1 //===-- ARMInstrVFP.td - VFP 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 VFP instruction set.
11 //===----------------------------------------------------------------------===//
13 def SDT_CMPFP0 : SDTypeProfile<0, 2, [SDTCisFP<0>, SDTCisVT<1, i32>]>;
14 def SDT_VMOVDRR : SDTypeProfile<1, 2, [SDTCisVT<0, f64>, SDTCisVT<1, i32>,
16 def SDT_VMOVRRD : SDTypeProfile<2, 1, [SDTCisVT<0, i32>, SDTCisSameAs<0, 1>,
19 def SDT_VMOVSR : SDTypeProfile<1, 1, [SDTCisVT<0, f32>, SDTCisVT<1, i32>]>;
21 def arm_fmstat : SDNode<"ARMISD::FMSTAT", SDTNone, [SDNPInGlue, SDNPOutGlue]>;
22 def arm_cmpfp : SDNode<"ARMISD::CMPFP", SDT_ARMFCmp, [SDNPOutGlue]>;
23 def arm_cmpfp0 : SDNode<"ARMISD::CMPFPw0", SDT_CMPFP0, [SDNPOutGlue]>;
24 def arm_fmdrr : SDNode<"ARMISD::VMOVDRR", SDT_VMOVDRR>;
25 def arm_fmrrd : SDNode<"ARMISD::VMOVRRD", SDT_VMOVRRD>;
26 def arm_vmovsr : SDNode<"ARMISD::VMOVSR", SDT_VMOVSR>;
28 def SDT_VMOVhr : SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisVT<1, i32>] >;
29 def SDT_VMOVrh : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisFP<1>] >;
30 def arm_vmovhr : SDNode<"ARMISD::VMOVhr", SDT_VMOVhr>;
31 def arm_vmovrh : SDNode<"ARMISD::VMOVrh", SDT_VMOVrh>;
33 //===----------------------------------------------------------------------===//
34 // Operand Definitions.
37 // 8-bit floating-point immediate encodings.
38 def FPImmOperand : AsmOperandClass {
40 let ParserMethod = "parseFPImm";
43 def vfp_f16imm : Operand<f16>,
44 PatLeaf<(f16 fpimm), [{
45 return ARM_AM::getFP16Imm(N->getValueAPF()) != -1;
46 }], SDNodeXForm<fpimm, [{
47 APFloat InVal = N->getValueAPF();
48 uint32_t enc = ARM_AM::getFP16Imm(InVal);
49 return CurDAG->getTargetConstant(enc, SDLoc(N), MVT::i32);
51 let PrintMethod = "printFPImmOperand";
52 let ParserMatchClass = FPImmOperand;
55 def vfp_f32imm_xform : SDNodeXForm<fpimm, [{
56 APFloat InVal = N->getValueAPF();
57 uint32_t enc = ARM_AM::getFP32Imm(InVal);
58 return CurDAG->getTargetConstant(enc, SDLoc(N), MVT::i32);
61 def gi_vfp_f32imm : GICustomOperandRenderer<"renderVFPF32Imm">,
62 GISDNodeXFormEquiv<vfp_f32imm_xform>;
64 def vfp_f32imm : Operand<f32>,
65 PatLeaf<(f32 fpimm), [{
66 return ARM_AM::getFP32Imm(N->getValueAPF()) != -1;
67 }], vfp_f32imm_xform> {
68 let PrintMethod = "printFPImmOperand";
69 let ParserMatchClass = FPImmOperand;
70 let GISelPredicateCode = [{
71 const auto &MO = MI.getOperand(1);
74 return ARM_AM::getFP32Imm(MO.getFPImm()->getValueAPF()) != -1;
78 def vfp_f64imm_xform : SDNodeXForm<fpimm, [{
79 APFloat InVal = N->getValueAPF();
80 uint32_t enc = ARM_AM::getFP64Imm(InVal);
81 return CurDAG->getTargetConstant(enc, SDLoc(N), MVT::i32);
84 def gi_vfp_f64imm : GICustomOperandRenderer<"renderVFPF64Imm">,
85 GISDNodeXFormEquiv<vfp_f64imm_xform>;
87 def vfp_f64imm : Operand<f64>,
88 PatLeaf<(f64 fpimm), [{
89 return ARM_AM::getFP64Imm(N->getValueAPF()) != -1;
90 }], vfp_f64imm_xform> {
91 let PrintMethod = "printFPImmOperand";
92 let ParserMatchClass = FPImmOperand;
93 let GISelPredicateCode = [{
94 const auto &MO = MI.getOperand(1);
97 return ARM_AM::getFP64Imm(MO.getFPImm()->getValueAPF()) != -1;
101 def alignedload16 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
102 return cast<LoadSDNode>(N)->getAlignment() >= 2;
105 def alignedload32 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
106 return cast<LoadSDNode>(N)->getAlignment() >= 4;
109 def alignedstore16 : PatFrag<(ops node:$val, node:$ptr),
110 (store node:$val, node:$ptr), [{
111 return cast<StoreSDNode>(N)->getAlignment() >= 2;
114 def alignedstore32 : PatFrag<(ops node:$val, node:$ptr),
115 (store node:$val, node:$ptr), [{
116 return cast<StoreSDNode>(N)->getAlignment() >= 4;
119 // The VCVT to/from fixed-point instructions encode the 'fbits' operand
120 // (the number of fixed bits) differently than it appears in the assembly
121 // source. It's encoded as "Size - fbits" where Size is the size of the
122 // fixed-point representation (32 or 16) and fbits is the value appearing
123 // in the assembly source, an integer in [0,16] or (0,32], depending on size.
124 def fbits32_asm_operand : AsmOperandClass { let Name = "FBits32"; }
125 def fbits32 : Operand<i32> {
126 let PrintMethod = "printFBits32";
127 let ParserMatchClass = fbits32_asm_operand;
130 def fbits16_asm_operand : AsmOperandClass { let Name = "FBits16"; }
131 def fbits16 : Operand<i32> {
132 let PrintMethod = "printFBits16";
133 let ParserMatchClass = fbits16_asm_operand;
136 //===----------------------------------------------------------------------===//
137 // Load / store Instructions.
140 let canFoldAsLoad = 1, isReMaterializable = 1 in {
142 def VLDRD : ADI5<0b1101, 0b01, (outs DPR:$Dd), (ins addrmode5:$addr),
143 IIC_fpLoad64, "vldr", "\t$Dd, $addr",
144 [(set DPR:$Dd, (f64 (alignedload32 addrmode5:$addr)))]>;
146 def VLDRS : ASI5<0b1101, 0b01, (outs SPR:$Sd), (ins addrmode5:$addr),
147 IIC_fpLoad32, "vldr", "\t$Sd, $addr",
148 [(set SPR:$Sd, (alignedload32 addrmode5:$addr))]> {
149 // Some single precision VFP instructions may be executed on both NEON and VFP
151 let D = VFPNeonDomain;
154 let isUnpredicable = 1 in
155 def VLDRH : AHI5<0b1101, 0b01, (outs HPR:$Sd), (ins addrmode5fp16:$addr),
156 IIC_fpLoad16, "vldr", ".16\t$Sd, $addr",
157 [(set HPR:$Sd, (alignedload16 addrmode5fp16:$addr))]>,
158 Requires<[HasFullFP16]>;
160 } // End of 'let canFoldAsLoad = 1, isReMaterializable = 1 in'
162 def VSTRD : ADI5<0b1101, 0b00, (outs), (ins DPR:$Dd, addrmode5:$addr),
163 IIC_fpStore64, "vstr", "\t$Dd, $addr",
164 [(alignedstore32 (f64 DPR:$Dd), addrmode5:$addr)]>;
166 def VSTRS : ASI5<0b1101, 0b00, (outs), (ins SPR:$Sd, addrmode5:$addr),
167 IIC_fpStore32, "vstr", "\t$Sd, $addr",
168 [(alignedstore32 SPR:$Sd, addrmode5:$addr)]> {
169 // Some single precision VFP instructions may be executed on both NEON and VFP
171 let D = VFPNeonDomain;
174 let isUnpredicable = 1 in
175 def VSTRH : AHI5<0b1101, 0b00, (outs), (ins HPR:$Sd, addrmode5fp16:$addr),
176 IIC_fpStore16, "vstr", ".16\t$Sd, $addr",
177 [(alignedstore16 HPR:$Sd, addrmode5fp16:$addr)]>,
178 Requires<[HasFullFP16]>;
180 //===----------------------------------------------------------------------===//
181 // Load / store multiple Instructions.
184 multiclass vfp_ldst_mult<string asm, bit L_bit,
185 InstrItinClass itin, InstrItinClass itin_upd> {
188 AXDI4<(outs), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
190 !strconcat(asm, "ia${p}\t$Rn, $regs"), "", []> {
191 let Inst{24-23} = 0b01; // Increment After
192 let Inst{21} = 0; // No writeback
193 let Inst{20} = L_bit;
196 AXDI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs,
198 IndexModeUpd, itin_upd,
199 !strconcat(asm, "ia${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
200 let Inst{24-23} = 0b01; // Increment After
201 let Inst{21} = 1; // Writeback
202 let Inst{20} = L_bit;
205 AXDI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs,
207 IndexModeUpd, itin_upd,
208 !strconcat(asm, "db${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
209 let Inst{24-23} = 0b10; // Decrement Before
210 let Inst{21} = 1; // Writeback
211 let Inst{20} = L_bit;
216 AXSI4<(outs), (ins GPR:$Rn, pred:$p, spr_reglist:$regs, variable_ops),
218 !strconcat(asm, "ia${p}\t$Rn, $regs"), "", []> {
219 let Inst{24-23} = 0b01; // Increment After
220 let Inst{21} = 0; // No writeback
221 let Inst{20} = L_bit;
223 // Some single precision VFP instructions may be executed on both NEON and
225 let D = VFPNeonDomain;
228 AXSI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, spr_reglist:$regs,
230 IndexModeUpd, itin_upd,
231 !strconcat(asm, "ia${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
232 let Inst{24-23} = 0b01; // Increment After
233 let Inst{21} = 1; // Writeback
234 let Inst{20} = L_bit;
236 // Some single precision VFP instructions may be executed on both NEON and
238 let D = VFPNeonDomain;
241 AXSI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, spr_reglist:$regs,
243 IndexModeUpd, itin_upd,
244 !strconcat(asm, "db${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
245 let Inst{24-23} = 0b10; // Decrement Before
246 let Inst{21} = 1; // Writeback
247 let Inst{20} = L_bit;
249 // Some single precision VFP instructions may be executed on both NEON and
251 let D = VFPNeonDomain;
255 let hasSideEffects = 0 in {
257 let mayLoad = 1, hasExtraDefRegAllocReq = 1 in
258 defm VLDM : vfp_ldst_mult<"vldm", 1, IIC_fpLoad_m, IIC_fpLoad_mu>;
260 let mayStore = 1, hasExtraSrcRegAllocReq = 1 in
261 defm VSTM : vfp_ldst_mult<"vstm", 0, IIC_fpStore_m, IIC_fpStore_mu>;
265 def : MnemonicAlias<"vldm", "vldmia">;
266 def : MnemonicAlias<"vstm", "vstmia">;
269 //===----------------------------------------------------------------------===//
270 // Lazy load / store multiple Instructions
273 def VLLDM : AXSI4<(outs), (ins GPRnopc:$Rn, pred:$p), IndexModeNone,
274 IIC_fpLoad_m, "vlldm${p}\t$Rn", "", []>,
275 Requires<[HasV8MMainline, Has8MSecExt]> {
276 let Inst{24-23} = 0b00;
286 def VLSTM : AXSI4<(outs), (ins GPRnopc:$Rn, pred:$p), IndexModeNone,
287 IIC_fpStore_m, "vlstm${p}\t$Rn", "", []>,
288 Requires<[HasV8MMainline, Has8MSecExt]> {
289 let Inst{24-23} = 0b00;
298 def : InstAlias<"vpush${p} $r", (VSTMDDB_UPD SP, pred:$p, dpr_reglist:$r), 0>,
300 def : InstAlias<"vpush${p} $r", (VSTMSDB_UPD SP, pred:$p, spr_reglist:$r), 0>,
302 def : InstAlias<"vpop${p} $r", (VLDMDIA_UPD SP, pred:$p, dpr_reglist:$r), 0>,
304 def : InstAlias<"vpop${p} $r", (VLDMSIA_UPD SP, pred:$p, spr_reglist:$r), 0>,
306 defm : VFPDTAnyInstAlias<"vpush${p}", "$r",
307 (VSTMSDB_UPD SP, pred:$p, spr_reglist:$r)>;
308 defm : VFPDTAnyInstAlias<"vpush${p}", "$r",
309 (VSTMDDB_UPD SP, pred:$p, dpr_reglist:$r)>;
310 defm : VFPDTAnyInstAlias<"vpop${p}", "$r",
311 (VLDMSIA_UPD SP, pred:$p, spr_reglist:$r)>;
312 defm : VFPDTAnyInstAlias<"vpop${p}", "$r",
313 (VLDMDIA_UPD SP, pred:$p, dpr_reglist:$r)>;
315 // FLDMX, FSTMX - Load and store multiple unknown precision registers for
317 // These instruction are deprecated so we don't want them to get selected.
318 // However, there is no UAL syntax for them, so we keep them around for
319 // (dis)assembly only.
320 multiclass vfp_ldstx_mult<string asm, bit L_bit> {
323 AXXI4<(outs), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
324 IndexModeNone, !strconcat(asm, "iax${p}\t$Rn, $regs"), "", []> {
325 let Inst{24-23} = 0b01; // Increment After
326 let Inst{21} = 0; // No writeback
327 let Inst{20} = L_bit;
330 AXXI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
331 IndexModeUpd, !strconcat(asm, "iax${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
332 let Inst{24-23} = 0b01; // Increment After
333 let Inst{21} = 1; // Writeback
334 let Inst{20} = L_bit;
337 AXXI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
338 IndexModeUpd, !strconcat(asm, "dbx${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
339 let Inst{24-23} = 0b10; // Decrement Before
340 let Inst{21} = 1; // Writeback
341 let Inst{20} = L_bit;
345 defm FLDM : vfp_ldstx_mult<"fldm", 1>;
346 defm FSTM : vfp_ldstx_mult<"fstm", 0>;
348 def : VFP2MnemonicAlias<"fldmeax", "fldmdbx">;
349 def : VFP2MnemonicAlias<"fldmfdx", "fldmiax">;
351 def : VFP2MnemonicAlias<"fstmeax", "fstmiax">;
352 def : VFP2MnemonicAlias<"fstmfdx", "fstmdbx">;
354 //===----------------------------------------------------------------------===//
355 // FP Binary Operations.
