xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86ScheduleBdVer2.td (revision e9ac41698b2f322d55ccf9da50a3596edb2c1800)
1//=- X86ScheduleBdVer2.td - X86 BdVer2 (Piledriver) Scheduling * 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 defines the machine model for AMD bdver2 (Piledriver) to support
10// instruction scheduling and other instruction cost heuristics.
11// Based on:
12//  * AMD Software Optimization Guide for AMD Family 15h Processors.
13//    https://support.amd.com/TechDocs/47414_15h_sw_opt_guide.pdf
14//  * The microarchitecture of Intel, AMD and VIA CPUs, By Agner Fog
15//    http://www.agner.org/optimize/microarchitecture.pdf
16//  * https://www.realworldtech.com/bulldozer/
17//    Yes, that is for Bulldozer aka bdver1, not Piledriver aka bdver2.
18//
19//===----------------------------------------------------------------------===//
20
21def BdVer2Model : SchedMachineModel {
22  let IssueWidth = 4; // Up to 4 IPC can be decoded, issued, retired.
23  let MicroOpBufferSize = 128; // RCU reorder buffer size, which is unconfirmed.
24  let LoopMicroOpBufferSize = -1; // There does not seem to be a loop buffer.
25  let LoadLatency = 4; // L1 data cache has a 4-cycle load-to-use latency.
26  let HighLatency = 25; // FIXME: any better choice?
27  let MispredictPenalty = 20; // Minimum branch misdirection penalty.
28
29  let PostRAScheduler = 1; // Enable Post RegAlloc Scheduler pass.
30
31  // FIXME: Incomplete. This flag is set to allow the scheduler to assign
32  //        a default model to unrecognized opcodes.
33  let CompleteModel = 0;
34} // SchedMachineModel
35
36let SchedModel = BdVer2Model in {
37
38
39//===----------------------------------------------------------------------===//
40// Pipes
41//===----------------------------------------------------------------------===//
42
43// There are total of eight pipes.
44
45//===----------------------------------------------------------------------===//
46// Integer execution pipes
47//
48
49// Two EX (ALU) pipes.
50def PdEX0  : ProcResource<1>; // ALU, Integer Pipe0
51def PdEX1  : ProcResource<1>; // ALU, Integer Pipe1
52def PdEX01 : ProcResGroup<[PdEX0, PdEX1]>;
53
54// Two AGLU pipes, identical.
55def PdAGLU01 : ProcResource<2>; // AGU, Integer Pipe[23]
56
57//===----------------------------------------------------------------------===//
58// Floating point execution pipes
59//
60
61// Four FPU pipes.
62
63def PdFPU0 : ProcResource<1>; // Vector/FPU Pipe0
64def PdFPU1 : ProcResource<1>; // Vector/FPU Pipe1
65def PdFPU2 : ProcResource<1>; // Vector/FPU Pipe2
66def PdFPU3 : ProcResource<1>; // Vector/FPU Pipe3
67
68// FPU grouping
69def PdFPU01 : ProcResGroup<[PdFPU0, PdFPU1]>;
70def PdFPU23 : ProcResGroup<[PdFPU2, PdFPU3]>;
71
72
73//===----------------------------------------------------------------------===//
74// RCU
75//===----------------------------------------------------------------------===//
76
77// The Retire Control Unit on Piledriver can retire up to 4 macro-ops per cycle.
78// On the other hand, the RCU reorder buffer size for Piledriver does not
79// seem be specified in any trustworthy source.
80// But as per https://www.realworldtech.com/bulldozer/6/ the Bulldozer had
81// RCU reorder buffer size of 128. So that is a good guess for now.
82def PdRCU : RetireControlUnit<128, 4>;
83
84
85//===----------------------------------------------------------------------===//
86// Pipelines
87//===----------------------------------------------------------------------===//
88
89// There are total of two pipelines, each one with it's own scheduler.
90
91//===----------------------------------------------------------------------===//
92// Integer Pipeline Scheduling
93//
94
95// There is one Integer Scheduler per core.
96
97// Integer physical register file has 96 registers of 64-bit.
98def PdIntegerPRF : RegisterFile<96, [GR64, CCR]>;
99
100// Unified Integer, Memory Scheduler has 40 entries.
101def PdEX : ProcResGroup<[PdEX0, PdEX1, PdAGLU01]> {
102  // Up to 4 IPC can be decoded, issued, retired.
103  let BufferSize = 40;
104}
105
106
107//===----------------------------------------------------------------------===//
108// FPU Pipeline Scheduling
109//
110
111// The FPU unit is shared between the two cores.
112
113// FP physical register file has 160 registers of 128-bit.
114// Operations on 256-bit data types are cracked into two COPs.
115def PdFpuPRF : RegisterFile<160, [VR64, VR128, VR256], [1, 1, 2]>;
116
117// Unified FP Scheduler has 64 entries,
118def PdFPU : ProcResGroup<[PdFPU0, PdFPU1, PdFPU2, PdFPU3]> {
119  // Up to 4 IPC can be decoded, issued, retired.
120  let BufferSize = 64;
121}
122
123
124//===----------------------------------------------------------------------===//
125// Functional units
126//===----------------------------------------------------------------------===//
127
128//===----------------------------------------------------------------------===//
129// Load-Store Units
130//
131
132let Super = PdAGLU01 in
133def PdLoad  : ProcResource<2> {
134  // For Piledriver, the load queue is 40 entries deep.
135  let BufferSize = 40;
136}
137
138def PdLoadQueue : LoadQueue<PdLoad>;
139
140let Super = PdAGLU01 in
141def PdStore : ProcResource<1> {
142  // For Piledriver, the store queue is 24 entries deep.
143  let BufferSize = 24;
144}
145
146def PdStoreQueue : StoreQueue<PdStore>;
147
148//===----------------------------------------------------------------------===//
149// Integer Execution Units
150//
151
152def PdDiv    : ProcResource<1>; // PdEX0; unpipelined integer division
153def PdCount  : ProcResource<1>; // PdEX0; POPCNT, LZCOUNT
154
155def PdMul    : ProcResource<1>; // PdEX1; integer multiplication
156def PdBranch : ProcResource<1>; // PdEX1; JMP, fused branches
157
158//===----------------------------------------------------------------------===//
159// Floating-Point Units
160//
161
162// Two FMAC/FPFMA units.
163def PdFPFMA  : ProcResource<2>; // PdFPU0, PdFPU1
164
165// One 128-bit integer multiply-accumulate unit.
166def PdFPMMA  : ProcResource<1>; // PdFPU0
167
168// One fp conversion unit.
169def PdFPCVT  : ProcResource<1>; // PdFPU0
170
171// One unit for shuffles, packs, permutes, shifts.
172def PdFPXBR  : ProcResource<1>; // PdFPU1
173
174// Two 128-bit packed integer units.
175def PdFPMAL  : ProcResource<2>; // PdFPU2, PdFPU3
176
177// One FP store unit.
178def PdFPSTO  : ProcResource<1>; // PdFPU3
179
180
181//===----------------------------------------------------------------------===//
182// Basic helper classes.
183//===----------------------------------------------------------------------===//
184
185// Many SchedWrites are defined in pairs with and without a folded load.
186// Instructions with folded loads are usually micro-fused, so they only appear
187// as two micro-ops when dispatched by the schedulers.
188// This multiclass defines the resource usage for variants with and without
189// folded loads.
190multiclass PdWriteRes<SchedWrite SchedRW,
191                      list<ProcResourceKind> ExePorts, int Lat = 1,
192                      list<int> Res = [], int UOps = 1> {
193  def : WriteRes<SchedRW, ExePorts> {
194    let Latency = Lat;
195    let ReleaseAtCycles = Res;
196    let NumMicroOps = UOps;
197  }
198}
199
200multiclass __pdWriteResPair<X86FoldableSchedWrite SchedRW,
201                            list<ProcResourceKind> ExePorts, int Lat,
202                            list<int> Res, int UOps,
203                            int LoadLat, int LoadRes, int LoadUOps> {
204  defm : PdWriteRes<SchedRW, ExePorts, Lat, Res, UOps>;
205
206  defm : PdWriteRes<SchedRW.Folded,
207                    !listconcat([PdLoad], ExePorts),
208                    !add(Lat, LoadLat),
209                    !if(!and(!empty(Res), !eq(LoadRes, 1)),
210                      [],
211                      !listconcat([LoadRes],
212                        !if(!empty(Res),
213                          !listsplat(1, !size(ExePorts)),
214                          Res))),
215                    !add(UOps, LoadUOps)>;
216}
217
218multiclass PdWriteResExPair<X86FoldableSchedWrite SchedRW,
219                            list<ProcResourceKind> ExePorts, int Lat = 1,
220                            list<int> Res = [], int UOps = 1,
221                            int LoadUOps = 0> {
222  defm : __pdWriteResPair<SchedRW, ExePorts, Lat, Res, UOps,
223                          /*LoadLat*/4, /*LoadRes*/3, LoadUOps>;
224}
225
226multiclass PdWriteResXMMPair<X86FoldableSchedWrite SchedRW,
227                             list<ProcResourceKind> ExePorts, int Lat = 1,
228                             list<int> Res = [], int UOps = 1,
229                             int LoadUOps = 0> {
230  defm : __pdWriteResPair<SchedRW, ExePorts, Lat, Res, UOps,
231                           /*LoadLat*/5, /*LoadRes*/3, LoadUOps>;
232}
233
234multiclass PdWriteResYMMPair<X86FoldableSchedWrite SchedRW,
235                             list<ProcResourceKind> ExePorts, int Lat,
236                             list<int> Res = [], int UOps = 2,
237                             int LoadUOps = 0> {
238  defm : __pdWriteResPair<SchedRW, ExePorts, Lat, Res, UOps,
239                           /*LoadLat*/5, /*LoadRes*/3, LoadUOps>;
240}
241
242//===----------------------------------------------------------------------===//
243// Here be dragons.
244//===----------------------------------------------------------------------===//
245
246// L1 data cache has a 4-cycle load-to-use latency, so ReadAfterLd registers
247// needn't be available until 4 cycles after the memory operand.
248def : ReadAdvance<ReadAfterLd, 4>;
249
250// Vector loads are 5 cycles, so ReadAfterVec*Ld registers needn't be available
251// until 5 cycles after the memory operand.
252def : ReadAdvance<ReadAfterVecLd, 5>;
253def : ReadAdvance<ReadAfterVecXLd, 5>;
254def : ReadAdvance<ReadAfterVecYLd, 5>;
255
256// Transfer from int domain to ivec domain incurs additional latency of 8..10cy
257// Reference: Agner, Microarchitecture, "AMD Bulldozer, Piledriver, Steamroller
258// and Excavator pipeline", "Data delay between different execution domains"
259def : ReadAdvance<ReadInt2Fpu, -10>;
260
261// A folded store needs a cycle on the PdStore for the store data.
262def : WriteRes<WriteRMW, [PdStore]>;
263
264////////////////////////////////////////////////////////////////////////////////
265// Loads, stores, and moves, not folded with other operations.
