1 //===-- ARMInstrVFP.td - VFP support for ARM ---------------*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the ARM VFP instruction set.
12 //===----------------------------------------------------------------------===//
14 def SDT_CMPFP0 : SDTypeProfile<0, 1, [SDTCisFP<0>]>;
15 def SDT_VMOVDRR : SDTypeProfile<1, 2, [SDTCisVT<0, f64>, SDTCisVT<1, i32>,
18 def arm_fmstat : SDNode<"ARMISD::FMSTAT", SDTNone, [SDNPInGlue, SDNPOutGlue]>;
19 def arm_cmpfp : SDNode<"ARMISD::CMPFP", SDT_ARMCmp, [SDNPOutGlue]>;
20 def arm_cmpfp0 : SDNode<"ARMISD::CMPFPw0", SDT_CMPFP0, [SDNPOutGlue]>;
21 def arm_fmdrr : SDNode<"ARMISD::VMOVDRR", SDT_VMOVDRR>;
23 //===----------------------------------------------------------------------===//
24 // Operand Definitions.
27 // 8-bit floating-point immediate encodings.
28 def FPImmOperand : AsmOperandClass {
30 let ParserMethod = "parseFPImm";
33 def vfp_f16imm : Operand<f16>,
34 PatLeaf<(f16 fpimm), [{
35 return ARM_AM::getFP16Imm(N->getValueAPF()) != -1;
36 }], SDNodeXForm<fpimm, [{
37 APFloat InVal = N->getValueAPF();
38 uint32_t enc = ARM_AM::getFP16Imm(InVal);
39 return CurDAG->getTargetConstant(enc, MVT::i32);
41 let PrintMethod = "printFPImmOperand";
42 let ParserMatchClass = FPImmOperand;
45 def vfp_f32imm : Operand<f32>,
46 PatLeaf<(f32 fpimm), [{
47 return ARM_AM::getFP32Imm(N->getValueAPF()) != -1;
48 }], SDNodeXForm<fpimm, [{
49 APFloat InVal = N->getValueAPF();
50 uint32_t enc = ARM_AM::getFP32Imm(InVal);
51 return CurDAG->getTargetConstant(enc, SDLoc(N), MVT::i32);
53 let PrintMethod = "printFPImmOperand";
54 let ParserMatchClass = FPImmOperand;
57 def vfp_f64imm : Operand<f64>,
58 PatLeaf<(f64 fpimm), [{
59 return ARM_AM::getFP64Imm(N->getValueAPF()) != -1;
60 }], SDNodeXForm<fpimm, [{
61 APFloat InVal = N->getValueAPF();
62 uint32_t enc = ARM_AM::getFP64Imm(InVal);
63 return CurDAG->getTargetConstant(enc, SDLoc(N), MVT::i32);
65 let PrintMethod = "printFPImmOperand";
66 let ParserMatchClass = FPImmOperand;
69 def alignedload32 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
70 return cast<LoadSDNode>(N)->getAlignment() >= 4;
73 def alignedstore32 : PatFrag<(ops node:$val, node:$ptr),
74 (store node:$val, node:$ptr), [{
75 return cast<StoreSDNode>(N)->getAlignment() >= 4;
78 // The VCVT to/from fixed-point instructions encode the 'fbits' operand
79 // (the number of fixed bits) differently than it appears in the assembly
80 // source. It's encoded as "Size - fbits" where Size is the size of the
81 // fixed-point representation (32 or 16) and fbits is the value appearing
82 // in the assembly source, an integer in [0,16] or (0,32], depending on size.
83 def fbits32_asm_operand : AsmOperandClass { let Name = "FBits32"; }
84 def fbits32 : Operand<i32> {
85 let PrintMethod = "printFBits32";
86 let ParserMatchClass = fbits32_asm_operand;
89 def fbits16_asm_operand : AsmOperandClass { let Name = "FBits16"; }
90 def fbits16 : Operand<i32> {
91 let PrintMethod = "printFBits16";
92 let ParserMatchClass = fbits16_asm_operand;
95 //===----------------------------------------------------------------------===//
96 // Load / store Instructions.
99 let canFoldAsLoad = 1, isReMaterializable = 1 in {
101 def VLDRD : ADI5<0b1101, 0b01, (outs DPR:$Dd), (ins addrmode5:$addr),
102 IIC_fpLoad64, "vldr", "\t$Dd, $addr",
103 [(set DPR:$Dd, (f64 (alignedload32 addrmode5:$addr)))]>;
105 def VLDRS : ASI5<0b1101, 0b01, (outs SPR:$Sd), (ins addrmode5:$addr),
106 IIC_fpLoad32, "vldr", "\t$Sd, $addr",
107 [(set SPR:$Sd, (alignedload32 addrmode5:$addr))]> {
108 // Some single precision VFP instructions may be executed on both NEON and VFP
110 let D = VFPNeonDomain;
113 def VLDRH : AHI5<0b1101, 0b01, (outs SPR:$Sd), (ins addrmode5fp16:$addr),
114 IIC_fpLoad16, "vldr", ".16\t$Sd, $addr",
116 Requires<[HasFullFP16]>;
118 } // End of 'let canFoldAsLoad = 1, isReMaterializable = 1 in'
120 def VSTRD : ADI5<0b1101, 0b00, (outs), (ins DPR:$Dd, addrmode5:$addr),
121 IIC_fpStore64, "vstr", "\t$Dd, $addr",
122 [(alignedstore32 (f64 DPR:$Dd), addrmode5:$addr)]>;
124 def VSTRS : ASI5<0b1101, 0b00, (outs), (ins SPR:$Sd, addrmode5:$addr),
125 IIC_fpStore32, "vstr", "\t$Sd, $addr",
126 [(alignedstore32 SPR:$Sd, addrmode5:$addr)]> {
127 // Some single precision VFP instructions may be executed on both NEON and VFP
129 let D = VFPNeonDomain;
132 def VSTRH : AHI5<0b1101, 0b00, (outs), (ins SPR:$Sd, addrmode5fp16:$addr),
133 IIC_fpStore16, "vstr", ".16\t$Sd, $addr",
135 Requires<[HasFullFP16]>;
137 //===----------------------------------------------------------------------===//
138 // Load / store multiple Instructions.
141 multiclass vfp_ldst_mult<string asm, bit L_bit,
142 InstrItinClass itin, InstrItinClass itin_upd> {
145 AXDI4<(outs), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
147 !strconcat(asm, "ia${p}\t$Rn, $regs"), "", []> {
148 let Inst{24-23} = 0b01; // Increment After
149 let Inst{21} = 0; // No writeback
150 let Inst{20} = L_bit;
153 AXDI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs,
155 IndexModeUpd, itin_upd,
156 !strconcat(asm, "ia${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
157 let Inst{24-23} = 0b01; // Increment After
158 let Inst{21} = 1; // Writeback
159 let Inst{20} = L_bit;
162 AXDI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs,
164 IndexModeUpd, itin_upd,
165 !strconcat(asm, "db${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
166 let Inst{24-23} = 0b10; // Decrement Before
167 let Inst{21} = 1; // Writeback
168 let Inst{20} = L_bit;
173 AXSI4<(outs), (ins GPR:$Rn, pred:$p, spr_reglist:$regs, variable_ops),
175 !strconcat(asm, "ia${p}\t$Rn, $regs"), "", []> {
176 let Inst{24-23} = 0b01; // Increment After
177 let Inst{21} = 0; // No writeback
178 let Inst{20} = L_bit;
180 // Some single precision VFP instructions may be executed on both NEON and
182 let D = VFPNeonDomain;
185 AXSI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, spr_reglist:$regs,
187 IndexModeUpd, itin_upd,
188 !strconcat(asm, "ia${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
189 let Inst{24-23} = 0b01; // Increment After
190 let Inst{21} = 1; // Writeback
191 let Inst{20} = L_bit;
193 // Some single precision VFP instructions may be executed on both NEON and
195 let D = VFPNeonDomain;
198 AXSI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, spr_reglist:$regs,
200 IndexModeUpd, itin_upd,
201 !strconcat(asm, "db${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
202 let Inst{24-23} = 0b10; // Decrement Before
203 let Inst{21} = 1; // Writeback
204 let Inst{20} = L_bit;
206 // Some single precision VFP instructions may be executed on both NEON and
208 let D = VFPNeonDomain;
212 let hasSideEffects = 0 in {
214 let mayLoad = 1, hasExtraDefRegAllocReq = 1 in
215 defm VLDM : vfp_ldst_mult<"vldm", 1, IIC_fpLoad_m, IIC_fpLoad_mu>;
217 let mayStore = 1, hasExtraSrcRegAllocReq = 1 in
218 defm VSTM : vfp_ldst_mult<"vstm", 0, IIC_fpStore_m, IIC_fpStore_mu>;
222 def : MnemonicAlias<"vldm", "vldmia">;
223 def : MnemonicAlias<"vstm", "vstmia">;
226 //===----------------------------------------------------------------------===//
227 // Lazy load / store multiple Instructions
230 def VLLDM : AXSI4<(outs), (ins GPRnopc:$Rn, pred:$p), IndexModeNone,
231 IIC_fpLoad_m, "vlldm${p}\t$Rn", "", []>,
232 Requires<[HasV8MMainline, Has8MSecExt]> {
233 let Inst{24-23} = 0b00;
243 def VLSTM : AXSI4<(outs), (ins GPRnopc:$Rn, pred:$p), IndexModeNone,
244 IIC_fpStore_m, "vlstm${p}\t$Rn", "", []>,
245 Requires<[HasV8MMainline, Has8MSecExt]> {
246 let Inst{24-23} = 0b00;
256 // FLDM/FSTM - Load / Store multiple single / double precision registers for
258 // These instructions are deprecated!
259 def : VFP2MnemonicAlias<"fldmias", "vldmia">;
260 def : VFP2MnemonicAlias<"fldmdbs", "vldmdb">;
261 def : VFP2MnemonicAlias<"fldmeas", "vldmdb">;
262 def : VFP2MnemonicAlias<"fldmfds", "vldmia">;
263 def : VFP2MnemonicAlias<"fldmiad", "vldmia">;
264 def : VFP2MnemonicAlias<"fldmdbd", "vldmdb">;
265 def : VFP2MnemonicAlias<"fldmead", "vldmdb">;
266 def : VFP2MnemonicAlias<"fldmfdd", "vldmia">;
268 def : VFP2MnemonicAlias<"fstmias", "vstmia">;
269 def : VFP2MnemonicAlias<"fstmdbs", "vstmdb">;
270 def : VFP2MnemonicAlias<"fstmeas", "vstmia">;
271 def : VFP2MnemonicAlias<"fstmfds", "vstmdb">;
272 def : VFP2MnemonicAlias<"fstmiad", "vstmia">;
273 def : VFP2MnemonicAlias<"fstmdbd", "vstmdb">;
274 def : VFP2MnemonicAlias<"fstmead", "vstmia">;
275 def : VFP2MnemonicAlias<"fstmfdd", "vstmdb">;
277 def : InstAlias<"vpush${p} $r", (VSTMDDB_UPD SP, pred:$p, dpr_reglist:$r), 0>,
279 def : InstAlias<"vpush${p} $r", (VSTMSDB_UPD SP, pred:$p, spr_reglist:$r), 0>,
281 def : InstAlias<"vpop${p} $r", (VLDMDIA_UPD SP, pred:$p, dpr_reglist:$r), 0>,
283 def : InstAlias<"vpop${p} $r", (VLDMSIA_UPD SP, pred:$p, spr_reglist:$r), 0>,
285 defm : VFPDTAnyInstAlias<"vpush${p}", "$r",
286 (VSTMSDB_UPD SP, pred:$p, spr_reglist:$r)>;
287 defm : VFPDTAnyInstAlias<"vpush${p}", "$r",
288 (VSTMDDB_UPD SP, pred:$p, dpr_reglist:$r)>;
289 defm : VFPDTAnyInstAlias<"vpop${p}", "$r",
290 (VLDMSIA_UPD SP, pred:$p, spr_reglist:$r)>;
291 defm : VFPDTAnyInstAlias<"vpop${p}", "$r",
292 (VLDMDIA_UPD SP, pred:$p, dpr_reglist:$r)>;
294 // FLDMX, FSTMX - Load and store multiple unknown precision registers for
296 // These instruction are deprecated so we don't want them to get selected.
