xref: /freebsd/contrib/llvm-project/llvm/lib/Target/Hexagon/HexagonPatterns.td (revision 13ec1e3155c7e9bf037b12af186351b7fa9b9450)
1//===- HexagonPatterns.td - Selection Patterns for Hexagon -*- tablegen -*-===//
2//
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
6//
7//===----------------------------------------------------------------------===//
8
9// Table of contents:
10//     (0) Definitions
11//     (1) Immediates
12//     (2) Type casts
13//     (3) Extend/truncate
14//     (4) Logical
15//     (5) Compare
16//     (6) Select
17//     (7) Insert/extract
18//     (8) Shift/permute
19//     (9) Arithmetic/bitwise
20//    (10) Bit
21//    (11) PIC
22//    (12) Load
23//    (13) Store
24//    (14) Memop
25//    (15) Call
26//    (16) Branch
27//    (17) Misc
28
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).
54
55
56// --(0) Definitions -----------------------------------------------------
57//
58
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
61// in the patterns.
62def AddrFI: ComplexPattern<i32, 1, "SelectAddrFI", [frameindex], []>;
63
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.
67def AddrGA: ComplexPattern<i32, 1, "SelectAddrGA", [], []>;
68def AddrGP: ComplexPattern<i32, 1, "SelectAddrGP", [], []>;
69def AnyImm: ComplexPattern<i32, 1, "SelectAnyImm", [], []>;
70def AnyInt: ComplexPattern<i32, 1, "SelectAnyInt", [], []>;
71
72// Global address or a constant being a multiple of 2^n.
73def AnyImm0: ComplexPattern<i32, 1, "SelectAnyImm0", [], []>;
74def AnyImm1: ComplexPattern<i32, 1, "SelectAnyImm1", [], []>;
75def AnyImm2: ComplexPattern<i32, 1, "SelectAnyImm2", [], []>;
76def AnyImm3: ComplexPattern<i32, 1, "SelectAnyImm3", [], []>;
77
78
79// Type helper frags.
80def V2I1:   PatLeaf<(v2i1    PredRegs:$R)>;
81def V4I1:   PatLeaf<(v4i1    PredRegs:$R)>;
82def V8I1:   PatLeaf<(v8i1    PredRegs:$R)>;
83def V4I8:   PatLeaf<(v4i8    IntRegs:$R)>;
84def V2I16:  PatLeaf<(v2i16   IntRegs:$R)>;
85
86def V8I8:   PatLeaf<(v8i8    DoubleRegs:$R)>;
87def V4I16:  PatLeaf<(v4i16   DoubleRegs:$R)>;
88def V2I32:  PatLeaf<(v2i32   DoubleRegs:$R)>;
89
90def HQ8:    PatLeaf<(VecQ8   HvxQR:$R)>;
91def HQ16:   PatLeaf<(VecQ16  HvxQR:$R)>;
92def HQ32:   PatLeaf<(VecQ32  HvxQR:$R)>;
93
94def HVI8:   PatLeaf<(VecI8   HvxVR:$R)>;
95def HVI16:  PatLeaf<(VecI16  HvxVR:$R)>;
96def HVI32:  PatLeaf<(VecI32  HvxVR:$R)>;
97
98def HWI8:   PatLeaf<(VecPI8  HvxWR:$R)>;
99def HWI16:  PatLeaf<(VecPI16 HvxWR:$R)>;
100def HWI32:  PatLeaf<(VecPI32 HvxWR:$R)>;
101
102def SDTVecLeaf:
103  SDTypeProfile<1, 0, [SDTCisVec<0>]>;
104def SDTVecVecIntOp:
105  SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisVec<1>, SDTCisSameAs<1,2>,
106                       SDTCisVT<3,i32>]>;
107
108def HexagonPTRUE:      SDNode<"HexagonISD::PTRUE",      SDTVecLeaf>;
109def HexagonPFALSE:     SDNode<"HexagonISD::PFALSE",     SDTVecLeaf>;
110def HexagonVALIGN:     SDNode<"HexagonISD::VALIGN",     SDTVecVecIntOp>;
111def HexagonVALIGNADDR: SDNode<"HexagonISD::VALIGNADDR", SDTIntUnaryOp>;
112
113def ptrue:  PatFrag<(ops), (HexagonPTRUE)>;
114def pfalse: PatFrag<(ops), (HexagonPFALSE)>;
115def pnot:   PatFrag<(ops node:$Pu), (xor node:$Pu, ptrue)>;
116
117def valign: PatFrag<(ops node:$Vt, node:$Vs, node:$Ru),
118                    (HexagonVALIGN node:$Vt, node:$Vs, node:$Ru)>;
119def valignaddr: PatFrag<(ops node:$Addr), (HexagonVALIGNADDR node:$Addr)>;
120
121// Pattern fragments to extract the low and high subregisters from a
122// 64-bit value.
123def LoReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG (i64 $Rs), isub_lo)>;
124def HiReg: OutPatFrag<(ops node:$Rs), (EXTRACT_SUBREG (i64 $Rs), isub_hi)>;
125
126def IsOrAdd: PatFrag<(ops node:$A, node:$B), (or node:$A, node:$B), [{
127  return isOrEquivalentToAdd(N);
128}]>;
129
130def IsPow2_32: PatLeaf<(i32 imm), [{
131  uint32_t V = N->getZExtValue();
132  return isPowerOf2_32(V);
133}]>;
134
135def IsPow2_64: PatLeaf<(i64 imm), [{
136  uint64_t V = N->getZExtValue();
137  return isPowerOf2_64(V);
138}]>;
139
140def IsNPow2_32: PatLeaf<(i32 imm), [{
141  uint32_t NV = ~N->getZExtValue();
142  return isPowerOf2_32(NV);
143}]>;
144
145def IsPow2_64L: PatLeaf<(i64 imm), [{
146  uint64_t V = N->getZExtValue();
147  return isPowerOf2_64(V) && Log2_64(V) < 32;
148}]>;
149
150def IsPow2_64H: PatLeaf<(i64 imm), [{
151  uint64_t V = N->getZExtValue();
152  return isPowerOf2_64(V) && Log2_64(V) >= 32;
153}]>;
154
155def IsNPow2_64L: PatLeaf<(i64 imm), [{
156  uint64_t NV = ~N->getZExtValue();
157  return isPowerOf2_64(NV) && Log2_64(NV) < 32;
158}]>;
159
160def IsNPow2_64H: PatLeaf<(i64 imm), [{
161  uint64_t NV = ~N->getZExtValue();
162  return isPowerOf2_64(NV) && Log2_64(NV) >= 32;
163}]>;
164
165class IsULE<int Width, int Arg>: PatLeaf<(i32 imm),
166  "uint64_t V = N->getZExtValue();" #
167  "return isUInt<" # Width # ">(V) && V <= " # Arg # ";"
168>;
169
170class IsUGT<int Width, int Arg>: PatLeaf<(i32 imm),
171  "uint64_t V = N->getZExtValue();" #
172  "return isUInt<" # Width # ">(V) && V > " # Arg # ";"
173>;
174
175def SDEC1: SDNodeXForm<imm, [{
176  int32_t V = N->getSExtValue();
177  return CurDAG->getTargetConstant(V-1, SDLoc(N), MVT::i32);
178}]>;
179
180def UDEC1: SDNodeXForm<imm, [{
181  uint32_t V = N->getZExtValue();
182  assert(V >= 1);
183  return CurDAG->getTargetConstant(V-1, SDLoc(N), MVT::i32);
184}]>;
185
186def UDEC32: SDNodeXForm<imm, [{
187  uint32_t V = N->getZExtValue();
188  assert(V >= 32);
189  return CurDAG->getTargetConstant(V-32, SDLoc(N), MVT::i32);
190}]>;
191
192class Subi<int From>: SDNodeXForm<imm,
193  "int32_t V = " # From # " - N->getSExtValue();" #
194  "return CurDAG->getTargetConstant(V, SDLoc(N), MVT::i32);"
195>;
196
197def Log2_32: SDNodeXForm<imm, [{
198  uint32_t V = N->getZExtValue();
199  return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
200}]>;
201
202def Log2_64: SDNodeXForm<imm, [{
203  uint64_t V = N->getZExtValue();
204  return CurDAG->getTargetConstant(Log2_64(V), SDLoc(N), MVT::i32);
205}]>;
206
207def LogN2_32: SDNodeXForm<imm, [{
208  uint32_t NV = ~N->getZExtValue();
209  return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
210}]>;
211
212def LogN2_64: SDNodeXForm<imm, [{
213  uint64_t NV = ~N->getZExtValue();
214  return CurDAG->getTargetConstant(Log2_64(NV), SDLoc(N), MVT::i32);
215}]>;
216
217def NegImm8: SDNodeXForm<imm, [{
218  int8_t NV = -N->getSExtValue();
219  return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
220}]>;
221
222def NegImm16: SDNodeXForm<imm, [{
223  int16_t NV = -N->getSExtValue();
224  return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
225}]>;
226
227def NegImm32: SDNodeXForm<imm, [{
228  int32_t NV = -N->getSExtValue();
229  return CurDAG->getTargetConstant(NV, SDLoc(N), MVT::i32);
230}]>;
231
232def SplatB: SDNodeXForm<imm, [{
233  uint32_t V = N->getZExtValue();
234  assert(isUInt<8>(V) || V >> 8 == 0xFFFFFF);
235  V &= 0xFF;
236  uint32_t S = V << 24 | V << 16 | V << 8 | V;
237  return CurDAG->getTargetConstant(S, SDLoc(N), MVT::i32);
238}]>;
239
240def SplatH: SDNodeXForm<imm, [{
241  uint32_t V = N->getZExtValue();
242  assert(isUInt<16>(V) || V >> 16 == 0xFFFF);
243  V &= 0xFFFF;
244  return CurDAG->getTargetConstant(V << 16 | V, SDLoc(N), MVT::i32);
245}]>;
246
247
248// Helpers for type promotions/contractions.
249def I1toI32:  OutPatFrag<(ops node:$Rs), (C2_muxii (i1 $Rs), 1, 0)>;
250def I32toI1:  OutPatFrag<(ops node:$Rs), (i1 (C2_cmpgtui (i32 $Rs), (i32 0)))>;
251def ToZext64: OutPatFrag<(ops node:$Rs), (i64 (A4_combineir 0, (i32 $Rs)))>;
252def ToSext64: OutPatFrag<(ops node:$Rs), (i64 (A2_sxtw (i32 $Rs)))>;
253def ToAext64: OutPatFrag<(ops node:$Rs),
254  (REG_SEQUENCE DoubleRegs, (i32 (IMPLICIT_DEF)), isub_hi, (i32 $Rs), isub_lo)>;
255
256def Combinew: OutPatFrag<(ops node:$Rs, node:$Rt),
257  (REG_SEQUENCE DoubleRegs, $Rs, isub_hi, $Rt, isub_lo)>;
258
259def addrga: PatLeaf<(i32 AddrGA:$Addr)>;
260def addrgp: PatLeaf<(i32 AddrGP:$Addr)>;
261def anyimm: PatLeaf<(i32 AnyImm:$Imm)>;
262def anyint: PatLeaf<(i32 AnyInt:$Imm)>;
263
264// Global address or an aligned constant.
265def anyimm0: PatLeaf<(i32 AnyImm0:$Addr)>;
266def anyimm1: PatLeaf<(i32 AnyImm1:$Addr)>;
267def anyimm2: PatLeaf<(i32 AnyImm2:$Addr)>;
268def anyimm3: PatLeaf<(i32 AnyImm3:$Addr)>;
269
270def f32ImmPred : PatLeaf<(f32 fpimm:$F)>;
271def f64ImmPred : PatLeaf<(f64 fpimm:$F)>;
272
273// This complex pattern is really only to detect various forms of
274// sign-extension i32->i64. The selected value will be of type i64
275// whose low word is the value being extended. The high word is
276// unspecified.
277def Usxtw:  ComplexPattern<i64, 1, "DetectUseSxtw", [], []>;
278
279def Aext64: PatFrag<(ops node:$Rs), (i64 (anyext node:$Rs))>;
280def Zext64: PatFrag<(ops node:$Rs), (i64 (zext node:$Rs))>;
281def Sext64: PatLeaf<(i64 Usxtw:$Rs)>;
282
283def azext: PatFrags<(ops node:$Rs), [(zext node:$Rs), (anyext node:$Rs)]>;
284def asext: PatFrags<(ops node:$Rs), [(sext node:$Rs), (anyext node:$Rs)]>;
285
286def: Pat<(IsOrAdd (i32 AddrFI:$Rs), s32_0ImmPred:$off),
287         (PS_fi (i32 AddrFI:$Rs), imm:$off)>;
288
289
290// Converters from unary/binary SDNode to PatFrag.
291class pf1<SDNode Op> : PatFrag<(ops node:$a), (Op node:$a)>;
292class pf2<SDNode Op> : PatFrag<(ops node:$a, node:$b), (Op node:$a, node:$b)>;
293
294class Not2<PatFrag P>
295  : PatFrag<(ops node:$A, node:$B), (P node:$A, (not node:$B))>;
296class VNot2<PatFrag P, PatFrag Not>
297  : PatFrag<(ops node:$A, node:$B), (P node:$A, (Not node:$B))>;
298
299// If there is a constant operand that feeds the and/or instruction,
300// do not generate the compound instructions.
301// It is not always profitable, as some times we end up with a transfer.
302// Check the below example.
303// ra = #65820; rb = lsr(rb, #8); rc ^= and (rb, ra)
304// Instead this is preferable.
305// ra = and (#65820, lsr(ra, #8)); rb = xor(rb, ra)
306class Su_ni1<PatFrag Op>
307  : PatFrag<Op.Operands, !head(Op.Fragments), [{
308            if (hasOneUse(N)){
309              // Check if Op1 is an immediate operand.
310              SDValue Op1 = N->getOperand(1);
311              return !isa<ConstantSDNode>(Op1);
312            }
313            return false;}],
314            Op.OperandTransform>;
315
316class Su<PatFrag Op>
317  : PatFrag<Op.Operands, !head(Op.Fragments), [{ return hasOneUse(N); }],
318            Op.OperandTransform>;
319
320// Main selection macros.
321
322class OpR_R_pat<InstHexagon MI, PatFrag Op, ValueType ResVT, PatFrag RegPred>
323  : Pat<(ResVT (Op RegPred:$Rs)), (MI RegPred:$Rs)>;
324
325class OpR_RI_pat<InstHexagon MI, PatFrag Op, ValueType ResType,
326                 PatFrag RegPred, PatFrag ImmPred>
327  : Pat<(ResType (Op RegPred:$Rs, ImmPred:$I)),
328        (MI RegPred:$Rs, imm:$I)>;
329
330class OpR_RR_pat<InstHexagon MI, PatFrag Op, ValueType ResType,
331                 PatFrag RsPred, PatFrag RtPred = RsPred>
332  : Pat<(ResType (Op RsPred:$Rs, RtPred:$Rt)),
333        (MI RsPred:$Rs, RtPred:$Rt)>;
334
335class AccRRI_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op,
336                 PatFrag RegPred, PatFrag ImmPred>
337  : Pat<(AccOp RegPred:$Rx, (Op RegPred:$Rs, ImmPred:$I)),
338        (MI RegPred:$Rx, RegPred:$Rs, imm:$I)>;
339
340class AccRRR_pat<InstHexagon MI, PatFrag AccOp, PatFrag Op,
341                 PatFrag RxPred, PatFrag RsPred, PatFrag RtPred>
342  : Pat<(AccOp RxPred:$Rx, (Op RsPred:$Rs, RtPred:$Rt)),
343        (MI RxPred:$Rx, RsPred:$Rs, RtPred:$Rt)>;
344
345multiclass SelMinMax_pats<PatFrag CmpOp, PatFrag Val,
346                          InstHexagon InstA, InstHexagon InstB> {
347  def: Pat<(select (i1 (CmpOp Val:$A, Val:$B)), Val:$A, Val:$B),
348           (InstA Val:$A, Val:$B)>;
349  def: Pat<(select (i1 (CmpOp Val:$A, Val:$B)), Val:$B, Val:$A),
350           (InstB Val:$A, Val:$B)>;
351}
352
353multiclass MinMax_pats<InstHexagon PickT, InstHexagon PickS,
354                       SDPatternOperator Sel, SDPatternOperator CmpOp,
355                       ValueType CmpType, PatFrag CmpPred> {
356  def: Pat<(Sel (CmpType (CmpOp CmpPred:$Vs, CmpPred:$Vt)),
357                CmpPred:$Vt, CmpPred:$Vs),
358           (PickT CmpPred:$Vs, CmpPred:$Vt)>;
359  def: Pat<(Sel (CmpType (CmpOp CmpPred:$Vs, CmpPred:$Vt)),
360                CmpPred:$Vs, CmpPred:$Vt),
361           (PickS CmpPred:$Vs, CmpPred:$Vt)>;
362}
363
364// Bitcasts between same-size vector types are no-ops, except for the
365// actual type change.
366multiclass NopCast_pat<ValueType Ty1, ValueType Ty2, RegisterClass RC> {
367  def: Pat<(Ty1 (bitconvert (Ty2 RC:$Val))), (Ty1 RC:$Val)>;
368  def: Pat<(Ty2 (bitconvert (Ty1 RC:$Val))), (Ty2 RC:$Val)>;
369}
370
371// Frags for commonly used SDNodes.
372def Add: pf2<add>;    def And: pf2<and>;    def Sra: pf2<sra>;
373def Sub: pf2<sub>;    def Or:  pf2<or>;     def Srl: pf2<srl>;
374def Mul: pf2<mul>;    def Xor: pf2<xor>;    def Shl: pf2<shl>;
375
376def Smin: pf2<smin>;  def Smax: pf2<smax>;
377def Umin: pf2<umin>;  def Umax: pf2<umax>;
378
379def Rol: pf2<rotl>;
380
381// --(1) Immediate -------------------------------------------------------
382//
383
384def Imm64Lo: SDNodeXForm<imm, [{
385  return CurDAG->getTargetConstant(int32_t (N->getSExtValue()),
386                                   SDLoc(N), MVT::i32);
387}]>;
388def Imm64Hi: SDNodeXForm<imm, [{
389  return CurDAG->getTargetConstant(int32_t (N->getSExtValue()>>32),
390                                   SDLoc(N), MVT::i32);
391}]>;
392
393
394def SDTHexagonCONST32
395  : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisPtrTy<0>]>;
396
397def HexagonJT:          SDNode<"HexagonISD::JT",          SDTIntUnaryOp>;
398def HexagonCP:          SDNode<"HexagonISD::CP",          SDTIntUnaryOp>;
399def HexagonCONST32:     SDNode<"HexagonISD::CONST32",     SDTHexagonCONST32>;
400def HexagonCONST32_GP:  SDNode<"HexagonISD::CONST32_GP",  SDTHexagonCONST32>;
401
402def TruncI64ToI32: SDNodeXForm<imm, [{
403  return CurDAG->getTargetConstant(N->getSExtValue(), SDLoc(N), MVT::i32);
404}]>;
405
406def: Pat<(s32_0ImmPred:$s16), (A2_tfrsi imm:$s16)>;
407def: Pat<(s8_0Imm64Pred:$s8), (A2_tfrpi (TruncI64ToI32 $s8))>;
408
409def: Pat<(HexagonCONST32    tglobaltlsaddr:$A), (A2_tfrsi imm:$A)>;
410def: Pat<(HexagonCONST32    bbl:$A),            (A2_tfrsi imm:$A)>;
411def: Pat<(HexagonCONST32    tglobaladdr:$A),    (A2_tfrsi imm:$A)>;
412def: Pat<(HexagonCONST32_GP tblockaddress:$A),  (A2_tfrsi imm:$A)>;
413def: Pat<(HexagonCONST32_GP tglobaladdr:$A),    (A2_tfrsi imm:$A)>;
414def: Pat<(HexagonJT         tjumptable:$A),     (A2_tfrsi imm:$A)>;
415def: Pat<(HexagonCP         tconstpool:$A),     (A2_tfrsi imm:$A)>;
416// The HVX load patterns also match CP directly. Make sure that if
417// the selection of this opcode changes, it's updated in all places.
418
419def: Pat<(i1 0),        (PS_false)>;
420def: Pat<(i1 1),        (PS_true)>;
421def: Pat<(i64 imm:$v),  (CONST64 imm:$v)>,
422     Requires<[UseSmallData,NotOptTinyCore]>;
423def: Pat<(i64 imm:$v),
424         (Combinew (A2_tfrsi (Imm64Hi $v)), (A2_tfrsi (Imm64Lo $v)))>;
425
426def ftoi : SDNodeXForm<fpimm, [{
427  APInt I = N->getValueAPF().bitcastToAPInt();
428  return CurDAG->getTargetConstant(I.getZExtValue(), SDLoc(N),
429                                   MVT::getIntegerVT(I.getBitWidth()));
430}]>;
431
432def: Pat<(f32ImmPred:$f), (A2_tfrsi (ftoi $f))>;
433def: Pat<(f64ImmPred:$f), (CONST64  (ftoi $f))>;
434
435def ToI32: OutPatFrag<(ops node:$V), (A2_tfrsi $V)>;
436
437// --(2) Type cast -------------------------------------------------------
438//
439
440def: OpR_R_pat<F2_conv_sf2df,      pf1<fpextend>,   f64, F32>;
441def: OpR_R_pat<F2_conv_df2sf,      pf1<fpround>,    f32, F64>;
442
443def: OpR_R_pat<F2_conv_w2sf,       pf1<sint_to_fp>, f32, I32>;
444def: OpR_R_pat<F2_conv_d2sf,       pf1<sint_to_fp>, f32, I64>;
445def: OpR_R_pat<F2_conv_w2df,       pf1<sint_to_fp>, f64, I32>;
446def: OpR_R_pat<F2_conv_d2df,       pf1<sint_to_fp>, f64, I64>;
447
448def: OpR_R_pat<F2_conv_uw2sf,      pf1<uint_to_fp>, f32, I32>;
449def: OpR_R_pat<F2_conv_ud2sf,      pf1<uint_to_fp>, f32, I64>;
450def: OpR_R_pat<F2_conv_uw2df,      pf1<uint_to_fp>, f64, I32>;
451def: OpR_R_pat<F2_conv_ud2df,      pf1<uint_to_fp>, f64, I64>;
452
453def: OpR_R_pat<F2_conv_sf2w_chop,  pf1<fp_to_sint>, i32, F32>;
454def: OpR_R_pat<F2_conv_df2w_chop,  pf1<fp_to_sint>, i32, F64>;
455def: OpR_R_pat<F2_conv_sf2d_chop,  pf1<fp_to_sint>, i64, F32>;
456def: OpR_R_pat<F2_conv_df2d_chop,  pf1<fp_to_sint>, i64, F64>;
457
458def: OpR_R_pat<F2_conv_sf2uw_chop, pf1<fp_to_uint>, i32, F32>;
459def: OpR_R_pat<F2_conv_df2uw_chop, pf1<fp_to_uint>, i32, F64>;
460def: OpR_R_pat<F2_conv_sf2ud_chop, pf1<fp_to_uint>, i64, F32>;
461def: OpR_R_pat<F2_conv_df2ud_chop, pf1<fp_to_uint>, i64, F64>;
462
463// Bitcast is different than [fp|sint|uint]_to_[sint|uint|fp].
464def: Pat<(i32 (bitconvert F32:$v)), (I32:$v)>;
465def: Pat<(f32 (bitconvert I32:$v)), (F32:$v)>;
466def: Pat<(i64 (bitconvert F64:$v)), (I64:$v)>;
467def: Pat<(f64 (bitconvert I64:$v)), (F64:$v)>;
468
469// Bit convert 32- and 64-bit types.
470// All of these are bitcastable to one another: i32, v2i16, v4i8.
471defm: NopCast_pat<i32,   v2i16, IntRegs>;
472defm: NopCast_pat<i32,    v4i8, IntRegs>;
473defm: NopCast_pat<v2i16,  v4i8, IntRegs>;
474// All of these are bitcastable to one another: i64, v2i32, v4i16, v8i8.
475defm: NopCast_pat<i64,   v2i32, DoubleRegs>;
476defm: NopCast_pat<i64,   v4i16, DoubleRegs>;
477defm: NopCast_pat<i64,    v8i8, DoubleRegs>;
478defm: NopCast_pat<v2i32, v4i16, DoubleRegs>;
479defm: NopCast_pat<v2i32,  v8i8, DoubleRegs>;
480defm: NopCast_pat<v4i16,  v8i8, DoubleRegs>;
481
482
483// --(3) Extend/truncate -------------------------------------------------
484//
485
486def: Pat<(sext_inreg I32:$Rs, i8),  (A2_sxtb I32:$Rs)>;
487def: Pat<(sext_inreg I32:$Rs, i16), (A2_sxth I32:$Rs)>;
488def: Pat<(sext_inreg I64:$Rs, i32), (A2_sxtw (LoReg $Rs))>;
489def: Pat<(sext_inreg I64:$Rs, i16), (A2_sxtw (A2_sxth (LoReg $Rs)))>;
490def: Pat<(sext_inreg I64:$Rs, i8),  (A2_sxtw (A2_sxtb (LoReg $Rs)))>;
491
492def: Pat<(i64 (sext I32:$Rs)), (A2_sxtw I32:$Rs)>;
493def: Pat<(Zext64 I32:$Rs),     (ToZext64 $Rs)>;
494def: Pat<(Aext64 I32:$Rs),     (ToZext64 $Rs)>;
495
496def: Pat<(i32 (trunc I64:$Rs)), (LoReg $Rs)>;
497def: Pat<(i1 (trunc I32:$Rs)),  (S2_tstbit_i I32:$Rs, 0)>;
498def: Pat<(i1 (trunc I64:$Rs)),  (S2_tstbit_i (LoReg $Rs), 0)>;
499
500let AddedComplexity = 20 in {
501  def: Pat<(and I32:$Rs, 255),   (A2_zxtb I32:$Rs)>;
502  def: Pat<(and I32:$Rs, 65535), (A2_zxth I32:$Rs)>;
503}
504
505// Extensions from i1 or vectors of i1.