358 let TwoOperandAliasConstraint = "$Dn = $Dd" in
359 def VADDD : ADbI<0b11100, 0b11, 0, 0,
360 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
361 IIC_fpALU64, "vadd", ".f64\t$Dd, $Dn, $Dm",
362 [(set DPR:$Dd, (fadd DPR:$Dn, (f64 DPR:$Dm)))]>,
363 Sched<[WriteFPALU64]>;
365 let TwoOperandAliasConstraint = "$Sn = $Sd" in
366 def VADDS : ASbIn<0b11100, 0b11, 0, 0,
367 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
368 IIC_fpALU32, "vadd", ".f32\t$Sd, $Sn, $Sm",
369 [(set SPR:$Sd, (fadd SPR:$Sn, SPR:$Sm))]>,
370 Sched<[WriteFPALU32]> {
371 // Some single precision VFP instructions may be executed on both NEON and
372 // VFP pipelines on A8.
373 let D = VFPNeonA8Domain;
376 let TwoOperandAliasConstraint = "$Sn = $Sd" in
377 def VADDH : AHbI<0b11100, 0b11, 0, 0,
378 (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
379 IIC_fpALU16, "vadd", ".f16\t$Sd, $Sn, $Sm",
380 [(set HPR:$Sd, (fadd HPR:$Sn, HPR:$Sm))]>,
381 Sched<[WriteFPALU32]>;
383 let TwoOperandAliasConstraint = "$Dn = $Dd" in
384 def VSUBD : ADbI<0b11100, 0b11, 1, 0,
385 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
386 IIC_fpALU64, "vsub", ".f64\t$Dd, $Dn, $Dm",
387 [(set DPR:$Dd, (fsub DPR:$Dn, (f64 DPR:$Dm)))]>,
388 Sched<[WriteFPALU64]>;
390 let TwoOperandAliasConstraint = "$Sn = $Sd" in
391 def VSUBS : ASbIn<0b11100, 0b11, 1, 0,
392 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
393 IIC_fpALU32, "vsub", ".f32\t$Sd, $Sn, $Sm",
394 [(set SPR:$Sd, (fsub SPR:$Sn, SPR:$Sm))]>,
395 Sched<[WriteFPALU32]>{
396 // Some single precision VFP instructions may be executed on both NEON and
397 // VFP pipelines on A8.
398 let D = VFPNeonA8Domain;
401 let TwoOperandAliasConstraint = "$Sn = $Sd" in
402 def VSUBH : AHbI<0b11100, 0b11, 1, 0,
403 (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
404 IIC_fpALU16, "vsub", ".f16\t$Sd, $Sn, $Sm",
405 [(set HPR:$Sd, (fsub HPR:$Sn, HPR:$Sm))]>,
406 Sched<[WriteFPALU32]>;
408 let TwoOperandAliasConstraint = "$Dn = $Dd" in
409 def VDIVD : ADbI<0b11101, 0b00, 0, 0,
410 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
411 IIC_fpDIV64, "vdiv", ".f64\t$Dd, $Dn, $Dm",
412 [(set DPR:$Dd, (fdiv DPR:$Dn, (f64 DPR:$Dm)))]>,
413 Sched<[WriteFPDIV64]>;
415 let TwoOperandAliasConstraint = "$Sn = $Sd" in
416 def VDIVS : ASbI<0b11101, 0b00, 0, 0,
417 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
418 IIC_fpDIV32, "vdiv", ".f32\t$Sd, $Sn, $Sm",
419 [(set SPR:$Sd, (fdiv SPR:$Sn, SPR:$Sm))]>,
420 Sched<[WriteFPDIV32]>;
422 let TwoOperandAliasConstraint = "$Sn = $Sd" in
423 def VDIVH : AHbI<0b11101, 0b00, 0, 0,
424 (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
425 IIC_fpDIV16, "vdiv", ".f16\t$Sd, $Sn, $Sm",
426 [(set HPR:$Sd, (fdiv HPR:$Sn, HPR:$Sm))]>,
427 Sched<[WriteFPDIV32]>;
429 let TwoOperandAliasConstraint = "$Dn = $Dd" in
430 def VMULD : ADbI<0b11100, 0b10, 0, 0,
431 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
432 IIC_fpMUL64, "vmul", ".f64\t$Dd, $Dn, $Dm",
433 [(set DPR:$Dd, (fmul DPR:$Dn, (f64 DPR:$Dm)))]>,
434 Sched<[WriteFPMUL64, ReadFPMUL, ReadFPMUL]>;
436 let TwoOperandAliasConstraint = "$Sn = $Sd" in
437 def VMULS : ASbIn<0b11100, 0b10, 0, 0,
438 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
439 IIC_fpMUL32, "vmul", ".f32\t$Sd, $Sn, $Sm",
440 [(set SPR:$Sd, (fmul SPR:$Sn, SPR:$Sm))]>,
441 Sched<[WriteFPMUL32, ReadFPMUL, ReadFPMUL]> {
442 // Some single precision VFP instructions may be executed on both NEON and
443 // VFP pipelines on A8.
444 let D = VFPNeonA8Domain;
447 let TwoOperandAliasConstraint = "$Sn = $Sd" in
448 def VMULH : AHbI<0b11100, 0b10, 0, 0,
449 (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
450 IIC_fpMUL16, "vmul", ".f16\t$Sd, $Sn, $Sm",
451 [(set HPR:$Sd, (fmul HPR:$Sn, HPR:$Sm))]>,
452 Sched<[WriteFPMUL32, ReadFPMUL, ReadFPMUL]>;
454 def VNMULD : ADbI<0b11100, 0b10, 1, 0,
455 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
456 IIC_fpMUL64, "vnmul", ".f64\t$Dd, $Dn, $Dm",
457 [(set DPR:$Dd, (fneg (fmul DPR:$Dn, (f64 DPR:$Dm))))]>,
458 Sched<[WriteFPMUL64, ReadFPMUL, ReadFPMUL]>;
460 def VNMULS : ASbI<0b11100, 0b10, 1, 0,
461 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
462 IIC_fpMUL32, "vnmul", ".f32\t$Sd, $Sn, $Sm",
463 [(set SPR:$Sd, (fneg (fmul SPR:$Sn, SPR:$Sm)))]>,
464 Sched<[WriteFPMUL32, ReadFPMUL, ReadFPMUL]> {
465 // Some single precision VFP instructions may be executed on both NEON and
466 // VFP pipelines on A8.
467 let D = VFPNeonA8Domain;
470 def VNMULH : AHbI<0b11100, 0b10, 1, 0,
471 (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
472 IIC_fpMUL16, "vnmul", ".f16\t$Sd, $Sn, $Sm",
473 [(set HPR:$Sd, (fneg (fmul HPR:$Sn, HPR:$Sm)))]>,
474 Sched<[WriteFPMUL32, ReadFPMUL, ReadFPMUL]>;
476 multiclass vsel_inst<string op, bits<2> opc, int CC> {
477 let DecoderNamespace = "VFPV8", PostEncoderMethod = "",
478 Uses = [CPSR], AddedComplexity = 4, isUnpredicable = 1 in {
479 def H : AHbInp<0b11100, opc, 0,
480 (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
481 NoItinerary, !strconcat("vsel", op, ".f16\t$Sd, $Sn, $Sm"),
482 [(set HPR:$Sd, (ARMcmov HPR:$Sm, HPR:$Sn, CC))]>,
483 Requires<[HasFullFP16]>;
485 def S : ASbInp<0b11100, opc, 0,
486 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
487 NoItinerary, !strconcat("vsel", op, ".f32\t$Sd, $Sn, $Sm"),
488 [(set SPR:$Sd, (ARMcmov SPR:$Sm, SPR:$Sn, CC))]>,
489 Requires<[HasFPARMv8]>;
491 def D : ADbInp<0b11100, opc, 0,
492 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
493 NoItinerary, !strconcat("vsel", op, ".f64\t$Dd, $Dn, $Dm"),
494 [(set DPR:$Dd, (ARMcmov (f64 DPR:$Dm), (f64 DPR:$Dn), CC))]>,
495 Requires<[HasFPARMv8, HasDPVFP]>;
499 // The CC constants here match ARMCC::CondCodes.
500 defm VSELGT : vsel_inst<"gt", 0b11, 12>;
501 defm VSELGE : vsel_inst<"ge", 0b10, 10>;
502 defm VSELEQ : vsel_inst<"eq", 0b00, 0>;
503 defm VSELVS : vsel_inst<"vs", 0b01, 6>;
505 multiclass vmaxmin_inst<string op, bit opc, SDNode SD> {
506 let DecoderNamespace = "VFPV8", PostEncoderMethod = "",
507 isUnpredicable = 1 in {
508 def H : AHbInp<0b11101, 0b00, opc,
509 (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
510 NoItinerary, !strconcat(op, ".f16\t$Sd, $Sn, $Sm"),
511 [(set HPR:$Sd, (SD HPR:$Sn, HPR:$Sm))]>,
512 Requires<[HasFullFP16]>;
514 def S : ASbInp<0b11101, 0b00, opc,
515 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
516 NoItinerary, !strconcat(op, ".f32\t$Sd, $Sn, $Sm"),
517 [(set SPR:$Sd, (SD SPR:$Sn, SPR:$Sm))]>,
518 Requires<[HasFPARMv8]>;
520 def D : ADbInp<0b11101, 0b00, opc,
521 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
522 NoItinerary, !strconcat(op, ".f64\t$Dd, $Dn, $Dm"),
523 [(set DPR:$Dd, (f64 (SD (f64 DPR:$Dn), (f64 DPR:$Dm))))]>,
524 Requires<[HasFPARMv8, HasDPVFP]>;
528 defm VMAXNM : vmaxmin_inst<"vmaxnm", 0, fmaxnum>;
529 defm VMINNM : vmaxmin_inst<"vminnm", 1, fminnum>;
531 // Match reassociated forms only if not sign dependent rounding.
532 def : Pat<(fmul (fneg DPR:$a), (f64 DPR:$b)),
533 (VNMULD DPR:$a, DPR:$b)>,
534 Requires<[NoHonorSignDependentRounding,HasDPVFP]>;
535 def : Pat<(fmul (fneg SPR:$a), SPR:$b),
536 (VNMULS SPR:$a, SPR:$b)>, Requires<[NoHonorSignDependentRounding]>;
538 // These are encoded as unary instructions.
539 let Defs = [FPSCR_NZCV] in {
540 def VCMPED : ADuI<0b11101, 0b11, 0b0100, 0b11, 0,
541 (outs), (ins DPR:$Dd, DPR:$Dm),
542 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, $Dm",
543 [(arm_cmpfp DPR:$Dd, (f64 DPR:$Dm), (i32 1))]>;
545 def VCMPES : ASuI<0b11101, 0b11, 0b0100, 0b11, 0,
546 (outs), (ins SPR:$Sd, SPR:$Sm),
547 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, $Sm",
548 [(arm_cmpfp SPR:$Sd, SPR:$Sm, (i32 1))]> {
549 // Some single precision VFP instructions may be executed on both NEON and
550 // VFP pipelines on A8.
551 let D = VFPNeonA8Domain;
554 def VCMPEH : AHuI<0b11101, 0b11, 0b0100, 0b11, 0,
555 (outs), (ins HPR:$Sd, HPR:$Sm),
556 IIC_fpCMP16, "vcmpe", ".f16\t$Sd, $Sm",
557 [(arm_cmpfp HPR:$Sd, HPR:$Sm, (i32 1))]>;
559 def VCMPD : ADuI<0b11101, 0b11, 0b0100, 0b01, 0,
560 (outs), (ins DPR:$Dd, DPR:$Dm),
561 IIC_fpCMP64, "vcmp", ".f64\t$Dd, $Dm",
562 [(arm_cmpfp DPR:$Dd, (f64 DPR:$Dm), (i32 0))]>;
564 def VCMPS : ASuI<0b11101, 0b11, 0b0100, 0b01, 0,
565 (outs), (ins SPR:$Sd, SPR:$Sm),
566 IIC_fpCMP32, "vcmp", ".f32\t$Sd, $Sm",
567 [(arm_cmpfp SPR:$Sd, SPR:$Sm, (i32 0))]> {
568 // Some single precision VFP instructions may be executed on both NEON and
569 // VFP pipelines on A8.
570 let D = VFPNeonA8Domain;
573 def VCMPH : AHuI<0b11101, 0b11, 0b0100, 0b01, 0,
574 (outs), (ins HPR:$Sd, HPR:$Sm),
575 IIC_fpCMP16, "vcmp", ".f16\t$Sd, $Sm",
576 [(arm_cmpfp HPR:$Sd, HPR:$Sm, (i32 0))]>;
577 } // Defs = [FPSCR_NZCV]
579 //===----------------------------------------------------------------------===//
580 // FP Unary Operations.
583 def VABSD : ADuI<0b11101, 0b11, 0b0000, 0b11, 0,
584 (outs DPR:$Dd), (ins DPR:$Dm),
585 IIC_fpUNA64, "vabs", ".f64\t$Dd, $Dm",
586 [(set DPR:$Dd, (fabs (f64 DPR:$Dm)))]>;
588 def VABSS : ASuIn<0b11101, 0b11, 0b0000, 0b11, 0,
589 (outs SPR:$Sd), (ins SPR:$Sm),
590 IIC_fpUNA32, "vabs", ".f32\t$Sd, $Sm",
591 [(set SPR:$Sd, (fabs SPR:$Sm))]> {
592 // Some single precision VFP instructions may be executed on both NEON and
593 // VFP pipelines on A8.
