1 //===-- ARMInstrMVE.td - MVE support for ARM ---------------*- tablegen -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file describes the ARM MVE instruction set.
11 //===----------------------------------------------------------------------===//
13 class ExpandImmAsmOp<string shift> : AsmOperandClass {
14 let Name = !strconcat("ExpandImm", shift);
15 let PredicateMethod = !strconcat("isExpImm<", shift, ">");
16 let RenderMethod = "addImmOperands";
18 class InvertedExpandImmAsmOp<string shift, string size> : AsmOperandClass {
19 let Name = !strconcat("InvertedExpandImm", shift, "_", size);
20 let PredicateMethod = !strconcat("isInvertedExpImm<", shift, ",", size, ">");
21 let RenderMethod = "addImmOperands";
24 class ExpandImm<string shift> : Operand<i32> {
25 let ParserMatchClass = ExpandImmAsmOp<shift>;
26 let EncoderMethod = !strconcat("getExpandedImmOpValue<",shift,",false>");
27 let DecoderMethod = !strconcat("DecodeExpandedImmOperand<",shift,">");
28 let PrintMethod = "printExpandedImmOperand";
30 class InvertedExpandImm<string shift, string size> : Operand<i32> {
31 let ParserMatchClass = InvertedExpandImmAsmOp<shift, size>;
32 let EncoderMethod = !strconcat("getExpandedImmOpValue<",shift,",true>");
33 let PrintMethod = "printExpandedImmOperand";
34 // No decoder method needed, because this operand type is only used
35 // by aliases (VAND and VORN)
38 def expzero00 : ExpandImm<"0">;
39 def expzero08 : ExpandImm<"8">;
40 def expzero16 : ExpandImm<"16">;
41 def expzero24 : ExpandImm<"24">;
43 def expzero00inv16 : InvertedExpandImm<"0", "16">;
44 def expzero08inv16 : InvertedExpandImm<"8", "16">;
46 def expzero00inv32 : InvertedExpandImm<"0", "32">;
47 def expzero08inv32 : InvertedExpandImm<"8", "32">;
48 def expzero16inv32 : InvertedExpandImm<"16", "32">;
49 def expzero24inv32 : InvertedExpandImm<"24", "32">;
52 def vpt_mask : Operand<i32> {
53 let PrintMethod = "printVPTMask";
54 let ParserMatchClass = it_mask_asmoperand;
55 let EncoderMethod = "getVPTMaskOpValue";
56 let DecoderMethod = "DecodeVPTMaskOperand";
59 // VPT/VCMP restricted predicate for sign invariant types
60 def pred_restricted_i_asmoperand : AsmOperandClass {
61 let Name = "CondCodeRestrictedI";
62 let RenderMethod = "addITCondCodeOperands";
63 let PredicateMethod = "isITCondCodeRestrictedI";
64 let ParserMethod = "parseITCondCode";
65 let DiagnosticString = "condition code for sign-independent integer "#
66 "comparison must be EQ or NE";
69 // VPT/VCMP restricted predicate for signed types
70 def pred_restricted_s_asmoperand : AsmOperandClass {
71 let Name = "CondCodeRestrictedS";
72 let RenderMethod = "addITCondCodeOperands";
73 let PredicateMethod = "isITCondCodeRestrictedS";
74 let ParserMethod = "parseITCondCode";
75 let DiagnosticString = "condition code for signed integer "#
76 "comparison must be EQ, NE, LT, GT, LE or GE";
79 // VPT/VCMP restricted predicate for unsigned types
80 def pred_restricted_u_asmoperand : AsmOperandClass {
81 let Name = "CondCodeRestrictedU";
82 let RenderMethod = "addITCondCodeOperands";
83 let PredicateMethod = "isITCondCodeRestrictedU";
84 let ParserMethod = "parseITCondCode";
85 let DiagnosticString = "condition code for unsigned integer "#
86 "comparison must be EQ, NE, HS or HI";
89 // VPT/VCMP restricted predicate for floating point
90 def pred_restricted_fp_asmoperand : AsmOperandClass {
91 let Name = "CondCodeRestrictedFP";
92 let RenderMethod = "addITCondCodeOperands";
93 let PredicateMethod = "isITCondCodeRestrictedFP";
94 let ParserMethod = "parseITCondCode";
95 let DiagnosticString = "condition code for floating-point "#
96 "comparison must be EQ, NE, LT, GT, LE or GE";
99 class VCMPPredicateOperand : Operand<i32>;
101 def pred_basic_i : VCMPPredicateOperand {
102 let PrintMethod = "printMandatoryRestrictedPredicateOperand";
103 let ParserMatchClass = pred_restricted_i_asmoperand;
104 let DecoderMethod = "DecodeRestrictedIPredicateOperand";
105 let EncoderMethod = "getRestrictedCondCodeOpValue";
108 def pred_basic_u : VCMPPredicateOperand {
109 let PrintMethod = "printMandatoryRestrictedPredicateOperand";
110 let ParserMatchClass = pred_restricted_u_asmoperand;
111 let DecoderMethod = "DecodeRestrictedUPredicateOperand";
112 let EncoderMethod = "getRestrictedCondCodeOpValue";
115 def pred_basic_s : VCMPPredicateOperand {
116 let PrintMethod = "printMandatoryRestrictedPredicateOperand";
117 let ParserMatchClass = pred_restricted_s_asmoperand;
118 let DecoderMethod = "DecodeRestrictedSPredicateOperand";
119 let EncoderMethod = "getRestrictedCondCodeOpValue";
122 def pred_basic_fp : VCMPPredicateOperand {
123 let PrintMethod = "printMandatoryRestrictedPredicateOperand";
124 let ParserMatchClass = pred_restricted_fp_asmoperand;
125 let DecoderMethod = "DecodeRestrictedFPPredicateOperand";
126 let EncoderMethod = "getRestrictedCondCodeOpValue";
129 // Register list operands for interleaving load/stores
130 def VecList2QAsmOperand : AsmOperandClass {
131 let Name = "VecListTwoMQ";
132 let ParserMethod = "parseVectorList";
133 let RenderMethod = "addMVEVecListOperands";
134 let DiagnosticString = "operand must be a list of two consecutive "#
135 "q-registers in range [q0,q7]";
138 def VecList2Q : RegisterOperand<QQPR, "printMVEVectorListTwoQ"> {
139 let ParserMatchClass = VecList2QAsmOperand;
140 let PrintMethod = "printMVEVectorList<2>";
143 def VecList4QAsmOperand : AsmOperandClass {
144 let Name = "VecListFourMQ";
145 let ParserMethod = "parseVectorList";
146 let RenderMethod = "addMVEVecListOperands";
147 let DiagnosticString = "operand must be a list of four consecutive "#
148 "q-registers in range [q0,q7]";
151 def VecList4Q : RegisterOperand<QQQQPR, "printMVEVectorListFourQ"> {
152 let ParserMatchClass = VecList4QAsmOperand;
153 let PrintMethod = "printMVEVectorList<4>";
156 // taddrmode_imm7 := reg[r0-r7] +/- (imm7 << shift)
157 class TMemImm7ShiftOffsetAsmOperand<int shift> : AsmOperandClass {
158 let Name = "TMemImm7Shift"#shift#"Offset";
159 let PredicateMethod = "isMemImm7ShiftedOffset<"#shift#",ARM::tGPRRegClassID>";
160 let RenderMethod = "addMemImmOffsetOperands";
163 class taddrmode_imm7<int shift> : MemOperand,
164 ComplexPattern<i32, 2, "SelectTAddrModeImm7<"#shift#">", []> {
165 let ParserMatchClass = TMemImm7ShiftOffsetAsmOperand<shift>;
166 // They are printed the same way as the T2 imm8 version
167 let PrintMethod = "printT2AddrModeImm8Operand<false>";
168 // This can also be the same as the T2 version.
169 let EncoderMethod = "getT2AddrModeImmOpValue<7,"#shift#">";
170 let DecoderMethod = "DecodeTAddrModeImm7<"#shift#">";
171 let MIOperandInfo = (ops tGPR:$base, i32imm:$offsimm);
174 // t2addrmode_imm7 := reg +/- (imm7)
175 class MemImm7ShiftOffsetAsmOperand<int shift> : AsmOperandClass {
176 let Name = "MemImm7Shift"#shift#"Offset";
177 let PredicateMethod = "isMemImm7ShiftedOffset<" # shift #
178 ",ARM::GPRnopcRegClassID>";
179 let RenderMethod = "addMemImmOffsetOperands";
182 def MemImm7Shift0OffsetAsmOperand : MemImm7ShiftOffsetAsmOperand<0>;
183 def MemImm7Shift1OffsetAsmOperand : MemImm7ShiftOffsetAsmOperand<1>;
184 def MemImm7Shift2OffsetAsmOperand : MemImm7ShiftOffsetAsmOperand<2>;
185 class T2AddrMode_Imm7<int shift> : MemOperand,
186 ComplexPattern<i32, 2, "SelectT2AddrModeImm7<"#shift#">", []> {
187 let EncoderMethod = "getT2AddrModeImmOpValue<7,"#shift#">";
188 let DecoderMethod = "DecodeT2AddrModeImm7<"#shift#", 0>";
189 let ParserMatchClass =
190 !cast<AsmOperandClass>("MemImm7Shift"#shift#"OffsetAsmOperand");
191 let MIOperandInfo = (ops GPRnopc:$base, i32imm:$offsimm);
194 class t2addrmode_imm7<int shift> : T2AddrMode_Imm7<shift> {
195 // They are printed the same way as the imm8 version
196 let PrintMethod = "printT2AddrModeImm8Operand<false>";
199 class MemImm7ShiftOffsetWBAsmOperand<int shift> : AsmOperandClass {
200 let Name = "MemImm7Shift"#shift#"OffsetWB";
201 let PredicateMethod = "isMemImm7ShiftedOffset<" # shift #
202 ",ARM::rGPRRegClassID>";
203 let RenderMethod = "addMemImmOffsetOperands";
206 def MemImm7Shift0OffsetWBAsmOperand : MemImm7ShiftOffsetWBAsmOperand<0>;
207 def MemImm7Shift1OffsetWBAsmOperand : MemImm7ShiftOffsetWBAsmOperand<1>;
208 def MemImm7Shift2OffsetWBAsmOperand : MemImm7ShiftOffsetWBAsmOperand<2>;
210 class t2addrmode_imm7_pre<int shift> : T2AddrMode_Imm7<shift> {
211 // They are printed the same way as the imm8 version
212 let PrintMethod = "printT2AddrModeImm8Operand<true>";
213 let ParserMatchClass =
214 !cast<AsmOperandClass>("MemImm7Shift"#shift#"OffsetWBAsmOperand");
215 let DecoderMethod = "DecodeT2AddrModeImm7<"#shift#", 1>";
216 let MIOperandInfo = (ops rGPR:$base, i32imm:$offsim);
219 class t2am_imm7shiftOffsetAsmOperand<int shift>
220 : AsmOperandClass { let Name = "Imm7Shift"#shift; }
221 def t2am_imm7shift0OffsetAsmOperand : t2am_imm7shiftOffsetAsmOperand<0>;
222 def t2am_imm7shift1OffsetAsmOperand : t2am_imm7shiftOffsetAsmOperand<1>;
223 def t2am_imm7shift2OffsetAsmOperand : t2am_imm7shiftOffsetAsmOperand<2>;
225 class t2am_imm7_offset<int shift> : MemOperand,
226 ComplexPattern<i32, 1, "SelectT2AddrModeImm7Offset<"#shift#">",
227 [], [SDNPWantRoot]> {
228 // They are printed the same way as the imm8 version
229 let PrintMethod = "printT2AddrModeImm8OffsetOperand";
230 let ParserMatchClass =
231 !cast<AsmOperandClass>("t2am_imm7shift"#shift#"OffsetAsmOperand");
232 let EncoderMethod = "getT2ScaledImmOpValue<7,"#shift#">";
233 let DecoderMethod = "DecodeT2Imm7<"#shift#">";
236 // Operands for gather/scatter loads of the form [Rbase, Qoffsets]
237 class MemRegRQOffsetAsmOperand<int shift> : AsmOperandClass {
238 let Name = "MemRegRQS"#shift#"Offset";
239 let PredicateMethod = "isMemRegRQOffset<"#shift#">";
240 let RenderMethod = "addMemRegRQOffsetOperands";
243 def MemRegRQS0OffsetAsmOperand : MemRegRQOffsetAsmOperand<0>;
244 def MemRegRQS1OffsetAsmOperand : MemRegRQOffsetAsmOperand<1>;
245 def MemRegRQS2OffsetAsmOperand : MemRegRQOffsetAsmOperand<2>;
246 def MemRegRQS3OffsetAsmOperand : MemRegRQOffsetAsmOperand<3>;
248 // mve_addr_rq_shift := reg + vreg{ << UXTW #shift}
249 class mve_addr_rq_shift<int shift> : MemOperand {
250 let EncoderMethod = "getMveAddrModeRQOpValue";
251 let PrintMethod = "printMveAddrModeRQOperand<"#shift#">";
252 let ParserMatchClass =
253 !cast<AsmOperandClass>("MemRegRQS"#shift#"OffsetAsmOperand");
254 let DecoderMethod = "DecodeMveAddrModeRQ";
255 let MIOperandInfo = (ops GPRnopc:$base, MQPR:$offsreg);
258 class MemRegQOffsetAsmOperand<int shift> : AsmOperandClass {
259 let Name = "MemRegQS"#shift#"Offset";
260 let PredicateMethod = "isMemRegQOffset<"#shift#">";
261 let RenderMethod = "addMemImmOffsetOperands";
264 def MemRegQS2OffsetAsmOperand : MemRegQOffsetAsmOperand<2>;
265 def MemRegQS3OffsetAsmOperand : MemRegQOffsetAsmOperand<3>;
267 // mve_addr_q_shift := vreg {+ #imm7s2/4}
268 class mve_addr_q_shift<int shift> : MemOperand {
269 let EncoderMethod = "getMveAddrModeQOpValue<"#shift#">";
270 // Can be printed same way as other reg + imm operands
271 let PrintMethod = "printT2AddrModeImm8Operand<false>";
272 let ParserMatchClass =
273 !cast<AsmOperandClass>("MemRegQS"#shift#"OffsetAsmOperand");
274 let DecoderMethod = "DecodeMveAddrModeQ<"#shift#">";
275 let MIOperandInfo = (ops MQPR:$base, i32imm:$imm);
278 // --------- Start of base classes for the instructions themselves
280 class MVE_MI<dag oops, dag iops, InstrItinClass itin, string asm,
281 string ops, string cstr, list<dag> pattern>
282 : Thumb2XI<oops, iops, AddrModeNone, 4, itin, !strconcat(asm, "\t", ops), cstr,
284 Requires<[HasMVEInt]> {
286 let DecoderNamespace = "MVE";
289 // MVE_p is used for most predicated instructions, to add the cluster
290 // of input operands that provides the VPT suffix (none, T or E) and
291 // the input predicate register.
292 class MVE_p<dag oops, dag iops, InstrItinClass itin, string iname,
293 string suffix, string ops, vpred_ops vpred, string cstr,
294 list<dag> pattern=[]>
295 : MVE_MI<oops, !con(iops, (ins vpred:$vp)), itin,
296 // If the instruction has a suffix, like vadd.f32, then the
297 // VPT predication suffix goes before the dot, so the full
298 // name has to be "vadd${vp}.f32".
299 !strconcat(iname, "${vp}",
300 !if(!eq(suffix, ""), "", !strconcat(".", suffix))),
301 ops, !strconcat(cstr, vpred.vpred_constraint), pattern> {
302 let Inst{31-29} = 0b111;
303 let Inst{27-26} = 0b11;
306 class MVE_f<dag oops, dag iops, InstrItinClass itin, string iname,
307 string suffix, string ops, vpred_ops vpred, string cstr,
308 list<dag> pattern=[]>
309 : MVE_p<oops, iops, itin, iname, suffix, ops, vpred, cstr, pattern> {
310 let Predicates = [HasMVEFloat];
313 class MVE_MI_with_pred<dag oops, dag iops, InstrItinClass itin, string asm,
314 string ops, string cstr, list<dag> pattern>
315 : Thumb2I<oops, iops, AddrModeNone, 4, itin, asm, !strconcat("\t", ops), cstr,
317 Requires<[HasV8_1MMainline, HasMVEInt]> {
319 let DecoderNamespace = "MVE";
322 class MVE_VMOV_lane_base<dag oops, dag iops, InstrItinClass itin, string asm,
323 string suffix, string ops, string cstr,
325 : Thumb2I<oops, iops, AddrModeNone, 4, itin, asm,
326 !if(!eq(suffix, ""), "", "." # suffix) # "\t" # ops,
328 Requires<[HasV8_1MMainline, HasMVEInt]> {
330 let DecoderNamespace = "MVE";
333 class MVE_ScalarShift<string iname, dag oops, dag iops, string asm, string cstr,
334 list<dag> pattern=[]>
335 : MVE_MI_with_pred<oops, iops, NoItinerary, iname, asm, cstr, pattern> {
336 let Inst{31-20} = 0b111010100101;
341 class MVE_ScalarShiftSingleReg<string iname, dag iops, string asm, string cstr,
342 list<dag> pattern=[]>
343 : MVE_ScalarShift<iname, (outs rGPR:$RdaDest), iops, asm, cstr, pattern> {
346 let Inst{19-16} = RdaDest{3-0};
349 class MVE_ScalarShiftSRegImm<string iname, bits<2> op5_4, list<dag> pattern=[]>
350 : MVE_ScalarShiftSingleReg<iname, (ins rGPR:$RdaSrc, long_shift:$imm),
351 "$RdaSrc, $imm", "$RdaDest = $RdaSrc", pattern> {
355 let Inst{14-12} = imm{4-2};
356 let Inst{11-8} = 0b1111;
357 let Inst{7-6} = imm{1-0};
358 let Inst{5-4} = op5_4{1-0};
359 let Inst{3-0} = 0b1111;
362 def MVE_SQSHL : MVE_ScalarShiftSRegImm<"sqshl", 0b11>;
363 def MVE_SRSHR : MVE_ScalarShiftSRegImm<"srshr", 0b10>;
364 def MVE_UQSHL : MVE_ScalarShiftSRegImm<"uqshl", 0b00>;
365 def MVE_URSHR : MVE_ScalarShiftSRegImm<"urshr", 0b01>;
367 class MVE_ScalarShiftSRegReg<string iname, bits<2> op5_4, list<dag> pattern=[]>
368 : MVE_ScalarShiftSingleReg<iname, (ins rGPR:$RdaSrc, rGPR:$Rm),
369 "$RdaSrc, $Rm", "$RdaDest = $RdaSrc", pattern> {
372 let Inst{15-12} = Rm{3-0};
373 let Inst{11-8} = 0b1111;
374 let Inst{7-6} = 0b00;
375 let Inst{5-4} = op5_4{1-0};
376 let Inst{3-0} = 0b1101;
378 let Unpredictable{8-6} = 0b111;
381 def MVE_SQRSHR : MVE_ScalarShiftSRegReg<"sqrshr", 0b10>;
382 def MVE_UQRSHL : MVE_ScalarShiftSRegReg<"uqrshl", 0b00>;
384 class MVE_ScalarShiftDoubleReg<string iname, dag iops, string asm,
385 string cstr, list<dag> pattern=[]>
386 : MVE_ScalarShift<iname, (outs tGPREven:$RdaLo, tGPROdd:$RdaHi),
387 iops, asm, cstr, pattern> {
391 let Inst{19-17} = RdaLo{3-1};
392 let Inst{11-9} = RdaHi{3-1};
395 class MVE_ScalarShiftDRegImm<string iname, bits<2> op5_4, bit op16,
396 list<dag> pattern=[]>
397 : MVE_ScalarShiftDoubleReg<
398 iname, (ins tGPREven:$RdaLo_src, tGPROdd:$RdaHi_src, long_shift:$imm),
399 "$RdaLo, $RdaHi, $imm", "$RdaLo = $RdaLo_src,$RdaHi = $RdaHi_src",
405 let Inst{14-12} = imm{4-2};
406 let Inst{7-6} = imm{1-0};
407 let Inst{5-4} = op5_4{1-0};
408 let Inst{3-0} = 0b1111;
411 class MVE_ScalarShiftDRegRegBase<string iname, dag iops, string asm,
412 bit op5, bit op16, list<dag> pattern=[]>
413 : MVE_ScalarShiftDoubleReg<
414 iname, iops, asm, "@earlyclobber $RdaHi,@earlyclobber $RdaLo,"
415 "$RdaLo = $RdaLo_src,$RdaHi = $RdaHi_src",
420 let Inst{15-12} = Rm{3-0};
424 let Inst{3-0} = 0b1101;
426 // Custom decoder method because of the following overlapping encodings:
429 // SQRSHRL and SQRSHR
430 // UQRSHLL and UQRSHL
431 let DecoderMethod = "DecodeMVEOverlappingLongShift";
434 class MVE_ScalarShiftDRegReg<string iname, bit op5, list<dag> pattern=[]>
435 : MVE_ScalarShiftDRegRegBase<
436 iname, (ins tGPREven:$RdaLo_src, tGPROdd:$RdaHi_src, rGPR:$Rm),
437 "$RdaLo, $RdaHi, $Rm", op5, 0b0, pattern> {
442 class MVE_ScalarShiftDRegRegWithSat<string iname, bit op5, list<dag> pattern=[]>
443 : MVE_ScalarShiftDRegRegBase<
444 iname, (ins tGPREven:$RdaLo_src, tGPROdd:$RdaHi_src, rGPR:$Rm, saturateop:$sat),
445 "$RdaLo, $RdaHi, $sat, $Rm", op5, 0b1, pattern> {
451 def MVE_ASRLr : MVE_ScalarShiftDRegReg<"asrl", 0b1, [(set tGPREven:$RdaLo, tGPROdd:$RdaHi,
452 (ARMasrl tGPREven:$RdaLo_src,
453 tGPROdd:$RdaHi_src, rGPR:$Rm))]>;
454 def MVE_ASRLi : MVE_ScalarShiftDRegImm<"asrl", 0b10, ?, [(set tGPREven:$RdaLo, tGPROdd:$RdaHi,
455 (ARMasrl tGPREven:$RdaLo_src,
456 tGPROdd:$RdaHi_src, (i32 long_shift:$imm)))]>;
457 def MVE_LSLLr : MVE_ScalarShiftDRegReg<"lsll", 0b0, [(set tGPREven:$RdaLo, tGPROdd:$RdaHi,
458 (ARMlsll tGPREven:$RdaLo_src,
459 tGPROdd:$RdaHi_src, rGPR:$Rm))]>;
460 def MVE_LSLLi : MVE_ScalarShiftDRegImm<"lsll", 0b00, ?, [(set tGPREven:$RdaLo, tGPROdd:$RdaHi,
461 (ARMlsll tGPREven:$RdaLo_src,
462 tGPROdd:$RdaHi_src, (i32 long_shift:$imm)))]>;
463 def MVE_LSRL : MVE_ScalarShiftDRegImm<"lsrl", 0b01, ?, [(set tGPREven:$RdaLo, tGPROdd:$RdaHi,
464 (ARMlsrl tGPREven:$RdaLo_src,
465 tGPROdd:$RdaHi_src, (i32 long_shift:$imm)))]>;
467 def MVE_SQRSHRL : MVE_ScalarShiftDRegRegWithSat<"sqrshrl", 0b1>;
468 def MVE_SQSHLL : MVE_ScalarShiftDRegImm<"sqshll", 0b11, 0b1>;
469 def MVE_SRSHRL : MVE_ScalarShiftDRegImm<"srshrl", 0b10, 0b1>;
471 def MVE_UQRSHLL : MVE_ScalarShiftDRegRegWithSat<"uqrshll", 0b0>;
472 def MVE_UQSHLL : MVE_ScalarShiftDRegImm<"uqshll", 0b00, 0b1>;
473 def MVE_URSHRL : MVE_ScalarShiftDRegImm<"urshrl", 0b01, 0b1>;
475 // start of mve_rDest instructions
477 class MVE_rDest<dag oops, dag iops, InstrItinClass itin,
478 string iname, string suffix,
479 string ops, string cstr, list<dag> pattern=[]>
480 // Always use vpred_n and not vpred_r: with the output register being
481 // a GPR and not a vector register, there can't be any question of
482 // what to put in its inactive lanes.
483 : MVE_p<oops, iops, itin, iname, suffix, ops, vpred_n, cstr, pattern> {
485 let Inst{25-23} = 0b101;
486 let Inst{11-9} = 0b111;
490 class MVE_VABAV<string suffix, bit U, bits<2> size, list<dag> pattern=[]>
491 : MVE_rDest<(outs rGPR:$Rda), (ins rGPR:$Rda_src, MQPR:$Qn, MQPR:$Qm),
492 NoItinerary, "vabav", suffix, "$Rda, $Qn, $Qm", "$Rda = $Rda_src",
500 let Inst{21-20} = size{1-0};
501 let Inst{19-17} = Qn{2-0};
503 let Inst{15-12} = Rda{3-0};
508 let Inst{3-1} = Qm{2-0};
510 let invalidForTailPredication = 1;
513 def MVE_VABAVs8 : MVE_VABAV<"s8", 0b0, 0b00>;
514 def MVE_VABAVs16 : MVE_VABAV<"s16", 0b0, 0b01>;
515 def MVE_VABAVs32 : MVE_VABAV<"s32", 0b0, 0b10>;
516 def MVE_VABAVu8 : MVE_VABAV<"u8", 0b1, 0b00>;
517 def MVE_VABAVu16 : MVE_VABAV<"u16", 0b1, 0b01>;
518 def MVE_VABAVu32 : MVE_VABAV<"u32", 0b1, 0b10>;
520 class MVE_VADDV<string iname, string suffix, dag iops, string cstr,
521 bit A, bit U, bits<2> size, list<dag> pattern=[]>
522 : MVE_rDest<(outs tGPREven:$Rda), iops, NoItinerary,
523 iname, suffix, "$Rda, $Qm", cstr, pattern> {
528 let Inst{22-20} = 0b111;
529 let Inst{19-18} = size{1-0};
530 let Inst{17-16} = 0b01;
531 let Inst{15-13} = Rda{3-1};
533 let Inst{8-6} = 0b100;
535 let Inst{3-1} = Qm{2-0};
537 let invalidForTailPredication = 1;
540 multiclass MVE_VADDV_A<string suffix, bit U, bits<2> size,
541 list<dag> pattern=[]> {
542 def acc : MVE_VADDV<"vaddva", suffix,
543 (ins tGPREven:$Rda_src, MQPR:$Qm), "$Rda = $Rda_src",
544 0b1, U, size, pattern>;
545 def no_acc : MVE_VADDV<"vaddv", suffix,
547 0b0, U, size, pattern>;
550 defm MVE_VADDVs8 : MVE_VADDV_A<"s8", 0b0, 0b00>;
551 defm MVE_VADDVs16 : MVE_VADDV_A<"s16", 0b0, 0b01>;
552 defm MVE_VADDVs32 : MVE_VADDV_A<"s32", 0b0, 0b10>;
553 defm MVE_VADDVu8 : MVE_VADDV_A<"u8", 0b1, 0b00>;
554 defm MVE_VADDVu16 : MVE_VADDV_A<"u16", 0b1, 0b01>;
555 defm MVE_VADDVu32 : MVE_VADDV_A<"u32", 0b1, 0b10>;
557 let Predicates = [HasMVEInt] in {
558 def : Pat<(i32 (vecreduce_add (v4i32 MQPR:$src))), (i32 (MVE_VADDVu32no_acc $src))>;
559 def : Pat<(i32 (vecreduce_add (v8i16 MQPR:$src))), (i32 (MVE_VADDVu16no_acc $src))>;
560 def : Pat<(i32 (vecreduce_add (v16i8 MQPR:$src))), (i32 (MVE_VADDVu8no_acc $src))>;
561 def : Pat<(i32 (add (i32 (vecreduce_add (v4i32 MQPR:$src1))), (i32 tGPR:$src2))),
562 (i32 (MVE_VADDVu32acc $src2, $src1))>;
563 def : Pat<(i32 (add (i32 (vecreduce_add (v8i16 MQPR:$src1))), (i32 tGPR:$src2))),
564 (i32 (MVE_VADDVu16acc $src2, $src1))>;
565 def : Pat<(i32 (add (i32 (vecreduce_add (v16i8 MQPR:$src1))), (i32 tGPR:$src2))),
566 (i32 (MVE_VADDVu8acc $src2, $src1))>;
570 class MVE_VADDLV<string iname, string suffix, dag iops, string cstr,
571 bit A, bit U, list<dag> pattern=[]>
572 : MVE_rDest<(outs tGPREven:$RdaLo, tGPROdd:$RdaHi), iops, NoItinerary, iname,
573 suffix, "$RdaLo, $RdaHi, $Qm", cstr, pattern> {
579 let Inst{22-20} = RdaHi{3-1};
580 let Inst{19-18} = 0b10;
581 let Inst{17-16} = 0b01;
582 let Inst{15-13} = RdaLo{3-1};
584 let Inst{8-6} = 0b100;
586 let Inst{3-1} = Qm{2-0};
588 let invalidForTailPredication = 1;
591 multiclass MVE_VADDLV_A<string suffix, bit U, list<dag> pattern=[]> {
592 def acc : MVE_VADDLV<"vaddlva", suffix,
593 (ins tGPREven:$RdaLo_src, tGPROdd:$RdaHi_src, MQPR:$Qm),
594 "$RdaLo = $RdaLo_src,$RdaHi = $RdaHi_src",
596 def no_acc : MVE_VADDLV<"vaddlv", suffix,
602 defm MVE_VADDLVs32 : MVE_VADDLV_A<"s32", 0b0>;
603 defm MVE_VADDLVu32 : MVE_VADDLV_A<"u32", 0b1>;
605 class MVE_VMINMAXNMV<string iname, string suffix, bit sz,
606 bit bit_17, bit bit_7, list<dag> pattern=[]>
607 : MVE_rDest<(outs rGPR:$RdaDest), (ins rGPR:$RdaSrc, MQPR:$Qm),
608 NoItinerary, iname, suffix, "$RdaSrc, $Qm",
609 "$RdaDest = $RdaSrc", pattern> {
614 let Inst{22-20} = 0b110;
615 let Inst{19-18} = 0b11;
616 let Inst{17} = bit_17;
618 let Inst{15-12} = RdaDest{3-0};
621 let Inst{6-5} = 0b00;
622 let Inst{3-1} = Qm{2-0};
625 let Predicates = [HasMVEFloat];
626 let invalidForTailPredication = 1;
629 multiclass MVE_VMINMAXNMV_fty<string iname, bit bit_7, list<dag> pattern=[]> {
630 def f32 : MVE_VMINMAXNMV<iname, "f32", 0b0, 0b1, bit_7, pattern>;
631 def f16 : MVE_VMINMAXNMV<iname, "f16", 0b1, 0b1, bit_7, pattern>;
634 defm MVE_VMINNMV : MVE_VMINMAXNMV_fty<"vminnmv", 0b1>;
635 defm MVE_VMAXNMV : MVE_VMINMAXNMV_fty<"vmaxnmv", 0b0>;
637 multiclass MVE_VMINMAXNMAV_fty<string iname, bit bit_7, list<dag> pattern=[]> {
638 def f32 : MVE_VMINMAXNMV<iname, "f32", 0b0, 0b0, bit_7, pattern>;
639 def f16 : MVE_VMINMAXNMV<iname, "f16", 0b1, 0b0, bit_7, pattern>;
642 defm MVE_VMINNMAV : MVE_VMINMAXNMAV_fty<"vminnmav", 0b1>;
643 defm MVE_VMAXNMAV : MVE_VMINMAXNMAV_fty<"vmaxnmav", 0b0>;
645 class MVE_VMINMAXV<string iname, string suffix, bit U, bits<2> size,
646 bit bit_17, bit bit_7, list<dag> pattern=[]>
647 : MVE_rDest<(outs rGPR:$RdaDest), (ins rGPR:$RdaSrc, MQPR:$Qm), NoItinerary,
648 iname, suffix, "$RdaSrc, $Qm", "$RdaDest = $RdaSrc", pattern> {
653 let Inst{22-20} = 0b110;
654 let Inst{19-18} = size{1-0};
655 let Inst{17} = bit_17;
657 let Inst{15-12} = RdaDest{3-0};
660 let Inst{6-5} = 0b00;
661 let Inst{3-1} = Qm{2-0};
663 let invalidForTailPredication = 1;
666 multiclass MVE_VMINMAXV_ty<string iname, bit bit_7, list<dag> pattern=[]> {
667 def s8 : MVE_VMINMAXV<iname, "s8", 0b0, 0b00, 0b1, bit_7>;
668 def s16 : MVE_VMINMAXV<iname, "s16", 0b0, 0b01, 0b1, bit_7>;
669 def s32 : MVE_VMINMAXV<iname, "s32", 0b0, 0b10, 0b1, bit_7>;
670 def u8 : MVE_VMINMAXV<iname, "u8", 0b1, 0b00, 0b1, bit_7>;
671 def u16 : MVE_VMINMAXV<iname, "u16", 0b1, 0b01, 0b1, bit_7>;
672 def u32 : MVE_VMINMAXV<iname, "u32", 0b1, 0b10, 0b1, bit_7>;
675 defm MVE_VMINV : MVE_VMINMAXV_ty<"vminv", 0b1>;
676 defm MVE_VMAXV : MVE_VMINMAXV_ty<"vmaxv", 0b0>;
678 let Predicates = [HasMVEInt] in {
679 def : Pat<(i32 (vecreduce_smax (v16i8 MQPR:$src))),
680 (i32 (MVE_VMAXVs8 (t2MVNi (i32 127)), $src))>;
681 def : Pat<(i32 (vecreduce_smax (v8i16 MQPR:$src))),
682 (i32 (MVE_VMAXVs16 (t2MOVi32imm (i32 -32768)), $src))>;
683 def : Pat<(i32 (vecreduce_smax (v4i32 MQPR:$src))),
684 (i32 (MVE_VMAXVs32 (t2MOVi (i32 -2147483648)), $src))>;
685 def : Pat<(i32 (vecreduce_umax (v16i8 MQPR:$src))),
686 (i32 (MVE_VMAXVu8 (t2MOVi (i32 0)), $src))>;
687 def : Pat<(i32 (vecreduce_umax (v8i16 MQPR:$src))),
688 (i32 (MVE_VMAXVu16 (t2MOVi (i32 0)), $src))>;
689 def : Pat<(i32 (vecreduce_umax (v4i32 MQPR:$src))),
690 (i32 (MVE_VMAXVu32 (t2MOVi (i32 0)), $src))>;
692 def : Pat<(i32 (vecreduce_smin (v16i8 MQPR:$src))),
693 (i32 (MVE_VMINVs8 (t2MOVi (i32 127)), $src))>;