506def: Pat<(i32 (azext I1:$Pu)), (C2_muxii I1:$Pu, 1, 0)>;
507def: Pat<(i64 (azext I1:$Pu)), (ToZext64 (C2_muxii I1:$Pu, 1, 0))>;
508def: Pat<(i32  (sext I1:$Pu)), (C2_muxii I1:$Pu, -1, 0)>;
509def: Pat<(i64  (sext I1:$Pu)), (Combinew (C2_muxii PredRegs:$Pu, -1, 0),
510                                         (C2_muxii PredRegs:$Pu, -1, 0))>;
511
512def: Pat<(v2i16 (sext V2I1:$Pu)), (S2_vtrunehb (C2_mask V2I1:$Pu))>;
513def: Pat<(v2i32 (sext V2I1:$Pu)), (C2_mask V2I1:$Pu)>;
514def: Pat<(v4i8  (sext V4I1:$Pu)), (S2_vtrunehb (C2_mask V4I1:$Pu))>;
515def: Pat<(v4i16 (sext V4I1:$Pu)), (C2_mask V4I1:$Pu)>;
516def: Pat<(v8i8  (sext V8I1:$Pu)), (C2_mask V8I1:$Pu)>;
517
518def Vsplatpi: OutPatFrag<(ops node:$V),
519                         (Combinew (A2_tfrsi $V), (A2_tfrsi $V))>;
520
521def: Pat<(v2i16 (azext V2I1:$Pu)),
522         (A2_andir (LoReg (C2_mask V2I1:$Pu)), (i32 0x00010001))>;
523def: Pat<(v2i32 (azext V2I1:$Pu)),
524         (A2_andp (C2_mask V2I1:$Pu), (A2_combineii (i32 1), (i32 1)))>;
525def: Pat<(v4i8 (azext V4I1:$Pu)),
526         (A2_andir (LoReg (C2_mask V4I1:$Pu)), (i32 0x01010101))>;
527def: Pat<(v4i16 (azext V4I1:$Pu)),
528         (A2_andp (C2_mask V4I1:$Pu), (Vsplatpi (i32 0x00010001)))>;
529def: Pat<(v8i8 (azext V8I1:$Pu)),
530         (A2_andp (C2_mask V8I1:$Pu), (Vsplatpi (i32 0x01010101)))>;
531
532def: Pat<(v4i16 (azext  V4I8:$Rs)),  (S2_vzxtbh V4I8:$Rs)>;
533def: Pat<(v2i32 (azext  V2I16:$Rs)), (S2_vzxthw V2I16:$Rs)>;
534def: Pat<(v4i16 (sext   V4I8:$Rs)),  (S2_vsxtbh V4I8:$Rs)>;
535def: Pat<(v2i32 (sext   V2I16:$Rs)), (S2_vsxthw V2I16:$Rs)>;
536
537def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i8)),
538         (Combinew (A2_sxtb (HiReg $Rs)), (A2_sxtb (LoReg $Rs)))>;
539
540def: Pat<(v2i32 (sext_inreg V2I32:$Rs, v2i16)),
541         (Combinew (A2_sxth (HiReg $Rs)), (A2_sxth (LoReg $Rs)))>;
542
543// Truncate: from vector B copy all 'E'ven 'B'yte elements:
544// A[0] = B[0];  A[1] = B[2];  A[2] = B[4];  A[3] = B[6];
545def: Pat<(v4i8 (trunc V4I16:$Rs)),
546         (S2_vtrunehb V4I16:$Rs)>;
547
548// Truncate: from vector B copy all 'O'dd 'B'yte elements:
549// A[0] = B[1];  A[1] = B[3];  A[2] = B[5];  A[3] = B[7];
550// S2_vtrunohb
551
552// Truncate: from vectors B and C copy all 'E'ven 'H'alf-word elements:
553// A[0] = B[0];  A[1] = B[2];  A[2] = C[0];  A[3] = C[2];
554// S2_vtruneh
555
556def: Pat<(v2i16 (trunc V2I32:$Rs)),
557         (A2_combine_ll (HiReg $Rs), (LoReg $Rs))>;
558
559
560// --(4) Logical ---------------------------------------------------------
561//
562
563def: Pat<(not I1:$Ps),      (C2_not I1:$Ps)>;
564def: Pat<(pnot V2I1:$Ps),   (C2_not V2I1:$Ps)>;
565def: Pat<(pnot V4I1:$Ps),   (C2_not V4I1:$Ps)>;
566def: Pat<(pnot V8I1:$Ps),   (C2_not V8I1:$Ps)>;
567def: Pat<(add I1:$Ps, -1),  (C2_not I1:$Ps)>;
568
569def: OpR_RR_pat<C2_and,         And, i1, I1>;
570def: OpR_RR_pat<C2_or,           Or, i1, I1>;
571def: OpR_RR_pat<C2_xor,         Xor, i1, I1>;
572def: OpR_RR_pat<C2_andn,  Not2<And>, i1, I1>;
573def: OpR_RR_pat<C2_orn,    Not2<Or>, i1, I1>;
574
575def: AccRRR_pat<C4_and_and,   And,       Su<And>, I1, I1, I1>;
576def: AccRRR_pat<C4_and_or,    And,       Su< Or>, I1, I1, I1>;
577def: AccRRR_pat<C4_or_and,     Or,       Su<And>, I1, I1, I1>;
578def: AccRRR_pat<C4_or_or,      Or,       Su< Or>, I1, I1, I1>;
579def: AccRRR_pat<C4_and_andn,  And, Su<Not2<And>>, I1, I1, I1>;
580def: AccRRR_pat<C4_and_orn,   And, Su<Not2< Or>>, I1, I1, I1>;
581def: AccRRR_pat<C4_or_andn,    Or, Su<Not2<And>>, I1, I1, I1>;
582def: AccRRR_pat<C4_or_orn,     Or, Su<Not2< Or>>, I1, I1, I1>;
583
584multiclass BoolvOpR_RR_pat<InstHexagon MI, PatFrag VOp> {
585  def: OpR_RR_pat<MI, VOp, v2i1, V2I1>;
586  def: OpR_RR_pat<MI, VOp, v4i1, V4I1>;
587  def: OpR_RR_pat<MI, VOp, v8i1, V8I1>;
588}
589
590multiclass BoolvAccRRR_pat<InstHexagon MI, PatFrag AccOp, PatFrag VOp> {
591  def: AccRRR_pat<MI, AccOp, VOp, V2I1, V2I1, V2I1>;
592  def: AccRRR_pat<MI, AccOp, VOp, V4I1, V4I1, V4I1>;
593  def: AccRRR_pat<MI, AccOp, VOp, V8I1, V8I1, V8I1>;
594}
595
596defm: BoolvOpR_RR_pat<C2_and,                    And>;
597defm: BoolvOpR_RR_pat<C2_or,                      Or>;
598defm: BoolvOpR_RR_pat<C2_xor,                    Xor>;
599defm: BoolvOpR_RR_pat<C2_andn,      VNot2<And, pnot>>;
600defm: BoolvOpR_RR_pat<C2_orn,       VNot2< Or, pnot>>;
601
602// op(Ps, op(Pt, Pu))
603defm: BoolvAccRRR_pat<C4_and_and,   And, Su<And>>;
604defm: BoolvAccRRR_pat<C4_and_or,    And, Su<Or>>;
605defm: BoolvAccRRR_pat<C4_or_and,    Or,  Su<And>>;
606defm: BoolvAccRRR_pat<C4_or_or,     Or,  Su<Or>>;
607
608// op(Ps, op(Pt, !Pu))
609defm: BoolvAccRRR_pat<C4_and_andn,  And, Su<VNot2<And, pnot>>>;
610defm: BoolvAccRRR_pat<C4_and_orn,   And, Su<VNot2< Or, pnot>>>;
611defm: BoolvAccRRR_pat<C4_or_andn,   Or,  Su<VNot2<And, pnot>>>;
612defm: BoolvAccRRR_pat<C4_or_orn,    Or,  Su<VNot2< Or, pnot>>>;
613
614
615// --(5) Compare ---------------------------------------------------------
616//
617
618// Avoid negated comparisons, i.e. those of form "Pd = !cmp(...)".
619// These cannot form compounds (e.g. J4_cmpeqi_tp0_jump_nt).
620
621def: OpR_RI_pat<C2_cmpeqi,    seteq,          i1, I32,  anyimm>;
622def: OpR_RI_pat<C2_cmpgti,    setgt,          i1, I32,  anyimm>;
623def: OpR_RI_pat<C2_cmpgtui,   setugt,         i1, I32,  anyimm>;
624
625def: Pat<(i1 (setge I32:$Rs, s32_0ImmPred:$s10)),
626         (C2_cmpgti I32:$Rs, (SDEC1 imm:$s10))>;
627def: Pat<(i1 (setuge I32:$Rs, u32_0ImmPred:$u9)),
628         (C2_cmpgtui I32:$Rs, (UDEC1 imm:$u9))>;
629
630def: Pat<(i1 (setlt I32:$Rs, s32_0ImmPred:$s10)),
631         (C2_not (C2_cmpgti I32:$Rs, (SDEC1 imm:$s10)))>;
632def: Pat<(i1 (setult I32:$Rs, u32_0ImmPred:$u9)),
633         (C2_not (C2_cmpgtui I32:$Rs, (UDEC1 imm:$u9)))>;
634
635// Patfrag to convert the usual comparison patfrags (e.g. setlt) to ones
636// that reverse the order of the operands.
637class RevCmp<PatFrag F>
638  : PatFrag<(ops node:$rhs, node:$lhs), !head(F.Fragments), F.PredicateCode,
639            F.OperandTransform>;
640
641def: OpR_RR_pat<C2_cmpeq,     seteq,          i1,   I32>;
642def: OpR_RR_pat<C2_cmpgt,     setgt,          i1,   I32>;
643def: OpR_RR_pat<C2_cmpgtu,    setugt,         i1,   I32>;
644def: OpR_RR_pat<C2_cmpgt,     RevCmp<setlt>,  i1,   I32>;
645def: OpR_RR_pat<C2_cmpgtu,    RevCmp<setult>, i1,   I32>;
646def: OpR_RR_pat<C2_cmpeqp,    seteq,          i1,   I64>;
647def: OpR_RR_pat<C2_cmpgtp,    setgt,          i1,   I64>;
648def: OpR_RR_pat<C2_cmpgtup,   setugt,         i1,   I64>;
649def: OpR_RR_pat<C2_cmpgtp,    RevCmp<setlt>,  i1,   I64>;
650def: OpR_RR_pat<C2_cmpgtup,   RevCmp<setult>, i1,   I64>;
651def: OpR_RR_pat<A2_vcmpbeq,   seteq,          i1,   V8I8>;
652def: OpR_RR_pat<A2_vcmpbeq,   seteq,          v8i1, V8I8>;
653def: OpR_RR_pat<A4_vcmpbgt,   RevCmp<setlt>,  i1,   V8I8>;
654def: OpR_RR_pat<A4_vcmpbgt,   RevCmp<setlt>,  v8i1, V8I8>;
655def: OpR_RR_pat<A4_vcmpbgt,   setgt,          i1,   V8I8>;
656def: OpR_RR_pat<A4_vcmpbgt,   setgt,          v8i1, V8I8>;
657def: OpR_RR_pat<A2_vcmpbgtu,  RevCmp<setult>, i1,   V8I8>;
658def: OpR_RR_pat<A2_vcmpbgtu,  RevCmp<setult>, v8i1, V8I8>;
659def: OpR_RR_pat<A2_vcmpbgtu,  setugt,         i1,   V8I8>;
660def: OpR_RR_pat<A2_vcmpbgtu,  setugt,         v8i1, V8I8>;
661def: OpR_RR_pat<A2_vcmpheq,   seteq,          i1,   V4I16>;
662def: OpR_RR_pat<A2_vcmpheq,   seteq,          v4i1, V4I16>;
663def: OpR_RR_pat<A2_vcmphgt,   RevCmp<setlt>,  i1,   V4I16>;
664def: OpR_RR_pat<A2_vcmphgt,   RevCmp<setlt>,  v4i1, V4I16>;
665def: OpR_RR_pat<A2_vcmphgt,   setgt,          i1,   V4I16>;
666def: OpR_RR_pat<A2_vcmphgt,   setgt,          v4i1, V4I16>;
667def: OpR_RR_pat<A2_vcmphgtu,  RevCmp<setult>, i1,   V4I16>;
668def: OpR_RR_pat<A2_vcmphgtu,  RevCmp<setult>, v4i1, V4I16>;
669def: OpR_RR_pat<A2_vcmphgtu,  setugt,         i1,   V4I16>;
670def: OpR_RR_pat<A2_vcmphgtu,  setugt,         v4i1, V4I16>;
671def: OpR_RR_pat<A2_vcmpweq,   seteq,          i1,   V2I32>;
672def: OpR_RR_pat<A2_vcmpweq,   seteq,          v2i1, V2I32>;
673def: OpR_RR_pat<A2_vcmpwgt,   RevCmp<setlt>,  i1,   V2I32>;
674def: OpR_RR_pat<A2_vcmpwgt,   RevCmp<setlt>,  v2i1, V2I32>;
675def: OpR_RR_pat<A2_vcmpwgt,   setgt,          i1,   V2I32>;
676def: OpR_RR_pat<A2_vcmpwgt,   setgt,          v2i1, V2I32>;
677def: OpR_RR_pat<A2_vcmpwgtu,  RevCmp<setult>, i1,   V2I32>;
678def: OpR_RR_pat<A2_vcmpwgtu,  RevCmp<setult>, v2i1, V2I32>;
679def: OpR_RR_pat<A2_vcmpwgtu,  setugt,         i1,   V2I32>;
680def: OpR_RR_pat<A2_vcmpwgtu,  setugt,         v2i1, V2I32>;
681
682def: OpR_RR_pat<F2_sfcmpeq,   seteq,          i1, F32>;
683def: OpR_RR_pat<F2_sfcmpgt,   setgt,          i1, F32>;
684def: OpR_RR_pat<F2_sfcmpge,   setge,          i1, F32>;
685def: OpR_RR_pat<F2_sfcmpeq,   setoeq,         i1, F32>;
686def: OpR_RR_pat<F2_sfcmpgt,   setogt,         i1, F32>;
687def: OpR_RR_pat<F2_sfcmpge,   setoge,         i1, F32>;
688def: OpR_RR_pat<F2_sfcmpgt,   RevCmp<setolt>, i1, F32>;
689def: OpR_RR_pat<F2_sfcmpge,   RevCmp<setole>, i1, F32>;
690def: OpR_RR_pat<F2_sfcmpgt,   RevCmp<setlt>,  i1, F32>;
691def: OpR_RR_pat<F2_sfcmpge,   RevCmp<setle>,  i1, F32>;
692def: OpR_RR_pat<F2_sfcmpuo,   setuo,          i1, F32>;
693
694def: OpR_RR_pat<F2_dfcmpeq,   seteq,          i1, F64>;
695def: OpR_RR_pat<F2_dfcmpgt,   setgt,          i1, F64>;
696def: OpR_RR_pat<F2_dfcmpge,   setge,          i1, F64>;
697def: OpR_RR_pat<F2_dfcmpeq,   setoeq,         i1, F64>;
698def: OpR_RR_pat<F2_dfcmpgt,   setogt,         i1, F64>;
699def: OpR_RR_pat<F2_dfcmpge,   setoge,         i1, F64>;
700def: OpR_RR_pat<F2_dfcmpgt,   RevCmp<setolt>, i1, F64>;
701def: OpR_RR_pat<F2_dfcmpge,   RevCmp<setole>, i1, F64>;
702def: OpR_RR_pat<F2_dfcmpgt,   RevCmp<setlt>,  i1, F64>;
703def: OpR_RR_pat<F2_dfcmpge,   RevCmp<setle>,  i1, F64>;
704def: OpR_RR_pat<F2_dfcmpuo,   setuo,          i1, F64>;
705
706// Avoid C4_cmpneqi, C4_cmpltei, C4_cmplteui, since they cannot form compounds.
707
708def: Pat<(i1 (setne I32:$Rs, anyimm:$u5)),
709         (C2_not (C2_cmpeqi I32:$Rs, imm:$u5))>;
710def: Pat<(i1 (setle I32:$Rs, anyimm:$u5)),
711         (C2_not (C2_cmpgti I32:$Rs, imm:$u5))>;
712def: Pat<(i1 (setule I32:$Rs, anyimm:$u5)),
713         (C2_not (C2_cmpgtui I32:$Rs, imm:$u5))>;
714
715class OpmR_RR_pat<PatFrag Output, PatFrag Op, ValueType ResType,
716                  PatFrag RsPred, PatFrag RtPred = RsPred>
717  : Pat<(ResType (Op RsPred:$Rs, RtPred:$Rt)),
718        (Output RsPred:$Rs, RtPred:$Rt)>;
719
720class Outn<InstHexagon MI>
721  : OutPatFrag<(ops node:$Rs, node:$Rt),
722               (C2_not (MI $Rs, $Rt))>;
723
724def: OpmR_RR_pat<Outn<C2_cmpeq>,    setne,          i1,   I32>;
725def: OpmR_RR_pat<Outn<C2_cmpgt>,    setle,          i1,   I32>;
726def: OpmR_RR_pat<Outn<C2_cmpgtu>,   setule,         i1,   I32>;
727def: OpmR_RR_pat<Outn<C2_cmpgt>,    RevCmp<setge>,  i1,   I32>;
728def: OpmR_RR_pat<Outn<C2_cmpgtu>,   RevCmp<setuge>, i1,   I32>;
729def: OpmR_RR_pat<Outn<C2_cmpeqp>,   setne,          i1,   I64>;
730def: OpmR_RR_pat<Outn<C2_cmpgtp>,   setle,          i1,   I64>;
731def: OpmR_RR_pat<Outn<C2_cmpgtup>,  setule,         i1,   I64>;
732def: OpmR_RR_pat<Outn<C2_cmpgtp>,   RevCmp<setge>,  i1,   I64>;
733def: OpmR_RR_pat<Outn<C2_cmpgtup>,  RevCmp<setuge>, i1,   I64>;
734def: OpmR_RR_pat<Outn<A2_vcmpbeq>,  setne,          v8i1, V8I8>;
735def: OpmR_RR_pat<Outn<A4_vcmpbgt>,  setle,          v8i1, V8I8>;
736def: OpmR_RR_pat<Outn<A2_vcmpbgtu>, setule,         v8i1, V8I8>;
737def: OpmR_RR_pat<Outn<A4_vcmpbgt>,  RevCmp<setge>,  v8i1, V8I8>;
738def: OpmR_RR_pat<Outn<A2_vcmpbgtu>, RevCmp<setuge>, v8i1, V8I8>;
739def: OpmR_RR_pat<Outn<A2_vcmpheq>,  setne,          v4i1, V4I16>;
740def: OpmR_RR_pat<Outn<A2_vcmphgt>,  setle,          v4i1, V4I16>;
741def: OpmR_RR_pat<Outn<A2_vcmphgtu>, setule,         v4i1, V4I16>;
742def: OpmR_RR_pat<Outn<A2_vcmphgt>,  RevCmp<setge>,  v4i1, V4I16>;
743def: OpmR_RR_pat<Outn<A2_vcmphgtu>, RevCmp<setuge>, v4i1, V4I16>;
744def: OpmR_RR_pat<Outn<A2_vcmpweq>,  setne,          v2i1, V2I32>;
745def: OpmR_RR_pat<Outn<A2_vcmpwgt>,  setle,          v2i1, V2I32>;
746def: OpmR_RR_pat<Outn<A2_vcmpwgtu>, setule,         v2i1, V2I32>;
747def: OpmR_RR_pat<Outn<A2_vcmpwgt>,  RevCmp<setge>,  v2i1, V2I32>;
748def: OpmR_RR_pat<Outn<A2_vcmpwgtu>, RevCmp<setuge>, v2i1, V2I32>;
749
750let AddedComplexity = 100 in {
751  def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt), 255), 0)),
752           (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt)>;
753  def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt), 255), 0)),
754           (C2_not (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt))>;
755  def: Pat<(i1 (seteq (and (xor I32:$Rs, I32:$Rt), 65535), 0)),
756           (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt)>;
757  def: Pat<(i1 (setne (and (xor I32:$Rs, I32:$Rt), 65535), 0)),
758           (C2_not (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt))>;
759}
760
761// PatFrag for AsserZext which takes the original type as a parameter.
762def SDTAssertZext: SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisSameAs<0,1>]>;
763def AssertZextSD: SDNode<"ISD::AssertZext", SDTAssertZext>;
764class AssertZext<ValueType T>: PatFrag<(ops node:$A), (AssertZextSD $A, T)>;
765
766multiclass Cmpb_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
767                      PatLeaf ImmPred, int Mask> {
768  def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
769           (MI I32:$Rs, imm:$I)>;
770  def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
771           (MI I32:$Rs, imm:$I)>;
772}
773
774multiclass CmpbN_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
775                     PatLeaf ImmPred, int Mask> {
776  def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
777           (C2_not (MI I32:$Rs, imm:$I))>;
778  def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
779           (C2_not (MI I32:$Rs, imm:$I))>;
780}
781
782multiclass CmpbND_pat<InstHexagon MI, PatFrag Op, PatFrag AssertExt,
783                      PatLeaf ImmPred, int Mask> {
784  def: Pat<(i1 (Op (and I32:$Rs, Mask), ImmPred:$I)),
785           (C2_not (MI I32:$Rs, (UDEC1 imm:$I)))>;
786  def: Pat<(i1 (Op (AssertExt I32:$Rs), ImmPred:$I)),
787           (C2_not (MI I32:$Rs, (UDEC1 imm:$I)))>;
788}
789
790let AddedComplexity = 200 in {
791  defm: Cmpb_pat  <A4_cmpbeqi,  seteq,  AssertZext<i8>,  IsUGT<8,31>,  255>;
792  defm: CmpbN_pat <A4_cmpbeqi,  setne,  AssertZext<i8>,  IsUGT<8,31>,  255>;
793  defm: Cmpb_pat  <A4_cmpbgtui, setugt, AssertZext<i8>,  IsUGT<32,31>, 255>;
794  defm: CmpbN_pat <A4_cmpbgtui, setule, AssertZext<i8>,  IsUGT<32,31>, 255>;
795  defm: Cmpb_pat  <A4_cmphgtui, setugt, AssertZext<i16>, IsUGT<32,31>, 65535>;
796  defm: CmpbN_pat <A4_cmphgtui, setule, AssertZext<i16>, IsUGT<32,31>, 65535>;
797  defm: CmpbND_pat<A4_cmpbgtui, setult, AssertZext<i8>,  IsUGT<32,32>, 255>;
798  defm: CmpbND_pat<A4_cmphgtui, setult, AssertZext<i16>, IsUGT<32,32>, 65535>;
799}
800
801def: Pat<(i32 (zext (i1 (seteq I32:$Rs, I32:$Rt)))),
802         (A4_rcmpeq I32:$Rs, I32:$Rt)>;
803def: Pat<(i32 (zext (i1 (setne I32:$Rs, I32:$Rt)))),
804         (A4_rcmpneq I32:$Rs, I32:$Rt)>;
805def: Pat<(i32 (zext (i1 (seteq I32:$Rs, anyimm:$s8)))),
806         (A4_rcmpeqi I32:$Rs, imm:$s8)>;
807def: Pat<(i32 (zext (i1 (setne I32:$Rs, anyimm:$s8)))),
808         (A4_rcmpneqi I32:$Rs, imm:$s8)>;
809
810def: Pat<(i1 (seteq I1:$Ps, (i1 -1))), (I1:$Ps)>;
811def: Pat<(i1 (setne I1:$Ps, (i1 -1))), (C2_not I1:$Ps)>;
812def: Pat<(i1 (seteq I1:$Ps, I1:$Pt)),  (C2_xor I1:$Ps, (C2_not I1:$Pt))>;
813def: Pat<(i1 (setne I1:$Ps, I1:$Pt)),  (C2_xor I1:$Ps, I1:$Pt)>;
814
815// Floating-point comparisons with checks for ordered/unordered status.
816
817class T3<InstHexagon MI1, InstHexagon MI2, InstHexagon MI3>
818  : OutPatFrag<(ops node:$Rs, node:$Rt),
819               (MI1 (MI2 $Rs, $Rt), (MI3 $Rs, $Rt))>;
820
821class Cmpuf<InstHexagon MI>:  T3<C2_or,  F2_sfcmpuo, MI>;
822class Cmpud<InstHexagon MI>:  T3<C2_or,  F2_dfcmpuo, MI>;
823
824class Cmpufn<InstHexagon MI>: T3<C2_orn, F2_sfcmpuo, MI>;
825class Cmpudn<InstHexagon MI>: T3<C2_orn, F2_dfcmpuo, MI>;
826
827def: OpmR_RR_pat<Cmpuf<F2_sfcmpeq>,  setueq,         i1, F32>;
828def: OpmR_RR_pat<Cmpuf<F2_sfcmpge>,  setuge,         i1, F32>;
829def: OpmR_RR_pat<Cmpuf<F2_sfcmpgt>,  setugt,         i1, F32>;
830def: OpmR_RR_pat<Cmpuf<F2_sfcmpge>,  RevCmp<setule>, i1, F32>;
831def: OpmR_RR_pat<Cmpuf<F2_sfcmpgt>,  RevCmp<setult>, i1, F32>;
832def: OpmR_RR_pat<Cmpufn<F2_sfcmpeq>, setune,         i1, F32>;
833
834def: OpmR_RR_pat<Cmpud<F2_dfcmpeq>,  setueq,         i1, F64>;
835def: OpmR_RR_pat<Cmpud<F2_dfcmpge>,  setuge,         i1, F64>;
836def: OpmR_RR_pat<Cmpud<F2_dfcmpgt>,  setugt,         i1, F64>;
837def: OpmR_RR_pat<Cmpud<F2_dfcmpge>,  RevCmp<setule>, i1, F64>;
838def: OpmR_RR_pat<Cmpud<F2_dfcmpgt>,  RevCmp<setult>, i1, F64>;
839def: OpmR_RR_pat<Cmpudn<F2_dfcmpeq>, setune,         i1, F64>;
840
841def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setone, i1, F32>;
842def: OpmR_RR_pat<Outn<F2_sfcmpeq>, setne,  i1, F32>;
843
844def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setone, i1, F64>;
845def: OpmR_RR_pat<Outn<F2_dfcmpeq>, setne,  i1, F64>;
846
847def: OpmR_RR_pat<Outn<F2_sfcmpuo>, seto,   i1, F32>;
848def: OpmR_RR_pat<Outn<F2_dfcmpuo>, seto,   i1, F64>;
849
850
851// --(6) Select ----------------------------------------------------------
852//
853
854def: Pat<(select I1:$Pu, I32:$Rs, I32:$Rt),
855         (C2_mux I1:$Pu, I32:$Rs, I32:$Rt)>;
856def: Pat<(select I1:$Pu, anyimm:$s8, I32:$Rs),
857         (C2_muxri I1:$Pu, imm:$s8, I32:$Rs)>;
858def: Pat<(select I1:$Pu, I32:$Rs, anyimm:$s8),
859         (C2_muxir I1:$Pu, I32:$Rs, imm:$s8)>;
860def: Pat<(select I1:$Pu, anyimm:$s8, s8_0ImmPred:$S8),
861         (C2_muxii I1:$Pu, imm:$s8, imm:$S8)>;
862
863def: Pat<(select (not I1:$Pu), I32:$Rs, I32:$Rt),
864         (C2_mux I1:$Pu, I32:$Rt, I32:$Rs)>;
865def: Pat<(select (not I1:$Pu), s8_0ImmPred:$S8, anyimm:$s8),
866         (C2_muxii I1:$Pu, imm:$s8, imm:$S8)>;
867def: Pat<(select (not I1:$Pu), anyimm:$s8, I32:$Rs),
868         (C2_muxir I1:$Pu, I32:$Rs, imm:$s8)>;
869def: Pat<(select (not I1:$Pu), I32:$Rs, anyimm:$s8),
870         (C2_muxri I1:$Pu, imm:$s8, I32:$Rs)>;
871
872// Map from a 64-bit select to an emulated 64-bit mux.
873// Hexagon does not support 64-bit MUXes; so emulate with combines.
