xref: /freebsd/contrib/llvm-project/llvm/lib/Target/ARM/ARMInstrVFP.td (revision 5e801ac66d24704442eba426ed13c3effb8a34e7)
1//===-- ARMInstrVFP.td - VFP support for ARM ---------------*- tablegen -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file describes the ARM VFP instruction set.
10//
11//===----------------------------------------------------------------------===//
12
13def SDT_CMPFP0  : SDTypeProfile<0, 1, [SDTCisFP<0>]>;
14def SDT_VMOVDRR : SDTypeProfile<1, 2, [SDTCisVT<0, f64>, SDTCisVT<1, i32>,
15                                       SDTCisSameAs<1, 2>]>;
16def SDT_VMOVRRD : SDTypeProfile<2, 1, [SDTCisVT<0, i32>, SDTCisSameAs<0, 1>,
17                                       SDTCisVT<2, f64>]>;
18
19def SDT_VMOVSR : SDTypeProfile<1, 1, [SDTCisVT<0, f32>, SDTCisVT<1, i32>]>;
20
21def arm_fmstat : SDNode<"ARMISD::FMSTAT",  SDTNone, [SDNPInGlue, SDNPOutGlue]>;
22def arm_cmpfp  : SDNode<"ARMISD::CMPFP",   SDT_ARMCmp, [SDNPOutGlue]>;
23def arm_cmpfp0 : SDNode<"ARMISD::CMPFPw0", SDT_CMPFP0, [SDNPOutGlue]>;
24def arm_cmpfpe : SDNode<"ARMISD::CMPFPE",  SDT_ARMCmp, [SDNPOutGlue]>;
25def arm_cmpfpe0: SDNode<"ARMISD::CMPFPEw0",SDT_CMPFP0, [SDNPOutGlue]>;
26def arm_fmdrr  : SDNode<"ARMISD::VMOVDRR", SDT_VMOVDRR>;
27def arm_fmrrd  : SDNode<"ARMISD::VMOVRRD", SDT_VMOVRRD>;
28def arm_vmovsr  : SDNode<"ARMISD::VMOVSR", SDT_VMOVSR>;
29
30def SDT_VMOVhr : SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisVT<1, i32>] >;
31def SDT_VMOVrh : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisFP<1>] >;
32def arm_vmovhr : SDNode<"ARMISD::VMOVhr", SDT_VMOVhr>;
33def arm_vmovrh : SDNode<"ARMISD::VMOVrh", SDT_VMOVrh>;
34
35//===----------------------------------------------------------------------===//
36// Operand Definitions.
37//
38
39// 8-bit floating-point immediate encodings.
40def FPImmOperand : AsmOperandClass {
41  let Name = "FPImm";
42  let ParserMethod = "parseFPImm";
43}
44
45def vfp_f16imm : Operand<f16>,
46                 PatLeaf<(f16 fpimm), [{
47      return ARM_AM::getFP16Imm(N->getValueAPF()) != -1;
48    }], SDNodeXForm<fpimm, [{
49      APFloat InVal = N->getValueAPF();
50      uint32_t enc = ARM_AM::getFP16Imm(InVal);
51      return CurDAG->getTargetConstant(enc, SDLoc(N), MVT::i32);
52    }]>> {
53  let PrintMethod = "printFPImmOperand";
54  let ParserMatchClass = FPImmOperand;
55}
56
57def vfp_f32f16imm_xform : SDNodeXForm<fpimm, [{
58      APFloat InVal = N->getValueAPF();
59      uint32_t enc = ARM_AM::getFP32FP16Imm(InVal);
60      return CurDAG->getTargetConstant(enc, SDLoc(N), MVT::i32);
61    }]>;
62
63def vfp_f32f16imm : PatLeaf<(f32 fpimm), [{
64      return ARM_AM::getFP32FP16Imm(N->getValueAPF()) != -1;
65    }], vfp_f32f16imm_xform>;
66
67def vfp_f32imm_xform : SDNodeXForm<fpimm, [{
68      APFloat InVal = N->getValueAPF();
69      uint32_t enc = ARM_AM::getFP32Imm(InVal);
70      return CurDAG->getTargetConstant(enc, SDLoc(N), MVT::i32);
71    }]>;
72
73def gi_vfp_f32imm : GICustomOperandRenderer<"renderVFPF32Imm">,
74                    GISDNodeXFormEquiv<vfp_f32imm_xform>;
75
76def vfp_f32imm : Operand<f32>,
77                 PatLeaf<(f32 fpimm), [{
78      return ARM_AM::getFP32Imm(N->getValueAPF()) != -1;
79    }], vfp_f32imm_xform> {
80  let PrintMethod = "printFPImmOperand";
81  let ParserMatchClass = FPImmOperand;
82  let GISelPredicateCode = [{
83      const auto &MO = MI.getOperand(1);
84      if (!MO.isFPImm())
85        return false;
86      return ARM_AM::getFP32Imm(MO.getFPImm()->getValueAPF()) != -1;
87    }];
88}
89
90def vfp_f64imm_xform : SDNodeXForm<fpimm, [{
91      APFloat InVal = N->getValueAPF();
92      uint32_t enc = ARM_AM::getFP64Imm(InVal);
93      return CurDAG->getTargetConstant(enc, SDLoc(N), MVT::i32);
94    }]>;
95
96def gi_vfp_f64imm : GICustomOperandRenderer<"renderVFPF64Imm">,
97                    GISDNodeXFormEquiv<vfp_f64imm_xform>;
98
99def vfp_f64imm : Operand<f64>,
100                 PatLeaf<(f64 fpimm), [{
101      return ARM_AM::getFP64Imm(N->getValueAPF()) != -1;
102    }], vfp_f64imm_xform> {
103  let PrintMethod = "printFPImmOperand";
104  let ParserMatchClass = FPImmOperand;
105  let GISelPredicateCode = [{
106      const auto &MO = MI.getOperand(1);
107      if (!MO.isFPImm())
108        return false;
109      return ARM_AM::getFP64Imm(MO.getFPImm()->getValueAPF()) != -1;
110    }];
111}
112
113def alignedload16 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
114  return cast<LoadSDNode>(N)->getAlignment() >= 2;
115}]>;
116
117def alignedload32 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
118  return cast<LoadSDNode>(N)->getAlignment() >= 4;
119}]>;
120
121def alignedstore16 : PatFrag<(ops node:$val, node:$ptr),
122                             (store node:$val, node:$ptr), [{
123  return cast<StoreSDNode>(N)->getAlignment() >= 2;
124}]>;
125
126def alignedstore32 : PatFrag<(ops node:$val, node:$ptr),
127                             (store node:$val, node:$ptr), [{
128  return cast<StoreSDNode>(N)->getAlignment() >= 4;
129}]>;
130
131// The VCVT to/from fixed-point instructions encode the 'fbits' operand
132// (the number of fixed bits) differently than it appears in the assembly
133// source. It's encoded as "Size - fbits" where Size is the size of the
134// fixed-point representation (32 or 16) and fbits is the value appearing
135// in the assembly source, an integer in [0,16] or (0,32], depending on size.
136def fbits32_asm_operand : AsmOperandClass { let Name = "FBits32"; }
137def fbits32 : Operand<i32> {
138  let PrintMethod = "printFBits32";
139  let ParserMatchClass = fbits32_asm_operand;
140}
141
142def fbits16_asm_operand : AsmOperandClass { let Name = "FBits16"; }
143def fbits16 : Operand<i32> {
144  let PrintMethod = "printFBits16";
145  let ParserMatchClass = fbits16_asm_operand;
146}
147
148//===----------------------------------------------------------------------===//
149//  Load / store Instructions.
150//
151
152let canFoldAsLoad = 1, isReMaterializable = 1 in {
153
154def VLDRD : ADI5<0b1101, 0b01, (outs DPR:$Dd), (ins addrmode5:$addr),
155                 IIC_fpLoad64, "vldr", "\t$Dd, $addr",
156                 [(set DPR:$Dd, (f64 (alignedload32 addrmode5:$addr)))]>,
157            Requires<[HasFPRegs]>;
158
159def VLDRS : ASI5<0b1101, 0b01, (outs SPR:$Sd), (ins addrmode5:$addr),
160                 IIC_fpLoad32, "vldr", "\t$Sd, $addr",
161                 [(set SPR:$Sd, (alignedload32 addrmode5:$addr))]>,
162            Requires<[HasFPRegs]> {
163  // Some single precision VFP instructions may be executed on both NEON and VFP
164  // pipelines.
165  let D = VFPNeonDomain;
166}
167
168let isUnpredicable = 1 in
169def VLDRH : AHI5<0b1101, 0b01, (outs HPR:$Sd), (ins addrmode5fp16:$addr),
170                 IIC_fpLoad16, "vldr", ".16\t$Sd, $addr",
171                 [(set HPR:$Sd, (f16 (alignedload16 addrmode5fp16:$addr)))]>,
172            Requires<[HasFPRegs16]>;
173
174} // End of 'let canFoldAsLoad = 1, isReMaterializable = 1 in'
175
176def : Pat<(bf16 (alignedload16 addrmode5fp16:$addr)),
177          (VLDRH addrmode5fp16:$addr)> {
178  let Predicates = [HasFPRegs16];
179}
180def : Pat<(bf16 (alignedload16 addrmode3:$addr)),
181          (COPY_TO_REGCLASS (LDRH addrmode3:$addr), HPR)> {
182  let Predicates = [HasNoFPRegs16, IsARM];
183}
184def : Pat<(bf16 (alignedload16 t2addrmode_imm12:$addr)),
185          (COPY_TO_REGCLASS (t2LDRHi12 t2addrmode_imm12:$addr), HPR)> {
186  let Predicates = [HasNoFPRegs16, IsThumb];
187}
188
189def VSTRD : ADI5<0b1101, 0b00, (outs), (ins DPR:$Dd, addrmode5:$addr),
190                 IIC_fpStore64, "vstr", "\t$Dd, $addr",
191                 [(alignedstore32 (f64 DPR:$Dd), addrmode5:$addr)]>,
192            Requires<[HasFPRegs]>;
193
194def VSTRS : ASI5<0b1101, 0b00, (outs), (ins SPR:$Sd, addrmode5:$addr),
195                 IIC_fpStore32, "vstr", "\t$Sd, $addr",
196                 [(alignedstore32 SPR:$Sd, addrmode5:$addr)]>,
197            Requires<[HasFPRegs]> {
198  // Some single precision VFP instructions may be executed on both NEON and VFP
199  // pipelines.
200  let D = VFPNeonDomain;
201}
202
203let isUnpredicable = 1 in
204def VSTRH : AHI5<0b1101, 0b00, (outs), (ins HPR:$Sd, addrmode5fp16:$addr),
205                 IIC_fpStore16, "vstr", ".16\t$Sd, $addr",
206                 [(alignedstore16 (f16 HPR:$Sd), addrmode5fp16:$addr)]>,
207            Requires<[HasFPRegs16]>;
208
209def : Pat<(alignedstore16 (bf16 HPR:$Sd), addrmode5fp16:$addr),
210          (VSTRH (bf16 HPR:$Sd), addrmode5fp16:$addr)> {
211  let Predicates = [HasFPRegs16];
212}
213def : Pat<(alignedstore16 (bf16 HPR:$Sd), addrmode3:$addr),
214          (STRH (COPY_TO_REGCLASS $Sd, GPR), addrmode3:$addr)> {
215  let Predicates = [HasNoFPRegs16, IsARM];
216}
217def : Pat<(alignedstore16 (bf16 HPR:$Sd), t2addrmode_imm12:$addr),
218          (t2STRHi12 (COPY_TO_REGCLASS $Sd, GPR), t2addrmode_imm12:$addr)> {
219  let Predicates = [HasNoFPRegs16, IsThumb];
220}
221
222//===----------------------------------------------------------------------===//
223//  Load / store multiple Instructions.
224//
225
226multiclass vfp_ldst_mult<string asm, bit L_bit,
227                         InstrItinClass itin, InstrItinClass itin_upd> {
228  let Predicates = [HasFPRegs] in {
229  // Double Precision
230  def DIA :
231    AXDI4<(outs), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
232          IndexModeNone, itin,
233          !strconcat(asm, "ia${p}\t$Rn, $regs"), "", []> {
234    let Inst{24-23} = 0b01;       // Increment After
235    let Inst{21}    = 0;          // No writeback
236    let Inst{20}    = L_bit;
237  }
238  def DIA_UPD :
239    AXDI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs,
240                               variable_ops),
241          IndexModeUpd, itin_upd,
242          !strconcat(asm, "ia${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
243    let Inst{24-23} = 0b01;       // Increment After
244    let Inst{21}    = 1;          // Writeback
245    let Inst{20}    = L_bit;
246  }
247  def DDB_UPD :
248    AXDI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs,
249                               variable_ops),
250          IndexModeUpd, itin_upd,
251          !strconcat(asm, "db${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
252    let Inst{24-23} = 0b10;       // Decrement Before
253    let Inst{21}    = 1;          // Writeback
254    let Inst{20}    = L_bit;
255  }
256
257  // Single Precision
258  def SIA :
259    AXSI4<(outs), (ins GPR:$Rn, pred:$p, spr_reglist:$regs, variable_ops),
260          IndexModeNone, itin,
261          !strconcat(asm, "ia${p}\t$Rn, $regs"), "", []> {
262    let Inst{24-23} = 0b01;       // Increment After
263    let Inst{21}    = 0;          // No writeback
264    let Inst{20}    = L_bit;
265
266    // Some single precision VFP instructions may be executed on both NEON and
267    // VFP pipelines.
268    let D = VFPNeonDomain;
269  }
270  def SIA_UPD :
271    AXSI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, spr_reglist:$regs,
272                               variable_ops),
273          IndexModeUpd, itin_upd,
274          !strconcat(asm, "ia${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
275    let Inst{24-23} = 0b01;       // Increment After
276    let Inst{21}    = 1;          // Writeback
277    let Inst{20}    = L_bit;
278
279    // Some single precision VFP instructions may be executed on both NEON and
280    // VFP pipelines.
281    let D = VFPNeonDomain;
282  }
283  def SDB_UPD :
284    AXSI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, spr_reglist:$regs,
285                               variable_ops),
286          IndexModeUpd, itin_upd,
287          !strconcat(asm, "db${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
288    let Inst{24-23} = 0b10;       // Decrement Before
289    let Inst{21}    = 1;          // Writeback
290    let Inst{20}    = L_bit;
291
292    // Some single precision VFP instructions may be executed on both NEON and
293    // VFP pipelines.
294    let D = VFPNeonDomain;
295  }
296  }
297}
298
299let hasSideEffects = 0 in {
300
301let mayLoad = 1, hasExtraDefRegAllocReq = 1 in
302defm VLDM : vfp_ldst_mult<"vldm", 1, IIC_fpLoad_m, IIC_fpLoad_mu>;
303
304let mayStore = 1, hasExtraSrcRegAllocReq = 1 in
305defm VSTM : vfp_ldst_mult<"vstm", 0, IIC_fpStore_m, IIC_fpStore_mu>;
306
307} // hasSideEffects
308
309def : MnemonicAlias<"vldm", "vldmia">;
310def : MnemonicAlias<"vstm", "vstmia">;
311
312
313//===----------------------------------------------------------------------===//
314//  Lazy load / store multiple Instructions
315//
316def VLLDM : AXSI4<(outs), (ins GPRnopc:$Rn, pred:$p), IndexModeNone,
317                  NoItinerary, "vlldm${p}\t$Rn", "", []>,
318            Requires<[HasV8MMainline, Has8MSecExt]> {
319    let Inst{24-23} = 0b00;
320    let Inst{22}    = 0;
321    let Inst{21}    = 1;
322    let Inst{20}    = 1;
323    let Inst{15-12} = 0;
324    let Inst{7-0}   = 0;
325    let mayLoad     = 1;
326    let Defs = [Q0, Q1, Q2, Q3, Q4, Q5, Q6, Q7, VPR, FPSCR, FPSCR_NZCV];
327}
328
329def VLSTM : AXSI4<(outs), (ins GPRnopc:$Rn, pred:$p), IndexModeNone,
330                  NoItinerary, "vlstm${p}\t$Rn", "", []>,
331            Requires<[HasV8MMainline, Has8MSecExt]> {
332    let Inst{24-23} = 0b00;
333    let Inst{22}    = 0;
334    let Inst{21}    = 1;
335    let Inst{20}    = 0;
336    let Inst{15-12} = 0;
337    let Inst{7-0}   = 0;
338    let mayStore    = 1;
339}
340
341def : InstAlias<"vpush${p} $r", (VSTMDDB_UPD SP, pred:$p, dpr_reglist:$r), 0>,
342                Requires<[HasFPRegs]>;
343def : InstAlias<"vpush${p} $r", (VSTMSDB_UPD SP, pred:$p, spr_reglist:$r), 0>,
344                Requires<[HasFPRegs]>;
345def : InstAlias<"vpop${p} $r",  (VLDMDIA_UPD SP, pred:$p, dpr_reglist:$r), 0>,
346                Requires<[HasFPRegs]>;
347def : InstAlias<"vpop${p} $r",  (VLDMSIA_UPD SP, pred:$p, spr_reglist:$r), 0>,
348                Requires<[HasFPRegs]>;
349defm : VFPDTAnyInstAlias<"vpush${p}", "$r",
350                         (VSTMSDB_UPD SP, pred:$p, spr_reglist:$r)>;
351defm : VFPDTAnyInstAlias<"vpush${p}", "$r",
352                         (VSTMDDB_UPD SP, pred:$p, dpr_reglist:$r)>;
353defm : VFPDTAnyInstAlias<"vpop${p}", "$r",
354                         (VLDMSIA_UPD SP, pred:$p, spr_reglist:$r)>;
355defm : VFPDTAnyInstAlias<"vpop${p}", "$r",
356                         (VLDMDIA_UPD SP, pred:$p, dpr_reglist:$r)>;
357
358// FLDMX, FSTMX - Load and store multiple unknown precision registers for
359// pre-armv6 cores.
360// These instruction are deprecated so we don't want them to get selected.
361// However, there is no UAL syntax for them, so we keep them around for
362// (dis)assembly only.
363multiclass vfp_ldstx_mult<string asm, bit L_bit> {
364  let Predicates = [HasFPRegs], hasNoSchedulingInfo = 1 in {
365  // Unknown precision
366  def XIA :
367    AXXI4<(outs), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
368          IndexModeNone, !strconcat(asm, "iax${p}\t$Rn, $regs"), "", []> {
369    let Inst{24-23} = 0b01;       // Increment After
370    let Inst{21}    = 0;          // No writeback
371    let Inst{20}    = L_bit;
372  }
373  def XIA_UPD :
374    AXXI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
375          IndexModeUpd, !strconcat(asm, "iax${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
376    let Inst{24-23} = 0b01;         // Increment After
377    let Inst{21}    = 1;            // Writeback
378    let Inst{20}    = L_bit;
379  }
380  def XDB_UPD :
381    AXXI4<(outs GPR:$wb), (ins GPR:$Rn, pred:$p, dpr_reglist:$regs, variable_ops),
382          IndexModeUpd, !strconcat(asm, "dbx${p}\t$Rn!, $regs"), "$Rn = $wb", []> {
383    let Inst{24-23} = 0b10;         // Decrement Before
384    let Inst{21}    = 1;            // Writeback
385    let Inst{20}    = L_bit;
386  }
387  }
388}
389
390defm FLDM : vfp_ldstx_mult<"fldm", 1>;
391defm FSTM : vfp_ldstx_mult<"fstm", 0>;
392
393def : VFP2MnemonicAlias<"fldmeax", "fldmdbx">;
394def : VFP2MnemonicAlias<"fldmfdx", "fldmiax">;
395
396def : VFP2MnemonicAlias<"fstmeax", "fstmiax">;
397def : VFP2MnemonicAlias<"fstmfdx", "fstmdbx">;
398
399//===----------------------------------------------------------------------===//
400// FP Binary Operations.
