forked from OSchip/llvm-project
428 lines
12 KiB
TableGen
428 lines
12 KiB
TableGen
//===-- PPCScheduleP9.td - PPC P9 Scheduling Definitions ---*- tablegen -*-===//
|
|
//
|
|
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
// See https://llvm.org/LICENSE.txt for license information.
|
|
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
//
|
|
// This file defines the itinerary class data for the POWER9 processor.
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
def P9Model : SchedMachineModel {
|
|
// The maximum number of instructions to be issued at the same time.
|
|
// While a value of 8 is technically correct since 8 instructions can be
|
|
// fetched from the instruction cache. However, only 6 instructions may be
|
|
// actually dispatched at a time.
|
|
let IssueWidth = 8;
|
|
|
|
// Load latency is 4 or 5 cycles depending on the load. This latency assumes
|
|
// that we have a cache hit. For a cache miss the load latency will be more.
|
|
// There are two instructions (lxvl, lxvll) that have a latency of 6 cycles.
|
|
// However it is not worth bumping this value up to 6 when the vast majority
|
|
// of instructions are 4 or 5 cycles.
|
|
let LoadLatency = 5;
|
|
|
|
// A total of 16 cycles to recover from a branch mispredict.
|
|
let MispredictPenalty = 16;
|
|
|
|
// Try to make sure we have at least 10 dispatch groups in a loop.
|
|
// A dispatch group is 6 instructions.
|
|
let LoopMicroOpBufferSize = 60;
|
|
|
|
// As iops are dispatched to a slice, they are held in an independent slice
|
|
// issue queue until all register sources and other dependencies have been
|
|
// resolved and they can be issued. Each of four execution slices has an
|
|
// 11-entry iop issue queue.
|
|
let MicroOpBufferSize = 44;
|
|
|
|
let CompleteModel = 1;
|
|
|
|
// Do not support SPE (Signal Processing Engine), prefixed instructions on
|
|
// Power 9, paired vector mem ops, MMA, PC relative mem ops, or instructions
|
|
// introduced in ISA 3.1.
|
|
let UnsupportedFeatures = [HasSPE, PrefixInstrs, PairedVectorMemops, MMA,
|
|
PCRelativeMemops, IsISA3_1, IsISAFuture];
|
|
}
|
|
|
|
let SchedModel = P9Model in {
|
|
|
|
// ***************** Processor Resources *****************
|
|
|
|
// Dispatcher slots:
|
|
// x0, x1, x2, and x3 are the dedicated slice dispatch ports, where each
|
|
// corresponds to one of the four execution slices.
|
|
def DISPx02 : ProcResource<2>;
|
|
def DISPx13 : ProcResource<2>;
|
|
// The xa and xb ports can be used to send an iop to either of the two slices
|
|
// of the superslice, but are restricted to iops with only two primary sources.
|
|
def DISPxab : ProcResource<2>;
|
|
// b0 and b1 are dedicated dispatch ports into the branch slice.
|
|
def DISPb01 : ProcResource<2>;
|
|
|
|
// Any non BR dispatch ports
|
|
def DISP_NBR
|
|
: ProcResGroup<[ DISPx02, DISPx13, DISPxab]>;
|
|
def DISP_SS : ProcResGroup<[ DISPx02, DISPx13]>;
|
|
|
|
// Issue Ports
|
|
// An instruction can go down one of two issue queues.
|
|
// Address Generation (AGEN) mainly for loads and stores.
|
|
// Execution (EXEC) for most other instructions.
|
|
// Some instructions cannot be run on just any issue queue and may require an
|
|
// Even or an Odd queue. The EXECE represents the even queues and the EXECO
|
|
// represents the odd queues.