874def: Pat<(select I1:$Pu, I64:$Rs, I64:$Rt),
875         (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
876                   (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
877
878def: Pat<(select I1:$Pu, F32:$Rs, f32ImmPred:$I),
879         (C2_muxir I1:$Pu, F32:$Rs, (ftoi $I))>;
880def: Pat<(select I1:$Pu, f32ImmPred:$I, F32:$Rt),
881         (C2_muxri I1:$Pu, (ftoi $I), F32:$Rt)>;
882def: Pat<(select I1:$Pu, F32:$Rs, F32:$Rt),
883         (C2_mux I1:$Pu, F32:$Rs, F32:$Rt)>;
884def: Pat<(select I1:$Pu, F64:$Rs, F64:$Rt),
885         (Combinew (C2_mux I1:$Pu, (HiReg $Rs), (HiReg $Rt)),
886                   (C2_mux I1:$Pu, (LoReg $Rs), (LoReg $Rt)))>;
887
888def: Pat<(select (i1 (setult F32:$Ra, F32:$Rb)), F32:$Rs, F32:$Rt),
889         (C2_mux (F2_sfcmpgt F32:$Rb, F32:$Ra), F32:$Rs, F32:$Rt)>;
890def: Pat<(select (i1 (setult F64:$Ra, F64:$Rb)), F64:$Rs, F64:$Rt),
891         (C2_vmux (F2_dfcmpgt F64:$Rb, F64:$Ra), F64:$Rs, F64:$Rt)>;
892
893def: Pat<(select (not I1:$Pu), f32ImmPred:$I, F32:$Rs),
894         (C2_muxir I1:$Pu, F32:$Rs, (ftoi $I))>;
895def: Pat<(select (not I1:$Pu), F32:$Rt, f32ImmPred:$I),
896         (C2_muxri I1:$Pu, (ftoi $I), F32:$Rt)>;
897
898def: Pat<(vselect V8I1:$Pu, V8I8:$Rs, V8I8:$Rt),
899         (C2_vmux V8I1:$Pu, V8I8:$Rs, V8I8:$Rt)>;
900def: Pat<(vselect V4I1:$Pu, V4I16:$Rs, V4I16:$Rt),
901         (C2_vmux V4I1:$Pu, V4I16:$Rs, V4I16:$Rt)>;
902def: Pat<(vselect V2I1:$Pu, V2I32:$Rs, V2I32:$Rt),
903         (C2_vmux V2I1:$Pu, V2I32:$Rs, V2I32:$Rt)>;
904
905def: Pat<(vselect (pnot V8I1:$Pu), V8I8:$Rs, V8I8:$Rt),
906         (C2_vmux V8I1:$Pu, V8I8:$Rt, V8I8:$Rs)>;
907def: Pat<(vselect (pnot V4I1:$Pu), V4I16:$Rs, V4I16:$Rt),
908         (C2_vmux V4I1:$Pu, V4I16:$Rt, V4I16:$Rs)>;
909def: Pat<(vselect (pnot V2I1:$Pu), V2I32:$Rs, V2I32:$Rt),
910         (C2_vmux V2I1:$Pu, V2I32:$Rt, V2I32:$Rs)>;
911
912
913// From LegalizeDAG.cpp: (Pu ? Pv : Pw) <=> (Pu & Pv) | (!Pu & Pw).
914def: Pat<(select I1:$Pu, I1:$Pv, I1:$Pw),
915         (C2_or (C2_and  I1:$Pu, I1:$Pv),
916                (C2_andn I1:$Pw, I1:$Pu))>;
917
918
919def IsPosHalf : PatLeaf<(i32 IntRegs:$a), [{
920  return isPositiveHalfWord(N);
921}]>;
922
923multiclass SelMinMax16_pats<PatFrag CmpOp, InstHexagon InstA,
924                            InstHexagon InstB> {
925  def: Pat<(sext_inreg (select (i1 (CmpOp IsPosHalf:$Rs, IsPosHalf:$Rt)),
926                               IsPosHalf:$Rs, IsPosHalf:$Rt), i16),
927           (InstA IntRegs:$Rs, IntRegs:$Rt)>;
928  def: Pat<(sext_inreg (select (i1 (CmpOp IsPosHalf:$Rs, IsPosHalf:$Rt)),
929                               IsPosHalf:$Rt, IsPosHalf:$Rs), i16),
930           (InstB IntRegs:$Rs, IntRegs:$Rt)>;
931}
932
933let AddedComplexity = 200 in {
934  defm: SelMinMax16_pats<setge,  A2_max,  A2_min>;
935  defm: SelMinMax16_pats<setgt,  A2_max,  A2_min>;
936  defm: SelMinMax16_pats<setle,  A2_min,  A2_max>;
937  defm: SelMinMax16_pats<setlt,  A2_min,  A2_max>;
938  defm: SelMinMax16_pats<setuge, A2_maxu, A2_minu>;
939  defm: SelMinMax16_pats<setugt, A2_maxu, A2_minu>;
940  defm: SelMinMax16_pats<setule, A2_minu, A2_maxu>;
941  defm: SelMinMax16_pats<setult, A2_minu, A2_maxu>;
942}
943
944def: OpR_RR_pat<A2_min,   Smin, i32, I32, I32>;
945def: OpR_RR_pat<A2_max,   Smax, i32, I32, I32>;
946def: OpR_RR_pat<A2_minu,  Umin, i32, I32, I32>;
947def: OpR_RR_pat<A2_maxu,  Umax, i32, I32, I32>;
948def: OpR_RR_pat<A2_minp,  Smin, i64, I64, I64>;
949def: OpR_RR_pat<A2_maxp,  Smax, i64, I64, I64>;
950def: OpR_RR_pat<A2_minup, Umin, i64, I64, I64>;
951def: OpR_RR_pat<A2_maxup, Umax, i64, I64, I64>;
952
953let AddedComplexity = 100 in {
954  defm: MinMax_pats<F2_sfmin, F2_sfmax, select, setogt, i1, F32>;
955  defm: MinMax_pats<F2_sfmin, F2_sfmax, select, setoge, i1, F32>;
956  defm: MinMax_pats<F2_sfmax, F2_sfmin, select, setolt, i1, F32>;
957  defm: MinMax_pats<F2_sfmax, F2_sfmin, select, setole, i1, F32>;
958}
959
960let AddedComplexity = 100, Predicates = [HasV67] in {
961  defm: MinMax_pats<F2_dfmin, F2_dfmax, select, setogt, i1, F64>;
962  defm: MinMax_pats<F2_dfmin, F2_dfmax, select, setoge, i1, F64>;
963  defm: MinMax_pats<F2_dfmax, F2_dfmin, select, setolt, i1, F64>;
964  defm: MinMax_pats<F2_dfmax, F2_dfmin, select, setole, i1, F64>;
965}
966
967def: OpR_RR_pat<A2_vminb,  Smin, v8i8,  V8I8>;
968def: OpR_RR_pat<A2_vmaxb,  Smax, v8i8,  V8I8>;
969def: OpR_RR_pat<A2_vminub, Umin, v8i8,  V8I8>;
970def: OpR_RR_pat<A2_vmaxub, Umax, v8i8,  V8I8>;
971
972def: OpR_RR_pat<A2_vminh,  Smin, v4i16, V4I16>;
973def: OpR_RR_pat<A2_vmaxh,  Smax, v4i16, V4I16>;
974def: OpR_RR_pat<A2_vminuh, Umin, v4i16, V4I16>;
975def: OpR_RR_pat<A2_vmaxuh, Umax, v4i16, V4I16>;
976
977def: OpR_RR_pat<A2_vminw,  Smin, v2i32, V2I32>;
978def: OpR_RR_pat<A2_vmaxw,  Smax, v2i32, V2I32>;
979def: OpR_RR_pat<A2_vminuw, Umin, v2i32, V2I32>;
980def: OpR_RR_pat<A2_vmaxuw, Umax, v2i32, V2I32>;
981
982// --(7) Insert/extract --------------------------------------------------
983//
984
985def SDTHexagonINSERT:
986  SDTypeProfile<1, 4, [SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>,
987                       SDTCisInt<0>, SDTCisVT<3, i32>, SDTCisVT<4, i32>]>;
988def HexagonINSERT:    SDNode<"HexagonISD::INSERT",   SDTHexagonINSERT>;
989
990let AddedComplexity = 10 in {
991  def: Pat<(HexagonINSERT I32:$Rs, I32:$Rt, u5_0ImmPred:$u1, u5_0ImmPred:$u2),
992           (S2_insert I32:$Rs, I32:$Rt, imm:$u1, imm:$u2)>;
993  def: Pat<(HexagonINSERT I64:$Rs, I64:$Rt, u6_0ImmPred:$u1, u6_0ImmPred:$u2),
994           (S2_insertp I64:$Rs, I64:$Rt, imm:$u1, imm:$u2)>;
995}
996def: Pat<(HexagonINSERT I32:$Rs, I32:$Rt, I32:$Width, I32:$Off),
997         (S2_insert_rp I32:$Rs, I32:$Rt, (Combinew $Width, $Off))>;
998def: Pat<(HexagonINSERT I64:$Rs, I64:$Rt, I32:$Width, I32:$Off),
999         (S2_insertp_rp I64:$Rs, I64:$Rt, (Combinew $Width, $Off))>;
1000
1001def SDTHexagonEXTRACTU
1002  : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisInt<0>, SDTCisInt<1>,
1003                  SDTCisVT<2, i32>, SDTCisVT<3, i32>]>;
1004def HexagonEXTRACTU:   SDNode<"HexagonISD::EXTRACTU",   SDTHexagonEXTRACTU>;
1005
1006let AddedComplexity = 10 in {
1007  def: Pat<(HexagonEXTRACTU I32:$Rs, u5_0ImmPred:$u5, u5_0ImmPred:$U5),
1008           (S2_extractu I32:$Rs, imm:$u5, imm:$U5)>;
1009  def: Pat<(HexagonEXTRACTU I64:$Rs, u6_0ImmPred:$u6, u6_0ImmPred:$U6),
1010           (S2_extractup I64:$Rs, imm:$u6, imm:$U6)>;
1011}
1012def: Pat<(HexagonEXTRACTU I32:$Rs, I32:$Width, I32:$Off),
1013         (S2_extractu_rp I32:$Rs, (Combinew $Width, $Off))>;
1014def: Pat<(HexagonEXTRACTU I64:$Rs, I32:$Width, I32:$Off),
1015         (S2_extractup_rp I64:$Rs, (Combinew $Width, $Off))>;
1016
1017def: Pat<(v4i8  (splat_vector anyint:$V)), (ToI32 (SplatB $V))>;
1018def: Pat<(v2i16 (splat_vector anyint:$V)), (ToI32 (SplatH $V))>;
1019def: Pat<(v8i8  (splat_vector anyint:$V)),
1020          (Combinew (ToI32 (SplatB $V)), (ToI32 (SplatB $V)))>;
1021def: Pat<(v4i16 (splat_vector anyint:$V)),
1022          (Combinew (ToI32 (SplatH $V)), (ToI32 (SplatH $V)))>;
1023let AddedComplexity = 10 in
1024def: Pat<(v2i32 (splat_vector s8_0ImmPred:$s8)),
1025         (A2_combineii imm:$s8, imm:$s8)>;
1026def: Pat<(v2i32 (splat_vector anyimm:$V)), (Combinew (ToI32 $V), (ToI32 $V))>;
1027
1028def: Pat<(v4i8  (splat_vector I32:$Rs)), (S2_vsplatrb I32:$Rs)>;
1029def: Pat<(v2i16 (splat_vector I32:$Rs)), (LoReg (S2_vsplatrh I32:$Rs))>;
1030def: Pat<(v4i16 (splat_vector I32:$Rs)), (S2_vsplatrh I32:$Rs)>;
1031def: Pat<(v2i32 (splat_vector I32:$Rs)), (Combinew I32:$Rs, I32:$Rs)>;
1032
1033let AddedComplexity = 10 in
1034def: Pat<(v8i8 (splat_vector I32:$Rs)), (S6_vsplatrbp I32:$Rs)>,
1035     Requires<[HasV62]>;
1036def: Pat<(v8i8 (splat_vector I32:$Rs)),
1037         (Combinew (S2_vsplatrb I32:$Rs), (S2_vsplatrb I32:$Rs))>;
1038
1039
1040// --(8) Shift/permute ---------------------------------------------------
1041//
1042
1043def SDTHexagonI64I32I32: SDTypeProfile<1, 2,
1044  [SDTCisVT<0, i64>, SDTCisVT<1, i32>, SDTCisSameAs<1, 2>]>;
1045
1046def HexagonCOMBINE:  SDNode<"HexagonISD::COMBINE",  SDTHexagonI64I32I32>;
1047
1048def: Pat<(HexagonCOMBINE I32:$Rs, I32:$Rt), (Combinew $Rs, $Rt)>;
1049
1050// The complexity of the combines involving immediates should be greater
1051// than the complexity of the combine with two registers.
1052let AddedComplexity = 50 in {
1053  def: Pat<(HexagonCOMBINE I32:$Rs, anyimm:$s8),
1054           (A4_combineri IntRegs:$Rs, imm:$s8)>;
1055  def: Pat<(HexagonCOMBINE anyimm:$s8, I32:$Rs),
1056           (A4_combineir imm:$s8, IntRegs:$Rs)>;
1057}
1058
1059// The complexity of the combine with two immediates should be greater than
1060// the complexity of a combine involving a register.
1061let AddedComplexity = 75 in {
1062  def: Pat<(HexagonCOMBINE s8_0ImmPred:$s8, anyimm:$u6),
1063           (A4_combineii imm:$s8, imm:$u6)>;
1064  def: Pat<(HexagonCOMBINE anyimm:$s8, s8_0ImmPred:$S8),
1065           (A2_combineii imm:$s8, imm:$S8)>;
1066}
1067
1068def: Pat<(bswap I32:$Rs),  (A2_swiz I32:$Rs)>;
1069def: Pat<(bswap I64:$Rss), (Combinew (A2_swiz (LoReg $Rss)),
1070                                     (A2_swiz (HiReg $Rss)))>;
1071
1072def: Pat<(shl s6_0ImmPred:$s6, I32:$Rt),  (S4_lsli imm:$s6, I32:$Rt)>;
1073def: Pat<(shl I32:$Rs, (i32 16)),         (A2_aslh I32:$Rs)>;
1074def: Pat<(sra I32:$Rs, (i32 16)),         (A2_asrh I32:$Rs)>;
1075
1076def: OpR_RI_pat<S2_asr_i_r,  Sra, i32,   I32,   u5_0ImmPred>;
1077def: OpR_RI_pat<S2_lsr_i_r,  Srl, i32,   I32,   u5_0ImmPred>;
1078def: OpR_RI_pat<S2_asl_i_r,  Shl, i32,   I32,   u5_0ImmPred>;
1079def: OpR_RI_pat<S2_asr_i_p,  Sra, i64,   I64,   u6_0ImmPred>;
1080def: OpR_RI_pat<S2_lsr_i_p,  Srl, i64,   I64,   u6_0ImmPred>;
1081def: OpR_RI_pat<S2_asl_i_p,  Shl, i64,   I64,   u6_0ImmPred>;
1082def: OpR_RI_pat<S2_asr_i_vh, Sra, v4i16, V4I16, u4_0ImmPred>;
1083def: OpR_RI_pat<S2_lsr_i_vh, Srl, v4i16, V4I16, u4_0ImmPred>;
1084def: OpR_RI_pat<S2_asl_i_vh, Shl, v4i16, V4I16, u4_0ImmPred>;
1085def: OpR_RI_pat<S2_asr_i_vh, Sra, v2i32, V2I32, u5_0ImmPred>;
1086def: OpR_RI_pat<S2_lsr_i_vh, Srl, v2i32, V2I32, u5_0ImmPred>;
1087def: OpR_RI_pat<S2_asl_i_vh, Shl, v2i32, V2I32, u5_0ImmPred>;
1088
1089def: OpR_RR_pat<S2_asr_r_r, Sra, i32, I32, I32>;
1090def: OpR_RR_pat<S2_lsr_r_r, Srl, i32, I32, I32>;
1091def: OpR_RR_pat<S2_asl_r_r, Shl, i32, I32, I32>;
1092def: OpR_RR_pat<S2_asr_r_p, Sra, i64, I64, I32>;
1093def: OpR_RR_pat<S2_lsr_r_p, Srl, i64, I64, I32>;
1094def: OpR_RR_pat<S2_asl_r_p, Shl, i64, I64, I32>;
1095
1096// Funnel shifts.
1097def IsMul8_U3: PatLeaf<(i32 imm), [{
1098  uint64_t V = N->getZExtValue();
1099  return V % 8 == 0 && isUInt<3>(V / 8);
1100}]>;
1101
1102def Divu8: SDNodeXForm<imm, [{
1103  return CurDAG->getTargetConstant(N->getZExtValue() / 8, SDLoc(N), MVT::i32);
1104}]>;
1105
1106// Funnel shift-left.
1107def FShl32i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1108  (HiReg (S2_asl_i_p (Combinew $Rs, $Rt), $S))>;
1109def FShl32r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1110  (HiReg (S2_asl_r_p (Combinew $Rs, $Rt), $Ru))>;
1111
1112def FShl64i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1113  (S2_lsr_i_p_or (S2_asl_i_p $Rs, $S),  $Rt, (Subi<64> $S))>;
1114def FShl64r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1115  (S2_lsr_r_p_or (S2_asl_r_p $Rs, $Ru), $Rt, (A2_subri 64, $Ru))>;
1116
1117// Combined SDNodeXForm: (Divu8 (Subi<64> $S))
1118def Divu64_8: SDNodeXForm<imm, [{
1119  return CurDAG->getTargetConstant((64 - N->getSExtValue()) / 8,
1120                                   SDLoc(N), MVT::i32);
1121}]>;
1122
1123// Special cases:
1124let AddedComplexity = 100 in {
1125  def: Pat<(fshl I32:$Rs, I32:$Rt, (i32 16)),
1126           (A2_combine_lh I32:$Rs, I32:$Rt)>;
1127  def: Pat<(fshl I64:$Rs, I64:$Rt, IsMul8_U3:$S),
1128           (S2_valignib I64:$Rs, I64:$Rt, (Divu64_8 $S))>;
1129}
1130
1131let Predicates = [HasV60], AddedComplexity = 50 in {
1132  def: OpR_RI_pat<S6_rol_i_r, Rol, i32, I32, u5_0ImmPred>;
1133  def: OpR_RI_pat<S6_rol_i_p, Rol, i64, I64, u6_0ImmPred>;
1134}
1135let AddedComplexity = 30 in {
1136  def: Pat<(rotl I32:$Rs, u5_0ImmPred:$S),          (FShl32i $Rs, $Rs, imm:$S)>;
1137  def: Pat<(rotl I64:$Rs, u6_0ImmPred:$S),          (FShl64i $Rs, $Rs, imm:$S)>;
1138  def: Pat<(fshl I32:$Rs, I32:$Rt, u5_0ImmPred:$S), (FShl32i $Rs, $Rt, imm:$S)>;
1139  def: Pat<(fshl I64:$Rs, I64:$Rt, u6_0ImmPred:$S), (FShl64i $Rs, $Rt, imm:$S)>;
1140}
1141def: Pat<(rotl I32:$Rs, I32:$Rt),           (FShl32r $Rs, $Rs, $Rt)>;
1142def: Pat<(rotl I64:$Rs, I32:$Rt),           (FShl64r $Rs, $Rs, $Rt)>;
1143def: Pat<(fshl I32:$Rs, I32:$Rt, I32:$Ru),  (FShl32r $Rs, $Rt, $Ru)>;
1144def: Pat<(fshl I64:$Rs, I64:$Rt, I32:$Ru),  (FShl64r $Rs, $Rt, $Ru)>;
1145
1146// Funnel shift-right.
1147def FShr32i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1148  (LoReg (S2_lsr_i_p (Combinew $Rs, $Rt), $S))>;
1149def FShr32r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1150  (LoReg (S2_lsr_r_p (Combinew $Rs, $Rt), $Ru))>;
1151
1152def FShr64i: OutPatFrag<(ops node:$Rs, node:$Rt, node:$S),
1153  (S2_asl_i_p_or (S2_lsr_i_p $Rt, $S),  $Rs, (Subi<64> $S))>;
1154def FShr64r: OutPatFrag<(ops node:$Rs, node:$Rt, node:$Ru),
1155  (S2_asl_r_p_or (S2_lsr_r_p $Rt, $Ru), $Rs, (A2_subri 64, $Ru))>;
1156
1157// Special cases:
1158let AddedComplexity = 100 in {
1159  def: Pat<(fshr I32:$Rs, I32:$Rt, (i32 16)),
1160           (A2_combine_lh I32:$Rs, I32:$Rt)>;
1161  def: Pat<(fshr I64:$Rs, I64:$Rt, IsMul8_U3:$S),
1162           (S2_valignib I64:$Rs, I64:$Rt, (Divu8 $S))>;
1163}
1164
1165let Predicates = [HasV60], AddedComplexity = 50 in {
1166  def: Pat<(rotr I32:$Rs, u5_0ImmPred:$S), (S6_rol_i_r I32:$Rs, (Subi<32> $S))>;
1167  def: Pat<(rotr I64:$Rs, u6_0ImmPred:$S), (S6_rol_i_p I64:$Rs, (Subi<64> $S))>;
1168}
1169let AddedComplexity = 30 in {
1170  def: Pat<(rotr I32:$Rs, u5_0ImmPred:$S),          (FShr32i $Rs, $Rs, imm:$S)>;
1171  def: Pat<(rotr I64:$Rs, u6_0ImmPred:$S),          (FShr64i $Rs, $Rs, imm:$S)>;
1172  def: Pat<(fshr I32:$Rs, I32:$Rt, u5_0ImmPred:$S), (FShr32i $Rs, $Rt, imm:$S)>;
1173  def: Pat<(fshr I64:$Rs, I64:$Rt, u6_0ImmPred:$S), (FShr64i $Rs, $Rt, imm:$S)>;
1174}
1175def: Pat<(rotr I32:$Rs, I32:$Rt),           (FShr32r $Rs, $Rs, $Rt)>;
1176def: Pat<(rotr I64:$Rs, I32:$Rt),           (FShr64r $Rs, $Rs, $Rt)>;
1177def: Pat<(fshr I32:$Rs, I32:$Rt, I32:$Ru),  (FShr32r $Rs, $Rt, $Ru)>;
1178def: Pat<(fshr I64:$Rs, I64:$Rt, I32:$Ru),  (FShr64r $Rs, $Rt, $Ru)>;
1179
1180
1181def: Pat<(sra (add (sra I32:$Rs, u5_0ImmPred:$u5), 1), (i32 1)),
1182         (S2_asr_i_r_rnd I32:$Rs, imm:$u5)>;
1183def: Pat<(sra (add (sra I64:$Rs, u6_0ImmPred:$u6), 1), (i32 1)),
1184         (S2_asr_i_p_rnd I64:$Rs, imm:$u6)>;
1185
1186// Prefer S2_addasl_rrri over S2_asl_i_r_acc.
1187let AddedComplexity = 120 in
1188def: Pat<(add I32:$Rt, (shl I32:$Rs, u3_0ImmPred:$u3)),
1189         (S2_addasl_rrri IntRegs:$Rt, IntRegs:$Rs, imm:$u3)>;
1190
1191let AddedComplexity = 100 in {
1192  def: AccRRI_pat<S2_asr_i_r_acc,   Add, Su<Sra>, I32, u5_0ImmPred>;
1193  def: AccRRI_pat<S2_asr_i_r_nac,   Sub, Su<Sra>, I32, u5_0ImmPred>;
1194  def: AccRRI_pat<S2_asr_i_r_and,   And, Su<Sra>, I32, u5_0ImmPred>;
1195  def: AccRRI_pat<S2_asr_i_r_or,    Or,  Su<Sra>, I32, u5_0ImmPred>;
1196
1197  def: AccRRI_pat<S2_asr_i_p_acc,   Add, Su<Sra>, I64, u6_0ImmPred>;
1198  def: AccRRI_pat<S2_asr_i_p_nac,   Sub, Su<Sra>, I64, u6_0ImmPred>;
1199  def: AccRRI_pat<S2_asr_i_p_and,   And, Su<Sra>, I64, u6_0ImmPred>;
1200  def: AccRRI_pat<S2_asr_i_p_or,    Or,  Su<Sra>, I64, u6_0ImmPred>;
1201
1202  def: AccRRI_pat<S2_lsr_i_r_acc,   Add, Su<Srl>, I32, u5_0ImmPred>;
1203  def: AccRRI_pat<S2_lsr_i_r_nac,   Sub, Su<Srl>, I32, u5_0ImmPred>;
1204  def: AccRRI_pat<S2_lsr_i_r_and,   And, Su<Srl>, I32, u5_0ImmPred>;
1205  def: AccRRI_pat<S2_lsr_i_r_or,    Or,  Su<Srl>, I32, u5_0ImmPred>;
1206  def: AccRRI_pat<S2_lsr_i_r_xacc,  Xor, Su<Srl>, I32, u5_0ImmPred>;
1207
1208  def: AccRRI_pat<S2_lsr_i_p_acc,   Add, Su<Srl>, I64, u6_0ImmPred>;
1209  def: AccRRI_pat<S2_lsr_i_p_nac,   Sub, Su<Srl>, I64, u6_0ImmPred>;
1210  def: AccRRI_pat<S2_lsr_i_p_and,   And, Su<Srl>, I64, u6_0ImmPred>;
1211  def: AccRRI_pat<S2_lsr_i_p_or,    Or,  Su<Srl>, I64, u6_0ImmPred>;
1212  def: AccRRI_pat<S2_lsr_i_p_xacc,  Xor, Su<Srl>, I64, u6_0ImmPred>;
1213
1214  def: AccRRI_pat<S2_asl_i_r_acc,   Add, Su<Shl>, I32, u5_0ImmPred>;
1215  def: AccRRI_pat<S2_asl_i_r_nac,   Sub, Su<Shl>, I32, u5_0ImmPred>;
1216  def: AccRRI_pat<S2_asl_i_r_and,   And, Su<Shl>, I32, u5_0ImmPred>;
1217  def: AccRRI_pat<S2_asl_i_r_or,    Or,  Su<Shl>, I32, u5_0ImmPred>;
1218  def: AccRRI_pat<S2_asl_i_r_xacc,  Xor, Su<Shl>, I32, u5_0ImmPred>;
1219
1220  def: AccRRI_pat<S2_asl_i_p_acc,   Add, Su<Shl>, I64, u6_0ImmPred>;
1221  def: AccRRI_pat<S2_asl_i_p_nac,   Sub, Su<Shl>, I64, u6_0ImmPred>;
1222  def: AccRRI_pat<S2_asl_i_p_and,   And, Su<Shl>, I64, u6_0ImmPred>;
1223  def: AccRRI_pat<S2_asl_i_p_or,    Or,  Su<Shl>, I64, u6_0ImmPred>;
1224  def: AccRRI_pat<S2_asl_i_p_xacc,  Xor, Su<Shl>, I64, u6_0ImmPred>;
1225
1226  let Predicates = [HasV60] in {
1227    def: AccRRI_pat<S6_rol_i_r_acc,   Add, Su<Rol>, I32, u5_0ImmPred>;
1228    def: AccRRI_pat<S6_rol_i_r_nac,   Sub, Su<Rol>, I32, u5_0ImmPred>;
1229    def: AccRRI_pat<S6_rol_i_r_and,   And, Su<Rol>, I32, u5_0ImmPred>;
1230    def: AccRRI_pat<S6_rol_i_r_or,    Or,  Su<Rol>, I32, u5_0ImmPred>;
1231    def: AccRRI_pat<S6_rol_i_r_xacc,  Xor, Su<Rol>, I32, u5_0ImmPred>;
1232
1233    def: AccRRI_pat<S6_rol_i_p_acc,   Add, Su<Rol>, I64, u6_0ImmPred>;
1234    def: AccRRI_pat<S6_rol_i_p_nac,   Sub, Su<Rol>, I64, u6_0ImmPred>;
1235    def: AccRRI_pat<S6_rol_i_p_and,   And, Su<Rol>, I64, u6_0ImmPred>;
1236    def: AccRRI_pat<S6_rol_i_p_or,    Or,  Su<Rol>, I64, u6_0ImmPred>;
1237    def: AccRRI_pat<S6_rol_i_p_xacc,  Xor, Su<Rol>, I64, u6_0ImmPred>;
1238  }
1239}
1240
1241let AddedComplexity = 100 in {
1242  def: AccRRR_pat<S2_asr_r_r_acc,   Add, Su<Sra>, I32, I32, I32>;
1243  def: AccRRR_pat<S2_asr_r_r_nac,   Sub, Su<Sra>, I32, I32, I32>;
1244  def: AccRRR_pat<S2_asr_r_r_and,   And, Su<Sra>, I32, I32, I32>;
1245  def: AccRRR_pat<S2_asr_r_r_or,    Or,  Su<Sra>, I32, I32, I32>;
1246
1247  def: AccRRR_pat<S2_asr_r_p_acc,   Add, Su<Sra>, I64, I64, I32>;
1248  def: AccRRR_pat<S2_asr_r_p_nac,   Sub, Su<Sra>, I64, I64, I32>;
1249  def: AccRRR_pat<S2_asr_r_p_and,   And, Su<Sra>, I64, I64, I32>;
1250  def: AccRRR_pat<S2_asr_r_p_or,    Or,  Su<Sra>, I64, I64, I32>;
1251  def: AccRRR_pat<S2_asr_r_p_xor,   Xor, Su<Sra>, I64, I64, I32>;
1252
1253  def: AccRRR_pat<S2_lsr_r_r_acc,   Add, Su<Srl>, I32, I32, I32>;
1254  def: AccRRR_pat<S2_lsr_r_r_nac,   Sub, Su<Srl>, I32, I32, I32>;
1255  def: AccRRR_pat<S2_lsr_r_r_and,   And, Su<Srl>, I32, I32, I32>;
1256  def: AccRRR_pat<S2_lsr_r_r_or,    Or,  Su<Srl>, I32, I32, I32>;
1257
1258  def: AccRRR_pat<S2_lsr_r_p_acc,   Add, Su<Srl>, I64, I64, I32>;
1259  def: AccRRR_pat<S2_lsr_r_p_nac,   Sub, Su<Srl>, I64, I64, I32>;
1260  def: AccRRR_pat<S2_lsr_r_p_and,   And, Su<Srl>, I64, I64, I32>;
1261  def: AccRRR_pat<S2_lsr_r_p_or,    Or,  Su<Srl>, I64, I64, I32>;
1262  def: AccRRR_pat<S2_lsr_r_p_xor,   Xor, Su<Srl>, I64, I64, I32>;
1263
1264  def: AccRRR_pat<S2_asl_r_r_acc,   Add, Su<Shl>, I32, I32, I32>;
1265  def: AccRRR_pat<S2_asl_r_r_nac,   Sub, Su<Shl>, I32, I32, I32>;
1266  def: AccRRR_pat<S2_asl_r_r_and,   And, Su<Shl>, I32, I32, I32>;
1267  def: AccRRR_pat<S2_asl_r_r_or,    Or,  Su<Shl>, I32, I32, I32>;
1268
1269  def: AccRRR_pat<S2_asl_r_p_acc,   Add, Su<Shl>, I64, I64, I32>;
1270  def: AccRRR_pat<S2_asl_r_p_nac,   Sub, Su<Shl>, I64, I64, I32>;
1271  def: AccRRR_pat<S2_asl_r_p_and,   And, Su<Shl>, I64, I64, I32>;
1272  def: AccRRR_pat<S2_asl_r_p_or,    Or,  Su<Shl>, I64, I64, I32>;
1273  def: AccRRR_pat<S2_asl_r_p_xor,   Xor, Su<Shl>, I64, I64, I32>;
1274}
1275
1276
1277class OpshIRI_pat<InstHexagon MI, PatFrag Op, PatFrag ShOp,
1278                  PatFrag RegPred, PatFrag ImmPred>
1279  : Pat<(Op anyimm:$u8, (ShOp RegPred:$Rs, ImmPred:$U5)),
1280        (MI anyimm:$u8, RegPred:$Rs, imm:$U5)>;
1281
1282let AddedComplexity = 200, Predicates = [UseCompound] in {
1283  def: OpshIRI_pat<S4_addi_asl_ri,  Add, Su<Shl>, I32, u5_0ImmPred>;
1284  def: OpshIRI_pat<S4_addi_lsr_ri,  Add, Su<Srl>, I32, u5_0ImmPred>;
1285  def: OpshIRI_pat<S4_subi_asl_ri,  Sub, Su<Shl>, I32, u5_0ImmPred>;
1286  def: OpshIRI_pat<S4_subi_lsr_ri,  Sub, Su<Srl>, I32, u5_0ImmPred>;
1287  def: OpshIRI_pat<S4_andi_asl_ri,  And, Su<Shl>, I32, u5_0ImmPred>;
1288  def: OpshIRI_pat<S4_andi_lsr_ri,  And, Su<Srl>, I32, u5_0ImmPred>;
1289  def: OpshIRI_pat<S4_ori_asl_ri,   Or,  Su<Shl>, I32, u5_0ImmPred>;
1290  def: OpshIRI_pat<S4_ori_lsr_ri,   Or,  Su<Srl>, I32, u5_0ImmPred>;
1291}
1292
1293// Prefer this pattern to S2_asl_i_p_or for the special case of joining
1294// two 32-bit words into a 64-bit word.
