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_FTOI : SDTypeProfile<1, 1, [SDTCisVT<0, f32>, SDTCisFP<1>]>;
15 def SDT_ITOF : SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisVT<1, f32>]>;
16 def SDT_CMPFP0 : SDTypeProfile<0, 1, [SDTCisFP<0>]>;
17 def SDT_VMOVDRR : SDTypeProfile<1, 2, [SDTCisVT<0, f64>, SDTCisVT<1, i32>,
20 def arm_ftoui : SDNode<"ARMISD::FTOUI", SDT_FTOI>;
21 def arm_ftosi : SDNode<"ARMISD::FTOSI", SDT_FTOI>;
22 def arm_sitof : SDNode<"ARMISD::SITOF", SDT_ITOF>;
23 def arm_uitof : SDNode<"ARMISD::UITOF", SDT_ITOF>;
24 def arm_fmstat : SDNode<"ARMISD::FMSTAT", SDTNone, [SDNPInGlue, SDNPOutGlue]>;
25 def arm_cmpfp : SDNode<"ARMISD::CMPFP", SDT_ARMCmp, [SDNPOutGlue]>;
26 def arm_cmpfp0 : SDNode<"ARMISD::CMPFPw0", SDT_CMPFP0, [SDNPOutGlue]>;
27 def arm_fmdrr : SDNode<"ARMISD::VMOVDRR", SDT_VMOVDRR>;
30 //===----------------------------------------------------------------------===//
31 // Operand Definitions.
34 // 8-bit floating-point immediate encodings.
35 def FPImmOperand : AsmOperandClass {
37 let ParserMethod = "parseFPImm";
40 def vfp_f32imm : Operand<f32>,
41 PatLeaf<(f32 fpimm), [{
42 return ARM_AM::getFP32Imm(N->getValueAPF()) != -1;
43 }], SDNodeXForm<fpimm, [{
44 APFloat InVal = N->getValueAPF();
45 uint32_t enc = ARM_AM::getFP32Imm(InVal);
46 return CurDAG->getTargetConstant(enc, MVT::i32);
48 let PrintMethod = "printFPImmOperand";
49 let ParserMatchClass = FPImmOperand;
52 def vfp_f64imm : Operand<f64>,
53 PatLeaf<(f64 fpimm), [{
54 return ARM_AM::getFP64Imm(N->getValueAPF()) != -1;
55 }], SDNodeXForm<fpimm, [{
56 APFloat InVal = N->getValueAPF();
57 uint32_t enc = ARM_AM::getFP64Imm(InVal);
58 return CurDAG->getTargetConstant(enc, MVT::i32);
60 let PrintMethod = "printFPImmOperand";
61 let ParserMatchClass = FPImmOperand;
64 def alignedload32 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
65 return cast<LoadSDNode>(N)->getAlignment() >= 4;
68 def alignedstore32 : PatFrag<(ops node:$val, node:$ptr),
69 (store node:$val, node:$ptr), [{
70 return cast<StoreSDNode>(N)->getAlignment() >= 4;
73 // The VCVT to/from fixed-point instructions encode the 'fbits' operand
74 // (the number of fixed bits) differently than it appears in the assembly
75 // source. It's encoded as "Size - fbits" where Size is the size of the
76 // fixed-point representation (32 or 16) and fbits is the value appearing
77 // in the assembly source, an integer in [0,16] or (0,32], depending on size.
78 def fbits32_asm_operand : AsmOperandClass { let Name = "FBits32"; }
79 def fbits32 : Operand<i32> {
80 let PrintMethod = "printFBits32";
81 let ParserMatchClass = fbits32_asm_operand;
84 def fbits16_asm_operand : AsmOperandClass { let Name = "FBits16"; }
85 def fbits16 : Operand<i32> {
86 let PrintMethod = "printFBits16";
87 let ParserMatchClass = fbits16_asm_operand;
90 //===----------------------------------------------------------------------===//
91 // Load / store Instructions.
94 let canFoldAsLoad = 1, isReMaterializable = 1 in {
96 def VLDRD : ADI5<0b1101, 0b01, (outs DPR:$Dd), (ins addrmode5:$addr),
97 IIC_fpLoad64, "vldr", "\t$Dd, $addr",
98 [(set DPR:$Dd, (f64 (alignedload32 addrmode5:$addr)))]>;
100 def VLDRS : ASI5<0b1101, 0b01, (outs SPR:$Sd), (ins addrmode5:$addr),
101 IIC_fpLoad32, "vldr", "\t$Sd, $addr",
102 [(set SPR:$Sd, (load addrmode5:$addr))]> {
103 // Some single precision VFP instructions may be executed on both NEON and VFP
105 let D = VFPNeonDomain;
108 } // End of 'let canFoldAsLoad = 1, isReMaterializable = 1 in'
110 def VSTRD : ADI5<0b1101, 0b00, (outs), (ins DPR:$Dd, addrmode5:$addr),
111 IIC_fpStore64, "vstr", "\t$Dd, $addr",
112 [(alignedstore32 (f64 DPR:$Dd), addrmode5:$addr)]>;
114 def VSTRS : ASI5<0b1101, 0b00, (outs), (ins SPR:$Sd, addrmode5:$addr),
115 IIC_fpStore32, "vstr", "\t$Sd, $addr",
116 [(store SPR:$Sd, addrmode5:$addr)]> {
117 // Some single precision VFP instructions may be executed on both NEON and VFP
119 let D = VFPNeonDomain;
122 //===----------------------------------------------------------------------===//
123 // Load / store multiple Instructions.
126 multiclass vfp_ldst_mult<string asm, bit L_bit,
127 InstrItinClass itin, InstrItinClass itin_upd> {
130 AXDI4<(outs), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
132 !strconcat(asm, "ia${p}\t$Rn, $regs"), "", []> {
133 let Inst{24-23} = 0b01; // Increment After
134 let Inst{21} = 0; // No writeback
135 let Inst{20} = L_bit;
138 AXDI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs,
140 IndexModeUpd, itin_upd,
141 !strconcat(asm, "ia${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
142 let Inst{24-23} = 0b01; // Increment After
143 let Inst{21} = 1; // Writeback
144 let Inst{20} = L_bit;
147 AXDI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs,
149 IndexModeUpd, itin_upd,
150 !strconcat(asm, "db${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
151 let Inst{24-23} = 0b10; // Decrement Before
152 let Inst{21} = 1; // Writeback
153 let Inst{20} = L_bit;
158 AXSI4<(outs), (ins GPR:$Rn, pred:$p, spr_reglist:$regs, variable_ops),
160 !strconcat(asm, "ia${p}\t$Rn, $regs"), "", []> {
161 let Inst{24-23} = 0b01; // Increment After
162 let Inst{21} = 0; // No writeback
163 let Inst{20} = L_bit;
165 // Some single precision VFP instructions may be executed on both NEON and
167 let D = VFPNeonDomain;
170 AXSI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, spr_reglist:$regs,
172 IndexModeUpd, itin_upd,
173 !strconcat(asm, "ia${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
174 let Inst{24-23} = 0b01; // Increment After
175 let Inst{21} = 1; // Writeback
176 let Inst{20} = L_bit;
178 // Some single precision VFP instructions may be executed on both NEON and
180 let D = VFPNeonDomain;
183 AXSI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, spr_reglist:$regs,
185 IndexModeUpd, itin_upd,
186 !strconcat(asm, "db${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
187 let Inst{24-23} = 0b10; // Decrement Before
188 let Inst{21} = 1; // Writeback
189 let Inst{20} = L_bit;
191 // Some single precision VFP instructions may be executed on both NEON and
193 let D = VFPNeonDomain;
197 let neverHasSideEffects = 1 in {
199 let mayLoad = 1, hasExtraDefRegAllocReq = 1 in
200 defm VLDM : vfp_ldst_mult<"vldm", 1, IIC_fpLoad_m, IIC_fpLoad_mu>;
202 let mayStore = 1, hasExtraSrcRegAllocReq = 1 in
203 defm VSTM : vfp_ldst_mult<"vstm", 0, IIC_fpStore_m, IIC_fpStore_mu>;
205 } // neverHasSideEffects
207 def : MnemonicAlias<"vldm", "vldmia">;
208 def : MnemonicAlias<"vstm", "vstmia">;
210 // FLDM/FSTM - Load / Store multiple single / double precision registers for
212 // These instructions are deprecated!
