llvm-for-llvmta/lib/Target/ARM/ARMScheduleM7.td

489 lines
18 KiB
TableGen

//=- ARMScheduleM7.td - ARM Cortex-M7 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 SchedRead/Write data for the ARM Cortex-M7 processor.
//
//===----------------------------------------------------------------------===//
def CortexM7Model : SchedMachineModel {
let IssueWidth = 2; // Dual issue for most instructions.
let MicroOpBufferSize = 0; // The Cortex-M7 is in-order.
let LoadLatency = 2; // Best case for load-use case.
let MispredictPenalty = 4; // Mispredict cost for forward branches is 6,
// but 4 works better
let CompleteModel = 0;
}
//===--------------------------------------------------------------------===//
// The Cortex-M7 has two ALU, two LOAD, a STORE, a MAC, a BRANCH and a VFP
// pipe. The stages relevant to scheduling are as follows:
//
// EX1: address generation shifts
// EX2: fast load data ALUs FP operation
// EX3: slow load data integer writeback FP operation
// EX4: store data FP writeback
//
// There are shifters in both EX1 and EX2, and some instructions can be
// flexibly allocated between them. EX2 is used as the "zero" point
// for scheduling, so simple ALU operations executing in EX2 will have
// ReadAdvance<0> (the default) for their source operands and Latency = 1.
def M7UnitLoad : ProcResource<2> { let BufferSize = 0; }
def M7UnitStore : ProcResource<1> { let BufferSize = 0; }
def M7UnitALU : ProcResource<2>;
def M7UnitShift1 : ProcResource<1> { let BufferSize = 0; }
def M7UnitShift2 : ProcResource<1> { let BufferSize = 0; }
def M7UnitMAC : ProcResource<1> { let BufferSize = 0; }
def M7UnitBranch : ProcResource<1> { let BufferSize = 0; }
def M7UnitVFP : ProcResource<1> { let BufferSize = 0; }
def M7UnitVPort : ProcResource<2> { let BufferSize = 0; }
def M7UnitSIMD : ProcResource<1> { let BufferSize = 0; }
//===---------------------------------------------------------------------===//
// Subtarget-specific SchedWrite types with map ProcResources and set latency.
let SchedModel = CortexM7Model in {
def : WriteRes<WriteALU, [M7UnitALU]> { let Latency = 1; }
// Basic ALU with shifts.
let Latency = 1 in {
def : WriteRes<WriteALUsi, [M7UnitALU, M7UnitShift1]>;
def : WriteRes<WriteALUsr, [M7UnitALU, M7UnitShift1]>;
def : WriteRes<WriteALUSsr, [M7UnitALU, M7UnitShift1]>;
}
// Compares.
def : WriteRes<WriteCMP, [M7UnitALU]> { let Latency = 1; }
def : WriteRes<WriteCMPsi, [M7UnitALU, M7UnitShift1]> { let Latency = 2; }
def : WriteRes<WriteCMPsr, [M7UnitALU, M7UnitShift1]> { let Latency = 2; }
// Multiplies.
let Latency = 2 in {
def : WriteRes<WriteMUL16, [M7UnitMAC]>;
def : WriteRes<WriteMUL32, [M7UnitMAC]>;
def : WriteRes<WriteMUL64Lo, [M7UnitMAC]>;
def : WriteRes<WriteMUL64Hi, []> { let NumMicroOps = 0; }
}
// Multiply-accumulates.
let Latency = 2 in {
def : WriteRes<WriteMAC16, [M7UnitMAC]>;
def : WriteRes<WriteMAC32, [M7UnitMAC]>;
def : WriteRes<WriteMAC64Lo, [M7UnitMAC]> { let Latency = 2; }
def : WriteRes<WriteMAC64Hi, []> { let NumMicroOps = 0; }
}
// Divisions.