1295let AddedComplexity = 200 in
1296def: Pat<(or (shl (Aext64 I32:$a), (i32 32)), (Zext64 I32:$b)),
1297         (Combinew I32:$a, I32:$b)>;
1298
1299def: Pat<(or (or (or (shl (Zext64 (and I32:$b, (i32 65535))), (i32 16)),
1300                     (Zext64 (and I32:$a, (i32 65535)))),
1301                 (shl (Aext64 (and I32:$c, (i32 65535))), (i32 32))),
1302             (shl (Aext64 I32:$d), (i32 48))),
1303         (Combinew (A2_combine_ll I32:$d, I32:$c),
1304                   (A2_combine_ll I32:$b, I32:$a))>;
1305
1306let AddedComplexity = 200 in {
1307  def: Pat<(or (shl I32:$Rt, (i32 16)), (and I32:$Rs, (i32 65535))),
1308           (A2_combine_ll I32:$Rt, I32:$Rs)>;
1309  def: Pat<(or (shl I32:$Rt, (i32 16)), (srl I32:$Rs, (i32 16))),
1310           (A2_combine_lh I32:$Rt, I32:$Rs)>;
1311  def: Pat<(or (and I32:$Rt, (i32 268431360)), (and I32:$Rs, (i32 65535))),
1312           (A2_combine_hl I32:$Rt, I32:$Rs)>;
1313  def: Pat<(or (and I32:$Rt, (i32 268431360)), (srl I32:$Rs, (i32 16))),
1314           (A2_combine_hh I32:$Rt, I32:$Rs)>;
1315}
1316
1317def SDTHexagonVShift
1318  : SDTypeProfile<1, 2, [SDTCisSameAs<0, 1>, SDTCisVec<0>, SDTCisVT<2, i32>]>;
1319
1320def HexagonVASL: SDNode<"HexagonISD::VASL", SDTHexagonVShift>;
1321def HexagonVASR: SDNode<"HexagonISD::VASR", SDTHexagonVShift>;
1322def HexagonVLSR: SDNode<"HexagonISD::VLSR", SDTHexagonVShift>;
1323
1324def: OpR_RI_pat<S2_asl_i_vw, pf2<HexagonVASL>, v2i32, V2I32, u5_0ImmPred>;
1325def: OpR_RI_pat<S2_asl_i_vh, pf2<HexagonVASL>, v4i16, V4I16, u4_0ImmPred>;
1326def: OpR_RI_pat<S2_asr_i_vw, pf2<HexagonVASR>, v2i32, V2I32, u5_0ImmPred>;
1327def: OpR_RI_pat<S2_asr_i_vh, pf2<HexagonVASR>, v4i16, V4I16, u4_0ImmPred>;
1328def: OpR_RI_pat<S2_lsr_i_vw, pf2<HexagonVLSR>, v2i32, V2I32, u5_0ImmPred>;
1329def: OpR_RI_pat<S2_lsr_i_vh, pf2<HexagonVLSR>, v4i16, V4I16, u4_0ImmPred>;
1330
1331def: OpR_RR_pat<S2_asl_r_vw, pf2<HexagonVASL>, v2i32, V2I32, I32>;
1332def: OpR_RR_pat<S2_asl_r_vh, pf2<HexagonVASL>, v4i16, V4I16, I32>;
1333def: OpR_RR_pat<S2_asr_r_vw, pf2<HexagonVASR>, v2i32, V2I32, I32>;
1334def: OpR_RR_pat<S2_asr_r_vh, pf2<HexagonVASR>, v4i16, V4I16, I32>;
1335def: OpR_RR_pat<S2_lsr_r_vw, pf2<HexagonVLSR>, v2i32, V2I32, I32>;
1336def: OpR_RR_pat<S2_lsr_r_vh, pf2<HexagonVLSR>, v4i16, V4I16, I32>;
1337
1338def: Pat<(sra V2I32:$b, (v2i32 (splat_vector u5_0ImmPred:$c))),
1339         (S2_asr_i_vw V2I32:$b, imm:$c)>;
1340def: Pat<(srl V2I32:$b, (v2i32 (splat_vector u5_0ImmPred:$c))),
1341         (S2_lsr_i_vw V2I32:$b, imm:$c)>;
1342def: Pat<(shl V2I32:$b, (v2i32 (splat_vector u5_0ImmPred:$c))),
1343         (S2_asl_i_vw V2I32:$b, imm:$c)>;
1344def: Pat<(sra V4I16:$b, (v4i16 (splat_vector u4_0ImmPred:$c))),
1345         (S2_asr_i_vh V4I16:$b, imm:$c)>;
1346def: Pat<(srl V4I16:$b, (v4i16 (splat_vector u4_0ImmPred:$c))),
1347         (S2_lsr_i_vh V4I16:$b, imm:$c)>;
1348def: Pat<(shl V4I16:$b, (v4i16 (splat_vector u4_0ImmPred:$c))),
1349         (S2_asl_i_vh V4I16:$b, imm:$c)>;
1350
1351def: Pat<(HexagonVASR V2I16:$Rs, u4_0ImmPred:$S),
1352         (LoReg (S2_asr_i_vh (ToAext64 $Rs), imm:$S))>;
1353def: Pat<(HexagonVASL V2I16:$Rs, u4_0ImmPred:$S),
1354         (LoReg (S2_asl_i_vh (ToAext64 $Rs), imm:$S))>;
1355def: Pat<(HexagonVLSR V2I16:$Rs, u4_0ImmPred:$S),
1356         (LoReg (S2_lsr_i_vh (ToAext64 $Rs), imm:$S))>;
1357def: Pat<(HexagonVASR V2I16:$Rs, I32:$Rt),
1358         (LoReg (S2_asr_i_vh (ToAext64 $Rs), I32:$Rt))>;
1359def: Pat<(HexagonVASL V2I16:$Rs, I32:$Rt),
1360         (LoReg (S2_asl_i_vh (ToAext64 $Rs), I32:$Rt))>;
1361def: Pat<(HexagonVLSR V2I16:$Rs, I32:$Rt),
1362         (LoReg (S2_lsr_i_vh (ToAext64 $Rs), I32:$Rt))>;
1363
1364
1365// --(9) Arithmetic/bitwise ----------------------------------------------
1366//
1367
1368def: Pat<(abs  I32:$Rs), (A2_abs   I32:$Rs)>;
1369def: Pat<(abs  I64:$Rs), (A2_absp  I64:$Rs)>;
1370def: Pat<(not  I32:$Rs), (A2_subri -1, I32:$Rs)>;
1371def: Pat<(not  I64:$Rs), (A2_notp  I64:$Rs)>;
1372def: Pat<(ineg I64:$Rs), (A2_negp  I64:$Rs)>;
1373
1374def: Pat<(fabs F32:$Rs), (S2_clrbit_i    F32:$Rs, 31)>;
1375def: Pat<(fneg F32:$Rs), (S2_togglebit_i F32:$Rs, 31)>;
1376
1377def: Pat<(fabs F64:$Rs),
1378         (Combinew (S2_clrbit_i (HiReg $Rs), 31),
1379                   (i32 (LoReg $Rs)))>;
1380def: Pat<(fneg F64:$Rs),
1381         (Combinew (S2_togglebit_i (HiReg $Rs), 31),
1382                   (i32 (LoReg $Rs)))>;
1383
1384def: Pat<(add I32:$Rs, anyimm:$s16),   (A2_addi   I32:$Rs,  imm:$s16)>;
1385def: Pat<(or  I32:$Rs, anyimm:$s10),   (A2_orir   I32:$Rs,  imm:$s10)>;
1386def: Pat<(and I32:$Rs, anyimm:$s10),   (A2_andir  I32:$Rs,  imm:$s10)>;
1387def: Pat<(sub anyimm:$s10, I32:$Rs),   (A2_subri  imm:$s10, I32:$Rs)>;
1388
1389def: OpR_RR_pat<A2_add,       Add,        i32,   I32>;
1390def: OpR_RR_pat<A2_sub,       Sub,        i32,   I32>;
1391def: OpR_RR_pat<A2_and,       And,        i32,   I32>;
1392def: OpR_RR_pat<A2_or,        Or,         i32,   I32>;
1393def: OpR_RR_pat<A2_xor,       Xor,        i32,   I32>;
1394def: OpR_RR_pat<A2_addp,      Add,        i64,   I64>;
1395def: OpR_RR_pat<A2_subp,      Sub,        i64,   I64>;
1396def: OpR_RR_pat<A2_andp,      And,        i64,   I64>;
1397def: OpR_RR_pat<A2_orp,       Or,         i64,   I64>;
1398def: OpR_RR_pat<A2_xorp,      Xor,        i64,   I64>;
1399def: OpR_RR_pat<A4_andnp,     Not2<And>,  i64,   I64>;
1400def: OpR_RR_pat<A4_ornp,      Not2<Or>,   i64,   I64>;
1401
1402def: OpR_RR_pat<A2_svaddh,    Add,        v2i16, V2I16>;
1403def: OpR_RR_pat<A2_svsubh,    Sub,        v2i16, V2I16>;
1404
1405def: OpR_RR_pat<A2_vaddub,    Add,        v8i8,  V8I8>;
1406def: OpR_RR_pat<A2_vaddh,     Add,        v4i16, V4I16>;
1407def: OpR_RR_pat<A2_vaddw,     Add,        v2i32, V2I32>;
1408def: OpR_RR_pat<A2_vsubub,    Sub,        v8i8,  V8I8>;
1409def: OpR_RR_pat<A2_vsubh,     Sub,        v4i16, V4I16>;
1410def: OpR_RR_pat<A2_vsubw,     Sub,        v2i32, V2I32>;
1411
1412def: OpR_RR_pat<A2_and,       And,        v4i8,  V4I8>;
1413def: OpR_RR_pat<A2_xor,       Xor,        v4i8,  V4I8>;
1414def: OpR_RR_pat<A2_or,        Or,         v4i8,  V4I8>;
1415def: OpR_RR_pat<A2_and,       And,        v2i16, V2I16>;
1416def: OpR_RR_pat<A2_xor,       Xor,        v2i16, V2I16>;
1417def: OpR_RR_pat<A2_or,        Or,         v2i16, V2I16>;
1418def: OpR_RR_pat<A2_andp,      And,        v8i8,  V8I8>;
1419def: OpR_RR_pat<A2_orp,       Or,         v8i8,  V8I8>;
1420def: OpR_RR_pat<A2_xorp,      Xor,        v8i8,  V8I8>;
1421def: OpR_RR_pat<A2_andp,      And,        v4i16, V4I16>;
1422def: OpR_RR_pat<A2_orp,       Or,         v4i16, V4I16>;
1423def: OpR_RR_pat<A2_xorp,      Xor,        v4i16, V4I16>;
1424def: OpR_RR_pat<A2_andp,      And,        v2i32, V2I32>;
1425def: OpR_RR_pat<A2_orp,       Or,         v2i32, V2I32>;
1426def: OpR_RR_pat<A2_xorp,      Xor,        v2i32, V2I32>;
1427
1428def: OpR_RR_pat<M2_mpyi,      Mul,        i32,   I32>;
1429def: OpR_RR_pat<M2_mpy_up,    pf2<mulhs>, i32,   I32>;
1430def: OpR_RR_pat<M2_mpyu_up,   pf2<mulhu>, i32,   I32>;
1431def: OpR_RI_pat<M2_mpysip,    Mul,        i32,   I32, u32_0ImmPred>;
1432def: OpR_RI_pat<M2_mpysmi,    Mul,        i32,   I32, s32_0ImmPred>;
1433
1434// Arithmetic on predicates.
1435def: OpR_RR_pat<C2_xor,       Add,        i1,    I1>;
1436def: OpR_RR_pat<C2_xor,       Add,        v2i1,  V2I1>;
1437def: OpR_RR_pat<C2_xor,       Add,        v4i1,  V4I1>;
1438def: OpR_RR_pat<C2_xor,       Add,        v8i1,  V8I1>;
1439def: OpR_RR_pat<C2_xor,       Sub,        i1,    I1>;
1440def: OpR_RR_pat<C2_xor,       Sub,        v2i1,  V2I1>;
1441def: OpR_RR_pat<C2_xor,       Sub,        v4i1,  V4I1>;
1442def: OpR_RR_pat<C2_xor,       Sub,        v8i1,  V8I1>;
1443def: OpR_RR_pat<C2_and,       Mul,        i1,    I1>;
1444def: OpR_RR_pat<C2_and,       Mul,        v2i1,  V2I1>;
1445def: OpR_RR_pat<C2_and,       Mul,        v4i1,  V4I1>;
1446def: OpR_RR_pat<C2_and,       Mul,        v8i1,  V8I1>;
1447
1448def: OpR_RR_pat<F2_sfadd,     pf2<fadd>,    f32, F32>;
1449def: OpR_RR_pat<F2_sfsub,     pf2<fsub>,    f32, F32>;
1450def: OpR_RR_pat<F2_sfmpy,     pf2<fmul>,    f32, F32>;
1451def: OpR_RR_pat<F2_sfmin,     pf2<fminnum>, f32, F32>;
1452def: OpR_RR_pat<F2_sfmax,     pf2<fmaxnum>, f32, F32>;
1453
1454let Predicates = [HasV66] in {
1455  def: OpR_RR_pat<F2_dfadd,     pf2<fadd>,    f64, F64>;
1456  def: OpR_RR_pat<F2_dfsub,     pf2<fsub>,    f64, F64>;
1457}
1458
1459def DfMpy: OutPatFrag<(ops node:$Rs, node:$Rt),
1460  (F2_dfmpyhh
1461    (F2_dfmpylh
1462      (F2_dfmpylh
1463        (F2_dfmpyll $Rs, $Rt),
1464      $Rs, $Rt),
1465    $Rt, $Rs),
1466  $Rs, $Rt)>;
1467
1468let Predicates = [HasV67,UseUnsafeMath], AddedComplexity = 50 in {
1469  def: Pat<(fmul F64:$Rs, F64:$Rt), (DfMpy $Rs, $Rt)>;
1470}
1471let Predicates = [HasV67] in {
1472  def: OpR_RR_pat<F2_dfmin,     pf2<fminnum>, f64, F64>;
1473  def: OpR_RR_pat<F2_dfmax,     pf2<fmaxnum>, f64, F64>;
1474
1475  def: Pat<(fmul F64:$Rs, F64:$Rt), (DfMpy (F2_dfmpyfix $Rs, $Rt),
1476                                           (F2_dfmpyfix $Rt, $Rs))>;
1477}
1478
1479// In expressions like a0*b0 + a1*b1 + ..., prefer to generate multiply-add,
1480// over add-add with individual multiplies as inputs.
1481let AddedComplexity = 10 in {
1482  def: AccRRI_pat<M2_macsip,    Add, Su<Mul>, I32, u32_0ImmPred>;
1483  def: AccRRI_pat<M2_macsin,    Sub, Su<Mul>, I32, u32_0ImmPred>;
1484  def: AccRRR_pat<M2_maci,      Add, Su<Mul>, I32, I32, I32>;
1485  let Predicates = [HasV66] in
1486  def: AccRRR_pat<M2_mnaci,     Sub, Su<Mul>, I32, I32, I32>;
1487}
1488
1489def: AccRRI_pat<M2_naccii,    Sub, Su<Add>, I32, s32_0ImmPred>;
1490def: AccRRI_pat<M2_accii,     Add, Su<Add>, I32, s32_0ImmPred>;
1491def: AccRRR_pat<M2_acci,      Add, Su<Add>, I32, I32, I32>;
1492
1493// Mulh for vectors
1494//
1495def: Pat<(v2i32 (mulhu V2I32:$Rss, V2I32:$Rtt)),
1496         (Combinew (M2_mpyu_up (HiReg $Rss), (HiReg $Rtt)),
1497                   (M2_mpyu_up (LoReg $Rss), (LoReg $Rtt)))>;
1498
1499def: Pat<(v2i32 (mulhs V2I32:$Rs, V2I32:$Rt)),
1500         (Combinew (M2_mpy_up (HiReg $Rs), (HiReg $Rt)),
1501                   (M2_mpy_up (LoReg $Rt), (LoReg $Rt)))>;
1502
1503def Mulhub:
1504  OutPatFrag<(ops node:$Rss, node:$Rtt),
1505             (Combinew (S2_vtrunohb (M5_vmpybuu (HiReg $Rss), (HiReg $Rtt))),
1506                       (S2_vtrunohb (M5_vmpybuu (LoReg $Rss), (LoReg $Rtt))))>;
1507
1508// Equivalent of byte-wise arithmetic shift right by 7 in v8i8.
1509def Asr7:
1510  OutPatFrag<(ops node:$Rss), (C2_mask (C2_not (A4_vcmpbgti $Rss, 0)))>;
1511
1512def: Pat<(v8i8 (mulhu V8I8:$Rss, V8I8:$Rtt)),
1513         (Mulhub $Rss, $Rtt)>;
1514
1515def: Pat<(v8i8 (mulhs V8I8:$Rss, V8I8:$Rtt)),
1516         (A2_vsubub
1517           (Mulhub $Rss, $Rtt),
1518           (A2_vaddub (A2_andp V8I8:$Rss, (Asr7 $Rtt)),
1519                      (A2_andp V8I8:$Rtt, (Asr7 $Rss))))>;
1520
1521def Mpysh:
1522  OutPatFrag<(ops node:$Rs, node:$Rt), (M2_vmpy2s_s0 $Rs, $Rt)>;
1523def Mpyshh:
1524  OutPatFrag<(ops node:$Rss, node:$Rtt), (Mpysh (HiReg $Rss), (HiReg $Rtt))>;
1525def Mpyshl:
1526  OutPatFrag<(ops node:$Rss, node:$Rtt), (Mpysh (LoReg $Rss), (LoReg $Rtt))>;
1527
1528def Mulhsh:
1529  OutPatFrag<(ops node:$Rss, node:$Rtt),
1530             (Combinew (A2_combine_hh (HiReg (Mpyshh $Rss, $Rtt)),
1531                                      (LoReg (Mpyshh $Rss, $Rtt))),
1532                       (A2_combine_hh (HiReg (Mpyshl $Rss, $Rtt)),
1533                                      (LoReg (Mpyshl $Rss, $Rtt))))>;
1534
1535def: Pat<(v4i16 (mulhs V4I16:$Rss, V4I16:$Rtt)), (Mulhsh $Rss, $Rtt)>;
1536
1537def: Pat<(v4i16 (mulhu V4I16:$Rss, V4I16:$Rtt)),
1538         (A2_vaddh
1539           (Mulhsh $Rss, $Rtt),
1540           (A2_vaddh (A2_andp V4I16:$Rss, (S2_asr_i_vh $Rtt, 15)),
1541                     (A2_andp V4I16:$Rtt, (S2_asr_i_vh $Rss, 15))))>;
1542
1543
1544def: Pat<(ineg (mul I32:$Rs, u8_0ImmPred:$u8)),
1545         (M2_mpysin IntRegs:$Rs, imm:$u8)>;
1546
1547def n8_0ImmPred: PatLeaf<(i32 imm), [{
1548  int64_t V = N->getSExtValue();
1549  return -255 <= V && V <= 0;
1550}]>;
1551
1552// Change the sign of the immediate for Rd=-mpyi(Rs,#u8)
1553def: Pat<(mul I32:$Rs, n8_0ImmPred:$n8),
1554         (M2_mpysin I32:$Rs, (NegImm8 imm:$n8))>;
1555
1556def: Pat<(add Sext64:$Rs, I64:$Rt),
1557         (A2_addsp (LoReg Sext64:$Rs), I64:$Rt)>;
1558
1559def: AccRRR_pat<M4_and_and,   And, Su_ni1<And>,  I32,  I32,  I32>;
1560def: AccRRR_pat<M4_and_or,    And, Su_ni1<Or>,   I32,  I32,  I32>;
1561def: AccRRR_pat<M4_and_xor,   And, Su<Xor>,      I32,  I32,  I32>;
1562def: AccRRR_pat<M4_or_and,    Or,  Su_ni1<And>,  I32,  I32,  I32>;
1563def: AccRRR_pat<M4_or_or,     Or,  Su_ni1<Or>,   I32,  I32,  I32>;
1564def: AccRRR_pat<M4_or_xor,    Or,  Su<Xor>,      I32,  I32,  I32>;
1565def: AccRRR_pat<M4_xor_and,   Xor, Su_ni1<And>,  I32,  I32,  I32>;
1566def: AccRRR_pat<M4_xor_or,    Xor, Su_ni1<Or>,   I32,  I32,  I32>;
1567def: AccRRR_pat<M2_xor_xacc,  Xor, Su<Xor>,      I32,  I32,  I32>;
1568def: AccRRR_pat<M4_xor_xacc,  Xor, Su<Xor>,      I64,  I64,  I64>;
1569
1570// For dags like (or (and (not _), _), (shl _, _)) where the "or" with
1571// one argument matches the patterns below, and with the other argument
1572// matches S2_asl_r_r_or, etc, prefer the patterns below.
1573let AddedComplexity = 110 in {  // greater than S2_asl_r_r_and/or/xor.
1574  def: AccRRR_pat<M4_and_andn,  And, Su<Not2<And>>, I32,  I32,  I32>;
1575  def: AccRRR_pat<M4_or_andn,   Or,  Su<Not2<And>>, I32,  I32,  I32>;
1576  def: AccRRR_pat<M4_xor_andn,  Xor, Su<Not2<And>>, I32,  I32,  I32>;
1577}
1578
1579// S4_addaddi and S4_subaddi don't have tied operands, so give them
1580// a bit of preference.
