1 //===-- PPCScheduleP7.td - PPC P7 Scheduling Definitions ---*- 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 itinerary class data for the POWER7 processor.
12 //===----------------------------------------------------------------------===//
15 // IBM POWER7 multicore server processor
16 // B. Sinharoy, et al.
17 // IBM J. Res. & Dev. (55) 3. May/June 2011.
19 // Scheduling for the P7 involves tracking two types of resources:
20 // 1. The dispatch bundle slots
21 // 2. The functional unit resources
24 def P7_DU1 : FuncUnit;
25 def P7_DU2 : FuncUnit;
26 def P7_DU3 : FuncUnit;
27 def P7_DU4 : FuncUnit;
28 def P7_DU5 : FuncUnit;
29 def P7_DU6 : FuncUnit;
31 def P7_LS1 : FuncUnit; // Load/Store pipeline 1
32 def P7_LS2 : FuncUnit; // Load/Store pipeline 2
34 def P7_FX1 : FuncUnit; // FX pipeline 1
35 def P7_FX2 : FuncUnit; // FX pipeline 2
37 // VS pipeline 1 (vector integer ops. always here)
38 def P7_VS1 : FuncUnit; // VS pipeline 1
39 // VS pipeline 2 (128-bit stores and perms. here)
40 def P7_VS2 : FuncUnit; // VS pipeline 2
42 def P7_CRU : FuncUnit; // CR unit (CR logicals and move-from-SPRs)
43 def P7_BRU : FuncUnit; // BR unit
46 // Each LSU pipeline can also execute FX add and logical instructions.
47 // Each LSU pipeline can complete a load or store in one cycle.
49 // Each store is broken into two parts, AGEN goes to the LSU while a
50 // "data steering" op. goes to the FXU or VSU.
52 // FX loads have a two cycle load-to-use latency (so one "bubble" cycle).
53 // VSU loads have a three cycle load-to-use latency (so two "bubble" cycle).
55 // Frequent FX ops. take only one cycle and results can be used again in the
56 // next cycle (there is a self-bypass). Getting results from the other FX
57 // pipeline takes an additional cycle.
59 // The VSU XS is similar to the POWER6, but with a pipeline length of 2 cycles
60 // (instead of 3 cycles on the POWER6). VSU XS handles vector FX-style ops.
61 // Dispatch of an instruction to VS1 that uses four single prec. inputs
62 // (either to a float or XC op). prevents dispatch in that cycle to VS2 of any
63 // floating point instruction.
65 // The VSU PM is similar to the POWER6, but with a pipeline length of 3 cycles
66 // (instead of 4 cycles on the POWER6). vsel is handled by the PM pipeline
67 // (unlike on the POWER6).
69 // FMA from the VSUs can forward results in 6 cycles. VS1 XS and vector FP
70 // share the same write-back, and have a 5-cycle latency difference, so the
71 // IFU/IDU will not dispatch an XS instructon 5 cycles after a vector FP
72 // op. has been dispatched to VS1.
74 // Three cycles after an L1 cache hit, a dependent VSU instruction can issue.
76 // Instruction dispatch groups have (at most) four non-branch instructions, and
77 // two branches. Unlike on the POWER4/5, a branch does not automatically
78 // end the dispatch group, but a second branch must be the last in the group.
80 def P7Itineraries : ProcessorItineraries<
81 [P7_DU1, P7_DU2, P7_DU3, P7_DU4, P7_DU5, P7_DU6,
82 P7_LS1, P7_LS2, P7_FX1, P7_FX2, P7_VS1, P7_VS2, P7_CRU, P7_BRU], [], [
83 InstrItinData<IIC_IntSimple , [InstrStage<1, [P7_DU1, P7_DU2,
85 InstrStage<1, [P7_FX1, P7_FX2,
88 InstrItinData<IIC_IntGeneral , [InstrStage<1, [P7_DU1, P7_DU2,
90 InstrStage<1, [P7_FX1, P7_FX2]>],
92 InstrItinData<IIC_IntISEL, [InstrStage<1, [P7_DU1], 0>,
93 InstrStage<1, [P7_FX1, P7_FX2], 0>,
94 InstrStage<1, [P7_BRU]>],
96 InstrItinData<IIC_IntCompare , [InstrStage<1, [P7_DU1, P7_DU2,
98 InstrStage<1, [P7_FX1, P7_FX2]>],
100 // FIXME: Add record-form itinerary data.
