mirror of
https://github.com/c64scene-ar/llvm-6502.git
synced 2024-08-23 02:29:18 +00:00
The SSE rsqrt instruction (a fast reciprocal square root estimate) was grouped in the same scheduling IIC_SSE_SQRT* class as the accurate (but very slow) SSE sqrt instruction. For code which uses rsqrt (possibly with newton-raphson iterations) this poor scheduling was affecting performances. This patch splits off the rsqrt instruction from the sqrt instruction scheduling classes and creates new IIC_SSE_RSQER* classes with latency values based on Agner's table. Differential Revision: http://reviews.llvm.org/D5370 Patch by Simon Pilgrim. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@218517 91177308-0d34-0410-b5e6-96231b3b80d8
234 lines
7.5 KiB
TableGen
234 lines
7.5 KiB
TableGen
//=- X86ScheduleSLM.td - X86 Silvermont 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 Intel Silvermont to support
|
|
// instruction scheduling and other instruction cost heuristics.
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
def SLMModel : SchedMachineModel {
|
|
// All x86 instructions are modeled as a single micro-op, and SLM can decode 2
|
|
// instructions per cycle.
|
|
let IssueWidth = 2;
|
|
let MicroOpBufferSize = 32; // Based on the reorder buffer.
|
|
let LoadLatency = 3;
|
|
let MispredictPenalty = 10;
|
|
let PostRAScheduler = 1;
|
|
|
|
// For small loops, expand by a small factor to hide the backedge cost.
|
|
let LoopMicroOpBufferSize = 10;
|
|
|
|
// FIXME: SSE4 is unimplemented. This flag is set to allow
|
|
// the scheduler to assign a default model to unrecognized opcodes.
|
|
let CompleteModel = 0;
|
|
}
|
|
|
|
let SchedModel = SLMModel in {
|
|
|
|
// Silvermont has 5 reservation stations for micro-ops
|
|
|
|
def IEC_RSV0 : ProcResource<1>;
|
|
def IEC_RSV1 : ProcResource<1>;
|
|
def FPC_RSV0 : ProcResource<1> { let BufferSize = 1; }
|
|
def FPC_RSV1 : ProcResource<1> { let BufferSize = 1; }
|
|
def MEC_RSV : ProcResource<1>;
|
|
|
|
// Many micro-ops are capable of issuing on multiple ports.
|
|
def IEC_RSV01 : ProcResGroup<[IEC_RSV0, IEC_RSV1]>;
|
|
def FPC_RSV01 : ProcResGroup<[FPC_RSV0, FPC_RSV1]>;
|
|
|
|
def SMDivider : ProcResource<1>;
|
|
def SMFPMultiplier : ProcResource<1>;
|
|
def SMFPDivider : ProcResource<1>;
|
|
|
|
// Loads are 3 cycles, so ReadAfterLd registers needn't be available until 3
|
|
// cycles after the memory operand.
|
|
def : ReadAdvance<ReadAfterLd, 3>;
|
|
|
|
// 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 SMWriteResPair<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 MEC_RSV and adds 3 cycles to the
|
|
// latency.
|
|
def : WriteRes<SchedRW.Folded, [MEC_RSV, ExePort]> {
|
|
let Latency = !add(Lat, 3);
|
|
}
|
|
}
|
|
|
|
// A folded store needs a cycle on MEC_RSV for the store data, but it does not
|
|
// need an extra port cycle to recompute the address.
|
|
def : WriteRes<WriteRMW, [MEC_RSV]>;
|
|
|
|
def : WriteRes<WriteStore, [IEC_RSV01, MEC_RSV]>;
|
|
def : WriteRes<WriteLoad, [MEC_RSV]> { let Latency = 3; }
|
|
def : WriteRes<WriteMove, [IEC_RSV01]>;
|
|
def : WriteRes<WriteZero, []>;
|
|
|
|
defm : SMWriteResPair<WriteALU, IEC_RSV01, 1>;
|
|
defm : SMWriteResPair<WriteIMul, IEC_RSV1, 3>;
|
|
defm : SMWriteResPair<WriteShift, IEC_RSV0, 1>;
|
|
defm : SMWriteResPair<WriteJump, IEC_RSV1, 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, [IEC_RSV1]>;
|
|
|
|
// This is quite rough, latency depends on the dividend.