213 def : VFP2MnemonicAlias<"fldmias", "vldmia">;
214 def : VFP2MnemonicAlias<"fldmdbs", "vldmdb">;
215 def : VFP2MnemonicAlias<"fldmeas", "vldmdb">;
216 def : VFP2MnemonicAlias<"fldmfds", "vldmia">;
217 def : VFP2MnemonicAlias<"fldmiad", "vldmia">;
218 def : VFP2MnemonicAlias<"fldmdbd", "vldmdb">;
219 def : VFP2MnemonicAlias<"fldmead", "vldmdb">;
220 def : VFP2MnemonicAlias<"fldmfdd", "vldmia">;
222 def : VFP2MnemonicAlias<"fstmias", "vstmia">;
223 def : VFP2MnemonicAlias<"fstmdbs", "vstmdb">;
224 def : VFP2MnemonicAlias<"fstmeas", "vstmia">;
225 def : VFP2MnemonicAlias<"fstmfds", "vstmdb">;
226 def : VFP2MnemonicAlias<"fstmiad", "vstmia">;
227 def : VFP2MnemonicAlias<"fstmdbd", "vstmdb">;
228 def : VFP2MnemonicAlias<"fstmead", "vstmia">;
229 def : VFP2MnemonicAlias<"fstmfdd", "vstmdb">;
231 def : InstAlias<"vpush${p} $r", (VSTMDDB_UPD SP, pred:$p, dpr_reglist:$r)>,
233 def : InstAlias<"vpush${p} $r", (VSTMSDB_UPD SP, pred:$p, spr_reglist:$r)>,
235 def : InstAlias<"vpop${p} $r", (VLDMDIA_UPD SP, pred:$p, dpr_reglist:$r)>,
237 def : InstAlias<"vpop${p} $r", (VLDMSIA_UPD SP, pred:$p, spr_reglist:$r)>,
239 defm : VFPDTAnyInstAlias<"vpush${p}", "$r",
240 (VSTMSDB_UPD SP, pred:$p, spr_reglist:$r)>;
241 defm : VFPDTAnyInstAlias<"vpush${p}", "$r",
242 (VSTMDDB_UPD SP, pred:$p, dpr_reglist:$r)>;
243 defm : VFPDTAnyInstAlias<"vpop${p}", "$r",
244 (VLDMSIA_UPD SP, pred:$p, spr_reglist:$r)>;
245 defm : VFPDTAnyInstAlias<"vpop${p}", "$r",
246 (VLDMDIA_UPD SP, pred:$p, dpr_reglist:$r)>;
248 // FLDMX, FSTMX - Load and store multiple unknown precision registers for
250 // These instruction are deprecated so we don't want them to get selected.
251 multiclass vfp_ldstx_mult<string asm, bit L_bit> {
254 AXXI4<(outs), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
255 IndexModeNone, !strconcat(asm, "iax${p}\t$Rn, $regs"), "", []> {
256 let Inst{24-23} = 0b01; // Increment After
257 let Inst{21} = 0; // No writeback
258 let Inst{20} = L_bit;
261 AXXI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
262 IndexModeUpd, !strconcat(asm, "iax${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
263 let Inst{24-23} = 0b01; // Increment After
264 let Inst{21} = 1; // Writeback
265 let Inst{20} = L_bit;
268 AXXI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
269 IndexModeUpd, !strconcat(asm, "dbx${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
270 let Inst{24-23} = 0b10; // Decrement Before
271 let Inst{21} = 1; // Writeback
272 let Inst{20} = L_bit;
276 defm FLDM : vfp_ldstx_mult<"fldm", 1>;
277 defm FSTM : vfp_ldstx_mult<"fstm", 0>;
279 def : VFP2MnemonicAlias<"fldmeax", "fldmdbx">;
280 def : VFP2MnemonicAlias<"fldmfdx", "fldmiax">;
282 def : VFP2MnemonicAlias<"fstmeax", "fstmiax">;
283 def : VFP2MnemonicAlias<"fstmfdx", "fstmdbx">;
285 //===----------------------------------------------------------------------===//
286 // FP Binary Operations.
289 let TwoOperandAliasConstraint = "$Dn = $Dd" in
290 def VADDD : ADbI<0b11100, 0b11, 0, 0,
291 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
292 IIC_fpALU64, "vadd", ".f64\t$Dd, $Dn, $Dm",
293 [(set DPR:$Dd, (fadd DPR:$Dn, (f64 DPR:$Dm)))]>;
295 let TwoOperandAliasConstraint = "$Sn = $Sd" in
296 def VADDS : ASbIn<0b11100, 0b11, 0, 0,
297 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
298 IIC_fpALU32, "vadd", ".f32\t$Sd, $Sn, $Sm",
299 [(set SPR:$Sd, (fadd SPR:$Sn, SPR:$Sm))]> {
300 // Some single precision VFP instructions may be executed on both NEON and
301 // VFP pipelines on A8.
302 let D = VFPNeonA8Domain;
305 let TwoOperandAliasConstraint = "$Dn = $Dd" in
306 def VSUBD : ADbI<0b11100, 0b11, 1, 0,
307 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
308 IIC_fpALU64, "vsub", ".f64\t$Dd, $Dn, $Dm",
309 [(set DPR:$Dd, (fsub DPR:$Dn, (f64 DPR:$Dm)))]>;
311 let TwoOperandAliasConstraint = "$Sn = $Sd" in
312 def VSUBS : ASbIn<0b11100, 0b11, 1, 0,
313 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
314 IIC_fpALU32, "vsub", ".f32\t$Sd, $Sn, $Sm",
315 [(set SPR:$Sd, (fsub SPR:$Sn, SPR:$Sm))]> {
316 // Some single precision VFP instructions may be executed on both NEON and
317 // VFP pipelines on A8.
318 let D = VFPNeonA8Domain;
321 let TwoOperandAliasConstraint = "$Dn = $Dd" in
322 def VDIVD : ADbI<0b11101, 0b00, 0, 0,
323 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
324 IIC_fpDIV64, "vdiv", ".f64\t$Dd, $Dn, $Dm",
325 [(set DPR:$Dd, (fdiv DPR:$Dn, (f64 DPR:$Dm)))]>;
327 let TwoOperandAliasConstraint = "$Sn = $Sd" in
328 def VDIVS : ASbI<0b11101, 0b00, 0, 0,
329 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
330 IIC_fpDIV32, "vdiv", ".f32\t$Sd, $Sn, $Sm",
331 [(set SPR:$Sd, (fdiv SPR:$Sn, SPR:$Sm))]>;
333 let TwoOperandAliasConstraint = "$Dn = $Dd" in
334 def VMULD : ADbI<0b11100, 0b10, 0, 0,
335 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
336 IIC_fpMUL64, "vmul", ".f64\t$Dd, $Dn, $Dm",
337 [(set DPR:$Dd, (fmul DPR:$Dn, (f64 DPR:$Dm)))]>;
339 let TwoOperandAliasConstraint = "$Sn = $Sd" in
340 def VMULS : ASbIn<0b11100, 0b10, 0, 0,
341 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
342 IIC_fpMUL32, "vmul", ".f32\t$Sd, $Sn, $Sm",
343 [(set SPR:$Sd, (fmul SPR:$Sn, SPR:$Sm))]> {
344 // Some single precision VFP instructions may be executed on both NEON and
345 // VFP pipelines on A8.
346 let D = VFPNeonA8Domain;
349 def VNMULD : ADbI<0b11100, 0b10, 1, 0,
350 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
351 IIC_fpMUL64, "vnmul", ".f64\t$Dd, $Dn, $Dm",
352 [(set DPR:$Dd, (fneg (fmul DPR:$Dn, (f64 DPR:$Dm))))]>;
354 def VNMULS : ASbI<0b11100, 0b10, 1, 0,
355 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
356 IIC_fpMUL32, "vnmul", ".f32\t$Sd, $Sn, $Sm",
357 [(set SPR:$Sd, (fneg (fmul SPR:$Sn, SPR:$Sm)))]> {
358 // Some single precision VFP instructions may be executed on both NEON and
359 // VFP pipelines on A8.
360 let D = VFPNeonA8Domain;
363 multiclass vsel_inst<string op, bits<2> opc, int CC> {
364 let DecoderNamespace = "VFPV8", PostEncoderMethod = "",
365 Uses = [CPSR], AddedComplexity = 4 in {
366 def S : ASbInp<0b11100, opc, 0,
367 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
368 NoItinerary, !strconcat("vsel", op, ".f32\t$Sd, $Sn, $Sm"),
369 [(set SPR:$Sd, (ARMcmov SPR:$Sm, SPR:$Sn, CC))]>,
370 Requires<[HasFPARMv8]>;
372 def D : ADbInp<0b11100, opc, 0,
373 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
374 NoItinerary, !strconcat("vsel", op, ".f64\t$Dd, $Dn, $Dm"),
375 [(set DPR:$Dd, (ARMcmov (f64 DPR:$Dm), (f64 DPR:$Dn), CC))]>,
376 Requires<[HasFPARMv8, HasDPVFP]>;
380 // The CC constants here match ARMCC::CondCodes.
381 defm VSELGT : vsel_inst<"gt", 0b11, 12>;
382 defm VSELGE : vsel_inst<"ge", 0b10, 10>;
383 defm VSELEQ : vsel_inst<"eq", 0b00, 0>;
384 defm VSELVS : vsel_inst<"vs", 0b01, 6>;
386 multiclass vmaxmin_inst<string op, bit opc, SDNode SD> {
387 let DecoderNamespace = "VFPV8", PostEncoderMethod = "" in {
388 def S : ASbInp<0b11101, 0b00, opc,
389 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
390 NoItinerary, !strconcat(op, ".f32\t$Sd, $Sn, $Sm"),
391 [(set SPR:$Sd, (SD SPR:$Sn, SPR:$Sm))]>,
392 Requires<[HasFPARMv8]>;
394 def D : ADbInp<0b11101, 0b00, opc,
395 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
396 NoItinerary, !strconcat(op, ".f64\t$Dd, $Dn, $Dm"),
397 [(set DPR:$Dd, (f64 (SD (f64 DPR:$Dn), (f64 DPR:$Dm))))]>,
398 Requires<[HasFPARMv8, HasDPVFP]>;
402 defm VMAXNM : vmaxmin_inst<"vmaxnm", 0, ARMvmaxnm>;
403 defm VMINNM : vmaxmin_inst<"vminnm", 1, ARMvminnm>;
405 // Match reassociated forms only if not sign dependent rounding.