297 multiclass vfp_ldstx_mult<string asm, bit L_bit> {
300 AXXI4<(outs), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
301 IndexModeNone, !strconcat(asm, "iax${p}\t$Rn, $regs"), "", []> {
302 let Inst{24-23} = 0b01; // Increment After
303 let Inst{21} = 0; // No writeback
304 let Inst{20} = L_bit;
307 AXXI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
308 IndexModeUpd, !strconcat(asm, "iax${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
309 let Inst{24-23} = 0b01; // Increment After
310 let Inst{21} = 1; // Writeback
311 let Inst{20} = L_bit;
314 AXXI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
315 IndexModeUpd, !strconcat(asm, "dbx${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
316 let Inst{24-23} = 0b10; // Decrement Before
317 let Inst{21} = 1; // Writeback
318 let Inst{20} = L_bit;
322 defm FLDM : vfp_ldstx_mult<"fldm", 1>;
323 defm FSTM : vfp_ldstx_mult<"fstm", 0>;
325 def : VFP2MnemonicAlias<"fldmeax", "fldmdbx">;
326 def : VFP2MnemonicAlias<"fldmfdx", "fldmiax">;
328 def : VFP2MnemonicAlias<"fstmeax", "fstmiax">;
329 def : VFP2MnemonicAlias<"fstmfdx", "fstmdbx">;
331 //===----------------------------------------------------------------------===//
332 // FP Binary Operations.
335 let TwoOperandAliasConstraint = "$Dn = $Dd" in
336 def VADDD : ADbI<0b11100, 0b11, 0, 0,
337 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
338 IIC_fpALU64, "vadd", ".f64\t$Dd, $Dn, $Dm",
339 [(set DPR:$Dd, (fadd DPR:$Dn, (f64 DPR:$Dm)))]>;
341 let TwoOperandAliasConstraint = "$Sn = $Sd" in
342 def VADDS : ASbIn<0b11100, 0b11, 0, 0,
343 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
344 IIC_fpALU32, "vadd", ".f32\t$Sd, $Sn, $Sm",
345 [(set SPR:$Sd, (fadd SPR:$Sn, SPR:$Sm))]> {
346 // Some single precision VFP instructions may be executed on both NEON and
347 // VFP pipelines on A8.
348 let D = VFPNeonA8Domain;
351 let TwoOperandAliasConstraint = "$Sn = $Sd" in
352 def VADDH : AHbI<0b11100, 0b11, 0, 0,
353 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
354 IIC_fpALU16, "vadd", ".f16\t$Sd, $Sn, $Sm",
357 let TwoOperandAliasConstraint = "$Dn = $Dd" in
358 def VSUBD : ADbI<0b11100, 0b11, 1, 0,
359 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
360 IIC_fpALU64, "vsub", ".f64\t$Dd, $Dn, $Dm",
361 [(set DPR:$Dd, (fsub DPR:$Dn, (f64 DPR:$Dm)))]>;
363 let TwoOperandAliasConstraint = "$Sn = $Sd" in
364 def VSUBS : ASbIn<0b11100, 0b11, 1, 0,
365 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
366 IIC_fpALU32, "vsub", ".f32\t$Sd, $Sn, $Sm",
367 [(set SPR:$Sd, (fsub SPR:$Sn, SPR:$Sm))]> {
368 // Some single precision VFP instructions may be executed on both NEON and
369 // VFP pipelines on A8.
370 let D = VFPNeonA8Domain;
373 let TwoOperandAliasConstraint = "$Sn = $Sd" in
374 def VSUBH : AHbI<0b11100, 0b11, 1, 0,
375 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
376 IIC_fpALU16, "vsub", ".f16\t$Sd, $Sn, $Sm",
379 let TwoOperandAliasConstraint = "$Dn = $Dd" in
380 def VDIVD : ADbI<0b11101, 0b00, 0, 0,
381 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
382 IIC_fpDIV64, "vdiv", ".f64\t$Dd, $Dn, $Dm",
383 [(set DPR:$Dd, (fdiv DPR:$Dn, (f64 DPR:$Dm)))]>;
385 let TwoOperandAliasConstraint = "$Sn = $Sd" in
386 def VDIVS : ASbI<0b11101, 0b00, 0, 0,
387 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
388 IIC_fpDIV32, "vdiv", ".f32\t$Sd, $Sn, $Sm",
389 [(set SPR:$Sd, (fdiv SPR:$Sn, SPR:$Sm))]>;
391 let TwoOperandAliasConstraint = "$Sn = $Sd" in
392 def VDIVH : AHbI<0b11101, 0b00, 0, 0,
393 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
394 IIC_fpDIV16, "vdiv", ".f16\t$Sd, $Sn, $Sm",
397 let TwoOperandAliasConstraint = "$Dn = $Dd" in
398 def VMULD : ADbI<0b11100, 0b10, 0, 0,
399 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
400 IIC_fpMUL64, "vmul", ".f64\t$Dd, $Dn, $Dm",
401 [(set DPR:$Dd, (fmul DPR:$Dn, (f64 DPR:$Dm)))]>;
403 let TwoOperandAliasConstraint = "$Sn = $Sd" in
404 def VMULS : ASbIn<0b11100, 0b10, 0, 0,
405 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
406 IIC_fpMUL32, "vmul", ".f32\t$Sd, $Sn, $Sm",
407 [(set SPR:$Sd, (fmul SPR:$Sn, SPR:$Sm))]> {
408 // Some single precision VFP instructions may be executed on both NEON and
409 // VFP pipelines on A8.
410 let D = VFPNeonA8Domain;
413 let TwoOperandAliasConstraint = "$Sn = $Sd" in
414 def VMULH : AHbI<0b11100, 0b10, 0, 0,
415 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
416 IIC_fpMUL16, "vmul", ".f16\t$Sd, $Sn, $Sm",
419 def VNMULD : ADbI<0b11100, 0b10, 1, 0,
420 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
421 IIC_fpMUL64, "vnmul", ".f64\t$Dd, $Dn, $Dm",
422 [(set DPR:$Dd, (fneg (fmul DPR:$Dn, (f64 DPR:$Dm))))]>;
424 def VNMULS : ASbI<0b11100, 0b10, 1, 0,
425 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
426 IIC_fpMUL32, "vnmul", ".f32\t$Sd, $Sn, $Sm",
427 [(set SPR:$Sd, (fneg (fmul SPR:$Sn, SPR:$Sm)))]> {
428 // Some single precision VFP instructions may be executed on both NEON and
429 // VFP pipelines on A8.
430 let D = VFPNeonA8Domain;
433 def VNMULH : AHbI<0b11100, 0b10, 1, 0,
434 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
435 IIC_fpMUL16, "vnmul", ".f16\t$Sd, $Sn, $Sm",
438 multiclass vsel_inst<string op, bits<2> opc, int CC> {
439 let DecoderNamespace = "VFPV8", PostEncoderMethod = "",
440 Uses = [CPSR], AddedComplexity = 4 in {
441 def H : AHbInp<0b11100, opc, 0,
442 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
443 NoItinerary, !strconcat("vsel", op, ".f16\t$Sd, $Sn, $Sm"),
445 Requires<[HasFullFP16]>;
447 def S : ASbInp<0b11100, opc, 0,
448 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
449 NoItinerary, !strconcat("vsel", op, ".f32\t$Sd, $Sn, $Sm"),
450 [(set SPR:$Sd, (ARMcmov SPR:$Sm, SPR:$Sn, CC))]>,
451 Requires<[HasFPARMv8]>;
453 def D : ADbInp<0b11100, opc, 0,
454 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
455 NoItinerary, !strconcat("vsel", op, ".f64\t$Dd, $Dn, $Dm"),
456 [(set DPR:$Dd, (ARMcmov (f64 DPR:$Dm), (f64 DPR:$Dn), CC))]>,
457 Requires<[HasFPARMv8, HasDPVFP]>;
461 // The CC constants here match ARMCC::CondCodes.
462 defm VSELGT : vsel_inst<"gt", 0b11, 12>;
463 defm VSELGE : vsel_inst<"ge", 0b10, 10>;
464 defm VSELEQ : vsel_inst<"eq", 0b00, 0>;
465 defm VSELVS : vsel_inst<"vs", 0b01, 6>;
467 multiclass vmaxmin_inst<string op, bit opc, SDNode SD> {
468 let DecoderNamespace = "VFPV8", PostEncoderMethod = "" in {
469 def H : AHbInp<0b11101, 0b00, opc,
470 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
471 NoItinerary, !strconcat(op, ".f16\t$Sd, $Sn, $Sm"),
473 Requires<[HasFullFP16]>;
475 def S : ASbInp<0b11101, 0b00, opc,
476 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
477 NoItinerary, !strconcat(op, ".f32\t$Sd, $Sn, $Sm"),
478 [(set SPR:$Sd, (SD SPR:$Sn, SPR:$Sm))]>,
479 Requires<[HasFPARMv8]>;
481 def D : ADbInp<0b11101, 0b00, opc,
482 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
483 NoItinerary, !strconcat(op, ".f64\t$Dd, $Dn, $Dm"),
484 [(set DPR:$Dd, (f64 (SD (f64 DPR:$Dn), (f64 DPR:$Dm))))]>,
485 Requires<[HasFPARMv8, HasDPVFP]>;
489 defm VMAXNM : vmaxmin_inst<"vmaxnm", 0, fmaxnum>;
490 defm VMINNM : vmaxmin_inst<"vminnm", 1, fminnum>;
492 // Match reassociated forms only if not sign dependent rounding.
493 def : Pat<(fmul (fneg DPR:$a), (f64 DPR:$b)),
494 (VNMULD DPR:$a, DPR:$b)>,
495 Requires<[NoHonorSignDependentRounding,HasDPVFP]>;
496 def : Pat<(fmul (fneg SPR:$a), SPR:$b),
497 (VNMULS SPR:$a, SPR:$b)>, Requires<[NoHonorSignDependentRounding]>;
499 // These are encoded as unary instructions.
500 let Defs = [FPSCR_NZCV] in {
501 def VCMPED : ADuI<0b11101, 0b11, 0b0100, 0b11, 0,
502 (outs), (ins DPR:$Dd, DPR:$Dm),
503 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, $Dm",
504 [(arm_cmpfp DPR:$Dd, (f64 DPR:$Dm))]>;
506 def VCMPES : ASuI<0b11101, 0b11, 0b0100, 0b11, 0,
507 (outs), (ins SPR:$Sd, SPR:$Sm),
508 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, $Sm",
509 [(arm_cmpfp SPR:$Sd, SPR:$Sm)]> {
510 // Some single precision VFP instructions may be executed on both NEON and
511 // VFP pipelines on A8.
512 let D = VFPNeonA8Domain;
515 def VCMPEH : AHuI<0b11101, 0b11, 0b0100, 0b11, 0,
516 (outs), (ins SPR:$Sd, SPR:$Sm),
517 IIC_fpCMP16, "vcmpe", ".f16\t$Sd, $Sm",
521 // FIXME: Verify encoding after integrated assembler is working.
522 def VCMPD : ADuI<0b11101, 0b11, 0b0100, 0b01, 0,
523 (outs), (ins DPR:$Dd, DPR:$Dm),
524 IIC_fpCMP64, "vcmp", ".f64\t$Dd, $Dm",
525 [/* For disassembly only; pattern left blank */]>;
527 def VCMPS : ASuI<0b11101, 0b11, 0b0100, 0b01, 0,
528 (outs), (ins SPR:$Sd, SPR:$Sm),
529 IIC_fpCMP32, "vcmp", ".f32\t$Sd, $Sm",
530 [/* For disassembly only; pattern left blank */]> {
531 // Some single precision VFP instructions may be executed on both NEON and
532 // VFP pipelines on A8.
533 let D = VFPNeonA8Domain;
536 def VCMPH : AHuI<0b11101, 0b11, 0b0100, 0b01, 0,
537 (outs), (ins SPR:$Sd, SPR:$Sm),
538 IIC_fpCMP16, "vcmp", ".f16\t$Sd, $Sm",
540 } // Defs = [FPSCR_NZCV]
542 //===----------------------------------------------------------------------===//
543 // FP Unary Operations.
