1 //==- HexagonPatterns.td - Target Description for Hexagon -*- tablegen -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
13 // (3) Extend/truncate
19 // (9) Arithmetic/bitwise
29 // Guidelines (in no particular order):
30 // 1. Avoid relying on pattern ordering to give preference to one pattern
31 // over another, prefer using AddedComplexity instead. The reason for
32 // this is to avoid unintended conseqeuences (caused by altering the
33 // order) when making changes. The current order of patterns in this
34 // file obviously does play some role, but none of the ordering was
35 // deliberately chosen (other than to create a logical structure of
36 // this file). When making changes, adding AddedComplexity to existing
37 // patterns may be needed.
38 // 2. Maintain the logical structure of the file, try to put new patterns
39 // in designated sections.
40 // 3. Do not use A2_combinew instruction directly, use Combinew fragment
41 // instead. It uses REG_SEQUENCE, which is more amenable to optimizations.
42 // 4. Most selection macros are based on PatFrags. For DAGs that involve
43 // SDNodes, use pf1/pf2 to convert them to PatFrags. Use common frags
44 // whenever possible (see the Definitions section). When adding new
45 // macro, try to make is general to enable reuse across sections.
46 // 5. Compound instructions (e.g. Rx+Rs*Rt) are generated under the condition
47 // that the nested operation has only one use. Having it separated in case
48 // of multiple uses avoids duplication of (processor) work.
49 // 6. The v4 vector instructions (64-bit) are treated as core instructions,
50 // for example, A2_vaddh is in the "arithmetic" section with A2_add.
51 // 7. When adding a pattern for an instruction with a constant-extendable
52 // operand, allow all possible kinds of inputs for the immediate value
53 // (see AnyImm/anyimm and their variants in the Definitions section).
56 // --(0) Definitions -----------------------------------------------------
59 // This complex pattern exists only to create a machine instruction operand
60 // of type "frame index". There doesn't seem to be a way to do that directly
62 def AddrFI: ComplexPattern<i32, 1, "SelectAddrFI", [frameindex], []>;
64 // These complex patterns are not strictly necessary, since global address
65 // folding will happen during DAG combining. For distinguishing between GA
66 // and GP, pat frags with HexagonCONST32 and HexagonCONST32_GP can be used.
67 def AddrGA: ComplexPattern<i32, 1, "SelectAddrGA", [], []>;
68 def AddrGP: ComplexPattern<i32, 1, "SelectAddrGP", [], []>;
69 def AnyImm: ComplexPattern<i32, 1, "SelectAnyImm", [], []>;
70 def AnyInt: ComplexPattern<i32, 1, "SelectAnyInt", [], []>;
72 // Global address or a constant being a multiple of 2^n.
73 def AnyImm0: ComplexPattern<i32, 1, "SelectAnyImm0", [], []>;
74 def AnyImm1: ComplexPattern<i32, 1, "SelectAnyImm1", [], []>;
75 def AnyImm2: ComplexPattern<i32, 1, "SelectAnyImm2", [], []>;
76 def AnyImm3: ComplexPattern<i32, 1, "SelectAnyImm3", [], []>;
80 def V2I1: PatLeaf<(v2i1 PredRegs:$R)>;
81 def V4I1: PatLeaf<(v4i1 PredRegs:$R)>;
82 def V8I1: PatLeaf<(v8i1 PredRegs:$R)>;
83 def V4I8: PatLeaf<(v4i8 IntRegs:$R)>;
84 def V2I16: PatLeaf<(v2i16 IntRegs:$R)>;
86 def V8I8: PatLeaf<(v8i8 DoubleRegs:$R)>;
87 def V4I16: PatLeaf<(v4i16 DoubleRegs:$R)>;
88 def V2I32: PatLeaf<(v2i32 DoubleRegs:$R)>;
90 def HQ8: PatLeaf<(VecQ8 HvxQR:$R)>;
91 def HQ16: PatLeaf<(VecQ16 HvxQR:$R)>;
92 def HQ32: PatLeaf<(VecQ32 HvxQR:$R)>;
94 def HVI8: PatLeaf<(VecI8 HvxVR:$R)>;
95 def HVI16: PatLeaf<(VecI16 HvxVR:$R)>;
96 def HVI32: PatLeaf<(VecI32 HvxVR:$R)>;
98 def HWI8: PatLeaf<(VecPI8 HvxWR:$R)>;
99 def HWI16: PatLeaf<(VecPI16 HvxWR:$R)>;
100 def HWI32: PatLeaf<(VecPI32 HvxWR:$R)>;
103 SDTypeProfile<1, 0, [SDTCisVec<0>]>;
105 SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisVec<1>, SDTCisSameAs<1,2>,
108 def HexagonPTRUE: SDNode<"HexagonISD::PTRUE", SDTVecLeaf>;
109 def HexagonPFALSE: SDNode<"HexagonISD::PFALSE", SDTVecLeaf>;
110 def HexagonVALIGN: SDNode<"HexagonISD::VALIGN", SDTVecVecIntOp>;
111 def HexagonVALIGNADDR: SDNode<"HexagonISD::VALIGNADDR", SDTIntUnaryOp>;
113 def ptrue: PatFrag<(ops), (HexagonPTRUE)>;
114 def pfalse: PatFrag<(ops), (HexagonPFALSE)>;
115 def pnot: PatFrag<(ops node:$Pu), (xor node:$Pu, ptrue)>;
117 def valign: PatFrag<(ops node:$Vt, node:$Vs, node:$Ru),
118 (HexagonVALIGN node:$Vt, node:$Vs, node:$Ru)>;
119 def valignaddr: PatFrag<(ops node:$Addr), (HexagonVALIGNADDR node:$Addr)>;
121 // Pattern fragments to extract the low and high subregisters from a
123 def LoReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG (i64 $Rs), isub_lo)>;
124 def HiReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG (i64 $Rs), isub_hi)>;
126 def IsOrAdd: PatFrag<(ops node:$A, node:$B), (or node:$A, node:$B), [{
127 return isOrEquivalentToAdd(N);
130 def IsPow2_32: PatLeaf<(i32 imm), [{
131 uint32_t V = N->getZExtValue();
132 return isPowerOf2_32(V);
135 def IsPow2_64: PatLeaf<(i64 imm), [{
136 uint64_t V = N->getZExtValue();
137 return isPowerOf2_64(V);
140 def IsNPow2_32: PatLeaf<(i32 imm), [{
141 uint32_t NV = ~N->getZExtValue();
142 return isPowerOf2_32(NV);
145 def IsPow2_64L: PatLeaf<(i64 imm), [{
146 uint64_t V = N->getZExtValue();
147 return isPowerOf2_64(V) && Log2_64(V) < 32;
150 def IsPow2_64H: PatLeaf<(i64 imm), [{
151 uint64_t V = N->getZExtValue();
152 return isPowerOf2_64(V) && Log2_64(V) >= 32;
155 def IsNPow2_64L: PatLeaf<(i64 imm), [{
156 uint64_t NV = ~N->getZExtValue();
157 return isPowerOf2_64(NV) && Log2_64(NV) < 32;
160 def IsNPow2_64H: PatLeaf<(i64 imm), [{
161 uint64_t NV = ~N->getZExtValue();
162 return isPowerOf2_64(NV) && Log2_64(NV) >= 32;
165 class IsULE<int Width, int Arg>: PatLeaf<(i32 imm),
166 "uint64_t V = N->getZExtValue();" #
167 "return isUInt<" # Width # ">(V) && V <= " # Arg # ";"
170 class IsUGT<int Width, int Arg>: PatLeaf<(i32 imm),
171 "uint64_t V = N->getZExtValue();" #
172 "return isUInt<" # Width # ">(V) && V > " # Arg # ";"
175 def SDEC1: SDNodeXForm<imm, [{
176 int32_t V = N->getSExtValue();
177 return CurDAG->getTargetConstant(V-1, SDLoc(N), MVT::i32);
180 def UDEC1: SDNodeXForm<imm, [{
181 uint32_t V = N->getZExtValue();
183 return CurDAG->getTargetConstant(V-1, SDLoc(N), MVT::i32);
186 def UDEC32: SDNodeXForm<imm, [{
187 uint32_t V = N->getZExtValue();
189 return CurDAG->getTargetConstant(V-32, SDLoc(N), MVT::i32);
192 class Subi<int From>: SDNodeXForm<imm,
193 "int32_t V = " # From # " - N->getSExtValue();" #
194 "return CurDAG->getTargetConstant(V, SDLoc(N), MVT::i32);"
197 def Log2_32: SDNodeXForm<imm, [{
198 uint32_t V = N->getZExtValue();
199 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
202 def Log2_64: SDNodeXForm<imm, [{
203 uint64_t V = N->getZExtValue();
204 return CurDAG->getTargetConstant(Log2_64(V), SDLoc(N), MVT::i32);
207 def LogN2_32: SDNodeXForm<imm, [{
208 uint32_t NV = ~N->getZExtValue();
209 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
212 def LogN2_64: SDNodeXForm<imm, [{
213 uint64_t NV = ~N->getZExtValue();
214 return CurDAG->getTargetConstant(Log2_64(NV), SDLoc(N), MVT::i32);
217 def NegImm8: SDNodeXForm<imm, [{
218 int8_t NV = -N->getSExtValue();
219 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
222 def NegImm16: SDNodeXForm<imm, [{
223 int16_t NV = -N->getSExtValue();
224 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
227 def NegImm32: SDNodeXForm<imm, [{
228 int32_t NV = -N->getSExtValue();
229 return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
233 // Helpers for type promotions/contractions.
234 def I1toI32: OutPatFrag<(ops node:$Rs), (C2_muxii (i1 $Rs), 1, 0)>;
235 def I32toI1: OutPatFrag<(ops node:$Rs), (i1 (C2_cmpgtui (i32 $Rs), (i32 0)))>;
236 def ToZext64: OutPatFrag<(ops node:$Rs), (i64 (A4_combineir 0, (i32 $Rs)))>;
237 def ToSext64: OutPatFrag<(ops node:$Rs), (i64 (A2_sxtw (i32 $Rs)))>;
238 def ToAext64: OutPatFrag<(ops node:$Rs),
239 (REG_SEQUENCE DoubleRegs, (i32 (IMPLICIT_DEF)), isub_hi, (i32 $Rs), isub_lo)>;
241 def Combinew: OutPatFrag<(ops node:$Rs, node:$Rt),
242 (REG_SEQUENCE DoubleRegs, $Rs, isub_hi, $Rt, isub_lo)>;
244 def addrga: PatLeaf<(i32 AddrGA:$Addr)>;
245 def addrgp: PatLeaf<(i32 AddrGP:$Addr)>;
246 def anyimm: PatLeaf<(i32 AnyImm:$Imm)>;
247 def anyint: PatLeaf<(i32 AnyInt:$Imm)>;
249 // Global address or an aligned constant.
250 def anyimm0: PatLeaf<(i32 AnyImm0:$Addr)>;
251 def anyimm1: PatLeaf<(i32 AnyImm1:$Addr)>;
252 def anyimm2: PatLeaf<(i32 AnyImm2:$Addr)>;
253 def anyimm3: PatLeaf<(i32 AnyImm3:$Addr)>;
255 def f32ImmPred : PatLeaf<(f32 fpimm:$F)>;
256 def f64ImmPred : PatLeaf<(f64 fpimm:$F)>;
258 // This complex pattern is really only to detect various forms of
259 // sign-extension i32->i64. The selected value will be of type i64
260 // whose low word is the value being extended. The high word is
262 def Usxtw: ComplexPattern<i64, 1, "DetectUseSxtw", [], []>;
264 def Aext64: PatFrag<(ops node:$Rs), (i64 (anyext node:$Rs))>;
265 def Zext64: PatFrag<(ops node:$Rs), (i64 (zext node:$Rs))>;
266 def Sext64: PatLeaf<(i64 Usxtw:$Rs)>;
268 def azext: PatFrags<(ops node:$Rs), [(zext node:$Rs), (anyext node:$Rs)]>;
269 def asext: PatFrags<(ops node:$Rs), [(sext node:$Rs), (anyext node:$Rs)]>;
271 def: Pat<(IsOrAdd (i32 AddrFI:$Rs), s32_0ImmPred:$off),
272 (PS_fi (i32 AddrFI:$Rs), imm:$off)>;
275 // Converters from unary/binary SDNode to PatFrag.
276 class pf1<SDNode Op> : PatFrag<(ops node:$a), (Op node:$a)>;
277 class pf2<SDNode Op> : PatFrag<(ops node:$a, node:$b), (Op node:$a, node:$b)>;
279 class Not2<PatFrag P>
280 : PatFrag<(ops node:$A, node:$B), (P node:$A, (not node:$B))>;
282 // If there is a constant operand that feeds the and/or instruction,
283 // do not generate the compound instructions.
284 // It is not always profitable, as some times we end up with a transfer.
285 // Check the below example.
286 // ra = #65820; rb = lsr(rb, #8); rc ^= and (rb, ra)
287 // Instead this is preferable.
288 // ra = and (#65820, lsr(ra, #8)); rb = xor(rb, ra)
289 class Su_ni1<PatFrag Op>
290 : PatFrag<Op.Operands, !head(Op.Fragments), [{
292 // Check if Op1 is an immediate operand.
293 SDValue Op1 = N->getOperand(1);
294 return !isa<ConstantSDNode>(Op1);
297 Op.OperandTransform>;
300 : PatFrag<Op.Operands, !head(Op.Fragments), [{ return hasOneUse(N); }],
301 Op.OperandTransform>;
303 // Main selection macros.
305 class OpR_R_pat<InstHexagon MI, PatFrag Op, ValueType ResVT, PatFrag RegPred>
306 : Pat<(ResVT (Op RegPred:$Rs)), (MI RegPred:$Rs)>;
308 class OpR_RI_pat<InstHexagon MI, PatFrag Op, ValueType ResType,
309 PatFrag RegPred, PatFrag ImmPred>
310 : Pat<(ResType (Op RegPred:$Rs, ImmPred:$I)),
311 (MI RegPred:$Rs, imm:$I)>;
313 class OpR_RR_pat<InstHexagon MI, PatFrag Op, ValueType ResType,
314 PatFrag RsPred, PatFrag RtPred = RsPred>
315 : Pat<(ResType (Op RsPred:$Rs, RtPred:$Rt)),
316 (MI RsPred:$Rs, RtPred:$Rt)>;
318 class AccRRI_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op,
319 PatFrag RegPred, PatFrag ImmPred>
320 : Pat<(AccOp RegPred:$Rx, (Op RegPred:$Rs, ImmPred:$I)),
321 (MI RegPred:$Rx, RegPred:$Rs, imm:$I)>;
323 class AccRRR_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op,
324 PatFrag RxPred, PatFrag RsPred, PatFrag RtPred>
325 : Pat<(AccOp RxPred:$Rx, (Op RsPred:$Rs, RtPred:$Rt)),
326 (MI RxPred:$Rx, RsPred:$Rs, RtPred:$Rt)>;
328 multiclass SelMinMax_pats<PatFrag CmpOp, PatFrag Val,
329 InstHexagon InstA, InstHexagon InstB> {
330 def: Pat<(select (i1 (CmpOp Val:$A, Val:$B)), Val:$A, Val:$B),
331 (InstA Val:$A, Val:$B)>;
332 def: Pat<(select (i1 (CmpOp Val:$A, Val:$B)), Val:$B, Val:$A),
333 (InstB Val:$A, Val:$B)>;
336 multiclass MinMax_pats<InstHexagon PickT, InstHexagon PickS,
337 PatFrag Sel, PatFrag CmpOp,
338 ValueType CmpType, PatFrag CmpPred> {
339 def: Pat<(Sel (CmpType (CmpOp CmpPred:$Vs, CmpPred:$Vt)),
340 CmpPred:$Vt, CmpPred:$Vs),
341 (PickT CmpPred:$Vs, CmpPred:$Vt)>;
342 def: Pat<(Sel (CmpType (CmpOp CmpPred:$Vs, CmpPred:$Vt)),
343 CmpPred:$Vs, CmpPred:$Vt),
344 (PickS CmpPred:$Vs, CmpPred:$Vt)>;
347 // Bitcasts between same-size vector types are no-ops, except for the
348 // actual type change.
349 multiclass NopCast_pat<ValueType Ty1, ValueType Ty2, RegisterClass RC> {
350 def: Pat<(Ty1 (bitconvert (Ty2 RC:$Val))), (Ty1 RC:$Val)>;
351 def: Pat<(Ty2 (bitconvert (Ty1 RC:$Val))), (Ty2 RC:$Val)>;
355 // Frags for commonly used SDNodes.
356 def Add: pf2<add>; def And: pf2<and>; def Sra: pf2<sra>;
357 def Sub: pf2<sub>; def Or: pf2<or>; def Srl: pf2<srl>;
358 def Mul: pf2<mul>; def Xor: pf2<xor>; def Shl: pf2<shl>;
362 // --(1) Immediate -------------------------------------------------------
365 def Imm64Lo: SDNodeXForm<imm, [{
366 return CurDAG->getTargetConstant(int32_t (N->getSExtValue()),
369 def Imm64Hi: SDNodeXForm<imm, [{
370 return CurDAG->getTargetConstant(int32_t (N->getSExtValue()>>32),
375 def SDTHexagonCONST32
376 : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisPtrTy<0>]>;
378 def HexagonJT: SDNode<"HexagonISD::JT", SDTIntUnaryOp>;
379 def HexagonCP: SDNode<"HexagonISD::CP", SDTIntUnaryOp>;
380 def HexagonCONST32: SDNode<"HexagonISD::CONST32", SDTHexagonCONST32>;
381 def HexagonCONST32_GP: SDNode<"HexagonISD::CONST32_GP", SDTHexagonCONST32>;
383 def TruncI64ToI32: SDNodeXForm<imm, [{
384 return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i32);
387 def: Pat<(s32_0ImmPred:$s16), (A2_tfrsi imm:$s16)>;
388 def: Pat<(s8_0Imm64Pred:$s8), (A2_tfrpi (TruncI64ToI32 $s8))>;
390 def: Pat<(HexagonCONST32 tglobaltlsaddr:$A), (A2_tfrsi imm:$A)>;
391 def: Pat<(HexagonCONST32 bbl:$A), (A2_tfrsi imm:$A)>;
392 def: Pat<(HexagonCONST32 tglobaladdr:$A), (A2_tfrsi imm:$A)>;
393 def: Pat<(HexagonCONST32_GP tblockaddress:$A), (A2_tfrsi imm:$A)>;
394 def: Pat<(HexagonCONST32_GP tglobaladdr:$A), (A2_tfrsi imm:$A)>;
395 def: Pat<(HexagonJT tjumptable:$A), (A2_tfrsi imm:$A)>;
396 def: Pat<(HexagonCP tconstpool:$A), (A2_tfrsi imm:$A)>;
397 // The HVX load patterns also match CP directly. Make sure that if
398 // the selection of this opcode changes, it's updated in all places.
400 def: Pat<(i1 0), (PS_false)>;
401 def: Pat<(i1 1), (PS_true)>;
402 def: Pat<(i64 imm:$v), (CONST64 imm:$v)>,
403 Requires<[UseSmallData,NotOptTinyCore]>;
404 def: Pat<(i64 imm:$v),
405 (Combinew (A2_tfrsi (Imm64Hi $v)), (A2_tfrsi (Imm64Lo $v)))>;
407 def ftoi : SDNodeXForm<fpimm, [{
408 APInt I = N->getValueAPF().bitcastToAPInt();
409 return CurDAG->getTargetConstant(I.getZExtValue(), SDLoc(N),
410 MVT::getIntegerVT(I.getBitWidth()));
413 def: Pat<(f32ImmPred:$f), (A2_tfrsi (ftoi $f))>;
414 def: Pat<(f64ImmPred:$f), (CONST64 (ftoi $f))>;
416 def ToI32: OutPatFrag<(ops node:$V), (A2_tfrsi $V)>;
418 // --(2) Type cast -------------------------------------------------------
421 def: OpR_R_pat<F2_conv_sf2df, pf1<fpextend>, f64, F32>;
422 def: OpR_R_pat<F2_conv_df2sf, pf1<fpround>, f32, F64>;
424 def: OpR_R_pat<F2_conv_w2sf, pf1<sint_to_fp>, f32, I32>;
425 def: OpR_R_pat<F2_conv_d2sf, pf1<sint_to_fp>, f32, I64>;
426 def: OpR_R_pat<F2_conv_w2df, pf1<sint_to_fp>, f64, I32>;
427 def: OpR_R_pat<F2_conv_d2df, pf1<sint_to_fp>, f64, I64>;
429 def: OpR_R_pat<F2_conv_uw2sf, pf1<uint_to_fp>, f32, I32>;
430 def: OpR_R_pat<F2_conv_ud2sf, pf1<uint_to_fp>, f32, I64>;
431 def: OpR_R_pat<F2_conv_uw2df, pf1<uint_to_fp>, f64, I32>;
432 def: OpR_R_pat<F2_conv_ud2df, pf1<uint_to_fp>, f64, I64>;
434 def: OpR_R_pat<F2_conv_sf2w_chop, pf1<fp_to_sint>, i32, F32>;
435 def: OpR_R_pat<F2_conv_df2w_chop, pf1<fp_to_sint>, i32, F64>;
436 def: OpR_R_pat<F2_conv_sf2d_chop, pf1<fp_to_sint>, i64, F32>;
437 def: OpR_R_pat<F2_conv_df2d_chop, pf1<fp_to_sint>, i64, F64>;
439 def: OpR_R_pat<F2_conv_sf2uw_chop, pf1<fp_to_uint>, i32, F32>;
440 def: OpR_R_pat<F2_conv_df2uw_chop, pf1<fp_to_uint>, i32, F64>;
441 def: OpR_R_pat<F2_conv_sf2ud_chop, pf1<fp_to_uint>, i64, F32>;
442 def: OpR_R_pat<F2_conv_df2ud_chop, pf1<fp_to_uint>, i64, F64>;
444 // Bitcast is different than [fp|sint|uint]_to_[sint|uint|fp].
445 def: Pat<(i32 (bitconvert F32:$v)), (I32:$v)>;
446 def: Pat<(f32 (bitconvert I32:$v)), (F32:$v)>;
447 def: Pat<(i64 (bitconvert F64:$v)), (I64:$v)>;
448 def: Pat<(f64 (bitconvert I64:$v)), (F64:$v)>;
450 // Bit convert 32- and 64-bit types.
451 // All of these are bitcastable to one another: i32, v2i16, v4i8.
452 defm: NopCast_pat<i32, v2i16, IntRegs>;
453 defm: NopCast_pat<i32, v4i8, IntRegs>;
454 defm: NopCast_pat<v2i16, v4i8, IntRegs>;
455 // All of these are bitcastable to one another: i64, v2i32, v4i16, v8i8.
456 defm: NopCast_pat<i64, v2i32, DoubleRegs>;
457 defm: NopCast_pat<i64, v4i16, DoubleRegs>;
458 defm: NopCast_pat<i64, v8i8, DoubleRegs>;
459 defm: NopCast_pat<v2i32, v4i16, DoubleRegs>;
460 defm: NopCast_pat<v2i32, v8i8, DoubleRegs>;
461 defm: NopCast_pat<v4i16, v8i8, DoubleRegs>;
464 // --(3) Extend/truncate -------------------------------------------------
467 def: Pat<(sext_inreg I32:$Rs, i8), (A2_sxtb I32:$Rs)>;
468 def: Pat<(sext_inreg I32:$Rs, i16), (A2_sxth I32:$Rs)>;
469 def: Pat<(sext_inreg I64:$Rs, i32), (A2_sxtw (LoReg $Rs))>;
470 def: Pat<(sext_inreg I64:$Rs, i16), (A2_sxtw (A2_sxth (LoReg $Rs)))>;
471 def: Pat<(sext_inreg I64:$Rs, i8), (A2_sxtw (A2_sxtb (LoReg $Rs)))>;
473 def: Pat<(i64 (sext I32:$Rs)), (A2_sxtw I32:$Rs)>;
474 def: Pat<(Zext64 I32:$Rs), (ToZext64 $Rs)>;
475 def: Pat<(Aext64 I32:$Rs), (ToZext64 $Rs)>;
477 def: Pat<(i32 (trunc I64:$Rs)), (LoReg $Rs)>;
478 def: Pat<(i1 (trunc I32:$Rs)), (S2_tstbit_i I32:$Rs, 0)>;
479 def: Pat<(i1 (trunc I64:$Rs)), (S2_tstbit_i (LoReg $Rs), 0)>;
481 let AddedComplexity = 20 in {
482 def: Pat<(and I32:$Rs, 255), (A2_zxtb I32:$Rs)>;
483 def: Pat<(and I32:$Rs, 65535), (A2_zxth I32:$Rs)>;
486 // Extensions from i1 or vectors of i1.
