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 = $wb", []> {
187 let Inst{24-23} = 0b10; // Decrement Before
188 let Inst{21} = 1; // Writeback
189 let Inst{20} = L_bit;
191 // Some single precision VFP instructions may be executed on both NEON and
193 let D = VFPNeonDomain;
197 let neverHasSideEffects = 1 in {
199 let mayLoad = 1, hasExtraDefRegAllocReq = 1 in
200 defm VLDM : vfp_ldst_mult<"vldm", 1, IIC_fpLoad_m, IIC_fpLoad_mu>;
202 let mayStore = 1, hasExtraSrcRegAllocReq = 1 in
203 defm VSTM : vfp_ldst_mult<"vstm", 0, IIC_fpLoad_m, IIC_fpLoad_mu>;
205 } // neverHasSideEffects
207 def : MnemonicAlias<"vldm", "vldmia">;
208 def : MnemonicAlias<"vstm", "vstmia">;
210 def : InstAlias<"vpush${p} $r", (VSTMDDB_UPD SP, pred:$p, dpr_reglist:$r)>,
212 def : InstAlias<"vpush${p} $r", (VSTMSDB_UPD SP, pred:$p, spr_reglist:$r)>,
214 def : InstAlias<"vpop${p} $r", (VLDMDIA_UPD SP, pred:$p, dpr_reglist:$r)>,
216 def : InstAlias<"vpop${p} $r", (VLDMSIA_UPD SP, pred:$p, spr_reglist:$r)>,
218 defm : VFPDTAnyInstAlias<"vpush${p}", "$r",
219 (VSTMSDB_UPD SP, pred:$p, spr_reglist:$r)>;
220 defm : VFPDTAnyInstAlias<"vpush${p}", "$r",
221 (VSTMDDB_UPD SP, pred:$p, dpr_reglist:$r)>;
222 defm : VFPDTAnyInstAlias<"vpop${p}", "$r",
223 (VLDMSIA_UPD SP, pred:$p, spr_reglist:$r)>;
224 defm : VFPDTAnyInstAlias<"vpop${p}", "$r",
225 (VLDMDIA_UPD SP, pred:$p, dpr_reglist:$r)>;
227 // FLDMX, FSTMX - Load and store multiple unknown precision registers for
229 // These instruction are deprecated so we don't want them to get selected.
230 multiclass vfp_ldstx_mult<string asm, bit L_bit> {
233 AXXI4<(outs), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
234 IndexModeNone, !strconcat(asm, "iax${p}\t$Rn, $regs"), "", []> {
235 let Inst{24-23} = 0b01; // Increment After
236 let Inst{21} = 0; // No writeback
237 let Inst{20} = L_bit;
240 AXXI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
241 IndexModeUpd, !strconcat(asm, "iax${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
242 let Inst{24-23} = 0b01; // Increment After
243 let Inst{21} = 1; // Writeback
244 let Inst{20} = L_bit;
247 AXXI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
248 IndexModeUpd, !strconcat(asm, "dbx${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
249 let Inst{24-23} = 0b10; // Decrement Before
251 let Inst{20} = L_bit;
255 defm FLDM : vfp_ldstx_mult<"fldm", 1>;
256 defm FSTM : vfp_ldstx_mult<"fstm", 0>;
258 //===----------------------------------------------------------------------===//
259 // FP Binary Operations.
262 let TwoOperandAliasConstraint = "$Dn = $Dd" in
263 def VADDD : ADbI<0b11100, 0b11, 0, 0,
264 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
265 IIC_fpALU64, "vadd", ".f64\t$Dd, $Dn, $Dm",
266 [(set DPR:$Dd, (fadd DPR:$Dn, (f64 DPR:$Dm)))]>;
268 let TwoOperandAliasConstraint = "$Sn = $Sd" in
269 def VADDS : ASbIn<0b11100, 0b11, 0, 0,
270 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
271 IIC_fpALU32, "vadd", ".f32\t$Sd, $Sn, $Sm",
272 [(set SPR:$Sd, (fadd SPR:$Sn, SPR:$Sm))]> {
273 // Some single precision VFP instructions may be executed on both NEON and
274 // VFP pipelines on A8.
275 let D = VFPNeonA8Domain;
278 let TwoOperandAliasConstraint = "$Dn = $Dd" in
279 def VSUBD : ADbI<0b11100, 0b11, 1, 0,
280 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
281 IIC_fpALU64, "vsub", ".f64\t$Dd, $Dn, $Dm",
282 [(set DPR:$Dd, (fsub DPR:$Dn, (f64 DPR:$Dm)))]>;
284 let TwoOperandAliasConstraint = "$Sn = $Sd" in
285 def VSUBS : ASbIn<0b11100, 0b11, 1, 0,
286 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
287 IIC_fpALU32, "vsub", ".f32\t$Sd, $Sn, $Sm",
288 [(set SPR:$Sd, (fsub SPR:$Sn, SPR:$Sm))]> {
289 // Some single precision VFP instructions may be executed on both NEON and
290 // VFP pipelines on A8.
291 let D = VFPNeonA8Domain;
294 let TwoOperandAliasConstraint = "$Dn = $Dd" in
295 def VDIVD : ADbI<0b11101, 0b00, 0, 0,
296 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
297 IIC_fpDIV64, "vdiv", ".f64\t$Dd, $Dn, $Dm",
298 [(set DPR:$Dd, (fdiv DPR:$Dn, (f64 DPR:$Dm)))]>;
300 let TwoOperandAliasConstraint = "$Sn = $Sd" in
301 def VDIVS : ASbI<0b11101, 0b00, 0, 0,
302 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
303 IIC_fpDIV32, "vdiv", ".f32\t$Sd, $Sn, $Sm",
304 [(set SPR:$Sd, (fdiv SPR:$Sn, SPR:$Sm))]>;
306 let TwoOperandAliasConstraint = "$Dn = $Dd" in
307 def VMULD : ADbI<0b11100, 0b10, 0, 0,
308 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
309 IIC_fpMUL64, "vmul", ".f64\t$Dd, $Dn, $Dm",
310 [(set DPR:$Dd, (fmul DPR:$Dn, (f64 DPR:$Dm)))]>;
312 let TwoOperandAliasConstraint = "$Sn = $Sd" in
313 def VMULS : ASbIn<0b11100, 0b10, 0, 0,
314 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
315 IIC_fpMUL32, "vmul", ".f32\t$Sd, $Sn, $Sm",
316 [(set SPR:$Sd, (fmul SPR:$Sn, SPR:$Sm))]> {
317 // Some single precision VFP instructions may be executed on both NEON and
318 // VFP pipelines on A8.
319 let D = VFPNeonA8Domain;
322 def VNMULD : ADbI<0b11100, 0b10, 1, 0,
323 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
324 IIC_fpMUL64, "vnmul", ".f64\t$Dd, $Dn, $Dm",
325 [(set DPR:$Dd, (fneg (fmul DPR:$Dn, (f64 DPR:$Dm))))]>;
327 def VNMULS : ASbI<0b11100, 0b10, 1, 0,
328 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
329 IIC_fpMUL32, "vnmul", ".f32\t$Sd, $Sn, $Sm",
330 [(set SPR:$Sd, (fneg (fmul SPR:$Sn, SPR:$Sm)))]> {
331 // Some single precision VFP instructions may be executed on both NEON and
332 // VFP pipelines on A8.
333 let D = VFPNeonA8Domain;
336 multiclass vsel_inst<string op, bits<2> opc, int CC> {
337 let DecoderNamespace = "VFPV8", PostEncoderMethod = "",
338 Uses = [CPSR], AddedComplexity = 4 in {
339 def S : ASbInp<0b11100, opc, 0,
340 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
341 NoItinerary, !strconcat("vsel", op, ".f32\t$Sd, $Sn, $Sm"),
342 [(set SPR:$Sd, (ARMcmov SPR:$Sm, SPR:$Sn, CC))]>,
345 def D : ADbInp<0b11100, opc, 0,
346 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
347 NoItinerary, !strconcat("vsel", op, ".f64\t$Dd, $Dn, $Dm"),
348 [(set DPR:$Dd, (ARMcmov (f64 DPR:$Dm), (f64 DPR:$Dn), CC))]>,
353 // The CC constants here match ARMCC::CondCodes.
