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 let ParserMatchClass = TMemImm7ShiftOffsetAsmOperand<shift>;
165 // They are printed the same way as the T2 imm8 version
166 let PrintMethod = "printT2AddrModeImm8Operand<false>";
167 // This can also be the same as the T2 version.
168 let EncoderMethod = "getT2AddrModeImmOpValue<7,"#shift#">";
169 let DecoderMethod = "DecodeTAddrModeImm7<"#shift#">";
170 let MIOperandInfo = (ops tGPR:$base, i32imm:$offsimm);
173 // t2addrmode_imm7 := reg +/- (imm7)
174 class MemImm7ShiftOffsetAsmOperand<int shift> : AsmOperandClass {
175 let Name = "MemImm7Shift"#shift#"Offset";
176 let PredicateMethod = "isMemImm7ShiftedOffset<" # shift #
177 ",ARM::GPRnopcRegClassID>";
178 let RenderMethod = "addMemImmOffsetOperands";
181 def MemImm7Shift0OffsetAsmOperand : MemImm7ShiftOffsetAsmOperand<0>;
182 def MemImm7Shift1OffsetAsmOperand : MemImm7ShiftOffsetAsmOperand<1>;
183 def MemImm7Shift2OffsetAsmOperand : MemImm7ShiftOffsetAsmOperand<2>;
184 class T2AddrMode_Imm7<int shift> : MemOperand,
185 ComplexPattern<i32, 2, "SelectT2AddrModeImm7<"#shift#">", []> {
186 let EncoderMethod = "getT2AddrModeImmOpValue<7,"#shift#">";
187 let DecoderMethod = "DecodeT2AddrModeImm7<"#shift#", 0>";
188 let ParserMatchClass =
189 !cast<AsmOperandClass>("MemImm7Shift"#shift#"OffsetAsmOperand");
190 let MIOperandInfo = (ops GPRnopc:$base, i32imm:$offsimm);
193 class t2addrmode_imm7<int shift> : T2AddrMode_Imm7<shift> {
194 // They are printed the same way as the imm8 version
195 let PrintMethod = "printT2AddrModeImm8Operand<false>";
198 class MemImm7ShiftOffsetWBAsmOperand<int shift> : AsmOperandClass {
199 let Name = "MemImm7Shift"#shift#"OffsetWB";
200 let PredicateMethod = "isMemImm7ShiftedOffset<" # shift #
201 ",ARM::rGPRRegClassID>";
202 let RenderMethod = "addMemImmOffsetOperands";
205 def MemImm7Shift0OffsetWBAsmOperand : MemImm7ShiftOffsetWBAsmOperand<0>;
206 def MemImm7Shift1OffsetWBAsmOperand : MemImm7ShiftOffsetWBAsmOperand<1>;
207 def MemImm7Shift2OffsetWBAsmOperand : MemImm7ShiftOffsetWBAsmOperand<2>;
209 class t2addrmode_imm7_pre<int shift> : T2AddrMode_Imm7<shift> {
210 // They are printed the same way as the imm8 version
211 let PrintMethod = "printT2AddrModeImm8Operand<true>";
212 let ParserMatchClass =
213 !cast<AsmOperandClass>("MemImm7Shift"#shift#"OffsetWBAsmOperand");
214 let DecoderMethod = "DecodeT2AddrModeImm7<"#shift#", 1>";
215 let MIOperandInfo = (ops rGPR:$base, i32imm:$offsim);
218 class t2am_imm7shiftOffsetAsmOperand<int shift>
219 : AsmOperandClass { let Name = "Imm7Shift"#shift; }
220 def t2am_imm7shift0OffsetAsmOperand : t2am_imm7shiftOffsetAsmOperand<0>;
221 def t2am_imm7shift1OffsetAsmOperand : t2am_imm7shiftOffsetAsmOperand<1>;
222 def t2am_imm7shift2OffsetAsmOperand : t2am_imm7shiftOffsetAsmOperand<2>;
224 class t2am_imm7_offset<int shift> : MemOperand {
225 // They are printed the same way as the imm8 version
226 let PrintMethod = "printT2AddrModeImm8OffsetOperand";
227 let ParserMatchClass =
228 !cast<AsmOperandClass>("t2am_imm7shift"#shift#"OffsetAsmOperand");
229 let EncoderMethod = "getT2ScaledImmOpValue<7,"#shift#">";
230 let DecoderMethod = "DecodeT2Imm7<"#shift#">";
233 // Operands for gather/scatter loads of the form [Rbase, Qoffsets]
234 class MemRegRQOffsetAsmOperand<int shift> : AsmOperandClass {
235 let Name = "MemRegRQS"#shift#"Offset";
236 let PredicateMethod = "isMemRegRQOffset<"#shift#">";
237 let RenderMethod = "addMemRegRQOffsetOperands";
240 def MemRegRQS0OffsetAsmOperand : MemRegRQOffsetAsmOperand<0>;
241 def MemRegRQS1OffsetAsmOperand : MemRegRQOffsetAsmOperand<1>;
242 def MemRegRQS2OffsetAsmOperand : MemRegRQOffsetAsmOperand<2>;
243 def MemRegRQS3OffsetAsmOperand : MemRegRQOffsetAsmOperand<3>;
245 // mve_addr_rq_shift := reg + vreg{ << UXTW #shift}
246 class mve_addr_rq_shift<int shift> : MemOperand {
247 let EncoderMethod = "getMveAddrModeRQOpValue";
248 let PrintMethod = "printMveAddrModeRQOperand<"#shift#">";
249 let ParserMatchClass =
250 !cast<AsmOperandClass>("MemRegRQS"#shift#"OffsetAsmOperand");
251 let DecoderMethod = "DecodeMveAddrModeRQ";
252 let MIOperandInfo = (ops GPRnopc:$base, MQPR:$offsreg);
255 class MemRegQOffsetAsmOperand<int shift> : AsmOperandClass {
256 let Name = "MemRegQS"#shift#"Offset";
257 let PredicateMethod = "isMemRegQOffset<"#shift#">";
258 let RenderMethod = "addMemImmOffsetOperands";
261 def MemRegQS2OffsetAsmOperand : MemRegQOffsetAsmOperand<2>;
262 def MemRegQS3OffsetAsmOperand : MemRegQOffsetAsmOperand<3>;
264 // mve_addr_q_shift := vreg {+ #imm7s2/4}
265 class mve_addr_q_shift<int shift> : MemOperand {
266 let EncoderMethod = "getMveAddrModeQOpValue<"#shift#">";
267 // Can be printed same way as other reg + imm operands
268 let PrintMethod = "printT2AddrModeImm8Operand<false>";
269 let ParserMatchClass =
270 !cast<AsmOperandClass>("MemRegQS"#shift#"OffsetAsmOperand");
271 let DecoderMethod = "DecodeMveAddrModeQ<"#shift#">";
272 let MIOperandInfo = (ops MQPR:$base, i32imm:$imm);
275 // --------- Start of base classes for the instructions themselves
277 class MVE_MI<dag oops, dag iops, InstrItinClass itin, string asm,
278 string ops, string cstr, list<dag> pattern>
279 : Thumb2XI<oops, iops, AddrModeNone, 4, itin, !strconcat(asm, "\t", ops), cstr,
281 Requires<[HasMVEInt]> {
283 let DecoderNamespace = "MVE";
286 // MVE_p is used for most predicated instructions, to add the cluster
287 // of input operands that provides the VPT suffix (none, T or E) and
288 // the input predicate register.
289 class MVE_p<dag oops, dag iops, InstrItinClass itin, string iname,
290 string suffix, string ops, vpred_ops vpred, string cstr,
291 list<dag> pattern=[]>
292 : MVE_MI<oops, !con(iops, (ins vpred:$vp)), itin,
293 // If the instruction has a suffix, like vadd.f32, then the
294 // VPT predication suffix goes before the dot, so the full
295 // name has to be "vadd${vp}.f32".
296 !strconcat(iname, "${vp}",
297 !if(!eq(suffix, ""), "", !strconcat(".", suffix))),
298 ops, !strconcat(cstr, vpred.vpred_constraint), pattern> {
299 let Inst{31-29} = 0b111;
300 let Inst{27-26} = 0b11;
303 class MVE_f<dag oops, dag iops, InstrItinClass itin, string iname,
304 string suffix, string ops, vpred_ops vpred, string cstr,
305 list<dag> pattern=[]>
306 : MVE_p<oops, iops, itin, iname, suffix, ops, vpred, cstr, pattern> {
307 let Predicates = [HasMVEFloat];
310 class MVE_MI_with_pred<dag oops, dag iops, InstrItinClass itin, string asm,
311 string ops, string cstr, list<dag> pattern>
312 : Thumb2I<oops, iops, AddrModeNone, 4, itin, asm, !strconcat("\t", ops), cstr,
314 Requires<[HasV8_1MMainline, HasMVEInt]> {
316 let DecoderNamespace = "MVE";
319 class MVE_VMOV_lane_base<dag oops, dag iops, InstrItinClass itin, string asm,
320 string suffix, string ops, string cstr,
322 : Thumb2I<oops, iops, AddrModeNone, 4, itin, asm,
323 !if(!eq(suffix, ""), "", "." # suffix) # "\t" # ops,
325 Requires<[HasV8_1MMainline, HasMVEInt]> {
327 let DecoderNamespace = "MVE";
330 class MVE_ScalarShift<string iname, dag oops, dag iops, string asm, string cstr,
331 list<dag> pattern=[]>
332 : MVE_MI_with_pred<oops, iops, NoItinerary, iname, asm, cstr, pattern> {
333 let Inst{31-20} = 0b111010100101;
338 class MVE_ScalarShiftSingleReg<string iname, dag iops, string asm, string cstr,
339 list<dag> pattern=[]>
340 : MVE_ScalarShift<iname, (outs rGPR:$RdaDest), iops, asm, cstr, pattern> {
343 let Inst{19-16} = RdaDest{3-0};
346 class MVE_ScalarShiftSRegImm<string iname, bits<2> op5_4, list<dag> pattern=[]>
347 : MVE_ScalarShiftSingleReg<iname, (ins rGPR:$RdaSrc, long_shift:$imm),
348 "$RdaSrc, $imm", "$RdaDest = $RdaSrc", pattern> {
352 let Inst{14-12} = imm{4-2};
353 let Inst{11-8} = 0b1111;
354 let Inst{7-6} = imm{1-0};
355 let Inst{5-4} = op5_4{1-0};
356 let Inst{3-0} = 0b1111;
359 def MVE_SQSHL : MVE_ScalarShiftSRegImm<"sqshl", 0b11>;
360 def MVE_SRSHR : MVE_ScalarShiftSRegImm<"srshr", 0b10>;
361 def MVE_UQSHL : MVE_ScalarShiftSRegImm<"uqshl", 0b00>;
362 def MVE_URSHR : MVE_ScalarShiftSRegImm<"urshr", 0b01>;
364 class MVE_ScalarShiftSRegReg<string iname, bits<2> op5_4, list<dag> pattern=[]>
365 : MVE_ScalarShiftSingleReg<iname, (ins rGPR:$RdaSrc, rGPR:$Rm),
366 "$RdaSrc, $Rm", "$RdaDest = $RdaSrc", pattern> {
369 let Inst{15-12} = Rm{3-0};
370 let Inst{11-8} = 0b1111;
371 let Inst{7-6} = 0b00;
372 let Inst{5-4} = op5_4{1-0};
373 let Inst{3-0} = 0b1101;
376 def MVE_SQRSHR : MVE_ScalarShiftSRegReg<"sqrshr", 0b10>;
377 def MVE_UQRSHL : MVE_ScalarShiftSRegReg<"uqrshl", 0b00>;
379 class MVE_ScalarShiftDoubleReg<string iname, dag iops, string asm,
380 string cstr, list<dag> pattern=[]>
381 : MVE_ScalarShift<iname, (outs tGPREven:$RdaLo, tGPROdd:$RdaHi),
382 iops, asm, cstr, pattern> {
386 let Inst{19-17} = RdaLo{3-1};
387 let Inst{11-9} = RdaHi{3-1};
390 class MVE_ScalarShiftDRegImm<string iname, bits<2> op5_4, bit op16,
391 list<dag> pattern=[]>
392 : MVE_ScalarShiftDoubleReg<
393 iname, (ins tGPREven:$RdaLo_src, tGPROdd:$RdaHi_src, long_shift:$imm),
394 "$RdaLo, $RdaHi, $imm", "$RdaLo = $RdaLo_src,$RdaHi = $RdaHi_src",
400 let Inst{14-12} = imm{4-2};
401 let Inst{7-6} = imm{1-0};
402 let Inst{5-4} = op5_4{1-0};
403 let Inst{3-0} = 0b1111;
406 class MVE_ScalarShiftDRegReg<string iname, bit op5, bit op16,
407 list<dag> pattern=[]>
408 : MVE_ScalarShiftDoubleReg<
409 iname, (ins tGPREven:$RdaLo_src, tGPROdd:$RdaHi_src, rGPR:$Rm),
410 "$RdaLo, $RdaHi, $Rm", "@earlyclobber $RdaHi,@earlyclobber $RdaLo,"
411 "$RdaLo = $RdaLo_src,$RdaHi = $RdaHi_src",
416 let Inst{15-12} = Rm{3-0};
417 let Inst{7-6} = 0b00;
420 let Inst{3-0} = 0b1101;
422 // Custom decoder method because of the following overlapping encodings:
425 // SQRSHRL and SQRSHR
426 // UQRSHLL and UQRSHL
427 let DecoderMethod = "DecodeMVEOverlappingLongShift";
430 def MVE_ASRLr : MVE_ScalarShiftDRegReg<"asrl", 0b1, 0b0, [(set tGPREven:$RdaLo, tGPROdd:$RdaHi,
431 (ARMasrl tGPREven:$RdaLo_src,
432 tGPROdd:$RdaHi_src, rGPR:$Rm))]>;
433 def MVE_ASRLi : MVE_ScalarShiftDRegImm<"asrl", 0b10, ?, [(set tGPREven:$RdaLo, tGPROdd:$RdaHi,
434 (ARMasrl tGPREven:$RdaLo_src,
435 tGPROdd:$RdaHi_src, (i32 imm:$imm)))]>;
436 def MVE_LSLLr : MVE_ScalarShiftDRegReg<"lsll", 0b0, 0b0, [(set tGPREven:$RdaLo, tGPROdd:$RdaHi,
437 (ARMlsll tGPREven:$RdaLo_src,
438 tGPROdd:$RdaHi_src, rGPR:$Rm))]>;
439 def MVE_LSLLi : MVE_ScalarShiftDRegImm<"lsll", 0b00, ?, [(set tGPREven:$RdaLo, tGPROdd:$RdaHi,
440 (ARMlsll tGPREven:$RdaLo_src,
441 tGPROdd:$RdaHi_src, (i32 imm:$imm)))]>;
442 def MVE_LSRL : MVE_ScalarShiftDRegImm<"lsrl", 0b01, ?, [(set tGPREven:$RdaLo, tGPROdd:$RdaHi,
443 (ARMlsrl tGPREven:$RdaLo_src,
444 tGPROdd:$RdaHi_src, (i32 imm:$imm)))]>;
446 def MVE_SQRSHRL : MVE_ScalarShiftDRegReg<"sqrshrl", 0b1, 0b1>;
447 def MVE_SQSHLL : MVE_ScalarShiftDRegImm<"sqshll", 0b11, 0b1>;
448 def MVE_SRSHRL : MVE_ScalarShiftDRegImm<"srshrl", 0b10, 0b1>;
450 def MVE_UQRSHLL : MVE_ScalarShiftDRegReg<"uqrshll", 0b0, 0b1>;
451 def MVE_UQSHLL : MVE_ScalarShiftDRegImm<"uqshll", 0b00, 0b1>;
452 def MVE_URSHRL : MVE_ScalarShiftDRegImm<"urshrl", 0b01, 0b1>;
454 // start of mve_rDest instructions
456 class MVE_rDest<dag oops, dag iops, InstrItinClass itin,
457 string iname, string suffix,
458 string ops, string cstr, list<dag> pattern=[]>
459 // Always use vpred_n and not vpred_r: with the output register being
460 // a GPR and not a vector register, there can't be any question of
461 // what to put in its inactive lanes.
462 : MVE_p<oops, iops, itin, iname, suffix, ops, vpred_n, cstr, pattern> {
464 let Inst{25-23} = 0b101;
465 let Inst{11-9} = 0b111;
469 class MVE_VABAV<string suffix, bit U, bits<2> size, list<dag> pattern=[]>
470 : MVE_rDest<(outs rGPR:$Rda), (ins rGPR:$Rda_src, MQPR:$Qn, MQPR:$Qm),
471 NoItinerary, "vabav", suffix, "$Rda, $Qn, $Qm", "$Rda = $Rda_src",
479 let Inst{21-20} = size{1-0};
480 let Inst{19-17} = Qn{2-0};
482 let Inst{15-12} = Rda{3-0};
487 let Inst{3-1} = Qm{2-0};
491 def MVE_VABAVs8 : MVE_VABAV<"s8", 0b0, 0b00>;
492 def MVE_VABAVs16 : MVE_VABAV<"s16", 0b0, 0b01>;
493 def MVE_VABAVs32 : MVE_VABAV<"s32", 0b0, 0b10>;
494 def MVE_VABAVu8 : MVE_VABAV<"u8", 0b1, 0b00>;
495 def MVE_VABAVu16 : MVE_VABAV<"u16", 0b1, 0b01>;
496 def MVE_VABAVu32 : MVE_VABAV<"u32", 0b1, 0b10>;
498 class MVE_VADDV<string iname, string suffix, dag iops, string cstr,
499 bit A, bit U, bits<2> size, list<dag> pattern=[]>
500 : MVE_rDest<(outs tGPREven:$Rda), iops, NoItinerary,
501 iname, suffix, "$Rda, $Qm", cstr, pattern> {
506 let Inst{22-20} = 0b111;
507 let Inst{19-18} = size{1-0};
508 let Inst{17-16} = 0b01;
509 let Inst{15-13} = Rda{3-1};
511 let Inst{8-6} = 0b100;
513 let Inst{3-1} = Qm{2-0};
517 multiclass MVE_VADDV_A<string suffix, bit U, bits<2> size,
518 list<dag> pattern=[]> {
519 def acc : MVE_VADDV<"vaddva", suffix,
520 (ins tGPREven:$Rda_src, MQPR:$Qm), "$Rda = $Rda_src",
521 0b1, U, size, pattern>;
522 def no_acc : MVE_VADDV<"vaddv", suffix,
524 0b0, U, size, pattern>;
527 defm MVE_VADDVs8 : MVE_VADDV_A<"s8", 0b0, 0b00>;
528 defm MVE_VADDVs16 : MVE_VADDV_A<"s16", 0b0, 0b01>;
529 defm MVE_VADDVs32 : MVE_VADDV_A<"s32", 0b0, 0b10>;
530 defm MVE_VADDVu8 : MVE_VADDV_A<"u8", 0b1, 0b00>;
531 defm MVE_VADDVu16 : MVE_VADDV_A<"u16", 0b1, 0b01>;
532 defm MVE_VADDVu32 : MVE_VADDV_A<"u32", 0b1, 0b10>;
534 class MVE_VADDLV<string iname, string suffix, dag iops, string cstr,
535 bit A, bit U, list<dag> pattern=[]>
536 : MVE_rDest<(outs tGPREven:$RdaLo, tGPROdd:$RdaHi), iops, NoItinerary, iname,
537 suffix, "$RdaLo, $RdaHi, $Qm", cstr, pattern> {
543 let Inst{22-20} = RdaHi{3-1};
544 let Inst{19-18} = 0b10;
545 let Inst{17-16} = 0b01;
546 let Inst{15-13} = RdaLo{3-1};
548 let Inst{8-6} = 0b100;
550 let Inst{3-1} = Qm{2-0};
554 multiclass MVE_VADDLV_A<string suffix, bit U, list<dag> pattern=[]> {
555 def acc : MVE_VADDLV<"vaddlva", suffix,
556 (ins tGPREven:$RdaLo_src, tGPROdd:$RdaHi_src, MQPR:$Qm),
557 "$RdaLo = $RdaLo_src,$RdaHi = $RdaHi_src",
559 def no_acc : MVE_VADDLV<"vaddlv", suffix,
565 defm MVE_VADDLVs32 : MVE_VADDLV_A<"s32", 0b0>;
566 defm MVE_VADDLVu32 : MVE_VADDLV_A<"u32", 0b1>;
568 class MVE_VMINMAXNMV<string iname, string suffix, bit sz,
569 bit bit_17, bit bit_7, list<dag> pattern=[]>
570 : MVE_rDest<(outs rGPR:$RdaDest), (ins rGPR:$RdaSrc, MQPR:$Qm),
571 NoItinerary, iname, suffix, "$RdaSrc, $Qm",
572 "$RdaDest = $RdaSrc", pattern> {
577 let Inst{22-20} = 0b110;
578 let Inst{19-18} = 0b11;
579 let Inst{17} = bit_17;
581 let Inst{15-12} = RdaDest{3-0};
584 let Inst{6-5} = 0b00;
585 let Inst{3-1} = Qm{2-0};
588 let Predicates = [HasMVEFloat];
591 multiclass MVE_VMINMAXNMV_fty<string iname, bit bit_7, list<dag> pattern=[]> {
592 def f32 : MVE_VMINMAXNMV<iname, "f32", 0b0, 0b1, bit_7, pattern>;
593 def f16 : MVE_VMINMAXNMV<iname, "f16", 0b1, 0b1, bit_7, pattern>;
596 defm MVE_VMINNMV : MVE_VMINMAXNMV_fty<"vminnmv", 0b1>;
597 defm MVE_VMAXNMV : MVE_VMINMAXNMV_fty<"vmaxnmv", 0b0>;
599 multiclass MVE_VMINMAXNMAV_fty<string iname, bit bit_7, list<dag> pattern=[]> {
600 def f32 : MVE_VMINMAXNMV<iname, "f32", 0b0, 0b0, bit_7, pattern>;
601 def f16 : MVE_VMINMAXNMV<iname, "f16", 0b1, 0b0, bit_7, pattern>;
604 defm MVE_VMINNMAV : MVE_VMINMAXNMAV_fty<"vminnmav", 0b1>;
605 defm MVE_VMAXNMAV : MVE_VMINMAXNMAV_fty<"vmaxnmav", 0b0>;
607 class MVE_VMINMAXV<string iname, string suffix, bit U, bits<2> size,
608 bit bit_17, bit bit_7, list<dag> pattern=[]>
609 : MVE_rDest<(outs rGPR:$RdaDest), (ins rGPR:$RdaSrc, MQPR:$Qm), NoItinerary,
610 iname, suffix, "$RdaSrc, $Qm", "$RdaDest = $RdaSrc", pattern> {
615 let Inst{22-20} = 0b110;
616 let Inst{19-18} = size{1-0};
617 let Inst{17} = bit_17;
619 let Inst{15-12} = RdaDest{3-0};
622 let Inst{6-5} = 0b00;
623 let Inst{3-1} = Qm{2-0};
627 multiclass MVE_VMINMAXV_ty<string iname, bit bit_7, list<dag> pattern=[]> {
628 def s8 : MVE_VMINMAXV<iname, "s8", 0b0, 0b00, 0b1, bit_7>;
629 def s16 : MVE_VMINMAXV<iname, "s16", 0b0, 0b01, 0b1, bit_7>;
630 def s32 : MVE_VMINMAXV<iname, "s32", 0b0, 0b10, 0b1, bit_7>;
631 def u8 : MVE_VMINMAXV<iname, "u8", 0b1, 0b00, 0b1, bit_7>;
632 def u16 : MVE_VMINMAXV<iname, "u16", 0b1, 0b01, 0b1, bit_7>;
633 def u32 : MVE_VMINMAXV<iname, "u32", 0b1, 0b10, 0b1, bit_7>;
636 defm MVE_VMINV : MVE_VMINMAXV_ty<"vminv", 0b1>;
637 defm MVE_VMAXV : MVE_VMINMAXV_ty<"vmaxv", 0b0>;
639 multiclass MVE_VMINMAXAV_ty<string iname, bit bit_7, list<dag> pattern=[]> {
640 def s8 : MVE_VMINMAXV<iname, "s8", 0b0, 0b00, 0b0, bit_7>;
641 def s16 : MVE_VMINMAXV<iname, "s16", 0b0, 0b01, 0b0, bit_7>;
642 def s32 : MVE_VMINMAXV<iname, "s32", 0b0, 0b10, 0b0, bit_7>;
645 defm MVE_VMINAV : MVE_VMINMAXAV_ty<"vminav", 0b1>;
646 defm MVE_VMAXAV : MVE_VMINMAXAV_ty<"vmaxav", 0b0>;
648 class MVE_VMLAMLSDAV<string iname, string suffix, dag iops, string cstr,
649 bit sz, bit bit_28, bit A, bit X, bit bit_8, bit bit_0,
650 list<dag> pattern=[]>
651 : MVE_rDest<(outs tGPREven:$RdaDest), iops, NoItinerary, iname, suffix,
652 "$RdaDest, $Qn, $Qm", cstr, pattern> {
657 let Inst{28} = bit_28;
658 let Inst{22-20} = 0b111;
659 let Inst{19-17} = Qn{2-0};
661 let Inst{15-13} = RdaDest{3-1};
664 let Inst{7-6} = 0b00;
666 let Inst{3-1} = Qm{2-0};
670 multiclass MVE_VMLAMLSDAV_X<string iname, string suffix, dag iops, string cstr,
671 bit sz, bit bit_28, bit A, bit bit_8, bit bit_0,
672 list<dag> pattern=[]> {
673 def _noexch : MVE_VMLAMLSDAV<iname, suffix, iops, cstr, sz,
674 bit_28, A, 0b0, bit_8, bit_0, pattern>;
675 def _exch : MVE_VMLAMLSDAV<iname # "x", suffix, iops, cstr, sz,
676 bit_28, A, 0b1, bit_8, bit_0, pattern>;
679 multiclass MVE_VMLAMLSDAV_XA<string iname, string suffix, bit sz, bit bit_28,
680 bit bit_8, bit bit_0, list<dag> pattern=[]> {
681 defm _noacc : MVE_VMLAMLSDAV_X<iname, suffix, (ins MQPR:$Qn, MQPR:$Qm), "",
682 sz, bit_28, 0b0, bit_8, bit_0, pattern>;
683 defm _acc : MVE_VMLAMLSDAV_X<iname # "a", suffix,
684 (ins tGPREven:$RdaSrc, MQPR:$Qn, MQPR:$Qm),
685 "$RdaDest = $RdaSrc",
686 sz, bit_28, 0b1, bit_8, bit_0, pattern>;
689 multiclass MVE_VMLADAV_multi<string suffix, bit sz, bit U, bit bit_8,
690 list<dag> pattern=[]> {
691 defm "" : MVE_VMLAMLSDAV_XA<"vmladav", suffix, sz, U, bit_8, 0b0, pattern>;
694 defm MVE_VMLADAVs16 : MVE_VMLADAV_multi<"s16", 0b0, 0b0, 0b0>;
695 defm MVE_VMLADAVs32 : MVE_VMLADAV_multi<"s32", 0b1, 0b0, 0b0>;
696 defm MVE_VMLADAVu16 : MVE_VMLADAV_multi<"u16", 0b0, 0b1, 0b0>;
697 defm MVE_VMLADAVu32 : MVE_VMLADAV_multi<"u32", 0b1, 0b1, 0b0>;
699 defm MVE_VMLADAVs8 : MVE_VMLADAV_multi<"s8", 0b0, 0b0, 0b1>;
700 defm MVE_VMLADAVu8 : MVE_VMLADAV_multi<"u8", 0b0, 0b1, 0b1>;
702 // vmlav aliases vmladav
703 foreach acc = ["_acc", "_noacc"] in {
704 foreach suffix = ["s8", "s16", "s32", "u8", "u16", "u32"] in {
705 def : MVEInstAlias<!strconcat("vmlav", !if(!eq(acc, "_acc"), "a", ""),
