2012-05-17 22:37:09 +00:00
|
|
|
//===- MachineScheduler.cpp - Machine Instruction Scheduler ---------------===//
|
2012-01-13 06:30:30 +00:00
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// MachineScheduler schedules machine instructions after phi elimination. It
|
|
|
|
// preserves LiveIntervals so it can be invoked before register allocation.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#define DEBUG_TYPE "misched"
|
|
|
|
|
2012-03-08 01:41:12 +00:00
|
|
|
#include "llvm/CodeGen/MachineScheduler.h"
|
2012-12-03 16:50:05 +00:00
|
|
|
#include "llvm/ADT/OwningPtr.h"
|
|
|
|
#include "llvm/ADT/PriorityQueue.h"
|
|
|
|
#include "llvm/Analysis/AliasAnalysis.h"
|
|
|
|
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
|
2012-01-13 06:30:30 +00:00
|
|
|
#include "llvm/CodeGen/Passes.h"
|
2012-06-06 20:29:31 +00:00
|
|
|
#include "llvm/CodeGen/RegisterClassInfo.h"
|
2012-11-28 05:13:24 +00:00
|
|
|
#include "llvm/CodeGen/ScheduleDFS.h"
|
2012-05-24 22:11:09 +00:00
|
|
|
#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
|
2012-01-13 06:30:30 +00:00
|
|
|
#include "llvm/Support/CommandLine.h"
|
|
|
|
#include "llvm/Support/Debug.h"
|
|
|
|
#include "llvm/Support/ErrorHandling.h"
|
|
|
|
#include "llvm/Support/raw_ostream.h"
|
2012-01-17 06:55:07 +00:00
|
|
|
#include <queue>
|
|
|
|
|
2012-01-13 06:30:30 +00:00
|
|
|
using namespace llvm;
|
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
namespace llvm {
|
|
|
|
cl::opt<bool> ForceTopDown("misched-topdown", cl::Hidden,
|
|
|
|
cl::desc("Force top-down list scheduling"));
|
|
|
|
cl::opt<bool> ForceBottomUp("misched-bottomup", cl::Hidden,
|
|
|
|
cl::desc("Force bottom-up list scheduling"));
|
|
|
|
}
|
2012-03-14 04:00:41 +00:00
|
|
|
|
2012-03-07 00:18:25 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
static cl::opt<bool> ViewMISchedDAGs("view-misched-dags", cl::Hidden,
|
|
|
|
cl::desc("Pop up a window to show MISched dags after they are processed"));
|
2012-03-19 18:38:38 +00:00
|
|
|
|
|
|
|
static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
|
|
|
|
cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
|
2012-03-07 00:18:25 +00:00
|
|
|
#else
|
|
|
|
static bool ViewMISchedDAGs = false;
|
|
|
|
#endif // NDEBUG
|
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
// Experimental heuristics
|
|
|
|
static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
|
2012-11-13 08:47:29 +00:00
|
|
|
cl::desc("Enable load clustering."), cl::init(true));
|
2012-11-12 19:40:10 +00:00
|
|
|
|
2012-11-12 19:52:20 +00:00
|
|
|
// Experimental heuristics
|
|
|
|
static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
|
2012-11-13 08:47:29 +00:00
|
|
|
cl::desc("Enable scheduling for macro fusion."), cl::init(true));
|
2012-11-12 19:52:20 +00:00
|
|
|
|
2012-01-14 02:17:06 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Machine Instruction Scheduling Pass and Registry
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2012-04-24 20:36:19 +00:00
|
|
|
MachineSchedContext::MachineSchedContext():
|
|
|
|
MF(0), MLI(0), MDT(0), PassConfig(0), AA(0), LIS(0) {
|
|
|
|
RegClassInfo = new RegisterClassInfo();
|
|
|
|
}
|
|
|
|
|
|
|
|
MachineSchedContext::~MachineSchedContext() {
|
|
|
|
delete RegClassInfo;
|
|
|
|
}
|
|
|
|
|
2012-01-13 06:30:30 +00:00
|
|
|
namespace {
|
2012-01-17 06:55:03 +00:00
|
|
|
/// MachineScheduler runs after coalescing and before register allocation.
|
2012-03-08 01:41:12 +00:00
|
|
|
class MachineScheduler : public MachineSchedContext,
|
|
|
|
public MachineFunctionPass {
|
2012-01-13 06:30:30 +00:00
|
|
|
public:
|
2012-01-17 06:55:03 +00:00
|
|
|
MachineScheduler();
|
2012-01-13 06:30:30 +00:00
|
|
|
|
|
|
|
virtual void getAnalysisUsage(AnalysisUsage &AU) const;
|
|
|
|
|
|
|
|
virtual void releaseMemory() {}
|
|
|
|
|
|
|
|
virtual bool runOnMachineFunction(MachineFunction&);
|
|
|
|
|
|
|
|
virtual void print(raw_ostream &O, const Module* = 0) const;
|
|
|
|
|
|
|
|
static char ID; // Class identification, replacement for typeinfo
|
|
|
|
};
|
|
|
|
} // namespace
|
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
char MachineScheduler::ID = 0;
|
2012-01-13 06:30:30 +00:00
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
char &llvm::MachineSchedulerID = MachineScheduler::ID;
|
2012-01-13 06:30:30 +00:00
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
INITIALIZE_PASS_BEGIN(MachineScheduler, "misched",
|
2012-01-13 06:30:30 +00:00
|
|
|
"Machine Instruction Scheduler", false, false)
|
|
|
|
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
|
2012-01-17 06:55:03 +00:00
|
|
|
INITIALIZE_PASS_END(MachineScheduler, "misched",
|
2012-01-13 06:30:30 +00:00
|
|
|
"Machine Instruction Scheduler", false, false)
|
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
MachineScheduler::MachineScheduler()
|
2012-03-08 01:41:12 +00:00
|
|
|
: MachineFunctionPass(ID) {
|
2012-01-17 06:55:03 +00:00
|
|
|
initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
|
2012-01-13 06:30:30 +00:00
|
|
|
}
|
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
|
2012-01-13 06:30:30 +00:00
|
|
|
AU.setPreservesCFG();
|
|
|
|
AU.addRequiredID(MachineDominatorsID);
|
|
|
|
AU.addRequired<MachineLoopInfo>();
|
|
|
|
AU.addRequired<AliasAnalysis>();
|
2012-03-09 00:52:20 +00:00
|
|
|
AU.addRequired<TargetPassConfig>();
|
2012-01-13 06:30:30 +00:00
|
|
|
AU.addRequired<SlotIndexes>();
|
|
|
|
AU.addPreserved<SlotIndexes>();
|
|
|
|
AU.addRequired<LiveIntervals>();
|
|
|
|
AU.addPreserved<LiveIntervals>();
|
|
|
|
MachineFunctionPass::getAnalysisUsage(AU);
|
|
|
|
}
|
|
|
|
|
|
|
|
MachinePassRegistry MachineSchedRegistry::Registry;
|
|
|
|
|
2012-03-09 00:52:20 +00:00
|
|
|
/// A dummy default scheduler factory indicates whether the scheduler
|
|
|
|
/// is overridden on the command line.
|
|
|
|
static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
|
|
|
|
return 0;
|
|
|
|
}
|
2012-01-13 06:30:30 +00:00
|
|
|
|
|
|
|
/// MachineSchedOpt allows command line selection of the scheduler.
|
|
|
|
static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
|
|
|
|
RegisterPassParser<MachineSchedRegistry> >
|
|
|
|
MachineSchedOpt("misched",
|
2012-03-09 00:52:20 +00:00
|
|
|
cl::init(&useDefaultMachineSched), cl::Hidden,
|
2012-01-13 06:30:30 +00:00
|
|
|
cl::desc("Machine instruction scheduler to use"));
|
|
|
|
|
2012-03-09 00:52:20 +00:00
|
|
|
static MachineSchedRegistry
|
2012-03-14 04:00:41 +00:00
|
|
|
DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
|
2012-03-09 00:52:20 +00:00
|
|
|
useDefaultMachineSched);
|
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
/// Forward declare the standard machine scheduler. This will be used as the
|
2012-03-09 00:52:20 +00:00
|
|
|
/// default scheduler if the target does not set a default.
|
2012-03-14 04:00:41 +00:00
|
|
|
static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C);
|
2012-03-09 00:52:20 +00:00
|
|
|
|
2012-04-24 18:04:34 +00:00
|
|
|
|
|
|
|
/// Decrement this iterator until reaching the top or a non-debug instr.
|
|
|
|
static MachineBasicBlock::iterator
|
|
|
|
priorNonDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator Beg) {
|
|
|
|
assert(I != Beg && "reached the top of the region, cannot decrement");
|
|
|
|
while (--I != Beg) {
|
|
|
|
if (!I->isDebugValue())
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return I;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// If this iterator is a debug value, increment until reaching the End or a
|
|
|
|
/// non-debug instruction.
|
|
|
|
static MachineBasicBlock::iterator
|
|
|
|
nextIfDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator End) {
|
2012-05-17 18:35:03 +00:00
|
|
|
for(; I != End; ++I) {
|
2012-04-24 18:04:34 +00:00
|
|
|
if (!I->isDebugValue())
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return I;
|
|
|
|
}
|
|
|
|
|
2012-03-14 04:00:38 +00:00
|
|
|
/// Top-level MachineScheduler pass driver.
|
|
|
|
///
|
|
|
|
/// Visit blocks in function order. Divide each block into scheduling regions
|
2012-03-14 04:00:41 +00:00
|
|
|
/// and visit them bottom-up. Visiting regions bottom-up is not required, but is
|
|
|
|
/// consistent with the DAG builder, which traverses the interior of the
|
|
|
|
/// scheduling regions bottom-up.
|
2012-03-14 04:00:38 +00:00
|
|
|
///
|
|
|
|
/// This design avoids exposing scheduling boundaries to the DAG builder,
|
2012-03-14 04:00:41 +00:00
|
|
|
/// simplifying the DAG builder's support for "special" target instructions.
|
|
|
|
/// At the same time the design allows target schedulers to operate across
|
2012-03-14 04:00:38 +00:00
|
|
|
/// scheduling boundaries, for example to bundle the boudary instructions
|
|
|
|
/// without reordering them. This creates complexity, because the target
|
|
|
|
/// scheduler must update the RegionBegin and RegionEnd positions cached by
|
|
|
|
/// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
|
|
|
|
/// design would be to split blocks at scheduling boundaries, but LLVM has a
|
|
|
|
/// general bias against block splitting purely for implementation simplicity.
|
2012-03-08 01:41:12 +00:00
|
|
|
bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
|
2012-05-10 21:06:21 +00:00
|
|
|
DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
|
|
|
|
|
2012-03-08 01:41:12 +00:00
|
|
|
// Initialize the context of the pass.
|
|
|
|
MF = &mf;
|
|
|
|
MLI = &getAnalysis<MachineLoopInfo>();
|
|
|
|
MDT = &getAnalysis<MachineDominatorTree>();
|
2012-03-09 00:52:20 +00:00
|
|
|
PassConfig = &getAnalysis<TargetPassConfig>();
|
2012-03-08 01:41:12 +00:00
|
|
|
AA = &getAnalysis<AliasAnalysis>();
|
|
|
|
|
|
|
|
LIS = &getAnalysis<LiveIntervals>();
|
|
|
|
const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
|
|
|
|
|
2012-04-24 20:36:19 +00:00
|
|
|
RegClassInfo->runOnMachineFunction(*MF);
|
2012-04-24 17:56:43 +00:00
|
|
|
|
2012-03-08 01:41:12 +00:00
|
|
|
// Select the scheduler, or set the default.
|
2012-03-09 00:52:20 +00:00
|
|
|
MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
|
|
|
|
if (Ctor == useDefaultMachineSched) {
|
|
|
|
// Get the default scheduler set by the target.
|
|
|
|
Ctor = MachineSchedRegistry::getDefault();
|
|
|
|
if (!Ctor) {
|
2012-03-14 04:00:41 +00:00
|
|
|
Ctor = createConvergingSched;
|
2012-03-09 00:52:20 +00:00
|
|
|
MachineSchedRegistry::setDefault(Ctor);
|
|
|
|
}
|
2012-03-08 01:41:12 +00:00
|
|
|
}
|
|
|
|
// Instantiate the selected scheduler.
|
|
|
|
OwningPtr<ScheduleDAGInstrs> Scheduler(Ctor(this));
|
|
|
|
|
|
|
|
// Visit all machine basic blocks.
|
2012-04-24 17:56:43 +00:00
|
|
|
//
|
|
|
|
// TODO: Visit blocks in global postorder or postorder within the bottom-up
|
|
|
|
// loop tree. Then we can optionally compute global RegPressure.
|
2012-03-08 01:41:12 +00:00
|
|
|
for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
|
|
|
|
MBB != MBBEnd; ++MBB) {
|
|
|
|
|
2012-03-09 08:02:51 +00:00
|
|
|
Scheduler->startBlock(MBB);
|
|
|
|
|
2012-03-08 01:41:12 +00:00
|
|
|
// Break the block into scheduling regions [I, RegionEnd), and schedule each
|
2012-07-23 08:51:15 +00:00
|
|
|
// region as soon as it is discovered. RegionEnd points the scheduling
|
2012-03-09 22:34:56 +00:00
|
|
|
// boundary at the bottom of the region. The DAG does not include RegionEnd,
|
|
|
|
// but the region does (i.e. the next RegionEnd is above the previous
|
|
|
|
// RegionBegin). If the current block has no terminator then RegionEnd ==
|
|
|
|
// MBB->end() for the bottom region.
|
|
|
|
//
|
|
|
|
// The Scheduler may insert instructions during either schedule() or
|
|
|
|
// exitRegion(), even for empty regions. So the local iterators 'I' and
|
|
|
|
// 'RegionEnd' are invalid across these calls.
|
2012-11-06 07:10:34 +00:00
|
|
|
unsigned RemainingInstrs = MBB->size();
|
2012-03-09 03:46:39 +00:00
|
|
|
for(MachineBasicBlock::iterator RegionEnd = MBB->end();
|
2012-03-09 22:34:56 +00:00
|
|
|
RegionEnd != MBB->begin(); RegionEnd = Scheduler->begin()) {
|
2012-04-24 17:56:43 +00:00
|
|
|
|
2012-03-09 08:02:51 +00:00
|
|
|
// Avoid decrementing RegionEnd for blocks with no terminator.
|
|
|
|
if (RegionEnd != MBB->end()
|
|
|
|
|| TII->isSchedulingBoundary(llvm::prior(RegionEnd), MBB, *MF)) {
|
|
|
|
--RegionEnd;
|
|
|
|
// Count the boundary instruction.
|
2012-11-06 07:10:34 +00:00
|
|
|
--RemainingInstrs;
|
2012-03-09 08:02:51 +00:00
|
|
|
}
|
|
|
|
|
2012-03-08 01:41:12 +00:00
|
|
|
// The next region starts above the previous region. Look backward in the
|
|
|
|
// instruction stream until we find the nearest boundary.
|
|
|
|
MachineBasicBlock::iterator I = RegionEnd;
|
2012-11-06 07:10:34 +00:00
|
|
|
for(;I != MBB->begin(); --I, --RemainingInstrs) {
|
2012-03-08 01:41:12 +00:00
|
|
|
if (TII->isSchedulingBoundary(llvm::prior(I), MBB, *MF))
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
// Notify the scheduler of the region, even if we may skip scheduling
|
|
|
|
// it. Perhaps it still needs to be bundled.
|
2012-11-06 07:10:34 +00:00
|
|
|
Scheduler->enterRegion(MBB, I, RegionEnd, RemainingInstrs);
|
2012-03-08 01:41:12 +00:00
|
|
|
|
|
|
|
// Skip empty scheduling regions (0 or 1 schedulable instructions).
|
|
|
|
if (I == RegionEnd || I == llvm::prior(RegionEnd)) {
|
|
|
|
// Close the current region. Bundle the terminator if needed.
