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:$R)>;
103 SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisVec<1>, SDTCisSameAs<1,2>,
106 def HexagonVALIGN: SDNode<"HexagonISD::VALIGN", SDTVecVecIntOp>;
107 def HexagonVALIGNADDR: SDNode<"HexagonISD::VALIGNADDR", SDTIntUnaryOp>;
109 def valign: PatFrag<(ops node:$Vt, node:$Vs, node:$Ru),
110 (HexagonVALIGN node:$Vt, node:$Vs, node:$Ru)>;
111 def valignaddr: PatFrag<(ops node:$Addr), (HexagonVALIGNADDR node:$Addr)>;
113 // Pattern fragments to extract the low and high subregisters from a
115 def LoReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG (i64 $Rs), isub_lo)>;
116 def HiReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG (i64 $Rs), isub_hi)>;
118 def IsOrAdd: PatFrag<(ops node:$A, node:$B), (or node:$A, node:$B), [{
119 return isOrEquivalentToAdd(N);
122 def IsPow2_32: PatLeaf<(i32 imm), [{
123 uint32_t V = N->getZExtValue();
124 return isPowerOf2_32(V);
127 def IsPow2_64: PatLeaf<(i64 imm), [{
128 uint64_t V = N->getZExtValue();
129 return isPowerOf2_64(V);
132 def IsNPow2_32: PatLeaf<(i32 imm), [{
133 uint32_t NV = ~N->getZExtValue();
134 return isPowerOf2_32(NV);
137 def IsPow2_64L: PatLeaf<(i64 imm), [{
138 uint64_t V = N->getZExtValue();
139 return isPowerOf2_64(V) && Log2_64(V) < 32;
142 def IsPow2_64H: PatLeaf<(i64 imm), [{
143 uint64_t V = N->getZExtValue();
144 return isPowerOf2_64(V) && Log2_64(V) >= 32;
147 def IsNPow2_64L: PatLeaf<(i64 imm), [{
148 uint64_t NV = ~N->getZExtValue();
149 return isPowerOf2_64(NV) && Log2_64(NV) < 32;
152 def IsNPow2_64H: PatLeaf<(i64 imm), [{
153 uint64_t NV = ~N->getZExtValue();
154 return isPowerOf2_64(NV) && Log2_64(NV) >= 32;
157 class IsUGT<int Width, int Arg>: PatLeaf<(i32 imm),
158 "uint64_t V = N->getZExtValue();" #
159 "return isUInt<" # Width # ">(V) && V > " # Arg # ";"
162 def SDEC1: SDNodeXForm<imm, [{
163 int32_t V = N->getSExtValue();
164 return CurDAG->getTargetConstant(V-1, SDLoc(N), MVT::i32);
167 def UDEC1: SDNodeXForm<imm, [{
168 uint32_t V = N->getZExtValue();
170 return CurDAG->getTargetConstant(V-1, SDLoc(N), MVT::i32);
173 def UDEC32: SDNodeXForm<imm, [{
174 uint32_t V = N->getZExtValue();
176 return CurDAG->getTargetConstant(V-32, SDLoc(N), MVT::i32);
179 class Subi<int From>: SDNodeXForm<imm,
180 "int32_t V = " # From # " - N->getSExtValue();" #
181 "return CurDAG->getTargetConstant(V, SDLoc(N), MVT::i32);"
184 def Log2_32: SDNodeXForm<imm, [{
185 uint32_t V = N->getZExtValue();
186 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
189 def Log2_64: SDNodeXForm<imm, [{
190 uint64_t V = N->getZExtValue();
191 return CurDAG->getTargetConstant(Log2_64(V), SDLoc(N), MVT::i32);
194 def LogN2_32: SDNodeXForm<imm, [{
195 uint32_t NV = ~N->getZExtValue();
196 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
199 def LogN2_64: SDNodeXForm<imm, [{
200 uint64_t NV = ~N->getZExtValue();
201 return CurDAG->getTargetConstant(Log2_64(NV), SDLoc(N), MVT::i32);
204 def NegImm8: SDNodeXForm<imm, [{
205 int8_t NV = -N->getSExtValue();
206 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
209 def NegImm16: SDNodeXForm<imm, [{
210 int16_t NV = -N->getSExtValue();
211 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
214 def NegImm32: SDNodeXForm<imm, [{
215 int32_t NV = -N->getSExtValue();
216 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
220 // Helpers for type promotions/contractions.
221 def I1toI32: OutPatFrag<(ops node:$Rs), (C2_muxii (i1 $Rs), 1, 0)>;
222 def I32toI1: OutPatFrag<(ops node:$Rs), (i1 (C2_cmpgtui (i32 $Rs), (i32 0)))>;
223 def ToZext64: OutPatFrag<(ops node:$Rs), (i64 (A4_combineir 0, (i32 $Rs)))>;
224 def ToSext64: OutPatFrag<(ops node:$Rs), (i64 (A2_sxtw (i32 $Rs)))>;
225 def ToAext64: OutPatFrag<(ops node:$Rs),
226 (REG_SEQUENCE DoubleRegs, (i32 (IMPLICIT_DEF)), isub_hi, (i32 $Rs), isub_lo)>;
228 def Combinew: OutPatFrag<(ops node:$Rs, node:$Rt),
229 (REG_SEQUENCE DoubleRegs, $Rs, isub_hi, $Rt, isub_lo)>;
231 def addrga: PatLeaf<(i32 AddrGA:$Addr)>;
232 def addrgp: PatLeaf<(i32 AddrGP:$Addr)>;
233 def anyimm: PatLeaf<(i32 AnyImm:$Imm)>;
234 def anyint: PatLeaf<(i32 AnyInt:$Imm)>;
236 // Global address or an aligned constant.
237 def anyimm0: PatLeaf<(i32 AnyImm0:$Addr)>;
238 def anyimm1: PatLeaf<(i32 AnyImm1:$Addr)>;
239 def anyimm2: PatLeaf<(i32 AnyImm2:$Addr)>;
240 def anyimm3: PatLeaf<(i32 AnyImm3:$Addr)>;
242 def f32ImmPred : PatLeaf<(f32 fpimm:$F)>;
243 def f64ImmPred : PatLeaf<(f64 fpimm:$F)>;
245 // This complex pattern is really only to detect various forms of
246 // sign-extension i32->i64. The selected value will be of type i64
247 // whose low word is the value being extended. The high word is
249 def Usxtw: ComplexPattern<i64, 1, "DetectUseSxtw", [], []>;
251 def Aext64: PatFrag<(ops node:$Rs), (i64 (anyext node:$Rs))>;
252 def Zext64: PatFrag<(ops node:$Rs), (i64 (zext node:$Rs))>;
253 def Sext64: PatLeaf<(i64 Usxtw:$Rs)>;
255 def azext: PatFrags<(ops node:$Rs), [(zext node:$Rs), (anyext node:$Rs)]>;
256 def asext: PatFrags<(ops node:$Rs), [(sext node:$Rs), (anyext node:$Rs)]>;
258 def: Pat<(IsOrAdd (i32 AddrFI:$Rs), s32_0ImmPred:$off),
259 (PS_fi (i32 AddrFI:$Rs), imm:$off)>;
262 // Converters from unary/binary SDNode to PatFrag.
263 class pf1<SDNode Op> : PatFrag<(ops node:$a), (Op node:$a)>;
264 class pf2<SDNode Op> : PatFrag<(ops node:$a, node:$b), (Op node:$a, node:$b)>;
266 class Not2<PatFrag P>
267 : PatFrag<(ops node:$A, node:$B), (P node:$A, (not node:$B))>;
269 // If there is a constant operand that feeds the and/or instruction,
270 // do not generate the compound instructions.
271 // It is not always profitable, as some times we end up with a transfer.
272 // Check the below example.
273 // ra = #65820; rb = lsr(rb, #8); rc ^= and (rb, ra)
274 // Instead this is preferable.
275 // ra = and (#65820, lsr(ra, #8)); rb = xor(rb, ra)
276 class Su_ni1<PatFrag Op>
277 : PatFrag<Op.Operands, !head(Op.Fragments), [{
279 // Check if Op1 is an immediate operand.
280 SDValue Op1 = N->getOperand(1);
281 return !dyn_cast<ConstantSDNode>(Op1);
284 Op.OperandTransform>;
287 : PatFrag<Op.Operands, !head(Op.Fragments), [{ return hasOneUse(N); }],
288 Op.OperandTransform>;
290 // Main selection macros.
292 class OpR_R_pat<InstHexagon MI, PatFrag Op, ValueType ResVT, PatFrag RegPred>
293 : Pat<(ResVT (Op RegPred:$Rs)), (MI RegPred:$Rs)>;
295 class OpR_RI_pat<InstHexagon MI, PatFrag Op, ValueType ResType,
296 PatFrag RegPred, PatFrag ImmPred>
297 : Pat<(ResType (Op RegPred:$Rs, ImmPred:$I)),
298 (MI RegPred:$Rs, imm:$I)>;
300 class OpR_RR_pat<InstHexagon MI, PatFrag Op, ValueType ResType,
301 PatFrag RsPred, PatFrag RtPred = RsPred>
302 : Pat<(ResType (Op RsPred:$Rs, RtPred:$Rt)),
303 (MI RsPred:$Rs, RtPred:$Rt)>;
305 class AccRRI_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op,
306 PatFrag RegPred, PatFrag ImmPred>
307 : Pat<(AccOp RegPred:$Rx, (Op RegPred:$Rs, ImmPred:$I)),
308 (MI RegPred:$Rx, RegPred:$Rs, imm:$I)>;
310 class AccRRR_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op,
311 PatFrag RxPred, PatFrag RsPred, PatFrag RtPred>
312 : Pat<(AccOp RxPred:$Rx, (Op RsPred:$Rs, RtPred:$Rt)),
313 (MI RxPred:$Rx, RsPred:$Rs, RtPred:$Rt)>;
315 multiclass SelMinMax_pats<PatFrag CmpOp, PatFrag Val,
316 InstHexagon InstA, InstHexagon InstB> {
317 def: Pat<(select (i1 (CmpOp Val:$A, Val:$B)), Val:$A, Val:$B),
318 (InstA Val:$A, Val:$B)>;
319 def: Pat<(select (i1 (CmpOp Val:$A, Val:$B)), Val:$B, Val:$A),
320 (InstB Val:$A, Val:$B)>;
324 // Frags for commonly used SDNodes.
325 def Add: pf2<add>; def And: pf2<and>; def Sra: pf2<sra>;
326 def Sub: pf2<sub>; def Or: pf2<or>; def Srl: pf2<srl>;
327 def Mul: pf2<mul>; def Xor: pf2<xor>; def Shl: pf2<shl>;
331 // --(1) Immediate -------------------------------------------------------
334 def SDTHexagonCONST32
335 : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisPtrTy<0>]>;
337 def HexagonJT: SDNode<"HexagonISD::JT", SDTIntUnaryOp>;
338 def HexagonCP: SDNode<"HexagonISD::CP", SDTIntUnaryOp>;
339 def HexagonCONST32: SDNode<"HexagonISD::CONST32", SDTHexagonCONST32>;
340 def HexagonCONST32_GP: SDNode<"HexagonISD::CONST32_GP", SDTHexagonCONST32>;
342 def TruncI64ToI32: SDNodeXForm<imm, [{
343 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i32);
346 def: Pat<(s32_0ImmPred:$s16), (A2_tfrsi imm:$s16)>;
347 def: Pat<(s8_0Imm64Pred:$s8), (A2_tfrpi (TruncI64ToI32 $s8))>;
349 def: Pat<(HexagonCONST32 tglobaltlsaddr:$A), (A2_tfrsi imm:$A)>;
350 def: Pat<(HexagonCONST32 bbl:$A), (A2_tfrsi imm:$A)>;
351 def: Pat<(HexagonCONST32 tglobaladdr:$A), (A2_tfrsi imm:$A)>;
352 def: Pat<(HexagonCONST32_GP tblockaddress:$A), (A2_tfrsi imm:$A)>;
353 def: Pat<(HexagonCONST32_GP tglobaladdr:$A), (A2_tfrsi imm:$A)>;
354 def: Pat<(HexagonJT tjumptable:$A), (A2_tfrsi imm:$A)>;
355 def: Pat<(HexagonCP tconstpool:$A), (A2_tfrsi imm:$A)>;
356 // The HVX load patterns also match CP directly. Make sure that if
357 // the selection of this opcode changes, it's updated in all places.
359 def: Pat<(i1 0), (PS_false)>;
360 def: Pat<(i1 1), (PS_true)>;
361 def: Pat<(i64 imm:$v), (CONST64 imm:$v)>;
363 def ftoi : SDNodeXForm<fpimm, [{
364 APInt I = N->getValueAPF().bitcastToAPInt();
365 return CurDAG->getTargetConstant(I.getZExtValue(), SDLoc(N),
366 MVT::getIntegerVT(I.getBitWidth()));
369 def: Pat<(f32ImmPred:$f), (A2_tfrsi (ftoi $f))>;
370 def: Pat<(f64ImmPred:$f), (CONST64 (ftoi $f))>;
372 def ToI32: OutPatFrag<(ops node:$V), (A2_tfrsi $V)>;
374 // --(2) Type cast -------------------------------------------------------
377 def: OpR_R_pat<F2_conv_sf2df, pf1<fpextend>, f64, F32>;
378 def: OpR_R_pat<F2_conv_df2sf, pf1<fpround>, f32, F64>;
380 def: OpR_R_pat<F2_conv_w2sf, pf1<sint_to_fp>, f32, I32>;
381 def: OpR_R_pat<F2_conv_d2sf, pf1<sint_to_fp>, f32, I64>;
382 def: OpR_R_pat<F2_conv_w2df, pf1<sint_to_fp>, f64, I32>;
383 def: OpR_R_pat<F2_conv_d2df, pf1<sint_to_fp>, f64, I64>;
385 def: OpR_R_pat<F2_conv_uw2sf, pf1<uint_to_fp>, f32, I32>;
386 def: OpR_R_pat<F2_conv_ud2sf, pf1<uint_to_fp>, f32, I64>;
387 def: OpR_R_pat<F2_conv_uw2df, pf1<uint_to_fp>, f64, I32>;
388 def: OpR_R_pat<F2_conv_ud2df, pf1<uint_to_fp>, f64, I64>;
390 def: OpR_R_pat<F2_conv_sf2w_chop, pf1<fp_to_sint>, i32, F32>;
391 def: OpR_R_pat<F2_conv_df2w_chop, pf1<fp_to_sint>, i32, F64>;
392 def: OpR_R_pat<F2_conv_sf2d_chop, pf1<fp_to_sint>, i64, F32>;
393 def: OpR_R_pat<F2_conv_df2d_chop, pf1<fp_to_sint>, i64, F64>;
395 def: OpR_R_pat<F2_conv_sf2uw_chop, pf1<fp_to_uint>, i32, F32>;
396 def: OpR_R_pat<F2_conv_df2uw_chop, pf1<fp_to_uint>, i32, F64>;
397 def: OpR_R_pat<F2_conv_sf2ud_chop, pf1<fp_to_uint>, i64, F32>;
398 def: OpR_R_pat<F2_conv_df2ud_chop, pf1<fp_to_uint>, i64, F64>;
400 // Bitcast is different than [fp|sint|uint]_to_[sint|uint|fp].
401 def: Pat<(i32 (bitconvert F32:$v)), (I32:$v)>;
402 def: Pat<(f32 (bitconvert I32:$v)), (F32:$v)>;
403 def: Pat<(i64 (bitconvert F64:$v)), (I64:$v)>;
404 def: Pat<(f64 (bitconvert I64:$v)), (F64:$v)>;
406 multiclass Cast_pat<ValueType Ta, ValueType Tb, RegisterClass RC> {
407 def: Pat<(Tb (bitconvert (Ta RC:$Rs))), (Tb RC:$Rs)>;
408 def: Pat<(Ta (bitconvert (Tb RC:$Rs))), (Ta RC:$Rs)>;
411 // Bit convert vector types to integers.
412 defm: Cast_pat<v4i8, i32, IntRegs>;
413 defm: Cast_pat<v2i16, i32, IntRegs>;
414 defm: Cast_pat<v8i8, i64, DoubleRegs>;
415 defm: Cast_pat<v4i16, i64, DoubleRegs>;
416 defm: Cast_pat<v2i32, i64, DoubleRegs>;
419 // --(3) Extend/truncate -------------------------------------------------
422 def: Pat<(sext_inreg I32:$Rs, i8), (A2_sxtb I32:$Rs)>;
423 def: Pat<(sext_inreg I32:$Rs, i16), (A2_sxth I32:$Rs)>;
424 def: Pat<(sext_inreg I64:$Rs, i32), (A2_sxtw (LoReg $Rs))>;
425 def: Pat<(sext_inreg I64:$Rs, i16), (A2_sxtw (A2_sxth (LoReg $Rs)))>;
426 def: Pat<(sext_inreg I64:$Rs, i8), (A2_sxtw (A2_sxtb (LoReg $Rs)))>;
428 def: Pat<(i64 (sext I32:$Rs)), (A2_sxtw I32:$Rs)>;
429 def: Pat<(Zext64 I32:$Rs), (ToZext64 $Rs)>;
430 def: Pat<(Aext64 I32:$Rs), (ToZext64 $Rs)>;
432 def: Pat<(i32 (trunc I64:$Rs)), (LoReg $Rs)>;
433 def: Pat<(i1 (trunc I32:$Rs)), (S2_tstbit_i I32:$Rs, 0)>;
434 def: Pat<(i1 (trunc I64:$Rs)), (S2_tstbit_i (LoReg $Rs), 0)>;
436 let AddedComplexity = 20 in {
437 def: Pat<(and I32:$Rs, 255), (A2_zxtb I32:$Rs)>;
438 def: Pat<(and I32:$Rs, 65535), (A2_zxth I32:$Rs)>;
441 // Extensions from i1 or vectors of i1.
442 def: Pat<(i32 (azext I1:$Pu)), (C2_muxii I1:$Pu, 1, 0)>;
443 def: Pat<(i64 (azext I1:$Pu)), (ToZext64 (C2_muxii I1:$Pu, 1, 0))>;
444 def: Pat<(i32 (sext I1:$Pu)), (C2_muxii I1:$Pu, -1, 0)>;
445 def: Pat<(i64 (sext I1:$Pu)), (Combinew (C2_muxii PredRegs:$Pu, -1, 0),
446 (C2_muxii PredRegs:$Pu, -1, 0))>;
448 def: Pat<(v2i16 (sext V2I1:$Pu)), (S2_vtrunehb (C2_mask V2I1:$Pu))>;
449 def: Pat<(v2i32 (sext V2I1:$Pu)), (C2_mask V2I1:$Pu)>;
450 def: Pat<(v4i8 (sext V4I1:$Pu)), (S2_vtrunehb (C2_mask V4I1:$Pu))>;
451 def: Pat<(v4i16 (sext V4I1:$Pu)), (C2_mask V4I1:$Pu)>;
452 def: Pat<(v8i8 (sext V8I1:$Pu)), (C2_mask V8I1:$Pu)>;
454 def Vsplatpi: OutPatFrag<(ops node:$V),
455 (Combinew (A2_tfrsi $V), (A2_tfrsi $V))>;
457 def: Pat<(v2i16 (azext V2I1:$Pu)),
458 (A2_andir (LoReg (C2_mask V2I1:$Pu)), (i32 0x00010001))>;
459 def: Pat<(v2i32 (azext V2I1:$Pu)),
460 (A2_andp (C2_mask V2I1:$Pu), (A2_combineii (i32 1), (i32 1)))>;
461 def: Pat<(v4i8 (azext V4I1:$Pu)),
462 (A2_andir (LoReg (C2_mask V4I1:$Pu)), (i32 0x01010101))>;
463 def: Pat<(v4i16 (azext V4I1:$Pu)),
464 (A2_andp (C2_mask V4I1:$Pu), (Vsplatpi (i32 0x00010001)))>;
465 def: Pat<(v8i8 (azext V8I1:$Pu)),
466 (A2_andp (C2_mask V8I1:$Pu), (Vsplatpi (i32 0x01010101)))>;
468 def: Pat<(v4i16 (azext V4I8:$Rs)), (S2_vzxtbh V4I8:$Rs)>;
469 def: Pat<(v2i32 (azext V2I16:$Rs)), (S2_vzxthw V2I16:$Rs)>;
470 def: Pat<(v4i16 (sext V4I8:$Rs)), (S2_vsxtbh V4I8:$Rs)>;
471 def: Pat<(v2i32 (sext V2I16:$Rs)), (S2_vsxthw V2I16:$Rs)>;
473 def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i8)),
474 (Combinew (A2_sxtb (HiReg $Rs)), (A2_sxtb (LoReg $Rs)))>;
476 def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i16)),
477 (Combinew (A2_sxth (HiReg $Rs)), (A2_sxth (LoReg $Rs)))>;
479 // Truncate: from vector B copy all 'E'ven 'B'yte elements:
480 // A[0] = B[0]; A[1] = B[2]; A[2] = B[4]; A[3] = B[6];
481 def: Pat<(v4i8 (trunc V4I16:$Rs)),
482 (S2_vtrunehb V4I16:$Rs)>;
484 // Truncate: from vector B copy all 'O'dd 'B'yte elements:
485 // A[0] = B[1]; A[1] = B[3]; A[2] = B[5]; A[3] = B[7];
488 // Truncate: from vectors B and C copy all 'E'ven 'H'alf-word elements:
489 // A[0] = B[0]; A[1] = B[2]; A[2] = C[0]; A[3] = C[2];
492 def: Pat<(v2i16 (trunc V2I32:$Rs)),
493 (A2_combine_ll (HiReg $Rs), (LoReg $Rs))>;
496 // --(4) Logical ---------------------------------------------------------
499 def: Pat<(not I1:$Ps), (C2_not I1:$Ps)>;
500 def: Pat<(not V8I1:$Ps), (C2_not V8I1:$Ps)>;
501 def: Pat<(add I1:$Ps, -1), (C2_not I1:$Ps)>;
503 multiclass BoolOpR_RR_pat<InstHexagon MI, PatFrag Op> {
504 def: OpR_RR_pat<MI, Op, i1, I1>;
505 def: OpR_RR_pat<MI, Op, v2i1, V2I1>;
506 def: OpR_RR_pat<MI, Op, v4i1, V4I1>;
507 def: OpR_RR_pat<MI, Op, v8i1, V8I1>;
510 multiclass BoolAccRRR_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op> {
511 def: AccRRR_pat<MI, AccOp, Op, I1, I1, I1>;
512 def: AccRRR_pat<MI, AccOp, Op, V2I1, V2I1, V2I1>;
513 def: AccRRR_pat<MI, AccOp, Op, V4I1, V4I1, V4I1>;
514 def: AccRRR_pat<MI, AccOp, Op, V8I1, V8I1, V8I1>;
517 defm: BoolOpR_RR_pat<C2_and, And>;
518 defm: BoolOpR_RR_pat<C2_or, Or>;
519 defm: BoolOpR_RR_pat<C2_xor, Xor>;
520 defm: BoolOpR_RR_pat<C2_andn, Not2<And>>;
521 defm: BoolOpR_RR_pat<C2_orn, Not2<Or>>;
523 // op(Ps, op(Pt, Pu))
524 defm: BoolAccRRR_pat<C4_and_and, And, Su<And>>;
525 defm: BoolAccRRR_pat<C4_and_or, And, Su<Or>>;
526 defm: BoolAccRRR_pat<C4_or_and, Or, Su<And>>;
527 defm: BoolAccRRR_pat<C4_or_or, Or, Su<Or>>;
529 // op(Ps, op(Pt, ~Pu))
530 defm: BoolAccRRR_pat<C4_and_andn, And, Su<Not2<And>>>;
531 defm: BoolAccRRR_pat<C4_and_orn, And, Su<Not2<Or>>>;
532 defm: BoolAccRRR_pat<C4_or_andn, Or, Su<Not2<And>>>;
533 defm: BoolAccRRR_pat<C4_or_orn, Or, Su<Not2<Or>>>;
536 // --(5) Compare ---------------------------------------------------------
539 // Avoid negated comparisons, i.e. those of form "Pd = !cmp(...)".
