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> {
337 let DecoderNamespace = "VFPV8", PostEncoderMethod = "" in {
338 def S : ASbInp<0b11100, opc, 0,
339 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
340 NoItinerary, !strconcat("vsel", op, ".f32\t$Sd, $Sn, $Sm"),
341 []>, Requires<[HasV8FP]>;
343 def D : ADbInp<0b11100, opc, 0,
344 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
345 NoItinerary, !strconcat("vsel", op, ".f64\t$Dd, $Dn, $Dm"),
346 []>, Requires<[HasV8FP]>;
350 defm VSELGT : vsel_inst<"gt", 0b11>;
351 defm VSELGE : vsel_inst<"ge", 0b10>;
352 defm VSELEQ : vsel_inst<"eq", 0b00>;
353 defm VSELVS : vsel_inst<"vs", 0b01>;
355 multiclass vmaxmin_inst<string op, bit opc> {
356 let DecoderNamespace = "VFPV8", PostEncoderMethod = "" in {
357 def S : ASbInp<0b11101, 0b00, opc,
358 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
359 NoItinerary, !strconcat(op, ".f32\t$Sd, $Sn, $Sm"),
360 []>, Requires<[HasV8FP]>;
362 def D : ADbInp<0b11101, 0b00, opc,
363 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
364 NoItinerary, !strconcat(op, ".f64\t$Dd, $Dn, $Dm"),
365 []>, Requires<[HasV8FP]>;
369 defm VMAXNM : vmaxmin_inst<"vmaxnm", 0>;
370 defm VMINNM : vmaxmin_inst<"vminnm", 1>;
372 // Match reassociated forms only if not sign dependent rounding.
373 def : Pat<(fmul (fneg DPR:$a), (f64 DPR:$b)),
374 (VNMULD DPR:$a, DPR:$b)>, Requires<[NoHonorSignDependentRounding]>;
375 def : Pat<(fmul (fneg SPR:$a), SPR:$b),
376 (VNMULS SPR:$a, SPR:$b)>, Requires<[NoHonorSignDependentRounding]>;
378 // These are encoded as unary instructions.
379 let Defs = [FPSCR_NZCV] in {
380 def VCMPED : ADuI<0b11101, 0b11, 0b0100, 0b11, 0,
381 (outs), (ins DPR:$Dd, DPR:$Dm),
382 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, $Dm",
383 [(arm_cmpfp DPR:$Dd, (f64 DPR:$Dm))]>;
385 def VCMPES : ASuI<0b11101, 0b11, 0b0100, 0b11, 0,
386 (outs), (ins SPR:$Sd, SPR:$Sm),
387 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, $Sm",
388 [(arm_cmpfp SPR:$Sd, SPR:$Sm)]> {
389 // Some single precision VFP instructions may be executed on both NEON and
390 // VFP pipelines on A8.
391 let D = VFPNeonA8Domain;
394 // FIXME: Verify encoding after integrated assembler is working.
395 def VCMPD : ADuI<0b11101, 0b11, 0b0100, 0b01, 0,
396 (outs), (ins DPR:$Dd, DPR:$Dm),
397 IIC_fpCMP64, "vcmp", ".f64\t$Dd, $Dm",
398 [/* For disassembly only; pattern left blank */]>;
400 def VCMPS : ASuI<0b11101, 0b11, 0b0100, 0b01, 0,
401 (outs), (ins SPR:$Sd, SPR:$Sm),
402 IIC_fpCMP32, "vcmp", ".f32\t$Sd, $Sm",
403 [/* For disassembly only; pattern left blank */]> {
404 // Some single precision VFP instructions may be executed on both NEON and
405 // VFP pipelines on A8.
406 let D = VFPNeonA8Domain;
408 } // Defs = [FPSCR_NZCV]
410 //===----------------------------------------------------------------------===//
411 // FP Unary Operations.
414 def VABSD : ADuI<0b11101, 0b11, 0b0000, 0b11, 0,
415 (outs DPR:$Dd), (ins DPR:$Dm),
416 IIC_fpUNA64, "vabs", ".f64\t$Dd, $Dm",
417 [(set DPR:$Dd, (fabs (f64 DPR:$Dm)))]>;
419 def VABSS : ASuIn<0b11101, 0b11, 0b0000, 0b11, 0,
420 (outs SPR:$Sd), (ins SPR:$Sm),
421 IIC_fpUNA32, "vabs", ".f32\t$Sd, $Sm",
422 [(set SPR:$Sd, (fabs SPR:$Sm))]> {
423 // Some single precision VFP instructions may be executed on both NEON and
424 // VFP pipelines on A8.
425 let D = VFPNeonA8Domain;
428 let Defs = [FPSCR_NZCV] in {
429 def VCMPEZD : ADuI<0b11101, 0b11, 0b0101, 0b11, 0,
430 (outs), (ins DPR:$Dd),
431 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, #0",
432 [(arm_cmpfp0 (f64 DPR:$Dd))]> {
433 let Inst{3-0} = 0b0000;
437 def VCMPEZS : ASuI<0b11101, 0b11, 0b0101, 0b11, 0,
438 (outs), (ins SPR:$Sd),
439 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, #0",
440 [(arm_cmpfp0 SPR:$Sd)]> {
441 let Inst{3-0} = 0b0000;
444 // Some single precision VFP instructions may be executed on both NEON and
445 // VFP pipelines on A8.
446 let D = VFPNeonA8Domain;
449 // FIXME: Verify encoding after integrated assembler is working.
450 def VCMPZD : ADuI<0b11101, 0b11, 0b0101, 0b01, 0,
451 (outs), (ins DPR:$Dd),
452 IIC_fpCMP64, "vcmp", ".f64\t$Dd, #0",
453 [/* For disassembly only; pattern left blank */]> {
454 let Inst{3-0} = 0b0000;
458 def VCMPZS : ASuI<0b11101, 0b11, 0b0101, 0b01, 0,
459 (outs), (ins SPR:$Sd),
460 IIC_fpCMP32, "vcmp", ".f32\t$Sd, #0",
461 [/* For disassembly only; pattern left blank */]> {
462 let Inst{3-0} = 0b0000;
465 // Some single precision VFP instructions may be executed on both NEON and
466 // VFP pipelines on A8.
467 let D = VFPNeonA8Domain;
469 } // Defs = [FPSCR_NZCV]
471 def VCVTDS : ASuI<0b11101, 0b11, 0b0111, 0b11, 0,
472 (outs DPR:$Dd), (ins SPR:$Sm),
473 IIC_fpCVTDS, "vcvt", ".f64.f32\t$Dd, $Sm",
474 [(set DPR:$Dd, (fextend SPR:$Sm))]> {
475 // Instruction operands.