266////////////////////////////////////////////////////////////////////////////////
267
268def : WriteRes<WriteLoad,    [PdLoad]> { let Latency = 5; let ReleaseAtCycles = [2]; }
269def : WriteRes<WriteStore,   [PdStore]>;
270def : WriteRes<WriteStoreNT, [PdStore]>;
271def : WriteRes<WriteMove,    [PdEX01]> { let ReleaseAtCycles = [2]; }
272defm : X86WriteResUnsupported<WriteVecMaskedGatherWriteback>;
273
274// Load/store MXCSR.
275// FIXME: These are copy and pasted from WriteLoad/Store.
276def : WriteRes<WriteLDMXCSR, [PdLoad]> { let Latency = 5; }
277def : WriteRes<WriteSTMXCSR, [PdStore]> { let NumMicroOps = 2; let ReleaseAtCycles = [18]; }
278
279// Treat misc copies as a move.
280def : InstRW<[WriteMove], (instrs COPY)>;
281
282////////////////////////////////////////////////////////////////////////////////
283// Idioms that clear a register, like xorps %xmm0, %xmm0.
284// These can often bypass execution ports completely.
285////////////////////////////////////////////////////////////////////////////////
286
287def : WriteRes<WriteZero, [/*No ExePorts*/]>;
288
289////////////////////////////////////////////////////////////////////////////////
290// Branches don't produce values, so they have no latency, but they still
291// consume resources. Indirect branches can fold loads.
292////////////////////////////////////////////////////////////////////////////////
293
294defm : PdWriteResExPair<WriteJump,  [PdEX1, PdBranch]>;
295
296////////////////////////////////////////////////////////////////////////////////
297// Special case scheduling classes.
298////////////////////////////////////////////////////////////////////////////////
299
300def : WriteRes<WriteSystem,     [PdEX01]> { let Latency = 100; }
301def : WriteRes<WriteMicrocoded, [PdEX01]> { let Latency = 100; }
302def : WriteRes<WriteFence,      [PdStore]>;
303
304def PdWriteXLAT : SchedWriteRes<[PdEX01]> {
305  let Latency = 6;
306}
307def : InstRW<[PdWriteXLAT], (instrs XLAT)>;
308
309def PdWriteLARrr : SchedWriteRes<[PdEX01]> {
310  let Latency = 184;
311  let ReleaseAtCycles = [375];
312  let NumMicroOps = 45;
313}
314def : InstRW<[PdWriteLARrr], (instregex "LAR(16|32|64)rr",
315                                        "LSL(16|32|64)rr")>;
316
317// Nops don't have dependencies, so there's no actual latency, but we set this
318// to '1' to tell the scheduler that the nop uses an ALU slot for a cycle.
319def : WriteRes<WriteNop, [PdEX01]> { let ReleaseAtCycles = [2]; }
320
321////////////////////////////////////////////////////////////////////////////////
322// Arithmetic.
323////////////////////////////////////////////////////////////////////////////////
324
325defm : PdWriteResExPair<WriteALU,     [PdEX01], 1, [2]>;
326
327def PdWriteALURMW : SchedWriteRes<[PdLoad, PdEX01, PdStore]> {
328  let Latency = 6;
329  let ReleaseAtCycles = [3, 2, 1];
330  let NumMicroOps = 1;
331}
332def : SchedAlias<WriteALURMW, PdWriteALURMW>;
333
334def PdWriteLXADD : SchedWriteRes<[PdEX01]> {
335  let Latency = 6;
336  let ReleaseAtCycles = [88];
337  let NumMicroOps = 4;
338}
339def : InstRW<[PdWriteLXADD], (instrs LXADD8, LXADD16, LXADD32, LXADD64)>;
340
341def PdWriteBMI1 : SchedWriteRes<[PdEX01]> {
342  let Latency = 2;
343  let ReleaseAtCycles = [2];
344  let NumMicroOps = 2;
345}
346def : InstRW<[PdWriteBMI1],
347             (instrs BLCFILL32rr, BLCFILL64rr, BLCI32rr, BLCI64rr,
348                     BLCIC32rr, BLCIC64rr, BLCMSK32rr, BLCMSK64rr,
349                     BLCS32rr, BLCS64rr, BLSFILL32rr, BLSFILL64rr,
350                     BLSIC32rr, BLSIC64rr, T1MSKC32rr, T1MSKC64rr,
351                     TZMSK32rr, TZMSK64rr)>;
352
353def PdWriteBMI1m : SchedWriteRes<[PdLoad, PdEX01]> {
354  let Latency = 6;
355  let ReleaseAtCycles = [3, 3];
356  let NumMicroOps = 2;
357}
358def : InstRW<[PdWriteBMI1m],
359             (instrs BLCFILL32rm, BLCFILL64rm, BLCI32rm, BLCI64rm,
360                     BLCIC32rm, BLCIC64rm, BLCMSK32rm, BLCMSK64rm,
361                     BLCS32rm, BLCS64rm, BLSFILL32rm, BLSFILL64rm,
362                     BLSIC32rm, BLSIC64rm, T1MSKC32rm, T1MSKC64rm,
363                     TZMSK32rm, TZMSK64rm)>;
364
365defm : PdWriteResExPair<WriteADC,    [PdEX01],                  1,  [2]>;
366
367def PdWriteADCSBB64ri32 : SchedWriteRes<[PdEX01]> {
368  let ReleaseAtCycles = [3];
369}
370def : InstRW<[PdWriteADCSBB64ri32], (instrs ADC64ri32, SBB64ri32)>;
371
372defm : PdWriteRes<WriteBSWAP32,      [PdEX01]>;
373defm : PdWriteRes<WriteBSWAP64,      [PdEX01]>;
374defm : PdWriteRes<WriteCMPXCHG,      [PdEX1],                   3,  [3],        5>;
375defm : PdWriteRes<WriteCMPXCHGRMW,   [PdEX1, PdStore, PdLoad],  3,  [44, 1, 1], 2>;
376defm : PdWriteRes<WriteXCHG,         [PdEX1],                   1,  [],         2>;
377
378def PdWriteCMPXCHG8rr : SchedWriteRes<[PdEX1]> {
379  let Latency = 3;
380  let ReleaseAtCycles = [3];
381  let NumMicroOps = 3;
382}
383def : InstRW<[PdWriteCMPXCHG8rr], (instrs CMPXCHG8rr)>;
384
385def PdWriteCMPXCHG8rm : SchedWriteRes<[PdEX1]> {
386  let Latency = 3;
387  let ReleaseAtCycles = [23];
388  let NumMicroOps = 5;
389}
390def : InstRW<[PdWriteCMPXCHG8rm], (instrs CMPXCHG8rm)>;
391
392def PdWriteCMPXCHG16rm_CMPXCHG32rm_CMPXCHG64rm : SchedWriteRes<[PdEX1]> {
393  let Latency = 3;
394  let ReleaseAtCycles = [21];
395  let NumMicroOps = 6;
396}
397def : InstRW<[PdWriteCMPXCHG16rm_CMPXCHG32rm_CMPXCHG64rm],
398             (instrs CMPXCHG16rm, CMPXCHG32rm, CMPXCHG64rm)>;
399
400def PdWriteCMPXCHG8B : SchedWriteRes<[PdEX1]> {
401  let Latency = 3;
402  let ReleaseAtCycles = [26];
403  let NumMicroOps = 18;
404}
405def : InstRW<[PdWriteCMPXCHG8B], (instrs CMPXCHG8B)>;
406
407def PdWriteCMPXCHG16B : SchedWriteRes<[PdEX1]> {
408  let Latency = 3;
409  let ReleaseAtCycles = [69];
410  let NumMicroOps = 22;
411}
412def : InstRW<[PdWriteCMPXCHG16B], (instrs CMPXCHG16B)>;
413
414def PdWriteXADDm : SchedWriteRes<[PdEX1]> {
415  let Latency = 6;
416  let ReleaseAtCycles = [20];
417  let NumMicroOps = 4;
418}
419def : InstRW<[PdWriteXADDm], (instrs XADD8rm, XADD16rm, XADD32rm, XADD64rm)>;
420
421defm : PdWriteResExPair<WriteIMul8,     [PdEX1, PdMul],          4,  [1, 4]>;
422defm : PdWriteResExPair<WriteIMul16,    [PdEX1, PdMul],          4,  [1, 5],    2>;
423defm : PdWriteResExPair<WriteIMul16Imm, [PdEX1, PdMul],          5,  [1, 5],    2>;
424defm : PdWriteResExPair<WriteIMul16Reg, [PdEX1, PdMul],          4,  [1, 2]>;
425defm : PdWriteResExPair<WriteIMul32,    [PdEX1, PdMul],          4,  [1, 4]>;
426defm : PdWriteResExPair<WriteIMul32Imm, [PdEX1, PdMul],          4,  [1, 2],    1, 1>;
427defm : PdWriteResExPair<WriteIMul32Reg, [PdEX1, PdMul],          4,  [1, 2]>;
428defm : PdWriteResExPair<WriteIMul64,    [PdEX1, PdMul],          6,  [1, 6]>;
429defm : PdWriteResExPair<WriteIMul64Imm, [PdEX1, PdMul],          6,  [1, 4],1, 1>;
430defm : PdWriteResExPair<WriteIMul64Reg, [PdEX1, PdMul],          6,  [1, 4]>;
431
432// BMI2 MULX
433defm : X86WriteResUnsupported<WriteIMulH>;
434defm : X86WriteResUnsupported<WriteIMulHLd>;
435defm : X86WriteResPairUnsupported<WriteMULX32>;
436defm : X86WriteResPairUnsupported<WriteMULX64>;
437
438defm : PdWriteResExPair<WriteDiv8,    [PdEX1, PdDiv],           12,  [1, 12]>;
439defm : PdWriteResExPair<WriteDiv16,   [PdEX1, PdDiv],           15,  [1, 15],   2>;
440defm : PdWriteResExPair<WriteDiv32,   [PdEX1, PdDiv],           14,  [1, 14],   2>;
441defm : PdWriteResExPair<WriteDiv64,   [PdEX1, PdDiv],           14,  [1, 14],   2>;
442
443defm : PdWriteResExPair<WriteIDiv8,   [PdEX1, PdDiv],           12,  [1, 12]>;
444defm : PdWriteResExPair<WriteIDiv16,  [PdEX1, PdDiv],           15,  [1, 17],   2>;
445defm : PdWriteResExPair<WriteIDiv32,  [PdEX1, PdDiv],           14,  [1, 25],   2>;
446defm : PdWriteResExPair<WriteIDiv64,  [PdEX1, PdDiv],           14,  [1, 14],   2>;
447
448defm : PdWriteResExPair<WriteCRC32,   [PdEX01],                  2,  [4],       3>;
449
450def PdWriteCRC32r32r16 : SchedWriteRes<[PdEX01]> {
451  let Latency = 5;
452  let ReleaseAtCycles = [10];
453  let NumMicroOps = 5;
454}
455def : InstRW<[PdWriteCRC32r32r16], (instrs CRC32r32r16)>;
456
457def PdWriteCRC32r32r32 : SchedWriteRes<[PdEX01]> {
458  let Latency = 6;
459  let ReleaseAtCycles = [12];
460  let NumMicroOps = 7;
461}
462def : InstRW<[PdWriteCRC32r32r32], (instrs CRC32r32r32)>;
463
464def PdWriteCRC32r64r64 : SchedWriteRes<[PdEX01]> {
465  let Latency = 10;
466  let ReleaseAtCycles = [17];
467  let NumMicroOps = 11;
468}
469def : InstRW<[PdWriteCRC32r64r64], (instrs CRC32r64r64)>;
470
471defm : PdWriteResExPair<WriteCMOV,    [PdEX01]>; // Conditional move.