594 let D = VFPNeonA8Domain;
597 def VABSH : AHuI<0b11101, 0b11, 0b0000, 0b11, 0,
598 (outs HPR:$Sd), (ins HPR:$Sm),
599 IIC_fpUNA16, "vabs", ".f16\t$Sd, $Sm",
600 [(set HPR:$Sd, (fabs (f16 HPR:$Sm)))]>;
602 let Defs = [FPSCR_NZCV] in {
603 def VCMPEZD : ADuI<0b11101, 0b11, 0b0101, 0b11, 0,
604 (outs), (ins DPR:$Dd),
605 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, #0",
606 [(arm_cmpfp0 (f64 DPR:$Dd), (i32 1))]> {
607 let Inst{3-0} = 0b0000;
611 def VCMPEZS : ASuI<0b11101, 0b11, 0b0101, 0b11, 0,
612 (outs), (ins SPR:$Sd),
613 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, #0",
614 [(arm_cmpfp0 SPR:$Sd, (i32 1))]> {
615 let Inst{3-0} = 0b0000;
618 // Some single precision VFP instructions may be executed on both NEON and
619 // VFP pipelines on A8.
620 let D = VFPNeonA8Domain;
623 def VCMPEZH : AHuI<0b11101, 0b11, 0b0101, 0b11, 0,
624 (outs), (ins HPR:$Sd),
625 IIC_fpCMP16, "vcmpe", ".f16\t$Sd, #0",
626 [(arm_cmpfp0 HPR:$Sd, (i32 1))]> {
627 let Inst{3-0} = 0b0000;
631 def VCMPZD : ADuI<0b11101, 0b11, 0b0101, 0b01, 0,
632 (outs), (ins DPR:$Dd),
633 IIC_fpCMP64, "vcmp", ".f64\t$Dd, #0",
634 [(arm_cmpfp0 (f64 DPR:$Dd), (i32 0))]> {
635 let Inst{3-0} = 0b0000;
639 def VCMPZS : ASuI<0b11101, 0b11, 0b0101, 0b01, 0,
640 (outs), (ins SPR:$Sd),
641 IIC_fpCMP32, "vcmp", ".f32\t$Sd, #0",
642 [(arm_cmpfp0 SPR:$Sd, (i32 0))]> {
643 let Inst{3-0} = 0b0000;
646 // Some single precision VFP instructions may be executed on both NEON and
647 // VFP pipelines on A8.
648 let D = VFPNeonA8Domain;
651 def VCMPZH : AHuI<0b11101, 0b11, 0b0101, 0b01, 0,
652 (outs), (ins HPR:$Sd),
653 IIC_fpCMP16, "vcmp", ".f16\t$Sd, #0",
654 [(arm_cmpfp0 HPR:$Sd, (i32 0))]> {
655 let Inst{3-0} = 0b0000;
658 } // Defs = [FPSCR_NZCV]
660 def VCVTDS : ASuI<0b11101, 0b11, 0b0111, 0b11, 0,
661 (outs DPR:$Dd), (ins SPR:$Sm),
662 IIC_fpCVTDS, "vcvt", ".f64.f32\t$Dd, $Sm",
663 [(set DPR:$Dd, (fpextend SPR:$Sm))]>,
664 Sched<[WriteFPCVT]> {
665 // Instruction operands.
669 // Encode instruction operands.
670 let Inst{3-0} = Sm{4-1};
672 let Inst{15-12} = Dd{3-0};
673 let Inst{22} = Dd{4};
675 let Predicates = [HasVFP2, HasDPVFP];
678 // Special case encoding: bits 11-8 is 0b1011.
679 def VCVTSD : VFPAI<(outs SPR:$Sd), (ins DPR:$Dm), VFPUnaryFrm,
680 IIC_fpCVTSD, "vcvt", ".f32.f64\t$Sd, $Dm",
681 [(set SPR:$Sd, (fpround DPR:$Dm))]>,
682 Sched<[WriteFPCVT]> {
683 // Instruction operands.
687 // Encode instruction operands.
688 let Inst{3-0} = Dm{3-0};
690 let Inst{15-12} = Sd{4-1};
691 let Inst{22} = Sd{0};
693 let Inst{27-23} = 0b11101;
694 let Inst{21-16} = 0b110111;
695 let Inst{11-8} = 0b1011;
696 let Inst{7-6} = 0b11;
699 let Predicates = [HasVFP2, HasDPVFP];
702 // Between half, single and double-precision.
703 def VCVTBHS: ASuI<0b11101, 0b11, 0b0010, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
704 /* FIXME */ IIC_fpCVTSH, "vcvtb", ".f32.f16\t$Sd, $Sm",
705 [/* Intentionally left blank, see patterns below */]>,
709 def : FullFP16Pat<(f32 (fpextend HPR:$Sm)),
710 (VCVTBHS (COPY_TO_REGCLASS HPR:$Sm, SPR))>;
711 def : FP16Pat<(f16_to_fp GPR:$a),
712 (VCVTBHS (COPY_TO_REGCLASS GPR:$a, SPR))>;
714 def VCVTBSH: ASuI<0b11101, 0b11, 0b0011, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
715 /* FIXME */ IIC_fpCVTHS, "vcvtb", ".f16.f32\t$Sd, $Sm",
716 [/* Intentionally left blank, see patterns below */]>,
720 def : FullFP16Pat<(f16 (fpround SPR:$Sm)),
721 (COPY_TO_REGCLASS (VCVTBSH SPR:$Sm), HPR)>;
722 def : FP16Pat<(fp_to_f16 SPR:$a),
723 (i32 (COPY_TO_REGCLASS (VCVTBSH SPR:$a), GPR))>;
725 def VCVTTHS: ASuI<0b11101, 0b11, 0b0010, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
726 /* FIXME */ IIC_fpCVTSH, "vcvtt", ".f32.f16\t$Sd, $Sm",
727 [/* For disassembly only; pattern left blank */]>,
731 def VCVTTSH: ASuI<0b11101, 0b11, 0b0011, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
732 /* FIXME */ IIC_fpCVTHS, "vcvtt", ".f16.f32\t$Sd, $Sm",
733 [/* For disassembly only; pattern left blank */]>,
737 def VCVTBHD : ADuI<0b11101, 0b11, 0b0010, 0b01, 0,
738 (outs DPR:$Dd), (ins SPR:$Sm),
739 NoItinerary, "vcvtb", ".f64.f16\t$Dd, $Sm",
740 [/* Intentionally left blank, see patterns below */]>,
741 Requires<[HasFPARMv8, HasDPVFP]>,
742 Sched<[WriteFPCVT]> {
743 // Instruction operands.
746 // Encode instruction operands.
747 let Inst{3-0} = Sm{4-1};
751 def : FullFP16Pat<(f64 (fpextend HPR:$Sm)),
752 (VCVTBHD (COPY_TO_REGCLASS HPR:$Sm, SPR))>,
753 Requires<[HasFPARMv8, HasDPVFP]>;
754 def : FP16Pat<(f64 (f16_to_fp GPR:$a)),
755 (VCVTBHD (COPY_TO_REGCLASS GPR:$a, SPR))>,
756 Requires<[HasFPARMv8, HasDPVFP]>;
758 def VCVTBDH : ADuI<0b11101, 0b11, 0b0011, 0b01, 0,
759 (outs SPR:$Sd), (ins DPR:$Dm),
760 NoItinerary, "vcvtb", ".f16.f64\t$Sd, $Dm",
761 [/* Intentionally left blank, see patterns below */]>,
762 Requires<[HasFPARMv8, HasDPVFP]> {
763 // Instruction operands.
767 // Encode instruction operands.
768 let Inst{3-0} = Dm{3-0};
770 let Inst{15-12} = Sd{4-1};
771 let Inst{22} = Sd{0};
774 def : FullFP16Pat<(f16 (fpround DPR:$Dm)),
775 (COPY_TO_REGCLASS (VCVTBDH DPR:$Dm), HPR)>,
776 Requires<[HasFPARMv8, HasDPVFP]>;
777 def : FP16Pat<(fp_to_f16 (f64 DPR:$a)),
778 (i32 (COPY_TO_REGCLASS (VCVTBDH DPR:$a), GPR))>,
779 Requires<[HasFPARMv8, HasDPVFP]>;
781 def VCVTTHD : ADuI<0b11101, 0b11, 0b0010, 0b11, 0,
782 (outs DPR:$Dd), (ins SPR:$Sm),
783 NoItinerary, "vcvtt", ".f64.f16\t$Dd, $Sm",
784 []>, Requires<[HasFPARMv8, HasDPVFP]> {
785 // Instruction operands.
788 // Encode instruction operands.
789 let Inst{3-0} = Sm{4-1};
793 def VCVTTDH : ADuI<0b11101, 0b11, 0b0011, 0b11, 0,
794 (outs SPR:$Sd), (ins DPR:$Dm),
795 NoItinerary, "vcvtt", ".f16.f64\t$Sd, $Dm",
796 []>, Requires<[HasFPARMv8, HasDPVFP]> {
797 // Instruction operands.
801 // Encode instruction operands.
802 let Inst{15-12} = Sd{4-1};
803 let Inst{22} = Sd{0};
804 let Inst{3-0} = Dm{3-0};
808 multiclass vcvt_inst<string opc, bits<2> rm,
809 SDPatternOperator node = null_frag> {
810 let PostEncoderMethod = "", DecoderNamespace = "VFPV8" in {
811 def SH : AHuInp<0b11101, 0b11, 0b1100, 0b11, 0,
812 (outs SPR:$Sd), (ins HPR:$Sm),
813 NoItinerary, !strconcat("vcvt", opc, ".s32.f16\t$Sd, $Sm"),
815 Requires<[HasFullFP16]> {
816 let Inst{17-16} = rm;
819 def UH : AHuInp<0b11101, 0b11, 0b1100, 0b01, 0,
820 (outs SPR:$Sd), (ins HPR:$Sm),
821 NoItinerary, !strconcat("vcvt", opc, ".u32.f16\t$Sd, $Sm"),
823 Requires<[HasFullFP16]> {
824 let Inst{17-16} = rm;
827 def SS : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
828 (outs SPR:$Sd), (ins SPR:$Sm),
829 NoItinerary, !strconcat("vcvt", opc, ".s32.f32\t$Sd, $Sm"),
831 Requires<[HasFPARMv8]> {
832 let Inst{17-16} = rm;
835 def US : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
836 (outs SPR:$Sd), (ins SPR:$Sm),
837 NoItinerary, !strconcat("vcvt", opc, ".u32.f32\t$Sd, $Sm"),
839 Requires<[HasFPARMv8]> {
840 let Inst{17-16} = rm;
843 def SD : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
844 (outs SPR:$Sd), (ins DPR:$Dm),
845 NoItinerary, !strconcat("vcvt", opc, ".s32.f64\t$Sd, $Dm"),
847 Requires<[HasFPARMv8, HasDPVFP]> {
850 let Inst{17-16} = rm;
852 // Encode instruction operands
853 let Inst{3-0} = Dm{3-0};
858 def UD : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
859 (outs SPR:$Sd), (ins DPR:$Dm),
860 NoItinerary, !strconcat("vcvt", opc, ".u32.f64\t$Sd, $Dm"),
862 Requires<[HasFPARMv8, HasDPVFP]> {
865 let Inst{17-16} = rm;
867 // Encode instruction operands
868 let Inst{3-0} = Dm{3-0};
874 let Predicates = [HasFPARMv8] in {
875 let Predicates = [HasFullFP16] in {
876 def : Pat<(i32 (fp_to_sint (node HPR:$a))),
878 (!cast<Instruction>(NAME#"SH") HPR:$a),
881 def : Pat<(i32 (fp_to_uint (node HPR:$a))),
883 (!cast<Instruction>(NAME#"UH") HPR:$a),
886 def : Pat<(i32 (fp_to_sint (node SPR:$a))),
888 (!cast<Instruction>(NAME#"SS") SPR:$a),
890 def : Pat<(i32 (fp_to_uint (node SPR:$a))),
892 (!cast<Instruction>(NAME#"US") SPR:$a),
895 let Predicates = [HasFPARMv8, HasDPVFP] in {
896 def : Pat<(i32 (fp_to_sint (node (f64 DPR:$a)))),
898 (!cast<Instruction>(NAME#"SD") DPR:$a),
900 def : Pat<(i32 (fp_to_uint (node (f64 DPR:$a)))),
902 (!cast<Instruction>(NAME#"UD") DPR:$a),
907 defm VCVTA : vcvt_inst<"a", 0b00, fround>;
908 defm VCVTN : vcvt_inst<"n", 0b01>;
909 defm VCVTP : vcvt_inst<"p", 0b10, fceil>;
910 defm VCVTM : vcvt_inst<"m", 0b11, ffloor>;
912 def VNEGD : ADuI<0b11101, 0b11, 0b0001, 0b01, 0,
913 (outs DPR:$Dd), (ins DPR:$Dm),
914 IIC_fpUNA64, "vneg", ".f64\t$Dd, $Dm",
915 [(set DPR:$Dd, (fneg (f64 DPR:$Dm)))]>;
917 def VNEGS : ASuIn<0b11101, 0b11, 0b0001, 0b01, 0,
918 (outs SPR:$Sd), (ins SPR:$Sm),
919 IIC_fpUNA32, "vneg", ".f32\t$Sd, $Sm",
920 [(set SPR:$Sd, (fneg SPR:$Sm))]> {
921 // Some single precision VFP instructions may be executed on both NEON and
922 // VFP pipelines on A8.