|
2012-03-09 22:34:56 +00:00
|
|
|
// This invalidates 'RegionEnd' and 'I'.
|
2012-03-08 01:41:12 +00:00
|
|
|
Scheduler->exitRegion();
|
|
|
|
continue;
|
|
|
|
}
|
2012-05-24 22:11:14 +00:00
|
|
|
DEBUG(dbgs() << "********** MI Scheduling **********\n");
|
2012-08-22 06:07:19 +00:00
|
|
|
DEBUG(dbgs() << MF->getName()
|
2012-03-08 01:41:12 +00:00
|
|
|
<< ":BB#" << MBB->getNumber() << "\n From: " << *I << " To: ";
|
|
|
|
if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
|
|
|
|
else dbgs() << "End";
|
2012-11-06 07:10:34 +00:00
|
|
|
dbgs() << " Remaining: " << RemainingInstrs << "\n");
|
2012-03-08 01:41:12 +00:00
|
|
|
|
2012-03-09 03:46:42 +00:00
|
|
|
// Schedule a region: possibly reorder instructions.
|
2012-03-09 22:34:56 +00:00
|
|
|
// This invalidates 'RegionEnd' and 'I'.
|
2012-03-08 01:41:12 +00:00
|
|
|
Scheduler->schedule();
|
2012-03-09 03:46:42 +00:00
|
|
|
|
|
|
|
// Close the current region.
|
2012-03-08 01:41:12 +00:00
|
|
|
Scheduler->exitRegion();
|
|
|
|
|
|
|
|
// Scheduling has invalidated the current iterator 'I'. Ask the
|
|
|
|
// scheduler for the top of it's scheduled region.
|
|
|
|
RegionEnd = Scheduler->begin();
|
|
|
|
}
|
2012-11-06 07:10:34 +00:00
|
|
|
assert(RemainingInstrs == 0 && "Instruction count mismatch!");
|
2012-03-08 01:41:12 +00:00
|
|
|
Scheduler->finishBlock();
|
|
|
|
}
|
2012-04-01 07:24:23 +00:00
|
|
|
Scheduler->finalizeSchedule();
|
2012-03-21 04:12:12 +00:00
|
|
|
DEBUG(LIS->print(dbgs()));
|
2012-03-08 01:41:12 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
void MachineScheduler::print(raw_ostream &O, const Module* m) const {
|
|
|
|
// unimplemented
|
|
|
|
}
|
|
|
|
|
2012-09-11 22:23:19 +00:00
|
|
|
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
|
2012-09-11 00:39:15 +00:00
|
|
|
void ReadyQueue::dump() {
|
|
|
|
dbgs() << Name << ": ";
|
|
|
|
for (unsigned i = 0, e = Queue.size(); i < e; ++i)
|
|
|
|
dbgs() << Queue[i]->NodeNum << " ";
|
|
|
|
dbgs() << "\n";
|
|
|
|
}
|
|
|
|
#endif
|
2012-03-14 04:00:41 +00:00
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// ScheduleDAGMI - Base class for MachineInstr scheduling with LiveIntervals
|
|
|
|
// preservation.
|
|
|
|
//===----------------------------------------------------------------------===//
|
2012-01-14 02:17:06 +00:00
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
|
2012-11-12 19:52:20 +00:00
|
|
|
if (SuccSU != &ExitSU) {
|
|
|
|
// Do not use WillCreateCycle, it assumes SD scheduling.
|
|
|
|
// If Pred is reachable from Succ, then the edge creates a cycle.
|
|
|
|
if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
|
|
|
|
return false;
|
|
|
|
Topo.AddPred(SuccSU, PredDep.getSUnit());
|
|
|
|
}
|
2012-11-12 19:40:10 +00:00
|
|
|
SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
|
|
|
|
// Return true regardless of whether a new edge needed to be inserted.
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2012-01-17 06:55:07 +00:00
|
|
|
/// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
|
|
|
|
/// NumPredsLeft reaches zero, release the successor node.
|
2012-05-24 22:11:09 +00:00
|
|
|
///
|
|
|
|
/// FIXME: Adjust SuccSU height based on MinLatency.
|
2012-03-14 04:00:41 +00:00
|
|
|
void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
|
2012-01-17 06:55:07 +00:00
|
|
|
SUnit *SuccSU = SuccEdge->getSUnit();
|
|
|
|
|
2012-11-12 19:28:57 +00:00
|
|
|
if (SuccEdge->isWeak()) {
|
|
|
|
--SuccSU->WeakPredsLeft;
|
2012-11-12 19:40:10 +00:00
|
|
|
if (SuccEdge->isCluster())
|
|
|
|
NextClusterSucc = SuccSU;
|
2012-11-12 19:28:57 +00:00
|
|
|
return;
|
|
|
|
}
|
2012-01-17 06:55:07 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
if (SuccSU->NumPredsLeft == 0) {
|
|
|
|
dbgs() << "*** Scheduling failed! ***\n";
|
|
|
|
SuccSU->dump(this);
|
|
|
|
dbgs() << " has been released too many times!\n";
|
|
|
|
llvm_unreachable(0);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
--SuccSU->NumPredsLeft;
|
|
|
|
if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
|
2012-03-14 04:00:41 +00:00
|
|
|
SchedImpl->releaseTopNode(SuccSU);
|
2012-01-17 06:55:07 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// releaseSuccessors - Call releaseSucc on each of SU's successors.
|
2012-03-14 04:00:41 +00:00
|
|
|
void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
|
2012-01-17 06:55:07 +00:00
|
|
|
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
releaseSucc(SU, &*I);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
/// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
|
|
|
|
/// NumSuccsLeft reaches zero, release the predecessor node.
|
2012-05-24 22:11:09 +00:00
|
|
|
///
|
|
|
|
/// FIXME: Adjust PredSU height based on MinLatency.
|
2012-03-14 04:00:41 +00:00
|
|
|
void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
|
|
|
|
SUnit *PredSU = PredEdge->getSUnit();
|
|
|
|
|
2012-11-12 19:28:57 +00:00
|
|
|
if (PredEdge->isWeak()) {
|
|
|
|
--PredSU->WeakSuccsLeft;
|
2012-11-12 19:40:10 +00:00
|
|
|
if (PredEdge->isCluster())
|
|
|
|
NextClusterPred = PredSU;
|
2012-11-12 19:28:57 +00:00
|
|
|
return;
|
|
|
|
}
|
2012-03-14 04:00:41 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
if (PredSU->NumSuccsLeft == 0) {
|
|
|
|
dbgs() << "*** Scheduling failed! ***\n";
|
|
|
|
PredSU->dump(this);
|
|
|
|
dbgs() << " has been released too many times!\n";
|
|
|
|
llvm_unreachable(0);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
--PredSU->NumSuccsLeft;
|
|
|
|
if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
|
|
|
|
SchedImpl->releaseBottomNode(PredSU);
|
|
|
|
}
|
|
|
|
|
|
|
|
/// releasePredecessors - Call releasePred on each of SU's predecessors.
|
|
|
|
void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
|
|
|
|
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
releasePred(SU, &*I);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
|
|
|
|
MachineBasicBlock::iterator InsertPos) {
|
2012-05-17 18:35:03 +00:00
|
|
|
// Advance RegionBegin if the first instruction moves down.
|
2012-03-21 04:12:10 +00:00
|
|
|
if (&*RegionBegin == MI)
|
2012-05-17 18:35:03 +00:00
|
|
|
++RegionBegin;
|
|
|
|
|
|
|
|
// Update the instruction stream.
|
2012-03-14 04:00:41 +00:00
|
|
|
BB->splice(InsertPos, BB, MI);
|
2012-05-17 18:35:03 +00:00
|
|
|
|
|
|
|
// Update LiveIntervals
|
2012-10-16 00:22:51 +00:00
|
|
|
LIS->handleMove(MI, /*UpdateFlags=*/true);
|
2012-05-17 18:35:03 +00:00
|
|
|
|
|
|
|
// Recede RegionBegin if an instruction moves above the first.
|
2012-03-14 04:00:41 +00:00
|
|
|
if (RegionBegin == InsertPos)
|
|
|
|
RegionBegin = MI;
|
|
|
|
}
|
|
|
|
|
2012-03-21 04:12:07 +00:00
|
|
|
bool ScheduleDAGMI::checkSchedLimit() {
|
|
|
|
#ifndef NDEBUG
|
|
|
|
if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
|
|
|
|
CurrentTop = CurrentBottom;
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
++NumInstrsScheduled;
|
|
|
|
#endif
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2012-04-24 17:56:43 +00:00
|
|
|
/// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
|
|
|
|
/// crossing a scheduling boundary. [begin, end) includes all instructions in
|
|
|
|
/// the region, including the boundary itself and single-instruction regions
|
|
|
|
/// that don't get scheduled.
|
|
|
|
void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
|
|
|
|
MachineBasicBlock::iterator begin,
|
|
|
|
MachineBasicBlock::iterator end,
|
|
|
|
unsigned endcount)
|
|
|
|
{
|
|
|
|
ScheduleDAGInstrs::enterRegion(bb, begin, end, endcount);
|
2012-05-10 21:06:10 +00:00
|
|
|
|
|
|
|
// For convenience remember the end of the liveness region.
|
|
|
|
LiveRegionEnd =
|
|
|
|
(RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Setup the register pressure trackers for the top scheduled top and bottom
|
|
|
|
// scheduled regions.
|
|
|
|
void ScheduleDAGMI::initRegPressure() {
|
|
|
|
TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
|
|
|
|
BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
|
|
|
|
|
|
|
|
// Close the RPTracker to finalize live ins.
|
|
|
|
RPTracker.closeRegion();
|
|
|
|
|
2012-05-24 22:11:14 +00:00
|
|
|
DEBUG(RPTracker.getPressure().dump(TRI));
|
|
|
|
|
2012-05-10 21:06:10 +00:00
|
|
|
// Initialize the live ins and live outs.
|
|
|
|
TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
|
|
|
|
BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
|
|
|
|
|
|
|
|
// Close one end of the tracker so we can call
|
|
|
|
// getMaxUpward/DownwardPressureDelta before advancing across any
|
|
|
|
// instructions. This converts currently live regs into live ins/outs.
|
|
|
|
TopRPTracker.closeTop();
|
|
|
|
BotRPTracker.closeBottom();
|
|
|
|
|
|
|
|
// Account for liveness generated by the region boundary.
|
|
|
|
if (LiveRegionEnd != RegionEnd)
|
|
|
|
BotRPTracker.recede();
|
|
|
|
|
|
|
|
assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
|
2012-05-17 18:35:10 +00:00
|
|
|
|
|
|
|
// Cache the list of excess pressure sets in this region. This will also track
|
|
|
|
// the max pressure in the scheduled code for these sets.
|
|
|
|
RegionCriticalPSets.clear();
|
|
|
|
std::vector<unsigned> RegionPressure = RPTracker.getPressure().MaxSetPressure;
|
|
|
|
for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
|
|
|
|
unsigned Limit = TRI->getRegPressureSetLimit(i);
|
2012-09-11 00:39:15 +00:00
|
|
|
DEBUG(dbgs() << TRI->getRegPressureSetName(i)
|
|
|
|
<< "Limit " << Limit
|
|
|
|
<< " Actual " << RegionPressure[i] << "\n");
|
2012-05-17 18:35:10 +00:00
|
|
|
if (RegionPressure[i] > Limit)
|
|
|
|
RegionCriticalPSets.push_back(PressureElement(i, 0));
|
|
|
|
}
|
|
|
|
DEBUG(dbgs() << "Excess PSets: ";
|
|
|
|
for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
|
|
|
|
dbgs() << TRI->getRegPressureSetName(
|
|
|
|
RegionCriticalPSets[i].PSetID) << " ";
|
|
|
|
dbgs() << "\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
// FIXME: When the pressure tracker deals in pressure differences then we won't
|
|
|
|
// iterate over all RegionCriticalPSets[i].
|
|
|
|
void ScheduleDAGMI::
|
|
|
|
updateScheduledPressure(std::vector<unsigned> NewMaxPressure) {
|
|
|
|
for (unsigned i = 0, e = RegionCriticalPSets.size(); i < e; ++i) {
|
|
|
|
unsigned ID = RegionCriticalPSets[i].PSetID;
|
|
|
|
int &MaxUnits = RegionCriticalPSets[i].UnitIncrease;
|
|
|
|
if ((int)NewMaxPressure[ID] > MaxUnits)
|
|
|
|
MaxUnits = NewMaxPressure[ID];
|
|
|
|
}
|
2012-04-24 17:56:43 +00:00
|
|
|
}
|
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
/// schedule - Called back from MachineScheduler::runOnMachineFunction
|
2012-04-24 17:56:43 +00:00
|
|
|
/// after setting up the current scheduling region. [RegionBegin, RegionEnd)
|
|
|
|
/// only includes instructions that have DAG nodes, not scheduling boundaries.
|
2012-09-11 00:39:15 +00:00
|
|
|
///
|
|
|
|
/// This is a skeletal driver, with all the functionality pushed into helpers,
|
|
|
|
/// so that it can be easilly extended by experimental schedulers. Generally,
|
|
|
|
/// implementing MachineSchedStrategy should be sufficient to implement a new
|
|
|
|
/// scheduling algorithm. However, if a scheduler further subclasses
|
|
|
|
/// ScheduleDAGMI then it will want to override this virtual method in order to
|
|
|
|
/// update any specialized state.
|
2012-03-14 04:00:41 +00:00
|
|
|
void ScheduleDAGMI::schedule() {
|
2012-09-11 00:39:15 +00:00
|
|
|
buildDAGWithRegPressure();
|
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
Topo.InitDAGTopologicalSorting();
|
|
|
|
|
2012-09-14 17:22:42 +00:00
|
|
|
postprocessDAG();
|
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
|
|
|
|
SUnits[su].dumpAll(this));
|
|
|
|
|
|
|
|
if (ViewMISchedDAGs) viewGraph();
|
|
|
|
|
|
|
|
initQueues();
|
|
|
|
|
|
|
|
bool IsTopNode = false;
|
|
|
|
while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
|
2012-10-08 18:53:53 +00:00
|
|
|
assert(!SU->isScheduled && "Node already scheduled");
|
2012-09-11 00:39:15 +00:00
|
|
|
if (!checkSchedLimit())
|
|
|
|
break;
|
|
|
|
|
|
|
|
scheduleMI(SU, IsTopNode);
|
|
|
|
|
|
|
|
updateQueues(SU, IsTopNode);
|
|
|
|
}
|
|
|
|
assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
|
|
|
|
|
|
|
|
placeDebugValues();
|
2012-11-07 07:05:09 +00:00
|
|
|
|
|
|
|
DEBUG({
|
2012-11-28 03:42:47 +00:00
|
|
|
unsigned BBNum = begin()->getParent()->getNumber();
|
2012-11-07 07:05:09 +00:00
|
|
|
dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
|
|
|
|
dumpSchedule();
|
|
|
|
dbgs() << '\n';
|
|
|
|
});
|
2012-09-11 00:39:15 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Build the DAG and setup three register pressure trackers.
|
|
|
|
void ScheduleDAGMI::buildDAGWithRegPressure() {
|
2012-05-10 21:06:10 +00:00
|
|
|
// Initialize the register pressure tracker used by buildSchedGraph.
|
|
|
|
RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
|
|
|
|
|
|
|
|
// Account for liveness generate by the region boundary.
|
|
|
|
if (LiveRegionEnd != RegionEnd)
|
|
|
|
RPTracker.recede();
|
2012-04-24 17:56:43 +00:00
|
|
|
|
2012-05-10 21:06:10 +00:00
|
|
|
// Build the DAG, and compute current register pressure.
|
2012-04-24 17:56:43 +00:00
|
|
|
buildSchedGraph(AA, &RPTracker);
|
2012-09-11 00:39:15 +00:00
|
|
|
if (ViewMISchedDAGs) viewGraph();
|
2012-01-17 06:55:07 +00:00
|
|
|
|
2012-05-10 21:06:10 +00:00
|
|
|
// Initialize top/bottom trackers after computing region pressure.