472
473def PdWriteCMOVm : SchedWriteRes<[PdLoad, PdEX01]> {
474  let Latency = 5;
475  let ReleaseAtCycles = [3, 3];
476  let NumMicroOps = 2;
477}
478
479def PdWriteCMOVmVar : SchedWriteVariant<[
480  SchedVar<MCSchedPredicate<CheckImmOperand_s<7, "X86::COND_BE">>, [PdWriteCMOVm]>,
481  SchedVar<MCSchedPredicate<CheckImmOperand_s<7, "X86::COND_A">>,  [PdWriteCMOVm]>,
482  SchedVar<MCSchedPredicate<CheckImmOperand_s<7, "X86::COND_L">>,  [PdWriteCMOVm]>,
483  SchedVar<MCSchedPredicate<CheckImmOperand_s<7, "X86::COND_GE">>, [PdWriteCMOVm]>,
484  SchedVar<MCSchedPredicate<CheckImmOperand_s<7, "X86::COND_LE">>, [PdWriteCMOVm]>,
485  SchedVar<MCSchedPredicate<CheckImmOperand_s<7, "X86::COND_G">>,  [PdWriteCMOVm]>,
486  SchedVar<NoSchedPred, [WriteCMOV.Folded]>
487]>;
488
489def : InstRW<[PdWriteCMOVmVar], (instrs CMOV16rm, CMOV32rm, CMOV64rm)>;
490
491defm : PdWriteRes<WriteFCMOV,        [PdFPU0, PdFPFMA]>; // x87 conditional move.
492
493def : WriteRes<WriteSETCC,           [PdEX01]>; // Setcc.
494def : WriteRes<WriteSETCCStore,      [PdEX01, PdStore]>;
495
496def PdWriteSETGEmSETGmSETLEmSETLm : SchedWriteRes<[PdEX01]> {
497  let ReleaseAtCycles = [2];
498  let NumMicroOps = 2;
499}
500
501def PdSETGEmSETGmSETLEmSETLm :  SchedWriteVariant<[
502  SchedVar<MCSchedPredicate<CheckImmOperand_s<5, "X86::COND_GE">>, [PdWriteSETGEmSETGmSETLEmSETLm]>,
503  SchedVar<MCSchedPredicate<CheckImmOperand_s<5, "X86::COND_G">>,  [PdWriteSETGEmSETGmSETLEmSETLm]>,
504  SchedVar<MCSchedPredicate<CheckImmOperand_s<5, "X86::COND_LE">>, [PdWriteSETGEmSETGmSETLEmSETLm]>,
505  SchedVar<MCSchedPredicate<CheckImmOperand_s<5, "X86::COND_L">>,  [PdWriteSETGEmSETGmSETLEmSETLm]>,
506  SchedVar<NoSchedPred,                                            [WriteSETCCStore]>
507]>;
508def : InstRW<[PdSETGEmSETGmSETLEmSETLm], (instrs SETCCm)>;
509
510defm : PdWriteRes<WriteLAHFSAHF,      [PdEX01],          2,  [4],       2>;
511
512def PdWriteLAHF : SchedWriteRes<[PdEX01]> {
513  let Latency = 2;
514  let ReleaseAtCycles = [4];
515  let NumMicroOps = 4;
516}
517def : InstRW<[PdWriteLAHF], (instrs LAHF)>;
518
519def PdWriteSAHF : SchedWriteRes<[PdEX01]> {
520  let Latency = 2;
521  let ReleaseAtCycles = [2];
522  let NumMicroOps = 2;
523}
524def : InstRW<[PdWriteSAHF], (instrs SAHF)>;
525
526defm : PdWriteRes<WriteBitTest,          [PdEX01],         1, [2],      1>;
527defm : PdWriteRes<WriteBitTestImmLd,     [PdEX01, PdLoad], 5, [2,  3],  1>;
528defm : PdWriteRes<WriteBitTestRegLd,     [PdEX01, PdLoad], 5, [7,  2],  7>;
529defm : PdWriteRes<WriteBitTestSet,       [PdEX01],         2, [2],      2>;
530defm : PdWriteRes<WriteBitTestSetImmLd,  [PdEX01, PdLoad], 6, [1,  1],  4>;
531defm : PdWriteRes<WriteBitTestSetRegLd,  [PdEX01, PdLoad], 6, [1,  1], 10>;
532
533def PdWriteBTSIm : SchedWriteRes<[PdEX01, PdLoad]> {
534  let Latency = 7;
535  let ReleaseAtCycles = [42, 1];
536  let NumMicroOps = 4;
537}
538def : SchedAlias<WriteBitTestSetImmRMW, PdWriteBTSIm>;
539def PdWriteBTSRm : SchedWriteRes<[PdEX01, PdLoad]> {
540  let Latency = 7;
541  let ReleaseAtCycles = [44, 1];
542  let NumMicroOps = 10;
543}
544def : SchedAlias<WriteBitTestSetRegRMW, PdWriteBTSRm>;
545
546// This is for simple LEAs with one or two input operands.
547def : WriteRes<WriteLEA,              [PdEX01]> { let ReleaseAtCycles = [2]; }
548
549// This write is used for slow LEA instructions.
550def PdWrite3OpsLEA : SchedWriteRes<[PdEX01]> {
551  let Latency = 2;
552  let ReleaseAtCycles = [2];
553}
554
555// On Piledriver, a slow LEA is either a 3Ops LEA (base, index, offset),
556// or an LEA with a `Scale` value different than 1.
557def PdSlowLEAPredicate : MCSchedPredicate<
558  CheckAny<[
559    // A 3-operand LEA (base, index, offset).
560    IsThreeOperandsLEAFn,
561    // An LEA with a "Scale" different than 1.
562    CheckAll<[
563      CheckIsImmOperand<2>,
564      CheckNot<CheckImmOperand<2, 1>>
565    ]>
566  ]>
567>;
568
569def PdWriteLEA : SchedWriteVariant<[
570    SchedVar<PdSlowLEAPredicate, [PdWrite3OpsLEA]>,
571    SchedVar<NoSchedPred,        [WriteLEA]>
572]>;
573
574def : InstRW<[PdWriteLEA], (instrs LEA32r, LEA64r, LEA64_32r)>;
575
576def PdWriteLEA16r : SchedWriteRes<[PdEX01]> {
577  let ReleaseAtCycles = [3];
578  let NumMicroOps = 2;
579}
580def : InstRW<[PdWriteLEA16r], (instrs LEA16r)>;
581
582// Bit counts.
583defm : PdWriteResExPair<WriteBSF,     [PdEX01],          3,  [6],     6, 2>;
584defm : PdWriteResExPair<WriteBSR,     [PdEX01],          4,  [8],     7, 2>;
585defm : PdWriteResExPair<WritePOPCNT,  [PdEX01],          4,  [4]>;
586defm : PdWriteResExPair<WriteLZCNT,   [PdEX0],           2,  [2],     2>;
587defm : PdWriteResExPair<WriteTZCNT,   [PdEX0],           2,  [2],     2>;
588
589// BMI1 BEXTR, BMI2 BZHI
590defm : PdWriteResExPair<WriteBEXTR,   [PdEX01],          2,  [2],    2>;
591defm : PdWriteResExPair<WriteBLS,     [PdEX01],          2,  [2],    2>;
592defm : PdWriteResExPair<WriteBZHI,    [PdEX01]>;
593
594def PdWriteBEXTRI : SchedWriteRes<[PdEX01]> {
595  let Latency = 2;
596  let ReleaseAtCycles = [4];
597  let NumMicroOps = 2;
598}
599def : InstRW<[PdWriteBEXTRI], (instrs BEXTRI32ri, BEXTRI64ri)>;
600
601def PdWriteBEXTRIm : SchedWriteRes<[PdEX01]> {
602  let Latency = 2;
603  let ReleaseAtCycles = [5];
604  let NumMicroOps = 2;
605}
606def : InstRW<[PdWriteBEXTRIm], (instrs BEXTRI32mi, BEXTRI64mi)>;
607
608////////////////////////////////////////////////////////////////////////////////
609// Integer shifts and rotates.
610////////////////////////////////////////////////////////////////////////////////
611
612defm : PdWriteResExPair<WriteShift,    [PdEX01], 1, [2]>;
613defm : PdWriteResExPair<WriteShiftCL,  [PdEX01]>;
614defm : PdWriteResExPair<WriteRotate,   [PdEX01], 1, [2]>;
615defm : PdWriteResExPair<WriteRotateCL, [PdEX01]>;
616
617def PdWriteRCL8rCL : SchedWriteRes<[PdEX01]> {
618  let Latency = 12;
619  let ReleaseAtCycles = [24];
620  let NumMicroOps = 26;
621}
622def : InstRW<[PdWriteRCL8rCL], (instrs RCL8rCL)>;
623
624def PdWriteRCR8ri : SchedWriteRes<[PdEX01]> {
625  let Latency = 12;
626  let ReleaseAtCycles = [23];
627  let NumMicroOps = 23;
628}
629def : InstRW<[PdWriteRCR8ri], (instrs RCR8ri)>;
630
631def PdWriteRCR8rCL : SchedWriteRes<[PdEX01]> {
632  let Latency = 11;
633  let ReleaseAtCycles = [22];
634  let NumMicroOps = 24;
635}
636def : InstRW<[PdWriteRCR8rCL], (instrs RCR8rCL)>;
637
638def PdWriteRCL16rCL : SchedWriteRes<[PdEX01]> {
639  let Latency = 10;
640  let ReleaseAtCycles = [20];
641  let NumMicroOps = 22;
642}
643def : InstRW<[PdWriteRCL16rCL], (instrs RCL16rCL)>;
644
645def PdWriteRCR16ri : SchedWriteRes<[PdEX01]> {
646  let Latency = 10;
647  let ReleaseAtCycles = [19];
648  let NumMicroOps = 19;
649}
650def : InstRW<[PdWriteRCR16ri], (instrs RCR16ri)>;
651
652def PdWriteRCL3264rCL : SchedWriteRes<[PdEX01]> {
653  let Latency = 7;
654  let ReleaseAtCycles = [14];
655  let NumMicroOps = 17;
656}
657def : InstRW<[PdWriteRCL3264rCL], (instrs RCL32rCL, RCL64rCL)>;
658
659def PdWriteRCR3264rCL : SchedWriteRes<[PdEX01]> {
660  let Latency = 7;
661  let ReleaseAtCycles = [13];
662  let NumMicroOps = 16;
663}
664def : InstRW<[PdWriteRCR3264rCL], (instrs RCR32rCL, RCR64rCL)>;
665
666def PdWriteRCR32riRCR64ri : SchedWriteRes<[PdEX01]> {
667  let Latency = 7;
668  let ReleaseAtCycles = [14];
669  let NumMicroOps = 15;
670}
671def : InstRW<[PdWriteRCR32riRCR64ri], (instrs RCR32ri, RCR64ri)>;
672
673
674def PdWriteRCR16rCL : SchedWriteRes<[PdEX01]> {
675  let Latency = 9;
676  let ReleaseAtCycles = [18];
677  let NumMicroOps = 20;
678}
679def : InstRW<[PdWriteRCR16rCL], (instrs RCR16rCL)>;
680
681def PdWriteRCL16ri : SchedWriteRes<[PdEX01]> {
682  let Latency = 11;
683  let ReleaseAtCycles = [21];
684  let NumMicroOps = 21;
685}
686def : InstRW<[PdWriteRCL16ri], (instrs RCL16ri)>;
687
688def PdWriteRCL3264ri : SchedWriteRes<[PdEX01]> {
689  let Latency = 8;
690  let ReleaseAtCycles = [15];
691  let NumMicroOps = 16;
692}
693def : InstRW<[PdWriteRCL3264ri], (instrs RCL32ri, RCL64ri)>;
694
695def PdWriteRCL8ri : SchedWriteRes<[PdEX01]> {
696  let Latency = 13;
697  let ReleaseAtCycles = [25];
698  let NumMicroOps = 25;
699}
700def : InstRW<[PdWriteRCL8ri], (instrs RCL8ri)>;
701
702// SHLD/SHRD.
