1 //=- X86SchedHaswell.td - X86 Haswell Scheduling -------------*- tablegen -*-=//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the machine model for Haswell to support instruction
11 // scheduling and other instruction cost heuristics.
13 //===----------------------------------------------------------------------===//
15 def HaswellModel : SchedMachineModel {
16 // All x86 instructions are modeled as a single micro-op, and HW can decode 4
17 // instructions per cycle.
19 let MicroOpBufferSize = 192; // Based on the reorder buffer.
21 let MispredictPenalty = 16;
23 // Based on the LSD (loop-stream detector) queue size and benchmarking data.
24 let LoopMicroOpBufferSize = 50;
26 // FIXME: SSE4 and AVX are unimplemented. This flag is set to allow
27 // the scheduler to assign a default model to unrecognized opcodes.
28 let CompleteModel = 0;
31 let SchedModel = HaswellModel in {
33 // Haswell can issue micro-ops to 8 different ports in one cycle.
35 // Ports 0, 1, 5, and 6 handle all computation.
36 // Port 4 gets the data half of stores. Store data can be available later than
37 // the store address, but since we don't model the latency of stores, we can
39 // Ports 2 and 3 are identical. They handle loads and the address half of
40 // stores. Port 7 can handle address calculations.
41 def HWPort0 : ProcResource<1>;
42 def HWPort1 : ProcResource<1>;
43 def HWPort2 : ProcResource<1>;
44 def HWPort3 : ProcResource<1>;
45 def HWPort4 : ProcResource<1>;
46 def HWPort5 : ProcResource<1>;
47 def HWPort6 : ProcResource<1>;
48 def HWPort7 : ProcResource<1>;
50 // Many micro-ops are capable of issuing on multiple ports.
51 def HWPort23 : ProcResGroup<[HWPort2, HWPort3]>;
52 def HWPort237 : ProcResGroup<[HWPort2, HWPort3, HWPort7]>;
53 def HWPort05 : ProcResGroup<[HWPort0, HWPort5]>;
54 def HWPort06 : ProcResGroup<[HWPort0, HWPort6]>;
55 def HWPort15 : ProcResGroup<[HWPort1, HWPort5]>;
56 def HWPort16 : ProcResGroup<[HWPort1, HWPort6]>;
57 def HWPort015 : ProcResGroup<[HWPort0, HWPort1, HWPort5]>;
58 def HWPort056: ProcResGroup<[HWPort0, HWPort5, HWPort6]>;
59 def HWPort0156: ProcResGroup<[HWPort0, HWPort1, HWPort5, HWPort6]>;
61 // 60 Entry Unified Scheduler
62 def HWPortAny : ProcResGroup<[HWPort0, HWPort1, HWPort2, HWPort3, HWPort4,
63 HWPort5, HWPort6, HWPort7]> {
67 // Integer division issued on port 0.
68 def HWDivider : ProcResource<1>;
70 // Loads are 4 cycles, so ReadAfterLd registers needn't be available until 4
71 // cycles after the memory operand.
72 def : ReadAdvance<ReadAfterLd, 4>;
74 // Many SchedWrites are defined in pairs with and without a folded load.
75 // Instructions with folded loads are usually micro-fused, so they only appear
76 // as two micro-ops when queued in the reservation station.
77 // This multiclass defines the resource usage for variants with and without
79 multiclass HWWriteResPair<X86FoldableSchedWrite SchedRW,
80 ProcResourceKind ExePort,
82 // Register variant is using a single cycle on ExePort.
83 def : WriteRes<SchedRW, [ExePort]> { let Latency = Lat; }
85 // Memory variant also uses a cycle on port 2/3 and adds 4 cycles to the
87 def : WriteRes<SchedRW.Folded, [HWPort23, ExePort]> {
88 let Latency = !add(Lat, 4);
92 // A folded store needs a cycle on port 4 for the store data, but it does not
93 // need an extra port 2/3 cycle to recompute the address.