694 def : Pat<(i32 (vecreduce_smin (v8i16 MQPR:$src))),
695 (i32 (MVE_VMINVs16 (t2MOVi16 (i32 32767)), $src))>;
696 def : Pat<(i32 (vecreduce_smin (v4i32 MQPR:$src))),
697 (i32 (MVE_VMINVs32 (t2MVNi (i32 -2147483648)), $src))>;
698 def : Pat<(i32 (vecreduce_umin (v16i8 MQPR:$src))),
699 (i32 (MVE_VMINVu8 (t2MOVi (i32 255)), $src))>;
700 def : Pat<(i32 (vecreduce_umin (v8i16 MQPR:$src))),
701 (i32 (MVE_VMINVu16 (t2MOVi16 (i32 65535)), $src))>;
702 def : Pat<(i32 (vecreduce_umin (v4i32 MQPR:$src))),
703 (i32 (MVE_VMINVu32 (t2MOVi (i32 4294967295)), $src))>;
707 multiclass MVE_VMINMAXAV_ty<string iname, bit bit_7, list<dag> pattern=[]> {
708 def s8 : MVE_VMINMAXV<iname, "s8", 0b0, 0b00, 0b0, bit_7>;
709 def s16 : MVE_VMINMAXV<iname, "s16", 0b0, 0b01, 0b0, bit_7>;
710 def s32 : MVE_VMINMAXV<iname, "s32", 0b0, 0b10, 0b0, bit_7>;
713 defm MVE_VMINAV : MVE_VMINMAXAV_ty<"vminav", 0b1>;
714 defm MVE_VMAXAV : MVE_VMINMAXAV_ty<"vmaxav", 0b0>;
716 class MVE_VMLAMLSDAV<string iname, string suffix, dag iops, string cstr,
717 bit sz, bit bit_28, bit A, bit X, bit bit_8, bit bit_0,
718 list<dag> pattern=[]>
719 : MVE_rDest<(outs tGPREven:$RdaDest), iops, NoItinerary, iname, suffix,
720 "$RdaDest, $Qn, $Qm", cstr, pattern> {
725 let Inst{28} = bit_28;
726 let Inst{22-20} = 0b111;
727 let Inst{19-17} = Qn{2-0};
729 let Inst{15-13} = RdaDest{3-1};
732 let Inst{7-6} = 0b00;
734 let Inst{3-1} = Qm{2-0};
736 let invalidForTailPredication = 1;
739 multiclass MVE_VMLAMLSDAV_A<string iname, string x, string suffix,
740 bit sz, bit bit_28, bit X, bit bit_8, bit bit_0,
741 list<dag> pattern=[]> {
742 def ""#x#suffix : MVE_VMLAMLSDAV<iname # x, suffix,
743 (ins MQPR:$Qn, MQPR:$Qm), "",
744 sz, bit_28, 0b0, X, bit_8, bit_0, pattern>;
745 def "a"#x#suffix : MVE_VMLAMLSDAV<iname # "a" # x, suffix,
746 (ins tGPREven:$RdaSrc, MQPR:$Qn, MQPR:$Qm),
747 "$RdaDest = $RdaSrc",
748 sz, bit_28, 0b1, X, bit_8, bit_0, pattern>;
751 multiclass MVE_VMLAMLSDAV_AX<string iname, string suffix, bit sz, bit bit_28,
752 bit bit_8, bit bit_0, list<dag> pattern=[]> {
753 defm "" : MVE_VMLAMLSDAV_A<iname, "", suffix, sz, bit_28,
754 0b0, bit_8, bit_0, pattern>;
755 defm "" : MVE_VMLAMLSDAV_A<iname, "x", suffix, sz, bit_28,
756 0b1, bit_8, bit_0, pattern>;
759 multiclass MVE_VMLADAV_multi<string suffix, bit sz, bit bit_8,
760 list<dag> pattern=[]> {
761 defm "" : MVE_VMLAMLSDAV_AX<"vmladav", "s"#suffix,
762 sz, 0b0, bit_8, 0b0, pattern>;
763 defm "" : MVE_VMLAMLSDAV_A<"vmladav", "", "u"#suffix,
764 sz, 0b1, 0b0, bit_8, 0b0, pattern>;
767 multiclass MVE_VMLSDAV_multi<string suffix, bit sz, bit bit_28,
768 list<dag> pattern=[]> {
769 defm "" : MVE_VMLAMLSDAV_AX<"vmlsdav", "s"#suffix,
770 sz, bit_28, 0b0, 0b1, pattern>;
773 defm MVE_VMLADAV : MVE_VMLADAV_multi< "8", 0b0, 0b1>;
774 defm MVE_VMLADAV : MVE_VMLADAV_multi<"16", 0b0, 0b0>;
775 defm MVE_VMLADAV : MVE_VMLADAV_multi<"32", 0b1, 0b0>;
777 defm MVE_VMLSDAV : MVE_VMLSDAV_multi< "8", 0b0, 0b1>;
778 defm MVE_VMLSDAV : MVE_VMLSDAV_multi<"16", 0b0, 0b0>;
779 defm MVE_VMLSDAV : MVE_VMLSDAV_multi<"32", 0b1, 0b0>;
781 // vmlav aliases vmladav
782 foreach acc = ["", "a"] in {
783 foreach suffix = ["s8", "s16", "s32", "u8", "u16", "u32"] in {
784 def : MVEInstAlias<"vmlav"#acc#"${vp}."#suffix#"\t$RdaDest, $Qn, $Qm",
785 (!cast<Instruction>("MVE_VMLADAV"#acc#suffix)
786 tGPREven:$RdaDest, MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
790 // Base class for VMLALDAV and VMLSLDAV, VRMLALDAVH, VRMLSLDAVH
791 class MVE_VMLALDAVBase<string iname, string suffix, dag iops, string cstr,
792 bit sz, bit bit_28, bit A, bit X, bit bit_8, bit bit_0,
793 list<dag> pattern=[]>
794 : MVE_rDest<(outs tGPREven:$RdaLoDest, tGPROdd:$RdaHiDest), iops, NoItinerary,
795 iname, suffix, "$RdaLoDest, $RdaHiDest, $Qn, $Qm", cstr, pattern> {
801 let Inst{28} = bit_28;
802 let Inst{22-20} = RdaHiDest{3-1};
803 let Inst{19-17} = Qn{2-0};
805 let Inst{15-13} = RdaLoDest{3-1};
808 let Inst{7-6} = 0b00;
810 let Inst{3-1} = Qm{2-0};
812 let invalidForTailPredication = 1;
815 multiclass MVE_VMLALDAVBase_A<string iname, string x, string suffix,
816 bit sz, bit bit_28, bit X, bit bit_8, bit bit_0,
817 list<dag> pattern=[]> {
818 def ""#x#suffix : MVE_VMLALDAVBase<
819 iname # x, suffix, (ins MQPR:$Qn, MQPR:$Qm), "",
820 sz, bit_28, 0b0, X, bit_8, bit_0, pattern>;
821 def "a"#x#suffix : MVE_VMLALDAVBase<
822 iname # "a" # x, suffix,
823 (ins tGPREven:$RdaLoSrc, tGPROdd:$RdaHiSrc, MQPR:$Qn, MQPR:$Qm),
824 "$RdaLoDest = $RdaLoSrc,$RdaHiDest = $RdaHiSrc",
825 sz, bit_28, 0b1, X, bit_8, bit_0, pattern>;
829 multiclass MVE_VMLALDAVBase_AX<string iname, string suffix, bit sz, bit bit_28,
830 bit bit_8, bit bit_0, list<dag> pattern=[]> {
831 defm "" : MVE_VMLALDAVBase_A<iname, "", suffix, sz,
832 bit_28, 0b0, bit_8, bit_0, pattern>;
833 defm "" : MVE_VMLALDAVBase_A<iname, "x", suffix, sz,
834 bit_28, 0b1, bit_8, bit_0, pattern>;
837 multiclass MVE_VRMLALDAVH_multi<string suffix, list<dag> pattern=[]> {
838 defm "" : MVE_VMLALDAVBase_AX<"vrmlaldavh", "s"#suffix,
839 0b0, 0b0, 0b1, 0b0, pattern>;
840 defm "" : MVE_VMLALDAVBase_A<"vrmlaldavh", "", "u"#suffix,
841 0b0, 0b1, 0b0, 0b1, 0b0, pattern>;
844 defm MVE_VRMLALDAVH : MVE_VRMLALDAVH_multi<"32">;
846 // vrmlalvh aliases for vrmlaldavh
847 def : MVEInstAlias<"vrmlalvh${vp}.s32\t$RdaLo, $RdaHi, $Qn, $Qm",
849 tGPREven:$RdaLo, tGPROdd:$RdaHi,
850 MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
851 def : MVEInstAlias<"vrmlalvha${vp}.s32\t$RdaLo, $RdaHi, $Qn, $Qm",
853 tGPREven:$RdaLo, tGPROdd:$RdaHi,
854 MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
855 def : MVEInstAlias<"vrmlalvh${vp}.u32\t$RdaLo, $RdaHi, $Qn, $Qm",
857 tGPREven:$RdaLo, tGPROdd:$RdaHi,
858 MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
859 def : MVEInstAlias<"vrmlalvha${vp}.u32\t$RdaLo, $RdaHi, $Qn, $Qm",
861 tGPREven:$RdaLo, tGPROdd:$RdaHi,
862 MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
864 multiclass MVE_VMLALDAV_multi<string suffix, bit sz, list<dag> pattern=[]> {
865 defm "" : MVE_VMLALDAVBase_AX<"vmlaldav", "s"#suffix, sz, 0b0, 0b0, 0b0, pattern>;
866 defm "" : MVE_VMLALDAVBase_A<"vmlaldav", "", "u"#suffix,
867 sz, 0b1, 0b0, 0b0, 0b0, pattern>;
870 defm MVE_VMLALDAV : MVE_VMLALDAV_multi<"16", 0b0>;
871 defm MVE_VMLALDAV : MVE_VMLALDAV_multi<"32", 0b1>;
873 // vmlalv aliases vmlaldav
874 foreach acc = ["", "a"] in {
875 foreach suffix = ["s16", "s32", "u16", "u32"] in {
876 def : MVEInstAlias<"vmlalv" # acc # "${vp}." # suffix #
877 "\t$RdaLoDest, $RdaHiDest, $Qn, $Qm",
878 (!cast<Instruction>("MVE_VMLALDAV"#acc#suffix)
879 tGPREven:$RdaLoDest, tGPROdd:$RdaHiDest,
880 MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
884 multiclass MVE_VMLSLDAV_multi<string iname, string suffix, bit sz,
885 bit bit_28, list<dag> pattern=[]> {
886 defm "" : MVE_VMLALDAVBase_AX<iname, suffix, sz, bit_28, 0b0, 0b1, pattern>;
889 defm MVE_VMLSLDAV : MVE_VMLSLDAV_multi<"vmlsldav", "s16", 0b0, 0b0>;
890 defm MVE_VMLSLDAV : MVE_VMLSLDAV_multi<"vmlsldav", "s32", 0b1, 0b0>;
891 defm MVE_VRMLSLDAVH : MVE_VMLSLDAV_multi<"vrmlsldavh", "s32", 0b0, 0b1>;
893 // end of mve_rDest instructions
895 // start of mve_comp instructions
897 class MVE_comp<InstrItinClass itin, string iname, string suffix,
898 string cstr, list<dag> pattern=[]>
899 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm), itin, iname, suffix,
900 "$Qd, $Qn, $Qm", vpred_r, cstr, pattern> {
905 let Inst{22} = Qd{3};
906 let Inst{19-17} = Qn{2-0};
908 let Inst{15-13} = Qd{2-0};
910 let Inst{10-9} = 0b11;
913 let Inst{3-1} = Qm{2-0};
917 class MVE_VMINMAXNM<string iname, string suffix, bit sz, bit bit_21,
918 list<dag> pattern=[]>
919 : MVE_comp<NoItinerary, iname, suffix, "", pattern> {
922 let Inst{25-24} = 0b11;
924 let Inst{21} = bit_21;
931 let Predicates = [HasMVEFloat];
934 def MVE_VMAXNMf32 : MVE_VMINMAXNM<"vmaxnm", "f32", 0b0, 0b0>;
935 def MVE_VMAXNMf16 : MVE_VMINMAXNM<"vmaxnm", "f16", 0b1, 0b0>;
937 let Predicates = [HasMVEFloat] in {
938 def : Pat<(v4f32 (fmaxnum (v4f32 MQPR:$val1), (v4f32 MQPR:$val2))),
939 (v4f32 (MVE_VMAXNMf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2)))>;
940 def : Pat<(v8f16 (fmaxnum (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
941 (v8f16 (MVE_VMAXNMf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
944 def MVE_VMINNMf32 : MVE_VMINMAXNM<"vminnm", "f32", 0b0, 0b1>;
945 def MVE_VMINNMf16 : MVE_VMINMAXNM<"vminnm", "f16", 0b1, 0b1>;
947 let Predicates = [HasMVEFloat] in {
948 def : Pat<(v4f32 (fminnum (v4f32 MQPR:$val1), (v4f32 MQPR:$val2))),
949 (v4f32 (MVE_VMINNMf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2)))>;
950 def : Pat<(v8f16 (fminnum (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
951 (v8f16 (MVE_VMINNMf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
955 class MVE_VMINMAX<string iname, string suffix, bit U, bits<2> size,
956 bit bit_4, list<dag> pattern=[]>
957 : MVE_comp<NoItinerary, iname, suffix, "", pattern> {
960 let Inst{25-24} = 0b11;
962 let Inst{21-20} = size{1-0};
969 multiclass MVE_VMINMAX_all_sizes<string iname, bit bit_4> {
970 def s8 : MVE_VMINMAX<iname, "s8", 0b0, 0b00, bit_4>;
971 def s16 : MVE_VMINMAX<iname, "s16", 0b0, 0b01, bit_4>;
972 def s32 : MVE_VMINMAX<iname, "s32", 0b0, 0b10, bit_4>;
973 def u8 : MVE_VMINMAX<iname, "u8", 0b1, 0b00, bit_4>;
974 def u16 : MVE_VMINMAX<iname, "u16", 0b1, 0b01, bit_4>;
975 def u32 : MVE_VMINMAX<iname, "u32", 0b1, 0b10, bit_4>;
978 defm MVE_VMAX : MVE_VMINMAX_all_sizes<"vmax", 0b0>;
979 defm MVE_VMIN : MVE_VMINMAX_all_sizes<"vmin", 0b1>;
981 let Predicates = [HasMVEInt] in {
982 def : Pat<(v16i8 (smin (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
983 (v16i8 (MVE_VMINs8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
984 def : Pat<(v8i16 (smin (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
985 (v8i16 (MVE_VMINs16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
986 def : Pat<(v4i32 (smin (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
987 (v4i32 (MVE_VMINs32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
989 def : Pat<(v16i8 (smax (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
990 (v16i8 (MVE_VMAXs8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
991 def : Pat<(v8i16 (smax (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
992 (v8i16 (MVE_VMAXs16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
993 def : Pat<(v4i32 (smax (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
994 (v4i32 (MVE_VMAXs32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
996 def : Pat<(v16i8 (umin (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
997 (v16i8 (MVE_VMINu8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
998 def : Pat<(v8i16 (umin (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
999 (v8i16 (MVE_VMINu16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1000 def : Pat<(v4i32 (umin (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1001 (v4i32 (MVE_VMINu32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1003 def : Pat<(v16i8 (umax (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1004 (v16i8 (MVE_VMAXu8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1005 def : Pat<(v8i16 (umax (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1006 (v8i16 (MVE_VMAXu16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1007 def : Pat<(v4i32 (umax (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1008 (v4i32 (MVE_VMAXu32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1011 // end of mve_comp instructions
1013 // start of mve_bit instructions
1015 class MVE_bit_arith<dag oops, dag iops, string iname, string suffix,
1016 string ops, string cstr, list<dag> pattern=[]>
1017 : MVE_p<oops, iops, NoItinerary, iname, suffix, ops, vpred_r, cstr, pattern> {
1021 let Inst{22} = Qd{3};
1022 let Inst{15-13} = Qd{2-0};
1023 let Inst{5} = Qm{3};
1024 let Inst{3-1} = Qm{2-0};
1027 def MVE_VBIC : MVE_bit_arith<(outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm),
1028 "vbic", "", "$Qd, $Qn, $Qm", ""> {
1032 let Inst{25-23} = 0b110;
1033 let Inst{21-20} = 0b01;
1034 let Inst{19-17} = Qn{2-0};
1036 let Inst{12-8} = 0b00001;
1037 let Inst{7} = Qn{3};
1043 class MVE_VREV<string iname, string suffix, bits<2> size, bits<2> bit_8_7, string cstr="">
1044 : MVE_bit_arith<(outs MQPR:$Qd), (ins MQPR:$Qm), iname,
1045 suffix, "$Qd, $Qm", cstr> {
1048 let Inst{25-23} = 0b111;
1049 let Inst{21-20} = 0b11;
1050 let Inst{19-18} = size;
1051 let Inst{17-16} = 0b00;
1052 let Inst{12-9} = 0b0000;
1053 let Inst{8-7} = bit_8_7;
1059 def MVE_VREV64_8 : MVE_VREV<"vrev64", "8", 0b00, 0b00, "@earlyclobber $Qd">;
1060 def MVE_VREV64_16 : MVE_VREV<"vrev64", "16", 0b01, 0b00, "@earlyclobber $Qd">;
1061 def MVE_VREV64_32 : MVE_VREV<"vrev64", "32", 0b10, 0b00, "@earlyclobber $Qd">;
1063 def MVE_VREV32_8 : MVE_VREV<"vrev32", "8", 0b00, 0b01>;
1064 def MVE_VREV32_16 : MVE_VREV<"vrev32", "16", 0b01, 0b01>;
1066 def MVE_VREV16_8 : MVE_VREV<"vrev16", "8", 0b00, 0b10>;
1068 let Predicates = [HasMVEInt] in {
1069 def : Pat<(v8i16 (bswap (v8i16 MQPR:$src))),
1070 (v8i16 (MVE_VREV16_8 (v8i16 MQPR:$src)))>;
1071 def : Pat<(v4i32 (bswap (v4i32 MQPR:$src))),
1072 (v4i32 (MVE_VREV32_8 (v4i32 MQPR:$src)))>;
1075 let Predicates = [HasMVEInt] in {
1076 def : Pat<(v4i32 (ARMvrev64 (v4i32 MQPR:$src))),
1077 (v4i32 (MVE_VREV64_32 (v4i32 MQPR:$src)))>;
1078 def : Pat<(v8i16 (ARMvrev64 (v8i16 MQPR:$src))),
1079 (v8i16 (MVE_VREV64_16 (v8i16 MQPR:$src)))>;
1080 def : Pat<(v16i8 (ARMvrev64 (v16i8 MQPR:$src))),
1081 (v16i8 (MVE_VREV64_8 (v16i8 MQPR:$src)))>;
1083 def : Pat<(v8i16 (ARMvrev32 (v8i16 MQPR:$src))),
1084 (v8i16 (MVE_VREV32_16 (v8i16 MQPR:$src)))>;
1085 def : Pat<(v16i8 (ARMvrev32 (v16i8 MQPR:$src))),
1086 (v16i8 (MVE_VREV32_8 (v16i8 MQPR:$src)))>;
1088 def : Pat<(v16i8 (ARMvrev16 (v16i8 MQPR:$src))),
1089 (v16i8 (MVE_VREV16_8 (v16i8 MQPR:$src)))>;
1091 def : Pat<(v4f32 (ARMvrev64 (v4f32 MQPR:$src))),
1092 (v4f32 (MVE_VREV64_32 (v4f32 MQPR:$src)))>;
1093 def : Pat<(v8f16 (ARMvrev64 (v8f16 MQPR:$src))),
1094 (v8f16 (MVE_VREV64_16 (v8f16 MQPR:$src)))>;
1095 def : Pat<(v8f16 (ARMvrev32 (v8f16 MQPR:$src))),
1096 (v8f16 (MVE_VREV32_16 (v8f16 MQPR:$src)))>;
1099 def MVE_VMVN : MVE_bit_arith<(outs MQPR:$Qd), (ins MQPR:$Qm),
1100 "vmvn", "", "$Qd, $Qm", ""> {
1102 let Inst{25-23} = 0b111;
1103 let Inst{21-16} = 0b110000;
1104 let Inst{12-6} = 0b0010111;
1109 let Predicates = [HasMVEInt] in {
1110 def : Pat<(v16i8 (vnotq (v16i8 MQPR:$val1))),
1111 (v16i8 (MVE_VMVN (v16i8 MQPR:$val1)))>;
1112 def : Pat<(v8i16 (vnotq (v8i16 MQPR:$val1))),
1113 (v8i16 (MVE_VMVN (v8i16 MQPR:$val1)))>;
1114 def : Pat<(v4i32 (vnotq (v4i32 MQPR:$val1))),
1115 (v4i32 (MVE_VMVN (v4i32 MQPR:$val1)))>;
1116 def : Pat<(v2i64 (vnotq (v2i64 MQPR:$val1))),
1117 (v2i64 (MVE_VMVN (v2i64 MQPR:$val1)))>;
1120 class MVE_bit_ops<string iname, bits<2> bit_21_20, bit bit_28>
1121 : MVE_bit_arith<(outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm),
1122 iname, "", "$Qd, $Qn, $Qm", ""> {
1125 let Inst{28} = bit_28;
1126 let Inst{25-23} = 0b110;
1127 let Inst{21-20} = bit_21_20;
1128 let Inst{19-17} = Qn{2-0};
1130 let Inst{12-8} = 0b00001;
1131 let Inst{7} = Qn{3};
1137 def MVE_VEOR : MVE_bit_ops<"veor", 0b00, 0b1>;
1138 def MVE_VORN : MVE_bit_ops<"vorn", 0b11, 0b0>;
1139 def MVE_VORR : MVE_bit_ops<"vorr", 0b10, 0b0>;
1140 def MVE_VAND : MVE_bit_ops<"vand", 0b00, 0b0>;
1142 // add ignored suffixes as aliases
1144 foreach s=["s8", "s16", "s32", "u8", "u16", "u32", "i8", "i16", "i32", "f16", "f32"] in {
1145 def : MVEInstAlias<"vbic${vp}." # s # "\t$QdSrc, $QnSrc, $QmSrc",
1146 (MVE_VBIC MQPR:$QdSrc, MQPR:$QnSrc, MQPR:$QmSrc, vpred_r:$vp)>;
1147 def : MVEInstAlias<"veor${vp}." # s # "\t$QdSrc, $QnSrc, $QmSrc",
1148 (MVE_VEOR MQPR:$QdSrc, MQPR:$QnSrc, MQPR:$QmSrc, vpred_r:$vp)>;
1149 def : MVEInstAlias<"vorn${vp}." # s # "\t$QdSrc, $QnSrc, $QmSrc",
1150 (MVE_VORN MQPR:$QdSrc, MQPR:$QnSrc, MQPR:$QmSrc, vpred_r:$vp)>;
1151 def : MVEInstAlias<"vorr${vp}." # s # "\t$QdSrc, $QnSrc, $QmSrc",
1152 (MVE_VORR MQPR:$QdSrc, MQPR:$QnSrc, MQPR:$QmSrc, vpred_r:$vp)>;
1153 def : MVEInstAlias<"vand${vp}." # s # "\t$QdSrc, $QnSrc, $QmSrc",
1154 (MVE_VAND MQPR:$QdSrc, MQPR:$QnSrc, MQPR:$QmSrc, vpred_r:$vp)>;
1157 let Predicates = [HasMVEInt] in {
1158 def : Pat<(v16i8 (and (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1159 (v16i8 (MVE_VAND (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1160 def : Pat<(v8i16 (and (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1161 (v8i16 (MVE_VAND (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1162 def : Pat<(v4i32 (and (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1163 (v4i32 (MVE_VAND (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1164 def : Pat<(v2i64 (and (v2i64 MQPR:$val1), (v2i64 MQPR:$val2))),
1165 (v2i64 (MVE_VAND (v2i64 MQPR:$val1), (v2i64 MQPR:$val2)))>;
1167 def : Pat<(v16i8 (or (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1168 (v16i8 (MVE_VORR (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1169 def : Pat<(v8i16 (or (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1170 (v8i16 (MVE_VORR (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1171 def : Pat<(v4i32 (or (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1172 (v4i32 (MVE_VORR (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1173 def : Pat<(v2i64 (or (v2i64 MQPR:$val1), (v2i64 MQPR:$val2))),
1174 (v2i64 (MVE_VORR (v2i64 MQPR:$val1), (v2i64 MQPR:$val2)))>;
1176 def : Pat<(v16i8 (xor (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1177 (v16i8 (MVE_VEOR (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1178 def : Pat<(v8i16 (xor (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1179 (v8i16 (MVE_VEOR (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1180 def : Pat<(v4i32 (xor (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1181 (v4i32 (MVE_VEOR (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1182 def : Pat<(v2i64 (xor (v2i64 MQPR:$val1), (v2i64 MQPR:$val2))),
1183 (v2i64 (MVE_VEOR (v2i64 MQPR:$val1), (v2i64 MQPR:$val2)))>;
1185 def : Pat<(v16i8 (and (v16i8 MQPR:$val1), (vnotq MQPR:$val2))),
1186 (v16i8 (MVE_VBIC (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1187 def : Pat<(v8i16 (and (v8i16 MQPR:$val1), (vnotq MQPR:$val2))),
1188 (v8i16 (MVE_VBIC (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1189 def : Pat<(v4i32 (and (v4i32 MQPR:$val1), (vnotq MQPR:$val2))),
1190 (v4i32 (MVE_VBIC (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1191 def : Pat<(v2i64 (and (v2i64 MQPR:$val1), (vnotq MQPR:$val2))),
1192 (v2i64 (MVE_VBIC (v2i64 MQPR:$val1), (v2i64 MQPR:$val2)))>;
1194 def : Pat<(v16i8 (or (v16i8 MQPR:$val1), (vnotq MQPR:$val2))),
1195 (v16i8 (MVE_VORN (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1196 def : Pat<(v8i16 (or (v8i16 MQPR:$val1), (vnotq MQPR:$val2))),
1197 (v8i16 (MVE_VORN (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1198 def : Pat<(v4i32 (or (v4i32 MQPR:$val1), (vnotq MQPR:$val2))),
1199 (v4i32 (MVE_VORN (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1200 def : Pat<(v2i64 (or (v2i64 MQPR:$val1), (vnotq MQPR:$val2))),
1201 (v2i64 (MVE_VORN (v2i64 MQPR:$val1), (v2i64 MQPR:$val2)))>;
1204 class MVE_bit_cmode<string iname, string suffix, bits<4> cmode, dag inOps>
1205 : MVE_p<(outs MQPR:$Qd), inOps, NoItinerary,
1206 iname, suffix, "$Qd, $imm", vpred_n, "$Qd = $Qd_src"> {
1210 let Inst{28} = imm{7};
1211 let Inst{27-23} = 0b11111;
1212 let Inst{22} = Qd{3};
1213 let Inst{21-19} = 0b000;
1214 let Inst{18-16} = imm{6-4};
1215 let Inst{15-13} = Qd{2-0};
1217 let Inst{11-8} = cmode;
1218 let Inst{7-6} = 0b01;
1220 let Inst{3-0} = imm{3-0};
1223 class MVE_VORR<string suffix, bits<4> cmode, ExpandImm imm_type>
1224 : MVE_bit_cmode<"vorr", suffix, cmode, (ins MQPR:$Qd_src, imm_type:$imm)> {
1228 def MVE_VORRIZ0v4i32 : MVE_VORR<"i32", 0b0001, expzero00>;
1229 def MVE_VORRIZ0v8i16 : MVE_VORR<"i16", 0b1001, expzero00>;
1230 def MVE_VORRIZ8v4i32 : MVE_VORR<"i32", 0b0011, expzero08>;
1231 def MVE_VORRIZ8v8i16 : MVE_VORR<"i16", 0b1011, expzero08>;
1232 def MVE_VORRIZ16v4i32 : MVE_VORR<"i32", 0b0101, expzero16>;
1233 def MVE_VORRIZ24v4i32 : MVE_VORR<"i32", 0b0111, expzero24>;
1235 def MVE_VORNIZ0v4i32 : MVEAsmPseudo<"vorn${vp}.i32\t$Qd, $imm",
1236 (ins MQPR:$Qd_src, expzero00inv32:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1237 def MVE_VORNIZ0v8i16 : MVEAsmPseudo<"vorn${vp}.i16\t$Qd, $imm",
1238 (ins MQPR:$Qd_src, expzero00inv16:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1239 def MVE_VORNIZ8v4i32 : MVEAsmPseudo<"vorn${vp}.i32\t$Qd, $imm",
1240 (ins MQPR:$Qd_src, expzero08inv32:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1241 def MVE_VORNIZ8v8i16 : MVEAsmPseudo<"vorn${vp}.i16\t$Qd, $imm",
1242 (ins MQPR:$Qd_src, expzero08inv16:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1243 def MVE_VORNIZ16v4i32 : MVEAsmPseudo<"vorn${vp}.i32\t$Qd, $imm",
1244 (ins MQPR:$Qd_src, expzero16inv32:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1245 def MVE_VORNIZ24v4i32 : MVEAsmPseudo<"vorn${vp}.i32\t$Qd, $imm",
1246 (ins MQPR:$Qd_src, expzero24inv32:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1248 def MVE_VMOV : MVEInstAlias<"vmov${vp}\t$Qd, $Qm",
1249 (MVE_VORR MQPR:$Qd, MQPR:$Qm, MQPR:$Qm, vpred_r:$vp)>;
1251 class MVE_VBIC<string suffix, bits<4> cmode, ExpandImm imm_type>
1252 : MVE_bit_cmode<"vbic", suffix, cmode, (ins MQPR:$Qd_src, imm_type:$imm)> {
1256 def MVE_VBICIZ0v4i32 : MVE_VBIC<"i32", 0b0001, expzero00>;
1257 def MVE_VBICIZ0v8i16 : MVE_VBIC<"i16", 0b1001, expzero00>;
1258 def MVE_VBICIZ8v4i32 : MVE_VBIC<"i32", 0b0011, expzero08>;
1259 def MVE_VBICIZ8v8i16 : MVE_VBIC<"i16", 0b1011, expzero08>;
1260 def MVE_VBICIZ16v4i32 : MVE_VBIC<"i32", 0b0101, expzero16>;
1261 def MVE_VBICIZ24v4i32 : MVE_VBIC<"i32", 0b0111, expzero24>;
1263 def MVE_VANDIZ0v4i32 : MVEAsmPseudo<"vand${vp}.i32\t$Qda, $imm",
1264 (ins MQPR:$Qda_src, expzero00inv32:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1265 def MVE_VANDIZ0v8i16 : MVEAsmPseudo<"vand${vp}.i16\t$Qda, $imm",
1266 (ins MQPR:$Qda_src, expzero00inv16:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1267 def MVE_VANDIZ8v4i32 : MVEAsmPseudo<"vand${vp}.i32\t$Qda, $imm",
1268 (ins MQPR:$Qda_src, expzero08inv32:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1269 def MVE_VANDIZ8v8i16 : MVEAsmPseudo<"vand${vp}.i16\t$Qda, $imm",
1270 (ins MQPR:$Qda_src, expzero08inv16:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1271 def MVE_VANDIZ16v4i32 : MVEAsmPseudo<"vand${vp}.i32\t$Qda, $imm",
1272 (ins MQPR:$Qda_src, expzero16inv32:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1273 def MVE_VANDIZ24v4i32 : MVEAsmPseudo<"vand${vp}.i32\t$Qda, $imm",
1274 (ins MQPR:$Qda_src, expzero24inv32:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1276 class MVE_VMOV_lane_direction {
1283 def MVE_VMOV_from_lane : MVE_VMOV_lane_direction {
1285 let oops = (outs rGPR:$Rt);
1286 let iops = (ins MQPR:$Qd);
1287 let ops = "$Rt, $Qd$Idx";
1290 def MVE_VMOV_to_lane : MVE_VMOV_lane_direction {
1292 let oops = (outs MQPR:$Qd);
1293 let iops = (ins MQPR:$Qd_src, rGPR:$Rt);
1294 let ops = "$Qd$Idx, $Rt";
1295 let cstr = "$Qd = $Qd_src";
1298 class MVE_VMOV_lane<string suffix, bit U, dag indexop,
1299 MVE_VMOV_lane_direction dir>
1300 : MVE_VMOV_lane_base<dir.oops, !con(dir.iops, indexop), NoItinerary,
1301 "vmov", suffix, dir.ops, dir.cstr, []> {
1305 let Inst{31-24} = 0b11101110;
1307 let Inst{20} = dir.bit_20;
1308 let Inst{19-17} = Qd{2-0};
1309 let Inst{15-12} = Rt{3-0};
1310 let Inst{11-8} = 0b1011;
1311 let Inst{7} = Qd{3};
1312 let Inst{4-0} = 0b10000;
1315 class MVE_VMOV_lane_32<MVE_VMOV_lane_direction dir>
1316 : MVE_VMOV_lane<"32", 0b0, (ins MVEVectorIndex<4>:$Idx), dir> {
1319 let Inst{6-5} = 0b00;
1320 let Inst{16} = Idx{1};
1321 let Inst{21} = Idx{0};
1323 let Predicates = [HasFPRegsV8_1M];
1326 class MVE_VMOV_lane_16<string suffix, bit U, MVE_VMOV_lane_direction dir>
1327 : MVE_VMOV_lane<suffix, U, (ins MVEVectorIndex<8>:$Idx), dir> {
1331 let Inst{16} = Idx{2};
1332 let Inst{21} = Idx{1};
1333 let Inst{6} = Idx{0};
1336 class MVE_VMOV_lane_8<string suffix, bit U, MVE_VMOV_lane_direction dir>
1337 : MVE_VMOV_lane<suffix, U, (ins MVEVectorIndex<16>:$Idx), dir> {
1340 let Inst{16} = Idx{3};
1341 let Inst{21} = Idx{2};
1342 let Inst{6} = Idx{1};
1343 let Inst{5} = Idx{0};
1346 def MVE_VMOV_from_lane_32 : MVE_VMOV_lane_32< MVE_VMOV_from_lane>;
1347 def MVE_VMOV_to_lane_32 : MVE_VMOV_lane_32< MVE_VMOV_to_lane>;
1348 def MVE_VMOV_from_lane_s16 : MVE_VMOV_lane_16<"s16", 0b0, MVE_VMOV_from_lane>;
1349 def MVE_VMOV_from_lane_u16 : MVE_VMOV_lane_16<"u16", 0b1, MVE_VMOV_from_lane>;