479 // Encode instruction operands.
480 let Inst{3-0} = Sm{4-1};
482 let Inst{15-12} = Dd{3-0};
483 let Inst{22} = Dd{4};
486 // Special case encoding: bits 11-8 is 0b1011.
487 def VCVTSD : VFPAI<(outs SPR:$Sd), (ins DPR:$Dm), VFPUnaryFrm,
488 IIC_fpCVTSD, "vcvt", ".f32.f64\t$Sd, $Dm",
489 [(set SPR:$Sd, (fround DPR:$Dm))]> {
490 // Instruction operands.
494 // Encode instruction operands.
495 let Inst{3-0} = Dm{3-0};
497 let Inst{15-12} = Sd{4-1};
498 let Inst{22} = Sd{0};
500 let Inst{27-23} = 0b11101;
501 let Inst{21-16} = 0b110111;
502 let Inst{11-8} = 0b1011;
503 let Inst{7-6} = 0b11;
507 // Between half, single and double-precision. For disassembly only.
509 // FIXME: Verify encoding after integrated assembler is working.
510 def VCVTBHS: ASuI<0b11101, 0b11, 0b0010, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
511 /* FIXME */ IIC_fpCVTSH, "vcvtb", ".f32.f16\t$Sd, $Sm",
512 [/* For disassembly only; pattern left blank */]>;
514 def VCVTBSH: ASuI<0b11101, 0b11, 0b0011, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
515 /* FIXME */ IIC_fpCVTHS, "vcvtb", ".f16.f32\t$Sd, $Sm",
516 [/* For disassembly only; pattern left blank */]>;
518 def : Pat<(f32_to_f16 SPR:$a),
519 (i32 (COPY_TO_REGCLASS (VCVTBSH SPR:$a), GPR))>;
521 def : Pat<(f16_to_f32 GPR:$a),
522 (VCVTBHS (COPY_TO_REGCLASS GPR:$a, SPR))>;
524 def VCVTTHS: ASuI<0b11101, 0b11, 0b0010, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
525 /* FIXME */ IIC_fpCVTSH, "vcvtt", ".f32.f16\t$Sd, $Sm",
526 [/* For disassembly only; pattern left blank */]>;
528 def VCVTTSH: ASuI<0b11101, 0b11, 0b0011, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
529 /* FIXME */ IIC_fpCVTHS, "vcvtt", ".f16.f32\t$Sd, $Sm",
530 [/* For disassembly only; pattern left blank */]>;
532 def VCVTBHD : ADuI<0b11101, 0b11, 0b0010, 0b01, 0,
533 (outs DPR:$Dd), (ins SPR:$Sm),
534 NoItinerary, "vcvtb", ".f64.f16\t$Dd, $Sm",
535 []>, Requires<[HasV8FP]> {
536 // Instruction operands.
539 // Encode instruction operands.
540 let Inst{3-0} = Sm{4-1};
544 def VCVTBDH : ADuI<0b11101, 0b11, 0b0011, 0b01, 0,
545 (outs SPR:$Sd), (ins DPR:$Dm),
546 NoItinerary, "vcvtb", ".f16.f64\t$Sd, $Dm",
547 []>, Requires<[HasV8FP]> {
548 // Instruction operands.
552 // Encode instruction operands.
553 let Inst{3-0} = Dm{3-0};
555 let Inst{15-12} = Sd{4-1};
556 let Inst{22} = Sd{0};
559 def VCVTTHD : ADuI<0b11101, 0b11, 0b0010, 0b11, 0,
560 (outs DPR:$Dd), (ins SPR:$Sm),
561 NoItinerary, "vcvtt", ".f64.f16\t$Dd, $Sm",
562 []>, Requires<[HasV8FP]> {
563 // Instruction operands.
566 // Encode instruction operands.
567 let Inst{3-0} = Sm{4-1};
571 def VCVTTDH : ADuI<0b11101, 0b11, 0b0011, 0b11, 0,
572 (outs SPR:$Sd), (ins DPR:$Dm),
573 NoItinerary, "vcvtt", ".f16.f64\t$Sd, $Dm",
574 []>, Requires<[HasV8FP]> {
575 // Instruction operands.
579 // Encode instruction operands.
580 let Inst{15-12} = Sd{4-1};
581 let Inst{22} = Sd{0};
582 let Inst{3-0} = Dm{3-0};
586 def VNEGD : ADuI<0b11101, 0b11, 0b0001, 0b01, 0,
587 (outs DPR:$Dd), (ins DPR:$Dm),
588 IIC_fpUNA64, "vneg", ".f64\t$Dd, $Dm",
589 [(set DPR:$Dd, (fneg (f64 DPR:$Dm)))]>;
591 def VNEGS : ASuIn<0b11101, 0b11, 0b0001, 0b01, 0,
592 (outs SPR:$Sd), (ins SPR:$Sm),
593 IIC_fpUNA32, "vneg", ".f32\t$Sd, $Sm",
594 [(set SPR:$Sd, (fneg SPR:$Sm))]> {
595 // Some single precision VFP instructions may be executed on both NEON and
596 // VFP pipelines on A8.
597 let D = VFPNeonA8Domain;
600 def VSQRTD : ADuI<0b11101, 0b11, 0b0001, 0b11, 0,
601 (outs DPR:$Dd), (ins DPR:$Dm),
602 IIC_fpSQRT64, "vsqrt", ".f64\t$Dd, $Dm",
603 [(set DPR:$Dd, (fsqrt (f64 DPR:$Dm)))]>;
605 def VSQRTS : ASuI<0b11101, 0b11, 0b0001, 0b11, 0,
606 (outs SPR:$Sd), (ins SPR:$Sm),
607 IIC_fpSQRT32, "vsqrt", ".f32\t$Sd, $Sm",
608 [(set SPR:$Sd, (fsqrt SPR:$Sm))]>;
610 let neverHasSideEffects = 1 in {
611 def VMOVD : ADuI<0b11101, 0b11, 0b0000, 0b01, 0,
612 (outs DPR:$Dd), (ins DPR:$Dm),
613 IIC_fpUNA64, "vmov", ".f64\t$Dd, $Dm", []>;
615 def VMOVS : ASuI<0b11101, 0b11, 0b0000, 0b01, 0,
616 (outs SPR:$Sd), (ins SPR:$Sm),
617 IIC_fpUNA32, "vmov", ".f32\t$Sd, $Sm", []>;
618 } // neverHasSideEffects
620 //===----------------------------------------------------------------------===//
621 // FP <-> GPR Copies. Int <-> FP Conversions.
