rt)>;
1793 // Add extra complexity to prefer these instructions over bitsset/bitsclr.
1794 // The reason is that tstbit/ntstbit can be folded into a compound instruction:
1795 // if ([!]tstbit(...)) jump ...
1796 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
1797 def: Pat<(i1 (seteq (and I32:$Rs, IsPow2_32:$u5), 0)),
1798 (S4_ntstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
1799 def: Pat<(i1 (setne (and I32:$Rs, IsPow2_32:$u5), 0)),
1800 (S2_tstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
1801 def: Pat<(i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1802 (S4_ntstbit_r I32:$Rs, I32:$Rt)>;
1803 def: Pat<(i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1804 (S2_tstbit_r I32:$Rs, I32:$Rt)>;
1807 def: Pat<(i1 (seteq (and I64:$Rs, IsPow2_64L:$u6), 0)),
1808 (S4_ntstbit_i (LoReg $Rs), (Log2_64 $u6))>;
1809 def: Pat<(i1 (seteq (and I64:$Rs, IsPow2_64H:$u6), 0)),
1810 (S4_ntstbit_i (HiReg $Rs), (UDEC32 (i32 (Log2_64 $u6))))>;
1811 def: Pat<(i1 (setne (and I64:$Rs, IsPow2_64L:$u6), 0)),
1812 (S2_tstbit_i (LoReg $Rs), (Log2_32 imm:$u6))>;
1813 def: Pat<(i1 (setne (and I64:$Rs, IsPow2_64H:$u6), 0)),
1814 (S2_tstbit_i (HiReg $Rs), (UDEC32 (i32 (Log2_32 imm:$u6))))>;
1816 // Do not increase complexity of these patterns. In the DAG, "cmp i8" may be
1817 // represented as a compare against "value & 0xFF", which is an exact match
1818 // for cmpb (same for cmph). The patterns below do not contain any additional
1819 // complexity that would make them preferable, and if they were actually used
1820 // instead of cmpb/cmph, they would result in a compare against register that
1821 // is loaded with the byte/half mask (i.e. 0xFF or 0xFFFF).
1822 def: Pat<(i1 (setne (and I32:$Rs, u6_0ImmPred:$u6), 0)),
1823 (C4_nbitsclri I32:$Rs, imm:$u6)>;
1824 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), 0)),
1825 (C4_nbitsclr I32:$Rs, I32:$Rt)>;
1826 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), I32:$Rt)),
1827 (C4_nbitsset I32:$Rs, I32:$Rt)>;
1829 // Special patterns to address certain cases where the "top-down" matching
1830 // algorithm would cause suboptimal selection.
1832 let AddedComplexity = 100 in {
1833 // Avoid A4_rcmp[n]eqi in these cases:
1834 def: Pat<(i32 (zext (i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1835 (I1toI32 (S4_ntstbit_r IntRegs:$Rs, IntRegs:$Rt))>;
1836 def: Pat<(i32 (zext (i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1837 (I1toI32 (S2_tstbit_r IntRegs:$Rs, IntRegs:$Rt))>;
1838 def: Pat<(i32 (zext (i1 (seteq (and I32:$Rs, IsPow2_32:$u5), 0)))),
1839 (I1toI32 (S4_ntstbit_i I32:$Rs, (Log2_32 imm:$u5)))>;
1840 def: Pat<(i32 (zext (i1 (setne (and I32:$Rs, IsPow2_32:$u5), 0)))),
1841 (I1toI32 (S2_tstbit_i I32:$Rs, (Log2_32 imm:$u5)))>;
1842 def: Pat<(i32 (zext (i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1843 (I1toI32 (S4_ntstbit_r I32:$Rs, I32:$Rt))>;
1844 def: Pat<(i32 (zext (i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1845 (I1toI32 (S2_tstbit_r I32:$Rs, I32:$Rt))>;
1848 // --(11) PIC ------------------------------------------------------------
1851 def SDT_HexagonAtGot
1852 : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
1853 def SDT_HexagonAtPcrel
1854 : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
1856 // AT_GOT address-of-GOT, address-of-global, offset-in-global
1857 def HexagonAtGot : SDNode<"HexagonISD::AT_GOT", SDT_HexagonAtGot>;
1858 // AT_PCREL address-of-global
1859 def HexagonAtPcrel : SDNode<"HexagonISD::AT_PCREL", SDT_HexagonAtPcrel>;
1861 def: Pat<(HexagonAtGot I32:$got, I32:$addr, (i32 0)),
1862 (L2_loadri_io I32:$got, imm:$addr)>;
1863 def: Pat<(HexagonAtGot I32:$got, I32:$addr, s30_2ImmPred:$off),
1864 (A2_addi (L2_loadri_io I32:$got, imm:$addr), imm:$off)>;
1865 def: Pat<(HexagonAtPcrel I32:$addr),
1866 (C4_addipc imm:$addr)>;
1868 // The HVX load patterns also match AT_PCREL directly. Make sure that
1869 // if the selection of this opcode changes, it's updated in all places.
1872 // --(12) Load -----------------------------------------------------------
1875 def extloadv2i8: PatFrag<(ops node:$ptr), (extload node:$ptr), [{
1876 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1878 def extloadv4i8: PatFrag<(ops node:$ptr), (extload node:$ptr), [{
1879 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1882 def zextloadv2i8: PatFrag<(ops node:$ptr), (zextload node:$ptr), [{
1883 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1885 def zextloadv4i8: PatFrag<(ops node:$ptr), (zextload node:$ptr), [{
1886 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1889 def sextloadv2i8: PatFrag<(ops node:$ptr), (sextload node:$ptr), [{
1890 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1892 def sextloadv4i8: PatFrag<(ops node:$ptr), (sextload node:$ptr), [{
1893 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1896 // Patterns to select load-indexed: Rs + Off.
1897 // - frameindex [+ imm],
1898 multiclass Loadxfi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1900 def: Pat<(VT (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
1901 (VT (MI AddrFI:$fi, imm:$Off))>;
1902 def: Pat<(VT (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
1903 (VT (MI AddrFI:$fi, imm:$Off))>;
1904 def: Pat<(VT (Load AddrFI:$fi)), (VT (MI AddrFI:$fi, 0))>;
1907 // Patterns to select load-indexed: Rs + Off.
1908 // - base reg [+ imm]
1909 multiclass Loadxgi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1911 def: Pat<(VT (Load (add I32:$Rs, ImmPred:$Off))),
1912 (VT (MI IntRegs:$Rs, imm:$Off))>;
1913 def: Pat<(VT (Load (IsOrAdd I32:$Rs, ImmPred:$Off))),
1914 (VT (MI IntRegs:$Rs, imm:$Off))>;
1915 def: Pat<(VT (Load I32:$Rs)), (VT (MI IntRegs:$Rs, 0))>;
1918 // Patterns to select load-indexed: Rs + Off. Combines Loadxfi + Loadxgi.
1919 multiclass Loadxi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1921 defm: Loadxfi_pat<Load, VT, ImmPred, MI>;
1922 defm: Loadxgi_pat<Load, VT, ImmPred, MI>;
1925 // Patterns to select load reg indexed: Rs + Off with a value modifier.
1926 // - frameindex [+ imm]
1927 multiclass Loadxfim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1928 PatLeaf ImmPred, InstHexagon MI> {
1929 def: Pat<(VT (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
1930 (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>;
1931 def: Pat<(VT (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
1932 (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>;
1933 def: Pat<(VT (Load AddrFI:$fi)), (VT (ValueMod (MI AddrFI:$fi, 0)))>;
1936 // Patterns to select load reg indexed: Rs + Off with a value modifier.
1937 // - base reg [+ imm]
1938 multiclass Loadxgim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1939 PatLeaf ImmPred, InstHexagon MI> {
1940 def: Pat<(VT (Load (add I32:$Rs, ImmPred:$Off))),
1941 (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
1942 def: Pat<(VT (Load (IsOrAdd I32:$Rs, ImmPred:$Off))),
1943 (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
1944 def: Pat<(VT (Load I32:$Rs)), (VT (ValueMod (MI IntRegs:$Rs, 0)))>;
1947 // Patterns to select load reg indexed: Rs + Off with a value modifier.
1948 // Combines Loadxfim + Loadxgim.
1949 multiclass Loadxim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1950 PatLeaf ImmPred, InstHexagon MI> {
1951 defm: Loadxfim_pat<Load, VT, ValueMod, ImmPred, MI>;
1952 defm: Loadxgim_pat<Load, VT, ValueMod, ImmPred, MI>;
1955 // Pattern to select load reg reg-indexed: Rs + Rt<<u2.
1956 class Loadxr_shl_pat<PatFrag Load, ValueType VT, InstHexagon MI>
1957 : Pat<(VT (Load (add I32:$Rs, (i32 (shl I32:$Rt, u2_0ImmPred:$u2))))),
1958 (VT (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2))>;
1960 // Pattern to select load reg reg-indexed: Rs + Rt<<0.
1961 class Loadxr_add_pat<PatFrag Load, ValueType VT, InstHexagon MI>
1962 : Pat<(VT (Load (add I32:$Rs, I32:$Rt))),
1963 (VT (MI IntRegs:$Rs, IntRegs:$Rt, 0))>;
1965 // Pattern to select load reg reg-indexed: Rs + Rt<<u2 with value modifier.
1966 class Loadxrm_shl_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1968 : Pat<(VT (Load (add I32:$Rs, (i32 (shl I32:$Rt, u2_0ImmPred:$u2))))),
1969 (VT (ValueMod (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2)))>;
1971 // Pattern to select load reg reg-indexed: Rs + Rt<<0 with value modifier.
1972 class Loadxrm_add_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1974 : Pat<(VT (Load (add I32:$Rs, I32:$Rt))),
1975 (VT (ValueMod (MI IntRegs:$Rs, IntRegs:$Rt, 0)))>;
1977 // Pattern to select load long-offset reg-indexed: Addr + Rt<<u2.
1978 // Don't match for u2==0, instead use reg+imm for those cases.