1581let AddedComplexity = 30, Predicates = [UseCompound] in {
1582  def: Pat<(add I32:$Rs, (Su<Add> I32:$Ru, anyimm:$s6)),
1583           (S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
1584  def: Pat<(add anyimm:$s6, (Su<Add> I32:$Rs, I32:$Ru)),
1585           (S4_addaddi IntRegs:$Rs, IntRegs:$Ru, imm:$s6)>;
1586  def: Pat<(add I32:$Rs, (Su<Sub> anyimm:$s6, I32:$Ru)),
1587           (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1588  def: Pat<(sub (Su<Add> I32:$Rs, anyimm:$s6), I32:$Ru),
1589           (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1590  def: Pat<(add (Su<Sub> I32:$Rs, I32:$Ru), anyimm:$s6),
1591           (S4_subaddi IntRegs:$Rs, imm:$s6, IntRegs:$Ru)>;
1592}
1593
1594let Predicates = [UseCompound] in
1595def: Pat<(or I32:$Ru, (Su<And> I32:$Rx, anyimm:$s10)),
1596         (S4_or_andix IntRegs:$Ru, IntRegs:$Rx, imm:$s10)>;
1597
1598def: Pat<(or I32:$Rx, (Su<And> I32:$Rs, anyimm:$s10)),
1599         (S4_or_andi IntRegs:$Rx, IntRegs:$Rs, imm:$s10)>;
1600def: Pat<(or I32:$Rx, (Su<Or> I32:$Rs, anyimm:$s10)),
1601         (S4_or_ori IntRegs:$Rx, IntRegs:$Rs, imm:$s10)>;
1602
1603
1604def: Pat<(i32 (trunc (sra (Su<Mul> Sext64:$Rs, Sext64:$Rt), (i32 32)))),
1605         (M2_mpy_up (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1606def: Pat<(i32 (trunc (srl (Su<Mul> Sext64:$Rs, Sext64:$Rt), (i32 32)))),
1607         (M2_mpy_up (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1608
1609def: Pat<(mul (Zext64 I32:$Rs), (Zext64 I32:$Rt)),
1610         (M2_dpmpyuu_s0 I32:$Rs, I32:$Rt)>;
1611def: Pat<(mul (Aext64 I32:$Rs), (Aext64 I32:$Rt)),
1612         (M2_dpmpyuu_s0 I32:$Rs, I32:$Rt)>;
1613def: Pat<(mul Sext64:$Rs, Sext64:$Rt),
1614         (M2_dpmpyss_s0 (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1615
1616def: Pat<(add I64:$Rx, (Su<Mul> Sext64:$Rs, Sext64:$Rt)),
1617         (M2_dpmpyss_acc_s0 I64:$Rx, (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1618def: Pat<(sub I64:$Rx, (Su<Mul> Sext64:$Rs, Sext64:$Rt)),
1619         (M2_dpmpyss_nac_s0 I64:$Rx, (LoReg Sext64:$Rs), (LoReg Sext64:$Rt))>;
1620def: Pat<(add I64:$Rx, (Su<Mul> (Aext64 I32:$Rs), (Aext64 I32:$Rt))),
1621         (M2_dpmpyuu_acc_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1622def: Pat<(add I64:$Rx, (Su<Mul> (Zext64 I32:$Rs), (Zext64 I32:$Rt))),
1623         (M2_dpmpyuu_acc_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1624def: Pat<(sub I64:$Rx, (Su<Mul> (Aext64 I32:$Rs), (Aext64 I32:$Rt))),
1625         (M2_dpmpyuu_nac_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1626def: Pat<(sub I64:$Rx, (Su<Mul> (Zext64 I32:$Rs), (Zext64 I32:$Rt))),
1627         (M2_dpmpyuu_nac_s0 I64:$Rx, I32:$Rs, I32:$Rt)>;
1628
1629// Add halfword.
1630def: Pat<(sext_inreg (add I32:$Rt, I32:$Rs), i16),
1631         (A2_addh_l16_ll I32:$Rt, I32:$Rs)>;
1632def: Pat<(sra (add (shl I32:$Rt, (i32 16)), I32:$Rs), (i32 16)),
1633         (A2_addh_l16_hl I32:$Rt, I32:$Rs)>;
1634def: Pat<(shl (add I32:$Rt, I32:$Rs), (i32 16)),
1635         (A2_addh_h16_ll I32:$Rt, I32:$Rs)>;
1636
1637// Subtract halfword.
1638def: Pat<(sext_inreg (sub I32:$Rt, I32:$Rs), i16),
1639         (A2_subh_l16_ll I32:$Rt, I32:$Rs)>;
1640def: Pat<(sra (add (shl I32:$Rt, (i32 16)), I32:$Rs), (i32 16)),
1641         (A2_addh_l16_hl I32:$Rt, I32:$Rs)>;
1642def: Pat<(shl (sub I32:$Rt, I32:$Rs), (i32 16)),
1643         (A2_subh_h16_ll I32:$Rt, I32:$Rs)>;
1644
1645def: Pat<(mul I64:$Rss, I64:$Rtt),
1646         (Combinew
1647           (M2_maci (M2_maci (HiReg (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt))),
1648                             (LoReg $Rss),
1649                             (HiReg $Rtt)),
1650                    (LoReg $Rtt),
1651                    (HiReg $Rss)),
1652           (i32 (LoReg (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt)))))>;
1653
1654def MulHU : OutPatFrag<(ops node:$Rss, node:$Rtt),
1655  (A2_addp
1656    (M2_dpmpyuu_acc_s0
1657      (S2_lsr_i_p
1658        (A2_addp
1659          (M2_dpmpyuu_acc_s0
1660            (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (LoReg $Rtt)), 32),
1661            (HiReg $Rss),
1662            (LoReg $Rtt)),
1663          (A4_combineir 0, (LoReg (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt))))),
1664        32),
1665      (HiReg $Rss),
1666      (HiReg $Rtt)),
1667    (S2_lsr_i_p (M2_dpmpyuu_s0 (LoReg $Rss), (HiReg $Rtt)), 32))>;
1668
1669// Multiply 64-bit unsigned and use upper result.
1670def : Pat <(mulhu I64:$Rss, I64:$Rtt), (MulHU $Rss, $Rtt)>;
1671
1672// Multiply 64-bit signed and use upper result.
1673//
1674// For two signed 64-bit integers A and B, let A' and B' denote A and B
1675// with the sign bit cleared. Then A = -2^63*s(A) + A', where s(A) is the
1676// sign bit of A (and identically for B). With this notation, the signed
1677// product A*B can be written as:
1678//   AB = (-2^63 s(A) + A') * (-2^63 s(B) + B')
1679//      = 2^126 s(A)s(B) - 2^63 [s(A)B'+s(B)A'] + A'B'
1680//      = 2^126 s(A)s(B) + 2^63 [s(A)B'+s(B)A'] + A'B' - 2*2^63 [s(A)B'+s(B)A']
1681//      = (unsigned product AB) - 2^64 [s(A)B'+s(B)A']
1682
1683// Clear the sign bit in a 64-bit register.
1684def ClearSign : OutPatFrag<(ops node:$Rss),
1685  (Combinew (S2_clrbit_i (HiReg $Rss), 31), (i32 (LoReg $Rss)))>;
1686
1687def : Pat <(mulhs I64:$Rss, I64:$Rtt),
1688  (A2_subp
1689    (MulHU $Rss, $Rtt),
1690    (A2_addp
1691      (A2_andp (S2_asr_i_p $Rss, 63), (ClearSign $Rtt)),
1692      (A2_andp (S2_asr_i_p $Rtt, 63), (ClearSign $Rss))))>;
1693
1694// Prefer these instructions over M2_macsip/M2_macsin: the macsi* instructions
1695// will put the immediate addend into a register, while these instructions will
1696// use it directly. Such a construct does not appear in the middle of a gep,
1697// where M2_macsip would be preferable.
1698let AddedComplexity = 20, Predicates = [UseCompound] in {
1699  def: Pat<(add (Su<Mul> I32:$Rs, u6_0ImmPred:$U6), anyimm:$u6),
1700           (M4_mpyri_addi imm:$u6, IntRegs:$Rs, imm:$U6)>;
1701  def: Pat<(add (Su<Mul> I32:$Rs, I32:$Rt), anyimm:$u6),
1702           (M4_mpyrr_addi imm:$u6, IntRegs:$Rs, IntRegs:$Rt)>;
1703}
1704
1705// Keep these instructions less preferable to M2_macsip/M2_macsin.
1706let Predicates = [UseCompound] in {
1707  def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, u6_2ImmPred:$u6_2)),
1708           (M4_mpyri_addr_u2 IntRegs:$Ru, imm:$u6_2, IntRegs:$Rs)>;
1709  def: Pat<(add I32:$Ru, (Su<Mul> I32:$Rs, anyimm:$u6)),
1710           (M4_mpyri_addr IntRegs:$Ru, IntRegs:$Rs, imm:$u6)>;
1711  def: Pat<(add I32:$Ru, (Su<Mul> I32:$Ry, I32:$Rs)),
1712           (M4_mpyrr_addr IntRegs:$Ru, IntRegs:$Ry, IntRegs:$Rs)>;
1713}
1714
1715def: Pat<(fma F32:$Rs, F32:$Rt, F32:$Rx),
1716         (F2_sffma F32:$Rx, F32:$Rs, F32:$Rt)>;
1717def: Pat<(fma (fneg F32:$Rs), F32:$Rt, F32:$Rx),
1718         (F2_sffms F32:$Rx, F32:$Rs, F32:$Rt)>;
1719
1720def: Pat<(mul V2I32:$Rs, V2I32:$Rt),
1721         (PS_vmulw V2I32:$Rs, V2I32:$Rt)>;
1722def: Pat<(add V2I32:$Rx, (mul V2I32:$Rs, V2I32:$Rt)),
1723         (PS_vmulw_acc V2I32:$Rx, V2I32:$Rs, V2I32:$Rt)>;
1724
1725// Add/subtract two v4i8: Hexagon does not have an insn for this one, so
1726// we use the double add v8i8, and use only the low part of the result.
1727def: Pat<(add V4I8:$Rs, V4I8:$Rt),
1728         (LoReg (A2_vaddub (ToAext64 $Rs), (ToAext64 $Rt)))>;
1729def: Pat<(sub V4I8:$Rs, V4I8:$Rt),
1730         (LoReg (A2_vsubub (ToAext64 $Rs), (ToAext64 $Rt)))>;
1731
1732// Use M2_vmpy2s_s0 for half-word vector multiply. It multiplies two
1733// half-words, and saturates the result to a 32-bit value, except the
1734// saturation never happens (it can only occur with scaling).
1735def: Pat<(v2i16 (mul V2I16:$Rs, V2I16:$Rt)),
1736         (LoReg (S2_vtrunewh (A2_combineii 0, 0),
1737                             (M2_vmpy2s_s0 V2I16:$Rs, V2I16:$Rt)))>;
1738def: Pat<(v4i16 (mul V4I16:$Rs, V4I16:$Rt)),
1739         (S2_vtrunewh (M2_vmpy2s_s0 (HiReg $Rs), (HiReg $Rt)),
1740                      (M2_vmpy2s_s0 (LoReg $Rs), (LoReg $Rt)))>;
1741
1742// Multiplies two v4i8 vectors.
1743def: Pat<(v4i8 (mul V4I8:$Rs, V4I8:$Rt)),
1744         (S2_vtrunehb (M5_vmpybuu V4I8:$Rs, V4I8:$Rt))>;
1745
1746// Multiplies two v8i8 vectors.
1747def: Pat<(v8i8 (mul V8I8:$Rs, V8I8:$Rt)),
1748         (Combinew (S2_vtrunehb (M5_vmpybuu (HiReg $Rs), (HiReg $Rt))),
1749                   (S2_vtrunehb (M5_vmpybuu (LoReg $Rs), (LoReg $Rt))))>;
1750
1751
1752// --(10) Bit ------------------------------------------------------------
1753//
1754
1755// Count leading zeros.
1756def: Pat<(i32 (ctlz I32:$Rs)),                (S2_cl0 I32:$Rs)>;
1757def: Pat<(i32 (trunc (ctlz I64:$Rss))),       (S2_cl0p I64:$Rss)>;
1758
1759// Count trailing zeros.
1760def: Pat<(i32 (cttz I32:$Rs)),                (S2_ct0 I32:$Rs)>;
1761def: Pat<(i32 (trunc (cttz I64:$Rss))),       (S2_ct0p I64:$Rss)>;
1762
1763// Count leading ones.
1764def: Pat<(i32 (ctlz (not I32:$Rs))),          (S2_cl1 I32:$Rs)>;
1765def: Pat<(i32 (trunc (ctlz (not I64:$Rss)))), (S2_cl1p I64:$Rss)>;
1766
1767// Count trailing ones.
1768def: Pat<(i32 (cttz (not I32:$Rs))),           (S2_ct1 I32:$Rs)>;
1769def: Pat<(i32 (trunc (cttz (not I64:$Rss)))), (S2_ct1p I64:$Rss)>;
1770
1771// Define leading/trailing patterns that require zero-extensions to 64 bits.
1772def: Pat<(i64 (ctlz I64:$Rss)),               (ToZext64 (S2_cl0p I64:$Rss))>;
1773def: Pat<(i64 (cttz I64:$Rss)),               (ToZext64 (S2_ct0p I64:$Rss))>;
1774def: Pat<(i64 (ctlz (not I64:$Rss))),         (ToZext64 (S2_cl1p I64:$Rss))>;
1775def: Pat<(i64 (cttz (not I64:$Rss))),         (ToZext64 (S2_ct1p I64:$Rss))>;
1776
1777def: Pat<(i64 (ctpop I64:$Rss)),  (ToZext64 (S5_popcountp I64:$Rss))>;
1778def: Pat<(i32 (ctpop I32:$Rs)),   (S5_popcountp (A4_combineir 0, I32:$Rs))>;
1779
1780def: Pat<(bitreverse I32:$Rs),    (S2_brev I32:$Rs)>;
1781def: Pat<(bitreverse I64:$Rss),   (S2_brevp I64:$Rss)>;
1782
1783let AddedComplexity = 20 in { // Complexity greater than and/or/xor
1784  def: Pat<(and I32:$Rs, IsNPow2_32:$V),
1785           (S2_clrbit_i IntRegs:$Rs, (LogN2_32 $V))>;
1786  def: Pat<(or I32:$Rs, IsPow2_32:$V),
1787           (S2_setbit_i IntRegs:$Rs, (Log2_32 $V))>;
1788  def: Pat<(xor I32:$Rs, IsPow2_32:$V),
1789           (S2_togglebit_i IntRegs:$Rs, (Log2_32 $V))>;
1790
1791  def: Pat<(and I32:$Rs, (not (shl 1, I32:$Rt))),
1792           (S2_clrbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1793  def: Pat<(or I32:$Rs, (shl 1, I32:$Rt)),
1794           (S2_setbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1795  def: Pat<(xor I32:$Rs, (shl 1, I32:$Rt)),
1796           (S2_togglebit_r IntRegs:$Rs, IntRegs:$Rt)>;
1797}
1798
1799// Clr/set/toggle bit for 64-bit values with immediate bit index.
1800let AddedComplexity = 20 in { // Complexity greater than and/or/xor
1801  def: Pat<(and I64:$Rss, IsNPow2_64L:$V),
1802           (Combinew (i32 (HiReg $Rss)),
1803                     (S2_clrbit_i (LoReg $Rss), (LogN2_64 $V)))>;
1804  def: Pat<(and I64:$Rss, IsNPow2_64H:$V),
1805           (Combinew (S2_clrbit_i (HiReg $Rss), (UDEC32 (i32 (LogN2_64 $V)))),
1806                     (i32 (LoReg $Rss)))>;
1807
1808  def: Pat<(or I64:$Rss, IsPow2_64L:$V),
1809           (Combinew (i32 (HiReg $Rss)),
1810                     (S2_setbit_i (LoReg $Rss), (Log2_64 $V)))>;
1811  def: Pat<(or I64:$Rss, IsPow2_64H:$V),
1812           (Combinew (S2_setbit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
1813                     (i32 (LoReg $Rss)))>;
1814
1815  def: Pat<(xor I64:$Rss, IsPow2_64L:$V),
1816           (Combinew (i32 (HiReg $Rss)),
1817                     (S2_togglebit_i (LoReg $Rss), (Log2_64 $V)))>;
1818  def: Pat<(xor I64:$Rss, IsPow2_64H:$V),
1819           (Combinew (S2_togglebit_i (HiReg $Rss), (UDEC32 (i32 (Log2_64 $V)))),
1820                     (i32 (LoReg $Rss)))>;
1821}
1822
1823
1824let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm.
1825  def: Pat<(i1 (setne (and (shl 1, u5_0ImmPred:$u5), I32:$Rs), 0)),
1826           (S2_tstbit_i IntRegs:$Rs, imm:$u5)>;
1827  def: Pat<(i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1828           (S2_tstbit_r IntRegs:$Rs, IntRegs:$Rt)>;
1829  def: Pat<(i1 (trunc I32:$Rs)),
1830           (S2_tstbit_i IntRegs:$Rs, 0)>;
1831  def: Pat<(i1 (trunc I64:$Rs)),
1832           (S2_tstbit_i (LoReg DoubleRegs:$Rs), 0)>;
1833}
1834
1835def: Pat<(and (srl I32:$Rs, u5_0ImmPred:$u5), 1),
1836         (I1toI32 (S2_tstbit_i I32:$Rs, imm:$u5))>;
1837def: Pat<(and (srl I64:$Rss, IsULE<32,31>:$u6), 1),
1838         (ToZext64 (I1toI32 (S2_tstbit_i (LoReg $Rss), imm:$u6)))>;
1839def: Pat<(and (srl I64:$Rss, IsUGT<32,31>:$u6), 1),
1840         (ToZext64 (I1toI32 (S2_tstbit_i (HiReg $Rss), (UDEC32 $u6))))>;
1841
1842def: Pat<(and (not (srl I32:$Rs, u5_0ImmPred:$u5)), 1),
1843         (I1toI32 (S4_ntstbit_i I32:$Rs, imm:$u5))>;
1844def: Pat<(and (not (srl I64:$Rss, IsULE<32,31>:$u6)), 1),
1845         (ToZext64 (I1toI32 (S4_ntstbit_i (LoReg $Rss), imm:$u6)))>;
1846def: Pat<(and (not (srl I64:$Rss, IsUGT<32,31>:$u6)), 1),
1847         (ToZext64 (I1toI32 (S4_ntstbit_i (HiReg $Rss), (UDEC32 $u6))))>;
1848
1849let AddedComplexity = 20 in { // Complexity greater than compare reg-imm.
1850  def: Pat<(i1 (seteq (and I32:$Rs, u6_0ImmPred:$u6), 0)),
1851           (C2_bitsclri IntRegs:$Rs, imm:$u6)>;
1852  def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), 0)),
1853           (C2_bitsclr IntRegs:$Rs, IntRegs:$Rt)>;
1854}
1855
1856let AddedComplexity = 10 in   // Complexity greater than compare reg-reg.
1857def: Pat<(i1 (seteq (and I32:$Rs, I32:$Rt), IntRegs:$Rt)),
1858         (C2_bitsset IntRegs:$Rs, IntRegs:$Rt)>;
1859
1860def SDTTestBit:
1861  SDTypeProfile<1, 2, [SDTCisVT<0, i1>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
1862def HexagonTSTBIT: SDNode<"HexagonISD::TSTBIT", SDTTestBit>;
1863
1864def: Pat<(HexagonTSTBIT I32:$Rs, u5_0ImmPred:$u5),
1865         (S2_tstbit_i I32:$Rs, imm:$u5)>;
1866def: Pat<(HexagonTSTBIT I32:$Rs, I32:$Rt),
1867         (S2_tstbit_r I32:$Rs, I32:$Rt)>;
1868
1869// Add extra complexity to prefer these instructions over bitsset/bitsclr.
1870// The reason is that tstbit/ntstbit can be folded into a compound instruction:
1871//   if ([!]tstbit(...)) jump ...
1872let AddedComplexity = 20 in {   // Complexity greater than cmp reg-imm.
1873  def: Pat<(i1 (seteq (and I32:$Rs, IsPow2_32:$u5), 0)),
1874           (S4_ntstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
1875  def: Pat<(i1 (setne (and I32:$Rs, IsPow2_32:$u5), 0)),
1876           (S2_tstbit_i I32:$Rs, (Log2_32 imm:$u5))>;
1877  def: Pat<(i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1878           (S4_ntstbit_r I32:$Rs, I32:$Rt)>;
1879  def: Pat<(i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)),
1880           (S2_tstbit_r I32:$Rs, I32:$Rt)>;
1881}
1882
1883def: Pat<(i1 (seteq (and I64:$Rs, IsPow2_64L:$u6), 0)),
1884         (S4_ntstbit_i (LoReg $Rs), (Log2_64 $u6))>;
1885def: Pat<(i1 (seteq (and I64:$Rs, IsPow2_64H:$u6), 0)),
1886         (S4_ntstbit_i (HiReg $Rs), (UDEC32 (i32 (Log2_64 $u6))))>;
1887def: Pat<(i1 (setne (and I64:$Rs, IsPow2_64L:$u6), 0)),
1888         (S2_tstbit_i (LoReg $Rs), (Log2_64 imm:$u6))>;
1889def: Pat<(i1 (setne (and I64:$Rs, IsPow2_64H:$u6), 0)),
1890         (S2_tstbit_i (HiReg $Rs), (UDEC32 (i32 (Log2_64 imm:$u6))))>;
1891
1892// Do not increase complexity of these patterns. In the DAG, "cmp i8" may be
1893// represented as a compare against "value & 0xFF", which is an exact match
1894// for cmpb (same for cmph). The patterns below do not contain any additional
1895// complexity that would make them preferable, and if they were actually used
1896// instead of cmpb/cmph, they would result in a compare against register that
1897// is loaded with the byte/half mask (i.e. 0xFF or 0xFFFF).
1898def: Pat<(i1 (setne (and I32:$Rs, u6_0ImmPred:$u6), 0)),
1899         (C4_nbitsclri I32:$Rs, imm:$u6)>;
1900def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), 0)),
1901         (C4_nbitsclr I32:$Rs, I32:$Rt)>;
1902def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), I32:$Rt)),
1903         (C4_nbitsset I32:$Rs, I32:$Rt)>;
1904
1905// Special patterns to address certain cases where the "top-down" matching
1906// algorithm would cause suboptimal selection.
1907
1908let AddedComplexity = 100 in {
1909  // Avoid A4_rcmp[n]eqi in these cases:
1910  def: Pat<(i32 (zext (i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1911           (I1toI32 (S4_ntstbit_r IntRegs:$Rs, IntRegs:$Rt))>;
1912  def: Pat<(i32 (zext (i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1913           (I1toI32 (S2_tstbit_r IntRegs:$Rs, IntRegs:$Rt))>;
1914  def: Pat<(i32 (zext (i1 (seteq (and I32:$Rs, IsPow2_32:$u5), 0)))),
1915           (I1toI32 (S4_ntstbit_i I32:$Rs, (Log2_32 imm:$u5)))>;
1916  def: Pat<(i32 (zext (i1 (setne (and I32:$Rs, IsPow2_32:$u5), 0)))),
1917           (I1toI32 (S2_tstbit_i I32:$Rs, (Log2_32 imm:$u5)))>;
1918  def: Pat<(i32 (zext (i1 (seteq (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1919           (I1toI32 (S4_ntstbit_r I32:$Rs, I32:$Rt))>;
1920  def: Pat<(i32 (zext (i1 (setne (and (shl 1, I32:$Rt), I32:$Rs), 0)))),
1921           (I1toI32 (S2_tstbit_r I32:$Rs, I32:$Rt))>;
1922}
1923
1924// --(11) PIC ------------------------------------------------------------
1925//
1926
1927def SDT_HexagonAtGot
1928  : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisVT<2, i32>]>;
1929def SDT_HexagonAtPcrel
1930  : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
1931
1932// AT_GOT address-of-GOT, address-of-global, offset-in-global
1933def HexagonAtGot       : SDNode<"HexagonISD::AT_GOT", SDT_HexagonAtGot>;
1934// AT_PCREL address-of-global
1935def HexagonAtPcrel     : SDNode<"HexagonISD::AT_PCREL", SDT_HexagonAtPcrel>;
1936
1937def: Pat<(HexagonAtGot I32:$got, I32:$addr, (i32 0)),
1938         (L2_loadri_io I32:$got, imm:$addr)>;
1939def: Pat<(HexagonAtGot I32:$got, I32:$addr, s30_2ImmPred:$off),
1940         (A2_addi (L2_loadri_io I32:$got, imm:$addr), imm:$off)>;
1941def: Pat<(HexagonAtPcrel I32:$addr),
1942         (C4_addipc imm:$addr)>;
1943
1944// The HVX load patterns also match AT_PCREL directly. Make sure that
1945// if the selection of this opcode changes, it's updated in all places.
1946
1947
1948// --(12) Load -----------------------------------------------------------
1949//
1950
1951def L1toI32:  OutPatFrag<(ops node:$Rs), (A2_subri 0, (i32 $Rs))>;
1952def L1toI64:  OutPatFrag<(ops node:$Rs), (ToSext64 (L1toI32 $Rs))>;
1953
1954def extloadv2i8: PatFrag<(ops node:$ptr), (extload node:$ptr), [{
1955  return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1956}]>;
1957def extloadv4i8: PatFrag<(ops node:$ptr), (extload node:$ptr), [{
1958  return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1959}]>;
1960
1961def zextloadv2i8: PatFrag<(ops node:$ptr), (zextload node:$ptr), [{
1962  return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1963}]>;
1964def zextloadv4i8: PatFrag<(ops node:$ptr), (zextload node:$ptr), [{
1965  return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1966}]>;
1967
1968def sextloadv2i8: PatFrag<(ops node:$ptr), (sextload node:$ptr), [{
1969  return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v2i8;
1970}]>;
1971def sextloadv4i8: PatFrag<(ops node:$ptr), (sextload node:$ptr), [{
1972  return cast<LoadSDNode>(N)->getMemoryVT() == MVT::v4i8;
1973}]>;
1974
1975// Patterns to select load-indexed: Rs + Off.
1976// - frameindex [+ imm],
1977multiclass Loadxfi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1978                       InstHexagon MI> {
1979  def: Pat<(VT (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
1980           (VT (MI AddrFI:$fi, imm:$Off))>;
1981  def: Pat<(VT (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
1982           (VT (MI AddrFI:$fi, imm:$Off))>;
1983  def: Pat<(VT (Load AddrFI:$fi)), (VT (MI AddrFI:$fi, 0))>;
1984}
1985
1986// Patterns to select load-indexed: Rs + Off.
1987// - base reg [+ imm]
1988multiclass Loadxgi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1989                       InstHexagon MI> {
1990  def: Pat<(VT (Load (add I32:$Rs, ImmPred:$Off))),
1991           (VT (MI IntRegs:$Rs, imm:$Off))>;
1992  def: Pat<(VT (Load (IsOrAdd I32:$Rs, ImmPred:$Off))),
1993           (VT (MI IntRegs:$Rs, imm:$Off))>;
1994  def: Pat<(VT (Load I32:$Rs)), (VT (MI IntRegs:$Rs, 0))>;
1995}
1996
1997// Patterns to select load-indexed: Rs + Off. Combines Loadxfi + Loadxgi.
1998multiclass Loadxi_pat<PatFrag Load, ValueType VT, PatLeaf ImmPred,
1999                      InstHexagon MI> {
2000  defm: Loadxfi_pat<Load, VT, ImmPred, MI>;
2001  defm: Loadxgi_pat<Load, VT, ImmPred, MI>;
2002}
2003
2004// Patterns to select load reg indexed: Rs + Off with a value modifier.
2005// - frameindex [+ imm]
2006multiclass Loadxfim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2007                        PatLeaf ImmPred, InstHexagon MI> {
2008  def: Pat<(VT (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
2009           (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>;
2010  def: Pat<(VT (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
2011           (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>;
2012  def: Pat<(VT (Load AddrFI:$fi)), (VT (ValueMod (MI AddrFI:$fi, 0)))>;
2013}
2014
2015// Patterns to select load reg indexed: Rs + Off with a value modifier.
2016// - base reg [+ imm]
2017multiclass Loadxgim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2018                        PatLeaf ImmPred, InstHexagon MI> {
2019  def: Pat<(VT (Load (add I32:$Rs, ImmPred:$Off))),
2020           (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
2021  def: Pat<(VT (Load (IsOrAdd I32:$Rs, ImmPred:$Off))),
2022           (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>;
2023  def: Pat<(VT (Load I32:$Rs)), (VT (ValueMod (MI IntRegs:$Rs, 0)))>;
2024}
2025
2026// Patterns to select load reg indexed: Rs + Off with a value modifier.
2027// Combines Loadxfim + Loadxgim.
