1 //===-- PPCScheduleP9.td - PPC P9 Scheduling Definitions ---*- tablegen -*-===//
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
7 //===----------------------------------------------------------------------===//
9 // This file defines the itinerary class data for the POWER9 processor.
11 //===----------------------------------------------------------------------===//
12 include "PPCInstrInfo.td"
14 def P9Model : SchedMachineModel {
15 // The maximum number of instructions to be issued at the same time.
16 // While a value of 8 is technically correct since 8 instructions can be
17 // fetched from the instruction cache. However, only 6 instructions may be
18 // actually dispatched at a time.
21 // Load latency is 4 or 5 cycles depending on the load. This latency assumes
22 // that we have a cache hit. For a cache miss the load latency will be more.
23 // There are two instructions (lxvl, lxvll) that have a latencty of 6 cycles.
24 // However it is not worth bumping this value up to 6 when the vast majority
25 // of instructions are 4 or 5 cycles.
28 // A total of 16 cycles to recover from a branch mispredict.
29 let MispredictPenalty = 16;
31 // Try to make sure we have at least 10 dispatch groups in a loop.
32 // A dispatch group is 6 instructions.
33 let LoopMicroOpBufferSize = 60;
35 // As iops are dispatched to a slice, they are held in an independent slice
36 // issue queue until all register sources and other dependencies have been
37 // resolved and they can be issued. Each of four execution slices has an
38 // 11-entry iop issue queue.
39 let MicroOpBufferSize = 44;
41 let CompleteModel = 1;
43 // Do not support QPX (Quad Processing eXtension) or SPE (Signal Procesing
44 // Engine) on Power 9.
45 let UnsupportedFeatures = [HasQPX, HasSPE];
49 let SchedModel = P9Model in {
51 // ***************** Processor Resources *****************
54 // x0, x1, x2, and x3 are the dedicated slice dispatch ports, where each
55 // corresponds to one of the four execution slices.
56 def DISPx02 : ProcResource<2>;
57 def DISPx13 : ProcResource<2>;
58 // The xa and xb ports can be used to send an iop to either of the two slices
59 // of the superslice, but are restricted to iops with only two primary sources.
60 def DISPxab : ProcResource<2>;
61 // b0 and b1 are dedicated dispatch ports into the branch slice.
62 def DISPb01 : ProcResource<2>;
64 // Any non BR dispatch ports
66 : ProcResGroup<[ DISPx02, DISPx13, DISPxab]>;
67 def DISP_SS : ProcResGroup<[ DISPx02, DISPx13]>;
70 // An instruction can go down one of two issue queues.
71 // Address Generation (AGEN) mainly for loads and stores.
72 // Execution (EXEC) for most other instructions.
73 // Some instructions cannot be run on just any issue queue and may require an
74 // Even or an Odd queue. The EXECE represents the even queues and the EXECO
75 // represents the odd queues.
76 def IP_AGEN : ProcResource<4>;
77 def IP_EXEC : ProcResource<4>;
78 def IP_EXECE : ProcResource<2> {
82 def IP_EXECO : ProcResource<2> {
88 // Four ALU (Fixed Point Arithmetic) units in total. Two even, two Odd.
89 def ALU : ProcResource<4>;
90 def ALUE : ProcResource<2> {
94 def ALUO : ProcResource<2> {
99 // Two DIV (Fixed Point Divide) units.
100 def DIV : ProcResource<2>;
102 // Four DP (Floating Point) units in total. Two even, two Odd.
103 def DP : ProcResource<4>;
104 def DPE : ProcResource<2> {
108 def DPO : ProcResource<2> {
113 // Four LS (Load or Store) units.
114 def LS : ProcResource<4>;
116 // Two PM (Permute) units.
117 def PM : ProcResource<2>;
119 // Only one DFU (Decimal Floating Point and Quad Precision) unit.
120 def DFU : ProcResource<1>;
122 // Only one Branch unit.
123 def BR : ProcResource<1> {
127 // Only one CY (Crypto) unit.
128 def CY : ProcResource<1>;
130 // ***************** SchedWriteRes Definitions *****************
133 // Dispatch Rules: '-' or 'V'
134 // Vector ('V') - vector iops (128-bit operand) take only one decode and
135 // dispatch slot but are dispatched to both the even and odd slices of a
137 def DISP_1C : SchedWriteRes<[DISP_NBR]> {
141 // Dispatch Rules: 'E'
142 // Even slice ('E')- certain operations must be sent only to an even slice.