101 InstrItinData<IIC_IntDivW , [InstrStage<1, [P7_DU1], 0>,
102 InstrStage<1, [P7_DU2], 0>,
103 InstrStage<36, [P7_FX1, P7_FX2]>],
105 InstrItinData<IIC_IntDivD , [InstrStage<1, [P7_DU1], 0>,
106 InstrStage<1, [P7_DU2], 0>,
107 InstrStage<68, [P7_FX1, P7_FX2]>],
109 InstrItinData<IIC_IntMulHW , [InstrStage<1, [P7_DU1, P7_DU2,
111 InstrStage<1, [P7_FX1, P7_FX2]>],
113 InstrItinData<IIC_IntMulHWU , [InstrStage<1, [P7_DU1, P7_DU2,
115 InstrStage<1, [P7_FX1, P7_FX2]>],
117 InstrItinData<IIC_IntMulLI , [InstrStage<1, [P7_DU1, P7_DU2,
119 InstrStage<1, [P7_FX1, P7_FX2]>],
121 InstrItinData<IIC_IntRotate , [InstrStage<1, [P7_DU1, P7_DU2,
123 InstrStage<1, [P7_FX1, P7_FX2]>],
125 InstrItinData<IIC_IntRotateD , [InstrStage<1, [P7_DU1, P7_DU2,
127 InstrStage<1, [P7_FX1, P7_FX2]>],
129 InstrItinData<IIC_IntShift , [InstrStage<1, [P7_DU1, P7_DU2,
131 InstrStage<1, [P7_FX1, P7_FX2]>],
133 InstrItinData<IIC_IntTrapW , [InstrStage<1, [P7_DU1, P7_DU2,
135 InstrStage<1, [P7_FX1, P7_FX2]>],
137 InstrItinData<IIC_IntTrapD , [InstrStage<1, [P7_DU1, P7_DU2,
139 InstrStage<1, [P7_FX1, P7_FX2]>],
141 InstrItinData<IIC_BrB , [InstrStage<1, [P7_DU5, P7_DU6], 0>,
142 InstrStage<1, [P7_BRU]>],
144 InstrItinData<IIC_BrCR , [InstrStage<1, [P7_DU1], 0>,
145 InstrStage<1, [P7_CRU]>],
147 InstrItinData<IIC_BrMCR , [InstrStage<1, [P7_DU5, P7_DU6], 0>,
148 InstrStage<1, [P7_BRU]>],
150 InstrItinData<IIC_BrMCRX , [InstrStage<1, [P7_DU5, P7_DU6], 0>,
151 InstrStage<1, [P7_BRU]>],
153 InstrItinData<IIC_LdStLoad , [InstrStage<1, [P7_DU1, P7_DU2,
155 InstrStage<1, [P7_LS1, P7_LS2]>],
157 InstrItinData<IIC_LdStLoadUpd , [InstrStage<1, [P7_DU1], 0>,
158 InstrStage<1, [P7_DU2], 0>,
159 InstrStage<1, [P7_LS1, P7_LS2], 0>,
160 InstrStage<1, [P7_FX1, P7_FX2]>],
162 InstrItinData<IIC_LdStLoadUpdX, [InstrStage<1, [P7_DU1], 0>,
163 InstrStage<1, [P7_DU2], 0>,
164 InstrStage<1, [P7_DU3], 0>,
165 InstrStage<1, [P7_DU4], 0>,
166 InstrStage<1, [P7_FX1, P7_FX2]>,
167 InstrStage<1, [P7_LS1, P7_LS2], 0>,
168 InstrStage<1, [P7_FX1, P7_FX2]>],
170 InstrItinData<IIC_LdStLD , [InstrStage<1, [P7_DU1, P7_DU2,
172 InstrStage<1, [P7_LS1, P7_LS2]>],
174 InstrItinData<IIC_LdStLDU , [InstrStage<1, [P7_DU1], 0>,
175 InstrStage<1, [P7_DU2], 0>,
176 InstrStage<1, [P7_LS1, P7_LS2], 0>,
177 InstrStage<1, [P7_FX1, P7_FX2]>],