706 "${vp}.", suffix, "\t$RdaDest, $Qn, $Qm"),
707 (!cast<Instruction>("MVE_VMLADAV"#suffix#acc#"_noexch")
708 tGPREven:$RdaDest, MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
712 multiclass MVE_VMLSDAV_multi<string suffix, bit sz, bit bit_28,
713 list<dag> pattern=[]> {
714 defm "" : MVE_VMLAMLSDAV_XA<"vmlsdav", suffix, sz, bit_28, 0b0, 0b1, pattern>;
717 defm MVE_VMLSDAVs8 : MVE_VMLSDAV_multi<"s8", 0, 0b1>;
718 defm MVE_VMLSDAVs16 : MVE_VMLSDAV_multi<"s16", 0, 0b0>;
719 defm MVE_VMLSDAVs32 : MVE_VMLSDAV_multi<"s32", 1, 0b0>;
721 // Base class for VMLALDAV and VMLSLDAV, VRMLALDAVH, VRMLSLDAVH
722 class MVE_VMLALDAVBase<string iname, string suffix, dag iops, string cstr,
723 bit sz, bit bit_28, bit A, bit X, bit bit_8, bit bit_0,
724 list<dag> pattern=[]>
725 : MVE_rDest<(outs tGPREven:$RdaLoDest, tGPROdd:$RdaHiDest), iops, NoItinerary,
726 iname, suffix, "$RdaLoDest, $RdaHiDest, $Qn, $Qm", cstr, pattern> {
732 let Inst{28} = bit_28;
733 let Inst{22-20} = RdaHiDest{3-1};
734 let Inst{19-17} = Qn{2-0};
736 let Inst{15-13} = RdaLoDest{3-1};
739 let Inst{7-6} = 0b00;
741 let Inst{3-1} = Qm{2-0};
745 multiclass MVE_VMLALDAVBase_X<string iname, string suffix, dag iops,
746 string cstr, bit sz, bit bit_28, bit A,
747 bit bit_8, bit bit_0, list<dag> pattern=[]> {
748 def _noexch : MVE_VMLALDAVBase<iname, suffix, iops, cstr, sz,
749 bit_28, A, 0b0, bit_8, bit_0, pattern>;
750 def _exch : MVE_VMLALDAVBase<iname # "x", suffix, iops, cstr, sz,
751 bit_28, A, 0b1, bit_8, bit_0, pattern>;
754 multiclass MVE_VMLALDAVBase_XA<string iname, string suffix, bit sz, bit bit_28,
755 bit bit_8, bit bit_0, list<dag> pattern=[]> {
756 defm _noacc : MVE_VMLALDAVBase_X<
757 iname, suffix, (ins MQPR:$Qn, MQPR:$Qm), "",
758 sz, bit_28, 0b0, bit_8, bit_0, pattern>;
759 defm _acc : MVE_VMLALDAVBase_X<
760 iname # "a", suffix, (ins tGPREven:$RdaLoSrc, tGPROdd:$RdaHiSrc,
762 "$RdaLoDest = $RdaLoSrc,$RdaHiDest = $RdaHiSrc",
763 sz, bit_28, 0b1, bit_8, bit_0, pattern>;
766 multiclass MVE_VRMLALDAVH_multi<string suffix, bit U, list<dag> pattern=[]> {
767 defm "" : MVE_VMLALDAVBase_XA<
768 "vrmlaldavh", suffix, 0b0, U, 0b1, 0b0, pattern>;
771 defm MVE_VRMLALDAVHs32 : MVE_VRMLALDAVH_multi<"s32", 0>;
772 defm MVE_VRMLALDAVHu32 : MVE_VRMLALDAVH_multi<"u32", 1>;
774 // vrmlalvh aliases for vrmlaldavh
775 def : MVEInstAlias<"vrmlalvh${vp}.s32\t$RdaLo, $RdaHi, $Qn, $Qm",
776 (MVE_VRMLALDAVHs32_noacc_noexch
777 tGPREven:$RdaLo, tGPROdd:$RdaHi,
778 MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
779 def : MVEInstAlias<"vrmlalvha${vp}.s32\t$RdaLo, $RdaHi, $Qn, $Qm",
780 (MVE_VRMLALDAVHs32_acc_noexch
781 tGPREven:$RdaLo, tGPROdd:$RdaHi,
782 MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
783 def : MVEInstAlias<"vrmlalvh${vp}.u32\t$RdaLo, $RdaHi, $Qn, $Qm",
784 (MVE_VRMLALDAVHu32_noacc_noexch
785 tGPREven:$RdaLo, tGPROdd:$RdaHi,
786 MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
787 def : MVEInstAlias<"vrmlalvha${vp}.u32\t$RdaLo, $RdaHi, $Qn, $Qm",
788 (MVE_VRMLALDAVHu32_acc_noexch
789 tGPREven:$RdaLo, tGPROdd:$RdaHi,
790 MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
792 multiclass MVE_VMLALDAV_multi<string suffix, bit sz, bit U,
793 list<dag> pattern=[]> {
794 defm "" : MVE_VMLALDAVBase_XA<"vmlaldav", suffix, sz, U, 0b0, 0b0, pattern>;
797 defm MVE_VMLALDAVs16 : MVE_VMLALDAV_multi<"s16", 0b0, 0b0>;
798 defm MVE_VMLALDAVs32 : MVE_VMLALDAV_multi<"s32", 0b1, 0b0>;
799 defm MVE_VMLALDAVu16 : MVE_VMLALDAV_multi<"u16", 0b0, 0b1>;
800 defm MVE_VMLALDAVu32 : MVE_VMLALDAV_multi<"u32", 0b1, 0b1>;
802 // vmlalv aliases vmlaldav
803 foreach acc = ["_acc", "_noacc"] in {
804 foreach suffix = ["s16", "s32", "u16", "u32"] in {
805 def : MVEInstAlias<!strconcat("vmlalv", !if(!eq(acc, "_acc"), "a", ""),
806 "${vp}.", suffix, "\t$RdaLoDest, $RdaHiDest, $Qn, $Qm"),
807 (!cast<Instruction>("MVE_VMLALDAV"#suffix#acc#"_noexch")
808 tGPREven:$RdaLoDest, tGPROdd:$RdaHiDest,
809 MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
813 multiclass MVE_VMLSLDAV_multi<string iname, string suffix, bit sz,
814 bit bit_28, list<dag> pattern=[]> {
815 defm "" : MVE_VMLALDAVBase_XA<iname, suffix, sz, bit_28, 0b0, 0b1, pattern>;
818 defm MVE_VMLSLDAVs16 : MVE_VMLSLDAV_multi<"vmlsldav", "s16", 0b0, 0b0>;
819 defm MVE_VMLSLDAVs32 : MVE_VMLSLDAV_multi<"vmlsldav", "s32", 0b1, 0b0>;
820 defm MVE_VRMLSLDAVHs32 : MVE_VMLSLDAV_multi<"vrmlsldavh", "s32", 0b0, 0b1>;
822 // end of mve_rDest instructions
824 // start of mve_comp instructions
826 class MVE_comp<InstrItinClass itin, string iname, string suffix,
827 string cstr, list<dag> pattern=[]>
828 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm), itin, iname, suffix,
829 "$Qd, $Qn, $Qm", vpred_r, cstr, pattern> {
834 let Inst{22} = Qd{3};
835 let Inst{19-17} = Qn{2-0};
837 let Inst{15-13} = Qd{2-0};
839 let Inst{10-9} = 0b11;
842 let Inst{3-1} = Qm{2-0};
846 class MVE_VMINMAXNM<string iname, string suffix, bit sz, bit bit_21,
847 list<dag> pattern=[]>
848 : MVE_comp<NoItinerary, iname, suffix, "", pattern> {
851 let Inst{25-24} = 0b11;
853 let Inst{21} = bit_21;
860 let Predicates = [HasMVEFloat];
863 def MVE_VMAXNMf32 : MVE_VMINMAXNM<"vmaxnm", "f32", 0b0, 0b0>;
864 def MVE_VMAXNMf16 : MVE_VMINMAXNM<"vmaxnm", "f16", 0b1, 0b0>;
866 let Predicates = [HasMVEFloat] in {
867 def : Pat<(v4f32 (fmaxnum (v4f32 MQPR:$val1), (v4f32 MQPR:$val2))),
868 (v4f32 (MVE_VMAXNMf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2)))>;
869 def : Pat<(v8f16 (fmaxnum (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
870 (v8f16 (MVE_VMAXNMf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
873 def MVE_VMINNMf32 : MVE_VMINMAXNM<"vminnm", "f32", 0b0, 0b1>;
874 def MVE_VMINNMf16 : MVE_VMINMAXNM<"vminnm", "f16", 0b1, 0b1>;
876 let Predicates = [HasMVEFloat] in {
877 def : Pat<(v4f32 (fminnum (v4f32 MQPR:$val1), (v4f32 MQPR:$val2))),
878 (v4f32 (MVE_VMINNMf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2)))>;
879 def : Pat<(v8f16 (fminnum (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
880 (v8f16 (MVE_VMINNMf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
884 class MVE_VMINMAX<string iname, string suffix, bit U, bits<2> size,
885 bit bit_4, list<dag> pattern=[]>
886 : MVE_comp<NoItinerary, iname, suffix, "", pattern> {
889 let Inst{25-24} = 0b11;
891 let Inst{21-20} = size{1-0};
898 multiclass MVE_VMINMAX_all_sizes<string iname, bit bit_4> {
899 def s8 : MVE_VMINMAX<iname, "s8", 0b0, 0b00, bit_4>;
900 def s16 : MVE_VMINMAX<iname, "s16", 0b0, 0b01, bit_4>;
901 def s32 : MVE_VMINMAX<iname, "s32", 0b0, 0b10, bit_4>;
902 def u8 : MVE_VMINMAX<iname, "u8", 0b1, 0b00, bit_4>;
903 def u16 : MVE_VMINMAX<iname, "u16", 0b1, 0b01, bit_4>;
904 def u32 : MVE_VMINMAX<iname, "u32", 0b1, 0b10, bit_4>;
907 defm MVE_VMAX : MVE_VMINMAX_all_sizes<"vmax", 0b0>;
908 defm MVE_VMIN : MVE_VMINMAX_all_sizes<"vmin", 0b1>;
910 let Predicates = [HasMVEInt] in {
911 def : Pat<(v16i8 (smin (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
912 (v16i8 (MVE_VMINs8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
913 def : Pat<(v8i16 (smin (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
914 (v8i16 (MVE_VMINs16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
915 def : Pat<(v4i32 (smin (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
916 (v4i32 (MVE_VMINs32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
918 def : Pat<(v16i8 (smax (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
919 (v16i8 (MVE_VMAXs8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
920 def : Pat<(v8i16 (smax (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
921 (v8i16 (MVE_VMAXs16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
922 def : Pat<(v4i32 (smax (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
923 (v4i32 (MVE_VMAXs32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
925 def : Pat<(v16i8 (umin (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
926 (v16i8 (MVE_VMINu8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
927 def : Pat<(v8i16 (umin (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
928 (v8i16 (MVE_VMINu16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
929 def : Pat<(v4i32 (umin (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
930 (v4i32 (MVE_VMINu32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
932 def : Pat<(v16i8 (umax (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
933 (v16i8 (MVE_VMAXu8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
934 def : Pat<(v8i16 (umax (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
935 (v8i16 (MVE_VMAXu16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
936 def : Pat<(v4i32 (umax (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
937 (v4i32 (MVE_VMAXu32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
940 // end of mve_comp instructions
942 // start of mve_imm_shift instructions
944 def MVE_VSHLC : MVE_p<(outs rGPR:$RdmDest, MQPR:$Qd),
945 (ins MQPR:$QdSrc, rGPR:$RdmSrc, long_shift:$imm),
946 NoItinerary, "vshlc", "", "$QdSrc, $RdmSrc, $imm",
947 vpred_n, "$RdmDest = $RdmSrc,$Qd = $QdSrc"> {
953 let Inst{25-23} = 0b101;
954 let Inst{22} = Qd{3};
956 let Inst{20-16} = imm{4-0};
957 let Inst{15-13} = Qd{2-0};
958 let Inst{12-4} = 0b011111100;
959 let Inst{3-0} = RdmDest{3-0};
962 class MVE_shift_imm<dag oops, dag iops, string iname, string suffix,
963 string ops, vpred_ops vpred, string cstr,
964 list<dag> pattern=[]>
965 : MVE_p<oops, iops, NoItinerary, iname, suffix, ops, vpred, cstr, pattern> {
969 let Inst{22} = Qd{3};
970 let Inst{15-13} = Qd{2-0};
972 let Inst{3-1} = Qm{2-0};
975 class MVE_VMOVL<string iname, string suffix, bits<2> sz, bit U,
976 list<dag> pattern=[]>
977 : MVE_shift_imm<(outs MQPR:$Qd), (ins MQPR:$Qm),
978 iname, suffix, "$Qd, $Qm", vpred_r, "",
981 let Inst{25-23} = 0b101;
983 let Inst{20-19} = sz{1-0};
984 let Inst{18-16} = 0b000;
985 let Inst{11-6} = 0b111101;
990 multiclass MVE_VMOVL_shift_half<string iname, string suffix, bits<2> sz, bit U,
991 list<dag> pattern=[]> {
992 def bh : MVE_VMOVL<!strconcat(iname, "b"), suffix, sz, U, pattern> {
995 def th : MVE_VMOVL<!strconcat(iname, "t"), suffix, sz, U, pattern> {
1000 defm MVE_VMOVLs8 : MVE_VMOVL_shift_half<"vmovl", "s8", 0b01, 0b0>;
1001 defm MVE_VMOVLu8 : MVE_VMOVL_shift_half<"vmovl", "u8", 0b01, 0b1>;
1002 defm MVE_VMOVLs16 : MVE_VMOVL_shift_half<"vmovl", "s16", 0b10, 0b0>;
1003 defm MVE_VMOVLu16 : MVE_VMOVL_shift_half<"vmovl", "u16", 0b10, 0b1>;
1005 class MVE_VSHLL_imm<string iname, string suffix, bit U, bit th,
1006 dag immops, list<dag> pattern=[]>
1007 : MVE_shift_imm<(outs MQPR:$Qd), !con((ins MQPR:$Qm), immops),
1008 iname, suffix, "$Qd, $Qm, $imm", vpred_r, "", pattern> {
1010 let Inst{25-23} = 0b101;
1013 let Inst{11-6} = 0b111101;
1018 // The immediate VSHLL instructions accept shift counts from 1 up to
1019 // the lane width (8 or 16), but the full-width shifts have an
1020 // entirely separate encoding, given below with 'lw' in the name.
1022 class MVE_VSHLL_imm8<string iname, string suffix,
1023 bit U, bit th, list<dag> pattern=[]>
1024 : MVE_VSHLL_imm<iname, suffix, U, th, (ins mve_shift_imm1_7:$imm), pattern> {
1026 let Inst{20-19} = 0b01;
1027 let Inst{18-16} = imm;
1030 class MVE_VSHLL_imm16<string iname, string suffix,
1031 bit U, bit th, list<dag> pattern=[]>
1032 : MVE_VSHLL_imm<iname, suffix, U, th, (ins mve_shift_imm1_15:$imm), pattern> {
1035 let Inst{19-16} = imm;
1038 def MVE_VSHLL_imms8bh : MVE_VSHLL_imm8 <"vshllb", "s8", 0b0, 0b0>;
1039 def MVE_VSHLL_imms8th : MVE_VSHLL_imm8 <"vshllt", "s8", 0b0, 0b1>;
1040 def MVE_VSHLL_immu8bh : MVE_VSHLL_imm8 <"vshllb", "u8", 0b1, 0b0>;
1041 def MVE_VSHLL_immu8th : MVE_VSHLL_imm8 <"vshllt", "u8", 0b1, 0b1>;
1042 def MVE_VSHLL_imms16bh : MVE_VSHLL_imm16<"vshllb", "s16", 0b0, 0b0>;
1043 def MVE_VSHLL_imms16th : MVE_VSHLL_imm16<"vshllt", "s16", 0b0, 0b1>;
1044 def MVE_VSHLL_immu16bh : MVE_VSHLL_imm16<"vshllb", "u16", 0b1, 0b0>;
1045 def MVE_VSHLL_immu16th : MVE_VSHLL_imm16<"vshllt", "u16", 0b1, 0b1>;
1047 class MVE_VSHLL_by_lane_width<string iname, string suffix, bits<2> size,
1048 bit U, string ops, list<dag> pattern=[]>
1049 : MVE_shift_imm<(outs MQPR:$Qd), (ins MQPR:$Qm),
1050 iname, suffix, ops, vpred_r, "", pattern> {
1052 let Inst{25-23} = 0b100;
1053 let Inst{21-20} = 0b11;
1054 let Inst{19-18} = size{1-0};
1055 let Inst{17-16} = 0b01;
1056 let Inst{11-6} = 0b111000;
1061 multiclass MVE_VSHLL_lw<string iname, string suffix, bits<2> sz, bit U,
1062 string ops, list<dag> pattern=[]> {
1063 def bh : MVE_VSHLL_by_lane_width<iname#"b", suffix, sz, U, ops, pattern> {
1066 def th : MVE_VSHLL_by_lane_width<iname#"t", suffix, sz, U, ops, pattern> {
1071 defm MVE_VSHLL_lws8 : MVE_VSHLL_lw<"vshll", "s8", 0b00, 0b0, "$Qd, $Qm, #8">;
1072 defm MVE_VSHLL_lws16 : MVE_VSHLL_lw<"vshll", "s16", 0b01, 0b0, "$Qd, $Qm, #16">;
1073 defm MVE_VSHLL_lwu8 : MVE_VSHLL_lw<"vshll", "u8", 0b00, 0b1, "$Qd, $Qm, #8">;
1074 defm MVE_VSHLL_lwu16 : MVE_VSHLL_lw<"vshll", "u16", 0b01, 0b1, "$Qd, $Qm, #16">;
1076 class MVE_VxSHRN<string iname, string suffix, bit bit_12, bit bit_28,
1077 dag immops, list<dag> pattern=[]>
1078 : MVE_shift_imm<(outs MQPR:$Qd), !con((ins MQPR:$QdSrc, MQPR:$Qm), immops),
1079 iname, suffix, "$Qd, $Qm, $imm", vpred_n, "$Qd = $QdSrc",
1083 let Inst{28} = bit_28;
1084 let Inst{25-23} = 0b101;
1086 let Inst{20-16} = imm{4-0};
1087 let Inst{12} = bit_12;
1088 let Inst{11-6} = 0b111111;
1093 def MVE_VRSHRNi16bh : MVE_VxSHRN<
1094 "vrshrnb", "i16", 0b0, 0b1, (ins shr_imm8:$imm)> {
1095 let Inst{20-19} = 0b01;
1097 def MVE_VRSHRNi16th : MVE_VxSHRN<
1098 "vrshrnt", "i16", 0b1, 0b1,(ins shr_imm8:$imm)> {
1099 let Inst{20-19} = 0b01;
1101 def MVE_VRSHRNi32bh : MVE_VxSHRN<
1102 "vrshrnb", "i32", 0b0, 0b1, (ins shr_imm16:$imm)> {
1105 def MVE_VRSHRNi32th : MVE_VxSHRN<
1106 "vrshrnt", "i32", 0b1, 0b1, (ins shr_imm16:$imm)> {
1110 def MVE_VSHRNi16bh : MVE_VxSHRN<
1111 "vshrnb", "i16", 0b0, 0b0, (ins shr_imm8:$imm)> {
1112 let Inst{20-19} = 0b01;
1114 def MVE_VSHRNi16th : MVE_VxSHRN<
1115 "vshrnt", "i16", 0b1, 0b0, (ins shr_imm8:$imm)> {
1116 let Inst{20-19} = 0b01;
1118 def MVE_VSHRNi32bh : MVE_VxSHRN<
1119 "vshrnb", "i32", 0b0, 0b0, (ins shr_imm16:$imm)> {
1122 def MVE_VSHRNi32th : MVE_VxSHRN<
1123 "vshrnt", "i32", 0b1, 0b0, (ins shr_imm16:$imm)> {
1127 class MVE_VxQRSHRUN<string iname, string suffix, bit bit_28, bit bit_12, dag immops,
1128 list<dag> pattern=[]>
1129 : MVE_shift_imm<(outs MQPR:$Qd), !con((ins MQPR:$QdSrc, MQPR:$Qm), immops),
1130 iname, suffix, "$Qd, $Qm, $imm", vpred_n, "$Qd = $QdSrc",
1134 let Inst{28} = bit_28;
1135 let Inst{25-23} = 0b101;
1137 let Inst{20-16} = imm{4-0};
1138 let Inst{12} = bit_12;
1139 let Inst{11-6} = 0b111111;
1144 def MVE_VQRSHRUNs16bh : MVE_VxQRSHRUN<
1145 "vqrshrunb", "s16", 0b1, 0b0, (ins shr_imm8:$imm)> {
1146 let Inst{20-19} = 0b01;
1148 def MVE_VQRSHRUNs16th : MVE_VxQRSHRUN<
1149 "vqrshrunt", "s16", 0b1, 0b1, (ins shr_imm8:$imm)> {
1150 let Inst{20-19} = 0b01;
1152 def MVE_VQRSHRUNs32bh : MVE_VxQRSHRUN<
1153 "vqrshrunb", "s32", 0b1, 0b0, (ins shr_imm16:$imm)> {
1156 def MVE_VQRSHRUNs32th : MVE_VxQRSHRUN<
1157 "vqrshrunt", "s32", 0b1, 0b1, (ins shr_imm16:$imm)> {
1161 def MVE_VQSHRUNs16bh : MVE_VxQRSHRUN<
1162 "vqshrunb", "s16", 0b0, 0b0, (ins shr_imm8:$imm)> {
1163 let Inst{20-19} = 0b01;
1165 def MVE_VQSHRUNs16th : MVE_VxQRSHRUN<
1166 "vqshrunt", "s16", 0b0, 0b1, (ins shr_imm8:$imm)> {
1167 let Inst{20-19} = 0b01;
1169 def MVE_VQSHRUNs32bh : MVE_VxQRSHRUN<
1170 "vqshrunb", "s32", 0b0, 0b0, (ins shr_imm16:$imm)> {
1173 def MVE_VQSHRUNs32th : MVE_VxQRSHRUN<
1174 "vqshrunt", "s32", 0b0, 0b1, (ins shr_imm16:$imm)> {
1178 class MVE_VxQRSHRN<string iname, string suffix, bit bit_0, bit bit_12,
1179 dag immops, list<dag> pattern=[]>
1180 : MVE_shift_imm<(outs MQPR:$Qd), !con((ins MQPR:$QdSrc, MQPR:$Qm), immops),
1181 iname, suffix, "$Qd, $Qm, $imm", vpred_n, "$Qd = $QdSrc",
1185 let Inst{25-23} = 0b101;
1187 let Inst{20-16} = imm{4-0};
1188 let Inst{12} = bit_12;
1189 let Inst{11-6} = 0b111101;
1191 let Inst{0} = bit_0;
1194 multiclass MVE_VxQRSHRN_types<string iname, bit bit_0, bit bit_12> {
1195 def s16 : MVE_VxQRSHRN<iname, "s16", bit_0, bit_12, (ins shr_imm8:$imm)> {
1197 let Inst{20-19} = 0b01;
1199 def u16 : MVE_VxQRSHRN<iname, "u16", bit_0, bit_12, (ins shr_imm8:$imm)> {
1201 let Inst{20-19} = 0b01;
1203 def s32 : MVE_VxQRSHRN<iname, "s32", bit_0, bit_12, (ins shr_imm16:$imm)> {
1207 def u32 : MVE_VxQRSHRN<iname, "u32", bit_0, bit_12, (ins shr_imm16:$imm)> {
1213 defm MVE_VQRSHRNbh : MVE_VxQRSHRN_types<"vqrshrnb", 0b1, 0b0>;
1214 defm MVE_VQRSHRNth : MVE_VxQRSHRN_types<"vqrshrnt", 0b1, 0b1>;
1215 defm MVE_VQSHRNbh : MVE_VxQRSHRN_types<"vqshrnb", 0b0, 0b0>;
1216 defm MVE_VQSHRNth : MVE_VxQRSHRN_types<"vqshrnt", 0b0, 0b1>;
1218 // end of mve_imm_shift instructions
1220 // start of mve_shift instructions
1222 class MVE_shift_by_vec<string iname, string suffix, bit U,
1223 bits<2> size, bit bit_4, bit bit_8>
1224 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qm, MQPR:$Qn), NoItinerary,
1225 iname, suffix, "$Qd, $Qm, $Qn", vpred_r, "", []> {
1226 // Shift instructions which take a vector of shift counts
1232 let Inst{25-24} = 0b11;
1234 let Inst{22} = Qd{3};
1235 let Inst{21-20} = size;
1236 let Inst{19-17} = Qn{2-0};
1238 let Inst{15-13} = Qd{2-0};
1239 let Inst{12-9} = 0b0010;
1240 let Inst{8} = bit_8;
1241 let Inst{7} = Qn{3};
1243 let Inst{5} = Qm{3};
1244 let Inst{4} = bit_4;
1245 let Inst{3-1} = Qm{2-0};
1249 multiclass mve_shift_by_vec_multi<string iname, bit bit_4, bit bit_8> {
1250 def s8 : MVE_shift_by_vec<iname, "s8", 0b0, 0b00, bit_4, bit_8>;
1251 def s16 : MVE_shift_by_vec<iname, "s16", 0b0, 0b01, bit_4, bit_8>;
1252 def s32 : MVE_shift_by_vec<iname, "s32", 0b0, 0b10, bit_4, bit_8>;
1253 def u8 : MVE_shift_by_vec<iname, "u8", 0b1, 0b00, bit_4, bit_8>;
1254 def u16 : MVE_shift_by_vec<iname, "u16", 0b1, 0b01, bit_4, bit_8>;
1255 def u32 : MVE_shift_by_vec<iname, "u32", 0b1, 0b10, bit_4, bit_8>;
1258 defm MVE_VSHL_by_vec : mve_shift_by_vec_multi<"vshl", 0b0, 0b0>;
1259 defm MVE_VQSHL_by_vec : mve_shift_by_vec_multi<"vqshl", 0b1, 0b0>;
1260 defm MVE_VQRSHL_by_vec : mve_shift_by_vec_multi<"vqrshl", 0b1, 0b1>;
1261 defm MVE_VRSHL_by_vec : mve_shift_by_vec_multi<"vrshl", 0b0, 0b1>;
1263 class MVE_shift_with_imm<string iname, string suffix, dag oops, dag iops,
1264 string ops, vpred_ops vpred, string cstr,
1265 list<dag> pattern=[]>
1266 : MVE_p<oops, iops, NoItinerary, iname, suffix, ops, vpred, cstr, pattern> {
1271 let Inst{22} = Qd{3};
1272 let Inst{15-13} = Qd{2-0};
1273 let Inst{12-11} = 0b00;
1274 let Inst{7-6} = 0b01;
1275 let Inst{5} = Qm{3};