546 def VABSD : ADuI<0b11101, 0b11, 0b0000, 0b11, 0,
547 (outs DPR:$Dd), (ins DPR:$Dm),
548 IIC_fpUNA64, "vabs", ".f64\t$Dd, $Dm",
549 [(set DPR:$Dd, (fabs (f64 DPR:$Dm)))]>;
551 def VABSS : ASuIn<0b11101, 0b11, 0b0000, 0b11, 0,
552 (outs SPR:$Sd), (ins SPR:$Sm),
553 IIC_fpUNA32, "vabs", ".f32\t$Sd, $Sm",
554 [(set SPR:$Sd, (fabs SPR:$Sm))]> {
555 // Some single precision VFP instructions may be executed on both NEON and
556 // VFP pipelines on A8.
557 let D = VFPNeonA8Domain;
560 def VABSH : AHuI<0b11101, 0b11, 0b0000, 0b11, 0,
561 (outs SPR:$Sd), (ins SPR:$Sm),
562 IIC_fpUNA16, "vabs", ".f16\t$Sd, $Sm",
565 let Defs = [FPSCR_NZCV] in {
566 def VCMPEZD : ADuI<0b11101, 0b11, 0b0101, 0b11, 0,
567 (outs), (ins DPR:$Dd),
568 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, #0",
569 [(arm_cmpfp0 (f64 DPR:$Dd))]> {
570 let Inst{3-0} = 0b0000;
574 def VCMPEZS : ASuI<0b11101, 0b11, 0b0101, 0b11, 0,
575 (outs), (ins SPR:$Sd),
576 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, #0",
577 [(arm_cmpfp0 SPR:$Sd)]> {
578 let Inst{3-0} = 0b0000;
581 // Some single precision VFP instructions may be executed on both NEON and
582 // VFP pipelines on A8.
583 let D = VFPNeonA8Domain;
586 def VCMPEZH : AHuI<0b11101, 0b11, 0b0101, 0b11, 0,
587 (outs), (ins SPR:$Sd),
588 IIC_fpCMP16, "vcmpe", ".f16\t$Sd, #0",
590 let Inst{3-0} = 0b0000;
594 // FIXME: Verify encoding after integrated assembler is working.
595 def VCMPZD : ADuI<0b11101, 0b11, 0b0101, 0b01, 0,
596 (outs), (ins DPR:$Dd),
597 IIC_fpCMP64, "vcmp", ".f64\t$Dd, #0",
598 [/* For disassembly only; pattern left blank */]> {
599 let Inst{3-0} = 0b0000;
603 def VCMPZS : ASuI<0b11101, 0b11, 0b0101, 0b01, 0,
604 (outs), (ins SPR:$Sd),
605 IIC_fpCMP32, "vcmp", ".f32\t$Sd, #0",
606 [/* For disassembly only; pattern left blank */]> {
607 let Inst{3-0} = 0b0000;
610 // Some single precision VFP instructions may be executed on both NEON and
611 // VFP pipelines on A8.
612 let D = VFPNeonA8Domain;
615 def VCMPZH : AHuI<0b11101, 0b11, 0b0101, 0b01, 0,
616 (outs), (ins SPR:$Sd),
617 IIC_fpCMP16, "vcmp", ".f16\t$Sd, #0",
619 let Inst{3-0} = 0b0000;
622 } // Defs = [FPSCR_NZCV]
624 def VCVTDS : ASuI<0b11101, 0b11, 0b0111, 0b11, 0,
625 (outs DPR:$Dd), (ins SPR:$Sm),
626 IIC_fpCVTDS, "vcvt", ".f64.f32\t$Dd, $Sm",
627 [(set DPR:$Dd, (fextend SPR:$Sm))]> {
628 // Instruction operands.
632 // Encode instruction operands.
633 let Inst{3-0} = Sm{4-1};
635 let Inst{15-12} = Dd{3-0};
636 let Inst{22} = Dd{4};
638 let Predicates = [HasVFP2, HasDPVFP];
641 // Special case encoding: bits 11-8 is 0b1011.
642 def VCVTSD : VFPAI<(outs SPR:$Sd), (ins DPR:$Dm), VFPUnaryFrm,
643 IIC_fpCVTSD, "vcvt", ".f32.f64\t$Sd, $Dm",
644 [(set SPR:$Sd, (fround DPR:$Dm))]> {
645 // Instruction operands.
649 // Encode instruction operands.
650 let Inst{3-0} = Dm{3-0};
652 let Inst{15-12} = Sd{4-1};
653 let Inst{22} = Sd{0};
655 let Inst{27-23} = 0b11101;
656 let Inst{21-16} = 0b110111;
657 let Inst{11-8} = 0b1011;
658 let Inst{7-6} = 0b11;
661 let Predicates = [HasVFP2, HasDPVFP];
664 // Between half, single and double-precision. For disassembly only.
666 // FIXME: Verify encoding after integrated assembler is working.
667 def VCVTBHS: ASuI<0b11101, 0b11, 0b0010, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
668 /* FIXME */ IIC_fpCVTSH, "vcvtb", ".f32.f16\t$Sd, $Sm",
669 [/* For disassembly only; pattern left blank */]>,
672 def VCVTBSH: ASuI<0b11101, 0b11, 0b0011, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
673 /* FIXME */ IIC_fpCVTHS, "vcvtb", ".f16.f32\t$Sd, $Sm",
674 [/* For disassembly only; pattern left blank */]>,
677 def VCVTTHS: ASuI<0b11101, 0b11, 0b0010, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
678 /* FIXME */ IIC_fpCVTSH, "vcvtt", ".f32.f16\t$Sd, $Sm",
679 [/* For disassembly only; pattern left blank */]>,
682 def VCVTTSH: ASuI<0b11101, 0b11, 0b0011, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
683 /* FIXME */ IIC_fpCVTHS, "vcvtt", ".f16.f32\t$Sd, $Sm",
684 [/* For disassembly only; pattern left blank */]>,
687 def VCVTBHD : ADuI<0b11101, 0b11, 0b0010, 0b01, 0,
688 (outs DPR:$Dd), (ins SPR:$Sm),
689 NoItinerary, "vcvtb", ".f64.f16\t$Dd, $Sm",
690 []>, Requires<[HasFPARMv8, HasDPVFP]> {
691 // Instruction operands.
694 // Encode instruction operands.
695 let Inst{3-0} = Sm{4-1};
699 def VCVTBDH : ADuI<0b11101, 0b11, 0b0011, 0b01, 0,
700 (outs SPR:$Sd), (ins DPR:$Dm),
701 NoItinerary, "vcvtb", ".f16.f64\t$Sd, $Dm",
702 []>, Requires<[HasFPARMv8, HasDPVFP]> {
703 // Instruction operands.
707 // Encode instruction operands.
708 let Inst{3-0} = Dm{3-0};
710 let Inst{15-12} = Sd{4-1};
711 let Inst{22} = Sd{0};
714 def VCVTTHD : ADuI<0b11101, 0b11, 0b0010, 0b11, 0,
715 (outs DPR:$Dd), (ins SPR:$Sm),
716 NoItinerary, "vcvtt", ".f64.f16\t$Dd, $Sm",
717 []>, Requires<[HasFPARMv8, HasDPVFP]> {
718 // Instruction operands.
721 // Encode instruction operands.
722 let Inst{3-0} = Sm{4-1};
726 def VCVTTDH : ADuI<0b11101, 0b11, 0b0011, 0b11, 0,
727 (outs SPR:$Sd), (ins DPR:$Dm),
728 NoItinerary, "vcvtt", ".f16.f64\t$Sd, $Dm",
729 []>, Requires<[HasFPARMv8, HasDPVFP]> {
730 // Instruction operands.
734 // Encode instruction operands.
735 let Inst{15-12} = Sd{4-1};
736 let Inst{22} = Sd{0};
737 let Inst{3-0} = Dm{3-0};
741 def : Pat<(fp_to_f16 SPR:$a),
742 (i32 (COPY_TO_REGCLASS (VCVTBSH SPR:$a), GPR))>;
744 def : Pat<(fp_to_f16 (f64 DPR:$a)),
745 (i32 (COPY_TO_REGCLASS (VCVTBDH DPR:$a), GPR))>;
747 def : Pat<(f16_to_fp GPR:$a),
748 (VCVTBHS (COPY_TO_REGCLASS GPR:$a, SPR))>;
750 def : Pat<(f64 (f16_to_fp GPR:$a)),
751 (VCVTBHD (COPY_TO_REGCLASS GPR:$a, SPR))>;
753 multiclass vcvt_inst<string opc, bits<2> rm,
754 SDPatternOperator node = null_frag> {
755 let PostEncoderMethod = "", DecoderNamespace = "VFPV8" in {
756 def SH : AHuInp<0b11101, 0b11, 0b1100, 0b11, 0,
757 (outs SPR:$Sd), (ins SPR:$Sm),
758 NoItinerary, !strconcat("vcvt", opc, ".s32.f16\t$Sd, $Sm"),
760 Requires<[HasFullFP16]> {
761 let Inst{17-16} = rm;
764 def UH : AHuInp<0b11101, 0b11, 0b1100, 0b01, 0,
765 (outs SPR:$Sd), (ins SPR:$Sm),
766 NoItinerary, !strconcat("vcvt", opc, ".u32.f16\t$Sd, $Sm"),
768 Requires<[HasFullFP16]> {
769 let Inst{17-16} = rm;
772 def SS : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
773 (outs SPR:$Sd), (ins SPR:$Sm),
774 NoItinerary, !strconcat("vcvt", opc, ".s32.f32\t$Sd, $Sm"),
776 Requires<[HasFPARMv8]> {
777 let Inst{17-16} = rm;
780 def US : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
781 (outs SPR:$Sd), (ins SPR:$Sm),
782 NoItinerary, !strconcat("vcvt", opc, ".u32.f32\t$Sd, $Sm"),
784 Requires<[HasFPARMv8]> {
785 let Inst{17-16} = rm;
788 def SD : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
789 (outs SPR:$Sd), (ins DPR:$Dm),
790 NoItinerary, !strconcat("vcvt", opc, ".s32.f64\t$Sd, $Dm"),
792 Requires<[HasFPARMv8, HasDPVFP]> {
795 let Inst{17-16} = rm;
797 // Encode instruction operands
798 let Inst{3-0} = Dm{3-0};
803 def UD : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
804 (outs SPR:$Sd), (ins DPR:$Dm),
805 NoItinerary, !strconcat("vcvt", opc, ".u32.f64\t$Sd, $Dm"),
807 Requires<[HasFPARMv8, HasDPVFP]> {
810 let Inst{17-16} = rm;
812 // Encode instruction operands
813 let Inst{3-0} = Dm{3-0};
819 let Predicates = [HasFPARMv8] in {
820 def : Pat<(i32 (fp_to_sint (node SPR:$a))),
822 (!cast<Instruction>(NAME#"SS") SPR:$a),
824 def : Pat<(i32 (fp_to_uint (node SPR:$a))),
826 (!cast<Instruction>(NAME#"US") SPR:$a),
829 let Predicates = [HasFPARMv8, HasDPVFP] in {
830 def : Pat<(i32 (fp_to_sint (node (f64 DPR:$a)))),
832 (!cast<Instruction>(NAME#"SD") DPR:$a),
834 def : Pat<(i32 (fp_to_uint (node (f64 DPR:$a)))),
836 (!cast<Instruction>(NAME#"UD") DPR:$a),
841 defm VCVTA : vcvt_inst<"a", 0b00, frnd>;
842 defm VCVTN : vcvt_inst<"n", 0b01>;
843 defm VCVTP : vcvt_inst<"p", 0b10, fceil>;
844 defm VCVTM : vcvt_inst<"m", 0b11, ffloor>;
846 def VNEGD : ADuI<0b11101, 0b11, 0b0001, 0b01, 0,
847 (outs DPR:$Dd), (ins DPR:$Dm),
848 IIC_fpUNA64, "vneg", ".f64\t$Dd, $Dm",
849 [(set DPR:$Dd, (fneg (f64 DPR:$Dm)))]>;
851 def VNEGS : ASuIn<0b11101, 0b11, 0b0001, 0b01, 0,
852 (outs SPR:$Sd), (ins SPR:$Sm),
853 IIC_fpUNA32, "vneg", ".f32\t$Sd, $Sm",
854 [(set SPR:$Sd, (fneg SPR:$Sm))]> {
855 // Some single precision VFP instructions may be executed on both NEON and
856 // VFP pipelines on A8.