94 def : WriteRes<WriteRMW, [HWPort4]>;
98 def : WriteRes<WriteStore, [HWPort237, HWPort4]>;
99 def : WriteRes<WriteLoad, [HWPort23]> { let Latency = 4; }
100 def : WriteRes<WriteMove, [HWPort0156]>;
101 def : WriteRes<WriteZero, []>;
103 defm : HWWriteResPair<WriteALU, HWPort0156, 1>;
104 defm : HWWriteResPair<WriteIMul, HWPort1, 3>;
105 def : WriteRes<WriteIMulH, []> { let Latency = 3; }
106 defm : HWWriteResPair<WriteShift, HWPort06, 1>;
107 defm : HWWriteResPair<WriteJump, HWPort06, 1>;
109 // This is for simple LEAs with one or two input operands.
110 // The complex ones can only execute on port 1, and they require two cycles on
111 // the port to read all inputs. We don't model that.
112 def : WriteRes<WriteLEA, [HWPort15]>;
114 // This is quite rough, latency depends on the dividend.
115 def : WriteRes<WriteIDiv, [HWPort0, HWDivider]> {
117 let ResourceCycles = [1, 10];
119 def : WriteRes<WriteIDivLd, [HWPort23, HWPort0, HWDivider]> {
121 let ResourceCycles = [1, 1, 10];
124 // Scalar and vector floating point.
125 defm : HWWriteResPair<WriteFAdd, HWPort1, 3>;
126 defm : HWWriteResPair<WriteFMul, HWPort0, 5>;
127 defm : HWWriteResPair<WriteFDiv, HWPort0, 12>; // 10-14 cycles.
128 defm : HWWriteResPair<WriteFRcp, HWPort0, 5>;
129 defm : HWWriteResPair<WriteFSqrt, HWPort0, 15>;
130 defm : HWWriteResPair<WriteCvtF2I, HWPort1, 3>;
131 defm : HWWriteResPair<WriteCvtI2F, HWPort1, 4>;
132 defm : HWWriteResPair<WriteCvtF2F, HWPort1, 3>;
133 defm : HWWriteResPair<WriteFShuffle, HWPort5, 1>;
134 defm : HWWriteResPair<WriteFBlend, HWPort015, 1>;
135 defm : HWWriteResPair<WriteFShuffle256, HWPort5, 3>;
137 def : WriteRes<WriteFVarBlend, [HWPort5]> {
139 let ResourceCycles = [2];
141 def : WriteRes<WriteFVarBlendLd, [HWPort5, HWPort23]> {
143 let ResourceCycles = [2, 1];
146 // Vector integer operations.
147 defm : HWWriteResPair<WriteVecShift, HWPort0, 1>;
148 defm : HWWriteResPair<WriteVecLogic, HWPort015, 1>;
149 defm : HWWriteResPair<WriteVecALU, HWPort15, 1>;
150 defm : HWWriteResPair<WriteVecIMul, HWPort0, 5>;
151 defm : HWWriteResPair<WriteShuffle, HWPort5, 1>;
152 defm : HWWriteResPair<WriteBlend, HWPort15, 1>;
153 defm : HWWriteResPair<WriteShuffle256, HWPort5, 3>;
155 def : WriteRes<WriteVarBlend, [HWPort5]> {
157 let ResourceCycles = [2];
159 def : WriteRes<WriteVarBlendLd, [HWPort5, HWPort23]> {
161 let ResourceCycles = [2, 1];
164 def : WriteRes<WriteVarVecShift, [HWPort0, HWPort5]> {
166 let ResourceCycles = [2, 1];
168 def : WriteRes<WriteVarVecShiftLd, [HWPort0, HWPort5, HWPort23]> {
170 let ResourceCycles = [2, 1, 1];
173 def : WriteRes<WriteMPSAD, [HWPort0, HWPort5]> {
175 let ResourceCycles = [1, 2];
177 def : WriteRes<WriteMPSADLd, [HWPort23, HWPort0, HWPort5]> {
179 let ResourceCycles = [1, 1, 2];
182 // String instructions.