487 def: Pat<(i32 (azext I1:$Pu)), (C2_muxii I1:$Pu, 1, 0)>;
488 def: Pat<(i64 (azext I1:$Pu)), (ToZext64 (C2_muxii I1:$Pu, 1, 0))>;
489 def: Pat<(i32 (sext I1:$Pu)), (C2_muxii I1:$Pu, -1, 0)>;
490 def: Pat<(i64 (sext I1:$Pu)), (Combinew (C2_muxii PredRegs:$Pu, -1, 0),
491 (C2_muxii PredRegs:$Pu, -1, 0))>;
493 def: Pat<(v2i16 (sext V2I1:$Pu)), (S2_vtrunehb (C2_mask V2I1:$Pu))>;
494 def: Pat<(v2i32 (sext V2I1:$Pu)), (C2_mask V2I1:$Pu)>;
495 def: Pat<(v4i8 (sext V4I1:$Pu)), (S2_vtrunehb (C2_mask V4I1:$Pu))>;
496 def: Pat<(v4i16 (sext V4I1:$Pu)), (C2_mask V4I1:$Pu)>;
497 def: Pat<(v8i8 (sext V8I1:$Pu)), (C2_mask V8I1:$Pu)>;
499 def Vsplatpi: OutPatFrag<(ops node:$V),
500 (Combinew (A2_tfrsi $V), (A2_tfrsi $V))>;
502 def: Pat<(v2i16 (azext V2I1:$Pu)),
503 (A2_andir (LoReg (C2_mask V2I1:$Pu)), (i32 0x00010001))>;
504 def: Pat<(v2i32 (azext V2I1:$Pu)),
505 (A2_andp (C2_mask V2I1:$Pu), (A2_combineii (i32 1), (i32 1)))>;
506 def: Pat<(v4i8 (azext V4I1:$Pu)),
507 (A2_andir (LoReg (C2_mask V4I1:$Pu)), (i32 0x01010101))>;
508 def: Pat<(v4i16 (azext V4I1:$Pu)),
509 (A2_andp (C2_mask V4I1:$Pu), (Vsplatpi (i32 0x00010001)))>;
510 def: Pat<(v8i8 (azext V8I1:$Pu)),
511 (A2_andp (C2_mask V8I1:$Pu), (Vsplatpi (i32 0x01010101)))>;
513 def: Pat<(v4i16 (azext V4I8:$Rs)), (S2_vzxtbh V4I8:$Rs)>;
514 def: Pat<(v2i32 (azext V2I16:$Rs)), (S2_vzxthw V2I16:$Rs)>;
515 def: Pat<(v4i16 (sext V4I8:$Rs)), (S2_vsxtbh V4I8:$Rs)>;
516 def: Pat<(v2i32 (sext V2I16:$Rs)), (S2_vsxthw V2I16:$Rs)>;
518 def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i8)),
519 (Combinew (A2_sxtb (HiReg $Rs)), (A2_sxtb (LoReg $Rs)))>;
521 def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i16)),
522 (Combinew (A2_sxth (HiReg $Rs)), (A2_sxth (LoReg $Rs)))>;
524 // Truncate: from vector B copy all 'E'ven 'B'yte elements:
525 // A[0] = B[0]; A[1] = B[2]; A[2] = B[4]; A[3] = B[6];
526 def: Pat<(v4i8 (trunc V4I16:$Rs)),
527 (S2_vtrunehb V4I16:$Rs)>;
529 // Truncate: from vector B copy all 'O'dd 'B'yte elements:
530 // A[0] = B[1]; A[1] = B[3]; A[2] = B[5]; A[3] = B[7];
533 // Truncate: from vectors B and C copy all 'E'ven 'H'alf-word elements:
534 // A[0] = B[0]; A[1] = B[2]; A[2] = C[0]; A[3] = C[2];
537 def: Pat<(v2i16 (trunc V2I32:$Rs)),
538 (A2_combine_ll (HiReg $Rs), (LoReg $Rs))>;
541 // --(4) Logical ---------------------------------------------------------
544 def: Pat<(not I1:$Ps), (C2_not I1:$Ps)>;
545 def: Pat<(pnot V2I1:$Ps), (C2_not V2I1:$Ps)>;
546 def: Pat<(pnot V4I1:$Ps), (C2_not V4I1:$Ps)>;
547 def: Pat<(pnot V8I1:$Ps), (C2_not V8I1:$Ps)>;
548 def: Pat<(add I1:$Ps, -1), (C2_not I1:$Ps)>;
550 multiclass BoolOpR_RR_pat<InstHexagon MI, PatFrag Op> {
551 def: OpR_RR_pat<MI, Op, i1, I1>;
552 def: OpR_RR_pat<MI, Op, v2i1, V2I1>;
553 def: OpR_RR_pat<MI, Op, v4i1, V4I1>;
554 def: OpR_RR_pat<MI, Op, v8i1, V8I1>;
557 multiclass BoolAccRRR_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op> {
558 def: AccRRR_pat<MI, AccOp, Op, I1, I1, I1>;
559 def: AccRRR_pat<MI, AccOp, Op, V2I1, V2I1, V2I1>;
560 def: AccRRR_pat<MI, AccOp, Op, V4I1, V4I1, V4I1>;
561 def: AccRRR_pat<MI, AccOp, Op, V8I1, V8I1, V8I1>;
564 defm: BoolOpR_RR_pat<C2_and, And>;
565 defm: BoolOpR_RR_pat<C2_or, Or>;
566 defm: BoolOpR_RR_pat<C2_xor, Xor>;
567 defm: BoolOpR_RR_pat<C2_andn, Not2<And>>;
568 defm: BoolOpR_RR_pat<C2_orn, Not2<Or>>;
570 // op(Ps, op(Pt, Pu))
571 defm: BoolAccRRR_pat<C4_and_and, And, Su<And>>;
572 defm: BoolAccRRR_pat<C4_and_or, And, Su<Or>>;
573 defm: BoolAccRRR_pat<C4_or_and, Or, Su<And>>;
574 defm: BoolAccRRR_pat<C4_or_or, Or, Su<Or>>;
576 // op(Ps, op(Pt, ~Pu))
577 defm: BoolAccRRR_pat<C4_and_andn, And, Su<Not2<And>>>;
578 defm: BoolAccRRR_pat<C4_and_orn, And, Su<Not2<Or>>>;
579 defm: BoolAccRRR_pat<C4_or_andn, Or, Su<Not2<And>>>;
580 defm: BoolAccRRR_pat<C4_or_orn, Or, Su<Not2<Or>>>;
583 // --(5) Compare ---------------------------------------------------------
586 // Avoid negated comparisons, i.e. those of form "Pd = !cmp(...)".
587 // These cannot form compounds (e.g. J4_cmpeqi_tp0_jump_nt).
589 def: OpR_RI_pat<C2_cmpeqi, seteq, i1, I32, anyimm>;
590 def: OpR_RI_pat<C2_cmpgti, setgt, i1, I32, anyimm>;
591 def: OpR_RI_pat<C2_cmpgtui, setugt, i1, I32, anyimm>;
593 def: Pat<(i1 (setge I32:$Rs, s32_0ImmPred:$s10)),
594 (C2_cmpgti I32:$Rs, (SDEC1 imm:$s10))>;
595 def: Pat<(i1 (setuge I32:$Rs, u32_0ImmPred:$u9)),
596 (C2_cmpgtui I32:$Rs, (UDEC1 imm:$u9))>;
598 def: Pat<(i1 (setlt I32:$Rs, s32_0ImmPred:$s10)),
599 (C2_not (C2_cmpgti I32:$Rs, (SDEC1 imm:$s10)))>;
600 def: Pat<(i1 (setult I32:$Rs, u32_0ImmPred:$u9)),
601 (C2_not (C2_cmpgtui I32:$Rs, (UDEC1 imm:$u9)))>;
603 // Patfrag to convert the usual comparison patfrags (e.g. setlt) to ones
604 // that reverse the order of the operands.
605 class RevCmp<PatFrag F>
606 : PatFrag<(ops node:$rhs, node:$lhs), !head(F.Fragments), F.PredicateCode,
609 def: OpR_RR_pat<C2_cmpeq, seteq, i1, I32>;
610 def: OpR_RR_pat<C2_cmpgt, setgt, i1, I32>;
611 def: OpR_RR_pat<C2_cmpgtu, setugt, i1, I32>;
612 def: OpR_RR_pat<C2_cmpgt, RevCmp<setlt>, i1, I32>;
613 def: OpR_RR_pat<C2_cmpgtu, RevCmp<setult>, i1, I32>;
614 def: OpR_RR_pat<C2_cmpeqp, seteq, i1, I64>;
615 def: OpR_RR_pat<C2_cmpgtp, setgt, i1, I64>;
616 def: OpR_RR_pat<C2_cmpgtup, setugt, i1, I64>;
617 def: OpR_RR_pat<C2_cmpgtp, RevCmp<setlt>, i1, I64>;
618 def: OpR_RR_pat<C2_cmpgtup, RevCmp<setult>, i1, I64>;
619 def: OpR_RR_pat<A2_vcmpbeq, seteq, i1, V8I8>;
620 def: OpR_RR_pat<A2_vcmpbeq, seteq, v8i1, V8I8>;
621 def: OpR_RR_pat<A4_vcmpbgt, RevCmp<setlt>, i1, V8I8>;
622 def: OpR_RR_pat<A4_vcmpbgt, RevCmp<setlt>, v8i1, V8I8>;
623 def: OpR_RR_pat<A4_vcmpbgt, setgt, i1, V8I8>;
624 def: OpR_RR_pat<A4_vcmpbgt, setgt, v8i1, V8I8>;
625 def: OpR_RR_pat<A2_vcmpbgtu, RevCmp<setult>, i1, V8I8>;
626 def: OpR_RR_pat<A2_vcmpbgtu, RevCmp<setult>, v8i1, V8I8>;
627 def: OpR_RR_pat<A2_vcmpbgtu, setugt, i1, V8I8>;
628 def: OpR_RR_pat<A2_vcmpbgtu, setugt, v8i1, V8I8>;
629 def: OpR_RR_pat<A2_vcmpheq, seteq, i1, V4I16>;
630 def: OpR_RR_pat<A2_vcmpheq, seteq, v4i1, V4I16>;
631 def: OpR_RR_pat<A2_vcmphgt, RevCmp<setlt>, i1, V4I16>;
632 def: OpR_RR_pat<A2_vcmphgt, RevCmp<setlt>, v4i1, V4I16>;
633 def: OpR_RR_pat<A2_vcmphgt, setgt, i1, V4I16>;
634 def: OpR_RR_pat<A2_vcmphgt, setgt, v4i1, V4I16>;
635 def: OpR_RR_pat<A2_vcmphgtu, RevCmp<setult>, i1, V4I16>;
636 def: OpR_RR_pat<A2_vcmphgtu, RevCmp<setult>, v4i1, V4I16>;
637 def: OpR_RR_pat<A2_vcmphgtu, setugt, i1, V4I16>;
638 def: OpR_RR_pat<A2_vcmphgtu, setugt, v4i1, V4I16>;
639 def: OpR_RR_pat<A2_vcmpweq, seteq, i1, V2I32>;
640 def: OpR_RR_pat<A2_vcmpweq, seteq, v2i1, V2I32>;
641 def: OpR_RR_pat<A2_vcmpwgt, RevCmp<setlt>, i1, V2I32>;
642 def: OpR_RR_pat<A2_vcmpwgt, RevCmp<setlt>, v2i1, V2I32>;
643 def: OpR_RR_pat<A2_vcmpwgt, setgt, i1, V2I32>;
644 def: OpR_RR_pat<A2_vcmpwgt, setgt, v2i1, V2I32>;
645 def: OpR_RR_pat<A2_vcmpwgtu, RevCmp<setult>, i1, V2I32>;
646 def: OpR_RR_pat<A2_vcmpwgtu, RevCmp<setult>, v2i1, V2I32>;
647 def: OpR_RR_pat<A2_vcmpwgtu, setugt, i1, V2I32>;
648 def: OpR_RR_pat<A2_vcmpwgtu, setugt, v2i1, V2I32>;
650 def: OpR_RR_pat<F2_sfcmpeq, seteq, i1, F32>;
651 def: OpR_RR_pat<F2_sfcmpgt, setgt, i1, F32>;
652 def: OpR_RR_pat<F2_sfcmpge, setge, i1, F32>;
653 def: OpR_RR_pat<F2_sfcmpeq, setoeq, i1, F32>;
654 def: OpR_RR_pat<F2_sfcmpgt, setogt, i1, F32>;
655 def: OpR_RR_pat<F2_sfcmpge, setoge, i1, F32>;
656 def: OpR_RR_pat<F2_sfcmpgt, RevCmp<setolt>, i1, F32>;
657 def: OpR_RR_pat<F2_sfcmpge, RevCmp<setole>, i1, F32>;
658 def: OpR_RR_pat<F2_sfcmpgt, RevCmp<setlt>, i1, F32>;
659 def: OpR_RR_pat<F2_sfcmpge, RevCmp<setle>, i1, F32>;
660 def: OpR_RR_pat<F2_sfcmpuo, setuo, i1, F32>;
662 def: OpR_RR_pat<F2_dfcmpeq, seteq, i1, F64>;
663 def: OpR_RR_pat<F2_dfcmpgt, setgt, i1, F64>;
664 def: OpR_RR_pat<F2_dfcmpge, setge, i1, F64>;
665 def: OpR_RR_pat<F2_dfcmpeq, setoeq, i1, F64>;
666 def: OpR_RR_pat<F2_dfcmpgt, setogt, i1, F64>;
667 def: OpR_RR_pat<F2_dfcmpge, setoge, i1, F64>;
668 def: OpR_RR_pat<F2_dfcmpgt, RevCmp<setolt>, i1, F64>;
669 def: OpR_RR_pat<F2_dfcmpge, RevCmp<setole>, i1, F64>;
670 def: OpR_RR_pat<F2_dfcmpgt, RevCmp<setlt>, i1, F64>;
671 def: OpR_RR_pat<F2_dfcmpge, RevCmp<setle>, i1, F64>;
672 def: OpR_RR_pat<F2_dfcmpuo, setuo, i1, F64>;
674 // Avoid C4_cmpneqi, C4_cmpltei, C4_cmplteui, since they cannot form compounds.
676 def: Pat<(i1 (setne I32:$Rs, anyimm:$u5)),
677 (C2_not (C2_cmpeqi I32:$Rs, imm:$u5))>;
678 def: Pat<(i1 (setle I32:$Rs, anyimm:$u5)),
679 (C2_not (C2_cmpgti I32:$Rs, imm:$u5))>;
680 def: Pat<(i1 (setule I32:$Rs, anyimm:$u5)),
681 (C2_not (C2_cmpgtui I32:$Rs, imm:$u5))>;
683 class OpmR_RR_pat<PatFrag Output, PatFrag Op, ValueType ResType,
684 PatFrag RsPred, PatFrag RtPred = RsPred>
685 : Pat<(ResType (Op RsPred:$Rs, RtPred:$Rt)),
686 (Output RsPred:$Rs, RtPred:$Rt)>;
688 class Outn<InstHexagon MI>
689 : OutPatFrag<(ops node:$Rs, node:$Rt),
690 (C2_not (MI $Rs, $Rt))>;
692 def: OpmR_RR_pat<Outn<C2_cmpeq>, setne, i1, I32>;
693 def: OpmR_RR_pat<Outn<C2_cmpgt>, setle, i1, I32>;
694 def: OpmR_RR_pat<Outn<C2_cmpgtu>, setule, i1, I32>;
695 def: OpmR_RR_pat<Outn<C2_cmpgt>, RevCmp<setge>, i1, I32>;
696 def: OpmR_RR_pat<Outn<C2_cmpgtu>, RevCmp<setuge>, i1, I32>;
697 def: OpmR_RR_pat<Outn<C2_cmpeqp>, setne, i1, I64>;
698 def: OpmR_RR_pat<Outn<C2_cmpgtp>, setle, i1, I64>;
699 def: OpmR_RR_pat<Outn<C2_cmpgtup>, setule, i1, I64>;
700 def: OpmR_RR_pat<Outn<C2_cmpgtp>, RevCmp<setge>, i1, I64>;
701 def: OpmR_RR_pat<Outn<C2_cmpgtup>, RevCmp<setuge>, i1, I64>;
702 def: OpmR_RR_pat<Outn<A2_vcmpbeq>, setne, v8i1, V8I8>;
703 def: OpmR_RR_pat<Outn<A4_vcmpbgt>, setle, v8i1, V8I8>;
704 def: OpmR_RR_pat<Outn<A2_vcmpbgtu>, setule, v8i1, V8I8>;
705 def: OpmR_RR_pat<Outn<A4_vcmpbgt>, RevCmp<setge>, v8i1, V8I8>;
706 def: OpmR_RR_pat<Outn<A2_vcmpbgtu>, RevCmp<setuge>, v8i1, V8I8>;
707 def: OpmR_RR_pat<Outn<A2_vcmpheq>, setne, v4i1, V4I16>;
708 def: OpmR_RR_pat<Outn<A2_vcmphgt>, setle, v4i1, V4I16>;
709 def: OpmR_RR_pat<Outn<A2_vcmphgtu>, setule, v4i1, V4I16>;
710 def: OpmR_RR_pat<Outn<A2_vcmphgt>, RevCmp<setge>, v4i1, V4I16>;
711 def: OpmR_RR_pat<Outn<A2_vcmphgtu>, RevCmp<setuge>, v4i1, V4I16>;
712 def: OpmR_RR_pat<Outn<A2_vcmpweq>, setne, v2i1, V2I32>;
713 def: OpmR_RR_pat<Outn<A2_vcmpwgt>, setle, v2i1, V2I32>;
714 def: OpmR_RR_pat<Outn<A2_vcmpwgtu>, setule, v2i1, V2I32>;
715 def: OpmR_RR_pat<Outn<A2_vcmpwgt>, RevCmp<setge>, v2i1, V2I32>;
716 def: OpmR_RR_pat<Outn<A2_vcmpwgtu>, RevCmp<setuge>, v2i1, V2I32>;
718 let AddedComplexity = 100 in {
719 def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt), 255), 0)),
720 (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt)>;
721 def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt), 255), 0)),
722 (C2_not (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt))>;
723 def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt), 65535), 0)),
724 (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt)>;
725 def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt), 65535), 0)),
726 (C2_not (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt))>;
729 // PatFrag for AsserZext which takes the original type as a parameter.
730 def SDTAssertZext: SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisSameAs<0,1>]>;
731 def AssertZextSD: SDNode<"ISD::AssertZext", SDTAssertZext>;
732 class AssertZext<ValueType T>: PatFrag<(ops node:$A), (AssertZextSD $A, T)>;
734 multiclass Cmpb_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
735 PatLeaf ImmPred, int Mask> {
736 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
737 (MI I32:$Rs, imm:$I)>;
738 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
739 (MI I32:$Rs, imm:$I)>;
742 multiclass CmpbN_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
743 PatLeaf ImmPred, int Mask> {
744 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
745 (C2_not (MI I32:$Rs, imm:$I))>;
746 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
747 (C2_not (MI I32:$Rs, imm:$I))>;
750 multiclass CmpbND_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
751 PatLeaf ImmPred, int Mask> {
752 def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
753 (C2_not (MI I32:$Rs, (UDEC1 imm:$I)))>;
754 def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
755 (C2_not (MI I32:$Rs, (UDEC1 imm:$I)))>;
758 let AddedComplexity = 200 in {
759 defm: Cmpb_pat <A4_cmpbeqi, seteq, AssertZext<i8>, IsUGT<8,31>, 255>;
760 defm: CmpbN_pat <A4_cmpbeqi, setne, AssertZext<i8>, IsUGT<8,31>, 255>;
761 defm: Cmpb_pat <A4_cmpbgtui, setugt, AssertZext<i8>, IsUGT<32,31>, 255>;
762 defm: CmpbN_pat <A4_cmpbgtui, setule, AssertZext<i8>, IsUGT<32,31>, 255>;
763 defm: Cmpb_pat <A4_cmphgtui, setugt, AssertZext<i16>, IsUGT<32,31>, 65535>;
764 defm: CmpbN_pat <A4_cmphgtui, setule, AssertZext<i16>, IsUGT<32,31>, 65535>;
765 defm: CmpbND_pat<A4_cmpbgtui, setult, AssertZext<i8>, IsUGT<32,32>, 255>;
766 defm: CmpbND_pat<A4_cmphgtui, setult, AssertZext<i16>, IsUGT<32,32>, 65535>;
769 def: Pat<(i32 (zext (i1 (seteq I32:$Rs, I32:$Rt)))),
770 (A4_rcmpeq I32:$Rs, I32:$Rt)>;
771 def: Pat<(i32 (zext (i1 (setne I32:$Rs, I32:$Rt)))),
772 (A4_rcmpneq I32:$Rs, I32:$Rt)>;
773 def: Pat<(i32 (zext (i1 (seteq I32:$Rs, anyimm:$s8)))),
774 (A4_rcmpeqi I32:$Rs, imm:$s8)>;
775 def: Pat<(i32 (zext (i1 (setne I32:$Rs, anyimm:$s8)))),
776 (A4_rcmpneqi I32:$Rs, imm:$s8)>;
778 def: Pat<(i1 (seteq I1:$Ps, (i1 -1))), (I1:$Ps)>;
779 def: Pat<(i1 (setne I1:$Ps, (i1 -1))), (C2_not I1:$Ps)>;
780 def: Pat<(i1 (seteq I1:$Ps, I1:$Pt)), (C2_xor I1:$Ps, (C2_not I1:$Pt))>;
781 def: Pat<(i1 (setne I1:$Ps, I1:$Pt)), (C2_xor I1:$Ps, I1:$Pt)>;
783 // Floating-point comparisons with checks for ordered/unordered status.
785 class T3<InstHexagon MI1, InstHexagon MI2, InstHexagon MI3>
786 : OutPatFrag<(ops node:$Rs, node:$Rt),
787 (MI1 (MI2 $Rs, $Rt), (MI3 $Rs, $Rt))>;
789 class Cmpuf<InstHexagon MI>: T3<C2_or, F2_sfcmpuo, MI>;
790 class Cmpud<InstHexagon MI>: T3<C2_or, F2_dfcmpuo, MI>;
792 class Cmpufn<InstHexagon MI>: T3<C2_orn, F2_sfcmpuo, MI>;
793 class Cmpudn<InstHexagon MI>: T3<C2_orn, F2_dfcmpuo, MI>;
795 def: OpmR_RR_pat<Cmpuf<F2_sfcmpeq>, setueq, i1, F32>;
796 def: OpmR_RR_pat<Cmpuf<F2_sfcmpge>, setuge, i1, F32>;
797 def: OpmR_RR_pat<Cmpuf<F2_sfcmpgt>, setugt, i1, F32>;
798 def: OpmR_RR_pat<Cmpuf<F2_sfcmpge>, RevCmp<setule>, i1, F32>;
799 def: OpmR_RR_pat<Cmpuf<F2_sfcmpgt>, RevCmp<setult>, i1, F32>;
800 def: OpmR_RR_pat<Cmpufn<F2_sfcmpeq>, setune, i1, F32>;
802 def: OpmR_RR_pat<Cmpud<F2_dfcmpeq>, setueq, i1, F64>;
803 def: OpmR_RR_pat<Cmpud<F2_dfcmpge>, setuge, i1, F64>;
804 def: OpmR_RR_pat<Cmpud<F2_dfcmpgt>, setugt, i1, F64>;
805 def: OpmR_RR_pat<Cmpud<F2_dfcmpge>, RevCmp<setule>, i1, F64>;
806 def: OpmR_RR_pat<Cmpud<F2_dfcmpgt>, RevCmp<setult>, i1, F64>;
807 def: OpmR_RR_pat<Cmpudn<F2_dfcmpeq>, setune, i1, F64>;
809 def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setone, i1, F32>;
810 def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setne, i1, F32>;
812 def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setone, i1, F64>;
813 def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setne, i1, F64>;
815 def: OpmR_RR_pat<Outn<F2_sfcmpuo>, seto, i1, F32>;
816 def: OpmR_RR_pat<Outn<F2_dfcmpuo>, seto, i1, F64>;
819 // --(6) Select ----------------------------------------------------------
822 def: Pat<(select I1:$Pu, I32:$Rs, I32:$Rt),
823 (C2_mux I1:$Pu, I32:$Rs, I32:$Rt)>;
824 def: Pat<(select I1:$Pu, anyimm:$s8, I32:$Rs),
825 (C2_muxri I1:$Pu, imm:$s8, I32:$Rs)>;
826 def: Pat<(select I1:$Pu, I32:$Rs, anyimm:$s8),
827 (C2_muxir I1:$Pu, I32:$Rs, imm:$s8)>;
828 def: Pat<(select I1:$Pu, anyimm:$s8, s8_0ImmPred:$S8),
829 (C2_muxii I1:$Pu, imm:$s8, imm:$S8)>;
831 def: Pat<(select (not I1:$Pu), I32:$Rs, I32:$Rt),
832 (C2_mux I1:$Pu, I32:$Rt, I32:$Rs)>;
833 def: Pat<(select (not I1:$Pu), s8_0ImmPred:$S8, anyimm:$s8),
834 (C2_muxii I1:$Pu, imm:$s8, imm:$S8)>;
835 def: Pat<(select (not I1:$Pu), anyimm:$s8, I32:$Rs),
836 (C2_muxir I1:$Pu, I32:$Rs, imm:$s8)>;
837 def: Pat<(select (not I1:$Pu), I32:$Rs, anyimm:$s8),
838 (C2_muxri I1:$Pu, imm:$s8, I32:$Rs)>;
840 // Map from a 64-bit select to an emulated 64-bit mux.
