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,
339 (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
340 NoItinerary, !strconcat("vsel", op, ".f32\t$Sd, $Sn, $Sm"),
341 []>, Requires<[HasV8FP]>;
343 def D : ADbInp<0b11100, opc,
344 (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
345 NoItinerary, !strconcat("vsel", op, ".f64\t$Dd, $Dn, $Dm"),
346 []>, Requires<[HasV8FP]>;
350 defm VSELGT : vsel_inst<"gt", 0b11>;
351 defm VSELGE : vsel_inst<"ge", 0b10>;
352 defm VSELEQ : vsel_inst<"eq", 0b00>;
353 defm VSELVS : vsel_inst<"vs", 0b01>;
355 // Match reassociated forms only if not sign dependent rounding.
356 def : Pat<(fmul (fneg DPR:$a), (f64 DPR:$b)),
357 (VNMULD DPR:$a, DPR:$b)>, Requires<[NoHonorSignDependentRounding]>;
358 def : Pat<(fmul (fneg SPR:$a), SPR:$b),
359 (VNMULS SPR:$a, SPR:$b)>, Requires<[NoHonorSignDependentRounding]>;
361 // These are encoded as unary instructions.
362 let Defs = [FPSCR_NZCV] in {
363 def VCMPED : ADuI<0b11101, 0b11, 0b0100, 0b11, 0,
364 (outs), (ins DPR:$Dd, DPR:$Dm),
365 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, $Dm",
366 [(arm_cmpfp DPR:$Dd, (f64 DPR:$Dm))]>;
368 def VCMPES : ASuI<0b11101, 0b11, 0b0100, 0b11, 0,
369 (outs), (ins SPR:$Sd, SPR:$Sm),
370 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, $Sm",
371 [(arm_cmpfp SPR:$Sd, SPR:$Sm)]> {
372 // Some single precision VFP instructions may be executed on both NEON and
373 // VFP pipelines on A8.
374 let D = VFPNeonA8Domain;
377 // FIXME: Verify encoding after integrated assembler is working.
378 def VCMPD : ADuI<0b11101, 0b11, 0b0100, 0b01, 0,
379 (outs), (ins DPR:$Dd, DPR:$Dm),
380 IIC_fpCMP64, "vcmp", ".f64\t$Dd, $Dm",
381 [/* For disassembly only; pattern left blank */]>;
383 def VCMPS : ASuI<0b11101, 0b11, 0b0100, 0b01, 0,
384 (outs), (ins SPR:$Sd, SPR:$Sm),
385 IIC_fpCMP32, "vcmp", ".f32\t$Sd, $Sm",
386 [/* For disassembly only; pattern left blank */]> {
387 // Some single precision VFP instructions may be executed on both NEON and
388 // VFP pipelines on A8.
389 let D = VFPNeonA8Domain;
391 } // Defs = [FPSCR_NZCV]
393 //===----------------------------------------------------------------------===//
394 // FP Unary Operations.
397 def VABSD : ADuI<0b11101, 0b11, 0b0000, 0b11, 0,
398 (outs DPR:$Dd), (ins DPR:$Dm),
399 IIC_fpUNA64, "vabs", ".f64\t$Dd, $Dm",
400 [(set DPR:$Dd, (fabs (f64 DPR:$Dm)))]>;
402 def VABSS : ASuIn<0b11101, 0b11, 0b0000, 0b11, 0,
403 (outs SPR:$Sd), (ins SPR:$Sm),
404 IIC_fpUNA32, "vabs", ".f32\t$Sd, $Sm",
405 [(set SPR:$Sd, (fabs SPR:$Sm))]> {
406 // Some single precision VFP instructions may be executed on both NEON and
407 // VFP pipelines on A8.
408 let D = VFPNeonA8Domain;
411 let Defs = [FPSCR_NZCV] in {
412 def VCMPEZD : ADuI<0b11101, 0b11, 0b0101, 0b11, 0,
413 (outs), (ins DPR:$Dd),
414 IIC_fpCMP64, "vcmpe", ".f64\t$Dd, #0",
415 [(arm_cmpfp0 (f64 DPR:$Dd))]> {
416 let Inst{3-0} = 0b0000;
420 def VCMPEZS : ASuI<0b11101, 0b11, 0b0101, 0b11, 0,
421 (outs), (ins SPR:$Sd),
422 IIC_fpCMP32, "vcmpe", ".f32\t$Sd, #0",
423 [(arm_cmpfp0 SPR:$Sd)]> {
424 let Inst{3-0} = 0b0000;
427 // Some single precision VFP instructions may be executed on both NEON and
428 // VFP pipelines on A8.
429 let D = VFPNeonA8Domain;
432 // FIXME: Verify encoding after integrated assembler is working.
433 def VCMPZD : ADuI<0b11101, 0b11, 0b0101, 0b01, 0,
434 (outs), (ins DPR:$Dd),
435 IIC_fpCMP64, "vcmp", ".f64\t$Dd, #0",
436 [/* For disassembly only; pattern left blank */]> {
437 let Inst{3-0} = 0b0000;
441 def VCMPZS : ASuI<0b11101, 0b11, 0b0101, 0b01, 0,
442 (outs), (ins SPR:$Sd),
443 IIC_fpCMP32, "vcmp", ".f32\t$Sd, #0",
444 [/* For disassembly only; pattern left blank */]> {
445 let Inst{3-0} = 0b0000;
448 // Some single precision VFP instructions may be executed on both NEON and
449 // VFP pipelines on A8.
450 let D = VFPNeonA8Domain;
452 } // Defs = [FPSCR_NZCV]
454 def VCVTDS : ASuI<0b11101, 0b11, 0b0111, 0b11, 0,
455 (outs DPR:$Dd), (ins SPR:$Sm),
456 IIC_fpCVTDS, "vcvt", ".f64.f32\t$Dd, $Sm",
457 [(set DPR:$Dd, (fextend SPR:$Sm))]> {
458 // Instruction operands.
462 // Encode instruction operands.
463 let Inst{3-0} = Sm{4-1};
465 let Inst{15-12} = Dd{3-0};
466 let Inst{22} = Dd{4};
469 // Special case encoding: bits 11-8 is 0b1011.
470 def VCVTSD : VFPAI<(outs SPR:$Sd), (ins DPR:$Dm), VFPUnaryFrm,
471 IIC_fpCVTSD, "vcvt", ".f32.f64\t$Sd, $Dm",
472 [(set SPR:$Sd, (fround DPR:$Dm))]> {
473 // Instruction operands.
477 // Encode instruction operands.