1277 let Inst{3-1} = Qm{2-0};
1281 class MVE_VSxI_imm<string iname, string suffix, bit bit_8, dag imm>
1282 : MVE_shift_with_imm<iname, suffix, (outs MQPR:$Qd),
1283 !con((ins MQPR:$Qd_src, MQPR:$Qm), imm),
1284 "$Qd, $Qm, $imm", vpred_n, "$Qd = $Qd_src"> {
1287 let Inst{25-24} = 0b11;
1288 let Inst{21-16} = imm;
1289 let Inst{10-9} = 0b10;
1290 let Inst{8} = bit_8;
1293 def MVE_VSRIimm8 : MVE_VSxI_imm<"vsri", "8", 0b0, (ins shr_imm8:$imm)> {
1294 let Inst{21-19} = 0b001;
1297 def MVE_VSRIimm16 : MVE_VSxI_imm<"vsri", "16", 0b0, (ins shr_imm16:$imm)> {
1298 let Inst{21-20} = 0b01;
1301 def MVE_VSRIimm32 : MVE_VSxI_imm<"vsri", "32", 0b0, (ins shr_imm32:$imm)> {
1305 def MVE_VSLIimm8 : MVE_VSxI_imm<"vsli", "8", 0b1, (ins imm0_7:$imm)> {
1306 let Inst{21-19} = 0b001;
1309 def MVE_VSLIimm16 : MVE_VSxI_imm<"vsli", "16", 0b1, (ins imm0_15:$imm)> {
1310 let Inst{21-20} = 0b01;
1313 def MVE_VSLIimm32 : MVE_VSxI_imm<"vsli", "32", 0b1,(ins imm0_31:$imm)> {
1317 class MVE_VQSHL_imm<string suffix, dag imm>
1318 : MVE_shift_with_imm<"vqshl", suffix, (outs MQPR:$Qd),
1319 !con((ins MQPR:$Qm), imm), "$Qd, $Qm, $imm",
1323 let Inst{25-24} = 0b11;
1324 let Inst{21-16} = imm;
1325 let Inst{10-8} = 0b111;
1328 def MVE_VSLIimms8 : MVE_VQSHL_imm<"s8", (ins imm0_7:$imm)> {
1330 let Inst{21-19} = 0b001;
1333 def MVE_VSLIimmu8 : MVE_VQSHL_imm<"u8", (ins imm0_7:$imm)> {
1335 let Inst{21-19} = 0b001;
1338 def MVE_VSLIimms16 : MVE_VQSHL_imm<"s16", (ins imm0_15:$imm)> {
1340 let Inst{21-20} = 0b01;
1343 def MVE_VSLIimmu16 : MVE_VQSHL_imm<"u16", (ins imm0_15:$imm)> {
1345 let Inst{21-20} = 0b01;
1348 def MVE_VSLIimms32 : MVE_VQSHL_imm<"s32", (ins imm0_31:$imm)> {
1353 def MVE_VSLIimmu32 : MVE_VQSHL_imm<"u32", (ins imm0_31:$imm)> {
1358 class MVE_VQSHLU_imm<string suffix, dag imm>
1359 : MVE_shift_with_imm<"vqshlu", suffix, (outs MQPR:$Qd),
1360 !con((ins MQPR:$Qm), imm), "$Qd, $Qm, $imm",
1365 let Inst{25-24} = 0b11;
1366 let Inst{21-16} = imm;
1367 let Inst{10-8} = 0b110;
1370 def MVE_VQSHLU_imms8 : MVE_VQSHLU_imm<"s8", (ins imm0_7:$imm)> {
1371 let Inst{21-19} = 0b001;
1374 def MVE_VQSHLU_imms16 : MVE_VQSHLU_imm<"s16", (ins imm0_15:$imm)> {
1375 let Inst{21-20} = 0b01;
1378 def MVE_VQSHLU_imms32 : MVE_VQSHLU_imm<"s32", (ins imm0_31:$imm)> {
1382 class MVE_VRSHR_imm<string suffix, dag imm>
1383 : MVE_shift_with_imm<"vrshr", suffix, (outs MQPR:$Qd),
1384 !con((ins MQPR:$Qm), imm), "$Qd, $Qm, $imm",
1388 let Inst{25-24} = 0b11;
1389 let Inst{21-16} = imm;
1390 let Inst{10-8} = 0b010;
1393 def MVE_VRSHR_imms8 : MVE_VRSHR_imm<"s8", (ins shr_imm8:$imm)> {
1395 let Inst{21-19} = 0b001;
1398 def MVE_VRSHR_immu8 : MVE_VRSHR_imm<"u8", (ins shr_imm8:$imm)> {
1400 let Inst{21-19} = 0b001;
1403 def MVE_VRSHR_imms16 : MVE_VRSHR_imm<"s16", (ins shr_imm16:$imm)> {
1405 let Inst{21-20} = 0b01;
1408 def MVE_VRSHR_immu16 : MVE_VRSHR_imm<"u16", (ins shr_imm16:$imm)> {
1410 let Inst{21-20} = 0b01;
1413 def MVE_VRSHR_imms32 : MVE_VRSHR_imm<"s32", (ins shr_imm32:$imm)> {
1418 def MVE_VRSHR_immu32 : MVE_VRSHR_imm<"u32", (ins shr_imm32:$imm)> {
1423 class MVE_VSHR_imm<string suffix, dag imm>
1424 : MVE_shift_with_imm<"vshr", suffix, (outs MQPR:$Qd),
1425 !con((ins MQPR:$Qm), imm), "$Qd, $Qm, $imm",
1429 let Inst{25-24} = 0b11;
1430 let Inst{21-16} = imm;
1431 let Inst{10-8} = 0b000;
1434 def MVE_VSHR_imms8 : MVE_VSHR_imm<"s8", (ins shr_imm8:$imm)> {
1436 let Inst{21-19} = 0b001;
1439 def MVE_VSHR_immu8 : MVE_VSHR_imm<"u8", (ins shr_imm8:$imm)> {
1441 let Inst{21-19} = 0b001;
1444 def MVE_VSHR_imms16 : MVE_VSHR_imm<"s16", (ins shr_imm16:$imm)> {
1446 let Inst{21-20} = 0b01;
1449 def MVE_VSHR_immu16 : MVE_VSHR_imm<"u16", (ins shr_imm16:$imm)> {
1451 let Inst{21-20} = 0b01;
1454 def MVE_VSHR_imms32 : MVE_VSHR_imm<"s32", (ins shr_imm32:$imm)> {
1459 def MVE_VSHR_immu32 : MVE_VSHR_imm<"u32", (ins shr_imm32:$imm)> {
1464 class MVE_VSHL_imm<string suffix, dag imm>
1465 : MVE_shift_with_imm<"vshl", suffix, (outs MQPR:$Qd),
1466 !con((ins MQPR:$Qm), imm), "$Qd, $Qm, $imm",
1471 let Inst{25-24} = 0b11;
1472 let Inst{21-16} = imm;
1473 let Inst{10-8} = 0b101;
1476 def MVE_VSHL_immi8 : MVE_VSHL_imm<"i8", (ins imm0_7:$imm)> {
1477 let Inst{21-19} = 0b001;
1480 def MVE_VSHL_immi16 : MVE_VSHL_imm<"i16", (ins imm0_15:$imm)> {
1481 let Inst{21-20} = 0b01;
1484 def MVE_VSHL_immi32 : MVE_VSHL_imm<"i32", (ins imm0_31:$imm)> {
1487 // end of mve_shift instructions
1489 // start of mve_bit instructions
1491 class MVE_bit_arith<dag oops, dag iops, string iname, string suffix,
1492 string ops, string cstr, list<dag> pattern=[]>
1493 : MVE_p<oops, iops, NoItinerary, iname, suffix, ops, vpred_r, cstr, pattern> {
1497 let Inst{22} = Qd{3};
1498 let Inst{15-13} = Qd{2-0};
1499 let Inst{5} = Qm{3};
1500 let Inst{3-1} = Qm{2-0};
1503 def MVE_VBIC : MVE_bit_arith<(outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm),
1504 "vbic", "", "$Qd, $Qn, $Qm", ""> {
1508 let Inst{25-23} = 0b110;
1509 let Inst{21-20} = 0b01;
1510 let Inst{19-17} = Qn{2-0};
1512 let Inst{12-8} = 0b00001;
1513 let Inst{7} = Qn{3};
1519 class MVE_VREV<string iname, string suffix, bits<2> size, bits<2> bit_8_7>
1520 : MVE_bit_arith<(outs MQPR:$Qd), (ins MQPR:$Qm), iname,
1521 suffix, "$Qd, $Qm", ""> {
1524 let Inst{25-23} = 0b111;
1525 let Inst{21-20} = 0b11;
1526 let Inst{19-18} = size;
1527 let Inst{17-16} = 0b00;
1528 let Inst{12-9} = 0b0000;
1529 let Inst{8-7} = bit_8_7;
1535 def MVE_VREV64_8 : MVE_VREV<"vrev64", "8", 0b00, 0b00>;
1536 def MVE_VREV64_16 : MVE_VREV<"vrev64", "16", 0b01, 0b00>;
1537 def MVE_VREV64_32 : MVE_VREV<"vrev64", "32", 0b10, 0b00>;
1539 def MVE_VREV32_8 : MVE_VREV<"vrev32", "8", 0b00, 0b01>;
1540 def MVE_VREV32_16 : MVE_VREV<"vrev32", "16", 0b01, 0b01>;
1542 def MVE_VREV16_8 : MVE_VREV<"vrev16", "8", 0b00, 0b10>;
1544 let Predicates = [HasMVEInt] in {
1545 def : Pat<(v4i32 (ARMvrev64 (v4i32 MQPR:$src))),
1546 (v4i32 (MVE_VREV64_32 (v4i32 MQPR:$src)))>;
1547 def : Pat<(v8i16 (ARMvrev64 (v8i16 MQPR:$src))),
1548 (v8i16 (MVE_VREV64_16 (v8i16 MQPR:$src)))>;
1549 def : Pat<(v16i8 (ARMvrev64 (v16i8 MQPR:$src))),
1550 (v16i8 (MVE_VREV64_8 (v16i8 MQPR:$src)))>;
1552 def : Pat<(v8i16 (ARMvrev32 (v8i16 MQPR:$src))),
1553 (v8i16 (MVE_VREV32_16 (v8i16 MQPR:$src)))>;
1554 def : Pat<(v16i8 (ARMvrev32 (v16i8 MQPR:$src))),
1555 (v16i8 (MVE_VREV32_8 (v16i8 MQPR:$src)))>;
1557 def : Pat<(v16i8 (ARMvrev16 (v16i8 MQPR:$src))),
1558 (v16i8 (MVE_VREV16_8 (v16i8 MQPR:$src)))>;
1560 def : Pat<(v4f32 (ARMvrev64 (v4f32 MQPR:$src))),
1561 (v4f32 (MVE_VREV64_32 (v4f32 MQPR:$src)))>;
1562 def : Pat<(v8f16 (ARMvrev64 (v8f16 MQPR:$src))),
1563 (v8f16 (MVE_VREV64_16 (v8f16 MQPR:$src)))>;
1564 def : Pat<(v8f16 (ARMvrev32 (v8f16 MQPR:$src))),
1565 (v8f16 (MVE_VREV32_16 (v8f16 MQPR:$src)))>;
1568 def MVE_VMVN : MVE_bit_arith<(outs MQPR:$Qd), (ins MQPR:$Qm),
1569 "vmvn", "", "$Qd, $Qm", ""> {
1571 let Inst{25-23} = 0b111;
1572 let Inst{21-16} = 0b110000;
1573 let Inst{12-6} = 0b0010111;
1578 let Predicates = [HasMVEInt] in {
1579 def : Pat<(v16i8 (vnotq (v16i8 MQPR:$val1))),
1580 (v16i8 (MVE_VMVN (v16i8 MQPR:$val1)))>;
1581 def : Pat<(v8i16 (vnotq (v8i16 MQPR:$val1))),
1582 (v8i16 (MVE_VMVN (v8i16 MQPR:$val1)))>;
1583 def : Pat<(v4i32 (vnotq (v4i32 MQPR:$val1))),
1584 (v4i32 (MVE_VMVN (v4i32 MQPR:$val1)))>;
1587 class MVE_bit_ops<string iname, bits<2> bit_21_20, bit bit_28>
1588 : MVE_bit_arith<(outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm),
1589 iname, "", "$Qd, $Qn, $Qm", ""> {
1592 let Inst{28} = bit_28;
1593 let Inst{25-23} = 0b110;
1594 let Inst{21-20} = bit_21_20;
1595 let Inst{19-17} = Qn{2-0};
1597 let Inst{12-8} = 0b00001;
1598 let Inst{7} = Qn{3};
1604 def MVE_VEOR : MVE_bit_ops<"veor", 0b00, 0b1>;
1605 def MVE_VORN : MVE_bit_ops<"vorn", 0b11, 0b0>;
1606 def MVE_VORR : MVE_bit_ops<"vorr", 0b10, 0b0>;
1607 def MVE_VAND : MVE_bit_ops<"vand", 0b00, 0b0>;
1609 // add ignored suffixes as aliases
1611 foreach s=["s8", "s16", "s32", "u8", "u16", "u32", "i8", "i16", "i32", "f16", "f32"] in {
1612 def : MVEInstAlias<"vbic${vp}." # s # "\t$QdSrc, $QnSrc, $QmSrc",
1613 (MVE_VBIC MQPR:$QdSrc, MQPR:$QnSrc, MQPR:$QmSrc, vpred_r:$vp)>;
1614 def : MVEInstAlias<"veor${vp}." # s # "\t$QdSrc, $QnSrc, $QmSrc",
1615 (MVE_VEOR MQPR:$QdSrc, MQPR:$QnSrc, MQPR:$QmSrc, vpred_r:$vp)>;
1616 def : MVEInstAlias<"vorn${vp}." # s # "\t$QdSrc, $QnSrc, $QmSrc",
1617 (MVE_VORN MQPR:$QdSrc, MQPR:$QnSrc, MQPR:$QmSrc, vpred_r:$vp)>;
1618 def : MVEInstAlias<"vorr${vp}." # s # "\t$QdSrc, $QnSrc, $QmSrc",
1619 (MVE_VORR MQPR:$QdSrc, MQPR:$QnSrc, MQPR:$QmSrc, vpred_r:$vp)>;
1620 def : MVEInstAlias<"vand${vp}." # s # "\t$QdSrc, $QnSrc, $QmSrc",
1621 (MVE_VAND MQPR:$QdSrc, MQPR:$QnSrc, MQPR:$QmSrc, vpred_r:$vp)>;
1624 let Predicates = [HasMVEInt] in {
1625 def : Pat<(v16i8 (and (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1626 (v16i8 (MVE_VAND (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1627 def : Pat<(v8i16 (and (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1628 (v8i16 (MVE_VAND (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1629 def : Pat<(v4i32 (and (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1630 (v4i32 (MVE_VAND (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1632 def : Pat<(v16i8 (or (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1633 (v16i8 (MVE_VORR (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1634 def : Pat<(v8i16 (or (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1635 (v8i16 (MVE_VORR (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1636 def : Pat<(v4i32 (or (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1637 (v4i32 (MVE_VORR (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1639 def : Pat<(v16i8 (xor (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1640 (v16i8 (MVE_VEOR (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1641 def : Pat<(v8i16 (xor (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1642 (v8i16 (MVE_VEOR (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1643 def : Pat<(v4i32 (xor (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1644 (v4i32 (MVE_VEOR (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1646 def : Pat<(v16i8 (and (v16i8 MQPR:$val1), (vnotq MQPR:$val2))),
1647 (v16i8 (MVE_VBIC (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1648 def : Pat<(v8i16 (and (v8i16 MQPR:$val1), (vnotq MQPR:$val2))),
1649 (v8i16 (MVE_VBIC (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1650 def : Pat<(v4i32 (and (v4i32 MQPR:$val1), (vnotq MQPR:$val2))),
1651 (v4i32 (MVE_VBIC (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1653 def : Pat<(v16i8 (or (v16i8 MQPR:$val1), (vnotq (v16i8 MQPR:$val2)))),
1654 (v16i8 (MVE_VORN (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1655 def : Pat<(v8i16 (or (v8i16 MQPR:$val1), (vnotq MQPR:$val2))),
1656 (v8i16 (MVE_VORN (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1657 def : Pat<(v4i32 (or (v4i32 MQPR:$val1), (vnotq MQPR:$val2))),
1658 (v4i32 (MVE_VORN (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1661 class MVE_bit_cmode<string iname, string suffix, bits<4> cmode, dag inOps>
1662 : MVE_p<(outs MQPR:$Qd), inOps, NoItinerary,
1663 iname, suffix, "$Qd, $imm", vpred_n, "$Qd = $Qd_src"> {
1667 let Inst{28} = imm{7};
1668 let Inst{27-23} = 0b11111;
1669 let Inst{22} = Qd{3};
1670 let Inst{21-19} = 0b000;
1671 let Inst{18-16} = imm{6-4};
1672 let Inst{15-13} = Qd{2-0};
1674 let Inst{11-8} = cmode;
1675 let Inst{7-6} = 0b01;
1677 let Inst{3-0} = imm{3-0};
1680 class MVE_VORR<string suffix, bits<4> cmode, ExpandImm imm_type>
1681 : MVE_bit_cmode<"vorr", suffix, cmode, (ins MQPR:$Qd_src, imm_type:$imm)> {
1685 def MVE_VORRIZ0v4i32 : MVE_VORR<"i32", 0b0001, expzero00>;
1686 def MVE_VORRIZ0v8i16 : MVE_VORR<"i16", 0b1001, expzero00>;
1687 def MVE_VORRIZ8v4i32 : MVE_VORR<"i32", 0b0011, expzero08>;
1688 def MVE_VORRIZ8v8i16 : MVE_VORR<"i16", 0b1011, expzero08>;
1689 def MVE_VORRIZ16v4i32 : MVE_VORR<"i32", 0b0101, expzero16>;
1690 def MVE_VORRIZ24v4i32 : MVE_VORR<"i32", 0b0111, expzero24>;
1692 def MVE_VORNIZ0v4i32 : MVEAsmPseudo<"vorn${vp}.i32\t$Qd, $imm",
1693 (ins MQPR:$Qd_src, expzero00inv32:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1694 def MVE_VORNIZ0v8i16 : MVEAsmPseudo<"vorn${vp}.i16\t$Qd, $imm",
1695 (ins MQPR:$Qd_src, expzero00inv16:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1696 def MVE_VORNIZ8v4i32 : MVEAsmPseudo<"vorn${vp}.i32\t$Qd, $imm",
1697 (ins MQPR:$Qd_src, expzero08inv32:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1698 def MVE_VORNIZ8v8i16 : MVEAsmPseudo<"vorn${vp}.i16\t$Qd, $imm",
1699 (ins MQPR:$Qd_src, expzero08inv16:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1700 def MVE_VORNIZ16v4i32 : MVEAsmPseudo<"vorn${vp}.i32\t$Qd, $imm",
1701 (ins MQPR:$Qd_src, expzero16inv32:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1702 def MVE_VORNIZ24v4i32 : MVEAsmPseudo<"vorn${vp}.i32\t$Qd, $imm",
1703 (ins MQPR:$Qd_src, expzero24inv32:$imm, vpred_n:$vp), (outs MQPR:$Qd)>;
1705 def MVE_VMOV : MVEInstAlias<"vmov${vp}\t$Qd, $Qm",
1706 (MVE_VORR MQPR:$Qd, MQPR:$Qm, MQPR:$Qm, vpred_r:$vp)>;
1708 class MVE_VBIC<string suffix, bits<4> cmode, ExpandImm imm_type>
1709 : MVE_bit_cmode<"vbic", suffix, cmode, (ins MQPR:$Qd_src, imm_type:$imm)> {
1713 def MVE_VBICIZ0v4i32 : MVE_VBIC<"i32", 0b0001, expzero00>;
1714 def MVE_VBICIZ0v8i16 : MVE_VBIC<"i16", 0b1001, expzero00>;
1715 def MVE_VBICIZ8v4i32 : MVE_VBIC<"i32", 0b0011, expzero08>;
1716 def MVE_VBICIZ8v8i16 : MVE_VBIC<"i16", 0b1011, expzero08>;
1717 def MVE_VBICIZ16v4i32 : MVE_VBIC<"i32", 0b0101, expzero16>;
1718 def MVE_VBICIZ24v4i32 : MVE_VBIC<"i32", 0b0111, expzero24>;
1720 def MVE_VANDIZ0v4i32 : MVEAsmPseudo<"vand${vp}.i32\t$Qda, $imm",
1721 (ins MQPR:$Qda_src, expzero00inv32:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1722 def MVE_VANDIZ0v8i16 : MVEAsmPseudo<"vand${vp}.i16\t$Qda, $imm",
1723 (ins MQPR:$Qda_src, expzero00inv16:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1724 def MVE_VANDIZ8v4i32 : MVEAsmPseudo<"vand${vp}.i32\t$Qda, $imm",
1725 (ins MQPR:$Qda_src, expzero08inv32:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1726 def MVE_VANDIZ8v8i16 : MVEAsmPseudo<"vand${vp}.i16\t$Qda, $imm",
1727 (ins MQPR:$Qda_src, expzero08inv16:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1728 def MVE_VANDIZ16v4i32 : MVEAsmPseudo<"vand${vp}.i32\t$Qda, $imm",
1729 (ins MQPR:$Qda_src, expzero16inv32:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1730 def MVE_VANDIZ24v4i32 : MVEAsmPseudo<"vand${vp}.i32\t$Qda, $imm",
1731 (ins MQPR:$Qda_src, expzero24inv32:$imm, vpred_n:$vp), (outs MQPR:$Qda)>;
1733 class MVE_VMOV_lane_direction {
1740 def MVE_VMOV_from_lane : MVE_VMOV_lane_direction {
1742 let oops = (outs rGPR:$Rt);
1743 let iops = (ins MQPR:$Qd);
1744 let ops = "$Rt, $Qd$Idx";
1747 def MVE_VMOV_to_lane : MVE_VMOV_lane_direction {
1749 let oops = (outs MQPR:$Qd);
1750 let iops = (ins MQPR:$Qd_src, rGPR:$Rt);
1751 let ops = "$Qd$Idx, $Rt";
1752 let cstr = "$Qd = $Qd_src";
1755 class MVE_VMOV_lane<string suffix, bit U, dag indexop,
1756 MVE_VMOV_lane_direction dir>
1757 : MVE_VMOV_lane_base<dir.oops, !con(dir.iops, indexop), NoItinerary,
1758 "vmov", suffix, dir.ops, dir.cstr, []> {
1762 let Inst{31-24} = 0b11101110;
1764 let Inst{20} = dir.bit_20;
1765 let Inst{19-17} = Qd{2-0};
1766 let Inst{15-12} = Rt{3-0};
1767 let Inst{11-8} = 0b1011;
1768 let Inst{7} = Qd{3};
1769 let Inst{4-0} = 0b10000;
1772 class MVE_VMOV_lane_32<MVE_VMOV_lane_direction dir>
1773 : MVE_VMOV_lane<"32", 0b0, (ins MVEVectorIndex<4>:$Idx), dir> {
1776 let Inst{6-5} = 0b00;
1777 let Inst{16} = Idx{1};
1778 let Inst{21} = Idx{0};
1780 let Predicates = [HasFPRegsV8_1M];
1783 class MVE_VMOV_lane_16<string suffix, bit U, MVE_VMOV_lane_direction dir>
1784 : MVE_VMOV_lane<suffix, U, (ins MVEVectorIndex<8>:$Idx), dir> {
1788 let Inst{16} = Idx{2};
1789 let Inst{21} = Idx{1};
1790 let Inst{6} = Idx{0};
1793 class MVE_VMOV_lane_8<string suffix, bit U, MVE_VMOV_lane_direction dir>
1794 : MVE_VMOV_lane<suffix, U, (ins MVEVectorIndex<16>:$Idx), dir> {
1797 let Inst{16} = Idx{3};
1798 let Inst{21} = Idx{2};
1799 let Inst{6} = Idx{1};
1800 let Inst{5} = Idx{0};
1803 def MVE_VMOV_from_lane_32 : MVE_VMOV_lane_32< MVE_VMOV_from_lane>;
1804 def MVE_VMOV_to_lane_32 : MVE_VMOV_lane_32< MVE_VMOV_to_lane>;
1805 def MVE_VMOV_from_lane_s16 : MVE_VMOV_lane_16<"s16", 0b0, MVE_VMOV_from_lane>;
1806 def MVE_VMOV_from_lane_u16 : MVE_VMOV_lane_16<"u16", 0b1, MVE_VMOV_from_lane>;
1807 def MVE_VMOV_to_lane_16 : MVE_VMOV_lane_16< "16", 0b0, MVE_VMOV_to_lane>;
1808 def MVE_VMOV_from_lane_s8 : MVE_VMOV_lane_8 < "s8", 0b0, MVE_VMOV_from_lane>;
1809 def MVE_VMOV_from_lane_u8 : MVE_VMOV_lane_8 < "u8", 0b1, MVE_VMOV_from_lane>;
1810 def MVE_VMOV_to_lane_8 : MVE_VMOV_lane_8 < "8", 0b0, MVE_VMOV_to_lane>;
1812 let Predicates = [HasMVEInt] in {
1813 def : Pat<(extractelt (v2f64 MQPR:$src), imm:$lane),
1814 (f64 (EXTRACT_SUBREG MQPR:$src, (DSubReg_f64_reg imm:$lane)))>;
1815 def : Pat<(insertelt (v2f64 MQPR:$src1), DPR:$src2, imm:$lane),
1816 (INSERT_SUBREG (v2f64 (COPY_TO_REGCLASS MQPR:$src1, MQPR)), DPR:$src2, (DSubReg_f64_reg imm:$lane))>;
1818 def : Pat<(extractelt (v4i32 MQPR:$src), imm:$lane),
1820 (i32 (EXTRACT_SUBREG MQPR:$src, (SSubReg_f32_reg imm:$lane))), rGPR)>;
1821 def : Pat<(insertelt (v4i32 MQPR:$src1), rGPR:$src2, imm:$lane),
1822 (MVE_VMOV_to_lane_32 MQPR:$src1, rGPR:$src2, imm:$lane)>;
1824 def : Pat<(vector_insert (v16i8 MQPR:$src1), rGPR:$src2, imm:$lane),
1825 (MVE_VMOV_to_lane_8 MQPR:$src1, rGPR:$src2, imm:$lane)>;
1826 def : Pat<(vector_insert (v8i16 MQPR:$src1), rGPR:$src2, imm:$lane),
1827 (MVE_VMOV_to_lane_16 MQPR:$src1, rGPR:$src2, imm:$lane)>;
1829 def : Pat<(ARMvgetlanes (v16i8 MQPR:$src), imm:$lane),
1830 (MVE_VMOV_from_lane_s8 MQPR:$src, imm:$lane)>;
1831 def : Pat<(ARMvgetlanes (v8i16 MQPR:$src), imm:$lane),
1832 (MVE_VMOV_from_lane_s16 MQPR:$src, imm:$lane)>;
1833 def : Pat<(ARMvgetlaneu (v16i8 MQPR:$src), imm:$lane),
1834 (MVE_VMOV_from_lane_u8 MQPR:$src, imm:$lane)>;
1835 def : Pat<(ARMvgetlaneu (v8i16 MQPR:$src), imm:$lane),
1836 (MVE_VMOV_from_lane_u16 MQPR:$src, imm:$lane)>;
1838 def : Pat<(v16i8 (scalar_to_vector GPR:$src)),
1839 (MVE_VMOV_to_lane_8 (v16i8 (IMPLICIT_DEF)), rGPR:$src, (i32 0))>;
1840 def : Pat<(v8i16 (scalar_to_vector GPR:$src)),
1841 (MVE_VMOV_to_lane_16 (v8i16 (IMPLICIT_DEF)), rGPR:$src, (i32 0))>;
1842 def : Pat<(v4i32 (scalar_to_vector GPR:$src)),
1843 (MVE_VMOV_to_lane_32 (v4i32 (IMPLICIT_DEF)), rGPR:$src, (i32 0))>;
1845 // Floating point patterns, still enabled under HasMVEInt
1846 def : Pat<(extractelt (v4f32 MQPR:$src), imm:$lane),
1847 (COPY_TO_REGCLASS (f32 (EXTRACT_SUBREG MQPR:$src, (SSubReg_f32_reg imm:$lane))), SPR)>;