401//
402
403let TwoOperandAliasConstraint = "$Dn = $Dd" in
404def VADDD  : ADbI<0b11100, 0b11, 0, 0,
405                  (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
406                  IIC_fpALU64, "vadd", ".f64\t$Dd, $Dn, $Dm",
407                  [(set DPR:$Dd, (fadd DPR:$Dn, (f64 DPR:$Dm)))]>,
408             Sched<[WriteFPALU64]>;
409
410let TwoOperandAliasConstraint = "$Sn = $Sd" in
411def VADDS  : ASbIn<0b11100, 0b11, 0, 0,
412                   (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
413                   IIC_fpALU32, "vadd", ".f32\t$Sd, $Sn, $Sm",
414                   [(set SPR:$Sd, (fadd SPR:$Sn, SPR:$Sm))]>,
415             Sched<[WriteFPALU32]> {
416  // Some single precision VFP instructions may be executed on both NEON and
417  // VFP pipelines on A8.
418  let D = VFPNeonA8Domain;
419}
420
421let TwoOperandAliasConstraint = "$Sn = $Sd" in
422def VADDH  : AHbI<0b11100, 0b11, 0, 0,
423                  (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
424                  IIC_fpALU16, "vadd", ".f16\t$Sd, $Sn, $Sm",
425                  [(set (f16 HPR:$Sd), (fadd (f16 HPR:$Sn), (f16 HPR:$Sm)))]>,
426             Sched<[WriteFPALU32]>;
427
428let TwoOperandAliasConstraint = "$Dn = $Dd" in
429def VSUBD  : ADbI<0b11100, 0b11, 1, 0,
430                  (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
431                  IIC_fpALU64, "vsub", ".f64\t$Dd, $Dn, $Dm",
432                  [(set DPR:$Dd, (fsub DPR:$Dn, (f64 DPR:$Dm)))]>,
433             Sched<[WriteFPALU64]>;
434
435let TwoOperandAliasConstraint = "$Sn = $Sd" in
436def VSUBS  : ASbIn<0b11100, 0b11, 1, 0,
437                   (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
438                   IIC_fpALU32, "vsub", ".f32\t$Sd, $Sn, $Sm",
439                   [(set SPR:$Sd, (fsub SPR:$Sn, SPR:$Sm))]>,
440             Sched<[WriteFPALU32]>{
441  // Some single precision VFP instructions may be executed on both NEON and
442  // VFP pipelines on A8.
443  let D = VFPNeonA8Domain;
444}
445
446let TwoOperandAliasConstraint = "$Sn = $Sd" in
447def VSUBH  : AHbI<0b11100, 0b11, 1, 0,
448                  (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
449                  IIC_fpALU16, "vsub", ".f16\t$Sd, $Sn, $Sm",
450                  [(set (f16 HPR:$Sd), (fsub (f16 HPR:$Sn), (f16 HPR:$Sm)))]>,
451            Sched<[WriteFPALU32]>;
452
453let TwoOperandAliasConstraint = "$Dn = $Dd" in
454def VDIVD  : ADbI<0b11101, 0b00, 0, 0,
455                  (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
456                  IIC_fpDIV64, "vdiv", ".f64\t$Dd, $Dn, $Dm",
457                  [(set DPR:$Dd, (fdiv DPR:$Dn, (f64 DPR:$Dm)))]>,
458             Sched<[WriteFPDIV64]>;
459
460let TwoOperandAliasConstraint = "$Sn = $Sd" in
461def VDIVS  : ASbI<0b11101, 0b00, 0, 0,
462                  (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
463                  IIC_fpDIV32, "vdiv", ".f32\t$Sd, $Sn, $Sm",
464                  [(set SPR:$Sd, (fdiv SPR:$Sn, SPR:$Sm))]>,
465             Sched<[WriteFPDIV32]>;
466
467let TwoOperandAliasConstraint = "$Sn = $Sd" in
468def VDIVH  : AHbI<0b11101, 0b00, 0, 0,
469                  (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
470                  IIC_fpDIV16, "vdiv", ".f16\t$Sd, $Sn, $Sm",
471                  [(set (f16 HPR:$Sd), (fdiv (f16 HPR:$Sn), (f16 HPR:$Sm)))]>,
472             Sched<[WriteFPDIV32]>;
473
474let TwoOperandAliasConstraint = "$Dn = $Dd" in
475def VMULD  : ADbI<0b11100, 0b10, 0, 0,
476                  (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
477                  IIC_fpMUL64, "vmul", ".f64\t$Dd, $Dn, $Dm",
478                  [(set DPR:$Dd, (fmul DPR:$Dn, (f64 DPR:$Dm)))]>,
479             Sched<[WriteFPMUL64, ReadFPMUL, ReadFPMUL]>;
480
481let TwoOperandAliasConstraint = "$Sn = $Sd" in
482def VMULS  : ASbIn<0b11100, 0b10, 0, 0,
483                   (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
484                   IIC_fpMUL32, "vmul", ".f32\t$Sd, $Sn, $Sm",
485                   [(set SPR:$Sd, (fmul SPR:$Sn, SPR:$Sm))]>,
486            Sched<[WriteFPMUL32, ReadFPMUL, ReadFPMUL]> {
487  // Some single precision VFP instructions may be executed on both NEON and
488  // VFP pipelines on A8.
489  let D = VFPNeonA8Domain;
490}
491
492let TwoOperandAliasConstraint = "$Sn = $Sd" in
493def VMULH  : AHbI<0b11100, 0b10, 0, 0,
494                  (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
495                  IIC_fpMUL16, "vmul", ".f16\t$Sd, $Sn, $Sm",
496                  [(set (f16 HPR:$Sd), (fmul (f16 HPR:$Sn), (f16 HPR:$Sm)))]>,
497             Sched<[WriteFPMUL32, ReadFPMUL, ReadFPMUL]>;
498
499def VNMULD : ADbI<0b11100, 0b10, 1, 0,
500                  (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
501                  IIC_fpMUL64, "vnmul", ".f64\t$Dd, $Dn, $Dm",
502                  [(set DPR:$Dd, (fneg (fmul DPR:$Dn, (f64 DPR:$Dm))))]>,
503             Sched<[WriteFPMUL64, ReadFPMUL, ReadFPMUL]>;
504
505def VNMULS : ASbI<0b11100, 0b10, 1, 0,
506                  (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
507                  IIC_fpMUL32, "vnmul", ".f32\t$Sd, $Sn, $Sm",
508                  [(set SPR:$Sd, (fneg (fmul SPR:$Sn, SPR:$Sm)))]>,
509            Sched<[WriteFPMUL32, ReadFPMUL, ReadFPMUL]> {
510  // Some single precision VFP instructions may be executed on both NEON and
511  // VFP pipelines on A8.
512  let D = VFPNeonA8Domain;
513}
514
515def VNMULH : AHbI<0b11100, 0b10, 1, 0,
516                  (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
517                  IIC_fpMUL16, "vnmul", ".f16\t$Sd, $Sn, $Sm",
518                  [(set (f16 HPR:$Sd), (fneg (fmul (f16 HPR:$Sn), (f16 HPR:$Sm))))]>,
519             Sched<[WriteFPMUL32, ReadFPMUL, ReadFPMUL]>;
520
521multiclass vsel_inst<string op, bits<2> opc, int CC> {
522  let DecoderNamespace = "VFPV8", PostEncoderMethod = "",
523      Uses = [CPSR], AddedComplexity = 4, isUnpredicable = 1 in {
524    def H : AHbInp<0b11100, opc, 0,
525                   (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
526                   NoItinerary, !strconcat("vsel", op, ".f16\t$Sd, $Sn, $Sm"),
527                   [(set (f16 HPR:$Sd), (ARMcmov (f16 HPR:$Sm), (f16 HPR:$Sn), CC))]>,
528                   Requires<[HasFullFP16]>;
529
530    def S : ASbInp<0b11100, opc, 0,
531                   (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
532                   NoItinerary, !strconcat("vsel", op, ".f32\t$Sd, $Sn, $Sm"),
533                   [(set SPR:$Sd, (ARMcmov SPR:$Sm, SPR:$Sn, CC))]>,
534                   Requires<[HasFPARMv8]>;
535
536    def D : ADbInp<0b11100, opc, 0,
537                   (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
538                   NoItinerary, !strconcat("vsel", op, ".f64\t$Dd, $Dn, $Dm"),
539                   [(set DPR:$Dd, (ARMcmov (f64 DPR:$Dm), (f64 DPR:$Dn), CC))]>,
540                   Requires<[HasFPARMv8, HasDPVFP]>;
541  }
542}
543
544// The CC constants here match ARMCC::CondCodes.
545defm VSELGT : vsel_inst<"gt", 0b11, 12>;
546defm VSELGE : vsel_inst<"ge", 0b10, 10>;
547defm VSELEQ : vsel_inst<"eq", 0b00, 0>;
548defm VSELVS : vsel_inst<"vs", 0b01, 6>;
549
550multiclass vmaxmin_inst<string op, bit opc, SDNode SD> {
551  let DecoderNamespace = "VFPV8", PostEncoderMethod = "",
552      isUnpredicable = 1 in {
553    def H : AHbInp<0b11101, 0b00, opc,
554                   (outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm),
555                   NoItinerary, !strconcat(op, ".f16\t$Sd, $Sn, $Sm"),
556                   [(set (f16 HPR:$Sd), (SD (f16 HPR:$Sn), (f16 HPR:$Sm)))]>,
557                   Requires<[HasFullFP16]>;
558
559    def S : ASbInp<0b11101, 0b00, opc,
560                   (outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
561                   NoItinerary, !strconcat(op, ".f32\t$Sd, $Sn, $Sm"),
562                   [(set SPR:$Sd, (SD SPR:$Sn, SPR:$Sm))]>,
563                   Requires<[HasFPARMv8]>;
564
565    def D : ADbInp<0b11101, 0b00, opc,
566                   (outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm),
567                   NoItinerary, !strconcat(op, ".f64\t$Dd, $Dn, $Dm"),
568                   [(set DPR:$Dd, (f64 (SD (f64 DPR:$Dn), (f64 DPR:$Dm))))]>,
569                   Requires<[HasFPARMv8, HasDPVFP]>;
570  }
571}
572
573defm VFP_VMAXNM : vmaxmin_inst<"vmaxnm", 0, fmaxnum>;
574defm VFP_VMINNM : vmaxmin_inst<"vminnm", 1, fminnum>;
575
576// Match reassociated forms only if not sign dependent rounding.
577def : Pat<(fmul (fneg DPR:$a), (f64 DPR:$b)),
578          (VNMULD DPR:$a, DPR:$b)>,
579          Requires<[NoHonorSignDependentRounding,HasDPVFP]>;
580def : Pat<(fmul (fneg SPR:$a), SPR:$b),
581          (VNMULS SPR:$a, SPR:$b)>, Requires<[NoHonorSignDependentRounding]>;
582
583// These are encoded as unary instructions.
584let Defs = [FPSCR_NZCV] in {
585def VCMPED : ADuI<0b11101, 0b11, 0b0100, 0b11, 0,
586                  (outs), (ins DPR:$Dd, DPR:$Dm),
587                  IIC_fpCMP64, "vcmpe", ".f64\t$Dd, $Dm",
588                  [(arm_cmpfpe DPR:$Dd, (f64 DPR:$Dm))]>;
589
590def VCMPES : ASuI<0b11101, 0b11, 0b0100, 0b11, 0,
591                  (outs), (ins SPR:$Sd, SPR:$Sm),
592                  IIC_fpCMP32, "vcmpe", ".f32\t$Sd, $Sm",
593                  [(arm_cmpfpe SPR:$Sd, SPR:$Sm)]> {
594  // Some single precision VFP instructions may be executed on both NEON and
595  // VFP pipelines on A8.
596  let D = VFPNeonA8Domain;
597}
598
599def VCMPEH : AHuI<0b11101, 0b11, 0b0100, 0b11, 0,
600                  (outs), (ins HPR:$Sd, HPR:$Sm),
601                  IIC_fpCMP16, "vcmpe", ".f16\t$Sd, $Sm",
602                  [(arm_cmpfpe (f16 HPR:$Sd), (f16 HPR:$Sm))]>;
603
604def VCMPD  : ADuI<0b11101, 0b11, 0b0100, 0b01, 0,
605                  (outs), (ins DPR:$Dd, DPR:$Dm),
606                  IIC_fpCMP64, "vcmp", ".f64\t$Dd, $Dm",
607                  [(arm_cmpfp DPR:$Dd, (f64 DPR:$Dm))]>;
608
609def VCMPS  : ASuI<0b11101, 0b11, 0b0100, 0b01, 0,
610                  (outs), (ins SPR:$Sd, SPR:$Sm),
611                  IIC_fpCMP32, "vcmp", ".f32\t$Sd, $Sm",
612                  [(arm_cmpfp SPR:$Sd, SPR:$Sm)]> {
613  // Some single precision VFP instructions may be executed on both NEON and
614  // VFP pipelines on A8.
615  let D = VFPNeonA8Domain;
616}
617
618def VCMPH  : AHuI<0b11101, 0b11, 0b0100, 0b01, 0,
619                  (outs), (ins HPR:$Sd, HPR:$Sm),
620                  IIC_fpCMP16, "vcmp", ".f16\t$Sd, $Sm",
621                  [(arm_cmpfp (f16 HPR:$Sd), (f16 HPR:$Sm))]>;
622} // Defs = [FPSCR_NZCV]
623
624//===----------------------------------------------------------------------===//
625// FP Unary Operations.
626//
627
628def VABSD  : ADuI<0b11101, 0b11, 0b0000, 0b11, 0,
629                  (outs DPR:$Dd), (ins DPR:$Dm),
630                  IIC_fpUNA64, "vabs", ".f64\t$Dd, $Dm",
631                  [(set DPR:$Dd, (fabs (f64 DPR:$Dm)))]>;
632
633def VABSS  : ASuIn<0b11101, 0b11, 0b0000, 0b11, 0,
634                   (outs SPR:$Sd), (ins SPR:$Sm),
635                   IIC_fpUNA32, "vabs", ".f32\t$Sd, $Sm",
636                   [(set SPR:$Sd, (fabs SPR:$Sm))]> {
637  // Some single precision VFP instructions may be executed on both NEON and
638  // VFP pipelines on A8.
639  let D = VFPNeonA8Domain;
640}
641
642def VABSH  : AHuI<0b11101, 0b11, 0b0000, 0b11, 0,
643                   (outs HPR:$Sd), (ins HPR:$Sm),
644                   IIC_fpUNA16, "vabs", ".f16\t$Sd, $Sm",
645                   [(set (f16 HPR:$Sd), (fabs (f16 HPR:$Sm)))]>;
646
647let Defs = [FPSCR_NZCV] in {
648def VCMPEZD : ADuI<0b11101, 0b11, 0b0101, 0b11, 0,
649                   (outs), (ins DPR:$Dd),
650                   IIC_fpCMP64, "vcmpe", ".f64\t$Dd, #0",
651                   [(arm_cmpfpe0 (f64 DPR:$Dd))]> {
652  let Inst{3-0} = 0b0000;
653  let Inst{5}   = 0;
654}
655
656def VCMPEZS : ASuI<0b11101, 0b11, 0b0101, 0b11, 0,
657                   (outs), (ins SPR:$Sd),
658                   IIC_fpCMP32, "vcmpe", ".f32\t$Sd, #0",
659                   [(arm_cmpfpe0 SPR:$Sd)]> {
660  let Inst{3-0} = 0b0000;
661  let Inst{5}   = 0;
662
663  // Some single precision VFP instructions may be executed on both NEON and
664  // VFP pipelines on A8.
665  let D = VFPNeonA8Domain;
666}
667
668def VCMPEZH : AHuI<0b11101, 0b11, 0b0101, 0b11, 0,
669                   (outs), (ins HPR:$Sd),
670                   IIC_fpCMP16, "vcmpe", ".f16\t$Sd, #0",
671                   [(arm_cmpfpe0 (f16 HPR:$Sd))]> {
672  let Inst{3-0} = 0b0000;
673  let Inst{5}   = 0;
674}
675
676def VCMPZD  : ADuI<0b11101, 0b11, 0b0101, 0b01, 0,
677                   (outs), (ins DPR:$Dd),
678                   IIC_fpCMP64, "vcmp", ".f64\t$Dd, #0",
679                   [(arm_cmpfp0 (f64 DPR:$Dd))]> {
680  let Inst{3-0} = 0b0000;
681  let Inst{5}   = 0;
682}
683
684def VCMPZS  : ASuI<0b11101, 0b11, 0b0101, 0b01, 0,
685                   (outs), (ins SPR:$Sd),
686                   IIC_fpCMP32, "vcmp", ".f32\t$Sd, #0",
687                   [(arm_cmpfp0 SPR:$Sd)]> {
688  let Inst{3-0} = 0b0000;
689  let Inst{5}   = 0;
690
691  // Some single precision VFP instructions may be executed on both NEON and
692  // VFP pipelines on A8.
693  let D = VFPNeonA8Domain;
694}
695
696def VCMPZH  : AHuI<0b11101, 0b11, 0b0101, 0b01, 0,
697                   (outs), (ins HPR:$Sd),
698                   IIC_fpCMP16, "vcmp", ".f16\t$Sd, #0",
699                   [(arm_cmpfp0 (f16 HPR:$Sd))]> {
700  let Inst{3-0} = 0b0000;
701  let Inst{5}   = 0;
702}
703} // Defs = [FPSCR_NZCV]
704
705def VCVTDS  : ASuI<0b11101, 0b11, 0b0111, 0b11, 0,
706                   (outs DPR:$Dd), (ins SPR:$Sm),
707                   IIC_fpCVTDS, "vcvt", ".f64.f32\t$Dd, $Sm",
708                   [(set DPR:$Dd, (fpextend SPR:$Sm))]>,
709             Sched<[WriteFPCVT]> {
710  // Instruction operands.
711  bits<5> Dd;
712  bits<5> Sm;
713
714  // Encode instruction operands.
715  let Inst{3-0}   = Sm{4-1};
716  let Inst{5}     = Sm{0};
717  let Inst{15-12} = Dd{3-0};
718  let Inst{22}    = Dd{4};
719
720  let Predicates = [HasVFP2, HasDPVFP];
721  let hasSideEffects = 0;
722}
723
724// Special case encoding: bits 11-8 is 0b1011.