// These cannot be dual-issued with any instructions.
def : WriteRes<WriteDIV, [M7UnitALU]> {
let Latency = 7;
let SingleIssue = 1;
}
// Loads/Stores.
def : WriteRes<WriteLd, [M7UnitLoad]> { let Latency = 1; }
def : WriteRes<WritePreLd, [M7UnitLoad]> { let Latency = 2; }
def : WriteRes<WriteST, [M7UnitStore]> { let Latency = 2; }
// Branches.
def : WriteRes<WriteBr, [M7UnitBranch]> { let Latency = 2; }
def : WriteRes<WriteBrL, [M7UnitBranch]> { let Latency = 2; }
def : WriteRes<WriteBrTbl, [M7UnitBranch]> { let Latency = 2; }
// Noop.
def : WriteRes<WriteNoop, []> { let Latency = 0; }
//===---------------------------------------------------------------------===//
// Sched definitions for floating-point instructions
//
// Floating point conversions.
def : WriteRes<WriteFPCVT, [M7UnitVFP, M7UnitVPort]> { let Latency = 3; }
def : WriteRes<WriteFPMOV, [M7UnitVPort]> { let Latency = 3; }
// The FP pipeline has a latency of 3 cycles.
// ALU operations (32/64-bit). These go down the FP pipeline.
def : WriteRes<WriteFPALU32, [M7UnitVFP, M7UnitVPort]> { let Latency = 3; }
def : WriteRes<WriteFPALU64, [M7UnitVFP, M7UnitVPort, M7UnitVPort]> {
let Latency = 4;
let BeginGroup = 1;
}
// Multiplication
def : WriteRes<WriteFPMUL32, [M7UnitVFP, M7UnitVPort]> { let Latency = 3; }
def : WriteRes<WriteFPMUL64, [M7UnitVFP, M7UnitVPort, M7UnitVPort]> {
let Latency = 7;
let BeginGroup = 1;
}
// Multiply-accumulate. FPMAC goes down the FP Pipeline.
def : WriteRes<WriteFPMAC32, [M7UnitVFP, M7UnitVPort]> { let Latency = 6; }
def : WriteRes<WriteFPMAC64, [M7UnitVFP, M7UnitVPort, M7UnitVPort]> {
let Latency = 11;
let BeginGroup = 1;
}
// Division. Effective scheduling latency is 3, though real latency is larger
def : WriteRes<WriteFPDIV32, [M7UnitVFP, M7UnitVPort]> { let Latency = 16; }
def : WriteRes<WriteFPDIV64, [M7UnitVFP, M7UnitVPort, M7UnitVPort]> {
let Latency = 30;
let BeginGroup = 1;
}
// Square-root. Effective scheduling latency is 3; real latency is larger
def : WriteRes<WriteFPSQRT32, [M7UnitVFP, M7UnitVPort]> { let Latency = 16; }
def : WriteRes<WriteFPSQRT64, [M7UnitVFP, M7UnitVPort, M7UnitVPort]> {
let Latency = 30;
let BeginGroup = 1;
}
def M7WriteShift2 : SchedWriteRes<[M7UnitALU, M7UnitShift2]> {}
// Not used for M7, but needing definitions anyway
def : WriteRes<WriteVLD1, []>;
def : WriteRes<WriteVLD2, []>;
def : WriteRes<WriteVLD3, []>;
def : WriteRes<WriteVLD4, []>;
def : WriteRes<WriteVST1, []>;
def : WriteRes<WriteVST2, []>;
def : WriteRes<WriteVST3, []>;
def : WriteRes<WriteVST4, []>;
def M7SingleIssue : SchedWriteRes<[]> {
let SingleIssue = 1;
let NumMicroOps = 0;
}
def M7Slot0Only : SchedWriteRes<[]> {
let BeginGroup = 1;
let NumMicroOps = 0;
}
// What pipeline stage operands need to be ready for depending on
// where they come from.
def : ReadAdvance<ReadALUsr, 0>;
def : ReadAdvance<ReadMUL, 0>;
def : ReadAdvance<ReadMAC, 1>;
def : ReadAdvance<ReadALU, 0>;
def : ReadAdvance<ReadFPMUL, 0>;
def : ReadAdvance<ReadFPMAC, 3>;
def M7Read_ISS : SchedReadAdvance<-1>; // operands needed at EX1
def M7Read_EX2 : SchedReadAdvance<1>; // operands needed at EX3
def M7Read_EX3 : SchedReadAdvance<2>; // operands needed at EX4
// Non general purpose instructions may not be dual issued. These
// use both issue units.
def M7NonGeneralPurpose : SchedWriteRes<[]> {
// Assume that these will go down the main ALU pipeline.