179 InstrItinData<IIC_LdStLDUX , [InstrStage<1, [P7_DU1], 0>,
180 InstrStage<1, [P7_DU2], 0>,
181 InstrStage<1, [P7_DU3], 0>,
182 InstrStage<1, [P7_DU4], 0>,
183 InstrStage<1, [P7_FX1, P7_FX2]>,
184 InstrStage<1, [P7_LS1, P7_LS2], 0>,
185 InstrStage<1, [P7_FX1, P7_FX2]>],
187 InstrItinData<IIC_LdStLFD , [InstrStage<1, [P7_DU1, P7_DU2,
189 InstrStage<1, [P7_LS1, P7_LS2]>],
191 InstrItinData<IIC_LdStLVecX , [InstrStage<1, [P7_DU1, P7_DU2,
193 InstrStage<1, [P7_LS1, P7_LS2]>],
195 InstrItinData<IIC_LdStLFDU , [InstrStage<1, [P7_DU1], 0>,
196 InstrStage<1, [P7_DU2], 0>,
197 InstrStage<1, [P7_LS1, P7_LS2], 0>,
198 InstrStage<1, [P7_FX1, P7_FX2]>],
200 InstrItinData<IIC_LdStLFDUX , [InstrStage<1, [P7_DU1], 0>,
201 InstrStage<1, [P7_DU2], 0>,
202 InstrStage<1, [P7_LS1, P7_LS2], 0>,
203 InstrStage<1, [P7_FX1, P7_FX2]>],
205 InstrItinData<IIC_LdStLHA , [InstrStage<1, [P7_DU1], 0>,
206 InstrStage<1, [P7_DU2], 0>,
207 InstrStage<1, [P7_LS1, P7_LS2]>,
208 InstrStage<1, [P7_FX1, P7_FX2]>],
210 InstrItinData<IIC_LdStLHAU , [InstrStage<1, [P7_DU1], 0>,
211 InstrStage<1, [P7_DU2], 0>,
212 InstrStage<1, [P7_LS1, P7_LS2], 0>,
213 InstrStage<1, [P7_FX1, P7_FX2]>,
214 InstrStage<1, [P7_FX1, P7_FX2]>],
216 InstrItinData<IIC_LdStLHAUX , [InstrStage<1, [P7_DU1], 0>,
217 InstrStage<1, [P7_DU2], 0>,
218 InstrStage<1, [P7_DU3], 0>,
219 InstrStage<1, [P7_DU4], 0>,
220 InstrStage<1, [P7_FX1, P7_FX2]>,
221 InstrStage<1, [P7_LS1, P7_LS2], 0>,
222 InstrStage<1, [P7_FX1, P7_FX2]>,
223 InstrStage<1, [P7_FX1, P7_FX2]>],
225 InstrItinData<IIC_LdStLWA , [InstrStage<1, [P7_DU1], 0>,
226 InstrStage<1, [P7_DU2], 0>,
227 InstrStage<1, [P7_LS1, P7_LS2]>,
228 InstrStage<1, [P7_FX1, P7_FX2]>],
230 InstrItinData<IIC_LdStLWARX, [InstrStage<1, [P7_DU1], 0>,
231 InstrStage<1, [P7_DU2], 0>,
232 InstrStage<1, [P7_DU3], 0>,
233 InstrStage<1, [P7_DU4], 0>,
234 InstrStage<1, [P7_LS1, P7_LS2]>],
236 InstrItinData<IIC_LdStLDARX, [InstrStage<1, [P7_DU1], 0>,
237 InstrStage<1, [P7_DU2], 0>,
238 InstrStage<1, [P7_DU3], 0>,
239 InstrStage<1, [P7_DU4], 0>,
240 InstrStage<1, [P7_LS1, P7_LS2]>],
242 InstrItinData<IIC_LdStLMW , [InstrStage<1, [P7_DU1, P7_DU2,
244 InstrStage<1, [P7_LS1, P7_LS2]>],
246 InstrItinData<IIC_LdStStore , [InstrStage<1, [P7_DU1, P7_DU2,
