2012-04-14 13:54:10 +00:00
|
|
|
//===- MachineScheduler.cpp - Machine Instruction Scheduler ---------------===//
|
|
|
|
//
|
|
|
|
// 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.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2017-06-10 13:44:06 +00:00
|
|
|
#include "llvm/CodeGen/MachineScheduler.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/ADT/ArrayRef.h"
|
|
|
|
#include "llvm/ADT/BitVector.h"
|
|
|
|
#include "llvm/ADT/DenseMap.h"
|
2013-04-08 18:41:23 +00:00
|
|
|
#include "llvm/ADT/PriorityQueue.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/ADT/STLExtras.h"
|
2017-06-10 13:44:06 +00:00
|
|
|
#include "llvm/ADT/SmallVector.h"
|
|
|
|
#include "llvm/ADT/iterator_range.h"
|
2013-04-08 18:41:23 +00:00
|
|
|
#include "llvm/Analysis/AliasAnalysis.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/CodeGen/LiveInterval.h"
|
2017-12-18 20:10:56 +00:00
|
|
|
#include "llvm/CodeGen/LiveIntervals.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/CodeGen/MachineBasicBlock.h"
|
2013-04-08 18:41:23 +00:00
|
|
|
#include "llvm/CodeGen/MachineDominators.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/CodeGen/MachineFunction.h"
|
|
|
|
#include "llvm/CodeGen/MachineFunctionPass.h"
|
|
|
|
#include "llvm/CodeGen/MachineInstr.h"
|
2013-04-08 18:41:23 +00:00
|
|
|
#include "llvm/CodeGen/MachineLoopInfo.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/CodeGen/MachineOperand.h"
|
|
|
|
#include "llvm/CodeGen/MachinePassRegistry.h"
|
2013-12-22 00:04:03 +00:00
|
|
|
#include "llvm/CodeGen/MachineRegisterInfo.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/CodeGen/MachineValueType.h"
|
2012-04-14 13:54:10 +00:00
|
|
|
#include "llvm/CodeGen/Passes.h"
|
2012-08-15 19:34:23 +00:00
|
|
|
#include "llvm/CodeGen/RegisterClassInfo.h"
|
2017-06-10 13:44:06 +00:00
|
|
|
#include "llvm/CodeGen/RegisterPressure.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/CodeGen/ScheduleDAG.h"
|
|
|
|
#include "llvm/CodeGen/ScheduleDAGInstrs.h"
|
|
|
|
#include "llvm/CodeGen/ScheduleDAGMutation.h"
|
2013-04-08 18:41:23 +00:00
|
|
|
#include "llvm/CodeGen/ScheduleDFS.h"
|
2012-08-15 19:34:23 +00:00
|
|
|
#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/CodeGen/SlotIndexes.h"
|
2017-12-18 20:10:56 +00:00
|
|
|
#include "llvm/CodeGen/TargetInstrInfo.h"
|
|
|
|
#include "llvm/CodeGen/TargetLowering.h"
|
2016-07-23 20:41:05 +00:00
|
|
|
#include "llvm/CodeGen/TargetPassConfig.h"
|
2017-12-18 20:10:56 +00:00
|
|
|
#include "llvm/CodeGen/TargetRegisterInfo.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/CodeGen/TargetSchedule.h"
|
2017-12-18 20:10:56 +00:00
|
|
|
#include "llvm/CodeGen/TargetSubtargetInfo.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/MC/LaneBitmask.h"
|
|
|
|
#include "llvm/Pass.h"
|
2012-04-14 13:54:10 +00:00
|
|
|
#include "llvm/Support/CommandLine.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include "llvm/Support/Compiler.h"
|
2012-04-14 13:54:10 +00:00
|
|
|
#include "llvm/Support/Debug.h"
|
|
|
|
#include "llvm/Support/ErrorHandling.h"
|
2013-04-08 18:41:23 +00:00
|
|
|
#include "llvm/Support/GraphWriter.h"
|
2012-04-14 13:54:10 +00:00
|
|
|
#include "llvm/Support/raw_ostream.h"
|
2017-04-16 16:01:22 +00:00
|
|
|
#include <algorithm>
|
|
|
|
#include <cassert>
|
|
|
|
#include <cstdint>
|
|
|
|
#include <iterator>
|
|
|
|
#include <limits>
|
|
|
|
#include <memory>
|
|
|
|
#include <string>
|
|
|
|
#include <tuple>
|
|
|
|
#include <utility>
|
|
|
|
#include <vector>
|
2012-04-14 13:54:10 +00:00
|
|
|
|
|
|
|
using namespace llvm;
|
|
|
|
|
2017-05-29 16:25:25 +00:00
|
|
|
#define DEBUG_TYPE "machine-scheduler"
|
2014-11-24 09:08:18 +00:00
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
namespace llvm {
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
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"));
|
2015-01-18 16:17:27 +00:00
|
|
|
cl::opt<bool>
|
|
|
|
DumpCriticalPathLength("misched-dcpl", cl::Hidden,
|
|
|
|
cl::desc("Print critical path length to stdout"));
|
2017-04-16 16:01:22 +00:00
|
|
|
|
|
|
|
} // end namespace llvm
|
2012-04-14 13:54:10 +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"));
|
|
|
|
|
2015-12-30 11:46:15 +00:00
|
|
|
/// In some situations a few uninteresting nodes depend on nearly all other
|
|
|
|
/// nodes in the graph, provide a cutoff to hide them.
|
|
|
|
static cl::opt<unsigned> ViewMISchedCutoff("view-misched-cutoff", cl::Hidden,
|
|
|
|
cl::desc("Hide nodes with more predecessor/successor than cutoff"));
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
|
|
|
|
cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
|
2014-11-24 09:08:18 +00:00
|
|
|
|
|
|
|
static cl::opt<std::string> SchedOnlyFunc("misched-only-func", cl::Hidden,
|
|
|
|
cl::desc("Only schedule this function"));
|
|
|
|
static cl::opt<unsigned> SchedOnlyBlock("misched-only-block", cl::Hidden,
|
2017-12-18 20:10:56 +00:00
|
|
|
cl::desc("Only schedule this MBB#"));
|
2012-04-14 13:54:10 +00:00
|
|
|
#else
|
|
|
|
static bool ViewMISchedDAGs = false;
|
|
|
|
#endif // NDEBUG
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
/// Avoid quadratic complexity in unusually large basic blocks by limiting the
|
|
|
|
/// size of the ready lists.
|
|
|
|
static cl::opt<unsigned> ReadyListLimit("misched-limit", cl::Hidden,
|
|
|
|
cl::desc("Limit ready list to N instructions"), cl::init(256));
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
static cl::opt<bool> EnableRegPressure("misched-regpressure", cl::Hidden,
|
|
|
|
cl::desc("Enable register pressure scheduling."), cl::init(true));
|
|
|
|
|
|
|
|
static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden,
|
|
|
|
cl::desc("Enable cyclic critical path analysis."), cl::init(true));
|
2013-06-10 20:36:52 +00:00
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
static cl::opt<bool> EnableMemOpCluster("misched-cluster", cl::Hidden,
|
|
|
|
cl::desc("Enable memop clustering."),
|
|
|
|
cl::init(true));
|
2013-04-08 18:41:23 +00:00
|
|
|
|
|
|
|
static cl::opt<bool> VerifyScheduling("verify-misched", cl::Hidden,
|
|
|
|
cl::desc("Verify machine instrs before and after machine scheduling"));
|
|
|
|
|
|
|
|
// DAG subtrees must have at least this many nodes.
|
|
|
|
static const unsigned MinSubtreeSize = 8;
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
// Pin the vtables to this file.
|
|
|
|
void MachineSchedStrategy::anchor() {}
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
void ScheduleDAGMutation::anchor() {}
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Machine Instruction Scheduling Pass and Registry
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2017-04-16 16:01:22 +00:00
|
|
|
MachineSchedContext::MachineSchedContext() {
|
2012-08-15 19:34:23 +00:00
|
|
|
RegClassInfo = new RegisterClassInfo();
|
|
|
|
}
|
|
|
|
|
|
|
|
MachineSchedContext::~MachineSchedContext() {
|
|
|
|
delete RegClassInfo;
|
|
|
|
}
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
namespace {
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
/// Base class for a machine scheduler class that can run at any point.
|
|
|
|
class MachineSchedulerBase : public MachineSchedContext,
|
|
|
|
public MachineFunctionPass {
|
|
|
|
public:
|
|
|
|
MachineSchedulerBase(char &ID): MachineFunctionPass(ID) {}
|
|
|
|
|
|
|
|
void print(raw_ostream &O, const Module* = nullptr) const override;
|
|
|
|
|
|
|
|
protected:
|
2015-12-30 11:46:15 +00:00
|
|
|
void scheduleRegions(ScheduleDAGInstrs &Scheduler, bool FixKillFlags);
|
2014-11-24 09:08:18 +00:00
|
|
|
};
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
/// MachineScheduler runs after coalescing and before register allocation.
|
2014-11-24 09:08:18 +00:00
|
|
|
class MachineScheduler : public MachineSchedulerBase {
|
2012-04-14 13:54:10 +00:00
|
|
|
public:
|
|
|
|
MachineScheduler();
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void getAnalysisUsage(AnalysisUsage &AU) const override;
|
|
|
|
|
|
|
|
bool runOnMachineFunction(MachineFunction&) override;
|
|
|
|
|
|
|
|
static char ID; // Class identification, replacement for typeinfo
|
|
|
|
|
|
|
|
protected:
|
|
|
|
ScheduleDAGInstrs *createMachineScheduler();
|
|
|
|
};
|
2012-04-14 13:54:10 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
/// PostMachineScheduler runs after shortly before code emission.
|
|
|
|
class PostMachineScheduler : public MachineSchedulerBase {
|
|
|
|
public:
|
|
|
|
PostMachineScheduler();
|
2012-04-14 13:54:10 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void getAnalysisUsage(AnalysisUsage &AU) const override;
|
2012-04-14 13:54:10 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
bool runOnMachineFunction(MachineFunction&) override;
|
2012-04-14 13:54:10 +00:00
|
|
|
|
|
|
|
static char ID; // Class identification, replacement for typeinfo
|
2013-12-22 00:04:03 +00:00
|
|
|
|
|
|
|
protected:
|
2014-11-24 09:08:18 +00:00
|
|
|
ScheduleDAGInstrs *createPostMachineScheduler();
|
2012-04-14 13:54:10 +00:00
|
|
|
};
|
2017-04-16 16:01:22 +00:00
|
|
|
|
|
|
|
} // end anonymous namespace
|
2012-04-14 13:54:10 +00:00
|
|
|
|
|
|
|
char MachineScheduler::ID = 0;
|
|
|
|
|
|
|
|
char &llvm::MachineSchedulerID = MachineScheduler::ID;
|
|
|
|
|
2017-05-29 16:25:25 +00:00
|
|
|
INITIALIZE_PASS_BEGIN(MachineScheduler, DEBUG_TYPE,
|
2012-04-14 13:54:10 +00:00
|
|
|
"Machine Instruction Scheduler", false, false)
|
2015-12-30 11:46:15 +00:00
|
|
|
INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
|
2017-04-16 16:01:22 +00:00
|
|
|
INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
|
2012-04-14 13:54:10 +00:00
|
|
|
INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
|
2017-05-29 16:25:25 +00:00
|
|
|
INITIALIZE_PASS_END(MachineScheduler, DEBUG_TYPE,
|
2012-04-14 13:54:10 +00:00
|
|
|
"Machine Instruction Scheduler", false, false)
|
|
|
|
|
2017-12-18 20:10:56 +00:00
|
|
|
MachineScheduler::MachineScheduler() : MachineSchedulerBase(ID) {
|
2012-04-14 13:54:10 +00:00
|
|
|
initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
|
|
|
|
}
|
|
|
|
|
|
|
|
void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
|
|
|
|
AU.setPreservesCFG();
|
|
|
|
AU.addRequiredID(MachineDominatorsID);
|
|
|
|
AU.addRequired<MachineLoopInfo>();
|
2015-12-30 11:46:15 +00:00
|
|
|
AU.addRequired<AAResultsWrapperPass>();
|
2012-04-14 13:54:10 +00:00
|
|
|
AU.addRequired<TargetPassConfig>();
|
|
|
|
AU.addRequired<SlotIndexes>();
|
|
|
|
AU.addPreserved<SlotIndexes>();
|
|
|
|
AU.addRequired<LiveIntervals>();
|
|
|
|
AU.addPreserved<LiveIntervals>();
|
|
|
|
MachineFunctionPass::getAnalysisUsage(AU);
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
char PostMachineScheduler::ID = 0;
|
|
|
|
|
|
|
|
char &llvm::PostMachineSchedulerID = PostMachineScheduler::ID;
|
|
|
|
|
|
|
|
INITIALIZE_PASS(PostMachineScheduler, "postmisched",
|
|
|
|
"PostRA Machine Instruction Scheduler", false, false)
|
|
|
|
|
2017-12-18 20:10:56 +00:00
|
|
|
PostMachineScheduler::PostMachineScheduler() : MachineSchedulerBase(ID) {
|
2014-11-24 09:08:18 +00:00
|
|
|
initializePostMachineSchedulerPass(*PassRegistry::getPassRegistry());
|
|
|
|
}
|
|
|
|
|
|
|
|
void PostMachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
|
|
|
|
AU.setPreservesCFG();
|
|
|
|
AU.addRequiredID(MachineDominatorsID);
|
|
|
|
AU.addRequired<MachineLoopInfo>();
|
|
|
|
AU.addRequired<TargetPassConfig>();
|
|
|
|
MachineFunctionPass::getAnalysisUsage(AU);
|
|
|
|
}
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
MachinePassRegistry MachineSchedRegistry::Registry;
|
|
|
|
|
|
|
|
/// A dummy default scheduler factory indicates whether the scheduler
|
|
|
|
/// is overridden on the command line.
|
|
|
|
static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
|
2014-11-24 09:08:18 +00:00
|
|
|
return nullptr;
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// MachineSchedOpt allows command line selection of the scheduler.
|
|
|
|
static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
|
2017-04-16 16:01:22 +00:00
|
|
|
RegisterPassParser<MachineSchedRegistry>>
|
2012-04-14 13:54:10 +00:00
|
|
|
MachineSchedOpt("misched",
|
|
|
|
cl::init(&useDefaultMachineSched), cl::Hidden,
|
|
|
|
cl::desc("Machine instruction scheduler to use"));
|
|
|
|
|
|
|
|
static MachineSchedRegistry
|
|
|
|
DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
|
|
|
|
useDefaultMachineSched);
|
|
|
|
|
2015-05-27 18:44:32 +00:00
|
|
|
static cl::opt<bool> EnableMachineSched(
|
|
|
|
"enable-misched",
|
|
|
|
cl::desc("Enable the machine instruction scheduling pass."), cl::init(true),
|
|
|
|
cl::Hidden);
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
static cl::opt<bool> EnablePostRAMachineSched(
|
|
|
|
"enable-post-misched",
|
|
|
|
cl::desc("Enable the post-ra machine instruction scheduling pass."),
|
|
|
|
cl::init(true), cl::Hidden);
|
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
/// Decrement this iterator until reaching the top or a non-debug instr.
|
2013-12-22 00:04:03 +00:00
|
|
|
static MachineBasicBlock::const_iterator
|
|
|
|
priorNonDebug(MachineBasicBlock::const_iterator I,
|
|
|
|
MachineBasicBlock::const_iterator Beg) {
|
2012-08-15 19:34:23 +00:00
|
|
|
assert(I != Beg && "reached the top of the region, cannot decrement");
|
|
|
|
while (--I != Beg) {
|
|
|
|
if (!I->isDebugValue())
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return I;
|
|
|
|
}
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
/// Non-const version.
|
|
|
|
static MachineBasicBlock::iterator
|
|
|
|
priorNonDebug(MachineBasicBlock::iterator I,
|
|
|
|
MachineBasicBlock::const_iterator Beg) {
|
2017-01-02 19:17:04 +00:00
|
|
|
return priorNonDebug(MachineBasicBlock::const_iterator(I), Beg)
|
|
|
|
.getNonConstIterator();
|
2013-12-22 00:04:03 +00:00
|
|
|
}
|
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
/// If this iterator is a debug value, increment until reaching the End or a
|
|
|
|
/// non-debug instruction.
|
2013-12-22 00:04:03 +00:00
|
|
|
static MachineBasicBlock::const_iterator
|
|
|
|
nextIfDebug(MachineBasicBlock::const_iterator I,
|
|
|
|
MachineBasicBlock::const_iterator End) {
|
2012-08-15 19:34:23 +00:00
|
|
|
for(; I != End; ++I) {
|
|
|
|
if (!I->isDebugValue())
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return I;
|
|
|
|
}
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
/// Non-const version.
|
|
|
|
static MachineBasicBlock::iterator
|
|
|
|
nextIfDebug(MachineBasicBlock::iterator I,
|
|
|
|
MachineBasicBlock::const_iterator End) {
|
2017-01-02 19:17:04 +00:00
|
|
|
return nextIfDebug(MachineBasicBlock::const_iterator(I), End)
|
|
|
|
.getNonConstIterator();
|
2013-12-22 00:04:03 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Instantiate a ScheduleDAGInstrs that will be owned by the caller.
|
|
|
|
ScheduleDAGInstrs *MachineScheduler::createMachineScheduler() {
|
|
|
|
// Select the scheduler, or set the default.
|
|
|
|
MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
|
|
|
|
if (Ctor != useDefaultMachineSched)
|
|
|
|
return Ctor(this);
|
|
|
|
|
|
|
|
// Get the default scheduler set by the target for this function.
|
|
|
|
ScheduleDAGInstrs *Scheduler = PassConfig->createMachineScheduler(this);
|
|
|
|
if (Scheduler)
|
|
|
|
return Scheduler;
|
|
|
|
|
|
|
|
// Default to GenericScheduler.
|
2014-11-24 09:08:18 +00:00
|
|
|
return createGenericSchedLive(this);
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Instantiate a ScheduleDAGInstrs for PostRA scheduling that will be owned by
|
|
|
|
/// the caller. We don't have a command line option to override the postRA
|
|
|
|
/// scheduler. The Target must configure it.
|
|
|
|
ScheduleDAGInstrs *PostMachineScheduler::createPostMachineScheduler() {
|
|
|
|
// Get the postRA scheduler set by the target for this function.
|
|
|
|
ScheduleDAGInstrs *Scheduler = PassConfig->createPostMachineScheduler(this);
|
|
|
|
if (Scheduler)
|
|
|
|
return Scheduler;
|
|
|
|
|
|
|
|
// Default to GenericScheduler.
|
|
|
|
return createGenericSchedPostRA(this);
|
2013-12-22 00:04:03 +00:00
|
|
|
}
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
/// Top-level MachineScheduler pass driver.
|
|
|
|
///
|
|
|
|
/// Visit blocks in function order. Divide each block into scheduling regions
|
|
|
|
/// 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.
|
|
|
|
///
|
|
|
|
/// This design avoids exposing scheduling boundaries to the DAG builder,
|
|
|
|
/// simplifying the DAG builder's support for "special" target instructions.
|
|
|
|
/// At the same time the design allows target schedulers to operate across
|
|
|
|
/// 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.
|
|
|
|
bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
|
2017-12-18 20:10:56 +00:00
|
|
|
if (skipFunction(mf.getFunction()))
|
2016-07-23 20:41:05 +00:00
|
|
|
return false;
|
|
|
|
|
2015-05-27 18:44:32 +00:00
|
|
|
if (EnableMachineSched.getNumOccurrences()) {
|
|
|
|
if (!EnableMachineSched)
|
|
|
|
return false;
|
|
|
|
} else if (!mf.getSubtarget().enableMachineScheduler())
|
|
|
|
return false;
|
|
|
|
|
2015-12-30 11:46:15 +00:00
|
|
|
DEBUG(dbgs() << "Before MISched:\n"; mf.print(dbgs()));
|
2012-08-15 19:34:23 +00:00
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
// Initialize the context of the pass.