354 defm VSELGT : vsel_inst<"gt", 0b11, 12>;
355 defm VSELGE : vsel_inst<"ge", 0b10, 10>;
356 defm VSELEQ : vsel_inst<"eq", 0b00, 0>;
357 defm VSELVS : vsel_inst<"vs", 0b01, 6>;
359 multiclass vmaxmin_inst<string op, bit opc> {
360 let DecoderNamespace = "VFPV8", PostEncoderMethod = "" in {
361 def S : ASbInp<0b11101, 0b00, opc,
362 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
363 NoItinerary, !strconcat(op, ".f32\t$Sd, $Sn, $Sm"),
364 []>, Requires<[HasV8FP]>;
366 def D : ADbInp<0b11101, 0b00, opc,
367 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
368 NoItinerary, !strconcat(op, ".f64\t$Dd, $Dn, $Dm"),
369 []>, Requires<[HasV8FP]>;
373 defm VMAXNM : vmaxmin_inst<"vmaxnm", 0>;
374 defm VMINNM : vmaxmin_inst<"vminnm", 1>;
376 // Match reassociated forms only if not sign dependent rounding.
377 def : Pat<(fmul (fneg DPR:$a), (f64 DPR:$b)),
378 (VNMULD DPR:$a, DPR:$b)>, Requires<[NoHonorSignDependentRounding]>;
379 def : Pat<(fmul (fneg SPR:$a), SPR:$b),
380 (VNMULS SPR:$a, SPR:$b)>, Requires<[NoHonorSignDependentRounding]>;
382 // These are encoded as unary instructions.
383 let Defs = [FPSCR_NZCV] in {
384 def VCMPED : ADuI<0b11101, 0b11, 0b0100, 0b11, 0,
385 (outs), (ins DPR:$Dd, DPR:$Dm),
386 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, $Dm",
387 [(arm_cmpfp DPR:$Dd, (f64 DPR:$Dm))]>;
389 def VCMPES : ASuI<0b11101, 0b11, 0b0100, 0b11, 0,
390 (outs), (ins SPR:$Sd, SPR:$Sm),
391 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, $Sm",
392 [(arm_cmpfp SPR:$Sd, SPR:$Sm)]> {
393 // Some single precision VFP instructions may be executed on both NEON and
394 // VFP pipelines on A8.
395 let D = VFPNeonA8Domain;
398 // FIXME: Verify encoding after integrated assembler is working.
399 def VCMPD : ADuI<0b11101, 0b11, 0b0100, 0b01, 0,
400 (outs), (ins DPR:$Dd, DPR:$Dm),
401 IIC_fpCMP64, "vcmp", ".f64\t$Dd, $Dm",
402 [/* For disassembly only; pattern left blank */]>;
404 def VCMPS : ASuI<0b11101, 0b11, 0b0100, 0b01, 0,
405 (outs), (ins SPR:$Sd, SPR:$Sm),
406 IIC_fpCMP32, "vcmp", ".f32\t$Sd, $Sm",
407 [/* For disassembly only; pattern left blank */]> {
408 // Some single precision VFP instructions may be executed on both NEON and
409 // VFP pipelines on A8.
410 let D = VFPNeonA8Domain;
412 } // Defs = [FPSCR_NZCV]
414 //===----------------------------------------------------------------------===//
415 // FP Unary Operations.
418 def VABSD : ADuI<0b11101, 0b11, 0b0000, 0b11, 0,
419 (outs DPR:$Dd), (ins DPR:$Dm),
420 IIC_fpUNA64, "vabs", ".f64\t$Dd, $Dm",
421 [(set DPR:$Dd, (fabs (f64 DPR:$Dm)))]>;
423 def VABSS : ASuIn<0b11101, 0b11, 0b0000, 0b11, 0,
424 (outs SPR:$Sd), (ins SPR:$Sm),
425 IIC_fpUNA32, "vabs", ".f32\t$Sd, $Sm",
426 [(set SPR:$Sd, (fabs SPR:$Sm))]> {
427 // Some single precision VFP instructions may be executed on both NEON and
428 // VFP pipelines on A8.
429 let D = VFPNeonA8Domain;
432 let Defs = [FPSCR_NZCV] in {
433 def VCMPEZD : ADuI<0b11101, 0b11, 0b0101, 0b11, 0,
434 (outs), (ins DPR:$Dd),
435 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, #0",
436 [(arm_cmpfp0 (f64 DPR:$Dd))]> {
437 let Inst{3-0} = 0b0000;
441 def VCMPEZS : ASuI<0b11101, 0b11, 0b0101, 0b11, 0,
442 (outs), (ins SPR:$Sd),
443 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, #0",
444 [(arm_cmpfp0 SPR:$Sd)]> {
445 let Inst{3-0} = 0b0000;
448 // Some single precision VFP instructions may be executed on both NEON and
449 // VFP pipelines on A8.
450 let D = VFPNeonA8Domain;
453 // FIXME: Verify encoding after integrated assembler is working.
454 def VCMPZD : ADuI<0b11101, 0b11, 0b0101, 0b01, 0,
455 (outs), (ins DPR:$Dd),
456 IIC_fpCMP64, "vcmp", ".f64\t$Dd, #0",
457 [/* For disassembly only; pattern left blank */]> {
458 let Inst{3-0} = 0b0000;
462 def VCMPZS : ASuI<0b11101, 0b11, 0b0101, 0b01, 0,
463 (outs), (ins SPR:$Sd),
464 IIC_fpCMP32, "vcmp", ".f32\t$Sd, #0",
465 [/* For disassembly only; pattern left blank */]> {
466 let Inst{3-0} = 0b0000;
469 // Some single precision VFP instructions may be executed on both NEON and
470 // VFP pipelines on A8.
471 let D = VFPNeonA8Domain;
473 } // Defs = [FPSCR_NZCV]
475 def VCVTDS : ASuI<0b11101, 0b11, 0b0111, 0b11, 0,
476 (outs DPR:$Dd), (ins SPR:$Sm),
477 IIC_fpCVTDS, "vcvt", ".f64.f32\t$Dd, $Sm",
478 [(set DPR:$Dd, (fextend SPR:$Sm))]> {
479 // Instruction operands.
483 // Encode instruction operands.
484 let Inst{3-0} = Sm{4-1};
486 let Inst{15-12} = Dd{3-0};
487 let Inst{22} = Dd{4};
490 // Special case encoding: bits 11-8 is 0b1011.
491 def VCVTSD : VFPAI<(outs SPR:$Sd), (ins DPR:$Dm), VFPUnaryFrm,
492 IIC_fpCVTSD, "vcvt", ".f32.f64\t$Sd, $Dm",
493 [(set SPR:$Sd, (fround DPR:$Dm))]> {
494 // Instruction operands.
498 // Encode instruction operands.
499 let Inst{3-0} = Dm{3-0};
501 let Inst{15-12} = Sd{4-1};
502 let Inst{22} = Sd{0};
504 let Inst{27-23} = 0b11101;
505 let Inst{21-16} = 0b110111;
506 let Inst{11-8} = 0b1011;
507 let Inst{7-6} = 0b11;
511 // Between half, single and double-precision. For disassembly only.
513 // FIXME: Verify encoding after integrated assembler is working.
