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 latency 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 SPE (Signal Processing Engine), prefixed instructions on
44 // Power 9, paired vector mem ops, MMA, PC relative mem ops, or instructions
45 // introduced in ISA 3.1.
46 let UnsupportedFeatures = [HasSPE, PrefixInstrs, PairedVectorMemops, MMA,
47 PCRelativeMemops, IsISA3_1];
50 let SchedModel = P9Model in {
52 // ***************** Processor Resources *****************
55 // x0, x1, x2, and x3 are the dedicated slice dispatch ports, where each
56 // corresponds to one of the four execution slices.
57 def DISPx02 : ProcResource<2>;
58 def DISPx13 : ProcResource<2>;
59 // The xa and xb ports can be used to send an iop to either of the two slices
60 // of the superslice, but are restricted to iops with only two primary sources.
61 def DISPxab : ProcResource<2>;
62 // b0 and b1 are dedicated dispatch ports into the branch slice.
63 def DISPb01 : ProcResource<2>;
65 // Any non BR dispatch ports
67 : ProcResGroup<[ DISPx02, DISPx13, DISPxab]>;
68 def DISP_SS : ProcResGroup<[ DISPx02, DISPx13]>;
71 // An instruction can go down one of two issue queues.
72 // Address Generation (AGEN) mainly for loads and stores.
73 // Execution (EXEC) for most other instructions.
74 // Some instructions cannot be run on just any issue queue and may require an
75 // Even or an Odd queue. The EXECE represents the even queues and the EXECO
76 // represents the odd queues.
77 def IP_AGEN : ProcResource<4>;
78 def IP_EXEC : ProcResource<4>;
79 def IP_EXECE : ProcResource<2> {
83 def IP_EXECO : ProcResource<2> {
89 // Four ALU (Fixed Point Arithmetic) units in total. Two even, two Odd.
90 def ALU : ProcResource<4>;
91 def ALUE : ProcResource<2> {
95 def ALUO : ProcResource<2> {
100 // Two DIV (Fixed Point Divide) units.
101 def DIV : ProcResource<2>;
103 // Four DP (Floating Point) units in total. Two even, two Odd.
104 def DP : ProcResource<4>;
105 def DPE : ProcResource<2> {
109 def DPO : ProcResource<2> {
114 // Four LS (Load or Store) units.
115 def LS : ProcResource<4>;
117 // Two PM (Permute) units.
118 def PM : ProcResource<2>;
120 // Only one DFU (Decimal Floating Point and Quad Precision) unit.
121 def DFU : ProcResource<1>;
123 // Only one Branch unit.
124 def BR : ProcResource<1> {
128 // Only one CY (Crypto) unit.
129 def CY : ProcResource<1>;
131 // ***************** SchedWriteRes Definitions *****************
134 // Dispatch Rules: '-' or 'V'
135 // Vector ('V') - vector iops (128-bit operand) take only one decode and
136 // dispatch slot but are dispatched to both the even and odd slices of a
138 def DISP_1C : SchedWriteRes<[DISP_NBR]> {
142 // Dispatch Rules: 'E'
143 // Even slice ('E')- certain operations must be sent only to an even slice.
144 // Also consumes odd dispatch slice slot of the same superslice at dispatch
145 def DISP_EVEN_1C : SchedWriteRes<[ DISPx02, DISPx13 ]> {
149 // Dispatch Rules: 'P'
150 // Paired ('P') - certain cracked and expanded iops are paired such that they
151 // must dispatch together to the same superslice.
152 def DISP_PAIR_1C : SchedWriteRes<[ DISP_SS, DISP_SS]> {
156 // Tuple Restricted ('R') - certain iops preclude dispatching more than one
157 // operation per slice for the super- slice to which they are dispatched
158 def DISP_3SLOTS_1C : SchedWriteRes<[DISPx02, DISPx13, DISPxab]> {
162 // Each execution and branch slice can receive up to two iops per cycle
163 def DISP_BR_1C : SchedWriteRes<[ DISPxab ]> {
169 def IP_AGEN_1C : SchedWriteRes<[IP_AGEN]> {
174 def IP_EXEC_1C : SchedWriteRes<[IP_EXEC]> {
179 def IP_EXECE_1C : SchedWriteRes<[IP_EXECE]> {
184 def IP_EXECO_1C : SchedWriteRes<[IP_EXECO]> {
192 // An ALU may take either 2 or 3 cycles to complete the operation.