183 // Packed Compare Implicit Length Strings, Return Mask
184 def : WriteRes<WritePCmpIStrM, [HWPort0]> {
186 let ResourceCycles = [3];
188 def : WriteRes<WritePCmpIStrMLd, [HWPort0, HWPort23]> {
190 let ResourceCycles = [3, 1];
193 // Packed Compare Explicit Length Strings, Return Mask
194 def : WriteRes<WritePCmpEStrM, [HWPort0, HWPort16, HWPort5]> {
196 let ResourceCycles = [3, 2, 4];
198 def : WriteRes<WritePCmpEStrMLd, [HWPort05, HWPort16, HWPort23]> {
200 let ResourceCycles = [6, 2, 1];
203 // Packed Compare Implicit Length Strings, Return Index
204 def : WriteRes<WritePCmpIStrI, [HWPort0]> {
206 let ResourceCycles = [3];
208 def : WriteRes<WritePCmpIStrILd, [HWPort0, HWPort23]> {
210 let ResourceCycles = [3, 1];
213 // Packed Compare Explicit Length Strings, Return Index
214 def : WriteRes<WritePCmpEStrI, [HWPort05, HWPort16]> {
216 let ResourceCycles = [6, 2];
218 def : WriteRes<WritePCmpEStrILd, [HWPort0, HWPort16, HWPort5, HWPort23]> {
220 let ResourceCycles = [3, 2, 2, 1];
224 def : WriteRes<WriteAESDecEnc, [HWPort5]> {
226 let ResourceCycles = [1];
228 def : WriteRes<WriteAESDecEncLd, [HWPort5, HWPort23]> {
230 let ResourceCycles = [1, 1];
233 def : WriteRes<WriteAESIMC, [HWPort5]> {
235 let ResourceCycles = [2];
237 def : WriteRes<WriteAESIMCLd, [HWPort5, HWPort23]> {
239 let ResourceCycles = [2, 1];
242 def : WriteRes<WriteAESKeyGen, [HWPort0, HWPort5]> {
244 let ResourceCycles = [2, 8];
246 def : WriteRes<WriteAESKeyGenLd, [HWPort0, HWPort5, HWPort23]> {
248 let ResourceCycles = [2, 7, 1];
251 // Carry-less multiplication instructions.
252 def : WriteRes<WriteCLMul, [HWPort0, HWPort5]> {
254 let ResourceCycles = [2, 1];
256 def : WriteRes<WriteCLMulLd, [HWPort0, HWPort5, HWPort23]> {
258 let ResourceCycles = [2, 1, 1];
261 def : WriteRes<WriteSystem, [HWPort0156]> { let Latency = 100; }
262 def : WriteRes<WriteMicrocoded, [HWPort0156]> { let Latency = 100; }
263 def : WriteRes<WriteFence, [HWPort23, HWPort4]>;
264 def : WriteRes<WriteNop, []>;
266 //================ Exceptions ================//
268 //-- Specific Scheduling Models --//
269 def Write2P0156_Lat2 : SchedWriteRes<[HWPort0156]> {
271 let ResourceCycles = [2];
273 def Write2P0156_Lat2Ld : SchedWriteRes<[HWPort0156, HWPort23]> {
275 let ResourceCycles = [2, 1];
278 def Write2P237_P4 : SchedWriteRes<[HWPort237, HWPort4]> {
280 let ResourceCycles = [2, 1];
283 def WriteP06 : SchedWriteRes<[HWPort06]>;
285 def WriteP0156_2P237_P4 : SchedWriteRes<[HWPort0156, HWPort237, HWPort4]> {
287 let ResourceCycles = [1, 2, 1];
290 def Write2P0156_2P237_P4 : SchedWriteRes<[HWPort0156, HWPort237, HWPort4]> {
292 let ResourceCycles = [2, 2, 1];
295 def Write3P0156_2P237_P4 : SchedWriteRes<[HWPort0156, HWPort237, HWPort4]> {
297 let ResourceCycles = [3, 2, 1];
302 // - mm: 64 bit mmx register.
303 // - x = 128 bit xmm register.
304 // - (x)mm = mmx or xmm register.
305 // - y = 256 bit ymm register.
306 // - v = any vector register.