703defm : PdWriteRes<WriteSHDrri,       [PdEX01],         3, [6], 6>;
704defm : PdWriteRes<WriteSHDrrcl,      [PdEX01],         3, [8], 7>;
705
706def PdWriteSHLD16rrCLSHLD32rrCLSHRD32rrCL : SchedWriteRes<[PdEX01]> {
707  let Latency = 3;
708  let ReleaseAtCycles = [6];
709  let NumMicroOps = 7;
710}
711def : InstRW<[PdWriteSHLD16rrCLSHLD32rrCLSHRD32rrCL], (instrs SHLD16rrCL,
712                                                              SHLD32rrCL,
713                                                              SHRD32rrCL)>;
714
715defm : PdWriteRes<WriteSHDmri,       [PdLoad, PdEX01], 4, [1, 22], 8>;
716defm : PdWriteRes<WriteSHDmrcl,      [PdLoad, PdEX01], 4, [1, 22], 8>;
717
718////////////////////////////////////////////////////////////////////////////////
719// Floating point. This covers both scalar and vector operations.
720////////////////////////////////////////////////////////////////////////////////
721
722defm : PdWriteRes<WriteFLD0,               [PdFPU1, PdFPSTO], 3>;
723defm : PdWriteRes<WriteFLD1,               [PdFPU1, PdFPSTO], 3>;
724defm : PdWriteRes<WriteFLDC,               [PdFPU1, PdFPSTO], 3>;
725
726defm : PdWriteRes<WriteFLoad,              [PdLoad, PdFPU01, PdFPFMA], 5, [3, 1, 3]>;
727defm : PdWriteRes<WriteFLoadX,             [PdLoad, PdFPU01, PdFPFMA], 5, [3, 1, 3]>;
728defm : PdWriteRes<WriteFLoadY,             [PdLoad, PdFPU01, PdFPFMA], 5, [3, 1, 3], 2>;
729
730defm : PdWriteRes<WriteFMaskedLoad,        [PdLoad, PdFPU01, PdFPFMA], 6, [3, 1, 4]>;
731defm : PdWriteRes<WriteFMaskedLoadY,       [PdLoad, PdFPU01, PdFPFMA], 6, [3, 2, 4], 2>;
732
733defm : PdWriteRes<WriteFStore,             [PdStore, PdFPU23, PdFPSTO], 2, [1,  3, 1]>;
734defm : PdWriteRes<WriteFStoreX,            [PdStore, PdFPU23, PdFPSTO], 1, [1,  3, 1]>;
735defm : PdWriteRes<WriteFStoreY,            [PdStore, PdFPU23, PdFPSTO], 1, [1, 36, 2], 4>;
736
737def PdWriteMOVHPm : SchedWriteRes<[PdStore, PdFPU23,  PdFPSTO]> {
738  let Latency = 2;
739  let ReleaseAtCycles = [1, 3, 1];
740  let NumMicroOps = 2;
741}
742def : InstRW<[PdWriteMOVHPm], (instrs MOVHPDmr, MOVHPSmr, VMOVHPDmr, VMOVHPSmr)>;
743
744def PdWriteVMOVUPDYmrVMOVUPSYmr : SchedWriteRes<[PdStore, PdFPU1,  PdFPSTO]> {
745  let NumMicroOps = 8;
746}
747def : InstRW<[PdWriteVMOVUPDYmrVMOVUPSYmr], (instrs VMOVUPDYmr, VMOVUPSYmr)>;
748
749defm : PdWriteRes<WriteFStoreNT,           [PdStore, PdFPU1,  PdFPSTO], 3>;
750defm : PdWriteRes<WriteFStoreNTX,          [PdStore, PdFPU1,  PdFPSTO], 3>;
751defm : PdWriteRes<WriteFStoreNTY,          [PdStore, PdFPU1,  PdFPSTO], 3, [2, 2, 2], 4>;
752
753defm : PdWriteRes<WriteFMaskedStore32,     [PdStore, PdFPU01, PdFPFMA], 6, [1, 1, 188], 18>;
754defm : PdWriteRes<WriteFMaskedStore64,     [PdStore, PdFPU01, PdFPFMA], 6, [1, 1, 188], 18>;
755defm : PdWriteRes<WriteFMaskedStore32Y,    [PdStore, PdFPU01, PdFPFMA], 6, [2, 2, 376], 34>;
756defm : PdWriteRes<WriteFMaskedStore64Y,    [PdStore, PdFPU01, PdFPFMA], 6, [2, 2, 376], 34>;
757
758defm : PdWriteRes<WriteFMove,              [PdFPU01, PdFPFMA]>;
759defm : PdWriteRes<WriteFMoveX,             [PdFPU01, PdFPFMA], 1, [1, 2]>;
760defm : PdWriteRes<WriteFMoveY,             [PdFPU01, PdFPFMA], 2, [2, 2], 2>;
761defm : X86WriteResUnsupported<WriteFMoveZ>;
762
763defm : PdWriteRes<WriteEMMS,               [PdFPU01, PdFPFMA], 2>;
764
765defm : PdWriteResXMMPair<WriteFAdd,         [PdFPU0, PdFPFMA],  5>;
766defm : PdWriteResXMMPair<WriteFAddX,        [PdFPU0, PdFPFMA],  5>;
767defm : PdWriteResYMMPair<WriteFAddY,        [PdFPU0, PdFPFMA],  5, [1, 2]>;
768defm : X86WriteResPairUnsupported<WriteFAddZ>;
769
770def PdWriteX87Add: SchedWriteRes<[PdLoad, PdFPU0, PdFPFMA]> {
771  let Latency = 5;
772  let ReleaseAtCycles = [3, 1, 10];
773}
774def : InstRW<[PdWriteX87Add], (instrs ADD_FI16m,  ADD_FI32m,  ADD_F32m,  ADD_F64m,
775                                      SUB_FI16m,  SUB_FI32m,  SUB_F32m,  SUB_F64m,
776                                      SUBR_FI16m, SUBR_FI32m, SUBR_F32m, SUBR_F64m)>;
777
778defm : PdWriteResXMMPair<WriteFAdd64,       [PdFPU0, PdFPFMA],  5>;
779defm : PdWriteResXMMPair<WriteFAdd64X,      [PdFPU0, PdFPFMA],  5>;
780defm : PdWriteResYMMPair<WriteFAdd64Y,      [PdFPU0, PdFPFMA],  5, [1, 2]>;
781defm : X86WriteResPairUnsupported<WriteFAdd64Z>;
782
783defm : PdWriteResXMMPair<WriteFCmp,         [PdFPU0, PdFPFMA],  2>;
784defm : PdWriteResXMMPair<WriteFCmpX,        [PdFPU0, PdFPFMA],  2>;
785defm : PdWriteResYMMPair<WriteFCmpY,        [PdFPU0, PdFPFMA],  2, [1, 2]>;
786defm : X86WriteResPairUnsupported<WriteFCmpZ>;
787
788defm : PdWriteResXMMPair<WriteFCmp64,       [PdFPU0, PdFPFMA],  2>;
789defm : PdWriteResXMMPair<WriteFCmp64X,      [PdFPU0, PdFPFMA],  2>;
790defm : PdWriteResYMMPair<WriteFCmp64Y,      [PdFPU0, PdFPFMA],  2, [1, 2]>;
791defm : X86WriteResPairUnsupported<WriteFCmp64Z>;
792
793defm : PdWriteResXMMPair<WriteFCom,         [PdFPU0, PdFPFMA, PdEX0], 1, [], 2>;
794defm : PdWriteResXMMPair<WriteFComX,        [PdFPU0, PdFPFMA, PdEX0], 1, [], 2>;
795
796def PdWriteFCOMPm : SchedWriteRes<[PdFPU1, PdFPFMA]> {
797  let Latency = 6;
798}
799def : InstRW<[PdWriteFCOMPm], (instrs FCOM32m, FCOM64m, FCOMP32m, FCOMP64m)>;
800
801def PdWriteTST_F_UCOM_FPPr : SchedWriteRes<[PdFPU1, PdFPFMA]>;
802def : InstRW<[PdWriteTST_F_UCOM_FPPr], (instrs TST_F, UCOM_FPPr)>;
803
804defm : PdWriteResXMMPair<WriteFMul,         [PdFPU1, PdFPFMA],  5>;
805defm : PdWriteResXMMPair<WriteFMulX,        [PdFPU1, PdFPFMA],  5>;
806defm : PdWriteResYMMPair<WriteFMulY,        [PdFPU1, PdFPFMA],  5, [1, 2]>;
807defm : X86WriteResPairUnsupported<WriteFMulZ>;
808
809def PdWriteX87Mul: SchedWriteRes<[PdLoad, PdFPU1, PdFPFMA]> {
810  let Latency = 5;
811  let ReleaseAtCycles = [3, 1, 10];
812}
813def : InstRW<[PdWriteX87Mul], (instrs MUL_FI16m, MUL_FI32m, MUL_F32m, MUL_F64m)>;
814
815defm : PdWriteResXMMPair<WriteFMul64,       [PdFPU1, PdFPFMA],  5>;
816defm : PdWriteResXMMPair<WriteFMul64X,      [PdFPU1, PdFPFMA],  5>;
817defm : PdWriteResYMMPair<WriteFMul64Y,      [PdFPU1, PdFPFMA],  5, [1, 2]>;
818defm : X86WriteResPairUnsupported<WriteFMul64Z>;
819
820defm : PdWriteResXMMPair<WriteFMA,          [PdFPU, PdFPFMA], 5, [1, 3]>;
821defm : PdWriteResXMMPair<WriteFMAX,         [PdFPU, PdFPFMA], 5, [1, 3]>;
822defm : PdWriteResYMMPair<WriteFMAY,         [PdFPU, PdFPFMA], 5, [1, 3]>;
823defm : X86WriteResPairUnsupported<WriteFMAZ>;
824
825defm : PdWriteResXMMPair<WriteDPPD,         [PdFPU1, PdFPFMA], 15, [1, 10], 15, 2>;
826
827defm : PdWriteResXMMPair<WriteDPPS,         [PdFPU1, PdFPFMA], 25, [1, 14],  16, 2>;
