mirror of
https://github.com/RPCS3/llvm-mirror.git
synced 2024-11-24 03:33:20 +01:00
b4a5e33ee0
sequence - target independent framework When the DAGcombiner selects instruction sequences it could increase the critical path or resource len. For example, on arm64 there are multiply-accumulate instructions (madd, msub). If e.g. the equivalent multiply-add sequence is not on the crictial path it makes sense to select it instead of the combined, single accumulate instruction (madd/msub). The reason is that the conversion from add+mul to the madd could lengthen the critical path by the latency of the multiply. But the DAGCombiner would always combine and select the madd/msub instruction. This patch uses machine trace metrics to estimate critical path length and resource length of an original instruction sequence vs a combined instruction sequence and picks the faster code based on its estimates. This patch only commits the target independent framework that evaluates and selects code sequences. The machine instruction combiner is turned off for all targets and expected to evolve over time by gradually handling DAGCombiner pattern in the target specific code. This framework lays the groundwork for fixing rdar://16319955 llvm-svn: 214666
399 lines
16 KiB
C++
399 lines
16 KiB
C++
//===- lib/CodeGen/MachineTraceMetrics.h - Super-scalar metrics -*- C++ -*-===//
|
|
//
|
|
// 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 interface for the MachineTraceMetrics analysis pass
|
|
// that estimates CPU resource usage and critical data dependency paths through
|
|
// preferred traces. This is useful for super-scalar CPUs where execution speed
|
|
// can be limited both by data dependencies and by limited execution resources.
|
|
//
|
|
// Out-of-order CPUs will often be executing instructions from multiple basic
|
|
// blocks at the same time. This makes it difficult to estimate the resource
|
|
// usage accurately in a single basic block. Resources can be estimated better
|
|
// by looking at a trace through the current basic block.
|
|
//
|
|
// For every block, the MachineTraceMetrics pass will pick a preferred trace
|
|
// that passes through the block. The trace is chosen based on loop structure,
|
|
// branch probabilities, and resource usage. The intention is to pick likely
|
|
// traces that would be the most affected by code transformations.
|
|
//
|
|
// It is expensive to compute a full arbitrary trace for every block, so to
|
|
// save some computations, traces are chosen to be convergent. This means that
|
|
// if the traces through basic blocks A and B ever cross when moving away from
|
|
// A and B, they never diverge again. This applies in both directions - If the
|
|
// traces meet above A and B, they won't diverge when going further back.
|
|
//
|
|
// Traces tend to align with loops. The trace through a block in an inner loop
|
|
// will begin at the loop entry block and end at a back edge. If there are
|
|
// nested loops, the trace may begin and end at those instead.
|
|
//
|
|
// For each trace, we compute the critical path length, which is the number of
|
|
// cycles required to execute the trace when execution is limited by data
|
|
// dependencies only. We also compute the resource height, which is the number
|
|
// of cycles required to execute all instructions in the trace when ignoring
|
|
// data dependencies.
|
|
//
|
|
// Every instruction in the current block has a slack - the number of cycles
|
|
// execution of the instruction can be delayed without extending the critical
|
|
// path.
