mirror of
https://github.com/c64scene-ar/llvm-6502.git
synced 2024-09-02 11:30:10 +00:00
The initial values were arbitrary. I want them to be more conservative. This represents the number of latency cycles hidden by OOO execution. In practice, I think it should be within a small factor of the complex floating point operation latency so the scheduler can make some attempt to hide latency even for smallish blocks. These are by no means the best values, just a starting point for tuning heuristics. Some benchmarks such as TSVC run faster with this lower value for SandyBridge. I haven't run anything on Haswell, but it's shouldn't be 2x SB. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179450 91177308-0d34-0410-b5e6-96231b3b80d8
127 lines
4.8 KiB
TableGen
127 lines
4.8 KiB
TableGen
//=- X86SchedHaswell.td - X86 Haswell Scheduling -------------*- tablegen -*-=//
|
|
//
|
|
// The LLVM Compiler Infrastructure
|
|
//
|
|
// This file is distributed under the University of Illinois Open Source
|
|
// License. See LICENSE.TXT for details.
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
//
|
|
// This file defines the machine model for Haswell to support instruction
|
|
// scheduling and other instruction cost heuristics.
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
def HaswellModel : SchedMachineModel {
|
|
// All x86 instructions are modeled as a single micro-op, and HW can decode 4
|
|
// instructions per cycle.
|
|
let IssueWidth = 4;
|
|
let MinLatency = 0; // 0 = Out-of-order execution.
|
|
let LoadLatency = 4;
|
|
let ILPWindow = 30;
|
|
let MispredictPenalty = 16;
|
|
}
|
|
|
|
let SchedModel = HaswellModel in {
|
|
|
|
// Haswell can issue micro-ops to 8 different ports in one cycle.
|
|
|
|
// Ports 0, 1, 5, 6 and 7 handle all computation.
|
|
// Port 4 gets the data half of stores. Store data can be available later than
|
|
// the store address, but since we don't model the latency of stores, we can
|
|
// ignore that.
|
|
// Ports 2 and 3 are identical. They handle loads and the address half of
|
|
// stores. Port 7 can handle address calculations.
|
|
def HWPort0 : ProcResource<1>;
|
|
def HWPort1 : ProcResource<1>;
|
|
def HWPort2 : ProcResource<1>;
|
|
def HWPort3 : ProcResource<1>;
|
|
def HWPort4 : ProcResource<1>;
|
|
def HWPort5 : ProcResource<1>;
|
|
def HWPort6 : ProcResource<1>;
|
|
def HWPort7 : ProcResource<1>;
|
|
|
|
// Many micro-ops are capable of issuing on multiple ports.
|
|
def HWPort23 : ProcResGroup<[HWPort2, HWPort3]>;
|
|
def HWPort237 : ProcResGroup<[HWPort2, HWPort3, HWPort7]>;
|
|
def HWPort05 : ProcResGroup<[HWPort0, HWPort5]>;
|
|
def HWPort056 : ProcResGroup<[HWPort0, HWPort5, HWPort6]>;
|
|
def HWPort15 : ProcResGroup<[HWPort1, HWPort5]>;
|
|
def HWPort015 : ProcResGroup<[HWPort0, HWPort1, HWPort5]>;
|
|
def HWPort0156: ProcResGroup<[HWPort0, HWPort1, HWPort5, HWPort6]>;
|
|
|
|
// Integer division issued on port 0.
|
|
def HWDivider : ProcResource<1>;
|
|
|
|
// Loads are 4 cycles, so ReadAfterLd registers needn't be available until 4
|
|
// cycles after the memory operand.
|
|
def : ReadAdvance<ReadAfterLd, 4>;
|
|
|
|
// Many SchedWrites are defined in pairs with and without a folded load.
|
|
// Instructions with folded loads are usually micro-fused, so they only appear
|
|
// as two micro-ops when queued in the reservation station.
|
|
// This multiclass defines the resource usage for variants with and without
|
|
// folded loads.
|
|
multiclass HWWriteResPair<X86FoldableSchedWrite SchedRW,
|
|
ProcResourceKind ExePort,
|
|
int Lat> {
|
|
// Register variant is using a single cycle on ExePort.
|
|
def : WriteRes<SchedRW, [ExePort]> { let Latency = Lat; }
|
|
|
|
// Memory variant also uses a cycle on port 2/3 and adds 4 cycles to the
|
|
// latency.
|
|
def : WriteRes<SchedRW.Folded, [HWPort23, ExePort]> {
|
|
let Latency = !add(Lat, 4);
|
|
}
|
|
}
|
|
|
|
// A folded store needs a cycle on port 4 for the store data, but it does not
|
|
// need an extra port 2/3 cycle to recompute the address.
|
|
def : WriteRes<WriteRMW, [HWPort4]>;
|
|
|
|
def : WriteRes<WriteStore, [HWPort237, HWPort4]>;
|
|
def : WriteRes<WriteLoad, [HWPort23]> { let Latency = 4; }
|
|
def : WriteRes<WriteMove, [HWPort0156]>;
|
|
def : WriteRes<WriteZero, []>;
|
|
|
|
defm : HWWriteResPair<WriteALU, HWPort0156, 1>;
|
|
defm : HWWriteResPair<WriteIMul, HWPort1, 3>;
|
|
defm : HWWriteResPair<WriteShift, HWPort056, 1>;
|
|
defm : HWWriteResPair<WriteJump, HWPort5, 1>;
|
|
|
|
// This is for simple LEAs with one or two input operands.
|
|
// The complex ones can only execute on port 1, and they require two cycles on
|
|
// the port to read all inputs. We don't model that.
|
|
def : WriteRes<WriteLEA, [HWPort15]>;
|
|
|
|
// This is quite rough, latency depends on the dividend.
|
|
def : WriteRes<WriteIDiv, [HWPort0, HWDivider]> {
|
|
let Latency = 25;
|
|
let ResourceCycles = [1, 10];
|
|
}
|
|
def : WriteRes<WriteIDivLd, [HWPort23, HWPort0, HWDivider]> {
|
|
let Latency = 29;
|
|
let ResourceCycles = [1, 1, 10];
|
|
}
|
|
|
|
// Scalar and vector floating point.
|
|
defm : HWWriteResPair<WriteFAdd, HWPort1, 3>;
|
|
defm : HWWriteResPair<WriteFMul, HWPort0, 5>;
|
|
defm : HWWriteResPair<WriteFDiv, HWPort0, 12>; // 10-14 cycles.
|
|
defm : HWWriteResPair<WriteFRcp, HWPort0, 5>;
|
|
defm : HWWriteResPair<WriteFSqrt, HWPort0, 15>;
|
|
defm : HWWriteResPair<WriteCvtF2I, HWPort1, 3>;
|
|
defm : HWWriteResPair<WriteCvtI2F, HWPort1, 4>;
|
|
defm : HWWriteResPair<WriteCvtF2F, HWPort1, 3>;
|
|
|
|
// Vector integer operations.
|
|
defm : HWWriteResPair<WriteVecShift, HWPort05, 1>;
|
|
defm : HWWriteResPair<WriteVecLogic, HWPort015, 1>;
|
|
defm : HWWriteResPair<WriteVecALU, HWPort15, 1>;
|
|
defm : HWWriteResPair<WriteVecIMul, HWPort0, 5>;
|
|
defm : HWWriteResPair<WriteShuffle, HWPort15, 1>;
|
|
|
|
def : WriteRes<WriteSystem, [HWPort0156]> { let Latency = 100; }
|
|
def : WriteRes<WriteMicrocoded, [HWPort0156]> { let Latency = 100; }
|
|
} // SchedModel
|