828defm : PdWriteResYMMPair<WriteDPPSY,        [PdFPU1, PdFPFMA], 27, [2, 25], /*or 29*/ 25, 4>;
829
830def PdWriteVDPPSrri : SchedWriteRes<[PdFPU1, PdFPFMA]> {
831  let Latency = 27;
832  let ReleaseAtCycles = [1, 14];
833  let NumMicroOps = 17;
834}
835def : InstRW<[PdWriteVDPPSrri], (instrs VDPPSrri)>;
836
837defm : PdWriteResXMMPair<WriteFRcp,         [PdFPU1, PdFPFMA],  5>;
838defm : PdWriteResXMMPair<WriteFRcpX,        [PdFPU1, PdFPFMA],  5>;
839defm : PdWriteResYMMPair<WriteFRcpY,        [PdFPU1, PdFPFMA],  5, [2, 1]>;
840defm : X86WriteResPairUnsupported<WriteFRcpZ>;
841
842defm : PdWriteResXMMPair<WriteFRsqrt,       [PdFPU1, PdFPFMA],  5, [1, 2]>;
843defm : PdWriteResXMMPair<WriteFRsqrtX,      [PdFPU1, PdFPFMA],  5>;
844defm : PdWriteResYMMPair<WriteFRsqrtY,      [PdFPU1, PdFPFMA],  5, [2, 2]>;
845defm : X86WriteResPairUnsupported<WriteFRsqrtZ>;
846
847defm : PdWriteResXMMPair<WriteFDiv,         [PdFPU1, PdFPFMA], 9, [1, 9]>;
848defm : PdWriteResXMMPair<WriteFDivX,        [PdFPU1, PdFPFMA], 9, [1, 9]>;
849defm : PdWriteResYMMPair<WriteFDivY,        [PdFPU1, PdFPFMA], 9, [2, 18]>;
850defm : X86WriteResPairUnsupported<WriteFDivZ>;
851
852def PdWriteX87Div: SchedWriteRes<[PdLoad, PdFPU0, PdFPFMA]> {
853  let Latency = 9;
854  let ReleaseAtCycles = [3, 1, 18];
855}
856def : InstRW<[PdWriteX87Div], (instrs DIV_FI16m,  DIV_FI32m,
857                                      DIVR_FI16m, DIVR_FI32m,
858                                      DIV_F32m,   DIV_F64m,
859                                      DIVR_F32m,  DIVR_F64m)>;
860
861defm : PdWriteResXMMPair<WriteFDiv64,       [PdFPU1, PdFPFMA], 9, [1, 9]>;
862defm : PdWriteResXMMPair<WriteFDiv64X,      [PdFPU1, PdFPFMA], 9, [1, 9]>;
863defm : PdWriteResYMMPair<WriteFDiv64Y,      [PdFPU1, PdFPFMA], 9, [2, 18]>;
864defm : X86WriteResPairUnsupported<WriteFDiv64Z>;
865
866defm : PdWriteResXMMPair<WriteFSqrt,        [PdFPU1, PdFPFMA], 9, [1, 9]>;
867defm : PdWriteResXMMPair<WriteFSqrtX,       [PdFPU1, PdFPFMA], 9, [1, 9]>;
868defm : PdWriteResYMMPair<WriteFSqrtY,       [PdFPU1, PdFPFMA], 9, [2, 18]>;
869defm : X86WriteResPairUnsupported<WriteFSqrtZ>;
870
871defm : PdWriteResXMMPair<WriteFSqrt64,      [PdFPU1, PdFPFMA], 9, [1, 9]>;
872defm : PdWriteResXMMPair<WriteFSqrt64X,     [PdFPU1, PdFPFMA], 9, [1, 9]>;
873defm : PdWriteResYMMPair<WriteFSqrt64Y,     [PdFPU1, PdFPFMA], 9, [2, 18]>;
874defm : X86WriteResPairUnsupported<WriteFSqrt64Z>;
875
876defm : PdWriteResXMMPair<WriteFSqrt80,      [PdFPU1, PdFPFMA],  1, [1, 18]>;
877defm : PdWriteResXMMPair<WriteFSign,        [PdFPU1, PdFPFMA],  1, [1, 4]>;
878
879defm : PdWriteResXMMPair<WriteFRnd,         [PdFPU1, PdFPSTO],  4, []>;
880defm : PdWriteResYMMPair<WriteFRndY,        [PdFPU1, PdFPSTO],  4, [2, 1], 2>;
881defm : X86WriteResPairUnsupported<WriteFRndZ>;
882
883def PdWriteVFRCZP : SchedWriteRes<[PdFPU1, PdFPSTO]> {
884  let Latency = 10;
885  let ReleaseAtCycles = [2, 1];
886  let NumMicroOps = 2;
887}
888def : InstRW<[PdWriteVFRCZP], (instrs VFRCZPDrr, VFRCZPSrr)>;
889
890def PdWriteVFRCZS : SchedWriteRes<[PdFPU1, PdFPSTO]> {
891  let Latency = 10;
892  let ReleaseAtCycles = [10, 1];
893  let NumMicroOps = 2;
894}
895def : InstRW<[PdWriteVFRCZS], (instrs VFRCZSDrr, VFRCZSSrr)>;
896
897def PdWriteVFRCZm : SchedWriteRes<[PdFPU1, PdFPSTO]> {
898  let Latency = 15;
899  let ReleaseAtCycles = [2, 1];
900  let NumMicroOps = 3;
901}
902def : InstRW<[PdWriteVFRCZm], (instrs VFRCZPDrm, VFRCZPSrm,
903                                      VFRCZSDrm, VFRCZSSrm)>;
904
905def PdWriteVFRCZY : SchedWriteRes<[PdFPU1, PdFPSTO]> {
906  let Latency = 10;
907  let ReleaseAtCycles = [3, 1];
908  let NumMicroOps = 4;
909}
910def : InstRW<[PdWriteVFRCZY], (instrs VFRCZPSYrr, VFRCZPDYrr)>;
911
912def PdWriteVFRCZYm : SchedWriteRes<[PdFPU1, PdFPSTO]> {
913  let Latency = 15;
914  let ReleaseAtCycles = [4, 1];
915  let NumMicroOps = 8;
916}
917def : InstRW<[PdWriteVFRCZYm], (instrs VFRCZPSYrm, VFRCZPDYrm)>;
918
919defm : PdWriteResXMMPair<WriteFLogic,       [PdFPU23, PdFPMAL],  2>;
920defm : PdWriteResYMMPair<WriteFLogicY,      [PdFPU23, PdFPMAL],  2, [2, 2]>;
921defm : X86WriteResPairUnsupported<WriteFLogicZ>;
922
923defm : PdWriteResXMMPair<WriteFTest,        [PdFPU0, PdFPFMA, PdEX0],  1, [], 2>;
924defm : PdWriteResYMMPair<WriteFTestY,       [PdFPU01, PdFPFMA, PdEX0], 1, [4, 4, 1], 4, 2>;
925defm : X86WriteResPairUnsupported<WriteFTestZ>;
926
927defm : PdWriteResXMMPair<WriteFShuffle,     [PdFPU01, PdFPFMA],  2, [1, 2]>;
928defm : PdWriteResYMMPair<WriteFShuffleY,    [PdFPU01, PdFPFMA],  2, [2, 4], 2>;
929defm : X86WriteResPairUnsupported<WriteFShuffleZ>;
930
931def PdWriteVBROADCASTF128 : SchedWriteRes<[PdFPU01, PdFPFMA]> {
932  let Latency = 7;
933  let ReleaseAtCycles = [1, 3];
934  let NumMicroOps = 2;
935}
936def : InstRW<[PdWriteVBROADCASTF128], (instrs VBROADCASTF128rm)>;
937
938defm : PdWriteResXMMPair<WriteFVarShuffle,  [PdFPU1, PdFPXBR],  3>;
939defm : PdWriteResYMMPair<WriteFVarShuffleY, [PdFPU1, PdFPXBR],  3, [2, 2], 2>;
940defm : X86WriteResPairUnsupported<WriteFVarShuffleZ>;
941
942defm : PdWriteResXMMPair<WriteFBlend,       [PdFPU23, PdFPMAL],  2>;
943defm : PdWriteResYMMPair<WriteFBlendY,      [PdFPU23, PdFPMAL],  2, [2, 2], 2>;
944defm : X86WriteResPairUnsupported<WriteFBlendZ>;
945
946defm : PdWriteResXMMPair<WriteFVarBlend,    [PdFPU1, PdFPXBR],  2>;
947defm : PdWriteResYMMPair<WriteFVarBlendY,   [PdFPU1, PdFPXBR],  2, [2, 2], 2>;
948defm : X86WriteResPairUnsupported<WriteFVarBlendZ>;
949
950defm : PdWriteResXMMPair<WriteFShuffle256,  [PdFPU01, PdFPFMA],  2, [1, 3], 2>;
951defm : X86WriteResPairUnsupported<WriteFVarShuffle256>;
952
953def PdWriteVEXTRACTF128rr : SchedWriteRes<[PdFPU01, PdFPFMA]> {
954  let Latency = 2;
955  let ReleaseAtCycles = [1, 2];
956}
957def : InstRW<[PdWriteVEXTRACTF128rr], (instrs VEXTRACTF128rr)>;
958
959def PdWriteVEXTRACTF128mr : SchedWriteRes<[PdFPU01, PdFPFMA]> {
960  let Latency = 7;
961  let ReleaseAtCycles = [1, 4];
962  let NumMicroOps = 2;
963}
964def : InstRW<[PdWriteVEXTRACTF128mr], (instrs VEXTRACTF128mr)>;
965
966def PdWriteVPERM2F128rr : SchedWriteRes<[PdFPU01, PdFPFMA]> {
967  let Latency = 4;
968  let ReleaseAtCycles = [1, 6];
969  let NumMicroOps = 8;
970}
971def : InstRW<[PdWriteVPERM2F128rr], (instrs VPERM2F128rr)>;
972
973def PdWriteVPERM2F128rm : SchedWriteRes<[PdFPU01, PdFPFMA]> {
974  let Latency = 8; // 4 + 4
975  let ReleaseAtCycles = [1, 8];
976  let NumMicroOps = 10;
977}
978def : InstRW<[PdWriteVPERM2F128rm], (instrs VPERM2F128rm)>;
979
980////////////////////////////////////////////////////////////////////////////////
981// Conversions.