540 // These cannot form compounds (e.g. J4_cmpeqi_tp0_jump_nt).
542 def: OpR_RI_pat<C2_cmpeqi, seteq, i1, I32, anyimm>;
543 def: OpR_RI_pat<C2_cmpgti, setgt, i1, I32, anyimm>;
544 def: OpR_RI_pat<C2_cmpgtui, setugt, i1, I32, anyimm>;
546 def: Pat<(i1 (setge I32:$Rs, s32_0ImmPred:$s10)),
547 (C2_cmpgti I32:$Rs, (SDEC1 imm:$s10))>;
548 def: Pat<(i1 (setuge I32:$Rs, u32_0ImmPred:$u9)),
549 (C2_cmpgtui I32:$Rs, (UDEC1 imm:$u9))>;
551 def: Pat<(i1 (setlt I32:$Rs, s32_0ImmPred:$s10)),
552 (C2_not (C2_cmpgti I32:$Rs, (SDEC1 imm:$s10)))>;
553 def: Pat<(i1 (setult I32:$Rs, u32_0ImmPred:$u9)),
554 (C2_not (C2_cmpgtui I32:$Rs, (UDEC1 imm:$u9)))>;
556 // Patfrag to convert the usual comparison patfrags (e.g. setlt) to ones
557 // that reverse the order of the operands.
558 class RevCmp<PatFrag F>
559 : PatFrag<(ops node:$rhs, node:$lhs), !head(F.Fragments), F.PredicateCode,
562 def: OpR_RR_pat<C2_cmpeq, seteq, i1, I32>;
563 def: OpR_RR_pat<C2_cmpgt, setgt, i1, I32>;
564 def: OpR_RR_pat<C2_cmpgtu, setugt, i1, I32>;
565 def: OpR_RR_pat<C2_cmpgt, RevCmp<setlt>, i1, I32>;
566 def: OpR_RR_pat<C2_cmpgtu, RevCmp<setult>, i1, I32>;
567 def: OpR_RR_pat<C2_cmpeqp, seteq, i1, I64>;
568 def: OpR_RR_pat<C2_cmpgtp, setgt, i1, I64>;
569 def: OpR_RR_pat<C2_cmpgtup, setugt, i1, I64>;
570 def: OpR_RR_pat<C2_cmpgtp, RevCmp<setlt>, i1, I64>;
571 def: OpR_RR_pat<C2_cmpgtup, RevCmp<setult>, i1, I64>;
572 def: OpR_RR_pat<A2_vcmpbeq, seteq, i1, V8I8>;
573 def: OpR_RR_pat<A2_vcmpbeq, seteq, v8i1, V8I8>;
574 def: OpR_RR_pat<A4_vcmpbgt, RevCmp<setlt>, i1, V8I8>;
575 def: OpR_RR_pat<A4_vcmpbgt, RevCmp<setlt>, v8i1, V8I8>;
576 def: OpR_RR_pat<A4_vcmpbgt, setgt, i1, V8I8>;
577 def: OpR_RR_pat<A4_vcmpbgt, setgt, v8i1, V8I8>;
578 def: OpR_RR_pat<A2_vcmpbgtu, RevCmp<setult>, i1, V8I8>;
579 def: OpR_RR_pat<A2_vcmpbgtu, RevCmp<setult>, v8i1, V8I8>;
580 def: OpR_RR_pat<A2_vcmpbgtu, setugt, i1, V8I8>;
581 def: OpR_RR_pat<A2_vcmpbgtu, setugt, v8i1, V8I8>;
582 def: OpR_RR_pat<A2_vcmpheq, seteq, i1, V4I16>;
583 def: OpR_RR_pat<A2_vcmpheq, seteq, v4i1, V4I16>;
584 def: OpR_RR_pat<A2_vcmphgt, RevCmp<setlt>, i1, V4I16>;
585 def: OpR_RR_pat<A2_vcmphgt, RevCmp<setlt>, v4i1, V4I16>;
586 def: OpR_RR_pat<A2_vcmphgt, setgt, i1, V4I16>;
587 def: OpR_RR_pat<A2_vcmphgt, setgt, v4i1, V4I16>;
588 def: OpR_RR_pat<A2_vcmphgtu, RevCmp<setult>, i1, V4I16>;
589 def: OpR_RR_pat<A2_vcmphgtu, RevCmp<setult>, v4i1, V4I16>;
590 def: OpR_RR_pat<A2_vcmphgtu, setugt, i1, V4I16>;
591 def: OpR_RR_pat<A2_vcmphgtu, setugt, v4i1, V4I16>;
592 def: OpR_RR_pat<A2_vcmpweq, seteq, i1, V2I32>;
593 def: OpR_RR_pat<A2_vcmpweq, seteq, v2i1, V2I32>;
594 def: OpR_RR_pat<A2_vcmpwgt, RevCmp<setlt>, i1, V2I32>;
595 def: OpR_RR_pat<A2_vcmpwgt, RevCmp<setlt>, v2i1, V2I32>;
596 def: OpR_RR_pat<A2_vcmpwgt, setgt, i1, V2I32>;
597 def: OpR_RR_pat<A2_vcmpwgt, setgt, v2i1, V2I32>;
598 def: OpR_RR_pat<A2_vcmpwgtu, RevCmp<setult>, i1, V2I32>;
599 def: OpR_RR_pat<A2_vcmpwgtu, RevCmp<setult>, v2i1, V2I32>;
600 def: OpR_RR_pat<A2_vcmpwgtu, setugt, i1, V2I32>;
601 def: OpR_RR_pat<A2_vcmpwgtu, setugt, v2i1, V2I32>;
603 def: OpR_RR_pat<F2_sfcmpeq, seteq, i1, F32>;
604 def: OpR_RR_pat<F2_sfcmpgt, setgt, i1, F32>;
605 def: OpR_RR_pat<F2_sfcmpge, setge, i1, F32>;
606 def: OpR_RR_pat<F2_sfcmpeq, setoeq, i1, F32>;
607 def: OpR_RR_pat<F2_sfcmpgt, setogt, i1, F32>;
608 def: OpR_RR_pat<F2_sfcmpge, setoge, i1, F32>;
609 def: OpR_RR_pat<F2_sfcmpgt, RevCmp<setolt>, i1, F32>;
610 def: OpR_RR_pat<F2_sfcmpge, RevCmp<setole>, i1, F32>;
611 def: OpR_RR_pat<F2_sfcmpgt, RevCmp<setlt>, i1, F32>;
612 def: OpR_RR_pat<F2_sfcmpge, RevCmp<setle>, i1, F32>;
613 def: OpR_RR_pat<F2_sfcmpuo, setuo, i1, F32>;
615 def: OpR_RR_pat<F2_dfcmpeq, seteq, i1, F64>;
616 def: OpR_RR_pat<F2_dfcmpgt, setgt, i1, F64>;
617 def: OpR_RR_pat<F2_dfcmpge, setge, i1, F64>;
618 def: OpR_RR_pat<F2_dfcmpeq, setoeq, i1, F64>;
619 def: OpR_RR_pat<F2_dfcmpgt, setogt, i1, F64>;
620 def: OpR_RR_pat<F2_dfcmpge, setoge, i1, F64>;
621 def: OpR_RR_pat<F2_dfcmpgt, RevCmp<setolt>, i1, F64>;
622 def: OpR_RR_pat<F2_dfcmpge, RevCmp<setole>, i1, F64>;
623 def: OpR_RR_pat<F2_dfcmpgt, RevCmp<setlt>, i1, F64>;
624 def: OpR_RR_pat<F2_dfcmpge, RevCmp<setle>, i1, F64>;
625 def: OpR_RR_pat<F2_dfcmpuo, setuo, i1, F64>;
627 // Avoid C4_cmpneqi, C4_cmpltei, C4_cmplteui, since they cannot form compounds.
629 def: Pat<(i1 (setne I32:$Rs, anyimm:$u5)),
630 (C2_not (C2_cmpeqi I32:$Rs, imm:$u5))>;
631 def: Pat<(i1 (setle I32:$Rs, anyimm:$u5)),
632 (C2_not (C2_cmpgti I32:$Rs, imm:$u5))>;
633 def: Pat<(i1 (setule I32:$Rs, anyimm:$u5)),
634 (C2_not (C2_cmpgtui I32:$Rs, imm:$u5))>;
636 class OpmR_RR_pat<PatFrag Output, PatFrag Op, ValueType ResType,
637 PatFrag RsPred, PatFrag RtPred = RsPred>
638 : Pat<(ResType (Op RsPred:$Rs, RtPred:$Rt)),
639 (Output RsPred:$Rs, RtPred:$Rt)>;
641 class Outn<InstHexagon MI>
642 : OutPatFrag<(ops node:$Rs, node:$Rt),
643 (C2_not (MI $Rs, $Rt))>;
645 def: OpmR_RR_pat<Outn<C2_cmpeq>, setne, i1, I32>;
646 def: OpmR_RR_pat<Outn<C2_cmpgt>, setle, i1, I32>;
647 def: OpmR_RR_pat<Outn<C2_cmpgtu>, setule, i1, I32>;
648 def: OpmR_RR_pat<Outn<C2_cmpgt>, RevCmp<setge>, i1, I32>;
649 def: OpmR_RR_pat<Outn<C2_cmpgtu>, RevCmp<setuge>, i1, I32>;
650 def: OpmR_RR_pat<Outn<C2_cmpeqp>, setne, i1, I64>;
651 def: OpmR_RR_pat<Outn<C2_cmpgtp>, setle, i1, I64>;
652 def: OpmR_RR_pat<Outn<C2_cmpgtup>, setule, i1, I64>;
653 def: OpmR_RR_pat<Outn<C2_cmpgtp>, RevCmp<setge>, i1, I64>;
654 def: OpmR_RR_pat<Outn<C2_cmpgtup>, RevCmp<setuge>, i1, I64>;
655 def: OpmR_RR_pat<Outn<A2_vcmpbeq>, setne, v8i1, V8I8>;
656 def: OpmR_RR_pat<Outn<A4_vcmpbgt>, setle, v8i1, V8I8>;
657 def: OpmR_RR_pat<Outn<A2_vcmpbgtu>, setule, v8i1, V8I8>;
658 def: OpmR_RR_pat<Outn<A4_vcmpbgt>, RevCmp<setge>, v8i1, V8I8>;
659 def: OpmR_RR_pat<Outn<A2_vcmpbgtu>, RevCmp<setuge>, v8i1, V8I8>;
660 def: OpmR_RR_pat<Outn<A2_vcmpheq>, setne, v4i1, V4I16>;
661 def: OpmR_RR_pat<Outn<A2_vcmphgt>, setle, v4i1, V4I16>;
662 def: OpmR_RR_pat<Outn<A2_vcmphgtu>, setule, v4i1, V4I16>;
663 def: OpmR_RR_pat<Outn<A2_vcmphgt>, RevCmp<setge>, v4i1, V4I16>;
664 def: OpmR_RR_pat<Outn<A2_vcmphgtu>, RevCmp<setuge>, v4i1, V4I16>;
665 def: OpmR_RR_pat<Outn<A2_vcmpweq>, setne, v2i1, V2I32>;
666 def: OpmR_RR_pat<Outn<A2_vcmpwgt>, setle, v2i1, V2I32>;
667 def: OpmR_RR_pat<Outn<A2_vcmpwgtu>, setule, v2i1, V2I32>;
668 def: OpmR_RR_pat<Outn<A2_vcmpwgt>, RevCmp<setge>, v2i1, V2I32>;
669 def: OpmR_RR_pat<Outn<A2_vcmpwgtu>, RevCmp<setuge>, v2i1, V2I32>;
671 let AddedComplexity = 100 in {
672 def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt), 255), 0)),
673 (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt)>;
674 def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt), 255), 0)),
675 (C2_not (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt))>;
676 def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt), 65535), 0)),
677 (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt)>;
678 def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt), 65535), 0)),
679 (C2_not (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt))>;
682 // PatFrag for AsserZext which takes the original type as a parameter.
683 def SDTAssertZext: SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisSameAs<0,1>]>;
684 def AssertZextSD: SDNode<"ISD::AssertZext", SDTAssertZext>;
685 class AssertZext<ValueType T>: PatFrag<(ops node:$A), (AssertZextSD $A, T)>;
687 multiclass Cmpb_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
688 PatLeaf ImmPred, int Mask> {
689 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
690 (MI I32:$Rs, imm:$I)>;
691 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
692 (MI I32:$Rs, imm:$I)>;
695 multiclass CmpbN_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
696 PatLeaf ImmPred, int Mask> {
697 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
698 (C2_not (MI I32:$Rs, imm:$I))>;
699 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
700 (C2_not (MI I32:$Rs, imm:$I))>;
703 multiclass CmpbND_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
704 PatLeaf ImmPred, int Mask> {
705 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
706 (C2_not (MI I32:$Rs, (UDEC1 imm:$I)))>;
707 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
708 (C2_not (MI I32:$Rs, (UDEC1 imm:$I)))>;
711 let AddedComplexity = 200 in {
712 defm: Cmpb_pat <A4_cmpbeqi, seteq, AssertZext<i8>, IsUGT<8,31>, 255>;
713 defm: CmpbN_pat <A4_cmpbeqi, setne, AssertZext<i8>, IsUGT<8,31>, 255>;
714 defm: Cmpb_pat <A4_cmpbgtui, setugt, AssertZext<i8>, IsUGT<32,31>, 255>;
715 defm: CmpbN_pat <A4_cmpbgtui, setule, AssertZext<i8>, IsUGT<32,31>, 255>;
716 defm: Cmpb_pat <A4_cmphgtui, setugt, AssertZext<i16>, IsUGT<32,31>, 65535>;
717 defm: CmpbN_pat <A4_cmphgtui, setule, AssertZext<i16>, IsUGT<32,31>, 65535>;
718 defm: CmpbND_pat<A4_cmpbgtui, setult, AssertZext<i8>, IsUGT<32,32>, 255>;
719 defm: CmpbND_pat<A4_cmphgtui, setult, AssertZext<i16>, IsUGT<32,32>, 65535>;
722 def: Pat<(i32 (zext (i1 (seteq I32:$Rs, I32:$Rt)))),
723 (A4_rcmpeq I32:$Rs, I32:$Rt)>;
724 def: Pat<(i32 (zext (i1 (setne I32:$Rs, I32:$Rt)))),
725 (A4_rcmpneq I32:$Rs, I32:$Rt)>;
726 def: Pat<(i32 (zext (i1 (seteq I32:$Rs, anyimm:$s8)))),
727 (A4_rcmpeqi I32:$Rs, imm:$s8)>;
728 def: Pat<(i32 (zext (i1 (setne I32:$Rs, anyimm:$s8)))),
729 (A4_rcmpneqi I32:$Rs, imm:$s8)>;
731 def: Pat<(i1 (seteq I1:$Ps, (i1 -1))), (I1:$Ps)>;
732 def: Pat<(i1 (setne I1:$Ps, (i1 -1))), (C2_not I1:$Ps)>;
733 def: Pat<(i1 (seteq I1:$Ps, I1:$Pt)), (C2_xor I1:$Ps, (C2_not I1:$Pt))>;
734 def: Pat<(i1 (setne I1:$Ps, I1:$Pt)), (C2_xor I1:$Ps, I1:$Pt)>;
736 // Floating-point comparisons with checks for ordered/unordered status.
738 class T3<InstHexagon MI1, InstHexagon MI2, InstHexagon MI3>
739 : OutPatFrag<(ops node:$Rs, node:$Rt),
740 (MI1 (MI2 $Rs, $Rt), (MI3 $Rs, $Rt))>;
742 class Cmpuf<InstHexagon MI>: T3<C2_or, F2_sfcmpuo, MI>;
743 class Cmpud<InstHexagon MI>: T3<C2_or, F2_dfcmpuo, MI>;
745 class Cmpufn<InstHexagon MI>: T3<C2_orn, F2_sfcmpuo, MI>;
746 class Cmpudn<InstHexagon MI>: T3<C2_orn, F2_dfcmpuo, MI>;
748 def: OpmR_RR_pat<Cmpuf<F2_sfcmpeq>, setueq, i1, F32>;
749 def: OpmR_RR_pat<Cmpuf<F2_sfcmpge>, setuge, i1, F32>;
750 def: OpmR_RR_pat<Cmpuf<F2_sfcmpgt>, setugt, i1, F32>;
751 def: OpmR_RR_pat<Cmpuf<F2_sfcmpge>, RevCmp<setule>, i1, F32>;
752 def: OpmR_RR_pat<Cmpuf<F2_sfcmpgt>, RevCmp<setult>, i1, F32>;
753 def: OpmR_RR_pat<Cmpufn<F2_sfcmpeq>, setune, i1, F32>;
755 def: OpmR_RR_pat<Cmpud<F2_dfcmpeq>, setueq, i1, F64>;
756 def: OpmR_RR_pat<Cmpud<F2_dfcmpge>, setuge, i1, F64>;
757 def: OpmR_RR_pat<Cmpud<F2_dfcmpgt>, setugt, i1, F64>;
758 def: OpmR_RR_pat<Cmpud<F2_dfcmpge>, RevCmp<setule>, i1, F64>;
759 def: OpmR_RR_pat<Cmpud<F2_dfcmpgt>, RevCmp<setult>, i1, F64>;
760 def: OpmR_RR_pat<Cmpudn<F2_dfcmpeq>, setune, i1, F64>;
762 def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setone, i1, F32>;
763 def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setne, i1, F32>;
765 def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setone, i1, F64>;
766 def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setne, i1, F64>;
768 def: OpmR_RR_pat<Outn<F2_sfcmpuo>, seto, i1, F32>;
769 def: OpmR_RR_pat<Outn<F2_dfcmpuo>, seto, i1, F64>;
772 // --(6) Select ----------------------------------------------------------
775 def: Pat<(select I1:$Pu, I32:$Rs, I32:$Rt),
776 (C2_mux I1:$Pu, I32:$Rs, I32:$Rt)>;
777 def: Pat<(select I1:$Pu, anyimm:$s8, I32:$Rs),
778 (C2_muxri I1:$Pu, imm:$s8, I32:$Rs)>;
779 def: Pat<(select I1:$Pu, I32:$Rs, anyimm:$s8),
780 (C2_muxir I1:$Pu, I32:$Rs, imm:$s8)>;
781 def: Pat<(select I1:$Pu, anyimm:$s8, s8_0ImmPred:$S8),
782 (C2_muxii I1:$Pu, imm:$s8, imm:$S8)>;
784 def: Pat<(select (not I1:$Pu), I32:$Rs, I32:$Rt),
785 (C2_mux I1:$Pu, I32:$Rt, I32:$Rs)>;
786 def: Pat<(select (not I1:$Pu), s8_0ImmPred:$S8, anyimm:$s8),
787 (C2_muxii I1:$Pu, imm:$s8, imm:$S8)>;
788 def: Pat<(select (not I1:$Pu), anyimm:$s8, I32:$Rs),
789 (C2_muxir I1:$Pu, I32:$Rs, imm:$s8)>;
790 def: Pat<(select (not I1:$Pu), I32:$Rs, anyimm:$s8),
791 (C2_muxri I1:$Pu, imm:$s8, I32:$Rs)>;
793 // Map from a 64-bit select to an emulated 64-bit mux.
794 // Hexagon does not support 64-bit MUXes; so emulate with combines.
795 def: Pat<(select I1:$Pu, I64:$Rs, I64:$Rt),
796 (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
797 (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
799 def: Pat<(select I1:$Pu, F32:$Rs, f32ImmPred:$I),
800 (C2_muxir I1:$Pu, F32:$Rs, (ftoi $I))>;
801 def: Pat<(select I1:$Pu, f32ImmPred:$I, F32:$Rt),
802 (C2_muxri I1:$Pu, (ftoi $I), F32:$Rt)>;
803 def: Pat<(select I1:$Pu, F32:$Rs, F32:$Rt),
804 (C2_mux I1:$Pu, F32:$Rs, F32:$Rt)>;
805 def: Pat<(select I1:$Pu, F64:$Rs, F64:$Rt),
806 (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
807 (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
809 def: Pat<(select (i1 (setult F32:$Ra, F32:$Rb)), F32:$Rs, F32:$Rt),
810 (C2_mux (F2_sfcmpgt F32:$Rb, F32:$Ra), F32:$Rs, F32:$Rt)>;
811 def: Pat<(select (i1 (setult F64:$Ra, F64:$Rb)), F64:$Rs, F64:$Rt),
812 (C2_vmux (F2_dfcmpgt F64:$Rb, F64:$Ra), F64:$Rs, F64:$Rt)>;
814 def: Pat<(select (not I1:$Pu), f32ImmPred:$I, F32:$Rs),
815 (C2_muxir I1:$Pu, F32:$Rs, (ftoi $I))>;
816 def: Pat<(select (not I1:$Pu), F32:$Rt, f32ImmPred:$I),
817 (C2_muxri I1:$Pu, (ftoi $I), F32:$Rt)>;
819 def: Pat<(select I1:$Pu, V4I8:$Rs, V4I8:$Rt),
820 (LoReg (C2_vmux I1:$Pu, (ToAext64 $Rs), (ToAext64 $Rt)))>;
821 def: Pat<(select I1:$Pu, V2I16:$Rs, V2I16:$Rt),
822 (LoReg (C2_vmux I1:$Pu, (ToAext64 $Rs), (ToAext64 $Rt)))>;
823 def: Pat<(select I1:$Pu, V2I32:$Rs, V2I32:$Rt),
824 (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
825 (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
827 def: Pat<(vselect V8I1:$Pu, V8I8:$Rs, V8I8:$Rt),
828 (C2_vmux V8I1:$Pu, V8I8:$Rs, V8I8:$Rt)>;
829 def: Pat<(vselect V4I1:$Pu, V4I16:$Rs, V4I16:$Rt),
830 (C2_vmux V4I1:$Pu, V4I16:$Rs, V4I16:$Rt)>;
831 def: Pat<(vselect V2I1:$Pu, V2I32:$Rs, V2I32:$Rt),
832 (C2_vmux V2I1:$Pu, V2I32:$Rs, V2I32:$Rt)>;
834 // From LegalizeDAG.cpp: (Pu ? Pv : Pw) <=> (Pu & Pv) | (!Pu & Pw).