1979 class Loadxu_pat<PatFrag Load, ValueType VT, PatFrag ImmPred, InstHexagon MI>
1980 : Pat<(VT (Load (add (shl IntRegs:$Rt, u2_0ImmPred:$u2), ImmPred:$Addr))),
1981 (VT (MI IntRegs:$Rt, imm:$u2, ImmPred:$Addr))>;
1983 class Loadxum_pat<PatFrag Load, ValueType VT, PatFrag ImmPred, PatFrag ValueMod,
1985 : Pat<(VT (Load (add (shl IntRegs:$Rt, u2_0ImmPred:$u2), ImmPred:$Addr))),
1986 (VT (ValueMod (MI IntRegs:$Rt, imm:$u2, ImmPred:$Addr)))>;
1988 // Pattern to select load absolute.
1989 class Loada_pat<PatFrag Load, ValueType VT, PatFrag Addr, InstHexagon MI>
1990 : Pat<(VT (Load Addr:$addr)), (MI Addr:$addr)>;
1992 // Pattern to select load absolute with value modifier.
1993 class Loadam_pat<PatFrag Load, ValueType VT, PatFrag Addr, PatFrag ValueMod,
1995 : Pat<(VT (Load Addr:$addr)), (ValueMod (MI Addr:$addr))>;
1998 let AddedComplexity = 20 in {
1999 defm: Loadxi_pat<extloadi1, i32, anyimm0, L2_loadrub_io>;
2000 defm: Loadxi_pat<extloadi8, i32, anyimm0, L2_loadrub_io>;
2001 defm: Loadxi_pat<extloadi16, i32, anyimm1, L2_loadruh_io>;
2002 defm: Loadxi_pat<extloadv2i8, v2i16, anyimm1, L2_loadbzw2_io>;
2003 defm: Loadxi_pat<extloadv4i8, v4i16, anyimm2, L2_loadbzw4_io>;
2004 defm: Loadxi_pat<sextloadi8, i32, anyimm0, L2_loadrb_io>;
2005 defm: Loadxi_pat<sextloadi16, i32, anyimm1, L2_loadrh_io>;
2006 defm: Loadxi_pat<sextloadv2i8, v2i16, anyimm1, L2_loadbsw2_io>;
2007 defm: Loadxi_pat<sextloadv4i8, v4i16, anyimm2, L2_loadbzw4_io>;
2008 defm: Loadxi_pat<zextloadi1, i32, anyimm0, L2_loadrub_io>;
2009 defm: Loadxi_pat<zextloadi8, i32, anyimm0, L2_loadrub_io>;
2010 defm: Loadxi_pat<zextloadi16, i32, anyimm1, L2_loadruh_io>;
2011 defm: Loadxi_pat<zextloadv2i8, v2i16, anyimm1, L2_loadbzw2_io>;
2012 defm: Loadxi_pat<zextloadv4i8, v4i16, anyimm2, L2_loadbzw4_io>;
2013 defm: Loadxi_pat<load, i32, anyimm2, L2_loadri_io>;
2014 defm: Loadxi_pat<load, v2i16, anyimm2, L2_loadri_io>;
2015 defm: Loadxi_pat<load, v4i8, anyimm2, L2_loadri_io>;
2016 defm: Loadxi_pat<load, i64, anyimm3, L2_loadrd_io>;
2017 defm: Loadxi_pat<load, v2i32, anyimm3, L2_loadrd_io>;
2018 defm: Loadxi_pat<load, v4i16, anyimm3, L2_loadrd_io>;
2019 defm: Loadxi_pat<load, v8i8, anyimm3, L2_loadrd_io>;
2020 defm: Loadxi_pat<load, f32, anyimm2, L2_loadri_io>;
2021 defm: Loadxi_pat<load, f64, anyimm3, L2_loadrd_io>;
2024 defm: Loadxi_pat<atomic_load_8 , i32, anyimm0, L2_loadrub_io>;
2025 defm: Loadxi_pat<atomic_load_16, i32, anyimm1, L2_loadruh_io>;
2026 defm: Loadxi_pat<atomic_load_32, i32, anyimm2, L2_loadri_io>;
2027 defm: Loadxi_pat<atomic_load_64, i64, anyimm3, L2_loadrd_io>;
2030 let AddedComplexity = 30 in {
2031 defm: Loadxim_pat<extloadi1, i64, ToAext64, anyimm0, L2_loadrub_io>;
2032 defm: Loadxim_pat<extloadi8, i64, ToAext64, anyimm0, L2_loadrub_io>;
2033 defm: Loadxim_pat<extloadi16, i64, ToAext64, anyimm1, L2_loadruh_io>;
2034 defm: Loadxim_pat<extloadi32, i64, ToAext64, anyimm2, L2_loadri_io>;
2035 defm: Loadxim_pat<zextloadi1, i64, ToZext64, anyimm0, L2_loadrub_io>;
2036 defm: Loadxim_pat<zextloadi8, i64, ToZext64, anyimm0, L2_loadrub_io>;
2037 defm: Loadxim_pat<zextloadi16, i64, ToZext64, anyimm1, L2_loadruh_io>;
2038 defm: Loadxim_pat<zextloadi32, i64, ToZext64, anyimm2, L2_loadri_io>;
2039 defm: Loadxim_pat<sextloadi8, i64, ToSext64, anyimm0, L2_loadrb_io>;
2040 defm: Loadxim_pat<sextloadi16, i64, ToSext64, anyimm1, L2_loadrh_io>;
2041 defm: Loadxim_pat<sextloadi32, i64, ToSext64, anyimm2, L2_loadri_io>;
2044 let AddedComplexity = 60 in {
2045 def: Loadxu_pat<extloadi8, i32, anyimm0, L4_loadrub_ur>;
2046 def: Loadxu_pat<extloadi16, i32, anyimm1, L4_loadruh_ur>;
2047 def: Loadxu_pat<extloadv2i8, v2i16, anyimm1, L4_loadbzw2_ur>;
2048 def: Loadxu_pat<extloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
2049 def: Loadxu_pat<sextloadi8, i32, anyimm0, L4_loadrb_ur>;
2050 def: Loadxu_pat<sextloadi16, i32, anyimm1, L4_loadrh_ur>;
2051 def: Loadxu_pat<sextloadv2i8, v2i16, anyimm1, L4_loadbsw2_ur>;
2052 def: Loadxu_pat<sextloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
2053 def: Loadxu_pat<zextloadi8, i32, anyimm0, L4_loadrub_ur>;
2054 def: Loadxu_pat<zextloadi16, i32, anyimm1, L4_loadruh_ur>;
2055 def: Loadxu_pat<zextloadv2i8, v2i16, anyimm1, L4_loadbzw2_ur>;
2056 def: Loadxu_pat<zextloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
2057 def: Loadxu_pat<load, i32, anyimm2, L4_loadri_ur>;
2058 def: Loadxu_pat<load, v2i16, anyimm2, L4_loadri_ur>;
2059 def: Loadxu_pat<load, v4i8, anyimm2, L4_loadri_ur>;
2060 def: Loadxu_pat<load, i64, anyimm3, L4_loadrd_ur>;
2061 def: Loadxu_pat<load, v2i32, anyimm3, L4_loadrd_ur>;
2062 def: Loadxu_pat<load, v4i16, anyimm3, L4_loadrd_ur>;
2063 def: Loadxu_pat<load, v8i8, anyimm3, L4_loadrd_ur>;
2064 def: Loadxu_pat<load, f32, anyimm2, L4_loadri_ur>;
2065 def: Loadxu_pat<load, f64, anyimm3, L4_loadrd_ur>;
2067 def: Loadxum_pat<sextloadi8, i64, anyimm0, ToSext64, L4_loadrb_ur>;
2068 def: Loadxum_pat<zextloadi8, i64, anyimm0, ToZext64, L4_loadrub_ur>;
2069 def: Loadxum_pat<extloadi8, i64, anyimm0, ToAext64, L4_loadrub_ur>;
2070 def: Loadxum_pat<sextloadi16, i64, anyimm1, ToSext64, L4_loadrh_ur>;
2071 def: Loadxum_pat<zextloadi16, i64, anyimm1, ToZext64, L4_loadruh_ur>;
2072 def: Loadxum_pat<extloadi16, i64, anyimm1, ToAext64, L4_loadruh_ur>;
2073 def: Loadxum_pat<sextloadi32, i64, anyimm2, ToSext64, L4_loadri_ur>;
2074 def: Loadxum_pat<zextloadi32, i64, anyimm2, ToZext64, L4_loadri_ur>;
2075 def: Loadxum_pat<extloadi32, i64, anyimm2, ToAext64, L4_loadri_ur>;
2078 let AddedComplexity = 40 in {
2079 def: Loadxr_shl_pat<extloadi8, i32, L4_loadrub_rr>;
2080 def: Loadxr_shl_pat<zextloadi8, i32, L4_loadrub_rr>;
2081 def: Loadxr_shl_pat<sextloadi8, i32, L4_loadrb_rr>;
2082 def: Loadxr_shl_pat<extloadi16, i32, L4_loadruh_rr>;
2083 def: Loadxr_shl_pat<zextloadi16, i32, L4_loadruh_rr>;
2084 def: Loadxr_shl_pat<sextloadi16, i32, L4_loadrh_rr>;
2085 def: Loadxr_shl_pat<load, i32, L4_loadri_rr>;
2086 def: Loadxr_shl_pat<load, v2i16, L4_loadri_rr>;
2087 def: Loadxr_shl_pat<load, v4i8, L4_loadri_rr>;
2088 def: Loadxr_shl_pat<load, i64, L4_loadrd_rr>;
2089 def: Loadxr_shl_pat<load, v2i32, L4_loadrd_rr>;
2090 def: Loadxr_shl_pat<load, v4i16, L4_loadrd_rr>;
2091 def: Loadxr_shl_pat<load, v8i8, L4_loadrd_rr>;
2092 def: Loadxr_shl_pat<load, f32, L4_loadri_rr>;
2093 def: Loadxr_shl_pat<load, f64, L4_loadrd_rr>;
2096 let AddedComplexity = 20 in {
2097 def: Loadxr_add_pat<extloadi8, i32, L4_loadrub_rr>;
2098 def: Loadxr_add_pat<zextloadi8, i32, L4_loadrub_rr>;
2099 def: Loadxr_add_pat<sextloadi8, i32, L4_loadrb_rr>;
2100 def: Loadxr_add_pat<extloadi16, i32, L4_loadruh_rr>;
2101 def: Loadxr_add_pat<zextloadi16, i32, L4_loadruh_rr>;
2102 def: Loadxr_add_pat<sextloadi16, i32, L4_loadrh_rr>;
2103 def: Loadxr_add_pat<load, i32, L4_loadri_rr>;
2104 def: Loadxr_add_pat<load, v2i16, L4_loadri_rr>;
2105 def: Loadxr_add_pat<load, v4i8, L4_loadri_rr>;
2106 def: Loadxr_add_pat<load, i64, L4_loadrd_rr>;
2107 def: Loadxr_add_pat<load, v2i32, L4_loadrd_rr>;