514 def VCVTBHS: ASuI<0b11101, 0b11, 0b0010, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
515 /* FIXME */ IIC_fpCVTSH, "vcvtb", ".f32.f16\t$Sd, $Sm",
516 [/* For disassembly only; pattern left blank */]>;
518 def VCVTBSH: ASuI<0b11101, 0b11, 0b0011, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
519 /* FIXME */ IIC_fpCVTHS, "vcvtb", ".f16.f32\t$Sd, $Sm",
520 [/* For disassembly only; pattern left blank */]>;
522 def : Pat<(f32_to_f16 SPR:$a),
523 (i32 (COPY_TO_REGCLASS (VCVTBSH SPR:$a), GPR))>;
525 def : Pat<(f16_to_f32 GPR:$a),
526 (VCVTBHS (COPY_TO_REGCLASS GPR:$a, SPR))>;
528 def VCVTTHS: ASuI<0b11101, 0b11, 0b0010, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
529 /* FIXME */ IIC_fpCVTSH, "vcvtt", ".f32.f16\t$Sd, $Sm",
530 [/* For disassembly only; pattern left blank */]>;
532 def VCVTTSH: ASuI<0b11101, 0b11, 0b0011, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
533 /* FIXME */ IIC_fpCVTHS, "vcvtt", ".f16.f32\t$Sd, $Sm",
534 [/* For disassembly only; pattern left blank */]>;
536 def VCVTBHD : ADuI<0b11101, 0b11, 0b0010, 0b01, 0,
537 (outs DPR:$Dd), (ins SPR:$Sm),
538 NoItinerary, "vcvtb", ".f64.f16\t$Dd, $Sm",
539 []>, Requires<[HasV8FP]> {
540 // Instruction operands.
543 // Encode instruction operands.
544 let Inst{3-0} = Sm{4-1};
548 def VCVTBDH : ADuI<0b11101, 0b11, 0b0011, 0b01, 0,
549 (outs SPR:$Sd), (ins DPR:$Dm),
550 NoItinerary, "vcvtb", ".f16.f64\t$Sd, $Dm",
551 []>, Requires<[HasV8FP]> {
552 // Instruction operands.
556 // Encode instruction operands.
557 let Inst{3-0} = Dm{3-0};
559 let Inst{15-12} = Sd{4-1};
560 let Inst{22} = Sd{0};
563 def VCVTTHD : ADuI<0b11101, 0b11, 0b0010, 0b11, 0,
564 (outs DPR:$Dd), (ins SPR:$Sm),
565 NoItinerary, "vcvtt", ".f64.f16\t$Dd, $Sm",
566 []>, Requires<[HasV8FP]> {
567 // Instruction operands.
570 // Encode instruction operands.
571 let Inst{3-0} = Sm{4-1};
575 def VCVTTDH : ADuI<0b11101, 0b11, 0b0011, 0b11, 0,
576 (outs SPR:$Sd), (ins DPR:$Dm),
577 NoItinerary, "vcvtt", ".f16.f64\t$Sd, $Dm",
578 []>, Requires<[HasV8FP]> {
579 // Instruction operands.
583 // Encode instruction operands.
584 let Inst{15-12} = Sd{4-1};
585 let Inst{22} = Sd{0};
586 let Inst{3-0} = Dm{3-0};
590 multiclass vcvt_inst<string opc, bits<2> rm> {
591 let PostEncoderMethod = "", DecoderNamespace = "VFPV8" in {
592 def SS : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
593 (outs SPR:$Sd), (ins SPR:$Sm),
594 NoItinerary, !strconcat("vcvt", opc, ".s32.f32\t$Sd, $Sm"),
595 []>, Requires<[HasV8FP]> {
596 let Inst{17-16} = rm;
599 def US : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
600 (outs SPR:$Sd), (ins SPR:$Sm),
601 NoItinerary, !strconcat("vcvt", opc, ".u32.f32\t$Sd, $Sm"),
602 []>, Requires<[HasV8FP]> {
603 let Inst{17-16} = rm;
606 def SD : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
607 (outs SPR:$Sd), (ins DPR:$Dm),
608 NoItinerary, !strconcat("vcvt", opc, ".s32.f64\t$Sd, $Dm"),
609 []>, Requires<[HasV8FP]> {
612 let Inst{17-16} = rm;
614 // Encode instruction operands
615 let Inst{3-0} = Dm{3-0};
620 def UD : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
621 (outs SPR:$Sd), (ins DPR:$Dm),
622 NoItinerary, !strconcat("vcvt", opc, ".u32.f64\t$Sd, $Dm"),
623 []>, Requires<[HasV8FP]> {
626 let Inst{17-16} = rm;
628 // Encode instruction operands
629 let Inst{3-0} = Dm{3-0};
636 defm VCVTA : vcvt_inst<"a", 0b00>;
637 defm VCVTN : vcvt_inst<"n", 0b01>;
638 defm VCVTP : vcvt_inst<"p", 0b10>;
639 defm VCVTM : vcvt_inst<"m", 0b11>;
641 def VNEGD : ADuI<0b11101, 0b11, 0b0001, 0b01, 0,
642 (outs DPR:$Dd), (ins DPR:$Dm),
643 IIC_fpUNA64, "vneg", ".f64\t$Dd, $Dm",
644 [(set DPR:$Dd, (fneg (f64 DPR:$Dm)))]>;
646 def VNEGS : ASuIn<0b11101, 0b11, 0b0001, 0b01, 0,
647 (outs SPR:$Sd), (ins SPR:$Sm),
648 IIC_fpUNA32, "vneg", ".f32\t$Sd, $Sm",
649 [(set SPR:$Sd, (fneg SPR:$Sm))]> {
650 // Some single precision VFP instructions may be executed on both NEON and
651 // VFP pipelines on A8.
652 let D = VFPNeonA8Domain;
655 multiclass vrint_inst_zrx<string opc, bit op, bit op2> {
656 def S : ASuI<0b11101, 0b11, 0b0110, 0b11, 0,
657 (outs SPR:$Sd), (ins SPR:$Sm),
658 NoItinerary, !strconcat("vrint", opc), ".f32\t$Sd, $Sm",
659 []>, Requires<[HasV8FP]> {
663 def D : ADuI<0b11101, 0b11, 0b0110, 0b11, 0,
664 (outs DPR:$Dd), (ins DPR:$Dm),
665 NoItinerary, !strconcat("vrint", opc), ".f64\t$Dd, $Dm",
666 []>, Requires<[HasV8FP]> {
672 defm VRINTZ : vrint_inst_zrx<"z", 0, 1>;
673 defm VRINTR : vrint_inst_zrx<"r", 0, 0>;
674 defm VRINTX : vrint_inst_zrx<"x", 1, 0>;
676 multiclass vrint_inst_anpm<string opc, bits<2> rm> {
677 let PostEncoderMethod = "", DecoderNamespace = "VFPV8" in {
678 def S : ASuInp<0b11101, 0b11, 0b1000, 0b01, 0,
679 (outs SPR:$Sd), (ins SPR:$Sm),
680 NoItinerary, !strconcat("vrint", opc, ".f32\t$Sd, $Sm"),
681 []>, Requires<[HasV8FP]> {
682 let Inst{17-16} = rm;
684 def D : ADuInp<0b11101, 0b11, 0b1000, 0b01, 0,
685 (outs DPR:$Dd), (ins DPR:$Dm),
686 NoItinerary, !strconcat("vrint", opc, ".f64\t$Dd, $Dm"),
687 []>, Requires<[HasV8FP]> {
688 let Inst{17-16} = rm;
692 def : InstAlias<!strconcat("vrint", opc, ".f32.f32\t$Sd, $Sm"),
693 (!cast<Instruction>(NAME#"S") SPR:$Sd, SPR:$Sm)>;
694 def : InstAlias<!strconcat("vrint", opc, ".f64.f64\t$Dd, $Dm"),
695 (!cast<Instruction>(NAME#"D") DPR:$Dd, DPR:$Dm)>;
698 defm VRINTA : vrint_inst_anpm<"a", 0b00>;
699 defm VRINTN : vrint_inst_anpm<"n", 0b01>;
700 defm VRINTP : vrint_inst_anpm<"p", 0b10>;
701 defm VRINTM : vrint_inst_anpm<"m", 0b11>;
703 def VSQRTD : ADuI<0b11101, 0b11, 0b0001, 0b11, 0,
704 (outs DPR:$Dd), (ins DPR:$Dm),
705 IIC_fpSQRT64, "vsqrt", ".f64\t$Dd, $Dm",
706 [(set DPR:$Dd, (fsqrt (f64 DPR:$Dm)))]>;
708 def VSQRTS : ASuI<0b11101, 0b11, 0b0001, 0b11, 0,
709 (outs SPR:$Sd), (ins SPR:$Sm),
710 IIC_fpSQRT32, "vsqrt", ".f32\t$Sd, $Sm",
711 [(set SPR:$Sd, (fsqrt SPR:$Sm))]>;
713 let neverHasSideEffects = 1 in {
714 def VMOVD : ADuI<0b11101, 0b11, 0b0000, 0b01, 0,
715 (outs DPR:$Dd), (ins DPR:$Dm),
716 IIC_fpUNA64, "vmov", ".f64\t$Dd, $Dm", []>;
718 def VMOVS : ASuI<0b11101, 0b11, 0b0000, 0b01, 0,
719 (outs SPR:$Sd), (ins SPR:$Sm),
720 IIC_fpUNA32, "vmov", ".f32\t$Sd, $Sm", []>;
721 } // neverHasSideEffects
723 //===----------------------------------------------------------------------===//
724 // FP <-> GPR Copies. Int <-> FP Conversions.