|
|
|
|
initRegPressure();
|
2012-09-11 00:39:15 +00:00
|
|
|
}
|
2012-05-10 21:06:10 +00:00
|
|
|
|
2012-09-14 17:22:42 +00:00
|
|
|
/// Apply each ScheduleDAGMutation step in order.
|
|
|
|
void ScheduleDAGMI::postprocessDAG() {
|
|
|
|
for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
|
|
|
|
Mutations[i]->apply(this);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-10-15 18:02:27 +00:00
|
|
|
// Release all DAG roots for scheduling.
|
2012-11-12 19:28:57 +00:00
|
|
|
//
|
|
|
|
// Nodes with unreleased weak edges can still be roots.
|
2012-10-15 18:02:27 +00:00
|
|
|
void ScheduleDAGMI::releaseRoots() {
|
|
|
|
SmallVector<SUnit*, 16> BotRoots;
|
|
|
|
|
|
|
|
for (std::vector<SUnit>::iterator
|
|
|
|
I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
|
2012-11-12 19:28:57 +00:00
|
|
|
SUnit *SU = &(*I);
|
2012-10-15 18:02:27 +00:00
|
|
|
// A SUnit is ready to top schedule if it has no predecessors.
|
2012-11-12 19:28:57 +00:00
|
|
|
if (!I->NumPredsLeft && SU != &EntrySU)
|
|
|
|
SchedImpl->releaseTopNode(SU);
|
2012-10-15 18:02:27 +00:00
|
|
|
// A SUnit is ready to bottom schedule if it has no successors.
|
2012-11-12 19:28:57 +00:00
|
|
|
if (!I->NumSuccsLeft && SU != &ExitSU)
|
|
|
|
BotRoots.push_back(SU);
|
2012-10-15 18:02:27 +00:00
|
|
|
}
|
|
|
|
// Release bottom roots in reverse order so the higher priority nodes appear
|
|
|
|
// first. This is more natural and slightly more efficient.
|
|
|
|
for (SmallVectorImpl<SUnit*>::const_reverse_iterator
|
|
|
|
I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I)
|
|
|
|
SchedImpl->releaseBottomNode(*I);
|
|
|
|
}
|
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
/// Identify DAG roots and setup scheduler queues.
|
|
|
|
void ScheduleDAGMI::initQueues() {
|
2012-11-12 19:40:10 +00:00
|
|
|
NextClusterSucc = NULL;
|
|
|
|
NextClusterPred = NULL;
|
2012-10-15 18:02:27 +00:00
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
// Initialize the strategy before modifying the DAG.
|
2012-03-14 04:00:41 +00:00
|
|
|
SchedImpl->initialize(this);
|
|
|
|
|
2012-11-12 19:28:57 +00:00
|
|
|
// Release all DAG roots for scheduling, not including EntrySU/ExitSU.
|
|
|
|
releaseRoots();
|
|
|
|
|
2012-01-17 06:55:07 +00:00
|
|
|
releaseSuccessors(&EntrySU);
|
2012-03-14 04:00:41 +00:00
|
|
|
releasePredecessors(&ExitSU);
|
2012-01-17 06:55:07 +00:00
|
|
|
|
2012-10-15 18:02:27 +00:00
|
|
|
SchedImpl->registerRoots();
|
|
|
|
|
2012-12-01 01:22:49 +00:00
|
|
|
// Advance past initial DebugValues.
|
|
|
|
assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
|
2012-04-24 18:04:34 +00:00
|
|
|
CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
|
2012-12-01 01:22:49 +00:00
|
|
|
TopRPTracker.setPos(CurrentTop);
|
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
CurrentBottom = RegionEnd;
|
2012-09-11 00:39:15 +00:00
|
|
|
}
|
2012-05-10 21:06:10 +00:00
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
/// Move an instruction and update register pressure.
|
|
|
|
void ScheduleDAGMI::scheduleMI(SUnit *SU, bool IsTopNode) {
|
|
|
|
// Move the instruction to its new location in the instruction stream.
|
|
|
|
MachineInstr *MI = SU->getInstr();
|
2012-05-10 21:06:10 +00:00
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
if (IsTopNode) {
|
|
|
|
assert(SU->isTopReady() && "node still has unscheduled dependencies");
|
|
|
|
if (&*CurrentTop == MI)
|
|
|
|
CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
|
|
|
|
else {
|
|
|
|
moveInstruction(MI, CurrentTop);
|
|
|
|
TopRPTracker.setPos(MI);
|
2012-03-14 04:00:41 +00:00
|
|
|
}
|
2012-09-11 00:39:15 +00:00
|
|
|
|
|
|
|
// Update top scheduled pressure.
|
|
|
|
TopRPTracker.advance();
|
|
|
|
assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
|
|
|
|
updateScheduledPressure(TopRPTracker.getPressure().MaxSetPressure);
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
assert(SU->isBottomReady() && "node still has unscheduled dependencies");
|
|
|
|
MachineBasicBlock::iterator priorII =
|
|
|
|
priorNonDebug(CurrentBottom, CurrentTop);
|
|
|
|
if (&*priorII == MI)
|
|
|
|
CurrentBottom = priorII;
|
2012-01-17 06:55:07 +00:00
|
|
|
else {
|
2012-09-11 00:39:15 +00:00
|
|
|
if (&*CurrentTop == MI) {
|
|
|
|
CurrentTop = nextIfDebug(++CurrentTop, priorII);
|
|
|
|
TopRPTracker.setPos(CurrentTop);
|
2012-03-14 04:00:41 +00:00
|
|
|
}
|
2012-09-11 00:39:15 +00:00
|
|
|
moveInstruction(MI, CurrentBottom);
|
|
|
|
CurrentBottom = MI;
|
2012-01-17 06:55:07 +00:00
|
|
|
}
|
2012-09-11 00:39:15 +00:00
|
|
|
// Update bottom scheduled pressure.
|
|
|
|
BotRPTracker.recede();
|
|
|
|
assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
|
|
|
|
updateScheduledPressure(BotRPTracker.getPressure().MaxSetPressure);
|
2012-01-17 06:55:07 +00:00
|
|
|
}
|
2012-09-11 00:39:15 +00:00
|
|
|
}
|
2012-04-24 18:04:37 +00:00
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
/// Update scheduler queues after scheduling an instruction.
|
|
|
|
void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
|
|
|
|
// Release dependent instructions for scheduling.
|
|
|
|
if (IsTopNode)
|
|
|
|
releaseSuccessors(SU);
|
|
|
|
else
|
|
|
|
releasePredecessors(SU);
|
|
|
|
|
|
|
|
SU->isScheduled = true;
|
|
|
|
|
|
|
|
// Notify the scheduling strategy after updating the DAG.
|
|
|
|
SchedImpl->schedNode(SU, IsTopNode);
|
2012-04-24 18:04:37 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Reinsert any remaining debug_values, just like the PostRA scheduler.
|
|
|
|
void ScheduleDAGMI::placeDebugValues() {
|
|
|
|
// If first instruction was a DBG_VALUE then put it back.
|
|
|
|
if (FirstDbgValue) {
|
|
|
|
BB->splice(RegionBegin, BB, FirstDbgValue);
|
|
|
|
RegionBegin = FirstDbgValue;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
|
|
|
|
DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
|
|
|
|
std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
|
|
|
|
MachineInstr *DbgValue = P.first;
|
|
|
|
MachineBasicBlock::iterator OrigPrevMI = P.second;
|
2012-12-01 01:22:38 +00:00
|
|
|
if (&*RegionBegin == DbgValue)
|
|
|
|
++RegionBegin;
|
2012-04-24 18:04:37 +00:00
|
|
|
BB->splice(++OrigPrevMI, BB, DbgValue);
|
|
|
|
if (OrigPrevMI == llvm::prior(RegionEnd))
|
|
|
|
RegionEnd = DbgValue;
|
|
|
|
}
|
|
|
|
DbgValues.clear();
|
|
|
|
FirstDbgValue = NULL;
|
2012-01-17 06:55:07 +00:00
|
|
|
}
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
|
|
|
|
void ScheduleDAGMI::dumpSchedule() const {
|
|
|
|
for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
|
|
|
|
if (SUnit *SU = getSUnit(&(*MI)))
|
|
|
|
SU->dump(this);
|
|
|
|
else
|
|
|
|
dbgs() << "Missing SUnit\n";
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2012-11-12 19:52:20 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// LoadClusterMutation - DAG post-processing to cluster loads.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
namespace {
|
|
|
|
/// \brief Post-process the DAG to create cluster edges between neighboring
|
|
|
|
/// loads.
|
|
|
|
class LoadClusterMutation : public ScheduleDAGMutation {
|
|
|
|
struct LoadInfo {
|
|
|
|
SUnit *SU;
|
|
|
|
unsigned BaseReg;
|
|
|
|
unsigned Offset;
|
|
|
|
LoadInfo(SUnit *su, unsigned reg, unsigned ofs)
|
|
|
|
: SU(su), BaseReg(reg), Offset(ofs) {}
|
|
|
|
};
|
|
|
|
static bool LoadInfoLess(const LoadClusterMutation::LoadInfo &LHS,
|
|
|
|
const LoadClusterMutation::LoadInfo &RHS);
|
|
|
|
|
|
|
|
const TargetInstrInfo *TII;
|
|
|
|
const TargetRegisterInfo *TRI;
|
|
|
|
public:
|
|
|
|
LoadClusterMutation(const TargetInstrInfo *tii,
|
|
|
|
const TargetRegisterInfo *tri)
|
|
|
|
: TII(tii), TRI(tri) {}
|
|
|
|
|
|
|
|
virtual void apply(ScheduleDAGMI *DAG);
|
|
|
|
protected:
|
|
|
|
void clusterNeighboringLoads(ArrayRef<SUnit*> Loads, ScheduleDAGMI *DAG);
|
|
|
|
};
|
|
|
|
} // anonymous
|
|
|
|
|
|
|
|
bool LoadClusterMutation::LoadInfoLess(
|
|
|
|
const LoadClusterMutation::LoadInfo &LHS,
|
|
|
|
const LoadClusterMutation::LoadInfo &RHS) {
|
|
|
|
if (LHS.BaseReg != RHS.BaseReg)
|
|
|
|
return LHS.BaseReg < RHS.BaseReg;
|
|
|
|
return LHS.Offset < RHS.Offset;
|
|
|
|
}
|
|
|
|
|
|
|
|
void LoadClusterMutation::clusterNeighboringLoads(ArrayRef<SUnit*> Loads,
|
|
|
|
ScheduleDAGMI *DAG) {
|
|
|
|
SmallVector<LoadClusterMutation::LoadInfo,32> LoadRecords;
|
|
|
|
for (unsigned Idx = 0, End = Loads.size(); Idx != End; ++Idx) {
|
|
|
|
SUnit *SU = Loads[Idx];
|
|
|
|
unsigned BaseReg;
|
|
|
|
unsigned Offset;
|
|
|
|
if (TII->getLdStBaseRegImmOfs(SU->getInstr(), BaseReg, Offset, TRI))
|
|
|
|
LoadRecords.push_back(LoadInfo(SU, BaseReg, Offset));
|
|
|
|
}
|
|
|
|
if (LoadRecords.size() < 2)
|
|
|
|
return;
|
|
|
|
std::sort(LoadRecords.begin(), LoadRecords.end(), LoadInfoLess);
|
|
|
|
unsigned ClusterLength = 1;
|
|
|
|
for (unsigned Idx = 0, End = LoadRecords.size(); Idx < (End - 1); ++Idx) {
|
|
|
|
if (LoadRecords[Idx].BaseReg != LoadRecords[Idx+1].BaseReg) {
|
|
|
|
ClusterLength = 1;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
SUnit *SUa = LoadRecords[Idx].SU;
|
|
|
|
SUnit *SUb = LoadRecords[Idx+1].SU;
|
2012-11-12 21:28:10 +00:00
|
|
|
if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
|
2012-11-12 19:40:10 +00:00
|
|
|
&& DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
|
|
|
|
<< SUb->NodeNum << ")\n");
|
|
|
|
// Copy successor edges from SUa to SUb. Interleaving computation
|
|
|
|
// dependent on SUa can prevent load combining due to register reuse.
|
|
|
|
// Predecessor edges do not need to be copied from SUb to SUa since nearby
|
|
|
|
// loads should have effectively the same inputs.
|
|
|
|
for (SUnit::const_succ_iterator
|
|
|
|
SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
|
|
|
|
if (SI->getSUnit() == SUb)
|
|
|
|
continue;
|
|
|
|
DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
|
|
|
|
DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
|
|
|
|
}
|
|
|
|
++ClusterLength;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
ClusterLength = 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// \brief Callback from DAG postProcessing to create cluster edges for loads.
|
|
|
|
void LoadClusterMutation::apply(ScheduleDAGMI *DAG) {
|
|
|
|
// Map DAG NodeNum to store chain ID.
|
|
|
|
DenseMap<unsigned, unsigned> StoreChainIDs;
|
|
|
|
// Map each store chain to a set of dependent loads.
|
|
|
|
SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
|
|
|
|
for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
|
|
|
|
SUnit *SU = &DAG->SUnits[Idx];
|
|
|
|
if (!SU->getInstr()->mayLoad())
|
|
|
|
continue;
|
|
|
|
unsigned ChainPredID = DAG->SUnits.size();
|
|
|
|
for (SUnit::const_pred_iterator
|
|
|
|
PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
|
|
|
|
if (PI->isCtrl()) {
|
|
|
|
ChainPredID = PI->getSUnit()->NodeNum;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Check if this chain-like pred has been seen
|
|
|
|
// before. ChainPredID==MaxNodeID for loads at the top of the schedule.
|
|
|
|
unsigned NumChains = StoreChainDependents.size();
|
|
|
|
std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
|
|
|
|
StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
|
|
|
|
if (Result.second)
|
|
|
|
StoreChainDependents.resize(NumChains + 1);
|
|
|
|
StoreChainDependents[Result.first->second].push_back(SU);
|
|
|
|
}
|
|
|
|
// Iterate over the store chains.
|
|
|
|
for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
|
|
|
|
clusterNeighboringLoads(StoreChainDependents[Idx], DAG);
|
|
|
|
}
|
|
|
|
|
2012-11-12 19:52:20 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// MacroFusion - DAG post-processing to encourage fusion of macro ops.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
/// \brief Post-process the DAG to create cluster edges between instructions
|
|
|
|
/// that may be fused by the processor into a single operation.
|
|
|
|
class MacroFusion : public ScheduleDAGMutation {
|
|
|
|
const TargetInstrInfo *TII;
|
|
|
|
public:
|
|
|
|
MacroFusion(const TargetInstrInfo *tii): TII(tii) {}
|
|
|
|
|
|
|
|
virtual void apply(ScheduleDAGMI *DAG);
|
|
|
|
};
|
|
|
|
} // anonymous
|
|
|
|
|
|
|
|
/// \brief Callback from DAG postProcessing to create cluster edges to encourage
|
|
|
|
/// fused operations.
|
|
|
|
void MacroFusion::apply(ScheduleDAGMI *DAG) {
|
|
|
|
// For now, assume targets can only fuse with the branch.
|
|
|
|
MachineInstr *Branch = DAG->ExitSU.getInstr();
|
|
|
|
if (!Branch)
|
|
|
|
return;
|
|
|
|
|
|
|
|
for (unsigned Idx = DAG->SUnits.size(); Idx > 0;) {
|
|
|
|
SUnit *SU = &DAG->SUnits[--Idx];
|
|
|
|
if (!TII->shouldScheduleAdjacent(SU->getInstr(), Branch))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Create a single weak edge from SU to ExitSU. The only effect is to cause
|
|
|
|
// bottom-up scheduling to heavily prioritize the clustered SU. There is no
|
|
|
|
// need to copy predecessor edges from ExitSU to SU, since top-down
|
|
|
|
// scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
|
|
|
|
// of SU, we could create an artificial edge from the deepest root, but it
|
|
|
|
// hasn't been needed yet.