982////////////////////////////////////////////////////////////////////////////////
983
984defm : PdWriteResXMMPair<WriteCvtSS2I,   [PdFPU0, PdFPCVT, PdFPSTO, PdFPFMA, PdEX0], 13, [], 2>;
985
986defm : PdWriteResXMMPair<WriteCvtPS2I,   [PdFPU0, PdFPCVT, PdFPSTO], 4>;
987defm : PdWriteResYMMPair<WriteCvtPS2IY,  [PdFPU0, PdFPCVT, PdFPSTO], 4, [1, 2, 1]>;
988defm : X86WriteResPairUnsupported<WriteCvtPS2IZ>;
989
990defm : PdWriteResXMMPair<WriteCvtSD2I,   [PdFPU0, PdFPCVT, PdFPSTO, PdFPFMA, PdEX0], 13, [], 2>;
991
992defm : PdWriteResXMMPair<WriteCvtPD2I,   [PdFPU0, PdFPCVT, PdFPSTO],          8, [],        2>;
993defm : PdWriteResYMMPair<WriteCvtPD2IY,  [PdFPU0, PdFPCVT, PdFPSTO, PdFPFMA], 8, [1, 2, 1, 1], 4>;
994defm : X86WriteResPairUnsupported<WriteCvtPD2IZ>;
995
996def PdWriteMMX_CVTTPD2PIrr : SchedWriteRes<[PdFPU0, PdFPCVT, PdFPSTO]> {
997  let Latency = 6;
998  let NumMicroOps = 2;
999}
1000def : InstRW<[PdWriteMMX_CVTTPD2PIrr], (instrs MMX_CVTTPD2PIrr)>;
1001
1002// FIXME: f+3 ST, LD+STC latency
1003defm : PdWriteResXMMPair<WriteCvtI2SS,   [PdFPU0, PdFPCVT, PdFPSTO], 4, [], 2>;
1004// FIXME: .Folded version is one NumMicroOp *less*..
1005
1006defm : PdWriteResXMMPair<WriteCvtI2PS,   [PdFPU0, PdFPCVT, PdFPSTO], 4>;
1007defm : PdWriteResYMMPair<WriteCvtI2PSY,  [PdFPU0, PdFPCVT, PdFPSTO], 4, [1, 2, 1]>;
1008defm : X86WriteResPairUnsupported<WriteCvtI2PSZ>;
1009
1010defm : PdWriteResXMMPair<WriteCvtI2SD,   [PdFPU0, PdFPCVT, PdFPSTO], 4, [], 2>;
1011// FIXME: .Folded version is one NumMicroOp *less*..
1012
1013def PdWriteCVTSI642SDrr_CVTSI642SSrr_CVTSI2SDr_CVTSI2SSrr : SchedWriteRes<[PdFPU0, PdFPCVT, PdFPSTO]> {
1014  let Latency = 13;
1015  let ReleaseAtCycles = [1, 3, 1];
1016  let NumMicroOps = 2;
1017}
1018def : InstRW<[PdWriteCVTSI642SDrr_CVTSI642SSrr_CVTSI2SDr_CVTSI2SSrr], (instrs CVTSI642SDrr, CVTSI642SSrr, CVTSI2SDrr, CVTSI2SSrr)>;
1019
1020defm : PdWriteResXMMPair<WriteCvtI2PD,   [PdFPU0, PdFPCVT, PdFPSTO], 8, [],     2>;
1021defm : PdWriteResYMMPair<WriteCvtI2PDY,  [PdFPU0, PdFPCVT, PdFPSTO], 8, [1, 2, 1], 4, 1>;
1022defm : X86WriteResPairUnsupported<WriteCvtI2PDZ>;
1023
1024defm : PdWriteResXMMPair<WriteCvtSS2SD,  [PdFPU0, PdFPCVT, PdFPSTO], 4, [1, 2, 1]>;
1025
1026defm : PdWriteResXMMPair<WriteCvtPS2PD,  [PdFPU0, PdFPCVT, PdFPSTO], 8, [],     2>;
1027defm : PdWriteResYMMPair<WriteCvtPS2PDY, [PdFPU0, PdFPCVT, PdFPSTO], 8, [1, 2, 1], 4, 1>;
1028defm : X86WriteResPairUnsupported<WriteCvtPS2PDZ>;
1029
1030defm : PdWriteResXMMPair<WriteCvtSD2SS,  [PdFPU0, PdFPCVT, PdFPSTO], 4, [1, 2, 1]>;
1031
1032defm : PdWriteResXMMPair<WriteCvtPD2PS,  [PdFPU0, PdFPCVT, PdFPSTO],          8, [],        2>;
1033defm : PdWriteResYMMPair<WriteCvtPD2PSY, [PdFPU0, PdFPCVT, PdFPSTO, PdFPFMA], 8, [1, 2, 1, 1], 4>;
1034defm : X86WriteResPairUnsupported<WriteCvtPD2PSZ>;
1035
1036def PdWriteMMX_CVTPD2PIrrMMX_CVTPI2PDrr : SchedWriteRes<[PdFPU0, PdFPCVT, PdFPSTO]> {
1037  let Latency = 6;
1038  let NumMicroOps = 2;
1039}
1040def : InstRW<[PdWriteMMX_CVTPD2PIrrMMX_CVTPI2PDrr], (instrs MMX_CVTPD2PIrr,
1041                                                            MMX_CVTPI2PDrr)>;
1042
1043def PdWriteMMX_CVTPI2PSrr : SchedWriteRes<[PdFPU0, PdFPCVT, PdFPSTO]> {
1044  let Latency = 4;
1045  let NumMicroOps = 2;
1046}
1047def : InstRW<[PdWriteMMX_CVTPI2PSrr], (instrs MMX_CVTPI2PSrr)>;
1048
1049defm : PdWriteResXMMPair<WriteCvtPH2PS,  [PdFPU0, PdFPCVT, PdFPSTO], 8, [1, 2, 1], 2, 1>;
1050defm : PdWriteResYMMPair<WriteCvtPH2PSY, [PdFPU0, PdFPCVT, PdFPSTO], 8, [1, 2, 1], 4, 3>;
1051defm : X86WriteResPairUnsupported<WriteCvtPH2PSZ>;
1052
1053defm : PdWriteRes<WriteCvtPS2PH,        [PdFPU0, PdFPCVT, PdFPSTO],          8, [1, 2, 1],    2>;
1054defm : PdWriteRes<WriteCvtPS2PHY,       [PdFPU0, PdFPCVT, PdFPSTO, PdFPFMA], 8, [1, 2, 1, 1], 4>;
1055defm : X86WriteResUnsupported<WriteCvtPS2PHZ>;
1056
1057defm : PdWriteRes<WriteCvtPS2PHSt,      [PdFPU0, PdFPCVT, PdFPSTO, PdStore],          4, [1, 2, 1, 1],    3>;
1058defm : PdWriteRes<WriteCvtPS2PHYSt,     [PdFPU0, PdFPCVT, PdFPSTO, PdFPFMA, PdStore], 4, [1, 2, 1, 1, 1], 4>;
1059defm : X86WriteResUnsupported<WriteCvtPS2PHZSt>;
1060
1061////////////////////////////////////////////////////////////////////////////////
1062// Vector integer operations.
1063////////////////////////////////////////////////////////////////////////////////
1064
1065defm : PdWriteRes<WriteVecLoad,             [PdLoad, PdFPU01, PdFPMAL], 5, [3, 1, 3]>;
1066defm : PdWriteRes<WriteVecLoadX,            [PdLoad, PdFPU01, PdFPMAL], 5, [3, 1, 3]>;
1067defm : PdWriteRes<WriteVecLoadY,            [PdLoad, PdFPU01, PdFPMAL], 5, [3, 2, 3], 2>;
1068
1069defm : PdWriteRes<WriteVecLoadNT,           [PdLoad, PdFPU01, PdFPMAL], 5, [3, 1, 4]>;
1070defm : PdWriteRes<WriteVecLoadNTY,          [PdLoad, PdFPU01, PdFPMAL], 5, [3, 2, 4]>;
1071
1072defm : PdWriteRes<WriteVecMaskedLoad,       [PdLoad, PdFPU01, PdFPMAL], 6, [3, 1, 2]>;
1073defm : PdWriteRes<WriteVecMaskedLoadY,      [PdLoad, PdFPU01, PdFPMAL], 6, [3, 2, 4], 2>;
1074
1075defm : PdWriteRes<WriteVecStore,            [PdStore, PdFPU23, PdFPSTO], 2, [1, 3,  1]>;
1076defm : PdWriteRes<WriteVecStoreX,           [PdStore, PdFPU23, PdFPSTO], 1, [1, 3,  1]>;
1077defm : PdWriteRes<WriteVecStoreY,           [PdStore, PdFPU23, PdFPSTO], 1, [2, 36, 2], 4>;
1078
1079def PdWriteVMOVDQUYmr : SchedWriteRes<[PdStore, PdFPU1,   PdFPSTO]> {
1080  let NumMicroOps = 8;
1081}
1082def : InstRW<[PdWriteVMOVDQUYmr], (instrs VMOVDQUYmr)>;
1083
1084defm : PdWriteRes<WriteVecStoreNT,          [PdStore, PdFPU1,   PdFPSTO], 2>;
1085defm : PdWriteRes<WriteVecStoreNTY,         [PdStore, PdFPU1,   PdFPSTO], 2, [2, 2, 2], 4>;
1086
1087defm : X86WriteResUnsupported<WriteVecMaskedStore32>;
1088defm : X86WriteResUnsupported<WriteVecMaskedStore32Y>;
1089defm : X86WriteResUnsupported<WriteVecMaskedStore64>;
1090defm : X86WriteResUnsupported<WriteVecMaskedStore64Y>;
1091
1092defm : PdWriteRes<WriteVecMove,             [PdFPU01, PdFPMAL], 2>;
1093defm : PdWriteRes<WriteVecMoveX,            [PdFPU01, PdFPMAL], 1, [1, 2]>;
1094defm : PdWriteRes<WriteVecMoveY,            [PdFPU01, PdFPMAL], 2, [2, 2], 2>;
1095defm : X86WriteResUnsupported<WriteVecMoveZ>;
1096
1097def PdWriteMOVDQArr : SchedWriteRes<[PdFPU01, PdFPMAL]> {
1098}
1099def : InstRW<[PdWriteMOVDQArr], (instrs MOVDQArr)>;
1100
1101def PdWriteMOVQ2DQrr : SchedWriteRes<[PdFPU01, PdFPMAL]> {
1102  let Latency = 4;
1103}
1104def : InstRW<[PdWriteMOVQ2DQrr], (instrs MMX_MOVQ2DQrr)>;
1105
1106defm : PdWriteRes<WriteVecMoveToGpr,        [PdFPU0, PdFPFMA, PdEX0], 11>;
1107defm : PdWriteRes<WriteVecMoveFromGpr,      [PdFPU01, PdFPFMA], 11, [1, 2], 2>;
1108
1109defm : PdWriteResXMMPair<WriteVecALU,        [PdFPU23, PdFPMAL], 2>;
1110defm : PdWriteResXMMPair<WriteVecALUX,       [PdFPU23, PdFPMAL], 2>;
1111defm : X86WriteResPairUnsupported<WriteVecALUY>;
1112defm : X86WriteResPairUnsupported<WriteVecALUZ>;
1113
1114defm : PdWriteResXMMPair<WriteVecShift,      [PdFPU1, PdFPXBR], 3>;
1115defm : PdWriteResXMMPair<WriteVecShiftX,     [PdFPU1, PdFPXBR], 3>;
1116defm : X86WriteResPairUnsupported<WriteVecShiftY>;
1117defm : X86WriteResPairUnsupported<WriteVecShiftZ>;
1118
1119defm : PdWriteResXMMPair<WriteVecShiftImm,   [PdFPU1, PdFPXBR], 2>;
1120defm : PdWriteResXMMPair<WriteVecShiftImmX,  [PdFPU1, PdFPXBR], 2>;
1121defm : X86WriteResPairUnsupported<WriteVecShiftImmY>;
1122defm : X86WriteResPairUnsupported<WriteVecShiftImmZ>;
1123
1124defm : PdWriteResXMMPair<WriteVecIMul,       [PdFPU0, PdFPMMA], 4>;
1125defm : PdWriteResXMMPair<WriteVecIMulX,      [PdFPU0, PdFPMMA], 4>;
1126defm : X86WriteResPairUnsupported<WriteVecIMulY>;
1127defm : X86WriteResPairUnsupported<WriteVecIMulZ>;
1128
1129defm : PdWriteResXMMPair<WritePMULLD,        [PdFPU0, PdFPU01, PdFPMMA, PdFPMAL], 5, [2, 1, 2, 1]>;
1130defm : X86WriteResPairUnsupported<WritePMULLDY>;
1131defm : X86WriteResPairUnsupported<WritePMULLDZ>;
1132
1133def PdWriteVPMACS : SchedWriteRes<[PdFPU0, PdFPMMA, PdFPMAL]> {
1134  let Latency = 4;
1135}
1136def : InstRW<[PdWriteVPMACS], (instrs VPMACSDQHrr, VPMACSDQLrr, VPMACSSDQHrr,
1137                                      VPMACSSDQLrr)>;
1138
1139// FIXME: Investigate RR vs RM differences.