727 def VMOVRS : AVConv2I<0b11100001, 0b1010,
728 (outs GPR:$Rt), (ins SPR:$Sn),
729 IIC_fpMOVSI, "vmov", "\t$Rt, $Sn",
730 [(set GPR:$Rt, (bitconvert SPR:$Sn))]> {
731 // Instruction operands.
735 // Encode instruction operands.
736 let Inst{19-16} = Sn{4-1};
738 let Inst{15-12} = Rt;
740 let Inst{6-5} = 0b00;
741 let Inst{3-0} = 0b0000;
743 // Some single precision VFP instructions may be executed on both NEON and VFP
745 let D = VFPNeonDomain;
748 // Bitcast i32 -> f32. NEON prefers to use VMOVDRR.
749 def VMOVSR : AVConv4I<0b11100000, 0b1010,
750 (outs SPR:$Sn), (ins GPR:$Rt),
751 IIC_fpMOVIS, "vmov", "\t$Sn, $Rt",
752 [(set SPR:$Sn, (bitconvert GPR:$Rt))]>,
753 Requires<[HasVFP2, UseVMOVSR]> {
754 // Instruction operands.
758 // Encode instruction operands.
759 let Inst{19-16} = Sn{4-1};
761 let Inst{15-12} = Rt;
763 let Inst{6-5} = 0b00;
764 let Inst{3-0} = 0b0000;
766 // Some single precision VFP instructions may be executed on both NEON and VFP
768 let D = VFPNeonDomain;
771 let neverHasSideEffects = 1 in {
772 def VMOVRRD : AVConv3I<0b11000101, 0b1011,
773 (outs GPR:$Rt, GPR:$Rt2), (ins DPR:$Dm),
774 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $Dm",
775 [/* FIXME: Can't write pattern for multiple result instr*/]> {
776 // Instruction operands.
781 // Encode instruction operands.
782 let Inst{3-0} = Dm{3-0};
784 let Inst{15-12} = Rt;
785 let Inst{19-16} = Rt2;
787 let Inst{7-6} = 0b00;
789 // Some single precision VFP instructions may be executed on both NEON and VFP
791 let D = VFPNeonDomain;
794 def VMOVRRS : AVConv3I<0b11000101, 0b1010,
795 (outs GPR:$Rt, GPR:$Rt2), (ins SPR:$src1, SPR:$src2),
796 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $src1, $src2",
797 [/* For disassembly only; pattern left blank */]> {
802 // Encode instruction operands.
803 let Inst{3-0} = src1{4-1};
804 let Inst{5} = src1{0};
805 let Inst{15-12} = Rt;
806 let Inst{19-16} = Rt2;
808 let Inst{7-6} = 0b00;
810 // Some single precision VFP instructions may be executed on both NEON and VFP
812 let D = VFPNeonDomain;
813 let DecoderMethod = "DecodeVMOVRRS";
815 } // neverHasSideEffects
820 def VMOVDRR : AVConv5I<0b11000100, 0b1011,
821 (outs DPR:$Dm), (ins GPR:$Rt, GPR:$Rt2),
822 IIC_fpMOVID, "vmov", "\t$Dm, $Rt, $Rt2",
823 [(set DPR:$Dm, (arm_fmdrr GPR:$Rt, GPR:$Rt2))]> {
824 // Instruction operands.
829 // Encode instruction operands.
830 let Inst{3-0} = Dm{3-0};
832 let Inst{15-12} = Rt;
833 let Inst{19-16} = Rt2;
835 let Inst{7-6} = 0b00;
837 // Some single precision VFP instructions may be executed on both NEON and VFP
839 let D = VFPNeonDomain;
842 let neverHasSideEffects = 1 in
843 def VMOVSRR : AVConv5I<0b11000100, 0b1010,
844 (outs SPR:$dst1, SPR:$dst2), (ins GPR:$src1, GPR:$src2),
845 IIC_fpMOVID, "vmov", "\t$dst1, $dst2, $src1, $src2",
846 [/* For disassembly only; pattern left blank */]> {
847 // Instruction operands.
852 // Encode instruction operands.
853 let Inst{3-0} = dst1{4-1};
854 let Inst{5} = dst1{0};
855 let Inst{15-12} = src1;
856 let Inst{19-16} = src2;
858 let Inst{7-6} = 0b00;
860 // Some single precision VFP instructions may be executed on both NEON and VFP
862 let D = VFPNeonDomain;
864 let DecoderMethod = "DecodeVMOVSRR";
870 // FMRX: SPR system reg -> GPR
872 // FMXR: GPR -> VFP system reg
877 class AVConv1IDs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
878 bits<4> opcod4, dag oops, dag iops,
879 InstrItinClass itin, string opc, string asm,
881 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
883 // Instruction operands.
887 // Encode instruction operands.
888 let Inst{3-0} = Sm{4-1};
890 let Inst{15-12} = Dd{3-0};
891 let Inst{22} = Dd{4};
894 class AVConv1InSs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
895 bits<4> opcod4, dag oops, dag iops,InstrItinClass itin,
896 string opc, string asm, list<dag> pattern>
897 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
899 // Instruction operands.
903 // Encode instruction operands.
904 let Inst{3-0} = Sm{4-1};
906 let Inst{15-12} = Sd{4-1};
907 let Inst{22} = Sd{0};
910 def VSITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
911 (outs DPR:$Dd), (ins SPR:$Sm),
912 IIC_fpCVTID, "vcvt", ".f64.s32\t$Dd, $Sm",
913 [(set DPR:$Dd, (f64 (arm_sitof SPR:$Sm)))]> {
914 let Inst{7} = 1; // s32
917 def VSITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
918 (outs SPR:$Sd),(ins SPR:$Sm),
919 IIC_fpCVTIS, "vcvt", ".f32.s32\t$Sd, $Sm",
920 [(set SPR:$Sd, (arm_sitof SPR:$Sm))]> {
921 let Inst{7} = 1; // s32
923 // Some single precision VFP instructions may be executed on both NEON and
924 // VFP pipelines on A8.
925 let D = VFPNeonA8Domain;
928 def VUITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
929 (outs DPR:$Dd), (ins SPR:$Sm),
930 IIC_fpCVTID, "vcvt", ".f64.u32\t$Dd, $Sm",
931 [(set DPR:$Dd, (f64 (arm_uitof SPR:$Sm)))]> {
932 let Inst{7} = 0; // u32
935 def VUITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
936 (outs SPR:$Sd), (ins SPR:$Sm),
937 IIC_fpCVTIS, "vcvt", ".f32.u32\t$Sd, $Sm",
938 [(set SPR:$Sd, (arm_uitof SPR:$Sm))]> {
939 let Inst{7} = 0; // u32
941 // Some single precision VFP instructions may be executed on both NEON and
942 // VFP pipelines on A8.
943 let D = VFPNeonA8Domain;
948 class AVConv1IsD_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
949 bits<4> opcod4, dag oops, dag iops,
950 InstrItinClass itin, string opc, string asm,
952 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
954 // Instruction operands.
958 // Encode instruction operands.
959 let Inst{3-0} = Dm{3-0};
961 let Inst{15-12} = Sd{4-1};
962 let Inst{22} = Sd{0};
965 class AVConv1InsS_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
966 bits<4> opcod4, dag oops, dag iops,
967 InstrItinClass itin, string opc, string asm,
969 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
971 // Instruction operands.