857 let D = VFPNeonA8Domain;
860 def VNEGH : AHuI<0b11101, 0b11, 0b0001, 0b01, 0,
861 (outs SPR:$Sd), (ins SPR:$Sm),
862 IIC_fpUNA16, "vneg", ".f16\t$Sd, $Sm",
865 multiclass vrint_inst_zrx<string opc, bit op, bit op2, SDPatternOperator node> {
866 def H : AHuI<0b11101, 0b11, 0b0110, 0b11, 0,
867 (outs SPR:$Sd), (ins SPR:$Sm),
868 NoItinerary, !strconcat("vrint", opc), ".f16\t$Sd, $Sm",
870 Requires<[HasFullFP16]> {
875 def S : ASuI<0b11101, 0b11, 0b0110, 0b11, 0,
876 (outs SPR:$Sd), (ins SPR:$Sm),
877 NoItinerary, !strconcat("vrint", opc), ".f32\t$Sd, $Sm",
878 [(set (f32 SPR:$Sd), (node (f32 SPR:$Sm)))]>,
879 Requires<[HasFPARMv8]> {
883 def D : ADuI<0b11101, 0b11, 0b0110, 0b11, 0,
884 (outs DPR:$Dd), (ins DPR:$Dm),
885 NoItinerary, !strconcat("vrint", opc), ".f64\t$Dd, $Dm",
886 [(set (f64 DPR:$Dd), (node (f64 DPR:$Dm)))]>,
887 Requires<[HasFPARMv8, HasDPVFP]> {
892 def : InstAlias<!strconcat("vrint", opc, "$p.f16.f16\t$Sd, $Sm"),
893 (!cast<Instruction>(NAME#"H") SPR:$Sd, SPR:$Sm, pred:$p), 0>,
894 Requires<[HasFullFP16]>;
895 def : InstAlias<!strconcat("vrint", opc, "$p.f32.f32\t$Sd, $Sm"),
896 (!cast<Instruction>(NAME#"S") SPR:$Sd, SPR:$Sm, pred:$p), 0>,
897 Requires<[HasFPARMv8]>;
898 def : InstAlias<!strconcat("vrint", opc, "$p.f64.f64\t$Dd, $Dm"),
899 (!cast<Instruction>(NAME#"D") DPR:$Dd, DPR:$Dm, pred:$p), 0>,
900 Requires<[HasFPARMv8,HasDPVFP]>;
903 defm VRINTZ : vrint_inst_zrx<"z", 0, 1, ftrunc>;
904 defm VRINTR : vrint_inst_zrx<"r", 0, 0, fnearbyint>;
905 defm VRINTX : vrint_inst_zrx<"x", 1, 0, frint>;
907 multiclass vrint_inst_anpm<string opc, bits<2> rm,
908 SDPatternOperator node = null_frag> {
909 let PostEncoderMethod = "", DecoderNamespace = "VFPV8" in {
910 def H : AHuInp<0b11101, 0b11, 0b1000, 0b01, 0,
911 (outs SPR:$Sd), (ins SPR:$Sm),
912 NoItinerary, !strconcat("vrint", opc, ".f16\t$Sd, $Sm"),
914 Requires<[HasFullFP16]> {
915 let Inst{17-16} = rm;
917 def S : ASuInp<0b11101, 0b11, 0b1000, 0b01, 0,
918 (outs SPR:$Sd), (ins SPR:$Sm),
919 NoItinerary, !strconcat("vrint", opc, ".f32\t$Sd, $Sm"),
920 [(set (f32 SPR:$Sd), (node (f32 SPR:$Sm)))]>,
921 Requires<[HasFPARMv8]> {
922 let Inst{17-16} = rm;
924 def D : ADuInp<0b11101, 0b11, 0b1000, 0b01, 0,
925 (outs DPR:$Dd), (ins DPR:$Dm),
926 NoItinerary, !strconcat("vrint", opc, ".f64\t$Dd, $Dm"),
927 [(set (f64 DPR:$Dd), (node (f64 DPR:$Dm)))]>,
928 Requires<[HasFPARMv8, HasDPVFP]> {
929 let Inst{17-16} = rm;
933 def : InstAlias<!strconcat("vrint", opc, ".f32.f32\t$Sd, $Sm"),
934 (!cast<Instruction>(NAME#"S") SPR:$Sd, SPR:$Sm), 0>,
935 Requires<[HasFPARMv8]>;
936 def : InstAlias<!strconcat("vrint", opc, ".f64.f64\t$Dd, $Dm"),
937 (!cast<Instruction>(NAME#"D") DPR:$Dd, DPR:$Dm), 0>,
938 Requires<[HasFPARMv8,HasDPVFP]>;
941 defm VRINTA : vrint_inst_anpm<"a", 0b00, frnd>;
942 defm VRINTN : vrint_inst_anpm<"n", 0b01>;
943 defm VRINTP : vrint_inst_anpm<"p", 0b10, fceil>;
944 defm VRINTM : vrint_inst_anpm<"m", 0b11, ffloor>;
946 def VSQRTD : ADuI<0b11101, 0b11, 0b0001, 0b11, 0,
947 (outs DPR:$Dd), (ins DPR:$Dm),
948 IIC_fpSQRT64, "vsqrt", ".f64\t$Dd, $Dm",
949 [(set DPR:$Dd, (fsqrt (f64 DPR:$Dm)))]>;
951 def VSQRTS : ASuI<0b11101, 0b11, 0b0001, 0b11, 0,
952 (outs SPR:$Sd), (ins SPR:$Sm),
953 IIC_fpSQRT32, "vsqrt", ".f32\t$Sd, $Sm",
954 [(set SPR:$Sd, (fsqrt SPR:$Sm))]>;
956 def VSQRTH : AHuI<0b11101, 0b11, 0b0001, 0b11, 0,
957 (outs SPR:$Sd), (ins SPR:$Sm),
958 IIC_fpSQRT16, "vsqrt", ".f16\t$Sd, $Sm",
961 let hasSideEffects = 0 in {
962 def VMOVD : ADuI<0b11101, 0b11, 0b0000, 0b01, 0,
963 (outs DPR:$Dd), (ins DPR:$Dm),
964 IIC_fpUNA64, "vmov", ".f64\t$Dd, $Dm", []>;
966 def VMOVS : ASuI<0b11101, 0b11, 0b0000, 0b01, 0,
967 (outs SPR:$Sd), (ins SPR:$Sm),
968 IIC_fpUNA32, "vmov", ".f32\t$Sd, $Sm", []>;
970 let PostEncoderMethod = "", DecoderNamespace = "VFPV8" in {
971 def VMOVH : ASuInp<0b11101, 0b11, 0b0000, 0b01, 0,
972 (outs SPR:$Sd), (ins SPR:$Sm),
973 IIC_fpUNA16, "vmovx.f16\t$Sd, $Sm", []>,
974 Requires<[HasFullFP16]>;
976 def VINSH : ASuInp<0b11101, 0b11, 0b0000, 0b11, 0,
977 (outs SPR:$Sd), (ins SPR:$Sm),
978 IIC_fpUNA16, "vins.f16\t$Sd, $Sm", []>,
979 Requires<[HasFullFP16]>;
980 } // PostEncoderMethod
983 //===----------------------------------------------------------------------===//
984 // FP <-> GPR Copies. Int <-> FP Conversions.
987 def VMOVRS : AVConv2I<0b11100001, 0b1010,
988 (outs GPR:$Rt), (ins SPR:$Sn),
989 IIC_fpMOVSI, "vmov", "\t$Rt, $Sn",
990 [(set GPR:$Rt, (bitconvert SPR:$Sn))]> {
991 // Instruction operands.
995 // Encode instruction operands.
996 let Inst{19-16} = Sn{4-1};
998 let Inst{15-12} = Rt;
1000 let Inst{6-5} = 0b00;
1001 let Inst{3-0} = 0b0000;
1003 // Some single precision VFP instructions may be executed on both NEON and VFP
1005 let D = VFPNeonDomain;
1008 // Bitcast i32 -> f32. NEON prefers to use VMOVDRR.
1009 def VMOVSR : AVConv4I<0b11100000, 0b1010,
1010 (outs SPR:$Sn), (ins GPR:$Rt),
1011 IIC_fpMOVIS, "vmov", "\t$Sn, $Rt",
1012 [(set SPR:$Sn, (bitconvert GPR:$Rt))]>,
1013 Requires<[HasVFP2, UseVMOVSR]> {
1014 // Instruction operands.
1018 // Encode instruction operands.
1019 let Inst{19-16} = Sn{4-1};
1020 let Inst{7} = Sn{0};
1021 let Inst{15-12} = Rt;
1023 let Inst{6-5} = 0b00;
1024 let Inst{3-0} = 0b0000;
1026 // Some single precision VFP instructions may be executed on both NEON and VFP
1028 let D = VFPNeonDomain;
1031 let hasSideEffects = 0 in {
1032 def VMOVRRD : AVConv3I<0b11000101, 0b1011,
1033 (outs GPR:$Rt, GPR:$Rt2), (ins DPR:$Dm),
1034 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $Dm",
1035 [/* FIXME: Can't write pattern for multiple result instr*/]> {
1036 // Instruction operands.
1041 // Encode instruction operands.
1042 let Inst{3-0} = Dm{3-0};
1043 let Inst{5} = Dm{4};
1044 let Inst{15-12} = Rt;
1045 let Inst{19-16} = Rt2;
1047 let Inst{7-6} = 0b00;
1049 // Some single precision VFP instructions may be executed on both NEON and VFP
1051 let D = VFPNeonDomain;
1053 // This instruction is equivalent to
1054 // $Rt = EXTRACT_SUBREG $Dm, ssub_0
1055 // $Rt2 = EXTRACT_SUBREG $Dm, ssub_1
1056 let isExtractSubreg = 1;
1059 def VMOVRRS : AVConv3I<0b11000101, 0b1010,
1060 (outs GPR:$Rt, GPR:$Rt2), (ins SPR:$src1, SPR:$src2),
1061 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $src1, $src2",
1062 [/* For disassembly only; pattern left blank */]> {
1067 // Encode instruction operands.
1068 let Inst{3-0} = src1{4-1};
1069 let Inst{5} = src1{0};
1070 let Inst{15-12} = Rt;
1071 let Inst{19-16} = Rt2;
1073 let Inst{7-6} = 0b00;
1075 // Some single precision VFP instructions may be executed on both NEON and VFP
1077 let D = VFPNeonDomain;
1078 let DecoderMethod = "DecodeVMOVRRS";
1082 // FMDHR: GPR -> SPR
1083 // FMDLR: GPR -> SPR
1085 def VMOVDRR : AVConv5I<0b11000100, 0b1011,
1086 (outs DPR:$Dm), (ins GPR:$Rt, GPR:$Rt2),
1087 IIC_fpMOVID, "vmov", "\t$Dm, $Rt, $Rt2",
1088 [(set DPR:$Dm, (arm_fmdrr GPR:$Rt, GPR:$Rt2))]> {
1089 // Instruction operands.
1094 // Encode instruction operands.
1095 let Inst{3-0} = Dm{3-0};
1096 let Inst{5} = Dm{4};
1097 let Inst{15-12} = Rt;
1098 let Inst{19-16} = Rt2;
1100 let Inst{7-6} = 0b00;
1102 // Some single precision VFP instructions may be executed on both NEON and VFP
1104 let D = VFPNeonDomain;
1106 // This instruction is equivalent to
1107 // $Dm = REG_SEQUENCE $Rt, ssub_0, $Rt2, ssub_1
1108 let isRegSequence = 1;
1111 // Hoist an fabs or a fneg of a value coming from integer registers
1112 // and do the fabs/fneg on the integer value. This is never a lose
1113 // and could enable the conversion to float to be removed completely.