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#ifndef LLVM_CODEGEN_MACHINE_TRACE_METRICS_H
|
|
#define LLVM_CODEGEN_MACHINE_TRACE_METRICS_H
|
|
|
|
#include "llvm/ADT/ArrayRef.h"
|
|
#include "llvm/ADT/DenseMap.h"
|
|
#include "llvm/CodeGen/MachineFunctionPass.h"
|
|
#include "llvm/CodeGen/TargetSchedule.h"
|
|
|
|
namespace llvm {
|
|
|
|
class InstrItineraryData;
|
|
class MachineBasicBlock;
|
|
class MachineInstr;
|
|
class MachineLoop;
|
|
class MachineLoopInfo;
|
|
class MachineRegisterInfo;
|
|
class TargetInstrInfo;
|
|
class TargetRegisterInfo;
|
|
class raw_ostream;
|
|
|
|
class MachineTraceMetrics : public MachineFunctionPass {
|
|
const MachineFunction *MF;
|
|
const TargetInstrInfo *TII;
|
|
const TargetRegisterInfo *TRI;
|
|
const MachineRegisterInfo *MRI;
|
|
const MachineLoopInfo *Loops;
|
|
TargetSchedModel SchedModel;
|
|
|
|
public:
|
|
class Ensemble;
|
|
class Trace;
|
|
static char ID;
|
|
MachineTraceMetrics();
|
|
void getAnalysisUsage(AnalysisUsage&) const override;
|
|
bool runOnMachineFunction(MachineFunction&) override;
|
|
void releaseMemory() override;
|
|
void verifyAnalysis() const override;
|
|
|
|
friend class Ensemble;
|
|
friend class Trace;
|
|
|
|
/// Per-basic block information that doesn't depend on the trace through the
|
|
/// block.
|
|
struct FixedBlockInfo {
|
|
/// The number of non-trivial instructions in the block.
|
|
/// Doesn't count PHI and COPY instructions that are likely to be removed.
|
|
unsigned InstrCount;
|
|
|
|
/// True when the block contains calls.
|
|
bool HasCalls;
|
|
|
|
FixedBlockInfo() : InstrCount(~0u), HasCalls(false) {}
|
|
|
|
/// Returns true when resource information for this block has been computed.
|
|
bool hasResources() const { return InstrCount != ~0u; }
|
|
|
|
/// Invalidate resource information.
|
|
void invalidate() { InstrCount = ~0u; }
|
|
};
|
|
|
|
/// Get the fixed resource information about MBB. Compute it on demand.
|
|
const FixedBlockInfo *getResources(const MachineBasicBlock*);
|
|
|
|
/// Get the scaled number of cycles used per processor resource in MBB.
|
|
/// This is an array with SchedModel.getNumProcResourceKinds() entries.
|
|
/// The getResources() function above must have been called first.
|
|
///
|
|
/// These numbers have already been scaled by SchedModel.getResourceFactor().
|
|
ArrayRef<unsigned> getProcResourceCycles(unsigned MBBNum) const;
|
|
|
|
/// A virtual register or regunit required by a basic block or its trace
|
|
/// successors.
|
|
struct LiveInReg {
|
|
/// The virtual register required, or a register unit.
|
|
unsigned Reg;
|
|
|
|
/// For virtual registers: Minimum height of the defining instruction.
|
|
/// For regunits: Height of the highest user in the trace.
|
|
unsigned Height;
|
|
|
|
LiveInReg(unsigned Reg, unsigned Height = 0) : Reg(Reg), Height(Height) {}
|
|
};
|
|
|
|
/// Per-basic block information that relates to a specific trace through the
|
|
/// block. Convergent traces means that only one of these is required per
|
|
/// block in a trace ensemble.
|
|
struct TraceBlockInfo {
|
|
/// Trace predecessor, or NULL for the first block in the trace.
|
|
/// Valid when hasValidDepth().
|
|
const MachineBasicBlock *Pred;
|
|
|
|
/// Trace successor, or NULL for the last block in the trace.
|
|
/// Valid when hasValidHeight().
|
|
const MachineBasicBlock *Succ;
|
|
|
|
/// The block number of the head of the trace. (When hasValidDepth()).
|
|
unsigned Head;
|
|
|
|
/// The block number of the tail of the trace. (When hasValidHeight()).
|
|
unsigned Tail;
|
|
|
|
/// Accumulated number of instructions in the trace above this block.
|
|
/// Does not include instructions in this block.
|
|
unsigned InstrDepth;
|
|
|
|
/// Accumulated number of instructions in the trace below this block.
|
|
/// Includes instructions in this block.