975 // Encode instruction operands.
976 let Inst{3-0} = Sm{4-1};
978 let Inst{15-12} = Sd{4-1};
979 let Inst{22} = Sd{0};
982 // Always set Z bit in the instruction, i.e. "round towards zero" variants.
983 def VTOSIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
984 (outs SPR:$Sd), (ins DPR:$Dm),
985 IIC_fpCVTDI, "vcvt", ".s32.f64\t$Sd, $Dm",
986 [(set SPR:$Sd, (arm_ftosi (f64 DPR:$Dm)))]> {
987 let Inst{7} = 1; // Z bit
990 def VTOSIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
991 (outs SPR:$Sd), (ins SPR:$Sm),
992 IIC_fpCVTSI, "vcvt", ".s32.f32\t$Sd, $Sm",
993 [(set SPR:$Sd, (arm_ftosi SPR:$Sm))]> {
994 let Inst{7} = 1; // Z bit
996 // Some single precision VFP instructions may be executed on both NEON and
997 // VFP pipelines on A8.
998 let D = VFPNeonA8Domain;
1001 def VTOUIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
1002 (outs SPR:$Sd), (ins DPR:$Dm),
1003 IIC_fpCVTDI, "vcvt", ".u32.f64\t$Sd, $Dm",
1004 [(set SPR:$Sd, (arm_ftoui (f64 DPR:$Dm)))]> {
1005 let Inst{7} = 1; // Z bit
1008 def VTOUIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
1009 (outs SPR:$Sd), (ins SPR:$Sm),
1010 IIC_fpCVTSI, "vcvt", ".u32.f32\t$Sd, $Sm",
1011 [(set SPR:$Sd, (arm_ftoui SPR:$Sm))]> {
1012 let Inst{7} = 1; // Z bit
1014 // Some single precision VFP instructions may be executed on both NEON and
1015 // VFP pipelines on A8.
1016 let D = VFPNeonA8Domain;
1019 // And the Z bit '0' variants, i.e. use the rounding mode specified by FPSCR.
1020 let Uses = [FPSCR] in {
1021 // FIXME: Verify encoding after integrated assembler is working.
1022 def VTOSIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
1023 (outs SPR:$Sd), (ins DPR:$Dm),
1024 IIC_fpCVTDI, "vcvtr", ".s32.f64\t$Sd, $Dm",
1025 [(set SPR:$Sd, (int_arm_vcvtr (f64 DPR:$Dm)))]>{
1026 let Inst{7} = 0; // Z bit
1029 def VTOSIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
1030 (outs SPR:$Sd), (ins SPR:$Sm),
1031 IIC_fpCVTSI, "vcvtr", ".s32.f32\t$Sd, $Sm",
1032 [(set SPR:$Sd, (int_arm_vcvtr SPR:$Sm))]> {
1033 let Inst{7} = 0; // Z bit
1036 def VTOUIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
1037 (outs SPR:$Sd), (ins DPR:$Dm),
1038 IIC_fpCVTDI, "vcvtr", ".u32.f64\t$Sd, $Dm",
1039 [(set SPR:$Sd, (int_arm_vcvtru(f64 DPR:$Dm)))]>{
1040 let Inst{7} = 0; // Z bit
1043 def VTOUIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
1044 (outs SPR:$Sd), (ins SPR:$Sm),
1045 IIC_fpCVTSI, "vcvtr", ".u32.f32\t$Sd, $Sm",
1046 [(set SPR:$Sd, (int_arm_vcvtru SPR:$Sm))]> {
1047 let Inst{7} = 0; // Z bit
1051 // Convert between floating-point and fixed-point
1052 // Data type for fixed-point naming convention:
1053 // S16 (U=0, sx=0) -> SH
1054 // U16 (U=1, sx=0) -> UH
1055 // S32 (U=0, sx=1) -> SL
1056 // U32 (U=1, sx=1) -> UL
1058 let Constraints = "$a = $dst" in {
1060 // FP to Fixed-Point:
1062 // Single Precision register
1063 class AVConv1XInsS_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
1064 bit op5, dag oops, dag iops, InstrItinClass itin,
1065 string opc, string asm, list<dag> pattern>
1066 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern>,
1067 Sched<[WriteCvtFP]> {
1069 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
1070 let Inst{22} = dst{0};
1071 let Inst{15-12} = dst{4-1};
1074 // Double Precision register
1075 class AVConv1XInsD_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
1076 bit op5, dag oops, dag iops, InstrItinClass itin,
1077 string opc, string asm, list<dag> pattern>
1078 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern>,
1079 Sched<[WriteCvtFP]> {
1081 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
1082 let Inst{22} = dst{4};
1083 let Inst{15-12} = dst{3-0};
1086 def VTOSHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 0,
1087 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1088 IIC_fpCVTSI, "vcvt", ".s16.f32\t$dst, $a, $fbits", []> {
1089 // Some single precision VFP instructions may be executed on both NEON and
1090 // VFP pipelines on A8.
1091 let D = VFPNeonA8Domain;
1094 def VTOUHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 0,
1095 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1096 IIC_fpCVTSI, "vcvt", ".u16.f32\t$dst, $a, $fbits", []> {
1097 // Some single precision VFP instructions may be executed on both NEON and
1098 // VFP pipelines on A8.
1099 let D = VFPNeonA8Domain;
1102 def VTOSLS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 1,
1103 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1104 IIC_fpCVTSI, "vcvt", ".s32.f32\t$dst, $a, $fbits", []> {
1105 // Some single precision VFP instructions may be executed on both NEON and
1106 // VFP pipelines on A8.
1107 let D = VFPNeonA8Domain;
1110 def VTOULS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 1,
1111 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1112 IIC_fpCVTSI, "vcvt", ".u32.f32\t$dst, $a, $fbits", []> {
1113 // Some single precision VFP instructions may be executed on both NEON and
1114 // VFP pipelines on A8.
1115 let D = VFPNeonA8Domain;
1118 def VTOSHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 0,
1119 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1120 IIC_fpCVTDI, "vcvt", ".s16.f64\t$dst, $a, $fbits", []>;
1122 def VTOUHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 0,
1123 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1124 IIC_fpCVTDI, "vcvt", ".u16.f64\t$dst, $a, $fbits", []>;
1126 def VTOSLD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 1,
1127 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1128 IIC_fpCVTDI, "vcvt", ".s32.f64\t$dst, $a, $fbits", []>;
1130 def VTOULD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 1,
1131 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1132 IIC_fpCVTDI, "vcvt", ".u32.f64\t$dst, $a, $fbits", []>;
1134 // Fixed-Point to FP:
1136 def VSHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 0,
1137 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1138 IIC_fpCVTIS, "vcvt", ".f32.s16\t$dst, $a, $fbits", []> {
1139 // Some single precision VFP instructions may be executed on both NEON and
1140 // VFP pipelines on A8.
1141 let D = VFPNeonA8Domain;
1144 def VUHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 0,
1145 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1146 IIC_fpCVTIS, "vcvt", ".f32.u16\t$dst, $a, $fbits", []> {
1147 // Some single precision VFP instructions may be executed on both NEON and
1148 // VFP pipelines on A8.
1149 let D = VFPNeonA8Domain;
1152 def VSLTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 1,
1153 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1154 IIC_fpCVTIS, "vcvt", ".f32.s32\t$dst, $a, $fbits", []> {
1155 // Some single precision VFP instructions may be executed on both NEON and
1156 // VFP pipelines on A8.
1157 let D = VFPNeonA8Domain;
1160 def VULTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 1,
1161 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1162 IIC_fpCVTIS, "vcvt", ".f32.u32\t$dst, $a, $fbits", []> {
1163 // Some single precision VFP instructions may be executed on both NEON and
1164 // VFP pipelines on A8.