|
|
|
|
bool Success = DAG->addEdge(&DAG->ExitSU, SDep(SU, SDep::Cluster));
|
|
|
|
(void)Success;
|
|
|
|
assert(Success && "No DAG nodes should be reachable from ExitSU");
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "Macro Fuse SU(" << SU->NodeNum << ")\n");
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
2012-03-14 04:00:41 +00:00
|
|
|
// ConvergingScheduler - Implementation of the standard MachineSchedStrategy.
|
2012-01-17 06:55:03 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
2012-03-14 04:00:41 +00:00
|
|
|
/// ConvergingScheduler shrinks the unscheduled zone using heuristics to balance
|
|
|
|
/// the schedule.
|
|
|
|
class ConvergingScheduler : public MachineSchedStrategy {
|
2012-11-07 07:05:09 +00:00
|
|
|
public:
|
|
|
|
/// Represent the type of SchedCandidate found within a single queue.
|
|
|
|
/// pickNodeBidirectional depends on these listed by decreasing priority.
|
|
|
|
enum CandReason {
|
2012-11-12 19:40:10 +00:00
|
|
|
NoCand, SingleExcess, SingleCritical, Cluster,
|
|
|
|
ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
|
|
|
|
TopDepthReduce, TopPathReduce, SingleMax, MultiPressure, NextDefUse,
|
|
|
|
NodeOrder};
|
2012-11-07 07:05:09 +00:00
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
static const char *getReasonStr(ConvergingScheduler::CandReason Reason);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/// Policy for scheduling the next instruction in the candidate's zone.
|
|
|
|
struct CandPolicy {
|
|
|
|
bool ReduceLatency;
|
|
|
|
unsigned ReduceResIdx;
|
|
|
|
unsigned DemandResIdx;
|
|
|
|
|
|
|
|
CandPolicy(): ReduceLatency(false), ReduceResIdx(0), DemandResIdx(0) {}
|
|
|
|
};
|
|
|
|
|
|
|
|
/// Status of an instruction's critical resource consumption.
|
|
|
|
struct SchedResourceDelta {
|
|
|
|
// Count critical resources in the scheduled region required by SU.
|
|
|
|
unsigned CritResources;
|
|
|
|
|
|
|
|
// Count critical resources from another region consumed by SU.
|
|
|
|
unsigned DemandedResources;
|
|
|
|
|
|
|
|
SchedResourceDelta(): CritResources(0), DemandedResources(0) {}
|
|
|
|
|
|
|
|
bool operator==(const SchedResourceDelta &RHS) const {
|
|
|
|
return CritResources == RHS.CritResources
|
|
|
|
&& DemandedResources == RHS.DemandedResources;
|
|
|
|
}
|
|
|
|
bool operator!=(const SchedResourceDelta &RHS) const {
|
|
|
|
return !operator==(RHS);
|
|
|
|
}
|
|
|
|
};
|
2012-05-10 21:06:16 +00:00
|
|
|
|
|
|
|
/// Store the state used by ConvergingScheduler heuristics, required for the
|
|
|
|
/// lifetime of one invocation of pickNode().
|
|
|
|
struct SchedCandidate {
|
2012-11-07 07:05:09 +00:00
|
|
|
CandPolicy Policy;
|
|
|
|
|
2012-05-10 21:06:16 +00:00
|
|
|
// The best SUnit candidate.
|
|
|
|
SUnit *SU;
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
// The reason for this candidate.
|
|
|
|
CandReason Reason;
|
|
|
|
|
2012-05-10 21:06:16 +00:00
|
|
|
// Register pressure values for the best candidate.
|
|
|
|
RegPressureDelta RPDelta;
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
// Critical resource consumption of the best candidate.
|
|
|
|
SchedResourceDelta ResDelta;
|
|
|
|
|
|
|
|
SchedCandidate(const CandPolicy &policy)
|
|
|
|
: Policy(policy), SU(NULL), Reason(NoCand) {}
|
|
|
|
|
|
|
|
bool isValid() const { return SU; }
|
|
|
|
|
|
|
|
// Copy the status of another candidate without changing policy.
|
|
|
|
void setBest(SchedCandidate &Best) {
|
|
|
|
assert(Best.Reason != NoCand && "uninitialized Sched candidate");
|
|
|
|
SU = Best.SU;
|
|
|
|
Reason = Best.Reason;
|
|
|
|
RPDelta = Best.RPDelta;
|
|
|
|
ResDelta = Best.ResDelta;
|
|
|
|
}
|
|
|
|
|
|
|
|
void initResourceDelta(const ScheduleDAGMI *DAG,
|
|
|
|
const TargetSchedModel *SchedModel);
|
|
|
|
};
|
|
|
|
|
|
|
|
/// Summarize the unscheduled region.
|
|
|
|
struct SchedRemainder {
|
|
|
|
// Critical path through the DAG in expected latency.
|
|
|
|
unsigned CriticalPath;
|
|
|
|
|
|
|
|
// Unscheduled resources
|
|
|
|
SmallVector<unsigned, 16> RemainingCounts;
|
|
|
|
// Critical resource for the unscheduled zone.
|
|
|
|
unsigned CritResIdx;
|
|
|
|
// Number of micro-ops left to schedule.
|
|
|
|
unsigned RemainingMicroOps;
|
|
|
|
|
|
|
|
void reset() {
|
|
|
|
CriticalPath = 0;
|
|
|
|
RemainingCounts.clear();
|
|
|
|
CritResIdx = 0;
|
|
|
|
RemainingMicroOps = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
SchedRemainder() { reset(); }
|
|
|
|
|
|
|
|
void init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel);
|
2012-12-18 20:52:56 +00:00
|
|
|
|
|
|
|
unsigned getMaxRemainingCount(const TargetSchedModel *SchedModel) const {
|
|
|
|
if (!SchedModel->hasInstrSchedModel())
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
return std::max(
|
|
|
|
RemainingMicroOps * SchedModel->getMicroOpFactor(),
|
|
|
|
RemainingCounts[CritResIdx]);
|
|
|
|
}
|
2012-05-10 21:06:16 +00:00
|
|
|
};
|
|
|
|
|
2012-05-24 22:11:12 +00:00
|
|
|
/// Each Scheduling boundary is associated with ready queues. It tracks the
|
2012-11-07 07:05:09 +00:00
|
|
|
/// current cycle in the direction of movement, and maintains the state
|
2012-05-24 22:11:12 +00:00
|
|
|
/// of "hazards" and other interlocks at the current cycle.
|
2012-05-24 22:11:09 +00:00
|
|
|
struct SchedBoundary {
|
2012-06-29 03:23:22 +00:00
|
|
|
ScheduleDAGMI *DAG;
|
2012-10-10 05:43:09 +00:00
|
|
|
const TargetSchedModel *SchedModel;
|
2012-11-07 07:05:09 +00:00
|
|
|
SchedRemainder *Rem;
|
2012-06-29 03:23:22 +00:00
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
ReadyQueue Available;
|
|
|
|
ReadyQueue Pending;
|
|
|
|
bool CheckPending;
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
// For heuristics, keep a list of the nodes that immediately depend on the
|
|
|
|
// most recently scheduled node.
|
|
|
|
SmallPtrSet<const SUnit*, 8> NextSUs;
|
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
ScheduleHazardRecognizer *HazardRec;
|
|
|
|
|
|
|
|
unsigned CurrCycle;
|
|
|
|
unsigned IssueCount;
|
|
|
|
|
|
|
|
/// MinReadyCycle - Cycle of the soonest available instruction.
|
|
|
|
unsigned MinReadyCycle;
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
// The expected latency of the critical path in this scheduled zone.
|
|
|
|
unsigned ExpectedLatency;
|
|
|
|
|
|
|
|
// Resources used in the scheduled zone beyond this boundary.
|
|
|
|
SmallVector<unsigned, 16> ResourceCounts;
|
|
|
|
|
|
|
|
// Cache the critical resources ID in this scheduled zone.
|
|
|
|
unsigned CritResIdx;
|
|
|
|
|
|
|
|
// Is the scheduled region resource limited vs. latency limited.
|
|
|
|
bool IsResourceLimited;
|
|
|
|
|
|
|
|
unsigned ExpectedCount;
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
2012-06-05 21:11:27 +00:00
|
|
|
// Remember the greatest min operand latency.
|
|
|
|
unsigned MaxMinLatency;
|
2012-11-07 07:05:09 +00:00
|
|
|
#endif
|
|
|
|
|
|
|
|
void reset() {
|
|
|
|
Available.clear();
|
|
|
|
Pending.clear();
|
|
|
|
CheckPending = false;
|
|
|
|
NextSUs.clear();
|
|
|
|
HazardRec = 0;
|
|
|
|
CurrCycle = 0;
|
|
|
|
IssueCount = 0;
|
|
|
|
MinReadyCycle = UINT_MAX;
|
|
|
|
ExpectedLatency = 0;
|
|
|
|
ResourceCounts.resize(1);
|
|
|
|
assert(!ResourceCounts[0] && "nonzero count for bad resource");
|
|
|
|
CritResIdx = 0;
|
|
|
|
IsResourceLimited = false;
|
|
|
|
ExpectedCount = 0;
|
|
|
|
#ifndef NDEBUG
|
|
|
|
MaxMinLatency = 0;
|
|
|
|
#endif
|
|
|
|
// Reserve a zero-count for invalid CritResIdx.
|
|
|
|
ResourceCounts.resize(1);
|
|
|
|
}
|
2012-06-05 21:11:27 +00:00
|
|
|
|
2012-05-24 22:11:12 +00:00
|
|
|
/// Pending queues extend the ready queues with the same ID and the
|
|
|
|
/// PendingFlag set.
|
|
|
|
SchedBoundary(unsigned ID, const Twine &Name):
|
2012-11-07 07:05:09 +00:00
|
|
|
DAG(0), SchedModel(0), Rem(0), Available(ID, Name+".A"),
|
|
|
|
Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P") {
|
|
|
|
reset();
|
|
|
|
}
|
2012-05-24 22:11:09 +00:00
|
|
|
|
|
|
|
~SchedBoundary() { delete HazardRec; }
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
void init(ScheduleDAGMI *dag, const TargetSchedModel *smodel,
|
|
|
|
SchedRemainder *rem);
|
2012-10-10 05:43:09 +00:00
|
|
|
|
2012-05-24 22:11:12 +00:00
|
|
|
bool isTop() const {
|
|
|
|
return Available.getID() == ConvergingScheduler::TopQID;
|
|
|
|
}
|
2012-05-24 22:11:09 +00:00
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
unsigned getUnscheduledLatency(SUnit *SU) const {
|
|
|
|
if (isTop())
|
|
|
|
return SU->getHeight();
|
2012-12-18 20:52:56 +00:00
|
|
|
return SU->getDepth() + SU->Latency;
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
unsigned getCriticalCount() const {
|
|
|
|
return ResourceCounts[CritResIdx];
|
|
|
|
}
|
|
|
|
|
2012-06-29 03:23:24 +00:00
|
|
|
bool checkHazard(SUnit *SU);
|
|
|
|
|
2012-12-18 20:52:56 +00:00
|
|
|
void setLatencyPolicy(CandPolicy &Policy);
|
2012-11-07 07:05:09 +00:00
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
void releaseNode(SUnit *SU, unsigned ReadyCycle);
|
|
|
|
|
|
|
|
void bumpCycle();
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
void countResource(unsigned PIdx, unsigned Cycles);
|
|
|
|
|
2012-06-29 03:23:22 +00:00
|
|
|
void bumpNode(SUnit *SU);
|
2012-06-05 21:11:27 +00:00
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
void releasePending();
|
|
|
|
|
|
|
|
void removeReady(SUnit *SU);
|
|
|
|
|
|
|
|
SUnit *pickOnlyChoice();
|
|
|
|
};
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
private:
|
2012-03-14 04:00:41 +00:00
|
|
|
ScheduleDAGMI *DAG;
|
2012-10-10 05:43:09 +00:00
|
|
|
const TargetSchedModel *SchedModel;
|
2012-05-10 21:06:16 +00:00
|
|
|
const TargetRegisterInfo *TRI;
|
2012-03-14 04:00:41 +00:00
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
// State of the top and bottom scheduled instruction boundaries.
|
2012-11-07 07:05:09 +00:00
|
|
|
SchedRemainder Rem;
|
2012-05-24 22:11:09 +00:00
|
|
|
SchedBoundary Top;
|
|
|
|
SchedBoundary Bot;
|
2012-03-14 04:00:41 +00:00
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
public:
|
2012-05-24 22:11:12 +00:00
|
|
|
/// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
|
2012-05-10 21:06:16 +00:00
|
|
|
enum {
|
|
|
|
TopQID = 1,
|
2012-05-24 22:11:12 +00:00
|
|
|
BotQID = 2,
|
|
|
|
LogMaxQID = 2
|
2012-05-10 21:06:16 +00:00
|
|
|
};
|
|
|
|
|
2012-05-24 22:11:12 +00:00
|
|
|
ConvergingScheduler():
|
2012-10-10 05:43:09 +00:00
|
|
|
DAG(0), SchedModel(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
|
2012-05-10 21:06:12 +00:00
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
virtual void initialize(ScheduleDAGMI *dag);
|
2012-01-17 06:55:03 +00:00
|
|
|
|
2012-05-10 21:06:16 +00:00
|
|
|
virtual SUnit *pickNode(bool &IsTopNode);
|
2012-01-17 06:55:03 +00:00
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
virtual void schedNode(SUnit *SU, bool IsTopNode);
|
|
|
|
|
|
|
|
virtual void releaseTopNode(SUnit *SU);
|
|
|
|
|
|
|
|
virtual void releaseBottomNode(SUnit *SU);
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
virtual void registerRoots();
|
|
|
|
|
2012-05-10 21:06:16 +00:00
|
|
|
protected:
|
2012-11-07 07:05:09 +00:00
|
|
|
void balanceZones(
|
|
|
|
ConvergingScheduler::SchedBoundary &CriticalZone,
|
|
|
|
ConvergingScheduler::SchedCandidate &CriticalCand,
|
|
|
|
ConvergingScheduler::SchedBoundary &OppositeZone,
|
|
|
|
ConvergingScheduler::SchedCandidate &OppositeCand);
|
|
|
|
|
|
|
|
void checkResourceLimits(ConvergingScheduler::SchedCandidate &TopCand,
|
|
|
|
ConvergingScheduler::SchedCandidate &BotCand);
|
|
|
|
|
|
|
|
void tryCandidate(SchedCandidate &Cand,
|
|
|
|
SchedCandidate &TryCand,
|
|
|
|
SchedBoundary &Zone,
|
|
|
|
const RegPressureTracker &RPTracker,
|
|
|
|
RegPressureTracker &TempTracker);
|
|
|
|
|
|
|
|
SUnit *pickNodeBidirectional(bool &IsTopNode);
|
|
|
|
|
|
|
|
void pickNodeFromQueue(SchedBoundary &Zone,
|
|
|
|
const RegPressureTracker &RPTracker,
|
|
|
|
SchedCandidate &Candidate);
|
2012-05-17 18:35:10 +00:00
|
|
|
|
2012-05-10 21:06:19 +00:00
|
|
|
#ifndef NDEBUG
|
2012-11-07 07:05:09 +00:00
|
|
|
void traceCandidate(const SchedCandidate &Cand, const SchedBoundary &Zone);
|
2012-05-10 21:06:19 +00:00
|
|
|
#endif
|
2012-03-14 04:00:41 +00:00
|
|
|
};
|
|
|
|
} // namespace
|
2012-01-17 06:55:03 +00:00
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
void ConvergingScheduler::SchedRemainder::
|
|
|
|
init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
|
|
|
|
reset();
|
|
|
|
if (!SchedModel->hasInstrSchedModel())
|
|
|
|
return;
|
|
|
|
RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
|
|
|
|
for (std::vector<SUnit>::iterator
|
|
|
|
I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
|
|
|
|
const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
|
|
|
|
RemainingMicroOps += SchedModel->getNumMicroOps(I->getInstr(), SC);
|
|
|
|
for (TargetSchedModel::ProcResIter
|
|
|
|
PI = SchedModel->getWriteProcResBegin(SC),
|
|
|
|
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
|
|
|
|
unsigned PIdx = PI->ProcResourceIdx;
|
|
|
|
unsigned Factor = SchedModel->getResourceFactor(PIdx);
|
|
|
|
RemainingCounts[PIdx] += (Factor * PI->Cycles);
|
|
|
|
}
|
|
|
|
}
|
2012-12-18 20:52:49 +00:00
|
|
|
for (unsigned PIdx = 0, PEnd = SchedModel->getNumProcResourceKinds();
|
|
|
|
PIdx != PEnd; ++PIdx) {
|
|
|
|
if ((int)(RemainingCounts[PIdx] - RemainingCounts[CritResIdx])
|
|
|
|
>= (int)SchedModel->getLatencyFactor()) {
|
|
|
|
CritResIdx = PIdx;
|
|
|
|
}
|
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void ConvergingScheduler::SchedBoundary::
|
|
|
|
init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
|
|
|
|
reset();
|
|
|
|
DAG = dag;
|
|
|
|
SchedModel = smodel;
|
|
|
|
Rem = rem;
|
|
|
|
if (SchedModel->hasInstrSchedModel())
|
|
|
|
ResourceCounts.resize(SchedModel->getNumProcResourceKinds());
|
|
|
|
}
|
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
void ConvergingScheduler::initialize(ScheduleDAGMI *dag) {
|
|
|
|
DAG = dag;
|
2012-10-10 05:43:09 +00:00
|
|
|
SchedModel = DAG->getSchedModel();
|
2012-05-24 22:11:09 +00:00
|
|
|
TRI = DAG->TRI;
|
2012-11-07 07:05:09 +00:00
|
|
|
Rem.init(DAG, SchedModel);
|
|
|
|
Top.init(DAG, SchedModel, &Rem);
|
|
|
|
Bot.init(DAG, SchedModel, &Rem);
|
|
|
|
|
|
|
|
// Initialize resource counts.