|
|
unsigned InstrHeight;
|
|
|
|
TraceBlockInfo() :
|
|
Pred(nullptr), Succ(nullptr),
|
|
InstrDepth(~0u), InstrHeight(~0u),
|
|
HasValidInstrDepths(false), HasValidInstrHeights(false) {}
|
|
|
|
/// Returns true if the depth resources have been computed from the trace
|
|
/// above this block.
|
|
bool hasValidDepth() const { return InstrDepth != ~0u; }
|
|
|
|
/// Returns true if the height resources have been computed from the trace
|
|
/// below this block.
|
|
bool hasValidHeight() const { return InstrHeight != ~0u; }
|
|
|
|
/// Invalidate depth resources when some block above this one has changed.
|
|
void invalidateDepth() { InstrDepth = ~0u; HasValidInstrDepths = false; }
|
|
|
|
/// Invalidate height resources when a block below this one has changed.
|
|
void invalidateHeight() { InstrHeight = ~0u; HasValidInstrHeights = false; }
|
|
|
|
/// Assuming that this is a dominator of TBI, determine if it contains
|
|
/// useful instruction depths. A dominating block can be above the current
|
|
/// trace head, and any dependencies from such a far away dominator are not
|
|
/// expected to affect the critical path.
|
|
///
|
|
/// Also returns true when TBI == this.
|
|
bool isUsefulDominator(const TraceBlockInfo &TBI) const {
|
|
// The trace for TBI may not even be calculated yet.
|
|
if (!hasValidDepth() || !TBI.hasValidDepth())
|
|
return false;
|
|
// Instruction depths are only comparable if the traces share a head.
|
|
if (Head != TBI.Head)
|
|
return false;
|
|
// It is almost always the case that TBI belongs to the same trace as
|
|
// this block, but rare convoluted cases involving irreducible control
|
|
// flow, a dominator may share a trace head without actually being on the
|
|
// same trace as TBI. This is not a big problem as long as it doesn't
|
|
// increase the instruction depth.
|
|
return HasValidInstrDepths && InstrDepth <= TBI.InstrDepth;
|
|
}
|
|
|
|
// Data-dependency-related information. Per-instruction depth and height
|
|
// are computed from data dependencies in the current trace, using
|
|
// itinerary data.
|
|
|
|
/// Instruction depths have been computed. This implies hasValidDepth().
|
|
bool HasValidInstrDepths;
|
|
|
|
/// Instruction heights have been computed. This implies hasValidHeight().
|
|
bool HasValidInstrHeights;
|
|
|
|
/// Critical path length. This is the number of cycles in the longest data
|
|
/// dependency chain through the trace. This is only valid when both
|
|
/// HasValidInstrDepths and HasValidInstrHeights are set.
|
|
unsigned CriticalPath;
|
|
|
|
/// Live-in registers. These registers are defined above the current block
|
|
/// and used by this block or a block below it.
|
|
/// This does not include PHI uses in the current block, but it does
|
|
/// include PHI uses in deeper blocks.
|
|
SmallVector<LiveInReg, 4> LiveIns;
|
|
|
|
void print(raw_ostream&) const;
|
|
};
|
|
|
|
/// InstrCycles represents the cycle height and depth of an instruction in a
|
|
/// trace.
|
|
struct InstrCycles {
|
|
/// Earliest issue cycle as determined by data dependencies and instruction
|
|
/// latencies from the beginning of the trace. Data dependencies from
|
|
/// before the trace are not included.
|
|
unsigned Depth;
|
|
|
|
/// Minimum number of cycles from this instruction is issued to the of the
|
|
/// trace, as determined by data dependencies and instruction latencies.
|
|
unsigned Height;
|
|
};
|
|
|
|
/// A trace represents a plausible sequence of executed basic blocks that
|
|
/// passes through the current basic block one. The Trace class serves as a
|
|
/// handle to internal cached data structures.