|
|
|
|
MF = &mf;
|
|
|
|
MLI = &getAnalysis<MachineLoopInfo>();
|
|
|
|
MDT = &getAnalysis<MachineDominatorTree>();
|
|
|
|
PassConfig = &getAnalysis<TargetPassConfig>();
|
2015-12-30 11:46:15 +00:00
|
|
|
AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
|
2012-04-14 13:54:10 +00:00
|
|
|
|
|
|
|
LIS = &getAnalysis<LiveIntervals>();
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
if (VerifyScheduling) {
|
2013-12-22 00:04:03 +00:00
|
|
|
DEBUG(LIS->dump());
|
2013-04-08 18:41:23 +00:00
|
|
|
MF->verify(this, "Before machine scheduling.");
|
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
RegClassInfo->runOnMachineFunction(*MF);
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
// Instantiate the selected scheduler for this target, function, and
|
|
|
|
// optimization level.
|
2014-11-24 09:08:18 +00:00
|
|
|
std::unique_ptr<ScheduleDAGInstrs> Scheduler(createMachineScheduler());
|
2015-12-30 11:46:15 +00:00
|
|
|
scheduleRegions(*Scheduler, false);
|
2014-11-24 09:08:18 +00:00
|
|
|
|
|
|
|
DEBUG(LIS->dump());
|
|
|
|
if (VerifyScheduling)
|
|
|
|
MF->verify(this, "After machine scheduling.");
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool PostMachineScheduler::runOnMachineFunction(MachineFunction &mf) {
|
2017-12-18 20:10:56 +00:00
|
|
|
if (skipFunction(mf.getFunction()))
|
2014-11-24 09:08:18 +00:00
|
|
|
return false;
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
if (EnablePostRAMachineSched.getNumOccurrences()) {
|
|
|
|
if (!EnablePostRAMachineSched)
|
|
|
|
return false;
|
|
|
|
} else if (!mf.getSubtarget().enablePostRAScheduler()) {
|
2014-11-24 09:08:18 +00:00
|
|
|
DEBUG(dbgs() << "Subtarget disables post-MI-sched.\n");
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
DEBUG(dbgs() << "Before post-MI-sched:\n"; mf.print(dbgs()));
|
|
|
|
|
|
|
|
// Initialize the context of the pass.
|
|
|
|
MF = &mf;
|
2017-12-18 20:10:56 +00:00
|
|
|
MLI = &getAnalysis<MachineLoopInfo>();
|
2014-11-24 09:08:18 +00:00
|
|
|
PassConfig = &getAnalysis<TargetPassConfig>();
|
|
|
|
|
|
|
|
if (VerifyScheduling)
|
|
|
|
MF->verify(this, "Before post machine scheduling.");
|
|
|
|
|
|
|
|
// Instantiate the selected scheduler for this target, function, and
|
|
|
|
// optimization level.
|
|
|
|
std::unique_ptr<ScheduleDAGInstrs> Scheduler(createPostMachineScheduler());
|
2015-12-30 11:46:15 +00:00
|
|
|
scheduleRegions(*Scheduler, true);
|
2014-11-24 09:08:18 +00:00
|
|
|
|
|
|
|
if (VerifyScheduling)
|
|
|
|
MF->verify(this, "After post machine scheduling.");
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Return true of the given instruction should not be included in a scheduling
|
|
|
|
/// region.
|
|
|
|
///
|
|
|
|
/// MachineScheduler does not currently support scheduling across calls. To
|
|
|
|
/// handle calls, the DAG builder needs to be modified to create register
|
|
|
|
/// anti/output dependencies on the registers clobbered by the call's regmask
|
|
|
|
/// operand. In PreRA scheduling, the stack pointer adjustment already prevents
|
|
|
|
/// scheduling across calls. In PostRA scheduling, we need the isCall to enforce
|
|
|
|
/// the boundary, but there would be no benefit to postRA scheduling across
|
|
|
|
/// calls this late anyway.
|
|
|
|
static bool isSchedBoundary(MachineBasicBlock::iterator MI,
|
|
|
|
MachineBasicBlock *MBB,
|
|
|
|
MachineFunction *MF,
|
2015-12-30 11:46:15 +00:00
|
|
|
const TargetInstrInfo *TII) {
|
2016-07-23 20:41:05 +00:00
|
|
|
return MI->isCall() || TII->isSchedulingBoundary(*MI, MBB, *MF);
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
|
|
|
|
2017-12-18 20:10:56 +00:00
|
|
|
/// A region of an MBB for scheduling.
|
|
|
|
namespace {
|
|
|
|
struct SchedRegion {
|
|
|
|
/// RegionBegin is the first instruction in the scheduling region, and
|
|
|
|
/// RegionEnd is either MBB->end() or the scheduling boundary after the
|
|
|
|
/// last instruction in the scheduling region. These iterators cannot refer
|
|
|
|
/// to instructions outside of the identified scheduling region because
|
|
|
|
/// those may be reordered before scheduling this region.
|
|
|
|
MachineBasicBlock::iterator RegionBegin;
|
|
|
|
MachineBasicBlock::iterator RegionEnd;
|
|
|
|
unsigned NumRegionInstrs;
|
|
|
|
|
|
|
|
SchedRegion(MachineBasicBlock::iterator B, MachineBasicBlock::iterator E,
|
|
|
|
unsigned N) :
|
|
|
|
RegionBegin(B), RegionEnd(E), NumRegionInstrs(N) {}
|
|
|
|
};
|
|
|
|
} // end anonymous namespace
|
|
|
|
|
|
|
|
using MBBRegionsVector = SmallVector<SchedRegion, 16>;
|
|
|
|
|
|
|
|
static void
|
|
|
|
getSchedRegions(MachineBasicBlock *MBB,
|
|
|
|
MBBRegionsVector &Regions,
|
|
|
|
bool RegionsTopDown) {
|
|
|
|
MachineFunction *MF = MBB->getParent();
|
|
|
|
const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
|
|
|
|
|
|
|
|
MachineBasicBlock::iterator I = nullptr;
|
|
|
|
for(MachineBasicBlock::iterator RegionEnd = MBB->end();
|
|
|
|
RegionEnd != MBB->begin(); RegionEnd = I) {
|
|
|
|
|
|
|
|
// Avoid decrementing RegionEnd for blocks with no terminator.
|
|
|
|
if (RegionEnd != MBB->end() ||
|
|
|
|
isSchedBoundary(&*std::prev(RegionEnd), &*MBB, MF, TII)) {
|
|
|
|
--RegionEnd;
|
|
|
|
}
|
|
|
|
|
|
|
|
// The next region starts above the previous region. Look backward in the
|
|
|
|
// instruction stream until we find the nearest boundary.
|
|
|
|
unsigned NumRegionInstrs = 0;
|
|
|
|
I = RegionEnd;
|
|
|
|
for (;I != MBB->begin(); --I) {
|
|
|
|
MachineInstr &MI = *std::prev(I);
|
|
|
|
if (isSchedBoundary(&MI, &*MBB, MF, TII))
|
|
|
|
break;
|
|
|
|
if (!MI.isDebugValue())
|
|
|
|
// MBB::size() uses instr_iterator to count. Here we need a bundle to
|
|
|
|
// count as a single instruction.
|
|
|
|
++NumRegionInstrs;
|
|
|
|
}
|
|
|
|
|
|
|
|
Regions.push_back(SchedRegion(I, RegionEnd, NumRegionInstrs));
|
|
|
|
}
|
|
|
|
|
|
|
|
if (RegionsTopDown)
|
|
|
|
std::reverse(Regions.begin(), Regions.end());
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
/// Main driver for both MachineScheduler and PostMachineScheduler.
|
2015-12-30 11:46:15 +00:00
|
|
|
void MachineSchedulerBase::scheduleRegions(ScheduleDAGInstrs &Scheduler,
|
|
|
|
bool FixKillFlags) {
|
2012-04-14 13:54:10 +00:00
|
|
|
// Visit all machine basic blocks.
|
2012-08-15 19:34:23 +00:00
|
|
|
//
|
|
|
|
// TODO: Visit blocks in global postorder or postorder within the bottom-up
|
|
|
|
// loop tree. Then we can optionally compute global RegPressure.
|
2012-04-14 13:54:10 +00:00
|
|
|
for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
|
|
|
|
MBB != MBBEnd; ++MBB) {
|
|
|
|
|
2015-12-30 11:46:15 +00:00
|
|
|
Scheduler.startBlock(&*MBB);
|
2014-11-24 09:08:18 +00:00
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
if (SchedOnlyFunc.getNumOccurrences() && SchedOnlyFunc != MF->getName())
|
|
|
|
continue;
|
|
|
|
if (SchedOnlyBlock.getNumOccurrences()
|
|
|
|
&& (int)SchedOnlyBlock != MBB->getNumber())
|
|
|
|
continue;
|
|
|
|
#endif
|
2012-04-14 13:54:10 +00:00
|
|
|
|
2017-12-18 20:10:56 +00:00
|
|
|
// Break the block into scheduling regions [I, RegionEnd). RegionEnd
|
|
|
|
// points to the scheduling 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.
|
|
|
|
//
|
|
|
|
// All the regions of MBB are first found and stored in MBBRegions, which
|
|
|
|
// will be processed (MBB) top-down if initialized with true.
|
2012-04-14 13:54:10 +00:00
|
|
|
//
|
|
|
|
// The Scheduler may insert instructions during either schedule() or
|
|
|
|
// exitRegion(), even for empty regions. So the local iterators 'I' and
|
2017-12-18 20:10:56 +00:00
|
|
|
// 'RegionEnd' are invalid across these calls. Instructions must not be
|
|
|
|
// added to other regions than the current one without updating MBBRegions.
|
|
|
|
|
|
|
|
MBBRegionsVector MBBRegions;
|
|
|
|
getSchedRegions(&*MBB, MBBRegions, Scheduler.doMBBSchedRegionsTopDown());
|
|
|
|
for (MBBRegionsVector::iterator R = MBBRegions.begin();
|
|
|
|
R != MBBRegions.end(); ++R) {
|
|
|
|
MachineBasicBlock::iterator I = R->RegionBegin;
|
|
|
|
MachineBasicBlock::iterator RegionEnd = R->RegionEnd;
|
|
|
|
unsigned NumRegionInstrs = R->NumRegionInstrs;
|
2012-04-14 13:54:10 +00:00
|
|
|
|
|
|
|
// Notify the scheduler of the region, even if we may skip scheduling
|
|
|
|
// it. Perhaps it still needs to be bundled.
|
2015-12-30 11:46:15 +00:00
|
|
|
Scheduler.enterRegion(&*MBB, I, RegionEnd, NumRegionInstrs);
|
2012-04-14 13:54:10 +00:00
|
|
|
|
|
|
|
// Skip empty scheduling regions (0 or 1 schedulable instructions).
|
2014-11-24 09:08:18 +00:00
|
|
|
if (I == RegionEnd || I == std::prev(RegionEnd)) {
|
2012-04-14 13:54:10 +00:00
|
|
|
// Close the current region. Bundle the terminator if needed.
|
|
|
|
// This invalidates 'RegionEnd' and 'I'.
|
2014-11-24 09:08:18 +00:00
|
|
|
Scheduler.exitRegion();
|
2012-04-14 13:54:10 +00:00
|
|
|
continue;
|
|
|
|
}
|
2015-12-30 11:46:15 +00:00
|
|
|
DEBUG(dbgs() << "********** MI Scheduling **********\n");
|
2017-12-18 20:10:56 +00:00
|
|
|
DEBUG(dbgs() << MF->getName() << ":" << printMBBReference(*MBB) << " "
|
|
|
|
<< MBB->getName() << "\n From: " << *I << " To: ";
|
2012-04-14 13:54:10 +00:00
|
|
|
if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
|
|
|
|
else dbgs() << "End";
|
2016-07-23 20:41:05 +00:00
|
|
|
dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n');
|
2015-01-18 16:17:27 +00:00
|
|
|
if (DumpCriticalPathLength) {
|
|
|
|
errs() << MF->getName();
|
2017-12-18 20:10:56 +00:00
|
|
|
errs() << ":%bb. " << MBB->getNumber();
|
2015-01-18 16:17:27 +00:00
|
|
|
errs() << " " << MBB->getName() << " \n";
|
|
|
|
}
|
2012-04-14 13:54:10 +00:00
|
|
|
|
|
|
|
// Schedule a region: possibly reorder instructions.
|
2017-12-18 20:10:56 +00:00
|
|
|
// This invalidates the original region iterators.
|
2014-11-24 09:08:18 +00:00
|
|
|
Scheduler.schedule();
|
2012-04-14 13:54:10 +00:00
|
|
|
|
|
|
|
// Close the current region.
|
2014-11-24 09:08:18 +00:00
|
|
|
Scheduler.exitRegion();
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2014-11-24 09:08:18 +00:00
|
|
|
Scheduler.finishBlock();
|
2015-12-30 11:46:15 +00:00
|
|
|
// FIXME: Ideally, no further passes should rely on kill flags. However,
|
|
|
|
// thumb2 size reduction is currently an exception, so the PostMIScheduler
|
|
|
|
// needs to do this.
|
|
|
|
if (FixKillFlags)
|
2017-05-29 16:25:25 +00:00
|
|
|
Scheduler.fixupKills(*MBB);
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2014-11-24 09:08:18 +00:00
|
|
|
Scheduler.finalizeSchedule();
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void MachineSchedulerBase::print(raw_ostream &O, const Module* m) const {
|
2012-04-14 13:54:10 +00:00
|
|
|
// unimplemented
|
|
|
|
}
|
|
|
|
|
2017-04-16 16:01:22 +00:00
|
|
|
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
|
2017-06-26 20:32:52 +00:00
|
|
|
LLVM_DUMP_METHOD void ReadyQueue::dump() const {
|
2015-12-30 11:46:15 +00:00
|
|
|
dbgs() << "Queue " << Name << ": ";
|
2017-06-26 20:32:52 +00:00
|
|
|
for (const SUnit *SU : Queue)
|
|
|
|
dbgs() << SU->NodeNum << " ";
|
2012-12-02 13:10:19 +00:00
|
|
|
dbgs() << "\n";
|
|
|
|
}
|
2017-04-16 16:01:22 +00:00
|
|
|
#endif
|
2012-04-14 13:54:10 +00:00
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
2014-11-24 09:08:18 +00:00
|
|
|
// ScheduleDAGMI - Basic machine instruction scheduling. This is
|
|
|
|
// independent of PreRA/PostRA scheduling and involves no extra book-keeping for
|
|
|
|
// virtual registers.
|
|
|
|
// ===----------------------------------------------------------------------===/
|
2012-04-14 13:54:10 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
// Provide a vtable anchor.
|
2017-04-16 16:01:22 +00:00
|
|
|
ScheduleDAGMI::~ScheduleDAGMI() = default;
|
2013-04-08 18:41:23 +00:00
|
|
|
|
2013-06-10 20:36:52 +00:00
|
|
|
bool ScheduleDAGMI::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
|
|
|
|
return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
|
|
|
|
}
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
|
|
|
|
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());
|
|
|
|
}
|
|
|
|
SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
|
|
|
|
// Return true regardless of whether a new edge needed to be inserted.
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
/// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
|
|
|
|
/// NumPredsLeft reaches zero, release the successor node.
|
2012-08-15 19:34:23 +00:00
|
|
|
///
|
|
|
|
/// FIXME: Adjust SuccSU height based on MinLatency.
|
2012-04-14 13:54:10 +00:00
|
|
|
void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
|
|
|
|
SUnit *SuccSU = SuccEdge->getSUnit();
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
if (SuccEdge->isWeak()) {
|
|
|
|
--SuccSU->WeakPredsLeft;
|
|
|
|
if (SuccEdge->isCluster())
|
|
|
|
NextClusterSucc = SuccSU;
|
|
|
|
return;
|
|
|
|
}
|
2012-04-14 13:54:10 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
if (SuccSU->NumPredsLeft == 0) {
|
|
|
|
dbgs() << "*** Scheduling failed! ***\n";
|
|
|
|
SuccSU->dump(this);
|
|
|
|
dbgs() << " has been released too many times!\n";
|
2014-11-24 09:08:18 +00:00
|
|
|
llvm_unreachable(nullptr);
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
|
|
|
#endif
|
2014-11-24 09:08:18 +00:00
|
|
|
// SU->TopReadyCycle was set to CurrCycle when it was scheduled. However,
|
|
|
|
// CurrCycle may have advanced since then.
|
|
|
|
if (SuccSU->TopReadyCycle < SU->TopReadyCycle + SuccEdge->getLatency())
|
|
|
|
SuccSU->TopReadyCycle = SU->TopReadyCycle + SuccEdge->getLatency();
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
--SuccSU->NumPredsLeft;
|
|
|
|
if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
|
|
|
|
SchedImpl->releaseTopNode(SuccSU);
|
|
|
|
}
|
|
|
|
|
|
|
|
/// releaseSuccessors - Call releaseSucc on each of SU's successors.
|
|
|
|
void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SDep &Succ : SU->Succs)
|
|
|
|
releaseSucc(SU, &Succ);
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
|
|
|
|
/// NumSuccsLeft reaches zero, release the predecessor node.
|
2012-08-15 19:34:23 +00:00
|
|
|
///
|
|
|
|
/// FIXME: Adjust PredSU height based on MinLatency.
|
2012-04-14 13:54:10 +00:00
|
|
|
void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
|
|
|
|
SUnit *PredSU = PredEdge->getSUnit();
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
if (PredEdge->isWeak()) {
|
|
|
|
--PredSU->WeakSuccsLeft;
|
|
|
|
if (PredEdge->isCluster())
|
|
|
|
NextClusterPred = PredSU;
|
|
|
|
return;
|
|
|
|
}
|
2012-04-14 13:54:10 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
if (PredSU->NumSuccsLeft == 0) {
|
|
|
|
dbgs() << "*** Scheduling failed! ***\n";
|
|
|
|
PredSU->dump(this);
|
|
|
|
dbgs() << " has been released too many times!\n";
|
2014-11-24 09:08:18 +00:00
|
|
|
llvm_unreachable(nullptr);
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
|
|
|
#endif
|
2014-11-24 09:08:18 +00:00
|
|
|
// SU->BotReadyCycle was set to CurrCycle when it was scheduled. However,
|
|
|
|
// CurrCycle may have advanced since then.
|
|
|
|
if (PredSU->BotReadyCycle < SU->BotReadyCycle + PredEdge->getLatency())
|
|
|
|
PredSU->BotReadyCycle = SU->BotReadyCycle + PredEdge->getLatency();
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
--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) {
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SDep &Pred : SU->Preds)
|
|
|
|
releasePred(SU, &Pred);
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
|
|
|
|
2017-12-18 20:10:56 +00:00
|
|
|
void ScheduleDAGMI::startBlock(MachineBasicBlock *bb) {
|
|
|
|
ScheduleDAGInstrs::startBlock(bb);
|
|
|
|
SchedImpl->enterMBB(bb);
|
|
|
|
}
|
|
|
|
|
|
|
|
void ScheduleDAGMI::finishBlock() {
|
|
|
|
SchedImpl->leaveMBB();
|
|
|
|
ScheduleDAGInstrs::finishBlock();
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +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 regioninstrs)
|
|
|
|
{
|
|
|
|
ScheduleDAGInstrs::enterRegion(bb, begin, end, regioninstrs);
|
|
|
|
|
|
|
|
SchedImpl->initPolicy(begin, end, regioninstrs);
|
|
|
|
}
|
|
|
|
|
2013-06-10 20:36:52 +00:00
|
|
|
/// This is normally called from the main scheduler loop but may also be invoked
|
|
|
|
/// by the scheduling strategy to perform additional code motion.