143 // Also consumes odd dispatch slice slot of the same superslice at dispatch
144 def DISP_EVEN_1C : SchedWriteRes<[ DISPx02, DISPx13 ]> {
148 // Dispatch Rules: 'P'
149 // Paired ('P') - certain cracked and expanded iops are paired such that they
150 // must dispatch together to the same superslice.
151 def DISP_PAIR_1C : SchedWriteRes<[ DISP_SS, DISP_SS]> {
155 // Tuple Restricted ('R') - certain iops preclude dispatching more than one
156 // operation per slice for the super- slice to which they are dispatched
157 def DISP_3SLOTS_1C : SchedWriteRes<[DISPx02, DISPx13, DISPxab]> {
161 // Each execution and branch slice can receive up to two iops per cycle
162 def DISP_BR_1C : SchedWriteRes<[ DISPxab ]> {
168 def IP_AGEN_1C : SchedWriteRes<[IP_AGEN]> {
173 def IP_EXEC_1C : SchedWriteRes<[IP_EXEC]> {
178 def IP_EXECE_1C : SchedWriteRes<[IP_EXECE]> {
183 def IP_EXECO_1C : SchedWriteRes<[IP_EXECO]> {
191 // An ALU may take either 2 or 3 cycles to complete the operation.
192 // However, the ALU unit is only ever busy for 1 cycle at a time and may
193 // receive new instructions each cycle.
194 def P9_ALU_2C : SchedWriteRes<[ALU]> {
198 def P9_ALUE_2C : SchedWriteRes<[ALUE]> {
202 def P9_ALUO_2C : SchedWriteRes<[ALUO]> {
206 def P9_ALU_3C : SchedWriteRes<[ALU]> {
210 def P9_ALUE_3C : SchedWriteRes<[ALUE]> {
214 def P9_ALUO_3C : SchedWriteRes<[ALUO]> {
219 // A DIV unit may take from 5 to 40 cycles to complete.
220 // Some DIV operations may keep the unit busy for up to 8 cycles.
221 def P9_DIV_5C : SchedWriteRes<[DIV]> {
225 def P9_DIV_12C : SchedWriteRes<[DIV]> {
229 def P9_DIV_16C_8 : SchedWriteRes<[DIV]> {
230 let ResourceCycles = [8];
234 def P9_DIV_24C_8 : SchedWriteRes<[DIV]> {
235 let ResourceCycles = [8];
239 def P9_DIV_40C_8 : SchedWriteRes<[DIV]> {
240 let ResourceCycles = [8];
245 // A DP unit may take from 2 to 36 cycles to complete.
246 // Some DP operations keep the unit busy for up to 10 cycles.
247 def P9_DP_5C : SchedWriteRes<[DP]> {
251 def P9_DP_7C : SchedWriteRes<[DP]> {
255 def P9_DPE_7C : SchedWriteRes<[DPE]> {
259 def P9_DPO_7C : SchedWriteRes<[DPO]> {
263 def P9_DP_22C_5 : SchedWriteRes<[DP]> {
264 let ResourceCycles = [5];
268 def P9_DPO_24C_8 : SchedWriteRes<[DPO]> {
269 let ResourceCycles = [8];
273 def P9_DPE_24C_8 : SchedWriteRes<[DPE]> {
274 let ResourceCycles = [8];
278 def P9_DP_26C_5 : SchedWriteRes<[DP]> {
279 let ResourceCycles = [5];
283 def P9_DPE_27C_10 : SchedWriteRes<[DP]> {
284 let ResourceCycles = [10];
288 def P9_DPO_27C_10 : SchedWriteRes<[DP]> {
289 let ResourceCycles = [10];
293 def P9_DP_33C_8 : SchedWriteRes<[DP]> {
294 let ResourceCycles = [8];
298 def P9_DPE_33C_8 : SchedWriteRes<[DPE]> {
299 let ResourceCycles = [8];
303 def P9_DPO_33C_8 : SchedWriteRes<[DPO]> {
304 let ResourceCycles = [8];
308 def P9_DP_36C_10 : SchedWriteRes<[DP]> {
309 let ResourceCycles = [10];
313 def P9_DPE_36C_10 : SchedWriteRes<[DP]> {
314 let ResourceCycles = [10];
318 def P9_DPO_36C_10 : SchedWriteRes<[DP]> {
319 let ResourceCycles = [10];
324 // Three cycle permute operations.
325 def P9_PM_3C : SchedWriteRes<[PM]> {
329 // Load and Store Units
330 // Loads can have 4, 5 or 6 cycles of latency.