1848 def : Pat<(insertelt (v4f32 MQPR:$src1), (f32 SPR:$src2), imm:$lane),
1849 (INSERT_SUBREG (v4f32 (COPY_TO_REGCLASS MQPR:$src1, MQPR)), SPR:$src2, (SSubReg_f32_reg imm:$lane))>;
1851 def : Pat<(insertelt (v8f16 MQPR:$src1), HPR:$src2, imm:$lane),
1852 (MVE_VMOV_to_lane_16 MQPR:$src1, (COPY_TO_REGCLASS HPR:$src2, rGPR), imm:$lane)>;
1853 def : Pat<(extractelt (v8f16 MQPR:$src), imm:$lane),
1854 (COPY_TO_REGCLASS (MVE_VMOV_from_lane_u16 MQPR:$src, imm:$lane), HPR)>;
1856 def : Pat<(v4f32 (scalar_to_vector SPR:$src)),
1857 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), SPR:$src, ssub_0)>;
1858 def : Pat<(v4f32 (scalar_to_vector GPR:$src)),
1859 (MVE_VMOV_to_lane_32 (v4f32 (IMPLICIT_DEF)), rGPR:$src, (i32 0))>;
1860 def : Pat<(v8f16 (scalar_to_vector HPR:$src)),
1861 (INSERT_SUBREG (v8f16 (IMPLICIT_DEF)), HPR:$src, ssub_0)>;
1862 def : Pat<(v8f16 (scalar_to_vector GPR:$src)),
1863 (MVE_VMOV_to_lane_16 (v8f16 (IMPLICIT_DEF)), rGPR:$src, (i32 0))>;
1866 // end of mve_bit instructions
1868 // start of MVE Integer instructions
1870 class MVE_int<string iname, string suffix, bits<2> size, list<dag> pattern=[]>
1871 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm), NoItinerary,
1872 iname, suffix, "$Qd, $Qn, $Qm", vpred_r, "", pattern> {
1877 let Inst{22} = Qd{3};
1878 let Inst{21-20} = size;
1879 let Inst{19-17} = Qn{2-0};
1880 let Inst{15-13} = Qd{2-0};
1881 let Inst{7} = Qn{3};
1883 let Inst{5} = Qm{3};
1884 let Inst{3-1} = Qm{2-0};
1887 class MVE_VMULt1<string suffix, bits<2> size, list<dag> pattern=[]>
1888 : MVE_int<"vmul", suffix, size, pattern> {
1891 let Inst{25-23} = 0b110;
1893 let Inst{12-8} = 0b01001;
1898 def MVE_VMULt1i8 : MVE_VMULt1<"i8", 0b00>;
1899 def MVE_VMULt1i16 : MVE_VMULt1<"i16", 0b01>;
1900 def MVE_VMULt1i32 : MVE_VMULt1<"i32", 0b10>;
1902 let Predicates = [HasMVEInt] in {
1903 def : Pat<(v16i8 (mul (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1904 (v16i8 (MVE_VMULt1i8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1905 def : Pat<(v8i16 (mul (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1906 (v8i16 (MVE_VMULt1i16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1907 def : Pat<(v4i32 (mul (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1908 (v4i32 (MVE_VMULt1i32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1911 class MVE_VQxDMULH<string iname, string suffix, bits<2> size, bit rounding,
1912 list<dag> pattern=[]>
1913 : MVE_int<iname, suffix, size, pattern> {
1915 let Inst{28} = rounding;
1916 let Inst{25-23} = 0b110;
1918 let Inst{12-8} = 0b01011;
1923 class MVE_VQDMULH<string suffix, bits<2> size, list<dag> pattern=[]>
1924 : MVE_VQxDMULH<"vqdmulh", suffix, size, 0b0, pattern>;
1925 class MVE_VQRDMULH<string suffix, bits<2> size, list<dag> pattern=[]>
1926 : MVE_VQxDMULH<"vqrdmulh", suffix, size, 0b1, pattern>;
1928 def MVE_VQDMULHi8 : MVE_VQDMULH<"s8", 0b00>;
1929 def MVE_VQDMULHi16 : MVE_VQDMULH<"s16", 0b01>;
1930 def MVE_VQDMULHi32 : MVE_VQDMULH<"s32", 0b10>;
1932 def MVE_VQRDMULHi8 : MVE_VQRDMULH<"s8", 0b00>;
1933 def MVE_VQRDMULHi16 : MVE_VQRDMULH<"s16", 0b01>;
1934 def MVE_VQRDMULHi32 : MVE_VQRDMULH<"s32", 0b10>;
1936 class MVE_VADDSUB<string iname, string suffix, bits<2> size, bit subtract,
1937 list<dag> pattern=[]>
1938 : MVE_int<iname, suffix, size, pattern> {
1940 let Inst{28} = subtract;
1941 let Inst{25-23} = 0b110;
1943 let Inst{12-8} = 0b01000;
1948 class MVE_VADD<string suffix, bits<2> size, list<dag> pattern=[]>
1949 : MVE_VADDSUB<"vadd", suffix, size, 0b0, pattern>;
1950 class MVE_VSUB<string suffix, bits<2> size, list<dag> pattern=[]>
1951 : MVE_VADDSUB<"vsub", suffix, size, 0b1, pattern>;
1953 def MVE_VADDi8 : MVE_VADD<"i8", 0b00>;
1954 def MVE_VADDi16 : MVE_VADD<"i16", 0b01>;
1955 def MVE_VADDi32 : MVE_VADD<"i32", 0b10>;
1957 let Predicates = [HasMVEInt] in {
1958 def : Pat<(v16i8 (add (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1959 (v16i8 (MVE_VADDi8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1960 def : Pat<(v8i16 (add (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1961 (v8i16 (MVE_VADDi16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1962 def : Pat<(v4i32 (add (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1963 (v4i32 (MVE_VADDi32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1966 def MVE_VSUBi8 : MVE_VSUB<"i8", 0b00>;
1967 def MVE_VSUBi16 : MVE_VSUB<"i16", 0b01>;
1968 def MVE_VSUBi32 : MVE_VSUB<"i32", 0b10>;
1970 let Predicates = [HasMVEInt] in {
1971 def : Pat<(v16i8 (sub (v16i8 MQPR:$val1), (v16i8 MQPR:$val2))),
1972 (v16i8 (MVE_VSUBi8 (v16i8 MQPR:$val1), (v16i8 MQPR:$val2)))>;
1973 def : Pat<(v8i16 (sub (v8i16 MQPR:$val1), (v8i16 MQPR:$val2))),
1974 (v8i16 (MVE_VSUBi16 (v8i16 MQPR:$val1), (v8i16 MQPR:$val2)))>;
1975 def : Pat<(v4i32 (sub (v4i32 MQPR:$val1), (v4i32 MQPR:$val2))),
1976 (v4i32 (MVE_VSUBi32 (v4i32 MQPR:$val1), (v4i32 MQPR:$val2)))>;
1979 class MVE_VQADDSUB<string iname, string suffix, bit U, bit subtract,
1980 bits<2> size, list<dag> pattern=[]>
1981 : MVE_int<iname, suffix, size, pattern> {
1984 let Inst{25-23} = 0b110;
1986 let Inst{12-10} = 0b000;
1987 let Inst{9} = subtract;
1993 class MVE_VQADD<string suffix, bit U, bits<2> size, list<dag> pattern=[]>
1994 : MVE_VQADDSUB<"vqadd", suffix, U, 0b0, size, pattern>;
1995 class MVE_VQSUB<string suffix, bit U, bits<2> size, list<dag> pattern=[]>
1996 : MVE_VQADDSUB<"vqsub", suffix, U, 0b1, size, pattern>;
1998 def MVE_VQADDs8 : MVE_VQADD<"s8", 0b0, 0b00>;
1999 def MVE_VQADDs16 : MVE_VQADD<"s16", 0b0, 0b01>;
2000 def MVE_VQADDs32 : MVE_VQADD<"s32", 0b0, 0b10>;
2001 def MVE_VQADDu8 : MVE_VQADD<"u8", 0b1, 0b00>;
2002 def MVE_VQADDu16 : MVE_VQADD<"u16", 0b1, 0b01>;
2003 def MVE_VQADDu32 : MVE_VQADD<"u32", 0b1, 0b10>;
2005 def MVE_VQSUBs8 : MVE_VQSUB<"s8", 0b0, 0b00>;
2006 def MVE_VQSUBs16 : MVE_VQSUB<"s16", 0b0, 0b01>;
2007 def MVE_VQSUBs32 : MVE_VQSUB<"s32", 0b0, 0b10>;
2008 def MVE_VQSUBu8 : MVE_VQSUB<"u8", 0b1, 0b00>;
2009 def MVE_VQSUBu16 : MVE_VQSUB<"u16", 0b1, 0b01>;
2010 def MVE_VQSUBu32 : MVE_VQSUB<"u32", 0b1, 0b10>;
2012 class MVE_VABD_int<string suffix, bit U, bits<2> size, list<dag> pattern=[]>
2013 : MVE_int<"vabd", suffix, size, pattern> {
2016 let Inst{25-23} = 0b110;
2018 let Inst{12-8} = 0b00111;
2023 def MVE_VABDs8 : MVE_VABD_int<"s8", 0b0, 0b00>;
2024 def MVE_VABDs16 : MVE_VABD_int<"s16", 0b0, 0b01>;
2025 def MVE_VABDs32 : MVE_VABD_int<"s32", 0b0, 0b10>;
2026 def MVE_VABDu8 : MVE_VABD_int<"u8", 0b1, 0b00>;
2027 def MVE_VABDu16 : MVE_VABD_int<"u16", 0b1, 0b01>;
2028 def MVE_VABDu32 : MVE_VABD_int<"u32", 0b1, 0b10>;
2030 class MVE_VRHADD<string suffix, bit U, bits<2> size, list<dag> pattern=[]>
2031 : MVE_int<"vrhadd", suffix, size, pattern> {
2034 let Inst{25-23} = 0b110;
2036 let Inst{12-8} = 0b00001;
2041 def MVE_VRHADDs8 : MVE_VRHADD<"s8", 0b0, 0b00>;
2042 def MVE_VRHADDs16 : MVE_VRHADD<"s16", 0b0, 0b01>;
2043 def MVE_VRHADDs32 : MVE_VRHADD<"s32", 0b0, 0b10>;
2044 def MVE_VRHADDu8 : MVE_VRHADD<"u8", 0b1, 0b00>;
2045 def MVE_VRHADDu16 : MVE_VRHADD<"u16", 0b1, 0b01>;
2046 def MVE_VRHADDu32 : MVE_VRHADD<"u32", 0b1, 0b10>;
2048 class MVE_VHADDSUB<string iname, string suffix, bit U, bit subtract,
2049 bits<2> size, list<dag> pattern=[]>
2050 : MVE_int<iname, suffix, size, pattern> {
2053 let Inst{25-23} = 0b110;
2055 let Inst{12-10} = 0b000;
2056 let Inst{9} = subtract;
2062 class MVE_VHADD<string suffix, bit U, bits<2> size,
2063 list<dag> pattern=[]>
2064 : MVE_VHADDSUB<"vhadd", suffix, U, 0b0, size, pattern>;
2065 class MVE_VHSUB<string suffix, bit U, bits<2> size,
2066 list<dag> pattern=[]>
2067 : MVE_VHADDSUB<"vhsub", suffix, U, 0b1, size, pattern>;
2069 def MVE_VHADDs8 : MVE_VHADD<"s8", 0b0, 0b00>;
2070 def MVE_VHADDs16 : MVE_VHADD<"s16", 0b0, 0b01>;
2071 def MVE_VHADDs32 : MVE_VHADD<"s32", 0b0, 0b10>;
2072 def MVE_VHADDu8 : MVE_VHADD<"u8", 0b1, 0b00>;
2073 def MVE_VHADDu16 : MVE_VHADD<"u16", 0b1, 0b01>;
2074 def MVE_VHADDu32 : MVE_VHADD<"u32", 0b1, 0b10>;
2076 def MVE_VHSUBs8 : MVE_VHSUB<"s8", 0b0, 0b00>;
2077 def MVE_VHSUBs16 : MVE_VHSUB<"s16", 0b0, 0b01>;
2078 def MVE_VHSUBs32 : MVE_VHSUB<"s32", 0b0, 0b10>;
2079 def MVE_VHSUBu8 : MVE_VHSUB<"u8", 0b1, 0b00>;
2080 def MVE_VHSUBu16 : MVE_VHSUB<"u16", 0b1, 0b01>;
2081 def MVE_VHSUBu32 : MVE_VHSUB<"u32", 0b1, 0b10>;
2083 class MVE_VDUP<string suffix, bit B, bit E, list<dag> pattern=[]>
2084 : MVE_p<(outs MQPR:$Qd), (ins rGPR:$Rt), NoItinerary,
2085 "vdup", suffix, "$Qd, $Rt", vpred_r, "", pattern> {
2090 let Inst{25-23} = 0b101;
2092 let Inst{21-20} = 0b10;
2093 let Inst{19-17} = Qd{2-0};
2095 let Inst{15-12} = Rt;
2096 let Inst{11-8} = 0b1011;
2097 let Inst{7} = Qd{3};
2100 let Inst{4-0} = 0b10000;
2103 def MVE_VDUP32 : MVE_VDUP<"32", 0b0, 0b0>;
2104 def MVE_VDUP16 : MVE_VDUP<"16", 0b0, 0b1>;
2105 def MVE_VDUP8 : MVE_VDUP<"8", 0b1, 0b0>;
2107 let Predicates = [HasMVEInt] in {
2108 def : Pat<(v16i8 (ARMvdup (i32 rGPR:$elem))),
2109 (MVE_VDUP8 rGPR:$elem)>;
2110 def : Pat<(v8i16 (ARMvdup (i32 rGPR:$elem))),
2111 (MVE_VDUP16 rGPR:$elem)>;
2112 def : Pat<(v4i32 (ARMvdup (i32 rGPR:$elem))),
2113 (MVE_VDUP32 rGPR:$elem)>;
2115 def : Pat<(v4i32 (ARMvduplane (v4i32 MQPR:$src), imm:$lane)),
2116 (MVE_VDUP32 (MVE_VMOV_from_lane_32 MQPR:$src, imm:$lane))>;
2117 // For the 16-bit and 8-bit vduplanes we don't care about the signedness
2118 // of the lane move operation as we only want the lowest 8/16 bits anyway.
2119 def : Pat<(v8i16 (ARMvduplane (v8i16 MQPR:$src), imm:$lane)),
2120 (MVE_VDUP16 (MVE_VMOV_from_lane_u16 MQPR:$src, imm:$lane))>;
2121 def : Pat<(v16i8 (ARMvduplane (v16i8 MQPR:$src), imm:$lane)),
2122 (MVE_VDUP8 (MVE_VMOV_from_lane_u8 MQPR:$src, imm:$lane))>;
2124 def : Pat<(v4f32 (ARMvdup (f32 SPR:$elem))),
2125 (v4f32 (MVE_VDUP32 (i32 (COPY_TO_REGCLASS (f32 SPR:$elem), rGPR))))>;
2126 def : Pat<(v8f16 (ARMvdup (f16 HPR:$elem))),
2127 (v8f16 (MVE_VDUP16 (i32 (COPY_TO_REGCLASS (f16 HPR:$elem), rGPR))))>;
2129 def : Pat<(v4f32 (ARMvduplane (v4f32 MQPR:$src), imm:$lane)),
2130 (MVE_VDUP32 (MVE_VMOV_from_lane_32 MQPR:$src, imm:$lane))>;
2131 def : Pat<(v8f16 (ARMvduplane (v8f16 MQPR:$src), imm:$lane)),
2132 (MVE_VDUP16 (MVE_VMOV_from_lane_u16 MQPR:$src, imm:$lane))>;
2136 class MVEIntSingleSrc<string iname, string suffix, bits<2> size,
2137 list<dag> pattern=[]>
2138 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qm), NoItinerary,
2139 iname, suffix, "$Qd, $Qm", vpred_r, "", pattern> {
2143 let Inst{22} = Qd{3};
2144 let Inst{19-18} = size{1-0};
2145 let Inst{15-13} = Qd{2-0};
2146 let Inst{5} = Qm{3};
2147 let Inst{3-1} = Qm{2-0};
2150 class MVE_VCLSCLZ<string iname, string suffix, bits<2> size,
2151 bit count_zeroes, list<dag> pattern=[]>
2152 : MVEIntSingleSrc<iname, suffix, size, pattern> {
2155 let Inst{25-23} = 0b111;
2156 let Inst{21-20} = 0b11;
2157 let Inst{17-16} = 0b00;
2158 let Inst{12-8} = 0b00100;
2159 let Inst{7} = count_zeroes;
2165 def MVE_VCLSs8 : MVE_VCLSCLZ<"vcls", "s8", 0b00, 0b0>;
2166 def MVE_VCLSs16 : MVE_VCLSCLZ<"vcls", "s16", 0b01, 0b0>;
2167 def MVE_VCLSs32 : MVE_VCLSCLZ<"vcls", "s32", 0b10, 0b0>;
2169 def MVE_VCLZs8 : MVE_VCLSCLZ<"vclz", "i8", 0b00, 0b1>;
2170 def MVE_VCLZs16 : MVE_VCLSCLZ<"vclz", "i16", 0b01, 0b1>;
2171 def MVE_VCLZs32 : MVE_VCLSCLZ<"vclz", "i32", 0b10, 0b1>;
2173 class MVE_VABSNEG_int<string iname, string suffix, bits<2> size, bit negate,
2174 list<dag> pattern=[]>
2175 : MVEIntSingleSrc<iname, suffix, size, pattern> {
2178 let Inst{25-23} = 0b111;
2179 let Inst{21-20} = 0b11;
2180 let Inst{17-16} = 0b01;
2181 let Inst{12-8} = 0b00011;
2182 let Inst{7} = negate;
2188 def MVE_VABSs8 : MVE_VABSNEG_int<"vabs", "s8", 0b00, 0b0>;
2189 def MVE_VABSs16 : MVE_VABSNEG_int<"vabs", "s16", 0b01, 0b0>;
2190 def MVE_VABSs32 : MVE_VABSNEG_int<"vabs", "s32", 0b10, 0b0>;
2192 let Predicates = [HasMVEInt] in {
2193 def : Pat<(v16i8 (abs (v16i8 MQPR:$v))),
2194 (v16i8 (MVE_VABSs8 $v))>;
2195 def : Pat<(v8i16 (abs (v8i16 MQPR:$v))),
2196 (v8i16 (MVE_VABSs16 $v))>;
2197 def : Pat<(v4i32 (abs (v4i32 MQPR:$v))),
2198 (v4i32 (MVE_VABSs32 $v))>;
2201 def MVE_VNEGs8 : MVE_VABSNEG_int<"vneg", "s8", 0b00, 0b1>;
2202 def MVE_VNEGs16 : MVE_VABSNEG_int<"vneg", "s16", 0b01, 0b1>;
2203 def MVE_VNEGs32 : MVE_VABSNEG_int<"vneg", "s32", 0b10, 0b1>;
2205 class MVE_VQABSNEG<string iname, string suffix, bits<2> size,
2206 bit negate, list<dag> pattern=[]>
2207 : MVEIntSingleSrc<iname, suffix, size, pattern> {
2210 let Inst{25-23} = 0b111;
2211 let Inst{21-20} = 0b11;
2212 let Inst{17-16} = 0b00;
2213 let Inst{12-8} = 0b00111;
2214 let Inst{7} = negate;
2220 def MVE_VQABSs8 : MVE_VQABSNEG<"vqabs", "s8", 0b00, 0b0>;
2221 def MVE_VQABSs16 : MVE_VQABSNEG<"vqabs", "s16", 0b01, 0b0>;
2222 def MVE_VQABSs32 : MVE_VQABSNEG<"vqabs", "s32", 0b10, 0b0>;
2224 def MVE_VQNEGs8 : MVE_VQABSNEG<"vqneg", "s8", 0b00, 0b1>;
2225 def MVE_VQNEGs16 : MVE_VQABSNEG<"vqneg", "s16", 0b01, 0b1>;
2226 def MVE_VQNEGs32 : MVE_VQABSNEG<"vqneg", "s32", 0b10, 0b1>;
2228 class MVE_mod_imm<string iname, string suffix, bits<4> cmode, bit op,
2229 dag iops, list<dag> pattern=[]>
2230 : MVE_p<(outs MQPR:$Qd), iops, NoItinerary, iname, suffix, "$Qd, $imm",
2231 vpred_r, "", pattern> {
2235 let Inst{28} = imm{7};
2236 let Inst{25-23} = 0b111;
2237 let Inst{22} = Qd{3};
2238 let Inst{21-19} = 0b000;
2239 let Inst{18-16} = imm{6-4};
2240 let Inst{15-13} = Qd{2-0};
2242 let Inst{11-8} = cmode{3-0};
2243 let Inst{7-6} = 0b01;
2246 let Inst{3-0} = imm{3-0};
2248 let DecoderMethod = "DecodeMVEModImmInstruction";
2251 let isReMaterializable = 1 in {
2252 let isAsCheapAsAMove = 1 in {
2253 def MVE_VMOVimmi8 : MVE_mod_imm<"vmov", "i8", {1,1,1,0}, 0b0, (ins nImmSplatI8:$imm)>;
2254 def MVE_VMOVimmi16 : MVE_mod_imm<"vmov", "i16", {1,0,?,0}, 0b0, (ins nImmSplatI16:$imm)> {
2255 let Inst{9} = imm{9};
2257 def MVE_VMOVimmi32 : MVE_mod_imm<"vmov", "i32", {?,?,?,?}, 0b0, (ins nImmVMOVI32:$imm)> {
2258 let Inst{11-8} = imm{11-8};
2260 def MVE_VMOVimmi64 : MVE_mod_imm<"vmov", "i64", {1,1,1,0}, 0b1, (ins nImmSplatI64:$imm)>;
2261 def MVE_VMOVimmf32 : MVE_mod_imm<"vmov", "f32", {1,1,1,1}, 0b0, (ins nImmVMOVF32:$imm)>;
2262 } // let isAsCheapAsAMove = 1
2264 def MVE_VMVNimmi16 : MVE_mod_imm<"vmvn", "i16", {1,0,?,0}, 0b1, (ins nImmSplatI16:$imm)> {
2265 let Inst{9} = imm{9};
2267 def MVE_VMVNimmi32 : MVE_mod_imm<"vmvn", "i32", {?,?,?,?}, 0b1, (ins nImmVMOVI32:$imm)> {
2268 let Inst{11-8} = imm{11-8};
2270 } // let isReMaterializable = 1
2272 let Predicates = [HasMVEInt] in {
2273 def : Pat<(v16i8 (ARMvmovImm timm:$simm)),
2274 (v16i8 (MVE_VMOVimmi8 nImmSplatI8:$simm))>;
2275 def : Pat<(v8i16 (ARMvmovImm timm:$simm)),
2276 (v8i16 (MVE_VMOVimmi16 nImmSplatI16:$simm))>;
2277 def : Pat<(v4i32 (ARMvmovImm timm:$simm)),
2278 (v4i32 (MVE_VMOVimmi32 nImmVMOVI32:$simm))>;
2280 def : Pat<(v8i16 (ARMvmvnImm timm:$simm)),
2281 (v8i16 (MVE_VMVNimmi16 nImmSplatI16:$simm))>;
2282 def : Pat<(v4i32 (ARMvmvnImm timm:$simm)),
2283 (v4i32 (MVE_VMVNimmi32 nImmVMOVI32:$simm))>;
2285 def : Pat<(v4f32 (ARMvmovFPImm timm:$simm)),
2286 (v4f32 (MVE_VMOVimmf32 nImmVMOVF32:$simm))>;
2289 class MVE_VMINMAXA<string iname, string suffix, bits<2> size,
2290 bit bit_12, list<dag> pattern=[]>
2291 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qd_src, MQPR:$Qm),
2292 NoItinerary, iname, suffix, "$Qd, $Qm", vpred_n, "$Qd = $Qd_src",
2298 let Inst{25-23} = 0b100;
2299 let Inst{22} = Qd{3};
2300 let Inst{21-20} = 0b11;
2301 let Inst{19-18} = size;
2302 let Inst{17-16} = 0b11;
2303 let Inst{15-13} = Qd{2-0};
2304 let Inst{12} = bit_12;
2305 let Inst{11-6} = 0b111010;
2306 let Inst{5} = Qm{3};
2308 let Inst{3-1} = Qm{2-0};
2312 def MVE_VMAXAs8 : MVE_VMINMAXA<"vmaxa", "s8", 0b00, 0b0>;
2313 def MVE_VMAXAs16 : MVE_VMINMAXA<"vmaxa", "s16", 0b01, 0b0>;
2314 def MVE_VMAXAs32 : MVE_VMINMAXA<"vmaxa", "s32", 0b10, 0b0>;
2316 def MVE_VMINAs8 : MVE_VMINMAXA<"vmina", "s8", 0b00, 0b1>;
2317 def MVE_VMINAs16 : MVE_VMINMAXA<"vmina", "s16", 0b01, 0b1>;
2318 def MVE_VMINAs32 : MVE_VMINMAXA<"vmina", "s32", 0b10, 0b1>;
2320 // end of MVE Integer instructions
2322 // start of MVE Floating Point instructions
2324 class MVE_float<string iname, string suffix, dag oops, dag iops, string ops,
2325 vpred_ops vpred, string cstr, list<dag> pattern=[]>
2326 : MVE_f<oops, iops, NoItinerary, iname, suffix, ops, vpred, cstr, pattern> {
2331 let Inst{5} = Qm{3};
2332 let Inst{3-1} = Qm{2-0};
2336 class MVE_VRINT<string rmode, bits<3> op, string suffix, bits<2> size,
2337 list<dag> pattern=[]>
2338 : MVE_float<!strconcat("vrint", rmode), suffix, (outs MQPR:$Qd),
2339 (ins MQPR:$Qm), "$Qd, $Qm", vpred_r, "", pattern> {
2343 let Inst{25-23} = 0b111;
2344 let Inst{22} = Qd{3};
2345 let Inst{21-20} = 0b11;
2346 let Inst{19-18} = size;
2347 let Inst{17-16} = 0b10;
2348 let Inst{15-13} = Qd{2-0};
2349 let Inst{11-10} = 0b01;
2350 let Inst{9-7} = op{2-0};
2355 multiclass MVE_VRINT_ops<string suffix, bits<2> size, list<dag> pattern=[]> {
2356 def N : MVE_VRINT<"n", 0b000, suffix, size, pattern>;
2357 def X : MVE_VRINT<"x", 0b001, suffix, size, pattern>;
2358 def A : MVE_VRINT<"a", 0b010, suffix, size, pattern>;
2359 def Z : MVE_VRINT<"z", 0b011, suffix, size, pattern>;
2360 def M : MVE_VRINT<"m", 0b101, suffix, size, pattern>;
2361 def P : MVE_VRINT<"p", 0b111, suffix, size, pattern>;
2364 defm MVE_VRINTf16 : MVE_VRINT_ops<"f16", 0b01>;
2365 defm MVE_VRINTf32 : MVE_VRINT_ops<"f32", 0b10>;
2367 let Predicates = [HasMVEFloat] in {
2368 def : Pat<(v4f32 (frint (v4f32 MQPR:$val1))),
2369 (v4f32 (MVE_VRINTf32X (v4f32 MQPR:$val1)))>;
2370 def : Pat<(v8f16 (frint (v8f16 MQPR:$val1))),
2371 (v8f16 (MVE_VRINTf16X (v8f16 MQPR:$val1)))>;
2372 def : Pat<(v4f32 (fround (v4f32 MQPR:$val1))),
2373 (v4f32 (MVE_VRINTf32A (v4f32 MQPR:$val1)))>;
2374 def : Pat<(v8f16 (fround (v8f16 MQPR:$val1))),
2375 (v8f16 (MVE_VRINTf16A (v8f16 MQPR:$val1)))>;
2376 def : Pat<(v4f32 (ftrunc (v4f32 MQPR:$val1))),
2377 (v4f32 (MVE_VRINTf32Z (v4f32 MQPR:$val1)))>;
2378 def : Pat<(v8f16 (ftrunc (v8f16 MQPR:$val1))),
2379 (v8f16 (MVE_VRINTf16Z (v8f16 MQPR:$val1)))>;
2380 def : Pat<(v4f32 (ffloor (v4f32 MQPR:$val1))),
2381 (v4f32 (MVE_VRINTf32M (v4f32 MQPR:$val1)))>;
2382 def : Pat<(v8f16 (ffloor (v8f16 MQPR:$val1))),
2383 (v8f16 (MVE_VRINTf16M (v8f16 MQPR:$val1)))>;
2384 def : Pat<(v4f32 (fceil (v4f32 MQPR:$val1))),
2385 (v4f32 (MVE_VRINTf32P (v4f32 MQPR:$val1)))>;
2386 def : Pat<(v8f16 (fceil (v8f16 MQPR:$val1))),
2387 (v8f16 (MVE_VRINTf16P (v8f16 MQPR:$val1)))>;
2390 class MVEFloatArithNeon<string iname, string suffix, bit size,
2391 dag oops, dag iops, string ops,
2392 vpred_ops vpred, string cstr, list<dag> pattern=[]>
2393 : MVE_float<iname, suffix, oops, iops, ops, vpred, cstr, pattern> {
2394 let Inst{20} = size;