|
2014-11-24 09:08:18 +00:00
|
|
|
void ScheduleDAGMI::moveInstruction(
|
|
|
|
MachineInstr *MI, MachineBasicBlock::iterator InsertPos) {
|
2012-08-15 19:34:23 +00:00
|
|
|
// Advance RegionBegin if the first instruction moves down.
|
2012-04-14 13:54:10 +00:00
|
|
|
if (&*RegionBegin == MI)
|
2012-08-15 19:34:23 +00:00
|
|
|
++RegionBegin;
|
|
|
|
|
|
|
|
// Update the instruction stream.
|
2012-04-14 13:54:10 +00:00
|
|
|
BB->splice(InsertPos, BB, MI);
|
2012-08-15 19:34:23 +00:00
|
|
|
|
|
|
|
// Update LiveIntervals
|
2014-11-24 09:08:18 +00:00
|
|
|
if (LIS)
|
2016-07-23 20:41:05 +00:00
|
|
|
LIS->handleMove(*MI, /*UpdateFlags=*/true);
|
2012-08-15 19:34:23 +00:00
|
|
|
|
|
|
|
// Recede RegionBegin if an instruction moves above the first.
|
2012-04-14 13:54:10 +00:00
|
|
|
if (RegionBegin == InsertPos)
|
|
|
|
RegionBegin = MI;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool ScheduleDAGMI::checkSchedLimit() {
|
|
|
|
#ifndef NDEBUG
|
|
|
|
if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
|
|
|
|
CurrentTop = CurrentBottom;
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
++NumInstrsScheduled;
|
|
|
|
#endif
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
/// Per-region scheduling driver, called back from
|
|
|
|
/// MachineScheduler::runOnMachineFunction. This is a simplified driver that
|
|
|
|
/// does not consider liveness or register pressure. It is useful for PostRA
|
|
|
|
/// scheduling and potentially other custom schedulers.
|
|
|
|
void ScheduleDAGMI::schedule() {
|
2015-12-30 11:46:15 +00:00
|
|
|
DEBUG(dbgs() << "ScheduleDAGMI::schedule starting\n");
|
|
|
|
DEBUG(SchedImpl->dumpPolicy());
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
// Build the DAG.
|
|
|
|
buildSchedGraph(AA);
|
|
|
|
|
|
|
|
Topo.InitDAGTopologicalSorting();
|
|
|
|
|
|
|
|
postprocessDAG();
|
|
|
|
|
|
|
|
SmallVector<SUnit*, 8> TopRoots, BotRoots;
|
|
|
|
findRootsAndBiasEdges(TopRoots, BotRoots);
|
|
|
|
|
|
|
|
// Initialize the strategy before modifying the DAG.
|
|
|
|
// This may initialize a DFSResult to be used for queue priority.
|
|
|
|
SchedImpl->initialize(this);
|
|
|
|
|
2017-01-02 19:17:04 +00:00
|
|
|
DEBUG(
|
|
|
|
if (EntrySU.getInstr() != nullptr)
|
|
|
|
EntrySU.dumpAll(this);
|
2017-06-26 20:32:52 +00:00
|
|
|
for (const SUnit &SU : SUnits)
|
|
|
|
SU.dumpAll(this);
|
2017-01-02 19:17:04 +00:00
|
|
|
if (ExitSU.getInstr() != nullptr)
|
|
|
|
ExitSU.dumpAll(this);
|
|
|
|
);
|
2014-11-24 09:08:18 +00:00
|
|
|
if (ViewMISchedDAGs) viewGraph();
|
|
|
|
|
|
|
|
// Initialize ready queues now that the DAG and priority data are finalized.
|
|
|
|
initQueues(TopRoots, BotRoots);
|
|
|
|
|
|
|
|
bool IsTopNode = false;
|
2015-12-30 11:46:15 +00:00
|
|
|
while (true) {
|
|
|
|
DEBUG(dbgs() << "** ScheduleDAGMI::schedule picking next node\n");
|
|
|
|
SUnit *SU = SchedImpl->pickNode(IsTopNode);
|
|
|
|
if (!SU) break;
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
assert(!SU->isScheduled && "Node already scheduled");
|
|
|
|
if (!checkSchedLimit())
|
|
|
|
break;
|
|
|
|
|
|
|
|
MachineInstr *MI = SU->getInstr();
|
|
|
|
if (IsTopNode) {
|
|
|
|
assert(SU->isTopReady() && "node still has unscheduled dependencies");
|
|
|
|
if (&*CurrentTop == MI)
|
|
|
|
CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
|
|
|
|
else
|
|
|
|
moveInstruction(MI, CurrentTop);
|
2016-07-23 20:41:05 +00:00
|
|
|
} else {
|
2014-11-24 09:08:18 +00:00
|
|
|
assert(SU->isBottomReady() && "node still has unscheduled dependencies");
|
|
|
|
MachineBasicBlock::iterator priorII =
|
|
|
|
priorNonDebug(CurrentBottom, CurrentTop);
|
|
|
|
if (&*priorII == MI)
|
|
|
|
CurrentBottom = priorII;
|
|
|
|
else {
|
|
|
|
if (&*CurrentTop == MI)
|
|
|
|
CurrentTop = nextIfDebug(++CurrentTop, priorII);
|
|
|
|
moveInstruction(MI, CurrentBottom);
|
|
|
|
CurrentBottom = MI;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Notify the scheduling strategy before updating the DAG.
|
|
|
|
// This sets the scheduled node's ReadyCycle to CurrCycle. When updateQueues
|
|
|
|
// runs, it can then use the accurate ReadyCycle time to determine whether
|
|
|
|
// newly released nodes can move to the readyQ.
|
|
|
|
SchedImpl->schedNode(SU, IsTopNode);
|
|
|
|
|
|
|
|
updateQueues(SU, IsTopNode);
|
|
|
|
}
|
|
|
|
assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
|
|
|
|
|
|
|
|
placeDebugValues();
|
|
|
|
|
|
|
|
DEBUG({
|
2017-12-18 20:10:56 +00:00
|
|
|
dbgs() << "*** Final schedule for "
|
|
|
|
<< printMBBReference(*begin()->getParent()) << " ***\n";
|
|
|
|
dumpSchedule();
|
|
|
|
dbgs() << '\n';
|
|
|
|
});
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Apply each ScheduleDAGMutation step in order.
|
|
|
|
void ScheduleDAGMI::postprocessDAG() {
|
2017-06-26 20:32:52 +00:00
|
|
|
for (auto &m : Mutations)
|
|
|
|
m->apply(this);
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void ScheduleDAGMI::
|
|
|
|
findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
|
|
|
|
SmallVectorImpl<SUnit*> &BotRoots) {
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SUnit &SU : SUnits) {
|
|
|
|
assert(!SU.isBoundaryNode() && "Boundary node should not be in SUnits");
|
2014-11-24 09:08:18 +00:00
|
|
|
|
|
|
|
// Order predecessors so DFSResult follows the critical path.
|
2017-06-26 20:32:52 +00:00
|
|
|
SU.biasCriticalPath();
|
2014-11-24 09:08:18 +00:00
|
|
|
|
|
|
|
// A SUnit is ready to top schedule if it has no predecessors.
|
2017-06-26 20:32:52 +00:00
|
|
|
if (!SU.NumPredsLeft)
|
|
|
|
TopRoots.push_back(&SU);
|
2014-11-24 09:08:18 +00:00
|
|
|
// A SUnit is ready to bottom schedule if it has no successors.
|
2017-06-26 20:32:52 +00:00
|
|
|
if (!SU.NumSuccsLeft)
|
|
|
|
BotRoots.push_back(&SU);
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
|
|
|
ExitSU.biasCriticalPath();
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Identify DAG roots and setup scheduler queues.
|
|
|
|
void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
|
|
|
|
ArrayRef<SUnit*> BotRoots) {
|
|
|
|
NextClusterSucc = nullptr;
|
|
|
|
NextClusterPred = nullptr;
|
|
|
|
|
|
|
|
// Release all DAG roots for scheduling, not including EntrySU/ExitSU.
|
|
|
|
//
|
|
|
|
// Nodes with unreleased weak edges can still be roots.
|
|
|
|
// Release top roots in forward order.
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SUnit *SU : TopRoots)
|
|
|
|
SchedImpl->releaseTopNode(SU);
|
|
|
|
|
2014-11-24 09:08:18 +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);
|
|
|
|
}
|
|
|
|
|
|
|
|
releaseSuccessors(&EntrySU);
|
|
|
|
releasePredecessors(&ExitSU);
|
|
|
|
|
|
|
|
SchedImpl->registerRoots();
|
|
|
|
|
|
|
|
// Advance past initial DebugValues.
|
|
|
|
CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
|
|
|
|
CurrentBottom = RegionEnd;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// 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;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// 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;
|
|
|
|
}
|
|
|
|
|
2017-04-16 16:01:22 +00:00
|
|
|
for (std::vector<std::pair<MachineInstr *, MachineInstr *>>::iterator
|
2014-11-24 09:08:18 +00:00
|
|
|
DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
|
|
|
|
std::pair<MachineInstr *, MachineInstr *> P = *std::prev(DI);
|
|
|
|
MachineInstr *DbgValue = P.first;
|
|
|
|
MachineBasicBlock::iterator OrigPrevMI = P.second;
|
|
|
|
if (&*RegionBegin == DbgValue)
|
|
|
|
++RegionBegin;
|
|
|
|
BB->splice(++OrigPrevMI, BB, DbgValue);
|
|
|
|
if (OrigPrevMI == std::prev(RegionEnd))
|
|
|
|
RegionEnd = DbgValue;
|
|
|
|
}
|
|
|
|
DbgValues.clear();
|
|
|
|
FirstDbgValue = nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
|
2017-04-16 16:01:22 +00:00
|
|
|
LLVM_DUMP_METHOD void ScheduleDAGMI::dumpSchedule() const {
|
2014-11-24 09:08:18 +00:00
|
|
|
for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
|
|
|
|
if (SUnit *SU = getSUnit(&(*MI)))
|
|
|
|
SU->dump(this);
|
|
|
|
else
|
|
|
|
dbgs() << "Missing SUnit\n";
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// ScheduleDAGMILive - Base class for MachineInstr scheduling with LiveIntervals
|
|
|
|
// preservation.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
ScheduleDAGMILive::~ScheduleDAGMILive() {
|
|
|
|
delete DFSResult;
|
|
|
|
}
|
|
|
|
|
2017-01-02 19:17:04 +00:00
|
|
|
void ScheduleDAGMILive::collectVRegUses(SUnit &SU) {
|
|
|
|
const MachineInstr &MI = *SU.getInstr();
|
|
|
|
for (const MachineOperand &MO : MI.operands()) {
|
|
|
|
if (!MO.isReg())
|
|
|
|
continue;
|
|
|
|
if (!MO.readsReg())
|
|
|
|
continue;
|
|
|
|
if (TrackLaneMasks && !MO.isUse())
|
|
|
|
continue;
|
|
|
|
|
|
|
|
unsigned Reg = MO.getReg();
|
|
|
|
if (!TargetRegisterInfo::isVirtualRegister(Reg))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Ignore re-defs.
|
|
|
|
if (TrackLaneMasks) {
|
|
|
|
bool FoundDef = false;
|
|
|
|
for (const MachineOperand &MO2 : MI.operands()) {
|
|
|
|
if (MO2.isReg() && MO2.isDef() && MO2.getReg() == Reg && !MO2.isDead()) {
|
|
|
|
FoundDef = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (FoundDef)
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Record this local VReg use.
|
|
|
|
VReg2SUnitMultiMap::iterator UI = VRegUses.find(Reg);
|
|
|
|
for (; UI != VRegUses.end(); ++UI) {
|
|
|
|
if (UI->SU == &SU)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (UI == VRegUses.end())
|
|
|
|
VRegUses.insert(VReg2SUnit(Reg, LaneBitmask::getNone(), &SU));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-08-15 19:34:23 +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.
|
2014-11-24 09:08:18 +00:00
|
|
|
void ScheduleDAGMILive::enterRegion(MachineBasicBlock *bb,
|
2012-08-15 19:34:23 +00:00
|
|
|
MachineBasicBlock::iterator begin,
|
|
|
|
MachineBasicBlock::iterator end,
|
2013-12-22 00:04:03 +00:00
|
|
|
unsigned regioninstrs)
|
2012-08-15 19:34:23 +00:00
|
|
|
{
|
2014-11-24 09:08:18 +00:00
|
|
|
// ScheduleDAGMI initializes SchedImpl's per-region policy.
|
|
|
|
ScheduleDAGMI::enterRegion(bb, begin, end, regioninstrs);
|
2012-08-15 19:34:23 +00:00
|
|
|
|
|
|
|
// For convenience remember the end of the liveness region.
|
2014-11-24 09:08:18 +00:00
|
|
|
LiveRegionEnd = (RegionEnd == bb->end()) ? RegionEnd : std::next(RegionEnd);
|
2013-12-22 00:04:03 +00:00
|
|
|
|
|
|
|
SUPressureDiffs.clear();
|
|
|
|
|
|
|
|
ShouldTrackPressure = SchedImpl->shouldTrackPressure();
|
2016-07-23 20:41:05 +00:00
|
|
|
ShouldTrackLaneMasks = SchedImpl->shouldTrackLaneMasks();
|
|
|
|
|
|
|
|
assert((!ShouldTrackLaneMasks || ShouldTrackPressure) &&
|
|
|
|
"ShouldTrackLaneMasks requires ShouldTrackPressure");
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Setup the register pressure trackers for the top scheduled top and bottom
|
|
|
|
// scheduled regions.
|
2014-11-24 09:08:18 +00:00
|
|
|
void ScheduleDAGMILive::initRegPressure() {
|
2017-01-02 19:17:04 +00:00
|
|
|
VRegUses.clear();
|
|
|
|
VRegUses.setUniverse(MRI.getNumVirtRegs());
|
|
|
|
for (SUnit &SU : SUnits)
|
|
|
|
collectVRegUses(SU);
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin,
|
|
|
|
ShouldTrackLaneMasks, false);
|
|
|
|
BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd,
|
|
|
|
ShouldTrackLaneMasks, false);
|
2012-08-15 19:34:23 +00:00
|
|
|
|
|
|
|
// Close the RPTracker to finalize live ins.
|
|
|
|
RPTracker.closeRegion();
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
DEBUG(RPTracker.dump());
|
2012-08-15 19:34:23 +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();
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
BotRPTracker.initLiveThru(RPTracker);
|
|
|
|
if (!BotRPTracker.getLiveThru().empty()) {
|
|
|
|
TopRPTracker.initLiveThru(BotRPTracker.getLiveThru());
|
|
|
|
DEBUG(dbgs() << "Live Thru: ";
|
|
|
|
dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI));
|
|
|
|
};
|
|
|
|
|
|
|
|
// For each live out vreg reduce the pressure change associated with other
|
|
|
|
// uses of the same vreg below the live-out reaching def.
|
|
|
|
updatePressureDiffs(RPTracker.getPressure().LiveOutRegs);
|
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
// Account for liveness generated by the region boundary.
|
2013-12-22 00:04:03 +00:00
|
|
|
if (LiveRegionEnd != RegionEnd) {
|
2016-07-23 20:41:05 +00:00
|
|
|
SmallVector<RegisterMaskPair, 8> LiveUses;
|
2013-12-22 00:04:03 +00:00
|
|
|
BotRPTracker.recede(&LiveUses);
|
|
|
|
updatePressureDiffs(LiveUses);
|
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
|
2015-12-30 11:46:15 +00:00
|
|
|
DEBUG(
|
|
|
|
dbgs() << "Top Pressure:\n";
|
|
|
|
dumpRegSetPressure(TopRPTracker.getRegSetPressureAtPos(), TRI);
|
|
|
|
dbgs() << "Bottom Pressure:\n";
|
|
|
|
dumpRegSetPressure(BotRPTracker.getRegSetPressureAtPos(), TRI);
|
|
|
|
);
|
|
|
|
|
2017-12-18 20:10:56 +00:00
|
|
|
assert((BotRPTracker.getPos() == RegionEnd ||
|
|
|
|
(RegionEnd->isDebugValue() &&
|
|
|
|
BotRPTracker.getPos() == priorNonDebug(RegionEnd, RegionBegin))) &&
|
|
|
|
"Can't find the region bottom");
|
2012-08-15 19:34:23 +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();
|
2013-04-08 18:41:23 +00:00
|
|
|
const std::vector<unsigned> &RegionPressure =
|
|
|
|
RPTracker.getPressure().MaxSetPressure;
|
2012-08-15 19:34:23 +00:00
|
|
|
for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
|
2013-12-22 00:04:03 +00:00
|
|
|
unsigned Limit = RegClassInfo->getRegPressureSetLimit(i);
|
|
|
|
if (RegionPressure[i] > Limit) {
|
|
|
|
DEBUG(dbgs() << TRI->getRegPressureSetName(i)
|
|
|
|
<< " Limit " << Limit
|
|
|
|
<< " Actual " << RegionPressure[i] << "\n");
|
|
|
|
RegionCriticalPSets.push_back(PressureChange(i));
|
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
|
|
|
DEBUG(dbgs() << "Excess PSets: ";
|
2017-06-26 20:32:52 +00:00
|
|
|
for (const PressureChange &RCPS : RegionCriticalPSets)
|
2012-08-15 19:34:23 +00:00
|
|
|
dbgs() << TRI->getRegPressureSetName(
|
2017-06-26 20:32:52 +00:00
|
|
|
RCPS.getPSet()) << " ";
|
2012-08-15 19:34:23 +00:00
|
|
|
dbgs() << "\n");
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void ScheduleDAGMILive::
|
2013-12-22 00:04:03 +00:00
|
|
|
updateScheduledPressure(const SUnit *SU,
|
|
|
|
const std::vector<unsigned> &NewMaxPressure) {
|
|
|
|
const PressureDiff &PDiff = getPressureDiff(SU);
|
|
|
|
unsigned CritIdx = 0, CritEnd = RegionCriticalPSets.size();
|
2017-06-26 20:32:52 +00:00
|
|
|
for (const PressureChange &PC : PDiff) {
|
|
|
|
if (!PC.isValid())
|
2013-12-22 00:04:03 +00:00
|
|
|
break;
|
2017-06-26 20:32:52 +00:00
|
|
|
unsigned ID = PC.getPSet();
|
2013-12-22 00:04:03 +00:00
|
|
|
while (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() < ID)
|
|
|
|
++CritIdx;
|
|
|
|
if (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() == ID) {
|
|
|
|
if ((int)NewMaxPressure[ID] > RegionCriticalPSets[CritIdx].getUnitInc()
|
2017-04-16 16:01:22 +00:00
|
|
|
&& NewMaxPressure[ID] <= (unsigned)std::numeric_limits<int16_t>::max())
|
2013-12-22 00:04:03 +00:00
|
|
|
RegionCriticalPSets[CritIdx].setUnitInc(NewMaxPressure[ID]);
|
|
|
|
}
|
|
|
|
unsigned Limit = RegClassInfo->getRegPressureSetLimit(ID);
|
|
|
|
if (NewMaxPressure[ID] >= Limit - 2) {
|
|
|
|
DEBUG(dbgs() << " " << TRI->getRegPressureSetName(ID) << ": "
|
2015-05-27 18:44:32 +00:00
|
|
|
<< NewMaxPressure[ID]
|
|
|
|
<< ((NewMaxPressure[ID] > Limit) ? " > " : " <= ") << Limit
|
|
|
|
<< "(+ " << BotRPTracker.getLiveThru()[ID] << " livethru)\n");
|
2013-12-22 00:04:03 +00:00
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Update the PressureDiff array for liveness after scheduling this
|
|
|
|
/// instruction.