193 // However, the ALU unit is only ever busy for 1 cycle at a time and may
194 // receive new instructions each cycle.
195 def P9_ALU_2C : SchedWriteRes<[ALU]> {
199 def P9_ALUE_2C : SchedWriteRes<[ALUE]> {
203 def P9_ALUO_2C : SchedWriteRes<[ALUO]> {
207 def P9_ALU_3C : SchedWriteRes<[ALU]> {
211 def P9_ALUE_3C : SchedWriteRes<[ALUE]> {
215 def P9_ALUO_3C : SchedWriteRes<[ALUO]> {
220 // A DIV unit may take from 5 to 40 cycles to complete.
221 // Some DIV operations may keep the unit busy for up to 8 cycles.
222 def P9_DIV_5C : SchedWriteRes<[DIV]> {
226 def P9_DIV_12C : SchedWriteRes<[DIV]> {
230 def P9_DIV_16C_8 : SchedWriteRes<[DIV]> {
231 let ResourceCycles = [8];
235 def P9_DIV_24C_8 : SchedWriteRes<[DIV]> {
236 let ResourceCycles = [8];
240 def P9_DIV_40C_8 : SchedWriteRes<[DIV]> {
241 let ResourceCycles = [8];
246 // A DP unit may take from 2 to 36 cycles to complete.
247 // Some DP operations keep the unit busy for up to 10 cycles.
248 def P9_DP_5C : SchedWriteRes<[DP]> {
252 def P9_DP_7C : SchedWriteRes<[DP]> {
256 def P9_DPE_7C : SchedWriteRes<[DPE]> {
260 def P9_DPO_7C : SchedWriteRes<[DPO]> {
264 def P9_DP_22C_5 : SchedWriteRes<[DP]> {
265 let ResourceCycles = [5];
269 def P9_DPO_24C_8 : SchedWriteRes<[DPO]> {
270 let ResourceCycles = [8];
274 def P9_DPE_24C_8 : SchedWriteRes<[DPE]> {
275 let ResourceCycles = [8];
279 def P9_DP_26C_5 : SchedWriteRes<[DP]> {
280 let ResourceCycles = [5];
284 def P9_DPE_27C_10 : SchedWriteRes<[DP]> {
285 let ResourceCycles = [10];
289 def P9_DPO_27C_10 : SchedWriteRes<[DP]> {
290 let ResourceCycles = [10];
294 def P9_DP_33C_8 : SchedWriteRes<[DP]> {
295 let ResourceCycles = [8];
299 def P9_DPE_33C_8 : SchedWriteRes<[DPE]> {
300 let ResourceCycles = [8];
304 def P9_DPO_33C_8 : SchedWriteRes<[DPO]> {
305 let ResourceCycles = [8];
309 def P9_DP_36C_10 : SchedWriteRes<[DP]> {
310 let ResourceCycles = [10];
314 def P9_DPE_36C_10 : SchedWriteRes<[DP]> {
315 let ResourceCycles = [10];
319 def P9_DPO_36C_10 : SchedWriteRes<[DP]> {
320 let ResourceCycles = [10];
325 // Three cycle permute operations.
326 def P9_PM_3C : SchedWriteRes<[PM]> {
330 // Load and Store Units
331 // Loads can have 4, 5 or 6 cycles of latency.
332 // Stores are listed as having a single cycle of latency. This is not
333 // completely accurate since it takes more than 1 cycle to actually store
334 // the value. However, since the store does not produce a result it can be
335 // considered complete after one cycle.
336 def P9_LS_1C : SchedWriteRes<[LS]> {
340 def P9_LS_4C : SchedWriteRes<[LS]> {
344 def P9_LS_5C : SchedWriteRes<[LS]> {
348 def P9_LS_6C : SchedWriteRes<[LS]> {
353 // Some of the most expensive ops use the DFU.