624 def VMOVRS : AVConv2I<0b11100001, 0b1010,
625 (outs GPR:$Rt), (ins SPR:$Sn),
626 IIC_fpMOVSI, "vmov", "\t$Rt, $Sn",
627 [(set GPR:$Rt, (bitconvert SPR:$Sn))]> {
628 // Instruction operands.
632 // Encode instruction operands.
633 let Inst{19-16} = Sn{4-1};
635 let Inst{15-12} = Rt;
637 let Inst{6-5} = 0b00;
638 let Inst{3-0} = 0b0000;
640 // Some single precision VFP instructions may be executed on both NEON and VFP
642 let D = VFPNeonDomain;
645 // Bitcast i32 -> f32. NEON prefers to use VMOVDRR.
646 def VMOVSR : AVConv4I<0b11100000, 0b1010,
647 (outs SPR:$Sn), (ins GPR:$Rt),
648 IIC_fpMOVIS, "vmov", "\t$Sn, $Rt",
649 [(set SPR:$Sn, (bitconvert GPR:$Rt))]>,
650 Requires<[HasVFP2, UseVMOVSR]> {
651 // Instruction operands.
655 // Encode instruction operands.
656 let Inst{19-16} = Sn{4-1};
658 let Inst{15-12} = Rt;
660 let Inst{6-5} = 0b00;
661 let Inst{3-0} = 0b0000;
663 // Some single precision VFP instructions may be executed on both NEON and VFP
665 let D = VFPNeonDomain;
668 let neverHasSideEffects = 1 in {
669 def VMOVRRD : AVConv3I<0b11000101, 0b1011,
670 (outs GPR:$Rt, GPR:$Rt2), (ins DPR:$Dm),
671 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $Dm",
672 [/* FIXME: Can't write pattern for multiple result instr*/]> {
673 // Instruction operands.
678 // Encode instruction operands.
679 let Inst{3-0} = Dm{3-0};
681 let Inst{15-12} = Rt;
682 let Inst{19-16} = Rt2;
684 let Inst{7-6} = 0b00;
686 // Some single precision VFP instructions may be executed on both NEON and VFP
688 let D = VFPNeonDomain;
691 def VMOVRRS : AVConv3I<0b11000101, 0b1010,
692 (outs GPR:$Rt, GPR:$Rt2), (ins SPR:$src1, SPR:$src2),
693 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $src1, $src2",
694 [/* For disassembly only; pattern left blank */]> {
699 // Encode instruction operands.
700 let Inst{3-0} = src1{4-1};
701 let Inst{5} = src1{0};
702 let Inst{15-12} = Rt;
703 let Inst{19-16} = Rt2;
705 let Inst{7-6} = 0b00;
707 // Some single precision VFP instructions may be executed on both NEON and VFP
709 let D = VFPNeonDomain;
710 let DecoderMethod = "DecodeVMOVRRS";
712 } // neverHasSideEffects
717 def VMOVDRR : AVConv5I<0b11000100, 0b1011,
718 (outs DPR:$Dm), (ins GPR:$Rt, GPR:$Rt2),
719 IIC_fpMOVID, "vmov", "\t$Dm, $Rt, $Rt2",
720 [(set DPR:$Dm, (arm_fmdrr GPR:$Rt, GPR:$Rt2))]> {
721 // Instruction operands.
726 // Encode instruction operands.
727 let Inst{3-0} = Dm{3-0};
729 let Inst{15-12} = Rt;
730 let Inst{19-16} = Rt2;
732 let Inst{7-6} = 0b00;
734 // Some single precision VFP instructions may be executed on both NEON and VFP
736 let D = VFPNeonDomain;
739 let neverHasSideEffects = 1 in
740 def VMOVSRR : AVConv5I<0b11000100, 0b1010,
741 (outs SPR:$dst1, SPR:$dst2), (ins GPR:$src1, GPR:$src2),
742 IIC_fpMOVID, "vmov", "\t$dst1, $dst2, $src1, $src2",
743 [/* For disassembly only; pattern left blank */]> {
744 // Instruction operands.
749 // Encode instruction operands.
750 let Inst{3-0} = dst1{4-1};
751 let Inst{5} = dst1{0};
752 let Inst{15-12} = src1;
753 let Inst{19-16} = src2;
755 let Inst{7-6} = 0b00;
757 // Some single precision VFP instructions may be executed on both NEON and VFP
759 let D = VFPNeonDomain;
761 let DecoderMethod = "DecodeVMOVSRR";
767 // FMRX: SPR system reg -> GPR
769 // FMXR: GPR -> VFP system reg
774 class AVConv1IDs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
775 bits<4> opcod4, dag oops, dag iops,
776 InstrItinClass itin, string opc, string asm,
778 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
780 // Instruction operands.
784 // Encode instruction operands.
785 let Inst{3-0} = Sm{4-1};
787 let Inst{15-12} = Dd{3-0};
788 let Inst{22} = Dd{4};
791 class AVConv1InSs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
792 bits<4> opcod4, dag oops, dag iops,InstrItinClass itin,
793 string opc, string asm, list<dag> pattern>
794 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
796 // Instruction operands.
800 // Encode instruction operands.
801 let Inst{3-0} = Sm{4-1};
803 let Inst{15-12} = Sd{4-1};
804 let Inst{22} = Sd{0};
807 def VSITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
808 (outs DPR:$Dd), (ins SPR:$Sm),
809 IIC_fpCVTID, "vcvt", ".f64.s32\t$Dd, $Sm",
810 [(set DPR:$Dd, (f64 (arm_sitof SPR:$Sm)))]> {
811 let Inst{7} = 1; // s32
814 def VSITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
815 (outs SPR:$Sd),(ins SPR:$Sm),
816 IIC_fpCVTIS, "vcvt", ".f32.s32\t$Sd, $Sm",
817 [(set SPR:$Sd, (arm_sitof SPR:$Sm))]> {
818 let Inst{7} = 1; // s32
820 // Some single precision VFP instructions may be executed on both NEON and
821 // VFP pipelines on A8.
822 let D = VFPNeonA8Domain;
825 def VUITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
826 (outs DPR:$Dd), (ins SPR:$Sm),
827 IIC_fpCVTID, "vcvt", ".f64.u32\t$Dd, $Sm",
828 [(set DPR:$Dd, (f64 (arm_uitof SPR:$Sm)))]> {
829 let Inst{7} = 0; // u32
832 def VUITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
833 (outs SPR:$Sd), (ins SPR:$Sm),
834 IIC_fpCVTIS, "vcvt", ".f32.u32\t$Sd, $Sm",
835 [(set SPR:$Sd, (arm_uitof SPR:$Sm))]> {
836 let Inst{7} = 0; // u32
838 // Some single precision VFP instructions may be executed on both NEON and
839 // VFP pipelines on A8.