725def VCVTSD  : VFPAI<(outs SPR:$Sd), (ins DPR:$Dm), VFPUnaryFrm,
726                    IIC_fpCVTSD, "vcvt", ".f32.f64\t$Sd, $Dm",
727                    [(set SPR:$Sd, (fpround DPR:$Dm))]>,
728              Sched<[WriteFPCVT]> {
729  // Instruction operands.
730  bits<5> Sd;
731  bits<5> Dm;
732
733  // Encode instruction operands.
734  let Inst{3-0}   = Dm{3-0};
735  let Inst{5}     = Dm{4};
736  let Inst{15-12} = Sd{4-1};
737  let Inst{22}    = Sd{0};
738
739  let Inst{27-23} = 0b11101;
740  let Inst{21-16} = 0b110111;
741  let Inst{11-8}  = 0b1011;
742  let Inst{7-6}   = 0b11;
743  let Inst{4}     = 0;
744
745  let Predicates = [HasVFP2, HasDPVFP];
746  let hasSideEffects = 0;
747}
748
749// Between half, single and double-precision.
750let hasSideEffects = 0 in
751def VCVTBHS: ASuI<0b11101, 0b11, 0b0010, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
752                 /* FIXME */ IIC_fpCVTSH, "vcvtb", ".f32.f16\t$Sd, $Sm",
753                 [/* Intentionally left blank, see patterns below */]>,
754                 Requires<[HasFP16]>,
755             Sched<[WriteFPCVT]>;
756
757def : FP16Pat<(f32 (fpextend (f16 HPR:$Sm))),
758              (VCVTBHS (COPY_TO_REGCLASS (f16 HPR:$Sm), SPR))>;
759def : FP16Pat<(f16_to_fp GPR:$a),
760              (VCVTBHS (COPY_TO_REGCLASS GPR:$a, SPR))>;
761
762let hasSideEffects = 0 in
763def VCVTBSH: ASuI<0b11101, 0b11, 0b0011, 0b01, 0, (outs SPR:$Sd), (ins SPR:$Sm),
764                 /* FIXME */ IIC_fpCVTHS, "vcvtb", ".f16.f32\t$Sd, $Sm",
765                 [/* Intentionally left blank, see patterns below */]>,
766                 Requires<[HasFP16]>,
767             Sched<[WriteFPCVT]>;
768
769def : FP16Pat<(f16 (fpround SPR:$Sm)),
770              (COPY_TO_REGCLASS (VCVTBSH SPR:$Sm), HPR)>;
771def : FP16Pat<(fp_to_f16 SPR:$a),
772              (i32 (COPY_TO_REGCLASS (VCVTBSH SPR:$a), GPR))>;
773def : FP16Pat<(insertelt (v8f16 MQPR:$src1), (f16 (fpround (f32 SPR:$src2))), imm_even:$lane),
774              (v8f16 (INSERT_SUBREG (v8f16 MQPR:$src1), (VCVTBSH SPR:$src2),
775                                    (SSubReg_f16_reg imm:$lane)))>;
776def : FP16Pat<(insertelt (v4f16 DPR:$src1), (f16 (fpround (f32 SPR:$src2))), imm_even:$lane),
777              (v4f16 (INSERT_SUBREG (v4f16 DPR:$src1), (VCVTBSH SPR:$src2),
778                                    (SSubReg_f16_reg imm:$lane)))>;
779
780let hasSideEffects = 0 in
781def VCVTTHS: ASuI<0b11101, 0b11, 0b0010, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
782                 /* FIXME */ IIC_fpCVTSH, "vcvtt", ".f32.f16\t$Sd, $Sm",
783                 [/* Intentionally left blank, see patterns below */]>,
784                 Requires<[HasFP16]>,
785             Sched<[WriteFPCVT]>;
786
787def : FP16Pat<(f32 (fpextend (extractelt (v8f16 MQPR:$src), imm_odd:$lane))),
788              (VCVTTHS (EXTRACT_SUBREG MQPR:$src, (SSubReg_f16_reg imm_odd:$lane)))>;
789def : FP16Pat<(f32 (fpextend (extractelt (v4f16 DPR:$src), imm_odd:$lane))),
790              (VCVTTHS (EXTRACT_SUBREG
791                (v2f32 (COPY_TO_REGCLASS (v4f16 DPR:$src), DPR_VFP2)),
792                (SSubReg_f16_reg imm_odd:$lane)))>;
793
794let hasSideEffects = 0 in
795def VCVTTSH: ASuI<0b11101, 0b11, 0b0011, 0b11, 0, (outs SPR:$Sd), (ins SPR:$Sm),
796                 /* FIXME */ IIC_fpCVTHS, "vcvtt", ".f16.f32\t$Sd, $Sm",
797                 [/* Intentionally left blank, see patterns below */]>,
798                 Requires<[HasFP16]>,
799            Sched<[WriteFPCVT]>;
800
801def : FP16Pat<(insertelt (v8f16 MQPR:$src1), (f16 (fpround (f32 SPR:$src2))), imm_odd:$lane),
802              (v8f16 (INSERT_SUBREG (v8f16 MQPR:$src1), (VCVTTSH SPR:$src2),
803                                    (SSubReg_f16_reg imm:$lane)))>;
804def : FP16Pat<(insertelt (v4f16 DPR:$src1), (f16 (fpround (f32 SPR:$src2))), imm_odd:$lane),
805              (v4f16 (INSERT_SUBREG (v4f16 DPR:$src1), (VCVTTSH SPR:$src2),
806                                    (SSubReg_f16_reg imm:$lane)))>;
807
808def VCVTBHD : ADuI<0b11101, 0b11, 0b0010, 0b01, 0,
809                   (outs DPR:$Dd), (ins SPR:$Sm),
810                   NoItinerary, "vcvtb", ".f64.f16\t$Dd, $Sm",
811                   [/* Intentionally left blank, see patterns below */]>,
812                   Requires<[HasFPARMv8, HasDPVFP]>,
813              Sched<[WriteFPCVT]> {
814  // Instruction operands.
815  bits<5> Sm;
816
817  // Encode instruction operands.
818  let Inst{3-0} = Sm{4-1};
819  let Inst{5}   = Sm{0};
820
821  let hasSideEffects = 0;
822}
823
824def : FullFP16Pat<(f64 (fpextend (f16 HPR:$Sm))),
825                  (VCVTBHD (COPY_TO_REGCLASS (f16 HPR:$Sm), SPR))>,
826                  Requires<[HasFPARMv8, HasDPVFP]>;
827def : FP16Pat<(f64 (f16_to_fp GPR:$a)),
828              (VCVTBHD (COPY_TO_REGCLASS GPR:$a, SPR))>,
829              Requires<[HasFPARMv8, HasDPVFP]>;
830
831def VCVTBDH : ADuI<0b11101, 0b11, 0b0011, 0b01, 0,
832                   (outs SPR:$Sd), (ins DPR:$Dm),
833                   NoItinerary, "vcvtb", ".f16.f64\t$Sd, $Dm",
834                   [/* Intentionally left blank, see patterns below */]>,
835                   Requires<[HasFPARMv8, HasDPVFP]> {
836  // Instruction operands.
837  bits<5> Sd;
838  bits<5> Dm;
839
840  // Encode instruction operands.
841  let Inst{3-0}     = Dm{3-0};
842  let Inst{5}       = Dm{4};
843  let Inst{15-12}   = Sd{4-1};
844  let Inst{22}      = Sd{0};
845
846  let hasSideEffects = 0;
847}
848
849def : FullFP16Pat<(f16 (fpround DPR:$Dm)),
850                  (COPY_TO_REGCLASS (VCVTBDH DPR:$Dm), HPR)>,
851                  Requires<[HasFPARMv8, HasDPVFP]>;
852def : FP16Pat<(fp_to_f16 (f64 DPR:$a)),
853              (i32 (COPY_TO_REGCLASS (VCVTBDH DPR:$a), GPR))>,
854                   Requires<[HasFPARMv8, HasDPVFP]>;
855
856def VCVTTHD : ADuI<0b11101, 0b11, 0b0010, 0b11, 0,
857                   (outs DPR:$Dd), (ins SPR:$Sm),
858                   NoItinerary, "vcvtt", ".f64.f16\t$Dd, $Sm",
859                   []>, Requires<[HasFPARMv8, HasDPVFP]> {
860  // Instruction operands.
861  bits<5> Sm;
862
863  // Encode instruction operands.
864  let Inst{3-0} = Sm{4-1};
865  let Inst{5}   = Sm{0};
866
867  let hasSideEffects = 0;
868}
869
870def VCVTTDH : ADuI<0b11101, 0b11, 0b0011, 0b11, 0,
871                   (outs SPR:$Sd), (ins DPR:$Dm),
872                   NoItinerary, "vcvtt", ".f16.f64\t$Sd, $Dm",
873                   []>, Requires<[HasFPARMv8, HasDPVFP]> {
874  // Instruction operands.
875  bits<5> Sd;
876  bits<5> Dm;
877
878  // Encode instruction operands.
879  let Inst{15-12} = Sd{4-1};
880  let Inst{22}    = Sd{0};
881  let Inst{3-0}   = Dm{3-0};
882  let Inst{5}     = Dm{4};
883
884  let hasSideEffects = 0;
885}
886
887multiclass vcvt_inst<string opc, bits<2> rm,
888                     SDPatternOperator node = null_frag> {
889  let PostEncoderMethod = "", DecoderNamespace = "VFPV8", hasSideEffects = 0 in {
890    def SH : AHuInp<0b11101, 0b11, 0b1100, 0b11, 0,
891                    (outs SPR:$Sd), (ins HPR:$Sm),
892                    NoItinerary, !strconcat("vcvt", opc, ".s32.f16\t$Sd, $Sm"),
893                    []>,
894                    Requires<[HasFullFP16]> {
895      let Inst{17-16} = rm;
896    }
897
898    def UH : AHuInp<0b11101, 0b11, 0b1100, 0b01, 0,
899                    (outs SPR:$Sd), (ins HPR:$Sm),
900                    NoItinerary, !strconcat("vcvt", opc, ".u32.f16\t$Sd, $Sm"),
901                    []>,
902                    Requires<[HasFullFP16]> {
903      let Inst{17-16} = rm;
904    }
905
906    def SS : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
907                    (outs SPR:$Sd), (ins SPR:$Sm),
908                    NoItinerary, !strconcat("vcvt", opc, ".s32.f32\t$Sd, $Sm"),
909                    []>,
910                    Requires<[HasFPARMv8]> {
911      let Inst{17-16} = rm;
912    }
913
914    def US : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
915                    (outs SPR:$Sd), (ins SPR:$Sm),
916                    NoItinerary, !strconcat("vcvt", opc, ".u32.f32\t$Sd, $Sm"),
917                    []>,
918                    Requires<[HasFPARMv8]> {
919      let Inst{17-16} = rm;
920    }
921
922    def SD : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
923                    (outs SPR:$Sd), (ins DPR:$Dm),
924                    NoItinerary, !strconcat("vcvt", opc, ".s32.f64\t$Sd, $Dm"),
925                    []>,
926                    Requires<[HasFPARMv8, HasDPVFP]> {
927      bits<5> Dm;
928
929      let Inst{17-16} = rm;
930
931      // Encode instruction operands.
932      let Inst{3-0} = Dm{3-0};
933      let Inst{5}   = Dm{4};
934      let Inst{8} = 1;
935    }
936
937    def UD : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
938                    (outs SPR:$Sd), (ins DPR:$Dm),
939                    NoItinerary, !strconcat("vcvt", opc, ".u32.f64\t$Sd, $Dm"),
940                    []>,
941                    Requires<[HasFPARMv8, HasDPVFP]> {
942      bits<5> Dm;
943
944      let Inst{17-16} = rm;
945
946      // Encode instruction operands
947      let Inst{3-0}  = Dm{3-0};
948      let Inst{5}    = Dm{4};
949      let Inst{8} = 1;
950    }
951  }
952
953  let Predicates = [HasFPARMv8] in {
954    let Predicates = [HasFullFP16] in {
955    def : Pat<(i32 (fp_to_sint (node (f16 HPR:$a)))),
956              (COPY_TO_REGCLASS
957                (!cast<Instruction>(NAME#"SH") (f16 HPR:$a)),
958                GPR)>;
959
960    def : Pat<(i32 (fp_to_uint (node (f16 HPR:$a)))),
961              (COPY_TO_REGCLASS
962                (!cast<Instruction>(NAME#"UH") (f16 HPR:$a)),
963                GPR)>;
964    }
965    def : Pat<(i32 (fp_to_sint (node SPR:$a))),
966              (COPY_TO_REGCLASS
967                (!cast<Instruction>(NAME#"SS") SPR:$a),
968                GPR)>;
969    def : Pat<(i32 (fp_to_uint (node SPR:$a))),
970              (COPY_TO_REGCLASS
971                (!cast<Instruction>(NAME#"US") SPR:$a),
972                GPR)>;
973  }
974  let Predicates = [HasFPARMv8, HasDPVFP] in {
975    def : Pat<(i32 (fp_to_sint (node (f64 DPR:$a)))),
976              (COPY_TO_REGCLASS
977                (!cast<Instruction>(NAME#"SD") DPR:$a),
978                GPR)>;
979    def : Pat<(i32 (fp_to_uint (node (f64 DPR:$a)))),
980              (COPY_TO_REGCLASS
981                (!cast<Instruction>(NAME#"UD") DPR:$a),
982                GPR)>;
983  }
984}
985
986defm VCVTA : vcvt_inst<"a", 0b00, fround>;
987defm VCVTN : vcvt_inst<"n", 0b01>;
988defm VCVTP : vcvt_inst<"p", 0b10, fceil>;
989defm VCVTM : vcvt_inst<"m", 0b11, ffloor>;
990
991def VNEGD  : ADuI<0b11101, 0b11, 0b0001, 0b01, 0,
992                  (outs DPR:$Dd), (ins DPR:$Dm),
993                  IIC_fpUNA64, "vneg", ".f64\t$Dd, $Dm",
994                  [(set DPR:$Dd, (fneg (f64 DPR:$Dm)))]>;
995
996def VNEGS  : ASuIn<0b11101, 0b11, 0b0001, 0b01, 0,
997                   (outs SPR:$Sd), (ins SPR:$Sm),
998                   IIC_fpUNA32, "vneg", ".f32\t$Sd, $Sm",
999                   [(set SPR:$Sd, (fneg SPR:$Sm))]> {
1000  // Some single precision VFP instructions may be executed on both NEON and
1001  // VFP pipelines on A8.
1002  let D = VFPNeonA8Domain;
1003}
1004
1005def VNEGH  : AHuI<0b11101, 0b11, 0b0001, 0b01, 0,
1006                  (outs HPR:$Sd), (ins HPR:$Sm),
1007                  IIC_fpUNA16, "vneg", ".f16\t$Sd, $Sm",
1008                  [(set (f16 HPR:$Sd), (fneg (f16 HPR:$Sm)))]>;
1009
1010multiclass vrint_inst_zrx<string opc, bit op, bit op2, SDPatternOperator node> {
1011  def H : AHuI<0b11101, 0b11, 0b0110, 0b11, 0,
1012               (outs HPR:$Sd), (ins HPR:$Sm),
1013               NoItinerary, !strconcat("vrint", opc), ".f16\t$Sd, $Sm",
1014               [(set (f16 HPR:$Sd), (node (f16 HPR:$Sm)))]>,
1015               Requires<[HasFullFP16]> {
1016    let Inst{7} = op2;
1017    let Inst{16} = op;
1018  }
1019
1020  def S : ASuI<0b11101, 0b11, 0b0110, 0b11, 0,
1021               (outs SPR:$Sd), (ins SPR:$Sm),
1022               NoItinerary, !strconcat("vrint", opc), ".f32\t$Sd, $Sm",
1023               [(set (f32 SPR:$Sd), (node (f32 SPR:$Sm)))]>,
1024               Requires<[HasFPARMv8]> {
1025    let Inst{7} = op2;
1026    let Inst{16} = op;
1027  }
1028  def D : ADuI<0b11101, 0b11, 0b0110, 0b11, 0,
1029                (outs DPR:$Dd), (ins DPR:$Dm),
1030                NoItinerary, !strconcat("vrint", opc), ".f64\t$Dd, $Dm",
1031                [(set (f64 DPR:$Dd), (node (f64 DPR:$Dm)))]>,
1032                Requires<[HasFPARMv8, HasDPVFP]> {
1033    let Inst{7} = op2;
1034    let Inst{16} = op;
1035  }
1036
1037  def : InstAlias<!strconcat("vrint", opc, "$p.f16.f16\t$Sd, $Sm"),
1038                  (!cast<Instruction>(NAME#"H") SPR:$Sd, SPR:$Sm, pred:$p), 0>,
1039        Requires<[HasFullFP16]>;
1040  def : InstAlias<!strconcat("vrint", opc, "$p.f32.f32\t$Sd, $Sm"),
1041                  (!cast<Instruction>(NAME#"S") SPR:$Sd, SPR:$Sm, pred:$p), 0>,
1042        Requires<[HasFPARMv8]>;
1043  def : InstAlias<!strconcat("vrint", opc, "$p.f64.f64\t$Dd, $Dm"),
1044                  (!cast<Instruction>(NAME#"D") DPR:$Dd, DPR:$Dm, pred:$p), 0>,
1045        Requires<[HasFPARMv8,HasDPVFP]>;
1046}
1047
1048defm VRINTZ : vrint_inst_zrx<"z", 0, 1, ftrunc>;
1049defm VRINTR : vrint_inst_zrx<"r", 0, 0, fnearbyint>;
1050defm VRINTX : vrint_inst_zrx<"x", 1, 0, frint>;
1051
1052multiclass vrint_inst_anpm<string opc, bits<2> rm,
1053                           SDPatternOperator node = null_frag> {
1054  let PostEncoderMethod = "", DecoderNamespace = "VFPV8",
1055      isUnpredicable = 1 in {
1056    def H : AHuInp<0b11101, 0b11, 0b1000, 0b01, 0,
1057                   (outs HPR:$Sd), (ins HPR:$Sm),
1058                   NoItinerary, !strconcat("vrint", opc, ".f16\t$Sd, $Sm"),
1059                   [(set (f16 HPR:$Sd), (node (f16 HPR:$Sm)))]>,
1060                   Requires<[HasFullFP16]> {
1061      let Inst{17-16} = rm;
1062    }
1063    def S : ASuInp<0b11101, 0b11, 0b1000, 0b01, 0,
1064                   (outs SPR:$Sd), (ins SPR:$Sm),
1065                   NoItinerary, !strconcat("vrint", opc, ".f32\t$Sd, $Sm"),
1066                   [(set (f32 SPR:$Sd), (node (f32 SPR:$Sm)))]>,
1067                   Requires<[HasFPARMv8]> {
1068      let Inst{17-16} = rm;
1069    }
1070    def D : ADuInp<0b11101, 0b11, 0b1000, 0b01, 0,
1071                   (outs DPR:$Dd), (ins DPR:$Dm),
1072                   NoItinerary, !strconcat("vrint", opc, ".f64\t$Dd, $Dm"),
1073                   [(set (f64 DPR:$Dd), (node (f64 DPR:$Dm)))]>,
1074                   Requires<[HasFPARMv8, HasDPVFP]> {
1075      let Inst{17-16} = rm;
1076    }
1077  }
1078
1079  def : InstAlias<!strconcat("vrint", opc, ".f32.f32\t$Sd, $Sm"),
1080                  (!cast<Instruction>(NAME#"S") SPR:$Sd, SPR:$Sm), 0>,
1081        Requires<[HasFPARMv8]>;
1082  def : InstAlias<!strconcat("vrint", opc, ".f64.f64\t$Dd, $Dm"),
1083                  (!cast<Instruction>(NAME#"D") DPR:$Dd, DPR:$Dm), 0>,
1084        Requires<[HasFPARMv8,HasDPVFP]>;
1085}
1086
1087defm VRINTA : vrint_inst_anpm<"a", 0b00, fround>;
1088defm VRINTN : vrint_inst_anpm<"n", 0b01, int_arm_neon_vrintn>;
1089defm VRINTP : vrint_inst_anpm<"p", 0b10, fceil>;
1090defm VRINTM : vrint_inst_anpm<"m", 0b11, ffloor>;
1091
1092def VSQRTD : ADuI<0b11101, 0b11, 0b0001, 0b11, 0,
1093                  (outs DPR:$Dd), (ins DPR:$Dm),
1094                  IIC_fpSQRT64, "vsqrt", ".f64\t$Dd, $Dm",
1095                  [(set DPR:$Dd, (fsqrt (f64 DPR:$Dm)))]>,
1096             Sched<[WriteFPSQRT64]>;
1097
1098def VSQRTS : ASuI<0b11101, 0b11, 0b0001, 0b11, 0,
1099                  (outs SPR:$Sd), (ins SPR:$Sm),
1100                  IIC_fpSQRT32, "vsqrt", ".f32\t$Sd, $Sm",
1101                  [(set SPR:$Sd, (fsqrt SPR:$Sm))]>,
1102             Sched<[WriteFPSQRT32]>;
1103
1104def VSQRTH : AHuI<0b11101, 0b11, 0b0001, 0b11, 0,
1105                  (outs HPR:$Sd), (ins HPR:$Sm),
1106                  IIC_fpSQRT16, "vsqrt", ".f16\t$Sd, $Sm",
1107                  [(set (f16 HPR:$Sd), (fsqrt (f16 HPR:$Sm)))]>;
1108
1109let hasSideEffects = 0 in {
1110let isMoveReg = 1 in {
1111def VMOVD  : ADuI<0b11101, 0b11, 0b0000, 0b01, 0,
1112                  (outs DPR:$Dd), (ins DPR:$Dm),
1113                  IIC_fpUNA64, "vmov", ".f64\t$Dd, $Dm", []>,
1114             Requires<[HasFPRegs64]>;
1115
1116def VMOVS  : ASuI<0b11101, 0b11, 0b0000, 0b01, 0,
1117                  (outs SPR:$Sd), (ins SPR:$Sm),
1118                  IIC_fpUNA32, "vmov", ".f32\t$Sd, $Sm", []>,
1119             Requires<[HasFPRegs]>;
1120} // isMoveReg
1121
1122let PostEncoderMethod = "", DecoderNamespace = "VFPV8", isUnpredicable = 1 in {
1123def VMOVH  : ASuInp<0b11101, 0b11, 0b0000, 0b01, 0,
1124                  (outs SPR:$Sd), (ins SPR:$Sm),
1125                  IIC_fpUNA16, "vmovx.f16\t$Sd, $Sm", []>,
1126             Requires<[HasFullFP16]>;
1127
1128def VINSH  : ASuInp<0b11101, 0b11, 0b0000, 0b11, 0,
1129                  (outs SPR:$Sd), (ins SPR:$Sda, SPR:$Sm),
1130                  IIC_fpUNA16, "vins.f16\t$Sd, $Sm", []>,
1131             Requires<[HasFullFP16]> {
1132  let Constraints = "$Sd = $Sda";
1133}
1134
1135} // PostEncoderMethod
1136} // hasSideEffects
1137
1138//===----------------------------------------------------------------------===//
1139// FP <-> GPR Copies.  Int <-> FP Conversions.