406 def : Pat<(fmul (fneg DPR:$a), (f64 DPR:$b)),
407 (VNMULD DPR:$a, DPR:$b)>,
408 Requires<[NoHonorSignDependentRounding,HasDPVFP]>;
409 def : Pat<(fmul (fneg SPR:$a), SPR:$b),
410 (VNMULS SPR:$a, SPR:$b)>, Requires<[NoHonorSignDependentRounding]>;
412 // These are encoded as unary instructions.
413 let Defs = [FPSCR_NZCV] in {
414 def VCMPED : ADuI<0b11101, 0b11, 0b0100, 0b11, 0,
415 (outs), (ins DPR:$Dd, DPR:$Dm),
416 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, $Dm",
417 [(arm_cmpfp DPR:$Dd, (f64 DPR:$Dm))]>;
419 def VCMPES : ASuI<0b11101, 0b11, 0b0100, 0b11, 0,
420 (outs), (ins SPR:$Sd, SPR:$Sm),
421 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, $Sm",
422 [(arm_cmpfp SPR:$Sd, SPR:$Sm)]> {
423 // Some single precision VFP instructions may be executed on both NEON and
424 // VFP pipelines on A8.
425 let D = VFPNeonA8Domain;
428 // FIXME: Verify encoding after integrated assembler is working.
429 def VCMPD : ADuI<0b11101, 0b11, 0b0100, 0b01, 0,
430 (outs), (ins DPR:$Dd, DPR:$Dm),
431 IIC_fpCMP64, "vcmp", ".f64\t$Dd, $Dm",
432 [/* For disassembly only; pattern left blank */]>;
434 def VCMPS : ASuI<0b11101, 0b11, 0b0100, 0b01, 0,
435 (outs), (ins SPR:$Sd, SPR:$Sm),
436 IIC_fpCMP32, "vcmp", ".f32\t$Sd, $Sm",
437 [/* For disassembly only; pattern left blank */]> {
438 // Some single precision VFP instructions may be executed on both NEON and
439 // VFP pipelines on A8.
440 let D = VFPNeonA8Domain;
442 } // Defs = [FPSCR_NZCV]
444 //===----------------------------------------------------------------------===//
445 // FP Unary Operations.
448 def VABSD : ADuI<0b11101, 0b11, 0b0000, 0b11, 0,
449 (outs DPR:$Dd), (ins DPR:$Dm),
450 IIC_fpUNA64, "vabs", ".f64\t$Dd, $Dm",
451 [(set DPR:$Dd, (fabs (f64 DPR:$Dm)))]>;
453 def VABSS : ASuIn<0b11101, 0b11, 0b0000, 0b11, 0,
454 (outs SPR:$Sd), (ins SPR:$Sm),
455 IIC_fpUNA32, "vabs", ".f32\t$Sd, $Sm",
456 [(set SPR:$Sd, (fabs SPR:$Sm))]> {
457 // Some single precision VFP instructions may be executed on both NEON and
458 // VFP pipelines on A8.
459 let D = VFPNeonA8Domain;
462 let Defs = [FPSCR_NZCV] in {
463 def VCMPEZD : ADuI<0b11101, 0b11, 0b0101, 0b11, 0,
464 (outs), (ins DPR:$Dd),
465 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, #0",
466 [(arm_cmpfp0 (f64 DPR:$Dd))]> {
467 let Inst{3-0} = 0b0000;
471 def VCMPEZS : ASuI<0b11101, 0b11, 0b0101, 0b11, 0,
472 (outs), (ins SPR:$Sd),
473 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, #0",
474 [(arm_cmpfp0 SPR:$Sd)]> {
475 let Inst{3-0} = 0b0000;
478 // Some single precision VFP instructions may be executed on both NEON and
479 // VFP pipelines on A8.
480 let D = VFPNeonA8Domain;
483 // FIXME: Verify encoding after integrated assembler is working.
484 def VCMPZD : ADuI<0b11101, 0b11, 0b0101, 0b01, 0,
485 (outs), (ins DPR:$Dd),
486 IIC_fpCMP64, "vcmp", ".f64\t$Dd, #0",
487 [/* For disassembly only; pattern left blank */]> {
488 let Inst{3-0} = 0b0000;
492 def VCMPZS : ASuI<0b11101, 0b11, 0b0101, 0b01, 0,
493 (outs), (ins SPR:$Sd),
494 IIC_fpCMP32, "vcmp", ".f32\t$Sd, #0",
495 [/* For disassembly only; pattern left blank */]> {
496 let Inst{3-0} = 0b0000;
499 // Some single precision VFP instructions may be executed on both NEON and
500 // VFP pipelines on A8.
501 let D = VFPNeonA8Domain;
503 } // Defs = [FPSCR_NZCV]
505 def VCVTDS : ASuI<0b11101, 0b11, 0b0111, 0b11, 0,
506 (outs DPR:$Dd), (ins SPR:$Sm),
507 IIC_fpCVTDS, "vcvt", ".f64.f32\t$Dd, $Sm",
508 [(set DPR:$Dd, (fextend SPR:$Sm))]> {
509 // Instruction operands.
513 // Encode instruction operands.
514 let Inst{3-0} = Sm{4-1};
516 let Inst{15-12} = Dd{3-0};
517 let Inst{22} = Dd{4};
520 // Special case encoding: bits 11-8 is 0b1011.
521 def VCVTSD : VFPAI<(outs SPR:$Sd), (ins DPR:$Dm), VFPUnaryFrm,
522 IIC_fpCVTSD, "vcvt", ".f32.f64\t$Sd, $Dm",
523 [(set SPR:$Sd, (fround DPR:$Dm))]> {
524 // Instruction operands.
528 // Encode instruction operands.
529 let Inst{3-0} = Dm{3-0};
531 let Inst{15-12} = Sd{4-1};
532 let Inst{22} = Sd{0};
534 let Inst{27-23} = 0b11101;
535 let Inst{21-16} = 0b110111;
536 let Inst{11-8} = 0b1011;
537 let Inst{7-6} = 0b11;
540 let Predicates = [HasVFP2, HasDPVFP];
543 // Between half, single and double-precision. For disassembly only.
545 // FIXME: Verify encoding after integrated assembler is working.
546 def VCVTBHS: ASuI<0b11101, 0b11, 0b0010, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
547 /* FIXME */ IIC_fpCVTSH, "vcvtb", ".f32.f16\t$Sd, $Sm",
548 [/* For disassembly only; pattern left blank */]>;
550 def VCVTBSH: ASuI<0b11101, 0b11, 0b0011, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
551 /* FIXME */ IIC_fpCVTHS, "vcvtb", ".f16.f32\t$Sd, $Sm",
552 [/* For disassembly only; pattern left blank */]>;
554 def : Pat<(f32_to_f16 SPR:$a),
555 (i32 (COPY_TO_REGCLASS (VCVTBSH SPR:$a), GPR))>;
557 def : Pat<(f16_to_f32 GPR:$a),
558 (VCVTBHS (COPY_TO_REGCLASS GPR:$a, SPR))>;
560 def VCVTTHS: ASuI<0b11101, 0b11, 0b0010, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
561 /* FIXME */ IIC_fpCVTSH, "vcvtt", ".f32.f16\t$Sd, $Sm",
562 [/* For disassembly only; pattern left blank */]>;
564 def VCVTTSH: ASuI<0b11101, 0b11, 0b0011, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
565 /* FIXME */ IIC_fpCVTHS, "vcvtt", ".f16.f32\t$Sd, $Sm",
566 [/* For disassembly only; pattern left blank */]>;
568 def VCVTBHD : ADuI<0b11101, 0b11, 0b0010, 0b01, 0,
569 (outs DPR:$Dd), (ins SPR:$Sm),
570 NoItinerary, "vcvtb", ".f64.f16\t$Dd, $Sm",
571 []>, Requires<[HasFPARMv8, HasDPVFP]> {
572 // Instruction operands.
575 // Encode instruction operands.
576 let Inst{3-0} = Sm{4-1};
580 def VCVTBDH : ADuI<0b11101, 0b11, 0b0011, 0b01, 0,
581 (outs SPR:$Sd), (ins DPR:$Dm),
582 NoItinerary, "vcvtb", ".f16.f64\t$Sd, $Dm",
583 []>, Requires<[HasFPARMv8, HasDPVFP]> {
584 // Instruction operands.
588 // Encode instruction operands.
589 let Inst{3-0} = Dm{3-0};
591 let Inst{15-12} = Sd{4-1};
592 let Inst{22} = Sd{0};
595 def VCVTTHD : ADuI<0b11101, 0b11, 0b0010, 0b11, 0,
596 (outs DPR:$Dd), (ins SPR:$Sm),
597 NoItinerary, "vcvtt", ".f64.f16\t$Dd, $Sm",
598 []>, Requires<[HasFPARMv8, HasDPVFP]> {
599 // Instruction operands.
602 // Encode instruction operands.
603 let Inst{3-0} = Sm{4-1};
607 def VCVTTDH : ADuI<0b11101, 0b11, 0b0011, 0b11, 0,
608 (outs SPR:$Sd), (ins DPR:$Dm),
609 NoItinerary, "vcvtt", ".f16.f64\t$Sd, $Dm",
610 []>, Requires<[HasFPARMv8, HasDPVFP]> {
611 // Instruction operands.
615 // Encode instruction operands.