|
|
def : WriteRes<WriteIDiv, [IEC_RSV01, SMDivider]> {
|
|
let Latency = 25;
|
|
let ResourceCycles = [1, 25];
|
|
}
|
|
def : WriteRes<WriteIDivLd, [MEC_RSV, IEC_RSV01, SMDivider]> {
|
|
let Latency = 29;
|
|
let ResourceCycles = [1, 1, 25];
|
|
}
|
|
|
|
// Scalar and vector floating point.
|
|
defm : SMWriteResPair<WriteFAdd, FPC_RSV1, 3>;
|
|
defm : SMWriteResPair<WriteFRcp, FPC_RSV0, 5>;
|
|
defm : SMWriteResPair<WriteFRsqrt, FPC_RSV0, 5>;
|
|
defm : SMWriteResPair<WriteFSqrt, FPC_RSV0, 15>;
|
|
defm : SMWriteResPair<WriteCvtF2I, FPC_RSV01, 4>;
|
|
defm : SMWriteResPair<WriteCvtI2F, FPC_RSV01, 4>;
|
|
defm : SMWriteResPair<WriteCvtF2F, FPC_RSV01, 4>;
|
|
defm : SMWriteResPair<WriteFShuffle, FPC_RSV0, 1>;
|
|
defm : SMWriteResPair<WriteFBlend, FPC_RSV0, 1>;
|
|
|
|
// This is quite rough, latency depends on precision
|
|
def : WriteRes<WriteFMul, [FPC_RSV0, SMFPMultiplier]> {
|
|
let Latency = 5;
|
|
let ResourceCycles = [1, 2];
|
|
}
|
|
def : WriteRes<WriteFMulLd, [MEC_RSV, FPC_RSV0, SMFPMultiplier]> {
|
|
let Latency = 8;
|
|
let ResourceCycles = [1, 1, 2];
|
|
}
|
|
|
|
def : WriteRes<WriteFDiv, [FPC_RSV0, SMFPDivider]> {
|
|
let Latency = 34;
|
|
let ResourceCycles = [1, 34];
|
|
}
|
|
def : WriteRes<WriteFDivLd, [MEC_RSV, FPC_RSV0, SMFPDivider]> {
|
|
let Latency = 37;
|
|
let ResourceCycles = [1, 1, 34];
|
|
}
|
|
|
|
// Vector integer operations.
|
|
defm : SMWriteResPair<WriteVecShift, FPC_RSV0, 1>;
|
|
defm : SMWriteResPair<WriteVecLogic, FPC_RSV01, 1>;
|
|
defm : SMWriteResPair<WriteVecALU, FPC_RSV01, 1>;
|
|
defm : SMWriteResPair<WriteVecIMul, FPC_RSV0, 4>;
|
|
defm : SMWriteResPair<WriteShuffle, FPC_RSV0, 1>;
|
|
defm : SMWriteResPair<WriteBlend, FPC_RSV0, 1>;
|
|
defm : SMWriteResPair<WriteMPSAD, FPC_RSV0, 7>;
|
|
|
|
// String instructions.