841 // Hexagon does not support 64-bit MUXes; so emulate with combines.
842 def: Pat<(select I1:$Pu, I64:$Rs, I64:$Rt),
843 (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
844 (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
846 def: Pat<(select I1:$Pu, F32:$Rs, f32ImmPred:$I),
847 (C2_muxir I1:$Pu, F32:$Rs, (ftoi $I))>;
848 def: Pat<(select I1:$Pu, f32ImmPred:$I, F32:$Rt),
849 (C2_muxri I1:$Pu, (ftoi $I), F32:$Rt)>;
850 def: Pat<(select I1:$Pu, F32:$Rs, F32:$Rt),
851 (C2_mux I1:$Pu, F32:$Rs, F32:$Rt)>;
852 def: Pat<(select I1:$Pu, F64:$Rs, F64:$Rt),
853 (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
854 (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
856 def: Pat<(select (i1 (setult F32:$Ra, F32:$Rb)), F32:$Rs, F32:$Rt),
857 (C2_mux (F2_sfcmpgt F32:$Rb, F32:$Ra), F32:$Rs, F32:$Rt)>;
858 def: Pat<(select (i1 (setult F64:$Ra, F64:$Rb)), F64:$Rs, F64:$Rt),
859 (C2_vmux (F2_dfcmpgt F64:$Rb, F64:$Ra), F64:$Rs, F64:$Rt)>;
861 def: Pat<(select (not I1:$Pu), f32ImmPred:$I, F32:$Rs),
862 (C2_muxir I1:$Pu, F32:$Rs, (ftoi $I))>;
863 def: Pat<(select (not I1:$Pu), F32:$Rt, f32ImmPred:$I),
864 (C2_muxri I1:$Pu, (ftoi $I), F32:$Rt)>;
866 def: Pat<(vselect V8I1:$Pu, V8I8:$Rs, V8I8:$Rt),
867 (C2_vmux V8I1:$Pu, V8I8:$Rs, V8I8:$Rt)>;
868 def: Pat<(vselect V4I1:$Pu, V4I16:$Rs, V4I16:$Rt),
869 (C2_vmux V4I1:$Pu, V4I16:$Rs, V4I16:$Rt)>;
870 def: Pat<(vselect V2I1:$Pu, V2I32:$Rs, V2I32:$Rt),
871 (C2_vmux V2I1:$Pu, V2I32:$Rs, V2I32:$Rt)>;
873 def: Pat<(vselect (pnot V8I1:$Pu), V8I8:$Rs, V8I8:$Rt),
874 (C2_vmux V8I1:$Pu, V8I8:$Rt, V8I8:$Rs)>;
875 def: Pat<(vselect (pnot V4I1:$Pu), V4I16:$Rs, V4I16:$Rt),
876 (C2_vmux V4I1:$Pu, V4I16:$Rt, V4I16:$Rs)>;
877 def: Pat<(vselect (pnot V2I1:$Pu), V2I32:$Rs, V2I32:$Rt),
878 (C2_vmux V2I1:$Pu, V2I32:$Rt, V2I32:$Rs)>;
881 // From LegalizeDAG.cpp: (Pu ? Pv : Pw) <=> (Pu & Pv) | (!Pu & Pw).
882 def: Pat<(select I1:$Pu, I1:$Pv, I1:$Pw),
883 (C2_or (C2_and I1:$Pu, I1:$Pv),
884 (C2_andn I1:$Pw, I1:$Pu))>;
887 def IsPosHalf : PatLeaf<(i32 IntRegs:$a), [{
888 return isPositiveHalfWord(N);
891 multiclass SelMinMax16_pats<PatFrag CmpOp, InstHexagon InstA,
893 def: Pat<(sext_inreg (select (i1 (CmpOp IsPosHalf:$Rs, IsPosHalf:$Rt)),
894 IsPosHalf:$Rs, IsPosHalf:$Rt), i16),
895 (InstA IntRegs:$Rs, IntRegs:$Rt)>;
896 def: Pat<(sext_inreg (select (i1 (CmpOp IsPosHalf:$Rs, IsPosHalf:$Rt)),
897 IsPosHalf:$Rt, IsPosHalf:$Rs), i16),
898 (InstB IntRegs:$Rs, IntRegs:$Rt)>;
901 let AddedComplexity = 200 in {
902 defm: SelMinMax16_pats<setge, A2_max, A2_min>;
903 defm: SelMinMax16_pats<setgt, A2_max, A2_min>;
904 defm: SelMinMax16_pats<setle, A2_min, A2_max>;
905 defm: SelMinMax16_pats<setlt, A2_min, A2_max>;
906 defm: SelMinMax16_pats<setuge, A2_maxu, A2_minu>;
907 defm: SelMinMax16_pats<setugt, A2_maxu, A2_minu>;
908 defm: SelMinMax16_pats<setule, A2_minu, A2_maxu>;
909 defm: SelMinMax16_pats<setult, A2_minu, A2_maxu>;
912 let AddedComplexity = 200 in {
913 defm: MinMax_pats<A2_min, A2_max, select, setgt, i1, I32>;
914 defm: MinMax_pats<A2_min, A2_max, select, setge, i1, I32>;
915 defm: MinMax_pats<A2_max, A2_min, select, setlt, i1, I32>;
916 defm: MinMax_pats<A2_max, A2_min, select, setle, i1, I32>;
917 defm: MinMax_pats<A2_minu, A2_maxu, select, setugt, i1, I32>;
918 defm: MinMax_pats<A2_minu, A2_maxu, select, setuge, i1, I32>;
919 defm: MinMax_pats<A2_maxu, A2_minu, select, setult, i1, I32>;
920 defm: MinMax_pats<A2_maxu, A2_minu, select, setule, i1, I32>;
922 defm: MinMax_pats<A2_minp, A2_maxp, select, setgt, i1, I64>;
923 defm: MinMax_pats<A2_minp, A2_maxp, select, setge, i1, I64>;
924 defm: MinMax_pats<A2_maxp, A2_minp, select, setlt, i1, I64>;
925 defm: MinMax_pats<A2_maxp, A2_minp, select, setle, i1, I64>;
926 defm: MinMax_pats<A2_minup, A2_maxup, select, setugt, i1, I64>;
927 defm: MinMax_pats<A2_minup, A2_maxup, select, setuge, i1, I64>;
928 defm: MinMax_pats<A2_maxup, A2_minup, select, setult, i1, I64>;
929 defm: MinMax_pats<A2_maxup, A2_minup, select, setule, i1, I64>;
932 let AddedComplexity = 100 in {
933 defm: MinMax_pats<F2_sfmin, F2_sfmax, select, setogt, i1, F32>;
934 defm: MinMax_pats<F2_sfmin, F2_sfmax, select, setoge, i1, F32>;
935 defm: MinMax_pats<F2_sfmax, F2_sfmin, select, setolt, i1, F32>;
936 defm: MinMax_pats<F2_sfmax, F2_sfmin, select, setole, i1, F32>;
939 let AddedComplexity = 100, Predicates = [HasV67] in {
940 defm: MinMax_pats<F2_dfmin, F2_dfmax, select, setogt, i1, F64>;
941 defm: MinMax_pats<F2_dfmin, F2_dfmax, select, setoge, i1, F64>;
942 defm: MinMax_pats<F2_dfmax, F2_dfmin, select, setolt, i1, F64>;
943 defm: MinMax_pats<F2_dfmax, F2_dfmin, select, setole, i1, F64>;
946 defm: MinMax_pats<A2_vminb, A2_vmaxb, vselect, setgt, v8i1, V8I8>;
947 defm: MinMax_pats<A2_vminb, A2_vmaxb, vselect, setge, v8i1, V8I8>;
948 defm: MinMax_pats<A2_vminh, A2_vmaxh, vselect, setgt, v4i1, V4I16>;
949 defm: MinMax_pats<A2_vminh, A2_vmaxh, vselect, setge, v4i1, V4I16>;
950 defm: MinMax_pats<A2_vminw, A2_vmaxw, vselect, setgt, v2i1, V2I32>;
951 defm: MinMax_pats<A2_vminw, A2_vmaxw, vselect, setge, v2i1, V2I32>;
952 defm: MinMax_pats<A2_vminub, A2_vmaxub, vselect, setugt, v8i1, V8I8>;
953 defm: MinMax_pats<A2_vminub, A2_vmaxub, vselect, setuge, v8i1, V8I8>;
954 defm: MinMax_pats<A2_vminuh, A2_vmaxuh, vselect, setugt, v4i1, V4I16>;
955 defm: MinMax_pats<A2_vminuh, A2_vmaxuh, vselect, setuge, v4i1, V4I16>;
956 defm: MinMax_pats<A2_vminuw, A2_vmaxuw, vselect, setugt, v2i1, V2I32>;
957 defm: MinMax_pats<A2_vminuw, A2_vmaxuw, vselect, setuge, v2i1, V2I32>;
959 // --(7) Insert/extract --------------------------------------------------
962 def SDTHexagonINSERT:
963 SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>,
964 SDTCisInt<0>, SDTCisVT<3, i32>, SDTCisVT<4, i32>]>;
965 def HexagonINSERT: SDNode<"HexagonISD::INSERT", SDTHexagonINSERT>;
967 let AddedComplexity = 10 in {
968 def: Pat<(HexagonINSERT I32:$Rs, I32:$Rt, u5_0ImmPred:$u1, u5_0ImmPred:$u2),
969 (S2_insert I32:$Rs, I32:$Rt, imm:$u1, imm:$u2)>;
970 def: Pat<(HexagonINSERT I64:$Rs, I64:$Rt, u6_0ImmPred:$u1, u6_0ImmPred:$u2),
971 (S2_insertp I64:$Rs, I64:$Rt, imm:$u1, imm:$u2)>;
973 def: Pat<(HexagonINSERT I32:$Rs, I32:$Rt, I32:$Width, I32:$Off),
974 (S2_insert_rp I32:$Rs, I32:$Rt, (Combinew $Width, $Off))>;
975 def: Pat<(HexagonINSERT I64:$Rs, I64:$Rt, I32:$Width, I32:$Off),
976 (S2_insertp_rp I64:$Rs, I64:$Rt, (Combinew $Width, $Off))>;
978 def SDTHexagonEXTRACTU
979 : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisInt<0>, SDTCisInt<1>,
980 SDTCisVT<2, i32>, SDTCisVT<3, i32>]>;
981 def HexagonEXTRACTU: SDNode<"HexagonISD::EXTRACTU", SDTHexagonEXTRACTU>;
983 let AddedComplexity = 10 in {
984 def: Pat<(HexagonEXTRACTU I32:$Rs, u5_0ImmPred:$u5, u5_0ImmPred:$U5),
985 (S2_extractu I32:$Rs, imm:$u5, imm:$U5)>;
986 def: Pat<(HexagonEXTRACTU I64:$Rs, u6_0ImmPred:$u6, u6_0ImmPred:$U6),
987 (S2_extractup I64:$Rs, imm:$u6, imm:$U6)>;
989 def: Pat<(HexagonEXTRACTU I32:$Rs, I32:$Width, I32:$Off),
990 (S2_extractu_rp I32:$Rs, (Combinew $Width, $Off))>;
991 def: Pat<(HexagonEXTRACTU I64:$Rs, I32:$Width, I32:$Off),
992 (S2_extractup_rp I64:$Rs, (Combinew $Width, $Off))>;
994 def SDTHexagonVSPLAT:
995 SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVT<1, i32>]>;
997 def HexagonVSPLAT: SDNode<"HexagonISD::VSPLAT", SDTHexagonVSPLAT>;
999 def: Pat<(v4i8 (HexagonVSPLAT I32:$Rs)), (S2_vsplatrb I32:$Rs)>;
1000 def: Pat<(v4i16 (HexagonVSPLAT I32:$Rs)), (S2_vsplatrh I32:$Rs)>;
1001 def: Pat<(v2i32 (HexagonVSPLAT s8_0ImmPred:$s8)),
1002 (A2_combineii imm:$s8, imm:$s8)>;
1003 def: Pat<(v2i32 (HexagonVSPLAT I32:$Rs)), (Combinew I32:$Rs, I32:$Rs)>;
1005 let AddedComplexity = 10 in
1006 def: Pat<(v8i8 (HexagonVSPLAT I32:$Rs)), (S6_vsplatrbp I32:$Rs)>,
1008 def: Pat<(v8i8 (HexagonVSPLAT I32:$Rs)),
1009 (Combinew (S2_vsplatrb I32:$Rs), (S2_vsplatrb I32:$Rs))>;
1012 // --(8) Shift/permute ---------------------------------------------------
1015 def SDTHexagonI64I32I32: SDTypeProfile<1, 2,
1016 [SDTCisVT<0, i64>, SDTCisVT<1, i32>, SDTCisSameAs<1, 2>]>;
1018 def HexagonCOMBINE: SDNode<"HexagonISD::COMBINE", SDTHexagonI64I32I32>;
1020 def: Pat<(HexagonCOMBINE I32:$Rs, I32:$Rt), (Combinew $Rs, $Rt)>;
1022 // The complexity of the combines involving immediates should be greater
1023 // than the complexity of the combine with two registers.
1024 let AddedComplexity = 50 in {
1025 def: Pat<(HexagonCOMBINE I32:$Rs, anyimm:$s8),
1026 (A4_combineri IntRegs:$Rs, imm:$s8)>;
1027 def: Pat<(HexagonCOMBINE anyimm:$s8, I32:$Rs),
1028 (A4_combineir imm:$s8, IntRegs:$Rs)>;
1031 // The complexity of the combine with two immediates should be greater than
1032 // the complexity of a combine involving a register.
1033 let AddedComplexity = 75 in {
1034 def: Pat<(HexagonCOMBINE s8_0ImmPred:$s8, anyimm:$u6),
1035 (A4_combineii imm:$s8, imm:$u6)>;
1036 def: Pat<(HexagonCOMBINE anyimm:$s8, s8_0ImmPred:$S8),
1037 (A2_combineii imm:$s8, imm:$S8)>;
1040 def: Pat<(bswap I32:$Rs), (A2_swiz I32:$Rs)>;
1041 def: Pat<(bswap I64:$Rss), (Combinew (A2_swiz (LoReg $Rss)),
1042 (A2_swiz (HiReg $Rss)))>;
1044 def: Pat<(shl s6_0ImmPred:$s6, I32:$Rt), (S4_lsli imm:$s6, I32:$Rt)>;
1045 def: Pat<(shl I32:$Rs, (i32 16)), (A2_aslh I32:$Rs)>;
1046 def: Pat<(sra I32:$Rs, (i32 16)), (A2_asrh I32:$Rs)>;
1048 def: OpR_RI_pat<S2_asr_i_r, Sra, i32, I32, u5_0ImmPred>;
1049 def: OpR_RI_pat<S2_lsr_i_r, Srl, i32, I32, u5_0ImmPred>;
1050 def: OpR_RI_pat<S2_asl_i_r, Shl, i32, I32, u5_0ImmPred>;
1051 def: OpR_RI_pat<S2_asr_i_p, Sra, i64, I64, u6_0ImmPred>;
1052 def: OpR_RI_pat<S2_lsr_i_p, Srl, i64, I64, u6_0ImmPred>;
1053 def: OpR_RI_pat<S2_asl_i_p, Shl, i64, I64, u6_0ImmPred>;
1054 def: OpR_RI_pat<S2_asr_i_vh, Sra, v4i16, V4I16, u4_0ImmPred>;
1055 def: OpR_RI_pat<S2_lsr_i_vh, Srl, v4i16, V4I16, u4_0ImmPred>;
1056 def: OpR_RI_pat<S2_asl_i_vh, Shl, v4i16, V4I16, u4_0ImmPred>;
1057 def: OpR_RI_pat<S2_asr_i_vh, Sra, v2i32, V2I32, u5_0ImmPred>;
1058 def: OpR_RI_pat<S2_lsr_i_vh, Srl, v2i32, V2I32, u5_0ImmPred>;
1059 def: OpR_RI_pat<S2_asl_i_vh, Shl, v2i32, V2I32, u5_0ImmPred>;
1061 def: OpR_RR_pat<S2_asr_r_r, Sra, i32, I32, I32>;
1062 def: OpR_RR_pat<S2_lsr_r_r, Srl, i32, I32, I32>;
1063 def: OpR_RR_pat<S2_asl_r_r, Shl, i32, I32, I32>;
1064 def: OpR_RR_pat<S2_asr_r_p, Sra, i64, I64, I32>;
1065 def: OpR_RR_pat<S2_lsr_r_p, Srl, i64, I64, I32>;
1066 def: OpR_RR_pat<S2_asl_r_p, Shl, i64, I64, I32>;
1069 def IsMul8_U3: PatLeaf<(i32 imm), [{
1070 uint64_t V = N->getZExtValue();
1071 return V % 8 == 0 && isUInt<3>(V / 8);
1074 def Divu8: SDNodeXForm<imm, [{
1075 return CurDAG->getTargetConstant(N->getZExtValue() / 8, SDLoc(N), MVT::i32);
1078 // Funnel shift-left.
1079 def FShl32i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1080 (HiReg (S2_asl_i_p (Combinew $Rs, $Rt), $S))>;
1081 def FShl32r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1082 (HiReg (S2_asl_r_p (Combinew $Rs, $Rt), $Ru))>;
1084 def FShl64i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1085 (S2_lsr_i_p_or (S2_asl_i_p $Rt, $S), $Rs, (Subi<64> $S))>;
1086 def FShl64r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1087 (S2_lsr_r_p_or (S2_asl_r_p $Rt, $Ru), $Rs, (A2_subri 64, $Ru))>;
1089 // Combined SDNodeXForm: (Divu8 (Subi<64> $S))
1090 def Divu64_8: SDNodeXForm<imm, [{
1091 return CurDAG->getTargetConstant((64 - N->getSExtValue()) / 8,
1092 SDLoc(N), MVT::i32);
1096 let AddedComplexity = 100 in {
1097 def: Pat<(fshl I32:$Rs, I32:$Rt, (i32 16)),
1098 (A2_combine_lh I32:$Rs, I32:$Rt)>;
1099 def: Pat<(fshl I64:$Rs, I64:$Rt, IsMul8_U3:$S),
1100 (S2_valignib I64:$Rs, I64:$Rt, (Divu64_8 $S))>;
1103 let Predicates = [HasV60], AddedComplexity = 50 in {
1104 def: OpR_RI_pat<S6_rol_i_r, Rol, i32, I32, u5_0ImmPred>;
1105 def: OpR_RI_pat<S6_rol_i_p, Rol, i64, I64, u6_0ImmPred>;
1107 let AddedComplexity = 30 in {
1108 def: Pat<(rotl I32:$Rs, u5_0ImmPred:$S), (FShl32i $Rs, $Rs, imm:$S)>;
1109 def: Pat<(rotl I64:$Rs, u6_0ImmPred:$S), (FShl64i $Rs, $Rs, imm:$S)>;
1110 def: Pat<(fshl I32:$Rs, I32:$Rt, u5_0ImmPred:$S), (FShl32i $Rs, $Rt, imm:$S)>;
1111 def: Pat<(fshl I64:$Rs, I64:$Rt, u6_0ImmPred:$S), (FShl64i $Rs, $Rt, imm:$S)>;
1113 def: Pat<(rotl I32:$Rs, I32:$Rt), (FShl32r $Rs, $Rs, $Rt)>;
1114 def: Pat<(rotl I64:$Rs, I32:$Rt), (FShl64r $Rs, $Rs, $Rt)>;
1115 def: Pat<(fshl I32:$Rs, I32:$Rt, I32:$Ru), (FShl32r $Rs, $Rt, $Ru)>;
1116 def: Pat<(fshl I64:$Rs, I64:$Rt, I32:$Ru), (FShl64r $Rs, $Rt, $Ru)>;
1118 // Funnel shift-right.
1119 def FShr32i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1120 (LoReg (S2_lsr_i_p (Combinew $Rs, $Rt), $S))>;
1121 def FShr32r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1122 (LoReg (S2_lsr_r_p (Combinew $Rs, $Rt), $Ru))>;
1124 def FShr64i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1125 (S2_asl_i_p_or (S2_lsr_i_p $Rt, $S), $Rs, (Subi<64> $S))>;
1126 def FShr64r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1127 (S2_asl_r_p_or (S2_lsr_r_p $Rt, $Ru), $Rs, (A2_subri 64, $Ru))>;
1130 let AddedComplexity = 100 in {
1131 def: Pat<(fshr I32:$Rs, I32:$Rt, (i32 16)),
1132 (A2_combine_lh I32:$Rs, I32:$Rt)>;
1133 def: Pat<(fshr I64:$Rs, I64:$Rt, IsMul8_U3:$S),
1134 (S2_valignib I64:$Rs, I64:$Rt, (Divu8 $S))>;
1137 let Predicates = [HasV60], AddedComplexity = 50 in {
1138 def: Pat<(rotr I32:$Rs, u5_0ImmPred:$S), (S6_rol_i_r I32:$Rs, (Subi<32> $S))>;
1139 def: Pat<(rotr I64:$Rs, u6_0ImmPred:$S), (S6_rol_i_p I64:$Rs, (Subi<64> $S))>;
1141 let AddedComplexity = 30 in {
1142 def: Pat<(rotr I32:$Rs, u5_0ImmPred:$S), (FShr32i $Rs, $Rs, imm:$S)>;
1143 def: Pat<(rotr I64:$Rs, u6_0ImmPred:$S), (FShr64i $Rs, $Rs, imm:$S)>;
1144 def: Pat<(fshr I32:$Rs, I32:$Rt, u5_0ImmPred:$S), (FShr32i $Rs, $Rt, imm:$S)>;
1145 def: Pat<(fshr I64:$Rs, I64:$Rt, u6_0ImmPred:$S), (FShr64i $Rs, $Rt, imm:$S)>;
1147 def: Pat<(rotr I32:$Rs, I32:$Rt), (FShr32r $Rs, $Rs, $Rt)>;
1148 def: Pat<(rotr I64:$Rs, I32:$Rt), (FShr64r $Rs, $Rs, $Rt)>;
1149 def: Pat<(fshr I32:$Rs, I32:$Rt, I32:$Ru), (FShr32r $Rs, $Rt, $Ru)>;
1150 def: Pat<(fshr I64:$Rs, I64:$Rt, I32:$Ru), (FShr64r $Rs, $Rt, $Ru)>;
1153 def: Pat<(sra (add (sra I32:$Rs, u5_0ImmPred:$u5), 1), (i32 1)),
1154 (S2_asr_i_r_rnd I32:$Rs, imm:$u5)>;
1155 def: Pat<(sra (add (sra I64:$Rs, u6_0ImmPred:$u6), 1), (i32 1)),
1156 (S2_asr_i_p_rnd I64:$Rs, imm:$u6)>;
1158 // Prefer S2_addasl_rrri over S2_asl_i_r_acc.