248 InstrStage<1, [P7_LS1, P7_LS2], 0>,
249 InstrStage<1, [P7_FX1, P7_FX2]>],
251 InstrItinData<IIC_LdStSTD , [InstrStage<1, [P7_DU1, P7_DU2,
253 InstrStage<1, [P7_LS1, P7_LS2], 0>,
254 InstrStage<1, [P7_FX1, P7_FX2]>],
256 InstrItinData<IIC_LdStSTDU , [InstrStage<1, [P7_DU1], 0>,
257 InstrStage<1, [P7_DU2], 0>,
258 InstrStage<1, [P7_LS1, P7_LS2], 0>,
259 InstrStage<1, [P7_FX1, P7_FX2]>,
260 InstrStage<1, [P7_FX1, P7_FX2]>],
262 InstrItinData<IIC_LdStSTDUX , [InstrStage<1, [P7_DU1], 0>,
263 InstrStage<1, [P7_DU2], 0>,
264 InstrStage<1, [P7_DU3], 0>,
265 InstrStage<1, [P7_DU4], 0>,
266 InstrStage<1, [P7_LS1, P7_LS2], 0>,
267 InstrStage<1, [P7_FX1, P7_FX2]>,
268 InstrStage<1, [P7_FX1, P7_FX2]>],
270 InstrItinData<IIC_LdStSTFD , [InstrStage<1, [P7_DU1, P7_DU2,
272 InstrStage<1, [P7_LS1, P7_LS2], 0>,
273 InstrStage<1, [P7_VS1, P7_VS2]>],
275 InstrItinData<IIC_LdStSTFDU , [InstrStage<1, [P7_DU1], 0>,
276 InstrStage<1, [P7_DU2], 0>,
277 InstrStage<1, [P7_LS1, P7_LS2], 0>,
278 InstrStage<1, [P7_FX1, P7_FX2], 0>,
279 InstrStage<1, [P7_VS1, P7_VS2]>],
281 InstrItinData<IIC_LdStSTVEBX , [InstrStage<1, [P7_DU1, P7_DU2,
283 InstrStage<1, [P7_LS1, P7_LS2], 0>,
284 InstrStage<1, [P7_VS2]>],
286 InstrItinData<IIC_LdStSTDCX , [InstrStage<1, [P7_DU1], 0>,
287 InstrStage<1, [P7_DU2], 0>,
288 InstrStage<1, [P7_DU3], 0>,
289 InstrStage<1, [P7_DU4], 0>,
290 InstrStage<1, [P7_LS1, P7_LS2]>],
292 InstrItinData<IIC_LdStSTWCX , [InstrStage<1, [P7_DU1], 0>,
293 InstrStage<1, [P7_DU2], 0>,
294 InstrStage<1, [P7_DU3], 0>,
295 InstrStage<1, [P7_DU4], 0>,
296 InstrStage<1, [P7_LS1, P7_LS2]>],
298 InstrItinData<IIC_BrMCRX , [InstrStage<1, [P7_DU1], 0>,
299 InstrStage<1, [P7_DU2], 0>,
300 InstrStage<1, [P7_DU3], 0>,
301 InstrStage<1, [P7_DU4], 0>,
302 InstrStage<1, [P7_CRU]>,
303 InstrStage<1, [P7_FX1, P7_FX2]>],
305 InstrItinData<IIC_SprMFCR , [InstrStage<1, [P7_DU1], 0>,
306 InstrStage<1, [P7_CRU]>],
308 InstrItinData<IIC_SprMFCRF , [InstrStage<1, [P7_DU1], 0>,
309 InstrStage<1, [P7_CRU]>],
311 InstrItinData<IIC_SprMTSPR , [InstrStage<1, [P7_DU1], 0>,
312 InstrStage<1, [P7_FX1]>],
314 InstrItinData<IIC_FPGeneral , [InstrStage<1, [P7_DU1, P7_DU2,
316 InstrStage<1, [P7_VS1, P7_VS2]>],
318 InstrItinData<IIC_FPAddSub , [InstrStage<1, [P7_DU1, P7_DU2,
320 InstrStage<1, [P7_VS1, P7_VS2]>],