2028multiclass Loadxim_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2029                       PatLeaf ImmPred, InstHexagon MI> {
2030  defm: Loadxfim_pat<Load, VT, ValueMod, ImmPred, MI>;
2031  defm: Loadxgim_pat<Load, VT, ValueMod, ImmPred, MI>;
2032}
2033
2034// Pattern to select load reg reg-indexed: Rs + Rt<<u2.
2035class Loadxr_shl_pat<PatFrag Load, ValueType VT, InstHexagon MI>
2036  : Pat<(VT (Load (add I32:$Rs, (i32 (shl I32:$Rt, u2_0ImmPred:$u2))))),
2037        (VT (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2))>;
2038
2039// Pattern to select load reg reg-indexed: Rs + Rt<<0.
2040class Loadxr_add_pat<PatFrag Load, ValueType VT, InstHexagon MI>
2041  : Pat<(VT (Load (add I32:$Rs, I32:$Rt))),
2042        (VT (MI IntRegs:$Rs, IntRegs:$Rt, 0))>;
2043
2044// Pattern to select load reg reg-indexed: Rs + Rt<<u2 with value modifier.
2045class Loadxrm_shl_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2046                      InstHexagon MI>
2047  : Pat<(VT (Load (add I32:$Rs, (i32 (shl I32:$Rt, u2_0ImmPred:$u2))))),
2048        (VT (ValueMod (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2)))>;
2049
2050// Pattern to select load reg reg-indexed: Rs + Rt<<0 with value modifier.
2051class Loadxrm_add_pat<PatFrag Load, ValueType VT, PatFrag ValueMod,
2052                      InstHexagon MI>
2053  : Pat<(VT (Load (add I32:$Rs, I32:$Rt))),
2054        (VT (ValueMod (MI IntRegs:$Rs, IntRegs:$Rt, 0)))>;
2055
2056// Pattern to select load long-offset reg-indexed: Addr + Rt<<u2.
2057// Don't match for u2==0, instead use reg+imm for those cases.
2058class Loadxu_pat<PatFrag Load, ValueType VT, PatFrag ImmPred, InstHexagon MI>
2059  : Pat<(VT (Load (add (shl IntRegs:$Rt, u2_0ImmPred:$u2), ImmPred:$Addr))),
2060        (VT (MI IntRegs:$Rt, imm:$u2, ImmPred:$Addr))>;
2061
2062class Loadxum_pat<PatFrag Load, ValueType VT, PatFrag ImmPred, PatFrag ValueMod,
2063                  InstHexagon MI>
2064  : Pat<(VT (Load (add (shl IntRegs:$Rt, u2_0ImmPred:$u2), ImmPred:$Addr))),
2065        (VT (ValueMod (MI IntRegs:$Rt, imm:$u2, ImmPred:$Addr)))>;
2066
2067// Pattern to select load absolute.
2068class Loada_pat<PatFrag Load, ValueType VT, PatFrag Addr, InstHexagon MI>
2069  : Pat<(VT (Load Addr:$addr)), (MI Addr:$addr)>;
2070
2071// Pattern to select load absolute with value modifier.
2072class Loadam_pat<PatFrag Load, ValueType VT, PatFrag Addr, PatFrag ValueMod,
2073                 InstHexagon MI>
2074  : Pat<(VT (Load Addr:$addr)), (ValueMod (MI Addr:$addr))>;
2075
2076
2077let AddedComplexity = 20 in {
2078  defm: Loadxi_pat<extloadi1,       i32,   anyimm0, L2_loadrub_io>;
2079  defm: Loadxi_pat<extloadi8,       i32,   anyimm0, L2_loadrub_io>;
2080  defm: Loadxi_pat<extloadi16,      i32,   anyimm1, L2_loadruh_io>;
2081  defm: Loadxi_pat<extloadv2i8,     v2i16, anyimm1, L2_loadbzw2_io>;
2082  defm: Loadxi_pat<extloadv4i8,     v4i16, anyimm2, L2_loadbzw4_io>;
2083  defm: Loadxi_pat<sextloadi8,      i32,   anyimm0, L2_loadrb_io>;
2084  defm: Loadxi_pat<sextloadi16,     i32,   anyimm1, L2_loadrh_io>;
2085  defm: Loadxi_pat<sextloadv2i8,    v2i16, anyimm1, L2_loadbsw2_io>;
2086  defm: Loadxi_pat<sextloadv4i8,    v4i16, anyimm2, L2_loadbzw4_io>;
2087  defm: Loadxi_pat<zextloadi1,      i32,   anyimm0, L2_loadrub_io>;
2088  defm: Loadxi_pat<zextloadi8,      i32,   anyimm0, L2_loadrub_io>;
2089  defm: Loadxi_pat<zextloadi16,     i32,   anyimm1, L2_loadruh_io>;
2090  defm: Loadxi_pat<zextloadv2i8,    v2i16, anyimm1, L2_loadbzw2_io>;
2091  defm: Loadxi_pat<zextloadv4i8,    v4i16, anyimm2, L2_loadbzw4_io>;
2092  defm: Loadxi_pat<load,            i32,   anyimm2, L2_loadri_io>;
2093  defm: Loadxi_pat<load,            v2i16, anyimm2, L2_loadri_io>;
2094  defm: Loadxi_pat<load,            v4i8,  anyimm2, L2_loadri_io>;
2095  defm: Loadxi_pat<load,            i64,   anyimm3, L2_loadrd_io>;
2096  defm: Loadxi_pat<load,            v2i32, anyimm3, L2_loadrd_io>;
2097  defm: Loadxi_pat<load,            v4i16, anyimm3, L2_loadrd_io>;
2098  defm: Loadxi_pat<load,            v8i8,  anyimm3, L2_loadrd_io>;
2099  defm: Loadxi_pat<load,            f32,   anyimm2, L2_loadri_io>;
2100  defm: Loadxi_pat<load,            f64,   anyimm3, L2_loadrd_io>;
2101  // No sextloadi1.
2102
2103  defm: Loadxi_pat<atomic_load_8 ,  i32, anyimm0, L2_loadrub_io>;
2104  defm: Loadxi_pat<atomic_load_16,  i32, anyimm1, L2_loadruh_io>;
2105  defm: Loadxi_pat<atomic_load_32,  i32, anyimm2, L2_loadri_io>;
2106  defm: Loadxi_pat<atomic_load_64,  i64, anyimm3, L2_loadrd_io>;
2107}
2108
2109let AddedComplexity = 30 in {
2110  // Loads of i1 are loading a byte, and the byte should be either 0 or 1.
2111  // It doesn't matter if it's sign- or zero-extended, so use zero-extension
2112  // everywhere.
2113  defm: Loadxim_pat<sextloadi1,   i32, L1toI32,  anyimm0, L2_loadrub_io>;
2114  defm: Loadxim_pat<extloadi1,    i64, ToAext64, anyimm0, L2_loadrub_io>;
2115  defm: Loadxim_pat<sextloadi1,   i64, L1toI64,  anyimm0, L2_loadrub_io>;
2116  defm: Loadxim_pat<zextloadi1,   i64, ToZext64, anyimm0, L2_loadrub_io>;
2117
2118  defm: Loadxim_pat<extloadi8,    i64, ToAext64, anyimm0, L2_loadrub_io>;
2119  defm: Loadxim_pat<extloadi16,   i64, ToAext64, anyimm1, L2_loadruh_io>;
2120  defm: Loadxim_pat<extloadi32,   i64, ToAext64, anyimm2, L2_loadri_io>;
2121  defm: Loadxim_pat<zextloadi8,   i64, ToZext64, anyimm0, L2_loadrub_io>;
2122  defm: Loadxim_pat<zextloadi16,  i64, ToZext64, anyimm1, L2_loadruh_io>;
2123  defm: Loadxim_pat<zextloadi32,  i64, ToZext64, anyimm2, L2_loadri_io>;
2124  defm: Loadxim_pat<sextloadi8,   i64, ToSext64, anyimm0, L2_loadrb_io>;
2125  defm: Loadxim_pat<sextloadi16,  i64, ToSext64, anyimm1, L2_loadrh_io>;
2126  defm: Loadxim_pat<sextloadi32,  i64, ToSext64, anyimm2, L2_loadri_io>;
2127}
2128
2129let AddedComplexity  = 60 in {
2130  def: Loadxu_pat<extloadi1,    i32,   anyimm0, L4_loadrub_ur>;
2131  def: Loadxu_pat<extloadi8,    i32,   anyimm0, L4_loadrub_ur>;
2132  def: Loadxu_pat<extloadi16,   i32,   anyimm1, L4_loadruh_ur>;
2133  def: Loadxu_pat<extloadv2i8,  v2i16, anyimm1, L4_loadbzw2_ur>;
2134  def: Loadxu_pat<extloadv4i8,  v4i16, anyimm2, L4_loadbzw4_ur>;
2135  def: Loadxu_pat<sextloadi8,   i32,   anyimm0, L4_loadrb_ur>;
2136  def: Loadxu_pat<sextloadi16,  i32,   anyimm1, L4_loadrh_ur>;
2137  def: Loadxu_pat<sextloadv2i8, v2i16, anyimm1, L4_loadbsw2_ur>;
2138  def: Loadxu_pat<sextloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
2139  def: Loadxu_pat<zextloadi1,   i32,   anyimm0, L4_loadrub_ur>;
2140  def: Loadxu_pat<zextloadi8,   i32,   anyimm0, L4_loadrub_ur>;
2141  def: Loadxu_pat<zextloadi16,  i32,   anyimm1, L4_loadruh_ur>;
2142  def: Loadxu_pat<zextloadv2i8, v2i16, anyimm1, L4_loadbzw2_ur>;
2143  def: Loadxu_pat<zextloadv4i8, v4i16, anyimm2, L4_loadbzw4_ur>;
2144  def: Loadxu_pat<load,         i32,   anyimm2, L4_loadri_ur>;
2145  def: Loadxu_pat<load,         v2i16, anyimm2, L4_loadri_ur>;
2146  def: Loadxu_pat<load,         v4i8,  anyimm2, L4_loadri_ur>;
2147  def: Loadxu_pat<load,         i64,   anyimm3, L4_loadrd_ur>;
2148  def: Loadxu_pat<load,         v2i32, anyimm3, L4_loadrd_ur>;
2149  def: Loadxu_pat<load,         v4i16, anyimm3, L4_loadrd_ur>;
2150  def: Loadxu_pat<load,         v8i8,  anyimm3, L4_loadrd_ur>;
2151  def: Loadxu_pat<load,         f32,   anyimm2, L4_loadri_ur>;
2152  def: Loadxu_pat<load,         f64,   anyimm3, L4_loadrd_ur>;
2153
2154  def: Loadxum_pat<sextloadi1,  i32, anyimm0, L1toI32,  L4_loadrub_ur>;
2155  def: Loadxum_pat<extloadi1,   i64, anyimm0, ToAext64, L4_loadrub_ur>;
2156  def: Loadxum_pat<sextloadi1,  i64, anyimm0, L1toI64,  L4_loadrub_ur>;
2157  def: Loadxum_pat<zextloadi1,  i64, anyimm0, ToZext64, L4_loadrub_ur>;
2158
2159  def: Loadxum_pat<sextloadi8,  i64, anyimm0, ToSext64, L4_loadrb_ur>;
2160  def: Loadxum_pat<zextloadi8,  i64, anyimm0, ToZext64, L4_loadrub_ur>;
2161  def: Loadxum_pat<extloadi8,   i64, anyimm0, ToAext64, L4_loadrub_ur>;
2162  def: Loadxum_pat<sextloadi16, i64, anyimm1, ToSext64, L4_loadrh_ur>;
2163  def: Loadxum_pat<zextloadi16, i64, anyimm1, ToZext64, L4_loadruh_ur>;
2164  def: Loadxum_pat<extloadi16,  i64, anyimm1, ToAext64, L4_loadruh_ur>;
2165  def: Loadxum_pat<sextloadi32, i64, anyimm2, ToSext64, L4_loadri_ur>;
2166  def: Loadxum_pat<zextloadi32, i64, anyimm2, ToZext64, L4_loadri_ur>;
2167  def: Loadxum_pat<extloadi32,  i64, anyimm2, ToAext64, L4_loadri_ur>;
2168}
2169
2170let AddedComplexity = 40 in {
2171  def: Loadxr_shl_pat<extloadi1,     i32,   L4_loadrub_rr>;
2172  def: Loadxr_shl_pat<extloadi8,     i32,   L4_loadrub_rr>;
2173  def: Loadxr_shl_pat<zextloadi1,    i32,   L4_loadrub_rr>;
2174  def: Loadxr_shl_pat<zextloadi8,    i32,   L4_loadrub_rr>;
2175  def: Loadxr_shl_pat<sextloadi8,    i32,   L4_loadrb_rr>;
2176  def: Loadxr_shl_pat<extloadi16,    i32,   L4_loadruh_rr>;
2177  def: Loadxr_shl_pat<zextloadi16,   i32,   L4_loadruh_rr>;
2178  def: Loadxr_shl_pat<sextloadi16,   i32,   L4_loadrh_rr>;
2179  def: Loadxr_shl_pat<load,          i32,   L4_loadri_rr>;
2180  def: Loadxr_shl_pat<load,          v2i16, L4_loadri_rr>;
2181  def: Loadxr_shl_pat<load,          v4i8,  L4_loadri_rr>;
2182  def: Loadxr_shl_pat<load,          i64,   L4_loadrd_rr>;
2183  def: Loadxr_shl_pat<load,          v2i32, L4_loadrd_rr>;
2184  def: Loadxr_shl_pat<load,          v4i16, L4_loadrd_rr>;
2185  def: Loadxr_shl_pat<load,          v8i8,  L4_loadrd_rr>;
2186  def: Loadxr_shl_pat<load,          f32,   L4_loadri_rr>;
2187  def: Loadxr_shl_pat<load,          f64,   L4_loadrd_rr>;
2188}
2189
2190let AddedComplexity = 20 in {
2191  def: Loadxr_add_pat<extloadi1,     i32,   L4_loadrub_rr>;
2192  def: Loadxr_add_pat<extloadi8,     i32,   L4_loadrub_rr>;
2193  def: Loadxr_add_pat<zextloadi8,    i32,   L4_loadrub_rr>;
2194  def: Loadxr_add_pat<zextloadi1,    i32,   L4_loadrub_rr>;
2195  def: Loadxr_add_pat<sextloadi8,    i32,   L4_loadrb_rr>;
2196  def: Loadxr_add_pat<extloadi16,    i32,   L4_loadruh_rr>;
2197  def: Loadxr_add_pat<zextloadi16,   i32,   L4_loadruh_rr>;
2198  def: Loadxr_add_pat<sextloadi16,   i32,   L4_loadrh_rr>;
2199  def: Loadxr_add_pat<load,          i32,   L4_loadri_rr>;
2200  def: Loadxr_add_pat<load,          v2i16, L4_loadri_rr>;
2201  def: Loadxr_add_pat<load,          v4i8,  L4_loadri_rr>;
2202  def: Loadxr_add_pat<load,          i64,   L4_loadrd_rr>;
2203  def: Loadxr_add_pat<load,          v2i32, L4_loadrd_rr>;
2204  def: Loadxr_add_pat<load,          v4i16, L4_loadrd_rr>;
2205  def: Loadxr_add_pat<load,          v8i8,  L4_loadrd_rr>;
2206  def: Loadxr_add_pat<load,          f32,   L4_loadri_rr>;
2207  def: Loadxr_add_pat<load,          f64,   L4_loadrd_rr>;
2208}
2209
2210let AddedComplexity = 40 in {
2211  def: Loadxrm_shl_pat<sextloadi1,   i32, L1toI32,  L4_loadrub_rr>;
2212  def: Loadxrm_shl_pat<extloadi1,    i64, ToAext64, L4_loadrub_rr>;
2213  def: Loadxrm_shl_pat<sextloadi1,   i64, L1toI64,  L4_loadrub_rr>;
2214  def: Loadxrm_shl_pat<zextloadi1,   i64, ToZext64, L4_loadrub_rr>;
2215
2216  def: Loadxrm_shl_pat<extloadi8,    i64, ToAext64, L4_loadrub_rr>;
2217  def: Loadxrm_shl_pat<zextloadi8,   i64, ToZext64, L4_loadrub_rr>;
2218  def: Loadxrm_shl_pat<sextloadi8,   i64, ToSext64, L4_loadrb_rr>;
2219  def: Loadxrm_shl_pat<extloadi16,   i64, ToAext64, L4_loadruh_rr>;
2220  def: Loadxrm_shl_pat<zextloadi16,  i64, ToZext64, L4_loadruh_rr>;
2221  def: Loadxrm_shl_pat<sextloadi16,  i64, ToSext64, L4_loadrh_rr>;
2222  def: Loadxrm_shl_pat<extloadi32,   i64, ToAext64, L4_loadri_rr>;
2223  def: Loadxrm_shl_pat<zextloadi32,  i64, ToZext64, L4_loadri_rr>;
2224  def: Loadxrm_shl_pat<sextloadi32,  i64, ToSext64, L4_loadri_rr>;
2225}
2226
2227let AddedComplexity = 30 in {
2228  def: Loadxrm_add_pat<sextloadi1,   i32, L1toI32,  L4_loadrub_rr>;
2229  def: Loadxrm_add_pat<extloadi1,    i64, ToAext64, L4_loadrub_rr>;
2230  def: Loadxrm_add_pat<sextloadi1,   i64, L1toI64,  L4_loadrub_rr>;
2231  def: Loadxrm_add_pat<zextloadi1,   i64, ToZext64, L4_loadrub_rr>;
2232
2233  def: Loadxrm_add_pat<extloadi8,    i64, ToAext64, L4_loadrub_rr>;
2234  def: Loadxrm_add_pat<zextloadi8,   i64, ToZext64, L4_loadrub_rr>;
2235  def: Loadxrm_add_pat<sextloadi8,   i64, ToSext64, L4_loadrb_rr>;
2236  def: Loadxrm_add_pat<extloadi16,   i64, ToAext64, L4_loadruh_rr>;
2237  def: Loadxrm_add_pat<zextloadi16,  i64, ToZext64, L4_loadruh_rr>;
2238  def: Loadxrm_add_pat<sextloadi16,  i64, ToSext64, L4_loadrh_rr>;
2239  def: Loadxrm_add_pat<extloadi32,   i64, ToAext64, L4_loadri_rr>;
2240  def: Loadxrm_add_pat<zextloadi32,  i64, ToZext64, L4_loadri_rr>;
2241  def: Loadxrm_add_pat<sextloadi32,  i64, ToSext64, L4_loadri_rr>;
2242}
2243
2244// Absolute address
2245
2246let AddedComplexity  = 60 in {
2247  def: Loada_pat<extloadi1,       i32,   anyimm0, PS_loadrubabs>;
2248  def: Loada_pat<zextloadi1,      i32,   anyimm0, PS_loadrubabs>;
2249  def: Loada_pat<extloadi8,       i32,   anyimm0, PS_loadrubabs>;
2250  def: Loada_pat<sextloadi8,      i32,   anyimm0, PS_loadrbabs>;
2251  def: Loada_pat<zextloadi8,      i32,   anyimm0, PS_loadrubabs>;
2252  def: Loada_pat<extloadi16,      i32,   anyimm1, PS_loadruhabs>;
2253  def: Loada_pat<sextloadi16,     i32,   anyimm1, PS_loadrhabs>;
2254  def: Loada_pat<zextloadi16,     i32,   anyimm1, PS_loadruhabs>;
2255  def: Loada_pat<load,            i32,   anyimm2, PS_loadriabs>;
2256  def: Loada_pat<load,            v2i16, anyimm2, PS_loadriabs>;
2257  def: Loada_pat<load,            v4i8,  anyimm2, PS_loadriabs>;
2258  def: Loada_pat<load,            i64,   anyimm3, PS_loadrdabs>;
2259  def: Loada_pat<load,            v2i32, anyimm3, PS_loadrdabs>;
2260  def: Loada_pat<load,            v4i16, anyimm3, PS_loadrdabs>;
2261  def: Loada_pat<load,            v8i8,  anyimm3, PS_loadrdabs>;
2262  def: Loada_pat<load,            f32,   anyimm2, PS_loadriabs>;
2263  def: Loada_pat<load,            f64,   anyimm3, PS_loadrdabs>;
2264
2265  def: Loada_pat<atomic_load_8,   i32, anyimm0, PS_loadrubabs>;
2266  def: Loada_pat<atomic_load_16,  i32, anyimm1, PS_loadruhabs>;
2267  def: Loada_pat<atomic_load_32,  i32, anyimm2, PS_loadriabs>;
2268  def: Loada_pat<atomic_load_64,  i64, anyimm3, PS_loadrdabs>;
2269}
2270
2271let AddedComplexity  = 30 in {
2272  def: Loadam_pat<load,           i1,  anyimm0, I32toI1,  PS_loadrubabs>;
2273  def: Loadam_pat<sextloadi1,     i32, anyimm0, L1toI32,  PS_loadrubabs>;
2274  def: Loadam_pat<extloadi1,      i64, anyimm0, ToZext64, PS_loadrubabs>;
2275  def: Loadam_pat<sextloadi1,     i64, anyimm0, L1toI64,  PS_loadrubabs>;
2276  def: Loadam_pat<zextloadi1,     i64, anyimm0, ToZext64, PS_loadrubabs>;
2277
2278  def: Loadam_pat<extloadi8,      i64, anyimm0, ToAext64, PS_loadrubabs>;
2279  def: Loadam_pat<sextloadi8,     i64, anyimm0, ToSext64, PS_loadrbabs>;
2280  def: Loadam_pat<zextloadi8,     i64, anyimm0, ToZext64, PS_loadrubabs>;
2281  def: Loadam_pat<extloadi16,     i64, anyimm1, ToAext64, PS_loadruhabs>;
2282  def: Loadam_pat<sextloadi16,    i64, anyimm1, ToSext64, PS_loadrhabs>;
2283  def: Loadam_pat<zextloadi16,    i64, anyimm1, ToZext64, PS_loadruhabs>;
2284  def: Loadam_pat<extloadi32,     i64, anyimm2, ToAext64, PS_loadriabs>;
2285  def: Loadam_pat<sextloadi32,    i64, anyimm2, ToSext64, PS_loadriabs>;
2286  def: Loadam_pat<zextloadi32,    i64, anyimm2, ToZext64, PS_loadriabs>;
2287}
2288
2289// GP-relative address
2290
2291let AddedComplexity  = 100 in {
2292  def: Loada_pat<extloadi1,       i32,   addrgp,  L2_loadrubgp>;
2293  def: Loada_pat<zextloadi1,      i32,   addrgp,  L2_loadrubgp>;
2294  def: Loada_pat<extloadi8,       i32,   addrgp,  L2_loadrubgp>;
2295  def: Loada_pat<sextloadi8,      i32,   addrgp,  L2_loadrbgp>;
2296  def: Loada_pat<zextloadi8,      i32,   addrgp,  L2_loadrubgp>;
2297  def: Loada_pat<extloadi16,      i32,   addrgp,  L2_loadruhgp>;
2298  def: Loada_pat<sextloadi16,     i32,   addrgp,  L2_loadrhgp>;
2299  def: Loada_pat<zextloadi16,     i32,   addrgp,  L2_loadruhgp>;
2300  def: Loada_pat<load,            i32,   addrgp,  L2_loadrigp>;
2301  def: Loada_pat<load,            v2i16, addrgp,  L2_loadrigp>;
2302  def: Loada_pat<load,            v4i8,  addrgp,  L2_loadrigp>;
2303  def: Loada_pat<load,            i64,   addrgp,  L2_loadrdgp>;
2304  def: Loada_pat<load,            v2i32, addrgp,  L2_loadrdgp>;
2305  def: Loada_pat<load,            v4i16, addrgp,  L2_loadrdgp>;
2306  def: Loada_pat<load,            v8i8,  addrgp,  L2_loadrdgp>;
2307  def: Loada_pat<load,            f32,   addrgp,  L2_loadrigp>;
2308  def: Loada_pat<load,            f64,   addrgp,  L2_loadrdgp>;
2309
2310  def: Loada_pat<atomic_load_8,   i32, addrgp,  L2_loadrubgp>;
2311  def: Loada_pat<atomic_load_16,  i32, addrgp,  L2_loadruhgp>;
2312  def: Loada_pat<atomic_load_32,  i32, addrgp,  L2_loadrigp>;
2313  def: Loada_pat<atomic_load_64,  i64, addrgp,  L2_loadrdgp>;
2314}
2315
2316let AddedComplexity  = 70 in {
2317  def: Loadam_pat<sextloadi1,     i32, addrgp,  L1toI32,  L2_loadrubgp>;
2318  def: Loadam_pat<extloadi1,      i64, addrgp,  ToAext64, L2_loadrubgp>;
2319  def: Loadam_pat<sextloadi1,     i64, addrgp,  L1toI64,  L2_loadrubgp>;
2320  def: Loadam_pat<zextloadi1,     i64, addrgp,  ToZext64, L2_loadrubgp>;
2321
2322  def: Loadam_pat<extloadi8,      i64, addrgp,  ToAext64, L2_loadrubgp>;
2323  def: Loadam_pat<sextloadi8,     i64, addrgp,  ToSext64, L2_loadrbgp>;
2324  def: Loadam_pat<zextloadi8,     i64, addrgp,  ToZext64, L2_loadrubgp>;
2325  def: Loadam_pat<extloadi16,     i64, addrgp,  ToAext64, L2_loadruhgp>;
2326  def: Loadam_pat<sextloadi16,    i64, addrgp,  ToSext64, L2_loadrhgp>;
2327  def: Loadam_pat<zextloadi16,    i64, addrgp,  ToZext64, L2_loadruhgp>;
2328  def: Loadam_pat<extloadi32,     i64, addrgp,  ToAext64, L2_loadrigp>;
2329  def: Loadam_pat<sextloadi32,    i64, addrgp,  ToSext64, L2_loadrigp>;
2330  def: Loadam_pat<zextloadi32,    i64, addrgp,  ToZext64, L2_loadrigp>;
2331
2332  def: Loadam_pat<load,           i1,  addrgp,  I32toI1,  L2_loadrubgp>;
2333}
2334
2335// Patterns for loads of i1:
2336def: Pat<(i1 (load AddrFI:$fi)),
2337         (C2_tfrrp (L2_loadrub_io AddrFI:$fi, 0))>;
2338def: Pat<(i1 (load (add I32:$Rs, anyimm0:$Off))),
2339         (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, imm:$Off))>;
2340def: Pat<(i1 (load I32:$Rs)),
2341         (C2_tfrrp (L2_loadrub_io IntRegs:$Rs, 0))>;
2342
2343
2344// --(13) Store ----------------------------------------------------------
2345//
2346
2347class Storepi_pat<PatFrag Store, PatFrag Value, PatFrag Offset, InstHexagon MI>
2348  : Pat<(Store Value:$Rt, I32:$Rx, Offset:$s4),
2349        (MI I32:$Rx, imm:$s4, Value:$Rt)>;
2350
2351def: Storepi_pat<post_truncsti8,  I32, s4_0ImmPred, S2_storerb_pi>;
2352def: Storepi_pat<post_truncsti16, I32, s4_1ImmPred, S2_storerh_pi>;
2353def: Storepi_pat<post_store,      I32, s4_2ImmPred, S2_storeri_pi>;
2354def: Storepi_pat<post_store,      I64, s4_3ImmPred, S2_storerd_pi>;
2355
2356// Patterns for generating stores, where the address takes different forms:
2357// - frameindex,
2358// - frameindex + offset,
2359// - base + offset,
2360// - simple (base address without offset).
2361// These would usually be used together (via Storexi_pat defined below), but
2362// in some cases one may want to apply different properties (such as
2363// AddedComplexity) to the individual patterns.
