1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-24 03:33:20 +01:00
llvm-mirror/include/llvm/MC/MCSchedule.h
Andrea Di Biagio 5d3783c0d0 [MC][X86] Correctly model additional operand latency caused by transfer delays from the integer to the floating point unit.
This patch adds a new ReadAdvance definition named ReadInt2Fpu.
ReadInt2Fpu allows x86 scheduling models to accurately describe delays caused by
data transfers from the integer unit to the floating point unit.
ReadInt2Fpu currently defaults to a delay of zero cycles (i.e. no delay) for all
x86 models excluding BtVer2. That means, this patch is only a functional change
for the Jaguar cpu model only.

Tablegen definitions for instructions (V)PINSR* have been updated to account for
the new ReadInt2Fpu. That read is mapped to the the GPR input operand.
On Jaguar, int-to-fpu transfers are modeled as a +6cy delay. Before this patch,
that extra delay was added to the opcode latency. In practice, the insert opcode
only executes for 1cy. Most of the actual latency is actually contributed by the
so-called operand-latency. According to the AMD SOG for family 16h, (V)PINSR*
latency is defined by expression f+1, where f is defined as a forwarding delay
from the integer unit to the fpu.

When printing instruction latency from MCA (see InstructionInfoView.cpp) and LLC
(only when flag -print-schedule is speified), we now need to account for any
extra forwarding delays. We do this by checking if scheduling classes declare
any negative ReadAdvance entries. Quoting a code comment in TargetSchedule.td:
"A negative advance effectively increases latency, which may be used for
cross-domain stalls". When computing the instruction latency for the purpose of
our scheduling tests, we now add any extra delay to the formula. This avoids
regressing existing codegen and mca schedule tests. It comes with the cost of an
extra (but very simple) hook in MCSchedModel.

Differential Revision: https://reviews.llvm.org/D57056

llvm-svn: 351965
2019-01-23 16:35:07 +00:00

386 lines
16 KiB
C++

//===-- llvm/MC/MCSchedule.h - Scheduling -----------------------*- C++ -*-===//
//
// 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 classes used to describe a subtarget's machine model
// for scheduling and other instruction cost heuristics.
//
//===----------------------------------------------------------------------===//
#ifndef LLVM_MC_MCSCHEDULE_H
#define LLVM_MC_MCSCHEDULE_H
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/Optional.h"
#include "llvm/Config/llvm-config.h"
#include "llvm/Support/DataTypes.h"
#include <cassert>
namespace llvm {
struct InstrItinerary;
class MCSubtargetInfo;
class MCInstrInfo;
class MCInst;
class InstrItineraryData;
/// Define a kind of processor resource that will be modeled by the scheduler.
struct MCProcResourceDesc {
const char *Name;
unsigned NumUnits; // Number of resource of this kind
unsigned SuperIdx; // Index of the resources kind that contains this kind.
// Number of resources that may be buffered.
//
// Buffered resources (BufferSize != 0) may be consumed at some indeterminate
// cycle after dispatch. This should be used for out-of-order cpus when
// instructions that use this resource can be buffered in a reservaton
// station.
//
// Unbuffered resources (BufferSize == 0) always consume their resource some
// fixed number of cycles after dispatch. If a resource is unbuffered, then
// the scheduler will avoid scheduling instructions with conflicting resources
// in the same cycle. This is for in-order cpus, or the in-order portion of
// an out-of-order cpus.
int BufferSize;
// If the resource has sub-units, a pointer to the first element of an array
// of `NumUnits` elements containing the ProcResourceIdx of the sub units.
// nullptr if the resource does not have sub-units.
const unsigned *SubUnitsIdxBegin;
bool operator==(const MCProcResourceDesc &Other) const {
return NumUnits == Other.NumUnits && SuperIdx == Other.SuperIdx
&& BufferSize == Other.BufferSize;
}
};
/// Identify one of the processor resource kinds consumed by a particular
/// scheduling class for the specified number of cycles.
struct MCWriteProcResEntry {
uint16_t ProcResourceIdx;
uint16_t Cycles;
bool operator==(const MCWriteProcResEntry &Other) const {
return ProcResourceIdx == Other.ProcResourceIdx && Cycles == Other.Cycles;
}
};
/// Specify the latency in cpu cycles for a particular scheduling class and def
/// index. -1 indicates an invalid latency. Heuristics would typically consider
/// an instruction with invalid latency to have infinite latency. Also identify
/// the WriteResources of this def. When the operand expands to a sequence of
/// writes, this ID is the last write in the sequence.