835 def: Pat<(select I1:$Pu, I1:$Pv, I1:$Pw),
836 (C2_or (C2_and I1:$Pu, I1:$Pv),
837 (C2_andn I1:$Pw, I1:$Pu))>;
840 def IsPosHalf : PatLeaf<(i32 IntRegs:$a), [{
841 return isPositiveHalfWord(N);
844 multiclass SelMinMax16_pats<PatFrag CmpOp, InstHexagon InstA,
846 def: Pat<(sext_inreg (select (i1 (CmpOp IsPosHalf:$Rs, IsPosHalf:$Rt)),
847 IsPosHalf:$Rs, IsPosHalf:$Rt), i16),
848 (InstA IntRegs:$Rs, IntRegs:$Rt)>;
849 def: Pat<(sext_inreg (select (i1 (CmpOp IsPosHalf:$Rs, IsPosHalf:$Rt)),
850 IsPosHalf:$Rt, IsPosHalf:$Rs), i16),
851 (InstB IntRegs:$Rs, IntRegs:$Rt)>;
854 let AddedComplexity = 200 in {
855 defm: SelMinMax16_pats<setge, A2_max, A2_min>;
856 defm: SelMinMax16_pats<setgt, A2_max, A2_min>;
857 defm: SelMinMax16_pats<setle, A2_min, A2_max>;
858 defm: SelMinMax16_pats<setlt, A2_min, A2_max>;
859 defm: SelMinMax16_pats<setuge, A2_maxu, A2_minu>;
860 defm: SelMinMax16_pats<setugt, A2_maxu, A2_minu>;
861 defm: SelMinMax16_pats<setule, A2_minu, A2_maxu>;
862 defm: SelMinMax16_pats<setult, A2_minu, A2_maxu>;
865 let AddedComplexity = 200 in {
866 defm: SelMinMax_pats<setge, I32, A2_max, A2_min>;
867 defm: SelMinMax_pats<setgt, I32, A2_max, A2_min>;
868 defm: SelMinMax_pats<setle, I32, A2_min, A2_max>;
869 defm: SelMinMax_pats<setlt, I32, A2_min, A2_max>;
870 defm: SelMinMax_pats<setuge, I32, A2_maxu, A2_minu>;
871 defm: SelMinMax_pats<setugt, I32, A2_maxu, A2_minu>;
872 defm: SelMinMax_pats<setule, I32, A2_minu, A2_maxu>;
873 defm: SelMinMax_pats<setult, I32, A2_minu, A2_maxu>;
875 defm: SelMinMax_pats<setge, I64, A2_maxp, A2_minp>;
876 defm: SelMinMax_pats<setgt, I64, A2_maxp, A2_minp>;
877 defm: SelMinMax_pats<setle, I64, A2_minp, A2_maxp>;
878 defm: SelMinMax_pats<setlt, I64, A2_minp, A2_maxp>;
879 defm: SelMinMax_pats<setuge, I64, A2_maxup, A2_minup>;
880 defm: SelMinMax_pats<setugt, I64, A2_maxup, A2_minup>;
881 defm: SelMinMax_pats<setule, I64, A2_minup, A2_maxup>;
882 defm: SelMinMax_pats<setult, I64, A2_minup, A2_maxup>;
885 let AddedComplexity = 100 in {
886 defm: SelMinMax_pats<setolt, F32, F2_sfmin, F2_sfmax>;
887 defm: SelMinMax_pats<setole, F32, F2_sfmin, F2_sfmax>;
888 defm: SelMinMax_pats<setogt, F32, F2_sfmax, F2_sfmin>;
889 defm: SelMinMax_pats<setoge, F32, F2_sfmax, F2_sfmin>;
893 // --(7) Insert/extract --------------------------------------------------
896 def SDTHexagonINSERT:
897 SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>,
898 SDTCisInt<0>, SDTCisVT<3, i32>, SDTCisVT<4, i32>]>;
899 def HexagonINSERT: SDNode<"HexagonISD::INSERT", SDTHexagonINSERT>;
901 let AddedComplexity = 10 in {
902 def: Pat<(HexagonINSERT I32:$Rs, I32:$Rt, u5_0ImmPred:$u1, u5_0ImmPred:$u2),
903 (S2_insert I32:$Rs, I32:$Rt, imm:$u1, imm:$u2)>;
904 def: Pat<(HexagonINSERT I64:$Rs, I64:$Rt, u6_0ImmPred:$u1, u6_0ImmPred:$u2),
905 (S2_insertp I64:$Rs, I64:$Rt, imm:$u1, imm:$u2)>;
907 def: Pat<(HexagonINSERT I32:$Rs, I32:$Rt, I32:$Width, I32:$Off),
908 (S2_insert_rp I32:$Rs, I32:$Rt, (Combinew $Width, $Off))>;
909 def: Pat<(HexagonINSERT I64:$Rs, I64:$Rt, I32:$Width, I32:$Off),
910 (S2_insertp_rp I64:$Rs, I64:$Rt, (Combinew $Width, $Off))>;
912 def SDTHexagonEXTRACTU
913 : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisInt<0>, SDTCisInt<1>,
914 SDTCisVT<2, i32>, SDTCisVT<3, i32>]>;
915 def HexagonEXTRACTU: SDNode<"HexagonISD::EXTRACTU", SDTHexagonEXTRACTU>;
917 let AddedComplexity = 10 in {
918 def: Pat<(HexagonEXTRACTU I32:$Rs, u5_0ImmPred:$u5, u5_0ImmPred:$U5),
919 (S2_extractu I32:$Rs, imm:$u5, imm:$U5)>;
920 def: Pat<(HexagonEXTRACTU I64:$Rs, u6_0ImmPred:$u6, u6_0ImmPred:$U6),
921 (S2_extractup I64:$Rs, imm:$u6, imm:$U6)>;
923 def: Pat<(HexagonEXTRACTU I32:$Rs, I32:$Width, I32:$Off),
924 (S2_extractu_rp I32:$Rs, (Combinew $Width, $Off))>;
925 def: Pat<(HexagonEXTRACTU I64:$Rs, I32:$Width, I32:$Off),
926 (S2_extractup_rp I64:$Rs, (Combinew $Width, $Off))>;
928 def SDTHexagonVSPLAT:
929 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVT<1, i32>]>;
931 def HexagonVSPLAT: SDNode<"HexagonISD::VSPLAT", SDTHexagonVSPLAT>;
933 def: Pat<(v4i8 (HexagonVSPLAT I32:$Rs)), (S2_vsplatrb I32:$Rs)>;
934 def: Pat<(v4i16 (HexagonVSPLAT I32:$Rs)), (S2_vsplatrh I32:$Rs)>;
935 def: Pat<(v2i32 (HexagonVSPLAT s8_0ImmPred:$s8)),
936 (A2_combineii imm:$s8, imm:$s8)>;
937 def: Pat<(v2i32 (HexagonVSPLAT I32:$Rs)), (Combinew I32:$Rs, I32:$Rs)>;
939 let AddedComplexity = 10 in
940 def: Pat<(v8i8 (HexagonVSPLAT I32:$Rs)), (S6_vsplatrbp I32:$Rs)>,
942 def: Pat<(v8i8 (HexagonVSPLAT I32:$Rs)),
943 (Combinew (S2_vsplatrb I32:$Rs), (S2_vsplatrb I32:$Rs))>;
946 // --(8) Shift/permute ---------------------------------------------------
949 def SDTHexagonI64I32I32: SDTypeProfile<1, 2,
950 [SDTCisVT<0, i64>, SDTCisVT<1, i32>, SDTCisSameAs<1, 2>]>;
952 def HexagonCOMBINE: SDNode<"HexagonISD::COMBINE", SDTHexagonI64I32I32>;
954 def: Pat<(HexagonCOMBINE I32:$Rs, I32:$Rt), (Combinew $Rs, $Rt)>;
956 // The complexity of the combines involving immediates should be greater
957 // than the complexity of the combine with two registers.
958 let AddedComplexity = 50 in {
959 def: Pat<(HexagonCOMBINE I32:$Rs, anyimm:$s8),
960 (A4_combineri IntRegs:$Rs, imm:$s8)>;
961 def: Pat<(HexagonCOMBINE anyimm:$s8, I32:$Rs),
962 (A4_combineir imm:$s8, IntRegs:$Rs)>;
965 // The complexity of the combine with two immediates should be greater than
966 // the complexity of a combine involving a register.
967 let AddedComplexity = 75 in {
968 def: Pat<(HexagonCOMBINE s8_0ImmPred:$s8, anyimm:$u6),
969 (A4_combineii imm:$s8, imm:$u6)>;
970 def: Pat<(HexagonCOMBINE anyimm:$s8, s8_0ImmPred:$S8),
971 (A2_combineii imm:$s8, imm:$S8)>;
974 def: Pat<(bswap I32:$Rs), (A2_swiz I32:$Rs)>;
975 def: Pat<(bswap I64:$Rss), (Combinew (A2_swiz (LoReg $Rss)),
976 (A2_swiz (HiReg $Rss)))>;
978 def: Pat<(shl s6_0ImmPred:$s6, I32:$Rt), (S4_lsli imm:$s6, I32:$Rt)>;
979 def: Pat<(shl I32:$Rs, (i32 16)), (A2_aslh I32:$Rs)>;
980 def: Pat<(sra I32:$Rs, (i32 16)), (A2_asrh I32:$Rs)>;
982 def: OpR_RI_pat<S2_asr_i_r, Sra, i32, I32, u5_0ImmPred>;
983 def: OpR_RI_pat<S2_lsr_i_r, Srl, i32, I32, u5_0ImmPred>;
984 def: OpR_RI_pat<S2_asl_i_r, Shl, i32, I32, u5_0ImmPred>;
985 def: OpR_RI_pat<S2_asr_i_p, Sra, i64, I64, u6_0ImmPred>;
986 def: OpR_RI_pat<S2_lsr_i_p, Srl, i64, I64, u6_0ImmPred>;
987 def: OpR_RI_pat<S2_asl_i_p, Shl, i64, I64, u6_0ImmPred>;
988 def: OpR_RI_pat<S2_asr_i_vh, Sra, v4i16, V4I16, u4_0ImmPred>;
989 def: OpR_RI_pat<S2_lsr_i_vh, Srl, v4i16, V4I16, u4_0ImmPred>;
990 def: OpR_RI_pat<S2_asl_i_vh, Shl, v4i16, V4I16, u4_0ImmPred>;
991 def: OpR_RI_pat<S2_asr_i_vh, Sra, v2i32, V2I32, u5_0ImmPred>;
992 def: OpR_RI_pat<S2_lsr_i_vh, Srl, v2i32, V2I32, u5_0ImmPred>;
993 def: OpR_RI_pat<S2_asl_i_vh, Shl, v2i32, V2I32, u5_0ImmPred>;
995 def: OpR_RR_pat<S2_asr_r_r, Sra, i32, I32, I32>;
996 def: OpR_RR_pat<S2_lsr_r_r, Srl, i32, I32, I32>;
997 def: OpR_RR_pat<S2_asl_r_r, Shl, i32, I32, I32>;
998 def: OpR_RR_pat<S2_asr_r_p, Sra, i64, I64, I32>;
999 def: OpR_RR_pat<S2_lsr_r_p, Srl, i64, I64, I32>;
1000 def: OpR_RR_pat<S2_asl_r_p, Shl, i64, I64, I32>;
1003 def IsMul8_U3: PatLeaf<(i32 imm), [{
1004 uint64_t V = N->getZExtValue();
1005 return V % 8 == 0 && isUInt<3>(V / 8);
1008 def Divu8: SDNodeXForm<imm, [{
1009 return CurDAG->getTargetConstant(N->getZExtValue() / 8, SDLoc(N), MVT::i32);
1012 // Funnel shift-left.
1013 def FShl32i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1014 (HiReg (S2_asl_i_p (Combinew $Rs, $Rt), $S))>;
1015 def FShl32r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1016 (HiReg (S2_asl_r_p (Combinew $Rs, $Rt), $Ru))>;
1018 def FShl64i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1019 (S2_lsr_i_p_or (S2_asl_i_p $Rt, $S), $Rs, (Subi<64> $S))>;
1020 def FShl64r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1021 (S2_lsr_r_p_or (S2_asl_r_p $Rt, $Ru), $Rs, (A2_subri 64, $Ru))>;
1023 // Combined SDNodeXForm: (Divu8 (Subi<64> $S))
1024 def Divu64_8: SDNodeXForm<imm, [{
1025 return CurDAG->getTargetConstant((64 - N->getSExtValue()) / 8,
1026 SDLoc(N), MVT::i32);
1030 let AddedComplexity = 100 in {
1031 def: Pat<(fshl I32:$Rs, I32:$Rt, (i32 16)),
1032 (A2_combine_hl I32:$Rs, I32:$Rt)>;
1033 def: Pat<(fshl I64:$Rs, I64:$Rt, IsMul8_U3:$S),
1034 (S2_valignib I64:$Rs, I64:$Rt, (Divu64_8 $S))>;
1037 let Predicates = [HasV60], AddedComplexity = 50 in {
1038 def: OpR_RI_pat<S6_rol_i_r, Rol, i32, I32, u5_0ImmPred>;
1039 def: OpR_RI_pat<S6_rol_i_p, Rol, i64, I64, u6_0ImmPred>;
1041 let AddedComplexity = 30 in {
1042 def: Pat<(rotl I32:$Rs, u5_0ImmPred:$S), (FShl32i $Rs, $Rs, imm:$S)>;
1043 def: Pat<(rotl I64:$Rs, u6_0ImmPred:$S), (FShl64i $Rs, $Rs, imm:$S)>;
1044 def: Pat<(fshl I32:$Rs, I32:$Rt, u5_0ImmPred:$S), (FShl32i $Rs, $Rt, imm:$S)>;
1045 def: Pat<(fshl I64:$Rs, I64:$Rt, u6_0ImmPred:$S), (FShl64i $Rs, $Rt, imm:$S)>;
1047 def: Pat<(rotl I32:$Rs, I32:$Rt), (FShl32r $Rs, $Rs, $Rt)>;
1048 def: Pat<(rotl I64:$Rs, I32:$Rt), (FShl64r $Rs, $Rs, $Rt)>;
1049 def: Pat<(fshl I32:$Rs, I32:$Rt, I32:$Ru), (FShl32r $Rs, $Rt, $Ru)>;
1050 def: Pat<(fshl I64:$Rs, I64:$Rt, I32:$Ru), (FShl64r $Rs, $Rt, $Ru)>;
1052 // Funnel shift-right.
1053 def FShr32i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1054 (LoReg (S2_lsr_i_p (Combinew $Rs, $Rt), $S))>;
1055 def FShr32r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1056 (LoReg (S2_lsr_r_p (Combinew $Rs, $Rt), $Ru))>;
1058 def FShr64i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1059 (S2_asl_i_p_or (S2_lsr_i_p $Rt, $S), $Rs, (Subi<64> $S))>;
1060 def FShr64r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1061 (S2_asl_r_p_or (S2_lsr_r_p $Rt, $Ru), $Rs, (A2_subri 64, $Ru))>;
1064 let AddedComplexity = 100 in {
1065 def: Pat<(fshr I32:$Rs, I32:$Rt, (i32 16)),
1066 (A2_combine_hl I32:$Rs, I32:$Rt)>;
1067 def: Pat<(fshr I64:$Rs, I64:$Rt, IsMul8_U3:$S),
1068 (S2_valignib I64:$Rs, I64:$Rt, (Divu8 $S))>;
1071 let Predicates = [HasV60], AddedComplexity = 50 in {
1072 def: Pat<(rotr I32:$Rs, u5_0ImmPred:$S), (S6_rol_i_r I32:$Rs, (Subi<32> $S))>;
1073 def: Pat<(rotr I64:$Rs, u6_0ImmPred:$S), (S6_rol_i_p I64:$Rs, (Subi<64> $S))>;
1075 let AddedComplexity = 30 in {
1076 def: Pat<(rotr I32:$Rs, u5_0ImmPred:$S), (FShr32i $Rs, $Rs, imm:$S)>;
1077 def: Pat<(rotr I64:$Rs, u6_0ImmPred:$S), (FShr64i $Rs, $Rs, imm:$S)>;
1078 def: Pat<(fshr I32:$Rs, I32:$Rt, u5_0ImmPred:$S), (FShr32i $Rs, $Rt, imm:$S)>;
1079 def: Pat<(fshr I64:$Rs, I64:$Rt, u6_0ImmPred:$S), (FShr64i $Rs, $Rt, imm:$S)>;
1081 def: Pat<(rotr I32:$Rs, I32:$Rt), (FShr32r $Rs, $Rs, $Rt)>;
1082 def: Pat<(rotr I64:$Rs, I32:$Rt), (FShr64r $Rs, $Rs, $Rt)>;
1083 def: Pat<(fshr I32:$Rs, I32:$Rt, I32:$Ru), (FShr32r $Rs, $Rt, $Ru)>;
1084 def: Pat<(fshr I64:$Rs, I64:$Rt, I32:$Ru), (FShr64r $Rs, $Rt, $Ru)>;
1087 def: Pat<(sra (add (sra I32:$Rs, u5_0ImmPred:$u5), 1), (i32 1)),
1088 (S2_asr_i_r_rnd I32:$Rs, imm:$u5)>;
1089 def: Pat<(sra (add (sra I64:$Rs, u6_0ImmPred:$u6), 1), (i32 1)),
1090 (S2_asr_i_p_rnd I64:$Rs, imm:$u6)>;
1092 // Prefer S2_addasl_rrri over S2_asl_i_r_acc.
1093 let AddedComplexity = 120 in
1094 def: Pat<(add I32:$Rt, (shl I32:$Rs, u3_0ImmPred:$u3)),
1095 (S2_addasl_rrri IntRegs:$Rt, IntRegs:$Rs, imm:$u3)>;
1097 let AddedComplexity = 100 in {
1098 def: AccRRI_pat<S2_asr_i_r_acc, Add, Su<Sra>, I32, u5_0ImmPred>;
1099 def: AccRRI_pat<S2_asr_i_r_nac, Sub, Su<Sra>, I32, u5_0ImmPred>;
1100 def: AccRRI_pat<S2_asr_i_r_and, And, Su<Sra>, I32, u5_0ImmPred>;
1101 def: AccRRI_pat<S2_asr_i_r_or, Or, Su<Sra>, I32, u5_0ImmPred>;
1103 def: AccRRI_pat<S2_asr_i_p_acc, Add, Su<Sra>, I64, u6_0ImmPred>;
1104 def: AccRRI_pat<S2_asr_i_p_nac, Sub, Su<Sra>, I64, u6_0ImmPred>;
1105 def: AccRRI_pat<S2_asr_i_p_and, And, Su<Sra>, I64, u6_0ImmPred>;
1106 def: AccRRI_pat<S2_asr_i_p_or, Or, Su<Sra>, I64, u6_0ImmPred>;
1108 def: AccRRI_pat<S2_lsr_i_r_acc, Add, Su<Srl>, I32, u5_0ImmPred>;
1109 def: AccRRI_pat<S2_lsr_i_r_nac, Sub, Su<Srl>, I32, u5_0ImmPred>;
1110 def: AccRRI_pat<S2_lsr_i_r_and, And, Su<Srl>, I32, u5_0ImmPred>;
1111 def: AccRRI_pat<S2_lsr_i_r_or, Or, Su<Srl>, I32, u5_0ImmPred>;
1112 def: AccRRI_pat<S2_lsr_i_r_xacc, Xor, Su<Srl>, I32, u5_0ImmPred>;
1114 def: AccRRI_pat<S2_lsr_i_p_acc, Add, Su<Srl>, I64, u6_0ImmPred>;
1115 def: AccRRI_pat<S2_lsr_i_p_nac, Sub, Su<Srl>, I64, u6_0ImmPred>;
1116 def: AccRRI_pat<S2_lsr_i_p_and, And, Su<Srl>, I64, u6_0ImmPred>;
1117 def: AccRRI_pat<S2_lsr_i_p_or, Or, Su<Srl>, I64, u6_0ImmPred>;
1118 def: AccRRI_pat<S2_lsr_i_p_xacc, Xor, Su<Srl>, I64, u6_0ImmPred>;
1120 def: AccRRI_pat<S2_asl_i_r_acc, Add, Su<Shl>, I32, u5_0ImmPred>;
1121 def: AccRRI_pat<S2_asl_i_r_nac, Sub, Su<Shl>, I32, u5_0ImmPred>;
1122 def: AccRRI_pat<S2_asl_i_r_and, And, Su<Shl>, I32, u5_0ImmPred>;
1123 def: AccRRI_pat<S2_asl_i_r_or, Or, Su<Shl>, I32, u5_0ImmPred>;
1124 def: AccRRI_pat<S2_asl_i_r_xacc, Xor, Su<Shl>, I32, u5_0ImmPred>;
1126 def: AccRRI_pat<S2_asl_i_p_acc, Add, Su<Shl>, I64, u6_0ImmPred>;
1127 def: AccRRI_pat<S2_asl_i_p_nac, Sub, Su<Shl>, I64, u6_0ImmPred>;
1128 def: AccRRI_pat<S2_asl_i_p_and, And, Su<Shl>, I64, u6_0ImmPred>;
1129 def: AccRRI_pat<S2_asl_i_p_or, Or, Su<Shl>, I64, u6_0ImmPred>;
1130 def: AccRRI_pat<S2_asl_i_p_xacc, Xor, Su<Shl>, I64, u6_0ImmPred>;
1132 let Predicates = [HasV60] in {
1133 def: AccRRI_pat<S6_rol_i_r_acc, Add, Su<Rol>, I32, u5_0ImmPred>;
1134 def: AccRRI_pat<S6_rol_i_r_nac, Sub, Su<Rol>, I32, u5_0ImmPred>;
1135 def: AccRRI_pat<S6_rol_i_r_and, And, Su<Rol>, I32, u5_0ImmPred>;
1136 def: AccRRI_pat<S6_rol_i_r_or, Or, Su<Rol>, I32, u5_0ImmPred>;
1137 def: AccRRI_pat<S6_rol_i_r_xacc, Xor, Su<Rol>, I32, u5_0ImmPred>;
1139 def: AccRRI_pat<S6_rol_i_p_acc, Add, Su<Rol>, I64, u6_0ImmPred>;
1140 def: AccRRI_pat<S6_rol_i_p_nac, Sub, Su<Rol>, I64, u6_0ImmPred>;
1141 def: AccRRI_pat<S6_rol_i_p_and, And, Su<Rol>, I64, u6_0ImmPred>;
1142 def: AccRRI_pat<S6_rol_i_p_or, Or, Su<Rol>, I64, u6_0ImmPred>;
1143 def: AccRRI_pat<S6_rol_i_p_xacc, Xor, Su<Rol>, I64, u6_0ImmPred>;
1147 let AddedComplexity = 100 in {
1148 def: AccRRR_pat<S2_asr_r_r_acc, Add, Su<Sra>, I32, I32, I32>;
1149 def: AccRRR_pat<S2_asr_r_r_nac, Sub, Su<Sra>, I32, I32, I32>;
1150 def: AccRRR_pat<S2_asr_r_r_and, And, Su<Sra>, I32, I32, I32>;
1151 def: AccRRR_pat<S2_asr_r_r_or, Or, Su<Sra>, I32, I32, I32>;
1153 def: AccRRR_pat<S2_asr_r_p_acc, Add, Su<Sra>, I64, I64, I32>;
1154 def: AccRRR_pat<S2_asr_r_p_nac, Sub, Su<Sra>, I64, I64, I32>;
1155 def: AccRRR_pat<S2_asr_r_p_and, And, Su<Sra>, I64, I64, I32>;
1156 def: AccRRR_pat<S2_asr_r_p_or, Or, Su<Sra>, I64, I64, I32>;
1157 def: AccRRR_pat<S2_asr_r_p_xor, Xor, Su<Sra>, I64, I64, I32>;
1159 def: AccRRR_pat<S2_lsr_r_r_acc, Add, Su<Srl>, I32, I32, I32>;
1160 def: AccRRR_pat<S2_lsr_r_r_nac, Sub, Su<Srl>, I32, I32, I32>;
1161 def: AccRRR_pat<S2_lsr_r_r_and, And, Su<Srl>, I32, I32, I32>;
1162 def: AccRRR_pat<S2_lsr_r_r_or, Or, Su<Srl>, I32, I32, I32>;
1164 def: AccRRR_pat<S2_lsr_r_p_acc, Add, Su<Srl>, I64, I64, I32>;
1165 def: AccRRR_pat<S2_lsr_r_p_nac, Sub, Su<Srl>, I64, I64, I32>;
1166 def: AccRRR_pat<S2_lsr_r_p_and, And, Su<Srl>, I64, I64, I32>;
1167 def: AccRRR_pat<S2_lsr_r_p_or, Or, Su<Srl>, I64, I64, I32>;
1168 def: AccRRR_pat<S2_lsr_r_p_xor, Xor, Su<Srl>, I64, I64, I32>;
1170 def: AccRRR_pat<S2_asl_r_r_acc, Add, Su<Shl>, I32, I32, I32>;
1171 def: AccRRR_pat<S2_asl_r_r_nac, Sub, Su<Shl>, I32, I32, I32>;
1172 def: AccRRR_pat<S2_asl_r_r_and, And, Su<Shl>, I32, I32, I32>;
1173 def: AccRRR_pat<S2_asl_r_r_or, Or, Su<Shl>, I32, I32, I32>;
1175 def: AccRRR_pat<S2_asl_r_p_acc, Add, Su<Shl>, I64, I64, I32>;
1176 def: AccRRR_pat<S2_asl_r_p_nac, Sub, Su<Shl>, I64, I64, I32>;
1177 def: AccRRR_pat<S2_asl_r_p_and, And, Su<Shl>, I64, I64, I32>;
1178 def: AccRRR_pat<S2_asl_r_p_or, Or, Su<Shl>, I64, I64, I32>;
1179 def: AccRRR_pat<S2_asl_r_p_xor, Xor, Su<Shl>, I64, I64, I32>;
1183 class OpshIRI_pat<InstHexagon MI, PatFrag Op, PatFrag ShOp,
1184 PatFrag RegPred, PatFrag ImmPred>
1185 : Pat<(Op anyimm:$u8, (ShOp RegPred:$Rs, ImmPred:$U5)),
1186 (MI anyimm:$u8, RegPred:$Rs, imm:$U5)>;
1188 let AddedComplexity = 200 in {
1189 def: OpshIRI_pat<S4_addi_asl_ri, Add, Su<Shl>, I32, u5_0ImmPred>;
1190 def: OpshIRI_pat<S4_addi_lsr_ri, Add, Su<Srl>, I32, u5_0ImmPred>;
1191 def: OpshIRI_pat<S4_subi_asl_ri, Sub, Su<Shl>, I32, u5_0ImmPred>;
1192 def: OpshIRI_pat<S4_subi_lsr_ri, Sub, Su<Srl>, I32, u5_0ImmPred>;
1193 def: OpshIRI_pat<S4_andi_asl_ri, And, Su<Shl>, I32, u5_0ImmPred>;
1194 def: OpshIRI_pat<S4_andi_lsr_ri, And, Su<Srl>, I32, u5_0ImmPred>;
1195 def: OpshIRI_pat<S4_ori_asl_ri, Or, Su<Shl>, I32, u5_0ImmPred>;
1196 def: OpshIRI_pat<S4_ori_lsr_ri, Or, Su<Srl>, I32, u5_0ImmPred>;
1199 // Prefer this pattern to S2_asl_i_p_or for the special case of joining
1200 // two 32-bit words into a 64-bit word.