840 let D = VFPNeonA8Domain;
845 class AVConv1IsD_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
846 bits<4> opcod4, dag oops, dag iops,
847 InstrItinClass itin, string opc, string asm,
849 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
851 // Instruction operands.
855 // Encode instruction operands.
856 let Inst{3-0} = Dm{3-0};
858 let Inst{15-12} = Sd{4-1};
859 let Inst{22} = Sd{0};
862 class AVConv1InsS_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
863 bits<4> opcod4, dag oops, dag iops,
864 InstrItinClass itin, string opc, string asm,
866 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
868 // Instruction operands.
872 // Encode instruction operands.
873 let Inst{3-0} = Sm{4-1};
875 let Inst{15-12} = Sd{4-1};
876 let Inst{22} = Sd{0};
879 // Always set Z bit in the instruction, i.e. "round towards zero" variants.
880 def VTOSIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
881 (outs SPR:$Sd), (ins DPR:$Dm),
882 IIC_fpCVTDI, "vcvt", ".s32.f64\t$Sd, $Dm",
883 [(set SPR:$Sd, (arm_ftosi (f64 DPR:$Dm)))]> {
884 let Inst{7} = 1; // Z bit
887 def VTOSIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
888 (outs SPR:$Sd), (ins SPR:$Sm),
889 IIC_fpCVTSI, "vcvt", ".s32.f32\t$Sd, $Sm",
890 [(set SPR:$Sd, (arm_ftosi SPR:$Sm))]> {
891 let Inst{7} = 1; // Z bit
893 // Some single precision VFP instructions may be executed on both NEON and
894 // VFP pipelines on A8.
895 let D = VFPNeonA8Domain;
898 def VTOUIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
899 (outs SPR:$Sd), (ins DPR:$Dm),
900 IIC_fpCVTDI, "vcvt", ".u32.f64\t$Sd, $Dm",
901 [(set SPR:$Sd, (arm_ftoui (f64 DPR:$Dm)))]> {
902 let Inst{7} = 1; // Z bit
905 def VTOUIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
906 (outs SPR:$Sd), (ins SPR:$Sm),
907 IIC_fpCVTSI, "vcvt", ".u32.f32\t$Sd, $Sm",
908 [(set SPR:$Sd, (arm_ftoui SPR:$Sm))]> {
909 let Inst{7} = 1; // Z bit
911 // Some single precision VFP instructions may be executed on both NEON and
912 // VFP pipelines on A8.
913 let D = VFPNeonA8Domain;
916 // And the Z bit '0' variants, i.e. use the rounding mode specified by FPSCR.
917 let Uses = [FPSCR] in {
918 // FIXME: Verify encoding after integrated assembler is working.
919 def VTOSIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
920 (outs SPR:$Sd), (ins DPR:$Dm),
921 IIC_fpCVTDI, "vcvtr", ".s32.f64\t$Sd, $Dm",
922 [(set SPR:$Sd, (int_arm_vcvtr (f64 DPR:$Dm)))]>{
923 let Inst{7} = 0; // Z bit
926 def VTOSIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
927 (outs SPR:$Sd), (ins SPR:$Sm),
928 IIC_fpCVTSI, "vcvtr", ".s32.f32\t$Sd, $Sm",
929 [(set SPR:$Sd, (int_arm_vcvtr SPR:$Sm))]> {
930 let Inst{7} = 0; // Z bit
933 def VTOUIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
934 (outs SPR:$Sd), (ins DPR:$Dm),
935 IIC_fpCVTDI, "vcvtr", ".u32.f64\t$Sd, $Dm",
936 [(set SPR:$Sd, (int_arm_vcvtru(f64 DPR:$Dm)))]>{
937 let Inst{7} = 0; // Z bit
940 def VTOUIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
941 (outs SPR:$Sd), (ins SPR:$Sm),
942 IIC_fpCVTSI, "vcvtr", ".u32.f32\t$Sd, $Sm",
943 [(set SPR:$Sd, (int_arm_vcvtru SPR:$Sm))]> {
944 let Inst{7} = 0; // Z bit
948 // Convert between floating-point and fixed-point
949 // Data type for fixed-point naming convention:
950 // S16 (U=0, sx=0) -> SH
951 // U16 (U=1, sx=0) -> UH
952 // S32 (U=0, sx=1) -> SL
953 // U32 (U=1, sx=1) -> UL
955 let Constraints = "$a = $dst" in {
957 // FP to Fixed-Point:
959 // Single Precision register
960 class AVConv1XInsS_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
961 bit op5, dag oops, dag iops, InstrItinClass itin,
962 string opc, string asm, list<dag> pattern>
963 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern>,
964 Sched<[WriteCvtFP]> {
966 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
967 let Inst{22} = dst{0};
968 let Inst{15-12} = dst{4-1};
971 // Double Precision register
972 class AVConv1XInsD_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
973 bit op5, dag oops, dag iops, InstrItinClass itin,
974 string opc, string asm, list<dag> pattern>
975 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern>,
976 Sched<[WriteCvtFP]> {
978 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
979 let Inst{22} = dst{4};
980 let Inst{15-12} = dst{3-0};
983 def VTOSHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 0,
984 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
985 IIC_fpCVTSI, "vcvt", ".s16.f32\t$dst, $a, $fbits", []> {
986 // Some single precision VFP instructions may be executed on both NEON and
987 // VFP pipelines on A8.
988 let D = VFPNeonA8Domain;
991 def VTOUHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 0,
992 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
993 IIC_fpCVTSI, "vcvt", ".u16.f32\t$dst, $a, $fbits", []> {
994 // Some single precision VFP instructions may be executed on both NEON and
995 // VFP pipelines on A8.
996 let D = VFPNeonA8Domain;
999 def VTOSLS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 1,
1000 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1001 IIC_fpCVTSI, "vcvt", ".s32.f32\t$dst, $a, $fbits", []> {
1002 // Some single precision VFP instructions may be executed on both NEON and
1003 // VFP pipelines on A8.
1004 let D = VFPNeonA8Domain;
1007 def VTOULS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 1,
1008 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1009 IIC_fpCVTSI, "vcvt", ".u32.f32\t$dst, $a, $fbits", []> {
1010 // Some single precision VFP instructions may be executed on both NEON and
1011 // VFP pipelines on A8.