923 let D = VFPNeonA8Domain;
926 def VNEGH : AHuI<0b11101, 0b11, 0b0001, 0b01, 0,
927 (outs HPR:$Sd), (ins HPR:$Sm),
928 IIC_fpUNA16, "vneg", ".f16\t$Sd, $Sm",
929 [(set HPR:$Sd, (fneg HPR:$Sm))]>;
931 multiclass vrint_inst_zrx<string opc, bit op, bit op2, SDPatternOperator node> {
932 def H : AHuI<0b11101, 0b11, 0b0110, 0b11, 0,
933 (outs SPR:$Sd), (ins SPR:$Sm),
934 NoItinerary, !strconcat("vrint", opc), ".f16\t$Sd, $Sm",
936 Requires<[HasFullFP16]> {
941 def S : ASuI<0b11101, 0b11, 0b0110, 0b11, 0,
942 (outs SPR:$Sd), (ins SPR:$Sm),
943 NoItinerary, !strconcat("vrint", opc), ".f32\t$Sd, $Sm",
944 [(set (f32 SPR:$Sd), (node (f32 SPR:$Sm)))]>,
945 Requires<[HasFPARMv8]> {
949 def D : ADuI<0b11101, 0b11, 0b0110, 0b11, 0,
950 (outs DPR:$Dd), (ins DPR:$Dm),
951 NoItinerary, !strconcat("vrint", opc), ".f64\t$Dd, $Dm",
952 [(set (f64 DPR:$Dd), (node (f64 DPR:$Dm)))]>,
953 Requires<[HasFPARMv8, HasDPVFP]> {
958 def : InstAlias<!strconcat("vrint", opc, "$p.f16.f16\t$Sd, $Sm"),
959 (!cast<Instruction>(NAME#"H") SPR:$Sd, SPR:$Sm, pred:$p), 0>,
960 Requires<[HasFullFP16]>;
961 def : InstAlias<!strconcat("vrint", opc, "$p.f32.f32\t$Sd, $Sm"),
962 (!cast<Instruction>(NAME#"S") SPR:$Sd, SPR:$Sm, pred:$p), 0>,
963 Requires<[HasFPARMv8]>;
964 def : InstAlias<!strconcat("vrint", opc, "$p.f64.f64\t$Dd, $Dm"),
965 (!cast<Instruction>(NAME#"D") DPR:$Dd, DPR:$Dm, pred:$p), 0>,
966 Requires<[HasFPARMv8,HasDPVFP]>;
969 defm VRINTZ : vrint_inst_zrx<"z", 0, 1, ftrunc>;
970 defm VRINTR : vrint_inst_zrx<"r", 0, 0, fnearbyint>;
971 defm VRINTX : vrint_inst_zrx<"x", 1, 0, frint>;
973 multiclass vrint_inst_anpm<string opc, bits<2> rm,
974 SDPatternOperator node = null_frag> {
975 let PostEncoderMethod = "", DecoderNamespace = "VFPV8",
976 isUnpredicable = 1 in {
977 def H : AHuInp<0b11101, 0b11, 0b1000, 0b01, 0,
978 (outs SPR:$Sd), (ins SPR:$Sm),
979 NoItinerary, !strconcat("vrint", opc, ".f16\t$Sd, $Sm"),
981 Requires<[HasFullFP16]> {
982 let Inst{17-16} = rm;
984 def S : ASuInp<0b11101, 0b11, 0b1000, 0b01, 0,
985 (outs SPR:$Sd), (ins SPR:$Sm),
986 NoItinerary, !strconcat("vrint", opc, ".f32\t$Sd, $Sm"),
987 [(set (f32 SPR:$Sd), (node (f32 SPR:$Sm)))]>,
988 Requires<[HasFPARMv8]> {
989 let Inst{17-16} = rm;
991 def D : ADuInp<0b11101, 0b11, 0b1000, 0b01, 0,
992 (outs DPR:$Dd), (ins DPR:$Dm),
993 NoItinerary, !strconcat("vrint", opc, ".f64\t$Dd, $Dm"),
994 [(set (f64 DPR:$Dd), (node (f64 DPR:$Dm)))]>,
995 Requires<[HasFPARMv8, HasDPVFP]> {
996 let Inst{17-16} = rm;
1000 def : InstAlias<!strconcat("vrint", opc, ".f32.f32\t$Sd, $Sm"),
1001 (!cast<Instruction>(NAME#"S") SPR:$Sd, SPR:$Sm), 0>,
1002 Requires<[HasFPARMv8]>;
1003 def : InstAlias<!strconcat("vrint", opc, ".f64.f64\t$Dd, $Dm"),
1004 (!cast<Instruction>(NAME#"D") DPR:$Dd, DPR:$Dm), 0>,
1005 Requires<[HasFPARMv8,HasDPVFP]>;
1008 defm VRINTA : vrint_inst_anpm<"a", 0b00, fround>;
1009 defm VRINTN : vrint_inst_anpm<"n", 0b01, int_arm_neon_vrintn>;
1010 defm VRINTP : vrint_inst_anpm<"p", 0b10, fceil>;
1011 defm VRINTM : vrint_inst_anpm<"m", 0b11, ffloor>;
1013 def VSQRTD : ADuI<0b11101, 0b11, 0b0001, 0b11, 0,
1014 (outs DPR:$Dd), (ins DPR:$Dm),
1015 IIC_fpSQRT64, "vsqrt", ".f64\t$Dd, $Dm",
1016 [(set DPR:$Dd, (fsqrt (f64 DPR:$Dm)))]>,
1017 Sched<[WriteFPSQRT64]>;
1019 def VSQRTS : ASuI<0b11101, 0b11, 0b0001, 0b11, 0,
1020 (outs SPR:$Sd), (ins SPR:$Sm),
1021 IIC_fpSQRT32, "vsqrt", ".f32\t$Sd, $Sm",
1022 [(set SPR:$Sd, (fsqrt SPR:$Sm))]>,
1023 Sched<[WriteFPSQRT32]>;
1025 def VSQRTH : AHuI<0b11101, 0b11, 0b0001, 0b11, 0,
1026 (outs HPR:$Sd), (ins HPR:$Sm),
1027 IIC_fpSQRT16, "vsqrt", ".f16\t$Sd, $Sm",
1028 [(set HPR:$Sd, (fsqrt (f16 HPR:$Sm)))]>;
1030 let hasSideEffects = 0 in {
1031 let isMoveReg = 1 in {
1032 def VMOVD : ADuI<0b11101, 0b11, 0b0000, 0b01, 0,
1033 (outs DPR:$Dd), (ins DPR:$Dm),
1034 IIC_fpUNA64, "vmov", ".f64\t$Dd, $Dm", []>;
1036 def VMOVS : ASuI<0b11101, 0b11, 0b0000, 0b01, 0,
1037 (outs SPR:$Sd), (ins SPR:$Sm),
1038 IIC_fpUNA32, "vmov", ".f32\t$Sd, $Sm", []>;
1041 let PostEncoderMethod = "", DecoderNamespace = "VFPV8", isUnpredicable = 1 in {
1042 def VMOVH : ASuInp<0b11101, 0b11, 0b0000, 0b01, 0,
1043 (outs SPR:$Sd), (ins SPR:$Sm),
1044 IIC_fpUNA16, "vmovx.f16\t$Sd, $Sm", []>,
1045 Requires<[HasFullFP16]>;
1047 def VINSH : ASuInp<0b11101, 0b11, 0b0000, 0b11, 0,
1048 (outs SPR:$Sd), (ins SPR:$Sm),
1049 IIC_fpUNA16, "vins.f16\t$Sd, $Sm", []>,
1050 Requires<[HasFullFP16]>;
1051 } // PostEncoderMethod
1054 //===----------------------------------------------------------------------===//
1055 // FP <-> GPR Copies. Int <-> FP Conversions.
1058 let isMoveReg = 1 in {
1059 def VMOVRS : AVConv2I<0b11100001, 0b1010,
1060 (outs GPR:$Rt), (ins SPR:$Sn),
1061 IIC_fpMOVSI, "vmov", "\t$Rt, $Sn",
1062 [(set GPR:$Rt, (bitconvert SPR:$Sn))]>,
1063 Sched<[WriteFPMOV]> {
1064 // Instruction operands.
1068 // Encode instruction operands.
1069 let Inst{19-16} = Sn{4-1};
1070 let Inst{7} = Sn{0};
1071 let Inst{15-12} = Rt;
1073 let Inst{6-5} = 0b00;
1074 let Inst{3-0} = 0b0000;
1076 // Some single precision VFP instructions may be executed on both NEON and VFP
1078 let D = VFPNeonDomain;
1081 // Bitcast i32 -> f32. NEON prefers to use VMOVDRR.
1082 def VMOVSR : AVConv4I<0b11100000, 0b1010,
1083 (outs SPR:$Sn), (ins GPR:$Rt),
1084 IIC_fpMOVIS, "vmov", "\t$Sn, $Rt",
1085 [(set SPR:$Sn, (bitconvert GPR:$Rt))]>,
1086 Requires<[HasVFP2, UseVMOVSR]>,
1087 Sched<[WriteFPMOV]> {
1088 // Instruction operands.
1092 // Encode instruction operands.
1093 let Inst{19-16} = Sn{4-1};
1094 let Inst{7} = Sn{0};
1095 let Inst{15-12} = Rt;
1097 let Inst{6-5} = 0b00;
1098 let Inst{3-0} = 0b0000;
1100 // Some single precision VFP instructions may be executed on both NEON and VFP
1102 let D = VFPNeonDomain;
1105 def : Pat<(arm_vmovsr GPR:$Rt), (VMOVSR GPR:$Rt)>, Requires<[HasVFP2, UseVMOVSR]>;
1107 let hasSideEffects = 0 in {
1108 def VMOVRRD : AVConv3I<0b11000101, 0b1011,
1109 (outs GPR:$Rt, GPR:$Rt2), (ins DPR:$Dm),
1110 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $Dm",
1111 [(set GPR:$Rt, GPR:$Rt2, (arm_fmrrd DPR:$Dm))]>,
1112 Sched<[WriteFPMOV]> {
1113 // Instruction operands.
1118 // Encode instruction operands.
1119 let Inst{3-0} = Dm{3-0};
1120 let Inst{5} = Dm{4};
1121 let Inst{15-12} = Rt;
1122 let Inst{19-16} = Rt2;
1124 let Inst{7-6} = 0b00;
1126 // Some single precision VFP instructions may be executed on both NEON and VFP
1128 let D = VFPNeonDomain;
1130 // This instruction is equivalent to
1131 // $Rt = EXTRACT_SUBREG $Dm, ssub_0
1132 // $Rt2 = EXTRACT_SUBREG $Dm, ssub_1
1133 let isExtractSubreg = 1;
1136 def VMOVRRS : AVConv3I<0b11000101, 0b1010,
1137 (outs GPR:$Rt, GPR:$Rt2), (ins SPR:$src1, SPR:$src2),
1138 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $src1, $src2",
1139 [/* For disassembly only; pattern left blank */]>,
1140 Sched<[WriteFPMOV]> {
1145 // Encode instruction operands.
1146 let Inst{3-0} = src1{4-1};
1147 let Inst{5} = src1{0};
1148 let Inst{15-12} = Rt;
1149 let Inst{19-16} = Rt2;
1151 let Inst{7-6} = 0b00;
1153 // Some single precision VFP instructions may be executed on both NEON and VFP
1155 let D = VFPNeonDomain;
1156 let DecoderMethod = "DecodeVMOVRRS";
1160 // FMDHR: GPR -> SPR
1161 // FMDLR: GPR -> SPR
1163 def VMOVDRR : AVConv5I<0b11000100, 0b1011,
1164 (outs DPR:$Dm), (ins GPR:$Rt, GPR:$Rt2),
1165 IIC_fpMOVID, "vmov", "\t$Dm, $Rt, $Rt2",
1166 [(set DPR:$Dm, (arm_fmdrr GPR:$Rt, GPR:$Rt2))]>,
1167 Sched<[WriteFPMOV]> {
1168 // Instruction operands.
1173 // Encode instruction operands.
1174 let Inst{3-0} = Dm{3-0};
1175 let Inst{5} = Dm{4};
1176 let Inst{15-12} = Rt;
1177 let Inst{19-16} = Rt2;
1179 let Inst{7-6} = 0b00;
1181 // Some single precision VFP instructions may be executed on both NEON and VFP
1183 let D = VFPNeonDomain;
1185 // This instruction is equivalent to
1186 // $Dm = REG_SEQUENCE $Rt, ssub_0, $Rt2, ssub_1
1187 let isRegSequence = 1;
1190 // Hoist an fabs or a fneg of a value coming from integer registers
1191 // and do the fabs/fneg on the integer value. This is never a lose
1192 // and could enable the conversion to float to be removed completely.
1193 def : Pat<(fabs (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1194 (VMOVDRR GPR:$Rl, (BFC GPR:$Rh, (i32 0x7FFFFFFF)))>,
1195 Requires<[IsARM, HasV6T2]>;
1196 def : Pat<(fabs (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1197 (VMOVDRR GPR:$Rl, (t2BFC GPR:$Rh, (i32 0x7FFFFFFF)))>,
1198 Requires<[IsThumb2, HasV6T2]>;
1199 def : Pat<(fneg (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1200 (VMOVDRR GPR:$Rl, (EORri GPR:$Rh, (i32 0x80000000)))>,
1202 def : Pat<(fneg (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1203 (VMOVDRR GPR:$Rl, (t2EORri GPR:$Rh, (i32 0x80000000)))>,
1204 Requires<[IsThumb2]>;
1206 let hasSideEffects = 0 in
1207 def VMOVSRR : AVConv5I<0b11000100, 0b1010,
1208 (outs SPR:$dst1, SPR:$dst2), (ins GPR:$src1, GPR:$src2),
1209 IIC_fpMOVID, "vmov", "\t$dst1, $dst2, $src1, $src2",
1210 [/* For disassembly only; pattern left blank */]>,
1211 Sched<[WriteFPMOV]> {
1212 // Instruction operands.
1217 // Encode instruction operands.