|
|
class Trace {
|
|
Ensemble &TE;
|
|
TraceBlockInfo &TBI;
|
|
|
|
unsigned getBlockNum() const { return &TBI - &TE.BlockInfo[0]; }
|
|
|
|
public:
|
|
explicit Trace(Ensemble &te, TraceBlockInfo &tbi) : TE(te), TBI(tbi) {}
|
|
void print(raw_ostream&) const;
|
|
|
|
/// Compute the total number of instructions in the trace.
|
|
unsigned getInstrCount() const {
|
|
return TBI.InstrDepth + TBI.InstrHeight;
|
|
}
|
|
|
|
/// Return the resource depth of the top/bottom of the trace center block.
|
|
/// This is the number of cycles required to execute all instructions from
|
|
/// the trace head to the trace center block. The resource depth only
|
|
/// considers execution resources, it ignores data dependencies.
|
|
/// When Bottom is set, instructions in the trace center block are included.
|
|
unsigned getResourceDepth(bool Bottom) const;
|
|
|
|
/// Return the resource length of the trace. This is the number of cycles
|
|
/// required to execute the instructions in the trace if they were all
|
|
/// independent, exposing the maximum instruction-level parallelism.
|
|
///
|
|
/// Any blocks in Extrablocks are included as if they were part of the
|
|
/// trace. Likewise, extra resources required by the specified scheduling
|
|
/// classes are included. For the caller to account for extra machine
|
|
/// instructions, it must first resolve each instruction's scheduling class.
|
|
unsigned getResourceLength(
|
|
ArrayRef<const MachineBasicBlock *> Extrablocks = None,
|
|
ArrayRef<const MCSchedClassDesc *> ExtraInstrs = None,
|
|
ArrayRef<const MCSchedClassDesc *> RemoveInstrs = None) const;
|
|
|
|
/// Return the length of the (data dependency) critical path through the
|
|
/// trace.
|
|
unsigned getCriticalPath() const { return TBI.CriticalPath; }
|
|
|
|
/// Return the depth and height of MI. The depth is only valid for
|
|
/// instructions in or above the trace center block. The height is only
|
|
/// valid for instructions in or below the trace center block.
|
|
InstrCycles getInstrCycles(const MachineInstr *MI) const {
|
|
return TE.Cycles.lookup(MI);
|
|
}
|
|
|
|
/// Return the slack of MI. This is the number of cycles MI can be delayed
|
|
/// before the critical path becomes longer.
|
|
/// MI must be an instruction in the trace center block.
|
|
unsigned getInstrSlack(const MachineInstr *MI) const;
|
|
|
|
/// Return the Depth of a PHI instruction in a trace center block successor.
|
|
/// The PHI does not have to be part of the trace.
|
|
unsigned getPHIDepth(const MachineInstr *PHI) const;
|
|
|
|
/// A dependence is useful if the basic block of the defining instruction
|
|
/// is part of the trace of the user instruction. It is assumed that DefMI
|
|
/// dominates UseMI (see also isUsefulDominator).
|
|
bool isDepInTrace(const MachineInstr *DefMI,
|
|
const MachineInstr *UseMI) const;
|
|
};
|
|
|
|
/// A trace ensemble is a collection of traces selected using the same
|
|
/// strategy, for example 'minimum resource height'. There is one trace for
|
|
/// every block in the function.