struct MCWriteLatencyEntry {
int16_t Cycles;
uint16_t WriteResourceID;
bool operator==(const MCWriteLatencyEntry &Other) const {
return Cycles == Other.Cycles && WriteResourceID == Other.WriteResourceID;
}
};
/// Specify the number of cycles allowed after instruction issue before a
/// particular use operand reads its registers. This effectively reduces the
/// write's latency. Here we allow negative cycles for corner cases where
/// latency increases. This rule only applies when the entry's WriteResource
/// matches the write's WriteResource.
///
/// MCReadAdvanceEntries are sorted first by operand index (UseIdx), then by
/// WriteResourceIdx.
struct MCReadAdvanceEntry {
unsigned UseIdx;
unsigned WriteResourceID;
int Cycles;
bool operator==(const MCReadAdvanceEntry &Other) const {
return UseIdx == Other.UseIdx && WriteResourceID == Other.WriteResourceID
&& Cycles == Other.Cycles;
}
};
/// Summarize the scheduling resources required for an instruction of a
/// particular scheduling class.
///
/// Defined as an aggregate struct for creating tables with initializer lists.
struct MCSchedClassDesc {
static const unsigned short InvalidNumMicroOps = (1U << 14) - 1;
static const unsigned short VariantNumMicroOps = InvalidNumMicroOps - 1;
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
const char* Name;
#endif
uint16_t NumMicroOps : 14;
bool BeginGroup : 1;
bool EndGroup : 1;
uint16_t WriteProcResIdx; // First index into WriteProcResTable.
uint16_t NumWriteProcResEntries;
uint16_t WriteLatencyIdx; // First index into WriteLatencyTable.
uint16_t NumWriteLatencyEntries;
uint16_t ReadAdvanceIdx; // First index into ReadAdvanceTable.
uint16_t NumReadAdvanceEntries;
bool isValid() const {
return NumMicroOps != InvalidNumMicroOps;
}
bool isVariant() const {
return NumMicroOps == VariantNumMicroOps;
}
};
/// Specify the cost of a register definition in terms of number of physical
/// register allocated at register renaming stage. For example, AMD Jaguar.
/// natively supports 128-bit data types, and operations on 256-bit registers
/// (i.e. YMM registers) are internally split into two COPs (complex operations)
/// and each COP updates a physical register. Basically, on Jaguar, a YMM
/// register write effectively consumes two physical registers. That means,
/// the cost of a YMM write in the BtVer2 model is 2.
struct MCRegisterCostEntry {
unsigned RegisterClassID;
unsigned Cost;
bool AllowMoveElimination;
};
/// A register file descriptor.
///
/// This struct allows to describe processor register files. In particular, it
/// helps describing the size of the register file, as well as the cost of
/// allocating a register file at register renaming stage.
/// FIXME: this struct can be extended to provide information about the number
/// of read/write ports to the register file. A value of zero for field
/// 'NumPhysRegs' means: this register file has an unbounded number of physical
/// registers.
struct MCRegisterFileDesc {
const char *Name;
uint16_t NumPhysRegs;
uint16_t NumRegisterCostEntries;
// Index of the first cost entry in MCExtraProcessorInfo::RegisterCostTable.
uint16_t RegisterCostEntryIdx;
// A value of zero means: there is no limit in the number of moves that can be
// eliminated every cycle.
uint16_t MaxMovesEliminatedPerCycle;
// Ture if this register file only knows how to optimize register moves from
// known zero registers.
bool AllowZeroMoveEliminationOnly;
};
/// Provide extra details about the machine processor.
///
/// This is a collection of "optional" processor information that is not
/// normally used by the LLVM machine schedulers, but that can be consumed by
/// external tools like llvm-mca to improve the quality of the peformance
/// analysis.
struct MCExtraProcessorInfo {
// Actual size of the reorder buffer in hardware.
unsigned ReorderBufferSize;
// Number of instructions retired per cycle.
unsigned MaxRetirePerCycle;
const MCRegisterFileDesc *RegisterFiles;
unsigned NumRegisterFiles;
const MCRegisterCostEntry *RegisterCostTable;
unsigned NumRegisterCostEntries;
unsigned LoadQueueID;
unsigned StoreQueueID;
};
/// Machine model for scheduling, bundling, and heuristics.