// In reality, many look likely to stall the whole pipeline.
let Latency = 3;
let SingleIssue = 1;
}
// List the non general purpose instructions.
def : InstRW<[M7NonGeneralPurpose], (instregex "t2MRS", "tSVC", "tBKPT",
"t2MSR", "t2DMB", "t2DSB", "t2ISB",
"t2HVC", "t2SMC", "t2UDF", "ERET",
"tHINT", "t2HINT", "t2CLREX", "BUNDLE")>;
//===---------------------------------------------------------------------===//
// Sched definitions for load/store
//
// Mark whether the loads/stores must be single-issue
// Address operands are needed earlier
// Data operands are needed later
def M7BaseUpdate : SchedWriteRes<[]> {
let Latency = 0; // Update is bypassable out of EX1
let NumMicroOps = 0;
}
def M7LoadLatency1 : SchedWriteRes<[]> {
let Latency = 1;
let NumMicroOps = 0;
}
def M7SlowLoad : SchedWriteRes<[M7UnitLoad]> { let Latency = 2; }
// Byte and half-word loads should have greater latency than other loads.
// So should load exclusive.
def : InstRW<[M7SlowLoad],
(instregex "t2LDR(B|H|SB|SH)pc")>;
def : InstRW<[M7SlowLoad, M7Read_ISS],
(instregex "t2LDR(B|H|SB|SH)T", "t2LDR(B|H|SB|SH)i",
"tLDR(B|H)i")>;
def : InstRW<[M7SlowLoad, M7Read_ISS, M7Read_ISS],
(instregex "t2LDR(B|H|SB|SH)s", "tLDR(B|H)r", "tLDR(SB|SH)")>;
def : InstRW<[M7SlowLoad, M7BaseUpdate, M7Read_ISS],
(instregex "t2LDR(B|H|SB|SH)_(POST|PRE)")>;
// Exclusive loads/stores cannot be dual-issued
def : InstRW<[WriteLd, M7Slot0Only, M7Read_ISS],
(instregex "t2LDREX$")>;
def : InstRW<[M7SlowLoad, M7Slot0Only, M7Read_ISS],
(instregex "t2LDREX(B|H)")>;
def : InstRW<[WriteST, M7SingleIssue, M7Read_EX2, M7Read_ISS],
(instregex "t2STREX(B|H)?$")>;
// Load/store multiples cannot be dual-issued. Note that default scheduling
// occurs around read/write times of individual registers in the list; read
// time for STM cannot be overridden because it is a variadic source operand.
def : InstRW<[WriteLd, M7SingleIssue, M7Read_ISS],
(instregex "(t|t2)LDM(DB|IA)$")>;
def : InstRW<[WriteST, M7SingleIssue, M7Read_ISS],
(instregex "(t|t2)STM(DB|IA)$")>;
def : InstRW<[M7BaseUpdate, WriteLd, M7SingleIssue, M7Read_ISS],
(instregex "(t|t2)LDM(DB|IA)_UPD$", "tPOP")>;
def : InstRW<[M7BaseUpdate, WriteST, M7SingleIssue, M7Read_ISS],
(instregex "(t|t2)STM(DB|IA)_UPD$", "tPUSH")>;
// Load/store doubles cannot be dual-issued.