1012 let D = VFPNeonA8Domain;
1015 def VTOSHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 0,
1016 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1017 IIC_fpCVTDI, "vcvt", ".s16.f64\t$dst, $a, $fbits", []>;
1019 def VTOUHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 0,
1020 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1021 IIC_fpCVTDI, "vcvt", ".u16.f64\t$dst, $a, $fbits", []>;
1023 def VTOSLD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 1,
1024 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1025 IIC_fpCVTDI, "vcvt", ".s32.f64\t$dst, $a, $fbits", []>;
1027 def VTOULD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 1,
1028 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1029 IIC_fpCVTDI, "vcvt", ".u32.f64\t$dst, $a, $fbits", []>;
1031 // Fixed-Point to FP:
1033 def VSHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 0,
1034 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1035 IIC_fpCVTIS, "vcvt", ".f32.s16\t$dst, $a, $fbits", []> {
1036 // Some single precision VFP instructions may be executed on both NEON and
1037 // VFP pipelines on A8.
1038 let D = VFPNeonA8Domain;
1041 def VUHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 0,
1042 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1043 IIC_fpCVTIS, "vcvt", ".f32.u16\t$dst, $a, $fbits", []> {
1044 // Some single precision VFP instructions may be executed on both NEON and
1045 // VFP pipelines on A8.
1046 let D = VFPNeonA8Domain;
1049 def VSLTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 1,
1050 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1051 IIC_fpCVTIS, "vcvt", ".f32.s32\t$dst, $a, $fbits", []> {
1052 // Some single precision VFP instructions may be executed on both NEON and
1053 // VFP pipelines on A8.
1054 let D = VFPNeonA8Domain;
1057 def VULTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 1,
1058 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1059 IIC_fpCVTIS, "vcvt", ".f32.u32\t$dst, $a, $fbits", []> {
1060 // Some single precision VFP instructions may be executed on both NEON and
1061 // VFP pipelines on A8.
1062 let D = VFPNeonA8Domain;
1065 def VSHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 0,
1066 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1067 IIC_fpCVTID, "vcvt", ".f64.s16\t$dst, $a, $fbits", []>;
1069 def VUHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 0,
1070 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1071 IIC_fpCVTID, "vcvt", ".f64.u16\t$dst, $a, $fbits", []>;
1073 def VSLTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 1,
1074 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1075 IIC_fpCVTID, "vcvt", ".f64.s32\t$dst, $a, $fbits", []>;
1077 def VULTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 1,
1078 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1079 IIC_fpCVTID, "vcvt", ".f64.u32\t$dst, $a, $fbits", []>;
1081 } // End of 'let Constraints = "$a = $dst" in'
1083 //===----------------------------------------------------------------------===//
1084 // FP Multiply-Accumulate Operations.
1087 def VMLAD : ADbI<0b11100, 0b00, 0, 0,
1088 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1089 IIC_fpMAC64, "vmla", ".f64\t$Dd, $Dn, $Dm",
1090 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1091 (f64 DPR:$Ddin)))]>,
1092 RegConstraint<"$Ddin = $Dd">,
1093 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1095 def VMLAS : ASbIn<0b11100, 0b00, 0, 0,
1096 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1097 IIC_fpMAC32, "vmla", ".f32\t$Sd, $Sn, $Sm",
1098 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
1100 RegConstraint<"$Sdin = $Sd">,
1101 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1102 // Some single precision VFP instructions may be executed on both NEON and
1103 // VFP pipelines on A8.
1104 let D = VFPNeonA8Domain;
1107 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1108 (VMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1109 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1110 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1111 (VMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1112 Requires<[HasVFP2,DontUseNEONForFP, UseFPVMLx,DontUseFusedMAC]>;
1114 def VMLSD : ADbI<0b11100, 0b00, 1, 0,
1115 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1116 IIC_fpMAC64, "vmls", ".f64\t$Dd, $Dn, $Dm",
1117 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1118 (f64 DPR:$Ddin)))]>,
1119 RegConstraint<"$Ddin = $Dd">,
1120 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1122 def VMLSS : ASbIn<0b11100, 0b00, 1, 0,
1123 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1124 IIC_fpMAC32, "vmls", ".f32\t$Sd, $Sn, $Sm",
1125 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1127 RegConstraint<"$Sdin = $Sd">,
1128 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1129 // Some single precision VFP instructions may be executed on both NEON and
1130 // VFP pipelines on A8.
1131 let D = VFPNeonA8Domain;
1134 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1135 (VMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
1136 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1137 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1138 (VMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
1139 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1141 def VNMLAD : ADbI<0b11100, 0b01, 1, 0,
1142 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1143 IIC_fpMAC64, "vnmla", ".f64\t$Dd, $Dn, $Dm",
1144 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1145 (f64 DPR:$Ddin)))]>,
1146 RegConstraint<"$Ddin = $Dd">,
1147 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1149 def VNMLAS : ASbI<0b11100, 0b01, 1, 0,
1150 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1151 IIC_fpMAC32, "vnmla", ".f32\t$Sd, $Sn, $Sm",
1152 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1154 RegConstraint<"$Sdin = $Sd">,
1155 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1156 // Some single precision VFP instructions may be executed on both NEON and
1157 // VFP pipelines on A8.
1158 let D = VFPNeonA8Domain;
1161 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
1162 (VNMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1163 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1164 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
1165 (VNMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1166 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1168 def VNMLSD : ADbI<0b11100, 0b01, 0, 0,
1169 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1170 IIC_fpMAC64, "vnmls", ".f64\t$Dd, $Dn, $Dm",
1171 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1172 (f64 DPR:$Ddin)))]>,
1173 RegConstraint<"$Ddin = $Dd">,
1174 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1176 def VNMLSS : ASbI<0b11100, 0b01, 0, 0,
1177 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1178 IIC_fpMAC32, "vnmls", ".f32\t$Sd, $Sn, $Sm",
1179 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
1180 RegConstraint<"$Sdin = $Sd">,
1181 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1182 // Some single precision VFP instructions may be executed on both NEON and
1183 // VFP pipelines on A8.
1184 let D = VFPNeonA8Domain;
1187 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
1188 (VNMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
1189 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1190 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
1191 (VNMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
1192 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1194 //===----------------------------------------------------------------------===//
1195 // Fused FP Multiply-Accumulate Operations.