2364class Storexi_fi_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2365  : Pat<(Store Value:$Rs, AddrFI:$fi), (MI AddrFI:$fi, 0, Value:$Rs)>;
2366
2367multiclass Storexi_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2368                              InstHexagon MI> {
2369  def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
2370           (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
2371  def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
2372           (MI AddrFI:$fi, imm:$Off, Value:$Rs)>;
2373}
2374
2375multiclass Storexi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2376                           InstHexagon MI> {
2377  def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
2378           (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
2379  def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
2380           (MI IntRegs:$Rs, imm:$Off, Value:$Rt)>;
2381}
2382
2383class Storexi_base_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2384  : Pat<(Store Value:$Rt, I32:$Rs),
2385        (MI IntRegs:$Rs, 0, Value:$Rt)>;
2386
2387// Patterns for generating stores, where the address takes different forms,
2388// and where the value being stored is transformed through the value modifier
2389// ValueMod.  The address forms are same as above.
2390class Storexim_fi_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2391                      InstHexagon MI>
2392  : Pat<(Store Value:$Rs, AddrFI:$fi),
2393        (MI AddrFI:$fi, 0, (ValueMod Value:$Rs))>;
2394
2395multiclass Storexim_fi_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2396                               PatFrag ValueMod, InstHexagon MI> {
2397  def: Pat<(Store Value:$Rs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
2398           (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
2399  def: Pat<(Store Value:$Rs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
2400           (MI AddrFI:$fi, imm:$Off, (ValueMod Value:$Rs))>;
2401}
2402
2403multiclass Storexim_add_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred,
2404                            PatFrag ValueMod, InstHexagon MI> {
2405  def: Pat<(Store Value:$Rt, (add I32:$Rs, ImmPred:$Off)),
2406           (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
2407  def: Pat<(Store Value:$Rt, (IsOrAdd I32:$Rs, ImmPred:$Off)),
2408           (MI IntRegs:$Rs, imm:$Off, (ValueMod Value:$Rt))>;
2409}
2410
2411class Storexim_base_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2412                        InstHexagon MI>
2413  : Pat<(Store Value:$Rt, I32:$Rs),
2414        (MI IntRegs:$Rs, 0, (ValueMod Value:$Rt))>;
2415
2416multiclass Storexi_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
2417                       InstHexagon MI> {
2418  defm: Storexi_fi_add_pat <Store, Value, ImmPred, MI>;
2419  def:  Storexi_fi_pat     <Store, Value,          MI>;
2420  defm: Storexi_add_pat    <Store, Value, ImmPred, MI>;
2421}
2422
2423multiclass Storexim_pat<PatFrag Store, PatFrag Value, PatLeaf ImmPred,
2424                        PatFrag ValueMod, InstHexagon MI> {
2425  defm: Storexim_fi_add_pat <Store, Value, ImmPred, ValueMod, MI>;
2426  def:  Storexim_fi_pat     <Store, Value,          ValueMod, MI>;
2427  defm: Storexim_add_pat    <Store, Value, ImmPred, ValueMod, MI>;
2428}
2429
2430// Reg<<S + Imm
2431class Storexu_shl_pat<PatFrag Store, PatFrag Value, PatFrag ImmPred, InstHexagon MI>
2432  : Pat<(Store Value:$Rt, (add (shl I32:$Ru, u2_0ImmPred:$u2), ImmPred:$A)),
2433        (MI IntRegs:$Ru, imm:$u2, ImmPred:$A, Value:$Rt)>;
2434
2435// Reg<<S + Reg
2436class Storexr_shl_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2437  : Pat<(Store Value:$Ru, (add I32:$Rs, (shl I32:$Rt, u2_0ImmPred:$u2))),
2438        (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2, Value:$Ru)>;
2439
2440// Reg + Reg
2441class Storexr_add_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2442  : Pat<(Store Value:$Ru, (add I32:$Rs, I32:$Rt)),
2443        (MI IntRegs:$Rs, IntRegs:$Rt, 0, Value:$Ru)>;
2444
2445class Storea_pat<PatFrag Store, PatFrag Value, PatFrag Addr, InstHexagon MI>
2446  : Pat<(Store Value:$val, Addr:$addr), (MI Addr:$addr, Value:$val)>;
2447
2448class Stoream_pat<PatFrag Store, PatFrag Value, PatFrag Addr, PatFrag ValueMod,
2449                  InstHexagon MI>
2450  : Pat<(Store Value:$val, Addr:$addr),
2451        (MI Addr:$addr, (ValueMod Value:$val))>;
2452
2453// Regular stores in the DAG have two operands: value and address.
2454// Atomic stores also have two, but they are reversed: address, value.
2455// To use atomic stores with the patterns, they need to have their operands
2456// swapped. This relies on the knowledge that the F.Fragment uses names
2457// "ptr" and "val".
2458class AtomSt<PatFrag F>
2459  : PatFrag<(ops node:$val, node:$ptr), !head(F.Fragments), F.PredicateCode,
2460            F.OperandTransform> {
2461  let IsAtomic = F.IsAtomic;
2462  let MemoryVT = F.MemoryVT;
2463}
2464
2465
2466def IMM_BYTE : SDNodeXForm<imm, [{
2467  // -1 can be represented as 255, etc.
2468  // assigning to a byte restores our desired signed value.
2469  int8_t imm = N->getSExtValue();
2470  return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2471}]>;
2472
2473def IMM_HALF : SDNodeXForm<imm, [{
2474  // -1 can be represented as 65535, etc.
2475  // assigning to a short restores our desired signed value.
2476  int16_t imm = N->getSExtValue();
2477  return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2478}]>;
2479
2480def IMM_WORD : SDNodeXForm<imm, [{
2481  // -1 can be represented as 4294967295, etc.
2482  // Currently, it's not doing this. But some optimization
2483  // might convert -1 to a large +ve number.
2484  // assigning to a word restores our desired signed value.
2485  int32_t imm = N->getSExtValue();
2486  return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32);
2487}]>;
2488
2489def ToImmByte : OutPatFrag<(ops node:$R), (IMM_BYTE $R)>;
2490def ToImmHalf : OutPatFrag<(ops node:$R), (IMM_HALF $R)>;
2491def ToImmWord : OutPatFrag<(ops node:$R), (IMM_WORD $R)>;
2492
2493// Even though the offset is not extendable in the store-immediate, we
2494// can still generate the fi# in the base address. If the final offset
2495// is not valid for the instruction, we will replace it with a scratch
2496// register.
2497class SmallStackStore<PatFrag Store>
2498  : PatFrag<(ops node:$Val, node:$Addr), (Store node:$Val, node:$Addr), [{
2499  return isSmallStackStore(cast<StoreSDNode>(N));
2500}]>;
2501
2502// This is the complement of SmallStackStore.
2503class LargeStackStore<PatFrag Store>
2504  : PatFrag<(ops node:$Val, node:$Addr), (Store node:$Val, node:$Addr), [{
2505  return !isSmallStackStore(cast<StoreSDNode>(N));
2506}]>;
2507
2508// Preferred addressing modes for various combinations of stored value
2509// and address computation.
2510// For stores where the address and value are both immediates, prefer
2511// store-immediate. The reason is that the constant-extender optimization
2512// can replace store-immediate with a store-register, but there is nothing
2513// to generate a store-immediate out of a store-register.
2514//
2515//         C     R     F    F+C   R+C   R+R   R<<S+C   R<<S+R
2516// --+-------+-----+-----+------+-----+-----+--------+--------
2517// C |   imm | imm | imm |  imm | imm |  rr |     ur |     rr
2518// R |  abs* |  io |  io |   io |  io |  rr |     ur |     rr
2519//
2520// (*) Absolute or GP-relative.
2521//
2522// Note that any expression can be matched by Reg. In particular, an immediate
2523// can always be placed in a register, so patterns checking for Imm should
2524// have a higher priority than the ones involving Reg that could also match.
2525// For example, *(p+4) could become r1=#4; memw(r0+r1<<#0) instead of the
2526// preferred memw(r0+#4). Similarly Reg+Imm or Reg+Reg should be tried before
2527// Reg alone.
2528//
2529// The order in which the different combinations are tried:
2530//
2531//         C     F     R    F+C   R+C   R+R   R<<S+C   R<<S+R
2532// --+-------+-----+-----+------+-----+-----+--------+--------
2533// C |     1 |   6 |   - |    5 |   9 |   - |      - |      -
2534// R |     2 |   8 |  12 |    7 |  10 |  11 |      3 |      4
2535
2536
2537// First, match the unusual case of doubleword store into Reg+Imm4, i.e.
2538// a store where the offset Imm4 is a multiple of 4, but not of 8. This
2539// implies that Reg is also a proper multiple of 4. To still generate a
2540// doubleword store, add 4 to Reg, and subtract 4 from the offset.
2541
2542def s30_2ProperPred  : PatLeaf<(i32 imm), [{
2543  int64_t v = (int64_t)N->getSExtValue();
2544  return isShiftedInt<30,2>(v) && !isShiftedInt<29,3>(v);
2545}]>;
2546def RoundTo8 : SDNodeXForm<imm, [{
2547  int32_t Imm = N->getSExtValue();
2548  return CurDAG->getTargetConstant(Imm & -8, SDLoc(N), MVT::i32);
2549}]>;
2550
2551let AddedComplexity = 150 in
2552def: Pat<(store I64:$Ru, (add I32:$Rs, s30_2ProperPred:$Off)),
2553         (S2_storerd_io (A2_addi I32:$Rs, 4), (RoundTo8 $Off), I64:$Ru)>;
2554
2555class Storexi_abs_pat<PatFrag Store, PatFrag Value, InstHexagon MI>
2556  : Pat<(Store Value:$val, anyimm:$addr),
2557        (MI (ToI32 $addr), 0, Value:$val)>;
2558class Storexim_abs_pat<PatFrag Store, PatFrag Value, PatFrag ValueMod,
2559                       InstHexagon MI>
2560  : Pat<(Store Value:$val, anyimm:$addr),
2561        (MI (ToI32 $addr), 0, (ValueMod Value:$val))>;
2562
2563let AddedComplexity = 140 in {
2564  def: Storexim_abs_pat<truncstorei8,  anyint, ToImmByte, S4_storeirb_io>;
2565  def: Storexim_abs_pat<truncstorei16, anyint, ToImmHalf, S4_storeirh_io>;
2566  def: Storexim_abs_pat<store,         anyint, ToImmWord, S4_storeiri_io>;
2567
2568  def: Storexi_abs_pat<truncstorei8,  anyimm, S4_storeirb_io>;
2569  def: Storexi_abs_pat<truncstorei16, anyimm, S4_storeirh_io>;
2570  def: Storexi_abs_pat<store,         anyimm, S4_storeiri_io>;
2571}
2572
2573// GP-relative address
2574let AddedComplexity = 120 in {
2575  def: Storea_pat<truncstorei8,               I32, addrgp, S2_storerbgp>;
2576  def: Storea_pat<truncstorei16,              I32, addrgp, S2_storerhgp>;
2577  def: Storea_pat<store,                      I32, addrgp, S2_storerigp>;
2578  def: Storea_pat<store,                     V4I8, addrgp, S2_storerigp>;
2579  def: Storea_pat<store,                    V2I16, addrgp, S2_storerigp>;
2580  def: Storea_pat<store,                      I64, addrgp, S2_storerdgp>;
2581  def: Storea_pat<store,                     V8I8, addrgp, S2_storerdgp>;
2582  def: Storea_pat<store,                    V4I16, addrgp, S2_storerdgp>;
2583  def: Storea_pat<store,                    V2I32, addrgp, S2_storerdgp>;
2584  def: Storea_pat<store,                      F32, addrgp, S2_storerigp>;
2585  def: Storea_pat<store,                      F64, addrgp, S2_storerdgp>;
2586  def: Storea_pat<AtomSt<atomic_store_8>,     I32, addrgp, S2_storerbgp>;
2587  def: Storea_pat<AtomSt<atomic_store_16>,    I32, addrgp, S2_storerhgp>;
2588  def: Storea_pat<AtomSt<atomic_store_32>,    I32, addrgp, S2_storerigp>;
2589  def: Storea_pat<AtomSt<atomic_store_32>,   V4I8, addrgp, S2_storerigp>;
2590  def: Storea_pat<AtomSt<atomic_store_32>,  V2I16, addrgp, S2_storerigp>;
2591  def: Storea_pat<AtomSt<atomic_store_64>,    I64, addrgp, S2_storerdgp>;
2592  def: Storea_pat<AtomSt<atomic_store_64>,   V8I8, addrgp, S2_storerdgp>;
2593  def: Storea_pat<AtomSt<atomic_store_64>,  V4I16, addrgp, S2_storerdgp>;
2594  def: Storea_pat<AtomSt<atomic_store_64>,  V2I32, addrgp, S2_storerdgp>;
2595
2596  def: Stoream_pat<truncstorei8,  I64, addrgp, LoReg,    S2_storerbgp>;
2597  def: Stoream_pat<truncstorei16, I64, addrgp, LoReg,    S2_storerhgp>;
2598  def: Stoream_pat<truncstorei32, I64, addrgp, LoReg,    S2_storerigp>;
2599  def: Stoream_pat<store,         I1,  addrgp, I1toI32,  S2_storerbgp>;
2600}
2601
2602// Absolute address
2603let AddedComplexity = 110 in {
2604  def: Storea_pat<truncstorei8,               I32, anyimm0, PS_storerbabs>;
2605  def: Storea_pat<truncstorei16,              I32, anyimm1, PS_storerhabs>;
2606  def: Storea_pat<store,                      I32, anyimm2, PS_storeriabs>;
2607  def: Storea_pat<store,                     V4I8, anyimm2, PS_storeriabs>;
2608  def: Storea_pat<store,                    V2I16, anyimm2, PS_storeriabs>;
2609  def: Storea_pat<store,                      I64, anyimm3, PS_storerdabs>;
2610  def: Storea_pat<store,                     V8I8, anyimm3, PS_storerdabs>;
2611  def: Storea_pat<store,                    V4I16, anyimm3, PS_storerdabs>;
2612  def: Storea_pat<store,                    V2I32, anyimm3, PS_storerdabs>;
2613  def: Storea_pat<store,                      F32, anyimm2, PS_storeriabs>;
2614  def: Storea_pat<store,                      F64, anyimm3, PS_storerdabs>;
2615  def: Storea_pat<AtomSt<atomic_store_8>,     I32, anyimm0, PS_storerbabs>;
2616  def: Storea_pat<AtomSt<atomic_store_16>,    I32, anyimm1, PS_storerhabs>;
2617  def: Storea_pat<AtomSt<atomic_store_32>,    I32, anyimm2, PS_storeriabs>;
2618  def: Storea_pat<AtomSt<atomic_store_32>,   V4I8, anyimm2, PS_storeriabs>;
2619  def: Storea_pat<AtomSt<atomic_store_32>,  V2I16, anyimm2, PS_storeriabs>;
2620  def: Storea_pat<AtomSt<atomic_store_64>,    I64, anyimm3, PS_storerdabs>;
2621  def: Storea_pat<AtomSt<atomic_store_64>,   V8I8, anyimm3, PS_storerdabs>;
2622  def: Storea_pat<AtomSt<atomic_store_64>,  V4I16, anyimm3, PS_storerdabs>;
2623  def: Storea_pat<AtomSt<atomic_store_64>,  V2I32, anyimm3, PS_storerdabs>;
2624
2625  def: Stoream_pat<truncstorei8,  I64, anyimm0, LoReg,    PS_storerbabs>;
2626  def: Stoream_pat<truncstorei16, I64, anyimm1, LoReg,    PS_storerhabs>;
2627  def: Stoream_pat<truncstorei32, I64, anyimm2, LoReg,    PS_storeriabs>;
2628  def: Stoream_pat<store,         I1,  anyimm0, I1toI32,  PS_storerbabs>;
2629}
2630
2631// Reg<<S + Imm
2632let AddedComplexity = 100 in {
2633  def: Storexu_shl_pat<truncstorei8,    I32, anyimm0, S4_storerb_ur>;
2634  def: Storexu_shl_pat<truncstorei16,   I32, anyimm1, S4_storerh_ur>;
2635  def: Storexu_shl_pat<store,           I32, anyimm2, S4_storeri_ur>;
2636  def: Storexu_shl_pat<store,          V4I8, anyimm2, S4_storeri_ur>;
2637  def: Storexu_shl_pat<store,         V2I16, anyimm2, S4_storeri_ur>;
2638  def: Storexu_shl_pat<store,           I64, anyimm3, S4_storerd_ur>;
2639  def: Storexu_shl_pat<store,          V8I8, anyimm3, S4_storerd_ur>;
2640  def: Storexu_shl_pat<store,         V4I16, anyimm3, S4_storerd_ur>;
2641  def: Storexu_shl_pat<store,         V2I32, anyimm3, S4_storerd_ur>;
2642  def: Storexu_shl_pat<store,           F32, anyimm2, S4_storeri_ur>;
2643  def: Storexu_shl_pat<store,           F64, anyimm3, S4_storerd_ur>;
2644
2645  def: Pat<(store I1:$Pu, (add (shl I32:$Rs, u2_0ImmPred:$u2), anyimm:$A)),
2646           (S4_storerb_ur IntRegs:$Rs, imm:$u2, imm:$A, (I1toI32 I1:$Pu))>;
2647}
2648
2649// Reg<<S + Reg
2650let AddedComplexity = 90 in {
2651  def: Storexr_shl_pat<truncstorei8,    I32, S4_storerb_rr>;
2652  def: Storexr_shl_pat<truncstorei16,   I32, S4_storerh_rr>;
2653  def: Storexr_shl_pat<store,           I32, S4_storeri_rr>;
2654  def: Storexr_shl_pat<store,          V4I8, S4_storeri_rr>;
2655  def: Storexr_shl_pat<store,         V2I16, S4_storeri_rr>;
2656  def: Storexr_shl_pat<store,           I64, S4_storerd_rr>;
2657  def: Storexr_shl_pat<store,          V8I8, S4_storerd_rr>;
2658  def: Storexr_shl_pat<store,         V4I16, S4_storerd_rr>;
2659  def: Storexr_shl_pat<store,         V2I32, S4_storerd_rr>;
2660  def: Storexr_shl_pat<store,           F32, S4_storeri_rr>;
2661  def: Storexr_shl_pat<store,           F64, S4_storerd_rr>;
2662
2663  def: Pat<(store I1:$Pu, (add (shl I32:$Rs, u2_0ImmPred:$u2), I32:$Rt)),
2664           (S4_storerb_ur IntRegs:$Rt, IntRegs:$Rs, imm:$u2, (I1toI32 I1:$Pu))>;
2665}
2666
2667class SS_<PatFrag F> : SmallStackStore<F>;
2668class LS_<PatFrag F> : LargeStackStore<F>;
2669
2670multiclass IMFA_<PatFrag S, PatFrag V, PatFrag O, PatFrag M, InstHexagon I> {
2671  defm: Storexim_fi_add_pat<S, V, O, M, I>;
2672}
2673multiclass IFA_<PatFrag S, PatFrag V, PatFrag O, InstHexagon I> {
2674  defm: Storexi_fi_add_pat<S, V, O, I>;
2675}
2676
2677// Fi+Imm, store-immediate
2678let AddedComplexity = 80 in {
2679  defm: IMFA_<SS_<truncstorei8>,  anyint, u6_0ImmPred, ToImmByte, S4_storeirb_io>;
2680  defm: IMFA_<SS_<truncstorei16>, anyint, u6_1ImmPred, ToImmHalf, S4_storeirh_io>;
2681  defm: IMFA_<SS_<store>,         anyint, u6_2ImmPred, ToImmWord, S4_storeiri_io>;
2682
2683  defm: IFA_<SS_<truncstorei8>,   anyimm, u6_0ImmPred, S4_storeirb_io>;
2684  defm: IFA_<SS_<truncstorei16>,  anyimm, u6_1ImmPred, S4_storeirh_io>;
2685  defm: IFA_<SS_<store>,          anyimm, u6_2ImmPred, S4_storeiri_io>;
2686
2687  // For large-stack stores, generate store-register (prefer explicit Fi
2688  // in the address).
2689  defm: IMFA_<LS_<truncstorei8>,   anyimm, u6_0ImmPred, ToI32, S2_storerb_io>;
2690  defm: IMFA_<LS_<truncstorei16>,  anyimm, u6_1ImmPred, ToI32, S2_storerh_io>;
2691  defm: IMFA_<LS_<store>,          anyimm, u6_2ImmPred, ToI32, S2_storeri_io>;
2692}
2693
2694// Fi, store-immediate
2695let AddedComplexity = 70 in {
2696  def: Storexim_fi_pat<SS_<truncstorei8>,  anyint, ToImmByte, S4_storeirb_io>;
2697  def: Storexim_fi_pat<SS_<truncstorei16>, anyint, ToImmHalf, S4_storeirh_io>;
2698  def: Storexim_fi_pat<SS_<store>,         anyint, ToImmWord, S4_storeiri_io>;
2699
2700  def: Storexi_fi_pat<SS_<truncstorei8>,   anyimm, S4_storeirb_io>;
2701  def: Storexi_fi_pat<SS_<truncstorei16>,  anyimm, S4_storeirh_io>;
2702  def: Storexi_fi_pat<SS_<store>,          anyimm, S4_storeiri_io>;
2703
2704  // For large-stack stores, generate store-register (prefer explicit Fi
2705  // in the address).