def : InstRW<[M7BaseUpdate, WriteST, M7SingleIssue,
M7Read_EX2, M7Read_EX2, M7Read_ISS],
(instregex "t2STRD_(PRE|POST)")>;
def : InstRW<[WriteST, M7SingleIssue, M7Read_EX2, M7Read_EX2, M7Read_ISS],
(instregex "t2STRDi")>;
def : InstRW<[WriteLd, M7LoadLatency1, M7SingleIssue, M7BaseUpdate, M7Read_ISS],
(instregex "t2LDRD_(PRE|POST)")>;
def : InstRW<[WriteLd, M7LoadLatency1, M7SingleIssue, M7Read_ISS],
(instregex "t2LDRDi")>;
// Word load / preload
def : InstRW<[WriteLd],
(instregex "t2LDRpc", "t2PL[DI]pci", "tLDRpci")>;
def : InstRW<[WriteLd, M7Read_ISS],
(instregex "t2LDR(i|T)", "t2PL[DI](W)?i", "tLDRi", "tLDRspi")>;
def : InstRW<[WriteLd, M7Read_ISS, M7Read_ISS],
(instregex "t2LDRs", "t2PL[DI](w)?s", "tLDRr")>;
def : InstRW<[WriteLd, M7BaseUpdate, M7Read_ISS],
(instregex "t2LDR_(POST|PRE)")>;
// Stores
def : InstRW<[M7BaseUpdate, WriteST, M7Read_EX2, M7Read_ISS],
(instregex "t2STR(B|H)?_(POST|PRE)")>;
def : InstRW<[WriteST, M7Read_EX2, M7Read_ISS, M7Read_ISS],
(instregex "t2STR(B|H)?s$", "tSTR(B|H)?r$")>;
def : InstRW<[WriteST, M7Read_EX2, M7Read_ISS],
(instregex "t2STR(B|H)?(i|T)", "tSTR(B|H)?i$", "tSTRspi")>;
// TBB/TBH - single-issue only; takes two cycles to issue
def M7TableLoad : SchedWriteRes<[M7UnitLoad]> {
let NumMicroOps = 2;
let SingleIssue = 1;
}
def : InstRW<[M7TableLoad, M7Read_ISS, M7Read_ISS], (instregex "t2TB")>;
// VFP loads and stores
def M7LoadSP : SchedWriteRes<[M7UnitLoad, M7UnitVPort]> { let Latency = 1; }
def M7LoadDP : SchedWriteRes<[M7UnitLoad, M7UnitVPort, M7UnitVPort]> {
let Latency = 2;
let SingleIssue = 1;
}
def M7StoreSP : SchedWriteRes<[M7UnitStore, M7UnitVPort]>;
def M7StoreDP : SchedWriteRes<[M7UnitStore, M7UnitVPort, M7UnitVPort]> {
let SingleIssue = 1;
}
def : InstRW<[M7LoadSP, M7Read_ISS], (instregex "VLDR(S|H)$")>;
def : InstRW<[M7LoadDP, M7Read_ISS], (instregex "VLDRD$")>;
def : InstRW<[M7StoreSP, M7Read_EX3, M7Read_ISS], (instregex "VSTR(S|H)$")>;
def : InstRW<[M7StoreDP, M7Read_EX3, M7Read_ISS], (instregex "VSTRD$")>;
// Load/store multiples cannot be dual-issued.
def : InstRW<[WriteLd, M7SingleIssue, M7Read_ISS],
(instregex "VLDM(S|D|Q)(DB|IA)$")>;
def : InstRW<[WriteST, M7SingleIssue, M7Read_ISS],
(instregex "VSTM(S|D|Q)(DB|IA)$")>;
def : InstRW<[M7BaseUpdate, WriteLd, M7SingleIssue, M7Read_ISS],
(instregex "VLDM(S|D|Q)(DB|IA)_UPD$")>;
def : InstRW<[M7BaseUpdate, WriteST, M7SingleIssue, M7Read_ISS],
(instregex "VSTM(S|D|Q)(DB|IA)_UPD$")>;
//===---------------------------------------------------------------------===//
// Sched definitions for ALU
//
// Shifted ALU operands are read a cycle early.
def M7Ex1ReadNoFastBypass : SchedReadAdvance<-1, [WriteLd, M7LoadLatency1]>;
def : InstRW<[WriteALUsi, M7Ex1ReadNoFastBypass, M7Read_ISS],
(instregex "t2(ADC|ADDS|ADD|BIC|EOR|ORN|ORR|RSBS|RSB|SBC|SUBS)rs$",
"t2(SUB|CMP|CMNz|TEQ|TST)rs$",
"t2MOVsr(a|l)")>;
def : InstRW<[WriteALUsi, M7Read_ISS],
(instregex "t2MVNs")>;
// Treat pure shift operations (except for RRX) as if they used the EX1
// shifter but have timing as if they used the EX2 shifter as they usually
// can choose the EX2 shifter when needed. Will miss a few dual-issue cases,
// but the results prove to be better than trying to get them exact.