331 // Stores are listed as having a single cycle of latency. This is not
332 // completely accurate since it takes more than 1 cycle to actually store
333 // the value. However, since the store does not produce a result it can be
334 // considered complete after one cycle.
335 def P9_LS_1C : SchedWriteRes<[LS]> {
339 def P9_LS_4C : SchedWriteRes<[LS]> {
343 def P9_LS_5C : SchedWriteRes<[LS]> {
347 def P9_LS_6C : SchedWriteRes<[LS]> {
352 // Some of the most expensive ops use the DFU.
353 // Can take from 12 cycles to 76 cycles to obtain a result.
354 // The unit may be busy for up to 62 cycles.
355 def P9_DFU_12C : SchedWriteRes<[DFU]> {
359 def P9_DFU_23C : SchedWriteRes<[DFU]> {
361 let ResourceCycles = [11];
364 def P9_DFU_24C : SchedWriteRes<[DFU]> {
366 let ResourceCycles = [12];
369 def P9_DFU_37C : SchedWriteRes<[DFU]> {
371 let ResourceCycles = [25];
374 def P9_DFU_58C : SchedWriteRes<[DFU]> {
376 let ResourceCycles = [44];
379 def P9_DFU_76C : SchedWriteRes<[DFU]> {
381 let ResourceCycles = [62];
384 // 2 or 5 cycle latencies for the branch unit.
385 def P9_BR_2C : SchedWriteRes<[BR]> {
389 def P9_BR_5C : SchedWriteRes<[BR]> {
393 // 6 cycle latency for the crypto unit
394 def P9_CY_6C : SchedWriteRes<[CY]> {
398 // ***************** WriteSeq Definitions *****************
400 // These are combinations of the resources listed above.
401 // The idea is that some cracked instructions cannot be done in parallel and
402 // so the latencies for their resources must be added.
403 def P9_LoadAndALUOp_6C : WriteSequence<[P9_LS_4C, P9_ALU_2C]>;
404 def P9_LoadAndALUOp_7C : WriteSequence<[P9_LS_5C, P9_ALU_2C]>;
405 def P9_LoadAndALU2Op_7C : WriteSequence<[P9_LS_4C, P9_ALU_3C]>;
406 def P9_LoadAndALU2Op_8C : WriteSequence<[P9_LS_5C, P9_ALU_3C]>;
407 def P9_LoadAndPMOp_8C : WriteSequence<[P9_LS_5C, P9_PM_3C]>;
408 def P9_LoadAndLoadOp_8C : WriteSequence<[P9_LS_4C, P9_LS_4C]>;
409 def P9_IntDivAndALUOp_18C_8 : WriteSequence<[P9_DIV_16C_8, P9_ALU_2C]>;
410 def P9_IntDivAndALUOp_26C_8 : WriteSequence<[P9_DIV_24C_8, P9_ALU_2C]>;
411 def P9_IntDivAndALUOp_42C_8 : WriteSequence<[P9_DIV_40C_8, P9_ALU_2C]>;
412 def P9_StoreAndALUOp_3C : WriteSequence<[P9_LS_1C, P9_ALU_2C]>;
413 def P9_ALUOpAndALUOp_4C : WriteSequence<[P9_ALU_2C, P9_ALU_2C]>;
414 def P9_ALU2OpAndALU2Op_6C : WriteSequence<[P9_ALU_3C, P9_ALU_3C]>;
415 def P9_ALUOpAndALUOpAndALUOp_6C :
416 WriteSequence<[P9_ALU_2C, P9_ALU_2C, P9_ALU_2C]>;
417 def P9_DPOpAndALUOp_7C : WriteSequence<[P9_DP_5C, P9_ALU_2C]>;
418 def P9_DPOpAndALU2Op_10C : WriteSequence<[P9_DP_7C, P9_ALU_3C]>;
419 def P9_DPOpAndALU2Op_25C_5 : WriteSequence<[P9_DP_22C_5, P9_ALU_3C]>;
420 def P9_DPOpAndALU2Op_29C_5 : WriteSequence<[P9_DP_26C_5, P9_ALU_3C]>;
421 def P9_DPOpAndALU2Op_36C_8 : WriteSequence<[P9_DP_33C_8, P9_ALU_3C]>;
422 def P9_DPOpAndALU2Op_39C_10 : WriteSequence<[P9_DP_36C_10, P9_ALU_3C]>;
423 def P9_BROpAndALUOp_7C : WriteSequence<[P9_BR_5C, P9_ALU_2C]>;
425 // Include the resource requirements of individual instructions.
426 include "P9InstrResources.td"