2108 def: Loadxr_add_pat<load, v4i16, L4_loadrd_rr>;
2109 def: Loadxr_add_pat<load, v8i8, L4_loadrd_rr>;
2110 def: Loadxr_add_pat<load, f32, L4_loadri_rr>;
2111 def: Loadxr_add_pat<load, f64, L4_loadrd_rr>;
2114 let AddedComplexity = 40 in {
2115 def: Loadxrm_shl_pat<extloadi8, i64, ToAext64, L4_loadrub_rr>;
2116 def: Loadxrm_shl_pat<zextloadi8, i64, ToZext64, L4_loadrub_rr>;
2117 def: Loadxrm_shl_pat<sextloadi8, i64, ToSext64, L4_loadrb_rr>;
2118 def: Loadxrm_shl_pat<extloadi16, i64, ToAext64, L4_loadruh_rr>;
2119 def: Loadxrm_shl_pat<zextloadi16, i64, ToZext64, L4_loadruh_rr>;
2120 def: Loadxrm_shl_pat<sextloadi16, i64, ToSext64, L4_loadrh_rr>;
2121 def: Loadxrm_shl_pat<extloadi32, i64, ToAext64, L4_loadri_rr>;
2122 def: Loadxrm_shl_pat<zextloadi32, i64, ToZext64, L4_loadri_rr>;
2123 def: Loadxrm_shl_pat<sextloadi32, i64, ToSext64, L4_loadri_rr>;
2126 let AddedComplexity = 20 in {
2127 def: Loadxrm_add_pat<extloadi8, i64, ToAext64, L4_loadrub_rr>;
2128 def: Loadxrm_add_pat<zextloadi8, i64, ToZext64, L4_loadrub_rr>;
2129 def: Loadxrm_add_pat<sextloadi8, i64, ToSext64, L4_loadrb_rr>;
2130 def: Loadxrm_add_pat<extloadi16, i64, ToAext64, L4_loadruh_rr>;
2131 def: Loadxrm_add_pat<zextloadi16, i64, ToZext64, L4_loadruh_rr>;
2132 def: Loadxrm_add_pat<sextloadi16, i64, ToSext64, L4_loadrh_rr>;
2133 def: Loadxrm_add_pat<extloadi32, i64, ToAext64, L4_loadri_rr>;
2134 def: Loadxrm_add_pat<zextloadi32, i64, ToZext64, L4_loadri_rr>;
2135 def: Loadxrm_add_pat<sextloadi32, i64, ToSext64, L4_loadri_rr>;
2140 let AddedComplexity = 60 in {
2141 def: Loada_pat<zextloadi1, i32, anyimm0, PS_loadrubabs>;
2142 def: Loada_pat<sextloadi8, i32, anyimm0, PS_loadrbabs>;
2143 def: Loada_pat<extloadi8, i32, anyimm0, PS_loadrubabs>;
2144 def: Loada_pat<zextloadi8, i32, anyimm0, PS_loadrubabs>;
2145 def: Loada_pat<sextloadi16, i32, anyimm1, PS_loadrhabs>;
2146 def: Loada_pat<extloadi16, i32, anyimm1, PS_loadruhabs>;
2147 def: Loada_pat<zextloadi16, i32, anyimm1, PS_loadruhabs>;
2148 def: Loada_pat<load, i32, anyimm2, PS_loadriabs>;
2149 def: Loada_pat<load, v2i16, anyimm2, PS_loadriabs>;
2150 def: Loada_pat<load, v4i8, anyimm2, PS_loadriabs>;
2151 def: Loada_pat<load, i64, anyimm3, PS_loadrdabs>;
2152 def: Loada_pat<load, v2i32, anyimm3, PS_loadrdabs>;
2153 def: Loada_pat<load, v4i16, anyimm3, PS_loadrdabs>;
2154 def: Loada_pat<load, v8i8, anyimm3, PS_loadrdabs>;
2155 def: Loada_pat<load, f32, anyimm2, PS_loadriabs>;
2156 def: Loada_pat<load, f64, anyimm3, PS_loadrdabs>;
2158 def: Loada_pat<atomic_load_8, i32, anyimm0, PS_loadrubabs>;
2159 def: Loada_pat<atomic_load_16, i32, anyimm1, PS_loadruhabs>;
2160 def: Loada_pat<atomic_load_32, i32, anyimm2, PS_loadriabs>;
2161 def: Loada_pat<atomic_load_64, i64, anyimm3, PS_loadrdabs>;
2164 let AddedComplexity = 30 in {
2165 def: Loadam_pat<extloadi8, i64, anyimm0, ToAext64, PS_loadrubabs>;
2166 def: Loadam_pat<sextloadi8, i64, anyimm0, ToSext64, PS_loadrbabs>;
2167 def: Loadam_pat<zextloadi8, i64, anyimm0, ToZext64, PS_loadrubabs>;
2168 def: Loadam_pat<extloadi16, i64, anyimm1, ToAext64, PS_loadruhabs>;
2169 def: Loadam_pat<sextloadi16, i64, anyimm1, ToSext64, PS_loadrhabs>;
2170 def: Loadam_pat<zextloadi16, i64, anyimm1, ToZext64, PS_loadruhabs>;
2171 def: Loadam_pat<extloadi32, i64, anyimm2, ToAext64, PS_loadriabs>;
2172 def: Loadam_pat<sextloadi32, i64, anyimm2, ToSext64, PS_loadriabs>;
2173 def: Loadam_pat<zextloadi32, i64, anyimm2, ToZext64, PS_loadriabs>;
2175 def: Loadam_pat<load, i1, anyimm0, I32toI1, PS_loadrubabs>;
2176 def: Loadam_pat<zextloadi1, i64, anyimm0, ToZext64, PS_loadrubabs>;
2179 // GP-relative address
2181 let AddedComplexity = 100 in {
2182 def: Loada_pat<extloadi1, i32, addrgp, L2_loadrubgp>;
2183 def: Loada_pat<zextloadi1, i32, addrgp, L2_loadrubgp>;
2184 def: Loada_pat<extloadi8, i32, addrgp, L2_loadrubgp>;
2185 def: Loada_pat<sextloadi8, i32, addrgp, L2_loadrbgp>;
2186 def: Loada_pat<zextloadi8, i32, addrgp, L2_loadrubgp>;
2187 def: Loada_pat<extloadi16, i32, addrgp, L2_loadruhgp>;
2188 def: Loada_pat<sextloadi16, i32, addrgp, L2_loadrhgp>;
2189 def: Loada_pat<zextloadi16, i32, addrgp, L2_loadruhgp>;
2190 def: Loada_pat<load, i32, addrgp, L2_loadrigp>;
2191 def: Loada_pat<load, v2i16, addrgp, L2_loadrigp>;
2192 def: Loada_pat<load, v4i8, addrgp, L2_loadrigp>;
2193 def: Loada_pat<load, i64, addrgp, L2_loadrdgp>;
2194 def: Loada_pat<load, v2i32, addrgp, L2_loadrdgp>;
2195 def: Loada_pat<load, v4i16, addrgp, L2_loadrdgp>;
2196 def: Loada_pat<load, v8i8, addrgp, L2_loadrdgp>;
2197 def: Loada_pat<load, f32, addrgp, L2_loadrigp>;
2198 def: Loada_pat<load, f64, addrgp, L2_loadrdgp>;
2200 def: Loada_pat<atomic_load_8, i32, addrgp, L2_loadrubgp>;
2201 def: Loada_pat<atomic_load_16, i32, addrgp, L2_loadruhgp>;
2202 def: Loada_pat<atomic_load_32, i32, addrgp, L2_loadrigp>;
2203 def: Loada_pat<atomic_load_64, i64, addrgp, L2_loadrdgp>;
2206 let AddedComplexity = 70 in {
2207 def: Loadam_pat<extloadi8, i64, addrgp, ToAext64, L2_loadrubgp>;
2208 def: Loadam_pat<sextloadi8, i64, addrgp, ToSext64, L2_loadrbgp>;
2209 def: Loadam_pat<zextloadi8, i64, addrgp, ToZext64, L2_loadrubgp>;
2210 def: Loadam_pat<extloadi16, i64, addrgp, ToAext64, L2_loadruhgp>;
2211 def: Loadam_pat<sextloadi16, i64, addrgp, ToSext64, L2_loadrhgp>;
2212 def: Loadam_pat<zextloadi16, i64, addrgp, ToZext64, L2_loadruhgp>;
2213 def: Loadam_pat<extloadi32, i64, addrgp, ToAext64, L2_loadrigp>;
2214 def: Loadam_pat<sextloadi32, i64, addrgp, ToSext64, L2_loadrigp>;
2215 def: Loadam_pat<zextloadi32, i64, addrgp, ToZext64, L2_loadrigp>;
2217 def: Loadam_pat<load, i1, addrgp, I32toI1, L2_loadrubgp>;
2218 def: Loadam_pat<zextloadi1, i64, addrgp, ToZext64, L2_loadrubgp>;
2222 // Sign-extending loads of i1 need to replicate the lowest bit throughout
2223 // the 32-bit value. Since the loaded value can only be 0 or 1, 0-v should
2225 let AddedComplexity = 20 in
2226 def: Pat<(i32 (sextloadi1 I32:$Rs)),
2227 (A2_subri 0, (L2_loadrub_io IntRegs:$Rs, 0))>;
2229 // Patterns for loads of i1:
2230 def: Pat<(i1 (load AddrFI:$fi)),
2231 (C2_tfrrp (L2_loadrub_io AddrFI:$fi, 0))>;
2232 def: Pat<(i1 (load (add I32:$Rs, anyimm0:$Off))),
2233 (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, imm:$Off))>;
2234 def: Pat<(i1 (load I32:$Rs)),
2235 (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, 0))>;
2238 // --(13) Store ----------------------------------------------------------
2241 class Storepi_pat<PatFrag Store, PatFrag Value, PatFrag Offset, InstHexagon MI>
2242 : Pat<(Store Value:$Rt, I32:$Rx, Offset:$s4),
2243 (MI I32:$Rx, imm:$s4, Value:$Rt)>;
2245 def: Storepi_pat<post_truncsti8, I32, s4_0ImmPred, S2_storerb_pi>;
2246 def: Storepi_pat<post_truncsti16, I32, s4_1ImmPred, S2_storerh_pi>;
2247 def: Storepi_pat<post_store, I32, s4_2ImmPred, S2_storeri_pi>;
2248 def: Storepi_pat<post_store, I64, s4_3ImmPred, S2_storerd_pi>;
2250 // Patterns for generating stores, where the address takes different forms:
2252 // - frameindex + offset,
2254 // - simple (base address without offset).