1140//
1141
1142let isMoveReg = 1 in {
1143def VMOVRS : AVConv2I<0b11100001, 0b1010,
1144                      (outs GPR:$Rt), (ins SPR:$Sn),
1145                      IIC_fpMOVSI, "vmov", "\t$Rt, $Sn",
1146                      [(set GPR:$Rt, (bitconvert SPR:$Sn))]>,
1147             Requires<[HasFPRegs]>,
1148             Sched<[WriteFPMOV]> {
1149  // Instruction operands.
1150  bits<4> Rt;
1151  bits<5> Sn;
1152
1153  // Encode instruction operands.
1154  let Inst{19-16} = Sn{4-1};
1155  let Inst{7}     = Sn{0};
1156  let Inst{15-12} = Rt;
1157
1158  let Inst{6-5}   = 0b00;
1159  let Inst{3-0}   = 0b0000;
1160
1161  // Some single precision VFP instructions may be executed on both NEON and VFP
1162  // pipelines.
1163  let D = VFPNeonDomain;
1164}
1165
1166// Bitcast i32 -> f32.  NEON prefers to use VMOVDRR.
1167def VMOVSR : AVConv4I<0b11100000, 0b1010,
1168                      (outs SPR:$Sn), (ins GPR:$Rt),
1169                      IIC_fpMOVIS, "vmov", "\t$Sn, $Rt",
1170                      [(set SPR:$Sn, (bitconvert GPR:$Rt))]>,
1171             Requires<[HasFPRegs, UseVMOVSR]>,
1172             Sched<[WriteFPMOV]> {
1173  // Instruction operands.
1174  bits<5> Sn;
1175  bits<4> Rt;
1176
1177  // Encode instruction operands.
1178  let Inst{19-16} = Sn{4-1};
1179  let Inst{7}     = Sn{0};
1180  let Inst{15-12} = Rt;
1181
1182  let Inst{6-5}   = 0b00;
1183  let Inst{3-0}   = 0b0000;
1184
1185  // Some single precision VFP instructions may be executed on both NEON and VFP
1186  // pipelines.
1187  let D = VFPNeonDomain;
1188}
1189} // isMoveReg
1190def : Pat<(arm_vmovsr GPR:$Rt), (VMOVSR GPR:$Rt)>, Requires<[HasVFP2, UseVMOVSR]>;
1191
1192let hasSideEffects = 0 in {
1193def VMOVRRD  : AVConv3I<0b11000101, 0b1011,
1194                        (outs GPR:$Rt, GPR:$Rt2), (ins DPR:$Dm),
1195                        IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $Dm",
1196                 [(set GPR:$Rt, GPR:$Rt2, (arm_fmrrd DPR:$Dm))]>,
1197               Requires<[HasFPRegs]>,
1198               Sched<[WriteFPMOV]> {
1199  // Instruction operands.
1200  bits<5> Dm;
1201  bits<4> Rt;
1202  bits<4> Rt2;
1203
1204  // Encode instruction operands.
1205  let Inst{3-0}   = Dm{3-0};
1206  let Inst{5}     = Dm{4};
1207  let Inst{15-12} = Rt;
1208  let Inst{19-16} = Rt2;
1209
1210  let Inst{7-6} = 0b00;
1211
1212  // Some single precision VFP instructions may be executed on both NEON and VFP
1213  // pipelines.
1214  let D = VFPNeonDomain;
1215
1216  // This instruction is equivalent to
1217  // $Rt = EXTRACT_SUBREG $Dm, ssub_0
1218  // $Rt2 = EXTRACT_SUBREG $Dm, ssub_1
1219  let isExtractSubreg = 1;
1220}
1221
1222def VMOVRRS  : AVConv3I<0b11000101, 0b1010,
1223                      (outs GPR:$Rt, GPR:$Rt2), (ins SPR:$src1, SPR:$src2),
1224                 IIC_fpMOVDI, "vmov", "\t$Rt, $Rt2, $src1, $src2",
1225                 [/* For disassembly only; pattern left blank */]>,
1226               Requires<[HasFPRegs]>,
1227               Sched<[WriteFPMOV]> {
1228  bits<5> src1;
1229  bits<4> Rt;
1230  bits<4> Rt2;
1231
1232  // Encode instruction operands.
1233  let Inst{3-0}   = src1{4-1};
1234  let Inst{5}     = src1{0};
1235  let Inst{15-12} = Rt;
1236  let Inst{19-16} = Rt2;
1237
1238  let Inst{7-6} = 0b00;
1239
1240  // Some single precision VFP instructions may be executed on both NEON and VFP
1241  // pipelines.
1242  let D = VFPNeonDomain;
1243  let DecoderMethod = "DecodeVMOVRRS";
1244}
1245} // hasSideEffects
1246
1247// FMDHR: GPR -> SPR
1248// FMDLR: GPR -> SPR
1249
1250def VMOVDRR : AVConv5I<0b11000100, 0b1011,
1251                      (outs DPR:$Dm), (ins GPR:$Rt, GPR:$Rt2),
1252                      IIC_fpMOVID, "vmov", "\t$Dm, $Rt, $Rt2",
1253                      [(set DPR:$Dm, (arm_fmdrr GPR:$Rt, GPR:$Rt2))]>,
1254              Requires<[HasFPRegs]>,
1255              Sched<[WriteFPMOV]> {
1256  // Instruction operands.
1257  bits<5> Dm;
1258  bits<4> Rt;
1259  bits<4> Rt2;
1260
1261  // Encode instruction operands.
1262  let Inst{3-0}   = Dm{3-0};
1263  let Inst{5}     = Dm{4};
1264  let Inst{15-12} = Rt;
1265  let Inst{19-16} = Rt2;
1266
1267  let Inst{7-6}   = 0b00;
1268
1269  // Some single precision VFP instructions may be executed on both NEON and VFP
1270  // pipelines.
1271  let D = VFPNeonDomain;
1272
1273  // This instruction is equivalent to
1274  // $Dm = REG_SEQUENCE $Rt, ssub_0, $Rt2, ssub_1
1275  let isRegSequence = 1;
1276}
1277
1278// Hoist an fabs or a fneg of a value coming from integer registers
1279// and do the fabs/fneg on the integer value. This is never a lose
1280// and could enable the conversion to float to be removed completely.
1281def : Pat<(fabs (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1282          (VMOVDRR GPR:$Rl, (BFC GPR:$Rh, (i32 0x7FFFFFFF)))>,
1283      Requires<[IsARM, HasV6T2]>;
1284def : Pat<(fabs (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1285          (VMOVDRR GPR:$Rl, (t2BFC GPR:$Rh, (i32 0x7FFFFFFF)))>,
1286      Requires<[IsThumb2, HasV6T2]>;
1287def : Pat<(fneg (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1288          (VMOVDRR GPR:$Rl, (EORri GPR:$Rh, (i32 0x80000000)))>,
1289      Requires<[IsARM]>;
1290def : Pat<(fneg (arm_fmdrr GPR:$Rl, GPR:$Rh)),
1291          (VMOVDRR GPR:$Rl, (t2EORri GPR:$Rh, (i32 0x80000000)))>,
1292      Requires<[IsThumb2]>;
1293
1294let hasSideEffects = 0 in
1295def VMOVSRR : AVConv5I<0b11000100, 0b1010,
1296                     (outs SPR:$dst1, SPR:$dst2), (ins GPR:$src1, GPR:$src2),
1297                IIC_fpMOVID, "vmov", "\t$dst1, $dst2, $src1, $src2",
1298                [/* For disassembly only; pattern left blank */]>,
1299              Requires<[HasFPRegs]>,
1300              Sched<[WriteFPMOV]> {
1301  // Instruction operands.
1302  bits<5> dst1;
1303  bits<4> src1;
1304  bits<4> src2;
1305
1306  // Encode instruction operands.
1307  let Inst{3-0}   = dst1{4-1};
1308  let Inst{5}     = dst1{0};
1309  let Inst{15-12} = src1;
1310  let Inst{19-16} = src2;
1311
1312  let Inst{7-6} = 0b00;
1313
1314  // Some single precision VFP instructions may be executed on both NEON and VFP
1315  // pipelines.
1316  let D = VFPNeonDomain;
1317
1318  let DecoderMethod = "DecodeVMOVSRR";
1319}
1320
1321// Move H->R, clearing top 16 bits
1322def VMOVRH : AVConv2I<0b11100001, 0b1001,
1323                      (outs rGPR:$Rt), (ins HPR:$Sn),
1324                      IIC_fpMOVSI, "vmov", ".f16\t$Rt, $Sn",
1325                      []>,
1326             Requires<[HasFPRegs16]>,
1327             Sched<[WriteFPMOV]> {
1328  // Instruction operands.
1329  bits<4> Rt;
1330  bits<5> Sn;
1331
1332  // Encode instruction operands.
1333  let Inst{19-16} = Sn{4-1};
1334  let Inst{7}     = Sn{0};
1335  let Inst{15-12} = Rt;
1336
1337  let Inst{6-5}   = 0b00;
1338  let Inst{3-0}   = 0b0000;
1339
1340  let isUnpredicable = 1;
1341}
1342
1343// Move R->H, clearing top 16 bits
1344def VMOVHR : AVConv4I<0b11100000, 0b1001,
1345                      (outs HPR:$Sn), (ins rGPR:$Rt),
1346                      IIC_fpMOVIS, "vmov", ".f16\t$Sn, $Rt",
1347                      []>,
1348             Requires<[HasFPRegs16]>,
1349             Sched<[WriteFPMOV]> {
1350  // Instruction operands.
1351  bits<5> Sn;
1352  bits<4> Rt;
1353
1354  // Encode instruction operands.
1355  let Inst{19-16} = Sn{4-1};
1356  let Inst{7}     = Sn{0};
1357  let Inst{15-12} = Rt;
1358
1359  let Inst{6-5}   = 0b00;
1360  let Inst{3-0}   = 0b0000;
1361
1362  let isUnpredicable = 1;
1363}
1364
1365def : FPRegs16Pat<(arm_vmovrh (f16 HPR:$Sn)), (VMOVRH (f16 HPR:$Sn))>;
1366def : FPRegs16Pat<(arm_vmovrh (bf16 HPR:$Sn)), (VMOVRH (bf16 HPR:$Sn))>;
1367def : FPRegs16Pat<(f16 (arm_vmovhr rGPR:$Rt)), (VMOVHR rGPR:$Rt)>;
1368def : FPRegs16Pat<(bf16 (arm_vmovhr rGPR:$Rt)), (VMOVHR rGPR:$Rt)>;
1369
1370// FMRDH: SPR -> GPR
1371// FMRDL: SPR -> GPR
1372// FMRRS: SPR -> GPR
1373// FMRX:  SPR system reg -> GPR
1374// FMSRR: GPR -> SPR
1375// FMXR:  GPR -> VFP system reg
1376
1377
1378// Int -> FP:
1379
1380class AVConv1IDs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1381                        bits<4> opcod4, dag oops, dag iops,
1382                        InstrItinClass itin, string opc, string asm,
1383                        list<dag> pattern>
1384  : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1385             pattern> {
1386  // Instruction operands.
1387  bits<5> Dd;
1388  bits<5> Sm;
1389
1390  // Encode instruction operands.
1391  let Inst{3-0}   = Sm{4-1};
1392  let Inst{5}     = Sm{0};
1393  let Inst{15-12} = Dd{3-0};
1394  let Inst{22}    = Dd{4};
1395
1396  let Predicates = [HasVFP2, HasDPVFP];
1397  let hasSideEffects = 0;
1398}
1399
1400class AVConv1InSs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1401                         bits<4> opcod4, dag oops, dag iops,InstrItinClass itin,
1402                         string opc, string asm, list<dag> pattern>
1403  : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1404              pattern> {
1405  // Instruction operands.
1406  bits<5> Sd;
1407  bits<5> Sm;
1408
1409  // Encode instruction operands.
1410  let Inst{3-0}   = Sm{4-1};
1411  let Inst{5}     = Sm{0};
1412  let Inst{15-12} = Sd{4-1};
1413  let Inst{22}    = Sd{0};
1414
1415  let hasSideEffects = 0;
1416}
1417
1418class AVConv1IHs_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1419                        bits<4> opcod4, dag oops, dag iops,
1420                        InstrItinClass itin, string opc, string asm,
1421                        list<dag> pattern>
1422  : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1423             pattern> {
1424  // Instruction operands.
1425  bits<5> Sd;
1426  bits<5> Sm;
1427
1428  // Encode instruction operands.
1429  let Inst{3-0}   = Sm{4-1};
1430  let Inst{5}     = Sm{0};
1431  let Inst{15-12} = Sd{4-1};
1432  let Inst{22}    = Sd{0};
1433
1434  let Predicates = [HasFullFP16];
1435  let hasSideEffects = 0;
1436}
1437
1438def VSITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
1439                               (outs DPR:$Dd), (ins SPR:$Sm),
1440                               IIC_fpCVTID, "vcvt", ".f64.s32\t$Dd, $Sm",
1441                               []>,
1442             Sched<[WriteFPCVT]> {
1443  let Inst{7} = 1; // s32
1444}
1445
1446let Predicates=[HasVFP2, HasDPVFP] in {
1447  def : VFPPat<(f64 (sint_to_fp GPR:$a)),
1448               (VSITOD (COPY_TO_REGCLASS GPR:$a, SPR))>;
1449
1450  def : VFPPat<(f64 (sint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1451               (VSITOD (VLDRS addrmode5:$a))>;
1452}
1453
1454def VSITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
1455                                (outs SPR:$Sd),(ins SPR:$Sm),
1456                                IIC_fpCVTIS, "vcvt", ".f32.s32\t$Sd, $Sm",
1457                                []>,
1458             Sched<[WriteFPCVT]> {
1459  let Inst{7} = 1; // s32
1460
1461  // Some single precision VFP instructions may be executed on both NEON and
1462  // VFP pipelines on A8.
1463  let D = VFPNeonA8Domain;
1464}
1465
1466def : VFPNoNEONPat<(f32 (sint_to_fp GPR:$a)),
1467                   (VSITOS (COPY_TO_REGCLASS GPR:$a, SPR))>;
1468
1469def : VFPNoNEONPat<(f32 (sint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1470                   (VSITOS (VLDRS addrmode5:$a))>;
1471
1472def VSITOH : AVConv1IHs_Encode<0b11101, 0b11, 0b1000, 0b1001,
1473                               (outs HPR:$Sd), (ins SPR:$Sm),
1474                               IIC_fpCVTIH, "vcvt", ".f16.s32\t$Sd, $Sm",
1475                               []>,
1476             Sched<[WriteFPCVT]> {
1477  let Inst{7} = 1; // s32
1478  let isUnpredicable = 1;
1479}
1480
1481def : VFPNoNEONPat<(f16 (sint_to_fp GPR:$a)),
1482                   (VSITOH (COPY_TO_REGCLASS GPR:$a, SPR))>;
1483
1484def VUITOD : AVConv1IDs_Encode<0b11101, 0b11, 0b1000, 0b1011,
1485                               (outs DPR:$Dd), (ins SPR:$Sm),
1486                               IIC_fpCVTID, "vcvt", ".f64.u32\t$Dd, $Sm",
1487                               []>,
1488             Sched<[WriteFPCVT]> {
1489  let Inst{7} = 0; // u32
1490}
1491
1492let Predicates=[HasVFP2, HasDPVFP] in {
1493  def : VFPPat<(f64 (uint_to_fp GPR:$a)),
1494               (VUITOD (COPY_TO_REGCLASS GPR:$a, SPR))>;
1495
1496  def : VFPPat<(f64 (uint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1497               (VUITOD (VLDRS addrmode5:$a))>;
1498}
1499
1500def VUITOS : AVConv1InSs_Encode<0b11101, 0b11, 0b1000, 0b1010,
1501                                (outs SPR:$Sd), (ins SPR:$Sm),
1502                                IIC_fpCVTIS, "vcvt", ".f32.u32\t$Sd, $Sm",
1503                                []>,
1504             Sched<[WriteFPCVT]> {
1505  let Inst{7} = 0; // u32
1506
1507  // Some single precision VFP instructions may be executed on both NEON and
1508  // VFP pipelines on A8.