478 let Inst{3-0} = Dm{3-0};
480 let Inst{15-12} = Sd{4-1};
481 let Inst{22} = Sd{0};
483 let Inst{27-23} = 0b11101;
484 let Inst{21-16} = 0b110111;
485 let Inst{11-8} = 0b1011;
486 let Inst{7-6} = 0b11;
490 // Between half, single and double-precision. For disassembly only.
492 // FIXME: Verify encoding after integrated assembler is working.
493 def VCVTBHS: ASuI<0b11101, 0b11, 0b0010, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
494 /* FIXME */ IIC_fpCVTSH, "vcvtb", ".f32.f16\t$Sd, $Sm",
495 [/* For disassembly only; pattern left blank */]>;
497 def VCVTBSH: ASuI<0b11101, 0b11, 0b0011, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
498 /* FIXME */ IIC_fpCVTHS, "vcvtb", ".f16.f32\t$Sd, $Sm",
499 [/* For disassembly only; pattern left blank */]>;
501 def : Pat<(f32_to_f16 SPR:$a),
502 (i32 (COPY_TO_REGCLASS (VCVTBSH SPR:$a), GPR))>;
504 def : Pat<(f16_to_f32 GPR:$a),
505 (VCVTBHS (COPY_TO_REGCLASS GPR:$a, SPR))>;
507 def VCVTTHS: ASuI<0b11101, 0b11, 0b0010, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
508 /* FIXME */ IIC_fpCVTSH, "vcvtt", ".f32.f16\t$Sd, $Sm",
509 [/* For disassembly only; pattern left blank */]>;
511 def VCVTTSH: ASuI<0b11101, 0b11, 0b0011, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
512 /* FIXME */ IIC_fpCVTHS, "vcvtt", ".f16.f32\t$Sd, $Sm",
513 [/* For disassembly only; pattern left blank */]>;
515 def VCVTBHD : ADuI<0b11101, 0b11, 0b0010, 0b01, 0,
516 (outs DPR:$Dd), (ins SPR:$Sm),
517 NoItinerary, "vcvtb", ".f64.f16\t$Dd, $Sm",
518 []>, Requires<[HasV8FP]> {
519 // Instruction operands.
522 // Encode instruction operands.
523 let Inst{3-0} = Sm{4-1};
527 def VCVTBDH : ADuI<0b11101, 0b11, 0b0011, 0b01, 0,
528 (outs SPR:$Sd), (ins DPR:$Dm),
529 NoItinerary, "vcvtb", ".f16.f64\t$Sd, $Dm",
530 []>, Requires<[HasV8FP]> {
531 // Instruction operands.
535 // Encode instruction operands.
536 let Inst{3-0} = Dm{3-0};
538 let Inst{15-12} = Sd{4-1};
539 let Inst{22} = Sd{0};
542 def VCVTTHD : ADuI<0b11101, 0b11, 0b0010, 0b11, 0,
543 (outs DPR:$Dd), (ins SPR:$Sm),
544 NoItinerary, "vcvtt", ".f64.f16\t$Dd, $Sm",
545 []>, Requires<[HasV8FP]> {
546 // Instruction operands.
549 // Encode instruction operands.
550 let Inst{3-0} = Sm{4-1};
554 def VCVTTDH : ADuI<0b11101, 0b11, 0b0011, 0b11, 0,
555 (outs SPR:$Sd), (ins DPR:$Dm),
556 NoItinerary, "vcvtt", ".f16.f64\t$Sd, $Dm",
557 []>, Requires<[HasV8FP]> {
558 // Instruction operands.
562 // Encode instruction operands.
563 let Inst{15-12} = Sd{4-1};
564 let Inst{22} = Sd{0};
565 let Inst{3-0} = Dm{3-0};
569 def VNEGD : ADuI<0b11101, 0b11, 0b0001, 0b01, 0,
570 (outs DPR:$Dd), (ins DPR:$Dm),
571 IIC_fpUNA64, "vneg", ".f64\t$Dd, $Dm",
572 [(set DPR:$Dd, (fneg (f64 DPR:$Dm)))]>;
574 def VNEGS : ASuIn<0b11101, 0b11, 0b0001, 0b01, 0,
575 (outs SPR:$Sd), (ins SPR:$Sm),
576 IIC_fpUNA32, "vneg", ".f32\t$Sd, $Sm",
577 [(set SPR:$Sd, (fneg SPR:$Sm))]> {
578 // Some single precision VFP instructions may be executed on both NEON and
579 // VFP pipelines on A8.
580 let D = VFPNeonA8Domain;
583 def VSQRTD : ADuI<0b11101, 0b11, 0b0001, 0b11, 0,
584 (outs DPR:$Dd), (ins DPR:$Dm),
585 IIC_fpSQRT64, "vsqrt", ".f64\t$Dd, $Dm",
586 [(set DPR:$Dd, (fsqrt (f64 DPR:$Dm)))]>;
588 def VSQRTS : ASuI<0b11101, 0b11, 0b0001, 0b11, 0,
589 (outs SPR:$Sd), (ins SPR:$Sm),
590 IIC_fpSQRT32, "vsqrt", ".f32\t$Sd, $Sm",
591 [(set SPR:$Sd, (fsqrt SPR:$Sm))]>;
593 let neverHasSideEffects = 1 in {
594 def VMOVD : ADuI<0b11101, 0b11, 0b0000, 0b01, 0,
595 (outs DPR:$Dd), (ins DPR:$Dm),
596 IIC_fpUNA64, "vmov", ".f64\t$Dd, $Dm", []>;
598 def VMOVS : ASuI<0b11101, 0b11, 0b0000, 0b01, 0,
599 (outs SPR:$Sd), (ins SPR:$Sm),
600 IIC_fpUNA32, "vmov", ".f32\t$Sd, $Sm", []>;
601 } // neverHasSideEffects
603 //===----------------------------------------------------------------------===//
604 // FP <-> GPR Copies. Int <-> FP Conversions.
607 def VMOVRS : AVConv2I<0b11100001, 0b1010,
608 (outs GPR:$Rt), (ins SPR:$Sn),
609 IIC_fpMOVSI, "vmov", "\t$Rt, $Sn",
610 [(set GPR:$Rt, (bitconvert SPR:$Sn))]> {
611 // Instruction operands.
615 // Encode instruction operands.
616 let Inst{19-16} = Sn{4-1};
618 let Inst{15-12} = Rt;
620 let Inst{6-5} = 0b00;
621 let Inst{3-0} = 0b0000;
623 // Some single precision VFP instructions may be executed on both NEON and VFP
625 let D = VFPNeonDomain;
628 // Bitcast i32 -> f32. NEON prefers to use VMOVDRR.