|
|
def IP_AGEN : ProcResource<4>;
|
|
def IP_EXEC : ProcResource<4>;
|
|
def IP_EXECE : ProcResource<2> {
|
|
//Even Exec Ports
|
|
let Super = IP_EXEC;
|
|
}
|
|
def IP_EXECO : ProcResource<2> {
|
|
//Odd Exec Ports
|
|
let Super = IP_EXEC;
|
|
}
|
|
|
|
// Pipeline Groups
|
|
// Four ALU (Fixed Point Arithmetic) units in total. Two even, two Odd.
|
|
def ALU : ProcResource<4>;
|
|
def ALUE : ProcResource<2> {
|
|
//Even ALU pipelines
|
|
let Super = ALU;
|
|
}
|
|
def ALUO : ProcResource<2> {
|
|
//Odd ALU pipelines
|
|
let Super = ALU;
|
|
}
|
|
|
|
// Two DIV (Fixed Point Divide) units.
|
|
def DIV : ProcResource<2>;
|
|
|
|
// Four DP (Floating Point) units in total. Two even, two Odd.
|
|
def DP : ProcResource<4>;
|
|
def DPE : ProcResource<2> {
|
|
//Even DP pipelines
|
|
let Super = DP;
|
|
}
|
|
def DPO : ProcResource<2> {
|
|
//Odd DP pipelines
|
|
let Super = DP;
|
|
}
|
|
|
|
// Four LS (Load or Store) units.
|
|
def LS : ProcResource<4>;
|
|
|
|
// Two PM (Permute) units.
|
|
def PM : ProcResource<2>;
|
|
|
|
// Only one DFU (Decimal Floating Point and Quad Precision) unit.
|
|
def DFU : ProcResource<1>;
|
|
|
|
// Only one Branch unit.
|
|
def BR : ProcResource<1> {
|
|
let BufferSize = 16;
|
|
}
|
|
|
|
// Only one CY (Crypto) unit.
|
|
def CY : ProcResource<1>;
|
|
|
|
// ***************** SchedWriteRes Definitions *****************
|
|
|
|
// Dispatcher
|
|
// Dispatch Rules: '-' or 'V'
|
|
// Vector ('V') - vector iops (128-bit operand) take only one decode and
|
|
// dispatch slot but are dispatched to both the even and odd slices of a
|
|
// superslice.
|
|
def DISP_1C : SchedWriteRes<[DISP_NBR]> {
|
|
let NumMicroOps = 0;
|
|
let Latency = 1;
|
|
}
|
|
// Dispatch Rules: 'E'
|
|
// Even slice ('E')- certain operations must be sent only to an even slice.
|
|
// Also consumes odd dispatch slice slot of the same superslice at dispatch
|
|
def DISP_EVEN_1C : SchedWriteRes<[ DISPx02, DISPx13 ]> {
|
|
let NumMicroOps = 0;
|
|
let Latency = 1;
|
|
}
|
|
// Dispatch Rules: 'P'
|
|
// Paired ('P') - certain cracked and expanded iops are paired such that they
|
|
// must dispatch together to the same superslice.
|
|
def DISP_PAIR_1C : SchedWriteRes<[ DISP_SS, DISP_SS]> {
|
|
let NumMicroOps = 0;
|
|
let Latency = 1;
|
|
}
|
|
// Tuple Restricted ('R') - certain iops preclude dispatching more than one
|
|
// operation per slice for the super- slice to which they are dispatched
|
|
def DISP_3SLOTS_1C : SchedWriteRes<[DISPx02, DISPx13, DISPxab]> {
|
|
let NumMicroOps = 0;
|
|
let Latency = 1;
|
|
}
|
|
// Each execution and branch slice can receive up to two iops per cycle
|
|
def DISP_BR_1C : SchedWriteRes<[ DISPxab ]> {
|
|
let NumMicroOps = 0;
|
|
let Latency = 1;
|
|
}
|
|
|
|
// Issue Ports
|
|
def IP_AGEN_1C : SchedWriteRes<[IP_AGEN]> {
|
|
let NumMicroOps = 0;
|
|
let Latency = 1;
|
|
}
|
|
|
|
def IP_EXEC_1C : SchedWriteRes<[IP_EXEC]> {
|
|
let NumMicroOps = 0;
|
|
let Latency = 1;
|
|
}
|
|
|
|
def IP_EXECE_1C : SchedWriteRes<[IP_EXECE]> {
|
|
let NumMicroOps = 0;
|
|
let Latency = 1;
|
|
}
|
|
|
|
def IP_EXECO_1C : SchedWriteRes<[IP_EXECO]> {
|
|
let NumMicroOps = 0;
|
|
let Latency = 1;
|
|
}
|
|
|
|
//Pipeline Groups
|
|
|
|
// ALU Units
|
|
// An ALU may take either 2 or 3 cycles to complete the operation.