|
2016-07-23 20:41:05 +00:00
|
|
|
void ScheduleDAGMILive::updatePressureDiffs(
|
|
|
|
ArrayRef<RegisterMaskPair> LiveUses) {
|
|
|
|
for (const RegisterMaskPair &P : LiveUses) {
|
|
|
|
unsigned Reg = P.RegUnit;
|
2013-12-22 00:04:03 +00:00
|
|
|
/// FIXME: Currently assuming single-use physregs.
|
|
|
|
if (!TRI->isVirtualRegister(Reg))
|
|
|
|
continue;
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
if (ShouldTrackLaneMasks) {
|
|
|
|
// If the register has just become live then other uses won't change
|
|
|
|
// this fact anymore => decrement pressure.
|
|
|
|
// If the register has just become dead then other uses make it come
|
|
|
|
// back to life => increment pressure.
|
2017-01-02 19:17:04 +00:00
|
|
|
bool Decrement = P.LaneMask.any();
|
2016-07-23 20:41:05 +00:00
|
|
|
|
|
|
|
for (const VReg2SUnit &V2SU
|
|
|
|
: make_range(VRegUses.find(Reg), VRegUses.end())) {
|
|
|
|
SUnit &SU = *V2SU.SU;
|
|
|
|
if (SU.isScheduled || &SU == &ExitSU)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
PressureDiff &PDiff = getPressureDiff(&SU);
|
|
|
|
PDiff.addPressureChange(Reg, Decrement, &MRI);
|
|
|
|
DEBUG(
|
|
|
|
dbgs() << " UpdateRegP: SU(" << SU.NodeNum << ") "
|
2017-12-18 20:10:56 +00:00
|
|
|
<< printReg(Reg, TRI) << ':' << PrintLaneMask(P.LaneMask)
|
2016-07-23 20:41:05 +00:00
|
|
|
<< ' ' << *SU.getInstr();
|
|
|
|
dbgs() << " to ";
|
|
|
|
PDiff.dump(*TRI);
|
|
|
|
);
|
|
|
|
}
|
|
|
|
} else {
|
2017-01-02 19:17:04 +00:00
|
|
|
assert(P.LaneMask.any());
|
2017-12-18 20:10:56 +00:00
|
|
|
DEBUG(dbgs() << " LiveReg: " << printVRegOrUnit(Reg, TRI) << "\n");
|
2016-07-23 20:41:05 +00:00
|
|
|
// This may be called before CurrentBottom has been initialized. However,
|
|
|
|
// BotRPTracker must have a valid position. We want the value live into the
|
|
|
|
// instruction or live out of the block, so ask for the previous
|
|
|
|
// instruction's live-out.
|
|
|
|
const LiveInterval &LI = LIS->getInterval(Reg);
|
|
|
|
VNInfo *VNI;
|
|
|
|
MachineBasicBlock::const_iterator I =
|
|
|
|
nextIfDebug(BotRPTracker.getPos(), BB->end());
|
|
|
|
if (I == BB->end())
|
|
|
|
VNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
|
|
|
|
else {
|
|
|
|
LiveQueryResult LRQ = LI.Query(LIS->getInstructionIndex(*I));
|
|
|
|
VNI = LRQ.valueIn();
|
|
|
|
}
|
|
|
|
// RegisterPressureTracker guarantees that readsReg is true for LiveUses.
|
|
|
|
assert(VNI && "No live value at use.");
|
|
|
|
for (const VReg2SUnit &V2SU
|
|
|
|
: make_range(VRegUses.find(Reg), VRegUses.end())) {
|
|
|
|
SUnit *SU = V2SU.SU;
|
|
|
|
// If this use comes before the reaching def, it cannot be a last use,
|
|
|
|
// so decrease its pressure change.
|
|
|
|
if (!SU->isScheduled && SU != &ExitSU) {
|
|
|
|
LiveQueryResult LRQ =
|
|
|
|
LI.Query(LIS->getInstructionIndex(*SU->getInstr()));
|
|
|
|
if (LRQ.valueIn() == VNI) {
|
|
|
|
PressureDiff &PDiff = getPressureDiff(SU);
|
|
|
|
PDiff.addPressureChange(Reg, true, &MRI);
|
|
|
|
DEBUG(
|
|
|
|
dbgs() << " UpdateRegP: SU(" << SU->NodeNum << ") "
|
|
|
|
<< *SU->getInstr();
|
|
|
|
dbgs() << " to ";
|
|
|
|
PDiff.dump(*TRI);
|
|
|
|
);
|
|
|
|
}
|
2015-12-30 11:46:15 +00:00
|
|
|
}
|
2013-06-10 20:36:52 +00:00
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
}
|
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
/// schedule - Called back from MachineScheduler::runOnMachineFunction
|
|
|
|
/// after setting up the current scheduling region. [RegionBegin, RegionEnd)
|
|
|
|
/// only includes instructions that have DAG nodes, not scheduling boundaries.
|
|
|
|
///
|
|
|
|
/// This is a skeletal driver, with all the functionality pushed into helpers,
|
2015-12-30 11:46:15 +00:00
|
|
|
/// so that it can be easily extended by experimental schedulers. Generally,
|
2012-12-02 13:10:19 +00:00
|
|
|
/// implementing MachineSchedStrategy should be sufficient to implement a new
|
|
|
|
/// scheduling algorithm. However, if a scheduler further subclasses
|
2014-11-24 09:08:18 +00:00
|
|
|
/// ScheduleDAGMILive then it will want to override this virtual method in order
|
|
|
|
/// to update any specialized state.
|
|
|
|
void ScheduleDAGMILive::schedule() {
|
2015-12-30 11:46:15 +00:00
|
|
|
DEBUG(dbgs() << "ScheduleDAGMILive::schedule starting\n");
|
|
|
|
DEBUG(SchedImpl->dumpPolicy());
|
2012-12-02 13:10:19 +00:00
|
|
|
buildDAGWithRegPressure();
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
Topo.InitDAGTopologicalSorting();
|
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
postprocessDAG();
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
SmallVector<SUnit*, 8> TopRoots, BotRoots;
|
|
|
|
findRootsAndBiasEdges(TopRoots, BotRoots);
|
|
|
|
|
|
|
|
// Initialize the strategy before modifying the DAG.
|
|
|
|
// This may initialize a DFSResult to be used for queue priority.
|
|
|
|
SchedImpl->initialize(this);
|
|
|
|
|
2015-12-30 11:46:15 +00:00
|
|
|
DEBUG(
|
2017-01-02 19:17:04 +00:00
|
|
|
if (EntrySU.getInstr() != nullptr)
|
|
|
|
EntrySU.dumpAll(this);
|
2015-12-30 11:46:15 +00:00
|
|
|
for (const SUnit &SU : SUnits) {
|
|
|
|
SU.dumpAll(this);
|
|
|
|
if (ShouldTrackPressure) {
|
|
|
|
dbgs() << " Pressure Diff : ";
|
|
|
|
getPressureDiff(&SU).dump(*TRI);
|
|
|
|
}
|
2017-04-16 16:01:22 +00:00
|
|
|
dbgs() << " Single Issue : ";
|
|
|
|
if (SchedModel.mustBeginGroup(SU.getInstr()) &&
|
|
|
|
SchedModel.mustEndGroup(SU.getInstr()))
|
|
|
|
dbgs() << "true;";
|
|
|
|
else
|
|
|
|
dbgs() << "false;";
|
2015-12-30 11:46:15 +00:00
|
|
|
dbgs() << '\n';
|
|
|
|
}
|
2017-01-02 19:17:04 +00:00
|
|
|
if (ExitSU.getInstr() != nullptr)
|
|
|
|
ExitSU.dumpAll(this);
|
2015-12-30 11:46:15 +00:00
|
|
|
);
|
2012-12-02 13:10:19 +00:00
|
|
|
if (ViewMISchedDAGs) viewGraph();
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
// Initialize ready queues now that the DAG and priority data are finalized.
|
|
|
|
initQueues(TopRoots, BotRoots);
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
bool IsTopNode = false;
|
2015-12-30 11:46:15 +00:00
|
|
|
while (true) {
|
|
|
|
DEBUG(dbgs() << "** ScheduleDAGMILive::schedule picking next node\n");
|
|
|
|
SUnit *SU = SchedImpl->pickNode(IsTopNode);
|
|
|
|
if (!SU) break;
|
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
assert(!SU->isScheduled && "Node already scheduled");
|
|
|
|
if (!checkSchedLimit())
|
|
|
|
break;
|
|
|
|
|
|
|
|
scheduleMI(SU, IsTopNode);
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
if (DFSResult) {
|
|
|
|
unsigned SubtreeID = DFSResult->getSubtreeID(SU);
|
|
|
|
if (!ScheduledTrees.test(SubtreeID)) {
|
|
|
|
ScheduledTrees.set(SubtreeID);
|
|
|
|
DFSResult->scheduleTree(SubtreeID);
|
|
|
|
SchedImpl->scheduleTree(SubtreeID);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Notify the scheduling strategy after updating the DAG.
|
|
|
|
SchedImpl->schedNode(SU, IsTopNode);
|
2015-05-27 18:44:32 +00:00
|
|
|
|
|
|
|
updateQueues(SU, IsTopNode);
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
|
|
|
assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
|
|
|
|
|
|
|
|
placeDebugValues();
|
|
|
|
|
|
|
|
DEBUG({
|
2017-12-18 20:10:56 +00:00
|
|
|
dbgs() << "*** Final schedule for "
|
|
|
|
<< printMBBReference(*begin()->getParent()) << " ***\n";
|
|
|
|
dumpSchedule();
|
|
|
|
dbgs() << '\n';
|
|
|
|
});
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Build the DAG and setup three register pressure trackers.
|
2014-11-24 09:08:18 +00:00
|
|
|
void ScheduleDAGMILive::buildDAGWithRegPressure() {
|
2013-12-22 00:04:03 +00:00
|
|
|
if (!ShouldTrackPressure) {
|
|
|
|
RPTracker.reset();
|
|
|
|
RegionCriticalPSets.clear();
|
|
|
|
buildSchedGraph(AA);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
// Initialize the register pressure tracker used by buildSchedGraph.
|
2013-12-22 00:04:03 +00:00
|
|
|
RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd,
|
2016-07-23 20:41:05 +00:00
|
|
|
ShouldTrackLaneMasks, /*TrackUntiedDefs=*/true);
|
2012-12-02 13:10:19 +00:00
|
|
|
|
|
|
|
// Account for liveness generate by the region boundary.
|
|
|
|
if (LiveRegionEnd != RegionEnd)
|
|
|
|
RPTracker.recede();
|
|
|
|
|
|
|
|
// Build the DAG, and compute current register pressure.
|
2016-07-23 20:41:05 +00:00
|
|
|
buildSchedGraph(AA, &RPTracker, &SUPressureDiffs, LIS, ShouldTrackLaneMasks);
|
2012-12-02 13:10:19 +00:00
|
|
|
|
|
|
|
// Initialize top/bottom trackers after computing region pressure.
|
|
|
|
initRegPressure();
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void ScheduleDAGMILive::computeDFSResult() {
|
2013-04-08 18:41:23 +00:00
|
|
|
if (!DFSResult)
|
|
|
|
DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
|
|
|
|
DFSResult->clear();
|
|
|
|
ScheduledTrees.clear();
|
|
|
|
DFSResult->resize(SUnits.size());
|
|
|
|
DFSResult->compute(SUnits);
|
|
|
|
ScheduledTrees.resize(DFSResult->getNumSubtrees());
|
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
/// Compute the max cyclic critical path through the DAG. The scheduling DAG
|
|
|
|
/// only provides the critical path for single block loops. To handle loops that
|
|
|
|
/// span blocks, we could use the vreg path latencies provided by
|
|
|
|
/// MachineTraceMetrics instead. However, MachineTraceMetrics is not currently
|
|
|
|
/// available for use in the scheduler.
|
|
|
|
///
|
|
|
|
/// The cyclic path estimation identifies a def-use pair that crosses the back
|
|
|
|
/// edge and considers the depth and height of the nodes. For example, consider
|
|
|
|
/// the following instruction sequence where each instruction has unit latency
|
|
|
|
/// and defines an epomymous virtual register:
|
|
|
|
///
|
|
|
|
/// a->b(a,c)->c(b)->d(c)->exit
|
|
|
|
///
|
|
|
|
/// The cyclic critical path is a two cycles: b->c->b
|
|
|
|
/// The acyclic critical path is four cycles: a->b->c->d->exit
|
|
|
|
/// LiveOutHeight = height(c) = len(c->d->exit) = 2
|
|
|
|
/// LiveOutDepth = depth(c) + 1 = len(a->b->c) + 1 = 3
|
|
|
|
/// LiveInHeight = height(b) + 1 = len(b->c->d->exit) + 1 = 4
|
|
|
|
/// LiveInDepth = depth(b) = len(a->b) = 1
|
|
|
|
///
|
|
|
|
/// LiveOutDepth - LiveInDepth = 3 - 1 = 2
|
|
|
|
/// LiveInHeight - LiveOutHeight = 4 - 2 = 2
|
|
|
|
/// CyclicCriticalPath = min(2, 2) = 2
|
2014-11-24 09:08:18 +00:00
|
|
|
///
|
|
|
|
/// This could be relevant to PostRA scheduling, but is currently implemented
|
|
|
|
/// assuming LiveIntervals.
|
|
|
|
unsigned ScheduleDAGMILive::computeCyclicCriticalPath() {
|
2013-12-22 00:04:03 +00:00
|
|
|
// This only applies to single block loop.
|
|
|
|
if (!BB->isSuccessor(BB))
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
unsigned MaxCyclicLatency = 0;
|
|
|
|
// Visit each live out vreg def to find def/use pairs that cross iterations.
|
2016-07-23 20:41:05 +00:00
|
|
|
for (const RegisterMaskPair &P : RPTracker.getPressure().LiveOutRegs) {
|
|
|
|
unsigned Reg = P.RegUnit;
|
2013-12-22 00:04:03 +00:00
|
|
|
if (!TRI->isVirtualRegister(Reg))
|
|
|
|
continue;
|
|
|
|
const LiveInterval &LI = LIS->getInterval(Reg);
|
|
|
|
const VNInfo *DefVNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
|
|
|
|
if (!DefVNI)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
MachineInstr *DefMI = LIS->getInstructionFromIndex(DefVNI->def);
|
|
|
|
const SUnit *DefSU = getSUnit(DefMI);
|
|
|
|
if (!DefSU)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
unsigned LiveOutHeight = DefSU->getHeight();
|
|
|
|
unsigned LiveOutDepth = DefSU->getDepth() + DefSU->Latency;
|
|
|
|
// Visit all local users of the vreg def.
|
2015-12-30 11:46:15 +00:00
|
|
|
for (const VReg2SUnit &V2SU
|
|
|
|
: make_range(VRegUses.find(Reg), VRegUses.end())) {
|
|
|
|
SUnit *SU = V2SU.SU;
|
|
|
|
if (SU == &ExitSU)
|
2013-12-22 00:04:03 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
// Only consider uses of the phi.
|
2016-07-23 20:41:05 +00:00
|
|
|
LiveQueryResult LRQ = LI.Query(LIS->getInstructionIndex(*SU->getInstr()));
|
2013-12-22 00:04:03 +00:00
|
|
|
if (!LRQ.valueIn()->isPHIDef())
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Assume that a path spanning two iterations is a cycle, which could
|
|
|
|
// overestimate in strange cases. This allows cyclic latency to be
|
|
|
|
// estimated as the minimum slack of the vreg's depth or height.
|
|
|
|
unsigned CyclicLatency = 0;
|
2015-12-30 11:46:15 +00:00
|
|
|
if (LiveOutDepth > SU->getDepth())
|
|
|
|
CyclicLatency = LiveOutDepth - SU->getDepth();
|
2013-12-22 00:04:03 +00:00
|
|
|
|
2015-12-30 11:46:15 +00:00
|
|
|
unsigned LiveInHeight = SU->getHeight() + DefSU->Latency;
|
2013-12-22 00:04:03 +00:00
|
|
|
if (LiveInHeight > LiveOutHeight) {
|
|
|
|
if (LiveInHeight - LiveOutHeight < CyclicLatency)
|
|
|
|
CyclicLatency = LiveInHeight - LiveOutHeight;
|
2016-07-23 20:41:05 +00:00
|
|
|
} else
|
2013-12-22 00:04:03 +00:00
|
|
|
CyclicLatency = 0;
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "Cyclic Path: SU(" << DefSU->NodeNum << ") -> SU("
|
2015-12-30 11:46:15 +00:00
|
|
|
<< SU->NodeNum << ") = " << CyclicLatency << "c\n");
|
2013-12-22 00:04:03 +00:00
|
|
|
if (CyclicLatency > MaxCyclicLatency)
|
|
|
|
MaxCyclicLatency = CyclicLatency;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "c\n");
|
|
|
|
return MaxCyclicLatency;
|
|
|
|
}
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
/// Release ExitSU predecessors and setup scheduler queues. Re-position
|
|
|
|
/// the Top RP tracker in case the region beginning has changed.
|
|
|
|
void ScheduleDAGMILive::initQueues(ArrayRef<SUnit*> TopRoots,
|
|
|
|
ArrayRef<SUnit*> BotRoots) {
|
|
|
|
ScheduleDAGMI::initQueues(TopRoots, BotRoots);
|
|
|
|
if (ShouldTrackPressure) {
|
|
|
|
assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
|
|
|
|
TopRPTracker.setPos(CurrentTop);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
/// Move an instruction and update register pressure.
|
2014-11-24 09:08:18 +00:00
|
|
|
void ScheduleDAGMILive::scheduleMI(SUnit *SU, bool IsTopNode) {
|
2012-12-02 13:10:19 +00:00
|
|
|
// Move the instruction to its new location in the instruction stream.
|
|
|
|
MachineInstr *MI = SU->getInstr();
|
2012-08-15 19:34:23 +00:00
|
|
|
|
2012-12-02 13:10:19 +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);
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
if (ShouldTrackPressure) {
|
|
|
|
// Update top scheduled pressure.
|
2016-07-23 20:41:05 +00:00
|
|
|
RegisterOperands RegOpers;
|
|
|
|
RegOpers.collect(*MI, *TRI, MRI, ShouldTrackLaneMasks, false);
|
|
|
|
if (ShouldTrackLaneMasks) {
|
|
|
|
// Adjust liveness and add missing dead+read-undef flags.
|
|
|
|
SlotIndex SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot();
|
|
|
|
RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI);
|
|
|
|
} else {
|
|
|
|
// Adjust for missing dead-def flags.