1114 def : Pat<(fabs (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1115 (VMOVDRR GPR:$Rl, (BFC GPR:$Rh, (i32 0x7FFFFFFF)))>,
1116 Requires<[IsARM, HasV6T2]>;
1117 def : Pat<(fabs (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1118 (VMOVDRR GPR:$Rl, (t2BFC GPR:$Rh, (i32 0x7FFFFFFF)))>,
1119 Requires<[IsThumb2, HasV6T2]>;
1120 def : Pat<(fneg (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1121 (VMOVDRR GPR:$Rl, (EORri GPR:$Rh, (i32 0x80000000)))>,
1123 def : Pat<(fneg (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1124 (VMOVDRR GPR:$Rl, (t2EORri GPR:$Rh, (i32 0x80000000)))>,
1125 Requires<[IsThumb2]>;
1127 let hasSideEffects = 0 in
1128 def VMOVSRR : AVConv5I<0b11000100, 0b1010,
1129 (outs SPR:$dst1, SPR:$dst2), (ins GPR:$src1, GPR:$src2),
1130 IIC_fpMOVID, "vmov", "\t$dst1, $dst2, $src1, $src2",
1131 [/* For disassembly only; pattern left blank */]> {
1132 // Instruction operands.
1137 // Encode instruction operands.
1138 let Inst{3-0} = dst1{4-1};
1139 let Inst{5} = dst1{0};
1140 let Inst{15-12} = src1;
1141 let Inst{19-16} = src2;
1143 let Inst{7-6} = 0b00;
1145 // Some single precision VFP instructions may be executed on both NEON and VFP
1147 let D = VFPNeonDomain;
1149 let DecoderMethod = "DecodeVMOVSRR";
1152 // Move H->R, clearing top 16 bits
1153 def VMOVRH : AVConv2I<0b11100001, 0b1001,
1154 (outs GPR:$Rt), (ins SPR:$Sn),
1155 IIC_fpMOVSI, "vmov", ".f16\t$Rt, $Sn",
1157 Requires<[HasFullFP16]> {
1158 // Instruction operands.
1162 // Encode instruction operands.
1163 let Inst{19-16} = Sn{4-1};
1164 let Inst{7} = Sn{0};
1165 let Inst{15-12} = Rt;
1167 let Inst{6-5} = 0b00;
1168 let Inst{3-0} = 0b0000;
1171 // Move R->H, clearing top 16 bits
1172 def VMOVHR : AVConv4I<0b11100000, 0b1001,
1173 (outs SPR:$Sn), (ins GPR:$Rt),
1174 IIC_fpMOVIS, "vmov", ".f16\t$Sn, $Rt",
1176 Requires<[HasFullFP16]> {
1177 // Instruction operands.
1181 // Encode instruction operands.
1182 let Inst{19-16} = Sn{4-1};
1183 let Inst{7} = Sn{0};
1184 let Inst{15-12} = Rt;
1186 let Inst{6-5} = 0b00;
1187 let Inst{3-0} = 0b0000;
1190 // FMRDH: SPR -> GPR
1191 // FMRDL: SPR -> GPR
1192 // FMRRS: SPR -> GPR
1193 // FMRX: SPR system reg -> GPR
1194 // FMSRR: GPR -> SPR
1195 // FMXR: GPR -> VFP system reg
1200 class AVConv1IDs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1201 bits<4> opcod4, dag oops, dag iops,
1202 InstrItinClass itin, string opc, string asm,
1204 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1206 // Instruction operands.
1210 // Encode instruction operands.
1211 let Inst{3-0} = Sm{4-1};
1212 let Inst{5} = Sm{0};
1213 let Inst{15-12} = Dd{3-0};
1214 let Inst{22} = Dd{4};
1216 let Predicates = [HasVFP2, HasDPVFP];
1219 class AVConv1InSs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1220 bits<4> opcod4, dag oops, dag iops,InstrItinClass itin,
1221 string opc, string asm, list<dag> pattern>
1222 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1224 // Instruction operands.
1228 // Encode instruction operands.
1229 let Inst{3-0} = Sm{4-1};
1230 let Inst{5} = Sm{0};
1231 let Inst{15-12} = Sd{4-1};
1232 let Inst{22} = Sd{0};
1235 class AVConv1IHs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1236 bits<4> opcod4, dag oops, dag iops,
1237 InstrItinClass itin, string opc, string asm,
1239 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1241 // Instruction operands.
1245 // Encode instruction operands.
1246 let Inst{3-0} = Sm{4-1};
1247 let Inst{5} = Sm{0};
1248 let Inst{15-12} = Sd{4-1};
1249 let Inst{22} = Sd{0};
1251 let Predicates = [HasFullFP16];
1254 def VSITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
1255 (outs DPR:$Dd), (ins SPR:$Sm),
1256 IIC_fpCVTID, "vcvt", ".f64.s32\t$Dd, $Sm",
1258 let Inst{7} = 1; // s32
1261 let Predicates=[HasVFP2, HasDPVFP] in {
1262 def : VFPPat<(f64 (sint_to_fp GPR:$a)),
1263 (VSITOD (COPY_TO_REGCLASS GPR:$a, SPR))>;
1265 def : VFPPat<(f64 (sint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1266 (VSITOD (VLDRS addrmode5:$a))>;
1269 def VSITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
1270 (outs SPR:$Sd),(ins SPR:$Sm),
1271 IIC_fpCVTIS, "vcvt", ".f32.s32\t$Sd, $Sm",
1273 let Inst{7} = 1; // s32
1275 // Some single precision VFP instructions may be executed on both NEON and
1276 // VFP pipelines on A8.
1277 let D = VFPNeonA8Domain;
1280 def : VFPNoNEONPat<(f32 (sint_to_fp GPR:$a)),
1281 (VSITOS (COPY_TO_REGCLASS GPR:$a, SPR))>;
1283 def : VFPNoNEONPat<(f32 (sint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1284 (VSITOS (VLDRS addrmode5:$a))>;
1286 def VSITOH : AVConv1IHs_Encode<0b11101, 0b11, 0b1000, 0b1001,
1287 (outs SPR:$Sd), (ins SPR:$Sm),
1288 IIC_fpCVTIH, "vcvt", ".f16.s32\t$Sd, $Sm",
1290 let Inst{7} = 1; // s32
1293 def VUITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
1294 (outs DPR:$Dd), (ins SPR:$Sm),
1295 IIC_fpCVTID, "vcvt", ".f64.u32\t$Dd, $Sm",
1297 let Inst{7} = 0; // u32
1300 let Predicates=[HasVFP2, HasDPVFP] in {
1301 def : VFPPat<(f64 (uint_to_fp GPR:$a)),
1302 (VUITOD (COPY_TO_REGCLASS GPR:$a, SPR))>;
1304 def : VFPPat<(f64 (uint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1305 (VUITOD (VLDRS addrmode5:$a))>;
1308 def VUITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
1309 (outs SPR:$Sd), (ins SPR:$Sm),
1310 IIC_fpCVTIS, "vcvt", ".f32.u32\t$Sd, $Sm",
1312 let Inst{7} = 0; // u32
1314 // Some single precision VFP instructions may be executed on both NEON and
1315 // VFP pipelines on A8.
1316 let D = VFPNeonA8Domain;
1319 def : VFPNoNEONPat<(f32 (uint_to_fp GPR:$a)),
1320 (VUITOS (COPY_TO_REGCLASS GPR:$a, SPR))>;
1322 def : VFPNoNEONPat<(f32 (uint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1323 (VUITOS (VLDRS addrmode5:$a))>;
1325 def VUITOH : AVConv1IHs_Encode<0b11101, 0b11, 0b1000, 0b1001,
1326 (outs SPR:$Sd), (ins SPR:$Sm),
1327 IIC_fpCVTIH, "vcvt", ".f16.u32\t$Sd, $Sm",
1329 let Inst{7} = 0; // u32
1334 class AVConv1IsD_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1335 bits<4> opcod4, dag oops, dag iops,
1336 InstrItinClass itin, string opc, string asm,
1338 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1340 // Instruction operands.
1344 // Encode instruction operands.
1345 let Inst{3-0} = Dm{3-0};
1346 let Inst{5} = Dm{4};
1347 let Inst{15-12} = Sd{4-1};
1348 let Inst{22} = Sd{0};
1350 let Predicates = [HasVFP2, HasDPVFP];
1353 class AVConv1InsS_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1354 bits<4> opcod4, dag oops, dag iops,
1355 InstrItinClass itin, string opc, string asm,
1357 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1359 // Instruction operands.
1363 // Encode instruction operands.
1364 let Inst{3-0} = Sm{4-1};
1365 let Inst{5} = Sm{0};
1366 let Inst{15-12} = Sd{4-1};
1367 let Inst{22} = Sd{0};
1370 class AVConv1IsH_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1371 bits<4> opcod4, dag oops, dag iops,
1372 InstrItinClass itin, string opc, string asm,
1374 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1376 // Instruction operands.
1380 // Encode instruction operands.
1381 let Inst{3-0} = Sm{4-1};
1382 let Inst{5} = Sm{0};
1383 let Inst{15-12} = Sd{4-1};
1384 let Inst{22} = Sd{0};
1386 let Predicates = [HasFullFP16];
1389 // Always set Z bit in the instruction, i.e. "round towards zero" variants.
1390 def VTOSIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
1391 (outs SPR:$Sd), (ins DPR:$Dm),
1392 IIC_fpCVTDI, "vcvt", ".s32.f64\t$Sd, $Dm",
1394 let Inst{7} = 1; // Z bit
1397 let Predicates=[HasVFP2, HasDPVFP] in {
1398 def : VFPPat<(i32 (fp_to_sint (f64 DPR:$a))),
1399 (COPY_TO_REGCLASS (VTOSIZD DPR:$a), GPR)>;
1401 def : VFPPat<(alignedstore32 (i32 (fp_to_sint (f64 DPR:$a))), addrmode5:$ptr),
1402 (VSTRS (VTOSIZD DPR:$a), addrmode5:$ptr)>;
1405 def VTOSIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
1406 (outs SPR:$Sd), (ins SPR:$Sm),
1407 IIC_fpCVTSI, "vcvt", ".s32.f32\t$Sd, $Sm",
1409 let Inst{7} = 1; // Z bit
1411 // Some single precision VFP instructions may be executed on both NEON and
1412 // VFP pipelines on A8.
1413 let D = VFPNeonA8Domain;
1416 def : VFPNoNEONPat<(i32 (fp_to_sint SPR:$a)),
1417 (COPY_TO_REGCLASS (VTOSIZS SPR:$a), GPR)>;
1419 def : VFPNoNEONPat<(alignedstore32 (i32 (fp_to_sint (f32 SPR:$a))),
1421 (VSTRS (VTOSIZS SPR:$a), addrmode5:$ptr)>;
1423 def VTOSIZH : AVConv1IsH_Encode<0b11101, 0b11, 0b1101, 0b1001,
1424 (outs SPR:$Sd), (ins SPR:$Sm),
1425 IIC_fpCVTHI, "vcvt", ".s32.f16\t$Sd, $Sm",
1427 let Inst{7} = 1; // Z bit
1430 def VTOUIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
1431 (outs SPR:$Sd), (ins DPR:$Dm),
1432 IIC_fpCVTDI, "vcvt", ".u32.f64\t$Sd, $Dm",
1434 let Inst{7} = 1; // Z bit
1437 let Predicates=[HasVFP2, HasDPVFP] in {
1438 def : VFPPat<(i32 (fp_to_uint (f64 DPR:$a))),
1439 (COPY_TO_REGCLASS (VTOUIZD DPR:$a), GPR)>;
1441 def : VFPPat<(alignedstore32 (i32 (fp_to_uint (f64 DPR:$a))), addrmode5:$ptr),
1442 (VSTRS (VTOUIZD DPR:$a), addrmode5:$ptr)>;
1445 def VTOUIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
1446 (outs SPR:$Sd), (ins SPR:$Sm),
1447 IIC_fpCVTSI, "vcvt", ".u32.f32\t$Sd, $Sm",
1449 let Inst{7} = 1; // Z bit
1451 // Some single precision VFP instructions may be executed on both NEON and
1452 // VFP pipelines on A8.
1453 let D = VFPNeonA8Domain;
1456 def : VFPNoNEONPat<(i32 (fp_to_uint SPR:$a)),
1457 (COPY_TO_REGCLASS (VTOUIZS SPR:$a), GPR)>;
1459 def : VFPNoNEONPat<(alignedstore32 (i32 (fp_to_uint (f32 SPR:$a))),
1461 (VSTRS (VTOUIZS SPR:$a), addrmode5:$ptr)>;
1463 def VTOUIZH : AVConv1IsH_Encode<0b11101, 0b11, 0b1100, 0b1001,
1464 (outs SPR:$Sd), (ins SPR:$Sm),
1465 IIC_fpCVTHI, "vcvt", ".u32.f16\t$Sd, $Sm",
1467 let Inst{7} = 1; // Z bit
1470 // And the Z bit '0' variants, i.e. use the rounding mode specified by FPSCR.