def : InstRW<[M7WriteShift2, M7Read_ISS], (instregex "t2RRX$")>;
def : InstRW<[WriteALUsi], (instregex "(t|t2)(LSL|LSR|ASR|ROR)")>;
// Instructions that use the shifter, but have normal timing.
def : InstRW<[WriteALUsi,M7Slot0Only], (instregex "t2(BFC|BFI)$")>;
// Instructions which are slot zero only but otherwise normal.
def : InstRW<[WriteALU, M7Slot0Only], (instregex "t2CLZ")>;
// MAC operations that don't have SchedRW set.
def : InstRW<[WriteMAC32, ReadMUL, ReadMUL, ReadMAC], (instregex "t2SML[AS]D")>;
// Divides are special because they stall for their latency, and so look like a
// single-cycle as far as scheduling opportunities go. By putting WriteALU
// first, we make the operand latency 1, but keep the instruction latency 7.
def : InstRW<[WriteALU, WriteDIV], (instregex "t2(S|U)DIV")>;
// DSP extension operations
def M7WriteSIMD1 : SchedWriteRes<[M7UnitSIMD, M7UnitALU]> {
let Latency = 1;
let BeginGroup = 1;
}
def M7WriteSIMD2 : SchedWriteRes<[M7UnitSIMD, M7UnitALU]> {
let Latency = 2;
let BeginGroup = 1;
}
def M7WriteShSIMD1 : SchedWriteRes<[M7UnitSIMD, M7UnitALU, M7UnitShift1]> {
let Latency = 1;
let BeginGroup = 1;
}
def M7WriteShSIMD0 : SchedWriteRes<[M7UnitSIMD, M7UnitALU, M7UnitShift1]> {
let Latency = 0; // Bypassable out of EX1
let BeginGroup = 1;
}
def M7WriteShSIMD2 : SchedWriteRes<[M7UnitSIMD, M7UnitALU, M7UnitShift1]> {
let Latency = 2;
let BeginGroup = 1;
}
def : InstRW<[M7WriteShSIMD2, M7Read_ISS],
(instregex "t2(S|U)SAT")>;
def : InstRW<[M7WriteSIMD1, ReadALU],
(instregex "(t|t2)(S|U)XT(B|H)")>;
def : InstRW<[M7WriteSIMD1, ReadALU, ReadALU],
(instregex "t2(S|SH|U|UH)(ADD16|ADD8|ASX|SAX|SUB16|SUB8)",
"t2SEL")>;
def : InstRW<[M7WriteSIMD2, ReadALU, ReadALU],
(instregex "t2(Q|UQ)(ADD|ASX|SAX|SUB)", "t2USAD8")>;
def : InstRW<[M7WriteShSIMD2, M7Read_ISS, M7Read_ISS],
(instregex "t2QD(ADD|SUB)")>;
def : InstRW<[M7WriteShSIMD0, M7Read_ISS],
(instregex "t2(RBIT|REV)", "tREV")>;
def : InstRW<[M7WriteShSIMD1, M7Read_ISS],
(instregex "t2(SBFX|UBFX)")>;
def : InstRW<[M7WriteShSIMD1, ReadALU, M7Read_ISS],
(instregex "t2PKH(BT|TB)", "t2(S|U)XTA")>;
def : InstRW<[M7WriteSIMD2, ReadALU, ReadALU, M7Read_EX2],
(instregex "t2USADA8")>;
// MSR/MRS
def : InstRW<[M7NonGeneralPurpose], (instregex "MSR", "MRS")>;
//===---------------------------------------------------------------------===//
// Sched definitions for FP operations
//
// Effective scheduling latency is really 3 for nearly all FP operations,
// even if their true latency is higher.
def M7WriteVFPLatOverride : SchedWriteRes<[]> {
let Latency = 3;
let NumMicroOps = 0;
}
def M7WriteVFPExtraVPort : SchedWriteRes<[M7UnitVPort]> {
let Latency = 3;
let NumMicroOps = 0;
}
// Instructions which are missing default schedules.