629 def VMOVSR : AVConv4I<0b11100000, 0b1010,
630 (outs SPR:$Sn), (ins GPR:$Rt),
631 IIC_fpMOVIS, "vmov", "\t$Sn, $Rt",
632 [(set SPR:$Sn, (bitconvert GPR:$Rt))]>,
633 Requires<[HasVFP2, UseVMOVSR]> {
634 // Instruction operands.
638 // Encode instruction operands.
639 let Inst{19-16} = Sn{4-1};
641 let Inst{15-12} = Rt;
643 let Inst{6-5} = 0b00;
644 let Inst{3-0} = 0b0000;
646 // Some single precision VFP instructions may be executed on both NEON and VFP
648 let D = VFPNeonDomain;
651 let neverHasSideEffects = 1 in {
652 def VMOVRRD : AVConv3I<0b11000101, 0b1011,
653 (outs GPR:$Rt, GPR:$Rt2), (ins DPR:$Dm),
654 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $Dm",
655 [/* FIXME: Can't write pattern for multiple result instr*/]> {
656 // Instruction operands.
661 // Encode instruction operands.
662 let Inst{3-0} = Dm{3-0};
664 let Inst{15-12} = Rt;
665 let Inst{19-16} = Rt2;
667 let Inst{7-6} = 0b00;
669 // Some single precision VFP instructions may be executed on both NEON and VFP
671 let D = VFPNeonDomain;
674 def VMOVRRS : AVConv3I<0b11000101, 0b1010,
675 (outs GPR:$Rt, GPR:$Rt2), (ins SPR:$src1, SPR:$src2),
676 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $src1, $src2",
677 [/* For disassembly only; pattern left blank */]> {
682 // Encode instruction operands.
683 let Inst{3-0} = src1{4-1};
684 let Inst{5} = src1{0};
685 let Inst{15-12} = Rt;
686 let Inst{19-16} = Rt2;
688 let Inst{7-6} = 0b00;
690 // Some single precision VFP instructions may be executed on both NEON and VFP
692 let D = VFPNeonDomain;
693 let DecoderMethod = "DecodeVMOVRRS";
695 } // neverHasSideEffects
700 def VMOVDRR : AVConv5I<0b11000100, 0b1011,
701 (outs DPR:$Dm), (ins GPR:$Rt, GPR:$Rt2),
702 IIC_fpMOVID, "vmov", "\t$Dm, $Rt, $Rt2",
703 [(set DPR:$Dm, (arm_fmdrr GPR:$Rt, GPR:$Rt2))]> {
704 // Instruction operands.
709 // Encode instruction operands.
710 let Inst{3-0} = Dm{3-0};
712 let Inst{15-12} = Rt;
713 let Inst{19-16} = Rt2;
715 let Inst{7-6} = 0b00;
717 // Some single precision VFP instructions may be executed on both NEON and VFP
719 let D = VFPNeonDomain;
722 let neverHasSideEffects = 1 in
723 def VMOVSRR : AVConv5I<0b11000100, 0b1010,
724 (outs SPR:$dst1, SPR:$dst2), (ins GPR:$src1, GPR:$src2),
725 IIC_fpMOVID, "vmov", "\t$dst1, $dst2, $src1, $src2",
726 [/* For disassembly only; pattern left blank */]> {
727 // Instruction operands.
732 // Encode instruction operands.
733 let Inst{3-0} = dst1{4-1};
734 let Inst{5} = dst1{0};
735 let Inst{15-12} = src1;
736 let Inst{19-16} = src2;
738 let Inst{7-6} = 0b00;
740 // Some single precision VFP instructions may be executed on both NEON and VFP
742 let D = VFPNeonDomain;
744 let DecoderMethod = "DecodeVMOVSRR";
750 // FMRX: SPR system reg -> GPR
752 // FMXR: GPR -> VFP system reg
757 class AVConv1IDs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
758 bits<4> opcod4, dag oops, dag iops,
759 InstrItinClass itin, string opc, string asm,
761 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
763 // Instruction operands.
767 // Encode instruction operands.
768 let Inst{3-0} = Sm{4-1};
770 let Inst{15-12} = Dd{3-0};
771 let Inst{22} = Dd{4};
774 class AVConv1InSs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
775 bits<4> opcod4, dag oops, dag iops,InstrItinClass itin,
776 string opc, string asm, list<dag> pattern>
777 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
779 // Instruction operands.
783 // Encode instruction operands.
784 let Inst{3-0} = Sm{4-1};
786 let Inst{15-12} = Sd{4-1};
787 let Inst{22} = Sd{0};
790 def VSITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
791 (outs DPR:$Dd), (ins SPR:$Sm),
792 IIC_fpCVTID, "vcvt", ".f64.s32\t$Dd, $Sm",
793 [(set DPR:$Dd, (f64 (arm_sitof SPR:$Sm)))]> {
794 let Inst{7} = 1; // s32
797 def VSITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
798 (outs SPR:$Sd),(ins SPR:$Sm),
799 IIC_fpCVTIS, "vcvt", ".f32.s32\t$Sd, $Sm",
800 [(set SPR:$Sd, (arm_sitof SPR:$Sm))]> {
801 let Inst{7} = 1; // s32
803 // Some single precision VFP instructions may be executed on both NEON and
804 // VFP pipelines on A8.
805 let D = VFPNeonA8Domain;
808 def VUITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
809 (outs DPR:$Dd), (ins SPR:$Sm),
810 IIC_fpCVTID, "vcvt", ".f64.u32\t$Dd, $Sm",
811 [(set DPR:$Dd, (f64 (arm_uitof SPR:$Sm)))]> {
812 let Inst{7} = 0; // u32
815 def VUITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
816 (outs SPR:$Sd), (ins SPR:$Sm),
817 IIC_fpCVTIS, "vcvt", ".f32.u32\t$Sd, $Sm",
818 [(set SPR:$Sd, (arm_uitof SPR:$Sm))]> {
819 let Inst{7} = 0; // u32
821 // Some single precision VFP instructions may be executed on both NEON and
822 // VFP pipelines on A8.
823 let D = VFPNeonA8Domain;
828 class AVConv1IsD_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
829 bits<4> opcod4, dag oops, dag iops,
830 InstrItinClass itin, string opc, string asm,
832 : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
834 // Instruction operands.
838 // Encode instruction operands.