1201 let AddedComplexity = 200 in
1202 def: Pat<(or (shl (Aext64 I32:$a), (i32 32)), (Zext64 I32:$b)),
1203 (Combinew I32:$a, I32:$b)>;
1205 def: Pat<(or (or (or (shl (Zext64 (and I32:$b, (i32 65535))), (i32 16)),
1206 (Zext64 (and I32:$a, (i32 65535)))),
1207 (shl (Aext64 (and I32:$c, (i32 65535))), (i32 32))),
1208 (shl (Aext64 I32:$d), (i32 48))),
1209 (Combinew (A2_combine_ll I32:$d, I32:$c),
1210 (A2_combine_ll I32:$b, I32:$a))>;
1212 let AddedComplexity = 200 in {
1213 def: Pat<(or (shl I32:$Rt, (i32 16)), (and I32:$Rs, (i32 65535))),
1214 (A2_combine_ll I32:$Rt, I32:$Rs)>;
1215 def: Pat<(or (shl I32:$Rt, (i32 16)), (srl I32:$Rs, (i32 16))),
1216 (A2_combine_lh I32:$Rt, I32:$Rs)>;
1217 def: Pat<(or (and I32:$Rt, (i32 268431360)), (and I32:$Rs, (i32 65535))),
1218 (A2_combine_hl I32:$Rt, I32:$Rs)>;
1219 def: Pat<(or (and I32:$Rt, (i32 268431360)), (srl I32:$Rs, (i32 16))),
1220 (A2_combine_hh I32:$Rt, I32:$Rs)>;
1223 def SDTHexagonVShift
1224 : SDTypeProfile<1, 2, [SDTCisSameAs<0, 1>, SDTCisVec<0>, SDTCisVT<2, i32>]>;
1226 def HexagonVASL: SDNode<"HexagonISD::VASL", SDTHexagonVShift>;
1227 def HexagonVASR: SDNode<"HexagonISD::VASR", SDTHexagonVShift>;
1228 def HexagonVLSR: SDNode<"HexagonISD::VLSR", SDTHexagonVShift>;
1230 def: OpR_RI_pat<S2_asl_i_vw, pf2<HexagonVASL>, v2i32, V2I32, u5_0ImmPred>;
1231 def: OpR_RI_pat<S2_asl_i_vh, pf2<HexagonVASL>, v4i16, V4I16, u4_0ImmPred>;
1232 def: OpR_RI_pat<S2_asr_i_vw, pf2<HexagonVASR>, v2i32, V2I32, u5_0ImmPred>;
1233 def: OpR_RI_pat<S2_asr_i_vh, pf2<HexagonVASR>, v4i16, V4I16, u4_0ImmPred>;
1234 def: OpR_RI_pat<S2_lsr_i_vw, pf2<HexagonVLSR>, v2i32, V2I32, u5_0ImmPred>;
1235 def: OpR_RI_pat<S2_lsr_i_vh, pf2<HexagonVLSR>, v4i16, V4I16, u4_0ImmPred>;
1237 def: OpR_RR_pat<S2_asl_r_vw, pf2<HexagonVASL>, v2i32, V2I32, I32>;
1238 def: OpR_RR_pat<S2_asl_r_vh, pf2<HexagonVASL>, v4i16, V4I16, I32>;
1239 def: OpR_RR_pat<S2_asr_r_vw, pf2<HexagonVASR>, v2i32, V2I32, I32>;
1240 def: OpR_RR_pat<S2_asr_r_vh, pf2<HexagonVASR>, v4i16, V4I16, I32>;
1241 def: OpR_RR_pat<S2_lsr_r_vw, pf2<HexagonVLSR>, v2i32, V2I32, I32>;
1242 def: OpR_RR_pat<S2_lsr_r_vh, pf2<HexagonVLSR>, v4i16, V4I16, I32>;
1244 def: Pat<(sra V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c))),
1245 (S2_asr_i_vw V2I32:$b, imm:$c)>;
1246 def: Pat<(srl V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c))),
1247 (S2_lsr_i_vw V2I32:$b, imm:$c)>;
1248 def: Pat<(shl V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c))),
1249 (S2_asl_i_vw V2I32:$b, imm:$c)>;
1250 def: Pat<(sra V4I16:$b, (v4i16 (HexagonVSPLAT u4_0ImmPred:$c))),
1251 (S2_asr_i_vh V4I16:$b, imm:$c)>;
1252 def: Pat<(srl V4I16:$b, (v4i16 (HexagonVSPLAT u4_0ImmPred:$c))),
1253 (S2_lsr_i_vh V4I16:$b, imm:$c)>;
1254 def: Pat<(shl V4I16:$b, (v4i16 (HexagonVSPLAT u4_0ImmPred:$c))),
1255 (S2_asl_i_vh V4I16:$b, imm:$c)>;
1257 def: Pat<(HexagonVASR V2I16:$Rs, u4_0ImmPred:$S),
1258 (LoReg (S2_asr_i_vh (ToAext64 $Rs), imm:$S))>;
1259 def: Pat<(HexagonVASL V2I16:$Rs, u4_0ImmPred:$S),
1260 (LoReg (S2_asl_i_vh (ToAext64 $Rs), imm:$S))>;
1261 def: Pat<(HexagonVLSR V2I16:$Rs, u4_0ImmPred:$S),
1262 (LoReg (S2_lsr_i_vh (ToAext64 $Rs), imm:$S))>;
1263 def: Pat<(HexagonVASR V2I16:$Rs, I32:$Rt),
1264 (LoReg (S2_asr_i_vh (ToAext64 $Rs), I32:$Rt))>;
1265 def: Pat<(HexagonVASL V2I16:$Rs, I32:$Rt),
1266 (LoReg (S2_asl_i_vh (ToAext64 $Rs), I32:$Rt))>;
1267 def: Pat<(HexagonVLSR V2I16:$Rs, I32:$Rt),
1268 (LoReg (S2_lsr_i_vh (ToAext64 $Rs), I32:$Rt))>;
1271 // --(9) Arithmetic/bitwise ----------------------------------------------
1274 def: Pat<(abs I32:$Rs), (A2_abs I32:$Rs)>;
1275 def: Pat<(abs I64:$Rs), (A2_absp I64:$Rs)>;
1276 def: Pat<(not I32:$Rs), (A2_subri -1, I32:$Rs)>;
1277 def: Pat<(not I64:$Rs), (A2_notp I64:$Rs)>;
1278 def: Pat<(ineg I64:$Rs), (A2_negp I64:$Rs)>;
1280 def: Pat<(fabs F32:$Rs), (S2_clrbit_i F32:$Rs, 31)>;
1281 def: Pat<(fneg F32:$Rs), (S2_togglebit_i F32:$Rs, 31)>;
1283 def: Pat<(fabs F64:$Rs),
1284 (Combinew (S2_clrbit_i (HiReg $Rs), 31),
1285 (i32 (LoReg $Rs)))>;
1286 def: Pat<(fneg F64:$Rs),
1287 (Combinew (S2_togglebit_i (HiReg $Rs), 31),
1288 (i32 (LoReg $Rs)))>;
1290 def: Pat<(add I32:$Rs, anyimm:$s16), (A2_addi I32:$Rs, imm:$s16)>;
1291 def: Pat<(or I32:$Rs, anyimm:$s10), (A2_orir I32:$Rs, imm:$s10)>;
1292 def: Pat<(and I32:$Rs, anyimm:$s10), (A2_andir I32:$Rs, imm:$s10)>;
1293 def: Pat<(sub anyimm:$s10, I32:$Rs), (A2_subri imm:$s10, I32:$Rs)>;
1295 def: OpR_RR_pat<A2_add, Add, i32, I32>;
1296 def: OpR_RR_pat<A2_sub, Sub, i32, I32>;
1297 def: OpR_RR_pat<A2_and, And, i32, I32>;
1298 def: OpR_RR_pat<A2_or, Or, i32, I32>;
1299 def: OpR_RR_pat<A2_xor, Xor, i32, I32>;
1300 def: OpR_RR_pat<A2_addp, Add, i64, I64>;
1301 def: OpR_RR_pat<A2_subp, Sub, i64, I64>;
1302 def: OpR_RR_pat<A2_andp, And, i64, I64>;
1303 def: OpR_RR_pat<A2_orp, Or, i64, I64>;
1304 def: OpR_RR_pat<A2_xorp, Xor, i64, I64>;
1305 def: OpR_RR_pat<A4_andnp, Not2<And>, i64, I64>;
1306 def: OpR_RR_pat<A4_ornp, Not2<Or>, i64, I64>;
1308 def: OpR_RR_pat<A2_svaddh, Add, v2i16, V2I16>;
1309 def: OpR_RR_pat<A2_svsubh, Sub, v2i16, V2I16>;
1311 def: OpR_RR_pat<A2_vaddub, Add, v8i8, V8I8>;
1312 def: OpR_RR_pat<A2_vaddh, Add, v4i16, V4I16>;
1313 def: OpR_RR_pat<A2_vaddw, Add, v2i32, V2I32>;
1314 def: OpR_RR_pat<A2_vsubub, Sub, v8i8, V8I8>;
1315 def: OpR_RR_pat<A2_vsubh, Sub, v4i16, V4I16>;
1316 def: OpR_RR_pat<A2_vsubw, Sub, v2i32, V2I32>;
1318 def: OpR_RR_pat<A2_and, And, v4i8, V4I8>;
1319 def: OpR_RR_pat<A2_xor, Xor, v4i8, V4I8>;
1320 def: OpR_RR_pat<A2_or, Or, v4i8, V4I8>;
1321 def: OpR_RR_pat<A2_and, And, v2i16, V2I16>;
1322 def: OpR_RR_pat<A2_xor, Xor, v2i16, V2I16>;
1323 def: OpR_RR_pat<A2_or, Or, v2i16, V2I16>;
1324 def: OpR_RR_pat<A2_andp, And, v8i8, V8I8>;
1325 def: OpR_RR_pat<A2_orp, Or, v8i8, V8I8>;
1326 def: OpR_RR_pat<A2_xorp, Xor, v8i8, V8I8>;
1327 def: OpR_RR_pat<A2_andp, And, v4i16, V4I16>;
1328 def: OpR_RR_pat<A2_orp, Or, v4i16, V4I16>;
1329 def: OpR_RR_pat<A2_xorp, Xor, v4i16, V4I16>;
1330 def: OpR_RR_pat<A2_andp, And, v2i32, V2I32>;
1331 def: OpR_RR_pat<A2_orp, Or, v2i32, V2I32>;
1332 def: OpR_RR_pat<A2_xorp, Xor, v2i32, V2I32>;
1334 def: OpR_RR_pat<M2_mpyi, Mul, i32, I32>;
1335 def: OpR_RR_pat<M2_mpy_up, pf2<mulhs>, i32, I32>;
1336 def: OpR_RR_pat<M2_mpyu_up, pf2<mulhu>, i32, I32>;
1337 def: OpR_RI_pat<M2_mpysip, Mul, i32, I32, u32_0ImmPred>;
1338 def: OpR_RI_pat<M2_mpysmi, Mul, i32, I32, s32_0ImmPred>;
1340 // Arithmetic on predicates.
1341 def: OpR_RR_pat<C2_xor, Add, i1, I1>;
1342 def: OpR_RR_pat<C2_xor, Add, v2i1, V2I1>;
1343 def: OpR_RR_pat<C2_xor, Add, v4i1, V4I1>;
1344 def: OpR_RR_pat<C2_xor, Add, v8i1, V8I1>;
1345 def: OpR_RR_pat<C2_xor, Sub, i1, I1>;
1346 def: OpR_RR_pat<C2_xor, Sub, v2i1, V2I1>;
1347 def: OpR_RR_pat<C2_xor, Sub, v4i1, V4I1>;
1348 def: OpR_RR_pat<C2_xor, Sub, v8i1, V8I1>;
1349 def: OpR_RR_pat<C2_and, Mul, i1, I1>;
1350 def: OpR_RR_pat<C2_and, Mul, v2i1, V2I1>;
1351 def: OpR_RR_pat<C2_and, Mul, v4i1, V4I1>;
1352 def: OpR_RR_pat<C2_and, Mul, v8i1, V8I1>;
1354 def: OpR_RR_pat<F2_sfadd, pf2<fadd>, f32, F32>;
1355 def: OpR_RR_pat<F2_sfsub, pf2<fsub>, f32, F32>;
1356 def: OpR_RR_pat<F2_sfmpy, pf2<fmul>, f32, F32>;
1357 def: OpR_RR_pat<F2_sfmin, pf2<fminnum>, f32, F32>;
1358 def: OpR_RR_pat<F2_sfmax, pf2<fmaxnum>, f32, F32>;
1360 let Predicates = [HasV66] in {
1361 def: OpR_RR_pat<F2_dfadd, pf2<fadd>, f64, F64>;
1362 def: OpR_RR_pat<F2_dfsub, pf2<fsub>, f64, F64>;
1365 // In expressions like a0*b0 + a1*b1 + ..., prefer to generate multiply-add,
1366 // over add-add with individual multiplies as inputs.
1367 let AddedComplexity = 10 in {
1368 def: AccRRI_pat<M2_macsip, Add, Su<Mul>, I32, u32_0ImmPred>;
1369 def: AccRRI_pat<M2_macsin, Sub, Su<Mul>, I32, u32_0ImmPred>;
1370 def: AccRRR_pat<M2_maci, Add, Su<Mul>, I32, I32, I32>;
1371 let Predicates = [HasV66] in
1372 def: AccRRR_pat<M2_mnaci, Sub, Su<Mul>, I32, I32, I32>;
1375 def: AccRRI_pat<M2_naccii, Sub, Su<Add>, I32, s32_0ImmPred>;
1376 def: AccRRI_pat<M2_accii, Add, Su<Add>, I32, s32_0ImmPred>;
1377 def: AccRRR_pat<M2_acci, Add, Su<Add>, I32, I32, I32>;
1381 def: Pat<(v2i32 (mulhu V2I32:$Rss, V2I32:$Rtt)),
1382 (Combinew (M2_mpyu_up (HiReg $Rss), (HiReg $Rtt)),
1383 (M2_mpyu_up (LoReg $Rss), (LoReg $Rtt)))>;
1385 def: Pat<(v2i32 (mulhs V2I32:$Rs, V2I32:$Rt)),
1386 (Combinew (M2_mpy_up (HiReg $Rs), (HiReg $Rt)),
1387 (M2_mpy_up (LoReg $Rt), (LoReg $Rt)))>;
1390 OutPatFrag<(ops node:$Rss, node:$Rtt),
1391 (Combinew (S2_vtrunohb (M5_vmpybuu (HiReg $Rss), (HiReg $Rtt))),
1392 (S2_vtrunohb (M5_vmpybuu (LoReg $Rss), (LoReg $Rtt))))>;
1394 // Equivalent of byte-wise arithmetic shift right by 7 in v8i8.
1396 OutPatFrag<(ops node:$Rss), (C2_mask (C2_not (A4_vcmpbgti $Rss, 0)))>;
1398 def: Pat<(v8i8 (mulhu V8I8:$Rss, V8I8:$Rtt)),
1399 (Mulhub $Rss, $Rtt)>;
1401 def: Pat<(v8i8 (mulhs V8I8:$Rss, V8I8:$Rtt)),
1403 (Mulhub $Rss, $Rtt),
1404 (A2_vaddub (A2_andp V8I8:$Rss, (Asr7 $Rtt)),
1405 (A2_andp V8I8:$Rtt, (Asr7 $Rss))))>;
1408 OutPatFrag<(ops node:$Rs, node:$Rt), (M2_vmpy2s_s0 $Rs, $Rt)>;
1410 OutPatFrag<(ops node:$Rss, node:$Rtt), (Mpysh (HiReg $Rss), (HiReg $Rtt))>;
1412 OutPatFrag<(ops node:$Rss, node:$Rtt), (Mpysh (LoReg $Rss), (LoReg $Rtt))>;
1415 OutPatFrag<(ops node:$Rss, node:$Rtt),
1416 (Combinew (A2_combine_hh (HiReg (Mpyshh $Rss, $Rtt)),
1417 (LoReg (Mpyshh $Rss, $Rtt))),
1418 (A2_combine_hh (HiReg (Mpyshl $Rss, $Rtt)),
1419 (LoReg (Mpyshl $Rss, $Rtt))))>;
1421 def: Pat<(v4i16 (mulhs V4I16:$Rss, V4I16:$Rtt)), (Mulhsh $Rss, $Rtt)>;
1423 def: Pat<(v4i16 (mulhu V4I16:$Rss, V4I16:$Rtt)),
1425 (Mulhsh $Rss, $Rtt),
1426 (A2_vaddh (A2_andp V4I16:$Rss, (S2_asr_i_vh $Rtt, 15)),
1427 (A2_andp V4I16:$Rtt, (S2_asr_i_vh $Rss, 15))))>;
1430 def: Pat<(ineg (mul I32:$Rs, u8_0ImmPred:$u8)),
1431 (M2_mpysin IntRegs:$Rs, imm:$u8)>;
1433 def n8_0ImmPred: PatLeaf<(i32 imm), [{
1434 int64_t V = N->getSExtValue();
1435 return -255 <= V && V <= 0;
1438 // Change the sign of the immediate for Rd=-mpyi(Rs,#u8)
1439 def: Pat<(mul I32:$Rs, n8_0ImmPred:$n8),
1440 (M2_mpysin I32:$Rs, (NegImm8 imm:$n8))>;
1442 def: Pat<(add Sext64:$Rs, I64:$Rt),
1443 (A2_addsp (LoReg Sext64:$Rs), I64:$Rt)>;
1445 def: AccRRR_pat<M4_and_and, And, Su_ni1<And>, I32, I32, I32>;
1446 def: AccRRR_pat<M4_and_or, And, Su_ni1<Or>, I32, I32, I32>;
1447 def: AccRRR_pat<M4_and_xor, And, Su<Xor>, I32, I32, I32>;
1448 def: AccRRR_pat<M4_or_and, Or, Su_ni1<And>, I32, I32, I32>;
1449 def: AccRRR_pat<M4_or_or, Or, Su_ni1<Or>, I32, I32, I32>;
1450 def: AccRRR_pat<M4_or_xor, Or, Su<Xor>, I32, I32, I32>;
1451 def: AccRRR_pat<M4_xor_and, Xor, Su_ni1<And>, I32, I32, I32>;
1452 def: AccRRR_pat<M4_xor_or, Xor, Su_ni1<Or>, I32, I32, I32>;
1453 def: AccRRR_pat<M2_xor_xacc, Xor, Su<Xor>, I32, I32, I32>;
1454 def: AccRRR_pat<M4_xor_xacc, Xor, Su<Xor>, I64, I64, I64>;
1456 // For dags like (or (and (not _), _), (shl _, _)) where the "or" with
1457 // one argument matches the patterns below, and with the other argument
1458 // matches S2_asl_r_r_or, etc, prefer the patterns below.
1459 let AddedComplexity = 110 in { // greater than S2_asl_r_r_and/or/xor.
1460 def: AccRRR_pat<M4_and_andn, And, Su<Not2<And>>, I32, I32, I32>;
1461 def: AccRRR_pat<M4_or_andn, Or, Su<Not2<And>>, I32, I32, I32>;
1462 def: AccRRR_pat<M4_xor_andn, Xor, Su<Not2<And>>, I32, I32, I32>;
1465 // S4_addaddi and S4_subaddi don't have tied operands, so give them
1466 // a bit of preference.