|
|
// However, the ALU unit is only ever busy for 1 cycle at a time and may
|
|
// receive new instructions each cycle.
|
|
def P9_ALU_2C : SchedWriteRes<[ALU]> {
|
|
let Latency = 2;
|
|
}
|
|
|
|
def P9_ALUE_2C : SchedWriteRes<[ALUE]> {
|
|
let Latency = 2;
|
|
}
|
|
|
|
def P9_ALUO_2C : SchedWriteRes<[ALUO]> {
|
|
let Latency = 2;
|
|
}
|
|
|
|
def P9_ALU_3C : SchedWriteRes<[ALU]> {
|
|
let Latency = 3;
|
|
}
|
|
|
|
def P9_ALUE_3C : SchedWriteRes<[ALUE]> {
|
|
let Latency = 3;
|
|
}
|
|
|
|
def P9_ALUO_3C : SchedWriteRes<[ALUO]> {
|
|
let Latency = 3;
|
|
}
|
|
|
|
// DIV Unit
|
|
// A DIV unit may take from 5 to 40 cycles to complete.
|
|
// Some DIV operations may keep the unit busy for up to 8 cycles.
|
|
def P9_DIV_5C : SchedWriteRes<[DIV]> {
|
|
let Latency = 5;
|
|
}
|
|
|
|
def P9_DIV_12C : SchedWriteRes<[DIV]> {
|
|
let Latency = 12;
|
|
}
|
|
|
|
def P9_DIV_16C_8 : SchedWriteRes<[DIV]> {
|
|
let ResourceCycles = [8];
|
|
let Latency = 16;
|
|
}
|
|
|
|
def P9_DIV_24C_8 : SchedWriteRes<[DIV]> {
|
|
let ResourceCycles = [8];
|
|
let Latency = 24;
|
|
}
|
|
|
|
def P9_DIV_40C_8 : SchedWriteRes<[DIV]> {
|
|
let ResourceCycles = [8];
|
|
let Latency = 40;
|
|
}
|
|
|
|
// DP Unit
|
|
// A DP unit may take from 2 to 36 cycles to complete.
|
|
// Some DP operations keep the unit busy for up to 10 cycles.