1165 let D = VFPNeonA8Domain;
1168 def VSHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 0,
1169 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1170 IIC_fpCVTID, "vcvt", ".f64.s16\t$dst, $a, $fbits", []>;
1172 def VUHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 0,
1173 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1174 IIC_fpCVTID, "vcvt", ".f64.u16\t$dst, $a, $fbits", []>;
1176 def VSLTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 1,
1177 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1178 IIC_fpCVTID, "vcvt", ".f64.s32\t$dst, $a, $fbits", []>;
1180 def VULTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 1,
1181 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1182 IIC_fpCVTID, "vcvt", ".f64.u32\t$dst, $a, $fbits", []>;
1184 } // End of 'let Constraints = "$a = $dst" in'
1186 //===----------------------------------------------------------------------===//
1187 // FP Multiply-Accumulate Operations.
1190 def VMLAD : ADbI<0b11100, 0b00, 0, 0,
1191 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1192 IIC_fpMAC64, "vmla", ".f64\t$Dd, $Dn, $Dm",
1193 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1194 (f64 DPR:$Ddin)))]>,
1195 RegConstraint<"$Ddin = $Dd">,
1196 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1198 def VMLAS : ASbIn<0b11100, 0b00, 0, 0,
1199 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1200 IIC_fpMAC32, "vmla", ".f32\t$Sd, $Sn, $Sm",
1201 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
1203 RegConstraint<"$Sdin = $Sd">,
1204 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1205 // Some single precision VFP instructions may be executed on both NEON and
1206 // VFP pipelines on A8.
1207 let D = VFPNeonA8Domain;
1210 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1211 (VMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1212 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1213 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1214 (VMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1215 Requires<[HasVFP2,DontUseNEONForFP, UseFPVMLx,DontUseFusedMAC]>;
1217 def VMLSD : ADbI<0b11100, 0b00, 1, 0,
1218 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1219 IIC_fpMAC64, "vmls", ".f64\t$Dd, $Dn, $Dm",
1220 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1221 (f64 DPR:$Ddin)))]>,
1222 RegConstraint<"$Ddin = $Dd">,
1223 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1225 def VMLSS : ASbIn<0b11100, 0b00, 1, 0,
1226 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1227 IIC_fpMAC32, "vmls", ".f32\t$Sd, $Sn, $Sm",
1228 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1230 RegConstraint<"$Sdin = $Sd">,
1231 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1232 // Some single precision VFP instructions may be executed on both NEON and
1233 // VFP pipelines on A8.
1234 let D = VFPNeonA8Domain;
1237 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1238 (VMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
1239 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1240 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1241 (VMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
1242 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1244 def VNMLAD : ADbI<0b11100, 0b01, 1, 0,
1245 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1246 IIC_fpMAC64, "vnmla", ".f64\t$Dd, $Dn, $Dm",
1247 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1248 (f64 DPR:$Ddin)))]>,
1249 RegConstraint<"$Ddin = $Dd">,
1250 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1252 def VNMLAS : ASbI<0b11100, 0b01, 1, 0,
1253 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1254 IIC_fpMAC32, "vnmla", ".f32\t$Sd, $Sn, $Sm",
1255 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1257 RegConstraint<"$Sdin = $Sd">,
1258 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1259 // Some single precision VFP instructions may be executed on both NEON and
1260 // VFP pipelines on A8.
1261 let D = VFPNeonA8Domain;
1264 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
1265 (VNMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1266 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1267 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
1268 (VNMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1269 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1271 def VNMLSD : ADbI<0b11100, 0b01, 0, 0,
1272 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1273 IIC_fpMAC64, "vnmls", ".f64\t$Dd, $Dn, $Dm",
1274 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1275 (f64 DPR:$Ddin)))]>,
1276 RegConstraint<"$Ddin = $Dd">,
1277 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1279 def VNMLSS : ASbI<0b11100, 0b01, 0, 0,
1280 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1281 IIC_fpMAC32, "vnmls", ".f32\t$Sd, $Sn, $Sm",
1282 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
1283 RegConstraint<"$Sdin = $Sd">,
1284 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1285 // Some single precision VFP instructions may be executed on both NEON and
1286 // VFP pipelines on A8.
1287 let D = VFPNeonA8Domain;
1290 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
1291 (VNMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
1292 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1293 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
1294 (VNMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
1295 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1297 //===----------------------------------------------------------------------===//
1298 // Fused FP Multiply-Accumulate Operations.
1300 def VFMAD : ADbI<0b11101, 0b10, 0, 0,
1301 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1302 IIC_fpFMAC64, "vfma", ".f64\t$Dd, $Dn, $Dm",
1303 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1304 (f64 DPR:$Ddin)))]>,
1305 RegConstraint<"$Ddin = $Dd">,
1306 Requires<[HasVFP4,UseFusedMAC]>;
1308 def VFMAS : ASbIn<0b11101, 0b10, 0, 0,
1309 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1310 IIC_fpFMAC32, "vfma", ".f32\t$Sd, $Sn, $Sm",
1311 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
1313 RegConstraint<"$Sdin = $Sd">,
1314 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1315 // Some single precision VFP instructions may be executed on both NEON and
1319 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1320 (VFMAD DPR:$dstin, DPR:$a, DPR:$b)>,
1321 Requires<[HasVFP4,UseFusedMAC]>;
1322 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1323 (VFMAS SPR:$dstin, SPR:$a, SPR:$b)>,
1324 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1326 // Match @llvm.fma.* intrinsics
1327 // (fma x, y, z) -> (vfms z, x, y)
1328 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, DPR:$Ddin)),
1329 (VFMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1330 Requires<[HasVFP4]>;
1331 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, SPR:$Sdin)),
1332 (VFMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1333 Requires<[HasVFP4]>;
1335 def VFMSD : ADbI<0b11101, 0b10, 1, 0,
1336 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1337 IIC_fpFMAC64, "vfms", ".f64\t$Dd, $Dn, $Dm",
1338 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1339 (f64 DPR:$Ddin)))]>,
1340 RegConstraint<"$Ddin = $Dd">,
1341 Requires<[HasVFP4,UseFusedMAC]>;
1343 def VFMSS : ASbIn<0b11101, 0b10, 1, 0,
1344 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1345 IIC_fpFMAC32, "vfms", ".f32\t$Sd, $Sn, $Sm",
1346 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1348 RegConstraint<"$Sdin = $Sd">,
1349 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1350 // Some single precision VFP instructions may be executed on both NEON and
1354 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1355 (VFMSD DPR:$dstin, DPR:$a, DPR:$b)>,
1356 Requires<[HasVFP4,UseFusedMAC]>;
1357 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1358 (VFMSS SPR:$dstin, SPR:$a, SPR:$b)>,
1359 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1361 // Match @llvm.fma.* intrinsics
1362 // (fma (fneg x), y, z) -> (vfms z, x, y)
1363 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin)),
1364 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1365 Requires<[HasVFP4]>;
1366 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin)),
1367 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1368 Requires<[HasVFP4]>;
1369 // (fma x, (fneg y), z) -> (vfms z, x, y)