1467 let AddedComplexity = 30 in {
1468 def: Pat<(add I32:$Rs, (Su<Add> I32:$Ru, anyimm:$s6)),
1469 (S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
1470 def: Pat<(add anyimm:$s6, (Su<Add> I32:$Rs, I32:$Ru)),
1471 (S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
1472 def: Pat<(add I32:$Rs, (Su<Sub> anyimm:$s6, I32:$Ru)),
1473 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1474 def: Pat<(sub (Su<Add> I32:$Rs, anyimm:$s6), I32:$Ru),
1475 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1476 def: Pat<(add (Su<Sub> I32:$Rs, I32:$Ru), anyimm:$s6),
1477 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1480 def: Pat<(or I32:$Ru, (Su<And> I32:$Rx, anyimm:$s10)),
1481 (S4_or_andix IntRegs:$Ru, IntRegs:$Rx, imm:$s10)>;
1482 def: Pat<(or I32:$Rx, (Su<And> I32:$Rs, anyimm:$s10)),
1483 (S4_or_andi IntRegs:$Rx, IntRegs:$Rs, imm:$s10)>;
1484 def: Pat<(or I32:$Rx, (Su<Or> I32:$Rs, anyimm:$s10)),
1485 (S4_or_ori IntRegs:$Rx, IntRegs:$Rs, imm:$s10)>;
1488 def: Pat<(i32 (trunc (sra (Su<Mul> Sext64:$Rs, Sext64:$Rt), (i32 32)))),
1489 (M2_mpy_up (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1490 def: Pat<(i32 (trunc (srl (Su<Mul> Sext64:$Rs, Sext64:$Rt), (i32 32)))),
1491 (M2_mpy_up (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1493 def: Pat<(mul (Zext64 I32:$Rs), (Zext64 I32:$Rt)),
1494 (M2_dpmpyuu_s0 I32:$Rs, I32:$Rt)>;
1495 def: Pat<(mul (Aext64 I32:$Rs), (Aext64 I32:$Rt)),
1496 (M2_dpmpyuu_s0 I32:$Rs, I32:$Rt)>;
1497 def: Pat<(mul Sext64:$Rs, Sext64:$Rt),
1498 (M2_dpmpyss_s0 (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1500 def: Pat<(add I64:$Rx, (Su<Mul> Sext64:$Rs, Sext64:$Rt)),
1501 (M2_dpmpyss_acc_s0 I64:$Rx, (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1502 def: Pat<(sub I64:$Rx, (Su<Mul> Sext64:$Rs, Sext64:$Rt)),
1503 (M2_dpmpyss_nac_s0 I64:$Rx, (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1504 def: Pat<(add I64:$Rx, (Su<Mul> (Aext64 I32:$Rs), (Aext64 I32:$Rt))),
1505 (M2_dpmpyuu_acc_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1506 def: Pat<(add I64:$Rx, (Su<Mul> (Zext64 I32:$Rs), (Zext64 I32:$Rt))),
1507 (M2_dpmpyuu_acc_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1508 def: Pat<(sub I64:$Rx, (Su<Mul> (Aext64 I32:$Rs), (Aext64 I32:$Rt))),
1509 (M2_dpmpyuu_nac_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1510 def: Pat<(sub I64:$Rx, (Su<Mul> (Zext64 I32:$Rs), (Zext64 I32:$Rt))),
1511 (M2_dpmpyuu_nac_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1514 def: Pat<(sext_inreg (add I32:$Rt, I32:$Rs), i16),
1515 (A2_addh_l16_ll I32:$Rt, I32:$Rs)>;
1516 def: Pat<(sra (add (shl I32:$Rt, (i32 16)), I32:$Rs), (i32 16)),
1517 (A2_addh_l16_hl I32:$Rt, I32:$Rs)>;
1518 def: Pat<(shl (add I32:$Rt, I32:$Rs), (i32 16)),
1519 (A2_addh_h16_ll I32:$Rt, I32:$Rs)>;
1521 // Subtract halfword.
1522 def: Pat<(sext_inreg (sub I32:$Rt, I32:$Rs), i16),
1523 (A2_subh_l16_ll I32:$Rt, I32:$Rs)>;
1524 def: Pat<(sra (add (shl I32:$Rt, (i32 16)), I32:$Rs), (i32 16)),
1525 (A2_addh_l16_hl I32:$Rt, I32:$Rs)>;
1526 def: Pat<(shl (sub I32:$Rt, I32:$Rs), (i32 16)),
1527 (A2_subh_h16_ll I32:$Rt, I32:$Rs)>;
1529 def: Pat<(mul I64:$Rss, I64:$Rtt),
1531 (M2_maci (M2_maci (HiReg (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt))),
1536 (i32 (LoReg (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt)))))>;
1538 def MulHU : OutPatFrag<(ops node:$Rss, node:$Rtt),
1544 (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt)), 32),
1547 (A4_combineir 0, (LoReg (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt))))),
1551 (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt)), 32))>;
1553 // Multiply 64-bit unsigned and use upper result.
1554 def : Pat <(mulhu I64:$Rss, I64:$Rtt), (MulHU $Rss, $Rtt)>;
1556 // Multiply 64-bit signed and use upper result.
1558 // For two signed 64-bit integers A and B, let A' and B' denote A and B
1559 // with the sign bit cleared. Then A = -2^63*s(A) + A', where s(A) is the
1560 // sign bit of A (and identically for B). With this notation, the signed
1561 // product A*B can be written as:
1562 // AB = (-2^63 s(A) + A') * (-2^63 s(B) + B')
1563 // = 2^126 s(A)s(B) - 2^63 [s(A)B'+s(B)A'] + A'B'
1564 // = 2^126 s(A)s(B) + 2^63 [s(A)B'+s(B)A'] + A'B' - 2*2^63 [s(A)B'+s(B)A']
1565 // = (unsigned product AB) - 2^64 [s(A)B'+s(B)A']
1567 // Clear the sign bit in a 64-bit register.
1568 def ClearSign : OutPatFrag<(ops node:$Rss),
1569 (Combinew (S2_clrbit_i (HiReg $Rss), 31), (i32 (LoReg $Rss)))>;
1571 def : Pat <(mulhs I64:$Rss, I64:$Rtt),
1575 (A2_andp (S2_asr_i_p $Rss, 63), (ClearSign $Rtt)),
1576 (A2_andp (S2_asr_i_p $Rtt, 63), (ClearSign $Rss))))>;
1578 // Prefer these instructions over M2_macsip/M2_macsin: the macsi* instructions
1579 // will put the immediate addend into a register, while these instructions will
1580 // use it directly. Such a construct does not appear in the middle of a gep,
1581 // where M2_macsip would be preferable.
1582 let AddedComplexity = 20 in {
1583 def: Pat<(add (Su<Mul> I32:$Rs, u6_0ImmPred:$U6), anyimm:$u6),
1584 (M4_mpyri_addi imm:$u6, IntRegs:$Rs, imm:$U6)>;
1585 def: Pat<(add (Su<Mul> I32:$Rs, I32:$Rt), anyimm:$u6),
1586 (M4_mpyrr_addi imm:$u6, IntRegs:$Rs, IntRegs:$Rt)>;
1589 // Keep these instructions less preferable to M2_macsip/M2_macsin.
1590 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, u6_2ImmPred:$u6_2)),
1591 (M4_mpyri_addr_u2 IntRegs:$Ru, imm:$u6_2, IntRegs:$Rs)>;
1592 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, anyimm:$u6)),
1593 (M4_mpyri_addr IntRegs:$Ru, IntRegs:$Rs, imm:$u6)>;
1594 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Ry, I32:$Rs)),
1595 (M4_mpyrr_addr IntRegs:$Ru, IntRegs:$Ry, IntRegs:$Rs)>;
1598 def: Pat<(fma F32:$Rs, F32:$Rt, F32:$Rx),
1599 (F2_sffma F32:$Rx, F32:$Rs, F32:$Rt)>;
1600 def: Pat<(fma (fneg F32:$Rs), F32:$Rt, F32:$Rx),
1601 (F2_sffms F32:$Rx, F32:$Rs, F32:$Rt)>;
1602 def: Pat<(fma F32:$Rs, (fneg F32:$Rt), F32:$Rx),
1603 (F2_sffms F32:$Rx, F32:$Rs, F32:$Rt)>;
1606 def: Pat<(mul V2I32:$Rs, V2I32:$Rt),
1607 (PS_vmulw V2I32:$Rs, V2I32:$Rt)>;
1608 def: Pat<(add V2I32:$Rx, (mul V2I32:$Rs, V2I32:$Rt)),
1609 (PS_vmulw_acc V2I32:$Rx, V2I32:$Rs, V2I32:$Rt)>;
1611 // Add/subtract two v4i8: Hexagon does not have an insn for this one, so
1612 // we use the double add v8i8, and use only the low part of the result.
1613 def: Pat<(add V4I8:$Rs, V4I8:$Rt),
1614 (LoReg (A2_vaddub (ToAext64 $Rs), (ToAext64 $Rt)))>;
1615 def: Pat<(sub V4I8:$Rs, V4I8:$Rt),
1616 (LoReg (A2_vsubub (ToAext64 $Rs), (ToAext64 $Rt)))>;
1618 // Use M2_vmpy2s_s0 for half-word vector multiply. It multiplies two
1619 // half-words, and saturates the result to a 32-bit value, except the
1620 // saturation never happens (it can only occur with scaling).
1621 def: Pat<(v2i16 (mul V2I16:$Rs, V2I16:$Rt)),
1622 (LoReg (S2_vtrunewh (A2_combineii 0, 0),
1623 (M2_vmpy2s_s0 V2I16:$Rs, V2I16:$Rt)))>;
1624 def: Pat<(v4i16 (mul V4I16:$Rs, V4I16:$Rt)),
1625 (S2_vtrunewh (M2_vmpy2s_s0 (HiReg $Rs), (HiReg $Rt)),
1626 (M2_vmpy2s_s0 (LoReg $Rs), (LoReg $Rt)))>;
1628 // Multiplies two v4i8 vectors.
1629 def: Pat<(v4i8 (mul V4I8:$Rs, V4I8:$Rt)),
1630 (S2_vtrunehb (M5_vmpybuu V4I8:$Rs, V4I8:$Rt))>;
1632 // Multiplies two v8i8 vectors.
1633 def: Pat<(v8i8 (mul V8I8:$Rs, V8I8:$Rt)),
1634 (Combinew (S2_vtrunehb (M5_vmpybuu (HiReg $Rs), (HiReg $Rt))),
1635 (S2_vtrunehb (M5_vmpybuu (LoReg $Rs), (LoReg $Rt))))>;
1638 // --(10) Bit ------------------------------------------------------------
1641 // Count leading zeros.
1642 def: Pat<(ctlz I32:$Rs), (S2_cl0 I32:$Rs)>;
1643 def: Pat<(i32 (trunc (ctlz I64:$Rss))), (S2_cl0p I64:$Rss)>;
1645 // Count trailing zeros.
1646 def: Pat<(cttz I32:$Rs), (S2_ct0 I32:$Rs)>;
1647 def: Pat<(i32 (trunc (cttz I64:$Rss))), (S2_ct0p I64:$Rss)>;
1649 // Count leading ones.
1650 def: Pat<(ctlz (not I32:$Rs)), (S2_cl1 I32:$Rs)>;
1651 def: Pat<(i32 (trunc (ctlz (not I64:$Rss)))), (S2_cl1p I64:$Rss)>;
1653 // Count trailing ones.
1654 def: Pat<(cttz (not I32:$Rs)), (S2_ct1 I32:$Rs)>;
1655 def: Pat<(i32 (trunc (cttz (not I64:$Rss)))), (S2_ct1p I64:$Rss)>;
1657 // Define leading/trailing patterns that require zero-extensions to 64 bits.
1658 def: Pat<(i64 (ctlz I64:$Rss)), (ToZext64 (S2_cl0p I64:$Rss))>;
1659 def: Pat<(i64 (cttz I64:$Rss)), (ToZext64 (S2_ct0p I64:$Rss))>;
1660 def: Pat<(i64 (ctlz (not I64:$Rss))), (ToZext64 (S2_cl1p I64:$Rss))>;
1661 def: Pat<(i64 (cttz (not I64:$Rss))), (ToZext64 (S2_ct1p I64:$Rss))>;
1663 def: Pat<(i64 (ctpop I64:$Rss)), (ToZext64 (S5_popcountp I64:$Rss))>;
1664 def: Pat<(i32 (ctpop I32:$Rs)), (S5_popcountp (A4_combineir 0, I32:$Rs))>;
1666 def: Pat<(bitreverse I32:$Rs), (S2_brev I32:$Rs)>;
1667 def: Pat<(bitreverse I64:$Rss), (S2_brevp I64:$Rss)>;
1669 let AddedComplexity = 20 in { // Complexity greater than and/or/xor
1670 def: Pat<(and I32:$Rs, IsNPow2_32:$V),
1671 (S2_clrbit_i IntRegs:$Rs, (LogN2_32 $V))>;
1672 def: Pat<(or I32:$Rs, IsPow2_32:$V),
1673 (S2_setbit_i IntRegs:$Rs, (Log2_32 $V))>;
1674 def: Pat<(xor I32:$Rs, IsPow2_32:$V),
1675 (S2_togglebit_i IntRegs:$Rs, (Log2_32 $V))>;
1677 def: Pat<(and I32:$Rs, (not (shl 1, I32:$Rt))),
1678 (S2_clrbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1679 def: Pat<(or I32:$Rs, (shl 1, I32:$Rt)),
1680 (S2_setbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1681 def: Pat<(xor I32:$Rs, (shl 1, I32:$Rt)),
1682 (S2_togglebit_r IntRegs:$Rs, IntRegs:$Rt)>;
1685 // Clr/set/toggle bit for 64-bit values with immediate bit index.
1686 let AddedComplexity = 20 in { // Complexity greater than and/or/xor
1687 def: Pat<(and I64:$Rss, IsNPow2_64L:$V),
1688 (Combinew (i32 (HiReg $Rss)),
1689 (S2_clrbit_i (LoReg $Rss), (LogN2_64 $V)))>;
1690 def: Pat<(and I64:$Rss, IsNPow2_64H:$V),
1691 (Combinew (S2_clrbit_i (HiReg $Rss), (UDEC32 (i32 (LogN2_64 $V)))),
1692 (i32 (LoReg $Rss)))>;
1694 def: Pat<(or I64:$Rss, IsPow2_64L:$V),
1695 (Combinew (i32 (HiReg $Rss)),
1696 (S2_setbit_i (LoReg $Rss), (Log2_64 $V)))>;
1697 def: Pat<(or I64:$Rss, IsPow2_64H:$V),
1698 (Combinew (S2_setbit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
1699 (i32 (LoReg $Rss)))>;
1701 def: Pat<(xor I64:$Rss, IsPow2_64L:$V),
1702 (Combinew (i32 (HiReg $Rss)),
1703 (S2_togglebit_i (LoReg $Rss), (Log2_64 $V)))>;
1704 def: Pat<(xor I64:$Rss, IsPow2_64H:$V),
1705 (Combinew (S2_togglebit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
1706 (i32 (LoReg $Rss)))>;
1709 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
1710 def: Pat<(i1 (setne (and (shl 1, u5_0ImmPred:$u5), I32:$Rs), 0)),
1711 (S2_tstbit_i IntRegs:$Rs, imm:$u5)>;
1712 def: Pat<(i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1713 (S2_tstbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1714 def: Pat<(i1 (trunc I32:$Rs)),
1715 (S2_tstbit_i IntRegs:$Rs, 0)>;
1716 def: Pat<(i1 (trunc I64:$Rs)),
1717 (S2_tstbit_i (LoReg DoubleRegs:$Rs), 0)>;
1720 let AddedComplexity = 20 in { // Complexity greater than compare reg-imm.
1721 def: Pat<(i1 (seteq (and I32:$Rs, u6_0ImmPred:$u6), 0)),
1722 (C2_bitsclri IntRegs:$Rs, imm:$u6)>;
1723 def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), 0)),
1724 (C2_bitsclr IntRegs:$Rs, IntRegs:$Rt)>;
1727 let AddedComplexity = 10 in // Complexity greater than compare reg-reg.
1728 def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), IntRegs:$Rt)),
1729 (C2_bitsset IntRegs:$Rs, IntRegs:$Rt)>;
1732 SDTypeProfile<1, 2, [SDTCisVT<0, i1>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
1733 def HexagonTSTBIT: SDNode<"HexagonISD::TSTBIT", SDTTestBit>;
1735 def: Pat<(HexagonTSTBIT I32:$Rs, u5_0ImmPred:$u5),
1736 (S2_tstbit_i I32:$Rs, imm:$u5)>;
1737 def: Pat<(HexagonTSTBIT I32:$Rs, I32:$Rt),
1738 (S2_tstbit_r I32:$Rs, I32:$Rt)>;
1740 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
1741 def: Pat<(i1 (seteq (and (shl 1, u5_0ImmPred:$u5), I32:$Rs), 0)),
1742 (S4_ntstbit_i I32:$Rs, imm:$u5)>;
1743 def: Pat<(i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1744 (S4_ntstbit_r I32:$Rs, I32:$Rt)>;
1747 // Add extra complexity to prefer these instructions over bitsset/bitsclr.
1748 // The reason is that tstbit/ntstbit can be folded into a compound instruction:
1749 // if ([!]tstbit(...)) jump ...
1750 let AddedComplexity = 100 in
1751 def: Pat<(i1 (setne (and I32:$Rs, (i32 IsPow2_32:$u5)), (i32 0))),
1752 (S2_tstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
1754 let AddedComplexity = 100 in
1755 def: Pat<(i1 (seteq (and I32:$Rs, (i32 IsPow2_32:$u5)), (i32 0))),
1756 (S4_ntstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
1758 // Do not increase complexity of these patterns. In the DAG, "cmp i8" may be
1759 // represented as a compare against "value & 0xFF", which is an exact match
1760 // for cmpb (same for cmph). The patterns below do not contain any additional
1761 // complexity that would make them preferable, and if they were actually used
1762 // instead of cmpb/cmph, they would result in a compare against register that
1763 // is loaded with the byte/half mask (i.e. 0xFF or 0xFFFF).
1764 def: Pat<(i1 (setne (and I32:$Rs, u6_0ImmPred:$u6), 0)),
1765 (C4_nbitsclri I32:$Rs, imm:$u6)>;
1766 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), 0)),
1767 (C4_nbitsclr I32:$Rs, I32:$Rt)>;
1768 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), I32:$Rt)),
1769 (C4_nbitsset I32:$Rs, I32:$Rt)>;
1771 // Special patterns to address certain cases where the "top-down" matching
1772 // algorithm would cause suboptimal selection.
1774 let AddedComplexity = 100 in {
1775 // Avoid A4_rcmp[n]eqi in these cases:
1776 def: Pat<(i32 (zext (i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1777 (I1toI32 (S2_tstbit_r IntRegs:$Rs, IntRegs:$Rt))>;
1778 def: Pat<(i32 (zext (i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1779 (I1toI32 (S4_ntstbit_r IntRegs:$Rs, IntRegs:$Rt))>;
1782 // --(11) PIC ------------------------------------------------------------
1785 def SDT_HexagonAtGot
1786 : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
1787 def SDT_HexagonAtPcrel
1788 : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
1790 // AT_GOT address-of-GOT, address-of-global, offset-in-global
1791 def HexagonAtGot : SDNode<"HexagonISD::AT_GOT", SDT_HexagonAtGot>;
1792 // AT_PCREL address-of-global
1793 def HexagonAtPcrel : SDNode<"HexagonISD::AT_PCREL", SDT_HexagonAtPcrel>;
1795 def: Pat<(HexagonAtGot I32:$got, I32:$addr, (i32 0)),
1796 (L2_loadri_io I32:$got, imm:$addr)>;
1797 def: Pat<(HexagonAtGot I32:$got, I32:$addr, s30_2ImmPred:$off),
1798 (A2_addi (L2_loadri_io I32:$got, imm:$addr), imm:$off)>;
1799 def: Pat<(HexagonAtPcrel I32:$addr),
1800 (C4_addipc imm:$addr)>;
1802 // The HVX load patterns also match AT_PCREL directly. Make sure that
1803 // if the selection of this opcode changes, it's updated in all places.
1806 // --(12) Load -----------------------------------------------------------
1809 def extloadv2i8: PatFrag<(ops node:$ptr), (extload node:$ptr), [{
1810 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1812 def extloadv4i8: PatFrag<(ops node:$ptr), (extload node:$ptr), [{
1813 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1816 def zextloadv2i8: PatFrag<(ops node:$ptr), (zextload node:$ptr), [{
1817 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1819 def zextloadv4i8: PatFrag<(ops node:$ptr), (zextload node:$ptr), [{
1820 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1823 def sextloadv2i8: PatFrag<(ops node:$ptr), (sextload node:$ptr), [{
1824 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1826 def sextloadv4i8: PatFrag<(ops node:$ptr), (sextload node:$ptr), [{
1827 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1830 // Patterns to select load-indexed: Rs + Off.
1831 // - frameindex [+ imm],
1832 multiclass Loadxfi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1834 def: Pat<(VT (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
1835 (VT (MI AddrFI:$fi, imm:$Off))>;
1836 def: Pat<(VT (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
1837 (VT (MI AddrFI:$fi, imm:$Off))>;
1838 def: Pat<(VT (Load AddrFI:$fi)), (VT (MI AddrFI:$fi, 0))>;
1841 // Patterns to select load-indexed: Rs + Off.
1842 // - base reg [+ imm]
1843 multiclass Loadxgi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1845 def: Pat<(VT (Load (add I32:$Rs, ImmPred:$Off))),
1846 (VT (MI IntRegs:$Rs, imm:$Off))>;
1847 def: Pat<(VT (Load (IsOrAdd I32:$Rs, ImmPred:$Off))),
1848 (VT (MI IntRegs:$Rs, imm:$Off))>;
1849 def: Pat<(VT (Load I32:$Rs)), (VT (MI IntRegs:$Rs, 0))>;
1852 // Patterns to select load-indexed: Rs + Off. Combines Loadxfi + Loadxgi.
1853 multiclass Loadxi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1855 defm: Loadxfi_pat<Load, VT, ImmPred, MI>;
1856 defm: Loadxgi_pat<Load, VT, ImmPred, MI>;
1859 // Patterns to select load reg indexed: Rs + Off with a value modifier.
1860 // - frameindex [+ imm]
1861 multiclass Loadxfim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1862 PatLeaf ImmPred, InstHexagon MI> {
1863 def: Pat<(VT (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
1864 (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>;
1865 def: Pat<(VT (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
1866 (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>;
1867 def: Pat<(VT (Load AddrFI:$fi)), (VT (ValueMod (MI AddrFI:$fi, 0)))>;
1870 // Patterns to select load reg indexed: Rs + Off with a value modifier.
1871 // - base reg [+ imm]
1872 multiclass Loadxgim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1873 PatLeaf ImmPred, InstHexagon MI> {
1874 def: Pat<(VT (Load (add I32:$Rs, ImmPred:$Off))),
1875 (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
1876 def: Pat<(VT (Load (IsOrAdd I32:$Rs, ImmPred:$Off))),
1877 (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
1878 def: Pat<(VT (Load I32:$Rs)), (VT (ValueMod (MI IntRegs:$Rs, 0)))>;
1881 // Patterns to select load reg indexed: Rs + Off with a value modifier.
1882 // Combines Loadxfim + Loadxgim.
1883 multiclass Loadxim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1884 PatLeaf ImmPred, InstHexagon MI> {
1885 defm: Loadxfim_pat<Load, VT, ValueMod, ImmPred, MI>;
1886 defm: Loadxgim_pat<Load, VT, ValueMod, ImmPred, MI>;
1889 // Pattern to select load reg reg-indexed: Rs + Rt<<u2.
1890 class Loadxr_shl_pat<PatFrag Load, ValueType VT, InstHexagon MI>
1891 : Pat<(VT (Load (add I32:$Rs, (i32 (shl I32:$Rt, u2_0ImmPred:$u2))))),
1892 (VT (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2))>;
1894 // Pattern to select load reg reg-indexed: Rs + Rt<<0.
1895 class Loadxr_add_pat<PatFrag Load, ValueType VT, InstHexagon MI>
1896 : Pat<(VT (Load (add I32:$Rs, I32:$Rt))),
1897 (VT (MI IntRegs:$Rs, IntRegs:$Rt, 0))>;
1899 // Pattern to select load reg reg-indexed: Rs + Rt<<u2 with value modifier.