1218 let Inst{3-0} = dst1{4-1};
1219 let Inst{5} = dst1{0};
1220 let Inst{15-12} = src1;
1221 let Inst{19-16} = src2;
1223 let Inst{7-6} = 0b00;
1225 // Some single precision VFP instructions may be executed on both NEON and VFP
1227 let D = VFPNeonDomain;
1229 let DecoderMethod = "DecodeVMOVSRR";
1232 // Move H->R, clearing top 16 bits
1233 def VMOVRH : AVConv2I<0b11100001, 0b1001,
1234 (outs GPR:$Rt), (ins HPR:$Sn),
1235 IIC_fpMOVSI, "vmov", ".f16\t$Rt, $Sn",
1236 [(set GPR:$Rt, (arm_vmovrh HPR:$Sn))]>,
1237 Requires<[HasFullFP16]>,
1238 Sched<[WriteFPMOV]> {
1239 // Instruction operands.
1243 // Encode instruction operands.
1244 let Inst{19-16} = Sn{4-1};
1245 let Inst{7} = Sn{0};
1246 let Inst{15-12} = Rt;
1248 let Inst{6-5} = 0b00;
1249 let Inst{3-0} = 0b0000;
1251 let isUnpredicable = 1;
1254 // Move R->H, clearing top 16 bits
1255 def VMOVHR : AVConv4I<0b11100000, 0b1001,
1256 (outs HPR:$Sn), (ins GPR:$Rt),
1257 IIC_fpMOVIS, "vmov", ".f16\t$Sn, $Rt",
1258 [(set HPR:$Sn, (arm_vmovhr GPR:$Rt))]>,
1259 Requires<[HasFullFP16]>,
1260 Sched<[WriteFPMOV]> {
1261 // Instruction operands.
1265 // Encode instruction operands.
1266 let Inst{19-16} = Sn{4-1};
1267 let Inst{7} = Sn{0};
1268 let Inst{15-12} = Rt;
1270 let Inst{6-5} = 0b00;
1271 let Inst{3-0} = 0b0000;
1273 let isUnpredicable = 1;
1276 // FMRDH: SPR -> GPR
1277 // FMRDL: SPR -> GPR
1278 // FMRRS: SPR -> GPR
1279 // FMRX: SPR system reg -> GPR
1280 // FMSRR: GPR -> SPR
1281 // FMXR: GPR -> VFP system reg
1286 class AVConv1IDs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1287 bits<4> opcod4, dag oops, dag iops,
1288 InstrItinClass itin, string opc, string asm,
1290 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1292 // Instruction operands.
1296 // Encode instruction operands.
1297 let Inst{3-0} = Sm{4-1};
1298 let Inst{5} = Sm{0};
1299 let Inst{15-12} = Dd{3-0};
1300 let Inst{22} = Dd{4};
1302 let Predicates = [HasVFP2, HasDPVFP];
1305 class AVConv1InSs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1306 bits<4> opcod4, dag oops, dag iops,InstrItinClass itin,
1307 string opc, string asm, list<dag> pattern>
1308 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1310 // Instruction operands.
1314 // Encode instruction operands.
1315 let Inst{3-0} = Sm{4-1};
1316 let Inst{5} = Sm{0};
1317 let Inst{15-12} = Sd{4-1};
1318 let Inst{22} = Sd{0};
1321 class AVConv1IHs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1322 bits<4> opcod4, dag oops, dag iops,
1323 InstrItinClass itin, string opc, string asm,
1325 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1327 // Instruction operands.
1331 // Encode instruction operands.
1332 let Inst{3-0} = Sm{4-1};
1333 let Inst{5} = Sm{0};
1334 let Inst{15-12} = Sd{4-1};
1335 let Inst{22} = Sd{0};
1337 let Predicates = [HasFullFP16];
1340 def VSITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
1341 (outs DPR:$Dd), (ins SPR:$Sm),
1342 IIC_fpCVTID, "vcvt", ".f64.s32\t$Dd, $Sm",
1344 Sched<[WriteFPCVT]> {
1345 let Inst{7} = 1; // s32
1348 let Predicates=[HasVFP2, HasDPVFP] in {
1349 def : VFPPat<(f64 (sint_to_fp GPR:$a)),
1350 (VSITOD (COPY_TO_REGCLASS GPR:$a, SPR))>;
1352 def : VFPPat<(f64 (sint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1353 (VSITOD (VLDRS addrmode5:$a))>;
1356 def VSITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
1357 (outs SPR:$Sd),(ins SPR:$Sm),
1358 IIC_fpCVTIS, "vcvt", ".f32.s32\t$Sd, $Sm",
1360 Sched<[WriteFPCVT]> {
1361 let Inst{7} = 1; // s32
1363 // Some single precision VFP instructions may be executed on both NEON and
1364 // VFP pipelines on A8.
1365 let D = VFPNeonA8Domain;
1368 def : VFPNoNEONPat<(f32 (sint_to_fp GPR:$a)),
1369 (VSITOS (COPY_TO_REGCLASS GPR:$a, SPR))>;
1371 def : VFPNoNEONPat<(f32 (sint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1372 (VSITOS (VLDRS addrmode5:$a))>;
1374 def VSITOH : AVConv1IHs_Encode<0b11101, 0b11, 0b1000, 0b1001,
1375 (outs HPR:$Sd), (ins SPR:$Sm),
1376 IIC_fpCVTIH, "vcvt", ".f16.s32\t$Sd, $Sm",
1378 Sched<[WriteFPCVT]> {
1379 let Inst{7} = 1; // s32
1380 let isUnpredicable = 1;
1383 def : VFPNoNEONPat<(f16 (sint_to_fp GPR:$a)),
1384 (VSITOH (COPY_TO_REGCLASS GPR:$a, SPR))>;
1386 def VUITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
1387 (outs DPR:$Dd), (ins SPR:$Sm),
1388 IIC_fpCVTID, "vcvt", ".f64.u32\t$Dd, $Sm",
1390 Sched<[WriteFPCVT]> {
1391 let Inst{7} = 0; // u32
1394 let Predicates=[HasVFP2, HasDPVFP] in {
1395 def : VFPPat<(f64 (uint_to_fp GPR:$a)),
1396 (VUITOD (COPY_TO_REGCLASS GPR:$a, SPR))>;
1398 def : VFPPat<(f64 (uint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1399 (VUITOD (VLDRS addrmode5:$a))>;
1402 def VUITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
1403 (outs SPR:$Sd), (ins SPR:$Sm),
1404 IIC_fpCVTIS, "vcvt", ".f32.u32\t$Sd, $Sm",
1406 Sched<[WriteFPCVT]> {
1407 let Inst{7} = 0; // u32
1409 // Some single precision VFP instructions may be executed on both NEON and
1410 // VFP pipelines on A8.
1411 let D = VFPNeonA8Domain;
1414 def : VFPNoNEONPat<(f32 (uint_to_fp GPR:$a)),
1415 (VUITOS (COPY_TO_REGCLASS GPR:$a, SPR))>;
1417 def : VFPNoNEONPat<(f32 (uint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1418 (VUITOS (VLDRS addrmode5:$a))>;
1420 def VUITOH : AVConv1IHs_Encode<0b11101, 0b11, 0b1000, 0b1001,
1421 (outs HPR:$Sd), (ins SPR:$Sm),
1422 IIC_fpCVTIH, "vcvt", ".f16.u32\t$Sd, $Sm",
1424 Sched<[WriteFPCVT]> {
1425 let Inst{7} = 0; // u32
1426 let isUnpredicable = 1;
1429 def : VFPNoNEONPat<(f16 (uint_to_fp GPR:$a)),
1430 (VUITOH (COPY_TO_REGCLASS GPR:$a, SPR))>;
1434 class AVConv1IsD_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1435 bits<4> opcod4, dag oops, dag iops,
1436 InstrItinClass itin, string opc, string asm,
1438 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1440 // Instruction operands.
1444 // Encode instruction operands.
1445 let Inst{3-0} = Dm{3-0};
1446 let Inst{5} = Dm{4};
1447 let Inst{15-12} = Sd{4-1};
1448 let Inst{22} = Sd{0};
1450 let Predicates = [HasVFP2, HasDPVFP];
1453 class AVConv1InsS_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1454 bits<4> opcod4, dag oops, dag iops,
1455 InstrItinClass itin, string opc, string asm,
1457 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1459 // Instruction operands.
1463 // Encode instruction operands.
1464 let Inst{3-0} = Sm{4-1};
1465 let Inst{5} = Sm{0};
1466 let Inst{15-12} = Sd{4-1};
1467 let Inst{22} = Sd{0};
1470 class AVConv1IsH_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1471 bits<4> opcod4, dag oops, dag iops,
1472 InstrItinClass itin, string opc, string asm,
1474 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1476 // Instruction operands.
1480 // Encode instruction operands.
1481 let Inst{3-0} = Sm{4-1};
1482 let Inst{5} = Sm{0};
1483 let Inst{15-12} = Sd{4-1};
1484 let Inst{22} = Sd{0};
1486 let Predicates = [HasFullFP16];
1489 // Always set Z bit in the instruction, i.e. "round towards zero" variants.
1490 def VTOSIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
1491 (outs SPR:$Sd), (ins DPR:$Dm),
1492 IIC_fpCVTDI, "vcvt", ".s32.f64\t$Sd, $Dm",
1494 Sched<[WriteFPCVT]> {
1495 let Inst{7} = 1; // Z bit
1498 let Predicates=[HasVFP2, HasDPVFP] in {
1499 def : VFPPat<(i32 (fp_to_sint (f64 DPR:$a))),
1500 (COPY_TO_REGCLASS (VTOSIZD DPR:$a), GPR)>;
1502 def : VFPPat<(alignedstore32 (i32 (fp_to_sint (f64 DPR:$a))), addrmode5:$ptr),
1503 (VSTRS (VTOSIZD DPR:$a), addrmode5:$ptr)>;
1506 def VTOSIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
1507 (outs SPR:$Sd), (ins SPR:$Sm),
1508 IIC_fpCVTSI, "vcvt", ".s32.f32\t$Sd, $Sm",
1510 Sched<[WriteFPCVT]> {
1511 let Inst{7} = 1; // Z bit
1513 // Some single precision VFP instructions may be executed on both NEON and
1514 // VFP pipelines on A8.
1515 let D = VFPNeonA8Domain;
1518 def : VFPNoNEONPat<(i32 (fp_to_sint SPR:$a)),
1519 (COPY_TO_REGCLASS (VTOSIZS SPR:$a), GPR)>;
1521 def : VFPNoNEONPat<(alignedstore32 (i32 (fp_to_sint (f32 SPR:$a))),
1523 (VSTRS (VTOSIZS SPR:$a), addrmode5:$ptr)>;
1525 def VTOSIZH : AVConv1IsH_Encode<0b11101, 0b11, 0b1101, 0b1001,
1526 (outs SPR:$Sd), (ins HPR:$Sm),
1527 IIC_fpCVTHI, "vcvt", ".s32.f16\t$Sd, $Sm",
1529 Sched<[WriteFPCVT]> {
1530 let Inst{7} = 1; // Z bit
1531 let isUnpredicable = 1;
1534 def : VFPNoNEONPat<(i32 (fp_to_sint HPR:$a)),
1535 (COPY_TO_REGCLASS (VTOSIZH HPR:$a), GPR)>;
1537 def VTOUIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
1538 (outs SPR:$Sd), (ins DPR:$Dm),
1539 IIC_fpCVTDI, "vcvt", ".u32.f64\t$Sd, $Dm",
1541 Sched<[WriteFPCVT]> {
1542 let Inst{7} = 1; // Z bit
1545 let Predicates=[HasVFP2, HasDPVFP] in {
1546 def : VFPPat<(i32 (fp_to_uint (f64 DPR:$a))),
1547 (COPY_TO_REGCLASS (VTOUIZD DPR:$a), GPR)>;
1549 def : VFPPat<(alignedstore32 (i32 (fp_to_uint (f64 DPR:$a))), addrmode5:$ptr),
1550 (VSTRS (VTOUIZD DPR:$a), addrmode5:$ptr)>;
1553 def VTOUIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
1554 (outs SPR:$Sd), (ins SPR:$Sm),
1555 IIC_fpCVTSI, "vcvt", ".u32.f32\t$Sd, $Sm",
1557 Sched<[WriteFPCVT]> {
1558 let Inst{7} = 1; // Z bit
1560 // Some single precision VFP instructions may be executed on both NEON and
1561 // VFP pipelines on A8.
1562 let D = VFPNeonA8Domain;
1565 def : VFPNoNEONPat<(i32 (fp_to_uint SPR:$a)),
1566 (COPY_TO_REGCLASS (VTOUIZS SPR:$a), GPR)>;
1568 def : VFPNoNEONPat<(alignedstore32 (i32 (fp_to_uint (f32 SPR:$a))),
1570 (VSTRS (VTOUIZS SPR:$a), addrmode5:$ptr)>;
1572 def VTOUIZH : AVConv1IsH_Encode<0b11101, 0b11, 0b1100, 0b1001,
1573 (outs SPR:$Sd), (ins HPR:$Sm),
1574 IIC_fpCVTHI, "vcvt", ".u32.f16\t$Sd, $Sm",
1576 Sched<[WriteFPCVT]> {
1577 let Inst{7} = 1; // Z bit
1578 let isUnpredicable = 1;
1581 def : VFPNoNEONPat<(i32 (fp_to_uint HPR:$a)),
1582 (COPY_TO_REGCLASS (VTOUIZH HPR:$a), GPR)>;
1584 // And the Z bit '0' variants, i.e. use the rounding mode specified by FPSCR.