1159 let AddedComplexity = 120 in
1160 def: Pat<(add I32:$Rt, (shl I32:$Rs, u3_0ImmPred:$u3)),
1161 (S2_addasl_rrri IntRegs:$Rt, IntRegs:$Rs, imm:$u3)>;
1163 let AddedComplexity = 100 in {
1164 def: AccRRI_pat<S2_asr_i_r_acc, Add, Su<Sra>, I32, u5_0ImmPred>;
1165 def: AccRRI_pat<S2_asr_i_r_nac, Sub, Su<Sra>, I32, u5_0ImmPred>;
1166 def: AccRRI_pat<S2_asr_i_r_and, And, Su<Sra>, I32, u5_0ImmPred>;
1167 def: AccRRI_pat<S2_asr_i_r_or, Or, Su<Sra>, I32, u5_0ImmPred>;
1169 def: AccRRI_pat<S2_asr_i_p_acc, Add, Su<Sra>, I64, u6_0ImmPred>;
1170 def: AccRRI_pat<S2_asr_i_p_nac, Sub, Su<Sra>, I64, u6_0ImmPred>;
1171 def: AccRRI_pat<S2_asr_i_p_and, And, Su<Sra>, I64, u6_0ImmPred>;
1172 def: AccRRI_pat<S2_asr_i_p_or, Or, Su<Sra>, I64, u6_0ImmPred>;
1174 def: AccRRI_pat<S2_lsr_i_r_acc, Add, Su<Srl>, I32, u5_0ImmPred>;
1175 def: AccRRI_pat<S2_lsr_i_r_nac, Sub, Su<Srl>, I32, u5_0ImmPred>;
1176 def: AccRRI_pat<S2_lsr_i_r_and, And, Su<Srl>, I32, u5_0ImmPred>;
1177 def: AccRRI_pat<S2_lsr_i_r_or, Or, Su<Srl>, I32, u5_0ImmPred>;
1178 def: AccRRI_pat<S2_lsr_i_r_xacc, Xor, Su<Srl>, I32, u5_0ImmPred>;
1180 def: AccRRI_pat<S2_lsr_i_p_acc, Add, Su<Srl>, I64, u6_0ImmPred>;
1181 def: AccRRI_pat<S2_lsr_i_p_nac, Sub, Su<Srl>, I64, u6_0ImmPred>;
1182 def: AccRRI_pat<S2_lsr_i_p_and, And, Su<Srl>, I64, u6_0ImmPred>;
1183 def: AccRRI_pat<S2_lsr_i_p_or, Or, Su<Srl>, I64, u6_0ImmPred>;
1184 def: AccRRI_pat<S2_lsr_i_p_xacc, Xor, Su<Srl>, I64, u6_0ImmPred>;
1186 def: AccRRI_pat<S2_asl_i_r_acc, Add, Su<Shl>, I32, u5_0ImmPred>;
1187 def: AccRRI_pat<S2_asl_i_r_nac, Sub, Su<Shl>, I32, u5_0ImmPred>;
1188 def: AccRRI_pat<S2_asl_i_r_and, And, Su<Shl>, I32, u5_0ImmPred>;
1189 def: AccRRI_pat<S2_asl_i_r_or, Or, Su<Shl>, I32, u5_0ImmPred>;
1190 def: AccRRI_pat<S2_asl_i_r_xacc, Xor, Su<Shl>, I32, u5_0ImmPred>;
1192 def: AccRRI_pat<S2_asl_i_p_acc, Add, Su<Shl>, I64, u6_0ImmPred>;
1193 def: AccRRI_pat<S2_asl_i_p_nac, Sub, Su<Shl>, I64, u6_0ImmPred>;
1194 def: AccRRI_pat<S2_asl_i_p_and, And, Su<Shl>, I64, u6_0ImmPred>;
1195 def: AccRRI_pat<S2_asl_i_p_or, Or, Su<Shl>, I64, u6_0ImmPred>;
1196 def: AccRRI_pat<S2_asl_i_p_xacc, Xor, Su<Shl>, I64, u6_0ImmPred>;
1198 let Predicates = [HasV60] in {
1199 def: AccRRI_pat<S6_rol_i_r_acc, Add, Su<Rol>, I32, u5_0ImmPred>;
1200 def: AccRRI_pat<S6_rol_i_r_nac, Sub, Su<Rol>, I32, u5_0ImmPred>;
1201 def: AccRRI_pat<S6_rol_i_r_and, And, Su<Rol>, I32, u5_0ImmPred>;
1202 def: AccRRI_pat<S6_rol_i_r_or, Or, Su<Rol>, I32, u5_0ImmPred>;
1203 def: AccRRI_pat<S6_rol_i_r_xacc, Xor, Su<Rol>, I32, u5_0ImmPred>;
1205 def: AccRRI_pat<S6_rol_i_p_acc, Add, Su<Rol>, I64, u6_0ImmPred>;
1206 def: AccRRI_pat<S6_rol_i_p_nac, Sub, Su<Rol>, I64, u6_0ImmPred>;
1207 def: AccRRI_pat<S6_rol_i_p_and, And, Su<Rol>, I64, u6_0ImmPred>;
1208 def: AccRRI_pat<S6_rol_i_p_or, Or, Su<Rol>, I64, u6_0ImmPred>;
1209 def: AccRRI_pat<S6_rol_i_p_xacc, Xor, Su<Rol>, I64, u6_0ImmPred>;
1213 let AddedComplexity = 100 in {
1214 def: AccRRR_pat<S2_asr_r_r_acc, Add, Su<Sra>, I32, I32, I32>;
1215 def: AccRRR_pat<S2_asr_r_r_nac, Sub, Su<Sra>, I32, I32, I32>;
1216 def: AccRRR_pat<S2_asr_r_r_and, And, Su<Sra>, I32, I32, I32>;
1217 def: AccRRR_pat<S2_asr_r_r_or, Or, Su<Sra>, I32, I32, I32>;
1219 def: AccRRR_pat<S2_asr_r_p_acc, Add, Su<Sra>, I64, I64, I32>;
1220 def: AccRRR_pat<S2_asr_r_p_nac, Sub, Su<Sra>, I64, I64, I32>;
1221 def: AccRRR_pat<S2_asr_r_p_and, And, Su<Sra>, I64, I64, I32>;
1222 def: AccRRR_pat<S2_asr_r_p_or, Or, Su<Sra>, I64, I64, I32>;
1223 def: AccRRR_pat<S2_asr_r_p_xor, Xor, Su<Sra>, I64, I64, I32>;
1225 def: AccRRR_pat<S2_lsr_r_r_acc, Add, Su<Srl>, I32, I32, I32>;
1226 def: AccRRR_pat<S2_lsr_r_r_nac, Sub, Su<Srl>, I32, I32, I32>;
1227 def: AccRRR_pat<S2_lsr_r_r_and, And, Su<Srl>, I32, I32, I32>;
1228 def: AccRRR_pat<S2_lsr_r_r_or, Or, Su<Srl>, I32, I32, I32>;
1230 def: AccRRR_pat<S2_lsr_r_p_acc, Add, Su<Srl>, I64, I64, I32>;
1231 def: AccRRR_pat<S2_lsr_r_p_nac, Sub, Su<Srl>, I64, I64, I32>;
1232 def: AccRRR_pat<S2_lsr_r_p_and, And, Su<Srl>, I64, I64, I32>;
1233 def: AccRRR_pat<S2_lsr_r_p_or, Or, Su<Srl>, I64, I64, I32>;
1234 def: AccRRR_pat<S2_lsr_r_p_xor, Xor, Su<Srl>, I64, I64, I32>;
1236 def: AccRRR_pat<S2_asl_r_r_acc, Add, Su<Shl>, I32, I32, I32>;
1237 def: AccRRR_pat<S2_asl_r_r_nac, Sub, Su<Shl>, I32, I32, I32>;
1238 def: AccRRR_pat<S2_asl_r_r_and, And, Su<Shl>, I32, I32, I32>;
1239 def: AccRRR_pat<S2_asl_r_r_or, Or, Su<Shl>, I32, I32, I32>;
1241 def: AccRRR_pat<S2_asl_r_p_acc, Add, Su<Shl>, I64, I64, I32>;
1242 def: AccRRR_pat<S2_asl_r_p_nac, Sub, Su<Shl>, I64, I64, I32>;
1243 def: AccRRR_pat<S2_asl_r_p_and, And, Su<Shl>, I64, I64, I32>;
1244 def: AccRRR_pat<S2_asl_r_p_or, Or, Su<Shl>, I64, I64, I32>;
1245 def: AccRRR_pat<S2_asl_r_p_xor, Xor, Su<Shl>, I64, I64, I32>;
1249 class OpshIRI_pat<InstHexagon MI, PatFrag Op, PatFrag ShOp,
1250 PatFrag RegPred, PatFrag ImmPred>
1251 : Pat<(Op anyimm:$u8, (ShOp RegPred:$Rs, ImmPred:$U5)),
1252 (MI anyimm:$u8, RegPred:$Rs, imm:$U5)>;
1254 let AddedComplexity = 200, Predicates = [UseCompound] in {
1255 def: OpshIRI_pat<S4_addi_asl_ri, Add, Su<Shl>, I32, u5_0ImmPred>;
1256 def: OpshIRI_pat<S4_addi_lsr_ri, Add, Su<Srl>, I32, u5_0ImmPred>;
1257 def: OpshIRI_pat<S4_subi_asl_ri, Sub, Su<Shl>, I32, u5_0ImmPred>;
1258 def: OpshIRI_pat<S4_subi_lsr_ri, Sub, Su<Srl>, I32, u5_0ImmPred>;
1259 def: OpshIRI_pat<S4_andi_asl_ri, And, Su<Shl>, I32, u5_0ImmPred>;
1260 def: OpshIRI_pat<S4_andi_lsr_ri, And, Su<Srl>, I32, u5_0ImmPred>;
1261 def: OpshIRI_pat<S4_ori_asl_ri, Or, Su<Shl>, I32, u5_0ImmPred>;
1262 def: OpshIRI_pat<S4_ori_lsr_ri, Or, Su<Srl>, I32, u5_0ImmPred>;
1265 // Prefer this pattern to S2_asl_i_p_or for the special case of joining
1266 // two 32-bit words into a 64-bit word.
1267 let AddedComplexity = 200 in
1268 def: Pat<(or (shl (Aext64 I32:$a), (i32 32)), (Zext64 I32:$b)),
1269 (Combinew I32:$a, I32:$b)>;
1271 def: Pat<(or (or (or (shl (Zext64 (and I32:$b, (i32 65535))), (i32 16)),
1272 (Zext64 (and I32:$a, (i32 65535)))),
1273 (shl (Aext64 (and I32:$c, (i32 65535))), (i32 32))),
1274 (shl (Aext64 I32:$d), (i32 48))),
1275 (Combinew (A2_combine_ll I32:$d, I32:$c),
1276 (A2_combine_ll I32:$b, I32:$a))>;
1278 let AddedComplexity = 200 in {
1279 def: Pat<(or (shl I32:$Rt, (i32 16)), (and I32:$Rs, (i32 65535))),
1280 (A2_combine_ll I32:$Rt, I32:$Rs)>;
1281 def: Pat<(or (shl I32:$Rt, (i32 16)), (srl I32:$Rs, (i32 16))),
1282 (A2_combine_lh I32:$Rt, I32:$Rs)>;
1283 def: Pat<(or (and I32:$Rt, (i32 268431360)), (and I32:$Rs, (i32 65535))),
1284 (A2_combine_hl I32:$Rt, I32:$Rs)>;
1285 def: Pat<(or (and I32:$Rt, (i32 268431360)), (srl I32:$Rs, (i32 16))),
1286 (A2_combine_hh I32:$Rt, I32:$Rs)>;
1289 def SDTHexagonVShift
1290 : SDTypeProfile<1, 2, [SDTCisSameAs<0, 1>, SDTCisVec<0>, SDTCisVT<2, i32>]>;
1292 def HexagonVASL: SDNode<"HexagonISD::VASL", SDTHexagonVShift>;
1293 def HexagonVASR: SDNode<"HexagonISD::VASR", SDTHexagonVShift>;
1294 def HexagonVLSR: SDNode<"HexagonISD::VLSR", SDTHexagonVShift>;
1296 def: OpR_RI_pat<S2_asl_i_vw, pf2<HexagonVASL>, v2i32, V2I32, u5_0ImmPred>;
1297 def: OpR_RI_pat<S2_asl_i_vh, pf2<HexagonVASL>, v4i16, V4I16, u4_0ImmPred>;
1298 def: OpR_RI_pat<S2_asr_i_vw, pf2<HexagonVASR>, v2i32, V2I32, u5_0ImmPred>;
1299 def: OpR_RI_pat<S2_asr_i_vh, pf2<HexagonVASR>, v4i16, V4I16, u4_0ImmPred>;
1300 def: OpR_RI_pat<S2_lsr_i_vw, pf2<HexagonVLSR>, v2i32, V2I32, u5_0ImmPred>;
1301 def: OpR_RI_pat<S2_lsr_i_vh, pf2<HexagonVLSR>, v4i16, V4I16, u4_0ImmPred>;
1303 def: OpR_RR_pat<S2_asl_r_vw, pf2<HexagonVASL>, v2i32, V2I32, I32>;
1304 def: OpR_RR_pat<S2_asl_r_vh, pf2<HexagonVASL>, v4i16, V4I16, I32>;
1305 def: OpR_RR_pat<S2_asr_r_vw, pf2<HexagonVASR>, v2i32, V2I32, I32>;
1306 def: OpR_RR_pat<S2_asr_r_vh, pf2<HexagonVASR>, v4i16, V4I16, I32>;
1307 def: OpR_RR_pat<S2_lsr_r_vw, pf2<HexagonVLSR>, v2i32, V2I32, I32>;
1308 def: OpR_RR_pat<S2_lsr_r_vh, pf2<HexagonVLSR>, v4i16, V4I16, I32>;
1310 def: Pat<(sra V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c))),
1311 (S2_asr_i_vw V2I32:$b, imm:$c)>;
1312 def: Pat<(srl V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c))),
1313 (S2_lsr_i_vw V2I32:$b, imm:$c)>;
1314 def: Pat<(shl V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c))),
1315 (S2_asl_i_vw V2I32:$b, imm:$c)>;
1316 def: Pat<(sra V4I16:$b, (v4i16 (HexagonVSPLAT u4_0ImmPred:$c))),
1317 (S2_asr_i_vh V4I16:$b, imm:$c)>;
1318 def: Pat<(srl V4I16:$b, (v4i16 (HexagonVSPLAT u4_0ImmPred:$c))),
1319 (S2_lsr_i_vh V4I16:$b, imm:$c)>;
1320 def: Pat<(shl V4I16:$b, (v4i16 (HexagonVSPLAT u4_0ImmPred:$c))),
1321 (S2_asl_i_vh V4I16:$b, imm:$c)>;
1323 def: Pat<(HexagonVASR V2I16:$Rs, u4_0ImmPred:$S),
1324 (LoReg (S2_asr_i_vh (ToAext64 $Rs), imm:$S))>;
1325 def: Pat<(HexagonVASL V2I16:$Rs, u4_0ImmPred:$S),
1326 (LoReg (S2_asl_i_vh (ToAext64 $Rs), imm:$S))>;
1327 def: Pat<(HexagonVLSR V2I16:$Rs, u4_0ImmPred:$S),
1328 (LoReg (S2_lsr_i_vh (ToAext64 $Rs), imm:$S))>;
1329 def: Pat<(HexagonVASR V2I16:$Rs, I32:$Rt),
1330 (LoReg (S2_asr_i_vh (ToAext64 $Rs), I32:$Rt))>;
1331 def: Pat<(HexagonVASL V2I16:$Rs, I32:$Rt),
1332 (LoReg (S2_asl_i_vh (ToAext64 $Rs), I32:$Rt))>;
1333 def: Pat<(HexagonVLSR V2I16:$Rs, I32:$Rt),
1334 (LoReg (S2_lsr_i_vh (ToAext64 $Rs), I32:$Rt))>;
1337 // --(9) Arithmetic/bitwise ----------------------------------------------
1340 def: Pat<(abs I32:$Rs), (A2_abs I32:$Rs)>;
1341 def: Pat<(abs I64:$Rs), (A2_absp I64:$Rs)>;
1342 def: Pat<(not I32:$Rs), (A2_subri -1, I32:$Rs)>;
1343 def: Pat<(not I64:$Rs), (A2_notp I64:$Rs)>;
1344 def: Pat<(ineg I64:$Rs), (A2_negp I64:$Rs)>;
1346 def: Pat<(fabs F32:$Rs), (S2_clrbit_i F32:$Rs, 31)>;
1347 def: Pat<(fneg F32:$Rs), (S2_togglebit_i F32:$Rs, 31)>;
1349 def: Pat<(fabs F64:$Rs),
1350 (Combinew (S2_clrbit_i (HiReg $Rs), 31),
1351 (i32 (LoReg $Rs)))>;
1352 def: Pat<(fneg F64:$Rs),
1353 (Combinew (S2_togglebit_i (HiReg $Rs), 31),
1354 (i32 (LoReg $Rs)))>;
1356 def: Pat<(add I32:$Rs, anyimm:$s16), (A2_addi I32:$Rs, imm:$s16)>;
1357 def: Pat<(or I32:$Rs, anyimm:$s10), (A2_orir I32:$Rs, imm:$s10)>;
1358 def: Pat<(and I32:$Rs, anyimm:$s10), (A2_andir I32:$Rs, imm:$s10)>;
1359 def: Pat<(sub anyimm:$s10, I32:$Rs), (A2_subri imm:$s10, I32:$Rs)>;
1361 def: OpR_RR_pat<A2_add, Add, i32, I32>;
1362 def: OpR_RR_pat<A2_sub, Sub, i32, I32>;
1363 def: OpR_RR_pat<A2_and, And, i32, I32>;
1364 def: OpR_RR_pat<A2_or, Or, i32, I32>;
1365 def: OpR_RR_pat<A2_xor, Xor, i32, I32>;
1366 def: OpR_RR_pat<A2_addp, Add, i64, I64>;
1367 def: OpR_RR_pat<A2_subp, Sub, i64, I64>;
1368 def: OpR_RR_pat<A2_andp, And, i64, I64>;
1369 def: OpR_RR_pat<A2_orp, Or, i64, I64>;
1370 def: OpR_RR_pat<A2_xorp, Xor, i64, I64>;
1371 def: OpR_RR_pat<A4_andnp, Not2<And>, i64, I64>;
1372 def: OpR_RR_pat<A4_ornp, Not2<Or>, i64, I64>;
1374 def: OpR_RR_pat<A2_svaddh, Add, v2i16, V2I16>;
1375 def: OpR_RR_pat<A2_svsubh, Sub, v2i16, V2I16>;
1377 def: OpR_RR_pat<A2_vaddub, Add, v8i8, V8I8>;
1378 def: OpR_RR_pat<A2_vaddh, Add, v4i16, V4I16>;
1379 def: OpR_RR_pat<A2_vaddw, Add, v2i32, V2I32>;
1380 def: OpR_RR_pat<A2_vsubub, Sub, v8i8, V8I8>;
1381 def: OpR_RR_pat<A2_vsubh, Sub, v4i16, V4I16>;
1382 def: OpR_RR_pat<A2_vsubw, Sub, v2i32, V2I32>;
1384 def: OpR_RR_pat<A2_and, And, v4i8, V4I8>;
1385 def: OpR_RR_pat<A2_xor, Xor, v4i8, V4I8>;
1386 def: OpR_RR_pat<A2_or, Or, v4i8, V4I8>;
1387 def: OpR_RR_pat<A2_and, And, v2i16, V2I16>;
1388 def: OpR_RR_pat<A2_xor, Xor, v2i16, V2I16>;
1389 def: OpR_RR_pat<A2_or, Or, v2i16, V2I16>;
1390 def: OpR_RR_pat<A2_andp, And, v8i8, V8I8>;
1391 def: OpR_RR_pat<A2_orp, Or, v8i8, V8I8>;
1392 def: OpR_RR_pat<A2_xorp, Xor, v8i8, V8I8>;
1393 def: OpR_RR_pat<A2_andp, And, v4i16, V4I16>;
1394 def: OpR_RR_pat<A2_orp, Or, v4i16, V4I16>;
1395 def: OpR_RR_pat<A2_xorp, Xor, v4i16, V4I16>;
1396 def: OpR_RR_pat<A2_andp, And, v2i32, V2I32>;
1397 def: OpR_RR_pat<A2_orp, Or, v2i32, V2I32>;
1398 def: OpR_RR_pat<A2_xorp, Xor, v2i32, V2I32>;
1400 def: OpR_RR_pat<M2_mpyi, Mul, i32, I32>;
1401 def: OpR_RR_pat<M2_mpy_up, pf2<mulhs>, i32, I32>;
1402 def: OpR_RR_pat<M2_mpyu_up, pf2<mulhu>, i32, I32>;
1403 def: OpR_RI_pat<M2_mpysip, Mul, i32, I32, u32_0ImmPred>;
1404 def: OpR_RI_pat<M2_mpysmi, Mul, i32, I32, s32_0ImmPred>;
1406 // Arithmetic on predicates.
1407 def: OpR_RR_pat<C2_xor, Add, i1, I1>;
1408 def: OpR_RR_pat<C2_xor, Add, v2i1, V2I1>;
1409 def: OpR_RR_pat<C2_xor, Add, v4i1, V4I1>;
1410 def: OpR_RR_pat<C2_xor, Add, v8i1, V8I1>;
1411 def: OpR_RR_pat<C2_xor, Sub, i1, I1>;
1412 def: OpR_RR_pat<C2_xor, Sub, v2i1, V2I1>;
1413 def: OpR_RR_pat<C2_xor, Sub, v4i1, V4I1>;
1414 def: OpR_RR_pat<C2_xor, Sub, v8i1, V8I1>;
1415 def: OpR_RR_pat<C2_and, Mul, i1, I1>;
1416 def: OpR_RR_pat<C2_and, Mul, v2i1, V2I1>;
1417 def: OpR_RR_pat<C2_and, Mul, v4i1, V4I1>;
1418 def: OpR_RR_pat<C2_and, Mul, v8i1, V8I1>;
1420 def: OpR_RR_pat<F2_sfadd, pf2<fadd>, f32, F32>;
1421 def: OpR_RR_pat<F2_sfsub, pf2<fsub>, f32, F32>;
1422 def: OpR_RR_pat<F2_sfmpy, pf2<fmul>, f32, F32>;
1423 def: OpR_RR_pat<F2_sfmin, pf2<fminnum>, f32, F32>;
1424 def: OpR_RR_pat<F2_sfmax, pf2<fmaxnum>, f32, F32>;
1426 let Predicates = [HasV66] in {
1427 def: OpR_RR_pat<F2_dfadd, pf2<fadd>, f64, F64>;
1428 def: OpR_RR_pat<F2_dfsub, pf2<fsub>, f64, F64>;
1431 def DfMpy: OutPatFrag<(ops node:$Rs, node:$Rt),
1435 (F2_dfmpyll $Rs, $Rt),
1440 let Predicates = [HasV67,UseUnsafeMath], AddedComplexity = 50 in {
1441 def: Pat<(fmul F64:$Rs, F64:$Rt), (DfMpy $Rs, $Rt)>;
1443 let Predicates = [HasV67] in {
1444 def: OpR_RR_pat<F2_dfmin, pf2<fminnum>, f64, F64>;
1445 def: OpR_RR_pat<F2_dfmax, pf2<fmaxnum>, f64, F64>;
1447 def: Pat<(fmul F64:$Rs, F64:$Rt), (DfMpy (F2_dfmpyfix $Rs, $Rt),
1448 (F2_dfmpyfix $Rt, $Rs))>;
1451 // In expressions like a0*b0 + a1*b1 + ..., prefer to generate multiply-add,
1452 // over add-add with individual multiplies as inputs.
1453 let AddedComplexity = 10 in {
1454 def: AccRRI_pat<M2_macsip, Add, Su<Mul>, I32, u32_0ImmPred>;
1455 def: AccRRI_pat<M2_macsin, Sub, Su<Mul>, I32, u32_0ImmPred>;
1456 def: AccRRR_pat<M2_maci, Add, Su<Mul>, I32, I32, I32>;
1457 let Predicates = [HasV66] in
1458 def: AccRRR_pat<M2_mnaci, Sub, Su<Mul>, I32, I32, I32>;
1461 def: AccRRI_pat<M2_naccii, Sub, Su<Add>, I32, s32_0ImmPred>;
1462 def: AccRRI_pat<M2_accii, Add, Su<Add>, I32, s32_0ImmPred>;
1463 def: AccRRR_pat<M2_acci, Add, Su<Add>, I32, I32, I32>;
1467 def: Pat<(v2i32 (mulhu V2I32:$Rss, V2I32:$Rtt)),
1468 (Combinew (M2_mpyu_up (HiReg $Rss), (HiReg $Rtt)),
1469 (M2_mpyu_up (LoReg $Rss), (LoReg $Rtt)))>;
1471 def: Pat<(v2i32 (mulhs V2I32:$Rs, V2I32:$Rt)),
1472 (Combinew (M2_mpy_up (HiReg $Rs), (HiReg $Rt)),
1473 (M2_mpy_up (LoReg $Rt), (LoReg $Rt)))>;
1476 OutPatFrag<(ops node:$Rss, node:$Rtt),
1477 (Combinew (S2_vtrunohb (M5_vmpybuu (HiReg $Rss), (HiReg $Rtt))),
1478 (S2_vtrunohb (M5_vmpybuu (LoReg $Rss), (LoReg $Rtt))))>;
1480 // Equivalent of byte-wise arithmetic shift right by 7 in v8i8.