1509  let D = VFPNeonA8Domain;
1510}
1511
1512def : VFPNoNEONPat<(f32 (uint_to_fp GPR:$a)),
1513                   (VUITOS (COPY_TO_REGCLASS GPR:$a, SPR))>;
1514
1515def : VFPNoNEONPat<(f32 (uint_to_fp (i32 (alignedload32 addrmode5:$a)))),
1516                   (VUITOS (VLDRS addrmode5:$a))>;
1517
1518def VUITOH : AVConv1IHs_Encode<0b11101, 0b11, 0b1000, 0b1001,
1519                                (outs HPR:$Sd), (ins SPR:$Sm),
1520                                IIC_fpCVTIH, "vcvt", ".f16.u32\t$Sd, $Sm",
1521                                []>,
1522             Sched<[WriteFPCVT]> {
1523  let Inst{7} = 0; // u32
1524  let isUnpredicable = 1;
1525}
1526
1527def : VFPNoNEONPat<(f16 (uint_to_fp GPR:$a)),
1528                   (VUITOH (COPY_TO_REGCLASS GPR:$a, SPR))>;
1529
1530// FP -> Int:
1531
1532class AVConv1IsD_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1533                        bits<4> opcod4, dag oops, dag iops,
1534                        InstrItinClass itin, string opc, string asm,
1535                        list<dag> pattern>
1536  : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1537             pattern> {
1538  // Instruction operands.
1539  bits<5> Sd;
1540  bits<5> Dm;
1541
1542  // Encode instruction operands.
1543  let Inst{3-0}   = Dm{3-0};
1544  let Inst{5}     = Dm{4};
1545  let Inst{15-12} = Sd{4-1};
1546  let Inst{22}    = Sd{0};
1547
1548  let Predicates = [HasVFP2, HasDPVFP];
1549  let hasSideEffects = 0;
1550}
1551
1552class AVConv1InsS_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1553                         bits<4> opcod4, dag oops, dag iops,
1554                         InstrItinClass itin, string opc, string asm,
1555                         list<dag> pattern>
1556  : AVConv1In<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1557              pattern> {
1558  // Instruction operands.
1559  bits<5> Sd;
1560  bits<5> Sm;
1561
1562  // Encode instruction operands.
1563  let Inst{3-0}   = Sm{4-1};
1564  let Inst{5}     = Sm{0};
1565  let Inst{15-12} = Sd{4-1};
1566  let Inst{22}    = Sd{0};
1567
1568  let hasSideEffects = 0;
1569}
1570
1571class AVConv1IsH_Encode<bits<5> opcod1, bits<2> opcod2, bits<4> opcod3,
1572                         bits<4> opcod4, dag oops, dag iops,
1573                         InstrItinClass itin, string opc, string asm,
1574                         list<dag> pattern>
1575  : AVConv1I<opcod1, opcod2, opcod3, opcod4, oops, iops, itin, opc, asm,
1576              pattern> {
1577  // Instruction operands.
1578  bits<5> Sd;
1579  bits<5> Sm;
1580
1581  // Encode instruction operands.
1582  let Inst{3-0}   = Sm{4-1};
1583  let Inst{5}     = Sm{0};
1584  let Inst{15-12} = Sd{4-1};
1585  let Inst{22}    = Sd{0};
1586
1587  let Predicates = [HasFullFP16];
1588  let hasSideEffects = 0;
1589}
1590
1591// Always set Z bit in the instruction, i.e. "round towards zero" variants.
1592def VTOSIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
1593                                (outs SPR:$Sd), (ins DPR:$Dm),
1594                                IIC_fpCVTDI, "vcvt", ".s32.f64\t$Sd, $Dm",
1595                                []>,
1596              Sched<[WriteFPCVT]> {
1597  let Inst{7} = 1; // Z bit
1598}
1599
1600let Predicates=[HasVFP2, HasDPVFP] in {
1601  def : VFPPat<(i32 (fp_to_sint (f64 DPR:$a))),
1602               (COPY_TO_REGCLASS (VTOSIZD DPR:$a), GPR)>;
1603  def : VFPPat<(i32 (fp_to_sint_sat (f64 DPR:$a), i32)),
1604               (COPY_TO_REGCLASS (VTOSIZD DPR:$a), GPR)>;
1605
1606  def : VFPPat<(alignedstore32 (i32 (fp_to_sint (f64 DPR:$a))), addrmode5:$ptr),
1607               (VSTRS (VTOSIZD DPR:$a), addrmode5:$ptr)>;
1608  def : VFPPat<(alignedstore32 (i32 (fp_to_sint_sat (f64 DPR:$a), i32)), addrmode5:$ptr),
1609               (VSTRS (VTOSIZD DPR:$a), addrmode5:$ptr)>;
1610}
1611
1612def VTOSIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
1613                                 (outs SPR:$Sd), (ins SPR:$Sm),
1614                                 IIC_fpCVTSI, "vcvt", ".s32.f32\t$Sd, $Sm",
1615                                 []>,
1616              Sched<[WriteFPCVT]> {
1617  let Inst{7} = 1; // Z bit
1618
1619  // Some single precision VFP instructions may be executed on both NEON and
1620  // VFP pipelines on A8.
1621  let D = VFPNeonA8Domain;
1622}
1623
1624def : VFPNoNEONPat<(i32 (fp_to_sint SPR:$a)),
1625                   (COPY_TO_REGCLASS (VTOSIZS SPR:$a), GPR)>;
1626def : VFPPat<(i32 (fp_to_sint_sat SPR:$a, i32)),
1627             (COPY_TO_REGCLASS (VTOSIZS SPR:$a), GPR)>;
1628
1629def : VFPNoNEONPat<(alignedstore32 (i32 (fp_to_sint (f32 SPR:$a))),
1630                                   addrmode5:$ptr),
1631                   (VSTRS (VTOSIZS SPR:$a), addrmode5:$ptr)>;
1632def : VFPPat<(alignedstore32 (i32 (fp_to_sint_sat (f32 SPR:$a), i32)),
1633                                   addrmode5:$ptr),
1634             (VSTRS (VTOSIZS SPR:$a), addrmode5:$ptr)>;
1635
1636def VTOSIZH : AVConv1IsH_Encode<0b11101, 0b11, 0b1101, 0b1001,
1637                                 (outs SPR:$Sd), (ins HPR:$Sm),
1638                                 IIC_fpCVTHI, "vcvt", ".s32.f16\t$Sd, $Sm",
1639                                 []>,
1640              Sched<[WriteFPCVT]> {
1641  let Inst{7} = 1; // Z bit
1642  let isUnpredicable = 1;
1643}
1644
1645def : VFPNoNEONPat<(i32 (fp_to_sint (f16 HPR:$a))),
1646                   (COPY_TO_REGCLASS (VTOSIZH (f16 HPR:$a)), GPR)>;
1647def : VFPPat<(i32 (fp_to_sint_sat (f16 HPR:$a), i32)),
1648             (COPY_TO_REGCLASS (VTOSIZH (f16 HPR:$a)), GPR)>;
1649
1650def VTOUIZD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
1651                               (outs SPR:$Sd), (ins DPR:$Dm),
1652                               IIC_fpCVTDI, "vcvt", ".u32.f64\t$Sd, $Dm",
1653                               []>,
1654              Sched<[WriteFPCVT]> {
1655  let Inst{7} = 1; // Z bit
1656}
1657
1658let Predicates=[HasVFP2, HasDPVFP] in {
1659  def : VFPPat<(i32 (fp_to_uint (f64 DPR:$a))),
1660               (COPY_TO_REGCLASS (VTOUIZD DPR:$a), GPR)>;
1661  def : VFPPat<(i32 (fp_to_uint_sat (f64 DPR:$a), i32)),
1662               (COPY_TO_REGCLASS (VTOUIZD DPR:$a), GPR)>;
1663
1664  def : VFPPat<(alignedstore32 (i32 (fp_to_uint (f64 DPR:$a))), addrmode5:$ptr),
1665               (VSTRS (VTOUIZD DPR:$a), addrmode5:$ptr)>;
1666  def : VFPPat<(alignedstore32 (i32 (fp_to_uint_sat (f64 DPR:$a), i32)), addrmode5:$ptr),
1667               (VSTRS (VTOUIZD DPR:$a), addrmode5:$ptr)>;
1668}
1669
1670def VTOUIZS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
1671                                 (outs SPR:$Sd), (ins SPR:$Sm),
1672                                 IIC_fpCVTSI, "vcvt", ".u32.f32\t$Sd, $Sm",
1673                                 []>,
1674              Sched<[WriteFPCVT]> {
1675  let Inst{7} = 1; // Z bit
1676
1677  // Some single precision VFP instructions may be executed on both NEON and
1678  // VFP pipelines on A8.
1679  let D = VFPNeonA8Domain;
1680}
1681
1682def : VFPNoNEONPat<(i32 (fp_to_uint SPR:$a)),
1683                   (COPY_TO_REGCLASS (VTOUIZS SPR:$a), GPR)>;
1684def : VFPPat<(i32 (fp_to_uint_sat SPR:$a, i32)),
1685             (COPY_TO_REGCLASS (VTOUIZS SPR:$a), GPR)>;
1686
1687def : VFPNoNEONPat<(alignedstore32 (i32 (fp_to_uint (f32 SPR:$a))),
1688                                   addrmode5:$ptr),
1689                  (VSTRS (VTOUIZS SPR:$a), addrmode5:$ptr)>;
1690def : VFPPat<(alignedstore32 (i32 (fp_to_uint_sat (f32 SPR:$a), i32)),
1691                                   addrmode5:$ptr),
1692             (VSTRS (VTOUIZS SPR:$a), addrmode5:$ptr)>;
1693
1694def VTOUIZH : AVConv1IsH_Encode<0b11101, 0b11, 0b1100, 0b1001,
1695                                 (outs SPR:$Sd), (ins HPR:$Sm),
1696                                 IIC_fpCVTHI, "vcvt", ".u32.f16\t$Sd, $Sm",
1697                                 []>,
1698              Sched<[WriteFPCVT]> {
1699  let Inst{7} = 1; // Z bit
1700  let isUnpredicable = 1;
1701}
1702
1703def : VFPNoNEONPat<(i32 (fp_to_uint (f16 HPR:$a))),
1704                   (COPY_TO_REGCLASS (VTOUIZH (f16 HPR:$a)), GPR)>;
1705def : VFPPat<(i32 (fp_to_uint_sat (f16 HPR:$a), i32)),
1706             (COPY_TO_REGCLASS (VTOUIZH (f16 HPR:$a)), GPR)>;
1707
1708// And the Z bit '0' variants, i.e. use the rounding mode specified by FPSCR.
1709let Uses = [FPSCR] in {
1710def VTOSIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1101, 0b1011,
1711                                (outs SPR:$Sd), (ins DPR:$Dm),
1712                                IIC_fpCVTDI, "vcvtr", ".s32.f64\t$Sd, $Dm",
1713                                [(set SPR:$Sd, (int_arm_vcvtr (f64 DPR:$Dm)))]>,
1714              Sched<[WriteFPCVT]> {
1715  let Inst{7} = 0; // Z bit
1716}
1717
1718def VTOSIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1101, 0b1010,
1719                                 (outs SPR:$Sd), (ins SPR:$Sm),
1720                                 IIC_fpCVTSI, "vcvtr", ".s32.f32\t$Sd, $Sm",
1721                                 [(set SPR:$Sd, (int_arm_vcvtr SPR:$Sm))]>,
1722              Sched<[WriteFPCVT]> {
1723  let Inst{7} = 0; // Z bit
1724}
1725
1726def VTOSIRH : AVConv1IsH_Encode<0b11101, 0b11, 0b1101, 0b1001,
1727                                 (outs SPR:$Sd), (ins SPR:$Sm),
1728                                 IIC_fpCVTHI, "vcvtr", ".s32.f16\t$Sd, $Sm",
1729                                 []>,
1730              Sched<[WriteFPCVT]> {
1731  let Inst{7} = 0; // Z bit
1732  let isUnpredicable = 1;
1733}
1734
1735def VTOUIRD : AVConv1IsD_Encode<0b11101, 0b11, 0b1100, 0b1011,
1736                                (outs SPR:$Sd), (ins DPR:$Dm),
1737                                IIC_fpCVTDI, "vcvtr", ".u32.f64\t$Sd, $Dm",
1738                                [(set SPR:$Sd, (int_arm_vcvtru(f64 DPR:$Dm)))]>,
1739              Sched<[WriteFPCVT]> {
1740  let Inst{7} = 0; // Z bit
1741}
1742
1743def VTOUIRS : AVConv1InsS_Encode<0b11101, 0b11, 0b1100, 0b1010,
1744                                 (outs SPR:$Sd), (ins SPR:$Sm),
1745                                 IIC_fpCVTSI, "vcvtr", ".u32.f32\t$Sd, $Sm",
1746                                 [(set SPR:$Sd, (int_arm_vcvtru SPR:$Sm))]>,
1747              Sched<[WriteFPCVT]> {
1748  let Inst{7} = 0; // Z bit
1749}
1750
1751def VTOUIRH : AVConv1IsH_Encode<0b11101, 0b11, 0b1100, 0b1001,
1752                                 (outs SPR:$Sd), (ins SPR:$Sm),
1753                                 IIC_fpCVTHI, "vcvtr", ".u32.f16\t$Sd, $Sm",
1754                                 []>,
1755              Sched<[WriteFPCVT]> {
1756  let Inst{7} = 0; // Z bit
1757  let isUnpredicable = 1;
1758}
1759}
1760
1761// v8.3-a Javascript Convert to Signed fixed-point
1762def VJCVT : AVConv1IsD_Encode<0b11101, 0b11, 0b1001, 0b1011,
1763                                (outs SPR:$Sd), (ins DPR:$Dm),
1764                                IIC_fpCVTDI, "vjcvt", ".s32.f64\t$Sd, $Dm",
1765                                []>,
1766            Requires<[HasFPARMv8, HasV8_3a]> {
1767  let Inst{7} = 1; // Z bit
1768}
1769
1770// Convert between floating-point and fixed-point
1771// Data type for fixed-point naming convention:
1772//   S16 (U=0, sx=0) -> SH
1773//   U16 (U=1, sx=0) -> UH
1774//   S32 (U=0, sx=1) -> SL
1775//   U32 (U=1, sx=1) -> UL
1776
1777let Constraints = "$a = $dst" in {
1778
1779// FP to Fixed-Point:
1780
1781// Single Precision register
1782class AVConv1XInsS_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
1783                          bit op5, dag oops, dag iops, InstrItinClass itin,
1784                          string opc, string asm, list<dag> pattern>
1785  : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern> {
1786  bits<5> dst;
1787  // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
1788  let Inst{22} = dst{0};
1789  let Inst{15-12} = dst{4-1};
1790
1791  let hasSideEffects = 0;
1792}
1793
1794// Double Precision register
1795class AVConv1XInsD_Encode<bits<5> op1, bits<2> op2, bits<4> op3, bits<4> op4,
1796                          bit op5, dag oops, dag iops, InstrItinClass itin,
1797                          string opc, string asm, list<dag> pattern>
1798  : AVConv1XI<op1, op2, op3, op4, op5, oops, iops, itin, opc, asm, pattern> {
1799  bits<5> dst;
1800  // if dp_operation then UInt(D:Vd) else UInt(Vd:D);
1801  let Inst{22} = dst{4};
1802  let Inst{15-12} = dst{3-0};
1803
1804  let hasSideEffects = 0;
1805  let Predicates = [HasVFP2, HasDPVFP];
1806}
1807
1808let isUnpredicable = 1 in {
1809
1810def VTOSHH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1001, 0,
1811                       (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1812                 IIC_fpCVTHI, "vcvt", ".s16.f16\t$dst, $a, $fbits", []>,
1813             Requires<[HasFullFP16]>,
1814             Sched<[WriteFPCVT]>;
1815
1816def VTOUHH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1001, 0,
1817                       (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1818                 IIC_fpCVTHI, "vcvt", ".u16.f16\t$dst, $a, $fbits", []>,
1819             Requires<[HasFullFP16]>,
1820             Sched<[WriteFPCVT]>;
1821
1822def VTOSLH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1001, 1,
1823                       (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1824                 IIC_fpCVTHI, "vcvt", ".s32.f16\t$dst, $a, $fbits", []>,
1825             Requires<[HasFullFP16]>,
1826             Sched<[WriteFPCVT]>;
1827
1828def VTOULH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1001, 1,
1829                       (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1830                 IIC_fpCVTHI, "vcvt", ".u32.f16\t$dst, $a, $fbits", []>,
1831             Requires<[HasFullFP16]>,
1832             Sched<[WriteFPCVT]>;
1833
1834} // End of 'let isUnpredicable = 1 in'
1835
1836def VTOSHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 0,
1837                       (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1838                 IIC_fpCVTSI, "vcvt", ".s16.f32\t$dst, $a, $fbits", []>,
1839             Sched<[WriteFPCVT]> {
1840  // Some single precision VFP instructions may be executed on both NEON and
1841  // VFP pipelines on A8.
1842  let D = VFPNeonA8Domain;
1843}
1844
1845def VTOUHS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 0,
1846                       (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1847                 IIC_fpCVTSI, "vcvt", ".u16.f32\t$dst, $a, $fbits", []>,
1848             Sched<[WriteFPCVT]> {
1849  // Some single precision VFP instructions may be executed on both NEON and
1850  // VFP pipelines on A8.
1851  let D = VFPNeonA8Domain;
1852}
1853
1854def VTOSLS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1110, 0b1010, 1,
1855                       (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1856                 IIC_fpCVTSI, "vcvt", ".s32.f32\t$dst, $a, $fbits", []>,
1857             Sched<[WriteFPCVT]> {
1858  // Some single precision VFP instructions may be executed on both NEON and
1859  // VFP pipelines on A8.
1860  let D = VFPNeonA8Domain;
1861}
1862
1863def VTOULS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1111, 0b1010, 1,
1864                       (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1865                 IIC_fpCVTSI, "vcvt", ".u32.f32\t$dst, $a, $fbits", []>,
1866             Sched<[WriteFPCVT]> {
1867  // Some single precision VFP instructions may be executed on both NEON and
1868  // VFP pipelines on A8.