|
|
def P9_DP_5C : SchedWriteRes<[DP]> {
|
|
let Latency = 5;
|
|
}
|
|
|
|
def P9_DP_7C : SchedWriteRes<[DP]> {
|
|
let Latency = 7;
|
|
}
|
|
|
|
def P9_DPE_7C : SchedWriteRes<[DPE]> {
|
|
let Latency = 7;
|
|
}
|
|
|
|
def P9_DPO_7C : SchedWriteRes<[DPO]> {
|
|
let Latency = 7;
|
|
}
|
|
|
|
def P9_DP_22C_5 : SchedWriteRes<[DP]> {
|
|
let ResourceCycles = [5];
|
|
let Latency = 22;
|
|
}
|
|
|
|
def P9_DPO_24C_8 : SchedWriteRes<[DPO]> {
|
|
let ResourceCycles = [8];
|
|
let Latency = 24;
|
|
}
|
|
|
|
def P9_DPE_24C_8 : SchedWriteRes<[DPE]> {
|
|
let ResourceCycles = [8];
|
|
let Latency = 24;
|
|
}
|
|
|
|
def P9_DP_26C_5 : SchedWriteRes<[DP]> {
|
|
let ResourceCycles = [5];
|
|
let Latency = 22;
|
|
}
|
|
|
|
def P9_DPE_27C_10 : SchedWriteRes<[DP]> {
|
|
let ResourceCycles = [10];
|
|
let Latency = 27;
|
|
}
|
|
|
|
def P9_DPO_27C_10 : SchedWriteRes<[DP]> {
|
|
let ResourceCycles = [10];
|
|
let Latency = 27;
|
|
}
|
|
|
|
def P9_DP_33C_8 : SchedWriteRes<[DP]> {
|
|
let ResourceCycles = [8];
|
|
let Latency = 33;
|
|
}
|
|
|
|
def P9_DPE_33C_8 : SchedWriteRes<[DPE]> {
|
|
let ResourceCycles = [8];
|
|
let Latency = 33;
|
|
}
|
|
|
|
def P9_DPO_33C_8 : SchedWriteRes<[DPO]> {
|
|
let ResourceCycles = [8];
|
|
let Latency = 33;
|
|
}
|
|
|
|
def P9_DP_36C_10 : SchedWriteRes<[DP]> {
|
|
let ResourceCycles = [10];
|
|
let Latency = 36;
|
|
}
|
|
|
|
def P9_DPE_36C_10 : SchedWriteRes<[DP]> {
|
|
let ResourceCycles = [10];
|
|
let Latency = 36;
|
|
}
|
|
|
|
def P9_DPO_36C_10 : SchedWriteRes<[DP]> {
|
|
let ResourceCycles = [10];
|
|
let Latency = 36;
|
|
}
|
|
|
|
// PM Unit
|
|
// Three cycle permute operations.
|
|
def P9_PM_3C : SchedWriteRes<[PM]> {
|
|
let Latency = 3;
|
|
}
|
|
|
|
// Load and Store Units
|
|
// Loads can have 4, 5 or 6 cycles of latency.
|
|
// Stores are listed as having a single cycle of latency. This is not
|
|
// completely accurate since it takes more than 1 cycle to actually store
|
|
// the value. However, since the store does not produce a result it can be
|
|
// considered complete after one cycle.
|
|
def P9_LS_1C : SchedWriteRes<[LS]> {
|
|
let Latency = 1;
|
|
}
|
|
|
|
def P9_LS_4C : SchedWriteRes<[LS]> {
|
|
let Latency = 4;
|
|
}
|
|
|
|
def P9_LS_5C : SchedWriteRes<[LS]> {
|
|
let Latency = 5;
|
|
}
|
|
|
|
def P9_LS_6C : SchedWriteRes<[LS]> {
|
|
let Latency = 6;
|
|
}
|
|
|
|
// DFU Unit
|
|
// Some of the most expensive ops use the DFU.
|
|
// Can take from 12 cycles to 76 cycles to obtain a result.
|
|
// The unit may be busy for up to 62 cycles.
|
|
def P9_DFU_12C : SchedWriteRes<[DFU]> {
|
|
let Latency = 12;
|
|
}
|
|
|
|
def P9_DFU_23C : SchedWriteRes<[DFU]> {
|
|
let Latency = 23;
|
|
let ResourceCycles = [11];
|
|
}
|
|
|
|
def P9_DFU_24C : SchedWriteRes<[DFU]> {
|
|
let Latency = 24;
|
|
let ResourceCycles = [12];
|
|
}
|
|
|
|
def P9_DFU_37C : SchedWriteRes<[DFU]> {
|
|
let Latency = 37;
|
|
let ResourceCycles = [25];
|
|
}
|
|
|
|
def P9_DFU_58C : SchedWriteRes<[DFU]> {
|
|
let Latency = 58;
|
|
let ResourceCycles = [44];
|
|
}
|
|
|
|
def P9_DFU_76C : SchedWriteRes<[DFU]> {
|
|
let Latency = 76;
|
|
let ResourceCycles = [62];
|
|
}
|
|
|
|
// 2 or 5 cycle latencies for the branch unit.