1585 let Uses = [FPSCR] in {
1586 def VTOSIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
1587 (outs SPR:$Sd), (ins DPR:$Dm),
1588 IIC_fpCVTDI, "vcvtr", ".s32.f64\t$Sd, $Dm",
1589 [(set SPR:$Sd, (int_arm_vcvtr (f64 DPR:$Dm)))]>,
1590 Sched<[WriteFPCVT]> {
1591 let Inst{7} = 0; // Z bit
1594 def VTOSIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
1595 (outs SPR:$Sd), (ins SPR:$Sm),
1596 IIC_fpCVTSI, "vcvtr", ".s32.f32\t$Sd, $Sm",
1597 [(set SPR:$Sd, (int_arm_vcvtr SPR:$Sm))]>,
1598 Sched<[WriteFPCVT]> {
1599 let Inst{7} = 0; // Z bit
1602 def VTOSIRH : AVConv1IsH_Encode<0b11101, 0b11, 0b1101, 0b1001,
1603 (outs SPR:$Sd), (ins SPR:$Sm),
1604 IIC_fpCVTHI, "vcvtr", ".s32.f16\t$Sd, $Sm",
1606 Sched<[WriteFPCVT]> {
1607 let Inst{7} = 0; // Z bit
1608 let isUnpredicable = 1;
1611 def VTOUIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
1612 (outs SPR:$Sd), (ins DPR:$Dm),
1613 IIC_fpCVTDI, "vcvtr", ".u32.f64\t$Sd, $Dm",
1614 [(set SPR:$Sd, (int_arm_vcvtru(f64 DPR:$Dm)))]>,
1615 Sched<[WriteFPCVT]> {
1616 let Inst{7} = 0; // Z bit
1619 def VTOUIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
1620 (outs SPR:$Sd), (ins SPR:$Sm),
1621 IIC_fpCVTSI, "vcvtr", ".u32.f32\t$Sd, $Sm",
1622 [(set SPR:$Sd, (int_arm_vcvtru SPR:$Sm))]>,
1623 Sched<[WriteFPCVT]> {
1624 let Inst{7} = 0; // Z bit
1627 def VTOUIRH : AVConv1IsH_Encode<0b11101, 0b11, 0b1100, 0b1001,
1628 (outs SPR:$Sd), (ins SPR:$Sm),
1629 IIC_fpCVTHI, "vcvtr", ".u32.f16\t$Sd, $Sm",
1631 Sched<[WriteFPCVT]> {
1632 let Inst{7} = 0; // Z bit
1633 let isUnpredicable = 1;
1637 // v8.3-a Javascript Convert to Signed fixed-point
1638 def VJCVT : AVConv1IsD_Encode<0b11101, 0b11, 0b1001, 0b1011,
1639 (outs SPR:$Sd), (ins DPR:$Dm),
1640 IIC_fpCVTDI, "vjcvt", ".s32.f64\t$Sd, $Dm",
1642 Requires<[HasFPARMv8, HasV8_3a]> {
1643 let Inst{7} = 1; // Z bit
1646 // Convert between floating-point and fixed-point
1647 // Data type for fixed-point naming convention:
1648 // S16 (U=0, sx=0) -> SH
1649 // U16 (U=1, sx=0) -> UH
1650 // S32 (U=0, sx=1) -> SL
1651 // U32 (U=1, sx=1) -> UL
1653 let Constraints = "$a = $dst" in {
1655 // FP to Fixed-Point:
1657 // Single Precision register
1658 class AVConv1XInsS_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
1659 bit op5, dag oops, dag iops, InstrItinClass itin,
1660 string opc, string asm, list<dag> pattern>
1661 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern> {
1663 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
1664 let Inst{22} = dst{0};
1665 let Inst{15-12} = dst{4-1};
1668 // Double Precision register
1669 class AVConv1XInsD_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
1670 bit op5, dag oops, dag iops, InstrItinClass itin,
1671 string opc, string asm, list<dag> pattern>
1672 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern> {
1674 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
1675 let Inst{22} = dst{4};
1676 let Inst{15-12} = dst{3-0};
1678 let Predicates = [HasVFP2, HasDPVFP];
1681 let isUnpredicable = 1 in {
1683 def VTOSHH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1001, 0,
1684 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1685 IIC_fpCVTHI, "vcvt", ".s16.f16\t$dst, $a, $fbits", []>,
1686 Requires<[HasFullFP16]>,
1687 Sched<[WriteFPCVT]>;
1689 def VTOUHH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1001, 0,
1690 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1691 IIC_fpCVTHI, "vcvt", ".u16.f16\t$dst, $a, $fbits", []>,
1692 Requires<[HasFullFP16]>,
1693 Sched<[WriteFPCVT]>;
1695 def VTOSLH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1001, 1,
1696 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1697 IIC_fpCVTHI, "vcvt", ".s32.f16\t$dst, $a, $fbits", []>,
1698 Requires<[HasFullFP16]>,
1699 Sched<[WriteFPCVT]>;
1701 def VTOULH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1001, 1,
1702 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1703 IIC_fpCVTHI, "vcvt", ".u32.f16\t$dst, $a, $fbits", []>,
1704 Requires<[HasFullFP16]>,
1705 Sched<[WriteFPCVT]>;
1707 } // End of 'let isUnpredicable = 1 in'
1709 def VTOSHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 0,
1710 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1711 IIC_fpCVTSI, "vcvt", ".s16.f32\t$dst, $a, $fbits", []>,
1712 Sched<[WriteFPCVT]> {
1713 // Some single precision VFP instructions may be executed on both NEON and
1714 // VFP pipelines on A8.
1715 let D = VFPNeonA8Domain;
1718 def VTOUHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 0,
1719 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1720 IIC_fpCVTSI, "vcvt", ".u16.f32\t$dst, $a, $fbits", []> {
1721 // Some single precision VFP instructions may be executed on both NEON and
1722 // VFP pipelines on A8.
1723 let D = VFPNeonA8Domain;
1726 def VTOSLS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 1,
1727 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1728 IIC_fpCVTSI, "vcvt", ".s32.f32\t$dst, $a, $fbits", []> {
1729 // Some single precision VFP instructions may be executed on both NEON and
1730 // VFP pipelines on A8.
1731 let D = VFPNeonA8Domain;
1734 def VTOULS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 1,
1735 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1736 IIC_fpCVTSI, "vcvt", ".u32.f32\t$dst, $a, $fbits", []> {
1737 // Some single precision VFP instructions may be executed on both NEON and
1738 // VFP pipelines on A8.
1739 let D = VFPNeonA8Domain;
1742 def VTOSHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 0,
1743 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1744 IIC_fpCVTDI, "vcvt", ".s16.f64\t$dst, $a, $fbits", []>,
1745 Sched<[WriteFPCVT]>;
1747 def VTOUHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 0,
1748 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1749 IIC_fpCVTDI, "vcvt", ".u16.f64\t$dst, $a, $fbits", []>,
1750 Sched<[WriteFPCVT]>;
1752 def VTOSLD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 1,
1753 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1754 IIC_fpCVTDI, "vcvt", ".s32.f64\t$dst, $a, $fbits", []>,
1755 Sched<[WriteFPCVT]>;
1757 def VTOULD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 1,
1758 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1759 IIC_fpCVTDI, "vcvt", ".u32.f64\t$dst, $a, $fbits", []>,
1760 Sched<[WriteFPCVT]>;
1762 // Fixed-Point to FP:
1764 let isUnpredicable = 1 in {
1766 def VSHTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1001, 0,
1767 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1768 IIC_fpCVTIH, "vcvt", ".f16.s16\t$dst, $a, $fbits", []>,
1769 Requires<[HasFullFP16]>,
1770 Sched<[WriteFPCVT]>;
1772 def VUHTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1001, 0,
1773 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1774 IIC_fpCVTIH, "vcvt", ".f16.u16\t$dst, $a, $fbits", []>,
1775 Requires<[HasFullFP16]>,
1776 Sched<[WriteFPCVT]>;
1778 def VSLTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1001, 1,
1779 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1780 IIC_fpCVTIH, "vcvt", ".f16.s32\t$dst, $a, $fbits", []>,
1781 Requires<[HasFullFP16]>,
1782 Sched<[WriteFPCVT]>;
1784 def VULTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1001, 1,
1785 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1786 IIC_fpCVTIH, "vcvt", ".f16.u32\t$dst, $a, $fbits", []>,
1787 Requires<[HasFullFP16]>,
1788 Sched<[WriteFPCVT]>;
1790 } // End of 'let isUnpredicable = 1 in'
1792 def VSHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 0,
1793 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1794 IIC_fpCVTIS, "vcvt", ".f32.s16\t$dst, $a, $fbits", []>,
1795 Sched<[WriteFPCVT]> {
1796 // Some single precision VFP instructions may be executed on both NEON and
1797 // VFP pipelines on A8.
1798 let D = VFPNeonA8Domain;
1801 def VUHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 0,
1802 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1803 IIC_fpCVTIS, "vcvt", ".f32.u16\t$dst, $a, $fbits", []>,
1804 Sched<[WriteFPCVT]> {
1805 // Some single precision VFP instructions may be executed on both NEON and
1806 // VFP pipelines on A8.
1807 let D = VFPNeonA8Domain;
1810 def VSLTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 1,
1811 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1812 IIC_fpCVTIS, "vcvt", ".f32.s32\t$dst, $a, $fbits", []>,
1813 Sched<[WriteFPCVT]> {
1814 // Some single precision VFP instructions may be executed on both NEON and
1815 // VFP pipelines on A8.
1816 let D = VFPNeonA8Domain;
1819 def VULTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 1,
1820 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1821 IIC_fpCVTIS, "vcvt", ".f32.u32\t$dst, $a, $fbits", []>,
1822 Sched<[WriteFPCVT]> {
1823 // Some single precision VFP instructions may be executed on both NEON and
1824 // VFP pipelines on A8.
1825 let D = VFPNeonA8Domain;
1828 def VSHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 0,
1829 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1830 IIC_fpCVTID, "vcvt", ".f64.s16\t$dst, $a, $fbits", []>,
1831 Sched<[WriteFPCVT]>;
1833 def VUHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 0,
1834 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1835 IIC_fpCVTID, "vcvt", ".f64.u16\t$dst, $a, $fbits", []>,
1836 Sched<[WriteFPCVT]>;
1838 def VSLTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 1,
1839 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1840 IIC_fpCVTID, "vcvt", ".f64.s32\t$dst, $a, $fbits", []>,
1841 Sched<[WriteFPCVT]>;
1843 def VULTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 1,
1844 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1845 IIC_fpCVTID, "vcvt", ".f64.u32\t$dst, $a, $fbits", []>,
1846 Sched<[WriteFPCVT]>;
1848 } // End of 'let Constraints = "$a = $dst" in'
1850 //===----------------------------------------------------------------------===//
1851 // FP Multiply-Accumulate Operations.
1854 def VMLAD : ADbI<0b11100, 0b00, 0, 0,
1855 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1856 IIC_fpMAC64, "vmla", ".f64\t$Dd, $Dn, $Dm",
1857 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1858 (f64 DPR:$Ddin)))]>,
1859 RegConstraint<"$Ddin = $Dd">,
1860 Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>,
1861 Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
1863 def VMLAS : ASbIn<0b11100, 0b00, 0, 0,
1864 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1865 IIC_fpMAC32, "vmla", ".f32\t$Sd, $Sn, $Sm",
1866 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
1868 RegConstraint<"$Sdin = $Sd">,
1869 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>,
1870 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
1871 // Some single precision VFP instructions may be executed on both NEON and
1872 // VFP pipelines on A8.
1873 let D = VFPNeonA8Domain;
1876 def VMLAH : AHbI<0b11100, 0b00, 0, 0,
1877 (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
1878 IIC_fpMAC16, "vmla", ".f16\t$Sd, $Sn, $Sm",
1879 [(set HPR:$Sd, (fadd_mlx (fmul_su HPR:$Sn, HPR:$Sm),
1881 RegConstraint<"$Sdin = $Sd">,
1882 Requires<[HasFullFP16,UseFPVMLx]>;
1884 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1885 (VMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1886 Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>;
1887 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1888 (VMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1889 Requires<[HasVFP2,DontUseNEONForFP, UseFPVMLx]>;
1890 def : Pat<(fadd_mlx HPR:$dstin, (fmul_su HPR:$a, HPR:$b)),
1891 (VMLAH HPR:$dstin, HPR:$a, HPR:$b)>,
1892 Requires<[HasFullFP16,DontUseNEONForFP, UseFPVMLx]>;
1895 def VMLSD : ADbI<0b11100, 0b00, 1, 0,
1896 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1897 IIC_fpMAC64, "vmls", ".f64\t$Dd, $Dn, $Dm",
1898 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1899 (f64 DPR:$Ddin)))]>,
1900 RegConstraint<"$Ddin = $Dd">,
1901 Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>,
1902 Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
1904 def VMLSS : ASbIn<0b11100, 0b00, 1, 0,
1905 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1906 IIC_fpMAC32, "vmls", ".f32\t$Sd, $Sn, $Sm",
1907 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1909 RegConstraint<"$Sdin = $Sd">,
1910 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>,
1911 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
1912 // Some single precision VFP instructions may be executed on both NEON and
1913 // VFP pipelines on A8.
1914 let D = VFPNeonA8Domain;
1917 def VMLSH : AHbI<0b11100, 0b00, 1, 0,
1918 (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
1919 IIC_fpMAC16, "vmls", ".f16\t$Sd, $Sn, $Sm",
1920 [(set HPR:$Sd, (fadd_mlx (fneg (fmul_su HPR:$Sn, HPR:$Sm)),
1922 RegConstraint<"$Sdin = $Sd">,
1923 Requires<[HasFullFP16,UseFPVMLx]>;
1925 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1926 (VMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
1927 Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>;
1928 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1929 (VMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
1930 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>;
1931 def : Pat<(fsub_mlx HPR:$dstin, (fmul_su HPR:$a, HPR:$b)),
1932 (VMLSH HPR:$dstin, HPR:$a, HPR:$b)>,
1933 Requires<[HasFullFP16,DontUseNEONForFP,UseFPVMLx]>;
1935 def VNMLAD : ADbI<0b11100, 0b01, 1, 0,
1936 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1937 IIC_fpMAC64, "vnmla", ".f64\t$Dd, $Dn, $Dm",
1938 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1939 (f64 DPR:$Ddin)))]>,
1940 RegConstraint<"$Ddin = $Dd">,
1941 Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>,
1942 Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
1944 def VNMLAS : ASbI<0b11100, 0b01, 1, 0,
1945 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1946 IIC_fpMAC32, "vnmla", ".f32\t$Sd, $Sn, $Sm",
1947 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1949 RegConstraint<"$Sdin = $Sd">,
1950 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>,
1951 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
1952 // Some single precision VFP instructions may be executed on both NEON and
1953 // VFP pipelines on A8.