|
2012-05-24 22:11:09 +00:00
|
|
|
|
2012-10-10 05:43:09 +00:00
|
|
|
// Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
|
|
|
|
// are disabled, then these HazardRecs will be disabled.
|
|
|
|
const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
|
2012-05-24 22:11:09 +00:00
|
|
|
const TargetMachine &TM = DAG->MF.getTarget();
|
|
|
|
Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
|
|
|
|
Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
|
|
|
|
|
|
|
|
assert((!ForceTopDown || !ForceBottomUp) &&
|
|
|
|
"-misched-topdown incompatible with -misched-bottomup");
|
|
|
|
}
|
|
|
|
|
|
|
|
void ConvergingScheduler::releaseTopNode(SUnit *SU) {
|
2012-06-05 21:11:27 +00:00
|
|
|
if (SU->isScheduled)
|
|
|
|
return;
|
|
|
|
|
2012-12-18 20:52:52 +00:00
|
|
|
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
|
2012-06-05 21:11:27 +00:00
|
|
|
I != E; ++I) {
|
|
|
|
unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
|
2012-08-23 00:39:43 +00:00
|
|
|
unsigned MinLatency = I->getMinLatency();
|
2012-06-05 21:11:27 +00:00
|
|
|
#ifndef NDEBUG
|
2012-08-23 00:39:43 +00:00
|
|
|
Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency);
|
2012-06-05 21:11:27 +00:00
|
|
|
#endif
|
2012-08-23 00:39:43 +00:00
|
|
|
if (SU->TopReadyCycle < PredReadyCycle + MinLatency)
|
|
|
|
SU->TopReadyCycle = PredReadyCycle + MinLatency;
|
2012-06-05 21:11:27 +00:00
|
|
|
}
|
|
|
|
Top.releaseNode(SU, SU->TopReadyCycle);
|
2012-05-24 22:11:09 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void ConvergingScheduler::releaseBottomNode(SUnit *SU) {
|
2012-06-05 21:11:27 +00:00
|
|
|
if (SU->isScheduled)
|
|
|
|
return;
|
|
|
|
|
|
|
|
assert(SU->getInstr() && "Scheduled SUnit must have instr");
|
|
|
|
|
|
|
|
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
|
|
|
|
I != E; ++I) {
|
2012-11-12 19:40:10 +00:00
|
|
|
if (I->isWeak())
|
|
|
|
continue;
|
2012-06-05 21:11:27 +00:00
|
|
|
unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
|
2012-08-23 00:39:43 +00:00
|
|
|
unsigned MinLatency = I->getMinLatency();
|
2012-06-05 21:11:27 +00:00
|
|
|
#ifndef NDEBUG
|
2012-08-23 00:39:43 +00:00
|
|
|
Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency);
|
2012-06-05 21:11:27 +00:00
|
|
|
#endif
|
2012-08-23 00:39:43 +00:00
|
|
|
if (SU->BotReadyCycle < SuccReadyCycle + MinLatency)
|
|
|
|
SU->BotReadyCycle = SuccReadyCycle + MinLatency;
|
2012-06-05 21:11:27 +00:00
|
|
|
}
|
|
|
|
Bot.releaseNode(SU, SU->BotReadyCycle);
|
2012-05-24 22:11:09 +00:00
|
|
|
}
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
void ConvergingScheduler::registerRoots() {
|
|
|
|
Rem.CriticalPath = DAG->ExitSU.getDepth();
|
|
|
|
// Some roots may not feed into ExitSU. Check all of them in case.
|
|
|
|
for (std::vector<SUnit*>::const_iterator
|
|
|
|
I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
|
|
|
|
if ((*I)->getDepth() > Rem.CriticalPath)
|
|
|
|
Rem.CriticalPath = (*I)->getDepth();
|
|
|
|
}
|
|
|
|
DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
|
|
|
|
}
|
|
|
|
|
2012-06-29 03:23:24 +00:00
|
|
|
/// Does this SU have a hazard within the current instruction group.
|
|
|
|
///
|
|
|
|
/// The scheduler supports two modes of hazard recognition. The first is the
|
|
|
|
/// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
|
|
|
|
/// supports highly complicated in-order reservation tables
|
|
|
|
/// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
|
|
|
|
///
|
|
|
|
/// The second is a streamlined mechanism that checks for hazards based on
|
|
|
|
/// simple counters that the scheduler itself maintains. It explicitly checks
|
|
|
|
/// for instruction dispatch limitations, including the number of micro-ops that
|
|
|
|
/// can dispatch per cycle.
|
|
|
|
///
|
|
|
|
/// TODO: Also check whether the SU must start a new group.
|
|
|
|
bool ConvergingScheduler::SchedBoundary::checkHazard(SUnit *SU) {
|
|
|
|
if (HazardRec->isEnabled())
|
|
|
|
return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
|
|
|
|
|
2012-10-10 05:43:09 +00:00
|
|
|
unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
|
2012-11-07 07:05:09 +00:00
|
|
|
if ((IssueCount > 0) && (IssueCount + uops > SchedModel->getIssueWidth())) {
|
|
|
|
DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
|
|
|
|
<< SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
|
2012-06-29 03:23:24 +00:00
|
|
|
return true;
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2012-06-29 03:23:24 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2012-12-18 20:52:56 +00:00
|
|
|
/// Compute the remaining latency to determine whether ILP should be increased.
|
|
|
|
void ConvergingScheduler::SchedBoundary::setLatencyPolicy(CandPolicy &Policy) {
|
|
|
|
// FIXME: compile time. In all, we visit four queues here one we should only
|
|
|
|
// need to visit the one that was last popped if we cache the result.
|
|
|
|
unsigned RemLatency = 0;
|
|
|
|
for (ReadyQueue::iterator I = Available.begin(), E = Available.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
unsigned L = getUnscheduledLatency(*I);
|
|
|
|
if (L > RemLatency)
|
|
|
|
RemLatency = L;
|
|
|
|
}
|
|
|
|
for (ReadyQueue::iterator I = Pending.begin(), E = Pending.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
unsigned L = getUnscheduledLatency(*I);
|
|
|
|
if (L > RemLatency)
|
|
|
|
RemLatency = L;
|
|
|
|
}
|
2013-01-09 03:36:49 +00:00
|
|
|
unsigned CriticalPathLimit = Rem->CriticalPath + SchedModel->getILPWindow();
|
|
|
|
if (RemLatency + ExpectedLatency >= CriticalPathLimit
|
2012-12-18 20:52:56 +00:00
|
|
|
&& RemLatency > Rem->getMaxRemainingCount(SchedModel)) {
|
|
|
|
Policy.ReduceLatency = true;
|
|
|
|
DEBUG(dbgs() << "Increase ILP: " << Available.getName() << '\n');
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
void ConvergingScheduler::SchedBoundary::releaseNode(SUnit *SU,
|
|
|
|
unsigned ReadyCycle) {
|
2012-11-07 07:05:09 +00:00
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
if (ReadyCycle < MinReadyCycle)
|
|
|
|
MinReadyCycle = ReadyCycle;
|
|
|
|
|
|
|
|
// Check for interlocks first. For the purpose of other heuristics, an
|
|
|
|
// instruction that cannot issue appears as if it's not in the ReadyQueue.
|
2012-06-29 03:23:24 +00:00
|
|
|
if (ReadyCycle > CurrCycle || checkHazard(SU))
|
2012-05-24 22:11:09 +00:00
|
|
|
Pending.push(SU);
|
|
|
|
else
|
|
|
|
Available.push(SU);
|
2012-11-07 07:05:09 +00:00
|
|
|
|
|
|
|
// Record this node as an immediate dependent of the scheduled node.
|
|
|
|
NextSUs.insert(SU);
|
2012-05-24 22:11:09 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Move the boundary of scheduled code by one cycle.
|
|
|
|
void ConvergingScheduler::SchedBoundary::bumpCycle() {
|
2012-10-10 05:43:09 +00:00
|
|
|
unsigned Width = SchedModel->getIssueWidth();
|
2012-06-29 03:23:22 +00:00
|
|
|
IssueCount = (IssueCount <= Width) ? 0 : IssueCount - Width;
|
2012-05-24 22:11:09 +00:00
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
unsigned NextCycle = CurrCycle + 1;
|
2012-05-24 22:11:09 +00:00
|
|
|
assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
|
2012-11-07 07:05:09 +00:00
|
|
|
if (MinReadyCycle > NextCycle) {
|
|
|
|
IssueCount = 0;
|
|
|
|
NextCycle = MinReadyCycle;
|
|
|
|
}
|
2012-05-24 22:11:09 +00:00
|
|
|
|
|
|
|
if (!HazardRec->isEnabled()) {
|
2012-06-05 21:11:27 +00:00
|
|
|
// Bypass HazardRec virtual calls.
|
2012-05-24 22:11:09 +00:00
|
|
|
CurrCycle = NextCycle;
|
|
|
|
}
|
|
|
|
else {
|
2012-06-05 21:11:27 +00:00
|
|
|
// Bypass getHazardType calls in case of long latency.
|
2012-05-24 22:11:09 +00:00
|
|
|
for (; CurrCycle != NextCycle; ++CurrCycle) {
|
|
|
|
if (isTop())
|
|
|
|
HazardRec->AdvanceCycle();
|
|
|
|
else
|
|
|
|
HazardRec->RecedeCycle();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
CheckPending = true;
|
2012-11-07 07:05:09 +00:00
|
|
|
IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
|
2012-05-24 22:11:09 +00:00
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
DEBUG(dbgs() << " *** " << Available.getName() << " cycle "
|
2012-05-24 22:11:09 +00:00
|
|
|
<< CurrCycle << '\n');
|
|
|
|
}
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
/// Add the given processor resource to this scheduled zone.
|
|
|
|
void ConvergingScheduler::SchedBoundary::countResource(unsigned PIdx,
|
|
|
|
unsigned Cycles) {
|
|
|
|
unsigned Factor = SchedModel->getResourceFactor(PIdx);
|
|
|
|
DEBUG(dbgs() << " " << SchedModel->getProcResource(PIdx)->Name
|
|
|
|
<< " +(" << Cycles << "x" << Factor
|
|
|
|
<< ") / " << SchedModel->getLatencyFactor() << '\n');
|
|
|
|
|
|
|
|
unsigned Count = Factor * Cycles;
|
|
|
|
ResourceCounts[PIdx] += Count;
|
|
|
|
assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
|
|
|
|
Rem->RemainingCounts[PIdx] -= Count;
|
|
|
|
|
|
|
|
// Check if this resource exceeds the current critical resource by a full
|
|
|
|
// cycle. If so, it becomes the critical resource.
|
|
|
|
if ((int)(ResourceCounts[PIdx] - ResourceCounts[CritResIdx])
|
|
|
|
>= (int)SchedModel->getLatencyFactor()) {
|
|
|
|
CritResIdx = PIdx;
|
|
|
|
DEBUG(dbgs() << " *** Critical resource "
|
|
|
|
<< SchedModel->getProcResource(PIdx)->Name << " x"
|
|
|
|
<< ResourceCounts[PIdx] << '\n');
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-06-05 21:11:27 +00:00
|
|
|
/// Move the boundary of scheduled code by one SUnit.
|
2012-06-29 03:23:22 +00:00
|
|
|
void ConvergingScheduler::SchedBoundary::bumpNode(SUnit *SU) {
|
2012-06-05 21:11:27 +00:00
|
|
|
// Update the reservation table.
|
|
|
|
if (HazardRec->isEnabled()) {
|
|
|
|
if (!isTop() && SU->isCall) {
|
|
|
|
// Calls are scheduled with their preceding instructions. For bottom-up
|
|
|
|
// scheduling, clear the pipeline state before emitting.
|
|
|
|
HazardRec->Reset();
|
|
|
|
}
|
|
|
|
HazardRec->EmitInstruction(SU);
|
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
// Update resource counts and critical resource.
|
|
|
|
if (SchedModel->hasInstrSchedModel()) {
|
|
|
|
const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
|
|
|
|
Rem->RemainingMicroOps -= SchedModel->getNumMicroOps(SU->getInstr(), SC);
|
|
|
|
for (TargetSchedModel::ProcResIter
|
|
|
|
PI = SchedModel->getWriteProcResBegin(SC),
|
|
|
|
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
|
|
|
|
countResource(PI->ProcResourceIdx, PI->Cycles);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (isTop()) {
|
|
|
|
if (SU->getDepth() > ExpectedLatency)
|
|
|
|
ExpectedLatency = SU->getDepth();
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
if (SU->getHeight() > ExpectedLatency)
|
|
|
|
ExpectedLatency = SU->getHeight();
|
|
|
|
}
|
|
|
|
|
|
|
|
IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
|
|
|
|
|
2012-06-29 03:23:24 +00:00
|
|
|
// Check the instruction group dispatch limit.
|
|
|
|
// TODO: Check if this SU must end a dispatch group.
|
2012-10-10 05:43:09 +00:00
|
|
|
IssueCount += SchedModel->getNumMicroOps(SU->getInstr());
|
2012-11-07 07:05:09 +00:00
|
|
|
|
|
|
|
// checkHazard prevents scheduling multiple instructions per cycle that exceed
|
|
|
|
// issue width. However, we commonly reach the maximum. In this case
|
|
|
|
// opportunistically bump the cycle to avoid uselessly checking everything in
|
|
|
|
// the readyQ. Furthermore, a single instruction may produce more than one
|
|
|
|
// cycle's worth of micro-ops.
|
2012-10-10 05:43:09 +00:00
|
|
|
if (IssueCount >= SchedModel->getIssueWidth()) {
|
2012-11-07 07:05:09 +00:00
|
|
|
DEBUG(dbgs() << " *** Max instrs at cycle " << CurrCycle << '\n');
|
2012-06-05 21:11:27 +00:00
|
|
|
bumpCycle();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
/// Release pending ready nodes in to the available queue. This makes them
|
|
|
|
/// visible to heuristics.
|
|
|
|
void ConvergingScheduler::SchedBoundary::releasePending() {
|
|
|
|
// If the available queue is empty, it is safe to reset MinReadyCycle.