1140defm : PdWriteRes<WriteMPSAD,   [PdFPU0, PdFPMMA], 8, [1, 4], 8>;
1141defm : PdWriteRes<WriteMPSADLd, [PdFPU0, PdFPMMA, PdLoad], 14, [1, 4, 3],  8>;
1142defm : X86WriteResPairUnsupported<WriteMPSADY>;
1143defm : X86WriteResPairUnsupported<WriteMPSADZ>;
1144
1145defm : PdWriteResXMMPair<WritePSADBW,        [PdFPU01, PdFPMAL], 4, [1, 2], 2>;
1146defm : PdWriteResXMMPair<WritePSADBWX,       [PdFPU01, PdFPMAL], 4, [1, 2], 2>;
1147defm : X86WriteResPairUnsupported<WritePSADBWY>;
1148defm : X86WriteResPairUnsupported<WritePSADBWZ>;
1149
1150defm : PdWriteResXMMPair<WritePHMINPOS,      [PdFPU0,  PdFPMAL], 4, [], 2>;
1151
1152defm : PdWriteResXMMPair<WriteShuffle,       [PdFPU1, PdFPXBR], 2>;
1153defm : PdWriteResXMMPair<WriteShuffleX,      [PdFPU1, PdFPXBR], 2>;
1154defm : PdWriteResYMMPair<WriteShuffleY,      [PdFPU1, PdFPXBR], 2, [2, 2]>;
1155defm : X86WriteResPairUnsupported<WriteShuffleZ>;
1156
1157defm : PdWriteResXMMPair<WriteVarShuffle,    [PdFPU1, PdFPXBR], 3>;
1158defm : PdWriteResXMMPair<WriteVarShuffleX,   [PdFPU1, PdFPXBR], 3>;
1159defm : X86WriteResPairUnsupported<WriteVarShuffleY>;
1160defm : X86WriteResPairUnsupported<WriteVarShuffleZ>;
1161
1162def PdWriteVPPERM : SchedWriteRes<[PdFPU1, PdFPXBR]> {
1163  let Latency = 2;
1164  let ReleaseAtCycles = [1, 1];
1165}
1166def : InstRW<[PdWriteVPPERM], (instrs VPPERMrrr, VPPERMrrr_REV)>;
1167
1168def PdWriteVPPERMLd : SchedWriteRes<[PdFPU1, PdFPXBR, PdLoad]> {
1169  let Latency = 7;
1170  let ReleaseAtCycles = [1, 1, 3];
1171}
1172def : InstRW<[PdWriteVPPERMLd], (instrs VPPERMrrm, VPPERMrmr)>;
1173
1174defm : PdWriteResXMMPair<WriteBlend,         [PdFPU23, PdFPMAL], 2>;
1175defm : X86WriteResPairUnsupported<WriteBlendY>;
1176defm : X86WriteResPairUnsupported<WriteBlendZ>;
1177
1178defm : PdWriteResXMMPair<WriteVarBlend,      [PdFPU1, PdFPXBR], 2>;
1179defm : X86WriteResPairUnsupported<WriteVarBlendY>;
1180defm : X86WriteResPairUnsupported<WriteVarBlendZ>;
1181
1182defm : PdWriteResXMMPair<WriteVecLogic,      [PdFPU23, PdFPMAL], 2>;
1183defm : PdWriteResXMMPair<WriteVecLogicX,     [PdFPU23, PdFPMAL], 2>;
1184defm : X86WriteResPairUnsupported<WriteVecLogicY>;
1185defm : X86WriteResPairUnsupported<WriteVecLogicZ>;
1186
1187defm : PdWriteResXMMPair<WriteVecTest,       [PdFPU0, PdFPFMA, PdEX0],  1, [], 2>;
1188defm : PdWriteResYMMPair<WriteVecTestY,      [PdFPU01, PdFPFMA, PdEX0], 1, [2, 4, 1], 4, 2>;
1189defm : X86WriteResPairUnsupported<WriteVecTestZ>;
1190
1191defm : PdWriteResXMMPair<WriteShuffle256,    [PdFPU01, PdFPMAL]>;
1192defm : PdWriteResXMMPair<WriteVPMOV256,      [PdFPU01, PdFPMAL]>;
1193defm : PdWriteResXMMPair<WriteVarShuffle256, [PdFPU01, PdFPMAL]>;
1194
1195defm : PdWriteResXMMPair<WriteVarVecShift,   [PdFPU1, PdFPXBR], 3>;
1196defm : X86WriteResPairUnsupported<WriteVarVecShiftY>;
1197defm : X86WriteResPairUnsupported<WriteVarVecShiftZ>;
1198
1199////////////////////////////////////////////////////////////////////////////////
1200// Vector insert/extract operations.
1201////////////////////////////////////////////////////////////////////////////////
1202
1203defm : PdWriteRes<WriteVecInsert,    [PdFPU01, PdFPMAL], 2, [1, 3], 2>;
1204defm : PdWriteRes<WriteVecInsertLd,  [PdFPU01, PdFPMAL, PdLoad], 6, [1, 4, 3], 2>;
1205
1206defm : PdWriteRes<WriteVecExtract,   [PdFPU0, PdFPFMA, PdEX0], 12, [1, 3, 1], 2>;
1207defm : PdWriteRes<WriteVecExtractSt, [PdFPU1, PdFPSTO, PdStore], 13, [2, 1, 1], 2>;
1208
1209def PdWriteEXTRQ : SchedWriteRes<[PdFPU01, PdFPMAL]> {
1210  let Latency = 3;
1211  let ReleaseAtCycles = [1, 3];
1212}
1213def : InstRW<[PdWriteEXTRQ], (instrs EXTRQ, EXTRQI)>;
1214
1215////////////////////////////////////////////////////////////////////////////////
1216// SSE42 String instructions.
1217////////////////////////////////////////////////////////////////////////////////
1218
1219defm : PdWriteResXMMPair<WritePCmpIStrI, [PdFPU1, PdFPFMA, PdEX0], 11, [1, 6, 1], 7, 1>;
1220defm : PdWriteResXMMPair<WritePCmpIStrM, [PdFPU1, PdFPFMA, PdEX0],  7, [1, 8, 1], 7, 2>;
1221
1222defm : PdWriteResXMMPair<WritePCmpEStrI, [PdFPU1, PdStore, PdLoad, PdFPMAL, PdFPFMA, PdEX0], 14, [1, 10, 10, 10, 1, 1], 27, 1>;
1223defm : PdWriteResXMMPair<WritePCmpEStrM, [PdFPU1, PdStore, PdLoad, PdFPMAL, PdFPFMA, PdEX0], 10, [1, 10, 10, 10, 1, 1], 27, 1>;
1224
1225////////////////////////////////////////////////////////////////////////////////
1226// MOVMSK Instructions.
1227////////////////////////////////////////////////////////////////////////////////
1228
1229defm : PdWriteRes<WriteFMOVMSK, [PdFPU0, PdFPFMA, PdEX0],   12, [], 2>;
1230
1231defm : PdWriteRes<WriteVecMOVMSK, [PdFPU0, PdFPFMA, PdEX0], 12, [], 2>;
1232defm : X86WriteResUnsupported<WriteVecMOVMSKY>;
1233// defm : X86WriteResUnsupported<WriteVecMOVMSKZ>;
1234
1235defm : PdWriteRes<WriteMMXMOVMSK, [PdFPU0, PdFPFMA, PdEX0], 10, [], 2>;
1236
1237////////////////////////////////////////////////////////////////////////////////
1238// AES Instructions.
1239////////////////////////////////////////////////////////////////////////////////
1240
1241defm : PdWriteResXMMPair<WriteAESIMC,    [PdFPU0, PdFPMMA], 5>;
1242defm : PdWriteResXMMPair<WriteAESKeyGen, [PdFPU0, PdFPMMA], 5>;
1243defm : PdWriteResXMMPair<WriteAESDecEnc, [PdFPU0, PdFPMMA], 9, [], 2>;
1244
1245////////////////////////////////////////////////////////////////////////////////
1246// Horizontal add/sub  instructions.