309 //=== Integer Instructions ===//
310 //-- Move instructions --//
314 def : InstRW<[WriteALULd], (instregex "MOV16rm")>;
318 def : InstRW<[WriteLoad], (instregex "MOV(S|Z)X32rm(8|16)")>;
322 def : InstRW<[Write2P0156_Lat2],
323 (instregex "CMOV(O|NO|B|AE|E|NE|BE|A|S|NS|P|NP|L|GE|LE|G)(16|32|64)rr")>;
325 def : InstRW<[Write2P0156_Lat2Ld, ReadAfterLd],
326 (instregex "CMOV(O|NO|B|AE|E|NE|BE|A|S|NS|P|NP|L|GE|LE|G)(16|32|64)rm")>;
330 def WriteXCHG : SchedWriteRes<[HWPort0156]> {
332 let ResourceCycles = [3];
335 def : InstRW<[WriteXCHG], (instregex "XCHG(8|16|32|64)rr", "XCHG(16|32|64)ar")>;
338 def WriteXCHGrm : SchedWriteRes<[]> {
342 def : InstRW<[WriteXCHGrm], (instregex "XCHG(8|16|32|64)rm")>;
345 def WriteXLAT : SchedWriteRes<[]> {
349 def : InstRW<[WriteXLAT], (instregex "XLAT")>;
353 def : InstRW<[Write2P237_P4], (instregex "PUSH(16|32)rmm")>;
356 def WritePushF : SchedWriteRes<[HWPort1, HWPort4, HWPort237, HWPort06]> {
359 def : InstRW<[WritePushF], (instregex "PUSHF(16|32)")>;
362 def WritePushA : SchedWriteRes<[]> {
363 let NumMicroOps = 19;
365 def : InstRW<[WritePushA], (instregex "PUSHA(16|32)")>;
369 def : InstRW<[Write2P237_P4], (instregex "POP(16|32)rmm")>;
372 def WritePopF : SchedWriteRes<[]> {
375 def : InstRW<[WritePopF], (instregex "POPF(16|32)")>;
378 def WritePopA : SchedWriteRes<[]> {
379 let NumMicroOps = 18;
381 def : InstRW<[WritePopA], (instregex "POPA(16|32)")>;
384 def : InstRW<[WriteP06], (instregex "(S|L)AHF")>;
388 def WriteBSwap32 : SchedWriteRes<[HWPort15]>;
389 def : InstRW<[WriteBSwap32], (instregex "BSWAP32r")>;
392 def WriteBSwap64 : SchedWriteRes<[HWPort06, HWPort15]> {
395 def : InstRW<[WriteBSwap64], (instregex "BSWAP64r")>;
398 // r16,m16 / r64,m64.
399 def : InstRW<[Write2P0156_Lat2Ld], (instregex "MOVBE(16|64)rm")>;
402 def WriteMoveBE32rm : SchedWriteRes<[HWPort15, HWPort23]> {
405 def : InstRW<[WriteMoveBE32rm], (instregex "MOVBE32rm")>;
408 def WriteMoveBE16mr : SchedWriteRes<[HWPort06, HWPort237, HWPort4]> {
411 def : InstRW<[WriteMoveBE16mr], (instregex "MOVBE16mr")>;
414 def WriteMoveBE32mr : SchedWriteRes<[HWPort15, HWPort237, HWPort4]> {
417 def : InstRW<[WriteMoveBE32mr], (instregex "MOVBE32mr")>;
420 def WriteMoveBE64mr : SchedWriteRes<[HWPort06, HWPort15, HWPort237, HWPort4]> {
423 def : InstRW<[WriteMoveBE64mr], (instregex "MOVBE64mr")>;
425 //-- Arithmetic instructions --//
429 def : InstRW<[Write2P0156_2P237_P4],
430 (instregex "(ADD|SUB)(8|16|32|64)m(r|i)",
431 "(ADD|SUB)(8|16|32|64)mi8", "(ADD|SUB)64mi32")>;
435 def : InstRW<[Write2P0156_Lat2], (instregex "(ADC|SBB)(8|16|32|64)r(r|i)",
436 "(ADC|SBB)(16|32|64)ri8",
438 "(ADC|SBB)(8|16|32|64)rr_REV")>;
441 def : InstRW<[Write2P0156_Lat2Ld, ReadAfterLd], (instregex "(ADC|SBB)(8|16|32|64)rm")>;
444 def : InstRW<[Write3P0156_2P237_P4],
445 (instregex "(ADC|SBB)(8|16|32|64)m(r|i)",
446 "(ADC|SBB)(16|32|64)mi8",
451 def : InstRW<[WriteP0156_2P237_P4],