1869  let D = VFPNeonA8Domain;
1870}
1871
1872def VTOSHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 0,
1873                       (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1874                 IIC_fpCVTDI, "vcvt", ".s16.f64\t$dst, $a, $fbits", []>,
1875             Sched<[WriteFPCVT]>;
1876
1877def VTOUHD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 0,
1878                       (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1879                 IIC_fpCVTDI, "vcvt", ".u16.f64\t$dst, $a, $fbits", []>,
1880             Sched<[WriteFPCVT]>;
1881
1882def VTOSLD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1110, 0b1011, 1,
1883                       (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1884                 IIC_fpCVTDI, "vcvt", ".s32.f64\t$dst, $a, $fbits", []>,
1885             Sched<[WriteFPCVT]>;
1886
1887def VTOULD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1111, 0b1011, 1,
1888                       (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1889                 IIC_fpCVTDI, "vcvt", ".u32.f64\t$dst, $a, $fbits", []>,
1890             Sched<[WriteFPCVT]>;
1891
1892// Fixed-Point to FP:
1893
1894let isUnpredicable = 1 in {
1895
1896def VSHTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1001, 0,
1897                       (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1898                 IIC_fpCVTIH, "vcvt", ".f16.s16\t$dst, $a, $fbits", []>,
1899             Requires<[HasFullFP16]>,
1900             Sched<[WriteFPCVT]>;
1901
1902def VUHTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1001, 0,
1903                       (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1904                 IIC_fpCVTIH, "vcvt", ".f16.u16\t$dst, $a, $fbits", []>,
1905             Requires<[HasFullFP16]>,
1906             Sched<[WriteFPCVT]>;
1907
1908def VSLTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1001, 1,
1909                       (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1910                 IIC_fpCVTIH, "vcvt", ".f16.s32\t$dst, $a, $fbits", []>,
1911             Requires<[HasFullFP16]>,
1912             Sched<[WriteFPCVT]>;
1913
1914def VULTOH : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1001, 1,
1915                       (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1916                 IIC_fpCVTIH, "vcvt", ".f16.u32\t$dst, $a, $fbits", []>,
1917             Requires<[HasFullFP16]>,
1918             Sched<[WriteFPCVT]>;
1919
1920} // End of 'let isUnpredicable = 1 in'
1921
1922def VSHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 0,
1923                       (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1924                 IIC_fpCVTIS, "vcvt", ".f32.s16\t$dst, $a, $fbits", []>,
1925             Sched<[WriteFPCVT]> {
1926  // Some single precision VFP instructions may be executed on both NEON and
1927  // VFP pipelines on A8.
1928  let D = VFPNeonA8Domain;
1929}
1930
1931def VUHTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 0,
1932                       (outs SPR:$dst), (ins SPR:$a, fbits16:$fbits),
1933                 IIC_fpCVTIS, "vcvt", ".f32.u16\t$dst, $a, $fbits", []>,
1934             Sched<[WriteFPCVT]> {
1935  // Some single precision VFP instructions may be executed on both NEON and
1936  // VFP pipelines on A8.
1937  let D = VFPNeonA8Domain;
1938}
1939
1940def VSLTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1010, 0b1010, 1,
1941                       (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1942                 IIC_fpCVTIS, "vcvt", ".f32.s32\t$dst, $a, $fbits", []>,
1943             Sched<[WriteFPCVT]> {
1944  // Some single precision VFP instructions may be executed on both NEON and
1945  // VFP pipelines on A8.
1946  let D = VFPNeonA8Domain;
1947}
1948
1949def VULTOS : AVConv1XInsS_Encode<0b11101, 0b11, 0b1011, 0b1010, 1,
1950                       (outs SPR:$dst), (ins SPR:$a, fbits32:$fbits),
1951                 IIC_fpCVTIS, "vcvt", ".f32.u32\t$dst, $a, $fbits", []>,
1952             Sched<[WriteFPCVT]> {
1953  // Some single precision VFP instructions may be executed on both NEON and
1954  // VFP pipelines on A8.
1955  let D = VFPNeonA8Domain;
1956}
1957
1958def VSHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 0,
1959                       (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1960                 IIC_fpCVTID, "vcvt", ".f64.s16\t$dst, $a, $fbits", []>,
1961             Sched<[WriteFPCVT]>;
1962
1963def VUHTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 0,
1964                       (outs DPR:$dst), (ins DPR:$a, fbits16:$fbits),
1965                 IIC_fpCVTID, "vcvt", ".f64.u16\t$dst, $a, $fbits", []>,
1966             Sched<[WriteFPCVT]>;
1967
1968def VSLTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1010, 0b1011, 1,
1969                       (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1970                 IIC_fpCVTID, "vcvt", ".f64.s32\t$dst, $a, $fbits", []>,
1971             Sched<[WriteFPCVT]>;
1972
1973def VULTOD : AVConv1XInsD_Encode<0b11101, 0b11, 0b1011, 0b1011, 1,
1974                       (outs DPR:$dst), (ins DPR:$a, fbits32:$fbits),
1975                 IIC_fpCVTID, "vcvt", ".f64.u32\t$dst, $a, $fbits", []>,
1976             Sched<[WriteFPCVT]>;
1977
1978} // End of 'let Constraints = "$a = $dst" in'
1979
1980// BFloat16  - Single precision, unary, predicated
1981class BF16_VCVT<string opc, bits<2> op7_6>
1982   : VFPAI<(outs SPR:$Sd), (ins SPR:$dst, SPR:$Sm),
1983           VFPUnaryFrm, NoItinerary,
1984           opc, ".bf16.f32\t$Sd, $Sm", []>,
1985      RegConstraint<"$dst = $Sd">,
1986      Requires<[HasBF16]>,
1987     Sched<[]> {
1988  bits<5> Sd;
1989  bits<5> Sm;
1990
1991  // Encode instruction operands.
1992  let Inst{3-0}   = Sm{4-1};
1993  let Inst{5}     = Sm{0};
1994  let Inst{15-12} = Sd{4-1};
1995  let Inst{22}    = Sd{0};
1996
1997  let Inst{27-23} = 0b11101; // opcode1
1998  let Inst{21-20} = 0b11;    // opcode2
1999  let Inst{19-16} = 0b0011;  // opcode3
2000  let Inst{11-8}  = 0b1001;
2001  let Inst{7-6}   = op7_6;
2002  let Inst{4}     = 0;
2003
2004  let DecoderNamespace = "VFPV8";
2005  let hasSideEffects = 0;
2006}
2007
2008def BF16_VCVTB : BF16_VCVT<"vcvtb", 0b01>;
2009def BF16_VCVTT : BF16_VCVT<"vcvtt", 0b11>;
2010
2011//===----------------------------------------------------------------------===//
2012// FP Multiply-Accumulate Operations.
2013//
2014
2015def VMLAD : ADbI<0b11100, 0b00, 0, 0,
2016                 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2017                 IIC_fpMAC64, "vmla", ".f64\t$Dd, $Dn, $Dm",
2018                 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
2019                                          (f64 DPR:$Ddin)))]>,
2020              RegConstraint<"$Ddin = $Dd">,
2021              Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>,
2022              Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2023
2024def VMLAS : ASbIn<0b11100, 0b00, 0, 0,
2025                  (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2026                  IIC_fpMAC32, "vmla", ".f32\t$Sd, $Sn, $Sm",
2027                  [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
2028                                           SPR:$Sdin))]>,
2029              RegConstraint<"$Sdin = $Sd">,
2030              Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>,
2031              Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2032  // Some single precision VFP instructions may be executed on both NEON and
2033  // VFP pipelines on A8.
2034  let D = VFPNeonA8Domain;
2035}
2036
2037def VMLAH : AHbI<0b11100, 0b00, 0, 0,
2038                  (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2039                  IIC_fpMAC16, "vmla", ".f16\t$Sd, $Sn, $Sm",
2040                  [(set (f16 HPR:$Sd), (fadd_mlx (fmul_su (f16 HPR:$Sn), (f16 HPR:$Sm)),
2041                                           (f16 HPR:$Sdin)))]>,
2042              RegConstraint<"$Sdin = $Sd">,
2043              Requires<[HasFullFP16,UseFPVMLx]>;
2044
2045def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
2046          (VMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
2047          Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>;
2048def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
2049          (VMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
2050          Requires<[HasVFP2,DontUseNEONForFP, UseFPVMLx]>;
2051def : Pat<(fadd_mlx HPR:$dstin, (fmul_su (f16 HPR:$a), HPR:$b)),
2052          (VMLAH HPR:$dstin, (f16 HPR:$a), HPR:$b)>,
2053          Requires<[HasFullFP16,DontUseNEONForFP, UseFPVMLx]>;
2054
2055
2056def VMLSD : ADbI<0b11100, 0b00, 1, 0,
2057                 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2058                 IIC_fpMAC64, "vmls", ".f64\t$Dd, $Dn, $Dm",
2059                 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
2060                                          (f64 DPR:$Ddin)))]>,
2061              RegConstraint<"$Ddin = $Dd">,
2062              Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>,
2063              Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2064
2065def VMLSS : ASbIn<0b11100, 0b00, 1, 0,
2066                  (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2067                  IIC_fpMAC32, "vmls", ".f32\t$Sd, $Sn, $Sm",
2068                  [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
2069                                           SPR:$Sdin))]>,
2070              RegConstraint<"$Sdin = $Sd">,
2071              Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>,
2072              Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2073  // Some single precision VFP instructions may be executed on both NEON and
2074  // VFP pipelines on A8.
2075  let D = VFPNeonA8Domain;
2076}
2077
2078def VMLSH : AHbI<0b11100, 0b00, 1, 0,
2079                  (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2080                  IIC_fpMAC16, "vmls", ".f16\t$Sd, $Sn, $Sm",
2081                  [(set (f16 HPR:$Sd), (fadd_mlx (fneg (fmul_su (f16 HPR:$Sn), (f16 HPR:$Sm))),
2082                                           (f16 HPR:$Sdin)))]>,
2083              RegConstraint<"$Sdin = $Sd">,
2084              Requires<[HasFullFP16,UseFPVMLx]>;
2085
2086def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
2087          (VMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
2088          Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>;
2089def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
2090          (VMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
2091          Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>;
2092def : Pat<(fsub_mlx HPR:$dstin, (fmul_su (f16 HPR:$a), HPR:$b)),
2093          (VMLSH HPR:$dstin, (f16 HPR:$a), HPR:$b)>,
2094          Requires<[HasFullFP16,DontUseNEONForFP,UseFPVMLx]>;
2095
2096def VNMLAD : ADbI<0b11100, 0b01, 1, 0,
2097                  (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2098                  IIC_fpMAC64, "vnmla", ".f64\t$Dd, $Dn, $Dm",
2099                  [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
2100                                          (f64 DPR:$Ddin)))]>,
2101                RegConstraint<"$Ddin = $Dd">,
2102                Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>,
2103                Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2104
2105def VNMLAS : ASbI<0b11100, 0b01, 1, 0,
2106                  (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2107                  IIC_fpMAC32, "vnmla", ".f32\t$Sd, $Sn, $Sm",
2108                  [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
2109                                           SPR:$Sdin))]>,
2110                RegConstraint<"$Sdin = $Sd">,
2111                Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>,
2112                Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2113  // Some single precision VFP instructions may be executed on both NEON and
2114  // VFP pipelines on A8.
2115  let D = VFPNeonA8Domain;
2116}
2117
2118def VNMLAH : AHbI<0b11100, 0b01, 1, 0,
2119                  (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2120                  IIC_fpMAC16, "vnmla", ".f16\t$Sd, $Sn, $Sm",
2121                  [(set (f16 HPR:$Sd), (fsub_mlx (fneg (fmul_su (f16 HPR:$Sn), (f16 HPR:$Sm))),
2122                                           (f16 HPR:$Sdin)))]>,
2123                RegConstraint<"$Sdin = $Sd">,
2124                Requires<[HasFullFP16,UseFPVMLx]>;
2125
2126// (-(a * b) - dst) -> -(dst + (a * b))
2127def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
2128          (VNMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
2129          Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>;
2130def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
2131          (VNMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
2132          Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>;
2133def : Pat<(fsub_mlx (fneg (fmul_su (f16 HPR:$a), HPR:$b)), HPR:$dstin),
2134          (VNMLAH HPR:$dstin, (f16 HPR:$a), HPR:$b)>,
2135          Requires<[HasFullFP16,DontUseNEONForFP,UseFPVMLx]>;
2136
2137// (-dst - (a * b)) -> -(dst + (a * b))
2138def : Pat<(fsub_mlx (fneg DPR:$dstin), (fmul_su DPR:$a, (f64 DPR:$b))),
2139          (VNMLAD DPR:$dstin, DPR:$a, DPR:$b)>,
2140          Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>;
2141def : Pat<(fsub_mlx (fneg SPR:$dstin), (fmul_su SPR:$a, SPR:$b)),
2142          (VNMLAS SPR:$dstin, SPR:$a, SPR:$b)>,
2143          Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>;
2144def : Pat<(fsub_mlx (fneg HPR:$dstin), (fmul_su (f16 HPR:$a), HPR:$b)),
2145          (VNMLAH HPR:$dstin, (f16 HPR:$a), HPR:$b)>,
2146          Requires<[HasFullFP16,DontUseNEONForFP,UseFPVMLx]>;
2147
2148def VNMLSD : ADbI<0b11100, 0b01, 0, 0,
2149                  (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2150                  IIC_fpMAC64, "vnmls", ".f64\t$Dd, $Dn, $Dm",
2151                  [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
2152                                           (f64 DPR:$Ddin)))]>,
2153               RegConstraint<"$Ddin = $Dd">,
2154               Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>,
2155               Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2156
2157def VNMLSS : ASbI<0b11100, 0b01, 0, 0,
2158                  (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2159                  IIC_fpMAC32, "vnmls", ".f32\t$Sd, $Sn, $Sm",
2160             [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
2161                         RegConstraint<"$Sdin = $Sd">,
2162                Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>,
2163             Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2164  // Some single precision VFP instructions may be executed on both NEON and
2165  // VFP pipelines on A8.
2166  let D = VFPNeonA8Domain;
2167}
2168
2169def VNMLSH : AHbI<0b11100, 0b01, 0, 0,
2170                  (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2171                  IIC_fpMAC16, "vnmls", ".f16\t$Sd, $Sn, $Sm",
2172             [(set (f16 HPR:$Sd), (fsub_mlx (fmul_su (f16 HPR:$Sn), (f16 HPR:$Sm)), (f16 HPR:$Sdin)))]>,
2173                         RegConstraint<"$Sdin = $Sd">,
2174                Requires<[HasFullFP16,UseFPVMLx]>;
2175
2176def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
2177          (VNMLSD DPR:$dstin, DPR:$a, DPR:$b)>,
2178          Requires<[HasVFP2,HasDPVFP,UseFPVMLx]>;
2179def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
2180          (VNMLSS SPR:$dstin, SPR:$a, SPR:$b)>,
2181          Requires<[HasVFP2,DontUseNEONForFP,UseFPVMLx]>;
2182def : Pat<(fsub_mlx (fmul_su (f16 HPR:$a), HPR:$b), HPR:$dstin),
2183          (VNMLSH HPR:$dstin, (f16 HPR:$a), HPR:$b)>,
2184          Requires<[HasFullFP16,DontUseNEONForFP,UseFPVMLx]>;
2185
2186//===----------------------------------------------------------------------===//
2187// Fused FP Multiply-Accumulate Operations.
2188//
2189def VFMAD : ADbI<0b11101, 0b10, 0, 0,
2190                 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2191                 IIC_fpFMAC64, "vfma", ".f64\t$Dd, $Dn, $Dm",
2192                 [(set DPR:$Dd, (fadd_mlx (fmul_su DPR:$Dn, DPR:$Dm),
2193                                          (f64 DPR:$Ddin)))]>,
2194              RegConstraint<"$Ddin = $Dd">,
2195              Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>,
2196            Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2197
2198def VFMAS : ASbIn<0b11101, 0b10, 0, 0,
2199                  (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2200                  IIC_fpFMAC32, "vfma", ".f32\t$Sd, $Sn, $Sm",
2201                  [(set SPR:$Sd, (fadd_mlx (fmul_su SPR:$Sn, SPR:$Sm),
2202                                           SPR:$Sdin))]>,
2203              RegConstraint<"$Sdin = $Sd">,
2204              Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>,
2205            Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2206  // Some single precision VFP instructions may be executed on both NEON and
2207  // VFP pipelines.
2208}
2209
2210def VFMAH : AHbI<0b11101, 0b10, 0, 0,
2211                  (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2212                  IIC_fpFMAC16, "vfma", ".f16\t$Sd, $Sn, $Sm",
2213                  [(set (f16 HPR:$Sd), (fadd_mlx (fmul_su (f16 HPR:$Sn), (f16 HPR:$Sm)),
2214                                           (f16 HPR:$Sdin)))]>,
2215              RegConstraint<"$Sdin = $Sd">,
2216              Requires<[HasFullFP16,UseFusedMAC]>,
2217            Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2218
2219def : Pat<(fadd_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
2220          (VFMAD DPR:$dstin, DPR:$a, DPR:$b)>,
2221          Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
2222def : Pat<(fadd_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
2223          (VFMAS SPR:$dstin, SPR:$a, SPR:$b)>,
2224          Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
2225def : Pat<(fadd_mlx HPR:$dstin, (fmul_su (f16 HPR:$a), HPR:$b)),
2226          (VFMAH HPR:$dstin, (f16 HPR:$a), HPR:$b)>,
2227          Requires<[HasFullFP16,DontUseNEONForFP,UseFusedMAC]>;
2228
2229// Match @llvm.fma.* intrinsics
2230// (fma x, y, z) -> (vfms z, x, y)
2231def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, DPR:$Ddin)),
2232          (VFMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2233      Requires<[HasVFP4,HasDPVFP]>;
2234def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, SPR:$Sdin)),
2235          (VFMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2236      Requires<[HasVFP4]>;
2237def : Pat<(f16 (fma HPR:$Sn, HPR:$Sm, (f16 HPR:$Sdin))),
2238          (VFMAH (f16 HPR:$Sdin), (f16 HPR:$Sn), (f16 HPR:$Sm))>,
2239      Requires<[HasFullFP16]>;
2240
2241def VFMSD : ADbI<0b11101, 0b10, 1, 0,
2242                 (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2243                 IIC_fpFMAC64, "vfms", ".f64\t$Dd, $Dn, $Dm",
2244                 [(set DPR:$Dd, (fadd_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
2245                                          (f64 DPR:$Ddin)))]>,
2246              RegConstraint<"$Ddin = $Dd">,
2247              Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>,
2248              Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2249
2250def VFMSS : ASbIn<0b11101, 0b10, 1, 0,
2251                  (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2252                  IIC_fpFMAC32, "vfms", ".f32\t$Sd, $Sn, $Sm",
2253                  [(set SPR:$Sd, (fadd_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
2254                                           SPR:$Sdin))]>,
2255              RegConstraint<"$Sdin = $Sd">,
2256              Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>,
2257              Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2258  // Some single precision VFP instructions may be executed on both NEON and
2259  // VFP pipelines.