1482 OutPatFrag<(ops node:$Rss), (C2_mask (C2_not (A4_vcmpbgti $Rss, 0)))>;
1484 def: Pat<(v8i8 (mulhu V8I8:$Rss, V8I8:$Rtt)),
1485 (Mulhub $Rss, $Rtt)>;
1487 def: Pat<(v8i8 (mulhs V8I8:$Rss, V8I8:$Rtt)),
1489 (Mulhub $Rss, $Rtt),
1490 (A2_vaddub (A2_andp V8I8:$Rss, (Asr7 $Rtt)),
1491 (A2_andp V8I8:$Rtt, (Asr7 $Rss))))>;
1494 OutPatFrag<(ops node:$Rs, node:$Rt), (M2_vmpy2s_s0 $Rs, $Rt)>;
1496 OutPatFrag<(ops node:$Rss, node:$Rtt), (Mpysh (HiReg $Rss), (HiReg $Rtt))>;
1498 OutPatFrag<(ops node:$Rss, node:$Rtt), (Mpysh (LoReg $Rss), (LoReg $Rtt))>;
1501 OutPatFrag<(ops node:$Rss, node:$Rtt),
1502 (Combinew (A2_combine_hh (HiReg (Mpyshh $Rss, $Rtt)),
1503 (LoReg (Mpyshh $Rss, $Rtt))),
1504 (A2_combine_hh (HiReg (Mpyshl $Rss, $Rtt)),
1505 (LoReg (Mpyshl $Rss, $Rtt))))>;
1507 def: Pat<(v4i16 (mulhs V4I16:$Rss, V4I16:$Rtt)), (Mulhsh $Rss, $Rtt)>;
1509 def: Pat<(v4i16 (mulhu V4I16:$Rss, V4I16:$Rtt)),
1511 (Mulhsh $Rss, $Rtt),
1512 (A2_vaddh (A2_andp V4I16:$Rss, (S2_asr_i_vh $Rtt, 15)),
1513 (A2_andp V4I16:$Rtt, (S2_asr_i_vh $Rss, 15))))>;
1516 def: Pat<(ineg (mul I32:$Rs, u8_0ImmPred:$u8)),
1517 (M2_mpysin IntRegs:$Rs, imm:$u8)>;
1519 def n8_0ImmPred: PatLeaf<(i32 imm), [{
1520 int64_t V = N->getSExtValue();
1521 return -255 <= V && V <= 0;
1524 // Change the sign of the immediate for Rd=-mpyi(Rs,#u8)
1525 def: Pat<(mul I32:$Rs, n8_0ImmPred:$n8),
1526 (M2_mpysin I32:$Rs, (NegImm8 imm:$n8))>;
1528 def: Pat<(add Sext64:$Rs, I64:$Rt),
1529 (A2_addsp (LoReg Sext64:$Rs), I64:$Rt)>;
1531 def: AccRRR_pat<M4_and_and, And, Su_ni1<And>, I32, I32, I32>;
1532 def: AccRRR_pat<M4_and_or, And, Su_ni1<Or>, I32, I32, I32>;
1533 def: AccRRR_pat<M4_and_xor, And, Su<Xor>, I32, I32, I32>;
1534 def: AccRRR_pat<M4_or_and, Or, Su_ni1<And>, I32, I32, I32>;
1535 def: AccRRR_pat<M4_or_or, Or, Su_ni1<Or>, I32, I32, I32>;
1536 def: AccRRR_pat<M4_or_xor, Or, Su<Xor>, I32, I32, I32>;
1537 def: AccRRR_pat<M4_xor_and, Xor, Su_ni1<And>, I32, I32, I32>;
1538 def: AccRRR_pat<M4_xor_or, Xor, Su_ni1<Or>, I32, I32, I32>;
1539 def: AccRRR_pat<M2_xor_xacc, Xor, Su<Xor>, I32, I32, I32>;
1540 def: AccRRR_pat<M4_xor_xacc, Xor, Su<Xor>, I64, I64, I64>;
1542 // For dags like (or (and (not _), _), (shl _, _)) where the "or" with
1543 // one argument matches the patterns below, and with the other argument
1544 // matches S2_asl_r_r_or, etc, prefer the patterns below.
1545 let AddedComplexity = 110 in { // greater than S2_asl_r_r_and/or/xor.
1546 def: AccRRR_pat<M4_and_andn, And, Su<Not2<And>>, I32, I32, I32>;
1547 def: AccRRR_pat<M4_or_andn, Or, Su<Not2<And>>, I32, I32, I32>;
1548 def: AccRRR_pat<M4_xor_andn, Xor, Su<Not2<And>>, I32, I32, I32>;
1551 // S4_addaddi and S4_subaddi don't have tied operands, so give them
1552 // a bit of preference.
1553 let AddedComplexity = 30, Predicates = [UseCompound] in {
1554 def: Pat<(add I32:$Rs, (Su<Add> I32:$Ru, anyimm:$s6)),
1555 (S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
1556 def: Pat<(add anyimm:$s6, (Su<Add> I32:$Rs, I32:$Ru)),
1557 (S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
1558 def: Pat<(add I32:$Rs, (Su<Sub> anyimm:$s6, I32:$Ru)),
1559 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1560 def: Pat<(sub (Su<Add> I32:$Rs, anyimm:$s6), I32:$Ru),
1561 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1562 def: Pat<(add (Su<Sub> I32:$Rs, I32:$Ru), anyimm:$s6),
1563 (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1566 let Predicates = [UseCompound] in
1567 def: Pat<(or I32:$Ru, (Su<And> I32:$Rx, anyimm:$s10)),
1568 (S4_or_andix IntRegs:$Ru, IntRegs:$Rx, imm:$s10)>;
1570 def: Pat<(or I32:$Rx, (Su<And> I32:$Rs, anyimm:$s10)),
1571 (S4_or_andi IntRegs:$Rx, IntRegs:$Rs, imm:$s10)>;
1572 def: Pat<(or I32:$Rx, (Su<Or> I32:$Rs, anyimm:$s10)),
1573 (S4_or_ori IntRegs:$Rx, IntRegs:$Rs, imm:$s10)>;
1576 def: Pat<(i32 (trunc (sra (Su<Mul> Sext64:$Rs, Sext64:$Rt), (i32 32)))),
1577 (M2_mpy_up (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1578 def: Pat<(i32 (trunc (srl (Su<Mul> Sext64:$Rs, Sext64:$Rt), (i32 32)))),
1579 (M2_mpy_up (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1581 def: Pat<(mul (Zext64 I32:$Rs), (Zext64 I32:$Rt)),
1582 (M2_dpmpyuu_s0 I32:$Rs, I32:$Rt)>;
1583 def: Pat<(mul (Aext64 I32:$Rs), (Aext64 I32:$Rt)),
1584 (M2_dpmpyuu_s0 I32:$Rs, I32:$Rt)>;
1585 def: Pat<(mul Sext64:$Rs, Sext64:$Rt),
1586 (M2_dpmpyss_s0 (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1588 def: Pat<(add I64:$Rx, (Su<Mul> Sext64:$Rs, Sext64:$Rt)),
1589 (M2_dpmpyss_acc_s0 I64:$Rx, (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1590 def: Pat<(sub I64:$Rx, (Su<Mul> Sext64:$Rs, Sext64:$Rt)),
1591 (M2_dpmpyss_nac_s0 I64:$Rx, (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1592 def: Pat<(add I64:$Rx, (Su<Mul> (Aext64 I32:$Rs), (Aext64 I32:$Rt))),
1593 (M2_dpmpyuu_acc_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1594 def: Pat<(add I64:$Rx, (Su<Mul> (Zext64 I32:$Rs), (Zext64 I32:$Rt))),
1595 (M2_dpmpyuu_acc_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1596 def: Pat<(sub I64:$Rx, (Su<Mul> (Aext64 I32:$Rs), (Aext64 I32:$Rt))),
1597 (M2_dpmpyuu_nac_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1598 def: Pat<(sub I64:$Rx, (Su<Mul> (Zext64 I32:$Rs), (Zext64 I32:$Rt))),
1599 (M2_dpmpyuu_nac_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1602 def: Pat<(sext_inreg (add I32:$Rt, I32:$Rs), i16),
1603 (A2_addh_l16_ll I32:$Rt, I32:$Rs)>;
1604 def: Pat<(sra (add (shl I32:$Rt, (i32 16)), I32:$Rs), (i32 16)),
1605 (A2_addh_l16_hl I32:$Rt, I32:$Rs)>;
1606 def: Pat<(shl (add I32:$Rt, I32:$Rs), (i32 16)),
1607 (A2_addh_h16_ll I32:$Rt, I32:$Rs)>;
1609 // Subtract halfword.
1610 def: Pat<(sext_inreg (sub I32:$Rt, I32:$Rs), i16),
1611 (A2_subh_l16_ll I32:$Rt, I32:$Rs)>;
1612 def: Pat<(sra (add (shl I32:$Rt, (i32 16)), I32:$Rs), (i32 16)),
1613 (A2_addh_l16_hl I32:$Rt, I32:$Rs)>;
1614 def: Pat<(shl (sub I32:$Rt, I32:$Rs), (i32 16)),
1615 (A2_subh_h16_ll I32:$Rt, I32:$Rs)>;
1617 def: Pat<(mul I64:$Rss, I64:$Rtt),
1619 (M2_maci (M2_maci (HiReg (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt))),
1624 (i32 (LoReg (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt)))))>;
1626 def MulHU : OutPatFrag<(ops node:$Rss, node:$Rtt),
1632 (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt)), 32),
1635 (A4_combineir 0, (LoReg (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt))))),
1639 (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt)), 32))>;
1641 // Multiply 64-bit unsigned and use upper result.
1642 def : Pat <(mulhu I64:$Rss, I64:$Rtt), (MulHU $Rss, $Rtt)>;
1644 // Multiply 64-bit signed and use upper result.
1646 // For two signed 64-bit integers A and B, let A' and B' denote A and B
1647 // with the sign bit cleared. Then A = -2^63*s(A) + A', where s(A) is the
1648 // sign bit of A (and identically for B). With this notation, the signed
1649 // product A*B can be written as:
1650 // AB = (-2^63 s(A) + A') * (-2^63 s(B) + B')
1651 // = 2^126 s(A)s(B) - 2^63 [s(A)B'+s(B)A'] + A'B'
1652 // = 2^126 s(A)s(B) + 2^63 [s(A)B'+s(B)A'] + A'B' - 2*2^63 [s(A)B'+s(B)A']
1653 // = (unsigned product AB) - 2^64 [s(A)B'+s(B)A']
1655 // Clear the sign bit in a 64-bit register.
1656 def ClearSign : OutPatFrag<(ops node:$Rss),
1657 (Combinew (S2_clrbit_i (HiReg $Rss), 31), (i32 (LoReg $Rss)))>;
1659 def : Pat <(mulhs I64:$Rss, I64:$Rtt),
1663 (A2_andp (S2_asr_i_p $Rss, 63), (ClearSign $Rtt)),
1664 (A2_andp (S2_asr_i_p $Rtt, 63), (ClearSign $Rss))))>;
1666 // Prefer these instructions over M2_macsip/M2_macsin: the macsi* instructions
1667 // will put the immediate addend into a register, while these instructions will
1668 // use it directly. Such a construct does not appear in the middle of a gep,
1669 // where M2_macsip would be preferable.
1670 let AddedComplexity = 20, Predicates = [UseCompound] in {
1671 def: Pat<(add (Su<Mul> I32:$Rs, u6_0ImmPred:$U6), anyimm:$u6),
1672 (M4_mpyri_addi imm:$u6, IntRegs:$Rs, imm:$U6)>;
1673 def: Pat<(add (Su<Mul> I32:$Rs, I32:$Rt), anyimm:$u6),
1674 (M4_mpyrr_addi imm:$u6, IntRegs:$Rs, IntRegs:$Rt)>;
1677 // Keep these instructions less preferable to M2_macsip/M2_macsin.
1678 let Predicates = [UseCompound] in {
1679 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, u6_2ImmPred:$u6_2)),
1680 (M4_mpyri_addr_u2 IntRegs:$Ru, imm:$u6_2, IntRegs:$Rs)>;
1681 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, anyimm:$u6)),
1682 (M4_mpyri_addr IntRegs:$Ru, IntRegs:$Rs, imm:$u6)>;
1683 def: Pat<(add I32:$Ru, (Su<Mul> I32:$Ry, I32:$Rs)),
1684 (M4_mpyrr_addr IntRegs:$Ru, IntRegs:$Ry, IntRegs:$Rs)>;
1687 def: Pat<(fma F32:$Rs, F32:$Rt, F32:$Rx),
1688 (F2_sffma F32:$Rx, F32:$Rs, F32:$Rt)>;
1689 def: Pat<(fma (fneg F32:$Rs), F32:$Rt, F32:$Rx),
1690 (F2_sffms F32:$Rx, F32:$Rs, F32:$Rt)>;
1691 def: Pat<(fma F32:$Rs, (fneg F32:$Rt), F32:$Rx),
1692 (F2_sffms F32:$Rx, F32:$Rs, F32:$Rt)>;
1694 def: Pat<(mul V2I32:$Rs, V2I32:$Rt),
1695 (PS_vmulw V2I32:$Rs, V2I32:$Rt)>;
1696 def: Pat<(add V2I32:$Rx, (mul V2I32:$Rs, V2I32:$Rt)),
1697 (PS_vmulw_acc V2I32:$Rx, V2I32:$Rs, V2I32:$Rt)>;
1699 // Add/subtract two v4i8: Hexagon does not have an insn for this one, so
1700 // we use the double add v8i8, and use only the low part of the result.
1701 def: Pat<(add V4I8:$Rs, V4I8:$Rt),
1702 (LoReg (A2_vaddub (ToAext64 $Rs), (ToAext64 $Rt)))>;
1703 def: Pat<(sub V4I8:$Rs, V4I8:$Rt),
1704 (LoReg (A2_vsubub (ToAext64 $Rs), (ToAext64 $Rt)))>;
1706 // Use M2_vmpy2s_s0 for half-word vector multiply. It multiplies two
1707 // half-words, and saturates the result to a 32-bit value, except the
1708 // saturation never happens (it can only occur with scaling).
1709 def: Pat<(v2i16 (mul V2I16:$Rs, V2I16:$Rt)),
1710 (LoReg (S2_vtrunewh (A2_combineii 0, 0),
1711 (M2_vmpy2s_s0 V2I16:$Rs, V2I16:$Rt)))>;
1712 def: Pat<(v4i16 (mul V4I16:$Rs, V4I16:$Rt)),
1713 (S2_vtrunewh (M2_vmpy2s_s0 (HiReg $Rs), (HiReg $Rt)),
1714 (M2_vmpy2s_s0 (LoReg $Rs), (LoReg $Rt)))>;
1716 // Multiplies two v4i8 vectors.
1717 def: Pat<(v4i8 (mul V4I8:$Rs, V4I8:$Rt)),
1718 (S2_vtrunehb (M5_vmpybuu V4I8:$Rs, V4I8:$Rt))>;
1720 // Multiplies two v8i8 vectors.
1721 def: Pat<(v8i8 (mul V8I8:$Rs, V8I8:$Rt)),
1722 (Combinew (S2_vtrunehb (M5_vmpybuu (HiReg $Rs), (HiReg $Rt))),
1723 (S2_vtrunehb (M5_vmpybuu (LoReg $Rs), (LoReg $Rt))))>;
1726 // --(10) Bit ------------------------------------------------------------
1729 // Count leading zeros.
1730 def: Pat<(i32 (ctlz I32:$Rs)), (S2_cl0 I32:$Rs)>;
1731 def: Pat<(i32 (trunc (ctlz I64:$Rss))), (S2_cl0p I64:$Rss)>;
1733 // Count trailing zeros.
1734 def: Pat<(i32 (cttz I32:$Rs)), (S2_ct0 I32:$Rs)>;
1735 def: Pat<(i32 (trunc (cttz I64:$Rss))), (S2_ct0p I64:$Rss)>;
1737 // Count leading ones.
1738 def: Pat<(i32 (ctlz (not I32:$Rs))), (S2_cl1 I32:$Rs)>;
1739 def: Pat<(i32 (trunc (ctlz (not I64:$Rss)))), (S2_cl1p I64:$Rss)>;
1741 // Count trailing ones.
1742 def: Pat<(i32 (cttz (not I32:$Rs))), (S2_ct1 I32:$Rs)>;
1743 def: Pat<(i32 (trunc (cttz (not I64:$Rss)))), (S2_ct1p I64:$Rss)>;
1745 // Define leading/trailing patterns that require zero-extensions to 64 bits.
1746 def: Pat<(i64 (ctlz I64:$Rss)), (ToZext64 (S2_cl0p I64:$Rss))>;
1747 def: Pat<(i64 (cttz I64:$Rss)), (ToZext64 (S2_ct0p I64:$Rss))>;
1748 def: Pat<(i64 (ctlz (not I64:$Rss))), (ToZext64 (S2_cl1p I64:$Rss))>;
1749 def: Pat<(i64 (cttz (not I64:$Rss))), (ToZext64 (S2_ct1p I64:$Rss))>;
1751 def: Pat<(i64 (ctpop I64:$Rss)), (ToZext64 (S5_popcountp I64:$Rss))>;
1752 def: Pat<(i32 (ctpop I32:$Rs)), (S5_popcountp (A4_combineir 0, I32:$Rs))>;
1754 def: Pat<(bitreverse I32:$Rs), (S2_brev I32:$Rs)>;
1755 def: Pat<(bitreverse I64:$Rss), (S2_brevp I64:$Rss)>;
1757 let AddedComplexity = 20 in { // Complexity greater than and/or/xor
1758 def: Pat<(and I32:$Rs, IsNPow2_32:$V),
1759 (S2_clrbit_i IntRegs:$Rs, (LogN2_32 $V))>;
1760 def: Pat<(or I32:$Rs, IsPow2_32:$V),
1761 (S2_setbit_i IntRegs:$Rs, (Log2_32 $V))>;
1762 def: Pat<(xor I32:$Rs, IsPow2_32:$V),
1763 (S2_togglebit_i IntRegs:$Rs, (Log2_32 $V))>;
1765 def: Pat<(and I32:$Rs, (not (shl 1, I32:$Rt))),
1766 (S2_clrbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1767 def: Pat<(or I32:$Rs, (shl 1, I32:$Rt)),
1768 (S2_setbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1769 def: Pat<(xor I32:$Rs, (shl 1, I32:$Rt)),
1770 (S2_togglebit_r IntRegs:$Rs, IntRegs:$Rt)>;
1773 // Clr/set/toggle bit for 64-bit values with immediate bit index.
1774 let AddedComplexity = 20 in { // Complexity greater than and/or/xor
1775 def: Pat<(and I64:$Rss, IsNPow2_64L:$V),
1776 (Combinew (i32 (HiReg $Rss)),
1777 (S2_clrbit_i (LoReg $Rss), (LogN2_64 $V)))>;
1778 def: Pat<(and I64:$Rss, IsNPow2_64H:$V),
1779 (Combinew (S2_clrbit_i (HiReg $Rss), (UDEC32 (i32 (LogN2_64 $V)))),
1780 (i32 (LoReg $Rss)))>;
1782 def: Pat<(or I64:$Rss, IsPow2_64L:$V),
1783 (Combinew (i32 (HiReg $Rss)),
1784 (S2_setbit_i (LoReg $Rss), (Log2_64 $V)))>;
1785 def: Pat<(or I64:$Rss, IsPow2_64H:$V),
1786 (Combinew (S2_setbit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
1787 (i32 (LoReg $Rss)))>;
1789 def: Pat<(xor I64:$Rss, IsPow2_64L:$V),
1790 (Combinew (i32 (HiReg $Rss)),
1791 (S2_togglebit_i (LoReg $Rss), (Log2_64 $V)))>;
1792 def: Pat<(xor I64:$Rss, IsPow2_64H:$V),
1793 (Combinew (S2_togglebit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
1794 (i32 (LoReg $Rss)))>;
1798 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
1799 def: Pat<(i1 (setne (and (shl 1, u5_0ImmPred:$u5), I32:$Rs), 0)),
1800 (S2_tstbit_i IntRegs:$Rs, imm:$u5)>;
1801 def: Pat<(i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1802 (S2_tstbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1803 def: Pat<(i1 (trunc I32:$Rs)),
1804 (S2_tstbit_i IntRegs:$Rs, 0)>;
1805 def: Pat<(i1 (trunc I64:$Rs)),
1806 (S2_tstbit_i (LoReg DoubleRegs:$Rs), 0)>;
1809 def: Pat<(and (srl I32:$Rs, u5_0ImmPred:$u5), 1),
1810 (I1toI32 (S2_tstbit_i I32:$Rs, imm:$u5))>;
1811 def: Pat<(and (srl I64:$Rss, IsULE<32,31>:$u6), 1),
1812 (ToZext64 (I1toI32 (S2_tstbit_i (LoReg $Rss), imm:$u6)))>;
1813 def: Pat<(and (srl I64:$Rss, IsUGT<32,31>:$u6), 1),
1814 (ToZext64 (I1toI32 (S2_tstbit_i (HiReg $Rss), (UDEC32 $u6))))>;
1816 def: Pat<(and (not (srl I32:$Rs, u5_0ImmPred:$u5)), 1),
1817 (I1toI32 (S4_ntstbit_i I32:$Rs, imm:$u5))>;
1818 def: Pat<(and (not (srl I64:$Rss, IsULE<32,31>:$u6)), 1),
1819 (ToZext64 (I1toI32 (S4_ntstbit_i (LoReg $Rss), imm:$u6)))>;
1820 def: Pat<(and (not (srl I64:$Rss, IsUGT<32,31>:$u6)), 1),
1821 (ToZext64 (I1toI32 (S4_ntstbit_i (HiReg $Rss), (UDEC32 $u6))))>;
1823 let AddedComplexity = 20 in { // Complexity greater than compare reg-imm.
1824 def: Pat<(i1 (seteq (and I32:$Rs, u6_0ImmPred:$u6), 0)),
1825 (C2_bitsclri IntRegs:$Rs, imm:$u6)>;
1826 def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), 0)),
1827 (C2_bitsclr IntRegs:$Rs, IntRegs:$Rt)>;
1830 let AddedComplexity = 10 in // Complexity greater than compare reg-reg.
1831 def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), IntRegs:$Rt)),
1832 (C2_bitsset IntRegs:$Rs, IntRegs:$Rt)>;
1835 SDTypeProfile<1, 2, [SDTCisVT<0, i1>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
1836 def HexagonTSTBIT: SDNode<"HexagonISD::TSTBIT", SDTTestBit>;
1838 def: Pat<(HexagonTSTBIT I32:$Rs, u5_0ImmPred:$u5),
1839 (S2_tstbit_i I32:$Rs, imm:$u5)>;
1840 def: Pat<(HexagonTSTBIT I32:$Rs, I32:$Rt),
1841 (S2_tstbit_r I32:$Rs, I32:$Rt)>;
1843 // Add extra complexity to prefer these instructions over bitsset/bitsclr.
1844 // The reason is that tstbit/ntstbit can be folded into a compound instruction:
1845 // if ([!]tstbit(...)) jump ...
1846 let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
1847 def: Pat<(i1 (seteq (and I32:$Rs, IsPow2_32:$u5), 0)),
1848 (S4_ntstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
1849 def: Pat<(i1 (setne (and I32:$Rs, IsPow2_32:$u5), 0)),
1850 (S2_tstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
1851 def: Pat<(i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1852 (S4_ntstbit_r I32:$Rs, I32:$Rt)>;
1853 def: Pat<(i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1854 (S2_tstbit_r I32:$Rs, I32:$Rt)>;
1857 def: Pat<(i1 (seteq (and I64:$Rs, IsPow2_64L:$u6), 0)),
1858 (S4_ntstbit_i (LoReg $Rs), (Log2_64 $u6))>;
1859 def: Pat<(i1 (seteq (and I64:$Rs, IsPow2_64H:$u6), 0)),
1860 (S4_ntstbit_i (HiReg $Rs), (UDEC32 (i32 (Log2_64 $u6))))>;
1861 def: Pat<(i1 (setne (and I64:$Rs, IsPow2_64L:$u6), 0)),
1862 (S2_tstbit_i (LoReg $Rs), (Log2_64 imm:$u6))>;
1863 def: Pat<(i1 (setne (and I64:$Rs, IsPow2_64H:$u6), 0)),
1864 (S2_tstbit_i (HiReg $Rs), (UDEC32 (i32 (Log2_64 imm:$u6))))>;
1866 // Do not increase complexity of these patterns. In the DAG, "cmp i8" may be
1867 // represented as a compare against "value & 0xFF", which is an exact match
1868 // for cmpb (same for cmph). The patterns below do not contain any additional
1869 // complexity that would make them preferable, and if they were actually used
1870 // instead of cmpb/cmph, they would result in a compare against register that
1871 // is loaded with the byte/half mask (i.e. 0xFF or 0xFFFF).
1872 def: Pat<(i1 (setne (and I32:$Rs, u6_0ImmPred:$u6), 0)),
1873 (C4_nbitsclri I32:$Rs, imm:$u6)>;
1874 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), 0)),
1875 (C4_nbitsclr I32:$Rs, I32:$Rt)>;
1876 def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), I32:$Rt)),
1877 (C4_nbitsset I32:$Rs, I32:$Rt)>;
1879 // Special patterns to address certain cases where the "top-down" matching
1880 // algorithm would cause suboptimal selection.
1882 let AddedComplexity = 100 in {
1883 // Avoid A4_rcmp[n]eqi in these cases:
1884 def: Pat<(i32 (zext (i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1885 (I1toI32 (S4_ntstbit_r IntRegs:$Rs, IntRegs:$Rt))>;
1886 def: Pat<(i32 (zext (i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1887 (I1toI32 (S2_tstbit_r IntRegs:$Rs, IntRegs:$Rt))>;
1888 def: Pat<(i32 (zext (i1 (seteq (and I32:$Rs, IsPow2_32:$u5), 0)))),
1889 (I1toI32 (S4_ntstbit_i I32:$Rs, (Log2_32 imm:$u5)))>;
1890 def: Pat<(i32 (zext (i1 (setne (and I32:$Rs, IsPow2_32:$u5), 0)))),
1891 (I1toI32 (S2_tstbit_i I32:$Rs, (Log2_32 imm:$u5)))>;
1892 def: Pat<(i32 (zext (i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1893 (I1toI32 (S4_ntstbit_r I32:$Rs, I32:$Rt))>;
1894 def: Pat<(i32 (zext (i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1895 (I1toI32 (S2_tstbit_r I32:$Rs, I32:$Rt))>;
1898 // --(11) PIC ------------------------------------------------------------
1901 def SDT_HexagonAtGot
1902 : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
1903 def SDT_HexagonAtPcrel
1904 : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
1906 // AT_GOT address-of-GOT, address-of-global, offset-in-global
1907 def HexagonAtGot : SDNode<"HexagonISD::AT_GOT", SDT_HexagonAtGot>;
1908 // AT_PCREL address-of-global
1909 def HexagonAtPcrel : SDNode<"HexagonISD::AT_PCREL", SDT_HexagonAtPcrel>;
1911 def: Pat<(HexagonAtGot I32:$got, I32:$addr, (i32 0)),
1912 (L2_loadri_io I32:$got, imm:$addr)>;
1913 def: Pat<(HexagonAtGot I32:$got, I32:$addr, s30_2ImmPred:$off),
1914 (A2_addi (L2_loadri_io I32:$got, imm:$addr), imm:$off)>;
1915 def: Pat<(HexagonAtPcrel I32:$addr),
1916 (C4_addipc imm:$addr)>;
1918 // The HVX load patterns also match AT_PCREL directly. Make sure that
1919 // if the selection of this opcode changes, it's updated in all places.