354 // Can take from 12 cycles to 76 cycles to obtain a result.
355 // The unit may be busy for up to 62 cycles.
356 def P9_DFU_12C : SchedWriteRes<[DFU]> {
360 def P9_DFU_23C : SchedWriteRes<[DFU]> {
362 let ResourceCycles = [11];
365 def P9_DFU_24C : SchedWriteRes<[DFU]> {
367 let ResourceCycles = [12];
370 def P9_DFU_37C : SchedWriteRes<[DFU]> {
372 let ResourceCycles = [25];
375 def P9_DFU_58C : SchedWriteRes<[DFU]> {
377 let ResourceCycles = [44];
380 def P9_DFU_76C : SchedWriteRes<[DFU]> {
382 let ResourceCycles = [62];
385 // 2 or 5 cycle latencies for the branch unit.
386 def P9_BR_2C : SchedWriteRes<[BR]> {
390 def P9_BR_5C : SchedWriteRes<[BR]> {
394 // 6 cycle latency for the crypto unit
395 def P9_CY_6C : SchedWriteRes<[CY]> {
399 // ***************** WriteSeq Definitions *****************
401 // These are combinations of the resources listed above.
402 // The idea is that some cracked instructions cannot be done in parallel and
403 // so the latencies for their resources must be added.
404 def P9_LoadAndALUOp_6C : WriteSequence<[P9_LS_4C, P9_ALU_2C]>;
405 def P9_LoadAndALUOp_7C : WriteSequence<[P9_LS_5C, P9_ALU_2C]>;
406 def P9_LoadAndALU2Op_7C : WriteSequence<[P9_LS_4C, P9_ALU_3C]>;
407 def P9_LoadAndALU2Op_8C : WriteSequence<[P9_LS_5C, P9_ALU_3C]>;
408 def P9_LoadAndPMOp_8C : WriteSequence<[P9_LS_5C, P9_PM_3C]>;
409 def P9_LoadAndLoadOp_8C : WriteSequence<[P9_LS_4C, P9_LS_4C]>;
410 def P9_IntDivAndALUOp_18C_8 : WriteSequence<[P9_DIV_16C_8, P9_ALU_2C]>;
411 def P9_IntDivAndALUOp_26C_8 : WriteSequence<[P9_DIV_24C_8, P9_ALU_2C]>;
412 def P9_IntDivAndALUOp_42C_8 : WriteSequence<[P9_DIV_40C_8, P9_ALU_2C]>;
413 def P9_StoreAndALUOp_3C : WriteSequence<[P9_LS_1C, P9_ALU_2C]>;
414 def P9_ALUOpAndALUOp_4C : WriteSequence<[P9_ALU_2C, P9_ALU_2C]>;
415 def P9_ALU2OpAndALU2Op_6C : WriteSequence<[P9_ALU_3C, P9_ALU_3C]>;
416 def P9_ALUOpAndALUOpAndALUOp_6C :
417 WriteSequence<[P9_ALU_2C, P9_ALU_2C, P9_ALU_2C]>;
418 def P9_DPOpAndALUOp_7C : WriteSequence<[P9_DP_5C, P9_ALU_2C]>;
419 def P9_DPOpAndALU2Op_10C : WriteSequence<[P9_DP_7C, P9_ALU_3C]>;
420 def P9_DPOpAndALU2Op_25C_5 : WriteSequence<[P9_DP_22C_5, P9_ALU_3C]>;
421 def P9_DPOpAndALU2Op_29C_5 : WriteSequence<[P9_DP_26C_5, P9_ALU_3C]>;
422 def P9_DPOpAndALU2Op_36C_8 : WriteSequence<[P9_DP_33C_8, P9_ALU_3C]>;
423 def P9_DPOpAndALU2Op_39C_10 : WriteSequence<[P9_DP_36C_10, P9_ALU_3C]>;
424 def P9_BROpAndALUOp_7C : WriteSequence<[P9_BR_5C, P9_ALU_2C]>;
426 // Include the resource requirements of individual instructions.
427 include "P9InstrResources.td"