616 let Inst{15-12} = Sd{4-1};
617 let Inst{22} = Sd{0};
618 let Inst{3-0} = Dm{3-0};
622 multiclass vcvt_inst<string opc, bits<2> rm> {
623 let PostEncoderMethod = "", DecoderNamespace = "VFPV8" in {
624 def SS : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
625 (outs SPR:$Sd), (ins SPR:$Sm),
626 NoItinerary, !strconcat("vcvt", opc, ".s32.f32\t$Sd, $Sm"),
627 []>, Requires<[HasFPARMv8]> {
628 let Inst{17-16} = rm;
631 def US : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
632 (outs SPR:$Sd), (ins SPR:$Sm),
633 NoItinerary, !strconcat("vcvt", opc, ".u32.f32\t$Sd, $Sm"),
634 []>, Requires<[HasFPARMv8]> {
635 let Inst{17-16} = rm;
638 def SD : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
639 (outs SPR:$Sd), (ins DPR:$Dm),
640 NoItinerary, !strconcat("vcvt", opc, ".s32.f64\t$Sd, $Dm"),
641 []>, Requires<[HasFPARMv8, HasDPVFP]> {
644 let Inst{17-16} = rm;
646 // Encode instruction operands
647 let Inst{3-0} = Dm{3-0};
652 def UD : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
653 (outs SPR:$Sd), (ins DPR:$Dm),
654 NoItinerary, !strconcat("vcvt", opc, ".u32.f64\t$Sd, $Dm"),
655 []>, Requires<[HasFPARMv8, HasDPVFP]> {
658 let Inst{17-16} = rm;
660 // Encode instruction operands
661 let Inst{3-0} = Dm{3-0};
668 defm VCVTA : vcvt_inst<"a", 0b00>;
669 defm VCVTN : vcvt_inst<"n", 0b01>;
670 defm VCVTP : vcvt_inst<"p", 0b10>;
671 defm VCVTM : vcvt_inst<"m", 0b11>;
673 def VNEGD : ADuI<0b11101, 0b11, 0b0001, 0b01, 0,
674 (outs DPR:$Dd), (ins DPR:$Dm),
675 IIC_fpUNA64, "vneg", ".f64\t$Dd, $Dm",
676 [(set DPR:$Dd, (fneg (f64 DPR:$Dm)))]>;
678 def VNEGS : ASuIn<0b11101, 0b11, 0b0001, 0b01, 0,
679 (outs SPR:$Sd), (ins SPR:$Sm),
680 IIC_fpUNA32, "vneg", ".f32\t$Sd, $Sm",
681 [(set SPR:$Sd, (fneg SPR:$Sm))]> {
682 // Some single precision VFP instructions may be executed on both NEON and
683 // VFP pipelines on A8.
684 let D = VFPNeonA8Domain;
687 multiclass vrint_inst_zrx<string opc, bit op, bit op2> {
688 def S : ASuI<0b11101, 0b11, 0b0110, 0b11, 0,
689 (outs SPR:$Sd), (ins SPR:$Sm),
690 NoItinerary, !strconcat("vrint", opc), ".f32\t$Sd, $Sm",
691 []>, Requires<[HasFPARMv8]> {
695 def D : ADuI<0b11101, 0b11, 0b0110, 0b11, 0,
696 (outs DPR:$Dd), (ins DPR:$Dm),
697 NoItinerary, !strconcat("vrint", opc), ".f64\t$Dd, $Dm",
698 []>, Requires<[HasFPARMv8, HasDPVFP]> {
703 def : InstAlias<!strconcat("vrint", opc, "$p.f32.f32\t$Sd, $Sm"),
704 (!cast<Instruction>(NAME#"S") SPR:$Sd, SPR:$Sm, pred:$p)>,
705 Requires<[HasFPARMv8]>;
706 def : InstAlias<!strconcat("vrint", opc, "$p.f64.f64\t$Dd, $Dm"),
707 (!cast<Instruction>(NAME#"D") DPR:$Dd, DPR:$Dm, pred:$p)>,
708 Requires<[HasFPARMv8,HasDPVFP]>;
711 defm VRINTZ : vrint_inst_zrx<"z", 0, 1>;
712 defm VRINTR : vrint_inst_zrx<"r", 0, 0>;
713 defm VRINTX : vrint_inst_zrx<"x", 1, 0>;
715 multiclass vrint_inst_anpm<string opc, bits<2> rm> {
716 let PostEncoderMethod = "", DecoderNamespace = "VFPV8" in {
717 def S : ASuInp<0b11101, 0b11, 0b1000, 0b01, 0,
718 (outs SPR:$Sd), (ins SPR:$Sm),
719 NoItinerary, !strconcat("vrint", opc, ".f32\t$Sd, $Sm"),
720 []>, Requires<[HasFPARMv8]> {
721 let Inst{17-16} = rm;
723 def D : ADuInp<0b11101, 0b11, 0b1000, 0b01, 0,
724 (outs DPR:$Dd), (ins DPR:$Dm),
725 NoItinerary, !strconcat("vrint", opc, ".f64\t$Dd, $Dm"),
726 []>, Requires<[HasFPARMv8, HasDPVFP]> {
727 let Inst{17-16} = rm;
731 def : InstAlias<!strconcat("vrint", opc, ".f32.f32\t$Sd, $Sm"),
732 (!cast<Instruction>(NAME#"S") SPR:$Sd, SPR:$Sm)>,
733 Requires<[HasFPARMv8]>;
734 def : InstAlias<!strconcat("vrint", opc, ".f64.f64\t$Dd, $Dm"),
735 (!cast<Instruction>(NAME#"D") DPR:$Dd, DPR:$Dm)>,
736 Requires<[HasFPARMv8,HasDPVFP]>;
739 defm VRINTA : vrint_inst_anpm<"a", 0b00>;
740 defm VRINTN : vrint_inst_anpm<"n", 0b01>;
741 defm VRINTP : vrint_inst_anpm<"p", 0b10>;
742 defm VRINTM : vrint_inst_anpm<"m", 0b11>;
744 def VSQRTD : ADuI<0b11101, 0b11, 0b0001, 0b11, 0,
745 (outs DPR:$Dd), (ins DPR:$Dm),
746 IIC_fpSQRT64, "vsqrt", ".f64\t$Dd, $Dm",
747 [(set DPR:$Dd, (fsqrt (f64 DPR:$Dm)))]>;
749 def VSQRTS : ASuI<0b11101, 0b11, 0b0001, 0b11, 0,
750 (outs SPR:$Sd), (ins SPR:$Sm),
751 IIC_fpSQRT32, "vsqrt", ".f32\t$Sd, $Sm",
752 [(set SPR:$Sd, (fsqrt SPR:$Sm))]>;
754 let neverHasSideEffects = 1 in {
755 def VMOVD : ADuI<0b11101, 0b11, 0b0000, 0b01, 0,
756 (outs DPR:$Dd), (ins DPR:$Dm),
757 IIC_fpUNA64, "vmov", ".f64\t$Dd, $Dm", []>;
759 def VMOVS : ASuI<0b11101, 0b11, 0b0000, 0b01, 0,
760 (outs SPR:$Sd), (ins SPR:$Sm),
761 IIC_fpUNA32, "vmov", ".f32\t$Sd, $Sm", []>;
762 } // neverHasSideEffects
764 //===----------------------------------------------------------------------===//
765 // FP <-> GPR Copies. Int <-> FP Conversions.
768 def VMOVRS : AVConv2I<0b11100001, 0b1010,
769 (outs GPR:$Rt), (ins SPR:$Sn),
770 IIC_fpMOVSI, "vmov", "\t$Rt, $Sn",
771 [(set GPR:$Rt, (bitconvert SPR:$Sn))]> {
772 // Instruction operands.
776 // Encode instruction operands.
777 let Inst{19-16} = Sn{4-1};
779 let Inst{15-12} = Rt;
781 let Inst{6-5} = 0b00;
782 let Inst{3-0} = 0b0000;
784 // Some single precision VFP instructions may be executed on both NEON and VFP
786 let D = VFPNeonDomain;
789 // Bitcast i32 -> f32. NEON prefers to use VMOVDRR.
790 def VMOVSR : AVConv4I<0b11100000, 0b1010,
791 (outs SPR:$Sn), (ins GPR:$Rt),
792 IIC_fpMOVIS, "vmov", "\t$Sn, $Rt",
793 [(set SPR:$Sn, (bitconvert GPR:$Rt))]>,
794 Requires<[HasVFP2, UseVMOVSR]> {
795 // Instruction operands.
799 // Encode instruction operands.
800 let Inst{19-16} = Sn{4-1};
802 let Inst{15-12} = Rt;
804 let Inst{6-5} = 0b00;
805 let Inst{3-0} = 0b0000;
807 // Some single precision VFP instructions may be executed on both NEON and VFP
809 let D = VFPNeonDomain;
812 let neverHasSideEffects = 1 in {
813 def VMOVRRD : AVConv3I<0b11000101, 0b1011,
814 (outs GPR:$Rt, GPR:$Rt2), (ins DPR:$Dm),
815 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $Dm",
816 [/* FIXME: Can't write pattern for multiple result instr*/]> {
817 // Instruction operands.
822 // Encode instruction operands.
823 let Inst{3-0} = Dm{3-0};
825 let Inst{15-12} = Rt;
826 let Inst{19-16} = Rt2;
828 let Inst{7-6} = 0b00;
830 // Some single precision VFP instructions may be executed on both NEON and VFP
832 let D = VFPNeonDomain;
835 def VMOVRRS : AVConv3I<0b11000101, 0b1010,
836 (outs GPR:$Rt, GPR:$Rt2), (ins SPR:$src1, SPR:$src2),
837 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $src1, $src2",
838 [/* For disassembly only; pattern left blank */]> {
843 // Encode instruction operands.