1900 class Loadxrm_shl_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1902 : Pat<(VT (Load (add I32:$Rs, (i32 (shl I32:$Rt, u2_0ImmPred:$u2))))),
1903 (VT (ValueMod (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2)))>;
1905 // Pattern to select load reg reg-indexed: Rs + Rt<<0 with value modifier.
1906 class Loadxrm_add_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1908 : Pat<(VT (Load (add I32:$Rs, I32:$Rt))),
1909 (VT (ValueMod (MI IntRegs:$Rs, IntRegs:$Rt, 0)))>;
1911 // Pattern to select load long-offset reg-indexed: Addr + Rt<<u2.
1912 // Don't match for u2==0, instead use reg+imm for those cases.
1913 class Loadxu_pat<PatFrag Load, ValueType VT, PatFrag ImmPred, InstHexagon MI>
1914 : Pat<(VT (Load (add (shl IntRegs:$Rt, u2_0ImmPred:$u2), ImmPred:$Addr))),
1915 (VT (MI IntRegs:$Rt, imm:$u2, ImmPred:$Addr))>;
1917 class Loadxum_pat<PatFrag Load, ValueType VT, PatFrag ImmPred, PatFrag ValueMod,
1919 : Pat<(VT (Load (add (shl IntRegs:$Rt, u2_0ImmPred:$u2), ImmPred:$Addr))),
1920 (VT (ValueMod (MI IntRegs:$Rt, imm:$u2, ImmPred:$Addr)))>;
1922 // Pattern to select load absolute.
1923 class Loada_pat<PatFrag Load, ValueType VT, PatFrag Addr, InstHexagon MI>
1924 : Pat<(VT (Load Addr:$addr)), (MI Addr:$addr)>;
1926 // Pattern to select load absolute with value modifier.
1927 class Loadam_pat<PatFrag Load, ValueType VT, PatFrag Addr, PatFrag ValueMod,
1929 : Pat<(VT (Load Addr:$addr)), (ValueMod (MI Addr:$addr))>;
1932 let AddedComplexity = 20 in {
1933 defm: Loadxi_pat<extloadi1, i32, anyimm0, L2_loadrub_io>;
1934 defm: Loadxi_pat<extloadi8, i32, anyimm0, L2_loadrub_io>;
1935 defm: Loadxi_pat<extloadi16, i32, anyimm1, L2_loadruh_io>;
1936 defm: Loadxi_pat<extloadv2i8, v2i16, anyimm1, L2_loadbzw2_io>;
1937 defm: Loadxi_pat<extloadv4i8, v4i16, anyimm2, L2_loadbzw4_io>;
1938 defm: Loadxi_pat<sextloadi8, i32, anyimm0, L2_loadrb_io>;
1939 defm: Loadxi_pat<sextloadi16, i32, anyimm1, L2_loadrh_io>;
1940 defm: Loadxi_pat<sextloadv2i8, v2i16, anyimm1, L2_loadbsw2_io>;
1941 defm: Loadxi_pat<sextloadv4i8, v4i16, anyimm2, L2_loadbzw4_io>;
1942 defm: Loadxi_pat<zextloadi1, i32, anyimm0, L2_loadrub_io>;
1943 defm: Loadxi_pat<zextloadi8, i32, anyimm0, L2_loadrub_io>;
1944 defm: Loadxi_pat<zextloadi16, i32, anyimm1, L2_loadruh_io>;
1945 defm: Loadxi_pat<zextloadv2i8, v2i16, anyimm1, L2_loadbzw2_io>;
1946 defm: Loadxi_pat<zextloadv4i8, v4i16, anyimm2, L2_loadbzw4_io>;
1947 defm: Loadxi_pat<load, i32, anyimm2, L2_loadri_io>;
1948 defm: Loadxi_pat<load, v2i16, anyimm2, L2_loadri_io>;
1949 defm: Loadxi_pat<load, v4i8, anyimm2, L2_loadri_io>;
1950 defm: Loadxi_pat<load, i64, anyimm3, L2_loadrd_io>;
1951 defm: Loadxi_pat<load, v2i32, anyimm3, L2_loadrd_io>;
1952 defm: Loadxi_pat<load, v4i16, anyimm3, L2_loadrd_io>;
1953 defm: Loadxi_pat<load, v8i8, anyimm3, L2_loadrd_io>;
1954 defm: Loadxi_pat<load, f32, anyimm2, L2_loadri_io>;
1955 defm: Loadxi_pat<load, f64, anyimm3, L2_loadrd_io>;
1958 defm: Loadxi_pat<atomic_load_8 , i32, anyimm0, L2_loadrub_io>;
1959 defm: Loadxi_pat<atomic_load_16, i32, anyimm1, L2_loadruh_io>;
1960 defm: Loadxi_pat<atomic_load_32, i32, anyimm2, L2_loadri_io>;
1961 defm: Loadxi_pat<atomic_load_64, i64, anyimm3, L2_loadrd_io>;
1964 let AddedComplexity = 30 in {
1965 defm: Loadxim_pat<extloadi1, i64, ToAext64, anyimm0, L2_loadrub_io>;
1966 defm: Loadxim_pat<extloadi8, i64, ToAext64, anyimm0, L2_loadrub_io>;
1967 defm: Loadxim_pat<extloadi16, i64, ToAext64, anyimm1, L2_loadruh_io>;
1968 defm: Loadxim_pat<extloadi32, i64, ToAext64, anyimm2, L2_loadri_io>;
1969 defm: Loadxim_pat<zextloadi1, i64, ToZext64, anyimm0, L2_loadrub_io>;
1970 defm: Loadxim_pat<zextloadi8, i64, ToZext64, anyimm0, L2_loadrub_io>;
1971 defm: Loadxim_pat<zextloadi16, i64, ToZext64, anyimm1, L2_loadruh_io>;
1972 defm: Loadxim_pat<zextloadi32, i64, ToZext64, anyimm2, L2_loadri_io>;
1973 defm: Loadxim_pat<sextloadi8, i64, ToSext64, anyimm0, L2_loadrb_io>;
1974 defm: Loadxim_pat<sextloadi16, i64, ToSext64, anyimm1, L2_loadrh_io>;
1975 defm: Loadxim_pat<sextloadi32, i64, ToSext64, anyimm2, L2_loadri_io>;
1978 let AddedComplexity = 60 in {
1979 def: Loadxu_pat<extloadi8, i32, anyimm0, L4_loadrub_ur>;
1980 def: Loadxu_pat<extloadi16, i32, anyimm1, L4_loadruh_ur>;
1981 def: Loadxu_pat<extloadv2i8, v2i16, anyimm1, L4_loadbzw2_ur>;
1982 def: Loadxu_pat<extloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
1983 def: Loadxu_pat<sextloadi8, i32, anyimm0, L4_loadrb_ur>;
1984 def: Loadxu_pat<sextloadi16, i32, anyimm1, L4_loadrh_ur>;
1985 def: Loadxu_pat<sextloadv2i8, v2i16, anyimm1, L4_loadbsw2_ur>;
1986 def: Loadxu_pat<sextloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
1987 def: Loadxu_pat<zextloadi8, i32, anyimm0, L4_loadrub_ur>;
1988 def: Loadxu_pat<zextloadi16, i32, anyimm1, L4_loadruh_ur>;
1989 def: Loadxu_pat<zextloadv2i8, v2i16, anyimm1, L4_loadbzw2_ur>;
1990 def: Loadxu_pat<zextloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
1991 def: Loadxu_pat<load, i32, anyimm2, L4_loadri_ur>;
1992 def: Loadxu_pat<load, v2i16, anyimm2, L4_loadri_ur>;
1993 def: Loadxu_pat<load, v4i8, anyimm2, L4_loadri_ur>;
1994 def: Loadxu_pat<load, i64, anyimm3, L4_loadrd_ur>;
1995 def: Loadxu_pat<load, v2i32, anyimm3, L4_loadrd_ur>;
1996 def: Loadxu_pat<load, v4i16, anyimm3, L4_loadrd_ur>;
1997 def: Loadxu_pat<load, v8i8, anyimm3, L4_loadrd_ur>;
1998 def: Loadxu_pat<load, f32, anyimm2, L4_loadri_ur>;
1999 def: Loadxu_pat<load, f64, anyimm3, L4_loadrd_ur>;
2001 def: Loadxum_pat<sextloadi8, i64, anyimm0, ToSext64, L4_loadrb_ur>;
2002 def: Loadxum_pat<zextloadi8, i64, anyimm0, ToZext64, L4_loadrub_ur>;
2003 def: Loadxum_pat<extloadi8, i64, anyimm0, ToAext64, L4_loadrub_ur>;
2004 def: Loadxum_pat<sextloadi16, i64, anyimm1, ToSext64, L4_loadrh_ur>;
2005 def: Loadxum_pat<zextloadi16, i64, anyimm1, ToZext64, L4_loadruh_ur>;
2006 def: Loadxum_pat<extloadi16, i64, anyimm1, ToAext64, L4_loadruh_ur>;
2007 def: Loadxum_pat<sextloadi32, i64, anyimm2, ToSext64, L4_loadri_ur>;
2008 def: Loadxum_pat<zextloadi32, i64, anyimm2, ToZext64, L4_loadri_ur>;
2009 def: Loadxum_pat<extloadi32, i64, anyimm2, ToAext64, L4_loadri_ur>;
2012 let AddedComplexity = 40 in {
2013 def: Loadxr_shl_pat<extloadi8, i32, L4_loadrub_rr>;
2014 def: Loadxr_shl_pat<zextloadi8, i32, L4_loadrub_rr>;
2015 def: Loadxr_shl_pat<sextloadi8, i32, L4_loadrb_rr>;
2016 def: Loadxr_shl_pat<extloadi16, i32, L4_loadruh_rr>;
2017 def: Loadxr_shl_pat<zextloadi16, i32, L4_loadruh_rr>;
2018 def: Loadxr_shl_pat<sextloadi16, i32, L4_loadrh_rr>;
2019 def: Loadxr_shl_pat<load, i32, L4_loadri_rr>;
2020 def: Loadxr_shl_pat<load, v2i16, L4_loadri_rr>;
2021 def: Loadxr_shl_pat<load, v4i8, L4_loadri_rr>;
2022 def: Loadxr_shl_pat<load, i64, L4_loadrd_rr>;
2023 def: Loadxr_shl_pat<load, v2i32, L4_loadrd_rr>;
2024 def: Loadxr_shl_pat<load, v4i16, L4_loadrd_rr>;
2025 def: Loadxr_shl_pat<load, v8i8, L4_loadrd_rr>;
2026 def: Loadxr_shl_pat<load, f32, L4_loadri_rr>;
2027 def: Loadxr_shl_pat<load, f64, L4_loadrd_rr>;
2030 let AddedComplexity = 20 in {
2031 def: Loadxr_add_pat<extloadi8, i32, L4_loadrub_rr>;
2032 def: Loadxr_add_pat<zextloadi8, i32, L4_loadrub_rr>;
2033 def: Loadxr_add_pat<sextloadi8, i32, L4_loadrb_rr>;
2034 def: Loadxr_add_pat<extloadi16, i32, L4_loadruh_rr>;
2035 def: Loadxr_add_pat<zextloadi16, i32, L4_loadruh_rr>;
2036 def: Loadxr_add_pat<sextloadi16, i32, L4_loadrh_rr>;
2037 def: Loadxr_add_pat<load, i32, L4_loadri_rr>;
2038 def: Loadxr_add_pat<load, v2i16, L4_loadri_rr>;
2039 def: Loadxr_add_pat<load, v4i8, L4_loadri_rr>;
2040 def: Loadxr_add_pat<load, i64, L4_loadrd_rr>;
2041 def: Loadxr_add_pat<load, v2i32, L4_loadrd_rr>;
2042 def: Loadxr_add_pat<load, v4i16, L4_loadrd_rr>;
2043 def: Loadxr_add_pat<load, v8i8, L4_loadrd_rr>;
2044 def: Loadxr_add_pat<load, f32, L4_loadri_rr>;
2045 def: Loadxr_add_pat<load, f64, L4_loadrd_rr>;
2048 let AddedComplexity = 40 in {
2049 def: Loadxrm_shl_pat<extloadi8, i64, ToAext64, L4_loadrub_rr>;
2050 def: Loadxrm_shl_pat<zextloadi8, i64, ToZext64, L4_loadrub_rr>;
2051 def: Loadxrm_shl_pat<sextloadi8, i64, ToSext64, L4_loadrb_rr>;
2052 def: Loadxrm_shl_pat<extloadi16, i64, ToAext64, L4_loadruh_rr>;
2053 def: Loadxrm_shl_pat<zextloadi16, i64, ToZext64, L4_loadruh_rr>;
2054 def: Loadxrm_shl_pat<sextloadi16, i64, ToSext64, L4_loadrh_rr>;
2055 def: Loadxrm_shl_pat<extloadi32, i64, ToAext64, L4_loadri_rr>;
2056 def: Loadxrm_shl_pat<zextloadi32, i64, ToZext64, L4_loadri_rr>;
2057 def: Loadxrm_shl_pat<sextloadi32, i64, ToSext64, L4_loadri_rr>;
2060 let AddedComplexity = 20 in {
2061 def: Loadxrm_add_pat<extloadi8, i64, ToAext64, L4_loadrub_rr>;
2062 def: Loadxrm_add_pat<zextloadi8, i64, ToZext64, L4_loadrub_rr>;
2063 def: Loadxrm_add_pat<sextloadi8, i64, ToSext64, L4_loadrb_rr>;
2064 def: Loadxrm_add_pat<extloadi16, i64, ToAext64, L4_loadruh_rr>;
2065 def: Loadxrm_add_pat<zextloadi16, i64, ToZext64, L4_loadruh_rr>;
2066 def: Loadxrm_add_pat<sextloadi16, i64, ToSext64, L4_loadrh_rr>;
2067 def: Loadxrm_add_pat<extloadi32, i64, ToAext64, L4_loadri_rr>;
2068 def: Loadxrm_add_pat<zextloadi32, i64, ToZext64, L4_loadri_rr>;
2069 def: Loadxrm_add_pat<sextloadi32, i64, ToSext64, L4_loadri_rr>;
2074 let AddedComplexity = 60 in {
2075 def: Loada_pat<zextloadi1, i32, anyimm0, PS_loadrubabs>;
2076 def: Loada_pat<sextloadi8, i32, anyimm0, PS_loadrbabs>;
2077 def: Loada_pat<extloadi8, i32, anyimm0, PS_loadrubabs>;
2078 def: Loada_pat<zextloadi8, i32, anyimm0, PS_loadrubabs>;
2079 def: Loada_pat<sextloadi16, i32, anyimm1, PS_loadrhabs>;
2080 def: Loada_pat<extloadi16, i32, anyimm1, PS_loadruhabs>;
2081 def: Loada_pat<zextloadi16, i32, anyimm1, PS_loadruhabs>;
2082 def: Loada_pat<load, i32, anyimm2, PS_loadriabs>;
2083 def: Loada_pat<load, v2i16, anyimm2, PS_loadriabs>;
2084 def: Loada_pat<load, v4i8, anyimm2, PS_loadriabs>;
2085 def: Loada_pat<load, i64, anyimm3, PS_loadrdabs>;
2086 def: Loada_pat<load, v2i32, anyimm3, PS_loadrdabs>;
2087 def: Loada_pat<load, v4i16, anyimm3, PS_loadrdabs>;
2088 def: Loada_pat<load, v8i8, anyimm3, PS_loadrdabs>;
2089 def: Loada_pat<load, f32, anyimm2, PS_loadriabs>;
2090 def: Loada_pat<load, f64, anyimm3, PS_loadrdabs>;
2092 def: Loada_pat<atomic_load_8, i32, anyimm0, PS_loadrubabs>;
2093 def: Loada_pat<atomic_load_16, i32, anyimm1, PS_loadruhabs>;
2094 def: Loada_pat<atomic_load_32, i32, anyimm2, PS_loadriabs>;
2095 def: Loada_pat<atomic_load_64, i64, anyimm3, PS_loadrdabs>;
2098 let AddedComplexity = 30 in {
2099 def: Loadam_pat<extloadi8, i64, anyimm0, ToAext64, PS_loadrubabs>;
2100 def: Loadam_pat<sextloadi8, i64, anyimm0, ToSext64, PS_loadrbabs>;
2101 def: Loadam_pat<zextloadi8, i64, anyimm0, ToZext64, PS_loadrubabs>;
2102 def: Loadam_pat<extloadi16, i64, anyimm1, ToAext64, PS_loadruhabs>;
2103 def: Loadam_pat<sextloadi16, i64, anyimm1, ToSext64, PS_loadrhabs>;
2104 def: Loadam_pat<zextloadi16, i64, anyimm1, ToZext64, PS_loadruhabs>;
2105 def: Loadam_pat<extloadi32, i64, anyimm2, ToAext64, PS_loadriabs>;
2106 def: Loadam_pat<sextloadi32, i64, anyimm2, ToSext64, PS_loadriabs>;
2107 def: Loadam_pat<zextloadi32, i64, anyimm2, ToZext64, PS_loadriabs>;
2109 def: Loadam_pat<load, i1, anyimm0, I32toI1, PS_loadrubabs>;
2110 def: Loadam_pat<zextloadi1, i64, anyimm0, ToZext64, PS_loadrubabs>;
2113 // GP-relative address
2115 let AddedComplexity = 100 in {
2116 def: Loada_pat<extloadi1, i32, addrgp, L2_loadrubgp>;
2117 def: Loada_pat<zextloadi1, i32, addrgp, L2_loadrubgp>;
2118 def: Loada_pat<extloadi8, i32, addrgp, L2_loadrubgp>;
2119 def: Loada_pat<sextloadi8, i32, addrgp, L2_loadrbgp>;
2120 def: Loada_pat<zextloadi8, i32, addrgp, L2_loadrubgp>;
2121 def: Loada_pat<extloadi16, i32, addrgp, L2_loadruhgp>;
2122 def: Loada_pat<sextloadi16, i32, addrgp, L2_loadrhgp>;
2123 def: Loada_pat<zextloadi16, i32, addrgp, L2_loadruhgp>;
2124 def: Loada_pat<load, i32, addrgp, L2_loadrigp>;
2125 def: Loada_pat<load, v2i16, addrgp, L2_loadrigp>;
2126 def: Loada_pat<load, v4i8, addrgp, L2_loadrigp>;
2127 def: Loada_pat<load, i64, addrgp, L2_loadrdgp>;
2128 def: Loada_pat<load, v2i32, addrgp, L2_loadrdgp>;
2129 def: Loada_pat<load, v4i16, addrgp, L2_loadrdgp>;
2130 def: Loada_pat<load, v8i8, addrgp, L2_loadrdgp>;
2131 def: Loada_pat<load, f32, addrgp, L2_loadrigp>;
2132 def: Loada_pat<load, f64, addrgp, L2_loadrdgp>;
2134 def: Loada_pat<atomic_load_8, i32, addrgp, L2_loadrubgp>;
2135 def: Loada_pat<atomic_load_16, i32, addrgp, L2_loadruhgp>;
2136 def: Loada_pat<atomic_load_32, i32, addrgp, L2_loadrigp>;
2137 def: Loada_pat<atomic_load_64, i64, addrgp, L2_loadrdgp>;
2140 let AddedComplexity = 70 in {
2141 def: Loadam_pat<extloadi8, i64, addrgp, ToAext64, L2_loadrubgp>;
2142 def: Loadam_pat<sextloadi8, i64, addrgp, ToSext64, L2_loadrbgp>;
2143 def: Loadam_pat<zextloadi8, i64, addrgp, ToZext64, L2_loadrubgp>;
2144 def: Loadam_pat<extloadi16, i64, addrgp, ToAext64, L2_loadruhgp>;
2145 def: Loadam_pat<sextloadi16, i64, addrgp, ToSext64, L2_loadrhgp>;
2146 def: Loadam_pat<zextloadi16, i64, addrgp, ToZext64, L2_loadruhgp>;
2147 def: Loadam_pat<extloadi32, i64, addrgp, ToAext64, L2_loadrigp>;
2148 def: Loadam_pat<sextloadi32, i64, addrgp, ToSext64, L2_loadrigp>;
2149 def: Loadam_pat<zextloadi32, i64, addrgp, ToZext64, L2_loadrigp>;
2151 def: Loadam_pat<load, i1, addrgp, I32toI1, L2_loadrubgp>;
2152 def: Loadam_pat<zextloadi1, i64, addrgp, ToZext64, L2_loadrubgp>;
2156 // Sign-extending loads of i1 need to replicate the lowest bit throughout
2157 // the 32-bit value. Since the loaded value can only be 0 or 1, 0-v should
2159 let AddedComplexity = 20 in
2160 def: Pat<(i32 (sextloadi1 I32:$Rs)),
2161 (A2_subri 0, (L2_loadrub_io IntRegs:$Rs, 0))>;
2163 // Patterns for loads of i1:
2164 def: Pat<(i1 (load AddrFI:$fi)),
2165 (C2_tfrrp (L2_loadrub_io AddrFI:$fi, 0))>;
2166 def: Pat<(i1 (load (add I32:$Rs, anyimm0:$Off))),
2167 (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, imm:$Off))>;
2168 def: Pat<(i1 (load I32:$Rs)),
2169 (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, 0))>;
2172 // --(13) Store ----------------------------------------------------------
2175 class Storepi_pat<PatFrag Store, PatFrag Value, PatFrag Offset, InstHexagon MI>
2176 : Pat<(Store Value:$Rt, I32:$Rx, Offset:$s4),
2177 (MI I32:$Rx, imm:$s4, Value:$Rt)>;
2179 def: Storepi_pat<post_truncsti8, I32, s4_0ImmPred, S2_storerb_pi>;
2180 def: Storepi_pat<post_truncsti16, I32, s4_1ImmPred, S2_storerh_pi>;
2181 def: Storepi_pat<post_store, I32, s4_2ImmPred, S2_storeri_pi>;
2182 def: Storepi_pat<post_store, I64, s4_3ImmPred, S2_storerd_pi>;
2184 // Patterns for generating stores, where the address takes different forms:
2186 // - frameindex + offset,
2188 // - simple (base address without offset).
2189 // These would usually be used together (via Storexi_pat defined below), but
2190 // in some cases one may want to apply different properties (such as
2191 // AddedComplexity) to the individual patterns.
2192 class Storexi_fi_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2193 : Pat<(Store Value:$Rs, AddrFI:$fi), (MI AddrFI:$fi, 0, Value:$Rs)>;
2195 multiclass Storexi_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2197 def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
2198 (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
2199 def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
2200 (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
2203 multiclass Storexi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2205 def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
2206 (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
2207 def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
2208 (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
2211 class Storexi_base_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2212 : Pat<(Store Value:$Rt, I32:$Rs),
2213 (MI IntRegs:$Rs, 0, Value:$Rt)>;
2215 // Patterns for generating stores, where the address takes different forms,
2216 // and where the value being stored is transformed through the value modifier
2217 // ValueMod. The address forms are same as above.