1370 def : Pat<(f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin)),
1371 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1372 Requires<[HasVFP4]>;
1373 def : Pat<(f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin)),
1374 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1375 Requires<[HasVFP4]>;
1377 def VFNMAD : ADbI<0b11101, 0b01, 1, 0,
1378 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1379 IIC_fpFMAC64, "vfnma", ".f64\t$Dd, $Dn, $Dm",
1380 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1381 (f64 DPR:$Ddin)))]>,
1382 RegConstraint<"$Ddin = $Dd">,
1383 Requires<[HasVFP4,UseFusedMAC]>;
1385 def VFNMAS : ASbI<0b11101, 0b01, 1, 0,
1386 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1387 IIC_fpFMAC32, "vfnma", ".f32\t$Sd, $Sn, $Sm",
1388 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1390 RegConstraint<"$Sdin = $Sd">,
1391 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1392 // Some single precision VFP instructions may be executed on both NEON and
1396 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
1397 (VFNMAD DPR:$dstin, DPR:$a, DPR:$b)>,
1398 Requires<[HasVFP4,UseFusedMAC]>;
1399 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
1400 (VFNMAS SPR:$dstin, SPR:$a, SPR:$b)>,
1401 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1403 // Match @llvm.fma.* intrinsics
1404 // (fneg (fma x, y, z)) -> (vfnma z, x, y)
1405 def : Pat<(fneg (fma (f64 DPR:$Dn), (f64 DPR:$Dm), (f64 DPR:$Ddin))),
1406 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1407 Requires<[HasVFP4]>;
1408 def : Pat<(fneg (fma (f32 SPR:$Sn), (f32 SPR:$Sm), (f32 SPR:$Sdin))),
1409 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1410 Requires<[HasVFP4]>;
1411 // (fma (fneg x), y, (fneg z)) -> (vfnma z, x, y)
1412 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, (fneg DPR:$Ddin))),
1413 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1414 Requires<[HasVFP4]>;
1415 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, (fneg SPR:$Sdin))),
1416 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1417 Requires<[HasVFP4]>;
1419 def VFNMSD : ADbI<0b11101, 0b01, 0, 0,
1420 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1421 IIC_fpFMAC64, "vfnms", ".f64\t$Dd, $Dn, $Dm",
1422 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1423 (f64 DPR:$Ddin)))]>,
1424 RegConstraint<"$Ddin = $Dd">,
1425 Requires<[HasVFP4,UseFusedMAC]>;
1427 def VFNMSS : ASbI<0b11101, 0b01, 0, 0,
1428 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1429 IIC_fpFMAC32, "vfnms", ".f32\t$Sd, $Sn, $Sm",
1430 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
1431 RegConstraint<"$Sdin = $Sd">,
1432 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1433 // Some single precision VFP instructions may be executed on both NEON and
1437 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
1438 (VFNMSD DPR:$dstin, DPR:$a, DPR:$b)>,
1439 Requires<[HasVFP4,UseFusedMAC]>;
1440 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
1441 (VFNMSS SPR:$dstin, SPR:$a, SPR:$b)>,
1442 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1444 // Match @llvm.fma.* intrinsics
1446 // (fma x, y, (fneg z)) -> (vfnms z, x, y))
1447 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, (fneg DPR:$Ddin))),
1448 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1449 Requires<[HasVFP4]>;
1450 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, (fneg SPR:$Sdin))),
1451 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1452 Requires<[HasVFP4]>;
1453 // (fneg (fma (fneg x), y, z)) -> (vfnms z, x, y)
1454 def : Pat<(fneg (f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin))),
1455 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1456 Requires<[HasVFP4]>;
1457 def : Pat<(fneg (f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin))),
1458 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1459 Requires<[HasVFP4]>;
1460 // (fneg (fma x, (fneg y), z) -> (vfnms z, x, y)
1461 def : Pat<(fneg (f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin))),
1462 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1463 Requires<[HasVFP4]>;
1464 def : Pat<(fneg (f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin))),
1465 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1466 Requires<[HasVFP4]>;
1468 //===----------------------------------------------------------------------===//
1469 // FP Conditional moves.
1472 let neverHasSideEffects = 1 in {
1473 def VMOVDcc : PseudoInst<(outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm, cmovpred:$p),
1475 [(set (f64 DPR:$Dd),
1476 (ARMcmov DPR:$Dn, DPR:$Dm, cmovpred:$p))]>,
1477 RegConstraint<"$Dn = $Dd">, Requires<[HasVFP2]>;
1479 def VMOVScc : PseudoInst<(outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm, cmovpred:$p),
1481 [(set (f32 SPR:$Sd),
1482 (ARMcmov SPR:$Sn, SPR:$Sm, cmovpred:$p))]>,
1483 RegConstraint<"$Sn = $Sd">, Requires<[HasVFP2]>;
1484 } // neverHasSideEffects
1486 //===----------------------------------------------------------------------===//
1487 // Move from VFP System Register to ARM core register.
1490 class MovFromVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
1492 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
1494 // Instruction operand.
1497 let Inst{27-20} = 0b11101111;
1498 let Inst{19-16} = opc19_16;
1499 let Inst{15-12} = Rt;
1500 let Inst{11-8} = 0b1010;
1502 let Inst{6-5} = 0b00;
1504 let Inst{3-0} = 0b0000;
1507 // APSR is the application level alias of CPSR. This FPSCR N, Z, C, V flags
1509 let Defs = [CPSR], Uses = [FPSCR_NZCV], Rt = 0b1111 /* apsr_nzcv */ in
1510 def FMSTAT : MovFromVFP<0b0001 /* fpscr */, (outs), (ins),
1511 "vmrs", "\tAPSR_nzcv, fpscr", [(arm_fmstat)]>;
1513 // Application level FPSCR -> GPR
1514 let hasSideEffects = 1, Uses = [FPSCR] in
1515 def VMRS : MovFromVFP<0b0001 /* fpscr */, (outs GPR:$Rt), (ins),
1516 "vmrs", "\t$Rt, fpscr",
1517 [(set GPR:$Rt, (int_arm_get_fpscr))]>;
1519 // System level FPEXC, FPSID -> GPR
1520 let Uses = [FPSCR] in {
1521 def VMRS_FPEXC : MovFromVFP<0b1000 /* fpexc */, (outs GPR:$Rt), (ins),
1522 "vmrs", "\t$Rt, fpexc", []>;
1523 def VMRS_FPSID : MovFromVFP<0b0000 /* fpsid */, (outs GPR:$Rt), (ins),
1524 "vmrs", "\t$Rt, fpsid", []>;
1525 def VMRS_MVFR0 : MovFromVFP<0b0111 /* mvfr0 */, (outs GPR:$Rt), (ins),
1526 "vmrs", "\t$Rt, mvfr0", []>;
1527 def VMRS_MVFR1 : MovFromVFP<0b0110 /* mvfr1 */, (outs GPR:$Rt), (ins),
1528 "vmrs", "\t$Rt, mvfr1", []>;
1529 def VMRS_FPINST : MovFromVFP<0b1001 /* fpinst */, (outs GPR:$Rt), (ins),
1530 "vmrs", "\t$Rt, fpinst", []>;
1531 def VMRS_FPINST2 : MovFromVFP<0b1010 /* fpinst2 */, (outs GPR:$Rt), (ins),
1532 "vmrs", "\t$Rt, fpinst2", []>;
1535 //===----------------------------------------------------------------------===//
1536 // Move from ARM core register to VFP System Register.
1539 class MovToVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
1541 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
1543 // Instruction operand.
1546 // Encode instruction operand.
1547 let Inst{15-12} = src;
1549 let Inst{27-20} = 0b11101110;
1550 let Inst{19-16} = opc19_16;
1551 let Inst{11-8} = 0b1010;
1556 let Defs = [FPSCR] in {
1557 // Application level GPR -> FPSCR
1558 def VMSR : MovToVFP<0b0001 /* fpscr */, (outs), (ins GPR:$src),
1559 "vmsr", "\tfpscr, $src", [(int_arm_set_fpscr GPR:$src)]>;
1560 // System level GPR -> FPEXC
1561 def VMSR_FPEXC : MovToVFP<0b1000 /* fpexc */, (outs), (ins GPR:$src),
1562 "vmsr", "\tfpexc, $src", []>;
1563 // System level GPR -> FPSID
1564 def VMSR_FPSID : MovToVFP<0b0000 /* fpsid */, (outs), (ins GPR:$src),
1565 "vmsr", "\tfpsid, $src", []>;
1567 def VMSR_FPINST : MovToVFP<0b1001 /* fpinst */, (outs), (ins GPR:$src),
1568 "vmsr", "\tfpinst, $src", []>;
1569 def VMSR_FPINST2 : MovToVFP<0b1010 /* fpinst2 */, (outs), (ins GPR:$src),
1570 "vmsr", "\tfpinst2, $src", []>;
1573 //===----------------------------------------------------------------------===//
1577 // Materialize FP immediates. VFP3 only.
1578 let isReMaterializable = 1 in {
1579 def FCONSTD : VFPAI<(outs DPR:$Dd), (ins vfp_f64imm:$imm),
1580 VFPMiscFrm, IIC_fpUNA64,
1581 "vmov", ".f64\t$Dd, $imm",
1582 [(set DPR:$Dd, vfp_f64imm:$imm)]>, Requires<[HasVFP3]> {
1586 let Inst{27-23} = 0b11101;
1587 let Inst{22} = Dd{4};
1588 let Inst{21-20} = 0b11;
1589 let Inst{19-16} = imm{7-4};
1590 let Inst{15-12} = Dd{3-0};
1591 let Inst{11-9} = 0b101;
1592 let Inst{8} = 1; // Double precision.