844 let Inst{3-0} = src1{4-1};
845 let Inst{5} = src1{0};
846 let Inst{15-12} = Rt;
847 let Inst{19-16} = Rt2;
849 let Inst{7-6} = 0b00;
851 // Some single precision VFP instructions may be executed on both NEON and VFP
853 let D = VFPNeonDomain;
854 let DecoderMethod = "DecodeVMOVRRS";
856 } // neverHasSideEffects
861 def VMOVDRR : AVConv5I<0b11000100, 0b1011,
862 (outs DPR:$Dm), (ins GPR:$Rt, GPR:$Rt2),
863 IIC_fpMOVID, "vmov", "\t$Dm, $Rt, $Rt2",
864 [(set DPR:$Dm, (arm_fmdrr GPR:$Rt, GPR:$Rt2))]> {
865 // Instruction operands.
870 // Encode instruction operands.
871 let Inst{3-0} = Dm{3-0};
873 let Inst{15-12} = Rt;
874 let Inst{19-16} = Rt2;
876 let Inst{7-6} = 0b00;
878 // Some single precision VFP instructions may be executed on both NEON and VFP
880 let D = VFPNeonDomain;
883 let neverHasSideEffects = 1 in
884 def VMOVSRR : AVConv5I<0b11000100, 0b1010,
885 (outs SPR:$dst1, SPR:$dst2), (ins GPR:$src1, GPR:$src2),
886 IIC_fpMOVID, "vmov", "\t$dst1, $dst2, $src1, $src2",
887 [/* For disassembly only; pattern left blank */]> {
888 // Instruction operands.
893 // Encode instruction operands.
894 let Inst{3-0} = dst1{4-1};
895 let Inst{5} = dst1{0};
896 let Inst{15-12} = src1;
897 let Inst{19-16} = src2;
899 let Inst{7-6} = 0b00;
901 // Some single precision VFP instructions may be executed on both NEON and VFP
903 let D = VFPNeonDomain;
905 let DecoderMethod = "DecodeVMOVSRR";
911 // FMRX: SPR system reg -> GPR
913 // FMXR: GPR -> VFP system reg
918 class AVConv1IDs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
919 bits<4> opcod4, dag oops, dag iops,
920 InstrItinClass itin, string opc, string asm,
922 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
924 // Instruction operands.
928 // Encode instruction operands.
929 let Inst{3-0} = Sm{4-1};
931 let Inst{15-12} = Dd{3-0};
932 let Inst{22} = Dd{4};
934 let Predicates = [HasVFP2, HasDPVFP];
937 class AVConv1InSs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
938 bits<4> opcod4, dag oops, dag iops,InstrItinClass itin,
939 string opc, string asm, list<dag> pattern>
940 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
942 // Instruction operands.
946 // Encode instruction operands.
947 let Inst{3-0} = Sm{4-1};
949 let Inst{15-12} = Sd{4-1};
950 let Inst{22} = Sd{0};
953 def VSITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
954 (outs DPR:$Dd), (ins SPR:$Sm),
955 IIC_fpCVTID, "vcvt", ".f64.s32\t$Dd, $Sm",
956 [(set DPR:$Dd, (f64 (arm_sitof SPR:$Sm)))]> {
957 let Inst{7} = 1; // s32
960 def VSITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
961 (outs SPR:$Sd),(ins SPR:$Sm),
962 IIC_fpCVTIS, "vcvt", ".f32.s32\t$Sd, $Sm",
963 [(set SPR:$Sd, (arm_sitof SPR:$Sm))]> {
964 let Inst{7} = 1; // s32
966 // Some single precision VFP instructions may be executed on both NEON and
967 // VFP pipelines on A8.
968 let D = VFPNeonA8Domain;
971 def VUITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
972 (outs DPR:$Dd), (ins SPR:$Sm),
973 IIC_fpCVTID, "vcvt", ".f64.u32\t$Dd, $Sm",
974 [(set DPR:$Dd, (f64 (arm_uitof SPR:$Sm)))]> {
975 let Inst{7} = 0; // u32
978 def VUITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
979 (outs SPR:$Sd), (ins SPR:$Sm),
980 IIC_fpCVTIS, "vcvt", ".f32.u32\t$Sd, $Sm",
981 [(set SPR:$Sd, (arm_uitof SPR:$Sm))]> {
982 let Inst{7} = 0; // u32
984 // Some single precision VFP instructions may be executed on both NEON and
985 // VFP pipelines on A8.
986 let D = VFPNeonA8Domain;
991 class AVConv1IsD_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
992 bits<4> opcod4, dag oops, dag iops,
993 InstrItinClass itin, string opc, string asm,
995 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
997 // Instruction operands.
1001 // Encode instruction operands.
1002 let Inst{3-0} = Dm{3-0};
1003 let Inst{5} = Dm{4};
1004 let Inst{15-12} = Sd{4-1};
1005 let Inst{22} = Sd{0};
1007 let Predicates = [HasVFP2, HasDPVFP];
1010 class AVConv1InsS_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1011 bits<4> opcod4, dag oops, dag iops,
1012 InstrItinClass itin, string opc, string asm,
1014 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1016 // Instruction operands.
1020 // Encode instruction operands.
1021 let Inst{3-0} = Sm{4-1};
1022 let Inst{5} = Sm{0};
1023 let Inst{15-12} = Sd{4-1};
1024 let Inst{22} = Sd{0};
1027 // Always set Z bit in the instruction, i.e. "round towards zero" variants.
1028 def VTOSIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
1029 (outs SPR:$Sd), (ins DPR:$Dm),
1030 IIC_fpCVTDI, "vcvt", ".s32.f64\t$Sd, $Dm",
1031 [(set SPR:$Sd, (arm_ftosi (f64 DPR:$Dm)))]> {
1032 let Inst{7} = 1; // Z bit
1035 def VTOSIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
1036 (outs SPR:$Sd), (ins SPR:$Sm),
1037 IIC_fpCVTSI, "vcvt", ".s32.f32\t$Sd, $Sm",
1038 [(set SPR:$Sd, (arm_ftosi SPR:$Sm))]> {
1039 let Inst{7} = 1; // Z bit
1041 // Some single precision VFP instructions may be executed on both NEON and
1042 // VFP pipelines on A8.
1043 let D = VFPNeonA8Domain;
1046 def VTOUIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
1047 (outs SPR:$Sd), (ins DPR:$Dm),
1048 IIC_fpCVTDI, "vcvt", ".u32.f64\t$Sd, $Dm",
1049 [(set SPR:$Sd, (arm_ftoui (f64 DPR:$Dm)))]> {
1050 let Inst{7} = 1; // Z bit
1053 def VTOUIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
1054 (outs SPR:$Sd), (ins SPR:$Sm),
1055 IIC_fpCVTSI, "vcvt", ".u32.f32\t$Sd, $Sm",
1056 [(set SPR:$Sd, (arm_ftoui SPR:$Sm))]> {
1057 let Inst{7} = 1; // Z bit
1059 // Some single precision VFP instructions may be executed on both NEON and
1060 // VFP pipelines on A8.
1061 let D = VFPNeonA8Domain;
1064 // And the Z bit '0' variants, i.e. use the rounding mode specified by FPSCR.
1065 let Uses = [FPSCR] in {
1066 // FIXME: Verify encoding after integrated assembler is working.
1067 def VTOSIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
1068 (outs SPR:$Sd), (ins DPR:$Dm),
1069 IIC_fpCVTDI, "vcvtr", ".s32.f64\t$Sd, $Dm",
1070 [(set SPR:$Sd, (int_arm_vcvtr (f64 DPR:$Dm)))]>{
1071 let Inst{7} = 0; // Z bit
1074 def VTOSIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
1075 (outs SPR:$Sd), (ins SPR:$Sm),
1076 IIC_fpCVTSI, "vcvtr", ".s32.f32\t$Sd, $Sm",
1077 [(set SPR:$Sd, (int_arm_vcvtr SPR:$Sm))]> {
1078 let Inst{7} = 0; // Z bit
1081 def VTOUIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
1082 (outs SPR:$Sd), (ins DPR:$Dm),
1083 IIC_fpCVTDI, "vcvtr", ".u32.f64\t$Sd, $Dm",
1084 [(set SPR:$Sd, (int_arm_vcvtru(f64 DPR:$Dm)))]>{
1085 let Inst{7} = 0; // Z bit
1088 def VTOUIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
1089 (outs SPR:$Sd), (ins SPR:$Sm),
1090 IIC_fpCVTSI, "vcvtr", ".u32.f32\t$Sd, $Sm",
1091 [(set SPR:$Sd, (int_arm_vcvtru SPR:$Sm))]> {
1092 let Inst{7} = 0; // Z bit
1096 // Convert between floating-point and fixed-point
1097 // Data type for fixed-point naming convention:
1098 // S16 (U=0, sx=0) -> SH
1099 // U16 (U=1, sx=0) -> UH
1100 // S32 (U=0, sx=1) -> SL
1101 // U32 (U=1, sx=1) -> UL
1103 let Constraints = "$a = $dst" in {
1105 // FP to Fixed-Point:
1107 // Single Precision register
1108 class AVConv1XInsS_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
1109 bit op5, dag oops, dag iops, InstrItinClass itin,
1110 string opc, string asm, list<dag> pattern>
1111 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern>,
1112 Sched<[WriteCvtFP]> {
1114 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
1115 let Inst{22} = dst{0};
1116 let Inst{15-12} = dst{4-1};
1119 // Double Precision register
1120 class AVConv1XInsD_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
1121 bit op5, dag oops, dag iops, InstrItinClass itin,
1122 string opc, string asm, list<dag> pattern>
1123 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern>,
1124 Sched<[WriteCvtFP]> {
1126 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
1127 let Inst{22} = dst{4};
1128 let Inst{15-12} = dst{3-0};
1130 let Predicates = [HasVFP2, HasDPVFP];
1133 def VTOSHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 0,
1134 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1135 IIC_fpCVTSI, "vcvt", ".s16.f32\t$dst, $a, $fbits", []> {
1136 // Some single precision VFP instructions may be executed on both NEON and
1137 // VFP pipelines on A8.