839 let Inst{3-0} = Dm{3-0};
841 let Inst{15-12} = Sd{4-1};
842 let Inst{22} = Sd{0};
845 class AVConv1InsS_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
846 bits<4> opcod4, dag oops, dag iops,
847 InstrItinClass itin, string opc, string asm,
849 : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
851 // Instruction operands.
855 // Encode instruction operands.
856 let Inst{3-0} = Sm{4-1};
858 let Inst{15-12} = Sd{4-1};
859 let Inst{22} = Sd{0};
862 // Always set Z bit in the instruction, i.e. "round towards zero" variants.
863 def VTOSIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
864 (outs SPR:$Sd), (ins DPR:$Dm),
865 IIC_fpCVTDI, "vcvt", ".s32.f64\t$Sd, $Dm",
866 [(set SPR:$Sd, (arm_ftosi (f64 DPR:$Dm)))]> {
867 let Inst{7} = 1; // Z bit
870 def VTOSIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
871 (outs SPR:$Sd), (ins SPR:$Sm),
872 IIC_fpCVTSI, "vcvt", ".s32.f32\t$Sd, $Sm",
873 [(set SPR:$Sd, (arm_ftosi SPR:$Sm))]> {
874 let Inst{7} = 1; // Z bit
876 // Some single precision VFP instructions may be executed on both NEON and
877 // VFP pipelines on A8.
878 let D = VFPNeonA8Domain;
881 def VTOUIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
882 (outs SPR:$Sd), (ins DPR:$Dm),
883 IIC_fpCVTDI, "vcvt", ".u32.f64\t$Sd, $Dm",
884 [(set SPR:$Sd, (arm_ftoui (f64 DPR:$Dm)))]> {
885 let Inst{7} = 1; // Z bit
888 def VTOUIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
889 (outs SPR:$Sd), (ins SPR:$Sm),
890 IIC_fpCVTSI, "vcvt", ".u32.f32\t$Sd, $Sm",
891 [(set SPR:$Sd, (arm_ftoui SPR:$Sm))]> {
892 let Inst{7} = 1; // Z bit
894 // Some single precision VFP instructions may be executed on both NEON and
895 // VFP pipelines on A8.
896 let D = VFPNeonA8Domain;
899 // And the Z bit '0' variants, i.e. use the rounding mode specified by FPSCR.
900 let Uses = [FPSCR] in {
901 // FIXME: Verify encoding after integrated assembler is working.
902 def VTOSIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
903 (outs SPR:$Sd), (ins DPR:$Dm),
904 IIC_fpCVTDI, "vcvtr", ".s32.f64\t$Sd, $Dm",
905 [(set SPR:$Sd, (int_arm_vcvtr (f64 DPR:$Dm)))]>{
906 let Inst{7} = 0; // Z bit
909 def VTOSIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
910 (outs SPR:$Sd), (ins SPR:$Sm),
911 IIC_fpCVTSI, "vcvtr", ".s32.f32\t$Sd, $Sm",
912 [(set SPR:$Sd, (int_arm_vcvtr SPR:$Sm))]> {
913 let Inst{7} = 0; // Z bit
916 def VTOUIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
917 (outs SPR:$Sd), (ins DPR:$Dm),
918 IIC_fpCVTDI, "vcvtr", ".u32.f64\t$Sd, $Dm",
919 [(set SPR:$Sd, (int_arm_vcvtru(f64 DPR:$Dm)))]>{
920 let Inst{7} = 0; // Z bit
923 def VTOUIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
924 (outs SPR:$Sd), (ins SPR:$Sm),
925 IIC_fpCVTSI, "vcvtr", ".u32.f32\t$Sd, $Sm",
926 [(set SPR:$Sd, (int_arm_vcvtru SPR:$Sm))]> {
927 let Inst{7} = 0; // Z bit
931 // Convert between floating-point and fixed-point
932 // Data type for fixed-point naming convention:
933 // S16 (U=0, sx=0) -> SH
934 // U16 (U=1, sx=0) -> UH
935 // S32 (U=0, sx=1) -> SL
936 // U32 (U=1, sx=1) -> UL
938 let Constraints = "$a = $dst" in {
940 // FP to Fixed-Point:
942 // Single Precision register
943 class AVConv1XInsS_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
944 bit op5, dag oops, dag iops, InstrItinClass itin,
945 string opc, string asm, list<dag> pattern>
946 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern>,
947 Sched<[WriteCvtFP]> {
949 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
950 let Inst{22} = dst{0};
951 let Inst{15-12} = dst{4-1};
954 // Double Precision register
955 class AVConv1XInsD_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
956 bit op5, dag oops, dag iops, InstrItinClass itin,
957 string opc, string asm, list<dag> pattern>
958 : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern>,
959 Sched<[WriteCvtFP]> {
961 // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
962 let Inst{22} = dst{4};
963 let Inst{15-12} = dst{3-0};
966 def VTOSHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 0,
967 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
968 IIC_fpCVTSI, "vcvt", ".s16.f32\t$dst, $a, $fbits", []> {
969 // Some single precision VFP instructions may be executed on both NEON and
970 // VFP pipelines on A8.
971 let D = VFPNeonA8Domain;
974 def VTOUHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 0,
975 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
976 IIC_fpCVTSI, "vcvt", ".u16.f32\t$dst, $a, $fbits", []> {
977 // Some single precision VFP instructions may be executed on both NEON and
978 // VFP pipelines on A8.
979 let D = VFPNeonA8Domain;
982 def VTOSLS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 1,
983 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
984 IIC_fpCVTSI, "vcvt", ".s32.f32\t$dst, $a, $fbits", []> {
985 // Some single precision VFP instructions may be executed on both NEON and
986 // VFP pipelines on A8.
987 let D = VFPNeonA8Domain;
990 def VTOULS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 1,
991 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
992 IIC_fpCVTSI, "vcvt", ".u32.f32\t$dst, $a, $fbits", []> {
993 // Some single precision VFP instructions may be executed on both NEON and
994 // VFP pipelines on A8.
995 let D = VFPNeonA8Domain;
998 def VTOSHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 0,
999 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1000 IIC_fpCVTDI, "vcvt", ".s16.f64\t$dst, $a, $fbits", []>;
1002 def VTOUHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 0,
1003 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1004 IIC_fpCVTDI, "vcvt", ".u16.f64\t$dst, $a, $fbits", []>;
1006 def VTOSLD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 1,
1007 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1008 IIC_fpCVTDI, "vcvt", ".s32.f64\t$dst, $a, $fbits", []>;
1010 def VTOULD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 1,
1011 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1012 IIC_fpCVTDI, "vcvt", ".u32.f64\t$dst, $a, $fbits", []>;
1014 // Fixed-Point to FP:
1016 def VSHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 0,
1017 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1018 IIC_fpCVTIS, "vcvt", ".f32.s16\t$dst, $a, $fbits", []> {
1019 // Some single precision VFP instructions may be executed on both NEON and
1020 // VFP pipelines on A8.