2398 class MVE_VMUL_fp<string suffix, bit size, list<dag> pattern=[]>
2399 : MVEFloatArithNeon<"vmul", suffix, size, (outs MQPR:$Qd),
2400 (ins MQPR:$Qn, MQPR:$Qm), "$Qd, $Qn, $Qm", vpred_r, "",
2406 let Inst{25-23} = 0b110;
2407 let Inst{22} = Qd{3};
2409 let Inst{19-17} = Qn{2-0};
2410 let Inst{15-13} = Qd{2-0};
2411 let Inst{12-8} = 0b01101;
2412 let Inst{7} = Qn{3};
2416 def MVE_VMULf32 : MVE_VMUL_fp<"f32", 0b0>;
2417 def MVE_VMULf16 : MVE_VMUL_fp<"f16", 0b1>;
2419 let Predicates = [HasMVEFloat] in {
2420 def : Pat<(v4f32 (fmul (v4f32 MQPR:$val1), (v4f32 MQPR:$val2))),
2421 (v4f32 (MVE_VMULf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2)))>;
2422 def : Pat<(v8f16 (fmul (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
2423 (v8f16 (MVE_VMULf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
2426 class MVE_VCMLA<string suffix, bit size, list<dag> pattern=[]>
2427 : MVEFloatArithNeon<"vcmla", suffix, size, (outs MQPR:$Qd),
2428 (ins MQPR:$Qd_src, MQPR:$Qn, MQPR:$Qm, complexrotateop:$rot),
2429 "$Qd, $Qn, $Qm, $rot", vpred_n, "$Qd = $Qd_src", pattern> {
2436 let Inst{24-23} = rot;
2437 let Inst{22} = Qd{3};
2439 let Inst{19-17} = Qn{2-0};
2440 let Inst{15-13} = Qd{2-0};
2441 let Inst{12-8} = 0b01000;
2442 let Inst{7} = Qn{3};
2446 def MVE_VCMLAf16 : MVE_VCMLA<"f16", 0b0>;
2447 def MVE_VCMLAf32 : MVE_VCMLA<"f32", 0b1>;
2449 class MVE_VADDSUBFMA_fp<string iname, string suffix, bit size, bit bit_4,
2450 bit bit_8, bit bit_21, dag iops=(ins),
2451 vpred_ops vpred=vpred_r, string cstr="",
2452 list<dag> pattern=[]>
2453 : MVEFloatArithNeon<iname, suffix, size, (outs MQPR:$Qd),
2454 !con(iops, (ins MQPR:$Qn, MQPR:$Qm)), "$Qd, $Qn, $Qm",
2455 vpred, cstr, pattern> {
2460 let Inst{25-23} = 0b110;
2461 let Inst{22} = Qd{3};
2462 let Inst{21} = bit_21;
2463 let Inst{19-17} = Qn{2-0};
2464 let Inst{15-13} = Qd{2-0};
2465 let Inst{11-9} = 0b110;
2466 let Inst{8} = bit_8;
2467 let Inst{7} = Qn{3};
2468 let Inst{4} = bit_4;
2471 def MVE_VFMAf32 : MVE_VADDSUBFMA_fp<"vfma", "f32", 0b0, 0b1, 0b0, 0b0,
2472 (ins MQPR:$Qd_src), vpred_n, "$Qd = $Qd_src">;
2473 def MVE_VFMAf16 : MVE_VADDSUBFMA_fp<"vfma", "f16", 0b1, 0b1, 0b0, 0b0,
2474 (ins MQPR:$Qd_src), vpred_n, "$Qd = $Qd_src">;
2476 def MVE_VFMSf32 : MVE_VADDSUBFMA_fp<"vfms", "f32", 0b0, 0b1, 0b0, 0b1,
2477 (ins MQPR:$Qd_src), vpred_n, "$Qd = $Qd_src">;
2478 def MVE_VFMSf16 : MVE_VADDSUBFMA_fp<"vfms", "f16", 0b1, 0b1, 0b0, 0b1,
2479 (ins MQPR:$Qd_src), vpred_n, "$Qd = $Qd_src">;
2481 def MVE_VADDf32 : MVE_VADDSUBFMA_fp<"vadd", "f32", 0b0, 0b0, 0b1, 0b0>;
2482 def MVE_VADDf16 : MVE_VADDSUBFMA_fp<"vadd", "f16", 0b1, 0b0, 0b1, 0b0>;
2484 let Predicates = [HasMVEFloat] in {
2485 def : Pat<(v4f32 (fadd (v4f32 MQPR:$val1), (v4f32 MQPR:$val2))),
2486 (v4f32 (MVE_VADDf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2)))>;
2487 def : Pat<(v8f16 (fadd (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
2488 (v8f16 (MVE_VADDf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
2491 def MVE_VSUBf32 : MVE_VADDSUBFMA_fp<"vsub", "f32", 0b0, 0b0, 0b1, 0b1>;
2492 def MVE_VSUBf16 : MVE_VADDSUBFMA_fp<"vsub", "f16", 0b1, 0b0, 0b1, 0b1>;
2494 let Predicates = [HasMVEFloat] in {
2495 def : Pat<(v4f32 (fsub (v4f32 MQPR:$val1), (v4f32 MQPR:$val2))),
2496 (v4f32 (MVE_VSUBf32 (v4f32 MQPR:$val1), (v4f32 MQPR:$val2)))>;
2497 def : Pat<(v8f16 (fsub (v8f16 MQPR:$val1), (v8f16 MQPR:$val2))),
2498 (v8f16 (MVE_VSUBf16 (v8f16 MQPR:$val1), (v8f16 MQPR:$val2)))>;
2501 class MVE_VCADD<string suffix, bit size, list<dag> pattern=[]>
2502 : MVEFloatArithNeon<"vcadd", suffix, size, (outs MQPR:$Qd),
2503 (ins MQPR:$Qn, MQPR:$Qm, complexrotateopodd:$rot),
2504 "$Qd, $Qn, $Qm, $rot", vpred_r, "", pattern> {
2513 let Inst{22} = Qd{3};
2515 let Inst{19-17} = Qn{2-0};
2516 let Inst{15-13} = Qd{2-0};
2517 let Inst{12-8} = 0b01000;
2518 let Inst{7} = Qn{3};
2522 def MVE_VCADDf16 : MVE_VCADD<"f16", 0b0>;
2523 def MVE_VCADDf32 : MVE_VCADD<"f32", 0b1>;
2525 class MVE_VABD_fp<string suffix, bit size>
2526 : MVE_float<"vabd", suffix, (outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm),
2527 "$Qd, $Qn, $Qm", vpred_r, ""> {
2532 let Inst{25-23} = 0b110;
2533 let Inst{22} = Qd{3};
2535 let Inst{20} = size;
2536 let Inst{19-17} = Qn{2-0};
2538 let Inst{15-13} = Qd{2-0};
2539 let Inst{11-8} = 0b1101;
2540 let Inst{7} = Qn{3};
2544 def MVE_VABDf32 : MVE_VABD_fp<"f32", 0b0>;
2545 def MVE_VABDf16 : MVE_VABD_fp<"f16", 0b1>;
2547 class MVE_VCVT_fix<string suffix, bit fsi, bit U, bit op,
2548 Operand imm_operand_type, list<dag> pattern=[]>
2549 : MVE_float<"vcvt", suffix,
2550 (outs MQPR:$Qd), (ins MQPR:$Qm, imm_operand_type:$imm6),
2551 "$Qd, $Qm, $imm6", vpred_r, "", pattern> {
2556 let Inst{25-23} = 0b111;
2557 let Inst{22} = Qd{3};
2559 let Inst{19-16} = imm6{3-0};
2560 let Inst{15-13} = Qd{2-0};
2561 let Inst{11-10} = 0b11;
2567 let DecoderMethod = "DecodeMVEVCVTt1fp";
2570 class MVE_VCVT_imm_asmop<int Bits> : AsmOperandClass {
2571 let PredicateMethod = "isImmediate<1," # Bits # ">";
2572 let DiagnosticString =
2573 "MVE fixed-point immediate operand must be between 1 and " # Bits;
2574 let Name = "MVEVcvtImm" # Bits;
2575 let RenderMethod = "addImmOperands";
2577 class MVE_VCVT_imm<int Bits>: Operand<i32> {
2578 let ParserMatchClass = MVE_VCVT_imm_asmop<Bits>;
2579 let EncoderMethod = "getNEONVcvtImm32OpValue";
2580 let DecoderMethod = "DecodeVCVTImmOperand";
2583 class MVE_VCVT_fix_f32<string suffix, bit U, bit op>
2584 : MVE_VCVT_fix<suffix, 0b1, U, op, MVE_VCVT_imm<32>> {
2585 let Inst{20} = imm6{4};
2587 class MVE_VCVT_fix_f16<string suffix, bit U, bit op>
2588 : MVE_VCVT_fix<suffix, 0b0, U, op, MVE_VCVT_imm<16>> {
2592 def MVE_VCVTf16s16_fix : MVE_VCVT_fix_f16<"f16.s16", 0b0, 0b0>;
2593 def MVE_VCVTs16f16_fix : MVE_VCVT_fix_f16<"s16.f16", 0b0, 0b1>;
2594 def MVE_VCVTf16u16_fix : MVE_VCVT_fix_f16<"f16.u16", 0b1, 0b0>;
2595 def MVE_VCVTu16f16_fix : MVE_VCVT_fix_f16<"u16.f16", 0b1, 0b1>;
2596 def MVE_VCVTf32s32_fix : MVE_VCVT_fix_f32<"f32.s32", 0b0, 0b0>;
2597 def MVE_VCVTs32f32_fix : MVE_VCVT_fix_f32<"s32.f32", 0b0, 0b1>;
2598 def MVE_VCVTf32u32_fix : MVE_VCVT_fix_f32<"f32.u32", 0b1, 0b0>;
2599 def MVE_VCVTu32f32_fix : MVE_VCVT_fix_f32<"u32.f32", 0b1, 0b1>;
2601 class MVE_VCVT_fp_int_anpm<string suffix, bits<2> size, bit op, string anpm,
2602 bits<2> rm, list<dag> pattern=[]>
2603 : MVE_float<!strconcat("vcvt", anpm), suffix, (outs MQPR:$Qd),
2604 (ins MQPR:$Qm), "$Qd, $Qm", vpred_r, "", pattern> {
2608 let Inst{25-23} = 0b111;
2609 let Inst{22} = Qd{3};
2610 let Inst{21-20} = 0b11;
2611 let Inst{19-18} = size;
2612 let Inst{17-16} = 0b11;
2613 let Inst{15-13} = Qd{2-0};
2614 let Inst{12-10} = 0b000;
2620 multiclass MVE_VCVT_fp_int_anpm_multi<string suffix, bits<2> size, bit op,
2621 list<dag> pattern=[]> {
2622 def a : MVE_VCVT_fp_int_anpm<suffix, size, op, "a", 0b00>;
2623 def n : MVE_VCVT_fp_int_anpm<suffix, size, op, "n", 0b01>;
2624 def p : MVE_VCVT_fp_int_anpm<suffix, size, op, "p", 0b10>;
2625 def m : MVE_VCVT_fp_int_anpm<suffix, size, op, "m", 0b11>;
2628 // This defines instructions such as MVE_VCVTu16f16a, with an explicit
2629 // rounding-mode suffix on the mnemonic. The class below will define
2630 // the bare MVE_VCVTu16f16 (with implied rounding toward zero).
2631 defm MVE_VCVTs16f16 : MVE_VCVT_fp_int_anpm_multi<"s16.f16", 0b01, 0b0>;
2632 defm MVE_VCVTu16f16 : MVE_VCVT_fp_int_anpm_multi<"u16.f16", 0b01, 0b1>;
2633 defm MVE_VCVTs32f32 : MVE_VCVT_fp_int_anpm_multi<"s32.f32", 0b10, 0b0>;
2634 defm MVE_VCVTu32f32 : MVE_VCVT_fp_int_anpm_multi<"u32.f32", 0b10, 0b1>;
2636 class MVE_VCVT_fp_int<string suffix, bits<2> size, bits<2> op,
2637 list<dag> pattern=[]>
2638 : MVE_float<"vcvt", suffix, (outs MQPR:$Qd),
2639 (ins MQPR:$Qm), "$Qd, $Qm", vpred_r, "", pattern> {
2643 let Inst{25-23} = 0b111;
2644 let Inst{22} = Qd{3};
2645 let Inst{21-20} = 0b11;
2646 let Inst{19-18} = size;
2647 let Inst{17-16} = 0b11;
2648 let Inst{15-13} = Qd{2-0};
2649 let Inst{12-9} = 0b0011;
2654 // The unsuffixed VCVT for float->int implicitly rounds toward zero,
2655 // which I reflect here in the llvm instruction names
2656 def MVE_VCVTs16f16z : MVE_VCVT_fp_int<"s16.f16", 0b01, 0b10>;
2657 def MVE_VCVTu16f16z : MVE_VCVT_fp_int<"u16.f16", 0b01, 0b11>;
2658 def MVE_VCVTs32f32z : MVE_VCVT_fp_int<"s32.f32", 0b10, 0b10>;
2659 def MVE_VCVTu32f32z : MVE_VCVT_fp_int<"u32.f32", 0b10, 0b11>;
2660 // Whereas VCVT for int->float rounds to nearest
2661 def MVE_VCVTf16s16n : MVE_VCVT_fp_int<"f16.s16", 0b01, 0b00>;
2662 def MVE_VCVTf16u16n : MVE_VCVT_fp_int<"f16.u16", 0b01, 0b01>;
2663 def MVE_VCVTf32s32n : MVE_VCVT_fp_int<"f32.s32", 0b10, 0b00>;
2664 def MVE_VCVTf32u32n : MVE_VCVT_fp_int<"f32.u32", 0b10, 0b01>;
2666 let Predicates = [HasMVEFloat] in {
2667 def : Pat<(v4i32 (fp_to_sint (v4f32 MQPR:$src))),
2668 (v4i32 (MVE_VCVTs32f32z (v4f32 MQPR:$src)))>;
2669 def : Pat<(v4i32 (fp_to_uint (v4f32 MQPR:$src))),
2670 (v4i32 (MVE_VCVTu32f32z (v4f32 MQPR:$src)))>;
2671 def : Pat<(v8i16 (fp_to_sint (v8f16 MQPR:$src))),
2672 (v8i16 (MVE_VCVTs16f16z (v8f16 MQPR:$src)))>;
2673 def : Pat<(v8i16 (fp_to_uint (v8f16 MQPR:$src))),
2674 (v8i16 (MVE_VCVTu16f16z (v8f16 MQPR:$src)))>;
2675 def : Pat<(v4f32 (sint_to_fp (v4i32 MQPR:$src))),
2676 (v4f32 (MVE_VCVTf32s32n (v4i32 MQPR:$src)))>;
2677 def : Pat<(v4f32 (uint_to_fp (v4i32 MQPR:$src))),
2678 (v4f32 (MVE_VCVTf32u32n (v4i32 MQPR:$src)))>;
2679 def : Pat<(v8f16 (sint_to_fp (v8i16 MQPR:$src))),
2680 (v8f16 (MVE_VCVTf16s16n (v8i16 MQPR:$src)))>;
2681 def : Pat<(v8f16 (uint_to_fp (v8i16 MQPR:$src))),
2682 (v8f16 (MVE_VCVTf16u16n (v8i16 MQPR:$src)))>;
2685 class MVE_VABSNEG_fp<string iname, string suffix, bits<2> size, bit negate,
2686 list<dag> pattern=[]>
2687 : MVE_float<iname, suffix, (outs MQPR:$Qd),
2688 (ins MQPR:$Qm), "$Qd, $Qm", vpred_r, "", pattern> {
2692 let Inst{25-23} = 0b111;
2693 let Inst{22} = Qd{3};
2694 let Inst{21-20} = 0b11;
2695 let Inst{19-18} = size;
2696 let Inst{17-16} = 0b01;
2697 let Inst{15-13} = Qd{2-0};
2698 let Inst{11-8} = 0b0111;
2699 let Inst{7} = negate;
2703 def MVE_VABSf16 : MVE_VABSNEG_fp<"vabs", "f16", 0b01, 0b0>;
2704 def MVE_VABSf32 : MVE_VABSNEG_fp<"vabs", "f32", 0b10, 0b0>;
2706 let Predicates = [HasMVEFloat] in {
2707 def : Pat<(v8f16 (fabs MQPR:$src)),
2708 (MVE_VABSf16 MQPR:$src)>;
2709 def : Pat<(v4f32 (fabs MQPR:$src)),
2710 (MVE_VABSf32 MQPR:$src)>;
2713 def MVE_VNEGf16 : MVE_VABSNEG_fp<"vneg", "f16", 0b01, 0b1>;
2714 def MVE_VNEGf32 : MVE_VABSNEG_fp<"vneg", "f32", 0b10, 0b1>;
2716 let Predicates = [HasMVEFloat] in {
2717 def : Pat<(v8f16 (fneg MQPR:$src)),
2718 (MVE_VNEGf16 MQPR:$src)>;
2719 def : Pat<(v4f32 (fneg MQPR:$src)),
2720 (MVE_VNEGf32 MQPR:$src)>;
2723 class MVE_VMAXMINNMA<string iname, string suffix, bit size, bit bit_12,
2724 list<dag> pattern=[]>
2725 : MVE_f<(outs MQPR:$Qd), (ins MQPR:$Qd_src, MQPR:$Qm),
2726 NoItinerary, iname, suffix, "$Qd, $Qm", vpred_n, "$Qd = $Qd_src",
2731 let Inst{28} = size;
2732 let Inst{25-23} = 0b100;
2733 let Inst{22} = Qd{3};
2734 let Inst{21-16} = 0b111111;
2735 let Inst{15-13} = Qd{2-0};
2736 let Inst{12} = bit_12;
2737 let Inst{11-6} = 0b111010;
2738 let Inst{5} = Qm{3};
2740 let Inst{3-1} = Qm{2-0};
2744 def MVE_VMAXNMAf32 : MVE_VMAXMINNMA<"vmaxnma", "f32", 0b0, 0b0>;
2745 def MVE_VMAXNMAf16 : MVE_VMAXMINNMA<"vmaxnma", "f16", 0b1, 0b0>;
2747 def MVE_VMINNMAf32 : MVE_VMAXMINNMA<"vminnma", "f32", 0b0, 0b1>;
2748 def MVE_VMINNMAf16 : MVE_VMAXMINNMA<"vminnma", "f16", 0b1, 0b1>;
2750 // end of MVE Floating Point instructions
2752 // start of MVE compares
2754 class MVE_VCMPqq<string suffix, bit bit_28, bits<2> bits_21_20,
2755 VCMPPredicateOperand predtype, list<dag> pattern=[]>
2756 : MVE_p<(outs VCCR:$P0), (ins MQPR:$Qn, MQPR:$Qm, predtype:$fc),
2757 NoItinerary, "vcmp", suffix, "$fc, $Qn, $Qm", vpred_n, "", pattern> {
2758 // Base class for comparing two vector registers
2763 let Inst{28} = bit_28;
2764 let Inst{25-22} = 0b1000;
2765 let Inst{21-20} = bits_21_20;
2766 let Inst{19-17} = Qn{2-0};
2767 let Inst{16-13} = 0b1000;
2768 let Inst{12} = fc{2};
2769 let Inst{11-8} = 0b1111;
2770 let Inst{7} = fc{0};
2772 let Inst{5} = Qm{3};
2774 let Inst{3-1} = Qm{2-0};
2775 let Inst{0} = fc{1};
2777 let Constraints = "";
2779 // We need a custom decoder method for these instructions because of
2780 // the output VCCR operand, which isn't encoded in the instruction
2781 // bits anywhere (there is only one choice for it) but has to be
2782 // included in the MC operands so that codegen will be able to track
2783 // its data flow between instructions, spill/reload it when
2784 // necessary, etc. There seems to be no way to get the Tablegen
2785 // decoder to emit an operand that isn't affected by any instruction
2787 let DecoderMethod = "DecodeMVEVCMP<false," # predtype.DecoderMethod # ">";
2790 class MVE_VCMPqqf<string suffix, bit size>
2791 : MVE_VCMPqq<suffix, size, 0b11, pred_basic_fp> {
2792 let Predicates = [HasMVEFloat];
2795 class MVE_VCMPqqi<string suffix, bits<2> size>
2796 : MVE_VCMPqq<suffix, 0b1, size, pred_basic_i> {
2801 class MVE_VCMPqqu<string suffix, bits<2> size>
2802 : MVE_VCMPqq<suffix, 0b1, size, pred_basic_u> {
2807 class MVE_VCMPqqs<string suffix, bits<2> size>
2808 : MVE_VCMPqq<suffix, 0b1, size, pred_basic_s> {
2812 def MVE_VCMPf32 : MVE_VCMPqqf<"f32", 0b0>;
2813 def MVE_VCMPf16 : MVE_VCMPqqf<"f16", 0b1>;
2815 def MVE_VCMPi8 : MVE_VCMPqqi<"i8", 0b00>;
2816 def MVE_VCMPi16 : MVE_VCMPqqi<"i16", 0b01>;
2817 def MVE_VCMPi32 : MVE_VCMPqqi<"i32", 0b10>;
2819 def MVE_VCMPu8 : MVE_VCMPqqu<"u8", 0b00>;
2820 def MVE_VCMPu16 : MVE_VCMPqqu<"u16", 0b01>;
2821 def MVE_VCMPu32 : MVE_VCMPqqu<"u32", 0b10>;
2823 def MVE_VCMPs8 : MVE_VCMPqqs<"s8", 0b00>;
2824 def MVE_VCMPs16 : MVE_VCMPqqs<"s16", 0b01>;
2825 def MVE_VCMPs32 : MVE_VCMPqqs<"s32", 0b10>;
2827 class MVE_VCMPqr<string suffix, bit bit_28, bits<2> bits_21_20,
2828 VCMPPredicateOperand predtype, list<dag> pattern=[]>
2829 : MVE_p<(outs VCCR:$P0), (ins MQPR:$Qn, GPRwithZR:$Rm, predtype:$fc),
2830 NoItinerary, "vcmp", suffix, "$fc, $Qn, $Rm", vpred_n, "", pattern> {
2831 // Base class for comparing a vector register with a scalar
2836 let Inst{28} = bit_28;
2837 let Inst{25-22} = 0b1000;
2838 let Inst{21-20} = bits_21_20;
2839 let Inst{19-17} = Qn{2-0};
2840 let Inst{16-13} = 0b1000;
2841 let Inst{12} = fc{2};
2842 let Inst{11-8} = 0b1111;
2843 let Inst{7} = fc{0};
2845 let Inst{5} = fc{1};
2847 let Inst{3-0} = Rm{3-0};
2849 let Constraints = "";
2850 // Custom decoder method, for the same reason as MVE_VCMPqq
2851 let DecoderMethod = "DecodeMVEVCMP<true," # predtype.DecoderMethod # ">";
2854 class MVE_VCMPqrf<string suffix, bit size>
2855 : MVE_VCMPqr<suffix, size, 0b11, pred_basic_fp> {
2856 let Predicates = [HasMVEFloat];
2859 class MVE_VCMPqri<string suffix, bits<2> size>
2860 : MVE_VCMPqr<suffix, 0b1, size, pred_basic_i> {
2865 class MVE_VCMPqru<string suffix, bits<2> size>
2866 : MVE_VCMPqr<suffix, 0b1, size, pred_basic_u> {
2871 class MVE_VCMPqrs<string suffix, bits<2> size>
2872 : MVE_VCMPqr<suffix, 0b1, size, pred_basic_s> {
2876 def MVE_VCMPf32r : MVE_VCMPqrf<"f32", 0b0>;
2877 def MVE_VCMPf16r : MVE_VCMPqrf<"f16", 0b1>;
2879 def MVE_VCMPi8r : MVE_VCMPqri<"i8", 0b00>;
2880 def MVE_VCMPi16r : MVE_VCMPqri<"i16", 0b01>;
2881 def MVE_VCMPi32r : MVE_VCMPqri<"i32", 0b10>;
2883 def MVE_VCMPu8r : MVE_VCMPqru<"u8", 0b00>;
2884 def MVE_VCMPu16r : MVE_VCMPqru<"u16", 0b01>;
2885 def MVE_VCMPu32r : MVE_VCMPqru<"u32", 0b10>;
2887 def MVE_VCMPs8r : MVE_VCMPqrs<"s8", 0b00>;
2888 def MVE_VCMPs16r : MVE_VCMPqrs<"s16", 0b01>;
2889 def MVE_VCMPs32r : MVE_VCMPqrs<"s32", 0b10>;
2891 // end of MVE compares
2893 // start of MVE_qDest_qSrc
2895 class MVE_qDest_qSrc<string iname, string suffix, dag oops, dag iops,
2896 string ops, vpred_ops vpred, string cstr,
2897 list<dag> pattern=[]>
2898 : MVE_p<oops, iops, NoItinerary, iname, suffix,
2899 ops, vpred, cstr, pattern> {
2903 let Inst{25-23} = 0b100;
2904 let Inst{22} = Qd{3};
2905 let Inst{15-13} = Qd{2-0};
2906 let Inst{11-9} = 0b111;
2908 let Inst{5} = Qm{3};
2910 let Inst{3-1} = Qm{2-0};
2913 class MVE_VQxDMLxDH<string iname, bit exch, bit round, bit subtract,
2914 string suffix, bits<2> size, list<dag> pattern=[]>
2915 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
2916 (ins MQPR:$Qd_src, MQPR:$Qn, MQPR:$Qm), "$Qd, $Qn, $Qm",
2917 vpred_n, "$Qd = $Qd_src", pattern> {
2920 let Inst{28} = subtract;
2921 let Inst{21-20} = size;
2922 let Inst{19-17} = Qn{2-0};
2924 let Inst{12} = exch;
2926 let Inst{7} = Qn{3};
2927 let Inst{0} = round;
2930 multiclass MVE_VQxDMLxDH_multi<string iname, bit exch,
2931 bit round, bit subtract> {
2932 def s8 : MVE_VQxDMLxDH<iname, exch, round, subtract, "s8", 0b00>;
2933 def s16 : MVE_VQxDMLxDH<iname, exch, round, subtract, "s16", 0b01>;
2934 def s32 : MVE_VQxDMLxDH<iname, exch, round, subtract, "s32", 0b10>;
2937 defm MVE_VQDMLADH : MVE_VQxDMLxDH_multi<"vqdmladh", 0b0, 0b0, 0b0>;
2938 defm MVE_VQDMLADHX : MVE_VQxDMLxDH_multi<"vqdmladhx", 0b1, 0b0, 0b0>;
2939 defm MVE_VQRDMLADH : MVE_VQxDMLxDH_multi<"vqrdmladh", 0b0, 0b1, 0b0>;
2940 defm MVE_VQRDMLADHX : MVE_VQxDMLxDH_multi<"vqrdmladhx", 0b1, 0b1, 0b0>;
2941 defm MVE_VQDMLSDH : MVE_VQxDMLxDH_multi<"vqdmlsdh", 0b0, 0b0, 0b1>;
2942 defm MVE_VQDMLSDHX : MVE_VQxDMLxDH_multi<"vqdmlsdhx", 0b1, 0b0, 0b1>;
2943 defm MVE_VQRDMLSDH : MVE_VQxDMLxDH_multi<"vqrdmlsdh", 0b0, 0b1, 0b1>;
2944 defm MVE_VQRDMLSDHX : MVE_VQxDMLxDH_multi<"vqrdmlsdhx", 0b1, 0b1, 0b1>;
2946 class MVE_VCMUL<string iname, string suffix, bit size, list<dag> pattern=[]>
2947 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
2948 (ins MQPR:$Qn, MQPR:$Qm, complexrotateop:$rot),
2949 "$Qd, $Qn, $Qm, $rot", vpred_r, "", pattern> {
2953 let Inst{28} = size;
2954 let Inst{21-20} = 0b11;
2955 let Inst{19-17} = Qn{2-0};
2957 let Inst{12} = rot{1};
2959 let Inst{7} = Qn{3};
2960 let Inst{0} = rot{0};
2962 let Predicates = [HasMVEFloat];
2965 def MVE_VCMULf16 : MVE_VCMUL<"vcmul", "f16", 0b0>;
2966 def MVE_VCMULf32 : MVE_VCMUL<"vcmul", "f32", 0b1>;
2968 class MVE_VMULL<string iname, string suffix, bit bit_28, bits<2> bits_21_20,
2969 bit T, list<dag> pattern=[]>
2970 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
2971 (ins MQPR:$Qn, MQPR:$Qm), "$Qd, $Qn, $Qm",
2972 vpred_r, "", pattern> {
2977 let Inst{28} = bit_28;
2978 let Inst{21-20} = bits_21_20;
2979 let Inst{19-17} = Qn{2-0};
2983 let Inst{7} = Qn{3};
2987 multiclass MVE_VMULL_multi<string iname, string suffix,
2988 bit bit_28, bits<2> bits_21_20> {
2989 def bh : MVE_VMULL<iname # "b", suffix, bit_28, bits_21_20, 0b0>;
2990 def th : MVE_VMULL<iname # "t", suffix, bit_28, bits_21_20, 0b1>;
2993 // For integer multiplies, bits 21:20 encode size, and bit 28 signedness.