1138 let D = VFPNeonA8Domain;
1141 def VTOUHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 0,
1142 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1143 IIC_fpCVTSI, "vcvt", ".u16.f32\t$dst, $a, $fbits", []> {
1144 // Some single precision VFP instructions may be executed on both NEON and
1145 // VFP pipelines on A8.
1146 let D = VFPNeonA8Domain;
1149 def VTOSLS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 1,
1150 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1151 IIC_fpCVTSI, "vcvt", ".s32.f32\t$dst, $a, $fbits", []> {
1152 // Some single precision VFP instructions may be executed on both NEON and
1153 // VFP pipelines on A8.
1154 let D = VFPNeonA8Domain;
1157 def VTOULS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 1,
1158 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1159 IIC_fpCVTSI, "vcvt", ".u32.f32\t$dst, $a, $fbits", []> {
1160 // Some single precision VFP instructions may be executed on both NEON and
1161 // VFP pipelines on A8.
1162 let D = VFPNeonA8Domain;
1165 def VTOSHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 0,
1166 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1167 IIC_fpCVTDI, "vcvt", ".s16.f64\t$dst, $a, $fbits", []>;
1169 def VTOUHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 0,
1170 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1171 IIC_fpCVTDI, "vcvt", ".u16.f64\t$dst, $a, $fbits", []>;
1173 def VTOSLD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 1,
1174 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1175 IIC_fpCVTDI, "vcvt", ".s32.f64\t$dst, $a, $fbits", []>;
1177 def VTOULD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 1,
1178 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1179 IIC_fpCVTDI, "vcvt", ".u32.f64\t$dst, $a, $fbits", []>;
1181 // Fixed-Point to FP:
1183 def VSHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 0,
1184 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1185 IIC_fpCVTIS, "vcvt", ".f32.s16\t$dst, $a, $fbits", []> {
1186 // Some single precision VFP instructions may be executed on both NEON and
1187 // VFP pipelines on A8.
1188 let D = VFPNeonA8Domain;
1191 def VUHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 0,
1192 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1193 IIC_fpCVTIS, "vcvt", ".f32.u16\t$dst, $a, $fbits", []> {
1194 // Some single precision VFP instructions may be executed on both NEON and
1195 // VFP pipelines on A8.
1196 let D = VFPNeonA8Domain;
1199 def VSLTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 1,
1200 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1201 IIC_fpCVTIS, "vcvt", ".f32.s32\t$dst, $a, $fbits", []> {
1202 // Some single precision VFP instructions may be executed on both NEON and
1203 // VFP pipelines on A8.
1204 let D = VFPNeonA8Domain;
1207 def VULTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 1,
1208 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1209 IIC_fpCVTIS, "vcvt", ".f32.u32\t$dst, $a, $fbits", []> {
1210 // Some single precision VFP instructions may be executed on both NEON and
1211 // VFP pipelines on A8.
1212 let D = VFPNeonA8Domain;
1215 def VSHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 0,
1216 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1217 IIC_fpCVTID, "vcvt", ".f64.s16\t$dst, $a, $fbits", []>;
1219 def VUHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 0,
1220 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1221 IIC_fpCVTID, "vcvt", ".f64.u16\t$dst, $a, $fbits", []>;
1223 def VSLTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 1,
1224 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1225 IIC_fpCVTID, "vcvt", ".f64.s32\t$dst, $a, $fbits", []>;
1227 def VULTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 1,
1228 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1229 IIC_fpCVTID, "vcvt", ".f64.u32\t$dst, $a, $fbits", []>;
1231 } // End of 'let Constraints = "$a = $dst" in'
1233 //===----------------------------------------------------------------------===//
1234 // FP Multiply-Accumulate Operations.
1237 def VMLAD : ADbI<0b11100, 0b00, 0, 0,
1238 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1239 IIC_fpMAC64, "vmla", ".f64\t$Dd, $Dn, $Dm",
1240 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1241 (f64 DPR:$Ddin)))]>,
1242 RegConstraint<"$Ddin = $Dd">,
1243 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1245 def VMLAS : ASbIn<0b11100, 0b00, 0, 0,
1246 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1247 IIC_fpMAC32, "vmla", ".f32\t$Sd, $Sn, $Sm",
1248 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
1250 RegConstraint<"$Sdin = $Sd">,
1251 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1252 // Some single precision VFP instructions may be executed on both NEON and
1253 // VFP pipelines on A8.
1254 let D = VFPNeonA8Domain;
1257 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1258 (VMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1259 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1260 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1261 (VMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1262 Requires<[HasVFP2,DontUseNEONForFP, UseFPVMLx,DontUseFusedMAC]>;
1264 def VMLSD : ADbI<0b11100, 0b00, 1, 0,
1265 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1266 IIC_fpMAC64, "vmls", ".f64\t$Dd, $Dn, $Dm",
1267 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1268 (f64 DPR:$Ddin)))]>,
1269 RegConstraint<"$Ddin = $Dd">,
1270 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1272 def VMLSS : ASbIn<0b11100, 0b00, 1, 0,
1273 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1274 IIC_fpMAC32, "vmls", ".f32\t$Sd, $Sn, $Sm",
1275 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1277 RegConstraint<"$Sdin = $Sd">,
1278 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1279 // Some single precision VFP instructions may be executed on both NEON and
1280 // VFP pipelines on A8.
1281 let D = VFPNeonA8Domain;
1284 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1285 (VMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
1286 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1287 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1288 (VMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
1289 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1291 def VNMLAD : ADbI<0b11100, 0b01, 1, 0,
1292 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1293 IIC_fpMAC64, "vnmla", ".f64\t$Dd, $Dn, $Dm",
1294 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1295 (f64 DPR:$Ddin)))]>,
1296 RegConstraint<"$Ddin = $Dd">,
1297 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1299 def VNMLAS : ASbI<0b11100, 0b01, 1, 0,
1300 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1301 IIC_fpMAC32, "vnmla", ".f32\t$Sd, $Sn, $Sm",
1302 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1304 RegConstraint<"$Sdin = $Sd">,
1305 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1306 // Some single precision VFP instructions may be executed on both NEON and
1307 // VFP pipelines on A8.
1308 let D = VFPNeonA8Domain;
1311 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
1312 (VNMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1313 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1314 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
1315 (VNMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1316 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1318 def VNMLSD : ADbI<0b11100, 0b01, 0, 0,
1319 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1320 IIC_fpMAC64, "vnmls", ".f64\t$Dd, $Dn, $Dm",
1321 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1322 (f64 DPR:$Ddin)))]>,
1323 RegConstraint<"$Ddin = $Dd">,
1324 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1326 def VNMLSS : ASbI<0b11100, 0b01, 0, 0,
1327 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1328 IIC_fpMAC32, "vnmls", ".f32\t$Sd, $Sn, $Sm",
1329 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
1330 RegConstraint<"$Sdin = $Sd">,
1331 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1332 // Some single precision VFP instructions may be executed on both NEON and
1333 // VFP pipelines on A8.
1334 let D = VFPNeonA8Domain;
1337 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
1338 (VNMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
1339 Requires<[HasVFP2,HasDPVFP,UseFPVMLx,DontUseFusedMAC]>;
1340 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
1341 (VNMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
1342 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1344 //===----------------------------------------------------------------------===//
1345 // Fused FP Multiply-Accumulate Operations.