|
|
|
|
if (Available.empty())
|
|
|
|
MinReadyCycle = UINT_MAX;
|
|
|
|
|
|
|
|
// Check to see if any of the pending instructions are ready to issue. If
|
|
|
|
// so, add them to the available queue.
|
|
|
|
for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
|
|
|
|
SUnit *SU = *(Pending.begin()+i);
|
2012-06-05 21:11:27 +00:00
|
|
|
unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
|
2012-05-24 22:11:09 +00:00
|
|
|
|
|
|
|
if (ReadyCycle < MinReadyCycle)
|
|
|
|
MinReadyCycle = ReadyCycle;
|
|
|
|
|
|
|
|
if (ReadyCycle > CurrCycle)
|
|
|
|
continue;
|
|
|
|
|
2012-06-29 03:23:24 +00:00
|
|
|
if (checkHazard(SU))
|
2012-05-24 22:11:09 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
Available.push(SU);
|
|
|
|
Pending.remove(Pending.begin()+i);
|
|
|
|
--i; --e;
|
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
DEBUG(if (!Pending.empty()) Pending.dump());
|
2012-05-24 22:11:09 +00:00
|
|
|
CheckPending = false;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Remove SU from the ready set for this boundary.
|
|
|
|
void ConvergingScheduler::SchedBoundary::removeReady(SUnit *SU) {
|
|
|
|
if (Available.isInQueue(SU))
|
|
|
|
Available.remove(Available.find(SU));
|
|
|
|
else {
|
|
|
|
assert(Pending.isInQueue(SU) && "bad ready count");
|
|
|
|
Pending.remove(Pending.find(SU));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// If this queue only has one ready candidate, return it. As a side effect,
|
2012-11-07 07:05:09 +00:00
|
|
|
/// defer any nodes that now hit a hazard, and advance the cycle until at least
|
|
|
|
/// one node is ready. If multiple instructions are ready, return NULL.
|
2012-05-24 22:11:09 +00:00
|
|
|
SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() {
|
|
|
|
if (CheckPending)
|
|
|
|
releasePending();
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
if (IssueCount > 0) {
|
|
|
|
// Defer any ready instrs that now have a hazard.
|
|
|
|
for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
|
|
|
|
if (checkHazard(*I)) {
|
|
|
|
Pending.push(*I);
|
|
|
|
I = Available.remove(I);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
++I;
|
|
|
|
}
|
|
|
|
}
|
2012-05-24 22:11:09 +00:00
|
|
|
for (unsigned i = 0; Available.empty(); ++i) {
|
2012-06-05 21:11:27 +00:00
|
|
|
assert(i <= (HazardRec->getMaxLookAhead() + MaxMinLatency) &&
|
|
|
|
"permanent hazard"); (void)i;
|
2012-05-24 22:11:09 +00:00
|
|
|
bumpCycle();
|
|
|
|
releasePending();
|
|
|
|
}
|
|
|
|
if (Available.size() == 1)
|
|
|
|
return *Available.begin();
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
/// Record the candidate policy for opposite zones with different critical
|
|
|
|
/// resources.
|
|
|
|
///
|
|
|
|
/// If the CriticalZone is latency limited, don't force a policy for the
|
2012-12-18 20:52:56 +00:00
|
|
|
/// candidates here. Instead, setLatencyPolicy sets ReduceLatency if needed.
|
2012-11-07 07:05:09 +00:00
|
|
|
void ConvergingScheduler::balanceZones(
|
|
|
|
ConvergingScheduler::SchedBoundary &CriticalZone,
|
|
|
|
ConvergingScheduler::SchedCandidate &CriticalCand,
|
|
|
|
ConvergingScheduler::SchedBoundary &OppositeZone,
|
|
|
|
ConvergingScheduler::SchedCandidate &OppositeCand) {
|
|
|
|
|
|
|
|
if (!CriticalZone.IsResourceLimited)
|
|
|
|
return;
|
2012-12-18 20:52:56 +00:00
|
|
|
assert(SchedModel->hasInstrSchedModel() && "required schedmodel");
|
2012-11-07 07:05:09 +00:00
|
|
|
|
|
|
|
SchedRemainder *Rem = CriticalZone.Rem;
|
|
|
|
|
|
|
|
// If the critical zone is overconsuming a resource relative to the
|
|
|
|
// remainder, try to reduce it.
|
|
|
|
unsigned RemainingCritCount =
|
|
|
|
Rem->RemainingCounts[CriticalZone.CritResIdx];
|
2012-12-18 20:52:56 +00:00
|
|
|
if ((int)(Rem->getMaxRemainingCount(SchedModel) - RemainingCritCount)
|
2012-11-07 07:05:09 +00:00
|
|
|
> (int)SchedModel->getLatencyFactor()) {
|
|
|
|
CriticalCand.Policy.ReduceResIdx = CriticalZone.CritResIdx;
|
|
|
|
DEBUG(dbgs() << "Balance " << CriticalZone.Available.getName() << " reduce "
|
|
|
|
<< SchedModel->getProcResource(CriticalZone.CritResIdx)->Name
|
|
|
|
<< '\n');
|
|
|
|
}
|
|
|
|
// If the other zone is underconsuming a resource relative to the full zone,
|
|
|
|
// try to increase it.
|
|
|
|
unsigned OppositeCount =
|
|
|
|
OppositeZone.ResourceCounts[CriticalZone.CritResIdx];
|
|
|
|
if ((int)(OppositeZone.ExpectedCount - OppositeCount)
|
|
|
|
> (int)SchedModel->getLatencyFactor()) {
|
|
|
|
OppositeCand.Policy.DemandResIdx = CriticalZone.CritResIdx;
|
|
|
|
DEBUG(dbgs() << "Balance " << OppositeZone.Available.getName() << " demand "
|
|
|
|
<< SchedModel->getProcResource(OppositeZone.CritResIdx)->Name
|
|
|
|
<< '\n');
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Determine if the scheduled zones exceed resource limits or critical path and
|
|
|
|
/// set each candidate's ReduceHeight policy accordingly.
|
|
|
|
void ConvergingScheduler::checkResourceLimits(
|
|
|
|
ConvergingScheduler::SchedCandidate &TopCand,
|
|
|
|
ConvergingScheduler::SchedCandidate &BotCand) {
|
|
|
|
|
2012-12-18 20:52:56 +00:00
|
|
|
// Set ReduceLatency to true if needed.
|
|
|
|
Bot.setLatencyPolicy(TopCand.Policy);
|
|
|
|
Top.setLatencyPolicy(BotCand.Policy);
|
2012-11-07 07:05:09 +00:00
|
|
|
|
|
|
|
// Handle resource-limited regions.
|
|
|
|
if (Top.IsResourceLimited && Bot.IsResourceLimited
|
|
|
|
&& Top.CritResIdx == Bot.CritResIdx) {
|
|
|
|
// If the scheduled critical resource in both zones is no longer the
|
|
|
|
// critical remaining resource, attempt to reduce resource height both ways.
|
|
|
|
if (Top.CritResIdx != Rem.CritResIdx) {
|
|
|
|
TopCand.Policy.ReduceResIdx = Top.CritResIdx;
|
|
|
|
BotCand.Policy.ReduceResIdx = Bot.CritResIdx;
|
|
|
|
DEBUG(dbgs() << "Reduce scheduled "
|
|
|
|
<< SchedModel->getProcResource(Top.CritResIdx)->Name << '\n');
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
// Handle latency-limited regions.
|
|
|
|
if (!Top.IsResourceLimited && !Bot.IsResourceLimited) {
|
|
|
|
// If the total scheduled expected latency exceeds the region's critical
|
|
|
|
// path then reduce latency both ways.
|
|
|
|
//
|
|
|
|
// Just because a zone is not resource limited does not mean it is latency
|
|
|
|
// limited. Unbuffered resource, such as max micro-ops may cause CurrCycle
|
|
|
|
// to exceed expected latency.
|
|
|
|
if ((Top.ExpectedLatency + Bot.ExpectedLatency >= Rem.CriticalPath)
|
|
|
|
&& (Rem.CriticalPath > Top.CurrCycle + Bot.CurrCycle)) {
|
|
|
|
TopCand.Policy.ReduceLatency = true;
|
|
|
|
BotCand.Policy.ReduceLatency = true;
|
|
|
|
DEBUG(dbgs() << "Reduce scheduled latency " << Top.ExpectedLatency
|
|
|
|
<< " + " << Bot.ExpectedLatency << '\n');
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
// The critical resource is different in each zone, so request balancing.
|
|
|
|
|
|
|
|
// Compute the cost of each zone.
|
|
|
|
Top.ExpectedCount = std::max(Top.ExpectedLatency, Top.CurrCycle);
|
|
|
|
Top.ExpectedCount = std::max(
|
|
|
|
Top.getCriticalCount(),
|
|
|
|
Top.ExpectedCount * SchedModel->getLatencyFactor());
|
|
|
|
Bot.ExpectedCount = std::max(Bot.ExpectedLatency, Bot.CurrCycle);
|
|
|
|
Bot.ExpectedCount = std::max(
|
|
|
|
Bot.getCriticalCount(),
|
|
|
|
Bot.ExpectedCount * SchedModel->getLatencyFactor());
|
|
|
|
|
|
|
|
balanceZones(Top, TopCand, Bot, BotCand);
|
|
|
|
balanceZones(Bot, BotCand, Top, TopCand);
|
|
|
|
}
|
|
|
|
|
|
|
|
void ConvergingScheduler::SchedCandidate::
|
|
|
|
initResourceDelta(const ScheduleDAGMI *DAG,
|
|
|
|
const TargetSchedModel *SchedModel) {
|
|
|
|
if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
|
|
|
|
return;
|
|
|
|
|
|
|
|
const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
|
|
|
|
for (TargetSchedModel::ProcResIter
|
|
|
|
PI = SchedModel->getWriteProcResBegin(SC),
|
|
|
|
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
|
|
|
|
if (PI->ProcResourceIdx == Policy.ReduceResIdx)
|
|
|
|
ResDelta.CritResources += PI->Cycles;
|
|
|
|
if (PI->ProcResourceIdx == Policy.DemandResIdx)
|
|
|
|
ResDelta.DemandedResources += PI->Cycles;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Return true if this heuristic determines order.
|
|
|
|
static bool tryLess(unsigned TryVal, unsigned CandVal,
|
|
|
|
ConvergingScheduler::SchedCandidate &TryCand,
|
|
|
|
ConvergingScheduler::SchedCandidate &Cand,
|
|
|
|
ConvergingScheduler::CandReason Reason) {
|
|
|
|
if (TryVal < CandVal) {
|
|
|
|
TryCand.Reason = Reason;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (TryVal > CandVal) {
|
|
|
|
if (Cand.Reason > Reason)
|
|
|
|
Cand.Reason = Reason;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
2012-11-12 19:40:10 +00:00
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
static bool tryGreater(unsigned TryVal, unsigned CandVal,
|
|
|
|
ConvergingScheduler::SchedCandidate &TryCand,
|
|
|
|
ConvergingScheduler::SchedCandidate &Cand,
|
|
|
|
ConvergingScheduler::CandReason Reason) {
|
|
|
|
if (TryVal > CandVal) {
|
|
|
|
TryCand.Reason = Reason;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (TryVal < CandVal) {
|
|
|
|
if (Cand.Reason > Reason)
|
|
|
|
Cand.Reason = Reason;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
|
|
|
|
return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
|
|
|
|
}
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
/// Apply a set of heursitics to a new candidate. Heuristics are currently
|
|
|
|
/// hierarchical. This may be more efficient than a graduated cost model because
|
|
|
|
/// we don't need to evaluate all aspects of the model for each node in the
|
|
|
|
/// queue. But it's really done to make the heuristics easier to debug and
|
|
|
|
/// statistically analyze.
|
|
|
|
///
|
|
|
|
/// \param Cand provides the policy and current best candidate.
|
|
|
|
/// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
|
|
|
|
/// \param Zone describes the scheduled zone that we are extending.
|
|
|
|
/// \param RPTracker describes reg pressure within the scheduled zone.
|
|
|
|
/// \param TempTracker is a scratch pressure tracker to reuse in queries.
|
|
|
|
void ConvergingScheduler::tryCandidate(SchedCandidate &Cand,
|
|
|
|
SchedCandidate &TryCand,
|
|
|
|
SchedBoundary &Zone,
|
|
|
|
const RegPressureTracker &RPTracker,
|
|
|
|
RegPressureTracker &TempTracker) {
|
|
|
|
|
|
|
|
// Always initialize TryCand's RPDelta.
|
|
|
|
TempTracker.getMaxPressureDelta(TryCand.SU->getInstr(), TryCand.RPDelta,
|
|
|
|
DAG->getRegionCriticalPSets(),
|
|
|
|
DAG->getRegPressure().MaxSetPressure);
|
|
|
|
|
|
|
|
// Initialize the candidate if needed.
|
|
|
|
if (!Cand.isValid()) {
|
|
|
|
TryCand.Reason = NodeOrder;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
// Avoid exceeding the target's limit.
|
|
|
|
if (tryLess(TryCand.RPDelta.Excess.UnitIncrease,
|
|
|
|
Cand.RPDelta.Excess.UnitIncrease, TryCand, Cand, SingleExcess))
|
|
|
|
return;
|
|
|
|
if (Cand.Reason == SingleExcess)
|
|
|
|
Cand.Reason = MultiPressure;
|
|
|
|
|
|
|
|
// Avoid increasing the max critical pressure in the scheduled region.
|
|
|
|
if (tryLess(TryCand.RPDelta.CriticalMax.UnitIncrease,
|
|
|
|
Cand.RPDelta.CriticalMax.UnitIncrease,
|
|
|
|
TryCand, Cand, SingleCritical))
|
|
|
|
return;
|
|
|
|
if (Cand.Reason == SingleCritical)
|
|
|
|
Cand.Reason = MultiPressure;
|
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
// Keep clustered nodes together to encourage downstream peephole
|
|
|
|
// optimizations which may reduce resource requirements.
|
|
|
|
//
|
|
|
|
// This is a best effort to set things up for a post-RA pass. Optimizations
|
|
|
|
// like generating loads of multiple registers should ideally be done within
|
|
|
|
// the scheduler pass by combining the loads during DAG postprocessing.
|
|
|
|
const SUnit *NextClusterSU =
|
|
|
|
Zone.isTop() ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
|
|
|
|
if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
|
|
|
|
TryCand, Cand, Cluster))
|
|
|
|
return;
|
|
|
|
// Currently, weak edges are for clustering, so we hard-code that reason.
|
|
|
|
// However, deferring the current TryCand will not change Cand's reason.
|
|
|
|
CandReason OrigReason = Cand.Reason;
|
|
|
|
if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
|
|
|
|
getWeakLeft(Cand.SU, Zone.isTop()),
|
|
|
|
TryCand, Cand, Cluster)) {
|
|
|
|
Cand.Reason = OrigReason;
|
|
|
|
return;
|
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
// Avoid critical resource consumption and balance the schedule.
|
|
|
|
TryCand.initResourceDelta(DAG, SchedModel);
|
|
|
|
if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
|
|
|
|
TryCand, Cand, ResourceReduce))
|
|
|
|
return;
|
|
|
|
if (tryGreater(TryCand.ResDelta.DemandedResources,
|
|
|
|
Cand.ResDelta.DemandedResources,
|
|
|
|
TryCand, Cand, ResourceDemand))
|
|
|
|
return;
|
|
|
|
|
|
|
|
// Avoid serializing long latency dependence chains.
|
|
|
|
if (Cand.Policy.ReduceLatency) {
|
|
|
|
if (Zone.isTop()) {
|
|
|
|
if (Cand.SU->getDepth() * SchedModel->getLatencyFactor()
|
|
|
|
> Zone.ExpectedCount) {
|
|
|
|
if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
|
|
|
|
TryCand, Cand, TopDepthReduce))
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
|
|
|
|
TryCand, Cand, TopPathReduce))
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
if (Cand.SU->getHeight() * SchedModel->getLatencyFactor()
|
|
|
|
> Zone.ExpectedCount) {
|
|
|
|
if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
|
|
|
|
TryCand, Cand, BotHeightReduce))
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
|
|
|
|
TryCand, Cand, BotPathReduce))
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Avoid increasing the max pressure of the entire region.