1922 // --(12) Load -----------------------------------------------------------
1925 def extloadv2i8: PatFrag<(ops node:$ptr), (extload node:$ptr), [{
1926 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1928 def extloadv4i8: PatFrag<(ops node:$ptr), (extload node:$ptr), [{
1929 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1932 def zextloadv2i8: PatFrag<(ops node:$ptr), (zextload node:$ptr), [{
1933 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1935 def zextloadv4i8: PatFrag<(ops node:$ptr), (zextload node:$ptr), [{
1936 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1939 def sextloadv2i8: PatFrag<(ops node:$ptr), (sextload node:$ptr), [{
1940 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1942 def sextloadv4i8: PatFrag<(ops node:$ptr), (sextload node:$ptr), [{
1943 return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1946 // Patterns to select load-indexed: Rs + Off.
1947 // - frameindex [+ imm],
1948 multiclass Loadxfi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1950 def: Pat<(VT (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
1951 (VT (MI AddrFI:$fi, imm:$Off))>;
1952 def: Pat<(VT (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
1953 (VT (MI AddrFI:$fi, imm:$Off))>;
1954 def: Pat<(VT (Load AddrFI:$fi)), (VT (MI AddrFI:$fi, 0))>;
1957 // Patterns to select load-indexed: Rs + Off.
1958 // - base reg [+ imm]
1959 multiclass Loadxgi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1961 def: Pat<(VT (Load (add I32:$Rs, ImmPred:$Off))),
1962 (VT (MI IntRegs:$Rs, imm:$Off))>;
1963 def: Pat<(VT (Load (IsOrAdd I32:$Rs, ImmPred:$Off))),
1964 (VT (MI IntRegs:$Rs, imm:$Off))>;
1965 def: Pat<(VT (Load I32:$Rs)), (VT (MI IntRegs:$Rs, 0))>;
1968 // Patterns to select load-indexed: Rs + Off. Combines Loadxfi + Loadxgi.
1969 multiclass Loadxi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1971 defm: Loadxfi_pat<Load, VT, ImmPred, MI>;
1972 defm: Loadxgi_pat<Load, VT, ImmPred, MI>;
1975 // Patterns to select load reg indexed: Rs + Off with a value modifier.
1976 // - frameindex [+ imm]
1977 multiclass Loadxfim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1978 PatLeaf ImmPred, InstHexagon MI> {
1979 def: Pat<(VT (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
1980 (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>;
1981 def: Pat<(VT (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
1982 (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>;
1983 def: Pat<(VT (Load AddrFI:$fi)), (VT (ValueMod (MI AddrFI:$fi, 0)))>;
1986 // Patterns to select load reg indexed: Rs + Off with a value modifier.
1987 // - base reg [+ imm]
1988 multiclass Loadxgim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
1989 PatLeaf ImmPred, InstHexagon MI> {
1990 def: Pat<(VT (Load (add I32:$Rs, ImmPred:$Off))),
1991 (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
1992 def: Pat<(VT (Load (IsOrAdd I32:$Rs, ImmPred:$Off))),
1993 (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
1994 def: Pat<(VT (Load I32:$Rs)), (VT (ValueMod (MI IntRegs:$Rs, 0)))>;
1997 // Patterns to select load reg indexed: Rs + Off with a value modifier.
1998 // Combines Loadxfim + Loadxgim.
1999 multiclass Loadxim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2000 PatLeaf ImmPred, InstHexagon MI> {
2001 defm: Loadxfim_pat<Load, VT, ValueMod, ImmPred, MI>;
2002 defm: Loadxgim_pat<Load, VT, ValueMod, ImmPred, MI>;
2005 // Pattern to select load reg reg-indexed: Rs + Rt<<u2.
2006 class Loadxr_shl_pat<PatFrag Load, ValueType VT, InstHexagon MI>
2007 : Pat<(VT (Load (add I32:$Rs, (i32 (shl I32:$Rt, u2_0ImmPred:$u2))))),
2008 (VT (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2))>;
2010 // Pattern to select load reg reg-indexed: Rs + Rt<<0.
2011 class Loadxr_add_pat<PatFrag Load, ValueType VT, InstHexagon MI>
2012 : Pat<(VT (Load (add I32:$Rs, I32:$Rt))),
2013 (VT (MI IntRegs:$Rs, IntRegs:$Rt, 0))>;
2015 // Pattern to select load reg reg-indexed: Rs + Rt<<u2 with value modifier.
2016 class Loadxrm_shl_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2018 : Pat<(VT (Load (add I32:$Rs, (i32 (shl I32:$Rt, u2_0ImmPred:$u2))))),
2019 (VT (ValueMod (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2)))>;
2021 // Pattern to select load reg reg-indexed: Rs + Rt<<0 with value modifier.
2022 class Loadxrm_add_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2024 : Pat<(VT (Load (add I32:$Rs, I32:$Rt))),
2025 (VT (ValueMod (MI IntRegs:$Rs, IntRegs:$Rt, 0)))>;
2027 // Pattern to select load long-offset reg-indexed: Addr + Rt<<u2.
2028 // Don't match for u2==0, instead use reg+imm for those cases.
2029 class Loadxu_pat<PatFrag Load, ValueType VT, PatFrag ImmPred, InstHexagon MI>
2030 : Pat<(VT (Load (add (shl IntRegs:$Rt, u2_0ImmPred:$u2), ImmPred:$Addr))),
2031 (VT (MI IntRegs:$Rt, imm:$u2, ImmPred:$Addr))>;
2033 class Loadxum_pat<PatFrag Load, ValueType VT, PatFrag ImmPred, PatFrag ValueMod,
2035 : Pat<(VT (Load (add (shl IntRegs:$Rt, u2_0ImmPred:$u2), ImmPred:$Addr))),
2036 (VT (ValueMod (MI IntRegs:$Rt, imm:$u2, ImmPred:$Addr)))>;
2038 // Pattern to select load absolute.
2039 class Loada_pat<PatFrag Load, ValueType VT, PatFrag Addr, InstHexagon MI>
2040 : Pat<(VT (Load Addr:$addr)), (MI Addr:$addr)>;
2042 // Pattern to select load absolute with value modifier.
2043 class Loadam_pat<PatFrag Load, ValueType VT, PatFrag Addr, PatFrag ValueMod,
2045 : Pat<(VT (Load Addr:$addr)), (ValueMod (MI Addr:$addr))>;
2048 let AddedComplexity = 20 in {
2049 defm: Loadxi_pat<extloadi1, i32, anyimm0, L2_loadrub_io>;
2050 defm: Loadxi_pat<extloadi8, i32, anyimm0, L2_loadrub_io>;
2051 defm: Loadxi_pat<extloadi16, i32, anyimm1, L2_loadruh_io>;
2052 defm: Loadxi_pat<extloadv2i8, v2i16, anyimm1, L2_loadbzw2_io>;
2053 defm: Loadxi_pat<extloadv4i8, v4i16, anyimm2, L2_loadbzw4_io>;
2054 defm: Loadxi_pat<sextloadi8, i32, anyimm0, L2_loadrb_io>;
2055 defm: Loadxi_pat<sextloadi16, i32, anyimm1, L2_loadrh_io>;
2056 defm: Loadxi_pat<sextloadv2i8, v2i16, anyimm1, L2_loadbsw2_io>;
2057 defm: Loadxi_pat<sextloadv4i8, v4i16, anyimm2, L2_loadbzw4_io>;
2058 defm: Loadxi_pat<zextloadi1, i32, anyimm0, L2_loadrub_io>;
2059 defm: Loadxi_pat<zextloadi8, i32, anyimm0, L2_loadrub_io>;
2060 defm: Loadxi_pat<zextloadi16, i32, anyimm1, L2_loadruh_io>;
2061 defm: Loadxi_pat<zextloadv2i8, v2i16, anyimm1, L2_loadbzw2_io>;
2062 defm: Loadxi_pat<zextloadv4i8, v4i16, anyimm2, L2_loadbzw4_io>;
2063 defm: Loadxi_pat<load, i32, anyimm2, L2_loadri_io>;
2064 defm: Loadxi_pat<load, v2i16, anyimm2, L2_loadri_io>;
2065 defm: Loadxi_pat<load, v4i8, anyimm2, L2_loadri_io>;
2066 defm: Loadxi_pat<load, i64, anyimm3, L2_loadrd_io>;
2067 defm: Loadxi_pat<load, v2i32, anyimm3, L2_loadrd_io>;
2068 defm: Loadxi_pat<load, v4i16, anyimm3, L2_loadrd_io>;
2069 defm: Loadxi_pat<load, v8i8, anyimm3, L2_loadrd_io>;
2070 defm: Loadxi_pat<load, f32, anyimm2, L2_loadri_io>;
2071 defm: Loadxi_pat<load, f64, anyimm3, L2_loadrd_io>;
2074 defm: Loadxi_pat<atomic_load_8 , i32, anyimm0, L2_loadrub_io>;
2075 defm: Loadxi_pat<atomic_load_16, i32, anyimm1, L2_loadruh_io>;
2076 defm: Loadxi_pat<atomic_load_32, i32, anyimm2, L2_loadri_io>;
2077 defm: Loadxi_pat<atomic_load_64, i64, anyimm3, L2_loadrd_io>;
2080 let AddedComplexity = 30 in {
2081 defm: Loadxim_pat<extloadi1, i64, ToAext64, anyimm0, L2_loadrub_io>;
2082 defm: Loadxim_pat<extloadi8, i64, ToAext64, anyimm0, L2_loadrub_io>;
2083 defm: Loadxim_pat<extloadi16, i64, ToAext64, anyimm1, L2_loadruh_io>;
2084 defm: Loadxim_pat<extloadi32, i64, ToAext64, anyimm2, L2_loadri_io>;
2085 defm: Loadxim_pat<zextloadi1, i64, ToZext64, anyimm0, L2_loadrub_io>;
2086 defm: Loadxim_pat<zextloadi8, i64, ToZext64, anyimm0, L2_loadrub_io>;
2087 defm: Loadxim_pat<zextloadi16, i64, ToZext64, anyimm1, L2_loadruh_io>;
2088 defm: Loadxim_pat<zextloadi32, i64, ToZext64, anyimm2, L2_loadri_io>;
2089 defm: Loadxim_pat<sextloadi8, i64, ToSext64, anyimm0, L2_loadrb_io>;
2090 defm: Loadxim_pat<sextloadi16, i64, ToSext64, anyimm1, L2_loadrh_io>;
2091 defm: Loadxim_pat<sextloadi32, i64, ToSext64, anyimm2, L2_loadri_io>;
2094 let AddedComplexity = 60 in {
2095 def: Loadxu_pat<extloadi8, i32, anyimm0, L4_loadrub_ur>;
2096 def: Loadxu_pat<extloadi16, i32, anyimm1, L4_loadruh_ur>;
2097 def: Loadxu_pat<extloadv2i8, v2i16, anyimm1, L4_loadbzw2_ur>;
2098 def: Loadxu_pat<extloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
2099 def: Loadxu_pat<sextloadi8, i32, anyimm0, L4_loadrb_ur>;
2100 def: Loadxu_pat<sextloadi16, i32, anyimm1, L4_loadrh_ur>;
2101 def: Loadxu_pat<sextloadv2i8, v2i16, anyimm1, L4_loadbsw2_ur>;
2102 def: Loadxu_pat<sextloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
2103 def: Loadxu_pat<zextloadi8, i32, anyimm0, L4_loadrub_ur>;
2104 def: Loadxu_pat<zextloadi16, i32, anyimm1, L4_loadruh_ur>;
2105 def: Loadxu_pat<zextloadv2i8, v2i16, anyimm1, L4_loadbzw2_ur>;
2106 def: Loadxu_pat<zextloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
2107 def: Loadxu_pat<load, i32, anyimm2, L4_loadri_ur>;
2108 def: Loadxu_pat<load, v2i16, anyimm2, L4_loadri_ur>;
2109 def: Loadxu_pat<load, v4i8, anyimm2, L4_loadri_ur>;
2110 def: Loadxu_pat<load, i64, anyimm3, L4_loadrd_ur>;
2111 def: Loadxu_pat<load, v2i32, anyimm3, L4_loadrd_ur>;
2112 def: Loadxu_pat<load, v4i16, anyimm3, L4_loadrd_ur>;
2113 def: Loadxu_pat<load, v8i8, anyimm3, L4_loadrd_ur>;
2114 def: Loadxu_pat<load, f32, anyimm2, L4_loadri_ur>;
2115 def: Loadxu_pat<load, f64, anyimm3, L4_loadrd_ur>;
2117 def: Loadxum_pat<sextloadi8, i64, anyimm0, ToSext64, L4_loadrb_ur>;
2118 def: Loadxum_pat<zextloadi8, i64, anyimm0, ToZext64, L4_loadrub_ur>;
2119 def: Loadxum_pat<extloadi8, i64, anyimm0, ToAext64, L4_loadrub_ur>;
2120 def: Loadxum_pat<sextloadi16, i64, anyimm1, ToSext64, L4_loadrh_ur>;
2121 def: Loadxum_pat<zextloadi16, i64, anyimm1, ToZext64, L4_loadruh_ur>;
2122 def: Loadxum_pat<extloadi16, i64, anyimm1, ToAext64, L4_loadruh_ur>;
2123 def: Loadxum_pat<sextloadi32, i64, anyimm2, ToSext64, L4_loadri_ur>;
2124 def: Loadxum_pat<zextloadi32, i64, anyimm2, ToZext64, L4_loadri_ur>;
2125 def: Loadxum_pat<extloadi32, i64, anyimm2, ToAext64, L4_loadri_ur>;
2128 let AddedComplexity = 40 in {
2129 def: Loadxr_shl_pat<extloadi8, i32, L4_loadrub_rr>;
2130 def: Loadxr_shl_pat<zextloadi8, i32, L4_loadrub_rr>;
2131 def: Loadxr_shl_pat<sextloadi8, i32, L4_loadrb_rr>;
2132 def: Loadxr_shl_pat<extloadi16, i32, L4_loadruh_rr>;
2133 def: Loadxr_shl_pat<zextloadi16, i32, L4_loadruh_rr>;
2134 def: Loadxr_shl_pat<sextloadi16, i32, L4_loadrh_rr>;
2135 def: Loadxr_shl_pat<load, i32, L4_loadri_rr>;
2136 def: Loadxr_shl_pat<load, v2i16, L4_loadri_rr>;
2137 def: Loadxr_shl_pat<load, v4i8, L4_loadri_rr>;
2138 def: Loadxr_shl_pat<load, i64, L4_loadrd_rr>;
2139 def: Loadxr_shl_pat<load, v2i32, L4_loadrd_rr>;
2140 def: Loadxr_shl_pat<load, v4i16, L4_loadrd_rr>;
2141 def: Loadxr_shl_pat<load, v8i8, L4_loadrd_rr>;
2142 def: Loadxr_shl_pat<load, f32, L4_loadri_rr>;
2143 def: Loadxr_shl_pat<load, f64, L4_loadrd_rr>;
2146 let AddedComplexity = 20 in {
2147 def: Loadxr_add_pat<extloadi8, i32, L4_loadrub_rr>;
2148 def: Loadxr_add_pat<zextloadi8, i32, L4_loadrub_rr>;
2149 def: Loadxr_add_pat<sextloadi8, i32, L4_loadrb_rr>;
2150 def: Loadxr_add_pat<extloadi16, i32, L4_loadruh_rr>;
2151 def: Loadxr_add_pat<zextloadi16, i32, L4_loadruh_rr>;
2152 def: Loadxr_add_pat<sextloadi16, i32, L4_loadrh_rr>;
2153 def: Loadxr_add_pat<load, i32, L4_loadri_rr>;
2154 def: Loadxr_add_pat<load, v2i16, L4_loadri_rr>;
2155 def: Loadxr_add_pat<load, v4i8, L4_loadri_rr>;
2156 def: Loadxr_add_pat<load, i64, L4_loadrd_rr>;
2157 def: Loadxr_add_pat<load, v2i32, L4_loadrd_rr>;
2158 def: Loadxr_add_pat<load, v4i16, L4_loadrd_rr>;
2159 def: Loadxr_add_pat<load, v8i8, L4_loadrd_rr>;
2160 def: Loadxr_add_pat<load, f32, L4_loadri_rr>;
2161 def: Loadxr_add_pat<load, f64, L4_loadrd_rr>;
2164 let AddedComplexity = 40 in {
2165 def: Loadxrm_shl_pat<extloadi8, i64, ToAext64, L4_loadrub_rr>;
2166 def: Loadxrm_shl_pat<zextloadi8, i64, ToZext64, L4_loadrub_rr>;
2167 def: Loadxrm_shl_pat<sextloadi8, i64, ToSext64, L4_loadrb_rr>;
2168 def: Loadxrm_shl_pat<extloadi16, i64, ToAext64, L4_loadruh_rr>;
2169 def: Loadxrm_shl_pat<zextloadi16, i64, ToZext64, L4_loadruh_rr>;
2170 def: Loadxrm_shl_pat<sextloadi16, i64, ToSext64, L4_loadrh_rr>;
2171 def: Loadxrm_shl_pat<extloadi32, i64, ToAext64, L4_loadri_rr>;
2172 def: Loadxrm_shl_pat<zextloadi32, i64, ToZext64, L4_loadri_rr>;
2173 def: Loadxrm_shl_pat<sextloadi32, i64, ToSext64, L4_loadri_rr>;
2176 let AddedComplexity = 20 in {
2177 def: Loadxrm_add_pat<extloadi8, i64, ToAext64, L4_loadrub_rr>;
2178 def: Loadxrm_add_pat<zextloadi8, i64, ToZext64, L4_loadrub_rr>;
2179 def: Loadxrm_add_pat<sextloadi8, i64, ToSext64, L4_loadrb_rr>;
2180 def: Loadxrm_add_pat<extloadi16, i64, ToAext64, L4_loadruh_rr>;
2181 def: Loadxrm_add_pat<zextloadi16, i64, ToZext64, L4_loadruh_rr>;
2182 def: Loadxrm_add_pat<sextloadi16, i64, ToSext64, L4_loadrh_rr>;
2183 def: Loadxrm_add_pat<extloadi32, i64, ToAext64, L4_loadri_rr>;
2184 def: Loadxrm_add_pat<zextloadi32, i64, ToZext64, L4_loadri_rr>;
2185 def: Loadxrm_add_pat<sextloadi32, i64, ToSext64, L4_loadri_rr>;
2190 let AddedComplexity = 60 in {
2191 def: Loada_pat<zextloadi1, i32, anyimm0, PS_loadrubabs>;
2192 def: Loada_pat<sextloadi8, i32, anyimm0, PS_loadrbabs>;
2193 def: Loada_pat<extloadi8, i32, anyimm0, PS_loadrubabs>;
2194 def: Loada_pat<zextloadi8, i32, anyimm0, PS_loadrubabs>;
2195 def: Loada_pat<sextloadi16, i32, anyimm1, PS_loadrhabs>;
2196 def: Loada_pat<extloadi16, i32, anyimm1, PS_loadruhabs>;
2197 def: Loada_pat<zextloadi16, i32, anyimm1, PS_loadruhabs>;
2198 def: Loada_pat<load, i32, anyimm2, PS_loadriabs>;
2199 def: Loada_pat<load, v2i16, anyimm2, PS_loadriabs>;
2200 def: Loada_pat<load, v4i8, anyimm2, PS_loadriabs>;
2201 def: Loada_pat<load, i64, anyimm3, PS_loadrdabs>;
2202 def: Loada_pat<load, v2i32, anyimm3, PS_loadrdabs>;
2203 def: Loada_pat<load, v4i16, anyimm3, PS_loadrdabs>;
2204 def: Loada_pat<load, v8i8, anyimm3, PS_loadrdabs>;
2205 def: Loada_pat<load, f32, anyimm2, PS_loadriabs>;
2206 def: Loada_pat<load, f64, anyimm3, PS_loadrdabs>;
2208 def: Loada_pat<atomic_load_8, i32, anyimm0, PS_loadrubabs>;
2209 def: Loada_pat<atomic_load_16, i32, anyimm1, PS_loadruhabs>;
2210 def: Loada_pat<atomic_load_32, i32, anyimm2, PS_loadriabs>;
2211 def: Loada_pat<atomic_load_64, i64, anyimm3, PS_loadrdabs>;
2214 let AddedComplexity = 30 in {
2215 def: Loadam_pat<extloadi8, i64, anyimm0, ToAext64, PS_loadrubabs>;
2216 def: Loadam_pat<sextloadi8, i64, anyimm0, ToSext64, PS_loadrbabs>;
2217 def: Loadam_pat<zextloadi8, i64, anyimm0, ToZext64, PS_loadrubabs>;
2218 def: Loadam_pat<extloadi16, i64, anyimm1, ToAext64, PS_loadruhabs>;
2219 def: Loadam_pat<sextloadi16, i64, anyimm1, ToSext64, PS_loadrhabs>;
2220 def: Loadam_pat<zextloadi16, i64, anyimm1, ToZext64, PS_loadruhabs>;
2221 def: Loadam_pat<extloadi32, i64, anyimm2, ToAext64, PS_loadriabs>;
2222 def: Loadam_pat<sextloadi32, i64, anyimm2, ToSext64, PS_loadriabs>;
2223 def: Loadam_pat<zextloadi32, i64, anyimm2, ToZext64, PS_loadriabs>;
2225 def: Loadam_pat<load, i1, anyimm0, I32toI1, PS_loadrubabs>;
2226 def: Loadam_pat<zextloadi1, i64, anyimm0, ToZext64, PS_loadrubabs>;
2229 // GP-relative address
2231 let AddedComplexity = 100 in {
2232 def: Loada_pat<extloadi1, i32, addrgp, L2_loadrubgp>;
2233 def: Loada_pat<zextloadi1, i32, addrgp, L2_loadrubgp>;
2234 def: Loada_pat<extloadi8, i32, addrgp, L2_loadrubgp>;
2235 def: Loada_pat<sextloadi8, i32, addrgp, L2_loadrbgp>;
2236 def: Loada_pat<zextloadi8, i32, addrgp, L2_loadrubgp>;
2237 def: Loada_pat<extloadi16, i32, addrgp, L2_loadruhgp>;
2238 def: Loada_pat<sextloadi16, i32, addrgp, L2_loadrhgp>;
2239 def: Loada_pat<zextloadi16, i32, addrgp, L2_loadruhgp>;
2240 def: Loada_pat<load, i32, addrgp, L2_loadrigp>;
2241 def: Loada_pat<load, v2i16, addrgp, L2_loadrigp>;
2242 def: Loada_pat<load, v4i8, addrgp, L2_loadrigp>;
2243 def: Loada_pat<load, i64, addrgp, L2_loadrdgp>;
2244 def: Loada_pat<load, v2i32, addrgp, L2_loadrdgp>;
2245 def: Loada_pat<load, v4i16, addrgp, L2_loadrdgp>;
2246 def: Loada_pat<load, v8i8, addrgp, L2_loadrdgp>;
2247 def: Loada_pat<load, f32, addrgp, L2_loadrigp>;
2248 def: Loada_pat<load, f64, addrgp, L2_loadrdgp>;
2250 def: Loada_pat<atomic_load_8, i32, addrgp, L2_loadrubgp>;
2251 def: Loada_pat<atomic_load_16, i32, addrgp, L2_loadruhgp>;
2252 def: Loada_pat<atomic_load_32, i32, addrgp, L2_loadrigp>;
2253 def: Loada_pat<atomic_load_64, i64, addrgp, L2_loadrdgp>;
2256 let AddedComplexity = 70 in {
2257 def: Loadam_pat<extloadi8, i64, addrgp, ToAext64, L2_loadrubgp>;
2258 def: Loadam_pat<sextloadi8, i64, addrgp, ToSext64, L2_loadrbgp>;
2259 def: Loadam_pat<zextloadi8, i64, addrgp, ToZext64, L2_loadrubgp>;
2260 def: Loadam_pat<extloadi16, i64, addrgp, ToAext64, L2_loadruhgp>;
2261 def: Loadam_pat<sextloadi16, i64, addrgp, ToSext64, L2_loadrhgp>;
2262 def: Loadam_pat<zextloadi16, i64, addrgp, ToZext64, L2_loadruhgp>;
2263 def: Loadam_pat<extloadi32, i64, addrgp, ToAext64, L2_loadrigp>;
2264 def: Loadam_pat<sextloadi32, i64, addrgp, ToSext64, L2_loadrigp>;
2265 def: Loadam_pat<zextloadi32, i64, addrgp, ToZext64, L2_loadrigp>;
2267 def: Loadam_pat<load, i1, addrgp, I32toI1, L2_loadrubgp>;
2268 def: Loadam_pat<zextloadi1, i64, addrgp, ToZext64, L2_loadrubgp>;
2272 // Sign-extending loads of i1 need to replicate the lowest bit throughout
2273 // the 32-bit value. Since the loaded value can only be 0 or 1, 0-v should
2275 let AddedComplexity = 20 in
2276 def: Pat<(i32 (sextloadi1 I32:$Rs)),
2277 (A2_subri 0, (L2_loadrub_io IntRegs:$Rs, 0))>;
2279 // Patterns for loads of i1:
2280 def: Pat<(i1 (load AddrFI:$fi)),
2281 (C2_tfrrp (L2_loadrub_io AddrFI:$fi, 0))>;
2282 def: Pat<(i1 (load (add I32:$Rs, anyimm0:$Off))),
2283 (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, imm:$Off))>;
2284 def: Pat<(i1 (load I32:$Rs)),
2285 (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, 0))>;
2288 // --(13) Store ----------------------------------------------------------
2291 class Storepi_pat<PatFrag Store, PatFrag Value, PatFrag Offset, InstHexagon MI>
2292 : Pat<(Store Value:$Rt, I32:$Rx, Offset:$s4),
2293 (MI I32:$Rx, imm:$s4, Value:$Rt)>;
2295 def: Storepi_pat<post_truncsti8, I32, s4_0ImmPred, S2_storerb_pi>;
2296 def: Storepi_pat<post_truncsti16, I32, s4_1ImmPred, S2_storerh_pi>;
2297 def: Storepi_pat<post_store, I32, s4_2ImmPred, S2_storeri_pi>;
2298 def: Storepi_pat<post_store, I64, s4_3ImmPred, S2_storerd_pi>;
2300 // Patterns for generating stores, where the address takes different forms:
2302 // - frameindex + offset,
2304 // - simple (base address without offset).