2218 class Storexim_fi_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2220 : Pat<(Store Value:$Rs, AddrFI:$fi),
2221 (MI AddrFI:$fi, 0, (ValueMod Value:$Rs))>;
2223 multiclass Storexim_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2224 PatFrag ValueMod, InstHexagon MI> {
2225 def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
2226 (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
2227 def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
2228 (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
2231 multiclass Storexim_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2232 PatFrag ValueMod, InstHexagon MI> {
2233 def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
2234 (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
2235 def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
2236 (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
2239 class Storexim_base_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2241 : Pat<(Store Value:$Rt, I32:$Rs),
2242 (MI IntRegs:$Rs, 0, (ValueMod Value:$Rt))>;
2244 multiclass Storexi_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
2246 defm: Storexi_fi_add_pat <Store, Value, ImmPred, MI>;
2247 def: Storexi_fi_pat <Store, Value, MI>;
2248 defm: Storexi_add_pat <Store, Value, ImmPred, MI>;
2251 multiclass Storexim_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
2252 PatFrag ValueMod, InstHexagon MI> {
2253 defm: Storexim_fi_add_pat <Store, Value, ImmPred, ValueMod, MI>;
2254 def: Storexim_fi_pat <Store, Value, ValueMod, MI>;
2255 defm: Storexim_add_pat <Store, Value, ImmPred, ValueMod, MI>;
2259 class Storexu_shl_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred, InstHexagon MI>
2260 : Pat<(Store Value:$Rt, (add (shl I32:$Ru, u2_0ImmPred:$u2), ImmPred:$A)),
2261 (MI IntRegs:$Ru, imm:$u2, ImmPred:$A, Value:$Rt)>;
2264 class Storexr_shl_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2265 : Pat<(Store Value:$Ru, (add I32:$Rs, (shl I32:$Rt, u2_0ImmPred:$u2))),
2266 (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2, Value:$Ru)>;
2269 class Storexr_add_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2270 : Pat<(Store Value:$Ru, (add I32:$Rs, I32:$Rt)),
2271 (MI IntRegs:$Rs, IntRegs:$Rt, 0, Value:$Ru)>;
2273 class Storea_pat<PatFrag Store, PatFrag Value, PatFrag Addr, InstHexagon MI>
2274 : Pat<(Store Value:$val, Addr:$addr), (MI Addr:$addr, Value:$val)>;
2276 class Stoream_pat<PatFrag Store, PatFrag Value, PatFrag Addr, PatFrag ValueMod,
2278 : Pat<(Store Value:$val, Addr:$addr),
2279 (MI Addr:$addr, (ValueMod Value:$val))>;
2281 // Regular stores in the DAG have two operands: value and address.
2282 // Atomic stores also have two, but they are reversed: address, value.
2283 // To use atomic stores with the patterns, they need to have their operands
2284 // swapped. This relies on the knowledge that the F.Fragment uses names
2286 class AtomSt<PatFrag F>
2287 : PatFrag<(ops node:$val, node:$ptr), !head(F.Fragments), F.PredicateCode,
2288 F.OperandTransform> {
2289 let IsAtomic = F.IsAtomic;
2290 let MemoryVT = F.MemoryVT;
2294 def IMM_BYTE : SDNodeXForm<imm, [{
2295 // -1 can be represented as 255, etc.
2296 // assigning to a byte restores our desired signed value.
2297 int8_t imm = N->getSExtValue();
2298 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2301 def IMM_HALF : SDNodeXForm<imm, [{
2302 // -1 can be represented as 65535, etc.
2303 // assigning to a short restores our desired signed value.
2304 int16_t imm = N->getSExtValue();
2305 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2308 def IMM_WORD : SDNodeXForm<imm, [{
2309 // -1 can be represented as 4294967295, etc.
2310 // Currently, it's not doing this. But some optimization
2311 // might convert -1 to a large +ve number.
2312 // assigning to a word restores our desired signed value.
2313 int32_t imm = N->getSExtValue();
2314 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2317 def ToImmByte : OutPatFrag<(ops node:$R), (IMM_BYTE $R)>;
2318 def ToImmHalf : OutPatFrag<(ops node:$R), (IMM_HALF $R)>;
2319 def ToImmWord : OutPatFrag<(ops node:$R), (IMM_WORD $R)>;
2321 // Even though the offset is not extendable in the store-immediate, we
2322 // can still generate the fi# in the base address. If the final offset
2323 // is not valid for the instruction, we will replace it with a scratch
2325 class SmallStackStore<PatFrag Store>
2326 : PatFrag<(ops node:$Val, node:$Addr), (Store node:$Val, node:$Addr), [{
2327 return isSmallStackStore(cast<StoreSDNode>(N));
2330 // This is the complement of SmallStackStore.
2331 class LargeStackStore<PatFrag Store>
2332 : PatFrag<(ops node:$Val, node:$Addr), (Store node:$Val, node:$Addr), [{
2333 return !isSmallStackStore(cast<StoreSDNode>(N));
2336 // Preferred addressing modes for various combinations of stored value
2337 // and address computation.
2338 // For stores where the address and value are both immediates, prefer
2339 // store-immediate. The reason is that the constant-extender optimization
2340 // can replace store-immediate with a store-register, but there is nothing
2341 // to generate a store-immediate out of a store-register.
2343 // C R F F+C R+C R+R R<<S+C R<<S+R
2344 // --+-------+-----+-----+------+-----+-----+--------+--------
2345 // C | imm | imm | imm | imm | imm | rr | ur | rr
2346 // R | abs* | io | io | io | io | rr | ur | rr
2348 // (*) Absolute or GP-relative.
2350 // Note that any expression can be matched by Reg. In particular, an immediate
2351 // can always be placed in a register, so patterns checking for Imm should
2352 // have a higher priority than the ones involving Reg that could also match.
2353 // For example, *(p+4) could become r1=#4; memw(r0+r1<<#0) instead of the
2354 // preferred memw(r0+#4). Similarly Reg+Imm or Reg+Reg should be tried before
2357 // The order in which the different combinations are tried:
2359 // C F R F+C R+C R+R R<<S+C R<<S+R
2360 // --+-------+-----+-----+------+-----+-----+--------+--------
2361 // C | 1 | 6 | - | 5 | 9 | - | - | -
2362 // R | 2 | 8 | 12 | 7 | 10 | 11 | 3 | 4
2365 // First, match the unusual case of doubleword store into Reg+Imm4, i.e.
2366 // a store where the offset Imm4 is a multiple of 4, but not of 8. This
2367 // implies that Reg is also a proper multiple of 4. To still generate a
2368 // doubleword store, add 4 to Reg, and subtract 4 from the offset.
2370 def s30_2ProperPred : PatLeaf<(i32 imm), [{
2371 int64_t v = (int64_t)N->getSExtValue();
2372 return isShiftedInt<30,2>(v) && !isShiftedInt<29,3>(v);
2374 def RoundTo8 : SDNodeXForm<imm, [{
2375 int32_t Imm = N->getSExtValue();
2376 return CurDAG->getTargetConstant(Imm & -8, SDLoc(N), MVT::i32);
2379 let AddedComplexity = 150 in
2380 def: Pat<(store I64:$Ru, (add I32:$Rs, s30_2ProperPred:$Off)),
2381 (S2_storerd_io (A2_addi I32:$Rs, 4), (RoundTo8 $Off), I64:$Ru)>;
2383 class Storexi_abs_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2384 : Pat<(Store Value:$val, anyimm:$addr),
2385 (MI (ToI32 $addr), 0, Value:$val)>;
2386 class Storexim_abs_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2388 : Pat<(Store Value:$val, anyimm:$addr),
2389 (MI (ToI32 $addr), 0, (ValueMod Value:$val))>;
2391 let AddedComplexity = 140 in {
2392 def: Storexim_abs_pat<truncstorei8, anyint, ToImmByte, S4_storeirb_io>;
2393 def: Storexim_abs_pat<truncstorei16, anyint, ToImmHalf, S4_storeirh_io>;
2394 def: Storexim_abs_pat<store, anyint, ToImmWord, S4_storeiri_io>;
2396 def: Storexi_abs_pat<truncstorei8, anyimm, S4_storeirb_io>;
2397 def: Storexi_abs_pat<truncstorei16, anyimm, S4_storeirh_io>;
2398 def: Storexi_abs_pat<store, anyimm, S4_storeiri_io>;
2401 // GP-relative address
2402 let AddedComplexity = 120 in {
2403 def: Storea_pat<truncstorei8, I32, addrgp, S2_storerbgp>;
2404 def: Storea_pat<truncstorei16, I32, addrgp, S2_storerhgp>;
2405 def: Storea_pat<store, I32, addrgp, S2_storerigp>;
2406 def: Storea_pat<store, V4I8, addrgp, S2_storerigp>;
2407 def: Storea_pat<store, V2I16, addrgp, S2_storerigp>;
2408 def: Storea_pat<store, I64, addrgp, S2_storerdgp>;
2409 def: Storea_pat<store, V8I8, addrgp, S2_storerdgp>;
2410 def: Storea_pat<store, V4I16, addrgp, S2_storerdgp>;
2411 def: Storea_pat<store, V2I32, addrgp, S2_storerdgp>;
2412 def: Storea_pat<store, F32, addrgp, S2_storerigp>;
2413 def: Storea_pat<store, F64, addrgp, S2_storerdgp>;
2414 def: Storea_pat<AtomSt<atomic_store_8>, I32, addrgp, S2_storerbgp>;
2415 def: Storea_pat<AtomSt<atomic_store_16>, I32, addrgp, S2_storerhgp>;
2416 def: Storea_pat<AtomSt<atomic_store_32>, I32, addrgp, S2_storerigp>;
2417 def: Storea_pat<AtomSt<atomic_store_32>, V4I8, addrgp, S2_storerigp>;
2418 def: Storea_pat<AtomSt<atomic_store_32>, V2I16, addrgp, S2_storerigp>;
2419 def: Storea_pat<AtomSt<atomic_store_64>, I64, addrgp, S2_storerdgp>;
2420 def: Storea_pat<AtomSt<atomic_store_64>, V8I8, addrgp, S2_storerdgp>;
2421 def: Storea_pat<AtomSt<atomic_store_64>, V4I16, addrgp, S2_storerdgp>;
2422 def: Storea_pat<AtomSt<atomic_store_64>, V2I32, addrgp, S2_storerdgp>;
2424 def: Stoream_pat<truncstorei8, I64, addrgp, LoReg, S2_storerbgp>;
2425 def: Stoream_pat<truncstorei16, I64, addrgp, LoReg, S2_storerhgp>;
2426 def: Stoream_pat<truncstorei32, I64, addrgp, LoReg, S2_storerigp>;
2427 def: Stoream_pat<store, I1, addrgp, I1toI32, S2_storerbgp>;
2431 let AddedComplexity = 110 in {
2432 def: Storea_pat<truncstorei8, I32, anyimm0, PS_storerbabs>;
2433 def: Storea_pat<truncstorei16, I32, anyimm1, PS_storerhabs>;
2434 def: Storea_pat<store, I32, anyimm2, PS_storeriabs>;
2435 def: Storea_pat<store, V4I8, anyimm2, PS_storeriabs>;
2436 def: Storea_pat<store, V2I16, anyimm2, PS_storeriabs>;
2437 def: Storea_pat<store, I64, anyimm3, PS_storerdabs>;
2438 def: Storea_pat<store, V8I8, anyimm3, PS_storerdabs>;
2439 def: Storea_pat<store, V4I16, anyimm3, PS_storerdabs>;
2440 def: Storea_pat<store, V2I32, anyimm3, PS_storerdabs>;
2441 def: Storea_pat<store, F32, anyimm2, PS_storeriabs>;
2442 def: Storea_pat<store, F64, anyimm3, PS_storerdabs>;
2443 def: Storea_pat<AtomSt<atomic_store_8>, I32, anyimm0, PS_storerbabs>;
2444 def: Storea_pat<AtomSt<atomic_store_16>, I32, anyimm1, PS_storerhabs>;
2445 def: Storea_pat<AtomSt<atomic_store_32>, I32, anyimm2, PS_storeriabs>;
2446 def: Storea_pat<AtomSt<atomic_store_32>, V4I8, anyimm2, PS_storeriabs>;
2447 def: Storea_pat<AtomSt<atomic_store_32>, V2I16, anyimm2, PS_storeriabs>;
2448 def: Storea_pat<AtomSt<atomic_store_64>, I64, anyimm3, PS_storerdabs>;
2449 def: Storea_pat<AtomSt<atomic_store_64>, V8I8, anyimm3, PS_storerdabs>;
2450 def: Storea_pat<AtomSt<atomic_store_64>, V4I16, anyimm3, PS_storerdabs>;
2451 def: Storea_pat<AtomSt<atomic_store_64>, V2I32, anyimm3, PS_storerdabs>;
2453 def: Stoream_pat<truncstorei8, I64, anyimm0, LoReg, PS_storerbabs>;
2454 def: Stoream_pat<truncstorei16, I64, anyimm1, LoReg, PS_storerhabs>;
2455 def: Stoream_pat<truncstorei32, I64, anyimm2, LoReg, PS_storeriabs>;
2456 def: Stoream_pat<store, I1, anyimm0, I1toI32, PS_storerbabs>;
2460 let AddedComplexity = 100 in {
2461 def: Storexu_shl_pat<truncstorei8, I32, anyimm0, S4_storerb_ur>;
2462 def: Storexu_shl_pat<truncstorei16, I32, anyimm1, S4_storerh_ur>;
2463 def: Storexu_shl_pat<store, I32, anyimm2, S4_storeri_ur>;
2464 def: Storexu_shl_pat<store, V4I8, anyimm2, S4_storeri_ur>;
2465 def: Storexu_shl_pat<store, V2I16, anyimm2, S4_storeri_ur>;
2466 def: Storexu_shl_pat<store, I64, anyimm3, S4_storerd_ur>;
2467 def: Storexu_shl_pat<store, V8I8, anyimm3, S4_storerd_ur>;
2468 def: Storexu_shl_pat<store, V4I16, anyimm3, S4_storerd_ur>;
2469 def: Storexu_shl_pat<store, V2I32, anyimm3, S4_storerd_ur>;
2470 def: Storexu_shl_pat<store, F32, anyimm2, S4_storeri_ur>;
2471 def: Storexu_shl_pat<store, F64, anyimm3, S4_storerd_ur>;
2473 def: Pat<(store I1:$Pu, (add (shl I32:$Rs, u2_0ImmPred:$u2), anyimm:$A)),
2474 (S4_storerb_ur IntRegs:$Rs, imm:$u2, imm:$A, (I1toI32 I1:$Pu))>;
2478 let AddedComplexity = 90 in {
2479 def: Storexr_shl_pat<truncstorei8, I32, S4_storerb_rr>;
2480 def: Storexr_shl_pat<truncstorei16, I32, S4_storerh_rr>;
2481 def: Storexr_shl_pat<store, I32, S4_storeri_rr>;
2482 def: Storexr_shl_pat<store, V4I8, S4_storeri_rr>;
2483 def: Storexr_shl_pat<store, V2I16, S4_storeri_rr>;
2484 def: Storexr_shl_pat<store, I64, S4_storerd_rr>;
2485 def: Storexr_shl_pat<store, V8I8, S4_storerd_rr>;
2486 def: Storexr_shl_pat<store, V4I16, S4_storerd_rr>;
2487 def: Storexr_shl_pat<store, V2I32, S4_storerd_rr>;
2488 def: Storexr_shl_pat<store, F32, S4_storeri_rr>;
2489 def: Storexr_shl_pat<store, F64, S4_storerd_rr>;
2491 def: Pat<(store I1:$Pu, (add (shl I32:$Rs, u2_0ImmPred:$u2), I32:$Rt)),
2492 (S4_storerb_ur IntRegs:$Rt, IntRegs:$Rs, imm:$u2, (I1toI32 I1:$Pu))>;
2495 class SS_<PatFrag F> : SmallStackStore<F>;
2496 class LS_<PatFrag F> : LargeStackStore<F>;
2498 multiclass IMFA_<PatFrag S, PatFrag V, PatFrag O, PatFrag M, InstHexagon I> {
2499 defm: Storexim_fi_add_pat<S, V, O, M, I>;
2501 multiclass IFA_<PatFrag S, PatFrag V, PatFrag O, InstHexagon I> {
2502 defm: Storexi_fi_add_pat<S, V, O, I>;
2505 // Fi+Imm, store-immediate
2506 let AddedComplexity = 80 in {
2507 defm: IMFA_<SS_<truncstorei8>, anyint, u6_0ImmPred, ToImmByte, S4_storeirb_io>;
2508 defm: IMFA_<SS_<truncstorei16>, anyint, u6_1ImmPred, ToImmHalf, S4_storeirh_io>;
2509 defm: IMFA_<SS_<store>, anyint, u6_2ImmPred, ToImmWord, S4_storeiri_io>;
2511 defm: IFA_<SS_<truncstorei8>, anyimm, u6_0ImmPred, S4_storeirb_io>;
2512 defm: IFA_<SS_<truncstorei16>, anyimm, u6_1ImmPred, S4_storeirh_io>;
2513 defm: IFA_<SS_<store>, anyimm, u6_2ImmPred, S4_storeiri_io>;
2515 // For large-stack stores, generate store-register (prefer explicit Fi
2517 defm: IMFA_<LS_<truncstorei8>, anyimm, u6_0ImmPred, ToI32, S2_storerb_io>;
2518 defm: IMFA_<LS_<truncstorei16>, anyimm, u6_1ImmPred, ToI32, S2_storerh_io>;
2519 defm: IMFA_<LS_<store>, anyimm, u6_2ImmPred, ToI32, S2_storeri_io>;
2522 // Fi, store-immediate
2523 let AddedComplexity = 70 in {
2524 def: Storexim_fi_pat<SS_<truncstorei8>, anyint, ToImmByte, S4_storeirb_io>;
2525 def: Storexim_fi_pat<SS_<truncstorei16>, anyint, ToImmHalf, S4_storeirh_io>;
2526 def: Storexim_fi_pat<SS_<store>, anyint, ToImmWord, S4_storeiri_io>;
2528 def: Storexi_fi_pat<SS_<truncstorei8>, anyimm, S4_storeirb_io>;
2529 def: Storexi_fi_pat<SS_<truncstorei16>, anyimm, S4_storeirh_io>;
2530 def: Storexi_fi_pat<SS_<store>, anyimm, S4_storeiri_io>;
2532 // For large-stack stores, generate store-register (prefer explicit Fi
2534 def: Storexim_fi_pat<LS_<truncstorei8>, anyimm, ToI32, S2_storerb_io>;
2535 def: Storexim_fi_pat<LS_<truncstorei16>, anyimm, ToI32, S2_storerh_io>;
2536 def: Storexim_fi_pat<LS_<store>, anyimm, ToI32, S2_storeri_io>;
2539 // Fi+Imm, Fi, store-register
2540 let AddedComplexity = 60 in {
2541 defm: Storexi_fi_add_pat<truncstorei8, I32, anyimm, S2_storerb_io>;
2542 defm: Storexi_fi_add_pat<truncstorei16, I32, anyimm, S2_storerh_io>;
2543 defm: Storexi_fi_add_pat<store, I32, anyimm, S2_storeri_io>;
2544 defm: Storexi_fi_add_pat<store, V4I8, anyimm, S2_storeri_io>;
2545 defm: Storexi_fi_add_pat<store, V2I16, anyimm, S2_storeri_io>;
2546 defm: Storexi_fi_add_pat<store, I64, anyimm, S2_storerd_io>;
2547 defm: Storexi_fi_add_pat<store, V8I8, anyimm, S2_storerd_io>;
2548 defm: Storexi_fi_add_pat<store, V4I16, anyimm, S2_storerd_io>;
2549 defm: Storexi_fi_add_pat<store, V2I32, anyimm, S2_storerd_io>;
2550 defm: Storexi_fi_add_pat<store, F32, anyimm, S2_storeri_io>;
2551 defm: Storexi_fi_add_pat<store, F64, anyimm, S2_storerd_io>;
2552 defm: Storexim_fi_add_pat<store, I1, anyimm, I1toI32, S2_storerb_io>;
2554 def: Storexi_fi_pat<truncstorei8, I32, S2_storerb_io>;
2555 def: Storexi_fi_pat<truncstorei16, I32, S2_storerh_io>;
2556 def: Storexi_fi_pat<store, I32, S2_storeri_io>;
2557 def: Storexi_fi_pat<store, V4I8, S2_storeri_io>;
2558 def: Storexi_fi_pat<store, V2I16, S2_storeri_io>;
2559 def: Storexi_fi_pat<store, I64, S2_storerd_io>;
2560 def: Storexi_fi_pat<store, V8I8, S2_storerd_io>;
2561 def: Storexi_fi_pat<store, V4I16, S2_storerd_io>;
2562 def: Storexi_fi_pat<store, V2I32, S2_storerd_io>;
2563 def: Storexi_fi_pat<store, F32, S2_storeri_io>;
2564 def: Storexi_fi_pat<store, F64, S2_storerd_io>;
2565 def: Storexim_fi_pat<store, I1, I1toI32, S2_storerb_io>;
2569 multiclass IMRA_<PatFrag S, PatFrag V, PatFrag O, PatFrag M, InstHexagon I> {
2570 defm: Storexim_add_pat<S, V, O, M, I>;
2572 multiclass IRA_<PatFrag S, PatFrag V, PatFrag O, InstHexagon I> {
2573 defm: Storexi_add_pat<S, V, O, I>;
2576 // Reg+Imm, store-immediate
2577 let AddedComplexity = 50 in {
2578 defm: IMRA_<truncstorei8, anyint, u6_0ImmPred, ToImmByte, S4_storeirb_io>;
2579 defm: IMRA_<truncstorei16, anyint, u6_1ImmPred, ToImmHalf, S4_storeirh_io>;
2580 defm: IMRA_<store, anyint, u6_2ImmPred, ToImmWord, S4_storeiri_io>;
2582 defm: IRA_<truncstorei8, anyimm, u6_0ImmPred, S4_storeirb_io>;
2583 defm: IRA_<truncstorei16, anyimm, u6_1ImmPred, S4_storeirh_io>;
2584 defm: IRA_<store, anyimm, u6_2ImmPred, S4_storeiri_io>;
2587 // Reg+Imm, store-register
2588 let AddedComplexity = 40 in {
2589 defm: Storexi_pat<truncstorei8, I32, anyimm0, S2_storerb_io>;
2590 defm: Storexi_pat<truncstorei16, I32, anyimm1, S2_storerh_io>;
2591 defm: Storexi_pat<store, I32, anyimm2, S2_storeri_io>;
2592 defm: Storexi_pat<store, V4I8, anyimm2, S2_storeri_io>;
2593 defm: Storexi_pat<store, V2I16, anyimm2, S2_storeri_io>;
2594 defm: Storexi_pat<store, I64, anyimm3, S2_storerd_io>;
2595 defm: Storexi_pat<store, V8I8, anyimm3, S2_storerd_io>;
2596 defm: Storexi_pat<store, V4I16, anyimm3, S2_storerd_io>;
2597 defm: Storexi_pat<store, V2I32, anyimm3, S2_storerd_io>;
2598 defm: Storexi_pat<store, F32, anyimm2, S2_storeri_io>;
2599 defm: Storexi_pat<store, F64, anyimm3, S2_storerd_io>;
2601 defm: Storexim_pat<truncstorei8, I64, anyimm0, LoReg, S2_storerb_io>;
2602 defm: Storexim_pat<truncstorei16, I64, anyimm1, LoReg, S2_storerh_io>;
2603 defm: Storexim_pat<truncstorei32, I64, anyimm2, LoReg, S2_storeri_io>;
2604 defm: Storexim_pat<store, I1, anyimm0, I1toI32, S2_storerb_io>;
2606 defm: Storexi_pat<AtomSt<atomic_store_8>, I32, anyimm0, S2_storerb_io>;
2607 defm: Storexi_pat<AtomSt<atomic_store_16>, I32, anyimm1, S2_storerh_io>;
2608 defm: Storexi_pat<AtomSt<atomic_store_32>, I32, anyimm2, S2_storeri_io>;
2609 defm: Storexi_pat<AtomSt<atomic_store_32>, V4I8, anyimm2, S2_storeri_io>;
2610 defm: Storexi_pat<AtomSt<atomic_store_32>, V2I16, anyimm2, S2_storeri_io>;
2611 defm: Storexi_pat<AtomSt<atomic_store_64>, I64, anyimm3, S2_storerd_io>;
2612 defm: Storexi_pat<AtomSt<atomic_store_64>, V8I8, anyimm3, S2_storerd_io>;
2613 defm: Storexi_pat<AtomSt<atomic_store_64>, V4I16, anyimm3, S2_storerd_io>;
2614 defm: Storexi_pat<AtomSt<atomic_store_64>, V2I32, anyimm3, S2_storerd_io>;
2618 let AddedComplexity = 30 in {
2619 def: Storexr_add_pat<truncstorei8, I32, S4_storerb_rr>;
2620 def: Storexr_add_pat<truncstorei16, I32, S4_storerh_rr>;
2621 def: Storexr_add_pat<store, I32, S4_storeri_rr>;
2622 def: Storexr_add_pat<store, V4I8, S4_storeri_rr>;
2623 def: Storexr_add_pat<store, V2I16, S4_storeri_rr>;
2624 def: Storexr_add_pat<store, I64, S4_storerd_rr>;
2625 def: Storexr_add_pat<store, V8I8, S4_storerd_rr>;
2626 def: Storexr_add_pat<store, V4I16, S4_storerd_rr>;
2627 def: Storexr_add_pat<store, V2I32, S4_storerd_rr>;
2628 def: Storexr_add_pat<store, F32, S4_storeri_rr>;
2629 def: Storexr_add_pat<store, F64, S4_storerd_rr>;