1471 let Uses = [FPSCR] in {
1472 // FIXME: Verify encoding after integrated assembler is working.
1473 def VTOSIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
1474 (outs SPR:$Sd), (ins DPR:$Dm),
1475 IIC_fpCVTDI, "vcvtr", ".s32.f64\t$Sd, $Dm",
1476 [(set SPR:$Sd, (int_arm_vcvtr (f64 DPR:$Dm)))]>{
1477 let Inst{7} = 0; // Z bit
1480 def VTOSIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
1481 (outs SPR:$Sd), (ins SPR:$Sm),
1482 IIC_fpCVTSI, "vcvtr", ".s32.f32\t$Sd, $Sm",
1483 [(set SPR:$Sd, (int_arm_vcvtr SPR:$Sm))]> {
1484 let Inst{7} = 0; // Z bit
1487 def VTOSIRH : AVConv1IsH_Encode<0b11101, 0b11, 0b1101, 0b1001,
1488 (outs SPR:$Sd), (ins SPR:$Sm),
1489 IIC_fpCVTHI, "vcvtr", ".s32.f16\t$Sd, $Sm",
1491 let Inst{7} = 0; // Z bit
1494 def VTOUIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
1495 (outs SPR:$Sd), (ins DPR:$Dm),
1496 IIC_fpCVTDI, "vcvtr", ".u32.f64\t$Sd, $Dm",
1497 [(set SPR:$Sd, (int_arm_vcvtru(f64 DPR:$Dm)))]>{
1498 let Inst{7} = 0; // Z bit
1501 def VTOUIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
1502 (outs SPR:$Sd), (ins SPR:$Sm),
1503 IIC_fpCVTSI, "vcvtr", ".u32.f32\t$Sd, $Sm",
1504 [(set SPR:$Sd, (int_arm_vcvtru SPR:$Sm))]> {
1505 let Inst{7} = 0; // Z bit
1508 def VTOUIRH : AVConv1IsH_Encode<0b11101, 0b11, 0b1100, 0b1001,
1509 (outs SPR:$Sd), (ins SPR:$Sm),
1510 IIC_fpCVTHI, "vcvtr", ".u32.f16\t$Sd, $Sm",
1512 let Inst{7} = 0; // Z bit
1516 // Convert between floating-point and fixed-point
1517 // Data type for fixed-point naming convention:
1518 // S16 (U=0, sx=0) -> SH
1519 // U16 (U=1, sx=0) -> UH
1520 // S32 (U=0, sx=1) -> SL
1521 // U32 (U=1, sx=1) -> UL
1523 let Constraints = "$a = $dst" in {
1525 // FP to Fixed-Point:
1527 // Single Precision register
1528 class AVConv1XInsS_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
1529 bit op5, dag oops, dag iops, InstrItinClass itin,
1530 string opc, string asm, list<dag> pattern>
1531 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern>,
1532 Sched<[WriteCvtFP]> {
1534 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
1535 let Inst{22} = dst{0};
1536 let Inst{15-12} = dst{4-1};
1539 // Double Precision register
1540 class AVConv1XInsD_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
1541 bit op5, dag oops, dag iops, InstrItinClass itin,
1542 string opc, string asm, list<dag> pattern>
1543 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern>,
1544 Sched<[WriteCvtFP]> {
1546 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
1547 let Inst{22} = dst{4};
1548 let Inst{15-12} = dst{3-0};
1550 let Predicates = [HasVFP2, HasDPVFP];
1553 def VTOSHH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1001, 0,
1554 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1555 IIC_fpCVTHI, "vcvt", ".s16.f16\t$dst, $a, $fbits", []>,
1556 Requires<[HasFullFP16]>;
1558 def VTOUHH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1001, 0,
1559 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1560 IIC_fpCVTHI, "vcvt", ".u16.f16\t$dst, $a, $fbits", []>,
1561 Requires<[HasFullFP16]>;
1563 def VTOSLH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1001, 1,
1564 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1565 IIC_fpCVTHI, "vcvt", ".s32.f16\t$dst, $a, $fbits", []>,
1566 Requires<[HasFullFP16]>;
1568 def VTOULH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1001, 1,
1569 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1570 IIC_fpCVTHI, "vcvt", ".u32.f16\t$dst, $a, $fbits", []>,
1571 Requires<[HasFullFP16]>;
1573 def VTOSHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 0,
1574 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1575 IIC_fpCVTSI, "vcvt", ".s16.f32\t$dst, $a, $fbits", []> {
1576 // Some single precision VFP instructions may be executed on both NEON and
1577 // VFP pipelines on A8.
1578 let D = VFPNeonA8Domain;
1581 def VTOUHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 0,
1582 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1583 IIC_fpCVTSI, "vcvt", ".u16.f32\t$dst, $a, $fbits", []> {
1584 // Some single precision VFP instructions may be executed on both NEON and
1585 // VFP pipelines on A8.
1586 let D = VFPNeonA8Domain;
1589 def VTOSLS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 1,
1590 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1591 IIC_fpCVTSI, "vcvt", ".s32.f32\t$dst, $a, $fbits", []> {
1592 // Some single precision VFP instructions may be executed on both NEON and
1593 // VFP pipelines on A8.
1594 let D = VFPNeonA8Domain;
1597 def VTOULS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 1,
1598 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1599 IIC_fpCVTSI, "vcvt", ".u32.f32\t$dst, $a, $fbits", []> {
1600 // Some single precision VFP instructions may be executed on both NEON and
1601 // VFP pipelines on A8.
1602 let D = VFPNeonA8Domain;
1605 def VTOSHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 0,
1606 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1607 IIC_fpCVTDI, "vcvt", ".s16.f64\t$dst, $a, $fbits", []>;
1609 def VTOUHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 0,
1610 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1611 IIC_fpCVTDI, "vcvt", ".u16.f64\t$dst, $a, $fbits", []>;
1613 def VTOSLD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 1,
1614 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1615 IIC_fpCVTDI, "vcvt", ".s32.f64\t$dst, $a, $fbits", []>;
1617 def VTOULD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 1,
1618 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1619 IIC_fpCVTDI, "vcvt", ".u32.f64\t$dst, $a, $fbits", []>;
1621 // Fixed-Point to FP:
1623 def VSHTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1001, 0,
1624 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1625 IIC_fpCVTIH, "vcvt", ".f16.s16\t$dst, $a, $fbits", []>,
1626 Requires<[HasFullFP16]>;
1628 def VUHTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1001, 0,
1629 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1630 IIC_fpCVTIH, "vcvt", ".f16.u16\t$dst, $a, $fbits", []>,
1631 Requires<[HasFullFP16]>;
1633 def VSLTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1001, 1,
1634 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1635 IIC_fpCVTIH, "vcvt", ".f16.s32\t$dst, $a, $fbits", []>,
1636 Requires<[HasFullFP16]>;
1638 def VULTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1001, 1,
1639 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1640 IIC_fpCVTIH, "vcvt", ".f16.u32\t$dst, $a, $fbits", []>,
1641 Requires<[HasFullFP16]>;
1643 def VSHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 0,
1644 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1645 IIC_fpCVTIS, "vcvt", ".f32.s16\t$dst, $a, $fbits", []> {
1646 // Some single precision VFP instructions may be executed on both NEON and
1647 // VFP pipelines on A8.
1648 let D = VFPNeonA8Domain;
1651 def VUHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 0,
1652 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1653 IIC_fpCVTIS, "vcvt", ".f32.u16\t$dst, $a, $fbits", []> {
1654 // Some single precision VFP instructions may be executed on both NEON and
1655 // VFP pipelines on A8.
1656 let D = VFPNeonA8Domain;
1659 def VSLTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 1,
1660 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1661 IIC_fpCVTIS, "vcvt", ".f32.s32\t$dst, $a, $fbits", []> {
1662 // Some single precision VFP instructions may be executed on both NEON and
1663 // VFP pipelines on A8.
1664 let D = VFPNeonA8Domain;
1667 def VULTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 1,
1668 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1669 IIC_fpCVTIS, "vcvt", ".f32.u32\t$dst, $a, $fbits", []> {
1670 // Some single precision VFP instructions may be executed on both NEON and
1671 // VFP pipelines on A8.
1672 let D = VFPNeonA8Domain;
1675 def VSHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 0,
1676 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1677 IIC_fpCVTID, "vcvt", ".f64.s16\t$dst, $a, $fbits", []>;
1679 def VUHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 0,
1680 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1681 IIC_fpCVTID, "vcvt", ".f64.u16\t$dst, $a, $fbits", []>;
1683 def VSLTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 1,
1684 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1685 IIC_fpCVTID, "vcvt", ".f64.s32\t$dst, $a, $fbits", []>;
1687 def VULTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 1,
1688 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1689 IIC_fpCVTID, "vcvt", ".f64.u32\t$dst, $a, $fbits", []>;
1691 } // End of 'let Constraints = "$a = $dst" in'
1693 //===----------------------------------------------------------------------===//
1694 // FP Multiply-Accumulate Operations.
1697 def VMLAD : ADbI<0b11100, 0b00, 0, 0,
1698 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1699 IIC_fpMAC64, "vmla", ".f64\t$Dd, $Dn, $Dm",
1700 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1701 (f64 DPR:$Ddin)))]>,
1702 RegConstraint<"$Ddin = $Dd">,
1703 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1705 def VMLAS : ASbIn<0b11100, 0b00, 0, 0,
1706 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1707 IIC_fpMAC32, "vmla", ".f32\t$Sd, $Sn, $Sm",
1708 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
1710 RegConstraint<"$Sdin = $Sd">,
1711 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1712 // Some single precision VFP instructions may be executed on both NEON and
1713 // VFP pipelines on A8.
1714 let D = VFPNeonA8Domain;
1717 def VMLAH : AHbI<0b11100, 0b00, 0, 0,
1718 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1719 IIC_fpMAC16, "vmla", ".f16\t$Sd, $Sn, $Sm",
1721 RegConstraint<"$Sdin = $Sd">,
1722 Requires<[HasFullFP16,UseFPVMLx,DontUseFusedMAC]>;
1724 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1725 (VMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1726 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1727 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1728 (VMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1729 Requires<[HasVFP2,DontUseNEONForFP, UseFPVMLx,DontUseFusedMAC]>;
1731 def VMLSD : ADbI<0b11100, 0b00, 1, 0,
1732 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1733 IIC_fpMAC64, "vmls", ".f64\t$Dd, $Dn, $Dm",
1734 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1735 (f64 DPR:$Ddin)))]>,
1736 RegConstraint<"$Ddin = $Dd">,
1737 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1739 def VMLSS : ASbIn<0b11100, 0b00, 1, 0,
1740 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1741 IIC_fpMAC32, "vmls", ".f32\t$Sd, $Sn, $Sm",
1742 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1744 RegConstraint<"$Sdin = $Sd">,
1745 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1746 // Some single precision VFP instructions may be executed on both NEON and
1747 // VFP pipelines on A8.
1748 let D = VFPNeonA8Domain;
1751 def VMLSH : AHbI<0b11100, 0b00, 1, 0,
1752 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1753 IIC_fpMAC16, "vmls", ".f16\t$Sd, $Sn, $Sm",
1755 RegConstraint<"$Sdin = $Sd">,
1756 Requires<[HasFullFP16,UseFPVMLx,DontUseFusedMAC]>;
1758 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1759 (VMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
1760 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1761 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1762 (VMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
1763 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1765 def VNMLAD : ADbI<0b11100, 0b01, 1, 0,
1766 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1767 IIC_fpMAC64, "vnmla", ".f64\t$Dd, $Dn, $Dm",
1768 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1769 (f64 DPR:$Ddin)))]>,
1770 RegConstraint<"$Ddin = $Dd">,
1771 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1773 def VNMLAS : ASbI<0b11100, 0b01, 1, 0,
1774 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1775 IIC_fpMAC32, "vnmla", ".f32\t$Sd, $Sn, $Sm",
1776 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1778 RegConstraint<"$Sdin = $Sd">,
1779 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1780 // Some single precision VFP instructions may be executed on both NEON and
1781 // VFP pipelines on A8.