1350 def MVE_VMOV_to_lane_16 : MVE_VMOV_lane_16< "16", 0b0, MVE_VMOV_to_lane>;
1351 def MVE_VMOV_from_lane_s8 : MVE_VMOV_lane_8 < "s8", 0b0, MVE_VMOV_from_lane>;
1352 def MVE_VMOV_from_lane_u8 : MVE_VMOV_lane_8 < "u8", 0b1, MVE_VMOV_from_lane>;
1353 def MVE_VMOV_to_lane_8 : MVE_VMOV_lane_8 < "8", 0b0, MVE_VMOV_to_lane>;
1355 let Predicates = [HasMVEInt] in {
1356 def : Pat<(extractelt (v2f64 MQPR:$src), imm:$lane),
1357 (f64 (EXTRACT_SUBREG MQPR:$src, (DSubReg_f64_reg imm:$lane)))>;
1358 def : Pat<(insertelt (v2f64 MQPR:$src1), DPR:$src2, imm:$lane),
1359 (INSERT_SUBREG (v2f64 (COPY_TO_REGCLASS MQPR:$src1, MQPR)), DPR:$src2, (DSubReg_f64_reg imm:$lane))>;
1361 def : Pat<(extractelt (v4i32 MQPR:$src), imm:$lane),
1363 (i32 (EXTRACT_SUBREG MQPR:$src, (SSubReg_f32_reg imm:$lane))), rGPR)>;
1364 def : Pat<(insertelt (v4i32 MQPR:$src1), rGPR:$src2, imm:$lane),
1365 (MVE_VMOV_to_lane_32 MQPR:$src1, rGPR:$src2, imm:$lane)>;
1367 def : Pat<(vector_insert (v16i8 MQPR:$src1), rGPR:$src2, imm:$lane),
1368 (MVE_VMOV_to_lane_8 MQPR:$src1, rGPR:$src2, imm:$lane)>;
1369 def : Pat<(vector_insert (v8i16 MQPR:$src1), rGPR:$src2, imm:$lane),
1370 (MVE_VMOV_to_lane_16 MQPR:$src1, rGPR:$src2, imm:$lane)>;
1372 def : Pat<(ARMvgetlanes (v16i8 MQPR:$src), imm:$lane),
1373 (MVE_VMOV_from_lane_s8 MQPR:$src, imm:$lane)>;
1374 def : Pat<(ARMvgetlanes (v8i16 MQPR:$src), imm:$lane),
1375 (MVE_VMOV_from_lane_s16 MQPR:$src, imm:$lane)>;
1376 def : Pat<(ARMvgetlaneu (v16i8 MQPR:$src), imm:$lane),
1377 (MVE_VMOV_from_lane_u8 MQPR:$src, imm:$lane)>;
1378 def : Pat<(ARMvgetlaneu (v8i16 MQPR:$src), imm:$lane),
1379 (MVE_VMOV_from_lane_u16 MQPR:$src, imm:$lane)>;
1381 def : Pat<(v16i8 (scalar_to_vector GPR:$src)),
1382 (MVE_VMOV_to_lane_8 (v16i8 (IMPLICIT_DEF)), rGPR:$src, (i32 0))>;
1383 def : Pat<(v8i16 (scalar_to_vector GPR:$src)),
1384 (MVE_VMOV_to_lane_16 (v8i16 (IMPLICIT_DEF)), rGPR:$src, (i32 0))>;
1385 def : Pat<(v4i32 (scalar_to_vector GPR:$src)),
1386 (MVE_VMOV_to_lane_32 (v4i32 (IMPLICIT_DEF)), rGPR:$src, (i32 0))>;
1388 // Floating point patterns, still enabled under HasMVEInt
1389 def : Pat<(extractelt (v4f32 MQPR:$src), imm:$lane),
1390 (COPY_TO_REGCLASS (f32 (EXTRACT_SUBREG MQPR:$src, (SSubReg_f32_reg imm:$lane))), SPR)>;
1391 def : Pat<(insertelt (v4f32 MQPR:$src1), (f32 SPR:$src2), imm:$lane),
1392 (INSERT_SUBREG (v4f32 (COPY_TO_REGCLASS MQPR:$src1, MQPR)), SPR:$src2, (SSubReg_f32_reg imm:$lane))>;
1394 def : Pat<(insertelt (v8f16 MQPR:$src1), HPR:$src2, imm:$lane),
1395 (MVE_VMOV_to_lane_16 MQPR:$src1, (COPY_TO_REGCLASS HPR:$src2, rGPR), imm:$lane)>;
1396 def : Pat<(extractelt (v8f16 MQPR:$src), imm_even:$lane),
1397 (EXTRACT_SUBREG MQPR:$src, (SSubReg_f16_reg imm_even:$lane))>;
1398 def : Pat<(extractelt (v8f16 MQPR:$src), imm_odd:$lane),
1400 (VMOVH (EXTRACT_SUBREG MQPR:$src, (SSubReg_f16_reg imm_odd:$lane))),
1403 def : Pat<(v4f32 (scalar_to_vector SPR:$src)),
1404 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), SPR:$src, ssub_0)>;
1405 def : Pat<(v4f32 (scalar_to_vector GPR:$src)),
1406 (MVE_VMOV_to_lane_32 (v4f32 (IMPLICIT_DEF)), rGPR:$src, (i32 0))>;
1407 def : Pat<(v8f16 (scalar_to_vector HPR:$src)),
1408 (INSERT_SUBREG (v8f16 (IMPLICIT_DEF)), HPR:$src, ssub_0)>;
1409 def : Pat<(v8f16 (scalar_to_vector GPR:$src)),
1410 (MVE_VMOV_to_lane_16 (v8f16 (IMPLICIT_DEF)), rGPR:$src, (i32 0))>;
1413 // end of mve_bit instructions
1415 // start of MVE Integer instructions
1417 class MVE_int<string iname, string suffix, bits<2> size, list<dag> pattern=[]>
1418 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm), NoItinerary,
1419 iname, suffix, "$Qd, $Qn, $Qm", vpred_r, "", pattern> {
1424 let Inst{22} = Qd{3};
1425 let Inst{21-20} = size;
1426 let Inst{19-17} = Qn{2-0};
1427 let Inst{15-13} = Qd{2-0};
1428 let Inst{7} = Qn{3};
1430 let Inst{5} = Qm{3};
1431 let Inst{3-1} = Qm{2-0};
1434 class MVE_VMULt1<string suffix, bits<2> size, list<dag> pattern=[]>
1435 : MVE_int<"vmul", suffix, size, pattern> {
1438 let Inst{25-23} = 0b110;
1440 let Inst{12-8} = 0b01001;
1445 def MVE_VMULt1i8 : MVE_VMULt1<"i8", 0b00>;
1446 def MVE_VMULt1i16 : MVE_VMULt1<"i16", 0b01>;
1447 def MVE_VMULt1i32 : MVE_VMULt1<"i32", 0b10>;
1449 let Predicates = [HasMVEInt] in {
1450 def : Pat<(v16i8 (mul (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1451 (v16i8 (MVE_VMULt1i8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1452 def : Pat<(v8i16 (mul (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1453 (v8i16 (MVE_VMULt1i16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1454 def : Pat<(v4i32 (mul (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1455 (v4i32 (MVE_VMULt1i32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1458 class MVE_VQxDMULH<string iname, string suffix, bits<2> size, bit rounding,
1459 list<dag> pattern=[]>
1460 : MVE_int<iname, suffix, size, pattern> {
1462 let Inst{28} = rounding;
1463 let Inst{25-23} = 0b110;
1465 let Inst{12-8} = 0b01011;
1470 class MVE_VQDMULH<string suffix, bits<2> size, list<dag> pattern=[]>
1471 : MVE_VQxDMULH<"vqdmulh", suffix, size, 0b0, pattern>;
1472 class MVE_VQRDMULH<string suffix, bits<2> size, list<dag> pattern=[]>
1473 : MVE_VQxDMULH<"vqrdmulh", suffix, size, 0b1, pattern>;
1475 def MVE_VQDMULHi8 : MVE_VQDMULH<"s8", 0b00>;
1476 def MVE_VQDMULHi16 : MVE_VQDMULH<"s16", 0b01>;
1477 def MVE_VQDMULHi32 : MVE_VQDMULH<"s32", 0b10>;
1479 def MVE_VQRDMULHi8 : MVE_VQRDMULH<"s8", 0b00>;
1480 def MVE_VQRDMULHi16 : MVE_VQRDMULH<"s16", 0b01>;
1481 def MVE_VQRDMULHi32 : MVE_VQRDMULH<"s32", 0b10>;
1483 class MVE_VADDSUB<string iname, string suffix, bits<2> size, bit subtract,
1484 list<dag> pattern=[]>
1485 : MVE_int<iname, suffix, size, pattern> {
1487 let Inst{28} = subtract;
1488 let Inst{25-23} = 0b110;
1490 let Inst{12-8} = 0b01000;
1495 class MVE_VADD<string suffix, bits<2> size, list<dag> pattern=[]>
1496 : MVE_VADDSUB<"vadd", suffix, size, 0b0, pattern>;
1497 class MVE_VSUB<string suffix, bits<2> size, list<dag> pattern=[]>
1498 : MVE_VADDSUB<"vsub", suffix, size, 0b1, pattern>;
1500 def MVE_VADDi8 : MVE_VADD<"i8", 0b00>;
1501 def MVE_VADDi16 : MVE_VADD<"i16", 0b01>;
1502 def MVE_VADDi32 : MVE_VADD<"i32", 0b10>;
1504 let Predicates = [HasMVEInt] in {
1505 def : Pat<(v16i8 (add (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1506 (v16i8 (MVE_VADDi8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1507 def : Pat<(v8i16 (add (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1508 (v8i16 (MVE_VADDi16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1509 def : Pat<(v4i32 (add (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1510 (v4i32 (MVE_VADDi32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1513 def MVE_VSUBi8 : MVE_VSUB<"i8", 0b00>;
1514 def MVE_VSUBi16 : MVE_VSUB<"i16", 0b01>;
1515 def MVE_VSUBi32 : MVE_VSUB<"i32", 0b10>;
1517 let Predicates = [HasMVEInt] in {
1518 def : Pat<(v16i8 (sub (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1519 (v16i8 (MVE_VSUBi8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1520 def : Pat<(v8i16 (sub (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1521 (v8i16 (MVE_VSUBi16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1522 def : Pat<(v4i32 (sub (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1523 (v4i32 (MVE_VSUBi32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1526 class MVE_VQADDSUB<string iname, string suffix, bit U, bit subtract,
1527 bits<2> size, ValueType vt>
1528 : MVE_int<iname, suffix, size, []> {
1531 let Inst{25-23} = 0b110;
1533 let Inst{12-10} = 0b000;
1534 let Inst{9} = subtract;
1542 class MVE_VQADD<string suffix, bit U, bits<2> size, ValueType VT>
1543 : MVE_VQADDSUB<"vqadd", suffix, U, 0b0, size, VT>;
1544 class MVE_VQSUB<string suffix, bit U, bits<2> size, ValueType VT>
1545 : MVE_VQADDSUB<"vqsub", suffix, U, 0b1, size, VT>;
1547 def MVE_VQADDs8 : MVE_VQADD<"s8", 0b0, 0b00, v16i8>;
1548 def MVE_VQADDs16 : MVE_VQADD<"s16", 0b0, 0b01, v8i16>;
1549 def MVE_VQADDs32 : MVE_VQADD<"s32", 0b0, 0b10, v4i32>;
1550 def MVE_VQADDu8 : MVE_VQADD<"u8", 0b1, 0b00, v16i8>;
1551 def MVE_VQADDu16 : MVE_VQADD<"u16", 0b1, 0b01, v8i16>;
1552 def MVE_VQADDu32 : MVE_VQADD<"u32", 0b1, 0b10, v4i32>;
1554 def MVE_VQSUBs8 : MVE_VQSUB<"s8", 0b0, 0b00, v16i8>;
1555 def MVE_VQSUBs16 : MVE_VQSUB<"s16", 0b0, 0b01, v8i16>;
1556 def MVE_VQSUBs32 : MVE_VQSUB<"s32", 0b0, 0b10, v4i32>;
1557 def MVE_VQSUBu8 : MVE_VQSUB<"u8", 0b1, 0b00, v16i8>;
1558 def MVE_VQSUBu16 : MVE_VQSUB<"u16", 0b1, 0b01, v8i16>;
1559 def MVE_VQSUBu32 : MVE_VQSUB<"u32", 0b1, 0b10, v4i32>;
1561 let Predicates = [HasMVEInt] in {
1562 foreach instr = [MVE_VQADDu8, MVE_VQADDu16, MVE_VQADDu32] in
1563 foreach VT = [instr.VT] in
1564 def : Pat<(VT (uaddsat (VT MQPR:$Qm), (VT MQPR:$Qn))),
1565 (VT (instr (VT MQPR:$Qm), (VT MQPR:$Qn)))>;
1566 foreach instr = [MVE_VQADDs8, MVE_VQADDs16, MVE_VQADDs32] in
1567 foreach VT = [instr.VT] in
1568 def : Pat<(VT (saddsat (VT MQPR:$Qm), (VT MQPR:$Qn))),
1569 (VT (instr (VT MQPR:$Qm), (VT MQPR:$Qn)))>;
1570 foreach instr = [MVE_VQSUBu8, MVE_VQSUBu16, MVE_VQSUBu32] in
1571 foreach VT = [instr.VT] in
1572 def : Pat<(VT (usubsat (VT MQPR:$Qm), (VT MQPR:$Qn))),
1573 (VT (instr (VT MQPR:$Qm), (VT MQPR:$Qn)))>;
1574 foreach instr = [MVE_VQSUBs8, MVE_VQSUBs16, MVE_VQSUBs32] in
1575 foreach VT = [instr.VT] in
1576 def : Pat<(VT (ssubsat (VT MQPR:$Qm), (VT MQPR:$Qn))),
1577 (VT (instr (VT MQPR:$Qm), (VT MQPR:$Qn)))>;
1581 class MVE_VABD_int<string suffix, bit U, bits<2> size, list<dag> pattern=[]>
1582 : MVE_int<"vabd", suffix, size, pattern> {
1585 let Inst{25-23} = 0b110;
1587 let Inst{12-8} = 0b00111;
1592 def MVE_VABDs8 : MVE_VABD_int<"s8", 0b0, 0b00>;
1593 def MVE_VABDs16 : MVE_VABD_int<"s16", 0b0, 0b01>;
1594 def MVE_VABDs32 : MVE_VABD_int<"s32", 0b0, 0b10>;
1595 def MVE_VABDu8 : MVE_VABD_int<"u8", 0b1, 0b00>;
1596 def MVE_VABDu16 : MVE_VABD_int<"u16", 0b1, 0b01>;
1597 def MVE_VABDu32 : MVE_VABD_int<"u32", 0b1, 0b10>;
1599 class MVE_VRHADD<string suffix, bit U, bits<2> size, list<dag> pattern=[]>
1600 : MVE_int<"vrhadd", suffix, size, pattern> {
1603 let Inst{25-23} = 0b110;
1605 let Inst{12-8} = 0b00001;
1610 def MVE_VRHADDs8 : MVE_VRHADD<"s8", 0b0, 0b00>;
1611 def MVE_VRHADDs16 : MVE_VRHADD<"s16", 0b0, 0b01>;
1612 def MVE_VRHADDs32 : MVE_VRHADD<"s32", 0b0, 0b10>;
1613 def MVE_VRHADDu8 : MVE_VRHADD<"u8", 0b1, 0b00>;
1614 def MVE_VRHADDu16 : MVE_VRHADD<"u16", 0b1, 0b01>;
1615 def MVE_VRHADDu32 : MVE_VRHADD<"u32", 0b1, 0b10>;
1617 class MVE_VHADDSUB<string iname, string suffix, bit U, bit subtract,
1618 bits<2> size, list<dag> pattern=[]>
1619 : MVE_int<iname, suffix, size, pattern> {
1622 let Inst{25-23} = 0b110;
1624 let Inst{12-10} = 0b000;
1625 let Inst{9} = subtract;
1631 class MVE_VHADD<string suffix, bit U, bits<2> size,
1632 list<dag> pattern=[]>
1633 : MVE_VHADDSUB<"vhadd", suffix, U, 0b0, size, pattern>;
1634 class MVE_VHSUB<string suffix, bit U, bits<2> size,
1635 list<dag> pattern=[]>
1636 : MVE_VHADDSUB<"vhsub", suffix, U, 0b1, size, pattern>;
1638 def MVE_VHADDs8 : MVE_VHADD<"s8", 0b0, 0b00>;
1639 def MVE_VHADDs16 : MVE_VHADD<"s16", 0b0, 0b01>;
1640 def MVE_VHADDs32 : MVE_VHADD<"s32", 0b0, 0b10>;
1641 def MVE_VHADDu8 : MVE_VHADD<"u8", 0b1, 0b00>;
1642 def MVE_VHADDu16 : MVE_VHADD<"u16", 0b1, 0b01>;
1643 def MVE_VHADDu32 : MVE_VHADD<"u32", 0b1, 0b10>;
1645 def MVE_VHSUBs8 : MVE_VHSUB<"s8", 0b0, 0b00>;
1646 def MVE_VHSUBs16 : MVE_VHSUB<"s16", 0b0, 0b01>;
1647 def MVE_VHSUBs32 : MVE_VHSUB<"s32", 0b0, 0b10>;
1648 def MVE_VHSUBu8 : MVE_VHSUB<"u8", 0b1, 0b00>;
1649 def MVE_VHSUBu16 : MVE_VHSUB<"u16", 0b1, 0b01>;
1650 def MVE_VHSUBu32 : MVE_VHSUB<"u32", 0b1, 0b10>;
1652 let Predicates = [HasMVEInt] in {
1653 def : Pat<(v16i8 (ARMvshrsImm
1654 (v16i8 (add (v16i8 MQPR:$v1), (v16i8 MQPR:$v2))), 1)),
1656 (v16i8 MQPR:$v1), (v16i8 MQPR:$v2)))>;
1657 def : Pat<(v8i16 (ARMvshrsImm
1658 (v8i16 (add (v8i16 MQPR:$v1), (v8i16 MQPR:$v2))), 1)),
1659 (v8i16 (MVE_VHADDs16
1660 (v8i16 MQPR:$v1), (v8i16 MQPR:$v2)))>;
1661 def : Pat<(v4i32 (ARMvshrsImm
1662 (v4i32 (add (v4i32 MQPR:$v1), (v4i32 MQPR:$v2))), 1)),
1663 (v4i32 (MVE_VHADDs32
1664 (v4i32 MQPR:$v1), (v4i32 MQPR:$v2)))>;
1666 def : Pat<(v16i8 (ARMvshruImm
1667 (v16i8 (add (v16i8 MQPR:$v1), (v16i8 MQPR:$v2))), 1)),
1669 (v16i8 MQPR:$v1), (v16i8 MQPR:$v2)))>;
1670 def : Pat<(v8i16 (ARMvshruImm
1671 (v8i16 (add (v8i16 MQPR:$v1), (v8i16 MQPR:$v2))), 1)),
1672 (v8i16 (MVE_VHADDu16
1673 (v8i16 MQPR:$v1), (v8i16 MQPR:$v2)))>;
1674 def : Pat<(v4i32 (ARMvshruImm
1675 (v4i32 (add (v4i32 MQPR:$v1), (v4i32 MQPR:$v2))), 1)),
1676 (v4i32 (MVE_VHADDu32
1677 (v4i32 MQPR:$v1), (v4i32 MQPR:$v2)))>;
1679 def : Pat<(v16i8 (ARMvshrsImm
1680 (v16i8 (sub (v16i8 MQPR:$v1), (v16i8 MQPR:$v2))), 1)),
1682 (v16i8 MQPR:$v1), (v16i8 MQPR:$v2)))>;
1683 def : Pat<(v8i16 (ARMvshrsImm
1684 (v8i16 (sub (v8i16 MQPR:$v1), (v8i16 MQPR:$v2))), 1)),
1685 (v8i16 (MVE_VHSUBs16
1686 (v8i16 MQPR:$v1), (v8i16 MQPR:$v2)))>;
1687 def : Pat<(v4i32 (ARMvshrsImm
1688 (v4i32 (sub (v4i32 MQPR:$v1), (v4i32 MQPR:$v2))), 1)),
1689 (v4i32 (MVE_VHSUBs32
1690 (v4i32 MQPR:$v1), (v4i32 MQPR:$v2)))>;
1692 def : Pat<(v16i8 (ARMvshruImm
1693 (v16i8 (sub (v16i8 MQPR:$v1), (v16i8 MQPR:$v2))), 1)),
1695 (v16i8 MQPR:$v1), (v16i8 MQPR:$v2)))>;
1696 def : Pat<(v8i16 (ARMvshruImm
1697 (v8i16 (sub (v8i16 MQPR:$v1), (v8i16 MQPR:$v2))), 1)),
1698 (v8i16 (MVE_VHSUBu16
1699 (v8i16 MQPR:$v1), (v8i16 MQPR:$v2)))>;
1700 def : Pat<(v4i32 (ARMvshruImm
1701 (v4i32 (sub (v4i32 MQPR:$v1), (v4i32 MQPR:$v2))), 1)),
1702 (v4i32 (MVE_VHSUBu32
1703 (v4i32 MQPR:$v1), (v4i32 MQPR:$v2)))>;
1706 class MVE_VDUP<string suffix, bit B, bit E, list<dag> pattern=[]>
1707 : MVE_p<(outs MQPR:$Qd), (ins rGPR:$Rt), NoItinerary,
1708 "vdup", suffix, "$Qd, $Rt", vpred_r, "", pattern> {
1713 let Inst{25-23} = 0b101;
1715 let Inst{21-20} = 0b10;
1716 let Inst{19-17} = Qd{2-0};
1718 let Inst{15-12} = Rt;
1719 let Inst{11-8} = 0b1011;
1720 let Inst{7} = Qd{3};
1723 let Inst{4-0} = 0b10000;
1726 def MVE_VDUP32 : MVE_VDUP<"32", 0b0, 0b0>;
1727 def MVE_VDUP16 : MVE_VDUP<"16", 0b0, 0b1>;
1728 def MVE_VDUP8 : MVE_VDUP<"8", 0b1, 0b0>;
1730 let Predicates = [HasMVEInt] in {
1731 def : Pat<(v16i8 (ARMvdup (i32 rGPR:$elem))),
1732 (MVE_VDUP8 rGPR:$elem)>;
1733 def : Pat<(v8i16 (ARMvdup (i32 rGPR:$elem))),
1734 (MVE_VDUP16 rGPR:$elem)>;
1735 def : Pat<(v4i32 (ARMvdup (i32 rGPR:$elem))),
1736 (MVE_VDUP32 rGPR:$elem)>;
1738 def : Pat<(v4i32 (ARMvduplane (v4i32 MQPR:$src), imm:$lane)),
1739 (MVE_VDUP32 (MVE_VMOV_from_lane_32 MQPR:$src, imm:$lane))>;
1740 // For the 16-bit and 8-bit vduplanes we don't care about the signedness
1741 // of the lane move operation as we only want the lowest 8/16 bits anyway.
1742 def : Pat<(v8i16 (ARMvduplane (v8i16 MQPR:$src), imm:$lane)),
1743 (MVE_VDUP16 (MVE_VMOV_from_lane_u16 MQPR:$src, imm:$lane))>;
1744 def : Pat<(v16i8 (ARMvduplane (v16i8 MQPR:$src), imm:$lane)),
1745 (MVE_VDUP8 (MVE_VMOV_from_lane_u8 MQPR:$src, imm:$lane))>;
1747 def : Pat<(v4f32 (ARMvdup (f32 SPR:$elem))),
1748 (v4f32 (MVE_VDUP32 (i32 (COPY_TO_REGCLASS (f32 SPR:$elem), rGPR))))>;
1749 def : Pat<(v8f16 (ARMvdup (f16 HPR:$elem))),
1750 (v8f16 (MVE_VDUP16 (i32 (COPY_TO_REGCLASS (f16 HPR:$elem), rGPR))))>;
1752 def : Pat<(v4f32 (ARMvduplane (v4f32 MQPR:$src), imm:$lane)),
1753 (MVE_VDUP32 (MVE_VMOV_from_lane_32 MQPR:$src, imm:$lane))>;
1754 def : Pat<(v8f16 (ARMvduplane (v8f16 MQPR:$src), imm:$lane)),
1755 (MVE_VDUP16 (MVE_VMOV_from_lane_u16 MQPR:$src, imm:$lane))>;
1759 class MVEIntSingleSrc<string iname, string suffix, bits<2> size,
1760 list<dag> pattern=[]>
1761 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qm), NoItinerary,
1762 iname, suffix, "$Qd, $Qm", vpred_r, "", pattern> {
1766 let Inst{22} = Qd{3};
1767 let Inst{19-18} = size{1-0};
1768 let Inst{15-13} = Qd{2-0};
1769 let Inst{5} = Qm{3};
1770 let Inst{3-1} = Qm{2-0};
1773 class MVE_VCLSCLZ<string iname, string suffix, bits<2> size,
1774 bit count_zeroes, list<dag> pattern=[]>
1775 : MVEIntSingleSrc<iname, suffix, size, pattern> {
1778 let Inst{25-23} = 0b111;
1779 let Inst{21-20} = 0b11;
1780 let Inst{17-16} = 0b00;
1781 let Inst{12-8} = 0b00100;
1782 let Inst{7} = count_zeroes;
1788 def MVE_VCLSs8 : MVE_VCLSCLZ<"vcls", "s8", 0b00, 0b0>;
1789 def MVE_VCLSs16 : MVE_VCLSCLZ<"vcls", "s16", 0b01, 0b0>;
1790 def MVE_VCLSs32 : MVE_VCLSCLZ<"vcls", "s32", 0b10, 0b0>;
1792 def MVE_VCLZs8 : MVE_VCLSCLZ<"vclz", "i8", 0b00, 0b1>;
1793 def MVE_VCLZs16 : MVE_VCLSCLZ<"vclz", "i16", 0b01, 0b1>;
1794 def MVE_VCLZs32 : MVE_VCLSCLZ<"vclz", "i32", 0b10, 0b1>;
1796 let Predicates = [HasMVEInt] in {
1797 def : Pat<(v16i8 ( ctlz (v16i8 MQPR:$val1))),
1798 (v16i8 ( MVE_VCLZs8 (v16i8 MQPR:$val1)))>;
1799 def : Pat<(v4i32 ( ctlz (v4i32 MQPR:$val1))),
1800 (v4i32 ( MVE_VCLZs32 (v4i32 MQPR:$val1)))>;
1801 def : Pat<(v8i16 ( ctlz (v8i16 MQPR:$val1))),
1802 (v8i16 ( MVE_VCLZs16 (v8i16 MQPR:$val1)))>;
1805 class MVE_VABSNEG_int<string iname, string suffix, bits<2> size, bit negate,
1806 list<dag> pattern=[]>
1807 : MVEIntSingleSrc<iname, suffix, size, pattern> {
1810 let Inst{25-23} = 0b111;
1811 let Inst{21-20} = 0b11;
1812 let Inst{17-16} = 0b01;
1813 let Inst{12-8} = 0b00011;
1814 let Inst{7} = negate;
1820 def MVE_VABSs8 : MVE_VABSNEG_int<"vabs", "s8", 0b00, 0b0>;
1821 def MVE_VABSs16 : MVE_VABSNEG_int<"vabs", "s16", 0b01, 0b0>;
1822 def MVE_VABSs32 : MVE_VABSNEG_int<"vabs", "s32", 0b10, 0b0>;
1824 let Predicates = [HasMVEInt] in {
1825 def : Pat<(v16i8 (abs (v16i8 MQPR:$v))),
1826 (v16i8 (MVE_VABSs8 $v))>;
1827 def : Pat<(v8i16 (abs (v8i16 MQPR:$v))),
1828 (v8i16 (MVE_VABSs16 $v))>;
1829 def : Pat<(v4i32 (abs (v4i32 MQPR:$v))),
1830 (v4i32 (MVE_VABSs32 $v))>;
1833 def MVE_VNEGs8 : MVE_VABSNEG_int<"vneg", "s8", 0b00, 0b1>;
1834 def MVE_VNEGs16 : MVE_VABSNEG_int<"vneg", "s16", 0b01, 0b1>;
1835 def MVE_VNEGs32 : MVE_VABSNEG_int<"vneg", "s32", 0b10, 0b1>;
1837 let Predicates = [HasMVEInt] in {
1838 def : Pat<(v16i8 (vnegq (v16i8 MQPR:$v))),
1839 (v16i8 (MVE_VNEGs8 $v))>;
1840 def : Pat<(v8i16 (vnegq (v8i16 MQPR:$v))),
1841 (v8i16 (MVE_VNEGs16 $v))>;
1842 def : Pat<(v4i32 (vnegq (v4i32 MQPR:$v))),
1843 (v4i32 (MVE_VNEGs32 $v))>;
1846 class MVE_VQABSNEG<string iname, string suffix, bits<2> size,
1847 bit negate, list<dag> pattern=[]>
1848 : MVEIntSingleSrc<iname, suffix, size, pattern> {
1851 let Inst{25-23} = 0b111;
1852 let Inst{21-20} = 0b11;
1853 let Inst{17-16} = 0b00;
1854 let Inst{12-8} = 0b00111;
1855 let Inst{7} = negate;
1861 def MVE_VQABSs8 : MVE_VQABSNEG<"vqabs", "s8", 0b00, 0b0>;
1862 def MVE_VQABSs16 : MVE_VQABSNEG<"vqabs", "s16", 0b01, 0b0>;
1863 def MVE_VQABSs32 : MVE_VQABSNEG<"vqabs", "s32", 0b10, 0b0>;
1865 def MVE_VQNEGs8 : MVE_VQABSNEG<"vqneg", "s8", 0b00, 0b1>;
1866 def MVE_VQNEGs16 : MVE_VQABSNEG<"vqneg", "s16", 0b01, 0b1>;
1867 def MVE_VQNEGs32 : MVE_VQABSNEG<"vqneg", "s32", 0b10, 0b1>;
1869 class MVE_mod_imm<string iname, string suffix, bits<4> cmode, bit op,
1870 dag iops, list<dag> pattern=[]>
1871 : MVE_p<(outs MQPR:$Qd), iops, NoItinerary, iname, suffix, "$Qd, $imm",
1872 vpred_r, "", pattern> {
1876 let Inst{28} = imm{7};
1877 let Inst{25-23} = 0b111;
1878 let Inst{22} = Qd{3};
1879 let Inst{21-19} = 0b000;
1880 let Inst{18-16} = imm{6-4};
1881 let Inst{15-13} = Qd{2-0};
1883 let Inst{11-8} = cmode{3-0};
1884 let Inst{7-6} = 0b01;
1887 let Inst{3-0} = imm{3-0};
1889 let DecoderMethod = "DecodeMVEModImmInstruction";
1892 let isReMaterializable = 1 in {
1893 let isAsCheapAsAMove = 1 in {
1894 def MVE_VMOVimmi8 : MVE_mod_imm<"vmov", "i8", {1,1,1,0}, 0b0, (ins nImmSplatI8:$imm)>;
1895 def MVE_VMOVimmi16 : MVE_mod_imm<"vmov", "i16", {1,0,?,0}, 0b0, (ins nImmSplatI16:$imm)> {
1896 let Inst{9} = imm{9};
1898 def MVE_VMOVimmi32 : MVE_mod_imm<"vmov", "i32", {?,?,?,?}, 0b0, (ins nImmVMOVI32:$imm)> {
1899 let Inst{11-8} = imm{11-8};
1901 def MVE_VMOVimmi64 : MVE_mod_imm<"vmov", "i64", {1,1,1,0}, 0b1, (ins nImmSplatI64:$imm)>;
1902 def MVE_VMOVimmf32 : MVE_mod_imm<"vmov", "f32", {1,1,1,1}, 0b0, (ins nImmVMOVF32:$imm)>;
1903 } // let isAsCheapAsAMove = 1
1905 def MVE_VMVNimmi16 : MVE_mod_imm<"vmvn", "i16", {1,0,?,0}, 0b1, (ins nImmSplatI16:$imm)> {
1906 let Inst{9} = imm{9};
1908 def MVE_VMVNimmi32 : MVE_mod_imm<"vmvn", "i32", {?,?,?,?}, 0b1, (ins nImmVMOVI32:$imm)> {
1909 let Inst{11-8} = imm{11-8};
1911 } // let isReMaterializable = 1
1913 let Predicates = [HasMVEInt] in {
1914 def : Pat<(v16i8 (ARMvmovImm timm:$simm)),
1915 (v16i8 (MVE_VMOVimmi8 nImmSplatI8:$simm))>;
1916 def : Pat<(v8i16 (ARMvmovImm timm:$simm)),
1917 (v8i16 (MVE_VMOVimmi16 nImmSplatI16:$simm))>;
1918 def : Pat<(v4i32 (ARMvmovImm timm:$simm)),
1919 (v4i32 (MVE_VMOVimmi32 nImmVMOVI32:$simm))>;
1921 def : Pat<(v8i16 (ARMvmvnImm timm:$simm)),
1922 (v8i16 (MVE_VMVNimmi16 nImmSplatI16:$simm))>;
1923 def : Pat<(v4i32 (ARMvmvnImm timm:$simm)),
1924 (v4i32 (MVE_VMVNimmi32 nImmVMOVI32:$simm))>;
1926 def : Pat<(v4f32 (ARMvmovFPImm timm:$simm)),
1927 (v4f32 (MVE_VMOVimmf32 nImmVMOVF32:$simm))>;
1930 class MVE_VMINMAXA<string iname, string suffix, bits<2> size,
1931 bit bit_12, list<dag> pattern=[]>
1932 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qd_src, MQPR:$Qm),
1933 NoItinerary, iname, suffix, "$Qd, $Qm", vpred_n, "$Qd = $Qd_src",
1939 let Inst{25-23} = 0b100;
1940 let Inst{22} = Qd{3};
1941 let Inst{21-20} = 0b11;
1942 let Inst{19-18} = size;
1943 let Inst{17-16} = 0b11;
1944 let Inst{15-13} = Qd{2-0};
1945 let Inst{12} = bit_12;
1946 let Inst{11-6} = 0b111010;
1947 let Inst{5} = Qm{3};
1949 let Inst{3-1} = Qm{2-0};
1953 def MVE_VMAXAs8 : MVE_VMINMAXA<"vmaxa", "s8", 0b00, 0b0>;
1954 def MVE_VMAXAs16 : MVE_VMINMAXA<"vmaxa", "s16", 0b01, 0b0>;
1955 def MVE_VMAXAs32 : MVE_VMINMAXA<"vmaxa", "s32", 0b10, 0b0>;
1957 def MVE_VMINAs8 : MVE_VMINMAXA<"vmina", "s8", 0b00, 0b1>;
1958 def MVE_VMINAs16 : MVE_VMINMAXA<"vmina", "s16", 0b01, 0b1>;
1959 def MVE_VMINAs32 : MVE_VMINMAXA<"vmina", "s32", 0b10, 0b1>;
1961 // end of MVE Integer instructions
1963 // start of mve_imm_shift instructions
1965 def MVE_VSHLC : MVE_p<(outs rGPR:$RdmDest, MQPR:$Qd),
1966 (ins MQPR:$QdSrc, rGPR:$RdmSrc, long_shift:$imm),
1967 NoItinerary, "vshlc", "", "$QdSrc, $RdmSrc, $imm",
1968 vpred_n, "$RdmDest = $RdmSrc,$Qd = $QdSrc"> {
1974 let Inst{25-23} = 0b101;
1975 let Inst{22} = Qd{3};
1977 let Inst{20-16} = imm{4-0};
1978 let Inst{15-13} = Qd{2-0};
1979 let Inst{12-4} = 0b011111100;
1980 let Inst{3-0} = RdmDest{3-0};
1983 class MVE_shift_imm<dag oops, dag iops, string iname, string suffix,
1984 string ops, vpred_ops vpred, string cstr,
1985 list<dag> pattern=[]>
1986 : MVE_p<oops, iops, NoItinerary, iname, suffix, ops, vpred, cstr, pattern> {
1990 let Inst{22} = Qd{3};
1991 let Inst{15-13} = Qd{2-0};
1992 let Inst{5} = Qm{3};
1993 let Inst{3-1} = Qm{2-0};
1996 class MVE_VMOVL<string iname, string suffix, bits<2> sz, bit U,
1997 list<dag> pattern=[]>
1998 : MVE_shift_imm<(outs MQPR:$Qd), (ins MQPR:$Qm),
1999 iname, suffix, "$Qd, $Qm", vpred_r, "",
2002 let Inst{25-23} = 0b101;
2004 let Inst{20-19} = sz{1-0};
2005 let Inst{18-16} = 0b000;
2006 let Inst{11-6} = 0b111101;
2011 multiclass MVE_VMOVL_shift_half<string iname, string suffix, bits<2> sz, bit U,
2012 list<dag> pattern=[]> {
2013 def bh : MVE_VMOVL<!strconcat(iname, "b"), suffix, sz, U, pattern> {
2016 def th : MVE_VMOVL<!strconcat(iname, "t"), suffix, sz, U, pattern> {
2021 defm MVE_VMOVLs8 : MVE_VMOVL_shift_half<"vmovl", "s8", 0b01, 0b0>;
2022 defm MVE_VMOVLu8 : MVE_VMOVL_shift_half<"vmovl", "u8", 0b01, 0b1>;
2023 defm MVE_VMOVLs16 : MVE_VMOVL_shift_half<"vmovl", "s16", 0b10, 0b0>;
2024 defm MVE_VMOVLu16 : MVE_VMOVL_shift_half<"vmovl", "u16", 0b10, 0b1>;
2026 let Predicates = [HasMVEInt] in {
2027 def : Pat<(sext_inreg (v4i32 MQPR:$src), v4i16),
2028 (MVE_VMOVLs16bh MQPR:$src)>;
2029 def : Pat<(sext_inreg (v8i16 MQPR:$src), v8i8),
2030 (MVE_VMOVLs8bh MQPR:$src)>;
2031 def : Pat<(sext_inreg (v4i32 MQPR:$src), v4i8),
2032 (MVE_VMOVLs16bh (MVE_VMOVLs8bh MQPR:$src))>;
2034 // zext_inreg 16 -> 32
2035 def : Pat<(and (v4i32 MQPR:$src), (v4i32 (ARMvmovImm (i32 0xCFF)))),
2036 (MVE_VMOVLu16bh MQPR:$src)>;
2037 // zext_inreg 8 -> 16
2038 def : Pat<(and (v8i16 MQPR:$src), (v8i16 (ARMvmovImm (i32 0x8FF)))),
2039 (MVE_VMOVLu8bh MQPR:$src)>;
2043 class MVE_VSHLL_imm<string iname, string suffix, bit U, bit th,
2044 dag immops, list<dag> pattern=[]>
2045 : MVE_shift_imm<(outs MQPR:$Qd), !con((ins MQPR:$Qm), immops),
2046 iname, suffix, "$Qd, $Qm, $imm", vpred_r, "", pattern> {
2048 let Inst{25-23} = 0b101;
2051 let Inst{11-6} = 0b111101;
2056 // The immediate VSHLL instructions accept shift counts from 1 up to
2057 // the lane width (8 or 16), but the full-width shifts have an
2058 // entirely separate encoding, given below with 'lw' in the name.