2260}
2261
2262def VFMSH : AHbI<0b11101, 0b10, 1, 0,
2263                  (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2264                  IIC_fpFMAC16, "vfms", ".f16\t$Sd, $Sn, $Sm",
2265                  [(set (f16 HPR:$Sd), (fadd_mlx (fneg (fmul_su (f16 HPR:$Sn), (f16 HPR:$Sm))),
2266                                           (f16 HPR:$Sdin)))]>,
2267              RegConstraint<"$Sdin = $Sd">,
2268              Requires<[HasFullFP16,UseFusedMAC]>,
2269              Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2270
2271def : Pat<(fsub_mlx DPR:$dstin, (fmul_su DPR:$a, (f64 DPR:$b))),
2272          (VFMSD DPR:$dstin, DPR:$a, DPR:$b)>,
2273          Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
2274def : Pat<(fsub_mlx SPR:$dstin, (fmul_su SPR:$a, SPR:$b)),
2275          (VFMSS SPR:$dstin, SPR:$a, SPR:$b)>,
2276          Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
2277def : Pat<(fsub_mlx HPR:$dstin, (fmul_su (f16 HPR:$a), HPR:$b)),
2278          (VFMSH HPR:$dstin, (f16 HPR:$a), HPR:$b)>,
2279          Requires<[HasFullFP16,DontUseNEONForFP,UseFusedMAC]>;
2280
2281// Match @llvm.fma.* intrinsics
2282// (fma (fneg x), y, z) -> (vfms z, x, y)
2283def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin)),
2284          (VFMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2285      Requires<[HasVFP4,HasDPVFP]>;
2286def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin)),
2287          (VFMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2288      Requires<[HasVFP4]>;
2289def : Pat<(f16 (fma (fneg (f16 HPR:$Sn)), (f16 HPR:$Sm), (f16 HPR:$Sdin))),
2290          (VFMSH (f16 HPR:$Sdin), (f16 HPR:$Sn), (f16 HPR:$Sm))>,
2291      Requires<[HasFullFP16]>;
2292
2293def VFNMAD : ADbI<0b11101, 0b01, 1, 0,
2294                  (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2295                  IIC_fpFMAC64, "vfnma", ".f64\t$Dd, $Dn, $Dm",
2296                  [(set DPR:$Dd,(fsub_mlx (fneg (fmul_su DPR:$Dn,DPR:$Dm)),
2297                                          (f64 DPR:$Ddin)))]>,
2298                RegConstraint<"$Ddin = $Dd">,
2299                Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>,
2300                Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2301
2302def VFNMAS : ASbI<0b11101, 0b01, 1, 0,
2303                  (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2304                  IIC_fpFMAC32, "vfnma", ".f32\t$Sd, $Sn, $Sm",
2305                  [(set SPR:$Sd, (fsub_mlx (fneg (fmul_su SPR:$Sn, SPR:$Sm)),
2306                                           SPR:$Sdin))]>,
2307                RegConstraint<"$Sdin = $Sd">,
2308                Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>,
2309                Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2310  // Some single precision VFP instructions may be executed on both NEON and
2311  // VFP pipelines.
2312}
2313
2314def VFNMAH : AHbI<0b11101, 0b01, 1, 0,
2315                  (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2316                  IIC_fpFMAC16, "vfnma", ".f16\t$Sd, $Sn, $Sm",
2317                  [(set (f16 HPR:$Sd), (fsub_mlx (fneg (fmul_su (f16 HPR:$Sn), (f16 HPR:$Sm))),
2318                                           (f16 HPR:$Sdin)))]>,
2319                RegConstraint<"$Sdin = $Sd">,
2320                Requires<[HasFullFP16,UseFusedMAC]>,
2321                Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2322
2323def : Pat<(fsub_mlx (fneg (fmul_su DPR:$a, (f64 DPR:$b))), DPR:$dstin),
2324          (VFNMAD DPR:$dstin, DPR:$a, DPR:$b)>,
2325          Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
2326def : Pat<(fsub_mlx (fneg (fmul_su SPR:$a, SPR:$b)), SPR:$dstin),
2327          (VFNMAS SPR:$dstin, SPR:$a, SPR:$b)>,
2328          Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
2329
2330// Match @llvm.fma.* intrinsics
2331// (fneg (fma x, y, z)) -> (vfnma z, x, y)
2332def : Pat<(fneg (fma (f64 DPR:$Dn), (f64 DPR:$Dm), (f64 DPR:$Ddin))),
2333          (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2334      Requires<[HasVFP4,HasDPVFP]>;
2335def : Pat<(fneg (fma (f32 SPR:$Sn), (f32 SPR:$Sm), (f32 SPR:$Sdin))),
2336          (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2337      Requires<[HasVFP4]>;
2338def : Pat<(fneg (fma (f16 HPR:$Sn), (f16 HPR:$Sm), (f16 (f16 HPR:$Sdin)))),
2339          (VFNMAH (f16 HPR:$Sdin), (f16 HPR:$Sn), (f16 HPR:$Sm))>,
2340      Requires<[HasFullFP16]>;
2341// (fma (fneg x), y, (fneg z)) -> (vfnma z, x, y)
2342def : Pat<(f64 (fma (fneg DPR:$Dn), DPR:$Dm, (fneg DPR:$Ddin))),
2343          (VFNMAD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2344      Requires<[HasVFP4,HasDPVFP]>;
2345def : Pat<(f32 (fma (fneg SPR:$Sn), SPR:$Sm, (fneg SPR:$Sdin))),
2346          (VFNMAS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2347      Requires<[HasVFP4]>;
2348def : Pat<(f16 (fma (fneg (f16 HPR:$Sn)), (f16 HPR:$Sm), (fneg (f16 HPR:$Sdin)))),
2349          (VFNMAH (f16 HPR:$Sdin), (f16 HPR:$Sn), (f16 HPR:$Sm))>,
2350      Requires<[HasFullFP16]>;
2351
2352def VFNMSD : ADbI<0b11101, 0b01, 0, 0,
2353                  (outs DPR:$Dd), (ins DPR:$Ddin, DPR:$Dn, DPR:$Dm),
2354                  IIC_fpFMAC64, "vfnms", ".f64\t$Dd, $Dn, $Dm",
2355                  [(set DPR:$Dd, (fsub_mlx (fmul_su DPR:$Dn, DPR:$Dm),
2356                                           (f64 DPR:$Ddin)))]>,
2357               RegConstraint<"$Ddin = $Dd">,
2358               Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>,
2359               Sched<[WriteFPMAC64, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2360
2361def VFNMSS : ASbI<0b11101, 0b01, 0, 0,
2362                  (outs SPR:$Sd), (ins SPR:$Sdin, SPR:$Sn, SPR:$Sm),
2363                  IIC_fpFMAC32, "vfnms", ".f32\t$Sd, $Sn, $Sm",
2364             [(set SPR:$Sd, (fsub_mlx (fmul_su SPR:$Sn, SPR:$Sm), SPR:$Sdin))]>,
2365                         RegConstraint<"$Sdin = $Sd">,
2366                  Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>,
2367                  Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]> {
2368  // Some single precision VFP instructions may be executed on both NEON and
2369  // VFP pipelines.
2370}
2371
2372def VFNMSH : AHbI<0b11101, 0b01, 0, 0,
2373                  (outs HPR:$Sd), (ins HPR:$Sdin, HPR:$Sn, HPR:$Sm),
2374                  IIC_fpFMAC16, "vfnms", ".f16\t$Sd, $Sn, $Sm",
2375             [(set (f16 HPR:$Sd), (fsub_mlx (fmul_su (f16 HPR:$Sn), (f16 HPR:$Sm)), (f16 HPR:$Sdin)))]>,
2376                         RegConstraint<"$Sdin = $Sd">,
2377                  Requires<[HasFullFP16,UseFusedMAC]>,
2378                  Sched<[WriteFPMAC32, ReadFPMAC, ReadFPMUL, ReadFPMUL]>;
2379
2380def : Pat<(fsub_mlx (fmul_su DPR:$a, (f64 DPR:$b)), DPR:$dstin),
2381          (VFNMSD DPR:$dstin, DPR:$a, DPR:$b)>,
2382          Requires<[HasVFP4,HasDPVFP,UseFusedMAC]>;
2383def : Pat<(fsub_mlx (fmul_su SPR:$a, SPR:$b), SPR:$dstin),
2384          (VFNMSS SPR:$dstin, SPR:$a, SPR:$b)>,
2385          Requires<[HasVFP4,DontUseNEONForFP,UseFusedMAC]>;
2386
2387// Match @llvm.fma.* intrinsics
2388
2389// (fma x, y, (fneg z)) -> (vfnms z, x, y))
2390def : Pat<(f64 (fma DPR:$Dn, DPR:$Dm, (fneg DPR:$Ddin))),
2391          (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2392      Requires<[HasVFP4,HasDPVFP]>;
2393def : Pat<(f32 (fma SPR:$Sn, SPR:$Sm, (fneg SPR:$Sdin))),
2394          (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2395      Requires<[HasVFP4]>;
2396def : Pat<(f16 (fma (f16 HPR:$Sn), (f16 HPR:$Sm), (fneg (f16 HPR:$Sdin)))),
2397          (VFNMSH (f16 HPR:$Sdin), (f16 HPR:$Sn), (f16 HPR:$Sm))>,
2398      Requires<[HasFullFP16]>;
2399// (fneg (fma (fneg x), y, z)) -> (vfnms z, x, y)
2400def : Pat<(fneg (f64 (fma (fneg DPR:$Dn), DPR:$Dm, DPR:$Ddin))),
2401          (VFNMSD DPR:$Ddin, DPR:$Dn, DPR:$Dm)>,
2402      Requires<[HasVFP4,HasDPVFP]>;
2403def : Pat<(fneg (f32 (fma (fneg SPR:$Sn), SPR:$Sm, SPR:$Sdin))),
2404          (VFNMSS SPR:$Sdin, SPR:$Sn, SPR:$Sm)>,
2405      Requires<[HasVFP4]>;
2406def : Pat<(fneg (f16 (fma (fneg (f16 HPR:$Sn)), (f16 HPR:$Sm), (f16 HPR:$Sdin)))),
2407          (VFNMSH (f16 HPR:$Sdin), (f16 HPR:$Sn), (f16 HPR:$Sm))>,
2408      Requires<[HasFullFP16]>;
2409
2410//===----------------------------------------------------------------------===//
2411// FP Conditional moves.
2412//
2413
2414let hasSideEffects = 0 in {
2415def VMOVDcc  : PseudoInst<(outs DPR:$Dd), (ins DPR:$Dn, DPR:$Dm, cmovpred:$p),
2416                    IIC_fpUNA64,
2417                    [(set (f64 DPR:$Dd),
2418                          (ARMcmov DPR:$Dn, DPR:$Dm, cmovpred:$p))]>,
2419               RegConstraint<"$Dn = $Dd">, Requires<[HasFPRegs64]>;
2420
2421def VMOVScc  : PseudoInst<(outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm, cmovpred:$p),
2422                    IIC_fpUNA32,
2423                    [(set (f32 SPR:$Sd),
2424                          (ARMcmov SPR:$Sn, SPR:$Sm, cmovpred:$p))]>,
2425               RegConstraint<"$Sn = $Sd">, Requires<[HasFPRegs]>;
2426
2427def VMOVHcc  : PseudoInst<(outs HPR:$Sd), (ins HPR:$Sn, HPR:$Sm, cmovpred:$p),
2428                    IIC_fpUNA16,
2429                    [(set (f16 HPR:$Sd),
2430                          (ARMcmov (f16 HPR:$Sn), (f16 HPR:$Sm), cmovpred:$p))]>,
2431               RegConstraint<"$Sd = $Sn">, Requires<[HasFPRegs]>;
2432} // hasSideEffects
2433
2434//===----------------------------------------------------------------------===//
2435// Move from VFP System Register to ARM core register.
2436//
2437
2438class MovFromVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
2439                 list<dag> pattern>:
2440  VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
2441
2442  // Instruction operand.
2443  bits<4> Rt;
2444
2445  let Inst{27-20} = 0b11101111;
2446  let Inst{19-16} = opc19_16;
2447  let Inst{15-12} = Rt;
2448  let Inst{11-8}  = 0b1010;
2449  let Inst{7}     = 0;
2450  let Inst{6-5}   = 0b00;
2451  let Inst{4}     = 1;
2452  let Inst{3-0}   = 0b0000;
2453  let Unpredictable{7-5} = 0b111;
2454  let Unpredictable{3-0} = 0b1111;
2455}
2456
2457let DecoderMethod = "DecodeForVMRSandVMSR" in {
2458 // APSR is the application level alias of CPSR. This FPSCR N, Z, C, V flags
2459 // to APSR.
2460 let Defs = [CPSR], Uses = [FPSCR_NZCV], Predicates = [HasFPRegs],
2461     Rt = 0b1111 /* apsr_nzcv */ in
2462 def FMSTAT : MovFromVFP<0b0001 /* fpscr */, (outs), (ins),
2463                         "vmrs", "\tAPSR_nzcv, fpscr", [(arm_fmstat)]>;
2464
2465 // Application level FPSCR -> GPR
2466 let hasSideEffects = 1, Uses = [FPSCR], Predicates = [HasFPRegs] in
2467 def VMRS :  MovFromVFP<0b0001 /* fpscr */, (outs GPRnopc:$Rt), (ins),
2468                        "vmrs", "\t$Rt, fpscr",
2469                        [(set GPRnopc:$Rt, (int_arm_get_fpscr))]>;
2470
2471 // System level FPEXC, FPSID -> GPR
2472 let Uses = [FPSCR] in {
2473   def VMRS_FPEXC : MovFromVFP<0b1000 /* fpexc */, (outs GPRnopc:$Rt), (ins),
2474                               "vmrs", "\t$Rt, fpexc", []>;
2475   def VMRS_FPSID : MovFromVFP<0b0000 /* fpsid */, (outs GPRnopc:$Rt), (ins),
2476                               "vmrs", "\t$Rt, fpsid", []>;
2477   def VMRS_MVFR0 : MovFromVFP<0b0111 /* mvfr0 */, (outs GPRnopc:$Rt), (ins),
2478                              "vmrs", "\t$Rt, mvfr0", []>;
2479   def VMRS_MVFR1 : MovFromVFP<0b0110 /* mvfr1 */, (outs GPRnopc:$Rt), (ins),
2480                               "vmrs", "\t$Rt, mvfr1", []>;
2481   let Predicates = [HasFPARMv8] in {
2482     def VMRS_MVFR2 : MovFromVFP<0b0101 /* mvfr2 */, (outs GPRnopc:$Rt), (ins),
2483                                 "vmrs", "\t$Rt, mvfr2", []>;
2484   }
2485   def VMRS_FPINST : MovFromVFP<0b1001 /* fpinst */, (outs GPRnopc:$Rt), (ins),
2486                                "vmrs", "\t$Rt, fpinst", []>;
2487   def VMRS_FPINST2 : MovFromVFP<0b1010 /* fpinst2 */, (outs GPRnopc:$Rt),
2488                                 (ins), "vmrs", "\t$Rt, fpinst2", []>;
2489   let Predicates = [HasV8_1MMainline, HasFPRegs] in {
2490     // System level FPSCR_NZCVQC -> GPR
2491     def VMRS_FPSCR_NZCVQC
2492       : MovFromVFP<0b0010 /* fpscr_nzcvqc */,
2493                    (outs GPR:$Rt), (ins cl_FPSCR_NZCV:$fpscr_in),
2494                    "vmrs", "\t$Rt, fpscr_nzcvqc", []>;
2495   }
2496 }
2497 let Predicates = [HasV8_1MMainline, Has8MSecExt] in {
2498   // System level FPSCR -> GPR, with context saving for security extensions
2499   def VMRS_FPCXTNS : MovFromVFP<0b1110 /* fpcxtns */, (outs GPR:$Rt), (ins),
2500                                 "vmrs", "\t$Rt, fpcxtns", []>;
2501 }
2502 let Predicates = [HasV8_1MMainline, Has8MSecExt] in {
2503   // System level FPSCR -> GPR, with context saving for security extensions
2504   def VMRS_FPCXTS : MovFromVFP<0b1111 /* fpcxts */, (outs GPR:$Rt), (ins),
2505                                "vmrs", "\t$Rt, fpcxts", []>;
2506 }
2507
2508 let Predicates = [HasV8_1MMainline, HasMVEInt] in {
2509   // System level VPR/P0 -> GPR
2510   let Uses = [VPR] in
2511   def VMRS_VPR : MovFromVFP<0b1100 /* vpr */, (outs GPR:$Rt), (ins),
2512                             "vmrs", "\t$Rt, vpr", []>;
2513
2514   def VMRS_P0  : MovFromVFP<0b1101 /* p0 */, (outs GPR:$Rt), (ins VCCR:$cond),
2515                             "vmrs", "\t$Rt, p0", []>;
2516 }
2517}
2518
2519//===----------------------------------------------------------------------===//
2520// Move from ARM core register to VFP System Register.
2521//
2522
2523class MovToVFP<bits<4> opc19_16, dag oops, dag iops, string opc, string asm,
2524               list<dag> pattern>:
2525  VFPAI<oops, iops, VFPMiscFrm, IIC_fpSTAT, opc, asm, pattern> {
2526
2527  // Instruction operand.
2528  bits<4> Rt;
2529
2530  let Inst{27-20} = 0b11101110;
2531  let Inst{19-16} = opc19_16;
2532  let Inst{15-12} = Rt;
2533  let Inst{11-8}  = 0b1010;
2534  let Inst{7}     = 0;
2535  let Inst{6-5}   = 0b00;
2536  let Inst{4}     = 1;
2537  let Inst{3-0}   = 0b0000;
2538  let Predicates = [HasVFP2];
2539  let Unpredictable{7-5} = 0b111;
2540  let Unpredictable{3-0} = 0b1111;
2541}
2542
2543let DecoderMethod = "DecodeForVMRSandVMSR" in {
2544 let Defs = [FPSCR] in {
2545   let Predicates = [HasFPRegs] in
2546   // Application level GPR -> FPSCR
2547   def VMSR : MovToVFP<0b0001 /* fpscr */, (outs), (ins GPRnopc:$Rt),
2548                       "vmsr", "\tfpscr, $Rt",
2549                       [(int_arm_set_fpscr GPRnopc:$Rt)]>;
2550   // System level GPR -> FPEXC
2551   def VMSR_FPEXC : MovToVFP<0b1000 /* fpexc */, (outs), (ins GPRnopc:$Rt),
2552                               "vmsr", "\tfpexc, $Rt", []>;
2553   // System level GPR -> FPSID
2554   def VMSR_FPSID : MovToVFP<0b0000 /* fpsid */, (outs), (ins GPRnopc:$Rt),
2555                             "vmsr", "\tfpsid, $Rt", []>;
2556   def VMSR_FPINST : MovToVFP<0b1001 /* fpinst */, (outs), (ins GPRnopc:$Rt),
2557                              "vmsr", "\tfpinst, $Rt", []>;
2558   def VMSR_FPINST2 : MovToVFP<0b1010 /* fpinst2 */, (outs), (ins GPRnopc:$Rt),
2559                               "vmsr", "\tfpinst2, $Rt", []>;
2560 }
2561 let Predicates = [HasV8_1MMainline, Has8MSecExt] in {
2562   // System level GPR -> FPSCR with context saving for security extensions
2563   def VMSR_FPCXTNS : MovToVFP<0b1110 /* fpcxtns */, (outs), (ins GPR:$Rt),
2564                               "vmsr", "\tfpcxtns, $Rt", []>;
2565 }
2566 let Predicates = [HasV8_1MMainline, Has8MSecExt] in {
2567   // System level GPR -> FPSCR with context saving for security extensions
2568   def VMSR_FPCXTS : MovToVFP<0b1111 /* fpcxts */, (outs), (ins GPR:$Rt),
2569                              "vmsr", "\tfpcxts, $Rt", []>;
2570 }
2571 let Predicates = [HasV8_1MMainline, HasFPRegs] in {
2572   // System level GPR -> FPSCR_NZCVQC
2573   def VMSR_FPSCR_NZCVQC
2574     : MovToVFP<0b0010 /* fpscr_nzcvqc */,
2575                (outs cl_FPSCR_NZCV:$fpscr_out), (ins GPR:$Rt),
2576                "vmsr", "\tfpscr_nzcvqc, $Rt", []>;
2577 }
2578
2579 let Predicates = [HasV8_1MMainline, HasMVEInt] in {
2580   // System level GPR -> VPR/P0
2581   let Defs = [VPR] in
2582   def VMSR_VPR : MovToVFP<0b1100 /* vpr */, (outs), (ins GPR:$Rt),
2583                           "vmsr", "\tvpr, $Rt", []>;
2584
2585   def VMSR_P0  : MovToVFP<0b1101 /* p0 */, (outs VCCR:$cond), (ins GPR:$Rt),
2586                           "vmsr", "\tp0, $Rt", []>;
2587 }
2588}
2589
2590//===----------------------------------------------------------------------===//
2591// Misc.