1954 let D = VFPNeonA8Domain;
1957 def VNMLAH : AHbI<0b11100, 0b01, 1, 0,
1958 (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
1959 IIC_fpMAC16, "vnmla", ".f16\t$Sd, $Sn, $Sm",
1960 [(set HPR:$Sd, (fsub_mlx (fneg (fmul_su HPR:$Sn, HPR:$Sm)),
1962 RegConstraint<"$Sdin = $Sd">,
1963 Requires<[HasFullFP16,UseFPVMLx]>;
1965 // (-(a * b) - dst) -> -(dst + (a * b))
1966 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
1967 (VNMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1968 Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>;
1969 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
1970 (VNMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1971 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>;
1972 def : Pat<(fsub_mlx (fneg (fmul_su HPR:$a, HPR:$b)), HPR:$dstin),
1973 (VNMLAH HPR:$dstin, HPR:$a, HPR:$b)>,
1974 Requires<[HasFullFP16,DontUseNEONForFP,UseFPVMLx]>;
1976 // (-dst - (a * b)) -> -(dst + (a * b))
1977 def : Pat<(fsub_mlx (fneg DPR:$dstin), (fmul_su DPR:$a, (f64 DPR:$b))),
1978 (VNMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1979 Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>;
1980 def : Pat<(fsub_mlx (fneg SPR:$dstin), (fmul_su SPR:$a, SPR:$b)),
1981 (VNMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1982 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>;
1983 def : Pat<(fsub_mlx (fneg HPR:$dstin), (fmul_su HPR:$a, HPR:$b)),
1984 (VNMLAH HPR:$dstin, HPR:$a, HPR:$b)>,
1985 Requires<[HasFullFP16,DontUseNEONForFP,UseFPVMLx]>;
1987 def VNMLSD : ADbI<0b11100, 0b01, 0, 0,
1988 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1989 IIC_fpMAC64, "vnmls", ".f64\t$Dd, $Dn, $Dm",
1990 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1991 (f64 DPR:$Ddin)))]>,
1992 RegConstraint<"$Ddin = $Dd">,
1993 Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>,
1994 Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
1996 def VNMLSS : ASbI<0b11100, 0b01, 0, 0,
1997 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1998 IIC_fpMAC32, "vnmls", ".f32\t$Sd, $Sn, $Sm",
1999 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
2000 RegConstraint<"$Sdin = $Sd">,
2001 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>,
2002 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2003 // Some single precision VFP instructions may be executed on both NEON and
2004 // VFP pipelines on A8.
2005 let D = VFPNeonA8Domain;
2008 def VNMLSH : AHbI<0b11100, 0b01, 0, 0,
2009 (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2010 IIC_fpMAC16, "vnmls", ".f16\t$Sd, $Sn, $Sm",
2011 [(set HPR:$Sd, (fsub_mlx (fmul_su HPR:$Sn, HPR:$Sm), HPR:$Sdin))]>,
2012 RegConstraint<"$Sdin = $Sd">,
2013 Requires<[HasFullFP16,UseFPVMLx]>;
2015 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
2016 (VNMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
2017 Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>;
2018 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
2019 (VNMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
2020 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>;
2021 def : Pat<(fsub_mlx (fmul_su HPR:$a, HPR:$b), HPR:$dstin),
2022 (VNMLSH HPR:$dstin, HPR:$a, HPR:$b)>,
2023 Requires<[HasFullFP16,DontUseNEONForFP,UseFPVMLx]>;
2025 //===----------------------------------------------------------------------===//
2026 // Fused FP Multiply-Accumulate Operations.
2028 def VFMAD : ADbI<0b11101, 0b10, 0, 0,
2029 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2030 IIC_fpFMAC64, "vfma", ".f64\t$Dd, $Dn, $Dm",
2031 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
2032 (f64 DPR:$Ddin)))]>,
2033 RegConstraint<"$Ddin = $Dd">,
2034 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>,
2035 Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2037 def VFMAS : ASbIn<0b11101, 0b10, 0, 0,
2038 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2039 IIC_fpFMAC32, "vfma", ".f32\t$Sd, $Sn, $Sm",
2040 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
2042 RegConstraint<"$Sdin = $Sd">,
2043 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>,
2044 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2045 // Some single precision VFP instructions may be executed on both NEON and
2049 def VFMAH : AHbI<0b11101, 0b10, 0, 0,
2050 (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2051 IIC_fpFMAC16, "vfma", ".f16\t$Sd, $Sn, $Sm",
2052 [(set HPR:$Sd, (fadd_mlx (fmul_su HPR:$Sn, HPR:$Sm),
2054 RegConstraint<"$Sdin = $Sd">,
2055 Requires<[HasFullFP16,UseFusedMAC]>,
2056 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2058 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
2059 (VFMAD DPR:$dstin, DPR:$a, DPR:$b)>,
2060 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
2061 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
2062 (VFMAS SPR:$dstin, SPR:$a, SPR:$b)>,
2063 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
2064 def : Pat<(fadd_mlx HPR:$dstin, (fmul_su HPR:$a, HPR:$b)),
2065 (VFMAH HPR:$dstin, HPR:$a, HPR:$b)>,
2066 Requires<[HasFullFP16,DontUseNEONForFP,UseFusedMAC]>;
2068 // Match @llvm.fma.* intrinsics
2069 // (fma x, y, z) -> (vfms z, x, y)
2070 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, DPR:$Ddin)),
2071 (VFMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2072 Requires<[HasVFP4,HasDPVFP]>;
2073 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, SPR:$Sdin)),
2074 (VFMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2075 Requires<[HasVFP4]>;
2077 def VFMSD : ADbI<0b11101, 0b10, 1, 0,
2078 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2079 IIC_fpFMAC64, "vfms", ".f64\t$Dd, $Dn, $Dm",
2080 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
2081 (f64 DPR:$Ddin)))]>,
2082 RegConstraint<"$Ddin = $Dd">,
2083 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>,
2084 Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2086 def VFMSS : ASbIn<0b11101, 0b10, 1, 0,
2087 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2088 IIC_fpFMAC32, "vfms", ".f32\t$Sd, $Sn, $Sm",
2089 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
2091 RegConstraint<"$Sdin = $Sd">,
2092 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>,
2093 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2094 // Some single precision VFP instructions may be executed on both NEON and
2098 def VFMSH : AHbI<0b11101, 0b10, 1, 0,
2099 (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2100 IIC_fpFMAC16, "vfms", ".f16\t$Sd, $Sn, $Sm",
2101 [(set HPR:$Sd, (fadd_mlx (fneg (fmul_su HPR:$Sn, HPR:$Sm)),
2103 RegConstraint<"$Sdin = $Sd">,
2104 Requires<[HasFullFP16,UseFusedMAC]>,
2105 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2107 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
2108 (VFMSD DPR:$dstin, DPR:$a, DPR:$b)>,
2109 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
2110 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
2111 (VFMSS SPR:$dstin, SPR:$a, SPR:$b)>,
2112 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
2113 def : Pat<(fsub_mlx HPR:$dstin, (fmul_su HPR:$a, HPR:$b)),
2114 (VFMSH HPR:$dstin, HPR:$a, HPR:$b)>,
2115 Requires<[HasFullFP16,DontUseNEONForFP,UseFusedMAC]>;
2117 // Match @llvm.fma.* intrinsics
2118 // (fma (fneg x), y, z) -> (vfms z, x, y)
2119 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin)),
2120 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2121 Requires<[HasVFP4,HasDPVFP]>;
2122 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin)),
2123 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2124 Requires<[HasVFP4]>;
2125 // (fma x, (fneg y), z) -> (vfms z, x, y)
2126 def : Pat<(f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin)),
2127 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2128 Requires<[HasVFP4,HasDPVFP]>;
2129 def : Pat<(f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin)),
2130 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2131 Requires<[HasVFP4]>;
2133 def VFNMAD : ADbI<0b11101, 0b01, 1, 0,
2134 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2135 IIC_fpFMAC64, "vfnma", ".f64\t$Dd, $Dn, $Dm",
2136 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
2137 (f64 DPR:$Ddin)))]>,
2138 RegConstraint<"$Ddin = $Dd">,
2139 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>,
2140 Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2142 def VFNMAS : ASbI<0b11101, 0b01, 1, 0,
2143 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2144 IIC_fpFMAC32, "vfnma", ".f32\t$Sd, $Sn, $Sm",
2145 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
2147 RegConstraint<"$Sdin = $Sd">,
2148 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>,
2149 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2150 // Some single precision VFP instructions may be executed on both NEON and
2154 def VFNMAH : AHbI<0b11101, 0b01, 1, 0,
2155 (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2156 IIC_fpFMAC16, "vfnma", ".f16\t$Sd, $Sn, $Sm",
2157 [(set HPR:$Sd, (fsub_mlx (fneg (fmul_su HPR:$Sn, HPR:$Sm)),
2159 RegConstraint<"$Sdin = $Sd">,
2160 Requires<[HasFullFP16,UseFusedMAC]>,
2161 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2163 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
2164 (VFNMAD DPR:$dstin, DPR:$a, DPR:$b)>,
2165 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
2166 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
2167 (VFNMAS SPR:$dstin, SPR:$a, SPR:$b)>,
2168 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
2170 // Match @llvm.fma.* intrinsics
2171 // (fneg (fma x, y, z)) -> (vfnma z, x, y)
2172 def : Pat<(fneg (fma (f64 DPR:$Dn), (f64 DPR:$Dm), (f64 DPR:$Ddin))),
2173 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2174 Requires<[HasVFP4,HasDPVFP]>;
2175 def : Pat<(fneg (fma (f32 SPR:$Sn), (f32 SPR:$Sm), (f32 SPR:$Sdin))),
2176 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2177 Requires<[HasVFP4]>;
2178 // (fma (fneg x), y, (fneg z)) -> (vfnma z, x, y)
2179 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, (fneg DPR:$Ddin))),
2180 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2181 Requires<[HasVFP4,HasDPVFP]>;
2182 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, (fneg SPR:$Sdin))),
2183 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2184 Requires<[HasVFP4]>;
2186 def VFNMSD : ADbI<0b11101, 0b01, 0, 0,
2187 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2188 IIC_fpFMAC64, "vfnms", ".f64\t$Dd, $Dn, $Dm",
2189 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
2190 (f64 DPR:$Ddin)))]>,
2191 RegConstraint<"$Ddin = $Dd">,
2192 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>,
2193 Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2195 def VFNMSS : ASbI<0b11101, 0b01, 0, 0,
2196 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2197 IIC_fpFMAC32, "vfnms", ".f32\t$Sd, $Sn, $Sm",
2198 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
2199 RegConstraint<"$Sdin = $Sd">,
2200 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>,
2201 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2202 // Some single precision VFP instructions may be executed on both NEON and
2206 def VFNMSH : AHbI<0b11101, 0b01, 0, 0,
2207 (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2208 IIC_fpFMAC16, "vfnms", ".f16\t$Sd, $Sn, $Sm",
2209 [(set HPR:$Sd, (fsub_mlx (fmul_su HPR:$Sn, HPR:$Sm), HPR:$Sdin))]>,
2210 RegConstraint<"$Sdin = $Sd">,
2211 Requires<[HasFullFP16,UseFusedMAC]>,
2212 Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2214 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
2215 (VFNMSD DPR:$dstin, DPR:$a, DPR:$b)>,
2216 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
2217 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
2218 (VFNMSS SPR:$dstin, SPR:$a, SPR:$b)>,
2219 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
2221 // Match @llvm.fma.* intrinsics
2223 // (fma x, y, (fneg z)) -> (vfnms z, x, y))
2224 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, (fneg DPR:$Ddin))),
2225 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2226 Requires<[HasVFP4,HasDPVFP]>;
2227 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, (fneg SPR:$Sdin))),
2228 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2229 Requires<[HasVFP4]>;
2230 // (fneg (fma (fneg x), y, z)) -> (vfnms z, x, y)
2231 def : Pat<(fneg (f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin))),
2232 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2233 Requires<[HasVFP4,HasDPVFP]>;
2234 def : Pat<(fneg (f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin))),
2235 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2236 Requires<[HasVFP4]>;
2237 // (fneg (fma x, (fneg y), z) -> (vfnms z, x, y)
2238 def : Pat<(fneg (f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin))),
2239 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2240 Requires<[HasVFP4,HasDPVFP]>;
2241 def : Pat<(fneg (f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin))),
2242 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2243 Requires<[HasVFP4]>;
2245 //===----------------------------------------------------------------------===//
2246 // FP Conditional moves.
2249 let hasSideEffects = 0 in {
2250 def VMOVDcc : PseudoInst<(outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm, cmovpred:$p),
2252 [(set (f64 DPR:$Dd),
2253 (ARMcmov DPR:$Dn, DPR:$Dm, cmovpred:$p))]>,
2254 RegConstraint<"$Dn = $Dd">, Requires<[HasVFP2,HasDPVFP]>;
2256 def VMOVScc : PseudoInst<(outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm, cmovpred:$p),
2258 [(set (f32 SPR:$Sd),
2259 (ARMcmov SPR:$Sn, SPR:$Sm, cmovpred:$p))]>,
2260 RegConstraint<"$Sn = $Sd">, Requires<[HasVFP2]>;
2263 //===----------------------------------------------------------------------===//
2264 // Move from VFP System Register to ARM core register.
2267 class MovFromVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
2269 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
2271 // Instruction operand.