|
|
// Packed Compare Implicit Length Strings, Return Mask
|
|
def : WriteRes<WritePCmpIStrM, [FPC_RSV0]> {
|
|
let Latency = 13;
|
|
let ResourceCycles = [13];
|
|
}
|
|
def : WriteRes<WritePCmpIStrMLd, [FPC_RSV0, MEC_RSV]> {
|
|
let Latency = 13;
|
|
let ResourceCycles = [13, 1];
|
|
}
|
|
|
|
// Packed Compare Explicit Length Strings, Return Mask
|
|
def : WriteRes<WritePCmpEStrM, [FPC_RSV0]> {
|
|
let Latency = 17;
|
|
let ResourceCycles = [17];
|
|
}
|
|
def : WriteRes<WritePCmpEStrMLd, [FPC_RSV0, MEC_RSV]> {
|
|
let Latency = 17;
|
|
let ResourceCycles = [17, 1];
|
|
}
|
|
|
|
// Packed Compare Implicit Length Strings, Return Index
|
|
def : WriteRes<WritePCmpIStrI, [FPC_RSV0]> {
|
|
let Latency = 17;
|
|
let ResourceCycles = [17];
|
|
}
|
|
def : WriteRes<WritePCmpIStrILd, [FPC_RSV0, MEC_RSV]> {
|
|
let Latency = 17;
|
|
let ResourceCycles = [17, 1];
|
|
}
|
|
|
|
// Packed Compare Explicit Length Strings, Return Index
|
|
def : WriteRes<WritePCmpEStrI, [FPC_RSV0]> {
|
|
let Latency = 21;
|
|
let ResourceCycles = [21];
|
|
}
|
|
def : WriteRes<WritePCmpEStrILd, [FPC_RSV0, MEC_RSV]> {
|
|
let Latency = 21;
|
|
let ResourceCycles = [21, 1];
|
|
}
|
|
|
|
// AES Instructions.
|
|
def : WriteRes<WriteAESDecEnc, [FPC_RSV0]> {
|
|
let Latency = 8;
|
|
let ResourceCycles = [5];
|
|
}
|
|
def : WriteRes<WriteAESDecEncLd, [FPC_RSV0, MEC_RSV]> {
|
|
let Latency = 8;
|
|
let ResourceCycles = [5, 1];
|
|
}
|
|
|
|
def : WriteRes<WriteAESIMC, [FPC_RSV0]> {
|
|
let Latency = 8;
|
|
let ResourceCycles = [5];
|
|
}
|
|
def : WriteRes<WriteAESIMCLd, [FPC_RSV0, MEC_RSV]> {
|
|
let Latency = 8;
|
|
let ResourceCycles = [5, 1];
|
|
}
|
|
|
|
def : WriteRes<WriteAESKeyGen, [FPC_RSV0]> {
|
|
let Latency = 8;
|
|
let ResourceCycles = [5];
|
|
}
|
|
def : WriteRes<WriteAESKeyGenLd, [FPC_RSV0, MEC_RSV]> {
|
|
let Latency = 8;
|
|
let ResourceCycles = [5, 1];
|
|
}
|
|
|
|
// Carry-less multiplication instructions.
|
|
def : WriteRes<WriteCLMul, [FPC_RSV0]> {
|
|
let Latency = 10;
|
|
let ResourceCycles = [10];
|
|
}
|
|
def : WriteRes<WriteCLMulLd, [FPC_RSV0, MEC_RSV]> {
|
|
let Latency = 10;
|
|
let ResourceCycles = [10, 1];
|
|
}
|
|
|
|
|
|
def : WriteRes<WriteSystem, [FPC_RSV0]> { let Latency = 100; }
|
|
def : WriteRes<WriteMicrocoded, [FPC_RSV0]> { let Latency = 100; }
|
|
def : WriteRes<WriteFence, [MEC_RSV]>;
|
|
def : WriteRes<WriteNop, []>;
|
|
|
|
// AVX is not supported on that architecture, but we should define the basic
|
|
// scheduling resources anyway.
|
|
def : WriteRes<WriteIMulH, [FPC_RSV0]>;
|
|
defm : SMWriteResPair<WriteVarBlend, FPC_RSV0, 1>;
|
|
defm : SMWriteResPair<WriteFVarBlend, FPC_RSV0, 1>;
|
|
defm : SMWriteResPair<WriteFShuffle256, FPC_RSV0, 1>;
|
|
defm : SMWriteResPair<WriteShuffle256, FPC_RSV0, 1>;
|
|
defm : SMWriteResPair<WriteVarVecShift, FPC_RSV0, 1>;
|
|
} // SchedModel
|