///
/// The machine model directly provides basic information about the
/// microarchitecture to the scheduler in the form of properties. It also
/// optionally refers to scheduler resource tables and itinerary
/// tables. Scheduler resource tables model the latency and cost for each
/// instruction type. Itinerary tables are an independent mechanism that
/// provides a detailed reservation table describing each cycle of instruction
/// execution. Subtargets may define any or all of the above categories of data
/// depending on the type of CPU and selected scheduler.
///
/// The machine independent properties defined here are used by the scheduler as
/// an abstract machine model. A real micro-architecture has a number of
/// buffers, queues, and stages. Declaring that a given machine-independent
/// abstract property corresponds to a specific physical property across all
/// subtargets can't be done. Nonetheless, the abstract model is
/// useful. Futhermore, subtargets typically extend this model with processor
/// specific resources to model any hardware features that can be exploited by
/// sceduling heuristics and aren't sufficiently represented in the abstract.
///
/// The abstract pipeline is built around the notion of an "issue point". This
/// is merely a reference point for counting machine cycles. The physical
/// machine will have pipeline stages that delay execution. The scheduler does
/// not model those delays because they are irrelevant as long as they are
/// consistent. Inaccuracies arise when instructions have different execution
/// delays relative to each other, in addition to their intrinsic latency. Those
/// special cases can be handled by TableGen constructs such as, ReadAdvance,
/// which reduces latency when reading data, and ResourceCycles, which consumes
/// a processor resource when writing data for a number of abstract
/// cycles.
///
/// TODO: One tool currently missing is the ability to add a delay to
/// ResourceCycles. That would be easy to add and would likely cover all cases
/// currently handled by the legacy itinerary tables.
///
/// A note on out-of-order execution and, more generally, instruction
/// buffers. Part of the CPU pipeline is always in-order. The issue point, which
/// is the point of reference for counting cycles, only makes sense as an
/// in-order part of the pipeline. Other parts of the pipeline are sometimes
/// falling behind and sometimes catching up. It's only interesting to model
/// those other, decoupled parts of the pipeline if they may be predictably
/// resource constrained in a way that the scheduler can exploit.
///
/// The LLVM machine model distinguishes between in-order constraints and
/// out-of-order constraints so that the target's scheduling strategy can apply
/// appropriate heuristics. For a well-balanced CPU pipeline, out-of-order
/// resources would not typically be treated as a hard scheduling
/// constraint. For example, in the GenericScheduler, a delay caused by limited
/// out-of-order resources is not directly reflected in the number of cycles
/// that the scheduler sees between issuing an instruction and its dependent
/// instructions. In other words, out-of-order resources don't directly increase
/// the latency between pairs of instructions. However, they can still be used
/// to detect potential bottlenecks across a sequence of instructions and bias
/// the scheduling heuristics appropriately.
struct MCSchedModel {
// IssueWidth is the maximum number of instructions that may be scheduled in
// the same per-cycle group. This is meant to be a hard in-order constraint
// (a.k.a. "hazard"). In the GenericScheduler strategy, no more than
// IssueWidth micro-ops can ever be scheduled in a particular cycle.
//
// In practice, IssueWidth is useful to model any bottleneck between the
// decoder (after micro-op expansion) and the out-of-order reservation
// stations or the decoder bandwidth itself. If the total number of
// reservation stations is also a bottleneck, or if any other pipeline stage
// has a bandwidth limitation, then that can be naturally modeled by adding an
// out-of-order processor resource.
unsigned IssueWidth;
static const unsigned DefaultIssueWidth = 1;
// MicroOpBufferSize is the number of micro-ops that the processor may buffer
// for out-of-order execution.
//
// "0" means operations that are not ready in this cycle are not considered
// for scheduling (they go in the pending queue). Latency is paramount. This
// may be more efficient if many instructions are pending in a schedule.
//
// "1" means all instructions are considered for scheduling regardless of
// whether they are ready in this cycle. Latency still causes issue stalls,
// but we balance those stalls against other heuristics.