1021 let D = VFPNeonA8Domain;
1024 def VUHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 0,
1025 (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1026 IIC_fpCVTIS, "vcvt", ".f32.u16\t$dst, $a, $fbits", []> {
1027 // Some single precision VFP instructions may be executed on both NEON and
1028 // VFP pipelines on A8.
1029 let D = VFPNeonA8Domain;
1032 def VSLTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 1,
1033 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1034 IIC_fpCVTIS, "vcvt", ".f32.s32\t$dst, $a, $fbits", []> {
1035 // Some single precision VFP instructions may be executed on both NEON and
1036 // VFP pipelines on A8.
1037 let D = VFPNeonA8Domain;
1040 def VULTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 1,
1041 (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1042 IIC_fpCVTIS, "vcvt", ".f32.u32\t$dst, $a, $fbits", []> {
1043 // Some single precision VFP instructions may be executed on both NEON and
1044 // VFP pipelines on A8.
1045 let D = VFPNeonA8Domain;
1048 def VSHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 0,
1049 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1050 IIC_fpCVTID, "vcvt", ".f64.s16\t$dst, $a, $fbits", []>;
1052 def VUHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 0,
1053 (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1054 IIC_fpCVTID, "vcvt", ".f64.u16\t$dst, $a, $fbits", []>;
1056 def VSLTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 1,
1057 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1058 IIC_fpCVTID, "vcvt", ".f64.s32\t$dst, $a, $fbits", []>;
1060 def VULTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 1,
1061 (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1062 IIC_fpCVTID, "vcvt", ".f64.u32\t$dst, $a, $fbits", []>;
1064 } // End of 'let Constraints = "$a = $dst" in'
1066 //===----------------------------------------------------------------------===//
1067 // FP Multiply-Accumulate Operations.
1070 def VMLAD : ADbI<0b11100, 0b00, 0, 0,
1071 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1072 IIC_fpMAC64, "vmla", ".f64\t$Dd, $Dn, $Dm",
1073 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1074 (f64 DPR:$Ddin)))]>,
1075 RegConstraint<"$Ddin = $Dd">,
1076 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1078 def VMLAS : ASbIn<0b11100, 0b00, 0, 0,
1079 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1080 IIC_fpMAC32, "vmla", ".f32\t$Sd, $Sn, $Sm",
1081 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
1083 RegConstraint<"$Sdin = $Sd">,
1084 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1085 // Some single precision VFP instructions may be executed on both NEON and
1086 // VFP pipelines on A8.
1087 let D = VFPNeonA8Domain;
1090 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1091 (VMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1092 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1093 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1094 (VMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1095 Requires<[HasVFP2,DontUseNEONForFP, UseFPVMLx,DontUseFusedMAC]>;
1097 def VMLSD : ADbI<0b11100, 0b00, 1, 0,
1098 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1099 IIC_fpMAC64, "vmls", ".f64\t$Dd, $Dn, $Dm",
1100 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1101 (f64 DPR:$Ddin)))]>,
1102 RegConstraint<"$Ddin = $Dd">,
1103 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1105 def VMLSS : ASbIn<0b11100, 0b00, 1, 0,
1106 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1107 IIC_fpMAC32, "vmls", ".f32\t$Sd, $Sn, $Sm",
1108 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1110 RegConstraint<"$Sdin = $Sd">,
1111 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1112 // Some single precision VFP instructions may be executed on both NEON and
1113 // VFP pipelines on A8.
1114 let D = VFPNeonA8Domain;
1117 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1118 (VMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
1119 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1120 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1121 (VMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
1122 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1124 def VNMLAD : ADbI<0b11100, 0b01, 1, 0,
1125 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1126 IIC_fpMAC64, "vnmla", ".f64\t$Dd, $Dn, $Dm",
1127 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1128 (f64 DPR:$Ddin)))]>,
1129 RegConstraint<"$Ddin = $Dd">,
1130 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1132 def VNMLAS : ASbI<0b11100, 0b01, 1, 0,
1133 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1134 IIC_fpMAC32, "vnmla", ".f32\t$Sd, $Sn, $Sm",
1135 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1137 RegConstraint<"$Sdin = $Sd">,
1138 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1139 // Some single precision VFP instructions may be executed on both NEON and
1140 // VFP pipelines on A8.
1141 let D = VFPNeonA8Domain;
1144 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
1145 (VNMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
1146 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1147 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
1148 (VNMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
1149 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1151 def VNMLSD : ADbI<0b11100, 0b01, 0, 0,
1152 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1153 IIC_fpMAC64, "vnmls", ".f64\t$Dd, $Dn, $Dm",
1154 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1155 (f64 DPR:$Ddin)))]>,
1156 RegConstraint<"$Ddin = $Dd">,
1157 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1159 def VNMLSS : ASbI<0b11100, 0b01, 0, 0,
1160 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1161 IIC_fpMAC32, "vnmls", ".f32\t$Sd, $Sn, $Sm",
1162 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
1163 RegConstraint<"$Sdin = $Sd">,
1164 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]> {
1165 // Some single precision VFP instructions may be executed on both NEON and
1166 // VFP pipelines on A8.
1167 let D = VFPNeonA8Domain;
1170 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
1171 (VNMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
1172 Requires<[HasVFP2,UseFPVMLx,DontUseFusedMAC]>;
1173 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
1174 (VNMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
1175 Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx,DontUseFusedMAC]>;
1177 //===----------------------------------------------------------------------===//
1178 // Fused FP Multiply-Accumulate Operations.