2060 class MVE_VSHLL_imm8<string iname, string suffix,
2061 bit U, bit th, list<dag> pattern=[]>
2062 : MVE_VSHLL_imm<iname, suffix, U, th, (ins mve_shift_imm1_7:$imm), pattern> {
2064 let Inst{20-19} = 0b01;
2065 let Inst{18-16} = imm;
2068 class MVE_VSHLL_imm16<string iname, string suffix,
2069 bit U, bit th, list<dag> pattern=[]>
2070 : MVE_VSHLL_imm<iname, suffix, U, th, (ins mve_shift_imm1_15:$imm), pattern> {
2073 let Inst{19-16} = imm;
2076 def MVE_VSHLL_imms8bh : MVE_VSHLL_imm8 <"vshllb", "s8", 0b0, 0b0>;
2077 def MVE_VSHLL_imms8th : MVE_VSHLL_imm8 <"vshllt", "s8", 0b0, 0b1>;
2078 def MVE_VSHLL_immu8bh : MVE_VSHLL_imm8 <"vshllb", "u8", 0b1, 0b0>;
2079 def MVE_VSHLL_immu8th : MVE_VSHLL_imm8 <"vshllt", "u8", 0b1, 0b1>;
2080 def MVE_VSHLL_imms16bh : MVE_VSHLL_imm16<"vshllb", "s16", 0b0, 0b0>;
2081 def MVE_VSHLL_imms16th : MVE_VSHLL_imm16<"vshllt", "s16", 0b0, 0b1>;
2082 def MVE_VSHLL_immu16bh : MVE_VSHLL_imm16<"vshllb", "u16", 0b1, 0b0>;
2083 def MVE_VSHLL_immu16th : MVE_VSHLL_imm16<"vshllt", "u16", 0b1, 0b1>;
2085 class MVE_VSHLL_by_lane_width<string iname, string suffix, bits<2> size,
2086 bit U, string ops, list<dag> pattern=[]>
2087 : MVE_shift_imm<(outs MQPR:$Qd), (ins MQPR:$Qm),
2088 iname, suffix, ops, vpred_r, "", pattern> {
2090 let Inst{25-23} = 0b100;
2091 let Inst{21-20} = 0b11;
2092 let Inst{19-18} = size{1-0};
2093 let Inst{17-16} = 0b01;
2094 let Inst{11-6} = 0b111000;
2099 multiclass MVE_VSHLL_lw<string iname, string suffix, bits<2> sz, bit U,
2100 string ops, list<dag> pattern=[]> {
2101 def bh : MVE_VSHLL_by_lane_width<iname#"b", suffix, sz, U, ops, pattern> {
2104 def th : MVE_VSHLL_by_lane_width<iname#"t", suffix, sz, U, ops, pattern> {
2109 defm MVE_VSHLL_lws8 : MVE_VSHLL_lw<"vshll", "s8", 0b00, 0b0, "$Qd, $Qm, #8">;
2110 defm MVE_VSHLL_lws16 : MVE_VSHLL_lw<"vshll", "s16", 0b01, 0b0, "$Qd, $Qm, #16">;
2111 defm MVE_VSHLL_lwu8 : MVE_VSHLL_lw<"vshll", "u8", 0b00, 0b1, "$Qd, $Qm, #8">;
2112 defm MVE_VSHLL_lwu16 : MVE_VSHLL_lw<"vshll", "u16", 0b01, 0b1, "$Qd, $Qm, #16">;
2114 class MVE_VxSHRN<string iname, string suffix, bit bit_12, bit bit_28,
2115 dag immops, list<dag> pattern=[]>
2116 : MVE_shift_imm<(outs MQPR:$Qd), !con((ins MQPR:$QdSrc, MQPR:$Qm), immops),
2117 iname, suffix, "$Qd, $Qm, $imm", vpred_n, "$Qd = $QdSrc",
2121 let Inst{28} = bit_28;
2122 let Inst{25-23} = 0b101;
2124 let Inst{20-16} = imm{4-0};
2125 let Inst{12} = bit_12;
2126 let Inst{11-6} = 0b111111;
2131 def MVE_VRSHRNi16bh : MVE_VxSHRN<
2132 "vrshrnb", "i16", 0b0, 0b1, (ins shr_imm8:$imm)> {
2133 let Inst{20-19} = 0b01;
2135 def MVE_VRSHRNi16th : MVE_VxSHRN<
2136 "vrshrnt", "i16", 0b1, 0b1,(ins shr_imm8:$imm)> {
2137 let Inst{20-19} = 0b01;
2139 def MVE_VRSHRNi32bh : MVE_VxSHRN<
2140 "vrshrnb", "i32", 0b0, 0b1, (ins shr_imm16:$imm)> {
2143 def MVE_VRSHRNi32th : MVE_VxSHRN<
2144 "vrshrnt", "i32", 0b1, 0b1, (ins shr_imm16:$imm)> {
2148 def MVE_VSHRNi16bh : MVE_VxSHRN<
2149 "vshrnb", "i16", 0b0, 0b0, (ins shr_imm8:$imm)> {
2150 let Inst{20-19} = 0b01;
2152 def MVE_VSHRNi16th : MVE_VxSHRN<
2153 "vshrnt", "i16", 0b1, 0b0, (ins shr_imm8:$imm)> {
2154 let Inst{20-19} = 0b01;
2156 def MVE_VSHRNi32bh : MVE_VxSHRN<
2157 "vshrnb", "i32", 0b0, 0b0, (ins shr_imm16:$imm)> {
2160 def MVE_VSHRNi32th : MVE_VxSHRN<
2161 "vshrnt", "i32", 0b1, 0b0, (ins shr_imm16:$imm)> {
2165 class MVE_VxQRSHRUN<string iname, string suffix, bit bit_28, bit bit_12, dag immops,
2166 list<dag> pattern=[]>
2167 : MVE_shift_imm<(outs MQPR:$Qd), !con((ins MQPR:$QdSrc, MQPR:$Qm), immops),
2168 iname, suffix, "$Qd, $Qm, $imm", vpred_n, "$Qd = $QdSrc",
2172 let Inst{28} = bit_28;
2173 let Inst{25-23} = 0b101;
2175 let Inst{20-16} = imm{4-0};
2176 let Inst{12} = bit_12;
2177 let Inst{11-6} = 0b111111;
2182 def MVE_VQRSHRUNs16bh : MVE_VxQRSHRUN<
2183 "vqrshrunb", "s16", 0b1, 0b0, (ins shr_imm8:$imm)> {
2184 let Inst{20-19} = 0b01;
2186 def MVE_VQRSHRUNs16th : MVE_VxQRSHRUN<
2187 "vqrshrunt", "s16", 0b1, 0b1, (ins shr_imm8:$imm)> {
2188 let Inst{20-19} = 0b01;
2190 def MVE_VQRSHRUNs32bh : MVE_VxQRSHRUN<
2191 "vqrshrunb", "s32", 0b1, 0b0, (ins shr_imm16:$imm)> {
2194 def MVE_VQRSHRUNs32th : MVE_VxQRSHRUN<
2195 "vqrshrunt", "s32", 0b1, 0b1, (ins shr_imm16:$imm)> {
2199 def MVE_VQSHRUNs16bh : MVE_VxQRSHRUN<
2200 "vqshrunb", "s16", 0b0, 0b0, (ins shr_imm8:$imm)> {
2201 let Inst{20-19} = 0b01;
2203 def MVE_VQSHRUNs16th : MVE_VxQRSHRUN<
2204 "vqshrunt", "s16", 0b0, 0b1, (ins shr_imm8:$imm)> {
2205 let Inst{20-19} = 0b01;
2207 def MVE_VQSHRUNs32bh : MVE_VxQRSHRUN<
2208 "vqshrunb", "s32", 0b0, 0b0, (ins shr_imm16:$imm)> {
2211 def MVE_VQSHRUNs32th : MVE_VxQRSHRUN<
2212 "vqshrunt", "s32", 0b0, 0b1, (ins shr_imm16:$imm)> {
2216 class MVE_VxQRSHRN<string iname, string suffix, bit bit_0, bit bit_12,
2217 dag immops, list<dag> pattern=[]>
2218 : MVE_shift_imm<(outs MQPR:$Qd), !con((ins MQPR:$QdSrc, MQPR:$Qm), immops),
2219 iname, suffix, "$Qd, $Qm, $imm", vpred_n, "$Qd = $QdSrc",
2223 let Inst{25-23} = 0b101;
2225 let Inst{20-16} = imm{4-0};
2226 let Inst{12} = bit_12;
2227 let Inst{11-6} = 0b111101;
2229 let Inst{0} = bit_0;
2232 multiclass MVE_VxQRSHRN_types<string iname, bit bit_0, bit bit_12> {
2233 def s16 : MVE_VxQRSHRN<iname, "s16", bit_0, bit_12, (ins shr_imm8:$imm)> {
2235 let Inst{20-19} = 0b01;
2237 def u16 : MVE_VxQRSHRN<iname, "u16", bit_0, bit_12, (ins shr_imm8:$imm)> {
2239 let Inst{20-19} = 0b01;
2241 def s32 : MVE_VxQRSHRN<iname, "s32", bit_0, bit_12, (ins shr_imm16:$imm)> {
2245 def u32 : MVE_VxQRSHRN<iname, "u32", bit_0, bit_12, (ins shr_imm16:$imm)> {
2251 defm MVE_VQRSHRNbh : MVE_VxQRSHRN_types<"vqrshrnb", 0b1, 0b0>;
2252 defm MVE_VQRSHRNth : MVE_VxQRSHRN_types<"vqrshrnt", 0b1, 0b1>;
2253 defm MVE_VQSHRNbh : MVE_VxQRSHRN_types<"vqshrnb", 0b0, 0b0>;
2254 defm MVE_VQSHRNth : MVE_VxQRSHRN_types<"vqshrnt", 0b0, 0b1>;
2256 // end of mve_imm_shift instructions
2258 // start of mve_shift instructions
2260 class MVE_shift_by_vec<string iname, string suffix, bit U,
2261 bits<2> size, bit bit_4, bit bit_8>
2262 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qm, MQPR:$Qn), NoItinerary,
2263 iname, suffix, "$Qd, $Qm, $Qn", vpred_r, "", []> {
2264 // Shift instructions which take a vector of shift counts
2270 let Inst{25-24} = 0b11;
2272 let Inst{22} = Qd{3};
2273 let Inst{21-20} = size;
2274 let Inst{19-17} = Qn{2-0};
2276 let Inst{15-13} = Qd{2-0};
2277 let Inst{12-9} = 0b0010;
2278 let Inst{8} = bit_8;
2279 let Inst{7} = Qn{3};
2281 let Inst{5} = Qm{3};
2282 let Inst{4} = bit_4;
2283 let Inst{3-1} = Qm{2-0};
2287 multiclass mve_shift_by_vec_multi<string iname, bit bit_4, bit bit_8> {
2288 def s8 : MVE_shift_by_vec<iname, "s8", 0b0, 0b00, bit_4, bit_8>;
2289 def s16 : MVE_shift_by_vec<iname, "s16", 0b0, 0b01, bit_4, bit_8>;
2290 def s32 : MVE_shift_by_vec<iname, "s32", 0b0, 0b10, bit_4, bit_8>;
2291 def u8 : MVE_shift_by_vec<iname, "u8", 0b1, 0b00, bit_4, bit_8>;
2292 def u16 : MVE_shift_by_vec<iname, "u16", 0b1, 0b01, bit_4, bit_8>;
2293 def u32 : MVE_shift_by_vec<iname, "u32", 0b1, 0b10, bit_4, bit_8>;
2296 defm MVE_VSHL_by_vec : mve_shift_by_vec_multi<"vshl", 0b0, 0b0>;
2297 defm MVE_VQSHL_by_vec : mve_shift_by_vec_multi<"vqshl", 0b1, 0b0>;
2298 defm MVE_VQRSHL_by_vec : mve_shift_by_vec_multi<"vqrshl", 0b1, 0b1>;
2299 defm MVE_VRSHL_by_vec : mve_shift_by_vec_multi<"vrshl", 0b0, 0b1>;
2301 let Predicates = [HasMVEInt] in {
2302 def : Pat<(v4i32 (ARMvshlu (v4i32 MQPR:$Qm), (v4i32 MQPR:$Qn))),
2303 (v4i32 (MVE_VSHL_by_vecu32 (v4i32 MQPR:$Qm), (v4i32 MQPR:$Qn)))>;
2304 def : Pat<(v8i16 (ARMvshlu (v8i16 MQPR:$Qm), (v8i16 MQPR:$Qn))),
2305 (v8i16 (MVE_VSHL_by_vecu16 (v8i16 MQPR:$Qm), (v8i16 MQPR:$Qn)))>;
2306 def : Pat<(v16i8 (ARMvshlu (v16i8 MQPR:$Qm), (v16i8 MQPR:$Qn))),
2307 (v16i8 (MVE_VSHL_by_vecu8 (v16i8 MQPR:$Qm), (v16i8 MQPR:$Qn)))>;
2309 def : Pat<(v4i32 (ARMvshls (v4i32 MQPR:$Qm), (v4i32 MQPR:$Qn))),
2310 (v4i32 (MVE_VSHL_by_vecs32 (v4i32 MQPR:$Qm), (v4i32 MQPR:$Qn)))>;
2311 def : Pat<(v8i16 (ARMvshls (v8i16 MQPR:$Qm), (v8i16 MQPR:$Qn))),
2312 (v8i16 (MVE_VSHL_by_vecs16 (v8i16 MQPR:$Qm), (v8i16 MQPR:$Qn)))>;
2313 def : Pat<(v16i8 (ARMvshls (v16i8 MQPR:$Qm), (v16i8 MQPR:$Qn))),
2314 (v16i8 (MVE_VSHL_by_vecs8 (v16i8 MQPR:$Qm), (v16i8 MQPR:$Qn)))>;
2317 class MVE_shift_with_imm<string iname, string suffix, dag oops, dag iops,
2318 string ops, vpred_ops vpred, string cstr,
2319 list<dag> pattern=[]>
2320 : MVE_p<oops, iops, NoItinerary, iname, suffix, ops, vpred, cstr, pattern> {
2325 let Inst{22} = Qd{3};
2326 let Inst{15-13} = Qd{2-0};
2327 let Inst{12-11} = 0b00;
2328 let Inst{7-6} = 0b01;
2329 let Inst{5} = Qm{3};
2331 let Inst{3-1} = Qm{2-0};
2335 class MVE_VSxI_imm<string iname, string suffix, bit bit_8, dag imm>
2336 : MVE_shift_with_imm<iname, suffix, (outs MQPR:$Qd),
2337 !con((ins MQPR:$Qd_src, MQPR:$Qm), imm),
2338 "$Qd, $Qm, $imm", vpred_n, "$Qd = $Qd_src"> {
2341 let Inst{25-24} = 0b11;
2342 let Inst{21-16} = imm;
2343 let Inst{10-9} = 0b10;
2344 let Inst{8} = bit_8;
2347 def MVE_VSRIimm8 : MVE_VSxI_imm<"vsri", "8", 0b0, (ins shr_imm8:$imm)> {
2348 let Inst{21-19} = 0b001;
2351 def MVE_VSRIimm16 : MVE_VSxI_imm<"vsri", "16", 0b0, (ins shr_imm16:$imm)> {
2352 let Inst{21-20} = 0b01;
2355 def MVE_VSRIimm32 : MVE_VSxI_imm<"vsri", "32", 0b0, (ins shr_imm32:$imm)> {
2359 def MVE_VSLIimm8 : MVE_VSxI_imm<"vsli", "8", 0b1, (ins imm0_7:$imm)> {
2360 let Inst{21-19} = 0b001;
2363 def MVE_VSLIimm16 : MVE_VSxI_imm<"vsli", "16", 0b1, (ins imm0_15:$imm)> {
2364 let Inst{21-20} = 0b01;
2367 def MVE_VSLIimm32 : MVE_VSxI_imm<"vsli", "32", 0b1,(ins imm0_31:$imm)> {
2371 class MVE_VQSHL_imm<string suffix, dag imm>
2372 : MVE_shift_with_imm<"vqshl", suffix, (outs MQPR:$Qd),
2373 !con((ins MQPR:$Qm), imm), "$Qd, $Qm, $imm",
2377 let Inst{25-24} = 0b11;
2378 let Inst{21-16} = imm;
2379 let Inst{10-8} = 0b111;
2382 def MVE_VSLIimms8 : MVE_VQSHL_imm<"s8", (ins imm0_7:$imm)> {
2384 let Inst{21-19} = 0b001;
2387 def MVE_VSLIimmu8 : MVE_VQSHL_imm<"u8", (ins imm0_7:$imm)> {
2389 let Inst{21-19} = 0b001;
2392 def MVE_VSLIimms16 : MVE_VQSHL_imm<"s16", (ins imm0_15:$imm)> {
2394 let Inst{21-20} = 0b01;
2397 def MVE_VSLIimmu16 : MVE_VQSHL_imm<"u16", (ins imm0_15:$imm)> {
2399 let Inst{21-20} = 0b01;
2402 def MVE_VSLIimms32 : MVE_VQSHL_imm<"s32", (ins imm0_31:$imm)> {
2407 def MVE_VSLIimmu32 : MVE_VQSHL_imm<"u32", (ins imm0_31:$imm)> {
2412 class MVE_VQSHLU_imm<string suffix, dag imm>
2413 : MVE_shift_with_imm<"vqshlu", suffix, (outs MQPR:$Qd),
2414 !con((ins MQPR:$Qm), imm), "$Qd, $Qm, $imm",
2419 let Inst{25-24} = 0b11;
2420 let Inst{21-16} = imm;
2421 let Inst{10-8} = 0b110;
2424 def MVE_VQSHLU_imms8 : MVE_VQSHLU_imm<"s8", (ins imm0_7:$imm)> {
2425 let Inst{21-19} = 0b001;
2428 def MVE_VQSHLU_imms16 : MVE_VQSHLU_imm<"s16", (ins imm0_15:$imm)> {
2429 let Inst{21-20} = 0b01;
2432 def MVE_VQSHLU_imms32 : MVE_VQSHLU_imm<"s32", (ins imm0_31:$imm)> {
2436 class MVE_VRSHR_imm<string suffix, dag imm>
2437 : MVE_shift_with_imm<"vrshr", suffix, (outs MQPR:$Qd),
2438 !con((ins MQPR:$Qm), imm), "$Qd, $Qm, $imm",
2442 let Inst{25-24} = 0b11;
2443 let Inst{21-16} = imm;
2444 let Inst{10-8} = 0b010;
2447 def MVE_VRSHR_imms8 : MVE_VRSHR_imm<"s8", (ins shr_imm8:$imm)> {
2449 let Inst{21-19} = 0b001;
2452 def MVE_VRSHR_immu8 : MVE_VRSHR_imm<"u8", (ins shr_imm8:$imm)> {
2454 let Inst{21-19} = 0b001;
2457 def MVE_VRSHR_imms16 : MVE_VRSHR_imm<"s16", (ins shr_imm16:$imm)> {
2459 let Inst{21-20} = 0b01;
2462 def MVE_VRSHR_immu16 : MVE_VRSHR_imm<"u16", (ins shr_imm16:$imm)> {
2464 let Inst{21-20} = 0b01;
2467 def MVE_VRSHR_imms32 : MVE_VRSHR_imm<"s32", (ins shr_imm32:$imm)> {
2472 def MVE_VRSHR_immu32 : MVE_VRSHR_imm<"u32", (ins shr_imm32:$imm)> {
2477 class MVE_VSHR_imm<string suffix, dag imm>
2478 : MVE_shift_with_imm<"vshr", suffix, (outs MQPR:$Qd),
2479 !con((ins MQPR:$Qm), imm), "$Qd, $Qm, $imm",
2483 let Inst{25-24} = 0b11;
2484 let Inst{21-16} = imm;
2485 let Inst{10-8} = 0b000;
2488 def MVE_VSHR_imms8 : MVE_VSHR_imm<"s8", (ins shr_imm8:$imm)> {
2490 let Inst{21-19} = 0b001;
2493 def MVE_VSHR_immu8 : MVE_VSHR_imm<"u8", (ins shr_imm8:$imm)> {
2495 let Inst{21-19} = 0b001;
2498 def MVE_VSHR_imms16 : MVE_VSHR_imm<"s16", (ins shr_imm16:$imm)> {
2500 let Inst{21-20} = 0b01;
2503 def MVE_VSHR_immu16 : MVE_VSHR_imm<"u16", (ins shr_imm16:$imm)> {
2505 let Inst{21-20} = 0b01;
2508 def MVE_VSHR_imms32 : MVE_VSHR_imm<"s32", (ins shr_imm32:$imm)> {
2513 def MVE_VSHR_immu32 : MVE_VSHR_imm<"u32", (ins shr_imm32:$imm)> {
2518 class MVE_VSHL_imm<string suffix, dag imm>
2519 : MVE_shift_with_imm<"vshl", suffix, (outs MQPR:$Qd),
2520 !con((ins MQPR:$Qm), imm), "$Qd, $Qm, $imm",
2525 let Inst{25-24} = 0b11;
2526 let Inst{21-16} = imm;
2527 let Inst{10-8} = 0b101;
2530 def MVE_VSHL_immi8 : MVE_VSHL_imm<"i8", (ins imm0_7:$imm)> {
2531 let Inst{21-19} = 0b001;
2534 def MVE_VSHL_immi16 : MVE_VSHL_imm<"i16", (ins imm0_15:$imm)> {
2535 let Inst{21-20} = 0b01;
2538 def MVE_VSHL_immi32 : MVE_VSHL_imm<"i32", (ins imm0_31:$imm)> {
2542 let Predicates = [HasMVEInt] in {
2543 def : Pat<(v4i32 (ARMvshlImm (v4i32 MQPR:$src), imm0_31:$imm)),
2544 (v4i32 (MVE_VSHL_immi32 (v4i32 MQPR:$src), imm0_31:$imm))>;
2545 def : Pat<(v8i16 (ARMvshlImm (v8i16 MQPR:$src), imm0_15:$imm)),
2546 (v8i16 (MVE_VSHL_immi16 (v8i16 MQPR:$src), imm0_15:$imm))>;
2547 def : Pat<(v16i8 (ARMvshlImm (v16i8 MQPR:$src), imm0_7:$imm)),
2548 (v16i8 (MVE_VSHL_immi8 (v16i8 MQPR:$src), imm0_7:$imm))>;
2550 def : Pat<(v4i32 (ARMvshruImm (v4i32 MQPR:$src), imm0_31:$imm)),
2551 (v4i32 (MVE_VSHR_immu32 (v4i32 MQPR:$src), imm0_31:$imm))>;
2552 def : Pat<(v8i16 (ARMvshruImm (v8i16 MQPR:$src), imm0_15:$imm)),
2553 (v8i16 (MVE_VSHR_immu16 (v8i16 MQPR:$src), imm0_15:$imm))>;
2554 def : Pat<(v16i8 (ARMvshruImm (v16i8 MQPR:$src), imm0_7:$imm)),
2555 (v16i8 (MVE_VSHR_immu8 (v16i8 MQPR:$src), imm0_7:$imm))>;
2557 def : Pat<(v4i32 (ARMvshrsImm (v4i32 MQPR:$src), imm0_31:$imm)),
2558 (v4i32 (MVE_VSHR_imms32 (v4i32 MQPR:$src), imm0_31:$imm))>;
2559 def : Pat<(v8i16 (ARMvshrsImm (v8i16 MQPR:$src), imm0_15:$imm)),
2560 (v8i16 (MVE_VSHR_imms16 (v8i16 MQPR:$src), imm0_15:$imm))>;
2561 def : Pat<(v16i8 (ARMvshrsImm (v16i8 MQPR:$src), imm0_7:$imm)),
2562 (v16i8 (MVE_VSHR_imms8 (v16i8 MQPR:$src), imm0_7:$imm))>;
2565 // end of mve_shift instructions
2567 // start of MVE Floating Point instructions
2569 class MVE_float<string iname, string suffix, dag oops, dag iops, string ops,
2570 vpred_ops vpred, string cstr, list<dag> pattern=[]>
2571 : MVE_f<oops, iops, NoItinerary, iname, suffix, ops, vpred, cstr, pattern> {
2576 let Inst{5} = Qm{3};
2577 let Inst{3-1} = Qm{2-0};
2581 class MVE_VRINT<string rmode, bits<3> op, string suffix, bits<2> size,
2582 list<dag> pattern=[]>
2583 : MVE_float<!strconcat("vrint", rmode), suffix, (outs MQPR:$Qd),
2584 (ins MQPR:$Qm), "$Qd, $Qm", vpred_r, "", pattern> {
2588 let Inst{25-23} = 0b111;
2589 let Inst{22} = Qd{3};
2590 let Inst{21-20} = 0b11;
2591 let Inst{19-18} = size;
2592 let Inst{17-16} = 0b10;
2593 let Inst{15-13} = Qd{2-0};
2594 let Inst{11-10} = 0b01;
2595 let Inst{9-7} = op{2-0};
2600 multiclass MVE_VRINT_ops<string suffix, bits<2> size, list<dag> pattern=[]> {
2601 def N : MVE_VRINT<"n", 0b000, suffix, size, pattern>;
2602 def X : MVE_VRINT<"x", 0b001, suffix, size, pattern>;
2603 def A : MVE_VRINT<"a", 0b010, suffix, size, pattern>;
2604 def Z : MVE_VRINT<"z", 0b011, suffix, size, pattern>;
2605 def M : MVE_VRINT<"m", 0b101, suffix, size, pattern>;
2606 def P : MVE_VRINT<"p", 0b111, suffix, size, pattern>;
2609 defm MVE_VRINTf16 : MVE_VRINT_ops<"f16", 0b01>;
2610 defm MVE_VRINTf32 : MVE_VRINT_ops<"f32", 0b10>;
2612 let Predicates = [HasMVEFloat] in {
2613 def : Pat<(v4f32 (frint (v4f32 MQPR:$val1))),
2614 (v4f32 (MVE_VRINTf32X (v4f32 MQPR:$val1)))>;
2615 def : Pat<(v8f16 (frint (v8f16 MQPR:$val1))),
2616 (v8f16 (MVE_VRINTf16X (v8f16 MQPR:$val1)))>;
2617 def : Pat<(v4f32 (fround (v4f32 MQPR:$val1))),
2618 (v4f32 (MVE_VRINTf32A (v4f32 MQPR:$val1)))>;
2619 def : Pat<(v8f16 (fround (v8f16 MQPR:$val1))),
2620 (v8f16 (MVE_VRINTf16A (v8f16 MQPR:$val1)))>;
2621 def : Pat<(v4f32 (ftrunc (v4f32 MQPR:$val1))),
2622 (v4f32 (MVE_VRINTf32Z (v4f32 MQPR:$val1)))>;
2623 def : Pat<(v8f16 (ftrunc (v8f16 MQPR:$val1))),
2624 (v8f16 (MVE_VRINTf16Z (v8f16 MQPR:$val1)))>;
2625 def : Pat<(v4f32 (ffloor (v4f32 MQPR:$val1))),
2626 (v4f32 (MVE_VRINTf32M (v4f32 MQPR:$val1)))>;
2627 def : Pat<(v8f16 (ffloor (v8f16 MQPR:$val1))),
2628 (v8f16 (MVE_VRINTf16M (v8f16 MQPR:$val1)))>;
2629 def : Pat<(v4f32 (fceil (v4f32 MQPR:$val1))),
2630 (v4f32 (MVE_VRINTf32P (v4f32 MQPR:$val1)))>;
2631 def : Pat<(v8f16 (fceil (v8f16 MQPR:$val1))),
2632 (v8f16 (MVE_VRINTf16P (v8f16 MQPR:$val1)))>;
2635 class MVEFloatArithNeon<string iname, string suffix, bit size,
2636 dag oops, dag iops, string ops,
2637 vpred_ops vpred, string cstr, list<dag> pattern=[]>
2638 : MVE_float<iname, suffix, oops, iops, ops, vpred, cstr, pattern> {
2639 let Inst{20} = size;
2643 class MVE_VMUL_fp<string suffix, bit size, list<dag> pattern=[]>
2644 : MVEFloatArithNeon<"vmul", suffix, size, (outs MQPR:$Qd),
2645 (ins MQPR:$Qn, MQPR:$Qm), "$Qd, $Qn, $Qm", vpred_r, "",
2651 let Inst{25-23} = 0b110;
2652 let Inst{22} = Qd{3};
2654 let Inst{19-17} = Qn{2-0};
2655 let Inst{15-13} = Qd{2-0};
2656 let Inst{12-8} = 0b01101;
2657 let Inst{7} = Qn{3};
2661 def MVE_VMULf32 : MVE_VMUL_fp<"f32", 0b0>;
2662 def MVE_VMULf16 : MVE_VMUL_fp<"f16", 0b1>;
2664 let Predicates = [HasMVEFloat] in {
2665 def : Pat<(v4f32 (fmul (v4f32 MQPR:$val1), (v4f32 MQPR:$val2))),
2666 (v4f32 (MVE_VMULf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2)))>;
2667 def : Pat<(v8f16 (fmul (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
2668 (v8f16 (MVE_VMULf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
2671 class MVE_VCMLA<string suffix, bit size, list<dag> pattern=[]>
2672 : MVEFloatArithNeon<"vcmla", suffix, size, (outs MQPR:$Qd),
2673 (ins MQPR:$Qd_src, MQPR:$Qn, MQPR:$Qm, complexrotateop:$rot),
2674 "$Qd, $Qn, $Qm, $rot", vpred_n, "$Qd = $Qd_src", pattern> {
2681 let Inst{24-23} = rot;
2682 let Inst{22} = Qd{3};
2684 let Inst{19-17} = Qn{2-0};
2685 let Inst{15-13} = Qd{2-0};
2686 let Inst{12-8} = 0b01000;
2687 let Inst{7} = Qn{3};
2691 def MVE_VCMLAf16 : MVE_VCMLA<"f16", 0b0>;
2692 def MVE_VCMLAf32 : MVE_VCMLA<"f32", 0b1>;
2694 class MVE_VADDSUBFMA_fp<string iname, string suffix, bit size, bit bit_4,
2695 bit bit_8, bit bit_21, dag iops=(ins),
2696 vpred_ops vpred=vpred_r, string cstr="",
2697 list<dag> pattern=[]>
2698 : MVEFloatArithNeon<iname, suffix, size, (outs MQPR:$Qd),
2699 !con(iops, (ins MQPR:$Qn, MQPR:$Qm)), "$Qd, $Qn, $Qm",
2700 vpred, cstr, pattern> {
2705 let Inst{25-23} = 0b110;
2706 let Inst{22} = Qd{3};
2707 let Inst{21} = bit_21;
2708 let Inst{19-17} = Qn{2-0};
2709 let Inst{15-13} = Qd{2-0};
2710 let Inst{11-9} = 0b110;
2711 let Inst{8} = bit_8;
2712 let Inst{7} = Qn{3};
2713 let Inst{4} = bit_4;
2716 def MVE_VFMAf32 : MVE_VADDSUBFMA_fp<"vfma", "f32", 0b0, 0b1, 0b0, 0b0,
2717 (ins MQPR:$Qd_src), vpred_n, "$Qd = $Qd_src">;
2718 def MVE_VFMAf16 : MVE_VADDSUBFMA_fp<"vfma", "f16", 0b1, 0b1, 0b0, 0b0,
2719 (ins MQPR:$Qd_src), vpred_n, "$Qd = $Qd_src">;
2721 def MVE_VFMSf32 : MVE_VADDSUBFMA_fp<"vfms", "f32", 0b0, 0b1, 0b0, 0b1,
2722 (ins MQPR:$Qd_src), vpred_n, "$Qd = $Qd_src">;
2723 def MVE_VFMSf16 : MVE_VADDSUBFMA_fp<"vfms", "f16", 0b1, 0b1, 0b0, 0b1,
2724 (ins MQPR:$Qd_src), vpred_n, "$Qd = $Qd_src">;
2726 let Predicates = [HasMVEFloat, UseFusedMAC] in {
2727 def : Pat<(v8f16 (fadd (v8f16 MQPR:$src1),
2728 (fmul (v8f16 MQPR:$src2),
2729 (v8f16 MQPR:$src3)))),
2730 (v8f16 (MVE_VFMAf16 $src1, $src2, $src3))>;
2731 def : Pat<(v4f32 (fadd (v4f32 MQPR:$src1),
2732 (fmul (v4f32 MQPR:$src2),
2733 (v4f32 MQPR:$src3)))),
2734 (v4f32 (MVE_VFMAf32 $src1, $src2, $src3))>;
2736 def : Pat<(v8f16 (fsub (v8f16 MQPR:$src1),
2737 (fmul (v8f16 MQPR:$src2),
2738 (v8f16 MQPR:$src3)))),
2739 (v8f16 (MVE_VFMSf16 $src1, $src2, $src3))>;
2740 def : Pat<(v4f32 (fsub (v4f32 MQPR:$src1),
2741 (fmul (v4f32 MQPR:$src2),
2742 (v4f32 MQPR:$src3)))),
2743 (v4f32 (MVE_VFMSf32 $src1, $src2, $src3))>;
2746 let Predicates = [HasMVEFloat] in {
2747 def : Pat<(v8f16 (fma (v8f16 MQPR:$src1), (v8f16 MQPR:$src2), (v8f16 MQPR:$src3))),
2748 (v8f16 (MVE_VFMAf16 $src3, $src1, $src2))>;
2749 def : Pat<(v4f32 (fma (v4f32 MQPR:$src1), (v4f32 MQPR:$src2), (v4f32 MQPR:$src3))),
2750 (v4f32 (MVE_VFMAf32 $src3, $src1, $src2))>;
2754 def MVE_VADDf32 : MVE_VADDSUBFMA_fp<"vadd", "f32", 0b0, 0b0, 0b1, 0b0>;
2755 def MVE_VADDf16 : MVE_VADDSUBFMA_fp<"vadd", "f16", 0b1, 0b0, 0b1, 0b0>;
2757 let Predicates = [HasMVEFloat] in {
2758 def : Pat<(v4f32 (fadd (v4f32 MQPR:$val1), (v4f32 MQPR:$val2))),
2759 (v4f32 (MVE_VADDf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2)))>;
2760 def : Pat<(v8f16 (fadd (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
2761 (v8f16 (MVE_VADDf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
2764 def MVE_VSUBf32 : MVE_VADDSUBFMA_fp<"vsub", "f32", 0b0, 0b0, 0b1, 0b1>;
2765 def MVE_VSUBf16 : MVE_VADDSUBFMA_fp<"vsub", "f16", 0b1, 0b0, 0b1, 0b1>;
2767 let Predicates = [HasMVEFloat] in {
2768 def : Pat<(v4f32 (fsub (v4f32 MQPR:$val1), (v4f32 MQPR:$val2))),
2769 (v4f32 (MVE_VSUBf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2)))>;
2770 def : Pat<(v8f16 (fsub (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
2771 (v8f16 (MVE_VSUBf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
2774 class MVE_VCADD<string suffix, bit size, string cstr="", list<dag> pattern=[]>
2775 : MVEFloatArithNeon<"vcadd", suffix, size, (outs MQPR:$Qd),
2776 (ins MQPR:$Qn, MQPR:$Qm, complexrotateopodd:$rot),
2777 "$Qd, $Qn, $Qm, $rot", vpred_r, cstr, pattern> {
2786 let Inst{22} = Qd{3};
2788 let Inst{19-17} = Qn{2-0};
2789 let Inst{15-13} = Qd{2-0};
2790 let Inst{12-8} = 0b01000;
2791 let Inst{7} = Qn{3};
2795 def MVE_VCADDf16 : MVE_VCADD<"f16", 0b0>;
2796 def MVE_VCADDf32 : MVE_VCADD<"f32", 0b1, "@earlyclobber $Qd">;
2798 class MVE_VABD_fp<string suffix, bit size>
2799 : MVE_float<"vabd", suffix, (outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm),
2800 "$Qd, $Qn, $Qm", vpred_r, ""> {
2805 let Inst{25-23} = 0b110;
2806 let Inst{22} = Qd{3};
2808 let Inst{20} = size;
2809 let Inst{19-17} = Qn{2-0};
2811 let Inst{15-13} = Qd{2-0};
2812 let Inst{11-8} = 0b1101;
2813 let Inst{7} = Qn{3};
2817 def MVE_VABDf32 : MVE_VABD_fp<"f32", 0b0>;
2818 def MVE_VABDf16 : MVE_VABD_fp<"f16", 0b1>;
2820 class MVE_VCVT_fix<string suffix, bit fsi, bit U, bit op,
2821 Operand imm_operand_type, list<dag> pattern=[]>
2822 : MVE_float<"vcvt", suffix,
2823 (outs MQPR:$Qd), (ins MQPR:$Qm, imm_operand_type:$imm6),
2824 "$Qd, $Qm, $imm6", vpred_r, "", pattern> {
2829 let Inst{25-23} = 0b111;
2830 let Inst{22} = Qd{3};
2832 let Inst{19-16} = imm6{3-0};
2833 let Inst{15-13} = Qd{2-0};
2834 let Inst{11-10} = 0b11;
2840 let DecoderMethod = "DecodeMVEVCVTt1fp";
2843 class MVE_VCVT_imm_asmop<int Bits> : AsmOperandClass {
2844 let PredicateMethod = "isImmediate<1," # Bits # ">";
2845 let DiagnosticString =
2846 "MVE fixed-point immediate operand must be between 1 and " # Bits;
2847 let Name = "MVEVcvtImm" # Bits;
2848 let RenderMethod = "addImmOperands";
2850 class MVE_VCVT_imm<int Bits>: Operand<i32> {
2851 let ParserMatchClass = MVE_VCVT_imm_asmop<Bits>;
2852 let EncoderMethod = "getNEONVcvtImm32OpValue";
2853 let DecoderMethod = "DecodeVCVTImmOperand";
2856 class MVE_VCVT_fix_f32<string suffix, bit U, bit op>
2857 : MVE_VCVT_fix<suffix, 0b1, U, op, MVE_VCVT_imm<32>> {
2858 let Inst{20} = imm6{4};
2860 class MVE_VCVT_fix_f16<string suffix, bit U, bit op>
2861 : MVE_VCVT_fix<suffix, 0b0, U, op, MVE_VCVT_imm<16>> {
2865 def MVE_VCVTf16s16_fix : MVE_VCVT_fix_f16<"f16.s16", 0b0, 0b0>;
2866 def MVE_VCVTs16f16_fix : MVE_VCVT_fix_f16<"s16.f16", 0b0, 0b1>;
2867 def MVE_VCVTf16u16_fix : MVE_VCVT_fix_f16<"f16.u16", 0b1, 0b0>;
2868 def MVE_VCVTu16f16_fix : MVE_VCVT_fix_f16<"u16.f16", 0b1, 0b1>;
2869 def MVE_VCVTf32s32_fix : MVE_VCVT_fix_f32<"f32.s32", 0b0, 0b0>;
2870 def MVE_VCVTs32f32_fix : MVE_VCVT_fix_f32<"s32.f32", 0b0, 0b1>;
2871 def MVE_VCVTf32u32_fix : MVE_VCVT_fix_f32<"f32.u32", 0b1, 0b0>;
2872 def MVE_VCVTu32f32_fix : MVE_VCVT_fix_f32<"u32.f32", 0b1, 0b1>;
2874 class MVE_VCVT_fp_int_anpm<string suffix, bits<2> size, bit op, string anpm,
2875 bits<2> rm, list<dag> pattern=[]>
2876 : MVE_float<!strconcat("vcvt", anpm), suffix, (outs MQPR:$Qd),
2877 (ins MQPR:$Qm), "$Qd, $Qm", vpred_r, "", pattern> {
2881 let Inst{25-23} = 0b111;
2882 let Inst{22} = Qd{3};
2883 let Inst{21-20} = 0b11;
2884 let Inst{19-18} = size;
2885 let Inst{17-16} = 0b11;
2886 let Inst{15-13} = Qd{2-0};
2887 let Inst{12-10} = 0b000;
2893 multiclass MVE_VCVT_fp_int_anpm_multi<string suffix, bits<2> size, bit op,
2894 list<dag> pattern=[]> {
2895 def a : MVE_VCVT_fp_int_anpm<suffix, size, op, "a", 0b00>;
2896 def n : MVE_VCVT_fp_int_anpm<suffix, size, op, "n", 0b01>;
2897 def p : MVE_VCVT_fp_int_anpm<suffix, size, op, "p", 0b10>;
2898 def m : MVE_VCVT_fp_int_anpm<suffix, size, op, "m", 0b11>;
2901 // This defines instructions such as MVE_VCVTu16f16a, with an explicit
2902 // rounding-mode suffix on the mnemonic. The class below will define
2903 // the bare MVE_VCVTu16f16 (with implied rounding toward zero).