1197 def VFMAD : ADbI<0b11101, 0b10, 0, 0,
1198 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1199 IIC_fpFMAC64, "vfma", ".f64\t$Dd, $Dn, $Dm",
1200 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1201 (f64 DPR:$Ddin)))]>,
1202 RegConstraint<"$Ddin = $Dd">,
1203 Requires<[HasVFP4,UseFusedMAC]>;
1205 def VFMAS : ASbIn<0b11101, 0b10, 0, 0,
1206 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1207 IIC_fpFMAC32, "vfma", ".f32\t$Sd, $Sn, $Sm",
1208 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
1210 RegConstraint<"$Sdin = $Sd">,
1211 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1212 // Some single precision VFP instructions may be executed on both NEON and
1216 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1217 (VFMAD DPR:$dstin, DPR:$a, DPR:$b)>,
1218 Requires<[HasVFP4,UseFusedMAC]>;
1219 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1220 (VFMAS SPR:$dstin, SPR:$a, SPR:$b)>,
1221 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1223 // Match @llvm.fma.* intrinsics
1224 // (fma x, y, z) -> (vfms z, x, y)
1225 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, DPR:$Ddin)),
1226 (VFMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1227 Requires<[HasVFP4]>;
1228 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, SPR:$Sdin)),
1229 (VFMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1230 Requires<[HasVFP4]>;
1232 def VFMSD : ADbI<0b11101, 0b10, 1, 0,
1233 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1234 IIC_fpFMAC64, "vfms", ".f64\t$Dd, $Dn, $Dm",
1235 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1236 (f64 DPR:$Ddin)))]>,
1237 RegConstraint<"$Ddin = $Dd">,
1238 Requires<[HasVFP4,UseFusedMAC]>;
1240 def VFMSS : ASbIn<0b11101, 0b10, 1, 0,
1241 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1242 IIC_fpFMAC32, "vfms", ".f32\t$Sd, $Sn, $Sm",
1243 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1245 RegConstraint<"$Sdin = $Sd">,
1246 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1247 // Some single precision VFP instructions may be executed on both NEON and
1251 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1252 (VFMSD DPR:$dstin, DPR:$a, DPR:$b)>,
1253 Requires<[HasVFP4,UseFusedMAC]>;
1254 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1255 (VFMSS SPR:$dstin, SPR:$a, SPR:$b)>,
1256 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1258 // Match @llvm.fma.* intrinsics
1259 // (fma (fneg x), y, z) -> (vfms z, x, y)
1260 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin)),
1261 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1262 Requires<[HasVFP4]>;
1263 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin)),
1264 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1265 Requires<[HasVFP4]>;
1266 // (fma x, (fneg y), z) -> (vfms z, x, y)
1267 def : Pat<(f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin)),
1268 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1269 Requires<[HasVFP4]>;
1270 def : Pat<(f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin)),
1271 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1272 Requires<[HasVFP4]>;
1274 def VFNMAD : ADbI<0b11101, 0b01, 1, 0,
1275 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1276 IIC_fpFMAC64, "vfnma", ".f64\t$Dd, $Dn, $Dm",
1277 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1278 (f64 DPR:$Ddin)))]>,
1279 RegConstraint<"$Ddin = $Dd">,
1280 Requires<[HasVFP4,UseFusedMAC]>;
1282 def VFNMAS : ASbI<0b11101, 0b01, 1, 0,
1283 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1284 IIC_fpFMAC32, "vfnma", ".f32\t$Sd, $Sn, $Sm",
1285 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1287 RegConstraint<"$Sdin = $Sd">,
1288 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1289 // Some single precision VFP instructions may be executed on both NEON and
1293 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
1294 (VFNMAD DPR:$dstin, DPR:$a, DPR:$b)>,
1295 Requires<[HasVFP4,UseFusedMAC]>;
1296 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
1297 (VFNMAS SPR:$dstin, SPR:$a, SPR:$b)>,
1298 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1300 // Match @llvm.fma.* intrinsics
1301 // (fneg (fma x, y, z)) -> (vfnma z, x, y)
1302 def : Pat<(fneg (fma (f64 DPR:$Dn), (f64 DPR:$Dm), (f64 DPR:$Ddin))),
1303 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1304 Requires<[HasVFP4]>;
1305 def : Pat<(fneg (fma (f32 SPR:$Sn), (f32 SPR:$Sm), (f32 SPR:$Sdin))),
1306 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1307 Requires<[HasVFP4]>;
1308 // (fma (fneg x), y, (fneg z)) -> (vfnma z, x, y)
1309 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, (fneg DPR:$Ddin))),
1310 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1311 Requires<[HasVFP4]>;
1312 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, (fneg SPR:$Sdin))),
1313 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1314 Requires<[HasVFP4]>;
1316 def VFNMSD : ADbI<0b11101, 0b01, 0, 0,
1317 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1318 IIC_fpFMAC64, "vfnms", ".f64\t$Dd, $Dn, $Dm",
1319 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1320 (f64 DPR:$Ddin)))]>,
1321 RegConstraint<"$Ddin = $Dd">,
1322 Requires<[HasVFP4,UseFusedMAC]>;
1324 def VFNMSS : ASbI<0b11101, 0b01, 0, 0,
1325 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1326 IIC_fpFMAC32, "vfnms", ".f32\t$Sd, $Sn, $Sm",
1327 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
1328 RegConstraint<"$Sdin = $Sd">,
1329 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1330 // Some single precision VFP instructions may be executed on both NEON and
1334 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
1335 (VFNMSD DPR:$dstin, DPR:$a, DPR:$b)>,
1336 Requires<[HasVFP4,UseFusedMAC]>;
1337 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
1338 (VFNMSS SPR:$dstin, SPR:$a, SPR:$b)>,
1339 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1341 // Match @llvm.fma.* intrinsics
1343 // (fma x, y, (fneg z)) -> (vfnms z, x, y))
1344 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, (fneg DPR:$Ddin))),
1345 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1346 Requires<[HasVFP4]>;
1347 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, (fneg SPR:$Sdin))),
1348 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1349 Requires<[HasVFP4]>;
1350 // (fneg (fma (fneg x), y, z)) -> (vfnms z, x, y)
1351 def : Pat<(fneg (f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin))),
1352 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1353 Requires<[HasVFP4]>;
1354 def : Pat<(fneg (f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin))),
1355 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1356 Requires<[HasVFP4]>;
1357 // (fneg (fma x, (fneg y), z) -> (vfnms z, x, y)
1358 def : Pat<(fneg (f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin))),
1359 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1360 Requires<[HasVFP4]>;
1361 def : Pat<(fneg (f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin))),
1362 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1363 Requires<[HasVFP4]>;
1365 //===----------------------------------------------------------------------===//
1366 // FP Conditional moves.