2255 // These would usually be used together (via Storexi_pat defined below), but
2256 // in some cases one may want to apply different properties (such as
2257 // AddedComplexity) to the individual patterns.
2258 class Storexi_fi_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2259 : Pat<(Store Value:$Rs, AddrFI:$fi), (MI AddrFI:$fi, 0, Value:$Rs)>;
2261 multiclass Storexi_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2263 def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
2264 (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
2265 def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
2266 (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
2269 multiclass Storexi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2271 def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
2272 (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
2273 def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
2274 (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
2277 class Storexi_base_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2278 : Pat<(Store Value:$Rt, I32:$Rs),
2279 (MI IntRegs:$Rs, 0, Value:$Rt)>;
2281 // Patterns for generating stores, where the address takes different forms,
2282 // and where the value being stored is transformed through the value modifier
2283 // ValueMod. The address forms are same as above.
2284 class Storexim_fi_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2286 : Pat<(Store Value:$Rs, AddrFI:$fi),
2287 (MI AddrFI:$fi, 0, (ValueMod Value:$Rs))>;
2289 multiclass Storexim_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2290 PatFrag ValueMod, InstHexagon MI> {
2291 def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
2292 (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
2293 def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
2294 (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
2297 multiclass Storexim_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2298 PatFrag ValueMod, InstHexagon MI> {
2299 def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
2300 (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
2301 def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
2302 (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
2305 class Storexim_base_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2307 : Pat<(Store Value:$Rt, I32:$Rs),
2308 (MI IntRegs:$Rs, 0, (ValueMod Value:$Rt))>;
2310 multiclass Storexi_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
2312 defm: Storexi_fi_add_pat <Store, Value, ImmPred, MI>;
2313 def: Storexi_fi_pat <Store, Value, MI>;
2314 defm: Storexi_add_pat <Store, Value, ImmPred, MI>;
2317 multiclass Storexim_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
2318 PatFrag ValueMod, InstHexagon MI> {
2319 defm: Storexim_fi_add_pat <Store, Value, ImmPred, ValueMod, MI>;
2320 def: Storexim_fi_pat <Store, Value, ValueMod, MI>;
2321 defm: Storexim_add_pat <Store, Value, ImmPred, ValueMod, MI>;
2325 class Storexu_shl_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred, InstHexagon MI>
2326 : Pat<(Store Value:$Rt, (add (shl I32:$Ru, u2_0ImmPred:$u2), ImmPred:$A)),
2327 (MI IntRegs:$Ru, imm:$u2, ImmPred:$A, Value:$Rt)>;
2330 class Storexr_shl_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2331 : Pat<(Store Value:$Ru, (add I32:$Rs, (shl I32:$Rt, u2_0ImmPred:$u2))),
2332 (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2, Value:$Ru)>;
2335 class Storexr_add_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2336 : Pat<(Store Value:$Ru, (add I32:$Rs, I32:$Rt)),
2337 (MI IntRegs:$Rs, IntRegs:$Rt, 0, Value:$Ru)>;
2339 class Storea_pat<PatFrag Store, PatFrag Value, PatFrag Addr, InstHexagon MI>
2340 : Pat<(Store Value:$val, Addr:$addr), (MI Addr:$addr, Value:$val)>;
2342 class Stoream_pat<PatFrag Store, PatFrag Value, PatFrag Addr, PatFrag ValueMod,
2344 : Pat<(Store Value:$val, Addr:$addr),
2345 (MI Addr:$addr, (ValueMod Value:$val))>;
2347 // Regular stores in the DAG have two operands: value and address.
2348 // Atomic stores also have two, but they are reversed: address, value.
2349 // To use atomic stores with the patterns, they need to have their operands
2350 // swapped. This relies on the knowledge that the F.Fragment uses names
2352 class AtomSt<PatFrag F>
2353 : PatFrag<(ops node:$val, node:$ptr), !head(F.Fragments), F.PredicateCode,
2354 F.OperandTransform> {
2355 let IsAtomic = F.IsAtomic;
2356 let MemoryVT = F.MemoryVT;
2360 def IMM_BYTE : SDNodeXForm<imm, [{
2361 // -1 can be represented as 255, etc.
2362 // assigning to a byte restores our desired signed value.
2363 int8_t imm = N->getSExtValue();
2364 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2367 def IMM_HALF : SDNodeXForm<imm, [{
2368 // -1 can be represented as 65535, etc.
2369 // assigning to a short restores our desired signed value.
2370 int16_t imm = N->getSExtValue();
2371 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2374 def IMM_WORD : SDNodeXForm<imm, [{
2375 // -1 can be represented as 4294967295, etc.
2376 // Currently, it's not doing this. But some optimization
2377 // might convert -1 to a large +ve number.
2378 // assigning to a word restores our desired signed value.
2379 int32_t imm = N->getSExtValue();
2380 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2383 def ToImmByte : OutPatFrag<(ops node:$R), (IMM_BYTE $R)>;
2384 def ToImmHalf : OutPatFrag<(ops node:$R), (IMM_HALF $R)>;
2385 def ToImmWord : OutPatFrag<(ops node:$R), (IMM_WORD $R)>;
2387 // Even though the offset is not extendable in the store-immediate, we
2388 // can still generate the fi# in the base address. If the final offset
2389 // is not valid for the instruction, we will replace it with a scratch
2391 class SmallStackStore<PatFrag Store>
2392 : PatFrag<(ops node:$Val, node:$Addr), (Store node:$Val, node:$Addr), [{
2393 return isSmallStackStore(cast<StoreSDNode>(N));
2396 // This is the complement of SmallStackStore.
2397 class LargeStackStore<PatFrag Store>
2398 : PatFrag<(ops node:$Val, node:$Addr), (Store node:$Val, node:$Addr), [{
2399 return !isSmallStackStore(cast<StoreSDNode>(N));
2402 // Preferred addressing modes for various combinations of stored value
2403 // and address computation.
2404 // For stores where the address and value are both immediates, prefer
2405 // store-immediate. The reason is that the constant-extender optimization
2406 // can replace store-immediate with a store-register, but there is nothing
2407 // to generate a store-immediate out of a store-register.
2409 // C R F F+C R+C R+R R<<S+C R<<S+R
2410 // --+-------+-----+-----+------+-----+-----+--------+--------
2411 // C | imm | imm | imm | imm | imm | rr | ur | rr
2412 // R | abs* | io | io | io | io | rr | ur | rr
2414 // (*) Absolute or GP-relative.
2416 // Note that any expression can be matched by Reg. In particular, an immediate
2417 // can always be placed in a register, so patterns checking for Imm should
2418 // have a higher priority than the ones involving Reg that could also match.
2419 // For example, *(p+4) could become r1=#4; memw(r0+r1<<#0) instead of the
2420 // preferred memw(r0+#4). Similarly Reg+Imm or Reg+Reg should be tried before
2423 // The order in which the different combinations are tried:
2425 // C F R F+C R+C R+R R<<S+C R<<S+R
2426 // --+-------+-----+-----+------+-----+-----+--------+--------
2427 // C | 1 | 6 | - | 5 | 9 | - | - | -
2428 // R | 2 | 8 | 12 | 7 | 10 | 11 | 3 | 4
2431 // First, match the unusual case of doubleword store into Reg+Imm4, i.e.
2432 // a store where the offset Imm4 is a multiple of 4, but not of 8. This
2433 // implies that Reg is also a proper multiple of 4. To still generate a
2434 // doubleword store, add 4 to Reg, and subtract 4 from the offset.