|
|
|
|
RegOpers.detectDeadDefs(*MI, *LIS);
|
|
|
|
}
|
|
|
|
|
|
|
|
TopRPTracker.advance(RegOpers);
|
2013-12-22 00:04:03 +00:00
|
|
|
assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
|
2015-12-30 11:46:15 +00:00
|
|
|
DEBUG(
|
|
|
|
dbgs() << "Top Pressure:\n";
|
|
|
|
dumpRegSetPressure(TopRPTracker.getRegSetPressureAtPos(), TRI);
|
|
|
|
);
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
updateScheduledPressure(SU, TopRPTracker.getPressure().MaxSetPressure);
|
|
|
|
}
|
2016-07-23 20:41:05 +00:00
|
|
|
} else {
|
2012-12-02 13:10:19 +00:00
|
|
|
assert(SU->isBottomReady() && "node still has unscheduled dependencies");
|
|
|
|
MachineBasicBlock::iterator priorII =
|
|
|
|
priorNonDebug(CurrentBottom, CurrentTop);
|
|
|
|
if (&*priorII == MI)
|
|
|
|
CurrentBottom = priorII;
|
2012-04-14 13:54:10 +00:00
|
|
|
else {
|
2012-12-02 13:10:19 +00:00
|
|
|
if (&*CurrentTop == MI) {
|
|
|
|
CurrentTop = nextIfDebug(++CurrentTop, priorII);
|
|
|
|
TopRPTracker.setPos(CurrentTop);
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2012-12-02 13:10:19 +00:00
|
|
|
moveInstruction(MI, CurrentBottom);
|
|
|
|
CurrentBottom = MI;
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
if (ShouldTrackPressure) {
|
2016-07-23 20:41:05 +00:00
|
|
|
RegisterOperands RegOpers;
|
|
|
|
RegOpers.collect(*MI, *TRI, MRI, ShouldTrackLaneMasks, false);
|
|
|
|
if (ShouldTrackLaneMasks) {
|
|
|
|
// Adjust liveness and add missing dead+read-undef flags.
|
|
|
|
SlotIndex SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot();
|
|
|
|
RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI);
|
|
|
|
} else {
|
|
|
|
// Adjust for missing dead-def flags.
|
|
|
|
RegOpers.detectDeadDefs(*MI, *LIS);
|
|
|
|
}
|
|
|
|
|
2017-12-18 20:10:56 +00:00
|
|
|
if (BotRPTracker.getPos() != CurrentBottom)
|
|
|
|
BotRPTracker.recedeSkipDebugValues();
|
2016-07-23 20:41:05 +00:00
|
|
|
SmallVector<RegisterMaskPair, 8> LiveUses;
|
|
|
|
BotRPTracker.recede(RegOpers, &LiveUses);
|
2013-12-22 00:04:03 +00:00
|
|
|
assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
|
2015-12-30 11:46:15 +00:00
|
|
|
DEBUG(
|
|
|
|
dbgs() << "Bottom Pressure:\n";
|
|
|
|
dumpRegSetPressure(BotRPTracker.getRegSetPressureAtPos(), TRI);
|
|
|
|
);
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
updateScheduledPressure(SU, BotRPTracker.getPressure().MaxSetPressure);
|
|
|
|
updatePressureDiffs(LiveUses);
|
|
|
|
}
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
2016-07-23 20:41:05 +00:00
|
|
|
// BaseMemOpClusterMutation - DAG post-processing to cluster loads or stores.
|
2013-04-08 18:41:23 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
/// \brief Post-process the DAG to create cluster edges between neighboring
|
2016-07-23 20:41:05 +00:00
|
|
|
/// loads or between neighboring stores.
|
|
|
|
class BaseMemOpClusterMutation : public ScheduleDAGMutation {
|
|
|
|
struct MemOpInfo {
|
2013-04-08 18:41:23 +00:00
|
|
|
SUnit *SU;
|
|
|
|
unsigned BaseReg;
|
2016-07-23 20:41:05 +00:00
|
|
|
int64_t Offset;
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
MemOpInfo(SUnit *su, unsigned reg, int64_t ofs)
|
|
|
|
: SU(su), BaseReg(reg), Offset(ofs) {}
|
2014-11-24 09:08:18 +00:00
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
bool operator<(const MemOpInfo&RHS) const {
|
2017-01-02 19:17:04 +00:00
|
|
|
return std::tie(BaseReg, Offset, SU->NodeNum) <
|
|
|
|
std::tie(RHS.BaseReg, RHS.Offset, RHS.SU->NodeNum);
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
2013-04-08 18:41:23 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
const TargetInstrInfo *TII;
|
|
|
|
const TargetRegisterInfo *TRI;
|
2016-07-23 20:41:05 +00:00
|
|
|
bool IsLoad;
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
public:
|
2016-07-23 20:41:05 +00:00
|
|
|
BaseMemOpClusterMutation(const TargetInstrInfo *tii,
|
|
|
|
const TargetRegisterInfo *tri, bool IsLoad)
|
|
|
|
: TII(tii), TRI(tri), IsLoad(IsLoad) {}
|
|
|
|
|
|
|
|
void apply(ScheduleDAGInstrs *DAGInstrs) override;
|
2013-04-08 18:41:23 +00:00
|
|
|
|
|
|
|
protected:
|
2016-07-23 20:41:05 +00:00
|
|
|
void clusterNeighboringMemOps(ArrayRef<SUnit *> MemOps, ScheduleDAGMI *DAG);
|
|
|
|
};
|
|
|
|
|
|
|
|
class StoreClusterMutation : public BaseMemOpClusterMutation {
|
|
|
|
public:
|
|
|
|
StoreClusterMutation(const TargetInstrInfo *tii,
|
|
|
|
const TargetRegisterInfo *tri)
|
|
|
|
: BaseMemOpClusterMutation(tii, tri, false) {}
|
|
|
|
};
|
|
|
|
|
|
|
|
class LoadClusterMutation : public BaseMemOpClusterMutation {
|
|
|
|
public:
|
|
|
|
LoadClusterMutation(const TargetInstrInfo *tii, const TargetRegisterInfo *tri)
|
|
|
|
: BaseMemOpClusterMutation(tii, tri, true) {}
|
2013-04-08 18:41:23 +00:00
|
|
|
};
|
2017-04-16 16:01:22 +00:00
|
|
|
|
|
|
|
} // end anonymous namespace
|
2013-04-08 18:41:23 +00:00
|
|
|
|
2017-01-02 19:17:04 +00:00
|
|
|
namespace llvm {
|
|
|
|
|
|
|
|
std::unique_ptr<ScheduleDAGMutation>
|
|
|
|
createLoadClusterDAGMutation(const TargetInstrInfo *TII,
|
|
|
|
const TargetRegisterInfo *TRI) {
|
2017-04-16 16:01:22 +00:00
|
|
|
return EnableMemOpCluster ? llvm::make_unique<LoadClusterMutation>(TII, TRI)
|
2017-01-02 19:17:04 +00:00
|
|
|
: nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::unique_ptr<ScheduleDAGMutation>
|
|
|
|
createStoreClusterDAGMutation(const TargetInstrInfo *TII,
|
|
|
|
const TargetRegisterInfo *TRI) {
|
2017-04-16 16:01:22 +00:00
|
|
|
return EnableMemOpCluster ? llvm::make_unique<StoreClusterMutation>(TII, TRI)
|
2017-01-02 19:17:04 +00:00
|
|
|
: nullptr;
|
|
|
|
}
|
|
|
|
|
2017-04-16 16:01:22 +00:00
|
|
|
} // end namespace llvm
|
2017-01-02 19:17:04 +00:00
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
void BaseMemOpClusterMutation::clusterNeighboringMemOps(
|
|
|
|
ArrayRef<SUnit *> MemOps, ScheduleDAGMI *DAG) {
|
|
|
|
SmallVector<MemOpInfo, 32> MemOpRecords;
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SUnit *SU : MemOps) {
|
2013-04-08 18:41:23 +00:00
|
|
|
unsigned BaseReg;
|
2016-07-23 20:41:05 +00:00
|
|
|
int64_t Offset;
|
|
|
|
if (TII->getMemOpBaseRegImmOfs(*SU->getInstr(), BaseReg, Offset, TRI))
|
|
|
|
MemOpRecords.push_back(MemOpInfo(SU, BaseReg, Offset));
|
2013-04-08 18:41:23 +00:00
|
|
|
}
|
2016-07-23 20:41:05 +00:00
|
|
|
if (MemOpRecords.size() < 2)
|
2013-04-08 18:41:23 +00:00
|
|
|
return;
|
2016-07-23 20:41:05 +00:00
|
|
|
|
|
|
|
std::sort(MemOpRecords.begin(), MemOpRecords.end());
|
2013-04-08 18:41:23 +00:00
|
|
|
unsigned ClusterLength = 1;
|
2016-07-23 20:41:05 +00:00
|
|
|
for (unsigned Idx = 0, End = MemOpRecords.size(); Idx < (End - 1); ++Idx) {
|
|
|
|
SUnit *SUa = MemOpRecords[Idx].SU;
|
|
|
|
SUnit *SUb = MemOpRecords[Idx+1].SU;
|
2017-12-18 20:10:56 +00:00
|
|
|
if (TII->shouldClusterMemOps(*SUa->getInstr(), MemOpRecords[Idx].BaseReg,
|
|
|
|
*SUb->getInstr(), MemOpRecords[Idx+1].BaseReg,
|
2016-07-23 20:41:05 +00:00
|
|
|
ClusterLength) &&
|
|
|
|
DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
|
|
|
|
DEBUG(dbgs() << "Cluster ld/st SU(" << SUa->NodeNum << ") - SU("
|
2013-04-08 18:41:23 +00:00
|
|
|
<< 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.
|
2017-06-26 20:32:52 +00:00
|
|
|
for (const SDep &Succ : SUa->Succs) {
|
|
|
|
if (Succ.getSUnit() == SUb)
|
2013-04-08 18:41:23 +00:00
|
|
|
continue;
|
2017-06-26 20:32:52 +00:00
|
|
|
DEBUG(dbgs() << " Copy Succ SU(" << Succ.getSUnit()->NodeNum << ")\n");
|
|
|
|
DAG->addEdge(Succ.getSUnit(), SDep(SUb, SDep::Artificial));
|
2013-04-08 18:41:23 +00:00
|
|
|
}
|
|
|
|
++ClusterLength;
|
2016-07-23 20:41:05 +00:00
|
|
|
} else
|
2013-04-08 18:41:23 +00:00
|
|
|
ClusterLength = 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// \brief Callback from DAG postProcessing to create cluster edges for loads.
|
2016-07-23 20:41:05 +00:00
|
|
|
void BaseMemOpClusterMutation::apply(ScheduleDAGInstrs *DAGInstrs) {
|
|
|
|
ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs);
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
// Map DAG NodeNum to store chain ID.
|
|
|
|
DenseMap<unsigned, unsigned> StoreChainIDs;
|
2016-07-23 20:41:05 +00:00
|
|
|
// Map each store chain to a set of dependent MemOps.
|
2013-04-08 18:41:23 +00:00
|
|
|
SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SUnit &SU : DAG->SUnits) {
|
|
|
|
if ((IsLoad && !SU.getInstr()->mayLoad()) ||
|
|
|
|
(!IsLoad && !SU.getInstr()->mayStore()))
|
2013-04-08 18:41:23 +00:00
|
|
|
continue;
|
2016-07-23 20:41:05 +00:00
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
unsigned ChainPredID = DAG->SUnits.size();
|
2017-06-26 20:32:52 +00:00
|
|
|
for (const SDep &Pred : SU.Preds) {
|
|
|
|
if (Pred.isCtrl()) {
|
|
|
|
ChainPredID = Pred.getSUnit()->NodeNum;
|
2013-04-08 18:41:23 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Check if this chain-like pred has been seen
|
2016-07-23 20:41:05 +00:00
|
|
|
// before. ChainPredID==MaxNodeID at the top of the schedule.
|
2013-04-08 18:41:23 +00:00
|
|
|
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);
|
2017-06-26 20:32:52 +00:00
|
|
|
StoreChainDependents[Result.first->second].push_back(&SU);
|
2013-04-08 18:41:23 +00:00
|
|
|
}
|
2016-07-23 20:41:05 +00:00
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
// Iterate over the store chains.
|
2017-06-26 20:32:52 +00:00
|
|
|
for (auto &SCD : StoreChainDependents)
|
|
|
|
clusterNeighboringMemOps(SCD, DAG);
|
2013-04-08 18:41:23 +00:00
|
|
|
}
|
|
|
|
|
2013-06-10 20:36:52 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// CopyConstrain - DAG post-processing to encourage copy elimination.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2013-06-10 20:36:52 +00:00
|
|
|
/// \brief Post-process the DAG to create weak edges from all uses of a copy to
|
|
|
|
/// the one use that defines the copy's source vreg, most likely an induction
|
|
|
|
/// variable increment.
|
|
|
|
class CopyConstrain : public ScheduleDAGMutation {
|
|
|
|
// Transient state.
|
|
|
|
SlotIndex RegionBeginIdx;
|
2017-12-18 20:10:56 +00:00
|
|
|
|
2013-06-10 20:36:52 +00:00
|
|
|
// RegionEndIdx is the slot index of the last non-debug instruction in the
|
|
|
|
// scheduling region. So we may have RegionBeginIdx == RegionEndIdx.
|
|
|
|
SlotIndex RegionEndIdx;
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2013-06-10 20:36:52 +00:00
|
|
|
public:
|
|
|
|
CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {}
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
void apply(ScheduleDAGInstrs *DAGInstrs) override;
|
2013-06-10 20:36:52 +00:00
|
|
|
|
|
|
|
protected:
|
2014-11-24 09:08:18 +00:00
|
|
|
void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG);
|
2013-06-10 20:36:52 +00:00
|
|
|
};
|
2017-04-16 16:01:22 +00:00
|
|
|
|
|
|
|
} // end anonymous namespace
|
2013-06-10 20:36:52 +00:00
|
|
|
|
2017-01-02 19:17:04 +00:00
|
|
|
namespace llvm {
|
|
|
|
|
|
|
|
std::unique_ptr<ScheduleDAGMutation>
|
|
|
|
createCopyConstrainDAGMutation(const TargetInstrInfo *TII,
|
2017-04-16 16:01:22 +00:00
|
|
|
const TargetRegisterInfo *TRI) {
|
|
|
|
return llvm::make_unique<CopyConstrain>(TII, TRI);
|
2017-01-02 19:17:04 +00:00
|
|
|
}
|
|
|
|
|
2017-04-16 16:01:22 +00:00
|
|
|
} // end namespace llvm
|
2017-01-02 19:17:04 +00:00
|
|
|
|
2013-06-10 20:36:52 +00:00
|
|
|
/// constrainLocalCopy handles two possibilities:
|
|
|
|
/// 1) Local src:
|
|
|
|
/// I0: = dst
|
|
|
|
/// I1: src = ...
|
|
|
|
/// I2: = dst
|
|
|
|
/// I3: dst = src (copy)
|
|
|
|
/// (create pred->succ edges I0->I1, I2->I1)
|
|
|
|
///
|
|
|
|
/// 2) Local copy:
|
|
|
|
/// I0: dst = src (copy)
|
|
|
|
/// I1: = dst
|
|
|
|
/// I2: src = ...
|
|
|
|
/// I3: = dst
|
|
|
|
/// (create pred->succ edges I1->I2, I3->I2)
|
|
|
|
///
|
|
|
|
/// Although the MachineScheduler is currently constrained to single blocks,
|
|
|
|
/// this algorithm should handle extended blocks. An EBB is a set of
|
|
|
|
/// contiguously numbered blocks such that the previous block in the EBB is
|
|
|
|
/// always the single predecessor.
|
2014-11-24 09:08:18 +00:00
|
|
|
void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG) {
|
2013-06-10 20:36:52 +00:00
|
|
|
LiveIntervals *LIS = DAG->getLIS();
|
|
|
|
MachineInstr *Copy = CopySU->getInstr();
|
|
|
|
|
|
|
|
// Check for pure vreg copies.
|
2016-07-23 20:41:05 +00:00
|
|
|
const MachineOperand &SrcOp = Copy->getOperand(1);
|
|
|
|
unsigned SrcReg = SrcOp.getReg();
|
|
|
|
if (!TargetRegisterInfo::isVirtualRegister(SrcReg) || !SrcOp.readsReg())
|
2013-06-10 20:36:52 +00:00
|
|
|
return;
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
const MachineOperand &DstOp = Copy->getOperand(0);
|
|
|
|
unsigned DstReg = DstOp.getReg();
|
|
|
|
if (!TargetRegisterInfo::isVirtualRegister(DstReg) || DstOp.isDead())
|
2013-06-10 20:36:52 +00:00
|
|
|
return;
|
|
|
|
|
|
|
|
// Check if either the dest or source is local. If it's live across a back
|
|
|
|
// edge, it's not local. Note that if both vregs are live across the back
|
|
|
|
// edge, we cannot successfully contrain the copy without cyclic scheduling.
|
2015-05-27 18:44:32 +00:00
|
|
|
// If both the copy's source and dest are local live intervals, then we
|
|
|
|
// should treat the dest as the global for the purpose of adding
|
|
|
|
// constraints. This adds edges from source's other uses to the copy.
|
|
|
|
unsigned LocalReg = SrcReg;
|
|
|
|
unsigned GlobalReg = DstReg;
|
2013-06-10 20:36:52 +00:00
|
|
|
LiveInterval *LocalLI = &LIS->getInterval(LocalReg);
|
|
|
|
if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) {
|
2015-05-27 18:44:32 +00:00
|
|
|
LocalReg = DstReg;
|
|
|
|
GlobalReg = SrcReg;
|
2013-06-10 20:36:52 +00:00
|
|
|
LocalLI = &LIS->getInterval(LocalReg);
|
|
|
|
if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx))
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
LiveInterval *GlobalLI = &LIS->getInterval(GlobalReg);
|
|
|
|
|
|
|
|
// Find the global segment after the start of the local LI.
|
|
|
|
LiveInterval::iterator GlobalSegment = GlobalLI->find(LocalLI->beginIndex());
|
|
|
|
// If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a
|
|
|
|
// local live range. We could create edges from other global uses to the local
|
|
|
|
// start, but the coalescer should have already eliminated these cases, so
|
|
|
|
// don't bother dealing with it.
|
|
|
|
if (GlobalSegment == GlobalLI->end())
|
|
|
|
return;
|
|
|
|
|
|
|
|
// If GlobalSegment is killed at the LocalLI->start, the call to find()
|
|
|
|
// returned the next global segment. But if GlobalSegment overlaps with
|
|
|
|
// LocalLI->start, then advance to the next segement. If a hole in GlobalLI
|
|
|
|
// exists in LocalLI's vicinity, GlobalSegment will be the end of the hole.
|
|
|
|
if (GlobalSegment->contains(LocalLI->beginIndex()))
|
|
|
|
++GlobalSegment;
|
|
|
|
|
|
|
|
if (GlobalSegment == GlobalLI->end())
|
|
|
|
return;
|
|
|
|
|
|
|
|
// Check if GlobalLI contains a hole in the vicinity of LocalLI.
|
|
|
|
if (GlobalSegment != GlobalLI->begin()) {
|
|
|
|
// Two address defs have no hole.
|
2014-11-24 09:08:18 +00:00
|
|
|
if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->end,
|
2013-06-10 20:36:52 +00:00
|
|
|
GlobalSegment->start)) {
|
|
|
|
return;
|
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
// If the prior global segment may be defined by the same two-address
|
|
|
|
// instruction that also defines LocalLI, then can't make a hole here.
|
2014-11-24 09:08:18 +00:00
|
|
|
if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->start,
|
2013-12-22 00:04:03 +00:00
|
|
|
LocalLI->beginIndex())) {
|
|
|
|
return;
|
|
|
|
}
|
2013-06-10 20:36:52 +00:00
|
|
|
// If GlobalLI has a prior segment, it must be live into the EBB. Otherwise
|
|
|
|
// it would be a disconnected component in the live range.