2994 // For polynomial multiplies, bits 21:20 take the unused value 0b11, and
2995 // bit 28 switches to encoding the size.
2997 defm MVE_VMULLs8 : MVE_VMULL_multi<"vmull", "s8", 0b0, 0b00>;
2998 defm MVE_VMULLs16 : MVE_VMULL_multi<"vmull", "s16", 0b0, 0b01>;
2999 defm MVE_VMULLs32 : MVE_VMULL_multi<"vmull", "s32", 0b0, 0b10>;
3000 defm MVE_VMULLu8 : MVE_VMULL_multi<"vmull", "u8", 0b1, 0b00>;
3001 defm MVE_VMULLu16 : MVE_VMULL_multi<"vmull", "u16", 0b1, 0b01>;
3002 defm MVE_VMULLu32 : MVE_VMULL_multi<"vmull", "u32", 0b1, 0b10>;
3003 defm MVE_VMULLp8 : MVE_VMULL_multi<"vmull", "p8", 0b0, 0b11>;
3004 defm MVE_VMULLp16 : MVE_VMULL_multi<"vmull", "p16", 0b1, 0b11>;
3006 class MVE_VxMULH<string iname, string suffix, bit U, bits<2> size,
3007 bit round, list<dag> pattern=[]>
3008 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
3009 (ins MQPR:$Qn, MQPR:$Qm), "$Qd, $Qn, $Qm",
3010 vpred_r, "", pattern> {
3014 let Inst{21-20} = size;
3015 let Inst{19-17} = Qn{2-0};
3017 let Inst{12} = round;
3019 let Inst{7} = Qn{3};
3023 def MVE_VMULHs8 : MVE_VxMULH<"vmulh", "s8", 0b0, 0b00, 0b0>;
3024 def MVE_VMULHs16 : MVE_VxMULH<"vmulh", "s16", 0b0, 0b01, 0b0>;
3025 def MVE_VMULHs32 : MVE_VxMULH<"vmulh", "s32", 0b0, 0b10, 0b0>;
3026 def MVE_VMULHu8 : MVE_VxMULH<"vmulh", "u8", 0b1, 0b00, 0b0>;
3027 def MVE_VMULHu16 : MVE_VxMULH<"vmulh", "u16", 0b1, 0b01, 0b0>;
3028 def MVE_VMULHu32 : MVE_VxMULH<"vmulh", "u32", 0b1, 0b10, 0b0>;
3030 def MVE_VRMULHs8 : MVE_VxMULH<"vrmulh", "s8", 0b0, 0b00, 0b1>;
3031 def MVE_VRMULHs16 : MVE_VxMULH<"vrmulh", "s16", 0b0, 0b01, 0b1>;
3032 def MVE_VRMULHs32 : MVE_VxMULH<"vrmulh", "s32", 0b0, 0b10, 0b1>;
3033 def MVE_VRMULHu8 : MVE_VxMULH<"vrmulh", "u8", 0b1, 0b00, 0b1>;
3034 def MVE_VRMULHu16 : MVE_VxMULH<"vrmulh", "u16", 0b1, 0b01, 0b1>;
3035 def MVE_VRMULHu32 : MVE_VxMULH<"vrmulh", "u32", 0b1, 0b10, 0b1>;
3037 class MVE_VxMOVxN<string iname, string suffix, bit bit_28, bit bit_17,
3038 bits<2> size, bit T, list<dag> pattern=[]>
3039 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
3040 (ins MQPR:$Qd_src, MQPR:$Qm), "$Qd, $Qm",
3041 vpred_n, "$Qd = $Qd_src", pattern> {
3043 let Inst{28} = bit_28;
3044 let Inst{21-20} = 0b11;
3045 let Inst{19-18} = size;
3046 let Inst{17} = bit_17;
3050 let Inst{7} = !if(!eq(bit_17, 0), 1, 0);
3054 multiclass MVE_VxMOVxN_halves<string iname, string suffix,
3055 bit bit_28, bit bit_17, bits<2> size> {
3056 def bh : MVE_VxMOVxN<iname # "b", suffix, bit_28, bit_17, size, 0b0>;
3057 def th : MVE_VxMOVxN<iname # "t", suffix, bit_28, bit_17, size, 0b1>;
3060 defm MVE_VMOVNi16 : MVE_VxMOVxN_halves<"vmovn", "i16", 0b1, 0b0, 0b00>;
3061 defm MVE_VMOVNi32 : MVE_VxMOVxN_halves<"vmovn", "i32", 0b1, 0b0, 0b01>;
3062 defm MVE_VQMOVNs16 : MVE_VxMOVxN_halves<"vqmovn", "s16", 0b0, 0b1, 0b00>;
3063 defm MVE_VQMOVNs32 : MVE_VxMOVxN_halves<"vqmovn", "s32", 0b0, 0b1, 0b01>;
3064 defm MVE_VQMOVNu16 : MVE_VxMOVxN_halves<"vqmovn", "u16", 0b1, 0b1, 0b00>;
3065 defm MVE_VQMOVNu32 : MVE_VxMOVxN_halves<"vqmovn", "u32", 0b1, 0b1, 0b01>;
3066 defm MVE_VQMOVUNs16 : MVE_VxMOVxN_halves<"vqmovun", "s16", 0b0, 0b0, 0b00>;
3067 defm MVE_VQMOVUNs32 : MVE_VxMOVxN_halves<"vqmovun", "s32", 0b0, 0b0, 0b01>;
3069 class MVE_VCVT_ff<string iname, string suffix, bit op, bit T,
3070 list<dag> pattern=[]>
3071 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd), (ins MQPR:$Qd_src, MQPR:$Qm),
3072 "$Qd, $Qm", vpred_n, "$Qd = $Qd_src", pattern> {
3074 let Inst{21-16} = 0b111111;
3076 let Inst{8-7} = 0b00;
3079 let Predicates = [HasMVEFloat];
3082 multiclass MVE_VCVT_ff_halves<string suffix, bit op> {
3083 def bh : MVE_VCVT_ff<"vcvtb", suffix, op, 0b0>;
3084 def th : MVE_VCVT_ff<"vcvtt", suffix, op, 0b1>;
3087 defm MVE_VCVTf16f32 : MVE_VCVT_ff_halves<"f16.f32", 0b0>;
3088 defm MVE_VCVTf32f16 : MVE_VCVT_ff_halves<"f32.f16", 0b1>;
3090 class MVE_VxCADD<string iname, string suffix, bits<2> size, bit halve,
3091 list<dag> pattern=[]>
3092 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
3093 (ins MQPR:$Qn, MQPR:$Qm, complexrotateopodd:$rot),
3094 "$Qd, $Qn, $Qm, $rot", vpred_r, "",
3099 let Inst{28} = halve;
3100 let Inst{21-20} = size;
3101 let Inst{19-17} = Qn{2-0};
3105 let Inst{7} = Qn{3};
3109 def MVE_VCADDi8 : MVE_VxCADD<"vcadd", "i8", 0b00, 0b1>;
3110 def MVE_VCADDi16 : MVE_VxCADD<"vcadd", "i16", 0b01, 0b1>;
3111 def MVE_VCADDi32 : MVE_VxCADD<"vcadd", "i32", 0b10, 0b1>;
3113 def MVE_VHCADDs8 : MVE_VxCADD<"vhcadd", "s8", 0b00, 0b0>;
3114 def MVE_VHCADDs16 : MVE_VxCADD<"vhcadd", "s16", 0b01, 0b0>;
3115 def MVE_VHCADDs32 : MVE_VxCADD<"vhcadd", "s32", 0b10, 0b0>;
3117 class MVE_VADCSBC<string iname, bit I, bit subtract,
3118 dag carryin, list<dag> pattern=[]>
3119 : MVE_qDest_qSrc<iname, "i32", (outs MQPR:$Qd, cl_FPSCR_NZCV:$carryout),
3120 !con((ins MQPR:$Qn, MQPR:$Qm), carryin),
3121 "$Qd, $Qn, $Qm", vpred_r, "", pattern> {
3124 let Inst{28} = subtract;
3125 let Inst{21-20} = 0b11;
3126 let Inst{19-17} = Qn{2-0};
3130 let Inst{7} = Qn{3};
3133 // Custom decoder method in order to add the FPSCR operand(s), which
3134 // Tablegen won't do right
3135 let DecoderMethod = "DecodeMVEVADCInstruction";
3138 def MVE_VADC : MVE_VADCSBC<"vadc", 0b0, 0b0, (ins cl_FPSCR_NZCV:$carryin)>;
3139 def MVE_VADCI : MVE_VADCSBC<"vadci", 0b1, 0b0, (ins)>;
3141 def MVE_VSBC : MVE_VADCSBC<"vsbc", 0b0, 0b1, (ins cl_FPSCR_NZCV:$carryin)>;
3142 def MVE_VSBCI : MVE_VADCSBC<"vsbci", 0b1, 0b1, (ins)>;
3144 class MVE_VQDMULL<string iname, string suffix, bit size, bit T,
3145 list<dag> pattern=[]>
3146 : MVE_qDest_qSrc<iname, suffix, (outs MQPR:$Qd),
3147 (ins MQPR:$Qn, MQPR:$Qm), "$Qd, $Qn, $Qm",
3148 vpred_r, "", pattern> {
3151 let Inst{28} = size;
3152 let Inst{21-20} = 0b11;
3153 let Inst{19-17} = Qn{2-0};
3157 let Inst{7} = Qn{3};
3161 multiclass MVE_VQDMULL_halves<string suffix, bit size> {
3162 def bh : MVE_VQDMULL<"vqdmullb", suffix, size, 0b0>;
3163 def th : MVE_VQDMULL<"vqdmullt", suffix, size, 0b1>;
3166 defm MVE_VQDMULLs16 : MVE_VQDMULL_halves<"s16", 0b0>;
3167 defm MVE_VQDMULLs32 : MVE_VQDMULL_halves<"s32", 0b1>;
3169 // end of mve_qDest_qSrc
3171 // start of mve_qDest_rSrc
3173 class MVE_qr_base<dag oops, dag iops, InstrItinClass itin, string iname,
3174 string suffix, string ops, vpred_ops vpred, string cstr,
3175 list<dag> pattern=[]>
3176 : MVE_p<oops, iops, NoItinerary, iname, suffix, ops, vpred, cstr, pattern> {
3181 let Inst{25-23} = 0b100;
3182 let Inst{22} = Qd{3};
3183 let Inst{19-17} = Qn{2-0};
3184 let Inst{15-13} = Qd{2-0};
3185 let Inst{11-9} = 0b111;
3186 let Inst{7} = Qn{3};
3189 let Inst{3-0} = Rm{3-0};
3192 class MVE_qDest_rSrc<string iname, string suffix, list<dag> pattern=[]>
3193 : MVE_qr_base<(outs MQPR:$Qd), (ins MQPR:$Qn, rGPR:$Rm),
3194 NoItinerary, iname, suffix, "$Qd, $Qn, $Rm", vpred_r, "",
3197 class MVE_qDestSrc_rSrc<string iname, string suffix, list<dag> pattern=[]>
3198 : MVE_qr_base<(outs MQPR:$Qd), (ins MQPR:$Qd_src, MQPR:$Qn, rGPR:$Rm),
3199 NoItinerary, iname, suffix, "$Qd, $Qn, $Rm", vpred_n, "$Qd = $Qd_src",
3202 class MVE_qDest_single_rSrc<string iname, string suffix, list<dag> pattern=[]>
3203 : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qd_src, rGPR:$Rm), NoItinerary, iname,
3204 suffix, "$Qd, $Rm", vpred_n, "$Qd = $Qd_src", pattern> {
3208 let Inst{22} = Qd{3};
3209 let Inst{15-13} = Qd{2-0};
3210 let Inst{3-0} = Rm{3-0};
3213 class MVE_VADDSUB_qr<string iname, string suffix, bits<2> size,
3214 bit bit_5, bit bit_12, bit bit_16,
3215 bit bit_28, list<dag> pattern=[]>
3216 : MVE_qDest_rSrc<iname, suffix, pattern> {
3218 let Inst{28} = bit_28;
3219 let Inst{21-20} = size;
3220 let Inst{16} = bit_16;
3221 let Inst{12} = bit_12;
3223 let Inst{5} = bit_5;
3226 multiclass MVE_VADDSUB_qr_sizes<string iname, string suffix,
3227 bit bit_5, bit bit_12, bit bit_16,
3228 bit bit_28, list<dag> pattern=[]> {
3229 def "8" : MVE_VADDSUB_qr<iname, suffix#"8", 0b00,
3230 bit_5, bit_12, bit_16, bit_28>;
3231 def "16" : MVE_VADDSUB_qr<iname, suffix#"16", 0b01,
3232 bit_5, bit_12, bit_16, bit_28>;
3233 def "32" : MVE_VADDSUB_qr<iname, suffix#"32", 0b10,
3234 bit_5, bit_12, bit_16, bit_28>;
3237 defm MVE_VADD_qr_i : MVE_VADDSUB_qr_sizes<"vadd", "i", 0b0, 0b0, 0b1, 0b0>;
3238 defm MVE_VQADD_qr_s : MVE_VADDSUB_qr_sizes<"vqadd", "s", 0b1, 0b0, 0b0, 0b0>;
3239 defm MVE_VQADD_qr_u : MVE_VADDSUB_qr_sizes<"vqadd", "u", 0b1, 0b0, 0b0, 0b1>;
3241 defm MVE_VSUB_qr_i : MVE_VADDSUB_qr_sizes<"vsub", "i", 0b0, 0b1, 0b1, 0b0>;
3242 defm MVE_VQSUB_qr_s : MVE_VADDSUB_qr_sizes<"vqsub", "s", 0b1, 0b1, 0b0, 0b0>;
3243 defm MVE_VQSUB_qr_u : MVE_VADDSUB_qr_sizes<"vqsub", "u", 0b1, 0b1, 0b0, 0b1>;
3245 class MVE_VQDMULL_qr<string iname, string suffix, bit size,
3246 bit T, list<dag> pattern=[]>
3247 : MVE_qDest_rSrc<iname, suffix, pattern> {
3249 let Inst{28} = size;
3250 let Inst{21-20} = 0b11;
3257 multiclass MVE_VQDMULL_qr_halves<string suffix, bit size> {
3258 def bh : MVE_VQDMULL_qr<"vqdmullb", suffix, size, 0b0>;
3259 def th : MVE_VQDMULL_qr<"vqdmullt", suffix, size, 0b1>;
3262 defm MVE_VQDMULL_qr_s16 : MVE_VQDMULL_qr_halves<"s16", 0b0>;
3263 defm MVE_VQDMULL_qr_s32 : MVE_VQDMULL_qr_halves<"s32", 0b1>;
3265 class MVE_VxADDSUB_qr<string iname, string suffix,
3266 bit bit_28, bits<2> bits_21_20, bit subtract,
3267 list<dag> pattern=[]>
3268 : MVE_qDest_rSrc<iname, suffix, pattern> {
3270 let Inst{28} = bit_28;
3271 let Inst{21-20} = bits_21_20;
3273 let Inst{12} = subtract;
3278 def MVE_VHADD_qr_s8 : MVE_VxADDSUB_qr<"vhadd", "s8", 0b0, 0b00, 0b0>;
3279 def MVE_VHADD_qr_s16 : MVE_VxADDSUB_qr<"vhadd", "s16", 0b0, 0b01, 0b0>;
3280 def MVE_VHADD_qr_s32 : MVE_VxADDSUB_qr<"vhadd", "s32", 0b0, 0b10, 0b0>;
3281 def MVE_VHADD_qr_u8 : MVE_VxADDSUB_qr<"vhadd", "u8", 0b1, 0b00, 0b0>;
3282 def MVE_VHADD_qr_u16 : MVE_VxADDSUB_qr<"vhadd", "u16", 0b1, 0b01, 0b0>;
3283 def MVE_VHADD_qr_u32 : MVE_VxADDSUB_qr<"vhadd", "u32", 0b1, 0b10, 0b0>;
3285 def MVE_VHSUB_qr_s8 : MVE_VxADDSUB_qr<"vhsub", "s8", 0b0, 0b00, 0b1>;
3286 def MVE_VHSUB_qr_s16 : MVE_VxADDSUB_qr<"vhsub", "s16", 0b0, 0b01, 0b1>;
3287 def MVE_VHSUB_qr_s32 : MVE_VxADDSUB_qr<"vhsub", "s32", 0b0, 0b10, 0b1>;
3288 def MVE_VHSUB_qr_u8 : MVE_VxADDSUB_qr<"vhsub", "u8", 0b1, 0b00, 0b1>;
3289 def MVE_VHSUB_qr_u16 : MVE_VxADDSUB_qr<"vhsub", "u16", 0b1, 0b01, 0b1>;
3290 def MVE_VHSUB_qr_u32 : MVE_VxADDSUB_qr<"vhsub", "u32", 0b1, 0b10, 0b1>;
3292 let Predicates = [HasMVEFloat] in {
3293 def MVE_VADD_qr_f32 : MVE_VxADDSUB_qr<"vadd", "f32", 0b0, 0b11, 0b0>;
3294 def MVE_VADD_qr_f16 : MVE_VxADDSUB_qr<"vadd", "f16", 0b1, 0b11, 0b0>;
3296 def MVE_VSUB_qr_f32 : MVE_VxADDSUB_qr<"vsub", "f32", 0b0, 0b11, 0b1>;
3297 def MVE_VSUB_qr_f16 : MVE_VxADDSUB_qr<"vsub", "f16", 0b1, 0b11, 0b1>;
3300 class MVE_VxSHL_qr<string iname, string suffix, bit U, bits<2> size,
3301 bit bit_7, bit bit_17, list<dag> pattern=[]>
3302 : MVE_qDest_single_rSrc<iname, suffix, pattern> {
3305 let Inst{25-23} = 0b100;
3306 let Inst{21-20} = 0b11;
3307 let Inst{19-18} = size;
3308 let Inst{17} = bit_17;
3310 let Inst{12-8} = 0b11110;
3311 let Inst{7} = bit_7;
3312 let Inst{6-4} = 0b110;
3315 multiclass MVE_VxSHL_qr_types<string iname, bit bit_7, bit bit_17> {
3316 def s8 : MVE_VxSHL_qr<iname, "s8", 0b0, 0b00, bit_7, bit_17>;
3317 def s16 : MVE_VxSHL_qr<iname, "s16", 0b0, 0b01, bit_7, bit_17>;
3318 def s32 : MVE_VxSHL_qr<iname, "s32", 0b0, 0b10, bit_7, bit_17>;
3319 def u8 : MVE_VxSHL_qr<iname, "u8", 0b1, 0b00, bit_7, bit_17>;
3320 def u16 : MVE_VxSHL_qr<iname, "u16", 0b1, 0b01, bit_7, bit_17>;
3321 def u32 : MVE_VxSHL_qr<iname, "u32", 0b1, 0b10, bit_7, bit_17>;
3324 defm MVE_VSHL_qr : MVE_VxSHL_qr_types<"vshl", 0b0, 0b0>;
3325 defm MVE_VRSHL_qr : MVE_VxSHL_qr_types<"vrshl", 0b0, 0b1>;
3326 defm MVE_VQSHL_qr : MVE_VxSHL_qr_types<"vqshl", 0b1, 0b0>;
3327 defm MVE_VQRSHL_qr : MVE_VxSHL_qr_types<"vqrshl", 0b1, 0b1>;
3329 class MVE_VBRSR<string iname, string suffix, bits<2> size, list<dag> pattern=[]>
3330 : MVE_qDest_rSrc<iname, suffix, pattern> {
3333 let Inst{21-20} = size;
3340 def MVE_VBRSR8 : MVE_VBRSR<"vbrsr", "8", 0b00>;
3341 def MVE_VBRSR16 : MVE_VBRSR<"vbrsr", "16", 0b01>;
3342 def MVE_VBRSR32 : MVE_VBRSR<"vbrsr", "32", 0b10>;
3344 class MVE_VMUL_qr_int<string iname, string suffix,
3345 bits<2> size, list<dag> pattern=[]>
3346 : MVE_qDest_rSrc<iname, suffix, pattern> {
3349 let Inst{21-20} = size;
3356 def MVE_VMUL_qr_i8 : MVE_VMUL_qr_int<"vmul", "i8", 0b00>;
3357 def MVE_VMUL_qr_i16 : MVE_VMUL_qr_int<"vmul", "i16", 0b01>;
3358 def MVE_VMUL_qr_i32 : MVE_VMUL_qr_int<"vmul", "i32", 0b10>;
3360 class MVE_VxxMUL_qr<string iname, string suffix,
3361 bit bit_28, bits<2> bits_21_20, list<dag> pattern=[]>
3362 : MVE_qDest_rSrc<iname, suffix, pattern> {
3364 let Inst{28} = bit_28;
3365 let Inst{21-20} = bits_21_20;
3372 def MVE_VQDMULH_qr_s8 : MVE_VxxMUL_qr<"vqdmulh", "s8", 0b0, 0b00>;
3373 def MVE_VQDMULH_qr_s16 : MVE_VxxMUL_qr<"vqdmulh", "s16", 0b0, 0b01>;
3374 def MVE_VQDMULH_qr_s32 : MVE_VxxMUL_qr<"vqdmulh", "s32", 0b0, 0b10>;
3376 def MVE_VQRDMULH_qr_s8 : MVE_VxxMUL_qr<"vqrdmulh", "s8", 0b1, 0b00>;
3377 def MVE_VQRDMULH_qr_s16 : MVE_VxxMUL_qr<"vqrdmulh", "s16", 0b1, 0b01>;
3378 def MVE_VQRDMULH_qr_s32 : MVE_VxxMUL_qr<"vqrdmulh", "s32", 0b1, 0b10>;
3380 let Predicates = [HasMVEFloat] in {
3381 def MVE_VMUL_qr_f16 : MVE_VxxMUL_qr<"vmul", "f16", 0b1, 0b11>;
3382 def MVE_VMUL_qr_f32 : MVE_VxxMUL_qr<"vmul", "f32", 0b0, 0b11>;
3385 class MVE_VFMAMLA_qr<string iname, string suffix,
3386 bit bit_28, bits<2> bits_21_20, bit S,
3387 list<dag> pattern=[]>
3388 : MVE_qDestSrc_rSrc<iname, suffix, pattern> {
3390 let Inst{28} = bit_28;
3391 let Inst{21-20} = bits_21_20;
3398 def MVE_VMLA_qr_s8 : MVE_VFMAMLA_qr<"vmla", "s8", 0b0, 0b00, 0b0>;
3399 def MVE_VMLA_qr_s16 : MVE_VFMAMLA_qr<"vmla", "s16", 0b0, 0b01, 0b0>;
3400 def MVE_VMLA_qr_s32 : MVE_VFMAMLA_qr<"vmla", "s32", 0b0, 0b10, 0b0>;
3401 def MVE_VMLA_qr_u8 : MVE_VFMAMLA_qr<"vmla", "u8", 0b1, 0b00, 0b0>;
3402 def MVE_VMLA_qr_u16 : MVE_VFMAMLA_qr<"vmla", "u16", 0b1, 0b01, 0b0>;
3403 def MVE_VMLA_qr_u32 : MVE_VFMAMLA_qr<"vmla", "u32", 0b1, 0b10, 0b0>;
3405 def MVE_VMLAS_qr_s8 : MVE_VFMAMLA_qr<"vmlas", "s8", 0b0, 0b00, 0b1>;
3406 def MVE_VMLAS_qr_s16 : MVE_VFMAMLA_qr<"vmlas", "s16", 0b0, 0b01, 0b1>;
3407 def MVE_VMLAS_qr_s32 : MVE_VFMAMLA_qr<"vmlas", "s32", 0b0, 0b10, 0b1>;
3408 def MVE_VMLAS_qr_u8 : MVE_VFMAMLA_qr<"vmlas", "u8", 0b1, 0b00, 0b1>;
3409 def MVE_VMLAS_qr_u16 : MVE_VFMAMLA_qr<"vmlas", "u16", 0b1, 0b01, 0b1>;
3410 def MVE_VMLAS_qr_u32 : MVE_VFMAMLA_qr<"vmlas", "u32", 0b1, 0b10, 0b1>;
3412 let Predicates = [HasMVEFloat] in {
3413 def MVE_VFMA_qr_f16 : MVE_VFMAMLA_qr<"vfma", "f16", 0b1, 0b11, 0b0>;
3414 def MVE_VFMA_qr_f32 : MVE_VFMAMLA_qr<"vfma", "f32", 0b0, 0b11, 0b0>;
3415 def MVE_VFMA_qr_Sf16 : MVE_VFMAMLA_qr<"vfmas", "f16", 0b1, 0b11, 0b1>;
3416 def MVE_VFMA_qr_Sf32 : MVE_VFMAMLA_qr<"vfmas", "f32", 0b0, 0b11, 0b1>;
3419 class MVE_VQDMLAH_qr<string iname, string suffix, bit U, bits<2> size,
3420 bit bit_5, bit bit_12, list<dag> pattern=[]>
3421 : MVE_qDestSrc_rSrc<iname, suffix, pattern> {
3424 let Inst{21-20} = size;
3426 let Inst{12} = bit_12;
3428 let Inst{5} = bit_5;
3431 multiclass MVE_VQDMLAH_qr_types<string iname, bit bit_5, bit bit_12> {
3432 def s8 : MVE_VQDMLAH_qr<iname, "s8", 0b0, 0b00, bit_5, bit_12>;
3433 def s16 : MVE_VQDMLAH_qr<iname, "s16", 0b0, 0b01, bit_5, bit_12>;
3434 def s32 : MVE_VQDMLAH_qr<iname, "s32", 0b0, 0b10, bit_5, bit_12>;
3437 defm MVE_VQDMLAH_qr : MVE_VQDMLAH_qr_types<"vqdmlah", 0b1, 0b0>;
3438 defm MVE_VQRDMLAH_qr : MVE_VQDMLAH_qr_types<"vqrdmlah", 0b0, 0b0>;
3439 defm MVE_VQDMLASH_qr : MVE_VQDMLAH_qr_types<"vqdmlash", 0b1, 0b1>;
3440 defm MVE_VQRDMLASH_qr : MVE_VQDMLAH_qr_types<"vqrdmlash", 0b0, 0b1>;
3442 class MVE_VxDUP<string iname, string suffix, bits<2> size, bit bit_12,
3443 list<dag> pattern=[]>
3444 : MVE_p<(outs MQPR:$Qd, tGPREven:$Rn),
3445 (ins tGPREven:$Rn_src, MVE_VIDUP_imm:$imm), NoItinerary,
3446 iname, suffix, "$Qd, $Rn, $imm", vpred_r, "$Rn = $Rn_src",
3453 let Inst{25-23} = 0b100;
3454 let Inst{22} = Qd{3};
3455 let Inst{21-20} = size;
3456 let Inst{19-17} = Rn{3-1};
3458 let Inst{15-13} = Qd{2-0};
3459 let Inst{12} = bit_12;
3460 let Inst{11-8} = 0b1111;
3461 let Inst{7} = imm{1};
3462 let Inst{6-1} = 0b110111;
3463 let Inst{0} = imm{0};
3466 def MVE_VIDUPu8 : MVE_VxDUP<"vidup", "u8", 0b00, 0b0>;
3467 def MVE_VIDUPu16 : MVE_VxDUP<"vidup", "u16", 0b01, 0b0>;
3468 def MVE_VIDUPu32 : MVE_VxDUP<"vidup", "u32", 0b10, 0b0>;
3470 def MVE_VDDUPu8 : MVE_VxDUP<"vddup", "u8", 0b00, 0b1>;
3471 def MVE_VDDUPu16 : MVE_VxDUP<"vddup", "u16", 0b01, 0b1>;
3472 def MVE_VDDUPu32 : MVE_VxDUP<"vddup", "u32", 0b10, 0b1>;
3474 class MVE_VxWDUP<string iname, string suffix, bits<2> size, bit bit_12,
3475 list<dag> pattern=[]>
3476 : MVE_p<(outs MQPR:$Qd, tGPREven:$Rn),
3477 (ins tGPREven:$Rn_src, tGPROdd:$Rm, MVE_VIDUP_imm:$imm), NoItinerary,
3478 iname, suffix, "$Qd, $Rn, $Rm, $imm", vpred_r, "$Rn = $Rn_src",
3486 let Inst{25-23} = 0b100;
3487 let Inst{22} = Qd{3};
3488 let Inst{21-20} = size;
3489 let Inst{19-17} = Rn{3-1};
3491 let Inst{15-13} = Qd{2-0};
3492 let Inst{12} = bit_12;
3493 let Inst{11-8} = 0b1111;
3494 let Inst{7} = imm{1};
3495 let Inst{6-4} = 0b110;
3496 let Inst{3-1} = Rm{3-1};
3497 let Inst{0} = imm{0};
3500 def MVE_VIWDUPu8 : MVE_VxWDUP<"viwdup", "u8", 0b00, 0b0>;
3501 def MVE_VIWDUPu16 : MVE_VxWDUP<"viwdup", "u16", 0b01, 0b0>;
3502 def MVE_VIWDUPu32 : MVE_VxWDUP<"viwdup", "u32", 0b10, 0b0>;
3504 def MVE_VDWDUPu8 : MVE_VxWDUP<"vdwdup", "u8", 0b00, 0b1>;
3505 def MVE_VDWDUPu16 : MVE_VxWDUP<"vdwdup", "u16", 0b01, 0b1>;
3506 def MVE_VDWDUPu32 : MVE_VxWDUP<"vdwdup", "u32", 0b10, 0b1>;
3508 class MVE_VCTP<string suffix, bits<2> size, list<dag> pattern=[]>
3509 : MVE_p<(outs VCCR:$P0), (ins rGPR:$Rn), NoItinerary, "vctp", suffix,
3510 "$Rn", vpred_n, "", pattern> {
3513 let Inst{28-27} = 0b10;
3514 let Inst{26-22} = 0b00000;
3515 let Inst{21-20} = size;
3516 let Inst{19-16} = Rn{3-0};
3517 let Inst{15-11} = 0b11101;
3518 let Inst{10-0} = 0b00000000001;
3519 let Unpredictable{10-0} = 0b11111111111;
3521 let Constraints = "";
3522 let DecoderMethod = "DecodeMveVCTP";
3525 def MVE_VCTP8 : MVE_VCTP<"8", 0b00>;
3526 def MVE_VCTP16 : MVE_VCTP<"16", 0b01>;
3527 def MVE_VCTP32 : MVE_VCTP<"32", 0b10>;
3528 def MVE_VCTP64 : MVE_VCTP<"64", 0b11>;
3530 // end of mve_qDest_rSrc
3532 // start of coproc mov
3534 class MVE_VMOV_64bit<dag oops, dag iops, bit to_qreg, string ops, string cstr>
3535 : MVE_VMOV_lane_base<oops, !con(iops, (ins MVEPairVectorIndex2:$idx,
3536 MVEPairVectorIndex0:$idx2)),
3537 NoItinerary, "vmov", "", ops, cstr, []> {
3544 let Inst{31-23} = 0b111011000;
3545 let Inst{22} = Qd{3};
3547 let Inst{20} = to_qreg;
3548 let Inst{19-16} = Rt2{3-0};
3549 let Inst{15-13} = Qd{2-0};
3550 let Inst{12-5} = 0b01111000;
3552 let Inst{3-0} = Rt{3-0};
3555 // The assembly syntax for these instructions mentions the vector
3556 // register name twice, e.g.