2274 let Inst{27-20} = 0b11101111;
2275 let Inst{19-16} = opc19_16;
2276 let Inst{15-12} = Rt;
2277 let Inst{11-8} = 0b1010;
2279 let Inst{6-5} = 0b00;
2281 let Inst{3-0} = 0b0000;
2284 // APSR is the application level alias of CPSR. This FPSCR N, Z, C, V flags
2286 let Defs = [CPSR], Uses = [FPSCR_NZCV], Rt = 0b1111 /* apsr_nzcv */ in
2287 def FMSTAT : MovFromVFP<0b0001 /* fpscr */, (outs), (ins),
2288 "vmrs", "\tAPSR_nzcv, fpscr", [(arm_fmstat)]>;
2290 let DecoderMethod = "DecodeForVMRSandVMSR" in {
2291 // Application level FPSCR -> GPR
2292 let hasSideEffects = 1, Uses = [FPSCR] in
2293 def VMRS : MovFromVFP<0b0001 /* fpscr */, (outs GPRnopc:$Rt), (ins),
2294 "vmrs", "\t$Rt, fpscr",
2295 [(set GPRnopc:$Rt, (int_arm_get_fpscr))]>;
2297 // System level FPEXC, FPSID -> GPR
2298 let Uses = [FPSCR] in {
2299 def VMRS_FPEXC : MovFromVFP<0b1000 /* fpexc */, (outs GPRnopc:$Rt), (ins),
2300 "vmrs", "\t$Rt, fpexc", []>;
2301 def VMRS_FPSID : MovFromVFP<0b0000 /* fpsid */, (outs GPRnopc:$Rt), (ins),
2302 "vmrs", "\t$Rt, fpsid", []>;
2303 def VMRS_MVFR0 : MovFromVFP<0b0111 /* mvfr0 */, (outs GPRnopc:$Rt), (ins),
2304 "vmrs", "\t$Rt, mvfr0", []>;
2305 def VMRS_MVFR1 : MovFromVFP<0b0110 /* mvfr1 */, (outs GPRnopc:$Rt), (ins),
2306 "vmrs", "\t$Rt, mvfr1", []>;
2307 let Predicates = [HasFPARMv8] in {
2308 def VMRS_MVFR2 : MovFromVFP<0b0101 /* mvfr2 */, (outs GPRnopc:$Rt), (ins),
2309 "vmrs", "\t$Rt, mvfr2", []>;
2311 def VMRS_FPINST : MovFromVFP<0b1001 /* fpinst */, (outs GPRnopc:$Rt), (ins),
2312 "vmrs", "\t$Rt, fpinst", []>;
2313 def VMRS_FPINST2 : MovFromVFP<0b1010 /* fpinst2 */, (outs GPRnopc:$Rt),
2314 (ins), "vmrs", "\t$Rt, fpinst2", []>;
2318 //===----------------------------------------------------------------------===//
2319 // Move from ARM core register to VFP System Register.
2322 class MovToVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
2324 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
2326 // Instruction operand.
2329 // Encode instruction operand.
2330 let Inst{15-12} = src;
2332 let Inst{27-20} = 0b11101110;
2333 let Inst{19-16} = opc19_16;
2334 let Inst{11-8} = 0b1010;
2339 let DecoderMethod = "DecodeForVMRSandVMSR" in {
2340 let Defs = [FPSCR] in {
2341 // Application level GPR -> FPSCR
2342 def VMSR : MovToVFP<0b0001 /* fpscr */, (outs), (ins GPRnopc:$src),
2343 "vmsr", "\tfpscr, $src",
2344 [(int_arm_set_fpscr GPRnopc:$src)]>;
2345 // System level GPR -> FPEXC
2346 def VMSR_FPEXC : MovToVFP<0b1000 /* fpexc */, (outs), (ins GPRnopc:$src),
2347 "vmsr", "\tfpexc, $src", []>;
2348 // System level GPR -> FPSID
2349 def VMSR_FPSID : MovToVFP<0b0000 /* fpsid */, (outs), (ins GPRnopc:$src),
2350 "vmsr", "\tfpsid, $src", []>;
2351 def VMSR_FPINST : MovToVFP<0b1001 /* fpinst */, (outs), (ins GPRnopc:$src),
2352 "vmsr", "\tfpinst, $src", []>;
2353 def VMSR_FPINST2 : MovToVFP<0b1010 /* fpinst2 */, (outs), (ins GPRnopc:$src),
2354 "vmsr", "\tfpinst2, $src", []>;
2358 //===----------------------------------------------------------------------===//
2362 // Materialize FP immediates. VFP3 only.
2363 let isReMaterializable = 1 in {
2364 def FCONSTD : VFPAI<(outs DPR:$Dd), (ins vfp_f64imm:$imm),
2365 VFPMiscFrm, IIC_fpUNA64,
2366 "vmov", ".f64\t$Dd, $imm",
2367 [(set DPR:$Dd, vfp_f64imm:$imm)]>,
2368 Requires<[HasVFP3,HasDPVFP]> {
2372 let Inst{27-23} = 0b11101;
2373 let Inst{22} = Dd{4};
2374 let Inst{21-20} = 0b11;
2375 let Inst{19-16} = imm{7-4};
2376 let Inst{15-12} = Dd{3-0};
2377 let Inst{11-9} = 0b101;
2378 let Inst{8} = 1; // Double precision.
2379 let Inst{7-4} = 0b0000;
2380 let Inst{3-0} = imm{3-0};
2383 def FCONSTS : VFPAI<(outs SPR:$Sd), (ins vfp_f32imm:$imm),
2384 VFPMiscFrm, IIC_fpUNA32,
2385 "vmov", ".f32\t$Sd, $imm",
2386 [(set SPR:$Sd, vfp_f32imm:$imm)]>, Requires<[HasVFP3]> {
2390 let Inst{27-23} = 0b11101;
2391 let Inst{22} = Sd{0};
2392 let Inst{21-20} = 0b11;
2393 let Inst{19-16} = imm{7-4};
2394 let Inst{15-12} = Sd{4-1};
2395 let Inst{11-9} = 0b101;
2396 let Inst{8} = 0; // Single precision.
2397 let Inst{7-4} = 0b0000;
2398 let Inst{3-0} = imm{3-0};
2401 def FCONSTH : VFPAI<(outs HPR:$Sd), (ins vfp_f16imm:$imm),
2402 VFPMiscFrm, IIC_fpUNA16,
2403 "vmov", ".f16\t$Sd, $imm",
2404 [(set HPR:$Sd, vfp_f16imm:$imm)]>,
2405 Requires<[HasFullFP16]> {
2409 let Inst{27-23} = 0b11101;
2410 let Inst{22} = Sd{0};
2411 let Inst{21-20} = 0b11;
2412 let Inst{19-16} = imm{7-4};
2413 let Inst{15-12} = Sd{4-1};
2414 let Inst{11-8} = 0b1001; // Half precision
2415 let Inst{7-4} = 0b0000;
2416 let Inst{3-0} = imm{3-0};
2418 let isUnpredicable = 1;
2422 //===----------------------------------------------------------------------===//
2423 // Assembler aliases.
2425 // A few mnemonic aliases for pre-unifixed syntax. We don't guarantee to
2426 // support them all, but supporting at least some of the basics is
2427 // good to be friendly.
2428 def : VFP2MnemonicAlias<"flds", "vldr">;
2429 def : VFP2MnemonicAlias<"fldd", "vldr">;
2430 def : VFP2MnemonicAlias<"fmrs", "vmov">;
2431 def : VFP2MnemonicAlias<"fmsr", "vmov">;
2432 def : VFP2MnemonicAlias<"fsqrts", "vsqrt">;
2433 def : VFP2MnemonicAlias<"fsqrtd", "vsqrt">;
2434 def : VFP2MnemonicAlias<"fadds", "vadd.f32">;
2435 def : VFP2MnemonicAlias<"faddd", "vadd.f64">;
2436 def : VFP2MnemonicAlias<"fmrdd", "vmov">;
2437 def : VFP2MnemonicAlias<"fmrds", "vmov">;
2438 def : VFP2MnemonicAlias<"fmrrd", "vmov">;
2439 def : VFP2MnemonicAlias<"fmdrr", "vmov">;
2440 def : VFP2MnemonicAlias<"fmuls", "vmul.f32">;
2441 def : VFP2MnemonicAlias<"fmuld", "vmul.f64">;
2442 def : VFP2MnemonicAlias<"fnegs", "vneg.f32">;
2443 def : VFP2MnemonicAlias<"fnegd", "vneg.f64">;
2444 def : VFP2MnemonicAlias<"ftosizd", "vcvt.s32.f64">;
2445 def : VFP2MnemonicAlias<"ftosid", "vcvtr.s32.f64">;
2446 def : VFP2MnemonicAlias<"ftosizs", "vcvt.s32.f32">;
2447 def : VFP2MnemonicAlias<"ftosis", "vcvtr.s32.f32">;
2448 def : VFP2MnemonicAlias<"ftouizd", "vcvt.u32.f64">;
2449 def : VFP2MnemonicAlias<"ftouid", "vcvtr.u32.f64">;
2450 def : VFP2MnemonicAlias<"ftouizs", "vcvt.u32.f32">;
2451 def : VFP2MnemonicAlias<"ftouis", "vcvtr.u32.f32">;
2452 def : VFP2MnemonicAlias<"fsitod", "vcvt.f64.s32">;
2453 def : VFP2MnemonicAlias<"fsitos", "vcvt.f32.s32">;
2454 def : VFP2MnemonicAlias<"fuitod", "vcvt.f64.u32">;
2455 def : VFP2MnemonicAlias<"fuitos", "vcvt.f32.u32">;
2456 def : VFP2MnemonicAlias<"fsts", "vstr">;
2457 def : VFP2MnemonicAlias<"fstd", "vstr">;
2458 def : VFP2MnemonicAlias<"fmacd", "vmla.f64">;
2459 def : VFP2MnemonicAlias<"fmacs", "vmla.f32">;
2460 def : VFP2MnemonicAlias<"fcpys", "vmov.f32">;
2461 def : VFP2MnemonicAlias<"fcpyd", "vmov.f64">;
2462 def : VFP2MnemonicAlias<"fcmps", "vcmp.f32">;
2463 def : VFP2MnemonicAlias<"fcmpd", "vcmp.f64">;
2464 def : VFP2MnemonicAlias<"fdivs", "vdiv.f32">;
2465 def : VFP2MnemonicAlias<"fdivd", "vdiv.f64">;
2466 def : VFP2MnemonicAlias<"fmrx", "vmrs">;
2467 def : VFP2MnemonicAlias<"fmxr", "vmsr">;
2469 // Be friendly and accept the old form of zero-compare
2470 def : VFP2DPInstAlias<"fcmpzd${p} $val", (VCMPZD DPR:$val, pred:$p)>;
2471 def : VFP2InstAlias<"fcmpzs${p} $val", (VCMPZS SPR:$val, pred:$p)>;
2474 def : VFP2InstAlias<"fmstat${p}", (FMSTAT pred:$p)>;
2475 def : VFP2InstAlias<"fadds${p} $Sd, $Sn, $Sm",
2476 (VADDS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
2477 def : VFP2DPInstAlias<"faddd${p} $Dd, $Dn, $Dm",
2478 (VADDD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
2479 def : VFP2InstAlias<"fsubs${p} $Sd, $Sn, $Sm",
2480 (VSUBS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
2481 def : VFP2DPInstAlias<"fsubd${p} $Dd, $Dn, $Dm",
2482 (VSUBD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
2484 // No need for the size suffix on VSQRT. It's implied by the register classes.
2485 def : VFP2InstAlias<"vsqrt${p} $Sd, $Sm", (VSQRTS SPR:$Sd, SPR:$Sm, pred:$p)>;
2486 def : VFP2DPInstAlias<"vsqrt${p} $Dd, $Dm", (VSQRTD DPR:$Dd, DPR:$Dm, pred:$p)>;
2488 // VLDR/VSTR accept an optional type suffix.
2489 def : VFP2InstAlias<"vldr${p}.32 $Sd, $addr",
2490 (VLDRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
2491 def : VFP2InstAlias<"vstr${p}.32 $Sd, $addr",
2492 (VSTRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
2493 def : VFP2InstAlias<"vldr${p}.64 $Dd, $addr",
2494 (VLDRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
2495 def : VFP2InstAlias<"vstr${p}.64 $Dd, $addr",
2496 (VSTRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
2498 // VMOV can accept optional 32-bit or less data type suffix suffix.
2499 def : VFP2InstAlias<"vmov${p}.8 $Rt, $Sn",
2500 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
2501 def : VFP2InstAlias<"vmov${p}.16 $Rt, $Sn",
2502 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
2503 def : VFP2InstAlias<"vmov${p}.32 $Rt, $Sn",
2504 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
2505 def : VFP2InstAlias<"vmov${p}.8 $Sn, $Rt",
2506 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
2507 def : VFP2InstAlias<"vmov${p}.16 $Sn, $Rt",
2508 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
2509 def : VFP2InstAlias<"vmov${p}.32 $Sn, $Rt",
2510 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
2512 def : VFP2InstAlias<"vmov${p}.f64 $Rt, $Rt2, $Dn",
2513 (VMOVRRD GPR:$Rt, GPR:$Rt2, DPR:$Dn, pred:$p)>;
2514 def : VFP2InstAlias<"vmov${p}.f64 $Dn, $Rt, $Rt2",
2515 (VMOVDRR DPR:$Dn, GPR:$Rt, GPR:$Rt2, pred:$p)>;
2517 // VMOVS doesn't need the .f32 to disambiguate from the NEON encoding the way
2519 def : VFP2InstAlias<"vmov${p} $Sd, $Sm",
2520 (VMOVS SPR:$Sd, SPR:$Sm, pred:$p)>;
2522 // FCONSTD/FCONSTS alias for vmov.f64/vmov.f32
2523 // These aliases provide added functionality over vmov.f instructions by
2524 // allowing users to write assembly containing encoded floating point constants
2525 // (e.g. #0x70 vs #1.0). Without these alises there is no way for the
2526 // assembler to accept encoded fp constants (but the equivalent fp-literal is
2527 // accepted directly by vmovf).
2528 def : VFP3InstAlias<"fconstd${p} $Dd, $val",
2529 (FCONSTD DPR:$Dd, vfp_f64imm:$val, pred:$p)>;
2530 def : VFP3InstAlias<"fconsts${p} $Sd, $val",
2531 (FCONSTS SPR:$Sd, vfp_f32imm:$val, pred:$p)>;