1247////////////////////////////////////////////////////////////////////////////////
1248
1249defm : PdWriteResXMMPair<WriteFHAdd,  [PdFPU0, PdFPFMA], 11, [1, 5],     3, 1>;
1250defm : PdWriteResYMMPair<WriteFHAddY, [PdFPU0, PdFPFMA], 11, [1, 8], 8, 2>;
1251defm : X86WriteResPairUnsupported<WriteFHAddZ>;
1252
1253defm : PdWriteResXMMPair<WritePHAdd,  [PdFPU01, PdFPMAL], 5, [1, 4], 3, 1>;
1254defm : PdWriteResXMMPair<WritePHAddX, [PdFPU01, PdFPMAL], 2, [1, 2]>;
1255defm : X86WriteResPairUnsupported<WritePHAddY>;
1256defm : X86WriteResPairUnsupported<WritePHAddZ>;
1257
1258def : InstRW<[WritePHAdd], (instrs PHADDDrr, PHSUBDrr,
1259                                   PHADDWrr, PHSUBWrr,
1260                                   PHADDSWrr, PHSUBSWrr,
1261                                   VPHADDDrr, VPHSUBDrr,
1262                                   VPHADDWrr, VPHSUBWrr,
1263                                   VPHADDSWrr, VPHSUBSWrr)>;
1264
1265def : InstRW<[WritePHAdd.Folded], (instrs PHADDDrm, PHSUBDrm,
1266                                          PHADDWrm, PHSUBWrm,
1267                                          PHADDSWrm, PHSUBSWrm,
1268                                          VPHADDDrm, VPHSUBDrm,
1269                                          VPHADDWrm, VPHSUBWrm,
1270                                          VPHADDSWrm, VPHSUBSWrm)>;
1271
1272////////////////////////////////////////////////////////////////////////////////
1273// Carry-less multiplication instructions.
1274////////////////////////////////////////////////////////////////////////////////
1275
1276defm : PdWriteResXMMPair<WriteCLMul, [PdFPU0, PdFPMMA], 12, [1, 7], 5, 1>;
1277
1278def PdWriteVPCLMULQDQrr : SchedWriteRes<[PdFPU0, PdFPMMA]> {
1279  let Latency = 12;
1280  let ReleaseAtCycles = [1, 7];
1281  let NumMicroOps = 6;
1282}
1283def : InstRW<[PdWriteVPCLMULQDQrr], (instrs VPCLMULQDQrr)>;
1284
1285////////////////////////////////////////////////////////////////////////////////
1286// SSE4A instructions.
1287////////////////////////////////////////////////////////////////////////////////
1288
1289def PdWriteINSERTQ : SchedWriteRes<[PdFPU01, PdFPMAL]> {
1290  let Latency = 3;
1291  let ReleaseAtCycles = [1, 2];
1292}
1293def : InstRW<[PdWriteINSERTQ], (instrs INSERTQ)>;
1294
1295def PdWriteINSERTQI : SchedWriteRes<[PdFPU01, PdFPMAL]> {
1296  let Latency = 3;
1297  let ReleaseAtCycles = [1, 3];
1298}
1299def : InstRW<[PdWriteINSERTQI], (instrs INSERTQI)>;
1300
1301////////////////////////////////////////////////////////////////////////////////
1302// AVX instructions.
1303////////////////////////////////////////////////////////////////////////////////
1304
1305def PdWriteVBROADCASTYLd : SchedWriteRes<[PdLoad, PdFPU01, PdFPFMA]> {
1306  let Latency = 6;
1307  let ReleaseAtCycles = [1, 2, 4];
1308  let NumMicroOps = 2;
1309}
1310def : InstRW<[PdWriteVBROADCASTYLd, ReadAfterLd], (instrs VBROADCASTSDYrm,
1311                                                          VBROADCASTSSYrm)>;
1312
1313def PdWriteVZEROALL : SchedWriteRes<[]> {
1314  let Latency = 90;
1315  let NumMicroOps = 32;
1316}
1317def : InstRW<[PdWriteVZEROALL], (instrs VZEROALL)>;
1318
1319def PdWriteVZEROUPPER : SchedWriteRes<[]> {
1320  let Latency = 46;
1321  let NumMicroOps = 16;
1322}
1323def : InstRW<[PdWriteVZEROUPPER], (instrs VZEROUPPER)>;
1324
1325///////////////////////////////////////////////////////////////////////////////
1326//  SchedWriteVariant definitions.
1327///////////////////////////////////////////////////////////////////////////////
1328
1329def PdWriteZeroLatency : SchedWriteRes<[]> {
1330  let Latency = 0;
1331}
1332
1333def PdWriteZeroIdiom : SchedWriteVariant<[
1334  SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [PdWriteZeroLatency]>,
1335  SchedVar<MCSchedPredicate<TruePred>,           [WriteALU]>
1336]>;
1337def : InstRW<[PdWriteZeroIdiom], (instrs SUB32rr, SUB64rr,
1338                                         XOR32rr, XOR64rr)>;
1339
1340def PdWriteFZeroIdiom : SchedWriteVariant<[
1341  SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [PdWriteZeroLatency]>,
1342  SchedVar<MCSchedPredicate<TruePred>,           [WriteFLogic]>
1343]>;
1344def : InstRW<[PdWriteFZeroIdiom], (instrs XORPSrr,  VXORPSrr,
1345                                          XORPDrr,  VXORPDrr,
1346                                          ANDNPSrr, VANDNPSrr,
1347                                          ANDNPDrr, VANDNPDrr)>;
1348
1349// VXORPSYrr, VXORPDYrr, VANDNPSYrr, VANDNPDYrr "zero-idioms" have latency of 1.
1350
1351def PdWriteVZeroIdiomLogic : SchedWriteVariant<[
1352  SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [PdWriteZeroLatency]>,
1353  SchedVar<MCSchedPredicate<TruePred>,           [WriteVecLogic]>
1354]>;
1355def : InstRW<[PdWriteVZeroIdiomLogic], (instrs MMX_PXORrr, MMX_PANDNrr)>;
1356
1357def PdWriteVZeroIdiomLogicX : SchedWriteVariant<[
1358  SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [PdWriteZeroLatency]>,
1359  SchedVar<MCSchedPredicate<TruePred>,           [WriteVecLogicX]>
1360]>;
1361def : InstRW<[PdWriteVZeroIdiomLogicX], (instrs PXORrr,  VPXORrr,
1362                                                PANDNrr, VPANDNrr)>;
1363
1364def PdWriteVZeroIdiomALU : SchedWriteVariant<[
1365  SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [PdWriteZeroLatency]>,
1366  SchedVar<MCSchedPredicate<TruePred>,           [WriteVecALU]>
1367]>;
1368def : InstRW<[PdWriteVZeroIdiomALU], (instrs MMX_PSUBBrr,   MMX_PSUBDrr,
1369                                             MMX_PSUBQrr,   MMX_PSUBWrr,
1370                                             MMX_PCMPGTBrr,
1371                                             MMX_PCMPGTDrr,
1372                                             MMX_PCMPGTWrr)>;
1373
1374def PdWriteVZeroIdiomALUX : SchedWriteVariant<[
1375    SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [PdWriteZeroLatency]>,
1376    SchedVar<MCSchedPredicate<TruePred>,           [WriteVecALUX]>
1377]>;
1378def : InstRW<[PdWriteVZeroIdiomALUX], (instrs PSUBBrr, VPSUBBrr,
1379                                              PSUBDrr, VPSUBDrr,
1380                                              PSUBQrr, VPSUBQrr,
1381                                              PSUBWrr, VPSUBWrr,
1382                                              PCMPGTBrr, VPCMPGTBrr,
1383                                              PCMPGTDrr, VPCMPGTDrr,
1384                                              PCMPGTWrr, VPCMPGTWrr)>;
1385
1386///////////////////////////////////////////////////////////////////////////////
1387// Dependency breaking instructions.
1388///////////////////////////////////////////////////////////////////////////////
1389
1390// VPCMPGTQ, but not PCMPGTQ!
1391
1392def : IsZeroIdiomFunction<[
1393  // GPR Zero-idioms.
1394  DepBreakingClass<[ SUB32rr, SUB64rr, XOR32rr, XOR64rr ], ZeroIdiomPredicate>,
1395
1396  // MMX Zero-idioms.
1397  DepBreakingClass<[
1398    MMX_PXORrr, MMX_PANDNrr, MMX_PSUBBrr,
1399    MMX_PSUBDrr, MMX_PSUBQrr, MMX_PSUBWrr,
1400    MMX_PSUBSBrr, MMX_PSUBSWrr, MMX_PSUBUSBrr, MMX_PSUBUSWrr,
1401    MMX_PCMPGTBrr, MMX_PCMPGTDrr, MMX_PCMPGTWrr
1402  ], ZeroIdiomPredicate>,
1403
1404  // SSE Zero-idioms.
1405  DepBreakingClass<[
1406    // fp variants.
1407    XORPSrr, XORPDrr, ANDNPSrr, ANDNPDrr,
1408
1409    // int variants.
1410    PXORrr, PANDNrr,
1411    PSUBBrr, PSUBWrr, PSUBDrr, PSUBQrr,
1412    PSUBSBrr, PSUBSWrr, PSUBUSBrr, PSUBUSWrr,
1413    PCMPGTBrr, PCMPGTDrr, PCMPGTWrr
1414  ], ZeroIdiomPredicate>,
1415
1416  // AVX Zero-idioms.
1417  DepBreakingClass<[
1418    // xmm fp variants.
1419    VXORPSrr, VXORPDrr, VANDNPSrr, VANDNPDrr,
1420
1421    // xmm int variants.
1422    VPXORrr, VPANDNrr,
1423    VPSUBBrr, VPSUBWrr, VPSUBDrr, VPSUBQrr,
1424    VPSUBSBrr, VPSUBSWrr, VPSUBUSBrr, VPSUBUSWrr,
1425    VPCMPGTBrr, VPCMPGTWrr, VPCMPGTDrr, VPCMPGTQrr,
1426
1427    // ymm variants.
1428    VXORPSYrr, VXORPDYrr, VANDNPSYrr, VANDNPDYrr
1429  ], ZeroIdiomPredicate>
1430]>;
1431
1432def : IsDepBreakingFunction<[
1433  // GPR
1434  DepBreakingClass<[ SBB32rr, SBB64rr ], ZeroIdiomPredicate>,
1435  DepBreakingClass<[ CMP32rr, CMP64rr ], CheckSameRegOperand<0, 1> >,
1436
1437  // MMX
1438  DepBreakingClass<[
1439    MMX_PCMPEQBrr, MMX_PCMPEQDrr, MMX_PCMPEQWrr
1440  ], ZeroIdiomPredicate>,
1441
1442  // SSE
1443  DepBreakingClass<[
1444    PCMPEQBrr, PCMPEQWrr, PCMPEQDrr
1445    // But not PCMPEQQrr.
1446  ], ZeroIdiomPredicate>,
1447
1448  // AVX
1449  DepBreakingClass<[
1450    VPCMPEQBrr, VPCMPEQWrr, VPCMPEQDrr
1451    // But not VPCMPEQQrr.
1452  ], ZeroIdiomPredicate>
1453]>;
1454
1455
1456} // SchedModel
1457