|
|
def P9_BR_2C : SchedWriteRes<[BR]> {
|
|
let Latency = 2;
|
|
}
|
|
|
|
def P9_BR_5C : SchedWriteRes<[BR]> {
|
|
let Latency = 5;
|
|
}
|
|
|
|
// 6 cycle latency for the crypto unit
|
|
def P9_CY_6C : SchedWriteRes<[CY]> {
|
|
let Latency = 6;
|
|
}
|
|
|
|
// ***************** WriteSeq Definitions *****************
|
|
|
|
// These are combinations of the resources listed above.
|
|
// The idea is that some cracked instructions cannot be done in parallel and
|
|
// so the latencies for their resources must be added.
|
|
def P9_LoadAndALUOp_6C : WriteSequence<[P9_LS_4C, P9_ALU_2C]>;
|
|
def P9_LoadAndALUOp_7C : WriteSequence<[P9_LS_5C, P9_ALU_2C]>;
|
|
def P9_LoadAndALU2Op_7C : WriteSequence<[P9_LS_4C, P9_ALU_3C]>;
|
|
def P9_LoadAndALU2Op_8C : WriteSequence<[P9_LS_5C, P9_ALU_3C]>;
|
|
def P9_LoadAndPMOp_8C : WriteSequence<[P9_LS_5C, P9_PM_3C]>;
|
|
def P9_IntDivAndALUOp_18C_8 : WriteSequence<[P9_DIV_16C_8, P9_ALU_2C]>;
|
|
def P9_IntDivAndALUOp_26C_8 : WriteSequence<[P9_DIV_24C_8, P9_ALU_2C]>;
|
|
def P9_IntDivAndALUOp_42C_8 : WriteSequence<[P9_DIV_40C_8, P9_ALU_2C]>;
|
|
def P9_StoreAndALUOp_3C : WriteSequence<[P9_LS_1C, P9_ALU_2C]>;
|
|
def P9_ALUOpAndALUOp_4C : WriteSequence<[P9_ALU_2C, P9_ALU_2C]>;
|
|
def P9_ALU2OpAndALU2Op_6C : WriteSequence<[P9_ALU_3C, P9_ALU_3C]>;
|
|
def P9_ALUOpAndALUOpAndALUOp_6C :
|
|
WriteSequence<[P9_ALU_2C, P9_ALU_2C, P9_ALU_2C]>;
|
|
def P9_DPOpAndALUOp_7C : WriteSequence<[P9_DP_5C, P9_ALU_2C]>;
|
|
def P9_DPOpAndALU2Op_10C : WriteSequence<[P9_DP_7C, P9_ALU_3C]>;
|
|
def P9_DPOpAndALU2Op_25C_5 : WriteSequence<[P9_DP_22C_5, P9_ALU_3C]>;
|
|
def P9_DPOpAndALU2Op_29C_5 : WriteSequence<[P9_DP_26C_5, P9_ALU_3C]>;
|
|
def P9_DPOpAndALU2Op_36C_8 : WriteSequence<[P9_DP_33C_8, P9_ALU_3C]>;
|
|
def P9_DPOpAndALU2Op_39C_10 : WriteSequence<[P9_DP_36C_10, P9_ALU_3C]>;
|
|
def P9_BROpAndALUOp_7C : WriteSequence<[P9_BR_5C, P9_ALU_2C]>;
|
|
|
|
// Include the resource requirements of individual instructions.
|
|
include "P9InstrResources.td"
|
|
|
|
}
|
|
|