2904 defm MVE_VCVTs16f16 : MVE_VCVT_fp_int_anpm_multi<"s16.f16", 0b01, 0b0>;
2905 defm MVE_VCVTu16f16 : MVE_VCVT_fp_int_anpm_multi<"u16.f16", 0b01, 0b1>;
2906 defm MVE_VCVTs32f32 : MVE_VCVT_fp_int_anpm_multi<"s32.f32", 0b10, 0b0>;
2907 defm MVE_VCVTu32f32 : MVE_VCVT_fp_int_anpm_multi<"u32.f32", 0b10, 0b1>;
2909 class MVE_VCVT_fp_int<string suffix, bits<2> size, bits<2> op,
2910 list<dag> pattern=[]>
2911 : MVE_float<"vcvt", suffix, (outs MQPR:$Qd),
2912 (ins MQPR:$Qm), "$Qd, $Qm", vpred_r, "", pattern> {
2916 let Inst{25-23} = 0b111;
2917 let Inst{22} = Qd{3};
2918 let Inst{21-20} = 0b11;
2919 let Inst{19-18} = size;
2920 let Inst{17-16} = 0b11;
2921 let Inst{15-13} = Qd{2-0};
2922 let Inst{12-9} = 0b0011;
2927 // The unsuffixed VCVT for float->int implicitly rounds toward zero,
2928 // which I reflect here in the llvm instruction names
2929 def MVE_VCVTs16f16z : MVE_VCVT_fp_int<"s16.f16", 0b01, 0b10>;
2930 def MVE_VCVTu16f16z : MVE_VCVT_fp_int<"u16.f16", 0b01, 0b11>;
2931 def MVE_VCVTs32f32z : MVE_VCVT_fp_int<"s32.f32", 0b10, 0b10>;
2932 def MVE_VCVTu32f32z : MVE_VCVT_fp_int<"u32.f32", 0b10, 0b11>;
2933 // Whereas VCVT for int->float rounds to nearest
2934 def MVE_VCVTf16s16n : MVE_VCVT_fp_int<"f16.s16", 0b01, 0b00>;
2935 def MVE_VCVTf16u16n : MVE_VCVT_fp_int<"f16.u16", 0b01, 0b01>;
2936 def MVE_VCVTf32s32n : MVE_VCVT_fp_int<"f32.s32", 0b10, 0b00>;
2937 def MVE_VCVTf32u32n : MVE_VCVT_fp_int<"f32.u32", 0b10, 0b01>;
2939 let Predicates = [HasMVEFloat] in {
2940 def : Pat<(v4i32 (fp_to_sint (v4f32 MQPR:$src))),
2941 (v4i32 (MVE_VCVTs32f32z (v4f32 MQPR:$src)))>;
2942 def : Pat<(v4i32 (fp_to_uint (v4f32 MQPR:$src))),
2943 (v4i32 (MVE_VCVTu32f32z (v4f32 MQPR:$src)))>;
2944 def : Pat<(v8i16 (fp_to_sint (v8f16 MQPR:$src))),
2945 (v8i16 (MVE_VCVTs16f16z (v8f16 MQPR:$src)))>;
2946 def : Pat<(v8i16 (fp_to_uint (v8f16 MQPR:$src))),
2947 (v8i16 (MVE_VCVTu16f16z (v8f16 MQPR:$src)))>;
2948 def : Pat<(v4f32 (sint_to_fp (v4i32 MQPR:$src))),
2949 (v4f32 (MVE_VCVTf32s32n (v4i32 MQPR:$src)))>;
2950 def : Pat<(v4f32 (uint_to_fp (v4i32 MQPR:$src))),
2951 (v4f32 (MVE_VCVTf32u32n (v4i32 MQPR:$src)))>;
2952 def : Pat<(v8f16 (sint_to_fp (v8i16 MQPR:$src))),
2953 (v8f16 (MVE_VCVTf16s16n (v8i16 MQPR:$src)))>;
2954 def : Pat<(v8f16 (uint_to_fp (v8i16 MQPR:$src))),
2955 (v8f16 (MVE_VCVTf16u16n (v8i16 MQPR:$src)))>;
2958 class MVE_VABSNEG_fp<string iname, string suffix, bits<2> size, bit negate,
2959 list<dag> pattern=[]>
2960 : MVE_float<iname, suffix, (outs MQPR:$Qd),
2961 (ins MQPR:$Qm), "$Qd, $Qm", vpred_r, "", pattern> {
2965 let Inst{25-23} = 0b111;
2966 let Inst{22} = Qd{3};
2967 let Inst{21-20} = 0b11;
2968 let Inst{19-18} = size;
2969 let Inst{17-16} = 0b01;
2970 let Inst{15-13} = Qd{2-0};
2971 let Inst{11-8} = 0b0111;
2972 let Inst{7} = negate;
2976 def MVE_VABSf16 : MVE_VABSNEG_fp<"vabs", "f16", 0b01, 0b0>;
2977 def MVE_VABSf32 : MVE_VABSNEG_fp<"vabs", "f32", 0b10, 0b0>;
2979 let Predicates = [HasMVEFloat] in {
2980 def : Pat<(v8f16 (fabs MQPR:$src)),
2981 (MVE_VABSf16 MQPR:$src)>;
2982 def : Pat<(v4f32 (fabs MQPR:$src)),
2983 (MVE_VABSf32 MQPR:$src)>;
2986 def MVE_VNEGf16 : MVE_VABSNEG_fp<"vneg", "f16", 0b01, 0b1>;
2987 def MVE_VNEGf32 : MVE_VABSNEG_fp<"vneg", "f32", 0b10, 0b1>;
2989 let Predicates = [HasMVEFloat] in {
2990 def : Pat<(v8f16 (fneg MQPR:$src)),
2991 (MVE_VNEGf16 MQPR:$src)>;
2992 def : Pat<(v4f32 (fneg MQPR:$src)),
2993 (MVE_VNEGf32 MQPR:$src)>;
2996 class MVE_VMAXMINNMA<string iname, string suffix, bit size, bit bit_12,
2997 list<dag> pattern=[]>
2998 : MVE_f<(outs MQPR:$Qd), (ins MQPR:$Qd_src, MQPR:$Qm),
2999 NoItinerary, iname, suffix, "$Qd, $Qm", vpred_n, "$Qd = $Qd_src",
3004 let Inst{28} = size;
3005 let Inst{25-23} = 0b100;
3006 let Inst{22} = Qd{3};
3007 let Inst{21-16} = 0b111111;
3008 let Inst{15-13} = Qd{2-0};
3009 let Inst{12} = bit_12;
3010 let Inst{11-6} = 0b111010;
3011 let Inst{5} = Qm{3};
3013 let Inst{3-1} = Qm{2-0};
3017 def MVE_VMAXNMAf32 : MVE_VMAXMINNMA<"vmaxnma", "f32", 0b0, 0b0>;
3018 def MVE_VMAXNMAf16 : MVE_VMAXMINNMA<"vmaxnma", "f16", 0b1, 0b0>;
3020 def MVE_VMINNMAf32 : MVE_VMAXMINNMA<"vminnma", "f32", 0b0, 0b1>;
3021 def MVE_VMINNMAf16 : MVE_VMAXMINNMA<"vminnma", "f16", 0b1, 0b1>;
3023 // end of MVE Floating Point instructions
3025 // start of MVE compares
3027 class MVE_VCMPqq<string suffix, bit bit_28, bits<2> bits_21_20,
3028 VCMPPredicateOperand predtype, list<dag> pattern=[]>
3029 : MVE_p<(outs VCCR:$P0), (ins MQPR:$Qn, MQPR:$Qm, predtype:$fc),
3030 NoItinerary, "vcmp", suffix, "$fc, $Qn, $Qm", vpred_n, "", pattern> {
3031 // Base class for comparing two vector registers
3036 let Inst{28} = bit_28;
3037 let Inst{25-22} = 0b1000;
3038 let Inst{21-20} = bits_21_20;
3039 let Inst{19-17} = Qn{2-0};
3040 let Inst{16-13} = 0b1000;
3041 let Inst{12} = fc{2};
3042 let Inst{11-8} = 0b1111;
3043 let Inst{7} = fc{0};
3045 let Inst{5} = Qm{3};
3047 let Inst{3-1} = Qm{2-0};
3048 let Inst{0} = fc{1};
3050 let Constraints = "";
3052 // We need a custom decoder method for these instructions because of
3053 // the output VCCR operand, which isn't encoded in the instruction
3054 // bits anywhere (there is only one choice for it) but has to be
3055 // included in the MC operands so that codegen will be able to track
3056 // its data flow between instructions, spill/reload it when
3057 // necessary, etc. There seems to be no way to get the Tablegen
3058 // decoder to emit an operand that isn't affected by any instruction
3060 let DecoderMethod = "DecodeMVEVCMP<false," # predtype.DecoderMethod # ">";
3063 class MVE_VCMPqqf<string suffix, bit size>
3064 : MVE_VCMPqq<suffix, size, 0b11, pred_basic_fp> {
3065 let Predicates = [HasMVEFloat];
3068 class MVE_VCMPqqi<string suffix, bits<2> size>
3069 : MVE_VCMPqq<suffix, 0b1, size, pred_basic_i> {
3074 class MVE_VCMPqqu<string suffix, bits<2> size>
3075 : MVE_VCMPqq<suffix, 0b1, size, pred_basic_u> {
3080 class MVE_VCMPqqs<string suffix, bits<2> size>
3081 : MVE_VCMPqq<suffix, 0b1, size, pred_basic_s> {
3085 def MVE_VCMPf32 : MVE_VCMPqqf<"f32", 0b0>;
3086 def MVE_VCMPf16 : MVE_VCMPqqf<"f16", 0b1>;
3088 def MVE_VCMPi8 : MVE_VCMPqqi<"i8", 0b00>;
3089 def MVE_VCMPi16 : MVE_VCMPqqi<"i16", 0b01>;
3090 def MVE_VCMPi32 : MVE_VCMPqqi<"i32", 0b10>;
3092 def MVE_VCMPu8 : MVE_VCMPqqu<"u8", 0b00>;
3093 def MVE_VCMPu16 : MVE_VCMPqqu<"u16", 0b01>;
3094 def MVE_VCMPu32 : MVE_VCMPqqu<"u32", 0b10>;
3096 def MVE_VCMPs8 : MVE_VCMPqqs<"s8", 0b00>;
3097 def MVE_VCMPs16 : MVE_VCMPqqs<"s16", 0b01>;
3098 def MVE_VCMPs32 : MVE_VCMPqqs<"s32", 0b10>;
3100 class MVE_VCMPqr<string suffix, bit bit_28, bits<2> bits_21_20,
3101 VCMPPredicateOperand predtype, list<dag> pattern=[]>
3102 : MVE_p<(outs VCCR:$P0), (ins MQPR:$Qn, GPRwithZR:$Rm, predtype:$fc),
3103 NoItinerary, "vcmp", suffix, "$fc, $Qn, $Rm", vpred_n, "", pattern> {
3104 // Base class for comparing a vector register with a scalar
3109 let Inst{28} = bit_28;
3110 let Inst{25-22} = 0b1000;
3111 let Inst{21-20} = bits_21_20;
3112 let Inst{19-17} = Qn{2-0};
3113 let Inst{16-13} = 0b1000;
3114 let Inst{12} = fc{2};
3115 let Inst{11-8} = 0b1111;
3116 let Inst{7} = fc{0};
3118 let Inst{5} = fc{1};
3120 let Inst{3-0} = Rm{3-0};
3122 let Constraints = "";
3123 // Custom decoder method, for the same reason as MVE_VCMPqq
3124 let DecoderMethod = "DecodeMVEVCMP<true," # predtype.DecoderMethod # ">";
3127 class MVE_VCMPqrf<string suffix, bit size>
3128 : MVE_VCMPqr<suffix, size, 0b11, pred_basic_fp> {
3129 let Predicates = [HasMVEFloat];
3132 class MVE_VCMPqri<string suffix, bits<2> size>
3133 : MVE_VCMPqr<suffix, 0b1, size, pred_basic_i> {
3138 class MVE_VCMPqru<string suffix, bits<2> size>
3139 : MVE_VCMPqr<suffix, 0b1, size, pred_basic_u> {
3144 class MVE_VCMPqrs<string suffix, bits<2> size>
3145 : MVE_VCMPqr<suffix, 0b1, size, pred_basic_s> {
3149 def MVE_VCMPf32r : MVE_VCMPqrf<"f32", 0b0>;
3150 def MVE_VCMPf16r : MVE_VCMPqrf<"f16", 0b1>;
3152 def MVE_VCMPi8r : MVE_VCMPqri<"i8", 0b00>;
3153 def MVE_VCMPi16r : MVE_VCMPqri<"i16", 0b01>;
3154 def MVE_VCMPi32r : MVE_VCMPqri<"i32", 0b10>;
3156 def MVE_VCMPu8r : MVE_VCMPqru<"u8", 0b00>;
3157 def MVE_VCMPu16r : MVE_VCMPqru<"u16", 0b01>;
3158 def MVE_VCMPu32r : MVE_VCMPqru<"u32", 0b10>;
3160 def MVE_VCMPs8r : MVE_VCMPqrs<"s8", 0b00>;
3161 def MVE_VCMPs16r : MVE_VCMPqrs<"s16", 0b01>;
3162 def MVE_VCMPs32r : MVE_VCMPqrs<"s32", 0b10>;
3164 multiclass unpred_vcmp_z<string suffix, int fc> {
3165 def i8 : Pat<(v16i1 (ARMvcmpz (v16i8 MQPR:$v1), (i32 fc))),
3166 (v16i1 (!cast<Instruction>("MVE_VCMP"#suffix#"8r") (v16i8 MQPR:$v1), ZR, fc))>;
3167 def i16 : Pat<(v8i1 (ARMvcmpz (v8i16 MQPR:$v1), (i32 fc))),
3168 (v8i1 (!cast<Instruction>("MVE_VCMP"#suffix#"16r") (v8i16 MQPR:$v1), ZR, fc))>;
3169 def i32 : Pat<(v4i1 (ARMvcmpz (v4i32 MQPR:$v1), (i32 fc))),
3170 (v4i1 (!cast<Instruction>("MVE_VCMP"#suffix#"32r") (v4i32 MQPR:$v1), ZR, fc))>;
3172 def : Pat<(v16i1 (and (v16i1 VCCR:$p1), (v16i1 (ARMvcmpz (v16i8 MQPR:$v1), (i32 fc))))),
3173 (v16i1 (!cast<Instruction>("MVE_VCMP"#suffix#"8r") (v16i8 MQPR:$v1), ZR, fc, 1, VCCR:$p1))>;
3174 def : Pat<(v8i1 (and (v8i1 VCCR:$p1), (v8i1 (ARMvcmpz (v8i16 MQPR:$v1), (i32 fc))))),
3175 (v8i1 (!cast<Instruction>("MVE_VCMP"#suffix#"16r") (v8i16 MQPR:$v1), ZR, fc, 1, VCCR:$p1))>;
3176 def : Pat<(v4i1 (and (v4i1 VCCR:$p1), (v4i1 (ARMvcmpz (v4i32 MQPR:$v1), (i32 fc))))),
3177 (v4i1 (!cast<Instruction>("MVE_VCMP"#suffix#"32r") (v4i32 MQPR:$v1), ZR, fc, 1, VCCR:$p1))>;
3180 multiclass unpred_vcmp_r<string suffix, int fc> {
3181 def i8 : Pat<(v16i1 (ARMvcmp (v16i8 MQPR:$v1), (v16i8 MQPR:$v2), (i32 fc))),
3182 (v16i1 (!cast<Instruction>("MVE_VCMP"#suffix#"8") (v16i8 MQPR:$v1), (v16i8 MQPR:$v2), fc))>;
3183 def i16 : Pat<(v8i1 (ARMvcmp (v8i16 MQPR:$v1), (v8i16 MQPR:$v2), (i32 fc))),
3184 (v8i1 (!cast<Instruction>("MVE_VCMP"#suffix#"16") (v8i16 MQPR:$v1), (v8i16 MQPR:$v2), fc))>;
3185 def i32 : Pat<(v4i1 (ARMvcmp (v4i32 MQPR:$v1), (v4i32 MQPR:$v2), (i32 fc))),
3186 (v4i1 (!cast<Instruction>("MVE_VCMP"#suffix#"32") (v4i32 MQPR:$v1), (v4i32 MQPR:$v2), fc))>;
3188 def i8r : Pat<(v16i1 (ARMvcmp (v16i8 MQPR:$v1), (v16i8 (ARMvdup GPR:$v2)), (i32 fc))),
3189 (v16i1 (!cast<Instruction>("MVE_VCMP"#suffix#"8r") (v16i8 MQPR:$v1), (i32 GPR:$v2), fc))>;
3190 def i16r : Pat<(v8i1 (ARMvcmp (v8i16 MQPR:$v1), (v8i16 (ARMvdup GPR:$v2)), (i32 fc))),
3191 (v8i1 (!cast<Instruction>("MVE_VCMP"#suffix#"16r") (v8i16 MQPR:$v1), (i32 GPR:$v2), fc))>;
3192 def i32r : Pat<(v4i1 (ARMvcmp (v4i32 MQPR:$v1), (v4i32 (ARMvdup GPR:$v2)), (i32 fc))),
3193 (v4i1 (!cast<Instruction>("MVE_VCMP"#suffix#"32r") (v4i32 MQPR:$v1), (i32 GPR:$v2), fc))>;
3195 def : Pat<(v16i1 (and (v16i1 VCCR:$p1), (v16i1 (ARMvcmp (v16i8 MQPR:$v1), (v16i8 MQPR:$v2), (i32 fc))))),
3196 (v16i1 (!cast<Instruction>("MVE_VCMP"#suffix#"8") (v16i8 MQPR:$v1), (v16i8 MQPR:$v2), fc, 1, VCCR:$p1))>;
3197 def : Pat<(v8i1 (and (v8i1 VCCR:$p1), (v8i1 (ARMvcmp (v8i16 MQPR:$v1), (v8i16 MQPR:$v2), (i32 fc))))),
3198 (v8i1 (!cast<Instruction>("MVE_VCMP"#suffix#"16") (v8i16 MQPR:$v1), (v8i16 MQPR:$v2), fc, 1, VCCR:$p1))>;
3199 def : Pat<(v4i1 (and (v4i1 VCCR:$p1), (v4i1 (ARMvcmp (v4i32 MQPR:$v1), (v4i32 MQPR:$v2), (i32 fc))))),
3200 (v4i1 (!cast<Instruction>("MVE_VCMP"#suffix#"32") (v4i32 MQPR:$v1), (v4i32 MQPR:$v2), fc, 1, VCCR:$p1))>;
3202 def : Pat<(v16i1 (and (v16i1 VCCR:$p1), (v16i1 (ARMvcmp (v16i8 MQPR:$v1), (v16i8 (ARMvdup GPR:$v2)), (i32 fc))))),
3203 (v16i1 (!cast<Instruction>("MVE_VCMP"#suffix#"8r") (v16i8 MQPR:$v1), (i32 GPR:$v2), fc, 1, VCCR:$p1))>;
3204 def : Pat<(v8i1 (and (v8i1 VCCR:$p1), (v8i1 (ARMvcmp (v8i16 MQPR:$v1), (v8i16 (ARMvdup GPR:$v2)), (i32 fc))))),
3205 (v8i1 (!cast<Instruction>("MVE_VCMP"#suffix#"16r") (v8i16 MQPR:$v1), (i32 GPR:$v2), fc, 1, VCCR:$p1))>;
3206 def : Pat<(v4i1 (and (v4i1 VCCR:$p1), (v4i1 (ARMvcmp (v4i32 MQPR:$v1), (v4i32 (ARMvdup GPR:$v2)), (i32 fc))))),
3207 (v4i1 (!cast<Instruction>("MVE_VCMP"#suffix#"32r") (v4i32 MQPR:$v1), (i32 GPR:$v2), fc, 1, VCCR:$p1))>;
3210 multiclass unpred_vcmpf_z<int fc> {
3211 def f16 : Pat<(v8i1 (ARMvcmpz (v8f16 MQPR:$v1), (i32 fc))),
3212 (v8i1 (MVE_VCMPf16r (v8f16 MQPR:$v1), ZR, fc))>;
3213 def f32 : Pat<(v4i1 (ARMvcmpz (v4f32 MQPR:$v1), (i32 fc))),
3214 (v4i1 (MVE_VCMPf32r (v4f32 MQPR:$v1), ZR, fc))>;
3216 def : Pat<(v8i1 (and (v8i1 VCCR:$p1), (v8i1 (ARMvcmpz (v8f16 MQPR:$v1), (i32 fc))))),
3217 (v8i1 (MVE_VCMPf32r (v8f16 MQPR:$v1), ZR, fc, 1, VCCR:$p1))>;
3218 def : Pat<(v4i1 (and (v4i1 VCCR:$p1), (v4i1 (ARMvcmpz (v4f32 MQPR:$v1), (i32 fc))))),
3219 (v4i1 (MVE_VCMPf32r (v4f32 MQPR:$v1), ZR, fc, 1, VCCR:$p1))>;
3222 multiclass unpred_vcmpf_r<int fc> {
3223 def f16 : Pat<(v8i1 (ARMvcmp (v8f16 MQPR:$v1), (v8f16 MQPR:$v2), (i32 fc))),
3224 (v8i1 (MVE_VCMPf16 (v8f16 MQPR:$v1), (v8f16 MQPR:$v2), fc))>;
3225 def f32 : Pat<(v4i1 (ARMvcmp (v4f32 MQPR:$v1), (v4f32 MQPR:$v2), (i32 fc))),
3226 (v4i1 (MVE_VCMPf32 (v4f32 MQPR:$v1), (v4f32 MQPR:$v2), fc))>;
3228 def f16r : Pat<(v8i1 (ARMvcmp (v8f16 MQPR:$v1), (v8f16 (ARMvdup HPR:$v2)), (i32 fc))),
3229 (v8i1 (MVE_VCMPf16r (v8f16 MQPR:$v1), (i32 (COPY_TO_REGCLASS (f16 HPR:$v2), rGPR)), fc))>;
3230 def f32r : Pat<(v4i1 (ARMvcmp (v4f32 MQPR:$v1), (v4f32 (ARMvdup SPR:$v2)), (i32 fc))),
3231 (v4i1 (MVE_VCMPf32r (v4f32 MQPR:$v1), (i32 (COPY_TO_REGCLASS (f32 SPR:$v2), rGPR)), fc))>;
3233 def : Pat<(v8i1 (and (v8i1 VCCR:$p1), (v8i1 (ARMvcmp (v8f16 MQPR:$v1), (v8f16 MQPR:$v2), (i32 fc))))),
3234 (v8i1 (MVE_VCMPf16 (v8f16 MQPR:$v1), (v8f16 MQPR:$v2), fc, 1, VCCR:$p1))>;
3235 def : Pat<(v4i1 (and (v4i1 VCCR:$p1), (v4i1 (ARMvcmp (v4f32 MQPR:$v1), (v4f32 MQPR:$v2), (i32 fc))))),
3236 (v4i1 (MVE_VCMPf32 (v4f32 MQPR:$v1), (v4f32 MQPR:$v2), fc, 1, VCCR:$p1))>;
3238 def : Pat<(v8i1 (and (v8i1 VCCR:$p1), (v8i1 (ARMvcmp (v8f16 MQPR:$v1), (v8f16 (ARMvdup HPR:$v2)), (i32 fc))))),
3239 (v8i1 (MVE_VCMPf16r (v8f16 MQPR:$v1), (i32 (COPY_TO_REGCLASS (f16 HPR:$v2), rGPR)), fc, 1, VCCR:$p1))>;
3240 def : Pat<(v4i1 (and (v4i1 VCCR:$p1), (v4i1 (ARMvcmp (v4f32 MQPR:$v1), (v4f32 (ARMvdup SPR:$v2)), (i32 fc))))),
3241 (v4i1 (MVE_VCMPf32r (v4f32 MQPR:$v1), (i32 (COPY_TO_REGCLASS (f32 SPR:$v2), rGPR)), fc, 1, VCCR:$p1))>;
3244 let Predicates = [HasMVEInt] in {
3245 defm MVE_VCEQZ : unpred_vcmp_z<"i", 0>;
3246 defm MVE_VCNEZ : unpred_vcmp_z<"i", 1>;
3247 defm MVE_VCGEZ : unpred_vcmp_z<"s", 10>;
3248 defm MVE_VCLTZ : unpred_vcmp_z<"s", 11>;
3249 defm MVE_VCGTZ : unpred_vcmp_z<"s", 12>;
3250 defm MVE_VCLEZ : unpred_vcmp_z<"s", 13>;
3251 defm MVE_VCGTUZ : unpred_vcmp_z<"u", 8>;
3252 defm MVE_VCGEUZ : unpred_vcmp_z<"u", 2>;
3254 defm MVE_VCEQ : unpred_vcmp_r<"i", 0>;
3255 defm MVE_VCNE : unpred_vcmp_r<"i", 1>;
3256 defm MVE_VCGE : unpred_vcmp_r<"s", 10>;
3257 defm MVE_VCLT : unpred_vcmp_r<"s", 11>;
3258 defm MVE_VCGT : unpred_vcmp_r<"s", 12>;
3259 defm MVE_VCLE : unpred_vcmp_r<"s", 13>;
3260 defm MVE_VCGTU : unpred_vcmp_r<"u", 8>;
3261 defm MVE_VCGEU : unpred_vcmp_r<"u", 2>;
3264 let Predicates = [HasMVEFloat] in {
3265 defm MVE_VFCEQZ : unpred_vcmpf_z<0>;
3266 defm MVE_VFCNEZ : unpred_vcmpf_z<1>;
3267 defm MVE_VFCGEZ : unpred_vcmpf_z<10>;
3268 defm MVE_VFCLTZ : unpred_vcmpf_z<11>;
3269 defm MVE_VFCGTZ : unpred_vcmpf_z<12>;
3270 defm MVE_VFCLEZ : unpred_vcmpf_z<13>;
3272 defm MVE_VFCEQ : unpred_vcmpf_r<0>;
3273 defm MVE_VFCNE : unpred_vcmpf_r<1>;
3274 defm MVE_VFCGE : unpred_vcmpf_r<10>;
3275 defm MVE_VFCLT : unpred_vcmpf_r<11>;
3276 defm MVE_VFCGT : unpred_vcmpf_r<12>;
3277 defm MVE_VFCLE : unpred_vcmpf_r<13>;
3281 // Extra "worst case" and/or/xor partterns, going into and out of GRP
3282 multiclass two_predops<SDPatternOperator opnode, Instruction insn> {
3283 def v16i1 : Pat<(v16i1 (opnode (v16i1 VCCR:$p1), (v16i1 VCCR:$p2))),
3284 (v16i1 (COPY_TO_REGCLASS
3285 (insn (i32 (COPY_TO_REGCLASS (v16i1 VCCR:$p1), rGPR)),
3286 (i32 (COPY_TO_REGCLASS (v16i1 VCCR:$p2), rGPR))),
3288 def v8i1 : Pat<(v8i1 (opnode (v8i1 VCCR:$p1), (v8i1 VCCR:$p2))),
3289 (v8i1 (COPY_TO_REGCLASS
3290 (insn (i32 (COPY_TO_REGCLASS (v8i1 VCCR:$p1), rGPR)),
3291 (i32 (COPY_TO_REGCLASS (v8i1 VCCR:$p2), rGPR))),
3293 def v4i1 : Pat<(v4i1 (opnode (v4i1 VCCR:$p1), (v4i1 VCCR:$p2))),
3294 (v4i1 (COPY_TO_REGCLASS
3295 (insn (i32 (COPY_TO_REGCLASS (v4i1 VCCR:$p1), rGPR)),
3296 (i32 (COPY_TO_REGCLASS (v4i1 VCCR:$p2), rGPR))),
3300 let Predicates = [HasMVEInt] in {
3301 defm POR : two_predops<or, t2ORRrr>;
3302 defm PAND : two_predops<and, t2ANDrr>;
3303 defm PEOR : two_predops<xor, t2EORrr>;
3306 // Occasionally we need to cast between a i32 and a boolean vector, for
3307 // example when moving between rGPR and VPR.P0 as part of predicate vector
3308 // shuffles. We also sometimes need to cast between different predicate
3309 // vector types (v4i1<>v8i1, etc.) also as part of lowering vector shuffles.
3311 def predicate_cast : SDNode<"ARMISD::PREDICATE_CAST", SDTUnaryOp>;
3313 let Predicates = [HasMVEInt] in {
3314 foreach VT = [ v4i1, v8i1, v16i1 ] in {
3315 def : Pat<(i32 (predicate_cast (VT VCCR:$src))),
3316 (i32 (COPY_TO_REGCLASS (VT VCCR:$src), VCCR))>;
3317 def : Pat<(VT (predicate_cast (i32 VCCR:$src))),
3318 (VT (COPY_TO_REGCLASS (i32 VCCR:$src), VCCR))>;
3320 foreach VT2 = [ v4i1, v8i1, v16i1 ] in
3321 def : Pat<(VT (predicate_cast (VT2 VCCR:$src))),
3322 (VT (COPY_TO_REGCLASS (VT2 VCCR:$src), VCCR))>;
3326 // end of MVE compares
3328 // start of MVE_qDest_qSrc
3330 class MVE_qDest_qSrc<string iname, string suffix, dag oops, dag iops,
3331 string ops, vpred_ops vpred, string cstr,
3332 list<dag> pattern=[]>
3333 : MVE_p<oops, iops, NoItinerary, iname, suffix,
3334 ops, vpred, cstr, pattern> {
3338 let Inst{25-23} = 0b100;
3339 let Inst{22} = Qd{3};
3340 let Inst{15-13} = Qd{2-0};
3341 let Inst{11-9} = 0b111;
3343 let Inst{5} = Qm{3};
3345 let Inst{3-1} = Qm{2-0};
3348 class MVE_VQxDMLxDH<string iname, bit exch, bit round, bit subtract,
3349 string suffix, bits<2> size, string cstr="", list<dag> pattern=[]>
3350 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
3351 (ins MQPR:$Qd_src, MQPR:$Qn, MQPR:$Qm), "$Qd, $Qn, $Qm",
3352 vpred_n, "$Qd = $Qd_src"#cstr, pattern> {
3355 let Inst{28} = subtract;
3356 let Inst{21-20} = size;
3357 let Inst{19-17} = Qn{2-0};
3359 let Inst{12} = exch;
3361 let Inst{7} = Qn{3};
3362 let Inst{0} = round;
3365 multiclass MVE_VQxDMLxDH_multi<string iname, bit exch,
3366 bit round, bit subtract> {
3367 def s8 : MVE_VQxDMLxDH<iname, exch, round, subtract, "s8", 0b00>;
3368 def s16 : MVE_VQxDMLxDH<iname, exch, round, subtract, "s16", 0b01>;
3369 def s32 : MVE_VQxDMLxDH<iname, exch, round, subtract, "s32", 0b10, ",@earlyclobber $Qd">;
3372 defm MVE_VQDMLADH : MVE_VQxDMLxDH_multi<"vqdmladh", 0b0, 0b0, 0b0>;
3373 defm MVE_VQDMLADHX : MVE_VQxDMLxDH_multi<"vqdmladhx", 0b1, 0b0, 0b0>;
3374 defm MVE_VQRDMLADH : MVE_VQxDMLxDH_multi<"vqrdmladh", 0b0, 0b1, 0b0>;
3375 defm MVE_VQRDMLADHX : MVE_VQxDMLxDH_multi<"vqrdmladhx", 0b1, 0b1, 0b0>;
3376 defm MVE_VQDMLSDH : MVE_VQxDMLxDH_multi<"vqdmlsdh", 0b0, 0b0, 0b1>;
3377 defm MVE_VQDMLSDHX : MVE_VQxDMLxDH_multi<"vqdmlsdhx", 0b1, 0b0, 0b1>;
3378 defm MVE_VQRDMLSDH : MVE_VQxDMLxDH_multi<"vqrdmlsdh", 0b0, 0b1, 0b1>;
3379 defm MVE_VQRDMLSDHX : MVE_VQxDMLxDH_multi<"vqrdmlsdhx", 0b1, 0b1, 0b1>;
3381 class MVE_VCMUL<string iname, string suffix, bit size, string cstr="", list<dag> pattern=[]>
3382 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
3383 (ins MQPR:$Qn, MQPR:$Qm, complexrotateop:$rot),
3384 "$Qd, $Qn, $Qm, $rot", vpred_r, cstr, pattern> {
3388 let Inst{28} = size;
3389 let Inst{21-20} = 0b11;
3390 let Inst{19-17} = Qn{2-0};
3392 let Inst{12} = rot{1};
3394 let Inst{7} = Qn{3};
3395 let Inst{0} = rot{0};
3397 let Predicates = [HasMVEFloat];
3400 def MVE_VCMULf16 : MVE_VCMUL<"vcmul", "f16", 0b0>;
3401 def MVE_VCMULf32 : MVE_VCMUL<"vcmul", "f32", 0b1, "@earlyclobber $Qd">;
3403 class MVE_VMULL<string iname, string suffix, bit bit_28, bits<2> bits_21_20,
3404 bit T, string cstr, list<dag> pattern=[]>
3405 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
3406 (ins MQPR:$Qn, MQPR:$Qm), "$Qd, $Qn, $Qm",
3407 vpred_r, cstr, pattern> {
3412 let Inst{28} = bit_28;
3413 let Inst{21-20} = bits_21_20;
3414 let Inst{19-17} = Qn{2-0};
3418 let Inst{7} = Qn{3};
3422 multiclass MVE_VMULL_multi<string iname, string suffix,
3423 bit bit_28, bits<2> bits_21_20, string cstr=""> {
3424 def bh : MVE_VMULL<iname # "b", suffix, bit_28, bits_21_20, 0b0, cstr>;
3425 def th : MVE_VMULL<iname # "t", suffix, bit_28, bits_21_20, 0b1, cstr>;
3428 // For integer multiplies, bits 21:20 encode size, and bit 28 signedness.
3429 // For polynomial multiplies, bits 21:20 take the unused value 0b11, and
3430 // bit 28 switches to encoding the size.