322 InstrItinData<IIC_FPCompare , [InstrStage<1, [P7_DU1, P7_DU2,
324 InstrStage<1, [P7_VS1, P7_VS2]>],
326 InstrItinData<IIC_FPDivD , [InstrStage<1, [P7_DU1, P7_DU2,
328 InstrStage<1, [P7_VS1, P7_VS2]>],
330 InstrItinData<IIC_FPDivS , [InstrStage<1, [P7_DU1, P7_DU2,
332 InstrStage<1, [P7_VS1, P7_VS2]>],
334 InstrItinData<IIC_FPSqrtD , [InstrStage<1, [P7_DU1, P7_DU2,
336 InstrStage<1, [P7_VS1, P7_VS2]>],
338 InstrItinData<IIC_FPSqrtS , [InstrStage<1, [P7_DU1, P7_DU2,
340 InstrStage<1, [P7_VS1, P7_VS2]>],
342 InstrItinData<IIC_FPFused , [InstrStage<1, [P7_DU1, P7_DU2,
344 InstrStage<1, [P7_VS1, P7_VS2]>],
346 InstrItinData<IIC_FPRes , [InstrStage<1, [P7_DU1, P7_DU2,
348 InstrStage<1, [P7_VS1, P7_VS2]>],
350 InstrItinData<IIC_VecGeneral , [InstrStage<1, [P7_DU1], 0>,
351 InstrStage<1, [P7_VS1]>],
353 InstrItinData<IIC_VecVSL , [InstrStage<1, [P7_DU1], 0>,
354 InstrStage<1, [P7_VS1]>],
356 InstrItinData<IIC_VecVSR , [InstrStage<1, [P7_DU1], 0>,
357 InstrStage<1, [P7_VS1]>],
359 InstrItinData<IIC_VecFP , [InstrStage<1, [P7_DU1], 0>,
360 InstrStage<1, [P7_VS1, P7_VS2]>],
362 InstrItinData<IIC_VecFPCompare, [InstrStage<1, [P7_DU1], 0>,
363 InstrStage<1, [P7_VS1, P7_VS2]>],
365 InstrItinData<IIC_VecFPRound , [InstrStage<1, [P7_DU1], 0>,
366 InstrStage<1, [P7_VS1, P7_VS2]>],
368 InstrItinData<IIC_VecComplex , [InstrStage<1, [P7_DU1], 0>,
369 InstrStage<1, [P7_VS1]>],
371 InstrItinData<IIC_VecPerm , [InstrStage<1, [P7_DU1, P7_DU2], 0>,
372 InstrStage<1, [P7_VS2]>],
376 // ===---------------------------------------------------------------------===//
377 // P7 machine model for scheduling and other instruction cost heuristics.
379 def P7Model : SchedMachineModel {
380 let IssueWidth = 6; // 4 (non-branch) instructions are dispatched per cycle.
381 // Note that the dispatch bundle size is 6 (including
382 // branches), but the total internal issue bandwidth per
383 // cycle (from all queues) is 8.
385 let LoadLatency = 3; // Optimistic load latency assuming bypass.
386 // This is overriden by OperandCycles if the
387 // Itineraries are queried instead.
388 let MispredictPenalty = 16;
390 // Try to make sure we have at least 10 dispatch groups in a loop.
391 let LoopMicroOpBufferSize = 40;
393 let CompleteModel = 0;
395 let Itineraries = P7Itineraries;