|
|
|
|
if (tryLess(TryCand.RPDelta.CurrentMax.UnitIncrease,
|
|
|
|
Cand.RPDelta.CurrentMax.UnitIncrease, TryCand, Cand, SingleMax))
|
|
|
|
return;
|
|
|
|
if (Cand.Reason == SingleMax)
|
|
|
|
Cand.Reason = MultiPressure;
|
|
|
|
|
|
|
|
// Prefer immediate defs/users of the last scheduled instruction. This is a
|
|
|
|
// nice pressure avoidance strategy that also conserves the processor's
|
|
|
|
// register renaming resources and keeps the machine code readable.
|
2012-11-12 19:40:10 +00:00
|
|
|
if (tryGreater(Zone.NextSUs.count(TryCand.SU), Zone.NextSUs.count(Cand.SU),
|
|
|
|
TryCand, Cand, NextDefUse))
|
2012-11-07 07:05:09 +00:00
|
|
|
return;
|
2012-11-12 19:40:10 +00:00
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
// Fall through to original instruction order.
|
|
|
|
if ((Zone.isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
|
|
|
|
|| (!Zone.isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
|
|
|
|
TryCand.Reason = NodeOrder;
|
|
|
|
}
|
2012-05-10 21:06:19 +00:00
|
|
|
}
|
|
|
|
|
2012-05-17 22:37:09 +00:00
|
|
|
/// pickNodeFromQueue helper that returns true if the LHS reg pressure effect is
|
|
|
|
/// more desirable than RHS from scheduling standpoint.
|
2012-05-17 18:35:10 +00:00
|
|
|
static bool compareRPDelta(const RegPressureDelta &LHS,
|
|
|
|
const RegPressureDelta &RHS) {
|
|
|
|
// Compare each component of pressure in decreasing order of importance
|
|
|
|
// without checking if any are valid. Invalid PressureElements are assumed to
|
|
|
|
// have UnitIncrease==0, so are neutral.
|
2012-05-24 22:11:01 +00:00
|
|
|
|
|
|
|
// Avoid increasing the max critical pressure in the scheduled region.
|
2012-11-07 07:05:09 +00:00
|
|
|
if (LHS.Excess.UnitIncrease != RHS.Excess.UnitIncrease) {
|
|
|
|
DEBUG(dbgs() << "RP excess top - bot: "
|
|
|
|
<< (LHS.Excess.UnitIncrease - RHS.Excess.UnitIncrease) << '\n');
|
2012-05-17 18:35:10 +00:00
|
|
|
return LHS.Excess.UnitIncrease < RHS.Excess.UnitIncrease;
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2012-05-24 22:11:01 +00:00
|
|
|
// Avoid increasing the max critical pressure in the scheduled region.
|
2012-11-07 07:05:09 +00:00
|
|
|
if (LHS.CriticalMax.UnitIncrease != RHS.CriticalMax.UnitIncrease) {
|
|
|
|
DEBUG(dbgs() << "RP critical top - bot: "
|
|
|
|
<< (LHS.CriticalMax.UnitIncrease - RHS.CriticalMax.UnitIncrease)
|
|
|
|
<< '\n');
|
2012-05-17 18:35:10 +00:00
|
|
|
return LHS.CriticalMax.UnitIncrease < RHS.CriticalMax.UnitIncrease;
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2012-05-24 22:11:01 +00:00
|
|
|
// Avoid increasing the max pressure of the entire region.
|
2012-11-07 07:05:09 +00:00
|
|
|
if (LHS.CurrentMax.UnitIncrease != RHS.CurrentMax.UnitIncrease) {
|
|
|
|
DEBUG(dbgs() << "RP current top - bot: "
|
|
|
|
<< (LHS.CurrentMax.UnitIncrease - RHS.CurrentMax.UnitIncrease)
|
|
|
|
<< '\n');
|
2012-05-17 18:35:10 +00:00
|
|
|
return LHS.CurrentMax.UnitIncrease < RHS.CurrentMax.UnitIncrease;
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2012-05-17 18:35:10 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
const char *ConvergingScheduler::getReasonStr(
|
|
|
|
ConvergingScheduler::CandReason Reason) {
|
|
|
|
switch (Reason) {
|
|
|
|
case NoCand: return "NOCAND ";
|
|
|
|
case SingleExcess: return "REG-EXCESS";
|
|
|
|
case SingleCritical: return "REG-CRIT ";
|
2012-11-12 19:40:10 +00:00
|
|
|
case Cluster: return "CLUSTER ";
|
2012-11-07 07:05:09 +00:00
|
|
|
case SingleMax: return "REG-MAX ";
|
|
|
|
case MultiPressure: return "REG-MULTI ";
|
|
|
|
case ResourceReduce: return "RES-REDUCE";
|
|
|
|
case ResourceDemand: return "RES-DEMAND";
|
|
|
|
case TopDepthReduce: return "TOP-DEPTH ";
|
|
|
|
case TopPathReduce: return "TOP-PATH ";
|
|
|
|
case BotHeightReduce:return "BOT-HEIGHT";
|
|
|
|
case BotPathReduce: return "BOT-PATH ";
|
|
|
|
case NextDefUse: return "DEF-USE ";
|
|
|
|
case NodeOrder: return "ORDER ";
|
|
|
|
};
|
2012-11-09 15:45:22 +00:00
|
|
|
llvm_unreachable("Unknown reason!");
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand,
|
|
|
|
const SchedBoundary &Zone) {
|
|
|
|
const char *Label = getReasonStr(Cand.Reason);
|
|
|
|
PressureElement P;
|
|
|
|
unsigned ResIdx = 0;
|
|
|
|
unsigned Latency = 0;
|
|
|
|
switch (Cand.Reason) {
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
case SingleExcess:
|
|
|
|
P = Cand.RPDelta.Excess;
|
|
|
|
break;
|
|
|
|
case SingleCritical:
|
|
|
|
P = Cand.RPDelta.CriticalMax;
|
|
|
|
break;
|
|
|
|
case SingleMax:
|
|
|
|
P = Cand.RPDelta.CurrentMax;
|
|
|
|
break;
|
|
|
|
case ResourceReduce:
|
|
|
|
ResIdx = Cand.Policy.ReduceResIdx;
|
|
|
|
break;
|
|
|
|
case ResourceDemand:
|
|
|
|
ResIdx = Cand.Policy.DemandResIdx;
|
|
|
|
break;
|
|
|
|
case TopDepthReduce:
|
|
|
|
Latency = Cand.SU->getDepth();
|
|
|
|
break;
|
|
|
|
case TopPathReduce:
|
|
|
|
Latency = Cand.SU->getHeight();
|
|
|
|
break;
|
|
|
|
case BotHeightReduce:
|
|
|
|
Latency = Cand.SU->getHeight();
|
|
|
|
break;
|
|
|
|
case BotPathReduce:
|
|
|
|
Latency = Cand.SU->getDepth();
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
dbgs() << Label << " " << Zone.Available.getName() << " ";
|
|
|
|
if (P.isValid())
|
|
|
|
dbgs() << TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease
|
|
|
|
<< " ";
|
|
|
|
else
|
|
|
|
dbgs() << " ";
|
|
|
|
if (ResIdx)
|
|
|
|
dbgs() << SchedModel->getProcResource(ResIdx)->Name << " ";
|
|
|
|
else
|
|
|
|
dbgs() << " ";
|
|
|
|
if (Latency)
|
|
|
|
dbgs() << Latency << " cycles ";
|
|
|
|
else
|
|
|
|
dbgs() << " ";
|
|
|
|
Cand.SU->dump(DAG);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2012-05-10 21:06:16 +00:00
|
|
|
/// Pick the best candidate from the top queue.
|
|
|
|
///
|
|
|
|
/// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
|
|
|
|
/// DAG building. To adjust for the current scheduling location we need to
|
|
|
|
/// maintain the number of vreg uses remaining to be top-scheduled.
|
2012-11-07 07:05:09 +00:00
|
|
|
void ConvergingScheduler::pickNodeFromQueue(SchedBoundary &Zone,
|
|
|
|
const RegPressureTracker &RPTracker,
|
|
|
|
SchedCandidate &Cand) {
|
|
|
|
ReadyQueue &Q = Zone.Available;
|
|
|
|
|
2012-05-24 22:11:12 +00:00
|
|
|
DEBUG(Q.dump());
|
2012-05-17 18:35:10 +00:00
|
|
|
|
2012-05-10 21:06:16 +00:00
|
|
|
// getMaxPressureDelta temporarily modifies the tracker.
|
|
|
|
RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
|
|
|
|
|
2012-05-24 22:11:03 +00:00
|
|
|
for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
|
2012-05-10 21:06:16 +00:00
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
SchedCandidate TryCand(Cand.Policy);
|
|
|
|
TryCand.SU = *I;
|
|
|
|
tryCandidate(Cand, TryCand, Zone, RPTracker, TempTracker);
|
|
|
|
if (TryCand.Reason != NoCand) {
|
|
|
|
// Initialize resource delta if needed in case future heuristics query it.
|
|
|
|
if (TryCand.ResDelta == SchedResourceDelta())
|
|
|
|
TryCand.initResourceDelta(DAG, SchedModel);
|
|
|
|
Cand.setBest(TryCand);
|
|
|
|
DEBUG(traceCandidate(Cand, Zone));
|
2012-05-10 21:06:16 +00:00
|
|
|
}
|
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void tracePick(const ConvergingScheduler::SchedCandidate &Cand,
|
|
|
|
bool IsTop) {
|
|
|
|
DEBUG(dbgs() << "Pick " << (IsTop ? "top" : "bot")
|
|
|
|
<< " SU(" << Cand.SU->NodeNum << ") "
|
|
|
|
<< ConvergingScheduler::getReasonStr(Cand.Reason) << '\n');
|
2012-05-10 21:06:16 +00:00
|
|
|
}
|
|
|
|
|
2012-05-17 18:35:10 +00:00
|
|
|
/// Pick the best candidate node from either the top or bottom queue.
|
2012-11-07 07:05:09 +00:00
|
|
|
SUnit *ConvergingScheduler::pickNodeBidirectional(bool &IsTopNode) {
|
2012-05-17 18:35:10 +00:00
|
|
|
// Schedule as far as possible in the direction of no choice. This is most
|
|
|
|
// efficient, but also provides the best heuristics for CriticalPSets.
|
2012-05-24 22:11:09 +00:00
|
|
|
if (SUnit *SU = Bot.pickOnlyChoice()) {
|
2012-05-17 18:35:10 +00:00
|
|
|
IsTopNode = false;
|
2012-05-24 22:11:09 +00:00
|
|
|
return SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
2012-05-24 22:11:09 +00:00
|
|
|
if (SUnit *SU = Top.pickOnlyChoice()) {
|
2012-05-17 18:35:10 +00:00
|
|
|
IsTopNode = true;
|
2012-05-24 22:11:09 +00:00
|
|
|
return SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
CandPolicy NoPolicy;
|
|
|
|
SchedCandidate BotCand(NoPolicy);
|
|
|
|
SchedCandidate TopCand(NoPolicy);
|
|
|
|
checkResourceLimits(TopCand, BotCand);
|
|
|
|
|
2012-05-17 18:35:10 +00:00
|
|
|
// Prefer bottom scheduling when heuristics are silent.
|
2012-11-07 07:05:09 +00:00
|
|
|
pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
|
|
|
|
assert(BotCand.Reason != NoCand && "failed to find the first candidate");
|
2012-05-17 18:35:10 +00:00
|
|
|
|
|
|
|
// If either Q has a single candidate that provides the least increase in
|
|
|
|
// Excess pressure, we can immediately schedule from that Q.
|
|
|
|
//
|
|
|
|
// RegionCriticalPSets summarizes the pressure within the scheduled region and
|
|
|
|
// affects picking from either Q. If scheduling in one direction must
|
|
|
|
// increase pressure for one of the excess PSets, then schedule in that
|
|
|
|
// direction first to provide more freedom in the other direction.
|
2012-11-07 07:05:09 +00:00
|
|
|
if (BotCand.Reason == SingleExcess || BotCand.Reason == SingleCritical) {
|
2012-05-17 18:35:10 +00:00
|
|
|
IsTopNode = false;
|
2012-11-07 07:05:09 +00:00
|
|
|
tracePick(BotCand, IsTopNode);
|
2012-05-24 22:11:09 +00:00
|
|
|
return BotCand.SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
|
|
|
// Check if the top Q has a better candidate.
|
2012-11-07 07:05:09 +00:00
|
|
|
pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
|
|
|
|
assert(TopCand.Reason != NoCand && "failed to find the first candidate");
|
2012-05-17 18:35:10 +00:00
|
|
|
|
|
|
|
// If either Q has a single candidate that minimizes pressure above the
|
|
|
|
// original region's pressure pick it.
|
2012-11-07 07:05:09 +00:00
|
|
|
if (TopCand.Reason <= SingleMax || BotCand.Reason <= SingleMax) {
|
|
|
|
if (TopCand.Reason < BotCand.Reason) {
|
|
|
|
IsTopNode = true;
|
|
|
|
tracePick(TopCand, IsTopNode);
|
|
|
|
return TopCand.SU;
|
|
|
|
}
|
2012-05-17 18:35:10 +00:00
|
|
|
IsTopNode = false;
|
2012-11-07 07:05:09 +00:00
|
|
|
tracePick(BotCand, IsTopNode);
|
2012-05-24 22:11:09 +00:00
|
|
|
return BotCand.SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
// Check for a salient pressure difference and pick the best from either side.
|
|
|
|
if (compareRPDelta(TopCand.RPDelta, BotCand.RPDelta)) {
|
2012-05-17 18:35:10 +00:00
|
|
|
IsTopNode = true;
|
2012-11-07 07:05:09 +00:00
|
|
|
tracePick(TopCand, IsTopNode);
|
2012-05-24 22:11:09 +00:00
|
|
|
return TopCand.SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
// Otherwise prefer the bottom candidate, in node order if all else failed.