2436 def s30_2ProperPred : PatLeaf<(i32 imm), [{
2437 int64_t v = (int64_t)N->getSExtValue();
2438 return isShiftedInt<30,2>(v) && !isShiftedInt<29,3>(v);
2440 def RoundTo8 : SDNodeXForm<imm, [{
2441 int32_t Imm = N->getSExtValue();
2442 return CurDAG->getTargetConstant(Imm & -8, SDLoc(N), MVT::i32);
2445 let AddedComplexity = 150 in
2446 def: Pat<(store I64:$Ru, (add I32:$Rs, s30_2ProperPred:$Off)),
2447 (S2_storerd_io (A2_addi I32:$Rs, 4), (RoundTo8 $Off), I64:$Ru)>;
2449 class Storexi_abs_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2450 : Pat<(Store Value:$val, anyimm:$addr),
2451 (MI (ToI32 $addr), 0, Value:$val)>;
2452 class Storexim_abs_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2454 : Pat<(Store Value:$val, anyimm:$addr),
2455 (MI (ToI32 $addr), 0, (ValueMod Value:$val))>;
2457 let AddedComplexity = 140 in {
2458 def: Storexim_abs_pat<truncstorei8, anyint, ToImmByte, S4_storeirb_io>;
2459 def: Storexim_abs_pat<truncstorei16, anyint, ToImmHalf, S4_storeirh_io>;
2460 def: Storexim_abs_pat<store, anyint, ToImmWord, S4_storeiri_io>;
2462 def: Storexi_abs_pat<truncstorei8, anyimm, S4_storeirb_io>;
2463 def: Storexi_abs_pat<truncstorei16, anyimm, S4_storeirh_io>;
2464 def: Storexi_abs_pat<store, anyimm, S4_storeiri_io>;
2467 // GP-relative address
2468 let AddedComplexity = 120 in {
2469 def: Storea_pat<truncstorei8, I32, addrgp, S2_storerbgp>;
2470 def: Storea_pat<truncstorei16, I32, addrgp, S2_storerhgp>;
2471 def: Storea_pat<store, I32, addrgp, S2_storerigp>;
2472 def: Storea_pat<store, V4I8, addrgp, S2_storerigp>;
2473 def: Storea_pat<store, V2I16, addrgp, S2_storerigp>;
2474 def: Storea_pat<store, I64, addrgp, S2_storerdgp>;
2475 def: Storea_pat<store, V8I8, addrgp, S2_storerdgp>;
2476 def: Storea_pat<store, V4I16, addrgp, S2_storerdgp>;
2477 def: Storea_pat<store, V2I32, addrgp, S2_storerdgp>;
2478 def: Storea_pat<store, F32, addrgp, S2_storerigp>;
2479 def: Storea_pat<store, F64, addrgp, S2_storerdgp>;
2480 def: Storea_pat<AtomSt<atomic_store_8>, I32, addrgp, S2_storerbgp>;
2481 def: Storea_pat<AtomSt<atomic_store_16>, I32, addrgp, S2_storerhgp>;
2482 def: Storea_pat<AtomSt<atomic_store_32>, I32, addrgp, S2_storerigp>;
2483 def: Storea_pat<AtomSt<atomic_store_32>, V4I8, addrgp, S2_storerigp>;
2484 def: Storea_pat<AtomSt<atomic_store_32>, V2I16, addrgp, S2_storerigp>;
2485 def: Storea_pat<AtomSt<atomic_store_64>, I64, addrgp, S2_storerdgp>;
2486 def: Storea_pat<AtomSt<atomic_store_64>, V8I8, addrgp, S2_storerdgp>;
2487 def: Storea_pat<AtomSt<atomic_store_64>, V4I16, addrgp, S2_storerdgp>;
2488 def: Storea_pat<AtomSt<atomic_store_64>, V2I32, addrgp, S2_storerdgp>;
2490 def: Stoream_pat<truncstorei8, I64, addrgp, LoReg, S2_storerbgp>;
2491 def: Stoream_pat<truncstorei16, I64, addrgp, LoReg, S2_storerhgp>;
2492 def: Stoream_pat<truncstorei32, I64, addrgp, LoReg, S2_storerigp>;
2493 def: Stoream_pat<store, I1, addrgp, I1toI32, S2_storerbgp>;
2497 let AddedComplexity = 110 in {
2498 def: Storea_pat<truncstorei8, I32, anyimm0, PS_storerbabs>;
2499 def: Storea_pat<truncstorei16, I32, anyimm1, PS_storerhabs>;
2500 def: Storea_pat<store, I32, anyimm2, PS_storeriabs>;
2501 def: Storea_pat<store, V4I8, anyimm2, PS_storeriabs>;
2502 def: Storea_pat<store, V2I16, anyimm2, PS_storeriabs>;
2503 def: Storea_pat<store, I64, anyimm3, PS_storerdabs>;
2504 def: Storea_pat<store, V8I8, anyimm3, PS_storerdabs>;
2505 def: Storea_pat<store, V4I16, anyimm3, PS_storerdabs>;
2506 def: Storea_pat<store, V2I32, anyimm3, PS_storerdabs>;
2507 def: Storea_pat<store, F32, anyimm2, PS_storeriabs>;
2508 def: Storea_pat<store, F64, anyimm3, PS_storerdabs>;
2509 def: Storea_pat<AtomSt<atomic_store_8>, I32, anyimm0, PS_storerbabs>;
2510 def: Storea_pat<AtomSt<atomic_store_16>, I32, anyimm1, PS_storerhabs>;
2511 def: Storea_pat<AtomSt<atomic_store_32>, I32, anyimm2, PS_storeriabs>;
2512 def: Storea_pat<AtomSt<atomic_store_32>, V4I8, anyimm2, PS_storeriabs>;
2513 def: Storea_pat<AtomSt<atomic_store_32>, V2I16, anyimm2, PS_storeriabs>;
2514 def: Storea_pat<AtomSt<atomic_store_64>, I64, anyimm3, PS_storerdabs>;
2515 def: Storea_pat<AtomSt<atomic_store_64>, V8I8, anyimm3, PS_storerdabs>;
2516 def: Storea_pat<AtomSt<atomic_store_64>, V4I16, anyimm3, PS_storerdabs>;
2517 def: Storea_pat<AtomSt<atomic_store_64>, V2I32, anyimm3, PS_storerdabs>;
2519 def: Stoream_pat<truncstorei8, I64, anyimm0, LoReg, PS_storerbabs>;
2520 def: Stoream_pat<truncstorei16, I64, anyimm1, LoReg, PS_storerhabs>;
2521 def: Stoream_pat<truncstorei32, I64, anyimm2, LoReg, PS_storeriabs>;
2522 def: Stoream_pat<store, I1, anyimm0, I1toI32, PS_storerbabs>;
2526 let AddedComplexity = 100 in {
2527 def: Storexu_shl_pat<truncstorei8, I32, anyimm0, S4_storerb_ur>;
2528 def: Storexu_shl_pat<truncstorei16, I32, anyimm1, S4_storerh_ur>;
2529 def: Storexu_shl_pat<store, I32, anyimm2, S4_storeri_ur>;
2530 def: Storexu_shl_pat<store, V4I8, anyimm2, S4_storeri_ur>;
2531 def: Storexu_shl_pat<store, V2I16, anyimm2, S4_storeri_ur>;
2532 def: Storexu_shl_pat<store, I64, anyimm3, S4_storerd_ur>;
2533 def: Storexu_shl_pat<store, V8I8, anyimm3, S4_storerd_ur>;
2534 def: Storexu_shl_pat<store, V4I16, anyimm3, S4_storerd_ur>;
2535 def: Storexu_shl_pat<store, V2I32, anyimm3, S4_storerd_ur>;
2536 def: Storexu_shl_pat<store, F32, anyimm2, S4_storeri_ur>;
2537 def: Storexu_shl_pat<store, F64, anyimm3, S4_storerd_ur>;
2539 def: Pat<(store I1:$Pu, (add (shl I32:$Rs, u2_0ImmPred:$u2), anyimm:$A)),
2540 (S4_storerb_ur IntRegs:$Rs, imm:$u2, imm:$A, (I1toI32 I1:$Pu))>;
2544 let AddedComplexity = 90 in {
2545 def: Storexr_shl_pat<truncstorei8, I32, S4_storerb_rr>;
2546 def: Storexr_shl_pat<truncstorei16, I32, S4_storerh_rr>;
2547 def: Storexr_shl_pat<store, I32, S4_storeri_rr>;
2548 def: Storexr_shl_pat<store, V4I8, S4_storeri_rr>;
2549 def: Storexr_shl_pat<store, V2I16, S4_storeri_rr>;
2550 def: Storexr_shl_pat<store, I64, S4_storerd_rr>;
2551 def: Storexr_shl_pat<store, V8I8, S4_storerd_rr>;
2552 def: Storexr_shl_pat<store, V4I16, S4_storerd_rr>;
2553 def: Storexr_shl_pat<store, V2I32, S4_storerd_rr>;
2554 def: Storexr_shl_pat<store, F32, S4_storeri_rr>;
2555 def: Storexr_shl_pat<store, F64, S4_storerd_rr>;
2557 def: Pat<(store I1:$Pu, (add (shl I32:$Rs, u2_0ImmPred:$u2), I32:$Rt)),
2558 (S4_storerb_ur IntRegs:$Rt, IntRegs:$Rs, imm:$u2, (I1toI32 I1:$Pu))>;
2561 class SS_<PatFrag F> : SmallStackStore<F>;
2562 class LS_<PatFrag F> : LargeStackStore<F>;
2564 multiclass IMFA_<PatFrag S, PatFrag V, PatFrag O, PatFrag M, InstHexagon I> {
2565 defm: Storexim_fi_add_pat<S, V, O, M, I>;
2567 multiclass IFA_<PatFrag S, PatFrag V, PatFrag O, InstHexagon I> {
2568 defm: Storexi_fi_add_pat<S, V, O, I>;
2571 // Fi+Imm, store-immediate
2572 let AddedComplexity = 80 in {
2573 defm: IMFA_<SS_<truncstorei8>, anyint, u6_0ImmPred, ToImmByte, S4_storeirb_io>;
2574 defm: IMFA_<SS_<truncstorei16>, anyint, u6_1ImmPred, ToImmHalf, S4_storeirh_io>;
2575 defm: IMFA_<SS_<store>, anyint, u6_2ImmPred, ToImmWord, S4_storeiri_io>;
2577 defm: IFA_<SS_<truncstorei8>, anyimm, u6_0ImmPred, S4_storeirb_io>;
2578 defm: IFA_<SS_<truncstorei16>, anyimm, u6_1ImmPred, S4_storeirh_io>;
2579 defm: IFA_<SS_<store>, anyimm, u6_2ImmPred, S4_storeiri_io>;
2581 // For large-stack stores, generate store-register (prefer explicit Fi
2583 defm: IMFA_<LS_<truncstorei8>, anyimm, u6_0ImmPred, ToI32, S2_storerb_io>;
2584 defm: IMFA_<LS_<truncstorei16>, anyimm, u6_1ImmPred, ToI32, S2_storerh_io>;
2585 defm: IMFA_<LS_<store>, anyimm, u6_2ImmPred, ToI32, S2_storeri_io>;
2588 // Fi, store-immediate
2589 let AddedComplexity = 70 in {
2590 def: Storexim_fi_pat<SS_<truncstorei8>, anyint, ToImmByte, S4_storeirb_io>;
2591 def: Storexim_fi_pat<SS_<truncstorei16>, anyint, ToImmHalf, S4_storeirh_io>;
2592 def: Storexim_fi_pat<SS_<store>, anyint, ToImmWord, S4_storeiri_io>;
2594 def: Storexi_fi_pat<SS_<truncstorei8>, anyimm, S4_storeirb_io>;
2595 def: Storexi_fi_pat<SS_<truncstorei16>, anyimm, S4_storeirh_io>;
2596 def: Storexi_fi_pat<SS_<store>, anyimm, S4_storeiri_io>;
2598 // For large-stack stores, generate store-register (prefer explicit Fi
2600 def: Storexim_fi_pat<LS_<truncstorei8>, anyimm, ToI32, S2_storerb_io>;
2601 def: Storexim_fi_pat<LS_<truncstorei16>, anyimm, ToI32, S2_storerh_io>;
2602 def: Storexim_fi_pat<LS_<store>, anyimm, ToI32, S2_storeri_io>;
2605 // Fi+Imm, Fi, store-register
2606 let AddedComplexity = 60 in {
2607 defm: Storexi_fi_add_pat<truncstorei8, I32, anyimm, S2_storerb_io>;
2608 defm: Storexi_fi_add_pat<truncstorei16, I32, anyimm, S2_storerh_io>;
2609 defm: Storexi_fi_add_pat<store, I32, anyimm, S2_storeri_io>;
2610 defm: Storexi_fi_add_pat<store, V4I8, anyimm, S2_storeri_io>;
2611 defm: Storexi_fi_add_pat<store, V2I16, anyimm, S2_storeri_io>;
2612 defm: Storexi_fi_add_pat<store, I64, anyimm, S2_storerd_io>;