2592//
2593
2594// Materialize FP immediates. VFP3 only.
2595let isReMaterializable = 1 in {
2596def FCONSTD : VFPAI<(outs DPR:$Dd), (ins vfp_f64imm:$imm),
2597                    VFPMiscFrm, IIC_fpUNA64,
2598                    "vmov", ".f64\t$Dd, $imm",
2599                    [(set DPR:$Dd, vfp_f64imm:$imm)]>,
2600              Requires<[HasVFP3,HasDPVFP]> {
2601  bits<5> Dd;
2602  bits<8> imm;
2603
2604  let Inst{27-23} = 0b11101;
2605  let Inst{22}    = Dd{4};
2606  let Inst{21-20} = 0b11;
2607  let Inst{19-16} = imm{7-4};
2608  let Inst{15-12} = Dd{3-0};
2609  let Inst{11-9}  = 0b101;
2610  let Inst{8}     = 1;          // Double precision.
2611  let Inst{7-4}   = 0b0000;
2612  let Inst{3-0}   = imm{3-0};
2613}
2614
2615def FCONSTS : VFPAI<(outs SPR:$Sd), (ins vfp_f32imm:$imm),
2616                     VFPMiscFrm, IIC_fpUNA32,
2617                     "vmov", ".f32\t$Sd, $imm",
2618                     [(set SPR:$Sd, vfp_f32imm:$imm)]>, Requires<[HasVFP3]> {
2619  bits<5> Sd;
2620  bits<8> imm;
2621
2622  let Inst{27-23} = 0b11101;
2623  let Inst{22}    = Sd{0};
2624  let Inst{21-20} = 0b11;
2625  let Inst{19-16} = imm{7-4};
2626  let Inst{15-12} = Sd{4-1};
2627  let Inst{11-9}  = 0b101;
2628  let Inst{8}     = 0;          // Single precision.
2629  let Inst{7-4}   = 0b0000;
2630  let Inst{3-0}   = imm{3-0};
2631}
2632
2633def FCONSTH : VFPAI<(outs HPR:$Sd), (ins vfp_f16imm:$imm),
2634                     VFPMiscFrm, IIC_fpUNA16,
2635                     "vmov", ".f16\t$Sd, $imm",
2636                     [(set (f16 HPR:$Sd), vfp_f16imm:$imm)]>,
2637              Requires<[HasFullFP16]> {
2638  bits<5> Sd;
2639  bits<8> imm;
2640
2641  let Inst{27-23} = 0b11101;
2642  let Inst{22}    = Sd{0};
2643  let Inst{21-20} = 0b11;
2644  let Inst{19-16} = imm{7-4};
2645  let Inst{15-12} = Sd{4-1};
2646  let Inst{11-8}  = 0b1001;     // Half precision
2647  let Inst{7-4}   = 0b0000;
2648  let Inst{3-0}   = imm{3-0};
2649
2650  let isUnpredicable = 1;
2651}
2652}
2653
2654def : Pat<(f32 (vfp_f32f16imm:$imm)),
2655          (f32 (COPY_TO_REGCLASS (f16 (FCONSTH (vfp_f32f16imm_xform (f32 $imm)))), SPR))> {
2656  let Predicates = [HasFullFP16];
2657}
2658
2659//===----------------------------------------------------------------------===//
2660// Assembler aliases.
2661//
2662// A few mnemonic aliases for pre-unifixed syntax. We don't guarantee to
2663// support them all, but supporting at least some of the basics is
2664// good to be friendly.
2665def : VFP2MnemonicAlias<"flds", "vldr">;
2666def : VFP2MnemonicAlias<"fldd", "vldr">;
2667def : VFP2MnemonicAlias<"fmrs", "vmov">;
2668def : VFP2MnemonicAlias<"fmsr", "vmov">;
2669def : VFP2MnemonicAlias<"fsqrts", "vsqrt">;
2670def : VFP2MnemonicAlias<"fsqrtd", "vsqrt">;
2671def : VFP2MnemonicAlias<"fadds", "vadd.f32">;
2672def : VFP2MnemonicAlias<"faddd", "vadd.f64">;
2673def : VFP2MnemonicAlias<"fmrdd", "vmov">;
2674def : VFP2MnemonicAlias<"fmrds", "vmov">;
2675def : VFP2MnemonicAlias<"fmrrd", "vmov">;
2676def : VFP2MnemonicAlias<"fmdrr", "vmov">;
2677def : VFP2MnemonicAlias<"fmuls", "vmul.f32">;
2678def : VFP2MnemonicAlias<"fmuld", "vmul.f64">;
2679def : VFP2MnemonicAlias<"fnegs", "vneg.f32">;
2680def : VFP2MnemonicAlias<"fnegd", "vneg.f64">;
2681def : VFP2MnemonicAlias<"ftosizd", "vcvt.s32.f64">;
2682def : VFP2MnemonicAlias<"ftosid", "vcvtr.s32.f64">;
2683def : VFP2MnemonicAlias<"ftosizs", "vcvt.s32.f32">;
2684def : VFP2MnemonicAlias<"ftosis", "vcvtr.s32.f32">;
2685def : VFP2MnemonicAlias<"ftouizd", "vcvt.u32.f64">;
2686def : VFP2MnemonicAlias<"ftouid", "vcvtr.u32.f64">;
2687def : VFP2MnemonicAlias<"ftouizs", "vcvt.u32.f32">;
2688def : VFP2MnemonicAlias<"ftouis", "vcvtr.u32.f32">;
2689def : VFP2MnemonicAlias<"fsitod", "vcvt.f64.s32">;
2690def : VFP2MnemonicAlias<"fsitos", "vcvt.f32.s32">;
2691def : VFP2MnemonicAlias<"fuitod", "vcvt.f64.u32">;
2692def : VFP2MnemonicAlias<"fuitos", "vcvt.f32.u32">;
2693def : VFP2MnemonicAlias<"fsts", "vstr">;
2694def : VFP2MnemonicAlias<"fstd", "vstr">;
2695def : VFP2MnemonicAlias<"fmacd", "vmla.f64">;
2696def : VFP2MnemonicAlias<"fmacs", "vmla.f32">;
2697def : VFP2MnemonicAlias<"fcpys", "vmov.f32">;
2698def : VFP2MnemonicAlias<"fcpyd", "vmov.f64">;
2699def : VFP2MnemonicAlias<"fcmps", "vcmp.f32">;
2700def : VFP2MnemonicAlias<"fcmpd", "vcmp.f64">;
2701def : VFP2MnemonicAlias<"fdivs", "vdiv.f32">;
2702def : VFP2MnemonicAlias<"fdivd", "vdiv.f64">;
2703def : VFP2MnemonicAlias<"fmrx", "vmrs">;
2704def : VFP2MnemonicAlias<"fmxr", "vmsr">;
2705
2706// Be friendly and accept the old form of zero-compare
2707def : VFP2DPInstAlias<"fcmpzd${p} $val", (VCMPZD DPR:$val, pred:$p)>;
2708def : VFP2InstAlias<"fcmpzs${p} $val", (VCMPZS SPR:$val, pred:$p)>;
2709
2710
2711def : InstAlias<"fmstat${p}", (FMSTAT pred:$p), 0>, Requires<[HasFPRegs]>;
2712def : VFP2InstAlias<"fadds${p} $Sd, $Sn, $Sm",
2713                    (VADDS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
2714def : VFP2DPInstAlias<"faddd${p} $Dd, $Dn, $Dm",
2715                      (VADDD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
2716def : VFP2InstAlias<"fsubs${p} $Sd, $Sn, $Sm",
2717                    (VSUBS SPR:$Sd, SPR:$Sn, SPR:$Sm, pred:$p)>;
2718def : VFP2DPInstAlias<"fsubd${p} $Dd, $Dn, $Dm",
2719                      (VSUBD DPR:$Dd, DPR:$Dn, DPR:$Dm, pred:$p)>;
2720
2721// No need for the size suffix on VSQRT. It's implied by the register classes.
2722def : VFP2InstAlias<"vsqrt${p} $Sd, $Sm", (VSQRTS SPR:$Sd, SPR:$Sm, pred:$p)>;
2723def : VFP2DPInstAlias<"vsqrt${p} $Dd, $Dm", (VSQRTD DPR:$Dd, DPR:$Dm, pred:$p)>;
2724
2725// VLDR/VSTR accept an optional type suffix.
2726def : VFP2InstAlias<"vldr${p}.32 $Sd, $addr",
2727                    (VLDRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
2728def : VFP2InstAlias<"vstr${p}.32 $Sd, $addr",
2729                    (VSTRS SPR:$Sd, addrmode5:$addr, pred:$p)>;
2730def : VFP2InstAlias<"vldr${p}.64 $Dd, $addr",
2731                    (VLDRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
2732def : VFP2InstAlias<"vstr${p}.64 $Dd, $addr",
2733                    (VSTRD DPR:$Dd, addrmode5:$addr, pred:$p)>;
2734
2735// VMOV can accept optional 32-bit or less data type suffix suffix.
2736def : VFP2InstAlias<"vmov${p}.8 $Rt, $Sn",
2737                    (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
2738def : VFP2InstAlias<"vmov${p}.16 $Rt, $Sn",
2739                    (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
2740def : VFP2InstAlias<"vmov${p}.32 $Rt, $Sn",
2741                    (VMOVRS GPR:$Rt, SPR:$Sn, pred:$p)>;
2742def : VFP2InstAlias<"vmov${p}.8 $Sn, $Rt",
2743                    (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
2744def : VFP2InstAlias<"vmov${p}.16 $Sn, $Rt",
2745                    (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
2746def : VFP2InstAlias<"vmov${p}.32 $Sn, $Rt",
2747                    (VMOVSR SPR:$Sn, GPR:$Rt, pred:$p)>;
2748
2749def : VFP2InstAlias<"vmov${p}.f64 $Rt, $Rt2, $Dn",
2750                    (VMOVRRD GPR:$Rt, GPR:$Rt2, DPR:$Dn, pred:$p)>;
2751def : VFP2InstAlias<"vmov${p}.f64 $Dn, $Rt, $Rt2",
2752                    (VMOVDRR DPR:$Dn, GPR:$Rt, GPR:$Rt2, pred:$p)>;
2753
2754// VMOVS doesn't need the .f32 to disambiguate from the NEON encoding the way
2755// VMOVD does.
2756def : VFP2InstAlias<"vmov${p} $Sd, $Sm",
2757                    (VMOVS SPR:$Sd, SPR:$Sm, pred:$p)>;
2758
2759// FCONSTD/FCONSTS alias for vmov.f64/vmov.f32
2760// These aliases provide added functionality over vmov.f instructions by
2761// allowing users to write assembly containing encoded floating point constants
2762// (e.g. #0x70 vs #1.0).  Without these alises there is no way for the
2763// assembler to accept encoded fp constants (but the equivalent fp-literal is
2764// accepted directly by vmovf).
2765def : VFP3InstAlias<"fconstd${p} $Dd, $val",
2766                    (FCONSTD DPR:$Dd, vfp_f64imm:$val, pred:$p)>;
2767def : VFP3InstAlias<"fconsts${p} $Sd, $val",
2768                    (FCONSTS SPR:$Sd, vfp_f32imm:$val, pred:$p)>;
2769
2770def VSCCLRMD : VFPXI<(outs), (ins pred:$p, fp_dreglist_with_vpr:$regs, variable_ops),
2771                      AddrModeNone, 4, IndexModeNone, VFPMiscFrm, NoItinerary,
2772                      "vscclrm{$p}\t$regs", "", []>, Sched<[]> {
2773  bits<13> regs;
2774  let Inst{31-23} = 0b111011001;
2775  let Inst{22} = regs{12};
2776  let Inst{21-16} = 0b011111;
2777  let Inst{15-12} = regs{11-8};
2778  let Inst{11-8} = 0b1011;
2779  let Inst{7-1} = regs{7-1};
2780  let Inst{0} = 0;
2781
2782  let DecoderMethod = "DecodeVSCCLRM";
2783
2784  list<Predicate> Predicates = [HasV8_1MMainline, Has8MSecExt];
2785}
2786
2787def VSCCLRMS : VFPXI<(outs), (ins pred:$p, fp_sreglist_with_vpr:$regs, variable_ops),
2788                      AddrModeNone, 4, IndexModeNone, VFPMiscFrm, NoItinerary,
2789                      "vscclrm{$p}\t$regs", "", []>, Sched<[]> {
2790  bits<13> regs;
2791  let Inst{31-23} = 0b111011001;
2792  let Inst{22} = regs{8};
2793  let Inst{21-16} = 0b011111;
2794  let Inst{15-12} = regs{12-9};
2795  let Inst{11-8} = 0b1010;
2796  let Inst{7-0} = regs{7-0};
2797
2798  let DecoderMethod = "DecodeVSCCLRM";
2799
2800  list<Predicate> Predicates = [HasV8_1MMainline, Has8MSecExt];
2801}
2802
2803//===----------------------------------------------------------------------===//
2804// Store VFP System Register to memory.
2805//
2806
2807class vfp_vstrldr<bit opc, bit P, bit W, bits<4> SysReg, string sysreg,
2808                  dag oops, dag iops, IndexMode im, string Dest, string cstr>
2809    : VFPI<oops, iops, AddrModeT2_i7s4, 4, im, VFPLdStFrm, IIC_fpSTAT,
2810           !if(opc,"vldr","vstr"), !strconcat("\t", sysreg, ", ", Dest), cstr, []>,
2811      Sched<[]> {
2812  bits<12> addr;
2813  let Inst{27-25} = 0b110;
2814  let Inst{24} = P;
2815  let Inst{23} = addr{7};
2816  let Inst{22} = SysReg{3};
2817  let Inst{21} = W;
2818  let Inst{20} = opc;
2819  let Inst{19-16} = addr{11-8};
2820  let Inst{15-13} = SysReg{2-0};
2821  let Inst{12-7} = 0b011111;
2822  let Inst{6-0} = addr{6-0};
2823  list<Predicate> Predicates = [HasFPRegs, HasV8_1MMainline];
2824  let mayLoad = opc;
2825  let mayStore = !if(opc, 0b0, 0b1);
2826  let hasSideEffects = 1;
2827}
2828
2829multiclass vfp_vstrldr_sysreg<bit opc, bits<4> SysReg, string sysreg,
2830                              dag oops=(outs), dag iops=(ins)> {
2831  def _off :
2832    vfp_vstrldr<opc, 1, 0, SysReg, sysreg,
2833                oops, !con(iops, (ins t2addrmode_imm7s4:$addr)),
2834                IndexModePost, "$addr", "" > {
2835    let DecoderMethod = "DecodeVSTRVLDR_SYSREG<false>";
2836  }
2837
2838  def _pre :
2839    vfp_vstrldr<opc, 1, 1, SysReg, sysreg,
2840                !con(oops, (outs GPRnopc:$wb)),
2841                !con(iops, (ins t2addrmode_imm7s4_pre:$addr)),
2842                IndexModePre, "$addr!", "$addr.base = $wb"> {
2843    let DecoderMethod = "DecodeVSTRVLDR_SYSREG<true>";
2844  }
2845
2846  def _post :
2847    vfp_vstrldr<opc, 0, 1, SysReg, sysreg,
2848                !con(oops, (outs GPRnopc:$wb)),
2849                !con(iops, (ins t2_addr_offset_none:$Rn,
2850                                t2am_imm7s4_offset:$addr)),
2851                IndexModePost, "$Rn$addr", "$Rn.base = $wb"> {
2852   bits<4> Rn;
2853   let Inst{19-16} = Rn{3-0};
2854   let DecoderMethod = "DecodeVSTRVLDR_SYSREG<true>";
2855 }
2856}
2857
2858let Defs = [FPSCR] in {
2859  defm VSTR_FPSCR          : vfp_vstrldr_sysreg<0b0,0b0001, "fpscr">;
2860  defm VSTR_FPSCR_NZCVQC   : vfp_vstrldr_sysreg<0b0,0b0010, "fpscr_nzcvqc">;
2861
2862  let Predicates = [HasV8_1MMainline, Has8MSecExt] in {
2863    defm VSTR_FPCXTNS      : vfp_vstrldr_sysreg<0b0,0b1110, "fpcxtns">;
2864    defm VSTR_FPCXTS       : vfp_vstrldr_sysreg<0b0,0b1111, "fpcxts">;
2865  }
2866}
2867
2868let Predicates = [HasV8_1MMainline, HasMVEInt] in {
2869  let Uses = [VPR] in {
2870    defm VSTR_VPR          : vfp_vstrldr_sysreg<0b0,0b1100, "vpr">;
2871  }
2872  defm VSTR_P0             : vfp_vstrldr_sysreg<0b0,0b1101, "p0",
2873                                                (outs), (ins VCCR:$P0)>;
2874
2875  let Defs = [VPR] in {
2876    defm VLDR_VPR          : vfp_vstrldr_sysreg<0b1,0b1100, "vpr">;
2877  }
2878  defm VLDR_P0             : vfp_vstrldr_sysreg<0b1,0b1101, "p0",
2879                                                (outs VCCR:$P0), (ins)>;
2880}
2881
2882let Uses = [FPSCR] in {
2883  defm VLDR_FPSCR          : vfp_vstrldr_sysreg<0b1,0b0001, "fpscr">;
2884  defm VLDR_FPSCR_NZCVQC   : vfp_vstrldr_sysreg<0b1,0b0010, "fpscr_nzcvqc">;
2885
2886  let Predicates = [HasV8_1MMainline, Has8MSecExt] in {
2887    defm VLDR_FPCXTNS      : vfp_vstrldr_sysreg<0b1,0b1110, "fpcxtns">;
2888    defm VLDR_FPCXTS       : vfp_vstrldr_sysreg<0b1,0b1111, "fpcxts">;
2889  }
2890}
2891