2305 // These would usually be used together (via Storexi_pat defined below), but
2306 // in some cases one may want to apply different properties (such as
2307 // AddedComplexity) to the individual patterns.
2308 class Storexi_fi_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2309 : Pat<(Store Value:$Rs, AddrFI:$fi), (MI AddrFI:$fi, 0, Value:$Rs)>;
2311 multiclass Storexi_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2313 def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
2314 (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
2315 def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
2316 (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
2319 multiclass Storexi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2321 def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
2322 (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
2323 def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
2324 (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
2327 class Storexi_base_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2328 : Pat<(Store Value:$Rt, I32:$Rs),
2329 (MI IntRegs:$Rs, 0, Value:$Rt)>;
2331 // Patterns for generating stores, where the address takes different forms,
2332 // and where the value being stored is transformed through the value modifier
2333 // ValueMod. The address forms are same as above.
2334 class Storexim_fi_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2336 : Pat<(Store Value:$Rs, AddrFI:$fi),
2337 (MI AddrFI:$fi, 0, (ValueMod Value:$Rs))>;
2339 multiclass Storexim_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2340 PatFrag ValueMod, InstHexagon MI> {
2341 def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
2342 (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
2343 def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
2344 (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
2347 multiclass Storexim_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2348 PatFrag ValueMod, InstHexagon MI> {
2349 def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
2350 (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
2351 def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
2352 (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
2355 class Storexim_base_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2357 : Pat<(Store Value:$Rt, I32:$Rs),
2358 (MI IntRegs:$Rs, 0, (ValueMod Value:$Rt))>;
2360 multiclass Storexi_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
2362 defm: Storexi_fi_add_pat <Store, Value, ImmPred, MI>;
2363 def: Storexi_fi_pat <Store, Value, MI>;
2364 defm: Storexi_add_pat <Store, Value, ImmPred, MI>;
2367 multiclass Storexim_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
2368 PatFrag ValueMod, InstHexagon MI> {
2369 defm: Storexim_fi_add_pat <Store, Value, ImmPred, ValueMod, MI>;
2370 def: Storexim_fi_pat <Store, Value, ValueMod, MI>;
2371 defm: Storexim_add_pat <Store, Value, ImmPred, ValueMod, MI>;
2375 class Storexu_shl_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred, InstHexagon MI>
2376 : Pat<(Store Value:$Rt, (add (shl I32:$Ru, u2_0ImmPred:$u2), ImmPred:$A)),
2377 (MI IntRegs:$Ru, imm:$u2, ImmPred:$A, Value:$Rt)>;
2380 class Storexr_shl_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2381 : Pat<(Store Value:$Ru, (add I32:$Rs, (shl I32:$Rt, u2_0ImmPred:$u2))),
2382 (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2, Value:$Ru)>;
2385 class Storexr_add_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2386 : Pat<(Store Value:$Ru, (add I32:$Rs, I32:$Rt)),
2387 (MI IntRegs:$Rs, IntRegs:$Rt, 0, Value:$Ru)>;
2389 class Storea_pat<PatFrag Store, PatFrag Value, PatFrag Addr, InstHexagon MI>
2390 : Pat<(Store Value:$val, Addr:$addr), (MI Addr:$addr, Value:$val)>;
2392 class Stoream_pat<PatFrag Store, PatFrag Value, PatFrag Addr, PatFrag ValueMod,
2394 : Pat<(Store Value:$val, Addr:$addr),
2395 (MI Addr:$addr, (ValueMod Value:$val))>;
2397 // Regular stores in the DAG have two operands: value and address.
2398 // Atomic stores also have two, but they are reversed: address, value.
2399 // To use atomic stores with the patterns, they need to have their operands
2400 // swapped. This relies on the knowledge that the F.Fragment uses names
2402 class AtomSt<PatFrag F>
2403 : PatFrag<(ops node:$val, node:$ptr), !head(F.Fragments), F.PredicateCode,
2404 F.OperandTransform> {
2405 let IsAtomic = F.IsAtomic;
2406 let MemoryVT = F.MemoryVT;
2410 def IMM_BYTE : SDNodeXForm<imm, [{
2411 // -1 can be represented as 255, etc.
2412 // assigning to a byte restores our desired signed value.
2413 int8_t imm = N->getSExtValue();
2414 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2417 def IMM_HALF : SDNodeXForm<imm, [{
2418 // -1 can be represented as 65535, etc.
2419 // assigning to a short restores our desired signed value.
2420 int16_t imm = N->getSExtValue();
2421 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2424 def IMM_WORD : SDNodeXForm<imm, [{
2425 // -1 can be represented as 4294967295, etc.
2426 // Currently, it's not doing this. But some optimization
2427 // might convert -1 to a large +ve number.
2428 // assigning to a word restores our desired signed value.
2429 int32_t imm = N->getSExtValue();
2430 return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2433 def ToImmByte : OutPatFrag<(ops node:$R), (IMM_BYTE $R)>;
2434 def ToImmHalf : OutPatFrag<(ops node:$R), (IMM_HALF $R)>;
2435 def ToImmWord : OutPatFrag<(ops node:$R), (IMM_WORD $R)>;
2437 // Even though the offset is not extendable in the store-immediate, we
2438 // can still generate the fi# in the base address. If the final offset
2439 // is not valid for the instruction, we will replace it with a scratch
2441 class SmallStackStore<PatFrag Store>
2442 : PatFrag<(ops node:$Val, node:$Addr), (Store node:$Val, node:$Addr), [{
2443 return isSmallStackStore(cast<StoreSDNode>(N));
2446 // This is the complement of SmallStackStore.
2447 class LargeStackStore<PatFrag Store>
2448 : PatFrag<(ops node:$Val, node:$Addr), (Store node:$Val, node:$Addr), [{
2449 return !isSmallStackStore(cast<StoreSDNode>(N));
2452 // Preferred addressing modes for various combinations of stored value
2453 // and address computation.
2454 // For stores where the address and value are both immediates, prefer
2455 // store-immediate. The reason is that the constant-extender optimization
2456 // can replace store-immediate with a store-register, but there is nothing
2457 // to generate a store-immediate out of a store-register.
2459 // C R F F+C R+C R+R R<<S+C R<<S+R
2460 // --+-------+-----+-----+------+-----+-----+--------+--------
2461 // C | imm | imm | imm | imm | imm | rr | ur | rr
2462 // R | abs* | io | io | io | io | rr | ur | rr
2464 // (*) Absolute or GP-relative.
2466 // Note that any expression can be matched by Reg. In particular, an immediate
2467 // can always be placed in a register, so patterns checking for Imm should
2468 // have a higher priority than the ones involving Reg that could also match.
2469 // For example, *(p+4) could become r1=#4; memw(r0+r1<<#0) instead of the
2470 // preferred memw(r0+#4). Similarly Reg+Imm or Reg+Reg should be tried before
2473 // The order in which the different combinations are tried:
2475 // C F R F+C R+C R+R R<<S+C R<<S+R
2476 // --+-------+-----+-----+------+-----+-----+--------+--------
2477 // C | 1 | 6 | - | 5 | 9 | - | - | -
2478 // R | 2 | 8 | 12 | 7 | 10 | 11 | 3 | 4
2481 // First, match the unusual case of doubleword store into Reg+Imm4, i.e.
2482 // a store where the offset Imm4 is a multiple of 4, but not of 8. This
2483 // implies that Reg is also a proper multiple of 4. To still generate a
2484 // doubleword store, add 4 to Reg, and subtract 4 from the offset.
2486 def s30_2ProperPred : PatLeaf<(i32 imm), [{
2487 int64_t v = (int64_t)N->getSExtValue();
2488 return isShiftedInt<30,2>(v) && !isShiftedInt<29,3>(v);
2490 def RoundTo8 : SDNodeXForm<imm, [{
2491 int32_t Imm = N->getSExtValue();
2492 return CurDAG->getTargetConstant(Imm & -8, SDLoc(N), MVT::i32);
2495 let AddedComplexity = 150 in
2496 def: Pat<(store I64:$Ru, (add I32:$Rs, s30_2ProperPred:$Off)),
2497 (S2_storerd_io (A2_addi I32:$Rs, 4), (RoundTo8 $Off), I64:$Ru)>;
2499 class Storexi_abs_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2500 : Pat<(Store Value:$val, anyimm:$addr),
2501 (MI (ToI32 $addr), 0, Value:$val)>;
2502 class Storexim_abs_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2504 : Pat<(Store Value:$val, anyimm:$addr),
2505 (MI (ToI32 $addr), 0, (ValueMod Value:$val))>;
2507 let AddedComplexity = 140 in {
2508 def: Storexim_abs_pat<truncstorei8, anyint, ToImmByte, S4_storeirb_io>;
2509 def: Storexim_abs_pat<truncstorei16, anyint, ToImmHalf, S4_storeirh_io>;
2510 def: Storexim_abs_pat<store, anyint, ToImmWord, S4_storeiri_io>;
2512 def: Storexi_abs_pat<truncstorei8, anyimm, S4_storeirb_io>;
2513 def: Storexi_abs_pat<truncstorei16, anyimm, S4_storeirh_io>;
2514 def: Storexi_abs_pat<store, anyimm, S4_storeiri_io>;
2517 // GP-relative address
2518 let AddedComplexity = 120 in {
2519 def: Storea_pat<truncstorei8, I32, addrgp, S2_storerbgp>;
2520 def: Storea_pat<truncstorei16, I32, addrgp, S2_storerhgp>;
2521 def: Storea_pat<store, I32, addrgp, S2_storerigp>;
2522 def: Storea_pat<store, V4I8, addrgp, S2_storerigp>;
2523 def: Storea_pat<store, V2I16, addrgp, S2_storerigp>;
2524 def: Storea_pat<store, I64, addrgp, S2_storerdgp>;
2525 def: Storea_pat<store, V8I8, addrgp, S2_storerdgp>;
2526 def: Storea_pat<store, V4I16, addrgp, S2_storerdgp>;
2527 def: Storea_pat<store, V2I32, addrgp, S2_storerdgp>;
2528 def: Storea_pat<store, F32, addrgp, S2_storerigp>;
2529 def: Storea_pat<store, F64, addrgp, S2_storerdgp>;
2530 def: Storea_pat<AtomSt<atomic_store_8>, I32, addrgp, S2_storerbgp>;
2531 def: Storea_pat<AtomSt<atomic_store_16>, I32, addrgp, S2_storerhgp>;
2532 def: Storea_pat<AtomSt<atomic_store_32>, I32, addrgp, S2_storerigp>;
2533 def: Storea_pat<AtomSt<atomic_store_32>, V4I8, addrgp, S2_storerigp>;
2534 def: Storea_pat<AtomSt<atomic_store_32>, V2I16, addrgp, S2_storerigp>;
2535 def: Storea_pat<AtomSt<atomic_store_64>, I64, addrgp, S2_storerdgp>;
2536 def: Storea_pat<AtomSt<atomic_store_64>, V8I8, addrgp, S2_storerdgp>;
2537 def: Storea_pat<AtomSt<atomic_store_64>, V4I16, addrgp, S2_storerdgp>;
2538 def: Storea_pat<AtomSt<atomic_store_64>, V2I32, addrgp, S2_storerdgp>;
2540 def: Stoream_pat<truncstorei8, I64, addrgp, LoReg, S2_storerbgp>;
2541 def: Stoream_pat<truncstorei16, I64, addrgp, LoReg, S2_storerhgp>;
2542 def: Stoream_pat<truncstorei32, I64, addrgp, LoReg, S2_storerigp>;
2543 def: Stoream_pat<store, I1, addrgp, I1toI32, S2_storerbgp>;
2547 let AddedComplexity = 110 in {
2548 def: Storea_pat<truncstorei8, I32, anyimm0, PS_storerbabs>;
2549 def: Storea_pat<truncstorei16, I32, anyimm1, PS_storerhabs>;
2550 def: Storea_pat<store, I32, anyimm2, PS_storeriabs>;
2551 def: Storea_pat<store, V4I8, anyimm2, PS_storeriabs>;
2552 def: Storea_pat<store, V2I16, anyimm2, PS_storeriabs>;
2553 def: Storea_pat<store, I64, anyimm3, PS_storerdabs>;
2554 def: Storea_pat<store, V8I8, anyimm3, PS_storerdabs>;
2555 def: Storea_pat<store, V4I16, anyimm3, PS_storerdabs>;
2556 def: Storea_pat<store, V2I32, anyimm3, PS_storerdabs>;
2557 def: Storea_pat<store, F32, anyimm2, PS_storeriabs>;
2558 def: Storea_pat<store, F64, anyimm3, PS_storerdabs>;
2559 def: Storea_pat<AtomSt<atomic_store_8>, I32, anyimm0, PS_storerbabs>;
2560 def: Storea_pat<AtomSt<atomic_store_16>, I32, anyimm1, PS_storerhabs>;
2561 def: Storea_pat<AtomSt<atomic_store_32>, I32, anyimm2, PS_storeriabs>;
2562 def: Storea_pat<AtomSt<atomic_store_32>, V4I8, anyimm2, PS_storeriabs>;
2563 def: Storea_pat<AtomSt<atomic_store_32>, V2I16, anyimm2, PS_storeriabs>;
2564 def: Storea_pat<AtomSt<atomic_store_64>, I64, anyimm3, PS_storerdabs>;
2565 def: Storea_pat<AtomSt<atomic_store_64>, V8I8, anyimm3, PS_storerdabs>;
2566 def: Storea_pat<AtomSt<atomic_store_64>, V4I16, anyimm3, PS_storerdabs>;
2567 def: Storea_pat<AtomSt<atomic_store_64>, V2I32, anyimm3, PS_storerdabs>;
2569 def: Stoream_pat<truncstorei8, I64, anyimm0, LoReg, PS_storerbabs>;
2570 def: Stoream_pat<truncstorei16, I64, anyimm1, LoReg, PS_storerhabs>;
2571 def: Stoream_pat<truncstorei32, I64, anyimm2, LoReg, PS_storeriabs>;
2572 def: Stoream_pat<store, I1, anyimm0, I1toI32, PS_storerbabs>;
2576 let AddedComplexity = 100 in {
2577 def: Storexu_shl_pat<truncstorei8, I32, anyimm0, S4_storerb_ur>;
2578 def: Storexu_shl_pat<truncstorei16, I32, anyimm1, S4_storerh_ur>;
2579 def: Storexu_shl_pat<store, I32, anyimm2, S4_storeri_ur>;
2580 def: Storexu_shl_pat<store, V4I8, anyimm2, S4_storeri_ur>;
2581 def: Storexu_shl_pat<store, V2I16, anyimm2, S4_storeri_ur>;
2582 def: Storexu_shl_pat<store, I64, anyimm3, S4_storerd_ur>;
2583 def: Storexu_shl_pat<store, V8I8, anyimm3, S4_storerd_ur>;
2584 def: Storexu_shl_pat<store, V4I16, anyimm3, S4_storerd_ur>;
2585 def: Storexu_shl_pat<store, V2I32, anyimm3, S4_storerd_ur>;
2586 def: Storexu_shl_pat<store, F32, anyimm2, S4_storeri_ur>;
2587 def: Storexu_shl_pat<store, F64, anyimm3, S4_storerd_ur>;
2589 def: Pat<(store I1:$Pu, (add (shl I32:$Rs, u2_0ImmPred:$u2), anyimm:$A)),
2590 (S4_storerb_ur IntRegs:$Rs, imm:$u2, imm:$A, (I1toI32 I1:$Pu))>;
2594 let AddedComplexity = 90 in {
2595 def: Storexr_shl_pat<truncstorei8, I32, S4_storerb_rr>;
2596 def: Storexr_shl_pat<truncstorei16, I32, S4_storerh_rr>;
2597 def: Storexr_shl_pat<store, I32, S4_storeri_rr>;
2598 def: Storexr_shl_pat<store, V4I8, S4_storeri_rr>;
2599 def: Storexr_shl_pat<store, V2I16, S4_storeri_rr>;
2600 def: Storexr_shl_pat<store, I64, S4_storerd_rr>;
2601 def: Storexr_shl_pat<store, V8I8, S4_storerd_rr>;
2602 def: Storexr_shl_pat<store, V4I16, S4_storerd_rr>;
2603 def: Storexr_shl_pat<store, V2I32, S4_storerd_rr>;
2604 def: Storexr_shl_pat<store, F32, S4_storeri_rr>;
2605 def: Storexr_shl_pat<store, F64, S4_storerd_rr>;
2607 def: Pat<(store I1:$Pu, (add (shl I32:$Rs, u2_0ImmPred:$u2), I32:$Rt)),
2608 (S4_storerb_ur IntRegs:$Rt, IntRegs:$Rs, imm:$u2, (I1toI32 I1:$Pu))>;
2611 class SS_<PatFrag F> : SmallStackStore<F>;
2612 class LS_<PatFrag F> : LargeStackStore<F>;
2614 multiclass IMFA_<PatFrag S, PatFrag V, PatFrag O, PatFrag M, InstHexagon I> {
2615 defm: Storexim_fi_add_pat<S, V, O, M, I>;
2617 multiclass IFA_<PatFrag S, PatFrag V, PatFrag O, InstHexagon I> {
2618 defm: Storexi_fi_add_pat<S, V, O, I>;
2621 // Fi+Imm, store-immediate
2622 let AddedComplexity = 80 in {
2623 defm: IMFA_<SS_<truncstorei8>, anyint, u6_0ImmPred, ToImmByte, S4_storeirb_io>;
2624 defm: IMFA_<SS_<truncstorei16>, anyint, u6_1ImmPred, ToImmHalf, S4_storeirh_io>;
2625 defm: IMFA_<SS_<store>, anyint, u6_2ImmPred, ToImmWord, S4_storeiri_io>;
2627 defm: IFA_<SS_<truncstorei8>, anyimm, u6_0ImmPred, S4_storeirb_io>;
2628 defm: IFA_<SS_<truncstorei16>, anyimm, u6_1ImmPred, S4_storeirh_io>;
2629 defm: IFA_<SS_<store>, anyimm, u6_2ImmPred, S4_storeiri_io>;
2631 // For large-stack stores, generate store-register (prefer explicit Fi
2633 defm: IMFA_<LS_<truncstorei8>, anyimm, u6_0ImmPred, ToI32, S2_storerb_io>;
2634 defm: IMFA_<LS_<truncstorei16>, anyimm, u6_1ImmPred, ToI32, S2_storerh_io>;
2635 defm: IMFA_<LS_<store>, anyimm, u6_2ImmPred, ToI32, S2_storeri_io>;
2638 // Fi, store-immediate
2639 let AddedComplexity = 70 in {
2640 def: Storexim_fi_pat<SS_<truncstorei8>, anyint, ToImmByte, S4_storeirb_io>;
2641 def: Storexim_fi_pat<SS_<truncstorei16>, anyint, ToImmHalf, S4_storeirh_io>;
2642 def: Storexim_fi_pat<SS_<store>, anyint, ToImmWord, S4_storeiri_io>;
2644 def: Storexi_fi_pat<SS_<truncstorei8>, anyimm, S4_storeirb_io>;
2645 def: Storexi_fi_pat<SS_<truncstorei16>, anyimm, S4_storeirh_io>;
2646 def: Storexi_fi_pat<SS_<store>, anyimm, S4_storeiri_io>;
2648 // For large-stack stores, generate store-register (prefer explicit Fi
2650 def: Storexim_fi_pat<LS_<truncstorei8>, anyimm, ToI32, S2_storerb_io>;
2651 def: Storexim_fi_pat<LS_<truncstorei16>, anyimm, ToI32, S2_storerh_io>;
2652 def: Storexim_fi_pat<LS_<store>, anyimm, ToI32, S2_storeri_io>;
2655 // Fi+Imm, Fi, store-register
2656 let AddedComplexity = 60 in {
2657 defm: Storexi_fi_add_pat<truncstorei8, I32, anyimm, S2_storerb_io>;
2658 defm: Storexi_fi_add_pat<truncstorei16, I32, anyimm, S2_storerh_io>;
2659 defm: Storexi_fi_add_pat<store, I32, anyimm, S2_storeri_io>;
2660 defm: Storexi_fi_add_pat<store, V4I8, anyimm, S2_storeri_io>;
2661 defm: Storexi_fi_add_pat<store, V2I16, anyimm, S2_storeri_io>;
2662 defm: Storexi_fi_add_pat<store, I64, anyimm, S2_storerd_io>;
2663 defm: Storexi_fi_add_pat<store, V8I8, anyimm, S2_storerd_io>;
2664 defm: Storexi_fi_add_pat<store, V4I16, anyimm, S2_storerd_io>;
2665 defm: Storexi_fi_add_pat<store, V2I32, anyimm, S2_storerd_io>;
2666 defm: Storexi_fi_add_pat<store, F32, anyimm, S2_storeri_io>;
2667 defm: Storexi_fi_add_pat<store, F64, anyimm, S2_storerd_io>;
2668 defm: Storexim_fi_add_pat<store, I1, anyimm, I1toI32, S2_storerb_io>;
2670 def: Storexi_fi_pat<truncstorei8, I32, S2_storerb_io>;
2671 def: Storexi_fi_pat<truncstorei16, I32, S2_storerh_io>;
2672 def: Storexi_fi_pat<store, I32, S2_storeri_io>;
2673 def: Storexi_fi_pat<store, V4I8, S2_storeri_io>;
2674 def: Storexi_fi_pat<store, V2I16, S2_storeri_io>;
2675 def: Storexi_fi_pat<store, I64, S2_storerd_io>;
2676 def: Storexi_fi_pat<store, V8I8, S2_storerd_io>;
2677 def: Storexi_fi_pat<store, V4I16, S2_storerd_io>;
2678 def: Storexi_fi_pat<store, V2I32, S2_storerd_io>;
2679 def: Storexi_fi_pat<store, F32, S2_storeri_io>;
2680 def: Storexi_fi_pat<store, F64, S2_storerd_io>;
2681 def: Storexim_fi_pat<store, I1, I1toI32, S2_storerb_io>;
2685 multiclass IMRA_<PatFrag S, PatFrag V, PatFrag O, PatFrag M, InstHexagon I> {
2686 defm: Storexim_add_pat<S, V, O, M, I>;
2688 multiclass IRA_<PatFrag S, PatFrag V, PatFrag O, InstHexagon I> {
2689 defm: Storexi_add_pat<S, V, O, I>;
2692 // Reg+Imm, store-immediate
2693 let AddedComplexity = 50 in {
2694 defm: IMRA_<truncstorei8, anyint, u6_0ImmPred, ToImmByte, S4_storeirb_io>;
2695 defm: IMRA_<truncstorei16, anyint, u6_1ImmPred, ToImmHalf, S4_storeirh_io>;
2696 defm: IMRA_<store, anyint, u6_2ImmPred, ToImmWord, S4_storeiri_io>;
2698 defm: IRA_<truncstorei8, anyimm, u6_0ImmPred, S4_storeirb_io>;
2699 defm: IRA_<truncstorei16, anyimm, u6_1ImmPred, S4_storeirh_io>;
2700 defm: IRA_<store, anyimm, u6_2ImmPred, S4_storeiri_io>;
2703 // Reg+Imm, store-register
2704 let AddedComplexity = 40 in {
2705 defm: Storexi_pat<truncstorei8, I32, anyimm0, S2_storerb_io>;
2706 defm: Storexi_pat<truncstorei16, I32, anyimm1, S2_storerh_io>;
2707 defm: Storexi_pat<store, I32, anyimm2, S2_storeri_io>;
2708 defm: Storexi_pat<store, V4I8, anyimm2, S2_storeri_io>;
2709 defm: Storexi_pat<store, V2I16, anyimm2, S2_storeri_io>;
2710 defm: Storexi_pat<store, I64, anyimm3, S2_storerd_io>;
2711 defm: Storexi_pat<store, V8I8, anyimm3, S2_storerd_io>;
2712 defm: Storexi_pat<store, V4I16, anyimm3, S2_storerd_io>;
2713 defm: Storexi_pat<store, V2I32, anyimm3, S2_storerd_io>;
2714 defm: Storexi_pat<store, F32, anyimm2, S2_storeri_io>;
2715 defm: Storexi_pat<store, F64, anyimm3, S2_storerd_io>;
2717 defm: Storexim_pat<truncstorei8, I64, anyimm0, LoReg, S2_storerb_io>;
2718 defm: Storexim_pat<truncstorei16, I64, anyimm1, LoReg, S2_storerh_io>;
2719 defm: Storexim_pat<truncstorei32, I64, anyimm2, LoReg, S2_storeri_io>;
2720 defm: Storexim_pat<store, I1, anyimm0, I1toI32, S2_storerb_io>;
2722 defm: Storexi_pat<AtomSt<atomic_store_8>, I32, anyimm0, S2_storerb_io>;
2723 defm: Storexi_pat<AtomSt<atomic_store_16>, I32, anyimm1, S2_storerh_io>;
2724 defm: Storexi_pat<AtomSt<atomic_store_32>, I32, anyimm2, S2_storeri_io>;
2725 defm: Storexi_pat<AtomSt<atomic_store_32>, V4I8, anyimm2, S2_storeri_io>;
2726 defm: Storexi_pat<AtomSt<atomic_store_32>, V2I16, anyimm2, S2_storeri_io>;
2727 defm: Storexi_pat<AtomSt<atomic_store_64>, I64, anyimm3, S2_storerd_io>;
2728 defm: Storexi_pat<AtomSt<atomic_store_64>, V8I8, anyimm3, S2_storerd_io>;
2729 defm: Storexi_pat<AtomSt<atomic_store_64>, V4I16, anyimm3, S2_storerd_io>;
2730 defm: Storexi_pat<AtomSt<atomic_store_64>, V2I32, anyimm3, S2_storerd_io>;