3558 // vmov q2[2], q2[0], r0, r1
3559 // vmov r0, r1, q2[2], q2[0]
3561 // which needs a bit of juggling with MC operand handling.
3563 // For the move _into_ a vector register, the MC operand list also has
3564 // to mention the register name twice: once as the output, and once as
3565 // an extra input to represent where the unchanged half of the output
3566 // register comes from (when this instruction is used in code
3567 // generation). So we arrange that the first mention of the vector reg
3568 // in the instruction is considered by the AsmMatcher to be the output
3569 // ($Qd), and the second one is the input ($QdSrc). Binding them
3570 // together with the existing 'tie' constraint is enough to enforce at
3571 // register allocation time that they have to be the same register.
3573 // For the move _from_ a vector register, there's no way to get round
3574 // the fact that both instances of that register name have to be
3575 // inputs. They have to be the same register again, but this time, we
3576 // can't use a tie constraint, because that has to be between an
3577 // output and an input operand. So this time, we have to arrange that
3578 // the q-reg appears just once in the MC operand list, in spite of
3579 // being mentioned twice in the asm syntax - which needs a custom
3580 // AsmMatchConverter.
3582 def MVE_VMOV_q_rr : MVE_VMOV_64bit<(outs MQPR:$Qd),
3583 (ins MQPR:$QdSrc, rGPR:$Rt, rGPR:$Rt2),
3584 0b1, "$Qd$idx, $QdSrc$idx2, $Rt, $Rt2",
3586 let DecoderMethod = "DecodeMVEVMOVDRegtoQ";
3589 def MVE_VMOV_rr_q : MVE_VMOV_64bit<(outs rGPR:$Rt, rGPR:$Rt2), (ins MQPR:$Qd),
3590 0b0, "$Rt, $Rt2, $Qd$idx, $Qd$idx2", ""> {
3591 let DecoderMethod = "DecodeMVEVMOVQtoDReg";
3592 let AsmMatchConverter = "cvtMVEVMOVQtoDReg";
3595 // end of coproc mov
3597 // start of MVE interleaving load/store
3599 // Base class for the family of interleaving/deinterleaving
3600 // load/stores with names like VLD20.8 and VST43.32.
3601 class MVE_vldst24_base<bit writeback, bit fourregs, bits<2> stage, bits<2> size,
3602 bit load, dag Oops, dag loadIops, dag wbIops,
3603 string iname, string ops,
3604 string cstr, list<dag> pattern=[]>
3605 : MVE_MI<Oops, !con(loadIops, wbIops), NoItinerary, iname, ops, cstr, pattern> {
3609 let Inst{31-22} = 0b1111110010;
3610 let Inst{21} = writeback;
3611 let Inst{20} = load;
3612 let Inst{19-16} = Rn;
3613 let Inst{15-13} = VQd{2-0};
3614 let Inst{12-9} = 0b1111;
3615 let Inst{8-7} = size;
3616 let Inst{6-5} = stage;
3617 let Inst{4-1} = 0b0000;
3618 let Inst{0} = fourregs;
3621 let mayStore = !eq(load,0);
3624 // A parameter class used to encapsulate all the ways the writeback
3625 // variants of VLD20 and friends differ from the non-writeback ones.
3626 class MVE_vldst24_writeback<bit b, dag Oo, dag Io,
3627 string sy="", string c="", string n=""> {
3633 string id_suffix = n;
3636 // Another parameter class that encapsulates the differences between VLD2x
3638 class MVE_vldst24_nvecs<int n, list<int> s, bit b, RegisterOperand vl> {
3640 list<int> stages = s;
3642 RegisterOperand VecList = vl;
3645 // A third parameter class that distinguishes VLDnn.8 from .16 from .32.
3646 class MVE_vldst24_lanesize<int i, bits<2> b> {
3648 bits<2> sizebits = b;
3651 // A base class for each direction of transfer: one for load, one for
3652 // store. I can't make these a fourth independent parametric tuple
3653 // class, because they have to take the nvecs tuple class as a
3654 // parameter, in order to find the right VecList operand type.
3656 class MVE_vld24_base<MVE_vldst24_nvecs n, bits<2> pat, bits<2> size,
3657 MVE_vldst24_writeback wb, string iname,
3658 list<dag> pattern=[]>
3659 : MVE_vldst24_base<wb.writeback, n.bit0, pat, size, 1,
3660 !con((outs n.VecList:$VQd), wb.Oops),
3661 (ins n.VecList:$VQdSrc), wb.Iops,
3662 iname, "$VQd, $Rn" # wb.syntax,
3663 wb.cstr # ",$VQdSrc = $VQd", pattern>;
3665 class MVE_vst24_base<MVE_vldst24_nvecs n, bits<2> pat, bits<2> size,
3666 MVE_vldst24_writeback wb, string iname,
3667 list<dag> pattern=[]>
3668 : MVE_vldst24_base<wb.writeback, n.bit0, pat, size, 0,
3669 wb.Oops, (ins n.VecList:$VQd), wb.Iops,
3670 iname, "$VQd, $Rn" # wb.syntax,
3673 // Actually define all the interleaving loads and stores, by a series
3674 // of nested foreaches over number of vectors (VLD2/VLD4); stage
3675 // within one of those series (VLDx0/VLDx1/VLDx2/VLDx3); size of
3676 // vector lane; writeback or no writeback.
3677 foreach n = [MVE_vldst24_nvecs<2, [0,1], 0, VecList2Q>,
3678 MVE_vldst24_nvecs<4, [0,1,2,3], 1, VecList4Q>] in
3679 foreach stage = n.stages in
3680 foreach s = [MVE_vldst24_lanesize< 8, 0b00>,
3681 MVE_vldst24_lanesize<16, 0b01>,
3682 MVE_vldst24_lanesize<32, 0b10>] in
3683 foreach wb = [MVE_vldst24_writeback<
3684 1, (outs rGPR:$wb), (ins t2_nosp_addr_offset_none:$Rn),
3685 "!", "$Rn.base = $wb", "_wb">,
3686 MVE_vldst24_writeback<0, (outs), (ins t2_addr_offset_none:$Rn)>] in {
3688 // For each case within all of those foreaches, define the actual
3689 // instructions. The def names are made by gluing together pieces
3690 // from all the parameter classes, and will end up being things like
3691 // MVE_VLD20_8 and MVE_VST43_16_wb.
3693 def "MVE_VLD" # n.nvecs # stage # "_" # s.lanesize # wb.id_suffix
3694 : MVE_vld24_base<n, stage, s.sizebits, wb,
3695 "vld" # n.nvecs # stage # "." # s.lanesize>;
3697 def "MVE_VST" # n.nvecs # stage # "_" # s.lanesize # wb.id_suffix
3698 : MVE_vst24_base<n, stage, s.sizebits, wb,
3699 "vst" # n.nvecs # stage # "." # s.lanesize>;
3702 // end of MVE interleaving load/store
3704 // start of MVE predicable load/store
3706 // A parameter class for the direction of transfer.
3707 class MVE_ldst_direction<bit b, dag Oo, dag Io, string c=""> {
3713 def MVE_ld: MVE_ldst_direction<1, (outs MQPR:$Qd), (ins), ",@earlyclobber $Qd">;
3714 def MVE_st: MVE_ldst_direction<0, (outs), (ins MQPR:$Qd)>;
3716 // A parameter class for the size of memory access in a load.
3717 class MVE_memsz<bits<2> e, int s, AddrMode m, string mn, list<string> types> {
3718 bits<2> encoding = e; // opcode bit(s) for encoding
3719 int shift = s; // shift applied to immediate load offset
3722 // For instruction aliases: define the complete list of type
3723 // suffixes at this size, and the canonical ones for loads and
3725 string MnemonicLetter = mn;
3726 int TypeBits = !shl(8, s);
3727 string CanonLoadSuffix = ".u" # TypeBits;
3728 string CanonStoreSuffix = "." # TypeBits;
3729 list<string> suffixes = !foreach(letter, types, "." # letter # TypeBits);
3732 // Instances of MVE_memsz.
3734 // (memD doesn't need an AddrMode, because those are only for
3735 // contiguous loads, and memD is only used by gather/scatters.)
3736 def MVE_memB: MVE_memsz<0b00, 0, AddrModeT2_i7, "b", ["", "u", "s"]>;
3737 def MVE_memH: MVE_memsz<0b01, 1, AddrModeT2_i7s2, "h", ["", "u", "s", "f"]>;
3738 def MVE_memW: MVE_memsz<0b10, 2, AddrModeT2_i7s4, "w", ["", "u", "s", "f"]>;
3739 def MVE_memD: MVE_memsz<0b11, 3, ?, "d", ["", "u", "s", "f"]>;
3741 // This is the base class for all the MVE loads and stores other than
3742 // the interleaving ones. All the non-interleaving loads/stores share
3743 // the characteristic that they operate on just one vector register,
3744 // so they are VPT-predicable.
3746 // The predication operand is vpred_n, for both loads and stores. For
3747 // store instructions, the reason is obvious: if there is no output
3748 // register, there can't be a need for an input parameter giving the
3749 // output register's previous value. Load instructions also don't need
3750 // that input parameter, because unlike MVE data processing
3751 // instructions, predicated loads are defined to set the inactive
3752 // lanes of the output register to zero, instead of preserving their
3754 class MVE_VLDRSTR_base<MVE_ldst_direction dir, bit U, bit P, bit W, bit opc,
3755 dag oops, dag iops, string asm, string suffix,
3756 string ops, string cstr, list<dag> pattern=[]>
3757 : MVE_p<oops, iops, NoItinerary, asm, suffix, ops, vpred_n, cstr, pattern> {
3765 let Inst{20} = dir.load;
3766 let Inst{15-13} = Qd{2-0};
3768 let Inst{11-9} = 0b111;
3770 let mayLoad = dir.load;
3771 let mayStore = !eq(dir.load,0);
3774 // Contiguous load and store instructions. These come in two main
3775 // categories: same-size loads/stores in which 128 bits of vector
3776 // register is transferred to or from 128 bits of memory in the most
3777 // obvious way, and widening loads / narrowing stores, in which the
3778 // size of memory accessed is less than the size of a vector register,
3779 // so the load instructions sign- or zero-extend each memory value
3780 // into a wider vector lane, and the store instructions truncate
3783 // The instruction mnemonics for these two classes look reasonably
3784 // similar, but the actual encodings are different enough to need two
3785 // separate base classes.
3787 // Contiguous, same size
3788 class MVE_VLDRSTR_cs<MVE_ldst_direction dir, MVE_memsz memsz, bit P, bit W,
3789 dag oops, dag iops, string asm, string suffix,
3790 IndexMode im, string ops, string cstr>
3791 : MVE_VLDRSTR_base<dir, 0, P, W, 1, oops, iops, asm, suffix, ops, cstr> {
3793 let Inst{23} = addr{7};
3794 let Inst{19-16} = addr{11-8};
3795 let Inst{8-7} = memsz.encoding;
3796 let Inst{6-0} = addr{6-0};
3799 // Contiguous, widening/narrowing
3800 class MVE_VLDRSTR_cw<MVE_ldst_direction dir, MVE_memsz memsz, bit U,
3801 bit P, bit W, bits<2> size, dag oops, dag iops,
3802 string asm, string suffix, IndexMode im,
3803 string ops, string cstr>
3804 : MVE_VLDRSTR_base<dir, U, P, W, 0, oops, iops, asm, suffix, ops, cstr> {
3806 let Inst{23} = addr{7};
3807 let Inst{19} = memsz.encoding{0}; // enough to tell 16- from 32-bit
3808 let Inst{18-16} = addr{10-8};
3809 let Inst{8-7} = size;
3810 let Inst{6-0} = addr{6-0};
3815 // Multiclass wrapper on each of the _cw and _cs base classes, to
3816 // generate three writeback modes (none, preindex, postindex).
3818 multiclass MVE_VLDRSTR_cw_m<MVE_ldst_direction dir, MVE_memsz memsz,
3819 string asm, string suffix, bit U, bits<2> size> {
3820 let AM = memsz.AM in {
3821 def "" : MVE_VLDRSTR_cw<
3822 dir, memsz, U, 1, 0, size,
3823 dir.Oops, !con(dir.Iops, (ins taddrmode_imm7<memsz.shift>:$addr)),
3824 asm, suffix, IndexModeNone, "$Qd, $addr", "">;
3826 def _pre : MVE_VLDRSTR_cw<
3827 dir, memsz, U, 1, 1, size,
3828 !con((outs tGPR:$wb), dir.Oops),
3829 !con(dir.Iops, (ins taddrmode_imm7<memsz.shift>:$addr)),
3830 asm, suffix, IndexModePre, "$Qd, $addr!", "$addr.base = $wb"> {
3831 let DecoderMethod = "DecodeMVE_MEM_1_pre<"#memsz.shift#">";
3834 def _post : MVE_VLDRSTR_cw<
3835 dir, memsz, U, 0, 1, size,
3836 !con((outs tGPR:$wb), dir.Oops),
3837 !con(dir.Iops, (ins t_addr_offset_none:$Rn,
3838 t2am_imm7_offset<memsz.shift>:$addr)),
3839 asm, suffix, IndexModePost, "$Qd, $Rn$addr", "$Rn.base = $wb"> {
3841 let Inst{18-16} = Rn{2-0};
3846 multiclass MVE_VLDRSTR_cs_m<MVE_ldst_direction dir, MVE_memsz memsz,
3847 string asm, string suffix> {
3848 let AM = memsz.AM in {
3849 def "" : MVE_VLDRSTR_cs<
3851 dir.Oops, !con(dir.Iops, (ins t2addrmode_imm7<memsz.shift>:$addr)),
3852 asm, suffix, IndexModeNone, "$Qd, $addr", "">;
3854 def _pre : MVE_VLDRSTR_cs<
3856 !con((outs rGPR:$wb), dir.Oops),
3857 !con(dir.Iops, (ins t2addrmode_imm7_pre<memsz.shift>:$addr)),
3858 asm, suffix, IndexModePre, "$Qd, $addr!", "$addr.base = $wb"> {
3859 let DecoderMethod = "DecodeMVE_MEM_2_pre<"#memsz.shift#">";
3862 def _post : MVE_VLDRSTR_cs<
3864 !con((outs rGPR:$wb), dir.Oops),
3865 // We need an !if here to select the base register class,
3866 // because it's legal to write back to SP in a load of this
3867 // type, but not in a store.
3868 !con(dir.Iops, (ins !if(dir.load, t2_addr_offset_none,
3869 t2_nosp_addr_offset_none):$Rn,
3870 t2am_imm7_offset<memsz.shift>:$addr)),
3871 asm, suffix, IndexModePost, "$Qd, $Rn$addr", "$Rn.base = $wb"> {
3873 let Inst{19-16} = Rn{3-0};
3878 // Now actually declare all the contiguous load/stores, via those
3879 // multiclasses. The instruction ids coming out of this are the bare
3880 // names shown in the defm, with _pre or _post appended for writeback,
3881 // e.g. MVE_VLDRBS16, MVE_VSTRB16_pre, MVE_VSTRHU16_post.
3883 defm MVE_VLDRBS16: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memB, "vldrb", "s16", 0, 0b01>;
3884 defm MVE_VLDRBS32: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memB, "vldrb", "s32", 0, 0b10>;
3885 defm MVE_VLDRBU16: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memB, "vldrb", "u16", 1, 0b01>;
3886 defm MVE_VLDRBU32: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memB, "vldrb", "u32", 1, 0b10>;
3887 defm MVE_VLDRHS32: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memH, "vldrh", "s32", 0, 0b10>;
3888 defm MVE_VLDRHU32: MVE_VLDRSTR_cw_m<MVE_ld, MVE_memH, "vldrh", "u32", 1, 0b10>;
3890 defm MVE_VLDRBU8: MVE_VLDRSTR_cs_m<MVE_ld, MVE_memB, "vldrb", "u8">;
3891 defm MVE_VLDRHU16: MVE_VLDRSTR_cs_m<MVE_ld, MVE_memH, "vldrh", "u16">;
3892 defm MVE_VLDRWU32: MVE_VLDRSTR_cs_m<MVE_ld, MVE_memW, "vldrw", "u32">;
3894 defm MVE_VSTRB16: MVE_VLDRSTR_cw_m<MVE_st, MVE_memB, "vstrb", "16", 0, 0b01>;
3895 defm MVE_VSTRB32: MVE_VLDRSTR_cw_m<MVE_st, MVE_memB, "vstrb", "32", 0, 0b10>;
3896 defm MVE_VSTRH32: MVE_VLDRSTR_cw_m<MVE_st, MVE_memH, "vstrh", "32", 0, 0b10>;
3898 defm MVE_VSTRBU8 : MVE_VLDRSTR_cs_m<MVE_st, MVE_memB, "vstrb", "8">;
3899 defm MVE_VSTRHU16: MVE_VLDRSTR_cs_m<MVE_st, MVE_memH, "vstrh", "16">;
3900 defm MVE_VSTRWU32: MVE_VLDRSTR_cs_m<MVE_st, MVE_memW, "vstrw", "32">;
3902 // Gather loads / scatter stores whose address operand is of the form
3903 // [Rn,Qm], i.e. a single GPR as the common base address, plus a
3904 // vector of offset from it. ('Load/store this sequence of elements of
3905 // the same array.')
3907 // Like the contiguous family, these loads and stores can widen the
3908 // loaded values / truncate the stored ones, or they can just
3909 // load/store the same size of memory and vector lane. But unlike the
3910 // contiguous family, there's no particular difference in encoding
3911 // between those two cases.
3913 // This family also comes with the option to scale the offset values
3914 // in Qm by the size of the loaded memory (i.e. to treat them as array
3915 // indices), or not to scale them (to treat them as plain byte offsets
3916 // in memory, so that perhaps the loaded values are unaligned). The
3917 // scaled instructions' address operand in assembly looks like
3918 // [Rn,Qm,UXTW #2] or similar.
3921 class MVE_VLDRSTR_rq<MVE_ldst_direction dir, MVE_memsz memsz, bit U,
3922 bits<2> size, bit os, string asm, string suffix, int shift>
3923 : MVE_VLDRSTR_base<dir, U, 0b0, 0b0, 0, dir.Oops,
3924 !con(dir.Iops, (ins mve_addr_rq_shift<shift>:$addr)),
3925 asm, suffix, "$Qd, $addr", dir.cstr> {
3928 let Inst{19-16} = addr{6-3};
3929 let Inst{8-7} = size;
3930 let Inst{6} = memsz.encoding{1};
3932 let Inst{4} = memsz.encoding{0};
3933 let Inst{3-1} = addr{2-0};
3937 // Multiclass that defines the scaled and unscaled versions of an
3938 // instruction, when the memory size is wider than a byte. The scaled
3939 // version gets the default name like MVE_VLDRBU16_rq; the unscaled /
3940 // potentially unaligned version gets a "_u" suffix, e.g.
3941 // MVE_VLDRBU16_rq_u.
3942 multiclass MVE_VLDRSTR_rq_w<MVE_ldst_direction dir, MVE_memsz memsz,
3943 string asm, string suffix, bit U, bits<2> size> {
3944 def _u : MVE_VLDRSTR_rq<dir, memsz, U, size, 0, asm, suffix, 0>;
3945 def "" : MVE_VLDRSTR_rq<dir, memsz, U, size, 1, asm, suffix, memsz.shift>;
3948 // Subclass of MVE_VLDRSTR_rq with the same API as that multiclass,
3949 // for use when the memory size is one byte, so there's no 'scaled'
3950 // version of the instruction at all. (This is encoded as if it were
3951 // unscaled, but named in the default way with no _u suffix.)
3952 class MVE_VLDRSTR_rq_b<MVE_ldst_direction dir, MVE_memsz memsz,
3953 string asm, string suffix, bit U, bits<2> size>
3954 : MVE_VLDRSTR_rq<dir, memsz, U, size, 0, asm, suffix, 0>;
3956 // Actually define all the loads and stores in this family.
3958 def MVE_VLDRBU8_rq : MVE_VLDRSTR_rq_b<MVE_ld, MVE_memB, "vldrb","u8", 1,0b00>;
3959 def MVE_VLDRBU16_rq: MVE_VLDRSTR_rq_b<MVE_ld, MVE_memB, "vldrb","u16", 1,0b01>;
3960 def MVE_VLDRBS16_rq: MVE_VLDRSTR_rq_b<MVE_ld, MVE_memB, "vldrb","s16", 0,0b01>;
3961 def MVE_VLDRBU32_rq: MVE_VLDRSTR_rq_b<MVE_ld, MVE_memB, "vldrb","u32", 1,0b10>;
3962 def MVE_VLDRBS32_rq: MVE_VLDRSTR_rq_b<MVE_ld, MVE_memB, "vldrb","s32", 0,0b10>;
3964 defm MVE_VLDRHU16_rq: MVE_VLDRSTR_rq_w<MVE_ld, MVE_memH, "vldrh","u16", 1,0b01>;
3965 defm MVE_VLDRHU32_rq: MVE_VLDRSTR_rq_w<MVE_ld, MVE_memH, "vldrh","u32", 1,0b10>;
3966 defm MVE_VLDRHS32_rq: MVE_VLDRSTR_rq_w<MVE_ld, MVE_memH, "vldrh","s32", 0,0b10>;
3967 defm MVE_VLDRWU32_rq: MVE_VLDRSTR_rq_w<MVE_ld, MVE_memW, "vldrw","u32", 1,0b10>;
3968 defm MVE_VLDRDU64_rq: MVE_VLDRSTR_rq_w<MVE_ld, MVE_memD, "vldrd","u64", 1,0b11>;
3970 def MVE_VSTRB8_rq : MVE_VLDRSTR_rq_b<MVE_st, MVE_memB, "vstrb","8", 0,0b00>;
3971 def MVE_VSTRB16_rq : MVE_VLDRSTR_rq_b<MVE_st, MVE_memB, "vstrb","16", 0,0b01>;
3972 def MVE_VSTRB32_rq : MVE_VLDRSTR_rq_b<MVE_st, MVE_memB, "vstrb","32", 0,0b10>;
3974 defm MVE_VSTRH16_rq : MVE_VLDRSTR_rq_w<MVE_st, MVE_memH, "vstrh","16", 0,0b01>;
3975 defm MVE_VSTRH32_rq : MVE_VLDRSTR_rq_w<MVE_st, MVE_memH, "vstrh","32", 0,0b10>;
3976 defm MVE_VSTRW32_rq : MVE_VLDRSTR_rq_w<MVE_st, MVE_memW, "vstrw","32", 0,0b10>;
3977 defm MVE_VSTRD64_rq : MVE_VLDRSTR_rq_w<MVE_st, MVE_memD, "vstrd","64", 0,0b11>;
3979 // Gather loads / scatter stores whose address operand is of the form
3980 // [Qm,#imm], i.e. a vector containing a full base address for each
3981 // loaded item, plus an immediate offset applied consistently to all
3982 // of them. ('Load/store the same field from this vector of pointers
3983 // to a structure type.')