2706  def: Storexim_fi_pat<LS_<truncstorei8>,  anyimm, ToI32, S2_storerb_io>;
2707  def: Storexim_fi_pat<LS_<truncstorei16>, anyimm, ToI32, S2_storerh_io>;
2708  def: Storexim_fi_pat<LS_<store>,         anyimm, ToI32, S2_storeri_io>;
2709}
2710
2711// Fi+Imm, Fi, store-register
2712let AddedComplexity = 60 in {
2713  defm: Storexi_fi_add_pat<truncstorei8,    I32, anyimm, S2_storerb_io>;
2714  defm: Storexi_fi_add_pat<truncstorei16,   I32, anyimm, S2_storerh_io>;
2715  defm: Storexi_fi_add_pat<store,           I32, anyimm, S2_storeri_io>;
2716  defm: Storexi_fi_add_pat<store,          V4I8, anyimm, S2_storeri_io>;
2717  defm: Storexi_fi_add_pat<store,         V2I16, anyimm, S2_storeri_io>;
2718  defm: Storexi_fi_add_pat<store,           I64, anyimm, S2_storerd_io>;
2719  defm: Storexi_fi_add_pat<store,          V8I8, anyimm, S2_storerd_io>;
2720  defm: Storexi_fi_add_pat<store,         V4I16, anyimm, S2_storerd_io>;
2721  defm: Storexi_fi_add_pat<store,         V2I32, anyimm, S2_storerd_io>;
2722  defm: Storexi_fi_add_pat<store,           F32, anyimm, S2_storeri_io>;
2723  defm: Storexi_fi_add_pat<store,           F64, anyimm, S2_storerd_io>;
2724  defm: Storexim_fi_add_pat<store, I1, anyimm, I1toI32, S2_storerb_io>;
2725
2726  def: Storexi_fi_pat<truncstorei8,     I32, S2_storerb_io>;
2727  def: Storexi_fi_pat<truncstorei16,    I32, S2_storerh_io>;
2728  def: Storexi_fi_pat<store,            I32, S2_storeri_io>;
2729  def: Storexi_fi_pat<store,           V4I8, S2_storeri_io>;
2730  def: Storexi_fi_pat<store,          V2I16, S2_storeri_io>;
2731  def: Storexi_fi_pat<store,            I64, S2_storerd_io>;
2732  def: Storexi_fi_pat<store,           V8I8, S2_storerd_io>;
2733  def: Storexi_fi_pat<store,          V4I16, S2_storerd_io>;
2734  def: Storexi_fi_pat<store,          V2I32, S2_storerd_io>;
2735  def: Storexi_fi_pat<store,            F32, S2_storeri_io>;
2736  def: Storexi_fi_pat<store,            F64, S2_storerd_io>;
2737  def: Storexim_fi_pat<store, I1, I1toI32, S2_storerb_io>;
2738}
2739
2740
2741multiclass IMRA_<PatFrag S, PatFrag V, PatFrag O, PatFrag M, InstHexagon I> {
2742  defm: Storexim_add_pat<S, V, O, M, I>;
2743}
2744multiclass IRA_<PatFrag S, PatFrag V, PatFrag O, InstHexagon I> {
2745  defm: Storexi_add_pat<S, V, O, I>;
2746}
2747
2748// Reg+Imm, store-immediate
2749let AddedComplexity = 50 in {
2750  defm: IMRA_<truncstorei8,   anyint, u6_0ImmPred, ToImmByte, S4_storeirb_io>;
2751  defm: IMRA_<truncstorei16,  anyint, u6_1ImmPred, ToImmHalf, S4_storeirh_io>;
2752  defm: IMRA_<store,          anyint, u6_2ImmPred, ToImmWord, S4_storeiri_io>;
2753
2754  defm: IRA_<truncstorei8,    anyimm, u6_0ImmPred, S4_storeirb_io>;
2755  defm: IRA_<truncstorei16,   anyimm, u6_1ImmPred, S4_storeirh_io>;
2756  defm: IRA_<store,           anyimm, u6_2ImmPred, S4_storeiri_io>;
2757}
2758
2759// Reg+Imm, store-register
2760let AddedComplexity = 40 in {
2761  defm: Storexi_pat<truncstorei8,     I32, anyimm0, S2_storerb_io>;
2762  defm: Storexi_pat<truncstorei16,    I32, anyimm1, S2_storerh_io>;
2763  defm: Storexi_pat<store,            I32, anyimm2, S2_storeri_io>;
2764  defm: Storexi_pat<store,           V4I8, anyimm2, S2_storeri_io>;
2765  defm: Storexi_pat<store,          V2I16, anyimm2, S2_storeri_io>;
2766  defm: Storexi_pat<store,            I64, anyimm3, S2_storerd_io>;
2767  defm: Storexi_pat<store,           V8I8, anyimm3, S2_storerd_io>;
2768  defm: Storexi_pat<store,          V4I16, anyimm3, S2_storerd_io>;
2769  defm: Storexi_pat<store,          V2I32, anyimm3, S2_storerd_io>;
2770  defm: Storexi_pat<store,            F32, anyimm2, S2_storeri_io>;
2771  defm: Storexi_pat<store,            F64, anyimm3, S2_storerd_io>;
2772
2773  defm: Storexim_pat<truncstorei8,  I64, anyimm0, LoReg,   S2_storerb_io>;
2774  defm: Storexim_pat<truncstorei16, I64, anyimm1, LoReg,   S2_storerh_io>;
2775  defm: Storexim_pat<truncstorei32, I64, anyimm2, LoReg,   S2_storeri_io>;
2776  defm: Storexim_pat<store,         I1,  anyimm0, I1toI32, S2_storerb_io>;
2777
2778  defm: Storexi_pat<AtomSt<atomic_store_8>,     I32, anyimm0, S2_storerb_io>;
2779  defm: Storexi_pat<AtomSt<atomic_store_16>,    I32, anyimm1, S2_storerh_io>;
2780  defm: Storexi_pat<AtomSt<atomic_store_32>,    I32, anyimm2, S2_storeri_io>;
2781  defm: Storexi_pat<AtomSt<atomic_store_32>,   V4I8, anyimm2, S2_storeri_io>;
2782  defm: Storexi_pat<AtomSt<atomic_store_32>,  V2I16, anyimm2, S2_storeri_io>;
2783  defm: Storexi_pat<AtomSt<atomic_store_64>,    I64, anyimm3, S2_storerd_io>;
2784  defm: Storexi_pat<AtomSt<atomic_store_64>,   V8I8, anyimm3, S2_storerd_io>;
2785  defm: Storexi_pat<AtomSt<atomic_store_64>,  V4I16, anyimm3, S2_storerd_io>;
2786  defm: Storexi_pat<AtomSt<atomic_store_64>,  V2I32, anyimm3, S2_storerd_io>;
2787}
2788
2789// Reg+Reg
2790let AddedComplexity = 30 in {
2791  def: Storexr_add_pat<truncstorei8,    I32, S4_storerb_rr>;
2792  def: Storexr_add_pat<truncstorei16,   I32, S4_storerh_rr>;
2793  def: Storexr_add_pat<store,           I32, S4_storeri_rr>;
2794  def: Storexr_add_pat<store,          V4I8, S4_storeri_rr>;
2795  def: Storexr_add_pat<store,         V2I16, S4_storeri_rr>;
2796  def: Storexr_add_pat<store,           I64, S4_storerd_rr>;
2797  def: Storexr_add_pat<store,          V8I8, S4_storerd_rr>;
2798  def: Storexr_add_pat<store,         V4I16, S4_storerd_rr>;
2799  def: Storexr_add_pat<store,         V2I32, S4_storerd_rr>;
2800  def: Storexr_add_pat<store,           F32, S4_storeri_rr>;
2801  def: Storexr_add_pat<store,           F64, S4_storerd_rr>;
2802
2803  def: Pat<(store I1:$Pu, (add I32:$Rs, I32:$Rt)),
2804           (S4_storerb_rr IntRegs:$Rs, IntRegs:$Rt, 0, (I1toI32 I1:$Pu))>;
2805}
2806
2807// Reg, store-immediate
2808let AddedComplexity = 20 in {
2809  def: Storexim_base_pat<truncstorei8,  anyint, ToImmByte, S4_storeirb_io>;
2810  def: Storexim_base_pat<truncstorei16, anyint, ToImmHalf, S4_storeirh_io>;
2811  def: Storexim_base_pat<store,         anyint, ToImmWord, S4_storeiri_io>;
2812
2813  def: Storexi_base_pat<truncstorei8,   anyimm, S4_storeirb_io>;
2814  def: Storexi_base_pat<truncstorei16,  anyimm, S4_storeirh_io>;
2815  def: Storexi_base_pat<store,          anyimm, S4_storeiri_io>;
2816}
2817
2818// Reg, store-register
2819let AddedComplexity = 10 in {
2820  def: Storexi_base_pat<truncstorei8,     I32, S2_storerb_io>;
2821  def: Storexi_base_pat<truncstorei16,    I32, S2_storerh_io>;
2822  def: Storexi_base_pat<store,            I32, S2_storeri_io>;
2823  def: Storexi_base_pat<store,           V4I8, S2_storeri_io>;
2824  def: Storexi_base_pat<store,          V2I16, S2_storeri_io>;
2825  def: Storexi_base_pat<store,            I64, S2_storerd_io>;
2826  def: Storexi_base_pat<store,           V8I8, S2_storerd_io>;
2827  def: Storexi_base_pat<store,          V4I16, S2_storerd_io>;
2828  def: Storexi_base_pat<store,          V2I32, S2_storerd_io>;
2829  def: Storexi_base_pat<store,            F32, S2_storeri_io>;
2830  def: Storexi_base_pat<store,            F64, S2_storerd_io>;
2831
2832  def: Storexim_base_pat<truncstorei8,  I64, LoReg,   S2_storerb_io>;
2833  def: Storexim_base_pat<truncstorei16, I64, LoReg,   S2_storerh_io>;
2834  def: Storexim_base_pat<truncstorei32, I64, LoReg,   S2_storeri_io>;
2835  def: Storexim_base_pat<store,         I1,  I1toI32, S2_storerb_io>;
2836
2837  def: Storexi_base_pat<AtomSt<atomic_store_8>,     I32, S2_storerb_io>;
2838  def: Storexi_base_pat<AtomSt<atomic_store_16>,    I32, S2_storerh_io>;
2839  def: Storexi_base_pat<AtomSt<atomic_store_32>,    I32, S2_storeri_io>;
2840  def: Storexi_base_pat<AtomSt<atomic_store_32>,   V4I8, S2_storeri_io>;
2841  def: Storexi_base_pat<AtomSt<atomic_store_32>,  V2I16, S2_storeri_io>;
2842  def: Storexi_base_pat<AtomSt<atomic_store_64>,    I64, S2_storerd_io>;
2843  def: Storexi_base_pat<AtomSt<atomic_store_64>,   V8I8, S2_storerd_io>;
2844  def: Storexi_base_pat<AtomSt<atomic_store_64>,  V4I16, S2_storerd_io>;
2845  def: Storexi_base_pat<AtomSt<atomic_store_64>,  V2I32, S2_storerd_io>;
2846}
2847
2848
2849// --(14) Memop ----------------------------------------------------------
2850//
2851
2852def m5_0Imm8Pred : PatLeaf<(i32 imm), [{
2853  int8_t V = N->getSExtValue();
2854  return -32 < V && V <= -1;
2855}]>;
2856
2857def m5_0Imm16Pred : PatLeaf<(i32 imm), [{
2858  int16_t V = N->getSExtValue();
2859  return -32 < V && V <= -1;
2860}]>;
2861
2862def m5_0ImmPred  : PatLeaf<(i32 imm), [{
2863  int64_t V = N->getSExtValue();
2864  return -31 <= V && V <= -1;
2865}]>;
2866
2867def IsNPow2_8 : PatLeaf<(i32 imm), [{
2868  uint8_t NV = ~N->getZExtValue();
2869  return isPowerOf2_32(NV);
2870}]>;
2871
2872def IsNPow2_16 : PatLeaf<(i32 imm), [{
2873  uint16_t NV = ~N->getZExtValue();
2874  return isPowerOf2_32(NV);
2875}]>;
2876
2877def Log2_8 : SDNodeXForm<imm, [{
2878  uint8_t V = N->getZExtValue();
2879  return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
2880}]>;
2881
2882def Log2_16 : SDNodeXForm<imm, [{
2883  uint16_t V = N->getZExtValue();
2884  return CurDAG->getTargetConstant(Log2_32(V), SDLoc(N), MVT::i32);
2885}]>;
2886
2887def LogN2_8 : SDNodeXForm<imm, [{
2888  uint8_t NV = ~N->getZExtValue();
2889  return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
2890}]>;
2891
2892def LogN2_16 : SDNodeXForm<imm, [{
2893  uint16_t NV = ~N->getZExtValue();
2894  return CurDAG->getTargetConstant(Log2_32(NV), SDLoc(N), MVT::i32);
2895}]>;
2896
2897def IdImm : SDNodeXForm<imm, [{ return SDValue(N, 0); }]>;
2898
2899multiclass Memopxr_base_pat<PatFrag Load, PatFrag Store, SDNode Oper,
2900                            InstHexagon MI> {
2901  // Addr: i32
2902  def: Pat<(Store (Oper (Load I32:$Rs), I32:$A), I32:$Rs),
2903           (MI I32:$Rs, 0, I32:$A)>;
2904  // Addr: fi
2905  def: Pat<(Store (Oper (Load AddrFI:$Rs), I32:$A), AddrFI:$Rs),
2906           (MI AddrFI:$Rs, 0, I32:$A)>;
2907}
2908
2909multiclass Memopxr_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2910                           SDNode Oper, InstHexagon MI> {
2911  // Addr: i32
2912  def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), I32:$A),
2913                  (add I32:$Rs, ImmPred:$Off)),
2914           (MI I32:$Rs, imm:$Off, I32:$A)>;
2915  def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), I32:$A),
2916                  (IsOrAdd I32:$Rs, ImmPred:$Off)),
2917           (MI I32:$Rs, imm:$Off, I32:$A)>;
2918  // Addr: fi
2919  def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), I32:$A),
2920                  (add AddrFI:$Rs, ImmPred:$Off)),
2921           (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
2922  def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), I32:$A),
2923                  (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
2924           (MI AddrFI:$Rs, imm:$Off, I32:$A)>;
2925}
2926
2927multiclass Memopxr_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
2928                       SDNode Oper, InstHexagon MI> {
2929  let Predicates = [UseMEMOPS] in {
2930    defm: Memopxr_base_pat <Load, Store,          Oper, MI>;
2931    defm: Memopxr_add_pat  <Load, Store, ImmPred, Oper, MI>;
2932  }
2933}
2934
2935let AddedComplexity = 200 in {
2936  // add reg
2937  defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, add,
2938        /*anyext*/  L4_add_memopb_io>;
2939  defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, add,
2940        /*sext*/    L4_add_memopb_io>;
2941  defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, add,
2942        /*zext*/    L4_add_memopb_io>;
2943  defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, add,
2944        /*anyext*/  L4_add_memoph_io>;
2945  defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, add,
2946        /*sext*/    L4_add_memoph_io>;
2947  defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, add,
2948        /*zext*/    L4_add_memoph_io>;
2949  defm: Memopxr_pat<load, store, u6_2ImmPred, add, L4_add_memopw_io>;
2950
2951  // sub reg
2952  defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, sub,
2953        /*anyext*/  L4_sub_memopb_io>;
2954  defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub,
2955        /*sext*/    L4_sub_memopb_io>;
2956  defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub,
2957        /*zext*/    L4_sub_memopb_io>;
2958  defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, sub,
2959        /*anyext*/  L4_sub_memoph_io>;
2960  defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub,
2961        /*sext*/    L4_sub_memoph_io>;
2962  defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub,
2963        /*zext*/    L4_sub_memoph_io>;
2964  defm: Memopxr_pat<load, store, u6_2ImmPred, sub, L4_sub_memopw_io>;
2965
2966  // and reg
2967  defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, and,
2968        /*anyext*/  L4_and_memopb_io>;
2969  defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, and,
2970        /*sext*/    L4_and_memopb_io>;
2971  defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, and,
2972        /*zext*/    L4_and_memopb_io>;
2973  defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, and,
2974        /*anyext*/  L4_and_memoph_io>;
2975  defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, and,
2976        /*sext*/    L4_and_memoph_io>;
2977  defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, and,
2978        /*zext*/    L4_and_memoph_io>;
2979  defm: Memopxr_pat<load, store, u6_2ImmPred, and, L4_and_memopw_io>;
2980
2981  // or reg
2982  defm: Memopxr_pat<extloadi8, truncstorei8, u6_0ImmPred, or,
2983        /*anyext*/  L4_or_memopb_io>;
2984  defm: Memopxr_pat<sextloadi8, truncstorei8, u6_0ImmPred, or,
2985        /*sext*/    L4_or_memopb_io>;
2986  defm: Memopxr_pat<zextloadi8, truncstorei8, u6_0ImmPred, or,
2987        /*zext*/    L4_or_memopb_io>;
2988  defm: Memopxr_pat<extloadi16, truncstorei16, u6_1ImmPred, or,
2989        /*anyext*/  L4_or_memoph_io>;
2990  defm: Memopxr_pat<sextloadi16, truncstorei16, u6_1ImmPred, or,
2991        /*sext*/    L4_or_memoph_io>;
2992  defm: Memopxr_pat<zextloadi16, truncstorei16, u6_1ImmPred, or,
2993        /*zext*/    L4_or_memoph_io>;
2994  defm: Memopxr_pat<load, store, u6_2ImmPred, or, L4_or_memopw_io>;
2995}
2996
2997
2998multiclass Memopxi_base_pat<PatFrag Load, PatFrag Store, SDNode Oper,
2999                            PatFrag Arg, SDNodeXForm ArgMod, InstHexagon MI> {
3000  // Addr: i32
3001  def: Pat<(Store (Oper (Load I32:$Rs), Arg:$A), I32:$Rs),
3002           (MI I32:$Rs, 0, (ArgMod Arg:$A))>;
3003  // Addr: fi
3004  def: Pat<(Store (Oper (Load AddrFI:$Rs), Arg:$A), AddrFI:$Rs),
3005           (MI AddrFI:$Rs, 0, (ArgMod Arg:$A))>;
3006}
3007
3008multiclass Memopxi_add_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
3009                           SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
3010                           InstHexagon MI> {
3011  // Addr: i32
3012  def: Pat<(Store (Oper (Load (add I32:$Rs, ImmPred:$Off)), Arg:$A),
3013                  (add I32:$Rs, ImmPred:$Off)),
3014           (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
3015  def: Pat<(Store (Oper (Load (IsOrAdd I32:$Rs, ImmPred:$Off)), Arg:$A),
3016                  (IsOrAdd I32:$Rs, ImmPred:$Off)),
3017           (MI I32:$Rs, imm:$Off, (ArgMod Arg:$A))>;
3018  // Addr: fi
3019  def: Pat<(Store (Oper (Load (add AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
3020                  (add AddrFI:$Rs, ImmPred:$Off)),
3021           (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
3022  def: Pat<(Store (Oper (Load (IsOrAdd AddrFI:$Rs, ImmPred:$Off)), Arg:$A),
3023                  (IsOrAdd AddrFI:$Rs, ImmPred:$Off)),
3024           (MI AddrFI:$Rs, imm:$Off, (ArgMod Arg:$A))>;
3025}
3026
3027multiclass Memopxi_pat<PatFrag Load, PatFrag Store, PatFrag ImmPred,
3028                       SDNode Oper, PatFrag Arg, SDNodeXForm ArgMod,
3029                       InstHexagon MI> {
3030  let Predicates = [UseMEMOPS] in {
3031    defm: Memopxi_base_pat <Load, Store,          Oper, Arg, ArgMod, MI>;
3032    defm: Memopxi_add_pat  <Load, Store, ImmPred, Oper, Arg, ArgMod, MI>;
3033  }
3034}
3035
3036let AddedComplexity = 220 in {
3037  // add imm
3038  defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
3039        /*anyext*/  IdImm, L4_iadd_memopb_io>;
3040  defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
3041        /*sext*/    IdImm, L4_iadd_memopb_io>;
3042  defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, u5_0ImmPred,
3043        /*zext*/    IdImm, L4_iadd_memopb_io>;
3044  defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
3045        /*anyext*/  IdImm, L4_iadd_memoph_io>;
3046  defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
3047        /*sext*/    IdImm, L4_iadd_memoph_io>;
3048  defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, u5_0ImmPred,
3049        /*zext*/    IdImm, L4_iadd_memoph_io>;
3050  defm: Memopxi_pat<load, store, u6_2ImmPred, add, u5_0ImmPred, IdImm,
3051                    L4_iadd_memopw_io>;
3052  defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
3053        /*anyext*/  NegImm8, L4_iadd_memopb_io>;
3054  defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
3055        /*sext*/    NegImm8, L4_iadd_memopb_io>;
3056  defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, m5_0Imm8Pred,
3057        /*zext*/    NegImm8, L4_iadd_memopb_io>;
3058  defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
3059        /*anyext*/  NegImm16, L4_iadd_memoph_io>;
3060  defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
3061        /*sext*/    NegImm16, L4_iadd_memoph_io>;
3062  defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, m5_0Imm16Pred,
3063        /*zext*/    NegImm16, L4_iadd_memoph_io>;
3064  defm: Memopxi_pat<load, store, u6_2ImmPred, sub, m5_0ImmPred, NegImm32,
3065                    L4_iadd_memopw_io>;
3066
3067  // sub imm
3068  defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
3069        /*anyext*/  IdImm, L4_isub_memopb_io>;
3070  defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
3071        /*sext*/    IdImm, L4_isub_memopb_io>;
3072  defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, sub, u5_0ImmPred,
3073        /*zext*/    IdImm, L4_isub_memopb_io>;
3074  defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
3075        /*anyext*/  IdImm, L4_isub_memoph_io>;
3076  defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
3077        /*sext*/    IdImm, L4_isub_memoph_io>;
3078  defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, sub, u5_0ImmPred,
3079        /*zext*/    IdImm, L4_isub_memoph_io>;
3080  defm: Memopxi_pat<load, store, u6_2ImmPred, sub, u5_0ImmPred, IdImm,
3081                    L4_isub_memopw_io>;
3082  defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
3083        /*anyext*/  NegImm8, L4_isub_memopb_io>;
3084  defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
3085        /*sext*/    NegImm8, L4_isub_memopb_io>;
3086  defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, add, m5_0Imm8Pred,
3087        /*zext*/    NegImm8, L4_isub_memopb_io>;
3088  defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
3089        /*anyext*/  NegImm16, L4_isub_memoph_io>;
3090  defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
3091        /*sext*/    NegImm16, L4_isub_memoph_io>;
3092  defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, add, m5_0Imm16Pred,
3093        /*zext*/    NegImm16, L4_isub_memoph_io>;
3094  defm: Memopxi_pat<load, store, u6_2ImmPred, add, m5_0ImmPred, NegImm32,
3095                    L4_isub_memopw_io>;
3096
3097  // clrbit imm
3098  defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
3099        /*anyext*/  LogN2_8, L4_iand_memopb_io>;
3100  defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
3101        /*sext*/    LogN2_8, L4_iand_memopb_io>;
3102  defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, and, IsNPow2_8,
3103        /*zext*/    LogN2_8, L4_iand_memopb_io>;
3104  defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
3105        /*anyext*/  LogN2_16, L4_iand_memoph_io>;
3106  defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
3107        /*sext*/    LogN2_16, L4_iand_memoph_io>;
3108  defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, and, IsNPow2_16,
3109        /*zext*/    LogN2_16, L4_iand_memoph_io>;
3110  defm: Memopxi_pat<load, store, u6_2ImmPred, and, IsNPow2_32,
3111		    LogN2_32, L4_iand_memopw_io>;
3112
3113  // setbit imm
3114  defm: Memopxi_pat<extloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3115        /*anyext*/  Log2_8, L4_ior_memopb_io>;
3116  defm: Memopxi_pat<sextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3117        /*sext*/    Log2_8, L4_ior_memopb_io>;
3118  defm: Memopxi_pat<zextloadi8, truncstorei8, u6_0ImmPred, or, IsPow2_32,
3119        /*zext*/    Log2_8, L4_ior_memopb_io>;
3120  defm: Memopxi_pat<extloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3121        /*anyext*/  Log2_16, L4_ior_memoph_io>;
3122  defm: Memopxi_pat<sextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3123        /*sext*/    Log2_16, L4_ior_memoph_io>;
3124  defm: Memopxi_pat<zextloadi16, truncstorei16, u6_1ImmPred, or, IsPow2_32,
3125        /*zext*/    Log2_16, L4_ior_memoph_io>;
3126  defm: Memopxi_pat<load, store, u6_2ImmPred, or, IsPow2_32,
3127		    Log2_32, L4_ior_memopw_io>;
3128}
3129
3130
3131// --(15) Call -----------------------------------------------------------
3132//
3133
3134// Pseudo instructions.
3135def SDT_SPCallSeqStart
3136  : SDCallSeqStart<[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3137def SDT_SPCallSeqEnd
3138  : SDCallSeqEnd<[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3139
3140def callseq_start: SDNode<"ISD::CALLSEQ_START", SDT_SPCallSeqStart,
3141                          [SDNPHasChain, SDNPOutGlue]>;
3142def callseq_end:   SDNode<"ISD::CALLSEQ_END",   SDT_SPCallSeqEnd,
3143                          [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
3144
3145def SDT_SPCall: SDTypeProfile<0, 1, [SDTCisVT<0, i32>]>;
3146
3147def HexagonTCRet: SDNode<"HexagonISD::TC_RETURN", SDT_SPCall,
3148                         [SDNPHasChain,  SDNPOptInGlue, SDNPVariadic]>;
3149def callv3: SDNode<"HexagonISD::CALL", SDT_SPCall,
3150                   [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
3151def callv3nr: SDNode<"HexagonISD::CALLnr", SDT_SPCall,
3152                     [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue, SDNPVariadic]>;
3153
3154def: Pat<(callseq_start timm:$amt, timm:$amt2),
3155         (ADJCALLSTACKDOWN imm:$amt, imm:$amt2)>;
3156def: Pat<(callseq_end timm:$amt1, timm:$amt2),
3157         (ADJCALLSTACKUP imm:$amt1, imm:$amt2)>;
3158
3159def: Pat<(HexagonTCRet tglobaladdr:$dst),   (PS_tailcall_i tglobaladdr:$dst)>;
3160def: Pat<(HexagonTCRet texternalsym:$dst),  (PS_tailcall_i texternalsym:$dst)>;
3161def: Pat<(HexagonTCRet I32:$dst),           (PS_tailcall_r I32:$dst)>;
3162
3163def: Pat<(callv3 I32:$dst),                 (J2_callr I32:$dst)>;
3164def: Pat<(callv3 tglobaladdr:$dst),         (J2_call tglobaladdr:$dst)>;
3165def: Pat<(callv3 texternalsym:$dst),        (J2_call texternalsym:$dst)>;
3166def: Pat<(callv3 tglobaltlsaddr:$dst),      (J2_call tglobaltlsaddr:$dst)>;
3167
3168def: Pat<(callv3nr I32:$dst),               (PS_callr_nr I32:$dst)>;
3169def: Pat<(callv3nr tglobaladdr:$dst),       (PS_call_nr tglobaladdr:$dst)>;
3170def: Pat<(callv3nr texternalsym:$dst),      (PS_call_nr texternalsym:$dst)>;
3171
3172def retflag : SDNode<"HexagonISD::RET_FLAG", SDTNone,
3173                     [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
3174def eh_return: SDNode<"HexagonISD::EH_RETURN", SDTNone, [SDNPHasChain]>;
3175
3176def: Pat<(retflag),   (PS_jmpret (i32 R31))>;
3177def: Pat<(eh_return), (EH_RETURN_JMPR (i32 R31))>;
3178
3179
3180// --(16) Branch ---------------------------------------------------------
3181//
3182
3183def: Pat<(br      bb:$dst),         (J2_jump  b30_2Imm:$dst)>;
3184def: Pat<(brind   I32:$dst),        (J2_jumpr I32:$dst)>;
3185
3186def: Pat<(brcond I1:$Pu, bb:$dst),
3187         (J2_jumpt I1:$Pu, bb:$dst)>;
3188def: Pat<(brcond (not I1:$Pu), bb:$dst),
3189         (J2_jumpf I1:$Pu, bb:$dst)>;
3190def: Pat<(brcond (i1 (setne I1:$Pu, -1)), bb:$dst),
3191         (J2_jumpf I1:$Pu, bb:$dst)>;
3192def: Pat<(brcond (i1 (seteq I1:$Pu, 0)), bb:$dst),
3193         (J2_jumpf I1:$Pu, bb:$dst)>;
3194def: Pat<(brcond (i1 (setne I1:$Pu, 0)), bb:$dst),
3195         (J2_jumpt I1:$Pu, bb:$dst)>;
3196
3197
3198// --(17) Misc -----------------------------------------------------------
3199
3200
3201// Generate code of the form 'C2_muxii(cmpbgtui(Rdd, C-1),0,1)'
3202// for C code of the form r = (c>='0' && c<='9') ? 1 : 0.
3203// The isdigit transformation relies on two 'clever' aspects:
3204// 1) The data type is unsigned which allows us to eliminate a zero test after
3205//    biasing the expression by 48. We are depending on the representation of
3206//    the unsigned types, and semantics.
3207// 2) The front end has converted <= 9 into < 10 on entry to LLVM.
3208//
3209// For the C code:
3210//   retval = (c >= '0' && c <= '9') ? 1 : 0;
3211// The code is transformed upstream of llvm into
3212//   retval = (c-48) < 10 ? 1 : 0;
3213
3214def u7_0PosImmPred : ImmLeaf<i32, [{
3215  // True if the immediate fits in an 7-bit unsigned field and is positive.
3216  return Imm > 0 && isUInt<7>(Imm);
3217}]>;
3218
3219let AddedComplexity = 139 in
3220def: Pat<(i32 (zext (i1 (setult (and I32:$Rs, 255), u7_0PosImmPred:$u7)))),
3221         (C2_muxii (A4_cmpbgtui IntRegs:$Rs, (UDEC1 imm:$u7)), 0, 1)>;
3222
3223let AddedComplexity = 100 in
3224def: Pat<(or (or (shl (HexagonINSERT (i32 (zextloadi8 (add I32:$b, 2))),
3225                                     (i32 (extloadi8  (add I32:$b, 3))),
3226                                     24, 8),
3227                      (i32 16)),
3228                 (shl (i32 (zextloadi8 (add I32:$b, 1))), (i32 8))),
3229             (zextloadi8 I32:$b)),
3230         (A2_swiz (L2_loadri_io I32:$b, 0))>;
3231
3232
3233// We need custom lowering of ISD::PREFETCH into HexagonISD::DCFETCH
3234// because the SDNode ISD::PREFETCH has properties MayLoad and MayStore.
3235// We don't really want either one here.
3236def SDTHexagonDCFETCH: SDTypeProfile<0, 2, [SDTCisPtrTy<0>,SDTCisInt<1>]>;
3237def HexagonDCFETCH: SDNode<"HexagonISD::DCFETCH", SDTHexagonDCFETCH,
3238                           [SDNPHasChain]>;
3239
3240def: Pat<(HexagonDCFETCH IntRegs:$Rs, u11_3ImmPred:$u11_3),
3241         (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
3242def: Pat<(HexagonDCFETCH (i32 (add IntRegs:$Rs, u11_3ImmPred:$u11_3)), (i32 0)),
3243         (Y2_dcfetchbo IntRegs:$Rs, imm:$u11_3)>;
3244
3245def SDTHexagonALLOCA
3246  : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>;
3247def HexagonALLOCA
3248  : SDNode<"HexagonISD::ALLOCA", SDTHexagonALLOCA, [SDNPHasChain]>;
3249
3250def: Pat<(HexagonALLOCA I32:$Rs, (i32 imm:$A)),
3251         (PS_alloca IntRegs:$Rs, imm:$A)>;
3252
3253def HexagonBARRIER: SDNode<"HexagonISD::BARRIER", SDTNone, [SDNPHasChain]>;
3254def: Pat<(HexagonBARRIER), (Y2_barrier)>;
3255
3256def: Pat<(trap), (PS_crash)>;
3257
3258// Read cycle counter.
3259def SDTInt64Leaf: SDTypeProfile<1, 0, [SDTCisVT<0, i64>]>;
3260def HexagonREADCYCLE: SDNode<"HexagonISD::READCYCLE", SDTInt64Leaf,
3261  [SDNPHasChain]>;
3262
3263def: Pat<(HexagonREADCYCLE), (A4_tfrcpp UPCYCLE)>;
3264
3265// The declared return value of the store-locked intrinsics is i32, but
3266// the instructions actually define i1. To avoid register copies from
3267// IntRegs to PredRegs and back, fold the entire pattern checking the
3268// result against true/false.
3269let AddedComplexity = 100 in {
3270  def: Pat<(i1 (setne (int_hexagon_S2_storew_locked I32:$Rs, I32:$Rt), 0)),
3271           (S2_storew_locked I32:$Rs, I32:$Rt)>;
3272  def: Pat<(i1 (seteq (int_hexagon_S2_storew_locked I32:$Rs, I32:$Rt), 0)),
3273           (C2_not (S2_storew_locked I32:$Rs, I32:$Rt))>;
3274  def: Pat<(i1 (setne (int_hexagon_S4_stored_locked I32:$Rs, I64:$Rt), 0)),
3275           (S4_stored_locked I32:$Rs, I64:$Rt)>;
3276  def: Pat<(i1 (seteq (int_hexagon_S4_stored_locked I32:$Rs, I64:$Rt), 0)),
3277           (C2_not (S4_stored_locked I32:$Rs, I64:$Rt))>;
3278}
3279