1782 let D = VFPNeonA8Domain;
1785 def VNMLAH : AHbI<0b11100, 0b01, 1, 0,
1786 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1787 IIC_fpMAC16, "vnmla", ".f16\t$Sd, $Sn, $Sm",
1789 RegConstraint<"$Sdin = $Sd">,
1790 Requires<[HasFullFP16,UseFPVMLx,DontUseFusedMAC]>;
1792 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
1793 (VNMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1794 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1795 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
1796 (VNMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1797 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1799 def VNMLSD : ADbI<0b11100, 0b01, 0, 0,
1800 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1801 IIC_fpMAC64, "vnmls", ".f64\t$Dd, $Dn, $Dm",
1802 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1803 (f64 DPR:$Ddin)))]>,
1804 RegConstraint<"$Ddin = $Dd">,
1805 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1807 def VNMLSS : ASbI<0b11100, 0b01, 0, 0,
1808 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1809 IIC_fpMAC32, "vnmls", ".f32\t$Sd, $Sn, $Sm",
1810 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
1811 RegConstraint<"$Sdin = $Sd">,
1812 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1813 // Some single precision VFP instructions may be executed on both NEON and
1814 // VFP pipelines on A8.
1815 let D = VFPNeonA8Domain;
1818 def VNMLSH : AHbI<0b11100, 0b01, 0, 0,
1819 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1820 IIC_fpMAC16, "vnmls", ".f16\t$Sd, $Sn, $Sm",
1822 RegConstraint<"$Sdin = $Sd">,
1823 Requires<[HasFullFP16,UseFPVMLx,DontUseFusedMAC]>;
1825 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
1826 (VNMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
1827 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1828 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
1829 (VNMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
1830 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1832 //===----------------------------------------------------------------------===//
1833 // Fused FP Multiply-Accumulate Operations.
1835 def VFMAD : ADbI<0b11101, 0b10, 0, 0,
1836 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1837 IIC_fpFMAC64, "vfma", ".f64\t$Dd, $Dn, $Dm",
1838 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1839 (f64 DPR:$Ddin)))]>,
1840 RegConstraint<"$Ddin = $Dd">,
1841 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1843 def VFMAS : ASbIn<0b11101, 0b10, 0, 0,
1844 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1845 IIC_fpFMAC32, "vfma", ".f32\t$Sd, $Sn, $Sm",
1846 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
1848 RegConstraint<"$Sdin = $Sd">,
1849 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1850 // Some single precision VFP instructions may be executed on both NEON and
1854 def VFMAH : AHbI<0b11101, 0b10, 0, 0,
1855 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1856 IIC_fpFMAC16, "vfma", ".f16\t$Sd, $Sn, $Sm",
1858 RegConstraint<"$Sdin = $Sd">,
1859 Requires<[HasFullFP16,UseFusedMAC]>;
1861 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1862 (VFMAD DPR:$dstin, DPR:$a, DPR:$b)>,
1863 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1864 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1865 (VFMAS SPR:$dstin, SPR:$a, SPR:$b)>,
1866 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1868 // Match @llvm.fma.* intrinsics
1869 // (fma x, y, z) -> (vfms z, x, y)
1870 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, DPR:$Ddin)),
1871 (VFMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1872 Requires<[HasVFP4,HasDPVFP]>;
1873 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, SPR:$Sdin)),
1874 (VFMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1875 Requires<[HasVFP4]>;
1877 def VFMSD : ADbI<0b11101, 0b10, 1, 0,
1878 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1879 IIC_fpFMAC64, "vfms", ".f64\t$Dd, $Dn, $Dm",
1880 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1881 (f64 DPR:$Ddin)))]>,
1882 RegConstraint<"$Ddin = $Dd">,
1883 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1885 def VFMSS : ASbIn<0b11101, 0b10, 1, 0,
1886 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1887 IIC_fpFMAC32, "vfms", ".f32\t$Sd, $Sn, $Sm",
1888 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1890 RegConstraint<"$Sdin = $Sd">,
1891 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1892 // Some single precision VFP instructions may be executed on both NEON and
1896 def VFMSH : AHbI<0b11101, 0b10, 1, 0,
1897 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1898 IIC_fpFMAC16, "vfms", ".f16\t$Sd, $Sn, $Sm",
1900 RegConstraint<"$Sdin = $Sd">,
1901 Requires<[HasFullFP16,UseFusedMAC]>;
1903 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1904 (VFMSD DPR:$dstin, DPR:$a, DPR:$b)>,
1905 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1906 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1907 (VFMSS SPR:$dstin, SPR:$a, SPR:$b)>,
1908 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1910 // Match @llvm.fma.* intrinsics
1911 // (fma (fneg x), y, z) -> (vfms z, x, y)
1912 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin)),
1913 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1914 Requires<[HasVFP4,HasDPVFP]>;
1915 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin)),
1916 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1917 Requires<[HasVFP4]>;
1918 // (fma x, (fneg y), z) -> (vfms z, x, y)
1919 def : Pat<(f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin)),
1920 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1921 Requires<[HasVFP4,HasDPVFP]>;
1922 def : Pat<(f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin)),
1923 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1924 Requires<[HasVFP4]>;
1926 def VFNMAD : ADbI<0b11101, 0b01, 1, 0,
1927 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1928 IIC_fpFMAC64, "vfnma", ".f64\t$Dd, $Dn, $Dm",
1929 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1930 (f64 DPR:$Ddin)))]>,
1931 RegConstraint<"$Ddin = $Dd">,
1932 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1934 def VFNMAS : ASbI<0b11101, 0b01, 1, 0,
1935 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1936 IIC_fpFMAC32, "vfnma", ".f32\t$Sd, $Sn, $Sm",
1937 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1939 RegConstraint<"$Sdin = $Sd">,
1940 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1941 // Some single precision VFP instructions may be executed on both NEON and
1945 def VFNMAH : AHbI<0b11101, 0b01, 1, 0,
1946 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1947 IIC_fpFMAC16, "vfnma", ".f16\t$Sd, $Sn, $Sm",
1949 RegConstraint<"$Sdin = $Sd">,
1950 Requires<[HasFullFP16,UseFusedMAC]>;
1952 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
1953 (VFNMAD DPR:$dstin, DPR:$a, DPR:$b)>,
1954 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1955 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
1956 (VFNMAS SPR:$dstin, SPR:$a, SPR:$b)>,
1957 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1959 // Match @llvm.fma.* intrinsics
1960 // (fneg (fma x, y, z)) -> (vfnma z, x, y)
1961 def : Pat<(fneg (fma (f64 DPR:$Dn), (f64 DPR:$Dm), (f64 DPR:$Ddin))),
1962 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1963 Requires<[HasVFP4,HasDPVFP]>;
1964 def : Pat<(fneg (fma (f32 SPR:$Sn), (f32 SPR:$Sm), (f32 SPR:$Sdin))),
1965 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1966 Requires<[HasVFP4]>;
1967 // (fma (fneg x), y, (fneg z)) -> (vfnma z, x, y)
1968 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, (fneg DPR:$Ddin))),
1969 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1970 Requires<[HasVFP4,HasDPVFP]>;
1971 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, (fneg SPR:$Sdin))),
1972 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1973 Requires<[HasVFP4]>;
1975 def VFNMSD : ADbI<0b11101, 0b01, 0, 0,
1976 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1977 IIC_fpFMAC64, "vfnms", ".f64\t$Dd, $Dn, $Dm",
1978 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1979 (f64 DPR:$Ddin)))]>,
1980 RegConstraint<"$Ddin = $Dd">,
1981 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1983 def VFNMSS : ASbI<0b11101, 0b01, 0, 0,
1984 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1985 IIC_fpFMAC32, "vfnms", ".f32\t$Sd, $Sn, $Sm",
1986 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
1987 RegConstraint<"$Sdin = $Sd">,
1988 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1989 // Some single precision VFP instructions may be executed on both NEON and
1993 def VFNMSH : AHbI<0b11101, 0b01, 0, 0,
1994 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1995 IIC_fpFMAC16, "vfnms", ".f16\t$Sd, $Sn, $Sm",
1997 RegConstraint<"$Sdin = $Sd">,
1998 Requires<[HasFullFP16,UseFusedMAC]>;
2000 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
2001 (VFNMSD DPR:$dstin, DPR:$a, DPR:$b)>,
2002 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
2003 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
2004 (VFNMSS SPR:$dstin, SPR:$a, SPR:$b)>,
2005 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
2007 // Match @llvm.fma.* intrinsics
2009 // (fma x, y, (fneg z)) -> (vfnms z, x, y))
2010 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, (fneg DPR:$Ddin))),
2011 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2012 Requires<[HasVFP4,HasDPVFP]>;
2013 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, (fneg SPR:$Sdin))),
2014 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2015 Requires<[HasVFP4]>;
2016 // (fneg (fma (fneg x), y, z)) -> (vfnms z, x, y)
2017 def : Pat<(fneg (f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin))),
2018 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2019 Requires<[HasVFP4,HasDPVFP]>;
2020 def : Pat<(fneg (f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin))),
2021 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2022 Requires<[HasVFP4]>;
2023 // (fneg (fma x, (fneg y), z) -> (vfnms z, x, y)
2024 def : Pat<(fneg (f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin))),
2025 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2026 Requires<[HasVFP4,HasDPVFP]>;
2027 def : Pat<(fneg (f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin))),
2028 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2029 Requires<[HasVFP4]>;
2031 //===----------------------------------------------------------------------===//
2032 // FP Conditional moves.
2035 let hasSideEffects = 0 in {
2036 def VMOVDcc : PseudoInst<(outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm, cmovpred:$p),
2038 [(set (f64 DPR:$Dd),
2039 (ARMcmov DPR:$Dn, DPR:$Dm, cmovpred:$p))]>,
2040 RegConstraint<"$Dn = $Dd">, Requires<[HasVFP2,HasDPVFP]>;
2042 def VMOVScc : PseudoInst<(outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm, cmovpred:$p),
2044 [(set (f32 SPR:$Sd),
2045 (ARMcmov SPR:$Sn, SPR:$Sm, cmovpred:$p))]>,
2046 RegConstraint<"$Sn = $Sd">, Requires<[HasVFP2]>;
2049 //===----------------------------------------------------------------------===//
2050 // Move from VFP System Register to ARM core register.
2053 class MovFromVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
2055 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
2057 // Instruction operand.
2060 let Inst{27-20} = 0b11101111;
2061 let Inst{19-16} = opc19_16;
2062 let Inst{15-12} = Rt;
2063 let Inst{11-8} = 0b1010;
2065 let Inst{6-5} = 0b00;
2067 let Inst{3-0} = 0b0000;
2070 // APSR is the application level alias of CPSR. This FPSCR N, Z, C, V flags
2072 let Defs = [CPSR], Uses = [FPSCR_NZCV], Rt = 0b1111 /* apsr_nzcv */ in
2073 def FMSTAT : MovFromVFP<0b0001 /* fpscr */, (outs), (ins),
2074 "vmrs", "\tAPSR_nzcv, fpscr", [(arm_fmstat)]>;
2076 // Application level FPSCR -> GPR
2077 let hasSideEffects = 1, Uses = [FPSCR] in
2078 def VMRS : MovFromVFP<0b0001 /* fpscr */, (outs GPR:$Rt), (ins),
2079 "vmrs", "\t$Rt, fpscr",
2080 [(set GPR:$Rt, (int_arm_get_fpscr))]>;
2082 // System level FPEXC, FPSID -> GPR
2083 let Uses = [FPSCR] in {
2084 def VMRS_FPEXC : MovFromVFP<0b1000 /* fpexc */, (outs GPR:$Rt), (ins),
2085 "vmrs", "\t$Rt, fpexc", []>;
2086 def VMRS_FPSID : MovFromVFP<0b0000 /* fpsid */, (outs GPR:$Rt), (ins),
2087 "vmrs", "\t$Rt, fpsid", []>;
2088 def VMRS_MVFR0 : MovFromVFP<0b0111 /* mvfr0 */, (outs GPR:$Rt), (ins),
2089 "vmrs", "\t$Rt, mvfr0", []>;
2090 def VMRS_MVFR1 : MovFromVFP<0b0110 /* mvfr1 */, (outs GPR:$Rt), (ins),
2091 "vmrs", "\t$Rt, mvfr1", []>;
2092 def VMRS_MVFR2 : MovFromVFP<0b0101 /* mvfr2 */, (outs GPR:$Rt), (ins),
2093 "vmrs", "\t$Rt, mvfr2", []>, Requires<[HasFPARMv8]>;
2094 def VMRS_FPINST : MovFromVFP<0b1001 /* fpinst */, (outs GPR:$Rt), (ins),
2095 "vmrs", "\t$Rt, fpinst", []>;
2096 def VMRS_FPINST2 : MovFromVFP<0b1010 /* fpinst2 */, (outs GPR:$Rt), (ins),
2097 "vmrs", "\t$Rt, fpinst2", []>;
2100 //===----------------------------------------------------------------------===//
2101 // Move from ARM core register to VFP System Register.