2613 defm: Storexi_fi_add_pat<store, V8I8, anyimm, S2_storerd_io>;
2614 defm: Storexi_fi_add_pat<store, V4I16, anyimm, S2_storerd_io>;
2615 defm: Storexi_fi_add_pat<store, V2I32, anyimm, S2_storerd_io>;
2616 defm: Storexi_fi_add_pat<store, F32, anyimm, S2_storeri_io>;
2617 defm: Storexi_fi_add_pat<store, F64, anyimm, S2_storerd_io>;
2618 defm: Storexim_fi_add_pat<store, I1, anyimm, I1toI32, S2_storerb_io>;
2620 def: Storexi_fi_pat<truncstorei8, I32, S2_storerb_io>;
2621 def: Storexi_fi_pat<truncstorei16, I32, S2_storerh_io>;
2622 def: Storexi_fi_pat<store, I32, S2_storeri_io>;
2623 def: Storexi_fi_pat<store, V4I8, S2_storeri_io>;
2624 def: Storexi_fi_pat<store, V2I16, S2_storeri_io>;
2625 def: Storexi_fi_pat<store, I64, S2_storerd_io>;
2626 def: Storexi_fi_pat<store, V8I8, S2_storerd_io>;
2627 def: Storexi_fi_pat<store, V4I16, S2_storerd_io>;
2628 def: Storexi_fi_pat<store, V2I32, S2_storerd_io>;
2629 def: Storexi_fi_pat<store, F32, S2_storeri_io>;
2630 def: Storexi_fi_pat<store, F64, S2_storerd_io>;
2631 def: Storexim_fi_pat<store, I1, I1toI32, S2_storerb_io>;
2635 multiclass IMRA_<PatFrag S, PatFrag V, PatFrag O, PatFrag M, InstHexagon I> {
2636 defm: Storexim_add_pat<S, V, O, M, I>;
2638 multiclass IRA_<PatFrag S, PatFrag V, PatFrag O, InstHexagon I> {
2639 defm: Storexi_add_pat<S, V, O, I>;
2642 // Reg+Imm, store-immediate
2643 let AddedComplexity = 50 in {
2644 defm: IMRA_<truncstorei8, anyint, u6_0ImmPred, ToImmByte, S4_storeirb_io>;
2645 defm: IMRA_<truncstorei16, anyint, u6_1ImmPred, ToImmHalf, S4_storeirh_io>;
2646 defm: IMRA_<store, anyint, u6_2ImmPred, ToImmWord, S4_storeiri_io>;
2648 defm: IRA_<truncstorei8, anyimm, u6_0ImmPred, S4_storeirb_io>;
2649 defm: IRA_<truncstorei16, anyimm, u6_1ImmPred, S4_storeirh_io>;
2650 defm: IRA_<store, anyimm, u6_2ImmPred, S4_storeiri_io>;
2653 // Reg+Imm, store-register
2654 let AddedComplexity = 40 in {
2655 defm: Storexi_pat<truncstorei8, I32, anyimm0, S2_storerb_io>;
2656 defm: Storexi_pat<truncstorei16, I32, anyimm1, S2_storerh_io>;
2657 defm: Storexi_pat<store, I32, anyimm2, S2_storeri_io>;
2658 defm: Storexi_pat<store, V4I8, anyimm2, S2_storeri_io>;
2659 defm: Storexi_pat<store, V2I16, anyimm2, S2_storeri_io>;
2660 defm: Storexi_pat<store, I64, anyimm3, S2_storerd_io>;
2661 defm: Storexi_pat<store, V8I8, anyimm3, S2_storerd_io>;
2662 defm: Storexi_pat<store, V4I16, anyimm3, S2_storerd_io>;
2663 defm: Storexi_pat<store, V2I32, anyimm3, S2_storerd_io>;
2664 defm: Storexi_pat<store, F32, anyimm2, S2_storeri_io>;
2665 defm: Storexi_pat<store, F64, anyimm3, S2_storerd_io>;
2667 defm: Storexim_pat<truncstorei8, I64, anyimm0, LoReg, S2_storerb_io>;
2668 defm: Storexim_pat<truncstorei16, I64, anyimm1, LoReg, S2_storerh_io>;
2669 defm: Storexim_pat<truncstorei32, I64, anyimm2, LoReg, S2_storeri_io>;
2670 defm: Storexim_pat<store, I1, anyimm0, I1toI32, S2_storerb_io>;
2672 defm: Storexi_pat<AtomSt<atomic_store_8>, I32, anyimm0, S2_storerb_io>;
2673 defm: Storexi_pat<AtomSt<atomic_store_16>, I32, anyimm1, S2_storerh_io>;
2674 defm: Storexi_pat<AtomSt<atomic_store_32>, I32, anyimm2, S2_storeri_io>;
2675 defm: Storexi_pat<AtomSt<atomic_store_32>, V4I8, anyimm2, S2_storeri_io>;
2676 defm: Storexi_pat<AtomSt<atomic_store_32>, V2I16, anyimm2, S2_storeri_io>;
2677 defm: Storexi_pat<AtomSt<atomic_store_64>, I64, anyimm3, S2_storerd_io>;
2678 defm: Storexi_pat<AtomSt<atomic_store_64>, V8I8, anyimm3, S2_storerd_io>;
2679 defm: Storexi_pat<AtomSt<atomic_store_64>, V4I16, anyimm3, S2_storerd_io>;
2680 defm: Storexi_pat<AtomSt<atomic_store_64>, V2I32, anyimm3, S2_storerd_io>;
2684 let AddedComplexity = 30 in {
2685 def: Storexr_add_pat<truncstorei8, I32, S4_storerb_rr>;
2686 def: Storexr_add_pat<truncstorei16, I32, S4_storerh_rr>;
2687 def: Storexr_add_pat<store, I32, S4_storeri_rr>;
2688 def: Storexr_add_pat<store, V4I8, S4_storeri_rr>;
2689 def: Storexr_add_pat<store, V2I16, S4_storeri_rr>;
2690 def: Storexr_add_pat<store, I64, S4_storerd_rr>;
2691 def: Storexr_add_pat<store, V8I8, S4_storerd_rr>;
2692 def: Storexr_add_pat<store, V4I16, S4_storerd_rr>;
2693 def: Storexr_add_pat<store, V2I32, S4_storerd_rr>;
2694 def: Storexr_add_pat<store, F32, S4_storeri_rr>;
2695 def: Storexr_add_pat<store, F64, S4_storerd_rr>;
2697 def: Pat<(store I1:$Pu, (add I32:$Rs, I32:$Rt)),
2698 (S4_storerb_rr IntRegs:$Rs, IntRegs:$Rt, 0, (I1toI32 I1:$Pu))>;
2701 // Reg, store-immediate
2702 let AddedComplexity = 20 in {
2703 def: Storexim_base_pat<truncstorei8, anyint, ToImmByte, S4_storeirb_io>;
2704 def: Storexim_base_pat<truncstorei16, anyint, ToImmHalf, S4_storeirh_io>;
2705 def: Storexim_base_pat<store, anyint, ToImmWord, S4_storeiri_io>;
2707 def: Storexi_base_pat<truncstorei8, anyimm, S4_storeirb_io>;
2708 def: Storexi_base_pat<truncstorei16, anyimm, S4_storeirh_io>;
2709 def: Storexi_base_pat<store, anyimm, S4_storeiri_io>;
2712 // Reg, store-register
2713 let AddedComplexity = 10 in {
2714 def: Storexi_base_pat<truncstorei8, I32, S2_storerb_io>;
2715 def: Storexi_base_pat<truncstorei16, I32, S2_storerh_io>;
2716 def: Storexi_base_pat<store, I32, S2_storeri_io>;
2717 def: Storexi_base_pat<store, V4I8, S2_storeri_io>;
2718 def: Storexi_base_pat<store, V2I16, S2_storeri_io>;
2719 def: Storexi_base_pat<store, I64, S2_storerd_io>;
2720 def: Storexi_base_pat<store, V8I8, S2_storerd_io>;
2721 def: Storexi_base_pat<store, V4I16, S2_storerd_io>;
2722 def: Storexi_base_pat<store, V2I32, S2_storerd_io>;
2723 def: Storexi_base_pat<store, F32, S2_storeri_io>;
2724 def: Storexi_base_pat<store, F64, S2_storerd_io>;
2726 def: Storexim_base_pat<truncstorei8, I64, LoReg, S2_storerb_io>;
2727 def: Storexim_base_pat<truncstorei16, I64, LoReg, S2_storerh_io>;
2728 def: Storexim_base_pat<truncstorei32, I64, LoReg, S2_storeri_io>;
2729 def: Storexim_base_pat<store, I1, I1toI32, S2_storerb_io>;
2731 def: Storexi_base_pat<AtomSt<atomic_store_8>, I32, S2_storerb_io>;
2732 def: Storexi_base_pat<AtomSt<atomic_store_16>, I32, S2_storerh_io>;
2733 def: Storexi_base_pat<AtomSt<atomic_store_32>, I32, S2_storeri_io>;
2734 def: Storexi_base_pat<AtomSt<atomic_store_32>, V4I8, S2_storeri_io>;
2735 def: Storexi_base_pat<AtomSt<atomic_store_32>, V2I16, S2_storeri_io>;
2736 def: Storexi_base_pat<AtomSt<atomic_store_64>, I64, S2_storerd_io>;
2737 def: Storexi_base_pat<AtomSt<atomic_store_64>, V8I8, S2_storerd_io>;
2738 def: Storexi_base_pat<AtomSt<atomic_store_64>, V4I16, S2_storerd_io>;
2739 def: Storexi_base_pat<AtomSt<atomic_store_64>, V2I32, S2_storerd_io>;
2743 // --(14) Memop ----------------------------------------------------------
2746 def m5_0Imm8Pred : PatLeaf<(i32 imm), [{
2747 int8_t V = N->getSExtValue();
2748 return -32 < V && V <= -1;
2751 def m5_0Imm16Pred : PatLeaf<(i32 imm), [{
2752 int16_t V = N->getSExtValue();
2753 return -32 < V && V <= -1;
2756 def m5_0ImmPred : PatLeaf<(i32 imm), [{
2757 int64_t V = N->getSExtValue();
2758 return -31 <= V && V <= -1;
2761 def IsNPow2_8 : PatLeaf<(i32 imm), [{
2762 uint8_t NV = ~N->getZExtValue();
2763 return isPowerOf2_32(NV);
2766 def IsNPow2_16 : PatLeaf<(i32 imm), [{
2767 uint16_t NV = ~N->getZExtValue();
2768 return isPowerOf2_32(NV);
2771 def Log2_8 : SDNodeXForm<imm, [{
2772 uint8_t V = N->getZExtValue();
2773 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
2776 def Log2_16 : SDNodeXForm<imm, [{
2777 uint16_t V = N->getZExtValue();
2778 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
2781 def LogN2_8 : SDNodeXForm<imm, [{
2782 uint8_t NV = ~N->getZExtValue();
2783 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
2786 def LogN2_16 : SDNodeXForm<imm, [{
2787 uint16_t NV = ~N->getZExtValue();
2788 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
2791 def IdImm : SDNodeXForm<imm, [{ return SDValue(N, 0); }]>;
2793 multiclass Memopxr_base_pat<PatFrag Load, PatFrag Store, SDNode Oper,
2796 def: Pat<(Store (Oper (Load I32:$Rs), I32:$A), I32:$Rs),
2797 (MI I32:$Rs, 0, I32:$A)>;
2799 def: Pat<(Store (Oper (Load AddrFI:$Rs), I32:$A), AddrFI:$Rs),
2800 (MI AddrFI:$Rs, 0, I32:$A)>;
2803 multiclass Memopxr_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2804 SDNode Oper, InstHexagon MI> {
2806 def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), I32:$A),
2807 (add I32:$Rs, ImmPred:$Off)),
2808 (MI I32:$Rs, imm:$Off, I32:$A)>;
2809 def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), I32:$A),
2810 (IsOrAdd I32:$Rs, ImmPred:$Off)),
2811 (MI I32:$Rs, imm:$Off, I32:$A)>;
2813 def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), I32:$A),
2814 (add AddrFI:$Rs, ImmPred:$Off)),
2815 (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
2816 def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), I32:$A),
2817 (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
2818 (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