3985 // This family requires the vector lane size to be at least 32 bits
3986 // (so there's room for an address in each lane at all). It has no
3987 // widening/narrowing variants. But it does support preindex
3988 // writeback, in which the address vector is updated to hold the
3989 // addresses actually loaded from.
3992 class MVE_VLDRSTR_qi<MVE_ldst_direction dir, MVE_memsz memsz, bit W, dag wbops,
3993 string asm, string wbAsm, string suffix, string cstr = "">
3994 : MVE_VLDRSTR_base<dir, 1, 1, W, 1, !con(wbops, dir.Oops),
3995 !con(dir.Iops, (ins mve_addr_q_shift<memsz.shift>:$addr)),
3996 asm, suffix, "$Qd, $addr" # wbAsm, cstr # dir.cstr> {
3998 let Inst{23} = addr{7};
3999 let Inst{19-17} = addr{10-8};
4001 let Inst{8} = memsz.encoding{0}; // enough to distinguish 32- from 64-bit
4003 let Inst{6-0} = addr{6-0};
4006 // Multiclass that generates the non-writeback and writeback variants.
4007 multiclass MVE_VLDRSTR_qi_m<MVE_ldst_direction dir, MVE_memsz memsz,
4008 string asm, string suffix> {
4009 def "" : MVE_VLDRSTR_qi<dir, memsz, 0, (outs), asm, "", suffix>;
4010 def _pre : MVE_VLDRSTR_qi<dir, memsz, 1, (outs MQPR:$wb), asm, "!", suffix,
4011 "$addr.base = $wb"> {
4012 let DecoderMethod="DecodeMVE_MEM_3_pre<"#memsz.shift#">";
4016 // Actual instruction definitions.
4017 defm MVE_VLDRWU32_qi: MVE_VLDRSTR_qi_m<MVE_ld, MVE_memW, "vldrw", "u32">;
4018 defm MVE_VLDRDU64_qi: MVE_VLDRSTR_qi_m<MVE_ld, MVE_memD, "vldrd", "u64">;
4019 defm MVE_VSTRW32_qi: MVE_VLDRSTR_qi_m<MVE_st, MVE_memW, "vstrw", "32">;
4020 defm MVE_VSTRD64_qi: MVE_VLDRSTR_qi_m<MVE_st, MVE_memD, "vstrd", "64">;
4022 // Define aliases for all the instructions where memory size and
4023 // vector lane size are the same. These are mnemonic aliases, so they
4024 // apply consistently across all of the above families - contiguous
4025 // loads, and both the rq and qi types of gather/scatter.
4027 // Rationale: As long as you're loading (for example) 16-bit memory
4028 // values into 16-bit vector lanes, you can think of them as signed or
4029 // unsigned integers, fp16 or just raw 16-bit blobs and it makes no
4030 // difference. So we permit all of vldrh.16, vldrh.u16, vldrh.s16,
4031 // vldrh.f16 and treat them all as equivalent to the canonical
4032 // spelling (which happens to be .u16 for loads, and just .16 for
4035 foreach vpt_cond = ["", "t", "e"] in
4036 foreach memsz = [MVE_memB, MVE_memH, MVE_memW, MVE_memD] in
4037 foreach suffix = memsz.suffixes in {
4039 // These foreaches are conceptually ifs, implemented by iterating a
4040 // dummy variable over a list with 0 or 1 elements depending on the
4041 // condition. The idea is to iterate over _nearly_ all the suffixes
4042 // in memsz.suffixes, but omit the one we want all the others to alias.
4044 foreach _ = !if(!ne(suffix, memsz.CanonLoadSuffix), [1], []<int>) in
4045 def : MnemonicAlias<
4046 "vldr" # memsz.MnemonicLetter # vpt_cond # suffix,
4047 "vldr" # memsz.MnemonicLetter # vpt_cond # memsz.CanonLoadSuffix>;
4049 foreach _ = !if(!ne(suffix, memsz.CanonStoreSuffix), [1], []<int>) in
4050 def : MnemonicAlias<
4051 "vstr" # memsz.MnemonicLetter # vpt_cond # suffix,
4052 "vstr" # memsz.MnemonicLetter # vpt_cond # memsz.CanonStoreSuffix>;
4055 // end of MVE predicable load/store
4057 class MVE_VPT<string suffix, bits<2> size, dag iops, string asm, list<dag> pattern=[]>
4058 : MVE_MI<(outs ), iops, NoItinerary, !strconcat("vpt", "${Mk}", ".", suffix), asm, "", pattern> {
4063 let Inst{31-23} = 0b111111100;
4064 let Inst{22} = Mk{3};
4065 let Inst{21-20} = size;
4066 let Inst{19-17} = Qn{2-0};
4068 let Inst{15-13} = Mk{2-0};
4069 let Inst{12} = fc{2};
4070 let Inst{11-8} = 0b1111;
4071 let Inst{7} = fc{0};
4074 let Defs = [VPR, P0];
4077 class MVE_VPTt1<string suffix, bits<2> size, dag iops>
4078 : MVE_VPT<suffix, size, iops, "$fc, $Qn, $Qm"> {
4083 let Inst{5} = Qm{3};
4084 let Inst{3-1} = Qm{2-0};
4085 let Inst{0} = fc{1};
4088 class MVE_VPTt1i<string suffix, bits<2> size>
4089 : MVE_VPTt1<suffix, size,
4090 (ins vpt_mask:$Mk, pred_basic_i:$fc, MQPR:$Qn, MQPR:$Qm)> {
4095 def MVE_VPTv4i32 : MVE_VPTt1i<"i32", 0b10>;
4096 def MVE_VPTv8i16 : MVE_VPTt1i<"i16", 0b01>;
4097 def MVE_VPTv16i8 : MVE_VPTt1i<"i8", 0b00>;
4099 class MVE_VPTt1u<string suffix, bits<2> size>
4100 : MVE_VPTt1<suffix, size,
4101 (ins vpt_mask:$Mk, pred_basic_u:$fc, MQPR:$Qn, MQPR:$Qm)> {
4106 def MVE_VPTv4u32 : MVE_VPTt1u<"u32", 0b10>;
4107 def MVE_VPTv8u16 : MVE_VPTt1u<"u16", 0b01>;
4108 def MVE_VPTv16u8 : MVE_VPTt1u<"u8", 0b00>;
4110 class MVE_VPTt1s<string suffix, bits<2> size>
4111 : MVE_VPTt1<suffix, size,
4112 (ins vpt_mask:$Mk, pred_basic_s:$fc, MQPR:$Qn, MQPR:$Qm)> {
4116 def MVE_VPTv4s32 : MVE_VPTt1s<"s32", 0b10>;
4117 def MVE_VPTv8s16 : MVE_VPTt1s<"s16", 0b01>;
4118 def MVE_VPTv16s8 : MVE_VPTt1s<"s8", 0b00>;
4120 class MVE_VPTt2<string suffix, bits<2> size, dag iops>
4121 : MVE_VPT<suffix, size, iops,
4128 let Inst{5} = fc{1};
4129 let Inst{3-0} = Rm{3-0};
4132 class MVE_VPTt2i<string suffix, bits<2> size>
4133 : MVE_VPTt2<suffix, size,
4134 (ins vpt_mask:$Mk, pred_basic_i:$fc, MQPR:$Qn, GPRwithZR:$Rm)> {
4139 def MVE_VPTv4i32r : MVE_VPTt2i<"i32", 0b10>;
4140 def MVE_VPTv8i16r : MVE_VPTt2i<"i16", 0b01>;
4141 def MVE_VPTv16i8r : MVE_VPTt2i<"i8", 0b00>;
4143 class MVE_VPTt2u<string suffix, bits<2> size>
4144 : MVE_VPTt2<suffix, size,
4145 (ins vpt_mask:$Mk, pred_basic_u:$fc, MQPR:$Qn, GPRwithZR:$Rm)> {
4150 def MVE_VPTv4u32r : MVE_VPTt2u<"u32", 0b10>;
4151 def MVE_VPTv8u16r : MVE_VPTt2u<"u16", 0b01>;
4152 def MVE_VPTv16u8r : MVE_VPTt2u<"u8", 0b00>;
4154 class MVE_VPTt2s<string suffix, bits<2> size>
4155 : MVE_VPTt2<suffix, size,
4156 (ins vpt_mask:$Mk, pred_basic_s:$fc, MQPR:$Qn, GPRwithZR:$Rm)> {
4160 def MVE_VPTv4s32r : MVE_VPTt2s<"s32", 0b10>;
4161 def MVE_VPTv8s16r : MVE_VPTt2s<"s16", 0b01>;
4162 def MVE_VPTv16s8r : MVE_VPTt2s<"s8", 0b00>;
4165 class MVE_VPTf<string suffix, bit size, dag iops, string asm, list<dag> pattern=[]>
4166 : MVE_MI<(outs ), iops, NoItinerary, !strconcat("vpt", "${Mk}", ".", suffix), asm,
4172 let Inst{31-29} = 0b111;
4173 let Inst{28} = size;
4174 let Inst{27-23} = 0b11100;
4175 let Inst{22} = Mk{3};
4176 let Inst{21-20} = 0b11;
4177 let Inst{19-17} = Qn{2-0};
4179 let Inst{15-13} = Mk{2-0};
4180 let Inst{12} = fc{2};
4181 let Inst{11-8} = 0b1111;
4182 let Inst{7} = fc{0};
4186 let Predicates = [HasMVEFloat];
4189 class MVE_VPTft1<string suffix, bit size>
4190 : MVE_VPTf<suffix, size, (ins vpt_mask:$Mk, pred_basic_fp:$fc, MQPR:$Qn, MQPR:$Qm),
4196 let Inst{5} = Qm{3};
4197 let Inst{3-1} = Qm{2-0};
4198 let Inst{0} = fc{1};
4201 def MVE_VPTv4f32 : MVE_VPTft1<"f32", 0b0>;
4202 def MVE_VPTv8f16 : MVE_VPTft1<"f16", 0b1>;
4204 class MVE_VPTft2<string suffix, bit size>
4205 : MVE_VPTf<suffix, size, (ins vpt_mask:$Mk, pred_basic_fp:$fc, MQPR:$Qn, GPRwithZR:$Rm),
4211 let Inst{5} = fc{1};
4212 let Inst{3-0} = Rm{3-0};
4215 def MVE_VPTv4f32r : MVE_VPTft2<"f32", 0b0>;
4216 def MVE_VPTv8f16r : MVE_VPTft2<"f16", 0b1>;
4218 def MVE_VPST : MVE_MI<(outs ), (ins vpt_mask:$Mk), NoItinerary,
4219 !strconcat("vpst", "${Mk}"), "", "", []> {
4222 let Inst{31-23} = 0b111111100;
4223 let Inst{22} = Mk{3};
4224 let Inst{21-16} = 0b110001;
4225 let Inst{15-13} = Mk{2-0};
4226 let Inst{12-0} = 0b0111101001101;
4227 let Unpredictable{12} = 0b1;
4228 let Unpredictable{7} = 0b1;
4229 let Unpredictable{5} = 0b1;
4234 def MVE_VPSEL : MVE_p<(outs MQPR:$Qd), (ins MQPR:$Qn, MQPR:$Qm), NoItinerary,
4235 "vpsel", "", "$Qd, $Qn, $Qm", vpred_n, "", []> {
4241 let Inst{25-23} = 0b100;
4242 let Inst{22} = Qd{3};
4243 let Inst{21-20} = 0b11;
4244 let Inst{19-17} = Qn{2-0};
4246 let Inst{15-13} = Qd{2-0};
4247 let Inst{12-9} = 0b0111;
4249 let Inst{7} = Qn{3};
4251 let Inst{5} = Qm{3};
4253 let Inst{3-1} = Qm{2-0};
4257 foreach suffix = ["s8", "s16", "s32", "u8", "u16", "u32",
4258 "i8", "i16", "i32", "f16", "f32"] in
4259 def : MVEInstAlias<"vpsel${vp}." # suffix # "\t$Qd, $Qn, $Qm",
4260 (MVE_VPSEL MQPR:$Qd, MQPR:$Qn, MQPR:$Qm, vpred_n:$vp)>;
4262 def MVE_VPNOT : MVE_p<(outs), (ins), NoItinerary,
4263 "vpnot", "", "", vpred_n, "", []> {
4264 let Inst{31-0} = 0b11111110001100010000111101001101;
4265 let Unpredictable{19-17} = 0b111;
4266 let Unpredictable{12} = 0b1;
4267 let Unpredictable{7} = 0b1;
4268 let Unpredictable{5} = 0b1;
4272 let Constraints = "";
4275 class MVE_loltp_start<dag iops, string asm, string ops, bits<2> size>
4276 : t2LOL<(outs GPRlr:$LR), iops, asm, ops> {
4278 let Predicates = [HasMVEInt];
4280 let Inst{21-20} = size;
4281 let Inst{19-16} = Rn{3-0};
4285 class MVE_DLSTP<string asm, bits<2> size>
4286 : MVE_loltp_start<(ins rGPR:$Rn), asm, "$LR, $Rn", size> {
4288 let Inst{11-1} = 0b00000000000;
4289 let Unpredictable{10-1} = 0b1111111111;
4292 class MVE_WLSTP<string asm, bits<2> size>
4293 : MVE_loltp_start<(ins rGPR:$Rn, wlslabel_u11:$label),
4294 asm, "$LR, $Rn, $label", size> {
4297 let Inst{11} = label{0};
4298 let Inst{10-1} = label{10-1};
4301 def MVE_DLSTP_8 : MVE_DLSTP<"dlstp.8", 0b00>;
4302 def MVE_DLSTP_16 : MVE_DLSTP<"dlstp.16", 0b01>;
4303 def MVE_DLSTP_32 : MVE_DLSTP<"dlstp.32", 0b10>;
4304 def MVE_DLSTP_64 : MVE_DLSTP<"dlstp.64", 0b11>;
4306 def MVE_WLSTP_8 : MVE_WLSTP<"wlstp.8", 0b00>;
4307 def MVE_WLSTP_16 : MVE_WLSTP<"wlstp.16", 0b01>;
4308 def MVE_WLSTP_32 : MVE_WLSTP<"wlstp.32", 0b10>;
4309 def MVE_WLSTP_64 : MVE_WLSTP<"wlstp.64", 0b11>;
4311 class MVE_loltp_end<dag oops, dag iops, string asm, string ops>
4312 : t2LOL<oops, iops, asm, ops> {
4313 let Predicates = [HasMVEInt];
4314 let Inst{22-21} = 0b00;
4315 let Inst{19-16} = 0b1111;
4319 def MVE_LETP : MVE_loltp_end<(outs GPRlr:$LRout),
4320 (ins GPRlr:$LRin, lelabel_u11:$label),
4321 "letp", "$LRin, $label"> {
4325 let Inst{11} = label{0};
4326 let Inst{10-1} = label{10-1};
4329 def MVE_LCTP : MVE_loltp_end<(outs), (ins pred:$p), "lctp${p}", ""> {
4332 let Inst{11-1} = 0b00000000000;
4333 let Unpredictable{21-20} = 0b11;
4334 let Unpredictable{11-1} = 0b11111111111;
4338 //===----------------------------------------------------------------------===//
4340 //===----------------------------------------------------------------------===//
4342 class MVE_unpred_vector_store_typed<ValueType Ty, Instruction RegImmInst,
4343 PatFrag StoreKind, int shift>
4344 : Pat<(StoreKind (Ty MQPR:$val), t2addrmode_imm7<shift>:$addr),
4345 (RegImmInst (Ty MQPR:$val), t2addrmode_imm7<shift>:$addr)>;
4347 multiclass MVE_unpred_vector_store<Instruction RegImmInst, PatFrag StoreKind,
4349 def : MVE_unpred_vector_store_typed<v16i8, RegImmInst, StoreKind, shift>;
4350 def : MVE_unpred_vector_store_typed<v8i16, RegImmInst, StoreKind, shift>;
4351 def : MVE_unpred_vector_store_typed<v8f16, RegImmInst, StoreKind, shift>;
4352 def : MVE_unpred_vector_store_typed<v4i32, RegImmInst, StoreKind, shift>;
4353 def : MVE_unpred_vector_store_typed<v4f32, RegImmInst, StoreKind, shift>;
4354 def : MVE_unpred_vector_store_typed<v2i64, RegImmInst, StoreKind, shift>;
4355 def : MVE_unpred_vector_store_typed<v2f64, RegImmInst, StoreKind, shift>;
4358 class MVE_unpred_vector_load_typed<ValueType Ty, Instruction RegImmInst,
4359 PatFrag LoadKind, int shift>
4360 : Pat<(Ty (LoadKind t2addrmode_imm7<shift>:$addr)),
4361 (Ty (RegImmInst t2addrmode_imm7<shift>:$addr))>;
4363 multiclass MVE_unpred_vector_load<Instruction RegImmInst, PatFrag LoadKind,
4365 def : MVE_unpred_vector_load_typed<v16i8, RegImmInst, LoadKind, shift>;
4366 def : MVE_unpred_vector_load_typed<v8i16, RegImmInst, LoadKind, shift>;
4367 def : MVE_unpred_vector_load_typed<v8f16, RegImmInst, LoadKind, shift>;
4368 def : MVE_unpred_vector_load_typed<v4i32, RegImmInst, LoadKind, shift>;
4369 def : MVE_unpred_vector_load_typed<v4f32, RegImmInst, LoadKind, shift>;
4370 def : MVE_unpred_vector_load_typed<v2i64, RegImmInst, LoadKind, shift>;
4371 def : MVE_unpred_vector_load_typed<v2f64, RegImmInst, LoadKind, shift>;
4374 let Predicates = [HasMVEInt, IsLE] in {
4375 defm : MVE_unpred_vector_store<MVE_VSTRBU8, byte_alignedstore, 0>;
4376 defm : MVE_unpred_vector_store<MVE_VSTRHU16, hword_alignedstore, 1>;
4377 defm : MVE_unpred_vector_store<MVE_VSTRWU32, alignedstore32, 2>;
4379 defm : MVE_unpred_vector_load<MVE_VLDRBU8, byte_alignedload, 0>;
4380 defm : MVE_unpred_vector_load<MVE_VLDRHU16, hword_alignedload, 1>;
4381 defm : MVE_unpred_vector_load<MVE_VLDRWU32, alignedload32, 2>;
4383 def : Pat<(v16i1 (load t2addrmode_imm7<2>:$addr)),
4384 (v16i1 (VLDR_P0_off t2addrmode_imm7<2>:$addr))>;
4385 def : Pat<(v8i1 (load t2addrmode_imm7<2>:$addr)),
4386 (v8i1 (VLDR_P0_off t2addrmode_imm7<2>:$addr))>;
4387 def : Pat<(v4i1 (load t2addrmode_imm7<2>:$addr)),
4388 (v4i1 (VLDR_P0_off t2addrmode_imm7<2>:$addr))>;
4391 let Predicates = [HasMVEInt, IsBE] in {
4392 def : MVE_unpred_vector_store_typed<v16i8, MVE_VSTRBU8, store, 0>;
4393 def : MVE_unpred_vector_store_typed<v8i16, MVE_VSTRHU16, alignedstore16, 1>;
4394 def : MVE_unpred_vector_store_typed<v8f16, MVE_VSTRHU16, alignedstore16, 1>;
4395 def : MVE_unpred_vector_store_typed<v4i32, MVE_VSTRWU32, alignedstore32, 2>;
4396 def : MVE_unpred_vector_store_typed<v4f32, MVE_VSTRWU32, alignedstore32, 2>;
4398 def : MVE_unpred_vector_load_typed<v16i8, MVE_VLDRBU8, load, 0>;
4399 def : MVE_unpred_vector_load_typed<v8i16, MVE_VLDRHU16, alignedload16, 1>;
4400 def : MVE_unpred_vector_load_typed<v8f16, MVE_VLDRHU16, alignedload16, 1>;
4401 def : MVE_unpred_vector_load_typed<v4i32, MVE_VLDRWU32, alignedload32, 2>;
4402 def : MVE_unpred_vector_load_typed<v4f32, MVE_VLDRWU32, alignedload32, 2>;
4406 // Widening/Narrowing Loads/Stores
4408 let Predicates = [HasMVEInt] in {
4409 def : Pat<(truncstorevi8 (v8i16 MQPR:$val), t2addrmode_imm7<1>:$addr),
4410 (MVE_VSTRB16 MQPR:$val, t2addrmode_imm7<1>:$addr)>;
4411 def : Pat<(truncstorevi8 (v4i32 MQPR:$val), t2addrmode_imm7<1>:$addr),
4412 (MVE_VSTRB32 MQPR:$val, t2addrmode_imm7<1>:$addr)>;
4413 def : Pat<(truncstorevi16 (v4i32 MQPR:$val), t2addrmode_imm7<2>:$addr),
4414 (MVE_VSTRH32 MQPR:$val, t2addrmode_imm7<2>:$addr)>;
4417 multiclass MVEExtLoad<string DestLanes, string DestElemBits,
4418 string SrcElemBits, string SrcElemType,
4420 def _Any : Pat<(!cast<ValueType>("v" # DestLanes # "i" # DestElemBits)
4421 (!cast<PatFrag>("extloadvi" # SrcElemBits) am:$addr)),
4422 (!cast<Instruction>("MVE_VLDR" # SrcElemType # "U" # DestElemBits)
4424 def _Z : Pat<(!cast<ValueType>("v" # DestLanes # "i" # DestElemBits)
4425 (!cast<PatFrag>("zextloadvi" # SrcElemBits) am:$addr)),
4426 (!cast<Instruction>("MVE_VLDR" # SrcElemType # "U" # DestElemBits)
4428 def _S : Pat<(!cast<ValueType>("v" # DestLanes # "i" # DestElemBits)
4429 (!cast<PatFrag>("sextloadvi" # SrcElemBits) am:$addr)),
4430 (!cast<Instruction>("MVE_VLDR" # SrcElemType # "S" # DestElemBits)
4434 let Predicates = [HasMVEInt] in {
4435 defm : MVEExtLoad<"4", "32", "8", "B", t2addrmode_imm7<1>>;
4436 defm : MVEExtLoad<"8", "16", "8", "B", t2addrmode_imm7<1>>;
4437 defm : MVEExtLoad<"4", "32", "16", "H", t2addrmode_imm7<2>>;
4441 // Bit convert patterns
4443 let Predicates = [HasMVEInt] in {
4444 def : Pat<(v2f64 (bitconvert (v2i64 QPR:$src))), (v2f64 QPR:$src)>;
4445 def : Pat<(v2i64 (bitconvert (v2f64 QPR:$src))), (v2i64 QPR:$src)>;
4447 def : Pat<(v4i32 (bitconvert (v4f32 QPR:$src))), (v4i32 QPR:$src)>;
4448 def : Pat<(v4f32 (bitconvert (v4i32 QPR:$src))), (v4f32 QPR:$src)>;
4450 def : Pat<(v8i16 (bitconvert (v8f16 QPR:$src))), (v8i16 QPR:$src)>;
4451 def : Pat<(v8f16 (bitconvert (v8i16 QPR:$src))), (v8f16 QPR:$src)>;
4454 let Predicates = [IsLE,HasMVEInt] in {
4455 def : Pat<(v2f64 (bitconvert (v4f32 QPR:$src))), (v2f64 QPR:$src)>;
4456 def : Pat<(v2f64 (bitconvert (v4i32 QPR:$src))), (v2f64 QPR:$src)>;
4457 def : Pat<(v2f64 (bitconvert (v8f16 QPR:$src))), (v2f64 QPR:$src)>;
4458 def : Pat<(v2f64 (bitconvert (v8i16 QPR:$src))), (v2f64 QPR:$src)>;
4459 def : Pat<(v2f64 (bitconvert (v16i8 QPR:$src))), (v2f64 QPR:$src)>;
4461 def : Pat<(v2i64 (bitconvert (v4f32 QPR:$src))), (v2i64 QPR:$src)>;
4462 def : Pat<(v2i64 (bitconvert (v4i32 QPR:$src))), (v2i64 QPR:$src)>;
4463 def : Pat<(v2i64 (bitconvert (v8f16 QPR:$src))), (v2i64 QPR:$src)>;
4464 def : Pat<(v2i64 (bitconvert (v8i16 QPR:$src))), (v2i64 QPR:$src)>;
4465 def : Pat<(v2i64 (bitconvert (v16i8 QPR:$src))), (v2i64 QPR:$src)>;
4467 def : Pat<(v4f32 (bitconvert (v2f64 QPR:$src))), (v4f32 QPR:$src)>;
4468 def : Pat<(v4f32 (bitconvert (v2i64 QPR:$src))), (v4f32 QPR:$src)>;
4469 def : Pat<(v4f32 (bitconvert (v8f16 QPR:$src))), (v4f32 QPR:$src)>;
4470 def : Pat<(v4f32 (bitconvert (v8i16 QPR:$src))), (v4f32 QPR:$src)>;
4471 def : Pat<(v4f32 (bitconvert (v16i8 QPR:$src))), (v4f32 QPR:$src)>;
4473 def : Pat<(v4i32 (bitconvert (v2f64 QPR:$src))), (v4i32 QPR:$src)>;
4474 def : Pat<(v4i32 (bitconvert (v2i64 QPR:$src))), (v4i32 QPR:$src)>;
4475 def : Pat<(v4i32 (bitconvert (v8f16 QPR:$src))), (v4i32 QPR:$src)>;
4476 def : Pat<(v4i32 (bitconvert (v8i16 QPR:$src))), (v4i32 QPR:$src)>;
4477 def : Pat<(v4i32 (bitconvert (v16i8 QPR:$src))), (v4i32 QPR:$src)>;
4479 def : Pat<(v8f16 (bitconvert (v2f64 QPR:$src))), (v8f16 QPR:$src)>;
4480 def : Pat<(v8f16 (bitconvert (v2i64 QPR:$src))), (v8f16 QPR:$src)>;
4481 def : Pat<(v8f16 (bitconvert (v4f32 QPR:$src))), (v8f16 QPR:$src)>;
4482 def : Pat<(v8f16 (bitconvert (v4i32 QPR:$src))), (v8f16 QPR:$src)>;
4483 def : Pat<(v8f16 (bitconvert (v16i8 QPR:$src))), (v8f16 QPR:$src)>;
4485 def : Pat<(v8i16 (bitconvert (v2f64 QPR:$src))), (v8i16 QPR:$src)>;
4486 def : Pat<(v8i16 (bitconvert (v2i64 QPR:$src))), (v8i16 QPR:$src)>;
4487 def : Pat<(v8i16 (bitconvert (v4f32 QPR:$src))), (v8i16 QPR:$src)>;
4488 def : Pat<(v8i16 (bitconvert (v4i32 QPR:$src))), (v8i16 QPR:$src)>;
4489 def : Pat<(v8i16 (bitconvert (v16i8 QPR:$src))), (v8i16 QPR:$src)>;
4491 def : Pat<(v16i8 (bitconvert (v2f64 QPR:$src))), (v16i8 QPR:$src)>;
4492 def : Pat<(v16i8 (bitconvert (v2i64 QPR:$src))), (v16i8 QPR:$src)>;
4493 def : Pat<(v16i8 (bitconvert (v4f32 QPR:$src))), (v16i8 QPR:$src)>;
4494 def : Pat<(v16i8 (bitconvert (v4i32 QPR:$src))), (v16i8 QPR:$src)>;
4495 def : Pat<(v16i8 (bitconvert (v8f16 QPR:$src))), (v16i8 QPR:$src)>;
4496 def : Pat<(v16i8 (bitconvert (v8i16 QPR:$src))), (v16i8 QPR:$src)>;