|
|
class Ensemble {
|
|
SmallVector<TraceBlockInfo, 4> BlockInfo;
|
|
DenseMap<const MachineInstr*, InstrCycles> Cycles;
|
|
SmallVector<unsigned, 0> ProcResourceDepths;
|
|
SmallVector<unsigned, 0> ProcResourceHeights;
|
|
friend class Trace;
|
|
|
|
void computeTrace(const MachineBasicBlock*);
|
|
void computeDepthResources(const MachineBasicBlock*);
|
|
void computeHeightResources(const MachineBasicBlock*);
|
|
unsigned computeCrossBlockCriticalPath(const TraceBlockInfo&);
|
|
void computeInstrDepths(const MachineBasicBlock*);
|
|
void computeInstrHeights(const MachineBasicBlock*);
|
|
void addLiveIns(const MachineInstr *DefMI, unsigned DefOp,
|
|
ArrayRef<const MachineBasicBlock*> Trace);
|
|
|
|
protected:
|
|
MachineTraceMetrics &MTM;
|
|
virtual const MachineBasicBlock *pickTracePred(const MachineBasicBlock*) =0;
|
|
virtual const MachineBasicBlock *pickTraceSucc(const MachineBasicBlock*) =0;
|
|
explicit Ensemble(MachineTraceMetrics*);
|
|
const MachineLoop *getLoopFor(const MachineBasicBlock*) const;
|
|
const TraceBlockInfo *getDepthResources(const MachineBasicBlock*) const;
|
|
const TraceBlockInfo *getHeightResources(const MachineBasicBlock*) const;
|
|
ArrayRef<unsigned> getProcResourceDepths(unsigned MBBNum) const;
|
|
ArrayRef<unsigned> getProcResourceHeights(unsigned MBBNum) const;
|
|
|
|
public:
|
|
virtual ~Ensemble();
|
|
virtual const char *getName() const =0;
|
|
void print(raw_ostream&) const;
|
|
void invalidate(const MachineBasicBlock *MBB);
|
|
void verify() const;
|
|
|
|
/// Get the trace that passes through MBB.
|
|
/// The trace is computed on demand.
|
|
Trace getTrace(const MachineBasicBlock *MBB);
|
|
};
|
|
|
|
/// Strategies for selecting traces.
|
|
enum Strategy {
|
|
/// Select the trace through a block that has the fewest instructions.
|
|
TS_MinInstrCount,
|
|
|
|
TS_NumStrategies
|
|
};
|
|
|
|
/// Get the trace ensemble representing the given trace selection strategy.
|
|
/// The returned Ensemble object is owned by the MachineTraceMetrics analysis,
|
|
/// and valid for the lifetime of the analysis pass.
|
|
Ensemble *getEnsemble(Strategy);
|
|
|
|
/// Invalidate cached information about MBB. This must be called *before* MBB
|
|
/// is erased, or the CFG is otherwise changed.
|
|
///
|
|
/// This invalidates per-block information about resource usage for MBB only,
|
|
/// and it invalidates per-trace information for any trace that passes
|
|
/// through MBB.
|
|
///
|
|
/// Call Ensemble::getTrace() again to update any trace handles.
|
|
void invalidate(const MachineBasicBlock *MBB);
|
|
|
|
private:
|
|
// One entry per basic block, indexed by block number.
|
|
SmallVector<FixedBlockInfo, 4> BlockInfo;
|
|
|
|
// Cycles consumed on each processor resource per block.
|
|
// The number of processor resource kinds is constant for a given subtarget,
|
|
// but it is not known at compile time. The number of cycles consumed by
|
|
// block B on processor resource R is at ProcResourceCycles[B*Kinds + R]
|
|
// where Kinds = SchedModel.getNumProcResourceKinds().
|
|
SmallVector<unsigned, 0> ProcResourceCycles;
|
|
|
|
// One ensemble per strategy.
|
|
Ensemble* Ensembles[TS_NumStrategies];
|
|
|
|
// Convert scaled resource usage to a cycle count that can be compared with
|
|
// latencies.
|
|
unsigned getCycles(unsigned Scaled) {
|
|
unsigned Factor = SchedModel.getLatencyFactor();
|
|
return (Scaled + Factor - 1) / Factor;
|
|
}
|
|
};
|
|
|
|
inline raw_ostream &operator<<(raw_ostream &OS,
|
|
const MachineTraceMetrics::Trace &Tr) {
|
|
Tr.print(OS);
|
|
return OS;
|
|
}
|
|
|
|
inline raw_ostream &operator<<(raw_ostream &OS,
|
|
const MachineTraceMetrics::Ensemble &En) {
|
|
En.print(OS);
|
|
return OS;
|
|
}
|
|
} // end namespace llvm
|
|
|
|
#endif
|