2821 multiclass Memopxr_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2822 SDNode Oper, InstHexagon MI> {
2823 let Predicates = [UseMEMOPS] in {
2824 defm: Memopxr_base_pat <Load, Store, Oper, MI>;
2825 defm: Memopxr_add_pat <Load, Store, ImmPred, Oper, MI>;
2829 let AddedComplexity = 200 in {
2831 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, add,
2832 /*anyext*/ L4_add_memopb_io>;
2833 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, add,
2834 /*sext*/ L4_add_memopb_io>;
2835 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, add,
2836 /*zext*/ L4_add_memopb_io>;
2837 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, add,
2838 /*anyext*/ L4_add_memoph_io>;
2839 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, add,
2840 /*sext*/ L4_add_memoph_io>;
2841 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, add,
2842 /*zext*/ L4_add_memoph_io>;
2843 defm: Memopxr_pat<load, store, u6_2ImmPred, add, L4_add_memopw_io>;
2846 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, sub,
2847 /*anyext*/ L4_sub_memopb_io>;
2848 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub,
2849 /*sext*/ L4_sub_memopb_io>;
2850 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub,
2851 /*zext*/ L4_sub_memopb_io>;
2852 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, sub,
2853 /*anyext*/ L4_sub_memoph_io>;
2854 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub,
2855 /*sext*/ L4_sub_memoph_io>;
2856 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub,
2857 /*zext*/ L4_sub_memoph_io>;
2858 defm: Memopxr_pat<load, store, u6_2ImmPred, sub, L4_sub_memopw_io>;
2861 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, and,
2862 /*anyext*/ L4_and_memopb_io>;
2863 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, and,
2864 /*sext*/ L4_and_memopb_io>;
2865 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, and,
2866 /*zext*/ L4_and_memopb_io>;
2867 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, and,
2868 /*anyext*/ L4_and_memoph_io>;
2869 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, and,
2870 /*sext*/ L4_and_memoph_io>;
2871 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, and,
2872 /*zext*/ L4_and_memoph_io>;
2873 defm: Memopxr_pat<load, store, u6_2ImmPred, and, L4_and_memopw_io>;
2876 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, or,
2877 /*anyext*/ L4_or_memopb_io>;
2878 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, or,
2879 /*sext*/ L4_or_memopb_io>;
2880 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, or,
2881 /*zext*/ L4_or_memopb_io>;
2882 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, or,
2883 /*anyext*/ L4_or_memoph_io>;
2884 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, or,
2885 /*sext*/ L4_or_memoph_io>;
2886 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, or,
2887 /*zext*/ L4_or_memoph_io>;
2888 defm: Memopxr_pat<load, store, u6_2ImmPred, or, L4_or_memopw_io>;
2892 multiclass Memopxi_base_pat<PatFrag Load, PatFrag Store, SDNode Oper,
2893 PatFrag Arg, SDNodeXForm ArgMod, InstHexagon MI> {
2895 def: Pat<(Store (Oper (Load I32:$Rs), Arg:$A), I32:$Rs),
2896 (MI I32:$Rs, 0, (ArgMod Arg:$A))>;
2898 def: Pat<(Store (Oper (Load AddrFI:$Rs), Arg:$A), AddrFI:$Rs),
2899 (MI AddrFI:$Rs, 0, (ArgMod Arg:$A))>;
2902 multiclass Memopxi_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2903 SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
2906 def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), Arg:$A),
2907 (add I32:$Rs, ImmPred:$Off)),
2908 (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2909 def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), Arg:$A),
2910 (IsOrAdd I32:$Rs, ImmPred:$Off)),
2911 (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2913 def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
2914 (add AddrFI:$Rs, ImmPred:$Off)),
2915 (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2916 def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
2917 (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
2918 (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2921 multiclass Memopxi_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2922 SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
2924 let Predicates = [UseMEMOPS] in {
2925 defm: Memopxi_base_pat <Load, Store, Oper, Arg, ArgMod, MI>;
2926 defm: Memopxi_add_pat <Load, Store, ImmPred, Oper, Arg, ArgMod, MI>;
2930 let AddedComplexity = 220 in {
2932 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
2933 /*anyext*/ IdImm, L4_iadd_memopb_io>;
2934 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
2935 /*sext*/ IdImm, L4_iadd_memopb_io>;
2936 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
2937 /*zext*/ IdImm, L4_iadd_memopb_io>;
2938 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
2939 /*anyext*/ IdImm, L4_iadd_memoph_io>;
2940 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
2941 /*sext*/ IdImm, L4_iadd_memoph_io>;
2942 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
2943 /*zext*/ IdImm, L4_iadd_memoph_io>;
2944 defm: Memopxi_pat<load, store, u6_2ImmPred, add, u5_0ImmPred, IdImm,
2946 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
2947 /*anyext*/ NegImm8, L4_iadd_memopb_io>;
2948 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
2949 /*sext*/ NegImm8, L4_iadd_memopb_io>;
2950 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
2951 /*zext*/ NegImm8, L4_iadd_memopb_io>;
2952 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
2953 /*anyext*/ NegImm16, L4_iadd_memoph_io>;
2954 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
2955 /*sext*/ NegImm16, L4_iadd_memoph_io>;
2956 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
2957 /*zext*/ NegImm16, L4_iadd_memoph_io>;
2958 defm: Memopxi_pat<load, store, u6_2ImmPred, sub, m5_0ImmPred, NegImm32,
2962 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
2963 /*anyext*/ IdImm, L4_isub_memopb_io>;
2964 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
2965 /*sext*/ IdImm, L4_isub_memopb_io>;
2966 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
2967 /*zext*/ IdImm, L4_isub_memopb_io>;
2968 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
2969 /*anyext*/ IdImm, L4_isub_memoph_io>;
2970 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
2971 /*sext*/ IdImm, L4_isub_memoph_io>;
2972 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
2973 /*zext*/ IdImm, L4_isub_memoph_io>;
2974 defm: Memopxi_pat<load, store, u6_2ImmPred, sub, u5_0ImmPred, IdImm,
2976 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
2977 /*anyext*/ NegImm8, L4_isub_memopb_io>;
2978 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
2979 /*sext*/ NegImm8, L4_isub_memopb_io>;
2980 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
2981 /*zext*/ NegImm8, L4_isub_memopb_io>;
2982 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
2983 /*anyext*/ NegImm16, L4_isub_memoph_io>;
2984 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
2985 /*sext*/ NegImm16, L4_isub_memoph_io>;
2986 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
2987 /*zext*/ NegImm16, L4_isub_memoph_io>;
2988 defm: Memopxi_pat<load, store, u6_2ImmPred, add, m5_0ImmPred, NegImm32,
2992 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
2993 /*anyext*/ LogN2_8, L4_iand_memopb_io>;
2994 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
2995 /*sext*/ LogN2_8, L4_iand_memopb_io>;
2996 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
2997 /*zext*/ LogN2_8, L4_iand_memopb_io>;
2998 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
2999 /*anyext*/ LogN2_16, L4_iand_memoph_io>;
3000 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
3001 /*sext*/ LogN2_16, L4_iand_memoph_io>;
3002 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
3003 /*zext*/ LogN2_16, L4_iand_memoph_io>;
3004 defm: Memopxi_pat<load, store, u6_2ImmPred, and, IsNPow2_32,
3005 LogN2_32, L4_iand_memopw_io>;
3008 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3009 /*anyext*/ Log2_8, L4_ior_memopb_io>;
3010 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3011 /*sext*/ Log2_8, L4_ior_memopb_io>;
3012 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3013 /*zext*/ Log2_8, L4_ior_memopb_io>;
3014 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3015 /*anyext*/ Log2_16, L4_ior_memoph_io>;
3016 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3017 /*sext*/ Log2_16, L4_ior_memoph_io>;
3018 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3019 /*zext*/ Log2_16, L4_ior_memoph_io>;
3020 defm: Memopxi_pat<load, store, u6_2ImmPred, or, IsPow2_32,
3021 Log2_32, L4_ior_memopw_io>;
3025 // --(15) Call -----------------------------------------------------------
3028 // Pseudo instructions.