2104 class MovToVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
2106 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
2108 // Instruction operand.
2111 // Encode instruction operand.
2112 let Inst{15-12} = src;
2114 let Inst{27-20} = 0b11101110;
2115 let Inst{19-16} = opc19_16;
2116 let Inst{11-8} = 0b1010;
2121 let Defs = [FPSCR] in {
2122 // Application level GPR -> FPSCR
2123 def VMSR : MovToVFP<0b0001 /* fpscr */, (outs), (ins GPR:$src),
2124 "vmsr", "\tfpscr, $src", [(int_arm_set_fpscr GPR:$src)]>;
2125 // System level GPR -> FPEXC
2126 def VMSR_FPEXC : MovToVFP<0b1000 /* fpexc */, (outs), (ins GPR:$src),
2127 "vmsr", "\tfpexc, $src", []>;
2128 // System level GPR -> FPSID
2129 def VMSR_FPSID : MovToVFP<0b0000 /* fpsid */, (outs), (ins GPR:$src),
2130 "vmsr", "\tfpsid, $src", []>;
2132 def VMSR_FPINST : MovToVFP<0b1001 /* fpinst */, (outs), (ins GPR:$src),
2133 "vmsr", "\tfpinst, $src", []>;
2134 def VMSR_FPINST2 : MovToVFP<0b1010 /* fpinst2 */, (outs), (ins GPR:$src),
2135 "vmsr", "\tfpinst2, $src", []>;
2138 //===----------------------------------------------------------------------===//
2142 // Materialize FP immediates. VFP3 only.
2143 let isReMaterializable = 1 in {
2144 def FCONSTD : VFPAI<(outs DPR:$Dd), (ins vfp_f64imm:$imm),
2145 VFPMiscFrm, IIC_fpUNA64,
2146 "vmov", ".f64\t$Dd, $imm",
2147 [(set DPR:$Dd, vfp_f64imm:$imm)]>,
2148 Requires<[HasVFP3,HasDPVFP]> {
2152 let Inst{27-23} = 0b11101;
2153 let Inst{22} = Dd{4};
2154 let Inst{21-20} = 0b11;
2155 let Inst{19-16} = imm{7-4};
2156 let Inst{15-12} = Dd{3-0};
2157 let Inst{11-9} = 0b101;
2158 let Inst{8} = 1; // Double precision.
2159 let Inst{7-4} = 0b0000;
2160 let Inst{3-0} = imm{3-0};
2163 def FCONSTS : VFPAI<(outs SPR:$Sd), (ins vfp_f32imm:$imm),
2164 VFPMiscFrm, IIC_fpUNA32,
2165 "vmov", ".f32\t$Sd, $imm",
2166 [(set SPR:$Sd, vfp_f32imm:$imm)]>, Requires<[HasVFP3]> {
2170 let Inst{27-23} = 0b11101;
2171 let Inst{22} = Sd{0};
2172 let Inst{21-20} = 0b11;
2173 let Inst{19-16} = imm{7-4};
2174 let Inst{15-12} = Sd{4-1};
2175 let Inst{11-9} = 0b101;
2176 let Inst{8} = 0; // Single precision.
2177 let Inst{7-4} = 0b0000;
2178 let Inst{3-0} = imm{3-0};
2181 def FCONSTH : VFPAI<(outs SPR:$Sd), (ins vfp_f16imm:$imm),
2182 VFPMiscFrm, IIC_fpUNA16,
2183 "vmov", ".f16\t$Sd, $imm",
2184 []>, Requires<[HasFullFP16]> {
2188 let Inst{27-23} = 0b11101;
2189 let Inst{22} = Sd{0};
2190 let Inst{21-20} = 0b11;
2191 let Inst{19-16} = imm{7-4};
2192 let Inst{15-12} = Sd{4-1};
2193 let Inst{11-8} = 0b1001; // Half precision
2194 let Inst{7-4} = 0b0000;
2195 let Inst{3-0} = imm{3-0};
2199 //===----------------------------------------------------------------------===//
2200 // Assembler aliases.
2202 // A few mnemonic aliases for pre-unifixed syntax. We don't guarantee to
2203 // support them all, but supporting at least some of the basics is
2204 // good to be friendly.
2205 def : VFP2MnemonicAlias<"flds", "vldr">;
2206 def : VFP2MnemonicAlias<"fldd", "vldr">;
2207 def : VFP2MnemonicAlias<"fmrs", "vmov">;
2208 def : VFP2MnemonicAlias<"fmsr", "vmov">;
2209 def : VFP2MnemonicAlias<"fsqrts", "vsqrt">;
2210 def : VFP2MnemonicAlias<"fsqrtd", "vsqrt">;
2211 def : VFP2MnemonicAlias<"fadds", "vadd.f32">;
2212 def : VFP2MnemonicAlias<"faddd", "vadd.f64">;
2213 def : VFP2MnemonicAlias<"fmrdd", "vmov">;
2214 def : VFP2MnemonicAlias<"fmrds", "vmov">;
2215 def : VFP2MnemonicAlias<"fmrrd", "vmov">;
2216 def : VFP2MnemonicAlias<"fmdrr", "vmov">;
2217 def : VFP2MnemonicAlias<"fmuls", "vmul.f32">;
2218 def : VFP2MnemonicAlias<"fmuld", "vmul.f64">;
2219 def : VFP2MnemonicAlias<"fnegs", "vneg.f32">;
2220 def : VFP2MnemonicAlias<"fnegd", "vneg.f64">;
2221 def : VFP2MnemonicAlias<"ftosizd", "vcvt.s32.f64">;
2222 def : VFP2MnemonicAlias<"ftosid", "vcvtr.s32.f64">;
2223 def : VFP2MnemonicAlias<"ftosizs", "vcvt.s32.f32">;
2224 def : VFP2MnemonicAlias<"ftosis", "vcvtr.s32.f32">;
2225 def : VFP2MnemonicAlias<"ftouizd", "vcvt.u32.f64">;
2226 def : VFP2MnemonicAlias<"ftouid", "vcvtr.u32.f64">;
2227 def : VFP2MnemonicAlias<"ftouizs", "vcvt.u32.f32">;
2228 def : VFP2MnemonicAlias<"ftouis", "vcvtr.u32.f32">;
2229 def : VFP2MnemonicAlias<"fsitod", "vcvt.f64.s32">;
2230 def : VFP2MnemonicAlias<"fsitos", "vcvt.f32.s32">;
2231 def : VFP2MnemonicAlias<"fuitod", "vcvt.f64.u32">;
2232 def : VFP2MnemonicAlias<"fuitos", "vcvt.f32.u32">;
2233 def : VFP2MnemonicAlias<"fsts", "vstr">;
2234 def : VFP2MnemonicAlias<"fstd", "vstr">;
2235 def : VFP2MnemonicAlias<"fmacd", "vmla.f64">;
2236 def : VFP2MnemonicAlias<"fmacs", "vmla.f32">;
2237 def : VFP2MnemonicAlias<"fcpys", "vmov.f32">;
2238 def : VFP2MnemonicAlias<"fcpyd", "vmov.f64">;
2239 def : VFP2MnemonicAlias<"fcmps", "vcmp.f32">;
2240 def : VFP2MnemonicAlias<"fcmpd", "vcmp.f64">;
2241 def : VFP2MnemonicAlias<"fdivs", "vdiv.f32">;
2242 def : VFP2MnemonicAlias<"fdivd", "vdiv.f64">;
2243 def : VFP2MnemonicAlias<"fmrx", "vmrs">;
2244 def : VFP2MnemonicAlias<"fmxr", "vmsr">;
2246 // Be friendly and accept the old form of zero-compare
2247 def : VFP2DPInstAlias<"fcmpzd${p} $val", (VCMPZD DPR:$val, pred:$p)>;
2248 def : VFP2InstAlias<"fcmpzs${p} $val", (VCMPZS SPR:$val, pred:$p)>;
2251 def : VFP2InstAlias<"fmstat${p}", (FMSTAT pred:$p)>;
2252 def : VFP2InstAlias<"fadds${p} $Sd, $Sn, $Sm",
2253 (VADDS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
2254 def : VFP2DPInstAlias<"faddd${p} $Dd, $Dn, $Dm",
2255 (VADDD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
2256 def : VFP2InstAlias<"fsubs${p} $Sd, $Sn, $Sm",
2257 (VSUBS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
2258 def : VFP2DPInstAlias<"fsubd${p} $Dd, $Dn, $Dm",
2259 (VSUBD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
2261 // No need for the size suffix on VSQRT. It's implied by the register classes.
2262 def : VFP2InstAlias<"vsqrt${p} $Sd, $Sm", (VSQRTS SPR:$Sd, SPR:$Sm, pred:$p)>;
2263 def : VFP2DPInstAlias<"vsqrt${p} $Dd, $Dm", (VSQRTD DPR:$Dd, DPR:$Dm, pred:$p)>;
2265 // VLDR/VSTR accept an optional type suffix.
2266 def : VFP2InstAlias<"vldr${p}.32 $Sd, $addr",
2267 (VLDRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
2268 def : VFP2InstAlias<"vstr${p}.32 $Sd, $addr",
2269 (VSTRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
2270 def : VFP2InstAlias<"vldr${p}.64 $Dd, $addr",
2271 (VLDRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
2272 def : VFP2InstAlias<"vstr${p}.64 $Dd, $addr",
2273 (VSTRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
2275 // VMOV can accept optional 32-bit or less data type suffix suffix.
2276 def : VFP2InstAlias<"vmov${p}.8 $Rt, $Sn",
2277 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
2278 def : VFP2InstAlias<"vmov${p}.16 $Rt, $Sn",
2279 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
2280 def : VFP2InstAlias<"vmov${p}.32 $Rt, $Sn",
2281 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
2282 def : VFP2InstAlias<"vmov${p}.8 $Sn, $Rt",
2283 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
2284 def : VFP2InstAlias<"vmov${p}.16 $Sn, $Rt",
2285 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
2286 def : VFP2InstAlias<"vmov${p}.32 $Sn, $Rt",
2287 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
2289 def : VFP2InstAlias<"vmov${p}.f64 $Rt, $Rt2, $Dn",
2290 (VMOVRRD GPR:$Rt, GPR:$Rt2, DPR:$Dn, pred:$p)>;
2291 def : VFP2InstAlias<"vmov${p}.f64 $Dn, $Rt, $Rt2",
2292 (VMOVDRR DPR:$Dn, GPR:$Rt, GPR:$Rt2, pred:$p)>;
2294 // VMOVS doesn't need the .f32 to disambiguate from the NEON encoding the way
2296 def : VFP2InstAlias<"vmov${p} $Sd, $Sm",
2297 (VMOVS SPR:$Sd, SPR:$Sm, pred:$p)>;
2299 // FCONSTD/FCONSTS alias for vmov.f64/vmov.f32
2300 // These aliases provide added functionality over vmov.f instructions by
2301 // allowing users to write assembly containing encoded floating point constants
2302 // (e.g. #0x70 vs #1.0). Without these alises there is no way for the
2303 // assembler to accept encoded fp constants (but the equivalent fp-literal is
2304 // accepted directly by vmovf).
2305 def : VFP3InstAlias<"fconstd${p} $Dd, $val",
2306 (FCONSTD DPR:$Dd, vfp_f64imm:$val, pred:$p)>;
2307 def : VFP3InstAlias<"fconsts${p} $Sd, $val",
2308 (FCONSTS SPR:$Sd, vfp_f32imm:$val, pred:$p)>;