452 (instregex "(INC|DEC|NOT|NEG)(8|16|32|64)m",
453 "(INC|DEC)64(16|32)m")>;
457 def WriteMul16 : SchedWriteRes<[HWPort1, HWPort0156]> {
461 def : InstRW<[WriteMul16], (instregex "IMUL16r", "MUL16r")>;
464 def WriteMul16Ld : SchedWriteRes<[HWPort1, HWPort0156, HWPort23]> {
468 def : InstRW<[WriteMul16Ld], (instregex "IMUL16m", "MUL16m")>;
471 def WriteMul32 : SchedWriteRes<[HWPort1, HWPort0156]> {
475 def : InstRW<[WriteMul32], (instregex "IMUL32r", "MUL32r")>;
478 def WriteMul32Ld : SchedWriteRes<[HWPort1, HWPort0156, HWPort23]> {
482 def : InstRW<[WriteMul32Ld], (instregex "IMUL32m", "MUL32m")>;
485 def WriteMul64 : SchedWriteRes<[HWPort1, HWPort6]> {
489 def : InstRW<[WriteMul64], (instregex "IMUL64r", "MUL64r")>;
492 def WriteMul64Ld : SchedWriteRes<[HWPort1, HWPort6, HWPort23]> {
496 def : InstRW<[WriteMul64Ld], (instregex "IMUL64m", "MUL64m")>;
499 def WriteMul16rri : SchedWriteRes<[HWPort1, HWPort0156]> {
503 def : InstRW<[WriteMul16rri], (instregex "IMUL16rri", "IMUL16rri8")>;
506 def WriteMul16rmi : SchedWriteRes<[HWPort1, HWPort0156, HWPort23]> {
510 def : InstRW<[WriteMul16rmi], (instregex "IMUL16rmi", "IMUL16rmi8")>;
514 def WriteMulX32 : SchedWriteRes<[HWPort1, HWPort056]> {
517 let ResourceCycles = [1, 2];
519 def : InstRW<[WriteMulX32], (instregex "MULX32rr")>;
522 def WriteMulX32Ld : SchedWriteRes<[HWPort1, HWPort056, HWPort23]> {
525 let ResourceCycles = [1, 2, 1];
527 def : InstRW<[WriteMulX32Ld], (instregex "MULX32rm")>;
530 def WriteMulX64 : SchedWriteRes<[HWPort1, HWPort6]> {
534 def : InstRW<[WriteMulX64], (instregex "MULX64rr")>;
537 def WriteMulX64Ld : SchedWriteRes<[HWPort1, HWPort6, HWPort23]> {
541 def : InstRW<[WriteMulX64Ld], (instregex "MULX64rm")>;
545 def WriteDiv8 : SchedWriteRes<[HWPort0, HWPort1, HWPort5, HWPort6]> {
549 def : InstRW<[WriteDiv8], (instregex "DIV8r")>;
552 def WriteDiv16 : SchedWriteRes<[HWPort0, HWPort1, HWPort5, HWPort6]> {
554 let NumMicroOps = 10;
556 def : InstRW<[WriteDiv16], (instregex "DIV16r")>;
559 def WriteDiv32 : SchedWriteRes<[HWPort0, HWPort1, HWPort5, HWPort6]> {
561 let NumMicroOps = 10;
563 def : InstRW<[WriteDiv32], (instregex "DIV32r")>;
566 def WriteDiv64 : SchedWriteRes<[HWPort0, HWPort1, HWPort5, HWPort6]> {
568 let NumMicroOps = 36;
570 def : InstRW<[WriteDiv64], (instregex "DIV64r")>;
574 def WriteIDiv8 : SchedWriteRes<[HWPort0, HWPort1, HWPort5, HWPort6]> {
578 def : InstRW<[WriteIDiv8], (instregex "IDIV8r")>;
581 def WriteIDiv16 : SchedWriteRes<[HWPort0, HWPort1, HWPort5, HWPort6]> {
583 let NumMicroOps = 10;
585 def : InstRW<[WriteIDiv16], (instregex "IDIV16r")>;
588 def WriteIDiv32 : SchedWriteRes<[HWPort0, HWPort1, HWPort5, HWPort6]> {
592 def : InstRW<[WriteIDiv32], (instregex "IDIV32r")>;
595 def WriteIDiv64 : SchedWriteRes<[HWPort0, HWPort1, HWPort5, HWPort6]> {
597 let NumMicroOps = 59;
599 def : InstRW<[WriteIDiv64], (instregex "IDIV64r")>;