2734 let AddedComplexity = 30 in {
2735 def: Storexr_add_pat<truncstorei8, I32, S4_storerb_rr>;
2736 def: Storexr_add_pat<truncstorei16, I32, S4_storerh_rr>;
2737 def: Storexr_add_pat<store, I32, S4_storeri_rr>;
2738 def: Storexr_add_pat<store, V4I8, S4_storeri_rr>;
2739 def: Storexr_add_pat<store, V2I16, S4_storeri_rr>;
2740 def: Storexr_add_pat<store, I64, S4_storerd_rr>;
2741 def: Storexr_add_pat<store, V8I8, S4_storerd_rr>;
2742 def: Storexr_add_pat<store, V4I16, S4_storerd_rr>;
2743 def: Storexr_add_pat<store, V2I32, S4_storerd_rr>;
2744 def: Storexr_add_pat<store, F32, S4_storeri_rr>;
2745 def: Storexr_add_pat<store, F64, S4_storerd_rr>;
2747 def: Pat<(store I1:$Pu, (add I32:$Rs, I32:$Rt)),
2748 (S4_storerb_rr IntRegs:$Rs, IntRegs:$Rt, 0, (I1toI32 I1:$Pu))>;
2751 // Reg, store-immediate
2752 let AddedComplexity = 20 in {
2753 def: Storexim_base_pat<truncstorei8, anyint, ToImmByte, S4_storeirb_io>;
2754 def: Storexim_base_pat<truncstorei16, anyint, ToImmHalf, S4_storeirh_io>;
2755 def: Storexim_base_pat<store, anyint, ToImmWord, S4_storeiri_io>;
2757 def: Storexi_base_pat<truncstorei8, anyimm, S4_storeirb_io>;
2758 def: Storexi_base_pat<truncstorei16, anyimm, S4_storeirh_io>;
2759 def: Storexi_base_pat<store, anyimm, S4_storeiri_io>;
2762 // Reg, store-register
2763 let AddedComplexity = 10 in {
2764 def: Storexi_base_pat<truncstorei8, I32, S2_storerb_io>;
2765 def: Storexi_base_pat<truncstorei16, I32, S2_storerh_io>;
2766 def: Storexi_base_pat<store, I32, S2_storeri_io>;
2767 def: Storexi_base_pat<store, V4I8, S2_storeri_io>;
2768 def: Storexi_base_pat<store, V2I16, S2_storeri_io>;
2769 def: Storexi_base_pat<store, I64, S2_storerd_io>;
2770 def: Storexi_base_pat<store, V8I8, S2_storerd_io>;
2771 def: Storexi_base_pat<store, V4I16, S2_storerd_io>;
2772 def: Storexi_base_pat<store, V2I32, S2_storerd_io>;
2773 def: Storexi_base_pat<store, F32, S2_storeri_io>;
2774 def: Storexi_base_pat<store, F64, S2_storerd_io>;
2776 def: Storexim_base_pat<truncstorei8, I64, LoReg, S2_storerb_io>;
2777 def: Storexim_base_pat<truncstorei16, I64, LoReg, S2_storerh_io>;
2778 def: Storexim_base_pat<truncstorei32, I64, LoReg, S2_storeri_io>;
2779 def: Storexim_base_pat<store, I1, I1toI32, S2_storerb_io>;
2781 def: Storexi_base_pat<AtomSt<atomic_store_8>, I32, S2_storerb_io>;
2782 def: Storexi_base_pat<AtomSt<atomic_store_16>, I32, S2_storerh_io>;
2783 def: Storexi_base_pat<AtomSt<atomic_store_32>, I32, S2_storeri_io>;
2784 def: Storexi_base_pat<AtomSt<atomic_store_32>, V4I8, S2_storeri_io>;
2785 def: Storexi_base_pat<AtomSt<atomic_store_32>, V2I16, S2_storeri_io>;
2786 def: Storexi_base_pat<AtomSt<atomic_store_64>, I64, S2_storerd_io>;
2787 def: Storexi_base_pat<AtomSt<atomic_store_64>, V8I8, S2_storerd_io>;
2788 def: Storexi_base_pat<AtomSt<atomic_store_64>, V4I16, S2_storerd_io>;
2789 def: Storexi_base_pat<AtomSt<atomic_store_64>, V2I32, S2_storerd_io>;
2793 // --(14) Memop ----------------------------------------------------------
2796 def m5_0Imm8Pred : PatLeaf<(i32 imm), [{
2797 int8_t V = N->getSExtValue();
2798 return -32 < V && V <= -1;
2801 def m5_0Imm16Pred : PatLeaf<(i32 imm), [{
2802 int16_t V = N->getSExtValue();
2803 return -32 < V && V <= -1;
2806 def m5_0ImmPred : PatLeaf<(i32 imm), [{
2807 int64_t V = N->getSExtValue();
2808 return -31 <= V && V <= -1;
2811 def IsNPow2_8 : PatLeaf<(i32 imm), [{
2812 uint8_t NV = ~N->getZExtValue();
2813 return isPowerOf2_32(NV);
2816 def IsNPow2_16 : PatLeaf<(i32 imm), [{
2817 uint16_t NV = ~N->getZExtValue();
2818 return isPowerOf2_32(NV);
2821 def Log2_8 : SDNodeXForm<imm, [{
2822 uint8_t V = N->getZExtValue();
2823 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
2826 def Log2_16 : SDNodeXForm<imm, [{
2827 uint16_t V = N->getZExtValue();
2828 return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
2831 def LogN2_8 : SDNodeXForm<imm, [{
2832 uint8_t NV = ~N->getZExtValue();
2833 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
2836 def LogN2_16 : SDNodeXForm<imm, [{
2837 uint16_t NV = ~N->getZExtValue();
2838 return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
2841 def IdImm : SDNodeXForm<imm, [{ return SDValue(N, 0); }]>;
2843 multiclass Memopxr_base_pat<PatFrag Load, PatFrag Store, SDNode Oper,
2846 def: Pat<(Store (Oper (Load I32:$Rs), I32:$A), I32:$Rs),
2847 (MI I32:$Rs, 0, I32:$A)>;
2849 def: Pat<(Store (Oper (Load AddrFI:$Rs), I32:$A), AddrFI:$Rs),
2850 (MI AddrFI:$Rs, 0, I32:$A)>;
2853 multiclass Memopxr_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2854 SDNode Oper, InstHexagon MI> {
2856 def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), I32:$A),
2857 (add I32:$Rs, ImmPred:$Off)),
2858 (MI I32:$Rs, imm:$Off, I32:$A)>;
2859 def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), I32:$A),
2860 (IsOrAdd I32:$Rs, ImmPred:$Off)),
2861 (MI I32:$Rs, imm:$Off, I32:$A)>;
2863 def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), I32:$A),
2864 (add AddrFI:$Rs, ImmPred:$Off)),
2865 (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
2866 def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), I32:$A),
2867 (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
2868 (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
2871 multiclass Memopxr_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2872 SDNode Oper, InstHexagon MI> {
2873 let Predicates = [UseMEMOPS] in {
2874 defm: Memopxr_base_pat <Load, Store, Oper, MI>;
2875 defm: Memopxr_add_pat <Load, Store, ImmPred, Oper, MI>;
2879 let AddedComplexity = 200 in {
2881 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, add,
2882 /*anyext*/ L4_add_memopb_io>;
2883 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, add,
2884 /*sext*/ L4_add_memopb_io>;
2885 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, add,
2886 /*zext*/ L4_add_memopb_io>;
2887 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, add,
2888 /*anyext*/ L4_add_memoph_io>;
2889 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, add,
2890 /*sext*/ L4_add_memoph_io>;
2891 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, add,
2892 /*zext*/ L4_add_memoph_io>;
2893 defm: Memopxr_pat<load, store, u6_2ImmPred, add, L4_add_memopw_io>;
2896 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, sub,
2897 /*anyext*/ L4_sub_memopb_io>;
2898 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub,
2899 /*sext*/ L4_sub_memopb_io>;
2900 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub,
2901 /*zext*/ L4_sub_memopb_io>;
2902 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, sub,
2903 /*anyext*/ L4_sub_memoph_io>;
2904 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub,
2905 /*sext*/ L4_sub_memoph_io>;
2906 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub,
2907 /*zext*/ L4_sub_memoph_io>;
2908 defm: Memopxr_pat<load, store, u6_2ImmPred, sub, L4_sub_memopw_io>;
2911 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, and,
2912 /*anyext*/ L4_and_memopb_io>;
2913 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, and,
2914 /*sext*/ L4_and_memopb_io>;
2915 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, and,
2916 /*zext*/ L4_and_memopb_io>;
2917 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, and,
2918 /*anyext*/ L4_and_memoph_io>;
2919 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, and,
2920 /*sext*/ L4_and_memoph_io>;
2921 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, and,
2922 /*zext*/ L4_and_memoph_io>;
2923 defm: Memopxr_pat<load, store, u6_2ImmPred, and, L4_and_memopw_io>;
2926 defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, or,
2927 /*anyext*/ L4_or_memopb_io>;
2928 defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, or,
2929 /*sext*/ L4_or_memopb_io>;
2930 defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, or,
2931 /*zext*/ L4_or_memopb_io>;
2932 defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, or,
2933 /*anyext*/ L4_or_memoph_io>;
2934 defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, or,
2935 /*sext*/ L4_or_memoph_io>;
2936 defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, or,
2937 /*zext*/ L4_or_memoph_io>;
2938 defm: Memopxr_pat<load, store, u6_2ImmPred, or, L4_or_memopw_io>;
2942 multiclass Memopxi_base_pat<PatFrag Load, PatFrag Store, SDNode Oper,
2943 PatFrag Arg, SDNodeXForm ArgMod, InstHexagon MI> {
2945 def: Pat<(Store (Oper (Load I32:$Rs), Arg:$A), I32:$Rs),
2946 (MI I32:$Rs, 0, (ArgMod Arg:$A))>;
2948 def: Pat<(Store (Oper (Load AddrFI:$Rs), Arg:$A), AddrFI:$Rs),
2949 (MI AddrFI:$Rs, 0, (ArgMod Arg:$A))>;
2952 multiclass Memopxi_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2953 SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
2956 def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), Arg:$A),
2957 (add I32:$Rs, ImmPred:$Off)),
2958 (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2959 def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), Arg:$A),
2960 (IsOrAdd I32:$Rs, ImmPred:$Off)),
2961 (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2963 def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
2964 (add AddrFI:$Rs, ImmPred:$Off)),
2965 (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2966 def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
2967 (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
2968 (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
2971 multiclass Memopxi_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2972 SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
2974 let Predicates = [UseMEMOPS] in {
2975 defm: Memopxi_base_pat <Load, Store, Oper, Arg, ArgMod, MI>;
2976 defm: Memopxi_add_pat <Load, Store, ImmPred, Oper, Arg, ArgMod, MI>;
2980 let AddedComplexity = 220 in {
2982 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
2983 /*anyext*/ IdImm, L4_iadd_memopb_io>;
2984 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
2985 /*sext*/ IdImm, L4_iadd_memopb_io>;
2986 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
2987 /*zext*/ IdImm, L4_iadd_memopb_io>;
2988 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
2989 /*anyext*/ IdImm, L4_iadd_memoph_io>;
2990 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
2991 /*sext*/ IdImm, L4_iadd_memoph_io>;
2992 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
2993 /*zext*/ IdImm, L4_iadd_memoph_io>;
2994 defm: Memopxi_pat<load, store, u6_2ImmPred, add, u5_0ImmPred, IdImm,
2996 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
2997 /*anyext*/ NegImm8, L4_iadd_memopb_io>;
2998 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
2999 /*sext*/ NegImm8, L4_iadd_memopb_io>;
3000 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
3001 /*zext*/ NegImm8, L4_iadd_memopb_io>;
3002 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
3003 /*anyext*/ NegImm16, L4_iadd_memoph_io>;
3004 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
3005 /*sext*/ NegImm16, L4_iadd_memoph_io>;
3006 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
3007 /*zext*/ NegImm16, L4_iadd_memoph_io>;
3008 defm: Memopxi_pat<load, store, u6_2ImmPred, sub, m5_0ImmPred, NegImm32,
3012 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
3013 /*anyext*/ IdImm, L4_isub_memopb_io>;
3014 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
3015 /*sext*/ IdImm, L4_isub_memopb_io>;
3016 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
3017 /*zext*/ IdImm, L4_isub_memopb_io>;
3018 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
3019 /*anyext*/ IdImm, L4_isub_memoph_io>;
3020 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
3021 /*sext*/ IdImm, L4_isub_memoph_io>;
3022 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
3023 /*zext*/ IdImm, L4_isub_memoph_io>;
3024 defm: Memopxi_pat<load, store, u6_2ImmPred, sub, u5_0ImmPred, IdImm,
3026 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
3027 /*anyext*/ NegImm8, L4_isub_memopb_io>;
3028 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
3029 /*sext*/ NegImm8, L4_isub_memopb_io>;
3030 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
3031 /*zext*/ NegImm8, L4_isub_memopb_io>;
3032 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
3033 /*anyext*/ NegImm16, L4_isub_memoph_io>;
3034 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
3035 /*sext*/ NegImm16, L4_isub_memoph_io>;
3036 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
3037 /*zext*/ NegImm16, L4_isub_memoph_io>;
3038 defm: Memopxi_pat<load, store, u6_2ImmPred, add, m5_0ImmPred, NegImm32,
3042 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
3043 /*anyext*/ LogN2_8, L4_iand_memopb_io>;
3044 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
3045 /*sext*/ LogN2_8, L4_iand_memopb_io>;
3046 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
3047 /*zext*/ LogN2_8, L4_iand_memopb_io>;
3048 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
3049 /*anyext*/ LogN2_16, L4_iand_memoph_io>;
3050 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
3051 /*sext*/ LogN2_16, L4_iand_memoph_io>;
3052 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
3053 /*zext*/ LogN2_16, L4_iand_memoph_io>;
3054 defm: Memopxi_pat<load, store, u6_2ImmPred, and, IsNPow2_32,
3055 LogN2_32, L4_iand_memopw_io>;
3058 defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3059 /*anyext*/ Log2_8, L4_ior_memopb_io>;
3060 defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3061 /*sext*/ Log2_8, L4_ior_memopb_io>;
3062 defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3063 /*zext*/ Log2_8, L4_ior_memopb_io>;
3064 defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3065 /*anyext*/ Log2_16, L4_ior_memoph_io>;
3066 defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3067 /*sext*/ Log2_16, L4_ior_memoph_io>;
3068 defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3069 /*zext*/ Log2_16, L4_ior_memoph_io>;
3070 defm: Memopxi_pat<load, store, u6_2ImmPred, or, IsPow2_32,
3071 Log2_32, L4_ior_memopw_io>;
3075 // --(15) Call -----------------------------------------------------------
3078 // Pseudo instructions.
3079 def SDT_SPCallSeqStart
3080 : SDCallSeqStart<[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3081 def SDT_SPCallSeqEnd
3082 : SDCallSeqEnd<[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3084 def callseq_start: SDNode<"ISD::CALLSEQ_START", SDT_SPCallSeqStart,
3085 [SDNPHasChain, SDNPOutGlue]>;
3086 def callseq_end: SDNode<"ISD::CALLSEQ_END", SDT_SPCallSeqEnd,
3087 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
3089 def SDT_SPCall: SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
3091 def HexagonTCRet: SDNode<"HexagonISD::TC_RETURN", SDT_SPCall,
3092 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
3093 def callv3: SDNode<"HexagonISD::CALL", SDT_SPCall,
3094 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
3095 def callv3nr: SDNode<"HexagonISD::CALLnr", SDT_SPCall,
3096 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
3098 def: Pat<(callseq_start timm:$amt, timm:$amt2),
3099 (ADJCALLSTACKDOWN imm:$amt, imm:$amt2)>;
3100 def: Pat<(callseq_end timm:$amt1, timm:$amt2),
3101 (ADJCALLSTACKUP imm:$amt1, imm:$amt2)>;
3103 def: Pat<(HexagonTCRet tglobaladdr:$dst), (PS_tailcall_i tglobaladdr:$dst)>;
3104 def: Pat<(HexagonTCRet texternalsym:$dst), (PS_tailcall_i texternalsym:$dst)>;
3105 def: Pat<(HexagonTCRet I32:$dst), (PS_tailcall_r I32:$dst)>;
3107 def: Pat<(callv3 I32:$dst), (J2_callr I32:$dst)>;
3108 def: Pat<(callv3 tglobaladdr:$dst), (J2_call tglobaladdr:$dst)>;
3109 def: Pat<(callv3 texternalsym:$dst), (J2_call texternalsym:$dst)>;
3110 def: Pat<(callv3 tglobaltlsaddr:$dst), (J2_call tglobaltlsaddr:$dst)>;
3112 def: Pat<(callv3nr I32:$dst), (PS_callr_nr I32:$dst)>;
3113 def: Pat<(callv3nr tglobaladdr:$dst), (PS_call_nr tglobaladdr:$dst)>;
3114 def: Pat<(callv3nr texternalsym:$dst), (PS_call_nr texternalsym:$dst)>;
3116 def retflag : SDNode<"HexagonISD::RET_FLAG", SDTNone,
3117 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
3118 def eh_return: SDNode<"HexagonISD::EH_RETURN", SDTNone, [SDNPHasChain]>;
3120 def: Pat<(retflag), (PS_jmpret (i32 R31))>;
3121 def: Pat<(eh_return), (EH_RETURN_JMPR (i32 R31))>;
3124 // --(16) Branch ---------------------------------------------------------
3127 def: Pat<(br bb:$dst), (J2_jump b30_2Imm:$dst)>;
3128 def: Pat<(brind I32:$dst), (J2_jumpr I32:$dst)>;
3130 def: Pat<(brcond I1:$Pu, bb:$dst),
3131 (J2_jumpt I1:$Pu, bb:$dst)>;
3132 def: Pat<(brcond (not I1:$Pu), bb:$dst),
3133 (J2_jumpf I1:$Pu, bb:$dst)>;
3134 def: Pat<(brcond (i1 (setne I1:$Pu, -1)), bb:$dst),
3135 (J2_jumpf I1:$Pu, bb:$dst)>;
3136 def: Pat<(brcond (i1 (seteq I1:$Pu, 0)), bb:$dst),
3137 (J2_jumpf I1:$Pu, bb:$dst)>;
3138 def: Pat<(brcond (i1 (setne I1:$Pu, 0)), bb:$dst),
3139 (J2_jumpt I1:$Pu, bb:$dst)>;
3142 // --(17) Misc -----------------------------------------------------------
3145 // Generate code of the form 'C2_muxii(cmpbgtui(Rdd, C-1),0,1)'
3146 // for C code of the form r = (c>='0' && c<='9') ? 1 : 0.
3147 // The isdigit transformation relies on two 'clever' aspects:
3148 // 1) The data type is unsigned which allows us to eliminate a zero test after
3149 // biasing the expression by 48. We are depending on the representation of
3150 // the unsigned types, and semantics.
3151 // 2) The front end has converted <= 9 into < 10 on entry to LLVM.
3154 // retval = (c >= '0' && c <= '9') ? 1 : 0;
3155 // The code is transformed upstream of llvm into
3156 // retval = (c-48) < 10 ? 1 : 0;
3158 def u7_0PosImmPred : ImmLeaf<i32, [{
3159 // True if the immediate fits in an 7-bit unsigned field and is positive.
3160 return Imm > 0 && isUInt<7>(Imm);
3163 let AddedComplexity = 139 in
3164 def: Pat<(i32 (zext (i1 (setult (and I32:$Rs, 255), u7_0PosImmPred:$u7)))),
3165 (C2_muxii (A4_cmpbgtui IntRegs:$Rs, (UDEC1 imm:$u7)), 0, 1)>;
3167 let AddedComplexity = 100 in
3168 def: Pat<(or (or (shl (HexagonINSERT (i32 (zextloadi8 (add I32:$b, 2))),
3169 (i32 (extloadi8 (add I32:$b, 3))),
3172 (shl (i32 (zextloadi8 (add I32:$b, 1))), (i32 8))),
3173 (zextloadi8 I32:$b)),
3174 (A2_swiz (L2_loadri_io I32:$b, 0))>;
3177 // We need custom lowering of ISD::PREFETCH into HexagonISD::DCFETCH
3178 // because the SDNode ISD::PREFETCH has properties MayLoad and MayStore.
3179 // We don't really want either one here.
3180 def SDTHexagonDCFETCH: SDTypeProfile<0, 2, [SDTCisPtrTy<0>,SDTCisInt<1>]>;
3181 def HexagonDCFETCH: SDNode<"HexagonISD::DCFETCH", SDTHexagonDCFETCH,
3184 def: Pat<(HexagonDCFETCH IntRegs:$Rs, u11_3ImmPred:$u11_3),
3185 (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
3186 def: Pat<(HexagonDCFETCH (i32 (add IntRegs:$Rs, u11_3ImmPred:$u11_3)), (i32 0)),
3187 (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
3189 def SDTHexagonALLOCA
3190 : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3192 : SDNode<"HexagonISD::ALLOCA", SDTHexagonALLOCA, [SDNPHasChain]>;
3194 def: Pat<(HexagonALLOCA I32:$Rs, (i32 imm:$A)),
3195 (PS_alloca IntRegs:$Rs, imm:$A)>;
3197 def HexagonBARRIER: SDNode<"HexagonISD::BARRIER", SDTNone, [SDNPHasChain]>;
3198 def: Pat<(HexagonBARRIER), (Y2_barrier)>;
3200 def: Pat<(trap), (PS_crash)>;
3202 // Read cycle counter.
3203 def SDTInt64Leaf: SDTypeProfile<1, 0, [SDTCisVT<0, i64>]>;
3204 def HexagonREADCYCLE: SDNode<"HexagonISD::READCYCLE", SDTInt64Leaf,
3207 def: Pat<(HexagonREADCYCLE), (A4_tfrcpp UPCYCLE)>;
3209 // The declared return value of the store-locked intrinsics is i32, but
3210 // the instructions actually define i1. To avoid register copies from
3211 // IntRegs to PredRegs and back, fold the entire pattern checking the
3212 // result against true/false.
3213 let AddedComplexity = 100 in {
3214 def: Pat<(i1 (setne (int_hexagon_S2_storew_locked I32:$Rs, I32:$Rt), 0)),
3215 (S2_storew_locked I32:$Rs, I32:$Rt)>;
3216 def: Pat<(i1 (seteq (int_hexagon_S2_storew_locked I32:$Rs, I32:$Rt), 0)),
3217 (C2_not (S2_storew_locked I32:$Rs, I32:$Rt))>;
3218 def: Pat<(i1 (setne (int_hexagon_S4_stored_locked I32:$Rs, I64:$Rt), 0)),
3219 (S4_stored_locked I32:$Rs, I64:$Rt)>;
3220 def: Pat<(i1 (seteq (int_hexagon_S4_stored_locked I32:$Rs, I64:$Rt), 0)),
3221 (C2_not (S4_stored_locked I32:$Rs, I64:$Rt))>;