|
|
|
|
if (TopCand.Reason < BotCand.Reason) {
|
2012-05-17 18:35:10 +00:00
|
|
|
IsTopNode = true;
|
2012-11-07 07:05:09 +00:00
|
|
|
tracePick(TopCand, IsTopNode);
|
2012-05-24 22:11:09 +00:00
|
|
|
return TopCand.SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
|
|
|
IsTopNode = false;
|
2012-11-07 07:05:09 +00:00
|
|
|
tracePick(BotCand, IsTopNode);
|
2012-05-24 22:11:09 +00:00
|
|
|
return BotCand.SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
|
2012-05-10 21:06:16 +00:00
|
|
|
SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) {
|
|
|
|
if (DAG->top() == DAG->bottom()) {
|
2012-05-24 22:11:09 +00:00
|
|
|
assert(Top.Available.empty() && Top.Pending.empty() &&
|
|
|
|
Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
|
2012-05-10 21:06:16 +00:00
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
SUnit *SU;
|
2012-10-08 18:53:53 +00:00
|
|
|
do {
|
|
|
|
if (ForceTopDown) {
|
|
|
|
SU = Top.pickOnlyChoice();
|
|
|
|
if (!SU) {
|
2012-11-07 07:05:09 +00:00
|
|
|
CandPolicy NoPolicy;
|
|
|
|
SchedCandidate TopCand(NoPolicy);
|
|
|
|
pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
|
|
|
|
assert(TopCand.Reason != NoCand && "failed to find the first candidate");
|
2012-10-08 18:53:53 +00:00
|
|
|
SU = TopCand.SU;
|
|
|
|
}
|
|
|
|
IsTopNode = true;
|
2012-05-24 23:11:17 +00:00
|
|
|
}
|
2012-10-08 18:53:53 +00:00
|
|
|
else if (ForceBottomUp) {
|
|
|
|
SU = Bot.pickOnlyChoice();
|
|
|
|
if (!SU) {
|
2012-11-07 07:05:09 +00:00
|
|
|
CandPolicy NoPolicy;
|
|
|
|
SchedCandidate BotCand(NoPolicy);
|
|
|
|
pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
|
|
|
|
assert(BotCand.Reason != NoCand && "failed to find the first candidate");
|
2012-10-08 18:53:53 +00:00
|
|
|
SU = BotCand.SU;
|
|
|
|
}
|
|
|
|
IsTopNode = false;
|
2012-05-24 23:11:17 +00:00
|
|
|
}
|
2012-10-08 18:53:53 +00:00
|
|
|
else {
|
2012-11-07 07:05:09 +00:00
|
|
|
SU = pickNodeBidirectional(IsTopNode);
|
2012-10-08 18:53:53 +00:00
|
|
|
}
|
|
|
|
} while (SU->isScheduled);
|
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
if (SU->isTopReady())
|
|
|
|
Top.removeReady(SU);
|
|
|
|
if (SU->isBottomReady())
|
|
|
|
Bot.removeReady(SU);
|
2012-05-25 02:02:39 +00:00
|
|
|
|
|
|
|
DEBUG(dbgs() << "*** " << (IsTopNode ? "Top" : "Bottom")
|
|
|
|
<< " Scheduling Instruction in cycle "
|
|
|
|
<< (IsTopNode ? Top.CurrCycle : Bot.CurrCycle) << '\n';
|
|
|
|
SU->dump(DAG));
|
2012-05-24 22:11:09 +00:00
|
|
|
return SU;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Update the scheduler's state after scheduling a node. This is the same node
|
|
|
|
/// that was just returned by pickNode(). However, ScheduleDAGMI needs to update
|
2012-06-05 21:11:27 +00:00
|
|
|
/// it's state based on the current cycle before MachineSchedStrategy does.
|
2012-05-24 22:11:09 +00:00
|
|
|
void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) {
|
2012-06-05 21:11:27 +00:00
|
|
|
if (IsTopNode) {
|
|
|
|
SU->TopReadyCycle = Top.CurrCycle;
|
2012-06-29 03:23:22 +00:00
|
|
|
Top.bumpNode(SU);
|
2012-05-10 21:06:16 +00:00
|
|
|
}
|
2012-06-05 21:11:27 +00:00
|
|
|
else {
|
|
|
|
SU->BotReadyCycle = Bot.CurrCycle;
|
2012-06-29 03:23:22 +00:00
|
|
|
Bot.bumpNode(SU);
|
2012-05-10 21:06:16 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
/// Create the standard converging machine scheduler. This will be used as the
|
|
|
|
/// default scheduler if the target does not set a default.
|
|
|
|
static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) {
|
2012-03-14 11:26:37 +00:00
|
|
|
assert((!ForceTopDown || !ForceBottomUp) &&
|
2012-03-14 04:00:41 +00:00
|
|
|
"-misched-topdown incompatible with -misched-bottomup");
|
2012-11-12 19:40:10 +00:00
|
|
|
ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new ConvergingScheduler());
|
|
|
|
// Register DAG post-processors.
|
|
|
|
if (EnableLoadCluster)
|
|
|
|
DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI));
|
2012-11-12 19:52:20 +00:00
|
|
|
if (EnableMacroFusion)
|
|
|
|
DAG->addMutation(new MacroFusion(DAG->TII));
|
2012-11-12 19:40:10 +00:00
|
|
|
return DAG;
|
2012-01-17 06:55:03 +00:00
|
|
|
}
|
2012-03-14 04:00:41 +00:00
|
|
|
static MachineSchedRegistry
|
|
|
|
ConvergingSchedRegistry("converge", "Standard converging scheduler.",
|
|
|
|
createConvergingSched);
|
2012-01-17 06:55:03 +00:00
|
|
|
|
2012-10-15 18:02:27 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// ILP Scheduler. Currently for experimental analysis of heuristics.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
/// \brief Order nodes by the ILP metric.
|
|
|
|
struct ILPOrder {
|
2012-11-28 05:13:28 +00:00
|
|
|
SchedDFSResult *DFSResult;
|
|
|
|
BitVector *ScheduledTrees;
|
2012-10-15 18:02:27 +00:00
|
|
|
bool MaximizeILP;
|
|
|
|
|
2012-11-28 05:13:28 +00:00
|
|
|
ILPOrder(SchedDFSResult *dfs, BitVector *schedtrees, bool MaxILP)
|
|
|
|
: DFSResult(dfs), ScheduledTrees(schedtrees), MaximizeILP(MaxILP) {}
|
2012-10-15 18:02:27 +00:00
|
|
|
|
|
|
|
/// \brief Apply a less-than relation on node priority.
|
2012-11-28 05:13:28 +00:00
|
|
|
///
|
|
|
|
/// (Return true if A comes after B in the Q.)
|
2012-10-15 18:02:27 +00:00
|
|
|
bool operator()(const SUnit *A, const SUnit *B) const {
|
2012-11-28 05:13:28 +00:00
|
|
|
unsigned SchedTreeA = DFSResult->getSubtreeID(A);
|
|
|
|
unsigned SchedTreeB = DFSResult->getSubtreeID(B);
|
|
|
|
if (SchedTreeA != SchedTreeB) {
|
|
|
|
// Unscheduled trees have lower priority.
|
|
|
|
if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
|
|
|
|
return ScheduledTrees->test(SchedTreeB);
|
|
|
|
|
|
|
|
// Trees with shallower connections have have lower priority.
|
|
|
|
if (DFSResult->getSubtreeLevel(SchedTreeA)
|
|
|
|
!= DFSResult->getSubtreeLevel(SchedTreeB)) {
|
|
|
|
return DFSResult->getSubtreeLevel(SchedTreeA)
|
|
|
|
< DFSResult->getSubtreeLevel(SchedTreeB);
|
|
|
|
}
|
|
|
|
}
|
2012-10-15 18:02:27 +00:00
|
|
|
if (MaximizeILP)
|
2012-11-28 05:13:28 +00:00
|
|
|
return DFSResult->getILP(A) < DFSResult->getILP(B);
|
2012-10-15 18:02:27 +00:00
|
|
|
else
|
2012-11-28 05:13:28 +00:00
|
|
|
return DFSResult->getILP(A) > DFSResult->getILP(B);
|
2012-10-15 18:02:27 +00:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
/// \brief Schedule based on the ILP metric.
|
|
|
|
class ILPScheduler : public MachineSchedStrategy {
|
2012-11-28 05:13:28 +00:00
|
|
|
/// In case all subtrees are eventually connected to a common root through
|
|
|
|
/// data dependence (e.g. reduction), place an upper limit on their size.
|
|
|
|
///
|
|
|
|
/// FIXME: A subtree limit is generally good, but in the situation commented
|
|
|
|
/// above, where multiple similar subtrees feed a common root, we should
|
|
|
|
/// only split at a point where the resulting subtrees will be balanced.
|
|
|
|
/// (a motivating test case must be found).
|
|
|
|
static const unsigned SubtreeLimit = 16;
|
|
|
|
|
|
|
|
SchedDFSResult DFSResult;
|
|
|
|
BitVector ScheduledTrees;
|
2012-10-15 18:02:27 +00:00
|
|
|
ILPOrder Cmp;
|
|
|
|
|
|
|
|
std::vector<SUnit*> ReadyQ;
|
|
|
|
public:
|
|
|
|
ILPScheduler(bool MaximizeILP)
|
2012-11-28 05:13:28 +00:00
|
|
|
: DFSResult(/*BottomUp=*/true, SubtreeLimit),
|
|
|
|
Cmp(&DFSResult, &ScheduledTrees, MaximizeILP) {}
|
2012-10-15 18:02:27 +00:00
|
|
|
|
|
|
|
virtual void initialize(ScheduleDAGMI *DAG) {
|
|
|
|
ReadyQ.clear();
|
2012-11-28 05:13:28 +00:00
|
|
|
DFSResult.clear();
|
|
|
|
DFSResult.resize(DAG->SUnits.size());
|
|
|
|
ScheduledTrees.clear();
|
2012-10-15 18:02:27 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
virtual void registerRoots() {
|
2012-11-28 05:13:28 +00:00
|
|
|
DFSResult.compute(ReadyQ);
|
|
|
|
ScheduledTrees.resize(DFSResult.getNumSubtrees());
|
2012-11-29 14:36:26 +00:00
|
|
|
// Restore the heap in ReadyQ with the updated DFS results.
|
|
|
|
std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
2012-10-15 18:02:27 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Implement MachineSchedStrategy interface.
|
|
|
|
/// -----------------------------------------
|
|
|
|
|
2012-11-28 05:13:28 +00:00
|
|
|
/// Callback to select the highest priority node from the ready Q.
|
2012-10-15 18:02:27 +00:00
|
|
|
virtual SUnit *pickNode(bool &IsTopNode) {
|
|
|
|
if (ReadyQ.empty()) return NULL;
|
|
|
|
pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
|
|
|
SUnit *SU = ReadyQ.back();
|
|
|
|
ReadyQ.pop_back();
|
|
|
|
IsTopNode = false;
|
2012-11-28 05:13:28 +00:00
|
|
|
DEBUG(dbgs() << "*** Scheduling " << "SU(" << SU->NodeNum << "): "
|
|
|
|
<< *SU->getInstr()
|
|
|
|
<< " ILP: " << DFSResult.getILP(SU)
|
|
|
|
<< " Tree: " << DFSResult.getSubtreeID(SU) << " @"
|
|
|
|
<< DFSResult.getSubtreeLevel(DFSResult.getSubtreeID(SU))<< '\n');
|
2012-10-15 18:02:27 +00:00
|
|
|
return SU;
|
|
|
|
}
|
|
|
|
|
2012-11-28 05:13:28 +00:00
|
|
|
/// Callback after a node is scheduled. Mark a newly scheduled tree, notify
|
|
|
|
/// DFSResults, and resort the priority Q.
|
|
|
|
virtual void schedNode(SUnit *SU, bool IsTopNode) {
|
|
|
|
assert(!IsTopNode && "SchedDFSResult needs bottom-up");
|
|
|
|
if (!ScheduledTrees.test(DFSResult.getSubtreeID(SU))) {
|
|
|
|
ScheduledTrees.set(DFSResult.getSubtreeID(SU));
|
|
|
|
DFSResult.scheduleTree(DFSResult.getSubtreeID(SU));
|
|
|
|
std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
|
|
|
}
|
|
|
|
}
|
2012-10-15 18:02:27 +00:00
|
|
|
|
|
|
|
virtual void releaseTopNode(SUnit *) { /*only called for top roots*/ }
|
|
|
|
|
|
|
|
virtual void releaseBottomNode(SUnit *SU) {
|
|
|
|
ReadyQ.push_back(SU);
|
|
|
|
std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
|
|
|
}
|
|
|
|
};
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
|
|
|
|
return new ScheduleDAGMI(C, new ILPScheduler(true));
|
|
|
|
}
|
|
|
|
static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
|
|
|
|
return new ScheduleDAGMI(C, new ILPScheduler(false));
|
|
|
|
}
|
|
|
|
static MachineSchedRegistry ILPMaxRegistry(
|
|
|
|
"ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
|
|
|
|
static MachineSchedRegistry ILPMinRegistry(
|
|
|
|
"ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
|
|
|
|
|
2012-01-14 02:17:06 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Machine Instruction Shuffler for Correctness Testing
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2012-01-13 06:30:30 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
namespace {
|
2012-03-14 04:00:41 +00:00
|
|
|
/// Apply a less-than relation on the node order, which corresponds to the
|
|
|
|
/// instruction order prior to scheduling. IsReverse implements greater-than.
|
|
|
|
template<bool IsReverse>
|
|
|
|
struct SUnitOrder {
|
2012-01-17 06:55:07 +00:00
|
|
|
bool operator()(SUnit *A, SUnit *B) const {
|
2012-03-14 04:00:41 +00:00
|
|
|
if (IsReverse)
|
|
|
|
return A->NodeNum > B->NodeNum;
|
|
|
|
else
|
|
|
|
return A->NodeNum < B->NodeNum;
|
2012-01-17 06:55:07 +00:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
2012-01-13 06:30:30 +00:00
|
|
|
/// Reorder instructions as much as possible.
|
2012-03-14 04:00:41 +00:00
|
|
|
class InstructionShuffler : public MachineSchedStrategy {
|
|
|
|
bool IsAlternating;
|
|
|
|
bool IsTopDown;
|
|
|
|
|
|
|
|
// Using a less-than relation (SUnitOrder<false>) for the TopQ priority
|
|
|
|
// gives nodes with a higher number higher priority causing the latest
|
|
|
|
// instructions to be scheduled first.
|
|
|
|
PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
|
|
|
|
TopQ;
|
|
|
|
// When scheduling bottom-up, use greater-than as the queue priority.
|
|
|
|
PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
|
|
|
|
BottomQ;
|
2012-01-13 06:30:30 +00:00
|
|
|
public:
|
2012-03-14 04:00:41 +00:00
|
|
|
InstructionShuffler(bool alternate, bool topdown)
|
|
|
|
: IsAlternating(alternate), IsTopDown(topdown) {}
|
2012-01-13 06:30:30 +00:00
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
virtual void initialize(ScheduleDAGMI *) {
|
|
|
|
TopQ.clear();
|
|
|
|
BottomQ.clear();
|
|
|
|
}
|
2012-01-17 06:55:07 +00:00
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
/// Implement MachineSchedStrategy interface.
|
|
|
|
/// -----------------------------------------
|
|
|
|
|
|
|
|
virtual SUnit *pickNode(bool &IsTopNode) {
|
|
|
|
SUnit *SU;
|
|
|
|
if (IsTopDown) {
|
|
|
|
do {
|
|
|
|
if (TopQ.empty()) return NULL;
|
|
|
|
SU = TopQ.top();
|
|
|
|
TopQ.pop();
|
|
|
|
} while (SU->isScheduled);
|
|
|
|
IsTopNode = true;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
do {
|
|
|
|
if (BottomQ.empty()) return NULL;
|
|
|
|
SU = BottomQ.top();
|
|
|
|
BottomQ.pop();
|
|
|
|
} while (SU->isScheduled);
|
|
|
|
IsTopNode = false;
|
|
|
|
}
|
|
|
|
if (IsAlternating)
|
|
|
|
IsTopDown = !IsTopDown;
|
2012-01-17 06:55:07 +00:00
|
|
|
return SU;
|
|
|
|
}
|
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
virtual void schedNode(SUnit *SU, bool IsTopNode) {}
|
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
virtual void releaseTopNode(SUnit *SU) {
|
|
|
|
TopQ.push(SU);
|
|
|
|
}
|
|
|
|
virtual void releaseBottomNode(SUnit *SU) {
|
|
|
|
BottomQ.push(SU);
|
2012-01-13 06:30:30 +00:00
|
|
|
}
|
|
|
|
};
|
|
|
|
} // namespace
|
|
|
|
|
2012-03-08 01:41:12 +00:00
|
|
|
static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
|
2012-03-14 04:00:41 +00:00
|
|
|
bool Alternate = !ForceTopDown && !ForceBottomUp;
|
|
|
|
bool TopDown = !ForceBottomUp;
|
2012-03-14 11:26:37 +00:00
|
|
|
assert((TopDown || !ForceTopDown) &&
|
2012-03-14 04:00:41 +00:00
|
|
|
"-misched-topdown incompatible with -misched-bottomup");
|
|
|
|
return new ScheduleDAGMI(C, new InstructionShuffler(Alternate, TopDown));
|
2012-01-13 06:30:30 +00:00
|
|
|
}
|
2012-03-14 04:00:41 +00:00
|
|
|
static MachineSchedRegistry ShufflerRegistry(
|
|
|
|
"shuffle", "Shuffle machine instructions alternating directions",
|
|
|
|
createInstructionShuffler);
|
2012-01-13 06:30:30 +00:00
|
|
|
#endif // !NDEBUG
|