1369 let neverHasSideEffects = 1 in {
1370 def VMOVDcc : ARMPseudoInst<(outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm, pred:$p),
1372 [/*(set DPR:$Dd, (ARMcmov DPR:$Dn, DPR:$Dm, imm:$cc))*/]>,
1373 RegConstraint<"$Dn = $Dd">;
1375 def VMOVScc : ARMPseudoInst<(outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm, pred:$p),
1377 [/*(set SPR:$Sd, (ARMcmov SPR:$Sn, SPR:$Sm, imm:$cc))*/]>,
1378 RegConstraint<"$Sn = $Sd">;
1379 } // neverHasSideEffects
1381 //===----------------------------------------------------------------------===//
1382 // Move from VFP System Register to ARM core register.
1385 class MovFromVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
1387 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
1389 // Instruction operand.
1392 let Inst{27-20} = 0b11101111;
1393 let Inst{19-16} = opc19_16;
1394 let Inst{15-12} = Rt;
1395 let Inst{11-8} = 0b1010;
1397 let Inst{6-5} = 0b00;
1399 let Inst{3-0} = 0b0000;
1402 // APSR is the application level alias of CPSR. This FPSCR N, Z, C, V flags
1404 let Defs = [CPSR], Uses = [FPSCR_NZCV], Rt = 0b1111 /* apsr_nzcv */ in
1405 def FMSTAT : MovFromVFP<0b0001 /* fpscr */, (outs), (ins),
1406 "vmrs", "\tAPSR_nzcv, fpscr", [(arm_fmstat)]>;
1408 // Application level FPSCR -> GPR
1409 let hasSideEffects = 1, Uses = [FPSCR] in
1410 def VMRS : MovFromVFP<0b0001 /* fpscr */, (outs GPR:$Rt), (ins),
1411 "vmrs", "\t$Rt, fpscr",
1412 [(set GPR:$Rt, (int_arm_get_fpscr))]>;
1414 // System level FPEXC, FPSID -> GPR
1415 let Uses = [FPSCR] in {
1416 def VMRS_FPEXC : MovFromVFP<0b1000 /* fpexc */, (outs GPR:$Rt), (ins),
1417 "vmrs", "\t$Rt, fpexc", []>;
1418 def VMRS_FPSID : MovFromVFP<0b0000 /* fpsid */, (outs GPR:$Rt), (ins),
1419 "vmrs", "\t$Rt, fpsid", []>;
1420 def VMRS_MVFR0 : MovFromVFP<0b0111 /* mvfr0 */, (outs GPR:$Rt), (ins),
1421 "vmrs", "\t$Rt, mvfr0", []>;
1422 def VMRS_MVFR1 : MovFromVFP<0b0110 /* mvfr1 */, (outs GPR:$Rt), (ins),
1423 "vmrs", "\t$Rt, mvfr1", []>;
1424 def VMRS_FPINST : MovFromVFP<0b1001 /* fpinst */, (outs GPR:$Rt), (ins),
1425 "vmrs", "\t$Rt, fpinst", []>;
1426 def VMRS_FPINST2 : MovFromVFP<0b1010 /* fpinst2 */, (outs GPR:$Rt), (ins),
1427 "vmrs", "\t$Rt, fpinst2", []>;
1430 //===----------------------------------------------------------------------===//
1431 // Move from ARM core register to VFP System Register.
1434 class MovToVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
1436 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
1438 // Instruction operand.
1441 // Encode instruction operand.
1442 let Inst{15-12} = src;
1444 let Inst{27-20} = 0b11101110;
1445 let Inst{19-16} = opc19_16;
1446 let Inst{11-8} = 0b1010;
1451 let Defs = [FPSCR] in {
1452 // Application level GPR -> FPSCR
1453 def VMSR : MovToVFP<0b0001 /* fpscr */, (outs), (ins GPR:$src),
1454 "vmsr", "\tfpscr, $src", [(int_arm_set_fpscr GPR:$src)]>;
1455 // System level GPR -> FPEXC
1456 def VMSR_FPEXC : MovToVFP<0b1000 /* fpexc */, (outs), (ins GPR:$src),
1457 "vmsr", "\tfpexc, $src", []>;
1458 // System level GPR -> FPSID
1459 def VMSR_FPSID : MovToVFP<0b0000 /* fpsid */, (outs), (ins GPR:$src),
1460 "vmsr", "\tfpsid, $src", []>;
1462 def VMSR_FPINST : MovToVFP<0b1001 /* fpinst */, (outs), (ins GPR:$src),
1463 "vmsr", "\tfpinst, $src", []>;
1464 def VMSR_FPINST2 : MovToVFP<0b1010 /* fpinst2 */, (outs), (ins GPR:$src),
1465 "vmsr", "\tfpinst2, $src", []>;
1468 //===----------------------------------------------------------------------===//
1472 // Materialize FP immediates. VFP3 only.
1473 let isReMaterializable = 1 in {
1474 def FCONSTD : VFPAI<(outs DPR:$Dd), (ins vfp_f64imm:$imm),
1475 VFPMiscFrm, IIC_fpUNA64,
1476 "vmov", ".f64\t$Dd, $imm",
1477 [(set DPR:$Dd, vfp_f64imm:$imm)]>, Requires<[HasVFP3]> {
1481 let Inst{27-23} = 0b11101;
1482 let Inst{22} = Dd{4};
1483 let Inst{21-20} = 0b11;
1484 let Inst{19-16} = imm{7-4};
1485 let Inst{15-12} = Dd{3-0};
1486 let Inst{11-9} = 0b101;
1487 let Inst{8} = 1; // Double precision.
1488 let Inst{7-4} = 0b0000;
1489 let Inst{3-0} = imm{3-0};
1492 def FCONSTS : VFPAI<(outs SPR:$Sd), (ins vfp_f32imm:$imm),
1493 VFPMiscFrm, IIC_fpUNA32,
1494 "vmov", ".f32\t$Sd, $imm",
1495 [(set SPR:$Sd, vfp_f32imm:$imm)]>, Requires<[HasVFP3]> {
1499 let Inst{27-23} = 0b11101;
1500 let Inst{22} = Sd{0};
1501 let Inst{21-20} = 0b11;
1502 let Inst{19-16} = imm{7-4};
1503 let Inst{15-12} = Sd{4-1};
1504 let Inst{11-9} = 0b101;
1505 let Inst{8} = 0; // Single precision.
1506 let Inst{7-4} = 0b0000;
1507 let Inst{3-0} = imm{3-0};
1511 //===----------------------------------------------------------------------===//
1512 // Assembler aliases.
1514 // A few mnemnoic aliases for pre-unifixed syntax. We don't guarantee to
1515 // support them all, but supporting at least some of the basics is
1516 // good to be friendly.