3432 defm MVE_VMULLs8 : MVE_VMULL_multi<"vmull", "s8", 0b0, 0b00>;
3433 defm MVE_VMULLs16 : MVE_VMULL_multi<"vmull", "s16", 0b0, 0b01>;
3434 defm MVE_VMULLs32 : MVE_VMULL_multi<"vmull", "s32", 0b0, 0b10, "@earlyclobber $Qd">;
3435 defm MVE_VMULLu8 : MVE_VMULL_multi<"vmull", "u8", 0b1, 0b00>;
3436 defm MVE_VMULLu16 : MVE_VMULL_multi<"vmull", "u16", 0b1, 0b01>;
3437 defm MVE_VMULLu32 : MVE_VMULL_multi<"vmull", "u32", 0b1, 0b10, "@earlyclobber $Qd">;
3438 defm MVE_VMULLp8 : MVE_VMULL_multi<"vmull", "p8", 0b0, 0b11>;
3439 defm MVE_VMULLp16 : MVE_VMULL_multi<"vmull", "p16", 0b1, 0b11>;
3441 class MVE_VxMULH<string iname, string suffix, bit U, bits<2> size,
3442 bit round, list<dag> pattern=[]>
3443 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
3444 (ins MQPR:$Qn, MQPR:$Qm), "$Qd, $Qn, $Qm",
3445 vpred_r, "", pattern> {
3449 let Inst{21-20} = size;
3450 let Inst{19-17} = Qn{2-0};
3452 let Inst{12} = round;
3454 let Inst{7} = Qn{3};
3458 def MVE_VMULHs8 : MVE_VxMULH<"vmulh", "s8", 0b0, 0b00, 0b0>;
3459 def MVE_VMULHs16 : MVE_VxMULH<"vmulh", "s16", 0b0, 0b01, 0b0>;
3460 def MVE_VMULHs32 : MVE_VxMULH<"vmulh", "s32", 0b0, 0b10, 0b0>;
3461 def MVE_VMULHu8 : MVE_VxMULH<"vmulh", "u8", 0b1, 0b00, 0b0>;
3462 def MVE_VMULHu16 : MVE_VxMULH<"vmulh", "u16", 0b1, 0b01, 0b0>;
3463 def MVE_VMULHu32 : MVE_VxMULH<"vmulh", "u32", 0b1, 0b10, 0b0>;
3465 def MVE_VRMULHs8 : MVE_VxMULH<"vrmulh", "s8", 0b0, 0b00, 0b1>;
3466 def MVE_VRMULHs16 : MVE_VxMULH<"vrmulh", "s16", 0b0, 0b01, 0b1>;
3467 def MVE_VRMULHs32 : MVE_VxMULH<"vrmulh", "s32", 0b0, 0b10, 0b1>;
3468 def MVE_VRMULHu8 : MVE_VxMULH<"vrmulh", "u8", 0b1, 0b00, 0b1>;
3469 def MVE_VRMULHu16 : MVE_VxMULH<"vrmulh", "u16", 0b1, 0b01, 0b1>;
3470 def MVE_VRMULHu32 : MVE_VxMULH<"vrmulh", "u32", 0b1, 0b10, 0b1>;
3472 class MVE_VxMOVxN<string iname, string suffix, bit bit_28, bit bit_17,
3473 bits<2> size, bit T, list<dag> pattern=[]>
3474 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
3475 (ins MQPR:$Qd_src, MQPR:$Qm), "$Qd, $Qm",
3476 vpred_n, "$Qd = $Qd_src", pattern> {
3478 let Inst{28} = bit_28;
3479 let Inst{21-20} = 0b11;
3480 let Inst{19-18} = size;
3481 let Inst{17} = bit_17;
3485 let Inst{7} = !if(!eq(bit_17, 0), 1, 0);
3489 multiclass MVE_VxMOVxN_halves<string iname, string suffix,
3490 bit bit_28, bit bit_17, bits<2> size> {
3491 def bh : MVE_VxMOVxN<iname # "b", suffix, bit_28, bit_17, size, 0b0>;
3492 def th : MVE_VxMOVxN<iname # "t", suffix, bit_28, bit_17, size, 0b1>;
3495 defm MVE_VMOVNi16 : MVE_VxMOVxN_halves<"vmovn", "i16", 0b1, 0b0, 0b00>;
3496 defm MVE_VMOVNi32 : MVE_VxMOVxN_halves<"vmovn", "i32", 0b1, 0b0, 0b01>;
3497 defm MVE_VQMOVNs16 : MVE_VxMOVxN_halves<"vqmovn", "s16", 0b0, 0b1, 0b00>;
3498 defm MVE_VQMOVNs32 : MVE_VxMOVxN_halves<"vqmovn", "s32", 0b0, 0b1, 0b01>;
3499 defm MVE_VQMOVNu16 : MVE_VxMOVxN_halves<"vqmovn", "u16", 0b1, 0b1, 0b00>;
3500 defm MVE_VQMOVNu32 : MVE_VxMOVxN_halves<"vqmovn", "u32", 0b1, 0b1, 0b01>;
3501 defm MVE_VQMOVUNs16 : MVE_VxMOVxN_halves<"vqmovun", "s16", 0b0, 0b0, 0b00>;
3502 defm MVE_VQMOVUNs32 : MVE_VxMOVxN_halves<"vqmovun", "s32", 0b0, 0b0, 0b01>;
3504 class MVE_VCVT_ff<string iname, string suffix, bit op, bit T,
3505 list<dag> pattern=[]>
3506 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd), (ins MQPR:$Qd_src, MQPR:$Qm),
3507 "$Qd, $Qm", vpred_n, "$Qd = $Qd_src", pattern> {
3509 let Inst{21-16} = 0b111111;
3511 let Inst{8-7} = 0b00;
3514 let Predicates = [HasMVEFloat];
3517 multiclass MVE_VCVT_ff_halves<string suffix, bit op> {
3518 def bh : MVE_VCVT_ff<"vcvtb", suffix, op, 0b0>;
3519 def th : MVE_VCVT_ff<"vcvtt", suffix, op, 0b1>;
3522 defm MVE_VCVTf16f32 : MVE_VCVT_ff_halves<"f16.f32", 0b0>;
3523 defm MVE_VCVTf32f16 : MVE_VCVT_ff_halves<"f32.f16", 0b1>;
3525 class MVE_VxCADD<string iname, string suffix, bits<2> size, bit halve,
3526 string cstr="", list<dag> pattern=[]>
3527 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
3528 (ins MQPR:$Qn, MQPR:$Qm, complexrotateopodd:$rot),
3529 "$Qd, $Qn, $Qm, $rot", vpred_r, cstr, pattern> {
3533 let Inst{28} = halve;
3534 let Inst{21-20} = size;
3535 let Inst{19-17} = Qn{2-0};
3539 let Inst{7} = Qn{3};
3543 def MVE_VCADDi8 : MVE_VxCADD<"vcadd", "i8", 0b00, 0b1>;
3544 def MVE_VCADDi16 : MVE_VxCADD<"vcadd", "i16", 0b01, 0b1>;
3545 def MVE_VCADDi32 : MVE_VxCADD<"vcadd", "i32", 0b10, 0b1, "@earlyclobber $Qd">;
3547 def MVE_VHCADDs8 : MVE_VxCADD<"vhcadd", "s8", 0b00, 0b0>;
3548 def MVE_VHCADDs16 : MVE_VxCADD<"vhcadd", "s16", 0b01, 0b0>;
3549 def MVE_VHCADDs32 : MVE_VxCADD<"vhcadd", "s32", 0b10, 0b0, "@earlyclobber $Qd">;
3551 class MVE_VADCSBC<string iname, bit I, bit subtract,
3552 dag carryin, list<dag> pattern=[]>
3553 : MVE_qDest_qSrc<iname, "i32", (outs MQPR:$Qd, cl_FPSCR_NZCV:$carryout),
3554 !con((ins MQPR:$Qn, MQPR:$Qm), carryin),
3555 "$Qd, $Qn, $Qm", vpred_r, "", pattern> {
3558 let Inst{28} = subtract;
3559 let Inst{21-20} = 0b11;
3560 let Inst{19-17} = Qn{2-0};
3564 let Inst{7} = Qn{3};
3567 // Custom decoder method in order to add the FPSCR operand(s), which
3568 // Tablegen won't do right
3569 let DecoderMethod = "DecodeMVEVADCInstruction";
3572 def MVE_VADC : MVE_VADCSBC<"vadc", 0b0, 0b0, (ins cl_FPSCR_NZCV:$carryin)>;
3573 def MVE_VADCI : MVE_VADCSBC<"vadci", 0b1, 0b0, (ins)>;
3575 def MVE_VSBC : MVE_VADCSBC<"vsbc", 0b0, 0b1, (ins cl_FPSCR_NZCV:$carryin)>;
3576 def MVE_VSBCI : MVE_VADCSBC<"vsbci", 0b1, 0b1, (ins)>;
3578 class MVE_VQDMULL<string iname, string suffix, bit size, bit T,
3579 string cstr="", list<dag> pattern=[]>
3580 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
3581 (ins MQPR:$Qn, MQPR:$Qm), "$Qd, $Qn, $Qm",
3582 vpred_r, cstr, pattern> {
3585 let Inst{28} = size;
3586 let Inst{21-20} = 0b11;
3587 let Inst{19-17} = Qn{2-0};
3591 let Inst{7} = Qn{3};
3595 multiclass MVE_VQDMULL_halves<string suffix, bit size, string cstr=""> {
3596 def bh : MVE_VQDMULL<"vqdmullb", suffix, size, 0b0, cstr>;
3597 def th : MVE_VQDMULL<"vqdmullt", suffix, size, 0b1, cstr>;
3600 defm MVE_VQDMULLs16 : MVE_VQDMULL_halves<"s16", 0b0>;
3601 defm MVE_VQDMULLs32 : MVE_VQDMULL_halves<"s32", 0b1, "@earlyclobber $Qd">;
3603 // end of mve_qDest_qSrc
3605 // start of mve_qDest_rSrc
3607 class MVE_qr_base<dag oops, dag iops, InstrItinClass itin, string iname,
3608 string suffix, string ops, vpred_ops vpred, string cstr,
3609 list<dag> pattern=[]>
3610 : MVE_p<oops, iops, NoItinerary, iname, suffix, ops, vpred, cstr, pattern> {
3615 let Inst{25-23} = 0b100;
3616 let Inst{22} = Qd{3};
3617 let Inst{19-17} = Qn{2-0};
3618 let Inst{15-13} = Qd{2-0};
3619 let Inst{11-9} = 0b111;
3620 let Inst{7} = Qn{3};
3623 let Inst{3-0} = Rm{3-0};
3626 class MVE_qDest_rSrc<string iname, string suffix, string cstr="", list<dag> pattern=[]>
3627 : MVE_qr_base<(outs MQPR:$Qd), (ins MQPR:$Qn, rGPR:$Rm),
3628 NoItinerary, iname, suffix, "$Qd, $Qn, $Rm", vpred_r, cstr,
3631 class MVE_qDestSrc_rSrc<string iname, string suffix, list<dag> pattern=[]>
3632 : MVE_qr_base<(outs MQPR:$Qd), (ins MQPR:$Qd_src, MQPR:$Qn, rGPR:$Rm),
3633 NoItinerary, iname, suffix, "$Qd, $Qn, $Rm", vpred_n, "$Qd = $Qd_src",
3636 class MVE_qDest_single_rSrc<string iname, string suffix, list<dag> pattern=[]>
3637 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qd_src, rGPR:$Rm), NoItinerary, iname,
3638 suffix, "$Qd, $Rm", vpred_n, "$Qd = $Qd_src", pattern> {
3642 let Inst{22} = Qd{3};
3643 let Inst{15-13} = Qd{2-0};
3644 let Inst{3-0} = Rm{3-0};
3647 class MVE_VADDSUB_qr<string iname, string suffix, bits<2> size,
3648 bit bit_5, bit bit_12, bit bit_16,
3649 bit bit_28, list<dag> pattern=[]>
3650 : MVE_qDest_rSrc<iname, suffix, "", pattern> {
3652 let Inst{28} = bit_28;
3653 let Inst{21-20} = size;
3654 let Inst{16} = bit_16;
3655 let Inst{12} = bit_12;
3657 let Inst{5} = bit_5;
3660 multiclass MVE_VADDSUB_qr_sizes<string iname, string suffix,
3661 bit bit_5, bit bit_12, bit bit_16,
3662 bit bit_28, list<dag> pattern=[]> {
3663 def "8" : MVE_VADDSUB_qr<iname, suffix#"8", 0b00,
3664 bit_5, bit_12, bit_16, bit_28>;
3665 def "16" : MVE_VADDSUB_qr<iname, suffix#"16", 0b01,
3666 bit_5, bit_12, bit_16, bit_28>;
3667 def "32" : MVE_VADDSUB_qr<iname, suffix#"32", 0b10,
3668 bit_5, bit_12, bit_16, bit_28>;
3671 defm MVE_VADD_qr_i : MVE_VADDSUB_qr_sizes<"vadd", "i", 0b0, 0b0, 0b1, 0b0>;
3672 defm MVE_VQADD_qr_s : MVE_VADDSUB_qr_sizes<"vqadd", "s", 0b1, 0b0, 0b0, 0b0>;
3673 defm MVE_VQADD_qr_u : MVE_VADDSUB_qr_sizes<"vqadd", "u", 0b1, 0b0, 0b0, 0b1>;
3675 defm MVE_VSUB_qr_i : MVE_VADDSUB_qr_sizes<"vsub", "i", 0b0, 0b1, 0b1, 0b0>;
3676 defm MVE_VQSUB_qr_s : MVE_VADDSUB_qr_sizes<"vqsub", "s", 0b1, 0b1, 0b0, 0b0>;
3677 defm MVE_VQSUB_qr_u : MVE_VADDSUB_qr_sizes<"vqsub", "u", 0b1, 0b1, 0b0, 0b1>;
3679 let Predicates = [HasMVEInt] in {
3680 def : Pat<(v16i8 (add (v16i8 MQPR:$val1), (v16i8 (ARMvdup GPR:$val2)))),
3681 (v16i8 (MVE_VADD_qr_i8 (v16i8 MQPR:$val1), (i32 GPR:$val2)))>;
3682 def : Pat<(v8i16 (add (v8i16 MQPR:$val1), (v8i16 (ARMvdup GPR:$val2)))),
3683 (v8i16 (MVE_VADD_qr_i16 (v8i16 MQPR:$val1), (i32 GPR:$val2)))>;
3684 def : Pat<(v4i32 (add (v4i32 MQPR:$val1), (v4i32 (ARMvdup GPR:$val2)))),
3685 (v4i32 (MVE_VADD_qr_i32 (v4i32 MQPR:$val1), (i32 GPR:$val2)))>;
3688 let Predicates = [HasMVEInt] in {
3689 def : Pat<(v16i8 (sub (v16i8 MQPR:$val1), (v16i8 (ARMvdup GPR:$val2)))),
3690 (v16i8 (MVE_VSUB_qr_i8 (v16i8 MQPR:$val1), (i32 GPR:$val2)))>;
3691 def : Pat<(v8i16 (sub (v8i16 MQPR:$val1), (v8i16 (ARMvdup GPR:$val2)))),
3692 (v8i16 (MVE_VSUB_qr_i16 (v8i16 MQPR:$val1), (i32 GPR:$val2)))>;
3693 def : Pat<(v4i32 (sub (v4i32 MQPR:$val1), (v4i32 (ARMvdup GPR:$val2)))),
3694 (v4i32 (MVE_VSUB_qr_i32 (v4i32 MQPR:$val1), (i32 GPR:$val2)))>;
3697 class MVE_VQDMULL_qr<string iname, string suffix, bit size,
3698 bit T, string cstr="", list<dag> pattern=[]>
3699 : MVE_qDest_rSrc<iname, suffix, cstr, pattern> {
3701 let Inst{28} = size;
3702 let Inst{21-20} = 0b11;
3709 multiclass MVE_VQDMULL_qr_halves<string suffix, bit size, string cstr=""> {
3710 def bh : MVE_VQDMULL_qr<"vqdmullb", suffix, size, 0b0, cstr>;
3711 def th : MVE_VQDMULL_qr<"vqdmullt", suffix, size, 0b1, cstr>;
3714 defm MVE_VQDMULL_qr_s16 : MVE_VQDMULL_qr_halves<"s16", 0b0>;
3715 defm MVE_VQDMULL_qr_s32 : MVE_VQDMULL_qr_halves<"s32", 0b1, "@earlyclobber $Qd">;
3717 class MVE_VxADDSUB_qr<string iname, string suffix,
3718 bit bit_28, bits<2> bits_21_20, bit subtract,
3719 list<dag> pattern=[]>
3720 : MVE_qDest_rSrc<iname, suffix, "", pattern> {
3722 let Inst{28} = bit_28;
3723 let Inst{21-20} = bits_21_20;
3725 let Inst{12} = subtract;
3730 def MVE_VHADD_qr_s8 : MVE_VxADDSUB_qr<"vhadd", "s8", 0b0, 0b00, 0b0>;
3731 def MVE_VHADD_qr_s16 : MVE_VxADDSUB_qr<"vhadd", "s16", 0b0, 0b01, 0b0>;
3732 def MVE_VHADD_qr_s32 : MVE_VxADDSUB_qr<"vhadd", "s32", 0b0, 0b10, 0b0>;
3733 def MVE_VHADD_qr_u8 : MVE_VxADDSUB_qr<"vhadd", "u8", 0b1, 0b00, 0b0>;
3734 def MVE_VHADD_qr_u16 : MVE_VxADDSUB_qr<"vhadd", "u16", 0b1, 0b01, 0b0>;
3735 def MVE_VHADD_qr_u32 : MVE_VxADDSUB_qr<"vhadd", "u32", 0b1, 0b10, 0b0>;
3737 def MVE_VHSUB_qr_s8 : MVE_VxADDSUB_qr<"vhsub", "s8", 0b0, 0b00, 0b1>;
3738 def MVE_VHSUB_qr_s16 : MVE_VxADDSUB_qr<"vhsub", "s16", 0b0, 0b01, 0b1>;
3739 def MVE_VHSUB_qr_s32 : MVE_VxADDSUB_qr<"vhsub", "s32", 0b0, 0b10, 0b1>;
3740 def MVE_VHSUB_qr_u8 : MVE_VxADDSUB_qr<"vhsub", "u8", 0b1, 0b00, 0b1>;
3741 def MVE_VHSUB_qr_u16 : MVE_VxADDSUB_qr<"vhsub", "u16", 0b1, 0b01, 0b1>;
3742 def MVE_VHSUB_qr_u32 : MVE_VxADDSUB_qr<"vhsub", "u32", 0b1, 0b10, 0b1>;
3744 let Predicates = [HasMVEFloat] in {
3745 def MVE_VADD_qr_f32 : MVE_VxADDSUB_qr<"vadd", "f32", 0b0, 0b11, 0b0>;
3746 def MVE_VADD_qr_f16 : MVE_VxADDSUB_qr<"vadd", "f16", 0b1, 0b11, 0b0>;
3748 def MVE_VSUB_qr_f32 : MVE_VxADDSUB_qr<"vsub", "f32", 0b0, 0b11, 0b1>;
3749 def MVE_VSUB_qr_f16 : MVE_VxADDSUB_qr<"vsub", "f16", 0b1, 0b11, 0b1>;
3752 class MVE_VxSHL_qr<string iname, string suffix, bit U, bits<2> size,
3753 bit bit_7, bit bit_17, list<dag> pattern=[]>
3754 : MVE_qDest_single_rSrc<iname, suffix, pattern> {
3757 let Inst{25-23} = 0b100;
3758 let Inst{21-20} = 0b11;
3759 let Inst{19-18} = size;
3760 let Inst{17} = bit_17;
3762 let Inst{12-8} = 0b11110;
3763 let Inst{7} = bit_7;
3764 let Inst{6-4} = 0b110;
3767 multiclass MVE_VxSHL_qr_types<string iname, bit bit_7, bit bit_17> {
3768 def s8 : MVE_VxSHL_qr<iname, "s8", 0b0, 0b00, bit_7, bit_17>;
3769 def s16 : MVE_VxSHL_qr<iname, "s16", 0b0, 0b01, bit_7, bit_17>;
3770 def s32 : MVE_VxSHL_qr<iname, "s32", 0b0, 0b10, bit_7, bit_17>;
3771 def u8 : MVE_VxSHL_qr<iname, "u8", 0b1, 0b00, bit_7, bit_17>;
3772 def u16 : MVE_VxSHL_qr<iname, "u16", 0b1, 0b01, bit_7, bit_17>;
3773 def u32 : MVE_VxSHL_qr<iname, "u32", 0b1, 0b10, bit_7, bit_17>;
3776 defm MVE_VSHL_qr : MVE_VxSHL_qr_types<"vshl", 0b0, 0b0>;
3777 defm MVE_VRSHL_qr : MVE_VxSHL_qr_types<"vrshl", 0b0, 0b1>;
3778 defm MVE_VQSHL_qr : MVE_VxSHL_qr_types<"vqshl", 0b1, 0b0>;
3779 defm MVE_VQRSHL_qr : MVE_VxSHL_qr_types<"vqrshl", 0b1, 0b1>;
3781 let Predicates = [HasMVEInt] in {
3782 def : Pat<(v4i32 (ARMvshlu (v4i32 MQPR:$Qm), (v4i32 (ARMvdup GPR:$Rm)))),
3783 (v4i32 (MVE_VSHL_qru32 (v4i32 MQPR:$Qm), GPR:$Rm))>;
3784 def : Pat<(v8i16 (ARMvshlu (v8i16 MQPR:$Qm), (v8i16 (ARMvdup GPR:$Rm)))),
3785 (v8i16 (MVE_VSHL_qru16 (v8i16 MQPR:$Qm), GPR:$Rm))>;
3786 def : Pat<(v16i8 (ARMvshlu (v16i8 MQPR:$Qm), (v16i8 (ARMvdup GPR:$Rm)))),
3787 (v16i8 (MVE_VSHL_qru8 (v16i8 MQPR:$Qm), GPR:$Rm))>;
3789 def : Pat<(v4i32 (ARMvshls (v4i32 MQPR:$Qm), (v4i32 (ARMvdup GPR:$Rm)))),
3790 (v4i32 (MVE_VSHL_qrs32 (v4i32 MQPR:$Qm), GPR:$Rm))>;
3791 def : Pat<(v8i16 (ARMvshls (v8i16 MQPR:$Qm), (v8i16 (ARMvdup GPR:$Rm)))),
3792 (v8i16 (MVE_VSHL_qrs16 (v8i16 MQPR:$Qm), GPR:$Rm))>;
3793 def : Pat<(v16i8 (ARMvshls (v16i8 MQPR:$Qm), (v16i8 (ARMvdup GPR:$Rm)))),
3794 (v16i8 (MVE_VSHL_qrs8 (v16i8 MQPR:$Qm), GPR:$Rm))>;
3797 class MVE_VBRSR<string iname, string suffix, bits<2> size, list<dag> pattern=[]>
3798 : MVE_qDest_rSrc<iname, suffix, "", pattern> {
3801 let Inst{21-20} = size;
3808 def MVE_VBRSR8 : MVE_VBRSR<"vbrsr", "8", 0b00>;
3809 def MVE_VBRSR16 : MVE_VBRSR<"vbrsr", "16", 0b01>;
3810 def MVE_VBRSR32 : MVE_VBRSR<"vbrsr", "32", 0b10>;
3812 let Predicates = [HasMVEInt] in {
3813 def : Pat<(v16i8 ( bitreverse (v16i8 MQPR:$val1))),
3814 (v16i8 ( MVE_VBRSR8 (v16i8 MQPR:$val1), (t2MOVi (i32 8)) ))>;
3816 def : Pat<(v4i32 ( bitreverse (v4i32 MQPR:$val1))),
3817 (v4i32 ( MVE_VBRSR32 (v4i32 MQPR:$val1), (t2MOVi (i32 32)) ))>;
3819 def : Pat<(v8i16 ( bitreverse (v8i16 MQPR:$val1))),
3820 (v8i16 ( MVE_VBRSR16 (v8i16 MQPR:$val1), (t2MOVi (i32 16)) ))>;
3823 class MVE_VMUL_qr_int<string iname, string suffix,
3824 bits<2> size, list<dag> pattern=[]>
3825 : MVE_qDest_rSrc<iname, suffix, "", pattern> {
3828 let Inst{21-20} = size;
3835 def MVE_VMUL_qr_i8 : MVE_VMUL_qr_int<"vmul", "i8", 0b00>;
3836 def MVE_VMUL_qr_i16 : MVE_VMUL_qr_int<"vmul", "i16", 0b01>;
3837 def MVE_VMUL_qr_i32 : MVE_VMUL_qr_int<"vmul", "i32", 0b10>;
3839 let Predicates = [HasMVEInt] in {
3840 def : Pat<(v16i8 (mul (v16i8 MQPR:$val1), (v16i8 (ARMvdup GPR:$val2)))),
3841 (v16i8 (MVE_VMUL_qr_i8 (v16i8 MQPR:$val1), (i32 GPR:$val2)))>;
3842 def : Pat<(v8i16 (mul (v8i16 MQPR:$val1), (v8i16 (ARMvdup GPR:$val2)))),
3843 (v8i16 (MVE_VMUL_qr_i16 (v8i16 MQPR:$val1), (i32 GPR:$val2)))>;
3844 def : Pat<(v4i32 (mul (v4i32 MQPR:$val1), (v4i32 (ARMvdup GPR:$val2)))),
3845 (v4i32 (MVE_VMUL_qr_i32 (v4i32 MQPR:$val1), (i32 GPR:$val2)))>;
3848 class MVE_VxxMUL_qr<string iname, string suffix,
3849 bit bit_28, bits<2> bits_21_20, list<dag> pattern=[]>
3850 : MVE_qDest_rSrc<iname, suffix, "", pattern> {
3852 let Inst{28} = bit_28;
3853 let Inst{21-20} = bits_21_20;
3860 def MVE_VQDMULH_qr_s8 : MVE_VxxMUL_qr<"vqdmulh", "s8", 0b0, 0b00>;
3861 def MVE_VQDMULH_qr_s16 : MVE_VxxMUL_qr<"vqdmulh", "s16", 0b0, 0b01>;
3862 def MVE_VQDMULH_qr_s32 : MVE_VxxMUL_qr<"vqdmulh", "s32", 0b0, 0b10>;
3864 def MVE_VQRDMULH_qr_s8 : MVE_VxxMUL_qr<"vqrdmulh", "s8", 0b1, 0b00>;
3865 def MVE_VQRDMULH_qr_s16 : MVE_VxxMUL_qr<"vqrdmulh", "s16", 0b1, 0b01>;
3866 def MVE_VQRDMULH_qr_s32 : MVE_VxxMUL_qr<"vqrdmulh", "s32", 0b1, 0b10>;
3868 let Predicates = [HasMVEFloat] in {
3869 def MVE_VMUL_qr_f16 : MVE_VxxMUL_qr<"vmul", "f16", 0b1, 0b11>;
3870 def MVE_VMUL_qr_f32 : MVE_VxxMUL_qr<"vmul", "f32", 0b0, 0b11>;
3873 class MVE_VFMAMLA_qr<string iname, string suffix,
3874 bit bit_28, bits<2> bits_21_20, bit S,
3875 list<dag> pattern=[]>
3876 : MVE_qDestSrc_rSrc<iname, suffix, pattern> {
3878 let Inst{28} = bit_28;
3879 let Inst{21-20} = bits_21_20;
3886 def MVE_VMLA_qr_s8 : MVE_VFMAMLA_qr<"vmla", "s8", 0b0, 0b00, 0b0>;
3887 def MVE_VMLA_qr_s16 : MVE_VFMAMLA_qr<"vmla", "s16", 0b0, 0b01, 0b0>;
3888 def MVE_VMLA_qr_s32 : MVE_VFMAMLA_qr<"vmla", "s32", 0b0, 0b10, 0b0>;
3889 def MVE_VMLA_qr_u8 : MVE_VFMAMLA_qr<"vmla", "u8", 0b1, 0b00, 0b0>;
3890 def MVE_VMLA_qr_u16 : MVE_VFMAMLA_qr<"vmla", "u16", 0b1, 0b01, 0b0>;
3891 def MVE_VMLA_qr_u32 : MVE_VFMAMLA_qr<"vmla", "u32", 0b1, 0b10, 0b0>;
3893 def MVE_VMLAS_qr_s8 : MVE_VFMAMLA_qr<"vmlas", "s8", 0b0, 0b00, 0b1>;
3894 def MVE_VMLAS_qr_s16 : MVE_VFMAMLA_qr<"vmlas", "s16", 0b0, 0b01, 0b1>;
3895 def MVE_VMLAS_qr_s32 : MVE_VFMAMLA_qr<"vmlas", "s32", 0b0, 0b10, 0b1>;
3896 def MVE_VMLAS_qr_u8 : MVE_VFMAMLA_qr<"vmlas", "u8", 0b1, 0b00, 0b1>;
3897 def MVE_VMLAS_qr_u16 : MVE_VFMAMLA_qr<"vmlas", "u16", 0b1, 0b01, 0b1>;
3898 def MVE_VMLAS_qr_u32 : MVE_VFMAMLA_qr<"vmlas", "u32", 0b1, 0b10, 0b1>;
3900 let Predicates = [HasMVEInt] in {
3901 def : Pat<(v4i32 (add (v4i32 MQPR:$src1),
3902 (v4i32 (mul (v4i32 MQPR:$src2),
3903 (v4i32 (ARMvdup (i32 rGPR:$x))))))),
3904 (v4i32 (MVE_VMLA_qr_u32 $src1, $src2, $x))>;
3905 def : Pat<(v8i16 (add (v8i16 MQPR:$src1),
3906 (v8i16 (mul (v8i16 MQPR:$src2),
3907 (v8i16 (ARMvdup (i32 rGPR:$x))))))),
3908 (v8i16 (MVE_VMLA_qr_u16 $src1, $src2, $x))>;
3909 def : Pat<(v16i8 (add (v16i8 MQPR:$src1),
3910 (v16i8 (mul (v16i8 MQPR:$src2),
3911 (v16i8 (ARMvdup (i32 rGPR:$x))))))),
3912 (v16i8 (MVE_VMLA_qr_u8 $src1, $src2, $x))>;
3915 let Predicates = [HasMVEFloat] in {
3916 def MVE_VFMA_qr_f16 : MVE_VFMAMLA_qr<"vfma", "f16", 0b1, 0b11, 0b0>;
3917 def MVE_VFMA_qr_f32 : MVE_VFMAMLA_qr<"vfma", "f32", 0b0, 0b11, 0b0>;
3918 def MVE_VFMA_qr_Sf16 : MVE_VFMAMLA_qr<"vfmas", "f16", 0b1, 0b11, 0b1>;
3919 def MVE_VFMA_qr_Sf32 : MVE_VFMAMLA_qr<"vfmas", "f32", 0b0, 0b11, 0b1>;
3922 class MVE_VQDMLAH_qr<string iname, string suffix, bit U, bits<2> size,
3923 bit bit_5, bit bit_12, list<dag> pattern=[]>
3924 : MVE_qDestSrc_rSrc<iname, suffix, pattern> {
3927 let Inst{21-20} = size;
3929 let Inst{12} = bit_12;
3931 let Inst{5} = bit_5;
3934 multiclass MVE_VQDMLAH_qr_types<string iname, bit bit_5, bit bit_12> {
3935 def s8 : MVE_VQDMLAH_qr<iname, "s8", 0b0, 0b00, bit_5, bit_12>;
3936 def s16 : MVE_VQDMLAH_qr<iname, "s16", 0b0, 0b01, bit_5, bit_12>;
3937 def s32 : MVE_VQDMLAH_qr<iname, "s32", 0b0, 0b10, bit_5, bit_12>;
3940 defm MVE_VQDMLAH_qr : MVE_VQDMLAH_qr_types<"vqdmlah", 0b1, 0b0>;
3941 defm MVE_VQRDMLAH_qr : MVE_VQDMLAH_qr_types<"vqrdmlah", 0b0, 0b0>;
3942 defm MVE_VQDMLASH_qr : MVE_VQDMLAH_qr_types<"vqdmlash", 0b1, 0b1>;
3943 defm MVE_VQRDMLASH_qr : MVE_VQDMLAH_qr_types<"vqrdmlash", 0b0, 0b1>;
3945 class MVE_VxDUP<string iname, string suffix, bits<2> size, bit bit_12,
3946 list<dag> pattern=[]>
3947 : MVE_p<(outs MQPR:$Qd, tGPREven:$Rn),
3948 (ins tGPREven:$Rn_src, MVE_VIDUP_imm:$imm), NoItinerary,
3949 iname, suffix, "$Qd, $Rn, $imm", vpred_r, "$Rn = $Rn_src",
3956 let Inst{25-23} = 0b100;
3957 let Inst{22} = Qd{3};
3958 let Inst{21-20} = size;
3959 let Inst{19-17} = Rn{3-1};
3961 let Inst{15-13} = Qd{2-0};
3962 let Inst{12} = bit_12;
3963 let Inst{11-8} = 0b1111;
3964 let Inst{7} = imm{1};
3965 let Inst{6-1} = 0b110111;
3966 let Inst{0} = imm{0};
3969 def MVE_VIDUPu8 : MVE_VxDUP<"vidup", "u8", 0b00, 0b0>;
3970 def MVE_VIDUPu16 : MVE_VxDUP<"vidup", "u16", 0b01, 0b0>;
3971 def MVE_VIDUPu32 : MVE_VxDUP<"vidup", "u32", 0b10, 0b0>;
3973 def MVE_VDDUPu8 : MVE_VxDUP<"vddup", "u8", 0b00, 0b1>;
3974 def MVE_VDDUPu16 : MVE_VxDUP<"vddup", "u16", 0b01, 0b1>;
3975 def MVE_VDDUPu32 : MVE_VxDUP<"vddup", "u32", 0b10, 0b1>;
3977 class MVE_VxWDUP<string iname, string suffix, bits<2> size, bit bit_12,
3978 list<dag> pattern=[]>
3979 : MVE_p<(outs MQPR:$Qd, tGPREven:$Rn),
3980 (ins tGPREven:$Rn_src, tGPROdd:$Rm, MVE_VIDUP_imm:$imm), NoItinerary,
3981 iname, suffix, "$Qd, $Rn, $Rm, $imm", vpred_r, "$Rn = $Rn_src",
3989 let Inst{25-23} = 0b100;
3990 let Inst{22} = Qd{3};
3991 let Inst{21-20} = size;
3992 let Inst{19-17} = Rn{3-1};
3994 let Inst{15-13} = Qd{2-0};
3995 let Inst{12} = bit_12;
3996 let Inst{11-8} = 0b1111;
3997 let Inst{7} = imm{1};
3998 let Inst{6-4} = 0b110;
3999 let Inst{3-1} = Rm{3-1};
4000 let Inst{0} = imm{0};
4003 def MVE_VIWDUPu8 : MVE_VxWDUP<"viwdup", "u8", 0b00, 0b0>;
4004 def MVE_VIWDUPu16 : MVE_VxWDUP<"viwdup", "u16", 0b01, 0b0>;
4005 def MVE_VIWDUPu32 : MVE_VxWDUP<"viwdup", "u32", 0b10, 0b0>;
4007 def MVE_VDWDUPu8 : MVE_VxWDUP<"vdwdup", "u8", 0b00, 0b1>;
4008 def MVE_VDWDUPu16 : MVE_VxWDUP<"vdwdup", "u16", 0b01, 0b1>;
4009 def MVE_VDWDUPu32 : MVE_VxWDUP<"vdwdup", "u32", 0b10, 0b1>;
4011 let hasSideEffects = 1 in
4012 class MVE_VCTP<string suffix, bits<2> size, list<dag> pattern=[]>
4013 : MVE_p<(outs VCCR:$P0), (ins rGPR:$Rn), NoItinerary, "vctp", suffix,
4014 "$Rn", vpred_n, "", pattern> {
4017 let Inst{28-27} = 0b10;
4018 let Inst{26-22} = 0b00000;
4019 let Inst{21-20} = size;
4020 let Inst{19-16} = Rn{3-0};
4021 let Inst{15-11} = 0b11101;
4022 let Inst{10-0} = 0b00000000001;
4023 let Unpredictable{10-0} = 0b11111111111;
4025 let Constraints = "";
4026 let DecoderMethod = "DecodeMveVCTP";
4029 def MVE_VCTP8 : MVE_VCTP<"8", 0b00>;
4030 def MVE_VCTP16 : MVE_VCTP<"16", 0b01>;
4031 def MVE_VCTP32 : MVE_VCTP<"32", 0b10>;
4032 def MVE_VCTP64 : MVE_VCTP<"64", 0b11>;
4034 let Predicates = [HasMVEInt] in {
4035 def : Pat<(int_arm_vctp8 rGPR:$Rn),
4036 (v16i1 (MVE_VCTP8 rGPR:$Rn))>;
4037 def : Pat<(int_arm_vctp16 rGPR:$Rn),
4038 (v8i1 (MVE_VCTP16 rGPR:$Rn))>;
4039 def : Pat<(int_arm_vctp32 rGPR:$Rn),
4040 (v4i1 (MVE_VCTP32 rGPR:$Rn))>;
4043 // end of mve_qDest_rSrc
4045 // start of coproc mov
4047 class MVE_VMOV_64bit<dag oops, dag iops, bit to_qreg, string ops, string cstr>
4048 : MVE_VMOV_lane_base<oops, !con(iops, (ins MVEPairVectorIndex2:$idx,
4049 MVEPairVectorIndex0:$idx2)),
4050 NoItinerary, "vmov", "", ops, cstr, []> {
4057 let Inst{31-23} = 0b111011000;
4058 let Inst{22} = Qd{3};
4060 let Inst{20} = to_qreg;
4061 let Inst{19-16} = Rt2{3-0};
4062 let Inst{15-13} = Qd{2-0};
4063 let Inst{12-5} = 0b01111000;
4065 let Inst{3-0} = Rt{3-0};
4068 // The assembly syntax for these instructions mentions the vector
4069 // register name twice, e.g.
4071 // vmov q2[2], q2[0], r0, r1
4072 // vmov r0, r1, q2[2], q2[0]
4074 // which needs a bit of juggling with MC operand handling.
4076 // For the move _into_ a vector register, the MC operand list also has
4077 // to mention the register name twice: once as the output, and once as
4078 // an extra input to represent where the unchanged half of the output
4079 // register comes from (when this instruction is used in code
4080 // generation). So we arrange that the first mention of the vector reg
4081 // in the instruction is considered by the AsmMatcher to be the output
4082 // ($Qd), and the second one is the input ($QdSrc). Binding them
4083 // together with the existing 'tie' constraint is enough to enforce at
4084 // register allocation time that they have to be the same register.
4086 // For the move _from_ a vector register, there's no way to get round
4087 // the fact that both instances of that register name have to be
4088 // inputs. They have to be the same register again, but this time, we
4089 // can't use a tie constraint, because that has to be between an
4090 // output and an input operand. So this time, we have to arrange that
4091 // the q-reg appears just once in the MC operand list, in spite of
4092 // being mentioned twice in the asm syntax - which needs a custom
4093 // AsmMatchConverter.
4095 def MVE_VMOV_q_rr : MVE_VMOV_64bit<(outs MQPR:$Qd),
4096 (ins MQPR:$QdSrc, rGPR:$Rt, rGPR:$Rt2),
4097 0b1, "$Qd$idx, $QdSrc$idx2, $Rt, $Rt2",
4099 let DecoderMethod = "DecodeMVEVMOVDRegtoQ";
4102 def MVE_VMOV_rr_q : MVE_VMOV_64bit<(outs rGPR:$Rt, rGPR:$Rt2), (ins MQPR:$Qd),
4103 0b0, "$Rt, $Rt2, $Qd$idx, $Qd$idx2", ""> {
4104 let DecoderMethod = "DecodeMVEVMOVQtoDReg";
4105 let AsmMatchConverter = "cvtMVEVMOVQtoDReg";
4108 // end of coproc mov
4110 // start of MVE interleaving load/store
4112 // Base class for the family of interleaving/deinterleaving
4113 // load/stores with names like VLD20.8 and VST43.32.
4114 class MVE_vldst24_base<bit writeback, bit fourregs, bits<2> stage, bits<2> size,
4115 bit load, dag Oops, dag loadIops, dag wbIops,
4116 string iname, string ops,
4117 string cstr, list<dag> pattern=[]>
4118 : MVE_MI<Oops, !con(loadIops, wbIops), NoItinerary, iname, ops, cstr, pattern> {
4122 let Inst{31-22} = 0b1111110010;
4123 let Inst{21} = writeback;
4124 let Inst{20} = load;
4125 let Inst{19-16} = Rn;
4126 let Inst{15-13} = VQd{2-0};
4127 let Inst{12-9} = 0b1111;
4128 let Inst{8-7} = size;
4129 let Inst{6-5} = stage;
4130 let Inst{4-1} = 0b0000;
4131 let Inst{0} = fourregs;
4134 let mayStore = !eq(load,0);
4137 // A parameter class used to encapsulate all the ways the writeback
4138 // variants of VLD20 and friends differ from the non-writeback ones.
4139 class MVE_vldst24_writeback<bit b, dag Oo, dag Io,
4140 string sy="", string c="", string n=""> {
4146 string id_suffix = n;
4149 // Another parameter class that encapsulates the differences between VLD2x
4151 class MVE_vldst24_nvecs<int n, list<int> s, bit b, RegisterOperand vl> {
4153 list<int> stages = s;
4155 RegisterOperand VecList = vl;
4158 // A third parameter class that distinguishes VLDnn.8 from .16 from .32.
4159 class MVE_vldst24_lanesize<int i, bits<2> b> {
4161 bits<2> sizebits = b;
4164 // A base class for each direction of transfer: one for load, one for
4165 // store. I can't make these a fourth independent parametric tuple
4166 // class, because they have to take the nvecs tuple class as a
4167 // parameter, in order to find the right VecList operand type.
4169 class MVE_vld24_base<MVE_vldst24_nvecs n, bits<2> pat, bits<2> size,
4170 MVE_vldst24_writeback wb, string iname,
4171 list<dag> pattern=[]>
4172 : MVE_vldst24_base<wb.writeback, n.bit0, pat, size, 1,
4173 !con((outs n.VecList:$VQd), wb.Oops),
4174 (ins n.VecList:$VQdSrc), wb.Iops,
4175 iname, "$VQd, $Rn" # wb.syntax,
4176 wb.cstr # ",$VQdSrc = $VQd", pattern>;
4178 class MVE_vst24_base<MVE_vldst24_nvecs n, bits<2> pat, bits<2> size,
4179 MVE_vldst24_writeback wb, string iname,
4180 list<dag> pattern=[]>
4181 : MVE_vldst24_base<wb.writeback, n.bit0, pat, size, 0,
4182 wb.Oops, (ins n.VecList:$VQd), wb.Iops,
4183 iname, "$VQd, $Rn" # wb.syntax,
4186 // Actually define all the interleaving loads and stores, by a series
4187 // of nested foreaches over number of vectors (VLD2/VLD4); stage
4188 // within one of those series (VLDx0/VLDx1/VLDx2/VLDx3); size of
4189 // vector lane; writeback or no writeback.