1180 def VFMAD : ADbI<0b11101, 0b10, 0, 0,
1181 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1182 IIC_fpFMAC64, "vfma", ".f64\t$Dd, $Dn, $Dm",
1183 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1184 (f64 DPR:$Ddin)))]>,
1185 RegConstraint<"$Ddin = $Dd">,
1186 Requires<[HasVFP4,UseFusedMAC]>;
1188 def VFMAS : ASbIn<0b11101, 0b10, 0, 0,
1189 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1190 IIC_fpFMAC32, "vfma", ".f32\t$Sd, $Sn, $Sm",
1191 [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
1193 RegConstraint<"$Sdin = $Sd">,
1194 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1195 // Some single precision VFP instructions may be executed on both NEON and
1199 def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1200 (VFMAD DPR:$dstin, DPR:$a, DPR:$b)>,
1201 Requires<[HasVFP4,UseFusedMAC]>;
1202 def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1203 (VFMAS SPR:$dstin, SPR:$a, SPR:$b)>,
1204 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1206 // Match @llvm.fma.* intrinsics
1207 // (fma x, y, z) -> (vfms z, x, y)
1208 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, DPR:$Ddin)),
1209 (VFMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1210 Requires<[HasVFP4]>;
1211 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, SPR:$Sdin)),
1212 (VFMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1213 Requires<[HasVFP4]>;
1215 def VFMSD : ADbI<0b11101, 0b10, 1, 0,
1216 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1217 IIC_fpFMAC64, "vfms", ".f64\t$Dd, $Dn, $Dm",
1218 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1219 (f64 DPR:$Ddin)))]>,
1220 RegConstraint<"$Ddin = $Dd">,
1221 Requires<[HasVFP4,UseFusedMAC]>;
1223 def VFMSS : ASbIn<0b11101, 0b10, 1, 0,
1224 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1225 IIC_fpFMAC32, "vfms", ".f32\t$Sd, $Sn, $Sm",
1226 [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1228 RegConstraint<"$Sdin = $Sd">,
1229 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1230 // Some single precision VFP instructions may be executed on both NEON and
1234 def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
1235 (VFMSD DPR:$dstin, DPR:$a, DPR:$b)>,
1236 Requires<[HasVFP4,UseFusedMAC]>;
1237 def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
1238 (VFMSS SPR:$dstin, SPR:$a, SPR:$b)>,
1239 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1241 // Match @llvm.fma.* intrinsics
1242 // (fma (fneg x), y, z) -> (vfms z, x, y)
1243 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin)),
1244 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1245 Requires<[HasVFP4]>;
1246 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin)),
1247 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1248 Requires<[HasVFP4]>;
1249 // (fma x, (fneg y), z) -> (vfms z, x, y)
1250 def : Pat<(f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin)),
1251 (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1252 Requires<[HasVFP4]>;
1253 def : Pat<(f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin)),
1254 (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1255 Requires<[HasVFP4]>;
1257 def VFNMAD : ADbI<0b11101, 0b01, 1, 0,
1258 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1259 IIC_fpFMAC64, "vfnma", ".f64\t$Dd, $Dn, $Dm",
1260 [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
1261 (f64 DPR:$Ddin)))]>,
1262 RegConstraint<"$Ddin = $Dd">,
1263 Requires<[HasVFP4,UseFusedMAC]>;
1265 def VFNMAS : ASbI<0b11101, 0b01, 1, 0,
1266 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1267 IIC_fpFMAC32, "vfnma", ".f32\t$Sd, $Sn, $Sm",
1268 [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
1270 RegConstraint<"$Sdin = $Sd">,
1271 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1272 // Some single precision VFP instructions may be executed on both NEON and
1276 def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
1277 (VFNMAD DPR:$dstin, DPR:$a, DPR:$b)>,
1278 Requires<[HasVFP4,UseFusedMAC]>;
1279 def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
1280 (VFNMAS SPR:$dstin, SPR:$a, SPR:$b)>,
1281 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1283 // Match @llvm.fma.* intrinsics
1284 // (fneg (fma x, y, z)) -> (vfnma z, x, y)
1285 def : Pat<(fneg (fma (f64 DPR:$Dn), (f64 DPR:$Dm), (f64 DPR:$Ddin))),
1286 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1287 Requires<[HasVFP4]>;
1288 def : Pat<(fneg (fma (f32 SPR:$Sn), (f32 SPR:$Sm), (f32 SPR:$Sdin))),
1289 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1290 Requires<[HasVFP4]>;
1291 // (fma (fneg x), y, (fneg z)) -> (vfnma z, x, y)
1292 def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, (fneg DPR:$Ddin))),
1293 (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1294 Requires<[HasVFP4]>;
1295 def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, (fneg SPR:$Sdin))),
1296 (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1297 Requires<[HasVFP4]>;
1299 def VFNMSD : ADbI<0b11101, 0b01, 0, 0,
1300 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
1301 IIC_fpFMAC64, "vfnms", ".f64\t$Dd, $Dn, $Dm",
1302 [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
1303 (f64 DPR:$Ddin)))]>,
1304 RegConstraint<"$Ddin = $Dd">,
1305 Requires<[HasVFP4,UseFusedMAC]>;
1307 def VFNMSS : ASbI<0b11101, 0b01, 0, 0,
1308 (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
1309 IIC_fpFMAC32, "vfnms", ".f32\t$Sd, $Sn, $Sm",
1310 [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
1311 RegConstraint<"$Sdin = $Sd">,
1312 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]> {
1313 // Some single precision VFP instructions may be executed on both NEON and
1317 def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
1318 (VFNMSD DPR:$dstin, DPR:$a, DPR:$b)>,
1319 Requires<[HasVFP4,UseFusedMAC]>;
1320 def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
1321 (VFNMSS SPR:$dstin, SPR:$a, SPR:$b)>,
1322 Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
1324 // Match @llvm.fma.* intrinsics
1326 // (fma x, y, (fneg z)) -> (vfnms z, x, y))
1327 def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, (fneg DPR:$Ddin))),
1328 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1329 Requires<[HasVFP4]>;
1330 def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, (fneg SPR:$Sdin))),
1331 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1332 Requires<[HasVFP4]>;
1333 // (fneg (fma (fneg x), y, z)) -> (vfnms z, x, y)
1334 def : Pat<(fneg (f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin))),
1335 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1336 Requires<[HasVFP4]>;
1337 def : Pat<(fneg (f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin))),
1338 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1339 Requires<[HasVFP4]>;
1340 // (fneg (fma x, (fneg y), z) -> (vfnms z, x, y)
1341 def : Pat<(fneg (f64 (fma DPR:$Dn, (fneg DPR:$Dm), DPR:$Ddin))),
1342 (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
1343 Requires<[HasVFP4]>;
1344 def : Pat<(fneg (f32 (fma SPR:$Sn, (fneg SPR:$Sm), SPR:$Sdin))),
1345 (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
1346 Requires<[HasVFP4]>;
1348 //===----------------------------------------------------------------------===//
1349 // FP Conditional moves.