2631 def: Pat<(store I1:$Pu, (add I32:$Rs, I32:$Rt)),
2632 (S4_storerb_rr IntRegs:$Rs, IntRegs:$Rt, 0, (I1toI32 I1:$Pu))>;
2635 // Reg, store-immediate
2636 let AddedComplexity = 20 in {
2637 def: Storexim_base_pat<truncstorei8, anyint, ToImmByte, S4_storeirb_io>;
2638 def: Storexim_base_pat<truncstorei16, anyint, ToImmHalf, S4_storeirh_io>;
2639 def: Storexim_base_pat<store, anyint, ToImmWord, S4_storeiri_io>;
2641 def: Storexi_base_pat<truncstorei8, anyimm, S4_storeirb_io>;
2642 def: Storexi_base_pat<truncstorei16, anyimm, S4_storeirh_io>;
2643 def: Storexi_base_pat<store, anyimm, S4_storeiri_io>;
2646 // Reg, store-register
2647 let AddedComplexity = 10 in {
2648 def: Storexi_base_pat<truncstorei8, I32, S2_storerb_io>;
2649 def: Storexi_base_pat<truncstorei16, I32, S2_storerh_io>;
2650 def: Storexi_base_pat<store, I32, S2_storeri_io>;
2651 def: Storexi_base_pat<store, V4I8, S2_storeri_io>;
2652 def: Storexi_base_pat<store, V2I16, S2_storeri_io>;
2653 def: Storexi_base_pat<store, I64, S2_storerd_io>;
2654 def: Storexi_base_pat<store, V8I8, S2_storerd_io>;
2655 def: Storexi_base_pat<store, V4I16, S2_storerd_io>;
2656 def: Storexi_base_pat<store, V2I32, S2_storerd_io>;
2657 def: Storexi_base_pat<store, F32, S2_storeri_io>;
2658 def: Storexi_base_pat<store, F64, S2_storerd_io>;
2660 def: Storexim_base_pat<truncstorei8, I64, LoReg, S2_storerb_io>;
2661 def: Storexim_base_pat<truncstorei16, I64, LoReg, S2_storerh_io>;
2662 def: Storexim_base_pat<truncstorei32, I64, LoReg, S2_storeri_io>;
2663 def: Storexim_base_pat<store, I1, I1toI32, S2_storerb_io>;
2665 def: Storexi_base_pat<AtomSt<atomic_store_8>, I32, S2_storerb_io>;
2666 def: Storexi_base_pat<AtomSt<atomic_store_16>, I32, S2_storerh_io>;
2667 def: Storexi_base_pat<AtomSt<atomic_store_32>, I32, S2_storeri_io>;
2668 def: Storexi_base_pat<AtomSt<atomic_store_32>, V4I8, S2_storeri_io>;
2669 def: Storexi_base_pat<AtomSt<atomic_store_32>, V2I16, S2_storeri_io>;
2670 def: Storexi_base_pat<AtomSt<atomic_store_64>, I64, S2_storerd_io>;
2671 def: Storexi_base_pat<AtomSt<atomic_store_64>, V8I8, S2_storerd_io>;
2672 def: Storexi_base_pat<AtomSt<atomic_store_64>, V4I16, S2_storerd_io>;
2673 def: Storexi_base_pat<AtomSt<atomic_store_64>, V2I32, S2_storerd_io>;
2677 // --(14) Memop ----------------------------------------------------------
2680 def m5_0Imm8Pred : PatLeaf<(i32 imm), [{
2681 int8_t V = N->getSExtValue();
2682 return -32 < V && V <= -1;
2685 def m5_0Imm16Pred : PatLeaf<(i32 imm), [{
2686 int16_t V = N->getSExtValue();
2687 return -32 < V && V <= -1;
2690 def m5_0ImmPred : PatLeaf<(i32 imm), [{
2691 int64_t V = N->getSExtValue();
2692 return -31 <= V && V <= -1;
2695 def IsNPow2_8 : PatLeaf<(i32 imm), [{
2696 uint8_t NV = ~N->getZExtValue();
2697 return isPowerOf2_32(NV);
2700 def IsNPow2_16 : PatLeaf<(i32 imm), [{
2701 uint16_t NV = ~N->getZExtValue();
2702 return isPowerOf2_32(NV);
2705 def Log2_8 : SDNodeXForm<imm, [{
2706 uint8_t V = N->getZExtValue();
2707 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
2710 def Log2_16 : SDNodeXForm<imm, [{
2711 uint16_t V = N->getZExtValue();
2712 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
2715 def LogN2_8 : SDNodeXForm<imm, [{
2716 uint8_t NV = ~N->getZExtValue();
2717 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
2720 def LogN2_16 : SDNodeXForm<imm, [{
2721 uint16_t NV = ~N->getZExtValue();
2722 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
2725 def IdImm : SDNodeXForm<imm, [{ return SDValue(N, 0); }]>;
2727 multiclass Memopxr_base_pat<PatFrag Load, PatFrag Store, SDNode Oper,
2730 def: Pat<(Store (Oper (Load I32:$Rs), I32:$A), I32:$Rs),
2731 (MI I32:$Rs, 0, I32:$A)>;
2733 def: Pat<(Store (Oper (Load AddrFI:$Rs), I32:$A), AddrFI:$Rs),
2734 (MI AddrFI:$Rs, 0, I32:$A)>;
2737 multiclass Memopxr_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2738 SDNode Oper, InstHexagon MI> {
2740 def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), I32:$A),
2741 (add I32:$Rs, ImmPred:$Off)),
2742 (MI I32:$Rs, imm:$Off, I32:$A)>;
2743 def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), I32:$A),
2744 (IsOrAdd I32:$Rs, ImmPred:$Off)),
2745 (MI I32:$Rs, imm:$Off, I32:$A)>;
2747 def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), I32:$A),
2748 (add AddrFI:$Rs, ImmPred:$Off)),
2749 (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
2750 def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), I32:$A),
2751 (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
2752 (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
2755 multiclass Memopxr_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2756 SDNode Oper, InstHexagon MI> {
2757 let Predicates = [UseMEMOPS] in {
2758 defm: Memopxr_base_pat <Load, Store, Oper, MI>;
2759 defm: Memopxr_add_pat <Load, Store, ImmPred, Oper, MI>;
2763 let AddedComplexity = 200 in {
2765 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, add,
2766 /*anyext*/ L4_add_memopb_io>;
2767 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, add,
2768 /*sext*/ L4_add_memopb_io>;
2769 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, add,
2770 /*zext*/ L4_add_memopb_io>;
2771 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, add,
2772 /*anyext*/ L4_add_memoph_io>;
2773 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, add,
2774 /*sext*/ L4_add_memoph_io>;
2775 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, add,
2776 /*zext*/ L4_add_memoph_io>;
2777 defm: Memopxr_pat<load, store, u6_2ImmPred, add, L4_add_memopw_io>;
2780 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, sub,
2781 /*anyext*/ L4_sub_memopb_io>;
2782 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub,
2783 /*sext*/ L4_sub_memopb_io>;
2784 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub,
2785 /*zext*/ L4_sub_memopb_io>;
2786 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, sub,
2787 /*anyext*/ L4_sub_memoph_io>;
2788 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub,
2789 /*sext*/ L4_sub_memoph_io>;
2790 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub,
2791 /*zext*/ L4_sub_memoph_io>;
2792 defm: Memopxr_pat<load, store, u6_2ImmPred, sub, L4_sub_memopw_io>;
2795 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, and,
2796 /*anyext*/ L4_and_memopb_io>;
2797 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, and,
2798 /*sext*/ L4_and_memopb_io>;
2799 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, and,
2800 /*zext*/ L4_and_memopb_io>;
2801 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, and,
2802 /*anyext*/ L4_and_memoph_io>;
2803 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, and,
2804 /*sext*/ L4_and_memoph_io>;
2805 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, and,
2806 /*zext*/ L4_and_memoph_io>;
2807 defm: Memopxr_pat<load, store, u6_2ImmPred, and, L4_and_memopw_io>;
2810 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, or,
2811 /*anyext*/ L4_or_memopb_io>;
2812 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, or,
2813 /*sext*/ L4_or_memopb_io>;
2814 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, or,
2815 /*zext*/ L4_or_memopb_io>;
2816 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, or,
2817 /*anyext*/ L4_or_memoph_io>;
2818 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, or,
2819 /*sext*/ L4_or_memoph_io>;
2820 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, or,
2821 /*zext*/ L4_or_memoph_io>;
2822 defm: Memopxr_pat<load, store, u6_2ImmPred, or, L4_or_memopw_io>;
2826 multiclass Memopxi_base_pat<PatFrag Load, PatFrag Store, SDNode Oper,
2827 PatFrag Arg, SDNodeXForm ArgMod, InstHexagon MI> {
2829 def: Pat<(Store (Oper (Load I32:$Rs), Arg:$A), I32:$Rs),
2830 (MI I32:$Rs, 0, (ArgMod Arg:$A))>;
2832 def: Pat<(Store (Oper (Load AddrFI:$Rs), Arg:$A), AddrFI:$Rs),
2833 (MI AddrFI:$Rs, 0, (ArgMod Arg:$A))>;
2836 multiclass Memopxi_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2837 SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
2840 def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), Arg:$A),
2841 (add I32:$Rs, ImmPred:$Off)),
2842 (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2843 def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), Arg:$A),
2844 (IsOrAdd I32:$Rs, ImmPred:$Off)),
2845 (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2847 def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
2848 (add AddrFI:$Rs, ImmPred:$Off)),
2849 (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2850 def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
2851 (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
2852 (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2855 multiclass Memopxi_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2856 SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
2858 let Predicates = [UseMEMOPS] in {
2859 defm: Memopxi_base_pat <Load, Store, Oper, Arg, ArgMod, MI>;
2860 defm: Memopxi_add_pat <Load, Store, ImmPred, Oper, Arg, ArgMod, MI>;
2864 let AddedComplexity = 220 in {
2866 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
2867 /*anyext*/ IdImm, L4_iadd_memopb_io>;
2868 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
2869 /*sext*/ IdImm, L4_iadd_memopb_io>;
2870 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
2871 /*zext*/ IdImm, L4_iadd_memopb_io>;
2872 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
2873 /*anyext*/ IdImm, L4_iadd_memoph_io>;
2874 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
2875 /*sext*/ IdImm, L4_iadd_memoph_io>;
2876 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
2877 /*zext*/ IdImm, L4_iadd_memoph_io>;
2878 defm: Memopxi_pat<load, store, u6_2ImmPred, add, u5_0ImmPred, IdImm,
2880 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
2881 /*anyext*/ NegImm8, L4_iadd_memopb_io>;
2882 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
2883 /*sext*/ NegImm8, L4_iadd_memopb_io>;
2884 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
2885 /*zext*/ NegImm8, L4_iadd_memopb_io>;
2886 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
2887 /*anyext*/ NegImm16, L4_iadd_memoph_io>;
2888 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
2889 /*sext*/ NegImm16, L4_iadd_memoph_io>;
2890 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
2891 /*zext*/ NegImm16, L4_iadd_memoph_io>;
2892 defm: Memopxi_pat<load, store, u6_2ImmPred, sub, m5_0ImmPred, NegImm32,
2896 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
2897 /*anyext*/ IdImm, L4_isub_memopb_io>;
2898 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
2899 /*sext*/ IdImm, L4_isub_memopb_io>;
2900 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
2901 /*zext*/ IdImm, L4_isub_memopb_io>;
2902 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
2903 /*anyext*/ IdImm, L4_isub_memoph_io>;
2904 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
2905 /*sext*/ IdImm, L4_isub_memoph_io>;
2906 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
2907 /*zext*/ IdImm, L4_isub_memoph_io>;
2908 defm: Memopxi_pat<load, store, u6_2ImmPred, sub, u5_0ImmPred, IdImm,
2910 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
2911 /*anyext*/ NegImm8, L4_isub_memopb_io>;
2912 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
2913 /*sext*/ NegImm8, L4_isub_memopb_io>;
2914 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
2915 /*zext*/ NegImm8, L4_isub_memopb_io>;
2916 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
2917 /*anyext*/ NegImm16, L4_isub_memoph_io>;
2918 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
2919 /*sext*/ NegImm16, L4_isub_memoph_io>;
2920 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
2921 /*zext*/ NegImm16, L4_isub_memoph_io>;
2922 defm: Memopxi_pat<load, store, u6_2ImmPred, add, m5_0ImmPred, NegImm32,
2926 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
2927 /*anyext*/ LogN2_8, L4_iand_memopb_io>;
2928 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
2929 /*sext*/ LogN2_8, L4_iand_memopb_io>;
2930 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
2931 /*zext*/ LogN2_8, L4_iand_memopb_io>;
2932 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
2933 /*anyext*/ LogN2_16, L4_iand_memoph_io>;
2934 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
2935 /*sext*/ LogN2_16, L4_iand_memoph_io>;
2936 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
2937 /*zext*/ LogN2_16, L4_iand_memoph_io>;
2938 defm: Memopxi_pat<load, store, u6_2ImmPred, and, IsNPow2_32,
2939 LogN2_32, L4_iand_memopw_io>;
2942 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
2943 /*anyext*/ Log2_8, L4_ior_memopb_io>;
2944 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
2945 /*sext*/ Log2_8, L4_ior_memopb_io>;
2946 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
2947 /*zext*/ Log2_8, L4_ior_memopb_io>;
2948 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
2949 /*anyext*/ Log2_16, L4_ior_memoph_io>;
2950 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
2951 /*sext*/ Log2_16, L4_ior_memoph_io>;
2952 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
2953 /*zext*/ Log2_16, L4_ior_memoph_io>;
2954 defm: Memopxi_pat<load, store, u6_2ImmPred, or, IsPow2_32,
2955 Log2_32, L4_ior_memopw_io>;
2959 // --(15) Call -----------------------------------------------------------
2962 // Pseudo instructions.
2963 def SDT_SPCallSeqStart
2964 : SDCallSeqStart<[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
2965 def SDT_SPCallSeqEnd
2966 : SDCallSeqEnd<[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
2968 def callseq_start: SDNode<"ISD::CALLSEQ_START", SDT_SPCallSeqStart,
2969 [SDNPHasChain, SDNPOutGlue]>;
2970 def callseq_end: SDNode<"ISD::CALLSEQ_END", SDT_SPCallSeqEnd,
2971 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
2973 def SDT_SPCall: SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
2975 def HexagonTCRet: SDNode<"HexagonISD::TC_RETURN", SDT_SPCall,
2976 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
2977 def callv3: SDNode<"HexagonISD::CALL", SDT_SPCall,
2978 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
2979 def callv3nr: SDNode<"HexagonISD::CALLnr", SDT_SPCall,
2980 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
2982 def: Pat<(callseq_start timm:$amt, timm:$amt2),
2983 (ADJCALLSTACKDOWN imm:$amt, imm:$amt2)>;
2984 def: Pat<(callseq_end timm:$amt1, timm:$amt2),
2985 (ADJCALLSTACKUP imm:$amt1, imm:$amt2)>;
2987 def: Pat<(HexagonTCRet tglobaladdr:$dst), (PS_tailcall_i tglobaladdr:$dst)>;
2988 def: Pat<(HexagonTCRet texternalsym:$dst), (PS_tailcall_i texternalsym:$dst)>;
2989 def: Pat<(HexagonTCRet I32:$dst), (PS_tailcall_r I32:$dst)>;
2991 def: Pat<(callv3 I32:$dst), (J2_callr I32:$dst)>;
2992 def: Pat<(callv3 tglobaladdr:$dst), (J2_call tglobaladdr:$dst)>;
2993 def: Pat<(callv3 texternalsym:$dst), (J2_call texternalsym:$dst)>;
2994 def: Pat<(callv3 tglobaltlsaddr:$dst), (J2_call tglobaltlsaddr:$dst)>;
2996 def: Pat<(callv3nr I32:$dst), (PS_callr_nr I32:$dst)>;
2997 def: Pat<(callv3nr tglobaladdr:$dst), (PS_call_nr tglobaladdr:$dst)>;
2998 def: Pat<(callv3nr texternalsym:$dst), (PS_call_nr texternalsym:$dst)>;
3000 def retflag : SDNode<"HexagonISD::RET_FLAG", SDTNone,
3001 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
3002 def eh_return: SDNode<"HexagonISD::EH_RETURN", SDTNone, [SDNPHasChain]>;
3004 def: Pat<(retflag), (PS_jmpret (i32 R31))>;
3005 def: Pat<(eh_return), (EH_RETURN_JMPR (i32 R31))>;
3008 // --(16) Branch ---------------------------------------------------------
3011 def: Pat<(br bb:$dst), (J2_jump b30_2Imm:$dst)>;
3012 def: Pat<(brind I32:$dst), (J2_jumpr I32:$dst)>;
3014 def: Pat<(brcond I1:$Pu, bb:$dst),
3015 (J2_jumpt I1:$Pu, bb:$dst)>;
3016 def: Pat<(brcond (not I1:$Pu), bb:$dst),
3017 (J2_jumpf I1:$Pu, bb:$dst)>;
3018 def: Pat<(brcond (i1 (setne I1:$Pu, -1)), bb:$dst),
3019 (J2_jumpf I1:$Pu, bb:$dst)>;
3020 def: Pat<(brcond (i1 (seteq I1:$Pu, 0)), bb:$dst),
3021 (J2_jumpf I1:$Pu, bb:$dst)>;
3022 def: Pat<(brcond (i1 (setne I1:$Pu, 0)), bb:$dst),
3023 (J2_jumpt I1:$Pu, bb:$dst)>;
3026 // --(17) Misc -----------------------------------------------------------
3029 // Generate code of the form 'C2_muxii(cmpbgtui(Rdd, C-1),0,1)'
3030 // for C code of the form r = (c>='0' && c<='9') ? 1 : 0.
3031 // The isdigit transformation relies on two 'clever' aspects:
3032 // 1) The data type is unsigned which allows us to eliminate a zero test after
3033 // biasing the expression by 48. We are depending on the representation of
3034 // the unsigned types, and semantics.
3035 // 2) The front end has converted <= 9 into < 10 on entry to LLVM.
3038 // retval = (c >= '0' && c <= '9') ? 1 : 0;
3039 // The code is transformed upstream of llvm into
3040 // retval = (c-48) < 10 ? 1 : 0;
3042 def u7_0PosImmPred : ImmLeaf<i32, [{
3043 // True if the immediate fits in an 7-bit unsigned field and is positive.
3044 return Imm > 0 && isUInt<7>(Imm);
3047 let AddedComplexity = 139 in
3048 def: Pat<(i32 (zext (i1 (setult (and I32:$Rs, 255), u7_0PosImmPred:$u7)))),
3049 (C2_muxii (A4_cmpbgtui IntRegs:$Rs, (UDEC1 imm:$u7)), 0, 1)>;
3051 let AddedComplexity = 100 in
3052 def: Pat<(or (or (shl (HexagonINSERT (i32 (zextloadi8 (add I32:$b, 2))),
3053 (i32 (extloadi8 (add I32:$b, 3))),
3056 (shl (i32 (zextloadi8 (add I32:$b, 1))), (i32 8))),
3057 (zextloadi8 I32:$b)),
3058 (A2_swiz (L2_loadri_io I32:$b, 0))>;
3061 // We need custom lowering of ISD::PREFETCH into HexagonISD::DCFETCH
3062 // because the SDNode ISD::PREFETCH has properties MayLoad and MayStore.
3063 // We don't really want either one here.
3064 def SDTHexagonDCFETCH: SDTypeProfile<0, 2, [SDTCisPtrTy<0>,SDTCisInt<1>]>;
3065 def HexagonDCFETCH: SDNode<"HexagonISD::DCFETCH", SDTHexagonDCFETCH,
3068 def: Pat<(HexagonDCFETCH IntRegs:$Rs, u11_3ImmPred:$u11_3),
3069 (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
3070 def: Pat<(HexagonDCFETCH (i32 (add IntRegs:$Rs, u11_3ImmPred:$u11_3)), (i32 0)),
3071 (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
3073 def SDTHexagonALLOCA
3074 : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3076 : SDNode<"HexagonISD::ALLOCA", SDTHexagonALLOCA, [SDNPHasChain]>;
3078 def: Pat<(HexagonALLOCA I32:$Rs, (i32 imm:$A)),
3079 (PS_alloca IntRegs:$Rs, imm:$A)>;
3081 def HexagonBARRIER: SDNode<"HexagonISD::BARRIER", SDTNone, [SDNPHasChain]>;
3082 def: Pat<(HexagonBARRIER), (Y2_barrier)>;
3084 def: Pat<(trap), (J2_trap0 (i32 0))>;
3086 // Read cycle counter.
3087 def SDTInt64Leaf: SDTypeProfile<1, 0, [SDTCisVT<0, i64>]>;
3088 def HexagonREADCYCLE: SDNode<"HexagonISD::READCYCLE", SDTInt64Leaf,
3091 def: Pat<(HexagonREADCYCLE), (A4_tfrcpp UPCYCLE)>;
3093 // The declared return value of the store-locked intrinsics is i32, but
3094 // the instructions actually define i1. To avoid register copies from
3095 // IntRegs to PredRegs and back, fold the entire pattern checking the
3096 // result against true/false.
3097 let AddedComplexity = 100 in {
3098 def: Pat<(i1 (setne (int_hexagon_S2_storew_locked I32:$Rs, I32:$Rt), 0)),
3099 (S2_storew_locked I32:$Rs, I32:$Rt)>;
3100 def: Pat<(i1 (seteq (int_hexagon_S2_storew_locked I32:$Rs, I32:$Rt), 0)),
3101 (C2_not (S2_storew_locked I32:$Rs, I32:$Rt))>;
3102 def: Pat<(i1 (setne (int_hexagon_S4_stored_locked I32:$Rs, I64:$Rt), 0)),
3103 (S4_stored_locked I32:$Rs, I64:$Rt)>;
3104 def: Pat<(i1 (seteq (int_hexagon_S4_stored_locked I32:$Rs, I64:$Rt), 0)),
3105 (C2_not (S4_stored_locked I32:$Rs, I64:$Rt))>;