|
2014-11-24 09:08:18 +00:00
|
|
|
assert(std::prev(GlobalSegment)->start < LocalLI->beginIndex() &&
|
2013-06-10 20:36:52 +00:00
|
|
|
"Disconnected LRG within the scheduling region.");
|
|
|
|
}
|
|
|
|
MachineInstr *GlobalDef = LIS->getInstructionFromIndex(GlobalSegment->start);
|
|
|
|
if (!GlobalDef)
|
|
|
|
return;
|
|
|
|
|
|
|
|
SUnit *GlobalSU = DAG->getSUnit(GlobalDef);
|
|
|
|
if (!GlobalSU)
|
|
|
|
return;
|
|
|
|
|
|
|
|
// GlobalDef is the bottom of the GlobalLI hole. Open the hole by
|
|
|
|
// constraining the uses of the last local def to precede GlobalDef.
|
|
|
|
SmallVector<SUnit*,8> LocalUses;
|
|
|
|
const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(LocalLI->endIndex());
|
|
|
|
MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(LastLocalVN->def);
|
|
|
|
SUnit *LastLocalSU = DAG->getSUnit(LastLocalDef);
|
2017-06-26 20:32:52 +00:00
|
|
|
for (const SDep &Succ : LastLocalSU->Succs) {
|
|
|
|
if (Succ.getKind() != SDep::Data || Succ.getReg() != LocalReg)
|
2013-06-10 20:36:52 +00:00
|
|
|
continue;
|
2017-06-26 20:32:52 +00:00
|
|
|
if (Succ.getSUnit() == GlobalSU)
|
2013-06-10 20:36:52 +00:00
|
|
|
continue;
|
2017-06-26 20:32:52 +00:00
|
|
|
if (!DAG->canAddEdge(GlobalSU, Succ.getSUnit()))
|
2013-06-10 20:36:52 +00:00
|
|
|
return;
|
2017-06-26 20:32:52 +00:00
|
|
|
LocalUses.push_back(Succ.getSUnit());
|
2013-06-10 20:36:52 +00:00
|
|
|
}
|
|
|
|
// Open the top of the GlobalLI hole by constraining any earlier global uses
|
|
|
|
// to precede the start of LocalLI.
|
|
|
|
SmallVector<SUnit*,8> GlobalUses;
|
|
|
|
MachineInstr *FirstLocalDef =
|
|
|
|
LIS->getInstructionFromIndex(LocalLI->beginIndex());
|
|
|
|
SUnit *FirstLocalSU = DAG->getSUnit(FirstLocalDef);
|
2017-06-26 20:32:52 +00:00
|
|
|
for (const SDep &Pred : GlobalSU->Preds) {
|
|
|
|
if (Pred.getKind() != SDep::Anti || Pred.getReg() != GlobalReg)
|
2013-06-10 20:36:52 +00:00
|
|
|
continue;
|
2017-06-26 20:32:52 +00:00
|
|
|
if (Pred.getSUnit() == FirstLocalSU)
|
2013-06-10 20:36:52 +00:00
|
|
|
continue;
|
2017-06-26 20:32:52 +00:00
|
|
|
if (!DAG->canAddEdge(FirstLocalSU, Pred.getSUnit()))
|
2013-06-10 20:36:52 +00:00
|
|
|
return;
|
2017-06-26 20:32:52 +00:00
|
|
|
GlobalUses.push_back(Pred.getSUnit());
|
2013-06-10 20:36:52 +00:00
|
|
|
}
|
|
|
|
DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n");
|
|
|
|
// Add the weak edges.
|
|
|
|
for (SmallVectorImpl<SUnit*>::const_iterator
|
|
|
|
I = LocalUses.begin(), E = LocalUses.end(); I != E; ++I) {
|
|
|
|
DEBUG(dbgs() << " Local use SU(" << (*I)->NodeNum << ") -> SU("
|
|
|
|
<< GlobalSU->NodeNum << ")\n");
|
|
|
|
DAG->addEdge(GlobalSU, SDep(*I, SDep::Weak));
|
|
|
|
}
|
|
|
|
for (SmallVectorImpl<SUnit*>::const_iterator
|
|
|
|
I = GlobalUses.begin(), E = GlobalUses.end(); I != E; ++I) {
|
|
|
|
DEBUG(dbgs() << " Global use SU(" << (*I)->NodeNum << ") -> SU("
|
|
|
|
<< FirstLocalSU->NodeNum << ")\n");
|
|
|
|
DAG->addEdge(FirstLocalSU, SDep(*I, SDep::Weak));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// \brief Callback from DAG postProcessing to create weak edges to encourage
|
|
|
|
/// copy elimination.
|
2016-07-23 20:41:05 +00:00
|
|
|
void CopyConstrain::apply(ScheduleDAGInstrs *DAGInstrs) {
|
|
|
|
ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs);
|
2014-11-24 09:08:18 +00:00
|
|
|
assert(DAG->hasVRegLiveness() && "Expect VRegs with LiveIntervals");
|
|
|
|
|
2013-06-10 20:36:52 +00:00
|
|
|
MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end());
|
|
|
|
if (FirstPos == DAG->end())
|
|
|
|
return;
|
2016-07-23 20:41:05 +00:00
|
|
|
RegionBeginIdx = DAG->getLIS()->getInstructionIndex(*FirstPos);
|
2013-06-10 20:36:52 +00:00
|
|
|
RegionEndIdx = DAG->getLIS()->getInstructionIndex(
|
2016-07-23 20:41:05 +00:00
|
|
|
*priorNonDebug(DAG->end(), DAG->begin()));
|
2013-06-10 20:36:52 +00:00
|
|
|
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SUnit &SU : DAG->SUnits) {
|
|
|
|
if (!SU.getInstr()->isCopy())
|
2013-06-10 20:36:52 +00:00
|
|
|
continue;
|
|
|
|
|
2017-06-26 20:32:52 +00:00
|
|
|
constrainLocalCopy(&SU, static_cast<ScheduleDAGMILive*>(DAG));
|
2013-06-10 20:36:52 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
2014-11-24 09:08:18 +00:00
|
|
|
// MachineSchedStrategy helpers used by GenericScheduler, GenericPostScheduler
|
|
|
|
// and possibly other custom schedulers.
|
2012-04-14 13:54:10 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
static const unsigned InvalidCycle = ~0U;
|
2012-08-15 19:34:23 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
SchedBoundary::~SchedBoundary() { delete HazardRec; }
|
2013-06-10 20:36:52 +00:00
|
|
|
|
2017-12-18 20:10:56 +00:00
|
|
|
/// Given a Count of resource usage and a Latency value, return true if a
|
|
|
|
/// SchedBoundary becomes resource limited.
|
|
|
|
static bool checkResourceLimit(unsigned LFactor, unsigned Count,
|
|
|
|
unsigned Latency) {
|
|
|
|
return (int)(Count - (Latency * LFactor)) > (int)LFactor;
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void SchedBoundary::reset() {
|
|
|
|
// A new HazardRec is created for each DAG and owned by SchedBoundary.
|
|
|
|
// Destroying and reconstructing it is very expensive though. So keep
|
|
|
|
// invalid, placeholder HazardRecs.
|
|
|
|
if (HazardRec && HazardRec->isEnabled()) {
|
|
|
|
delete HazardRec;
|
|
|
|
HazardRec = nullptr;
|
|
|
|
}
|
|
|
|
Available.clear();
|
|
|
|
Pending.clear();
|
|
|
|
CheckPending = false;
|
|
|
|
CurrCycle = 0;
|
|
|
|
CurrMOps = 0;
|
2017-04-16 16:01:22 +00:00
|
|
|
MinReadyCycle = std::numeric_limits<unsigned>::max();
|
2014-11-24 09:08:18 +00:00
|
|
|
ExpectedLatency = 0;
|
|
|
|
DependentLatency = 0;
|
|
|
|
RetiredMOps = 0;
|
|
|
|
MaxExecutedResCount = 0;
|
|
|
|
ZoneCritResIdx = 0;
|
|
|
|
IsResourceLimited = false;
|
|
|
|
ReservedCycles.clear();
|
2012-08-15 19:34:23 +00:00
|
|
|
#ifndef NDEBUG
|
2014-11-24 09:08:18 +00:00
|
|
|
// Track the maximum number of stall cycles that could arise either from the
|
|
|
|
// latency of a DAG edge or the number of cycles that a processor resource is
|
|
|
|
// reserved (SchedBoundary::ReservedCycles).
|
|
|
|
MaxObservedStall = 0;
|
2012-08-15 19:34:23 +00:00
|
|
|
#endif
|
2014-11-24 09:08:18 +00:00
|
|
|
// Reserve a zero-count for invalid CritResIdx.
|
|
|
|
ExecutedResCounts.resize(1);
|
|
|
|
assert(!ExecutedResCounts[0] && "nonzero count for bad resource");
|
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void SchedRemainder::
|
2012-12-02 13:10:19 +00:00
|
|
|
init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
|
|
|
|
reset();
|
|
|
|
if (!SchedModel->hasInstrSchedModel())
|
|
|
|
return;
|
|
|
|
RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SUnit &SU : DAG->SUnits) {
|
|
|
|
const MCSchedClassDesc *SC = DAG->getSchedClass(&SU);
|
|
|
|
RemIssueCount += SchedModel->getNumMicroOps(SU.getInstr(), SC)
|
2013-12-22 00:04:03 +00:00
|
|
|
* SchedModel->getMicroOpFactor();
|
2012-12-02 13:10:19 +00:00
|
|
|
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);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void SchedBoundary::
|
2012-12-02 13:10:19 +00:00
|
|
|
init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
|
|
|
|
reset();
|
|
|
|
DAG = dag;
|
|
|
|
SchedModel = smodel;
|
|
|
|
Rem = rem;
|
2014-11-24 09:08:18 +00:00
|
|
|
if (SchedModel->hasInstrSchedModel()) {
|
2013-12-22 00:04:03 +00:00
|
|
|
ExecutedResCounts.resize(SchedModel->getNumProcResourceKinds());
|
2014-11-24 09:08:18 +00:00
|
|
|
ReservedCycles.resize(SchedModel->getNumProcResourceKinds(), InvalidCycle);
|
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
/// Compute the stall cycles based on this SUnit's ready time. Heuristics treat
|
|
|
|
/// these "soft stalls" differently than the hard stall cycles based on CPU
|
|
|
|
/// resources and computed by checkHazard(). A fully in-order model
|
|
|
|
/// (MicroOpBufferSize==0) will not make use of this since instructions are not
|
|
|
|
/// available for scheduling until they are ready. However, a weaker in-order
|
|
|
|
/// model may use this for heuristics. For example, if a processor has in-order
|
|
|
|
/// behavior when reading certain resources, this may come into play.
|
|
|
|
unsigned SchedBoundary::getLatencyStallCycles(SUnit *SU) {
|
|
|
|
if (!SU->isUnbuffered)
|
|
|
|
return 0;
|
2013-12-22 00:04:03 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
|
|
|
|
if (ReadyCycle > CurrCycle)
|
|
|
|
return ReadyCycle - CurrCycle;
|
|
|
|
return 0;
|
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
/// Compute the next cycle at which the given processor resource can be
|
|
|
|
/// scheduled.
|
|
|
|
unsigned SchedBoundary::
|
|
|
|
getNextResourceCycle(unsigned PIdx, unsigned Cycles) {
|
|
|
|
unsigned NextUnreserved = ReservedCycles[PIdx];
|
|
|
|
// If this resource has never been used, always return cycle zero.
|
|
|
|
if (NextUnreserved == InvalidCycle)
|
|
|
|
return 0;
|
|
|
|
// For bottom-up scheduling add the cycles needed for the current operation.
|
|
|
|
if (!isTop())
|
|
|
|
NextUnreserved += Cycles;
|
|
|
|
return NextUnreserved;
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
|
|
|
|
2012-08-15 19:34:23 +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.
|
2014-11-24 09:08:18 +00:00
|
|
|
bool SchedBoundary::checkHazard(SUnit *SU) {
|
|
|
|
if (HazardRec->isEnabled()
|
|
|
|
&& HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard) {
|
|
|
|
return true;
|
|
|
|
}
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
|
2013-12-22 00:04:03 +00:00
|
|
|
if ((CurrMOps > 0) && (CurrMOps + uops > SchedModel->getIssueWidth())) {
|
2012-12-02 13:10:19 +00:00
|
|
|
DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
|
|
|
|
<< SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
|
2012-08-15 19:34:23 +00:00
|
|
|
return true;
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2017-04-16 16:01:22 +00:00
|
|
|
|
|
|
|
if (CurrMOps > 0 &&
|
|
|
|
((isTop() && SchedModel->mustBeginGroup(SU->getInstr())) ||
|
|
|
|
(!isTop() && SchedModel->mustEndGroup(SU->getInstr())))) {
|
|
|
|
DEBUG(dbgs() << " hazard: SU(" << SU->NodeNum << ") must "
|
|
|
|
<< (isTop()? "begin" : "end") << " group\n");
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
if (SchedModel->hasInstrSchedModel() && SU->hasReservedResource) {
|
|
|
|
const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
|
2017-12-18 20:10:56 +00:00
|
|
|
for (const MCWriteProcResEntry &PE :
|
|
|
|
make_range(SchedModel->getWriteProcResBegin(SC),
|
|
|
|
SchedModel->getWriteProcResEnd(SC))) {
|
|
|
|
unsigned ResIdx = PE.ProcResourceIdx;
|
|
|
|
unsigned Cycles = PE.Cycles;
|
|
|
|
unsigned NRCycle = getNextResourceCycle(ResIdx, Cycles);
|
2014-11-24 09:08:18 +00:00
|
|
|
if (NRCycle > CurrCycle) {
|
|
|
|
#ifndef NDEBUG
|
2017-12-18 20:10:56 +00:00
|
|
|
MaxObservedStall = std::max(Cycles, MaxObservedStall);
|
2014-11-24 09:08:18 +00:00
|
|
|
#endif
|
|
|
|
DEBUG(dbgs() << " SU(" << SU->NodeNum << ") "
|
2017-12-18 20:10:56 +00:00
|
|
|
<< SchedModel->getResourceName(ResIdx)
|
2014-11-24 09:08:18 +00:00
|
|
|
<< "=" << NRCycle << "c\n");
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
// Find the unscheduled node in ReadySUs with the highest latency.
|
2014-11-24 09:08:18 +00:00
|
|
|
unsigned SchedBoundary::
|
2013-12-22 00:04:03 +00:00
|
|
|
findMaxLatency(ArrayRef<SUnit*> ReadySUs) {
|
2014-11-24 09:08:18 +00:00
|
|
|
SUnit *LateSU = nullptr;
|
2013-04-08 18:41:23 +00:00
|
|
|
unsigned RemLatency = 0;
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SUnit *SU : ReadySUs) {
|
|
|
|
unsigned L = getUnscheduledLatency(SU);
|
2013-12-22 00:04:03 +00:00
|
|
|
if (L > RemLatency) {
|
2013-04-08 18:41:23 +00:00
|
|
|
RemLatency = L;
|
2017-06-26 20:32:52 +00:00
|
|
|
LateSU = SU;
|
2013-12-22 00:04:03 +00:00
|
|
|
}
|
2013-04-08 18:41:23 +00:00
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
if (LateSU) {
|
|
|
|
DEBUG(dbgs() << Available.getName() << " RemLatency SU("
|
|
|
|
<< LateSU->NodeNum << ") " << RemLatency << "c\n");
|
2013-04-08 18:41:23 +00:00
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
return RemLatency;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Count resources in this zone and the remaining unscheduled
|
|
|
|
// instruction. Return the max count, scaled. Set OtherCritIdx to the critical
|
|
|
|
// resource index, or zero if the zone is issue limited.
|
2014-11-24 09:08:18 +00:00
|
|
|
unsigned SchedBoundary::
|
2013-12-22 00:04:03 +00:00
|
|
|
getOtherResourceCount(unsigned &OtherCritIdx) {
|
|
|
|
OtherCritIdx = 0;
|
|
|
|
if (!SchedModel->hasInstrSchedModel())
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
unsigned OtherCritCount = Rem->RemIssueCount
|
|
|
|
+ (RetiredMOps * SchedModel->getMicroOpFactor());
|
|
|
|
DEBUG(dbgs() << " " << Available.getName() << " + Remain MOps: "
|
|
|
|
<< OtherCritCount / SchedModel->getMicroOpFactor() << '\n');
|
|
|
|
for (unsigned PIdx = 1, PEnd = SchedModel->getNumProcResourceKinds();
|
|
|
|
PIdx != PEnd; ++PIdx) {
|
|
|
|
unsigned OtherCount = getResourceCount(PIdx) + Rem->RemainingCounts[PIdx];
|
|
|
|
if (OtherCount > OtherCritCount) {
|
|
|
|
OtherCritCount = OtherCount;
|
|
|
|
OtherCritIdx = PIdx;
|
|
|
|
}
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
if (OtherCritIdx) {
|
|
|
|
DEBUG(dbgs() << " " << Available.getName() << " + Remain CritRes: "
|
|
|
|
<< OtherCritCount / SchedModel->getResourceFactor(OtherCritIdx)
|
2014-11-24 09:08:18 +00:00
|
|
|
<< " " << SchedModel->getResourceName(OtherCritIdx) << "\n");
|
2013-12-22 00:04:03 +00:00
|
|
|
}
|
|
|
|
return OtherCritCount;
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void SchedBoundary::releaseNode(SUnit *SU, unsigned ReadyCycle) {
|
|
|
|
assert(SU->getInstr() && "Scheduled SUnit must have instr");
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
// ReadyCycle was been bumped up to the CurrCycle when this node was
|
|
|
|
// scheduled, but CurrCycle may have been eagerly advanced immediately after
|
|
|
|
// scheduling, so may now be greater than ReadyCycle.
|
|
|
|
if (ReadyCycle > CurrCycle)
|
|
|
|
MaxObservedStall = std::max(ReadyCycle - CurrCycle, MaxObservedStall);
|
|
|
|
#endif
|
2013-12-22 00:04:03 +00:00
|
|
|
|
2012-08-15 19:34:23 +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.
|
2013-12-22 00:04:03 +00:00
|
|
|
bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
|
2016-07-23 20:41:05 +00:00
|
|
|
if ((!IsBuffered && ReadyCycle > CurrCycle) || checkHazard(SU) ||
|
|
|
|
Available.size() >= ReadyListLimit)
|
2012-08-15 19:34:23 +00:00
|
|
|
Pending.push(SU);
|
|
|
|
else
|
|
|
|
Available.push(SU);
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
/// Move the boundary of scheduled code by one cycle.
|
2014-11-24 09:08:18 +00:00
|
|
|
void SchedBoundary::bumpCycle(unsigned NextCycle) {
|
2013-12-22 00:04:03 +00:00
|
|
|
if (SchedModel->getMicroOpBufferSize() == 0) {
|
2017-04-16 16:01:22 +00:00
|
|
|
assert(MinReadyCycle < std::numeric_limits<unsigned>::max() &&
|
|
|
|
"MinReadyCycle uninitialized");
|
2013-12-22 00:04:03 +00:00
|
|
|
if (MinReadyCycle > NextCycle)
|
|
|
|
NextCycle = MinReadyCycle;
|
|
|
|
}
|
|
|
|
// Update the current micro-ops, which will issue in the next cycle.
|
|
|
|
unsigned DecMOps = SchedModel->getIssueWidth() * (NextCycle - CurrCycle);
|
|
|
|
CurrMOps = (CurrMOps <= DecMOps) ? 0 : CurrMOps - DecMOps;
|
|
|
|
|
|
|
|
// Decrement DependentLatency based on the next cycle.