1352 let neverHasSideEffects = 1 in {
1353 def VMOVDcc : ARMPseudoInst<(outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm, pred:$p),
1355 [/*(set DPR:$Dd, (ARMcmov DPR:$Dn, DPR:$Dm, imm:$cc))*/]>,
1356 RegConstraint<"$Dn = $Dd">;
1358 def VMOVScc : ARMPseudoInst<(outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm, pred:$p),
1360 [/*(set SPR:$Sd, (ARMcmov SPR:$Sn, SPR:$Sm, imm:$cc))*/]>,
1361 RegConstraint<"$Sn = $Sd">;
1362 } // neverHasSideEffects
1364 //===----------------------------------------------------------------------===//
1365 // Move from VFP System Register to ARM core register.
1368 class MovFromVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
1370 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
1372 // Instruction operand.
1375 let Inst{27-20} = 0b11101111;
1376 let Inst{19-16} = opc19_16;
1377 let Inst{15-12} = Rt;
1378 let Inst{11-8} = 0b1010;
1380 let Inst{6-5} = 0b00;
1382 let Inst{3-0} = 0b0000;
1385 // APSR is the application level alias of CPSR. This FPSCR N, Z, C, V flags
1387 let Defs = [CPSR], Uses = [FPSCR_NZCV], Rt = 0b1111 /* apsr_nzcv */ in
1388 def FMSTAT : MovFromVFP<0b0001 /* fpscr */, (outs), (ins),
1389 "vmrs", "\tAPSR_nzcv, fpscr", [(arm_fmstat)]>;
1391 // Application level FPSCR -> GPR
1392 let hasSideEffects = 1, Uses = [FPSCR] in
1393 def VMRS : MovFromVFP<0b0001 /* fpscr */, (outs GPR:$Rt), (ins),
1394 "vmrs", "\t$Rt, fpscr",
1395 [(set GPR:$Rt, (int_arm_get_fpscr))]>;
1397 // System level FPEXC, FPSID -> GPR
1398 let Uses = [FPSCR] in {
1399 def VMRS_FPEXC : MovFromVFP<0b1000 /* fpexc */, (outs GPR:$Rt), (ins),
1400 "vmrs", "\t$Rt, fpexc", []>;
1401 def VMRS_FPSID : MovFromVFP<0b0000 /* fpsid */, (outs GPR:$Rt), (ins),
1402 "vmrs", "\t$Rt, fpsid", []>;
1403 def VMRS_MVFR0 : MovFromVFP<0b0111 /* mvfr0 */, (outs GPR:$Rt), (ins),
1404 "vmrs", "\t$Rt, mvfr0", []>;
1405 def VMRS_MVFR1 : MovFromVFP<0b0110 /* mvfr1 */, (outs GPR:$Rt), (ins),
1406 "vmrs", "\t$Rt, mvfr1", []>;
1407 def VMRS_FPINST : MovFromVFP<0b1001 /* fpinst */, (outs GPR:$Rt), (ins),
1408 "vmrs", "\t$Rt, fpinst", []>;
1409 def VMRS_FPINST2 : MovFromVFP<0b1010 /* fpinst2 */, (outs GPR:$Rt), (ins),
1410 "vmrs", "\t$Rt, fpinst2", []>;
1413 //===----------------------------------------------------------------------===//
1414 // Move from ARM core register to VFP System Register.
1417 class MovToVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
1419 VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
1421 // Instruction operand.
1424 // Encode instruction operand.
1425 let Inst{15-12} = src;
1427 let Inst{27-20} = 0b11101110;
1428 let Inst{19-16} = opc19_16;
1429 let Inst{11-8} = 0b1010;
1434 let Defs = [FPSCR] in {
1435 // Application level GPR -> FPSCR
1436 def VMSR : MovToVFP<0b0001 /* fpscr */, (outs), (ins GPR:$src),
1437 "vmsr", "\tfpscr, $src", [(int_arm_set_fpscr GPR:$src)]>;
1438 // System level GPR -> FPEXC
1439 def VMSR_FPEXC : MovToVFP<0b1000 /* fpexc */, (outs), (ins GPR:$src),
1440 "vmsr", "\tfpexc, $src", []>;
1441 // System level GPR -> FPSID
1442 def VMSR_FPSID : MovToVFP<0b0000 /* fpsid */, (outs), (ins GPR:$src),
1443 "vmsr", "\tfpsid, $src", []>;
1445 def VMSR_FPINST : MovToVFP<0b1001 /* fpinst */, (outs), (ins GPR:$src),
1446 "vmsr", "\tfpinst, $src", []>;
1447 def VMSR_FPINST2 : MovToVFP<0b1010 /* fpinst2 */, (outs), (ins GPR:$src),
1448 "vmsr", "\tfpinst2, $src", []>;
1451 //===----------------------------------------------------------------------===//
1455 // Materialize FP immediates. VFP3 only.
1456 let isReMaterializable = 1 in {
1457 def FCONSTD : VFPAI<(outs DPR:$Dd), (ins vfp_f64imm:$imm),
1458 VFPMiscFrm, IIC_fpUNA64,
1459 "vmov", ".f64\t$Dd, $imm",
1460 [(set DPR:$Dd, vfp_f64imm:$imm)]>, Requires<[HasVFP3]> {
1464 let Inst{27-23} = 0b11101;
1465 let Inst{22} = Dd{4};
1466 let Inst{21-20} = 0b11;
1467 let Inst{19-16} = imm{7-4};
1468 let Inst{15-12} = Dd{3-0};
1469 let Inst{11-9} = 0b101;
1470 let Inst{8} = 1; // Double precision.
1471 let Inst{7-4} = 0b0000;
1472 let Inst{3-0} = imm{3-0};
1475 def FCONSTS : VFPAI<(outs SPR:$Sd), (ins vfp_f32imm:$imm),
1476 VFPMiscFrm, IIC_fpUNA32,
1477 "vmov", ".f32\t$Sd, $imm",
1478 [(set SPR:$Sd, vfp_f32imm:$imm)]>, Requires<[HasVFP3]> {
1482 let Inst{27-23} = 0b11101;
1483 let Inst{22} = Sd{0};
1484 let Inst{21-20} = 0b11;
1485 let Inst{19-16} = imm{7-4};
1486 let Inst{15-12} = Sd{4-1};
1487 let Inst{11-9} = 0b101;
1488 let Inst{8} = 0; // Single precision.
1489 let Inst{7-4} = 0b0000;
1490 let Inst{3-0} = imm{3-0};
1494 //===----------------------------------------------------------------------===//
1495 // Assembler aliases.
1497 // A few mnemnoic aliases for pre-unifixed syntax. We don't guarantee to
1498 // support them all, but supporting at least some of the basics is
1499 // good to be friendly.