4190 foreach n = [MVE_vldst24_nvecs<2, [0,1], 0, VecList2Q>,
4191 MVE_vldst24_nvecs<4, [0,1,2,3], 1, VecList4Q>] in
4192 foreach stage = n.stages in
4193 foreach s = [MVE_vldst24_lanesize< 8, 0b00>,
4194 MVE_vldst24_lanesize<16, 0b01>,
4195 MVE_vldst24_lanesize<32, 0b10>] in
4196 foreach wb = [MVE_vldst24_writeback<
4197 1, (outs rGPR:$wb), (ins t2_nosp_addr_offset_none:$Rn),
4198 "!", "$Rn.base = $wb", "_wb">,
4199 MVE_vldst24_writeback<0, (outs), (ins t2_addr_offset_none:$Rn)>] in {
4201 // For each case within all of those foreaches, define the actual
4202 // instructions. The def names are made by gluing together pieces
4203 // from all the parameter classes, and will end up being things like
4204 // MVE_VLD20_8 and MVE_VST43_16_wb.
4206 def "MVE_VLD" # n.nvecs # stage # "_" # s.lanesize # wb.id_suffix
4207 : MVE_vld24_base<n, stage, s.sizebits, wb,
4208 "vld" # n.nvecs # stage # "." # s.lanesize>;
4210 def "MVE_VST" # n.nvecs # stage # "_" # s.lanesize # wb.id_suffix
4211 : MVE_vst24_base<n, stage, s.sizebits, wb,
4212 "vst" # n.nvecs # stage # "." # s.lanesize>;
4215 // end of MVE interleaving load/store
4217 // start of MVE predicable load/store
4219 // A parameter class for the direction of transfer.
4220 class MVE_ldst_direction<bit b, dag Oo, dag Io, string c=""> {
4226 def MVE_ld: MVE_ldst_direction<1, (outs MQPR:$Qd), (ins), ",@earlyclobber $Qd">;
4227 def MVE_st: MVE_ldst_direction<0, (outs), (ins MQPR:$Qd)>;
4229 // A parameter class for the size of memory access in a load.
4230 class MVE_memsz<bits<2> e, int s, AddrMode m, string mn, list<string> types> {
4231 bits<2> encoding = e; // opcode bit(s) for encoding
4232 int shift = s; // shift applied to immediate load offset
4235 // For instruction aliases: define the complete list of type
4236 // suffixes at this size, and the canonical ones for loads and
4238 string MnemonicLetter = mn;
4239 int TypeBits = !shl(8, s);
4240 string CanonLoadSuffix = ".u" # TypeBits;
4241 string CanonStoreSuffix = "." # TypeBits;
4242 list<string> suffixes = !foreach(letter, types, "." # letter # TypeBits);
4245 // Instances of MVE_memsz.
4247 // (memD doesn't need an AddrMode, because those are only for
4248 // contiguous loads, and memD is only used by gather/scatters.)
4249 def MVE_memB: MVE_memsz<0b00, 0, AddrModeT2_i7, "b", ["", "u", "s"]>;
4250 def MVE_memH: MVE_memsz<0b01, 1, AddrModeT2_i7s2, "h", ["", "u", "s", "f"]>;
4251 def MVE_memW: MVE_memsz<0b10, 2, AddrModeT2_i7s4, "w", ["", "u", "s", "f"]>;
4252 def MVE_memD: MVE_memsz<0b11, 3, ?, "d", ["", "u", "s", "f"]>;
4254 // This is the base class for all the MVE loads and stores other than
4255 // the interleaving ones. All the non-interleaving loads/stores share
4256 // the characteristic that they operate on just one vector register,
4257 // so they are VPT-predicable.
4259 // The predication operand is vpred_n, for both loads and stores. For
4260 // store instructions, the reason is obvious: if there is no output
4261 // register, there can't be a need for an input parameter giving the
4262 // output register's previous value. Load instructions also don't need
4263 // that input parameter, because unlike MVE data processing
4264 // instructions, predicated loads are defined to set the inactive
4265 // lanes of the output register to zero, instead of preserving their
4267 class MVE_VLDRSTR_base<MVE_ldst_direction dir, bit U, bit P, bit W, bit opc,
4268 dag oops, dag iops, string asm, string suffix,
4269 string ops, string cstr, list<dag> pattern=[]>
4270 : MVE_p<oops, iops, NoItinerary, asm, suffix, ops, vpred_n, cstr, pattern> {
4278 let Inst{20} = dir.load;
4279 let Inst{15-13} = Qd{2-0};
4281 let Inst{11-9} = 0b111;
4283 let mayLoad = dir.load;
4284 let mayStore = !eq(dir.load,0);
4287 // Contiguous load and store instructions. These come in two main
4288 // categories: same-size loads/stores in which 128 bits of vector
4289 // register is transferred to or from 128 bits of memory in the most
4290 // obvious way, and widening loads / narrowing stores, in which the
4291 // size of memory accessed is less than the size of a vector register,
4292 // so the load instructions sign- or zero-extend each memory value
4293 // into a wider vector lane, and the store instructions truncate
4296 // The instruction mnemonics for these two classes look reasonably
4297 // similar, but the actual encodings are different enough to need two
4298 // separate base classes.
4300 // Contiguous, same size
4301 class MVE_VLDRSTR_cs<MVE_ldst_direction dir, MVE_memsz memsz, bit P, bit W,
4302 dag oops, dag iops, string asm, string suffix,
4303 IndexMode im, string ops, string cstr>
4304 : MVE_VLDRSTR_base<dir, 0, P, W, 1, oops, iops, asm, suffix, ops, cstr> {
4306 let Inst{23} = addr{7};
4307 let Inst{19-16} = addr{11-8};
4308 let Inst{8-7} = memsz.encoding;
4309 let Inst{6-0} = addr{6-0};
4312 // Contiguous, widening/narrowing
4313 class MVE_VLDRSTR_cw<MVE_ldst_direction dir, MVE_memsz memsz, bit U,
4314 bit P, bit W, bits<2> size, dag oops, dag iops,
4315 string asm, string suffix, IndexMode im,
4316 string ops, string cstr>
4317 : MVE_VLDRSTR_base<dir, U, P, W, 0, oops, iops, asm, suffix, ops, cstr> {
4319 let Inst{23} = addr{7};
4320 let Inst{19} = memsz.encoding{0}; // enough to tell 16- from 32-bit
4321 let Inst{18-16} = addr{10-8};
4322 let Inst{8-7} = size;
4323 let Inst{6-0} = addr{6-0};
4328 // Multiclass wrapper on each of the _cw and _cs base classes, to
4329 // generate three writeback modes (none, preindex, postindex).
4331 multiclass MVE_VLDRSTR_cw_m<MVE_ldst_direction dir, MVE_memsz memsz,
4332 string asm, string suffix, bit U, bits<2> size> {
4333 let AM = memsz.AM in {
4334 def "" : MVE_VLDRSTR_cw<
4335 dir, memsz, U, 1, 0, size,
4336 dir.Oops, !con(dir.Iops, (ins taddrmode_imm7<memsz.shift>:$addr)),
4337 asm, suffix, IndexModeNone, "$Qd, $addr", "">;
4339 def _pre : MVE_VLDRSTR_cw<
4340 dir, memsz, U, 1, 1, size,
4341 !con((outs tGPR:$wb), dir.Oops),
4342 !con(dir.Iops, (ins taddrmode_imm7<memsz.shift>:$addr)),
4343 asm, suffix, IndexModePre, "$Qd, $addr!", "$addr.base = $wb"> {
4344 let DecoderMethod = "DecodeMVE_MEM_1_pre<"#memsz.shift#">";
4347 def _post : MVE_VLDRSTR_cw<
4348 dir, memsz, U, 0, 1, size,
4349 !con((outs tGPR:$wb), dir.Oops),
4350 !con(dir.Iops, (ins t_addr_offset_none:$Rn,
4351 t2am_imm7_offset<memsz.shift>:$addr)),
4352 asm, suffix, IndexModePost, "$Qd, $Rn$addr", "$Rn.base = $wb"> {
4354 let Inst{18-16} = Rn{2-0};
4359 multiclass MVE_VLDRSTR_cs_m<MVE_ldst_direction dir, MVE_memsz memsz,
4360 string asm, string suffix> {
4361 let AM = memsz.AM in {
4362 def "" : MVE_VLDRSTR_cs<
4364 dir.Oops, !con(dir.Iops, (ins t2addrmode_imm7<memsz.shift>:$addr)),
4365 asm, suffix, IndexModeNone, "$Qd, $addr", "">;
4367 def _pre : MVE_VLDRSTR_cs<
4369 !con((outs rGPR:$wb), dir.Oops),
4370 !con(dir.Iops, (ins t2addrmode_imm7_pre<memsz.shift>:$addr)),
4371 asm, suffix, IndexModePre, "$Qd, $addr!", "$addr.base = $wb"> {
4372 let DecoderMethod = "DecodeMVE_MEM_2_pre<"#memsz.shift#">";
4375 def _post : MVE_VLDRSTR_cs<
4377 !con((outs rGPR:$wb), dir.Oops),
4378 // We need an !if here to select the base register class,
4379 // because it's legal to write back to SP in a load of this
4380 // type, but not in a store.
4381 !con(dir.Iops, (ins !if(dir.load, t2_addr_offset_none,
4382 t2_nosp_addr_offset_none):$Rn,
4383 t2am_imm7_offset<memsz.shift>:$addr)),
4384 asm, suffix, IndexModePost, "$Qd, $Rn$addr", "$Rn.base = $wb"> {
4386 let Inst{19-16} = Rn{3-0};
4391 // Now actually declare all the contiguous load/stores, via those
4392 // multiclasses. The instruction ids coming out of this are the bare
4393 // names shown in the defm, with _pre or _post appended for writeback,
4394 // e.g. MVE_VLDRBS16, MVE_VSTRB16_pre, MVE_VSTRHU16_post.
4396 defm MVE_VLDRBS16: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memB, "vldrb", "s16", 0, 0b01>;
4397 defm MVE_VLDRBS32: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memB, "vldrb", "s32", 0, 0b10>;
4398 defm MVE_VLDRBU16: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memB, "vldrb", "u16", 1, 0b01>;
4399 defm MVE_VLDRBU32: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memB, "vldrb", "u32", 1, 0b10>;
4400 defm MVE_VLDRHS32: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memH, "vldrh", "s32", 0, 0b10>;
4401 defm MVE_VLDRHU32: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memH, "vldrh", "u32", 1, 0b10>;
4403 defm MVE_VLDRBU8: MVE_VLDRSTR_cs_m<MVE_ld, MVE_memB, "vldrb", "u8">;
4404 defm MVE_VLDRHU16: MVE_VLDRSTR_cs_m<MVE_ld, MVE_memH, "vldrh", "u16">;
4405 defm MVE_VLDRWU32: MVE_VLDRSTR_cs_m<MVE_ld, MVE_memW, "vldrw", "u32">;
4407 defm MVE_VSTRB16: MVE_VLDRSTR_cw_m<MVE_st, MVE_memB, "vstrb", "16", 0, 0b01>;
4408 defm MVE_VSTRB32: MVE_VLDRSTR_cw_m<MVE_st, MVE_memB, "vstrb", "32", 0, 0b10>;
4409 defm MVE_VSTRH32: MVE_VLDRSTR_cw_m<MVE_st, MVE_memH, "vstrh", "32", 0, 0b10>;
4411 defm MVE_VSTRBU8 : MVE_VLDRSTR_cs_m<MVE_st, MVE_memB, "vstrb", "8">;
4412 defm MVE_VSTRHU16: MVE_VLDRSTR_cs_m<MVE_st, MVE_memH, "vstrh", "16">;
4413 defm MVE_VSTRWU32: MVE_VLDRSTR_cs_m<MVE_st, MVE_memW, "vstrw", "32">;
4415 // Gather loads / scatter stores whose address operand is of the form
4416 // [Rn,Qm], i.e. a single GPR as the common base address, plus a
4417 // vector of offset from it. ('Load/store this sequence of elements of
4418 // the same array.')
4420 // Like the contiguous family, these loads and stores can widen the
4421 // loaded values / truncate the stored ones, or they can just
4422 // load/store the same size of memory and vector lane. But unlike the
4423 // contiguous family, there's no particular difference in encoding
4424 // between those two cases.
4426 // This family also comes with the option to scale the offset values
4427 // in Qm by the size of the loaded memory (i.e. to treat them as array
4428 // indices), or not to scale them (to treat them as plain byte offsets
4429 // in memory, so that perhaps the loaded values are unaligned). The
4430 // scaled instructions' address operand in assembly looks like
4431 // [Rn,Qm,UXTW #2] or similar.
4434 class MVE_VLDRSTR_rq<MVE_ldst_direction dir, MVE_memsz memsz, bit U,
4435 bits<2> size, bit os, string asm, string suffix, int shift>
4436 : MVE_VLDRSTR_base<dir, U, 0b0, 0b0, 0, dir.Oops,
4437 !con(dir.Iops, (ins mve_addr_rq_shift<shift>:$addr)),
4438 asm, suffix, "$Qd, $addr", dir.cstr> {
4441 let Inst{19-16} = addr{6-3};
4442 let Inst{8-7} = size;
4443 let Inst{6} = memsz.encoding{1};
4445 let Inst{4} = memsz.encoding{0};
4446 let Inst{3-1} = addr{2-0};
4450 // Multiclass that defines the scaled and unscaled versions of an
4451 // instruction, when the memory size is wider than a byte. The scaled
4452 // version gets the default name like MVE_VLDRBU16_rq; the unscaled /
4453 // potentially unaligned version gets a "_u" suffix, e.g.
4454 // MVE_VLDRBU16_rq_u.
4455 multiclass MVE_VLDRSTR_rq_w<MVE_ldst_direction dir, MVE_memsz memsz,
4456 string asm, string suffix, bit U, bits<2> size> {
4457 def _u : MVE_VLDRSTR_rq<dir, memsz, U, size, 0, asm, suffix, 0>;
4458 def "" : MVE_VLDRSTR_rq<dir, memsz, U, size, 1, asm, suffix, memsz.shift>;
4461 // Subclass of MVE_VLDRSTR_rq with the same API as that multiclass,
4462 // for use when the memory size is one byte, so there's no 'scaled'
4463 // version of the instruction at all. (This is encoded as if it were
4464 // unscaled, but named in the default way with no _u suffix.)
4465 class MVE_VLDRSTR_rq_b<MVE_ldst_direction dir, MVE_memsz memsz,
4466 string asm, string suffix, bit U, bits<2> size>
4467 : MVE_VLDRSTR_rq<dir, memsz, U, size, 0, asm, suffix, 0>;
4469 // Actually define all the loads and stores in this family.
4471 def MVE_VLDRBU8_rq : MVE_VLDRSTR_rq_b<MVE_ld, MVE_memB, "vldrb","u8", 1,0b00>;
4472 def MVE_VLDRBU16_rq: MVE_VLDRSTR_rq_b<MVE_ld, MVE_memB, "vldrb","u16", 1,0b01>;
4473 def MVE_VLDRBS16_rq: MVE_VLDRSTR_rq_b<MVE_ld, MVE_memB, "vldrb","s16", 0,0b01>;
4474 def MVE_VLDRBU32_rq: MVE_VLDRSTR_rq_b<MVE_ld, MVE_memB, "vldrb","u32", 1,0b10>;
4475 def MVE_VLDRBS32_rq: MVE_VLDRSTR_rq_b<MVE_ld, MVE_memB, "vldrb","s32", 0,0b10>;
4477 defm MVE_VLDRHU16_rq: MVE_VLDRSTR_rq_w<MVE_ld, MVE_memH, "vldrh","u16", 1,0b01>;
4478 defm MVE_VLDRHU32_rq: MVE_VLDRSTR_rq_w<MVE_ld, MVE_memH, "vldrh","u32", 1,0b10>;
4479 defm MVE_VLDRHS32_rq: MVE_VLDRSTR_rq_w<MVE_ld, MVE_memH, "vldrh","s32", 0,0b10>;
4480 defm MVE_VLDRWU32_rq: MVE_VLDRSTR_rq_w<MVE_ld, MVE_memW, "vldrw","u32", 1,0b10>;
4481 defm MVE_VLDRDU64_rq: MVE_VLDRSTR_rq_w<MVE_ld, MVE_memD, "vldrd","u64", 1,0b11>;
4483 def MVE_VSTRB8_rq : MVE_VLDRSTR_rq_b<MVE_st, MVE_memB, "vstrb","8", 0,0b00>;
4484 def MVE_VSTRB16_rq : MVE_VLDRSTR_rq_b<MVE_st, MVE_memB, "vstrb","16", 0,0b01>;
4485 def MVE_VSTRB32_rq : MVE_VLDRSTR_rq_b<MVE_st, MVE_memB, "vstrb","32", 0,0b10>;
4487 defm MVE_VSTRH16_rq : MVE_VLDRSTR_rq_w<MVE_st, MVE_memH, "vstrh","16", 0,0b01>;
4488 defm MVE_VSTRH32_rq : MVE_VLDRSTR_rq_w<MVE_st, MVE_memH, "vstrh","32", 0,0b10>;
4489 defm MVE_VSTRW32_rq : MVE_VLDRSTR_rq_w<MVE_st, MVE_memW, "vstrw","32", 0,0b10>;
4490 defm MVE_VSTRD64_rq : MVE_VLDRSTR_rq_w<MVE_st, MVE_memD, "vstrd","64", 0,0b11>;
4492 // Gather loads / scatter stores whose address operand is of the form
4493 // [Qm,#imm], i.e. a vector containing a full base address for each
4494 // loaded item, plus an immediate offset applied consistently to all
4495 // of them. ('Load/store the same field from this vector of pointers
4496 // to a structure type.')
4498 // This family requires the vector lane size to be at least 32 bits
4499 // (so there's room for an address in each lane at all). It has no
4500 // widening/narrowing variants. But it does support preindex
4501 // writeback, in which the address vector is updated to hold the
4502 // addresses actually loaded from.
4505 class MVE_VLDRSTR_qi<MVE_ldst_direction dir, MVE_memsz memsz, bit W, dag wbops,
4506 string asm, string wbAsm, string suffix, string cstr = "">
4507 : MVE_VLDRSTR_base<dir, 1, 1, W, 1, !con(wbops, dir.Oops),
4508 !con(dir.Iops, (ins mve_addr_q_shift<memsz.shift>:$addr)),
4509 asm, suffix, "$Qd, $addr" # wbAsm, cstr # dir.cstr> {
4511 let Inst{23} = addr{7};
4512 let Inst{19-17} = addr{10-8};
4514 let Inst{8} = memsz.encoding{0}; // enough to distinguish 32- from 64-bit
4516 let Inst{6-0} = addr{6-0};
4519 // Multiclass that generates the non-writeback and writeback variants.
4520 multiclass MVE_VLDRSTR_qi_m<MVE_ldst_direction dir, MVE_memsz memsz,
4521 string asm, string suffix> {
4522 def "" : MVE_VLDRSTR_qi<dir, memsz, 0, (outs), asm, "", suffix>;
4523 def _pre : MVE_VLDRSTR_qi<dir, memsz, 1, (outs MQPR:$wb), asm, "!", suffix,
4524 "$addr.base = $wb"> {
4525 let DecoderMethod="DecodeMVE_MEM_3_pre<"#memsz.shift#">";
4529 // Actual instruction definitions.
4530 defm MVE_VLDRWU32_qi: MVE_VLDRSTR_qi_m<MVE_ld, MVE_memW, "vldrw", "u32">;
4531 defm MVE_VLDRDU64_qi: MVE_VLDRSTR_qi_m<MVE_ld, MVE_memD, "vldrd", "u64">;
4532 defm MVE_VSTRW32_qi: MVE_VLDRSTR_qi_m<MVE_st, MVE_memW, "vstrw", "32">;
4533 defm MVE_VSTRD64_qi: MVE_VLDRSTR_qi_m<MVE_st, MVE_memD, "vstrd", "64">;
4535 // Define aliases for all the instructions where memory size and
4536 // vector lane size are the same. These are mnemonic aliases, so they
4537 // apply consistently across all of the above families - contiguous
4538 // loads, and both the rq and qi types of gather/scatter.
4540 // Rationale: As long as you're loading (for example) 16-bit memory
4541 // values into 16-bit vector lanes, you can think of them as signed or
4542 // unsigned integers, fp16 or just raw 16-bit blobs and it makes no
4543 // difference. So we permit all of vldrh.16, vldrh.u16, vldrh.s16,
4544 // vldrh.f16 and treat them all as equivalent to the canonical
4545 // spelling (which happens to be .u16 for loads, and just .16 for
4548 foreach vpt_cond = ["", "t", "e"] in
4549 foreach memsz = [MVE_memB, MVE_memH, MVE_memW, MVE_memD] in
4550 foreach suffix = memsz.suffixes in {
4552 // These foreaches are conceptually ifs, implemented by iterating a
4553 // dummy variable over a list with 0 or 1 elements depending on the
4554 // condition. The idea is to iterate over _nearly_ all the suffixes
4555 // in memsz.suffixes, but omit the one we want all the others to alias.
4557 foreach _ = !if(!ne(suffix, memsz.CanonLoadSuffix), [1], []<int>) in
4558 def : MnemonicAlias<
4559 "vldr" # memsz.MnemonicLetter # vpt_cond # suffix,
4560 "vldr" # memsz.MnemonicLetter # vpt_cond # memsz.CanonLoadSuffix>;
4562 foreach _ = !if(!ne(suffix, memsz.CanonStoreSuffix), [1], []<int>) in
4563 def : MnemonicAlias<
4564 "vstr" # memsz.MnemonicLetter # vpt_cond # suffix,
4565 "vstr" # memsz.MnemonicLetter # vpt_cond # memsz.CanonStoreSuffix>;
4568 // end of MVE predicable load/store
4570 class MVE_VPT<string suffix, bits<2> size, dag iops, string asm, list<dag> pattern=[]>
4571 : MVE_MI<(outs ), iops, NoItinerary, !strconcat("vpt", "${Mk}", ".", suffix), asm, "", pattern> {
4576 let Inst{31-23} = 0b111111100;
4577 let Inst{22} = Mk{3};
4578 let Inst{21-20} = size;
4579 let Inst{19-17} = Qn{2-0};
4581 let Inst{15-13} = Mk{2-0};
4582 let Inst{12} = fc{2};
4583 let Inst{11-8} = 0b1111;
4584 let Inst{7} = fc{0};
4590 class MVE_VPTt1<string suffix, bits<2> size, dag iops>
4591 : MVE_VPT<suffix, size, iops, "$fc, $Qn, $Qm"> {
4596 let Inst{5} = Qm{3};
4597 let Inst{3-1} = Qm{2-0};
4598 let Inst{0} = fc{1};
4601 class MVE_VPTt1i<string suffix, bits<2> size>
4602 : MVE_VPTt1<suffix, size,
4603 (ins vpt_mask:$Mk, MQPR:$Qn, MQPR:$Qm, pred_basic_i:$fc)> {
4608 def MVE_VPTv4i32 : MVE_VPTt1i<"i32", 0b10>;
4609 def MVE_VPTv8i16 : MVE_VPTt1i<"i16", 0b01>;
4610 def MVE_VPTv16i8 : MVE_VPTt1i<"i8", 0b00>;
4612 class MVE_VPTt1u<string suffix, bits<2> size>
4613 : MVE_VPTt1<suffix, size,
4614 (ins vpt_mask:$Mk, MQPR:$Qn, MQPR:$Qm, pred_basic_u:$fc)> {
4619 def MVE_VPTv4u32 : MVE_VPTt1u<"u32", 0b10>;
4620 def MVE_VPTv8u16 : MVE_VPTt1u<"u16", 0b01>;
4621 def MVE_VPTv16u8 : MVE_VPTt1u<"u8", 0b00>;
4623 class MVE_VPTt1s<string suffix, bits<2> size>
4624 : MVE_VPTt1<suffix, size,
4625 (ins vpt_mask:$Mk, MQPR:$Qn, MQPR:$Qm, pred_basic_s:$fc)> {
4629 def MVE_VPTv4s32 : MVE_VPTt1s<"s32", 0b10>;
4630 def MVE_VPTv8s16 : MVE_VPTt1s<"s16", 0b01>;
4631 def MVE_VPTv16s8 : MVE_VPTt1s<"s8", 0b00>;
4633 class MVE_VPTt2<string suffix, bits<2> size, dag iops>
4634 : MVE_VPT<suffix, size, iops,
4641 let Inst{5} = fc{1};
4642 let Inst{3-0} = Rm{3-0};
4645 class MVE_VPTt2i<string suffix, bits<2> size>
4646 : MVE_VPTt2<suffix, size,
4647 (ins vpt_mask:$Mk, MQPR:$Qn, GPRwithZR:$Rm, pred_basic_i:$fc)> {
4652 def MVE_VPTv4i32r : MVE_VPTt2i<"i32", 0b10>;
4653 def MVE_VPTv8i16r : MVE_VPTt2i<"i16", 0b01>;
4654 def MVE_VPTv16i8r : MVE_VPTt2i<"i8", 0b00>;
4656 class MVE_VPTt2u<string suffix, bits<2> size>
4657 : MVE_VPTt2<suffix, size,
4658 (ins vpt_mask:$Mk, MQPR:$Qn, GPRwithZR:$Rm, pred_basic_u:$fc)> {
4663 def MVE_VPTv4u32r : MVE_VPTt2u<"u32", 0b10>;
4664 def MVE_VPTv8u16r : MVE_VPTt2u<"u16", 0b01>;
4665 def MVE_VPTv16u8r : MVE_VPTt2u<"u8", 0b00>;
4667 class MVE_VPTt2s<string suffix, bits<2> size>
4668 : MVE_VPTt2<suffix, size,
4669 (ins vpt_mask:$Mk, MQPR:$Qn, GPRwithZR:$Rm, pred_basic_s:$fc)> {
4673 def MVE_VPTv4s32r : MVE_VPTt2s<"s32", 0b10>;
4674 def MVE_VPTv8s16r : MVE_VPTt2s<"s16", 0b01>;
4675 def MVE_VPTv16s8r : MVE_VPTt2s<"s8", 0b00>;
4678 class MVE_VPTf<string suffix, bit size, dag iops, string asm, list<dag> pattern=[]>
4679 : MVE_MI<(outs ), iops, NoItinerary, !strconcat("vpt", "${Mk}", ".", suffix), asm,
4685 let Inst{31-29} = 0b111;
4686 let Inst{28} = size;
4687 let Inst{27-23} = 0b11100;
4688 let Inst{22} = Mk{3};
4689 let Inst{21-20} = 0b11;
4690 let Inst{19-17} = Qn{2-0};
4692 let Inst{15-13} = Mk{2-0};
4693 let Inst{12} = fc{2};
4694 let Inst{11-8} = 0b1111;
4695 let Inst{7} = fc{0};
4699 let Predicates = [HasMVEFloat];
4702 class MVE_VPTft1<string suffix, bit size>
4703 : MVE_VPTf<suffix, size, (ins vpt_mask:$Mk, MQPR:$Qn, MQPR:$Qm, pred_basic_fp:$fc),
4709 let Inst{5} = Qm{3};
4710 let Inst{3-1} = Qm{2-0};
4711 let Inst{0} = fc{1};
4714 def MVE_VPTv4f32 : MVE_VPTft1<"f32", 0b0>;
4715 def MVE_VPTv8f16 : MVE_VPTft1<"f16", 0b1>;
4717 class MVE_VPTft2<string suffix, bit size>
4718 : MVE_VPTf<suffix, size, (ins vpt_mask:$Mk, MQPR:$Qn, GPRwithZR:$Rm, pred_basic_fp:$fc),
4724 let Inst{5} = fc{1};
4725 let Inst{3-0} = Rm{3-0};
4728 def MVE_VPTv4f32r : MVE_VPTft2<"f32", 0b0>;
4729 def MVE_VPTv8f16r : MVE_VPTft2<"f16", 0b1>;
4731 def MVE_VPST : MVE_MI<(outs ), (ins vpt_mask:$Mk), NoItinerary,
4732 !strconcat("vpst", "${Mk}"), "", "", []> {
4735 let Inst{31-23} = 0b111111100;
4736 let Inst{22} = Mk{3};
4737 let Inst{21-16} = 0b110001;
4738 let Inst{15-13} = Mk{2-0};
4739 let Inst{12-0} = 0b0111101001101;
4740 let Unpredictable{12} = 0b1;
4741 let Unpredictable{7} = 0b1;
4742 let Unpredictable{5} = 0b1;
4747 def MVE_VPSEL : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm), NoItinerary,
4748 "vpsel", "", "$Qd, $Qn, $Qm", vpred_n, "", []> {
4754 let Inst{25-23} = 0b100;
4755 let Inst{22} = Qd{3};
4756 let Inst{21-20} = 0b11;
4757 let Inst{19-17} = Qn{2-0};
4759 let Inst{15-13} = Qd{2-0};
4760 let Inst{12-9} = 0b0111;
4762 let Inst{7} = Qn{3};
4764 let Inst{5} = Qm{3};
4766 let Inst{3-1} = Qm{2-0};
4770 foreach suffix = ["s8", "s16", "s32", "u8", "u16", "u32",
4771 "i8", "i16", "i32", "f16", "f32"] in
4772 def : MVEInstAlias<"vpsel${vp}." # suffix # "\t$Qd, $Qn, $Qm",
4773 (MVE_VPSEL MQPR:$Qd, MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
4775 let Predicates = [HasMVEInt] in {
4776 def : Pat<(v16i8 (vselect (v16i1 VCCR:$pred), (v16i8 MQPR:$v1), (v16i8 MQPR:$v2))),
4777 (v16i8 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0, VCCR:$pred))>;
4778 def : Pat<(v8i16 (vselect (v8i1 VCCR:$pred), (v8i16 MQPR:$v1), (v8i16 MQPR:$v2))),
4779 (v8i16 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0, VCCR:$pred))>;
4780 def : Pat<(v4i32 (vselect (v4i1 VCCR:$pred), (v4i32 MQPR:$v1), (v4i32 MQPR:$v2))),
4781 (v4i32 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0, VCCR:$pred))>;
4783 def : Pat<(v8f16 (vselect (v8i1 VCCR:$pred), (v8f16 MQPR:$v1), (v8f16 MQPR:$v2))),
4784 (v8f16 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0, VCCR:$pred))>;
4785 def : Pat<(v4f32 (vselect (v4i1 VCCR:$pred), (v4f32 MQPR:$v1), (v4f32 MQPR:$v2))),
4786 (v4f32 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0, VCCR:$pred))>;
4788 def : Pat<(v16i8 (vselect (v16i8 MQPR:$pred), (v16i8 MQPR:$v1), (v16i8 MQPR:$v2))),
4789 (v16i8 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0,
4790 (MVE_VCMPi8 (v16i8 MQPR:$pred), (MVE_VMOVimmi8 0), 1)))>;
4791 def : Pat<(v8i16 (vselect (v8i16 MQPR:$pred), (v8i16 MQPR:$v1), (v8i16 MQPR:$v2))),
4792 (v8i16 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0,
4793 (MVE_VCMPi16 (v8i16 MQPR:$pred), (MVE_VMOVimmi16 0), 1)))>;
4794 def : Pat<(v4i32 (vselect (v4i32 MQPR:$pred), (v4i32 MQPR:$v1), (v4i32 MQPR:$v2))),
4795 (v4i32 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0,
4796 (MVE_VCMPi32 (v4i32 MQPR:$pred), (MVE_VMOVimmi32 0), 1)))>;
4798 def : Pat<(v8f16 (vselect (v8i16 MQPR:$pred), (v8f16 MQPR:$v1), (v8f16 MQPR:$v2))),
4799 (v8f16 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0,
4800 (MVE_VCMPi16 (v8i16 MQPR:$pred), (MVE_VMOVimmi16 0), 1)))>;
4801 def : Pat<(v4f32 (vselect (v4i32 MQPR:$pred), (v4f32 MQPR:$v1), (v4f32 MQPR:$v2))),
4802 (v4f32 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0,
4803 (MVE_VCMPi32 (v4i32 MQPR:$pred), (MVE_VMOVimmi32 0), 1)))>;
4806 def : Pat<(v16i8 (zext (v16i1 VCCR:$pred))),
4807 (v16i8 (MVE_VPSEL (MVE_VMOVimmi8 1), (MVE_VMOVimmi8 0), 0, VCCR:$pred))>;
4808 def : Pat<(v8i16 (zext (v8i1 VCCR:$pred))),
4809 (v8i16 (MVE_VPSEL (MVE_VMOVimmi16 1), (MVE_VMOVimmi16 0), 0, VCCR:$pred))>;
4810 def : Pat<(v4i32 (zext (v4i1 VCCR:$pred))),
4811 (v4i32 (MVE_VPSEL (MVE_VMOVimmi32 1), (MVE_VMOVimmi32 0), 0, VCCR:$pred))>;
4813 def : Pat<(v16i8 (sext (v16i1 VCCR:$pred))),
4814 (v16i8 (MVE_VPSEL (MVE_VMOVimmi8 255), (MVE_VMOVimmi8 0), 0, VCCR:$pred))>;
4815 def : Pat<(v8i16 (sext (v8i1 VCCR:$pred))),
4816 (v8i16 (MVE_VPSEL (MVE_VMOVimmi8 255), (MVE_VMOVimmi16 0), 0, VCCR:$pred))>;
4817 def : Pat<(v4i32 (sext (v4i1 VCCR:$pred))),
4818 (v4i32 (MVE_VPSEL (MVE_VMOVimmi8 255), (MVE_VMOVimmi32 0), 0, VCCR:$pred))>;
4820 def : Pat<(v16i8 (anyext (v16i1 VCCR:$pred))),
4821 (v16i8 (MVE_VPSEL (MVE_VMOVimmi8 1), (MVE_VMOVimmi8 0), 0, VCCR:$pred))>;
4822 def : Pat<(v8i16 (anyext (v8i1 VCCR:$pred))),
4823 (v8i16 (MVE_VPSEL (MVE_VMOVimmi16 1), (MVE_VMOVimmi16 0), 0, VCCR:$pred))>;
4824 def : Pat<(v4i32 (anyext (v4i1 VCCR:$pred))),
4825 (v4i32 (MVE_VPSEL (MVE_VMOVimmi32 1), (MVE_VMOVimmi32 0), 0, VCCR:$pred))>;
4827 def : Pat<(v16i1 (trunc (v16i8 MQPR:$v1))),
4828 (v16i1 (MVE_VCMPi32r (v16i8 MQPR:$v1), ZR, 1))>;
4829 def : Pat<(v8i1 (trunc (v8i16 MQPR:$v1))),
4830 (v8i1 (MVE_VCMPi32r (v8i16 MQPR:$v1), ZR, 1))>;
4831 def : Pat<(v4i1 (trunc (v4i32 MQPR:$v1))),
4832 (v4i1 (MVE_VCMPi32r (v4i32 MQPR:$v1), ZR, 1))>;
4835 let Predicates = [HasMVEFloat] in {
4837 // 112 is 1.0 in float
4838 def : Pat<(v4f32 (uint_to_fp (v4i1 VCCR:$pred))),
4839 (v4f32 (MVE_VPSEL (v4f32 (MVE_VMOVimmf32 112)), (v4f32 (MVE_VMOVimmi32 0)), 0, VCCR:$pred))>;
4840 // 2620 in 1.0 in half
4841 def : Pat<(v8f16 (uint_to_fp (v8i1 VCCR:$pred))),
4842 (v8f16 (MVE_VPSEL (v8f16 (MVE_VMOVimmi16 2620)), (v8f16 (MVE_VMOVimmi16 0)), 0, VCCR:$pred))>;
4843 // 240 is -1.0 in float
4844 def : Pat<(v4f32 (sint_to_fp (v4i1 VCCR:$pred))),
4845 (v4f32 (MVE_VPSEL (v4f32 (MVE_VMOVimmf32 240)), (v4f32 (MVE_VMOVimmi32 0)), 0, VCCR:$pred))>;
4846 // 2748 is -1.0 in half
4847 def : Pat<(v8f16 (sint_to_fp (v8i1 VCCR:$pred))),
4848 (v8f16 (MVE_VPSEL (v8f16 (MVE_VMOVimmi16 2748)), (v8f16 (MVE_VMOVimmi16 0)), 0, VCCR:$pred))>;
4850 def : Pat<(v4i1 (fp_to_uint (v4f32 MQPR:$v1))),
4851 (v4i1 (MVE_VCMPf32r (v4f32 MQPR:$v1), ZR, 1))>;
4852 def : Pat<(v8i1 (fp_to_uint (v8f16 MQPR:$v1))),