|
|
|
|
if ((NextCycle - CurrCycle) > DependentLatency)
|
|
|
|
DependentLatency = 0;
|
|
|
|
else
|
|
|
|
DependentLatency -= (NextCycle - CurrCycle);
|
2012-08-15 19:34:23 +00:00
|
|
|
|
|
|
|
if (!HazardRec->isEnabled()) {
|
|
|
|
// Bypass HazardRec virtual calls.
|
|
|
|
CurrCycle = NextCycle;
|
2016-07-23 20:41:05 +00:00
|
|
|
} else {
|
2012-08-15 19:34:23 +00:00
|
|
|
// Bypass getHazardType calls in case of long latency.
|
|
|
|
for (; CurrCycle != NextCycle; ++CurrCycle) {
|
|
|
|
if (isTop())
|
|
|
|
HazardRec->AdvanceCycle();
|
|
|
|
else
|
|
|
|
HazardRec->RecedeCycle();
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
|
|
|
CheckPending = true;
|
2013-12-22 00:04:03 +00:00
|
|
|
IsResourceLimited =
|
2017-12-18 20:10:56 +00:00
|
|
|
checkResourceLimit(SchedModel->getLatencyFactor(), getCriticalCount(),
|
|
|
|
getScheduledLatency());
|
2012-08-15 19:34:23 +00:00
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() << '\n');
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void SchedBoundary::incExecutedResources(unsigned PIdx, unsigned Count) {
|
2013-12-22 00:04:03 +00:00
|
|
|
ExecutedResCounts[PIdx] += Count;
|
|
|
|
if (ExecutedResCounts[PIdx] > MaxExecutedResCount)
|
|
|
|
MaxExecutedResCount = ExecutedResCounts[PIdx];
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
/// Add the given processor resource to this scheduled zone.
|
2013-12-22 00:04:03 +00:00
|
|
|
///
|
|
|
|
/// \param Cycles indicates the number of consecutive (non-pipelined) cycles
|
|
|
|
/// during which this resource is consumed.
|
|
|
|
///
|
|
|
|
/// \return the next cycle at which the instruction may execute without
|
|
|
|
/// oversubscribing resources.
|
2014-11-24 09:08:18 +00:00
|
|
|
unsigned SchedBoundary::
|
|
|
|
countResource(unsigned PIdx, unsigned Cycles, unsigned NextCycle) {
|
2012-12-02 13:10:19 +00:00
|
|
|
unsigned Factor = SchedModel->getResourceFactor(PIdx);
|
|
|
|
unsigned Count = Factor * Cycles;
|
2014-11-24 09:08:18 +00:00
|
|
|
DEBUG(dbgs() << " " << SchedModel->getResourceName(PIdx)
|
2013-12-22 00:04:03 +00:00
|
|
|
<< " +" << Cycles << "x" << Factor << "u\n");
|
|
|
|
|
|
|
|
// Update Executed resources counts.
|
|
|
|
incExecutedResources(PIdx, Count);
|
2012-12-02 13:10:19 +00:00
|
|
|
assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
|
|
|
|
Rem->RemainingCounts[PIdx] -= Count;
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
// Check if this resource exceeds the current critical resource. If so, it
|
|
|
|
// becomes the critical resource.
|
|
|
|
if (ZoneCritResIdx != PIdx && (getResourceCount(PIdx) > getCriticalCount())) {
|
|
|
|
ZoneCritResIdx = PIdx;
|
2012-12-02 13:10:19 +00:00
|
|
|
DEBUG(dbgs() << " *** Critical resource "
|
2014-11-24 09:08:18 +00:00
|
|
|
<< SchedModel->getResourceName(PIdx) << ": "
|
2013-12-22 00:04:03 +00:00
|
|
|
<< getResourceCount(PIdx) / SchedModel->getLatencyFactor() << "c\n");
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2014-11-24 09:08:18 +00:00
|
|
|
// For reserved resources, record the highest cycle using the resource.
|
|
|
|
unsigned NextAvailable = getNextResourceCycle(PIdx, Cycles);
|
|
|
|
if (NextAvailable > CurrCycle) {
|
|
|
|
DEBUG(dbgs() << " Resource conflict: "
|
|
|
|
<< SchedModel->getProcResource(PIdx)->Name << " reserved until @"
|
|
|
|
<< NextAvailable << "\n");
|
|
|
|
}
|
|
|
|
return NextAvailable;
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
/// Move the boundary of scheduled code by one SUnit.
|
2014-11-24 09:08:18 +00:00
|
|
|
void SchedBoundary::bumpNode(SUnit *SU) {
|
2012-08-15 19:34:23 +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();
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
HazardRec->EmitInstruction(SU);
|
|
|
|
}
|
2014-11-24 09:08:18 +00:00
|
|
|
// checkHazard should prevent scheduling multiple instructions per cycle that
|
|
|
|
// exceed the issue width.
|
2013-12-22 00:04:03 +00:00
|
|
|
const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
|
|
|
|
unsigned IncMOps = SchedModel->getNumMicroOps(SU->getInstr());
|
2014-11-24 09:08:18 +00:00
|
|
|
assert(
|
|
|
|
(CurrMOps == 0 || (CurrMOps + IncMOps) <= SchedModel->getIssueWidth()) &&
|
|
|
|
"Cannot schedule this instruction's MicroOps in the current cycle.");
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
|
|
|
|
DEBUG(dbgs() << " Ready @" << ReadyCycle << "c\n");
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
unsigned NextCycle = CurrCycle;
|
2013-12-22 00:04:03 +00:00
|
|
|
switch (SchedModel->getMicroOpBufferSize()) {
|
|
|
|
case 0:
|
|
|
|
assert(ReadyCycle <= CurrCycle && "Broken PendingQueue");
|
|
|
|
break;
|
|
|
|
case 1:
|
|
|
|
if (ReadyCycle > NextCycle) {
|
|
|
|
NextCycle = ReadyCycle;
|
|
|
|
DEBUG(dbgs() << " *** Stall until: " << ReadyCycle << "\n");
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
// We don't currently model the OOO reorder buffer, so consider all
|
2014-11-24 09:08:18 +00:00
|
|
|
// scheduled MOps to be "retired". We do loosely model in-order resource
|
|
|
|
// latency. If this instruction uses an in-order resource, account for any
|
|
|
|
// likely stall cycles.
|
|
|
|
if (SU->isUnbuffered && ReadyCycle > NextCycle)
|
|
|
|
NextCycle = ReadyCycle;
|
2013-12-22 00:04:03 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
RetiredMOps += IncMOps;
|
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
// Update resource counts and critical resource.
|
|
|
|
if (SchedModel->hasInstrSchedModel()) {
|
2013-12-22 00:04:03 +00:00
|
|
|
unsigned DecRemIssue = IncMOps * SchedModel->getMicroOpFactor();
|
|
|
|
assert(Rem->RemIssueCount >= DecRemIssue && "MOps double counted");
|
|
|
|
Rem->RemIssueCount -= DecRemIssue;
|
|
|
|
if (ZoneCritResIdx) {
|
|
|
|
// Scale scheduled micro-ops for comparing with the critical resource.
|
|
|
|
unsigned ScaledMOps =
|
|
|
|
RetiredMOps * SchedModel->getMicroOpFactor();
|
|
|
|
|
|
|
|
// If scaled micro-ops are now more than the previous critical resource by
|
|
|
|
// a full cycle, then micro-ops issue becomes critical.
|
|
|
|
if ((int)(ScaledMOps - getResourceCount(ZoneCritResIdx))
|
|
|
|
>= (int)SchedModel->getLatencyFactor()) {
|
|
|
|
ZoneCritResIdx = 0;
|
|
|
|
DEBUG(dbgs() << " *** Critical resource NumMicroOps: "
|
|
|
|
<< ScaledMOps / SchedModel->getLatencyFactor() << "c\n");
|
|
|
|
}
|
|
|
|
}
|
2012-12-02 13:10:19 +00:00
|
|
|
for (TargetSchedModel::ProcResIter
|
|
|
|
PI = SchedModel->getWriteProcResBegin(SC),
|
|
|
|
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
|
2013-12-22 00:04:03 +00:00
|
|
|
unsigned RCycle =
|
2014-11-24 09:08:18 +00:00
|
|
|
countResource(PI->ProcResourceIdx, PI->Cycles, NextCycle);
|
2013-12-22 00:04:03 +00:00
|
|
|
if (RCycle > NextCycle)
|
|
|
|
NextCycle = RCycle;
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2014-11-24 09:08:18 +00:00
|
|
|
if (SU->hasReservedResource) {
|
|
|
|
// For reserved resources, record the highest cycle using the resource.
|
|
|
|
// For top-down scheduling, this is the cycle in which we schedule this
|
|
|
|
// instruction plus the number of cycles the operations reserves the
|
|
|
|
// resource. For bottom-up is it simply the instruction's cycle.
|
|
|
|
for (TargetSchedModel::ProcResIter
|
|
|
|
PI = SchedModel->getWriteProcResBegin(SC),
|
|
|
|
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
|
|
|
|
unsigned PIdx = PI->ProcResourceIdx;
|
|
|
|
if (SchedModel->getProcResource(PIdx)->BufferSize == 0) {
|
|
|
|
if (isTop()) {
|
|
|
|
ReservedCycles[PIdx] =
|
|
|
|
std::max(getNextResourceCycle(PIdx, 0), NextCycle + PI->Cycles);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
ReservedCycles[PIdx] = NextCycle;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
// Update ExpectedLatency and DependentLatency.
|
|
|
|
unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency;
|
|
|
|
unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency;
|
|
|
|
if (SU->getDepth() > TopLatency) {
|
|
|
|
TopLatency = SU->getDepth();
|
|
|
|
DEBUG(dbgs() << " " << Available.getName()
|
|
|
|
<< " TopLatency SU(" << SU->NodeNum << ") " << TopLatency << "c\n");
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
if (SU->getHeight() > BotLatency) {
|
|
|
|
BotLatency = SU->getHeight();
|
|
|
|
DEBUG(dbgs() << " " << Available.getName()
|
|
|
|
<< " BotLatency SU(" << SU->NodeNum << ") " << BotLatency << "c\n");
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
// If we stall for any reason, bump the cycle.
|
2017-12-18 20:10:56 +00:00
|
|
|
if (NextCycle > CurrCycle)
|
2013-12-22 00:04:03 +00:00
|
|
|
bumpCycle(NextCycle);
|
2017-12-18 20:10:56 +00:00
|
|
|
else
|
2013-12-22 00:04:03 +00:00
|
|
|
// After updating ZoneCritResIdx and ExpectedLatency, check if we're
|
2014-11-24 09:08:18 +00:00
|
|
|
// resource limited. If a stall occurred, bumpCycle does this.
|
2013-12-22 00:04:03 +00:00
|
|
|
IsResourceLimited =
|
2017-12-18 20:10:56 +00:00
|
|
|
checkResourceLimit(SchedModel->getLatencyFactor(), getCriticalCount(),
|
|
|
|
getScheduledLatency());
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
// Update CurrMOps after calling bumpCycle to handle stalls, since bumpCycle
|
|
|
|
// resets CurrMOps. Loop to handle instructions with more MOps than issue in
|
|
|
|
// one cycle. Since we commonly reach the max MOps here, opportunistically
|
|
|
|
// bump the cycle to avoid uselessly checking everything in the readyQ.
|
|
|
|
CurrMOps += IncMOps;
|
2017-04-16 16:01:22 +00:00
|
|
|
|
|
|
|
// Bump the cycle count for issue group constraints.
|
|
|
|
// This must be done after NextCycle has been adjust for all other stalls.
|
|
|
|
// Calling bumpCycle(X) will reduce CurrMOps by one issue group and set
|
|
|
|
// currCycle to X.
|
|
|
|
if ((isTop() && SchedModel->mustEndGroup(SU->getInstr())) ||
|
|
|
|
(!isTop() && SchedModel->mustBeginGroup(SU->getInstr()))) {
|
|
|
|
DEBUG(dbgs() << " Bump cycle to "
|
|
|
|
<< (isTop() ? "end" : "begin") << " group\n");
|
|
|
|
bumpCycle(++NextCycle);
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
while (CurrMOps >= SchedModel->getIssueWidth()) {
|
|
|
|
DEBUG(dbgs() << " *** Max MOps " << CurrMOps
|
|
|
|
<< " at cycle " << CurrCycle << '\n');
|
|
|
|
bumpCycle(++NextCycle);
|
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
DEBUG(dumpScheduledState());
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Release pending ready nodes in to the available queue. This makes them
|
|
|
|
/// visible to heuristics.
|
2014-11-24 09:08:18 +00:00
|
|
|
void SchedBoundary::releasePending() {
|
2012-08-15 19:34:23 +00:00
|
|
|
// If the available queue is empty, it is safe to reset MinReadyCycle.
|
|
|
|
if (Available.empty())
|
2017-04-16 16:01:22 +00:00
|
|
|
MinReadyCycle = std::numeric_limits<unsigned>::max();
|
2012-08-15 19:34:23 +00:00
|
|
|
|
|
|
|
// Check to see if any of the pending instructions are ready to issue. If
|
|
|
|
// so, add them to the available queue.
|
2013-12-22 00:04:03 +00:00
|
|
|
bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
|
2012-08-15 19:34:23 +00:00
|
|
|
for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
|
|
|
|
SUnit *SU = *(Pending.begin()+i);
|
|
|
|
unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
|
|
|
|
|
|
|
|
if (ReadyCycle < MinReadyCycle)
|
|
|
|
MinReadyCycle = ReadyCycle;
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
if (!IsBuffered && ReadyCycle > CurrCycle)
|
2012-08-15 19:34:23 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
if (checkHazard(SU))
|
|
|
|
continue;
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
if (Available.size() >= ReadyListLimit)
|
|
|
|
break;
|
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
Available.push(SU);
|
|
|
|
Pending.remove(Pending.begin()+i);
|
|
|
|
--i; --e;
|
|
|
|
}
|
|
|
|
CheckPending = false;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Remove SU from the ready set for this boundary.
|
2014-11-24 09:08:18 +00:00
|
|
|
void SchedBoundary::removeReady(SUnit *SU) {
|
2012-08-15 19:34:23 +00:00
|
|
|
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,
|
2012-12-02 13:10:19 +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.
|
2014-11-24 09:08:18 +00:00
|
|
|
SUnit *SchedBoundary::pickOnlyChoice() {
|
2012-08-15 19:34:23 +00:00
|
|
|
if (CheckPending)
|
|
|
|
releasePending();
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
if (CurrMOps > 0) {
|
2012-12-02 13:10:19 +00:00
|
|
|
// 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;
|
|
|
|
}
|
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
for (unsigned i = 0; Available.empty(); ++i) {
|
2014-11-24 09:08:18 +00:00
|
|
|
// FIXME: Re-enable assert once PR20057 is resolved.
|
|
|
|
// assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedStall) &&
|
|
|
|
// "permanent hazard");
|
|
|
|
(void)i;
|
2013-12-22 00:04:03 +00:00
|
|
|
bumpCycle(CurrCycle + 1);
|
2012-08-15 19:34:23 +00:00
|
|
|
releasePending();
|
|
|
|
}
|
2016-07-23 20:41:05 +00:00
|
|
|
|
|
|
|
DEBUG(Pending.dump());
|
|
|
|
DEBUG(Available.dump());
|
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
if (Available.size() == 1)
|
|
|
|
return *Available.begin();
|
2014-11-24 09:08:18 +00:00
|
|
|
return nullptr;
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
|
|
|
|
2017-04-16 16:01:22 +00:00
|
|
|
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
|
2013-12-22 00:04:03 +00:00
|
|
|
// This is useful information to dump after bumpNode.
|
|
|
|
// Note that the Queue contents are more useful before pickNodeFromQueue.
|
2017-06-26 20:32:52 +00:00
|
|
|
LLVM_DUMP_METHOD void SchedBoundary::dumpScheduledState() const {
|
2013-12-22 00:04:03 +00:00
|
|
|
unsigned ResFactor;
|
|
|
|
unsigned ResCount;
|
|
|
|
if (ZoneCritResIdx) {
|
|
|
|
ResFactor = SchedModel->getResourceFactor(ZoneCritResIdx);
|
|
|
|
ResCount = getResourceCount(ZoneCritResIdx);
|
2016-07-23 20:41:05 +00:00
|
|
|
} else {
|
2013-12-22 00:04:03 +00:00
|
|
|
ResFactor = SchedModel->getMicroOpFactor();
|
2017-12-18 20:10:56 +00:00
|
|
|
ResCount = RetiredMOps * ResFactor;
|
2013-12-22 00:04:03 +00:00
|
|
|
}
|
|
|
|
unsigned LFactor = SchedModel->getLatencyFactor();
|
|
|
|
dbgs() << Available.getName() << " @" << CurrCycle << "c\n"
|
|
|
|
<< " Retired: " << RetiredMOps;
|
|
|
|
dbgs() << "\n Executed: " << getExecutedCount() / LFactor << "c";
|
|
|
|
dbgs() << "\n Critical: " << ResCount / LFactor << "c, "
|
2014-11-24 09:08:18 +00:00
|
|
|
<< ResCount / ResFactor << " "
|
|
|
|
<< SchedModel->getResourceName(ZoneCritResIdx)
|
2013-12-22 00:04:03 +00:00
|
|
|
<< "\n ExpectedLatency: " << ExpectedLatency << "c\n"
|
|
|
|
<< (IsResourceLimited ? " - Resource" : " - Latency")
|
|
|
|
<< " limited.\n";
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2013-12-22 00:04:03 +00:00
|
|
|
#endif
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// GenericScheduler - Generic implementation of MachineSchedStrategy.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
void GenericSchedulerBase::SchedCandidate::
|
2012-12-02 13:10:19 +00:00
|
|
|
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;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
/// Set the CandPolicy given a scheduling zone given the current resources and
|
|
|
|
/// latencies inside and outside the zone.
|
2016-07-23 20:41:05 +00:00
|
|
|
void GenericSchedulerBase::setPolicy(CandPolicy &Policy, bool IsPostRA,
|
2014-11-24 09:08:18 +00:00
|
|
|
SchedBoundary &CurrZone,
|
|
|
|
SchedBoundary *OtherZone) {
|
2015-06-21 13:59:01 +00:00
|
|
|
// Apply preemptive heuristics based on the total latency and resources
|
2014-11-24 09:08:18 +00:00
|
|
|
// inside and outside this zone. Potential stalls should be considered before
|
|
|
|
// following this policy.
|
|
|
|
|
|
|
|
// Compute remaining latency. We need this both to determine whether the
|
|
|
|
// overall schedule has become latency-limited and whether the instructions
|
|
|
|
// outside this zone are resource or latency limited.
|
|
|
|
//
|
|
|
|
// The "dependent" latency is updated incrementally during scheduling as the
|
|
|
|
// max height/depth of scheduled nodes minus the cycles since it was
|
|
|
|
// scheduled:
|
|
|
|
// DLat = max (N.depth - (CurrCycle - N.ReadyCycle) for N in Zone
|
|
|
|
//
|
|
|
|
// The "independent" latency is the max ready queue depth:
|
|
|
|
// ILat = max N.depth for N in Available|Pending
|
|
|
|
//
|
|
|
|
// RemainingLatency is the greater of independent and dependent latency.
|
|
|
|
unsigned RemLatency = CurrZone.getDependentLatency();
|
|
|
|
RemLatency = std::max(RemLatency,
|
|
|
|
CurrZone.findMaxLatency(CurrZone.Available.elements()));
|
|
|
|
RemLatency = std::max(RemLatency,
|
|
|
|
CurrZone.findMaxLatency(CurrZone.Pending.elements()));
|
|
|
|
|
|
|
|
// Compute the critical resource outside the zone.