1500 def : VFP2MnemonicAlias<"flds", "vldr">;
1501 def : VFP2MnemonicAlias<"fldd", "vldr">;
1502 def : VFP2MnemonicAlias<"fmrs", "vmov">;
1503 def : VFP2MnemonicAlias<"fmsr", "vmov">;
1504 def : VFP2MnemonicAlias<"fsqrts", "vsqrt">;
1505 def : VFP2MnemonicAlias<"fsqrtd", "vsqrt">;
1506 def : VFP2MnemonicAlias<"fadds", "vadd.f32">;
1507 def : VFP2MnemonicAlias<"faddd", "vadd.f64">;
1508 def : VFP2MnemonicAlias<"fmrdd", "vmov">;
1509 def : VFP2MnemonicAlias<"fmrds", "vmov">;
1510 def : VFP2MnemonicAlias<"fmrrd", "vmov">;
1511 def : VFP2MnemonicAlias<"fmdrr", "vmov">;
1512 def : VFP2MnemonicAlias<"fmuls", "vmul.f32">;
1513 def : VFP2MnemonicAlias<"fmuld", "vmul.f64">;
1514 def : VFP2MnemonicAlias<"fnegs", "vneg.f32">;
1515 def : VFP2MnemonicAlias<"fnegd", "vneg.f64">;
1516 def : VFP2MnemonicAlias<"ftosizd", "vcvt.s32.f64">;
1517 def : VFP2MnemonicAlias<"ftosid", "vcvtr.s32.f64">;
1518 def : VFP2MnemonicAlias<"ftosizs", "vcvt.s32.f32">;
1519 def : VFP2MnemonicAlias<"ftosis", "vcvtr.s32.f32">;
1520 def : VFP2MnemonicAlias<"ftouizd", "vcvt.u32.f64">;
1521 def : VFP2MnemonicAlias<"ftouid", "vcvtr.u32.f64">;
1522 def : VFP2MnemonicAlias<"ftouizs", "vcvt.u32.f32">;
1523 def : VFP2MnemonicAlias<"ftouis", "vcvtr.u32.f32">;
1524 def : VFP2MnemonicAlias<"fsitod", "vcvt.f64.s32">;
1525 def : VFP2MnemonicAlias<"fsitos", "vcvt.f32.s32">;
1526 def : VFP2MnemonicAlias<"fuitod", "vcvt.f64.u32">;
1527 def : VFP2MnemonicAlias<"fuitos", "vcvt.f32.u32">;
1528 def : VFP2MnemonicAlias<"fsts", "vstr">;
1529 def : VFP2MnemonicAlias<"fstd", "vstr">;
1530 def : VFP2MnemonicAlias<"fmacd", "vmla.f64">;
1531 def : VFP2MnemonicAlias<"fmacs", "vmla.f32">;
1532 def : VFP2MnemonicAlias<"fcpys", "vmov.f32">;
1533 def : VFP2MnemonicAlias<"fcpyd", "vmov.f64">;
1534 def : VFP2MnemonicAlias<"fcmps", "vcmp.f32">;
1535 def : VFP2MnemonicAlias<"fcmpd", "vcmp.f64">;
1536 def : VFP2MnemonicAlias<"fdivs", "vdiv.f32">;
1537 def : VFP2MnemonicAlias<"fdivd", "vdiv.f64">;
1538 def : VFP2MnemonicAlias<"fmrx", "vmrs">;
1539 def : VFP2MnemonicAlias<"fmxr", "vmsr">;
1541 // Be friendly and accept the old form of zero-compare
1542 def : VFP2InstAlias<"fcmpzd${p} $val", (VCMPZD DPR:$val, pred:$p)>;
1543 def : VFP2InstAlias<"fcmpzs${p} $val", (VCMPZS SPR:$val, pred:$p)>;
1546 def : VFP2InstAlias<"fmstat${p}", (FMSTAT pred:$p)>;
1547 def : VFP2InstAlias<"fadds${p} $Sd, $Sn, $Sm",
1548 (VADDS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
1549 def : VFP2InstAlias<"faddd${p} $Dd, $Dn, $Dm",
1550 (VADDD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
1551 def : VFP2InstAlias<"fsubs${p} $Sd, $Sn, $Sm",
1552 (VSUBS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
1553 def : VFP2InstAlias<"fsubd${p} $Dd, $Dn, $Dm",
1554 (VSUBD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
1556 // No need for the size suffix on VSQRT. It's implied by the register classes.
1557 def : VFP2InstAlias<"vsqrt${p} $Sd, $Sm", (VSQRTS SPR:$Sd, SPR:$Sm, pred:$p)>;
1558 def : VFP2InstAlias<"vsqrt${p} $Dd, $Dm", (VSQRTD DPR:$Dd, DPR:$Dm, pred:$p)>;
1560 // VLDR/VSTR accept an optional type suffix.
1561 def : VFP2InstAlias<"vldr${p}.32 $Sd, $addr",
1562 (VLDRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
1563 def : VFP2InstAlias<"vstr${p}.32 $Sd, $addr",
1564 (VSTRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
1565 def : VFP2InstAlias<"vldr${p}.64 $Dd, $addr",
1566 (VLDRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
1567 def : VFP2InstAlias<"vstr${p}.64 $Dd, $addr",
1568 (VSTRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
1570 // VMOV can accept optional 32-bit or less data type suffix suffix.
1571 def : VFP2InstAlias<"vmov${p}.8 $Rt, $Sn",
1572 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1573 def : VFP2InstAlias<"vmov${p}.16 $Rt, $Sn",
1574 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1575 def : VFP2InstAlias<"vmov${p}.32 $Rt, $Sn",
1576 (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
1577 def : VFP2InstAlias<"vmov${p}.8 $Sn, $Rt",
1578 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1579 def : VFP2InstAlias<"vmov${p}.16 $Sn, $Rt",
1580 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1581 def : VFP2InstAlias<"vmov${p}.32 $Sn, $Rt",
1582 (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
1584 def : VFP2InstAlias<"vmov${p}.f64 $Rt, $Rt2, $Dn",
1585 (VMOVRRD GPR:$Rt, GPR:$Rt2, DPR:$Dn, pred:$p)>;
1586 def : VFP2InstAlias<"vmov${p}.f64 $Dn, $Rt, $Rt2",
1587 (VMOVDRR DPR:$Dn, GPR:$Rt, GPR:$Rt2, pred:$p)>;
1589 // VMOVS doesn't need the .f32 to disambiguate from the NEON encoding the way
1591 def : VFP2InstAlias<"vmov${p} $Sd, $Sm",
1592 (VMOVS SPR:$Sd, SPR:$Sm, pred:$p)>;