def : InstRW<[WriteFPALU32],
(instregex "V(ABS|CVT.*|NEG|FP_VMAX.*|FP_VMIN.*|RINT.*)S$")>;
def : InstRW<[M7WriteVFPLatOverride, WriteFPALU64],
(instregex "V(ABS|CVT.*|NEG|FP_VMAX.*|FP_VMIN.*|RINT.*)D$")>;
// VCMP
def M7WriteVCMPS : SchedWriteRes<[M7UnitVFP, M7UnitVPort]> { let Latency = 0; }
def M7WriteVCMPD : SchedWriteRes<[M7UnitVFP, M7UnitVPort, M7UnitVPort]> {
let Latency = 0;
let BeginGroup = 1;
}
def : InstRW<[M7WriteVCMPS], (instregex "VCMPS$")>;
def : InstRW<[M7WriteVCMPD], (instregex "VCMPD$")>;
// VMRS/VMSR
def M7VMRS : SchedWriteRes<[M7UnitVFP, M7UnitVPort]> { let SingleIssue = 1; }
def M7VMSR : SchedWriteRes<[M7UnitVFP, M7UnitVPort]> { let SingleIssue = 1; }
def : InstRW<[M7VMRS], (instregex "FMSTAT")>;
def : InstRW<[M7VMSR], (instregex "VMSR")>;
// VSEL cannot bypass in its implied $cpsr operand; model as earlier read
def : InstRW<[WriteFPALU32, M7Slot0Only, ReadALU, ReadALU, M7Read_ISS],
(instregex "VSEL.*S$")>;
def : InstRW<[M7WriteVFPLatOverride, WriteFPALU64, M7Slot0Only,
ReadALU, ReadALU, M7Read_ISS],
(instregex "VSEL.*D$")>;
// VMOV
def : InstRW<[WriteFPMOV],
(instregex "VMOV(H|S)$", "FCONST(H|S)")>;
def : InstRW<[WriteFPMOV, M7WriteVFPExtraVPort, M7Slot0Only],
(instregex "VMOVD$")>;
def : InstRW<[WriteFPMOV, M7WriteVFPExtraVPort, M7Slot0Only],
(instregex "FCONSTD")>;
def : InstRW<[WriteFPMOV, M7WriteVFPExtraVPort, M7SingleIssue],
(instregex "VMOV(DRR|RRD|RRS|SRR)")>;
// Larger-latency overrides.
def : InstRW<[M7WriteVFPLatOverride, WriteFPDIV32], (instregex "VDIVS")>;
def : InstRW<[M7WriteVFPLatOverride, WriteFPDIV64], (instregex "VDIVD")>;
def : InstRW<[M7WriteVFPLatOverride, WriteFPSQRT32], (instregex "VSQRTS")>;
def : InstRW<[M7WriteVFPLatOverride, WriteFPSQRT64], (instregex "VSQRTD")>;
def : InstRW<[M7WriteVFPLatOverride, WriteFPMUL64],
(instregex "V(MUL|NMUL)D")>;
def : InstRW<[M7WriteVFPLatOverride, WriteFPALU64],
(instregex "V(ADD|SUB)D")>;
// Multiply-accumulate. Chained SP timing is correct; rest need overrides
// Double-precision chained MAC stalls the pipeline behind it for 3 cycles,
// making it appear to have 3 cycle latency for scheduling.
def : InstRW<[M7WriteVFPLatOverride, WriteFPMAC64,
ReadFPMAC, ReadFPMUL, ReadFPMUL],
(instregex "V(N)?ML(A|S)D$")>;
// Single-precision fused MACs look like latency 5 with advance of 2.
def M7WriteVFPLatOverride5 : SchedWriteRes<[]> {
let Latency = 5;
let NumMicroOps = 0;
}
def M7ReadFPMAC2 : SchedReadAdvance<2>;
def : InstRW<[M7WriteVFPLatOverride5, WriteFPMAC32,
M7ReadFPMAC2, ReadFPMUL, ReadFPMUL],
(instregex "VF(N)?M(A|S)S$")>;
// Double-precision fused MAC stalls the pipeline behind it for 2 cycles, making
// it appear to have 3 cycle latency for scheduling.
def : InstRW<[M7WriteVFPLatOverride, WriteFPMAC64,
ReadFPMAC, ReadFPMUL, ReadFPMUL],
(instregex "VF(N)?M(A|S)D$")>;
} // SchedModel = CortexM7Model