//
// "> 1" means the processor is out-of-order. This is a machine independent
// estimate of highly machine specific characteristics such as the register
// renaming pool and reorder buffer.
unsigned MicroOpBufferSize;
static const unsigned DefaultMicroOpBufferSize = 0;
// LoopMicroOpBufferSize is the number of micro-ops that the processor may
// buffer for optimized loop execution. More generally, this represents the
// optimal number of micro-ops in a loop body. A loop may be partially
// unrolled to bring the count of micro-ops in the loop body closer to this
// number.
unsigned LoopMicroOpBufferSize;
static const unsigned DefaultLoopMicroOpBufferSize = 0;
// LoadLatency is the expected latency of load instructions.
unsigned LoadLatency;
static const unsigned DefaultLoadLatency = 4;
// HighLatency is the expected latency of "very high latency" operations.
// See TargetInstrInfo::isHighLatencyDef().
// By default, this is set to an arbitrarily high number of cycles
// likely to have some impact on scheduling heuristics.
unsigned HighLatency;
static const unsigned DefaultHighLatency = 10;
// MispredictPenalty is the typical number of extra cycles the processor
// takes to recover from a branch misprediction.
unsigned MispredictPenalty;
static const unsigned DefaultMispredictPenalty = 10;
bool PostRAScheduler; // default value is false
bool CompleteModel;
unsigned ProcID;
const MCProcResourceDesc *ProcResourceTable;
const MCSchedClassDesc *SchedClassTable;
unsigned NumProcResourceKinds;
unsigned NumSchedClasses;
// Instruction itinerary tables used by InstrItineraryData.
friend class InstrItineraryData;
const InstrItinerary *InstrItineraries;
const MCExtraProcessorInfo *ExtraProcessorInfo;
bool hasExtraProcessorInfo() const { return ExtraProcessorInfo; }
unsigned getProcessorID() const { return ProcID; }
/// Does this machine model include instruction-level scheduling.
bool hasInstrSchedModel() const { return SchedClassTable; }
const MCExtraProcessorInfo &getExtraProcessorInfo() const {
assert(hasExtraProcessorInfo() &&
"No extra information available for this model");
return *ExtraProcessorInfo;
}
/// Return true if this machine model data for all instructions with a
/// scheduling class (itinerary class or SchedRW list).
bool isComplete() const { return CompleteModel; }
/// Return true if machine supports out of order execution.
bool isOutOfOrder() const { return MicroOpBufferSize > 1; }
unsigned getNumProcResourceKinds() const {
return NumProcResourceKinds;
}
const MCProcResourceDesc *getProcResource(unsigned ProcResourceIdx) const {
assert(hasInstrSchedModel() && "No scheduling machine model");
assert(ProcResourceIdx < NumProcResourceKinds && "bad proc resource idx");
return &ProcResourceTable[ProcResourceIdx];
}
const MCSchedClassDesc *getSchedClassDesc(unsigned SchedClassIdx) const {
assert(hasInstrSchedModel() && "No scheduling machine model");
assert(SchedClassIdx < NumSchedClasses && "bad scheduling class idx");
return &SchedClassTable[SchedClassIdx];
}
/// Returns the latency value for the scheduling class.
static int computeInstrLatency(const MCSubtargetInfo &STI,
const MCSchedClassDesc &SCDesc);
int computeInstrLatency(const MCSubtargetInfo &STI, unsigned SClass) const;
int computeInstrLatency(const MCSubtargetInfo &STI, const MCInstrInfo &MCII,
const MCInst &Inst) const;
// Returns the reciprocal throughput information from a MCSchedClassDesc.
static double
getReciprocalThroughput(const MCSubtargetInfo &STI,
const MCSchedClassDesc &SCDesc);
static double
getReciprocalThroughput(unsigned SchedClass, const InstrItineraryData &IID);
double
getReciprocalThroughput(const MCSubtargetInfo &STI, const MCInstrInfo &MCII,
const MCInst &Inst) const;
/// Returns the maximum forwarding delay for register reads dependent on
/// writes of scheduling class WriteResourceIdx.
static unsigned getForwardingDelayCycles(ArrayRef<MCReadAdvanceEntry> Entries,
unsigned WriteResourceIdx = 0);
/// Returns the default initialized model.
static const MCSchedModel &GetDefaultSchedModel() { return Default; }
static const MCSchedModel Default;
};
} // namespace llvm
#endif