3029 def SDT_SPCallSeqStart
3030 : SDCallSeqStart<[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3031 def SDT_SPCallSeqEnd
3032 : SDCallSeqEnd<[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3034 def callseq_start: SDNode<"ISD::CALLSEQ_START", SDT_SPCallSeqStart,
3035 [SDNPHasChain, SDNPOutGlue]>;
3036 def callseq_end: SDNode<"ISD::CALLSEQ_END", SDT_SPCallSeqEnd,
3037 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
3039 def SDT_SPCall: SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
3041 def HexagonTCRet: SDNode<"HexagonISD::TC_RETURN", SDT_SPCall,
3042 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
3043 def callv3: SDNode<"HexagonISD::CALL", SDT_SPCall,
3044 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
3045 def callv3nr: SDNode<"HexagonISD::CALLnr", SDT_SPCall,
3046 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
3048 def: Pat<(callseq_start timm:$amt, timm:$amt2),
3049 (ADJCALLSTACKDOWN imm:$amt, imm:$amt2)>;
3050 def: Pat<(callseq_end timm:$amt1, timm:$amt2),
3051 (ADJCALLSTACKUP imm:$amt1, imm:$amt2)>;
3053 def: Pat<(HexagonTCRet tglobaladdr:$dst), (PS_tailcall_i tglobaladdr:$dst)>;
3054 def: Pat<(HexagonTCRet texternalsym:$dst), (PS_tailcall_i texternalsym:$dst)>;
3055 def: Pat<(HexagonTCRet I32:$dst), (PS_tailcall_r I32:$dst)>;
3057 def: Pat<(callv3 I32:$dst), (J2_callr I32:$dst)>;
3058 def: Pat<(callv3 tglobaladdr:$dst), (J2_call tglobaladdr:$dst)>;
3059 def: Pat<(callv3 texternalsym:$dst), (J2_call texternalsym:$dst)>;
3060 def: Pat<(callv3 tglobaltlsaddr:$dst), (J2_call tglobaltlsaddr:$dst)>;
3062 def: Pat<(callv3nr I32:$dst), (PS_callr_nr I32:$dst)>;
3063 def: Pat<(callv3nr tglobaladdr:$dst), (PS_call_nr tglobaladdr:$dst)>;
3064 def: Pat<(callv3nr texternalsym:$dst), (PS_call_nr texternalsym:$dst)>;
3066 def retflag : SDNode<"HexagonISD::RET_FLAG", SDTNone,
3067 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
3068 def eh_return: SDNode<"HexagonISD::EH_RETURN", SDTNone, [SDNPHasChain]>;
3070 def: Pat<(retflag), (PS_jmpret (i32 R31))>;
3071 def: Pat<(eh_return), (EH_RETURN_JMPR (i32 R31))>;
3074 // --(16) Branch ---------------------------------------------------------
3077 def: Pat<(br bb:$dst), (J2_jump b30_2Imm:$dst)>;
3078 def: Pat<(brind I32:$dst), (J2_jumpr I32:$dst)>;
3080 def: Pat<(brcond I1:$Pu, bb:$dst),
3081 (J2_jumpt I1:$Pu, bb:$dst)>;
3082 def: Pat<(brcond (not I1:$Pu), bb:$dst),
3083 (J2_jumpf I1:$Pu, bb:$dst)>;
3084 def: Pat<(brcond (i1 (setne I1:$Pu, -1)), bb:$dst),
3085 (J2_jumpf I1:$Pu, bb:$dst)>;
3086 def: Pat<(brcond (i1 (seteq I1:$Pu, 0)), bb:$dst),
3087 (J2_jumpf I1:$Pu, bb:$dst)>;
3088 def: Pat<(brcond (i1 (setne I1:$Pu, 0)), bb:$dst),
3089 (J2_jumpt I1:$Pu, bb:$dst)>;
3092 // --(17) Misc -----------------------------------------------------------
3095 // Generate code of the form 'C2_muxii(cmpbgtui(Rdd, C-1),0,1)'
3096 // for C code of the form r = (c>='0' && c<='9') ? 1 : 0.
3097 // The isdigit transformation relies on two 'clever' aspects:
3098 // 1) The data type is unsigned which allows us to eliminate a zero test after
3099 // biasing the expression by 48. We are depending on the representation of
3100 // the unsigned types, and semantics.
3101 // 2) The front end has converted <= 9 into < 10 on entry to LLVM.
3104 // retval = (c >= '0' && c <= '9') ? 1 : 0;
3105 // The code is transformed upstream of llvm into
3106 // retval = (c-48) < 10 ? 1 : 0;
3108 def u7_0PosImmPred : ImmLeaf<i32, [{
3109 // True if the immediate fits in an 7-bit unsigned field and is positive.
3110 return Imm > 0 && isUInt<7>(Imm);
3113 let AddedComplexity = 139 in
3114 def: Pat<(i32 (zext (i1 (setult (and I32:$Rs, 255), u7_0PosImmPred:$u7)))),
3115 (C2_muxii (A4_cmpbgtui IntRegs:$Rs, (UDEC1 imm:$u7)), 0, 1)>;
3117 let AddedComplexity = 100 in
3118 def: Pat<(or (or (shl (HexagonINSERT (i32 (zextloadi8 (add I32:$b, 2))),
3119 (i32 (extloadi8 (add I32:$b, 3))),
3122 (shl (i32 (zextloadi8 (add I32:$b, 1))), (i32 8))),
3123 (zextloadi8 I32:$b)),
3124 (A2_swiz (L2_loadri_io I32:$b, 0))>;
3127 // We need custom lowering of ISD::PREFETCH into HexagonISD::DCFETCH
3128 // because the SDNode ISD::PREFETCH has properties MayLoad and MayStore.
3129 // We don't really want either one here.
3130 def SDTHexagonDCFETCH: SDTypeProfile<0, 2, [SDTCisPtrTy<0>,SDTCisInt<1>]>;
3131 def HexagonDCFETCH: SDNode<"HexagonISD::DCFETCH", SDTHexagonDCFETCH,
3134 def: Pat<(HexagonDCFETCH IntRegs:$Rs, u11_3ImmPred:$u11_3),
3135 (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
3136 def: Pat<(HexagonDCFETCH (i32 (add IntRegs:$Rs, u11_3ImmPred:$u11_3)), (i32 0)),
3137 (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
3139 def SDTHexagonALLOCA
3140 : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3142 : SDNode<"HexagonISD::ALLOCA", SDTHexagonALLOCA, [SDNPHasChain]>;
3144 def: Pat<(HexagonALLOCA I32:$Rs, (i32 imm:$A)),
3145 (PS_alloca IntRegs:$Rs, imm:$A)>;
3147 def HexagonBARRIER: SDNode<"HexagonISD::BARRIER", SDTNone, [SDNPHasChain]>;
3148 def: Pat<(HexagonBARRIER), (Y2_barrier)>;
3150 def: Pat<(trap), (PS_crash)>;
3152 // Read cycle counter.
3153 def SDTInt64Leaf: SDTypeProfile<1, 0, [SDTCisVT<0, i64>]>;
3154 def HexagonREADCYCLE: SDNode<"HexagonISD::READCYCLE", SDTInt64Leaf,
3157 def: Pat<(HexagonREADCYCLE), (A4_tfrcpp UPCYCLE)>;
3159 // The declared return value of the store-locked intrinsics is i32, but
3160 // the instructions actually define i1. To avoid register copies from
3161 // IntRegs to PredRegs and back, fold the entire pattern checking the
3162 // result against true/false.
3163 let AddedComplexity = 100 in {
3164 def: Pat<(i1 (setne (int_hexagon_S2_storew_locked I32:$Rs, I32:$Rt), 0)),
3165 (S2_storew_locked I32:$Rs, I32:$Rt)>;
3166 def: Pat<(i1 (seteq (int_hexagon_S2_storew_locked I32:$Rs, I32:$Rt), 0)),
3167 (C2_not (S2_storew_locked I32:$Rs, I32:$Rt))>;
3168 def: Pat<(i1 (setne (int_hexagon_S4_stored_locked I32:$Rs, I64:$Rt), 0)),
3169 (S4_stored_locked I32:$Rs, I64:$Rt)>;
3170 def: Pat<(i1 (seteq (int_hexagon_S4_stored_locked I32:$Rs, I64:$Rt), 0)),
3171 (C2_not (S4_stored_locked I32:$Rs, I64:$Rt))>;