1593 let Inst{7-4} = 0b0000;
1594 let Inst{3-0} = imm{3-0};
1597 def FCONSTS : VFPAI<(outs SPR:$Sd), (ins vfp_f32imm:$imm),
1598 VFPMiscFrm, IIC_fpUNA32,
1599 "vmov", ".f32\t$Sd, $imm",
1600 [(set SPR:$Sd, vfp_f32imm:$imm)]>, Requires<[HasVFP3]> {
1604 let Inst{27-23} = 0b11101;
1605 let Inst{22} = Sd{0};
1606 let Inst{21-20} = 0b11;
1607 let Inst{19-16} = imm{7-4};
1608 let Inst{15-12} = Sd{4-1};
1609 let Inst{11-9} = 0b101;
1610 let Inst{8} = 0; // Single precision.
1611 let Inst{7-4} = 0b0000;
1612 let Inst{3-0} = imm{3-0};
1616 //===----------------------------------------------------------------------===//
1617 // Assembler aliases.
1619 // A few mnemnoic aliases for pre-unifixed syntax. We don't guarantee to
1620 // support them all, but supporting at least some of the basics is
1621 // good to be friendly.
1622 def : VFP2MnemonicAlias<"flds", "vldr">;
1623 def : VFP2MnemonicAlias<"fldd", "vldr">;
1624 def : VFP2MnemonicAlias<"fmrs", "vmov">;
1625 def : VFP2MnemonicAlias<"fmsr", "vmov">;
1626 def : VFP2MnemonicAlias<"fsqrts", "vsqrt">;
1627 def : VFP2MnemonicAlias<"fsqrtd", "vsqrt">;
1628 def : VFP2MnemonicAlias<"fadds", "vadd.f32">;
1629 def : VFP2MnemonicAlias<"faddd", "vadd.f64">;
1630 def : VFP2MnemonicAlias<"fmrdd", "vmov">;
1631 def : VFP2MnemonicAlias<"fmrds", "vmov">;
1632 def : VFP2MnemonicAlias<"fmrrd", "vmov">;
1633 def : VFP2MnemonicAlias<"fmdrr", "vmov">;
1634 def : VFP2MnemonicAlias<"fmuls", "vmul.f32">;
1635 def : VFP2MnemonicAlias<"fmuld", "vmul.f64">;
1636 def : VFP2MnemonicAlias<"fnegs", "vneg.f32">;
1637 def : VFP2MnemonicAlias<"fnegd", "vneg.f64">;
1638 def : VFP2MnemonicAlias<"ftosizd", "vcvt.s32.f64">;
1639 def : VFP2MnemonicAlias<"ftosid", "vcvtr.s32.f64">;
1640 def : VFP2MnemonicAlias<"ftosizs", "vcvt.s32.f32">;
1641 def : VFP2MnemonicAlias<"ftosis", "vcvtr.s32.f32">;
1642 def : VFP2MnemonicAlias<"ftouizd", "vcvt.u32.f64">;
1643 def : VFP2MnemonicAlias<"ftouid", "vcvtr.u32.f64">;
1644 def : VFP2MnemonicAlias<"ftouizs", "vcvt.u32.f32">;
1645 def : VFP2MnemonicAlias<"ftouis", "vcvtr.u32.f32">;
1646 def : VFP2MnemonicAlias<"fsitod", "vcvt.f64.s32">;
1647 def : VFP2MnemonicAlias<"fsitos", "vcvt.f32.s32">;
1648 def : VFP2MnemonicAlias<"fuitod", "vcvt.f64.u32">;
1649 def : VFP2MnemonicAlias<"fuitos", "vcvt.f32.u32">;
1650 def : VFP2MnemonicAlias<"fsts", "vstr">;
1651 def : VFP2MnemonicAlias<"fstd", "vstr">;
1652 def : VFP2MnemonicAlias<"fmacd", "vmla.f64">;
1653 def : VFP2MnemonicAlias<"fmacs", "vmla.f32">;
1654 def : VFP2MnemonicAlias<"fcpys", "vmov.f32">;
1655 def : VFP2MnemonicAlias<"fcpyd", "vmov.f64">;
1656 def : VFP2MnemonicAlias<"fcmps", "vcmp.f32">;
1657 def : VFP2MnemonicAlias<"fcmpd", "vcmp.f64">;
1658 def : VFP2MnemonicAlias<"fdivs", "vdiv.f32">;
1659 def : VFP2MnemonicAlias<"fdivd", "vdiv.f64">;
1660 def : VFP2MnemonicAlias<"fmrx", "vmrs">;
1661 def : VFP2MnemonicAlias<"fmxr", "vmsr">;
1663 // Be friendly and accept the old form of zero-compare
1664 def : VFP2InstAlias<"fcmpzd${p} $val", (VCMPZD DPR:$val, pred:$p)>;
1665 def : VFP2InstAlias<"fcmpzs${p} $val", (VCMPZS SPR:$val, pred:$p)>;
1668 def : VFP2InstAlias<"fmstat${p}", (FMSTAT pred:$p)>;
1669 def : VFP2InstAlias<"fadds${p} $Sd, $Sn, $Sm",
1670 (VADDS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
1671 def : VFP2InstAlias<"faddd${p} $Dd, $Dn, $Dm",
1672 (VADDD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
1673 def : VFP2InstAlias<"fsubs${p} $Sd, $Sn, $Sm",
1674 (VSUBS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
1675 def : VFP2InstAlias<"fsubd${p} $Dd, $Dn, $Dm",
1676 (VSUBD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
1678 // No need for the size suffix on VSQRT. It's implied by the register classes.
1679 def : VFP2InstAlias<"vsqrt${p} $Sd, $Sm", (VSQRTS SPR:$Sd, SPR:$Sm, pred:$p)>;
1680 def : VFP2InstAlias<"vsqrt${p} $Dd, $Dm", (VSQRTD DPR:$Dd, DPR:$Dm, pred:$p)>;
1682 // VLDR/VSTR accept an optional type suffix.
1683 def : VFP2InstAlias<"vldr${p}.32 $Sd, $addr",
1684 (VLDRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
1685 def : VFP2InstAlias<"vstr${p}.32 $Sd, $addr",
1686 (VSTRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
1687 def : VFP2InstAlias<"vldr${p}.64 $Dd, $addr",
1688 (VLDRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
1689 def : VFP2InstAlias<"vstr${p}.64 $Dd, $addr",
1690 (VSTRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
1692 // VMOV can accept optional 32-bit or less data type suffix suffix.
1693 def : VFP2InstAlias<"vmov${p}.8 $Rt, $Sn",
1694 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1695 def : VFP2InstAlias<"vmov${p}.16 $Rt, $Sn",
1696 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1697 def : VFP2InstAlias<"vmov${p}.32 $Rt, $Sn",
1698 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1699 def : VFP2InstAlias<"vmov${p}.8 $Sn, $Rt",
1700 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1701 def : VFP2InstAlias<"vmov${p}.16 $Sn, $Rt",
1702 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1703 def : VFP2InstAlias<"vmov${p}.32 $Sn, $Rt",
1704 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1706 def : VFP2InstAlias<"vmov${p}.f64 $Rt, $Rt2, $Dn",
1707 (VMOVRRD GPR:$Rt, GPR:$Rt2, DPR:$Dn, pred:$p)>;
1708 def : VFP2InstAlias<"vmov${p}.f64 $Dn, $Rt, $Rt2",
1709 (VMOVDRR DPR:$Dn, GPR:$Rt, GPR:$Rt2, pred:$p)>;
1711 // VMOVS doesn't need the .f32 to disambiguate from the NEON encoding the way
1713 def : VFP2InstAlias<"vmov${p} $Sd, $Sm",
1714 (VMOVS SPR:$Sd, SPR:$Sm, pred:$p)>;