4853 (v8i1 (MVE_VCMPf16r (v8f16 MQPR:$v1), ZR, 1))>;
4854 def : Pat<(v4i1 (fp_to_sint (v4f32 MQPR:$v1))),
4855 (v4i1 (MVE_VCMPf32r (v4f32 MQPR:$v1), ZR, 1))>;
4856 def : Pat<(v8i1 (fp_to_sint (v8f16 MQPR:$v1))),
4857 (v8i1 (MVE_VCMPf16r (v8f16 MQPR:$v1), ZR, 1))>;
4860 def MVE_VPNOT : MVE_p<(outs VCCR:$P0), (ins VCCR:$P0_in), NoItinerary,
4861 "vpnot", "", "", vpred_n, "", []> {
4862 let Inst{31-0} = 0b11111110001100010000111101001101;
4863 let Unpredictable{19-17} = 0b111;
4864 let Unpredictable{12} = 0b1;
4865 let Unpredictable{7} = 0b1;
4866 let Unpredictable{5} = 0b1;
4868 let Constraints = "";
4869 let DecoderMethod = "DecodeMVEVPNOT";
4872 let Predicates = [HasMVEInt] in {
4873 def : Pat<(v4i1 (xor (v4i1 VCCR:$pred), (v4i1 (predicate_cast (i32 65535))))),
4874 (v4i1 (MVE_VPNOT (v4i1 VCCR:$pred)))>;
4875 def : Pat<(v8i1 (xor (v8i1 VCCR:$pred), (v8i1 (predicate_cast (i32 65535))))),
4876 (v8i1 (MVE_VPNOT (v8i1 VCCR:$pred)))>;
4877 def : Pat<(v16i1 (xor (v16i1 VCCR:$pred), (v16i1 (predicate_cast (i32 65535))))),
4878 (v16i1 (MVE_VPNOT (v16i1 VCCR:$pred)))>;
4882 class MVE_loltp_start<dag iops, string asm, string ops, bits<2> size>
4883 : t2LOL<(outs GPRlr:$LR), iops, asm, ops> {
4885 let Predicates = [HasMVEInt];
4887 let Inst{21-20} = size;
4888 let Inst{19-16} = Rn{3-0};
4892 class MVE_DLSTP<string asm, bits<2> size>
4893 : MVE_loltp_start<(ins rGPR:$Rn), asm, "$LR, $Rn", size> {
4895 let Inst{11-1} = 0b00000000000;
4896 let Unpredictable{10-1} = 0b1111111111;
4899 class MVE_WLSTP<string asm, bits<2> size>
4900 : MVE_loltp_start<(ins rGPR:$Rn, wlslabel_u11:$label),
4901 asm, "$LR, $Rn, $label", size> {
4904 let Inst{11} = label{0};
4905 let Inst{10-1} = label{10-1};
4908 def MVE_DLSTP_8 : MVE_DLSTP<"dlstp.8", 0b00>;
4909 def MVE_DLSTP_16 : MVE_DLSTP<"dlstp.16", 0b01>;
4910 def MVE_DLSTP_32 : MVE_DLSTP<"dlstp.32", 0b10>;
4911 def MVE_DLSTP_64 : MVE_DLSTP<"dlstp.64", 0b11>;
4913 def MVE_WLSTP_8 : MVE_WLSTP<"wlstp.8", 0b00>;
4914 def MVE_WLSTP_16 : MVE_WLSTP<"wlstp.16", 0b01>;
4915 def MVE_WLSTP_32 : MVE_WLSTP<"wlstp.32", 0b10>;
4916 def MVE_WLSTP_64 : MVE_WLSTP<"wlstp.64", 0b11>;
4918 class MVE_loltp_end<dag oops, dag iops, string asm, string ops>
4919 : t2LOL<oops, iops, asm, ops> {
4920 let Predicates = [HasMVEInt];
4921 let Inst{22-21} = 0b00;
4922 let Inst{19-16} = 0b1111;
4926 def MVE_LETP : MVE_loltp_end<(outs GPRlr:$LRout),
4927 (ins GPRlr:$LRin, lelabel_u11:$label),
4928 "letp", "$LRin, $label"> {
4932 let Inst{11} = label{0};
4933 let Inst{10-1} = label{10-1};
4936 def MVE_LCTP : MVE_loltp_end<(outs), (ins pred:$p), "lctp${p}", ""> {
4939 let Inst{11-1} = 0b00000000000;
4940 let Unpredictable{21-20} = 0b11;
4941 let Unpredictable{11-1} = 0b11111111111;
4945 //===----------------------------------------------------------------------===//
4947 //===----------------------------------------------------------------------===//
4949 class MVE_vector_store_typed<ValueType Ty, Instruction RegImmInst,
4950 PatFrag StoreKind, int shift>
4951 : Pat<(StoreKind (Ty MQPR:$val), t2addrmode_imm7<shift>:$addr),
4952 (RegImmInst (Ty MQPR:$val), t2addrmode_imm7<shift>:$addr)>;
4953 class MVE_vector_maskedstore_typed<ValueType Ty, Instruction RegImmInst,
4954 PatFrag StoreKind, int shift>
4955 : Pat<(StoreKind (Ty MQPR:$val), t2addrmode_imm7<shift>:$addr, VCCR:$pred),
4956 (RegImmInst (Ty MQPR:$val), t2addrmode_imm7<shift>:$addr, (i32 1), VCCR:$pred)>;
4958 multiclass MVE_vector_store<Instruction RegImmInst, PatFrag StoreKind,
4960 def : MVE_vector_store_typed<v16i8, RegImmInst, StoreKind, shift>;
4961 def : MVE_vector_store_typed<v8i16, RegImmInst, StoreKind, shift>;
4962 def : MVE_vector_store_typed<v8f16, RegImmInst, StoreKind, shift>;
4963 def : MVE_vector_store_typed<v4i32, RegImmInst, StoreKind, shift>;
4964 def : MVE_vector_store_typed<v4f32, RegImmInst, StoreKind, shift>;
4965 def : MVE_vector_store_typed<v2i64, RegImmInst, StoreKind, shift>;
4966 def : MVE_vector_store_typed<v2f64, RegImmInst, StoreKind, shift>;
4969 class MVE_vector_load_typed<ValueType Ty, Instruction RegImmInst,
4970 PatFrag LoadKind, int shift>
4971 : Pat<(Ty (LoadKind t2addrmode_imm7<shift>:$addr)),
4972 (Ty (RegImmInst t2addrmode_imm7<shift>:$addr))>;
4973 class MVE_vector_maskedload_typed<ValueType Ty, Instruction RegImmInst,
4974 PatFrag LoadKind, int shift>
4975 : Pat<(Ty (LoadKind t2addrmode_imm7<shift>:$addr, VCCR:$pred, (Ty NEONimmAllZerosV))),
4976 (Ty (RegImmInst t2addrmode_imm7<shift>:$addr, (i32 1), VCCR:$pred))>;
4978 multiclass MVE_vector_load<Instruction RegImmInst, PatFrag LoadKind,
4980 def : MVE_vector_load_typed<v16i8, RegImmInst, LoadKind, shift>;
4981 def : MVE_vector_load_typed<v8i16, RegImmInst, LoadKind, shift>;
4982 def : MVE_vector_load_typed<v8f16, RegImmInst, LoadKind, shift>;
4983 def : MVE_vector_load_typed<v4i32, RegImmInst, LoadKind, shift>;
4984 def : MVE_vector_load_typed<v4f32, RegImmInst, LoadKind, shift>;
4985 def : MVE_vector_load_typed<v2i64, RegImmInst, LoadKind, shift>;
4986 def : MVE_vector_load_typed<v2f64, RegImmInst, LoadKind, shift>;
4989 class MVE_vector_offset_store_typed<ValueType Ty, Instruction Opcode,
4990 PatFrag StoreKind, int shift>
4991 : Pat<(StoreKind (Ty MQPR:$Rt), tGPR:$Rn, t2am_imm7_offset<shift>:$addr),
4992 (Opcode MQPR:$Rt, tGPR:$Rn, t2am_imm7_offset<shift>:$addr)>;
4994 multiclass MVE_vector_offset_store<Instruction RegImmInst, PatFrag StoreKind,
4996 def : MVE_vector_offset_store_typed<v16i8, RegImmInst, StoreKind, shift>;
4997 def : MVE_vector_offset_store_typed<v8i16, RegImmInst, StoreKind, shift>;
4998 def : MVE_vector_offset_store_typed<v8f16, RegImmInst, StoreKind, shift>;
4999 def : MVE_vector_offset_store_typed<v4i32, RegImmInst, StoreKind, shift>;
5000 def : MVE_vector_offset_store_typed<v4f32, RegImmInst, StoreKind, shift>;
5001 def : MVE_vector_offset_store_typed<v2i64, RegImmInst, StoreKind, shift>;
5002 def : MVE_vector_offset_store_typed<v2f64, RegImmInst, StoreKind, shift>;
5005 def aligned32_pre_store : PatFrag<(ops node:$val, node:$ptr, node:$offset),
5006 (pre_store node:$val, node:$ptr, node:$offset), [{
5007 return cast<StoreSDNode>(N)->getAlignment() >= 4;
5009 def aligned32_post_store : PatFrag<(ops node:$val, node:$ptr, node:$offset),
5010 (post_store node:$val, node:$ptr, node:$offset), [{
5011 return cast<StoreSDNode>(N)->getAlignment() >= 4;
5013 def aligned16_pre_store : PatFrag<(ops node:$val, node:$ptr, node:$offset),
5014 (pre_store node:$val, node:$ptr, node:$offset), [{
5015 return cast<StoreSDNode>(N)->getAlignment() >= 2;
5017 def aligned16_post_store : PatFrag<(ops node:$val, node:$ptr, node:$offset),
5018 (post_store node:$val, node:$ptr, node:$offset), [{
5019 return cast<StoreSDNode>(N)->getAlignment() >= 2;
5022 def alignedmaskedload32 : PatFrag<(ops node:$ptr, node:$pred, node:$passthru),
5023 (masked_ld node:$ptr, node:$pred, node:$passthru), [{
5024 return cast<MaskedLoadSDNode>(N)->getAlignment() >= 4;
5026 def alignedmaskedload16 : PatFrag<(ops node:$ptr, node:$pred, node:$passthru),
5027 (masked_ld node:$ptr, node:$pred, node:$passthru), [{
5028 return cast<MaskedLoadSDNode>(N)->getAlignment() >= 2;
5030 def maskedload : PatFrag<(ops node:$ptr, node:$pred, node:$passthru),
5031 (masked_ld node:$ptr, node:$pred, node:$passthru)>;
5033 def alignedmaskedstore32 : PatFrag<(ops node:$val, node:$ptr, node:$pred),
5034 (masked_st node:$val, node:$ptr, node:$pred), [{
5035 return cast<MaskedStoreSDNode>(N)->getAlignment() >= 4;
5037 def alignedmaskedstore16 : PatFrag<(ops node:$val, node:$ptr, node:$pred),
5038 (masked_st node:$val, node:$ptr, node:$pred), [{
5039 return cast<MaskedStoreSDNode>(N)->getAlignment() >= 2;
5041 def maskedstore : PatFrag<(ops node:$val, node:$ptr, node:$pred),
5042 (masked_st node:$val, node:$ptr, node:$pred)>;
5044 let Predicates = [HasMVEInt, IsLE] in {
5046 defm : MVE_vector_store<MVE_VSTRBU8, byte_alignedstore, 0>;
5047 defm : MVE_vector_store<MVE_VSTRHU16, hword_alignedstore, 1>;
5048 defm : MVE_vector_store<MVE_VSTRWU32, alignedstore32, 2>;
5051 defm : MVE_vector_load<MVE_VLDRBU8, byte_alignedload, 0>;
5052 defm : MVE_vector_load<MVE_VLDRHU16, hword_alignedload, 1>;
5053 defm : MVE_vector_load<MVE_VLDRWU32, alignedload32, 2>;
5055 // Pre/post inc stores
5056 defm : MVE_vector_offset_store<MVE_VSTRBU8_pre, pre_store, 0>;
5057 defm : MVE_vector_offset_store<MVE_VSTRBU8_post, post_store, 0>;
5058 defm : MVE_vector_offset_store<MVE_VSTRHU16_pre, aligned16_pre_store, 1>;
5059 defm : MVE_vector_offset_store<MVE_VSTRHU16_post, aligned16_post_store, 1>;
5060 defm : MVE_vector_offset_store<MVE_VSTRWU32_pre, aligned32_pre_store, 2>;
5061 defm : MVE_vector_offset_store<MVE_VSTRWU32_post, aligned32_post_store, 2>;
5063 // Unaligned masked stores (aligned are below)
5064 def : Pat<(maskedstore (v4i32 MQPR:$val), t2addrmode_imm7<0>:$addr, VCCR:$pred),
5065 (MVE_VSTRBU8 MQPR:$val, t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)>;
5066 def : Pat<(maskedstore (v4f32 MQPR:$val), t2addrmode_imm7<0>:$addr, VCCR:$pred),
5067 (MVE_VSTRBU8 MQPR:$val, t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)>;
5068 def : Pat<(maskedstore (v8i16 MQPR:$val), t2addrmode_imm7<0>:$addr, VCCR:$pred),
5069 (MVE_VSTRBU8 MQPR:$val, t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)>;
5070 def : Pat<(maskedstore (v8f16 MQPR:$val), t2addrmode_imm7<0>:$addr, VCCR:$pred),
5071 (MVE_VSTRBU8 MQPR:$val, t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)>;
5073 // Unaligned masked loads
5074 def : Pat<(v4i32 (maskedload t2addrmode_imm7<0>:$addr, VCCR:$pred, (v4i32 NEONimmAllZerosV))),
5075 (v4i32 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred))>;
5076 def : Pat<(v4f32 (maskedload t2addrmode_imm7<0>:$addr, VCCR:$pred, (v4f32 NEONimmAllZerosV))),
5077 (v4f32 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred))>;
5078 def : Pat<(v8i16 (maskedload t2addrmode_imm7<0>:$addr, VCCR:$pred, (v8i16 NEONimmAllZerosV))),
5079 (v8i16 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred))>;
5080 def : Pat<(v8f16 (maskedload t2addrmode_imm7<0>:$addr, VCCR:$pred, (v8f16 NEONimmAllZerosV))),
5081 (v8f16 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred))>;
5084 let Predicates = [HasMVEInt, IsBE] in {
5086 def : MVE_vector_store_typed<v16i8, MVE_VSTRBU8, store, 0>;
5087 def : MVE_vector_store_typed<v8i16, MVE_VSTRHU16, alignedstore16, 1>;
5088 def : MVE_vector_store_typed<v8f16, MVE_VSTRHU16, alignedstore16, 1>;
5089 def : MVE_vector_store_typed<v4i32, MVE_VSTRWU32, alignedstore32, 2>;
5090 def : MVE_vector_store_typed<v4f32, MVE_VSTRWU32, alignedstore32, 2>;
5093 def : MVE_vector_load_typed<v16i8, MVE_VLDRBU8, load, 0>;
5094 def : MVE_vector_load_typed<v8i16, MVE_VLDRHU16, alignedload16, 1>;
5095 def : MVE_vector_load_typed<v8f16, MVE_VLDRHU16, alignedload16, 1>;
5096 def : MVE_vector_load_typed<v4i32, MVE_VLDRWU32, alignedload32, 2>;
5097 def : MVE_vector_load_typed<v4f32, MVE_VLDRWU32, alignedload32, 2>;
5099 // Other unaligned loads/stores need to go though a VREV
5100 def : Pat<(v2f64 (load t2addrmode_imm7<0>:$addr)),
5101 (v2f64 (MVE_VREV64_8 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr)))>;
5102 def : Pat<(v2i64 (load t2addrmode_imm7<0>:$addr)),
5103 (v2i64 (MVE_VREV64_8 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr)))>;
5104 def : Pat<(v4i32 (load t2addrmode_imm7<0>:$addr)),
5105 (v4i32 (MVE_VREV32_8 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr)))>;
5106 def : Pat<(v4f32 (load t2addrmode_imm7<0>:$addr)),
5107 (v4f32 (MVE_VREV32_8 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr)))>;
5108 def : Pat<(v8i16 (load t2addrmode_imm7<0>:$addr)),
5109 (v8i16 (MVE_VREV16_8 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr)))>;
5110 def : Pat<(v8f16 (load t2addrmode_imm7<0>:$addr)),
5111 (v8f16 (MVE_VREV16_8 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr)))>;
5112 def : Pat<(store (v2f64 MQPR:$val), t2addrmode_imm7<0>:$addr),
5113 (MVE_VSTRBU8 (MVE_VREV64_8 MQPR:$val), t2addrmode_imm7<0>:$addr)>;
5114 def : Pat<(store (v2i64 MQPR:$val), t2addrmode_imm7<0>:$addr),
5115 (MVE_VSTRBU8 (MVE_VREV64_8 MQPR:$val), t2addrmode_imm7<0>:$addr)>;
5116 def : Pat<(store (v4i32 MQPR:$val), t2addrmode_imm7<0>:$addr),
5117 (MVE_VSTRBU8 (MVE_VREV32_8 MQPR:$val), t2addrmode_imm7<0>:$addr)>;
5118 def : Pat<(store (v4f32 MQPR:$val), t2addrmode_imm7<0>:$addr),
5119 (MVE_VSTRBU8 (MVE_VREV32_8 MQPR:$val), t2addrmode_imm7<0>:$addr)>;
5120 def : Pat<(store (v8i16 MQPR:$val), t2addrmode_imm7<0>:$addr),
5121 (MVE_VSTRBU8 (MVE_VREV16_8 MQPR:$val), t2addrmode_imm7<0>:$addr)>;
5122 def : Pat<(store (v8f16 MQPR:$val), t2addrmode_imm7<0>:$addr),
5123 (MVE_VSTRBU8 (MVE_VREV16_8 MQPR:$val), t2addrmode_imm7<0>:$addr)>;
5125 // Pre/Post inc stores
5126 def : MVE_vector_offset_store_typed<v16i8, MVE_VSTRBU8_pre, pre_store, 0>;
5127 def : MVE_vector_offset_store_typed<v16i8, MVE_VSTRBU8_post, post_store, 0>;
5128 def : MVE_vector_offset_store_typed<v8i16, MVE_VSTRHU16_pre, aligned16_pre_store, 1>;
5129 def : MVE_vector_offset_store_typed<v8i16, MVE_VSTRHU16_post, aligned16_post_store, 1>;
5130 def : MVE_vector_offset_store_typed<v8f16, MVE_VSTRHU16_pre, aligned16_pre_store, 1>;
5131 def : MVE_vector_offset_store_typed<v8f16, MVE_VSTRHU16_post, aligned16_post_store, 1>;
5132 def : MVE_vector_offset_store_typed<v4i32, MVE_VSTRWU32_pre, aligned32_pre_store, 2>;
5133 def : MVE_vector_offset_store_typed<v4i32, MVE_VSTRWU32_post, aligned32_post_store, 2>;
5134 def : MVE_vector_offset_store_typed<v4f32, MVE_VSTRWU32_pre, aligned32_pre_store, 2>;
5135 def : MVE_vector_offset_store_typed<v4f32, MVE_VSTRWU32_post, aligned32_post_store, 2>;
5137 // Unaligned masked stores (aligned are below)
5138 def : Pat<(maskedstore (v4i32 MQPR:$val), t2addrmode_imm7<0>:$addr, VCCR:$pred),
5139 (MVE_VSTRBU8 (MVE_VREV32_8 MQPR:$val), t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)>;
5140 def : Pat<(maskedstore (v4f32 MQPR:$val), t2addrmode_imm7<0>:$addr, VCCR:$pred),
5141 (MVE_VSTRBU8 (MVE_VREV32_8 MQPR:$val), t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)>;
5142 def : Pat<(maskedstore (v8i16 MQPR:$val), t2addrmode_imm7<0>:$addr, VCCR:$pred),
5143 (MVE_VSTRBU8 (MVE_VREV16_8 MQPR:$val), t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)>;
5144 def : Pat<(maskedstore (v8f16 MQPR:$val), t2addrmode_imm7<0>:$addr, VCCR:$pred),
5145 (MVE_VSTRBU8 (MVE_VREV16_8 MQPR:$val), t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)>;
5146 // Unaligned masked loads
5147 def : Pat<(v4i32 (maskedload t2addrmode_imm7<0>:$addr, VCCR:$pred, (v4i32 NEONimmAllZerosV))),
5148 (v4i32 (MVE_VREV32_8 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)))>;
5149 def : Pat<(v4f32 (maskedload t2addrmode_imm7<0>:$addr, VCCR:$pred, (v4f32 NEONimmAllZerosV))),
5150 (v4f32 (MVE_VREV32_8 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)))>;
5151 def : Pat<(v8i16 (maskedload t2addrmode_imm7<0>:$addr, VCCR:$pred, (v8i16 NEONimmAllZerosV))),
5152 (v8i16 (MVE_VREV16_8 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)))>;
5153 def : Pat<(v8f16 (maskedload t2addrmode_imm7<0>:$addr, VCCR:$pred, (v8f16 NEONimmAllZerosV))),
5154 (v8f16 (MVE_VREV16_8 (MVE_VLDRBU8 t2addrmode_imm7<0>:$addr, (i32 1), VCCR:$pred)))>;
5157 let Predicates = [HasMVEInt] in {
5158 // Aligned masked store, shared between LE and BE
5159 def : MVE_vector_maskedstore_typed<v16i8, MVE_VSTRBU8, maskedstore, 0>;
5160 def : MVE_vector_maskedstore_typed<v8i16, MVE_VSTRHU16, alignedmaskedstore16, 1>;
5161 def : MVE_vector_maskedstore_typed<v8f16, MVE_VSTRHU16, alignedmaskedstore16, 1>;
5162 def : MVE_vector_maskedstore_typed<v4i32, MVE_VSTRWU32, alignedmaskedstore32, 2>;
5163 def : MVE_vector_maskedstore_typed<v4f32, MVE_VSTRWU32, alignedmaskedstore32, 2>;
5164 // Aligned masked loads
5165 def : MVE_vector_maskedload_typed<v16i8, MVE_VLDRBU8, maskedload, 0>;
5166 def : MVE_vector_maskedload_typed<v8i16, MVE_VLDRHU16, alignedmaskedload16, 1>;
5167 def : MVE_vector_maskedload_typed<v8f16, MVE_VLDRHU16, alignedmaskedload16, 1>;
5168 def : MVE_vector_maskedload_typed<v4i32, MVE_VLDRWU32, alignedmaskedload32, 2>;
5169 def : MVE_vector_maskedload_typed<v4f32, MVE_VLDRWU32, alignedmaskedload32, 2>;
5172 // Widening/Narrowing Loads/Stores
5174 let MinAlignment = 2 in {
5175 def truncstorevi16_align2 : PatFrag<(ops node:$val, node:$ptr),
5176 (truncstorevi16 node:$val, node:$ptr)>;
5177 def post_truncstvi16_align2 : PatFrag<(ops node:$val, node:$base, node:$offset),
5178 (post_truncstvi16 node:$val, node:$base, node:$offset)>;
5179 def pre_truncstvi16_align2 : PatFrag<(ops node:$val, node:$base, node:$offset),
5180 (pre_truncstvi16 node:$val, node:$base, node:$offset)>;
5183 let Predicates = [HasMVEInt] in {
5184 def : Pat<(truncstorevi8 (v8i16 MQPR:$val), taddrmode_imm7<0>:$addr),
5185 (MVE_VSTRB16 MQPR:$val, taddrmode_imm7<0>:$addr)>;
5186 def : Pat<(truncstorevi8 (v4i32 MQPR:$val), taddrmode_imm7<0>:$addr),
5187 (MVE_VSTRB32 MQPR:$val, taddrmode_imm7<0>:$addr)>;
5188 def : Pat<(truncstorevi16_align2 (v4i32 MQPR:$val), taddrmode_imm7<1>:$addr),
5189 (MVE_VSTRH32 MQPR:$val, taddrmode_imm7<1>:$addr)>;
5191 def : Pat<(post_truncstvi8 (v8i16 MQPR:$Rt), tGPR:$Rn, t2am_imm7_offset<0>:$addr),
5192 (MVE_VSTRB16_post MQPR:$Rt, tGPR:$Rn, t2am_imm7_offset<0>:$addr)>;
5193 def : Pat<(post_truncstvi8 (v4i32 MQPR:$Rt), tGPR:$Rn, t2am_imm7_offset<0>:$addr),
5194 (MVE_VSTRB32_post MQPR:$Rt, tGPR:$Rn, t2am_imm7_offset<0>:$addr)>;
5195 def : Pat<(post_truncstvi16_align2 (v4i32 MQPR:$Rt), tGPR:$Rn, t2am_imm7_offset<1>:$addr),
5196 (MVE_VSTRH32_post MQPR:$Rt, tGPR:$Rn, t2am_imm7_offset<1>:$addr)>;
5198 def : Pat<(pre_truncstvi8 (v8i16 MQPR:$Rt), tGPR:$Rn, t2am_imm7_offset<0>:$addr),
5199 (MVE_VSTRB16_pre MQPR:$Rt, tGPR:$Rn, t2am_imm7_offset<0>:$addr)>;
5200 def : Pat<(pre_truncstvi8 (v4i32 MQPR:$Rt), tGPR:$Rn, t2am_imm7_offset<0>:$addr),
5201 (MVE_VSTRB32_pre MQPR:$Rt, tGPR:$Rn, t2am_imm7_offset<0>:$addr)>;
5202 def : Pat<(pre_truncstvi16_align2 (v4i32 MQPR:$Rt), tGPR:$Rn, t2am_imm7_offset<1>:$addr),
5203 (MVE_VSTRH32_pre MQPR:$Rt, tGPR:$Rn, t2am_imm7_offset<1>:$addr)>;
5207 let MinAlignment = 2 in {
5208 def extloadvi16_align2 : PatFrag<(ops node:$ptr), (extloadvi16 node:$ptr)>;
5209 def sextloadvi16_align2 : PatFrag<(ops node:$ptr), (sextloadvi16 node:$ptr)>;
5210 def zextloadvi16_align2 : PatFrag<(ops node:$ptr), (zextloadvi16 node:$ptr)>;
5213 multiclass MVEExtLoad<string DestLanes, string DestElemBits,
5214 string SrcElemBits, string SrcElemType,
5215 string Align, Operand am> {
5216 def _Any : Pat<(!cast<ValueType>("v" # DestLanes # "i" # DestElemBits)
5217 (!cast<PatFrag>("extloadvi" # SrcElemBits # Align) am:$addr)),
5218 (!cast<Instruction>("MVE_VLDR" # SrcElemType # "U" # DestElemBits)
5220 def _Z : Pat<(!cast<ValueType>("v" # DestLanes # "i" # DestElemBits)
5221 (!cast<PatFrag>("zextloadvi" # SrcElemBits # Align) am:$addr)),
5222 (!cast<Instruction>("MVE_VLDR" # SrcElemType # "U" # DestElemBits)
5224 def _S : Pat<(!cast<ValueType>("v" # DestLanes # "i" # DestElemBits)
5225 (!cast<PatFrag>("sextloadvi" # SrcElemBits # Align) am:$addr)),
5226 (!cast<Instruction>("MVE_VLDR" # SrcElemType # "S" # DestElemBits)
5230 let Predicates = [HasMVEInt] in {
5231 defm : MVEExtLoad<"4", "32", "8", "B", "", taddrmode_imm7<0>>;
5232 defm : MVEExtLoad<"8", "16", "8", "B", "", taddrmode_imm7<0>>;
5233 defm : MVEExtLoad<"4", "32", "16", "H", "_align2", taddrmode_imm7<1>>;
5237 // Bit convert patterns
5239 let Predicates = [HasMVEInt] in {
5240 def : Pat<(v2f64 (bitconvert (v2i64 MQPR:$src))), (v2f64 MQPR:$src)>;
5241 def : Pat<(v2i64 (bitconvert (v2f64 MQPR:$src))), (v2i64 MQPR:$src)>;
5243 def : Pat<(v4i32 (bitconvert (v4f32 MQPR:$src))), (v4i32 MQPR:$src)>;
5244 def : Pat<(v4f32 (bitconvert (v4i32 MQPR:$src))), (v4f32 MQPR:$src)>;
5246 def : Pat<(v8i16 (bitconvert (v8f16 MQPR:$src))), (v8i16 MQPR:$src)>;
5247 def : Pat<(v8f16 (bitconvert (v8i16 MQPR:$src))), (v8f16 MQPR:$src)>;
5250 let Predicates = [IsLE,HasMVEInt] in {
5251 def : Pat<(v2f64 (bitconvert (v4f32 MQPR:$src))), (v2f64 MQPR:$src)>;
5252 def : Pat<(v2f64 (bitconvert (v4i32 MQPR:$src))), (v2f64 MQPR:$src)>;
5253 def : Pat<(v2f64 (bitconvert (v8f16 MQPR:$src))), (v2f64 MQPR:$src)>;
5254 def : Pat<(v2f64 (bitconvert (v8i16 MQPR:$src))), (v2f64 MQPR:$src)>;
5255 def : Pat<(v2f64 (bitconvert (v16i8 MQPR:$src))), (v2f64 MQPR:$src)>;
5257 def : Pat<(v2i64 (bitconvert (v4f32 MQPR:$src))), (v2i64 MQPR:$src)>;
5258 def : Pat<(v2i64 (bitconvert (v4i32 MQPR:$src))), (v2i64 MQPR:$src)>;
5259 def : Pat<(v2i64 (bitconvert (v8f16 MQPR:$src))), (v2i64 MQPR:$src)>;
5260 def : Pat<(v2i64 (bitconvert (v8i16 MQPR:$src))), (v2i64 MQPR:$src)>;
5261 def : Pat<(v2i64 (bitconvert (v16i8 MQPR:$src))), (v2i64 MQPR:$src)>;
5263 def : Pat<(v4f32 (bitconvert (v2f64 MQPR:$src))), (v4f32 MQPR:$src)>;
5264 def : Pat<(v4f32 (bitconvert (v2i64 MQPR:$src))), (v4f32 MQPR:$src)>;
5265 def : Pat<(v4f32 (bitconvert (v8f16 MQPR:$src))), (v4f32 MQPR:$src)>;
5266 def : Pat<(v4f32 (bitconvert (v8i16 MQPR:$src))), (v4f32 MQPR:$src)>;
5267 def : Pat<(v4f32 (bitconvert (v16i8 MQPR:$src))), (v4f32 MQPR:$src)>;
5269 def : Pat<(v4i32 (bitconvert (v2f64 MQPR:$src))), (v4i32 MQPR:$src)>;
5270 def : Pat<(v4i32 (bitconvert (v2i64 MQPR:$src))), (v4i32 MQPR:$src)>;
5271 def : Pat<(v4i32 (bitconvert (v8f16 MQPR:$src))), (v4i32 MQPR:$src)>;
5272 def : Pat<(v4i32 (bitconvert (v8i16 MQPR:$src))), (v4i32 MQPR:$src)>;
5273 def : Pat<(v4i32 (bitconvert (v16i8 MQPR:$src))), (v4i32 MQPR:$src)>;
5275 def : Pat<(v8f16 (bitconvert (v2f64 MQPR:$src))), (v8f16 MQPR:$src)>;
5276 def : Pat<(v8f16 (bitconvert (v2i64 MQPR:$src))), (v8f16 MQPR:$src)>;
5277 def : Pat<(v8f16 (bitconvert (v4f32 MQPR:$src))), (v8f16 MQPR:$src)>;
5278 def : Pat<(v8f16 (bitconvert (v4i32 MQPR:$src))), (v8f16 MQPR:$src)>;
5279 def : Pat<(v8f16 (bitconvert (v16i8 MQPR:$src))), (v8f16 MQPR:$src)>;
5281 def : Pat<(v8i16 (bitconvert (v2f64 MQPR:$src))), (v8i16 MQPR:$src)>;
5282 def : Pat<(v8i16 (bitconvert (v2i64 MQPR:$src))), (v8i16 MQPR:$src)>;
5283 def : Pat<(v8i16 (bitconvert (v4f32 MQPR:$src))), (v8i16 MQPR:$src)>;
5284 def : Pat<(v8i16 (bitconvert (v4i32 MQPR:$src))), (v8i16 MQPR:$src)>;
5285 def : Pat<(v8i16 (bitconvert (v16i8 MQPR:$src))), (v8i16 MQPR:$src)>;
5287 def : Pat<(v16i8 (bitconvert (v2f64 MQPR:$src))), (v16i8 MQPR:$src)>;
5288 def : Pat<(v16i8 (bitconvert (v2i64 MQPR:$src))), (v16i8 MQPR:$src)>;
5289 def : Pat<(v16i8 (bitconvert (v4f32 MQPR:$src))), (v16i8 MQPR:$src)>;
5290 def : Pat<(v16i8 (bitconvert (v4i32 MQPR:$src))), (v16i8 MQPR:$src)>;
5291 def : Pat<(v16i8 (bitconvert (v8f16 MQPR:$src))), (v16i8 MQPR:$src)>;
5292 def : Pat<(v16i8 (bitconvert (v8i16 MQPR:$src))), (v16i8 MQPR:$src)>;
5295 let Predicates = [IsBE,HasMVEInt] in {
5296 def : Pat<(v2f64 (bitconvert (v4f32 MQPR:$src))), (v2f64 (MVE_VREV64_32 MQPR:$src))>;
5297 def : Pat<(v2f64 (bitconvert (v4i32 MQPR:$src))), (v2f64 (MVE_VREV64_32 MQPR:$src))>;
5298 def : Pat<(v2f64 (bitconvert (v8f16 MQPR:$src))), (v2f64 (MVE_VREV64_16 MQPR:$src))>;
5299 def : Pat<(v2f64 (bitconvert (v8i16 MQPR:$src))), (v2f64 (MVE_VREV64_16 MQPR:$src))>;
5300 def : Pat<(v2f64 (bitconvert (v16i8 MQPR:$src))), (v2f64 (MVE_VREV64_8 MQPR:$src))>;
5302 def : Pat<(v2i64 (bitconvert (v4f32 MQPR:$src))), (v2i64 (MVE_VREV64_32 MQPR:$src))>;
5303 def : Pat<(v2i64 (bitconvert (v4i32 MQPR:$src))), (v2i64 (MVE_VREV64_32 MQPR:$src))>;
5304 def : Pat<(v2i64 (bitconvert (v8f16 MQPR:$src))), (v2i64 (MVE_VREV64_16 MQPR:$src))>;
5305 def : Pat<(v2i64 (bitconvert (v8i16 MQPR:$src))), (v2i64 (MVE_VREV64_16 MQPR:$src))>;
5306 def : Pat<(v2i64 (bitconvert (v16i8 MQPR:$src))), (v2i64 (MVE_VREV64_8 MQPR:$src))>;
5308 def : Pat<(v4f32 (bitconvert (v2f64 MQPR:$src))), (v4f32 (MVE_VREV64_32 MQPR:$src))>;
5309 def : Pat<(v4f32 (bitconvert (v2i64 MQPR:$src))), (v4f32 (MVE_VREV64_32 MQPR:$src))>;
5310 def : Pat<(v4f32 (bitconvert (v8f16 MQPR:$src))), (v4f32 (MVE_VREV32_16 MQPR:$src))>;
5311 def : Pat<(v4f32 (bitconvert (v8i16 MQPR:$src))), (v4f32 (MVE_VREV32_16 MQPR:$src))>;
5312 def : Pat<(v4f32 (bitconvert (v16i8 MQPR:$src))), (v4f32 (MVE_VREV32_8 MQPR:$src))>;
5314 def : Pat<(v4i32 (bitconvert (v2f64 MQPR:$src))), (v4i32 (MVE_VREV64_32 MQPR:$src))>;
5315 def : Pat<(v4i32 (bitconvert (v2i64 MQPR:$src))), (v4i32 (MVE_VREV64_32 MQPR:$src))>;
5316 def : Pat<(v4i32 (bitconvert (v8f16 MQPR:$src))), (v4i32 (MVE_VREV32_16 MQPR:$src))>;
5317 def : Pat<(v4i32 (bitconvert (v8i16 MQPR:$src))), (v4i32 (MVE_VREV32_16 MQPR:$src))>;
5318 def : Pat<(v4i32 (bitconvert (v16i8 MQPR:$src))), (v4i32 (MVE_VREV32_8 MQPR:$src))>;
5320 def : Pat<(v8f16 (bitconvert (v2f64 MQPR:$src))), (v8f16 (MVE_VREV64_16 MQPR:$src))>;
5321 def : Pat<(v8f16 (bitconvert (v2i64 MQPR:$src))), (v8f16 (MVE_VREV64_16 MQPR:$src))>;
5322 def : Pat<(v8f16 (bitconvert (v4f32 MQPR:$src))), (v8f16 (MVE_VREV32_16 MQPR:$src))>;
5323 def : Pat<(v8f16 (bitconvert (v4i32 MQPR:$src))), (v8f16 (MVE_VREV32_16 MQPR:$src))>;
5324 def : Pat<(v8f16 (bitconvert (v16i8 MQPR:$src))), (v8f16 (MVE_VREV16_8 MQPR:$src))>;
5326 def : Pat<(v8i16 (bitconvert (v2f64 MQPR:$src))), (v8i16 (MVE_VREV64_16 MQPR:$src))>;
5327 def : Pat<(v8i16 (bitconvert (v2i64 MQPR:$src))), (v8i16 (MVE_VREV64_16 MQPR:$src))>;
5328 def : Pat<(v8i16 (bitconvert (v4f32 MQPR:$src))), (v8i16 (MVE_VREV32_16 MQPR:$src))>;
5329 def : Pat<(v8i16 (bitconvert (v4i32 MQPR:$src))), (v8i16 (MVE_VREV32_16 MQPR:$src))>;
5330 def : Pat<(v8i16 (bitconvert (v16i8 MQPR:$src))), (v8i16 (MVE_VREV16_8 MQPR:$src))>;
5332 def : Pat<(v16i8 (bitconvert (v2f64 MQPR:$src))), (v16i8 (MVE_VREV64_8 MQPR:$src))>;
5333 def : Pat<(v16i8 (bitconvert (v2i64 MQPR:$src))), (v16i8 (MVE_VREV64_8 MQPR:$src))>;
5334 def : Pat<(v16i8 (bitconvert (v4f32 MQPR:$src))), (v16i8 (MVE_VREV32_8 MQPR:$src))>;
5335 def : Pat<(v16i8 (bitconvert (v4i32 MQPR:$src))), (v16i8 (MVE_VREV32_8 MQPR:$src))>;
5336 def : Pat<(v16i8 (bitconvert (v8f16 MQPR:$src))), (v16i8 (MVE_VREV16_8 MQPR:$src))>;
5337 def : Pat<(v16i8 (bitconvert (v8i16 MQPR:$src))), (v16i8 (MVE_VREV16_8 MQPR:$src))>;