1517 def : VFP2MnemonicAlias<"flds", "vldr">;
1518 def : VFP2MnemonicAlias<"fldd", "vldr">;
1519 def : VFP2MnemonicAlias<"fmrs", "vmov">;
1520 def : VFP2MnemonicAlias<"fmsr", "vmov">;
1521 def : VFP2MnemonicAlias<"fsqrts", "vsqrt">;
1522 def : VFP2MnemonicAlias<"fsqrtd", "vsqrt">;
1523 def : VFP2MnemonicAlias<"fadds", "vadd.f32">;
1524 def : VFP2MnemonicAlias<"faddd", "vadd.f64">;
1525 def : VFP2MnemonicAlias<"fmrdd", "vmov">;
1526 def : VFP2MnemonicAlias<"fmrds", "vmov">;
1527 def : VFP2MnemonicAlias<"fmrrd", "vmov">;
1528 def : VFP2MnemonicAlias<"fmdrr", "vmov">;
1529 def : VFP2MnemonicAlias<"fmuls", "vmul.f32">;
1530 def : VFP2MnemonicAlias<"fmuld", "vmul.f64">;
1531 def : VFP2MnemonicAlias<"fnegs", "vneg.f32">;
1532 def : VFP2MnemonicAlias<"fnegd", "vneg.f64">;
1533 def : VFP2MnemonicAlias<"ftosizd", "vcvt.s32.f64">;
1534 def : VFP2MnemonicAlias<"ftosid", "vcvtr.s32.f64">;
1535 def : VFP2MnemonicAlias<"ftosizs", "vcvt.s32.f32">;
1536 def : VFP2MnemonicAlias<"ftosis", "vcvtr.s32.f32">;
1537 def : VFP2MnemonicAlias<"ftouizd", "vcvt.u32.f64">;
1538 def : VFP2MnemonicAlias<"ftouid", "vcvtr.u32.f64">;
1539 def : VFP2MnemonicAlias<"ftouizs", "vcvt.u32.f32">;
1540 def : VFP2MnemonicAlias<"ftouis", "vcvtr.u32.f32">;
1541 def : VFP2MnemonicAlias<"fsitod", "vcvt.f64.s32">;
1542 def : VFP2MnemonicAlias<"fsitos", "vcvt.f32.s32">;
1543 def : VFP2MnemonicAlias<"fuitod", "vcvt.f64.u32">;
1544 def : VFP2MnemonicAlias<"fuitos", "vcvt.f32.u32">;
1545 def : VFP2MnemonicAlias<"fsts", "vstr">;
1546 def : VFP2MnemonicAlias<"fstd", "vstr">;
1547 def : VFP2MnemonicAlias<"fmacd", "vmla.f64">;
1548 def : VFP2MnemonicAlias<"fmacs", "vmla.f32">;
1549 def : VFP2MnemonicAlias<"fcpys", "vmov.f32">;
1550 def : VFP2MnemonicAlias<"fcpyd", "vmov.f64">;
1551 def : VFP2MnemonicAlias<"fcmps", "vcmp.f32">;
1552 def : VFP2MnemonicAlias<"fcmpd", "vcmp.f64">;
1553 def : VFP2MnemonicAlias<"fdivs", "vdiv.f32">;
1554 def : VFP2MnemonicAlias<"fdivd", "vdiv.f64">;
1555 def : VFP2MnemonicAlias<"fmrx", "vmrs">;
1556 def : VFP2MnemonicAlias<"fmxr", "vmsr">;
1558 // Be friendly and accept the old form of zero-compare
1559 def : VFP2InstAlias<"fcmpzd${p} $val", (VCMPZD DPR:$val, pred:$p)>;
1560 def : VFP2InstAlias<"fcmpzs${p} $val", (VCMPZS SPR:$val, pred:$p)>;
1563 def : VFP2InstAlias<"fmstat${p}", (FMSTAT pred:$p)>;
1564 def : VFP2InstAlias<"fadds${p} $Sd, $Sn, $Sm",
1565 (VADDS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
1566 def : VFP2InstAlias<"faddd${p} $Dd, $Dn, $Dm",
1567 (VADDD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
1568 def : VFP2InstAlias<"fsubs${p} $Sd, $Sn, $Sm",
1569 (VSUBS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
1570 def : VFP2InstAlias<"fsubd${p} $Dd, $Dn, $Dm",
1571 (VSUBD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
1573 // No need for the size suffix on VSQRT. It's implied by the register classes.
1574 def : VFP2InstAlias<"vsqrt${p} $Sd, $Sm", (VSQRTS SPR:$Sd, SPR:$Sm, pred:$p)>;
1575 def : VFP2InstAlias<"vsqrt${p} $Dd, $Dm", (VSQRTD DPR:$Dd, DPR:$Dm, pred:$p)>;
1577 // VLDR/VSTR accept an optional type suffix.
1578 def : VFP2InstAlias<"vldr${p}.32 $Sd, $addr",
1579 (VLDRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
1580 def : VFP2InstAlias<"vstr${p}.32 $Sd, $addr",
1581 (VSTRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
1582 def : VFP2InstAlias<"vldr${p}.64 $Dd, $addr",
1583 (VLDRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
1584 def : VFP2InstAlias<"vstr${p}.64 $Dd, $addr",
1585 (VSTRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
1587 // VMOV can accept optional 32-bit or less data type suffix suffix.
1588 def : VFP2InstAlias<"vmov${p}.8 $Rt, $Sn",
1589 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1590 def : VFP2InstAlias<"vmov${p}.16 $Rt, $Sn",
1591 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1592 def : VFP2InstAlias<"vmov${p}.32 $Rt, $Sn",
1593 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1594 def : VFP2InstAlias<"vmov${p}.8 $Sn, $Rt",
1595 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1596 def : VFP2InstAlias<"vmov${p}.16 $Sn, $Rt",
1597 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1598 def : VFP2InstAlias<"vmov${p}.32 $Sn, $Rt",
1599 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1601 def : VFP2InstAlias<"vmov${p}.f64 $Rt, $Rt2, $Dn",
1602 (VMOVRRD GPR:$Rt, GPR:$Rt2, DPR:$Dn, pred:$p)>;
1603 def : VFP2InstAlias<"vmov${p}.f64 $Dn, $Rt, $Rt2",
1604 (VMOVDRR DPR:$Dn, GPR:$Rt, GPR:$Rt2, pred:$p)>;
1606 // VMOVS doesn't need the .f32 to disambiguate from the NEON encoding the way
1608 def : VFP2InstAlias<"vmov${p} $Sd, $Sm",
1609 (VMOVS SPR:$Sd, SPR:$Sm, pred:$p)>;