1347 def VFMAD : ADbI<0b11101, 0b10, 0, 0,
1348 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1349 IIC_fpFMAC64, "vfma", ".f64\t$Dd, $Dn, $Dm",
1350 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1351 (f64 DPR:$Ddin)))]>,
1352 RegConstraint<"$Ddin = $Dd">,
1353 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1355 def VFMAS : ASbIn<0b11101, 0b10, 0, 0,
1356 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1357 IIC_fpFMAC32, "vfma", ".f32\t$Sd, $Sn, $Sm",
1358 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
1360 RegConstraint<"$Sdin = $Sd">,
1361 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1362 // Some single precision VFP instructions may be executed on both NEON and
1366 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1367 (VFMAD DPR:$dstin, DPR:$a, DPR:$b)>,
1368 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1369 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1370 (VFMAS SPR:$dstin, SPR:$a, SPR:$b)>,
1371 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1373 // Match @llvm.fma.* intrinsics
1374 // (fma x, y, z) -> (vfms z, x, y)
1375 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, DPR:$Ddin)),
1376 (VFMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1377 Requires<[HasVFP4,HasDPVFP]>;
1378 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, SPR:$Sdin)),
1379 (VFMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1380 Requires<[HasVFP4]>;
1382 def VFMSD : ADbI<0b11101, 0b10, 1, 0,
1383 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1384 IIC_fpFMAC64, "vfms", ".f64\t$Dd, $Dn, $Dm",
1385 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1386 (f64 DPR:$Ddin)))]>,
1387 RegConstraint<"$Ddin = $Dd">,
1388 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1390 def VFMSS : ASbIn<0b11101, 0b10, 1, 0,
1391 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1392 IIC_fpFMAC32, "vfms", ".f32\t$Sd, $Sn, $Sm",
1393 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1395 RegConstraint<"$Sdin = $Sd">,
1396 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1397 // Some single precision VFP instructions may be executed on both NEON and
1401 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1402 (VFMSD DPR:$dstin, DPR:$a, DPR:$b)>,
1403 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1404 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1405 (VFMSS SPR:$dstin, SPR:$a, SPR:$b)>,
1406 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1408 // Match @llvm.fma.* intrinsics
1409 // (fma (fneg x), y, z) -> (vfms z, x, y)
1410 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin)),
1411 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1412 Requires<[HasVFP4,HasDPVFP]>;
1413 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin)),
1414 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1415 Requires<[HasVFP4]>;
1416 // (fma x, (fneg y), z) -> (vfms z, x, y)
1417 def : Pat<(f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin)),
1418 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1419 Requires<[HasVFP4,HasDPVFP]>;
1420 def : Pat<(f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin)),
1421 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1422 Requires<[HasVFP4]>;
1424 def VFNMAD : ADbI<0b11101, 0b01, 1, 0,
1425 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1426 IIC_fpFMAC64, "vfnma", ".f64\t$Dd, $Dn, $Dm",
1427 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1428 (f64 DPR:$Ddin)))]>,
1429 RegConstraint<"$Ddin = $Dd">,
1430 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1432 def VFNMAS : ASbI<0b11101, 0b01, 1, 0,
1433 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1434 IIC_fpFMAC32, "vfnma", ".f32\t$Sd, $Sn, $Sm",
1435 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1437 RegConstraint<"$Sdin = $Sd">,
1438 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1439 // Some single precision VFP instructions may be executed on both NEON and
1443 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
1444 (VFNMAD DPR:$dstin, DPR:$a, DPR:$b)>,
1445 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1446 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
1447 (VFNMAS SPR:$dstin, SPR:$a, SPR:$b)>,
1448 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1450 // Match @llvm.fma.* intrinsics
1451 // (fneg (fma x, y, z)) -> (vfnma z, x, y)
1452 def : Pat<(fneg (fma (f64 DPR:$Dn), (f64 DPR:$Dm), (f64 DPR:$Ddin))),
1453 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1454 Requires<[HasVFP4,HasDPVFP]>;
1455 def : Pat<(fneg (fma (f32 SPR:$Sn), (f32 SPR:$Sm), (f32 SPR:$Sdin))),
1456 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1457 Requires<[HasVFP4]>;
1458 // (fma (fneg x), y, (fneg z)) -> (vfnma z, x, y)
1459 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, (fneg DPR:$Ddin))),
1460 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1461 Requires<[HasVFP4,HasDPVFP]>;
1462 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, (fneg SPR:$Sdin))),
1463 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1464 Requires<[HasVFP4]>;
1466 def VFNMSD : ADbI<0b11101, 0b01, 0, 0,
1467 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1468 IIC_fpFMAC64, "vfnms", ".f64\t$Dd, $Dn, $Dm",
1469 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1470 (f64 DPR:$Ddin)))]>,
1471 RegConstraint<"$Ddin = $Dd">,
1472 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1474 def VFNMSS : ASbI<0b11101, 0b01, 0, 0,
1475 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1476 IIC_fpFMAC32, "vfnms", ".f32\t$Sd, $Sn, $Sm",
1477 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
1478 RegConstraint<"$Sdin = $Sd">,
1479 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1480 // Some single precision VFP instructions may be executed on both NEON and
1484 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
1485 (VFNMSD DPR:$dstin, DPR:$a, DPR:$b)>,
1486 Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
1487 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
1488 (VFNMSS SPR:$dstin, SPR:$a, SPR:$b)>,
1489 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1491 // Match @llvm.fma.* intrinsics
1493 // (fma x, y, (fneg z)) -> (vfnms z, x, y))
1494 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, (fneg DPR:$Ddin))),
1495 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1496 Requires<[HasVFP4,HasDPVFP]>;
1497 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, (fneg SPR:$Sdin))),
1498 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1499 Requires<[HasVFP4]>;
1500 // (fneg (fma (fneg x), y, z)) -> (vfnms z, x, y)
1501 def : Pat<(fneg (f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin))),
1502 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1503 Requires<[HasVFP4,HasDPVFP]>;
1504 def : Pat<(fneg (f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin))),
1505 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1506 Requires<[HasVFP4]>;
1507 // (fneg (fma x, (fneg y), z) -> (vfnms z, x, y)
1508 def : Pat<(fneg (f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin))),
1509 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1510 Requires<[HasVFP4,HasDPVFP]>;
1511 def : Pat<(fneg (f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin))),
1512 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1513 Requires<[HasVFP4]>;
1515 //===----------------------------------------------------------------------===//
1516 // FP Conditional moves.
1519 let neverHasSideEffects = 1 in {
1520 def VMOVDcc : PseudoInst<(outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm, cmovpred:$p),
1522 [(set (f64 DPR:$Dd),
1523 (ARMcmov DPR:$Dn, DPR:$Dm, cmovpred:$p))]>,
1524 RegConstraint<"$Dn = $Dd">, Requires<[HasVFP2,HasDPVFP]>;
1526 def VMOVScc : PseudoInst<(outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm, cmovpred:$p),
1528 [(set (f32 SPR:$Sd),
1529 (ARMcmov SPR:$Sn, SPR:$Sm, cmovpred:$p))]>,
1530 RegConstraint<"$Sn = $Sd">, Requires<[HasVFP2]>;
1531 } // neverHasSideEffects
1533 //===----------------------------------------------------------------------===//
1534 // Move from VFP System Register to ARM core register.
1537 class MovFromVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
1539 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
1541 // Instruction operand.
1544 let Inst{27-20} = 0b11101111;
1545 let Inst{19-16} = opc19_16;
1546 let Inst{15-12} = Rt;
1547 let Inst{11-8} = 0b1010;
1549 let Inst{6-5} = 0b00;
1551 let Inst{3-0} = 0b0000;
1554 // APSR is the application level alias of CPSR. This FPSCR N, Z, C, V flags
1556 let Defs = [CPSR], Uses = [FPSCR_NZCV], Rt = 0b1111 /* apsr_nzcv */ in
1557 def FMSTAT : MovFromVFP<0b0001 /* fpscr */, (outs), (ins),
1558 "vmrs", "\tAPSR_nzcv, fpscr", [(arm_fmstat)]>;
1560 // Application level FPSCR -> GPR
1561 let hasSideEffects = 1, Uses = [FPSCR] in
1562 def VMRS : MovFromVFP<0b0001 /* fpscr */, (outs GPR:$Rt), (ins),
1563 "vmrs", "\t$Rt, fpscr",
1564 [(set GPR:$Rt, (int_arm_get_fpscr))]>;
1566 // System level FPEXC, FPSID -> GPR
1567 let Uses = [FPSCR] in {
1568 def VMRS_FPEXC : MovFromVFP<0b1000 /* fpexc */, (outs GPR:$Rt), (ins),
1569 "vmrs", "\t$Rt, fpexc", []>;
1570 def VMRS_FPSID : MovFromVFP<0b0000 /* fpsid */, (outs GPR:$Rt), (ins),
1571 "vmrs", "\t$Rt, fpsid", []>;
1572 def VMRS_MVFR0 : MovFromVFP<0b0111 /* mvfr0 */, (outs GPR:$Rt), (ins),
1573 "vmrs", "\t$Rt, mvfr0", []>;
1574 def VMRS_MVFR1 : MovFromVFP<0b0110 /* mvfr1 */, (outs GPR:$Rt), (ins),
1575 "vmrs", "\t$Rt, mvfr1", []>;
1576 def VMRS_MVFR2 : MovFromVFP<0b0101 /* mvfr2 */, (outs GPR:$Rt), (ins),
1577 "vmrs", "\t$Rt, mvfr2", []>, Requires<[HasFPARMv8]>;
1578 def VMRS_FPINST : MovFromVFP<0b1001 /* fpinst */, (outs GPR:$Rt), (ins),
1579 "vmrs", "\t$Rt, fpinst", []>;
1580 def VMRS_FPINST2 : MovFromVFP<0b1010 /* fpinst2 */, (outs GPR:$Rt), (ins),
1581 "vmrs", "\t$Rt, fpinst2", []>;
1584 //===----------------------------------------------------------------------===//
1585 // Move from ARM core register to VFP System Register.
1588 class MovToVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
1590 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
1592 // Instruction operand.
1595 // Encode instruction operand.
1596 let Inst{15-12} = src;
1598 let Inst{27-20} = 0b11101110;
1599 let Inst{19-16} = opc19_16;
1600 let Inst{11-8} = 0b1010;
1605 let Defs = [FPSCR] in {
1606 // Application level GPR -> FPSCR
1607 def VMSR : MovToVFP<0b0001 /* fpscr */, (outs), (ins GPR:$src),
1608 "vmsr", "\tfpscr, $src", [(int_arm_set_fpscr GPR:$src)]>;
1609 // System level GPR -> FPEXC
1610 def VMSR_FPEXC : MovToVFP<0b1000 /* fpexc */, (outs), (ins GPR:$src),
1611 "vmsr", "\tfpexc, $src", []>;
1612 // System level GPR -> FPSID
1613 def VMSR_FPSID : MovToVFP<0b0000 /* fpsid */, (outs), (ins GPR:$src),
1614 "vmsr", "\tfpsid, $src", []>;
1616 def VMSR_FPINST : MovToVFP<0b1001 /* fpinst */, (outs), (ins GPR:$src),
1617 "vmsr", "\tfpinst, $src", []>;
1618 def VMSR_FPINST2 : MovToVFP<0b1010 /* fpinst2 */, (outs), (ins GPR:$src),
1619 "vmsr", "\tfpinst2, $src", []>;
1622 //===----------------------------------------------------------------------===//
1626 // Materialize FP immediates. VFP3 only.
1627 let isReMaterializable = 1 in {
1628 def FCONSTD : VFPAI<(outs DPR:$Dd), (ins vfp_f64imm:$imm),
1629 VFPMiscFrm, IIC_fpUNA64,
1630 "vmov", ".f64\t$Dd, $imm",
1631 [(set DPR:$Dd, vfp_f64imm:$imm)]>,
1632 Requires<[HasVFP3,HasDPVFP]> {
1636 let Inst{27-23} = 0b11101;
1637 let Inst{22} = Dd{4};
1638 let Inst{21-20} = 0b11;
1639 let Inst{19-16} = imm{7-4};
1640 let Inst{15-12} = Dd{3-0};
1641 let Inst{11-9} = 0b101;
1642 let Inst{8} = 1; // Double precision.
1643 let Inst{7-4} = 0b0000;
1644 let Inst{3-0} = imm{3-0};
1647 def FCONSTS : VFPAI<(outs SPR:$Sd), (ins vfp_f32imm:$imm),
1648 VFPMiscFrm, IIC_fpUNA32,
1649 "vmov", ".f32\t$Sd, $imm",
1650 [(set SPR:$Sd, vfp_f32imm:$imm)]>, Requires<[HasVFP3]> {
1654 let Inst{27-23} = 0b11101;
1655 let Inst{22} = Sd{0};
1656 let Inst{21-20} = 0b11;
1657 let Inst{19-16} = imm{7-4};
1658 let Inst{15-12} = Sd{4-1};
1659 let Inst{11-9} = 0b101;
1660 let Inst{8} = 0; // Single precision.
1661 let Inst{7-4} = 0b0000;
1662 let Inst{3-0} = imm{3-0};
1666 //===----------------------------------------------------------------------===//
1667 // Assembler aliases.
1669 // A few mnemonic aliases for pre-unifixed syntax. We don't guarantee to
1670 // support them all, but supporting at least some of the basics is
1671 // good to be friendly.
1672 def : VFP2MnemonicAlias<"flds", "vldr">;
1673 def : VFP2MnemonicAlias<"fldd", "vldr">;
1674 def : VFP2MnemonicAlias<"fmrs", "vmov">;
1675 def : VFP2MnemonicAlias<"fmsr", "vmov">;
1676 def : VFP2MnemonicAlias<"fsqrts", "vsqrt">;
1677 def : VFP2MnemonicAlias<"fsqrtd", "vsqrt">;
1678 def : VFP2MnemonicAlias<"fadds", "vadd.f32">;
1679 def : VFP2MnemonicAlias<"faddd", "vadd.f64">;
1680 def : VFP2MnemonicAlias<"fmrdd", "vmov">;
1681 def : VFP2MnemonicAlias<"fmrds", "vmov">;
1682 def : VFP2MnemonicAlias<"fmrrd", "vmov">;
1683 def : VFP2MnemonicAlias<"fmdrr", "vmov">;
1684 def : VFP2MnemonicAlias<"fmuls", "vmul.f32">;
1685 def : VFP2MnemonicAlias<"fmuld", "vmul.f64">;
1686 def : VFP2MnemonicAlias<"fnegs", "vneg.f32">;
1687 def : VFP2MnemonicAlias<"fnegd", "vneg.f64">;
1688 def : VFP2MnemonicAlias<"ftosizd", "vcvt.s32.f64">;
1689 def : VFP2MnemonicAlias<"ftosid", "vcvtr.s32.f64">;
1690 def : VFP2MnemonicAlias<"ftosizs", "vcvt.s32.f32">;
1691 def : VFP2MnemonicAlias<"ftosis", "vcvtr.s32.f32">;
1692 def : VFP2MnemonicAlias<"ftouizd", "vcvt.u32.f64">;
1693 def : VFP2MnemonicAlias<"ftouid", "vcvtr.u32.f64">;
1694 def : VFP2MnemonicAlias<"ftouizs", "vcvt.u32.f32">;
1695 def : VFP2MnemonicAlias<"ftouis", "vcvtr.u32.f32">;
1696 def : VFP2MnemonicAlias<"fsitod", "vcvt.f64.s32">;
1697 def : VFP2MnemonicAlias<"fsitos", "vcvt.f32.s32">;
1698 def : VFP2MnemonicAlias<"fuitod", "vcvt.f64.u32">;
1699 def : VFP2MnemonicAlias<"fuitos", "vcvt.f32.u32">;
1700 def : VFP2MnemonicAlias<"fsts", "vstr">;
1701 def : VFP2MnemonicAlias<"fstd", "vstr">;
1702 def : VFP2MnemonicAlias<"fmacd", "vmla.f64">;
1703 def : VFP2MnemonicAlias<"fmacs", "vmla.f32">;
1704 def : VFP2MnemonicAlias<"fcpys", "vmov.f32">;
1705 def : VFP2MnemonicAlias<"fcpyd", "vmov.f64">;
1706 def : VFP2MnemonicAlias<"fcmps", "vcmp.f32">;
1707 def : VFP2MnemonicAlias<"fcmpd", "vcmp.f64">;
1708 def : VFP2MnemonicAlias<"fdivs", "vdiv.f32">;
1709 def : VFP2MnemonicAlias<"fdivd", "vdiv.f64">;
1710 def : VFP2MnemonicAlias<"fmrx", "vmrs">;
1711 def : VFP2MnemonicAlias<"fmxr", "vmsr">;
1713 // Be friendly and accept the old form of zero-compare
1714 def : VFP2DPInstAlias<"fcmpzd${p} $val", (VCMPZD DPR:$val, pred:$p)>;
1715 def : VFP2InstAlias<"fcmpzs${p} $val", (VCMPZS SPR:$val, pred:$p)>;
1718 def : VFP2InstAlias<"fmstat${p}", (FMSTAT pred:$p)>;
1719 def : VFP2InstAlias<"fadds${p} $Sd, $Sn, $Sm",
1720 (VADDS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
1721 def : VFP2DPInstAlias<"faddd${p} $Dd, $Dn, $Dm",
1722 (VADDD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
1723 def : VFP2InstAlias<"fsubs${p} $Sd, $Sn, $Sm",
1724 (VSUBS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
1725 def : VFP2DPInstAlias<"fsubd${p} $Dd, $Dn, $Dm",
1726 (VSUBD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
1728 // No need for the size suffix on VSQRT. It's implied by the register classes.
1729 def : VFP2InstAlias<"vsqrt${p} $Sd, $Sm", (VSQRTS SPR:$Sd, SPR:$Sm, pred:$p)>;
1730 def : VFP2DPInstAlias<"vsqrt${p} $Dd, $Dm", (VSQRTD DPR:$Dd, DPR:$Dm, pred:$p)>;
1732 // VLDR/VSTR accept an optional type suffix.
1733 def : VFP2InstAlias<"vldr${p}.32 $Sd, $addr",
1734 (VLDRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
1735 def : VFP2InstAlias<"vstr${p}.32 $Sd, $addr",
1736 (VSTRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
1737 def : VFP2InstAlias<"vldr${p}.64 $Dd, $addr",
1738 (VLDRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
1739 def : VFP2InstAlias<"vstr${p}.64 $Dd, $addr",
1740 (VSTRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
1742 // VMOV can accept optional 32-bit or less data type suffix suffix.
1743 def : VFP2InstAlias<"vmov${p}.8 $Rt, $Sn",
1744 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1745 def : VFP2InstAlias<"vmov${p}.16 $Rt, $Sn",
1746 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1747 def : VFP2InstAlias<"vmov${p}.32 $Rt, $Sn",
1748 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1749 def : VFP2InstAlias<"vmov${p}.8 $Sn, $Rt",
1750 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1751 def : VFP2InstAlias<"vmov${p}.16 $Sn, $Rt",
1752 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1753 def : VFP2InstAlias<"vmov${p}.32 $Sn, $Rt",
1754 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1756 def : VFP2InstAlias<"vmov${p}.f64 $Rt, $Rt2, $Dn",
1757 (VMOVRRD GPR:$Rt, GPR:$Rt2, DPR:$Dn, pred:$p)>;
1758 def : VFP2InstAlias<"vmov${p}.f64 $Dn, $Rt, $Rt2",
1759 (VMOVDRR DPR:$Dn, GPR:$Rt, GPR:$Rt2, pred:$p)>;
1761 // VMOVS doesn't need the .f32 to disambiguate from the NEON encoding the way
1763 def : VFP2InstAlias<"vmov${p} $Sd, $Sm",
1764 (VMOVS SPR:$Sd, SPR:$Sm, pred:$p)>;
1766 // FCONSTD/FCONSTS alias for vmov.f64/vmov.f32
1767 // These aliases provide added functionality over vmov.f instructions by
1768 // allowing users to write assembly containing encoded floating point constants
1769 // (e.g. #0x70 vs #1.0). Without these alises there is no way for the
1770 // assembler to accept encoded fp constants (but the equivalent fp-literal is
1771 // accepted directly by vmovf).
1772 def : VFP3InstAlias<"fconstd${p} $Dd, $val",
1773 (FCONSTD DPR:$Dd, vfp_f64imm:$val, pred:$p)>;
1774 def : VFP3InstAlias<"fconsts${p} $Sd, $val",
1775 (FCONSTS SPR:$Sd, vfp_f32imm:$val, pred:$p)>;