|
|
|
|
unsigned OtherCritIdx = 0;
|
|
|
|
unsigned OtherCount =
|
|
|
|
OtherZone ? OtherZone->getOtherResourceCount(OtherCritIdx) : 0;
|
|
|
|
|
|
|
|
bool OtherResLimited = false;
|
2017-12-18 20:10:56 +00:00
|
|
|
if (SchedModel->hasInstrSchedModel())
|
|
|
|
OtherResLimited = checkResourceLimit(SchedModel->getLatencyFactor(),
|
|
|
|
OtherCount, RemLatency);
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
// Schedule aggressively for latency in PostRA mode. We don't check for
|
|
|
|
// acyclic latency during PostRA, and highly out-of-order processors will
|
|
|
|
// skip PostRA scheduling.
|
|
|
|
if (!OtherResLimited) {
|
|
|
|
if (IsPostRA || (RemLatency + CurrZone.getCurrCycle() > Rem.CriticalPath)) {
|
|
|
|
Policy.ReduceLatency |= true;
|
|
|
|
DEBUG(dbgs() << " " << CurrZone.Available.getName()
|
|
|
|
<< " RemainingLatency " << RemLatency << " + "
|
|
|
|
<< CurrZone.getCurrCycle() << "c > CritPath "
|
|
|
|
<< Rem.CriticalPath << "\n");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// If the same resource is limiting inside and outside the zone, do nothing.
|
|
|
|
if (CurrZone.getZoneCritResIdx() == OtherCritIdx)
|
|
|
|
return;
|
|
|
|
|
|
|
|
DEBUG(
|
|
|
|
if (CurrZone.isResourceLimited()) {
|
|
|
|
dbgs() << " " << CurrZone.Available.getName() << " ResourceLimited: "
|
|
|
|
<< SchedModel->getResourceName(CurrZone.getZoneCritResIdx())
|
|
|
|
<< "\n";
|
|
|
|
}
|
|
|
|
if (OtherResLimited)
|
|
|
|
dbgs() << " RemainingLimit: "
|
|
|
|
<< SchedModel->getResourceName(OtherCritIdx) << "\n";
|
|
|
|
if (!CurrZone.isResourceLimited() && !OtherResLimited)
|
|
|
|
dbgs() << " Latency limited both directions.\n");
|
|
|
|
|
|
|
|
if (CurrZone.isResourceLimited() && !Policy.ReduceResIdx)
|
|
|
|
Policy.ReduceResIdx = CurrZone.getZoneCritResIdx();
|
|
|
|
|
|
|
|
if (OtherResLimited)
|
|
|
|
Policy.DemandResIdx = OtherCritIdx;
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
const char *GenericSchedulerBase::getReasonStr(
|
|
|
|
GenericSchedulerBase::CandReason Reason) {
|
|
|
|
switch (Reason) {
|
|
|
|
case NoCand: return "NOCAND ";
|
2016-07-23 20:41:05 +00:00
|
|
|
case Only1: return "ONLY1 ";
|
|
|
|
case PhysRegCopy: return "PREG-COPY ";
|
2014-11-24 09:08:18 +00:00
|
|
|
case RegExcess: return "REG-EXCESS";
|
|
|
|
case RegCritical: return "REG-CRIT ";
|
|
|
|
case Stall: return "STALL ";
|
|
|
|
case Cluster: return "CLUSTER ";
|
|
|
|
case Weak: return "WEAK ";
|
|
|
|
case RegMax: return "REG-MAX ";
|
|
|
|
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 ";
|
|
|
|
};
|
|
|
|
llvm_unreachable("Unknown reason!");
|
|
|
|
}
|
|
|
|
|
|
|
|
void GenericSchedulerBase::traceCandidate(const SchedCandidate &Cand) {
|
|
|
|
PressureChange P;
|
|
|
|
unsigned ResIdx = 0;
|
|
|
|
unsigned Latency = 0;
|
|
|
|
switch (Cand.Reason) {
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
case RegExcess:
|
|
|
|
P = Cand.RPDelta.Excess;
|
|
|
|
break;
|
|
|
|
case RegCritical:
|
|
|
|
P = Cand.RPDelta.CriticalMax;
|
|
|
|
break;
|
|
|
|
case RegMax:
|
|
|
|
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;
|
|
|
|
}
|
2015-12-30 11:46:15 +00:00
|
|
|
dbgs() << " Cand SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
|
2014-11-24 09:08:18 +00:00
|
|
|
if (P.isValid())
|
|
|
|
dbgs() << " " << TRI->getRegPressureSetName(P.getPSet())
|
|
|
|
<< ":" << P.getUnitInc() << " ";
|
|
|
|
else
|
|
|
|
dbgs() << " ";
|
|
|
|
if (ResIdx)
|
|
|
|
dbgs() << " " << SchedModel->getProcResource(ResIdx)->Name << " ";
|
|
|
|
else
|
|
|
|
dbgs() << " ";
|
|
|
|
if (Latency)
|
|
|
|
dbgs() << " " << Latency << " cycles ";
|
|
|
|
else
|
|
|
|
dbgs() << " ";
|
|
|
|
dbgs() << '\n';
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/// Return true if this heuristic determines order.
|
|
|
|
static bool tryLess(int TryVal, int CandVal,
|
|
|
|
GenericSchedulerBase::SchedCandidate &TryCand,
|
|
|
|
GenericSchedulerBase::SchedCandidate &Cand,
|
|
|
|
GenericSchedulerBase::CandReason Reason) {
|
|
|
|
if (TryVal < CandVal) {
|
|
|
|
TryCand.Reason = Reason;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (TryVal > CandVal) {
|
|
|
|
if (Cand.Reason > Reason)
|
|
|
|
Cand.Reason = Reason;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool tryGreater(int TryVal, int CandVal,
|
|
|
|
GenericSchedulerBase::SchedCandidate &TryCand,
|
|
|
|
GenericSchedulerBase::SchedCandidate &Cand,
|
|
|
|
GenericSchedulerBase::CandReason Reason) {
|
|
|
|
if (TryVal > CandVal) {
|
|
|
|
TryCand.Reason = Reason;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (TryVal < CandVal) {
|
|
|
|
if (Cand.Reason > Reason)
|
|
|
|
Cand.Reason = Reason;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool tryLatency(GenericSchedulerBase::SchedCandidate &TryCand,
|
|
|
|
GenericSchedulerBase::SchedCandidate &Cand,
|
|
|
|
SchedBoundary &Zone) {
|
|
|
|
if (Zone.isTop()) {
|
|
|
|
if (Cand.SU->getDepth() > Zone.getScheduledLatency()) {
|
|
|
|
if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
|
|
|
|
TryCand, Cand, GenericSchedulerBase::TopDepthReduce))
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
|
|
|
|
TryCand, Cand, GenericSchedulerBase::TopPathReduce))
|
|
|
|
return true;
|
2016-07-23 20:41:05 +00:00
|
|
|
} else {
|
2014-11-24 09:08:18 +00:00
|
|
|
if (Cand.SU->getHeight() > Zone.getScheduledLatency()) {
|
|
|
|
if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
|
|
|
|
TryCand, Cand, GenericSchedulerBase::BotHeightReduce))
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
|
|
|
|
TryCand, Cand, GenericSchedulerBase::BotPathReduce))
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
static void tracePick(GenericSchedulerBase::CandReason Reason, bool IsTop) {
|
2014-11-24 09:08:18 +00:00
|
|
|
DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ")
|
2016-07-23 20:41:05 +00:00
|
|
|
<< GenericSchedulerBase::getReasonStr(Reason) << '\n');
|
|
|
|
}
|
|
|
|
|
|
|
|
static void tracePick(const GenericSchedulerBase::SchedCandidate &Cand) {
|
|
|
|
tracePick(Cand.Reason, Cand.AtTop);
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void GenericScheduler::initialize(ScheduleDAGMI *dag) {
|
|
|
|
assert(dag->hasVRegLiveness() &&
|
|
|
|
"(PreRA)GenericScheduler needs vreg liveness");
|
|
|
|
DAG = static_cast<ScheduleDAGMILive*>(dag);
|
|
|
|
SchedModel = DAG->getSchedModel();
|
|
|
|
TRI = DAG->TRI;
|
|
|
|
|
|
|
|
Rem.init(DAG, SchedModel);
|
|
|
|
Top.init(DAG, SchedModel, &Rem);
|
|
|
|
Bot.init(DAG, SchedModel, &Rem);
|
|
|
|
|
|
|
|
// Initialize resource counts.
|
|
|
|
|
|
|
|
// Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
|
|
|
|
// are disabled, then these HazardRecs will be disabled.
|
|
|
|
const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
|
|
|
|
if (!Top.HazardRec) {
|
|
|
|
Top.HazardRec =
|
2015-01-18 16:17:27 +00:00
|
|
|
DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer(
|
|
|
|
Itin, DAG);
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
|
|
|
if (!Bot.HazardRec) {
|
|
|
|
Bot.HazardRec =
|
2015-01-18 16:17:27 +00:00
|
|
|
DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer(
|
|
|
|
Itin, DAG);
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
2016-07-23 20:41:05 +00:00
|
|
|
TopCand.SU = nullptr;
|
|
|
|
BotCand.SU = nullptr;
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Initialize the per-region scheduling policy.
|
|
|
|
void GenericScheduler::initPolicy(MachineBasicBlock::iterator Begin,
|
|
|
|
MachineBasicBlock::iterator End,
|
|
|
|
unsigned NumRegionInstrs) {
|
2017-12-18 20:10:56 +00:00
|
|
|
const MachineFunction &MF = *Begin->getMF();
|
2015-01-18 16:17:27 +00:00
|
|
|
const TargetLowering *TLI = MF.getSubtarget().getTargetLowering();
|
2014-11-24 09:08:18 +00:00
|
|
|
|
|
|
|
// Avoid setting up the register pressure tracker for small regions to save
|
|
|
|
// compile time. As a rough heuristic, only track pressure when the number of
|
|
|
|
// schedulable instructions exceeds half the integer register file.
|
|
|
|
RegionPolicy.ShouldTrackPressure = true;
|
|
|
|
for (unsigned VT = MVT::i32; VT > (unsigned)MVT::i1; --VT) {
|
|
|
|
MVT::SimpleValueType LegalIntVT = (MVT::SimpleValueType)VT;
|
|
|
|
if (TLI->isTypeLegal(LegalIntVT)) {
|
|
|
|
unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs(
|
|
|
|
TLI->getRegClassFor(LegalIntVT));
|
|
|
|
RegionPolicy.ShouldTrackPressure = NumRegionInstrs > (NIntRegs / 2);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// For generic targets, we default to bottom-up, because it's simpler and more
|
|
|
|
// compile-time optimizations have been implemented in that direction.
|
|
|
|
RegionPolicy.OnlyBottomUp = true;
|
|
|
|
|
|
|
|
// Allow the subtarget to override default policy.
|
2016-07-23 20:41:05 +00:00
|
|
|
MF.getSubtarget().overrideSchedPolicy(RegionPolicy, NumRegionInstrs);
|
2014-11-24 09:08:18 +00:00
|
|
|
|
|
|
|
// After subtarget overrides, apply command line options.
|
|
|
|
if (!EnableRegPressure)
|
|
|
|
RegionPolicy.ShouldTrackPressure = false;
|
|
|
|
|
|
|
|
// Check -misched-topdown/bottomup can force or unforce scheduling direction.
|
|
|
|
// e.g. -misched-bottomup=false allows scheduling in both directions.
|
|
|
|
assert((!ForceTopDown || !ForceBottomUp) &&
|
|
|
|
"-misched-topdown incompatible with -misched-bottomup");
|
|
|
|
if (ForceBottomUp.getNumOccurrences() > 0) {
|
|
|
|
RegionPolicy.OnlyBottomUp = ForceBottomUp;
|
|
|
|
if (RegionPolicy.OnlyBottomUp)
|
|
|
|
RegionPolicy.OnlyTopDown = false;
|
|
|
|
}
|
|
|
|
if (ForceTopDown.getNumOccurrences() > 0) {
|
|
|
|
RegionPolicy.OnlyTopDown = ForceTopDown;
|
|
|
|
if (RegionPolicy.OnlyTopDown)
|
|
|
|
RegionPolicy.OnlyBottomUp = false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-06-26 20:32:52 +00:00
|
|
|
void GenericScheduler::dumpPolicy() const {
|
2017-04-16 16:01:22 +00:00
|
|
|
// Cannot completely remove virtual function even in release mode.
|
|
|
|
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
|
2015-12-30 11:46:15 +00:00
|
|
|
dbgs() << "GenericScheduler RegionPolicy: "
|
|
|
|
<< " ShouldTrackPressure=" << RegionPolicy.ShouldTrackPressure
|
|
|
|
<< " OnlyTopDown=" << RegionPolicy.OnlyTopDown
|
|
|
|
<< " OnlyBottomUp=" << RegionPolicy.OnlyBottomUp
|
|
|
|
<< "\n";
|
2017-04-16 16:01:22 +00:00
|
|
|
#endif
|
2015-12-30 11:46:15 +00:00
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
/// Set IsAcyclicLatencyLimited if the acyclic path is longer than the cyclic
|
|
|
|
/// critical path by more cycles than it takes to drain the instruction buffer.
|
|
|
|
/// We estimate an upper bounds on in-flight instructions as:
|
|
|
|
///
|
|
|
|
/// CyclesPerIteration = max( CyclicPath, Loop-Resource-Height )
|
|
|
|
/// InFlightIterations = AcyclicPath / CyclesPerIteration
|
|
|
|
/// InFlightResources = InFlightIterations * LoopResources
|
|
|
|
///
|
|
|
|
/// TODO: Check execution resources in addition to IssueCount.
|
|
|
|
void GenericScheduler::checkAcyclicLatency() {
|
|
|
|
if (Rem.CyclicCritPath == 0 || Rem.CyclicCritPath >= Rem.CriticalPath)
|
|
|
|
return;
|
2013-12-22 00:04:03 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
// Scaled number of cycles per loop iteration.
|
|
|
|
unsigned IterCount =
|
|
|
|
std::max(Rem.CyclicCritPath * SchedModel->getLatencyFactor(),
|
|
|
|
Rem.RemIssueCount);
|
|
|
|
// Scaled acyclic critical path.
|
|
|
|
unsigned AcyclicCount = Rem.CriticalPath * SchedModel->getLatencyFactor();
|
|
|
|
// InFlightCount = (AcyclicPath / IterCycles) * InstrPerLoop
|
|
|
|
unsigned InFlightCount =
|
|
|
|
(AcyclicCount * Rem.RemIssueCount + IterCount-1) / IterCount;
|
|
|
|
unsigned BufferLimit =
|
|
|
|
SchedModel->getMicroOpBufferSize() * SchedModel->getMicroOpFactor();
|
|
|
|
|
|
|
|
Rem.IsAcyclicLatencyLimited = InFlightCount > BufferLimit;
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "IssueCycles="
|
|
|
|
<< Rem.RemIssueCount / SchedModel->getLatencyFactor() << "c "
|
|
|
|
<< "IterCycles=" << IterCount / SchedModel->getLatencyFactor()
|
|
|
|
<< "c NumIters=" << (AcyclicCount + IterCount-1) / IterCount
|
|
|
|
<< " InFlight=" << InFlightCount / SchedModel->getMicroOpFactor()
|
|
|
|
<< "m BufferLim=" << SchedModel->getMicroOpBufferSize() << "m\n";
|
|
|
|
if (Rem.IsAcyclicLatencyLimited)
|
|
|
|
dbgs() << " ACYCLIC LATENCY LIMIT\n");
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2013-04-08 18:41:23 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void GenericScheduler::registerRoots() {
|
|
|
|
Rem.CriticalPath = DAG->ExitSU.getDepth();
|
|
|
|
|
|
|
|
// Some roots may not feed into ExitSU. Check all of them in case.
|
2017-06-26 20:32:52 +00:00
|
|
|
for (const SUnit *SU : Bot.Available) {
|
|
|
|
if (SU->getDepth() > Rem.CriticalPath)
|
|
|
|
Rem.CriticalPath = SU->getDepth();
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2015-01-18 16:17:27 +00:00
|
|
|
DEBUG(dbgs() << "Critical Path(GS-RR ): " << Rem.CriticalPath << '\n');
|
|
|
|
if (DumpCriticalPathLength) {
|
|
|
|
errs() << "Critical Path(GS-RR ): " << Rem.CriticalPath << " \n";
|
|
|
|
}
|
2014-11-24 09:08:18 +00:00
|
|
|
|
2017-04-16 16:01:22 +00:00
|
|
|
if (EnableCyclicPath && SchedModel->getMicroOpBufferSize() > 0) {
|
2014-11-24 09:08:18 +00:00
|
|
|
Rem.CyclicCritPath = DAG->computeCyclicCriticalPath();
|
|
|
|
checkAcyclicLatency();
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
static bool tryPressure(const PressureChange &TryP,
|
|
|
|
const PressureChange &CandP,
|
2014-11-24 09:08:18 +00:00
|
|
|
GenericSchedulerBase::SchedCandidate &TryCand,
|
|
|
|
GenericSchedulerBase::SchedCandidate &Cand,
|
2015-12-30 11:46:15 +00:00
|
|
|
GenericSchedulerBase::CandReason Reason,
|
|
|
|
const TargetRegisterInfo *TRI,
|
|
|
|
const MachineFunction &MF) {
|
2013-12-22 00:04:03 +00:00
|
|
|
// If one candidate decreases and the other increases, go with it.
|
|
|
|
// Invalid candidates have UnitInc==0.
|
2015-01-18 16:17:27 +00:00
|
|
|
if (tryGreater(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand,
|
|
|
|
Reason)) {
|
2013-12-22 00:04:03 +00:00
|
|
|
return true;
|
|
|
|
}
|
2016-07-23 20:41:05 +00:00
|
|
|
// Do not compare the magnitude of pressure changes between top and bottom
|
|
|
|
// boundary.
|
|
|
|
if (Cand.AtTop != TryCand.AtTop)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// If both candidates affect the same set in the same boundary, go with the
|
|
|
|
// smallest increase.
|
|
|
|
unsigned TryPSet = TryP.getPSetOrMax();
|
|
|
|
unsigned CandPSet = CandP.getPSetOrMax();
|
|
|
|
if (TryPSet == CandPSet) {
|
|
|
|
return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand,
|
|
|
|
Reason);
|
|
|
|
}
|
2015-12-30 11:46:15 +00:00
|
|
|
|
|
|
|
int TryRank = TryP.isValid() ? TRI->getRegPressureSetScore(MF, TryPSet) :
|
|
|
|
std::numeric_limits<int>::max();
|
|
|
|
|
|
|
|
int CandRank = CandP.isValid() ? TRI->getRegPressureSetScore(MF, CandPSet) :
|
|
|
|
std::numeric_limits<int>::max();
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
// If the candidates are decreasing pressure, reverse priority.
|
|
|
|
if (TryP.getUnitInc() < 0)
|
|
|
|
std::swap(TryRank, CandRank);
|
|
|
|
return tryGreater(TryRank, CandRank, TryCand, Cand, Reason);
|
|
|
|
}
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
|
|
|
|
return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
|
|
|
|
}
|
|
|
|
|
2013-06-10 20:36:52 +00:00
|
|
|
/// Minimize physical register live ranges. Regalloc wants them adjacent to
|
|
|
|
/// their physreg def/use.
|
|
|
|
///
|
|
|
|
/// FIXME: This is an unnecessary check on the critical path. Most are root/leaf
|
|
|
|
/// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled
|
|
|
|
/// with the operation that produces or consumes the physreg. We'll do this when
|
|
|
|
/// regalloc has support for parallel copies.
|
|
|
|
static int biasPhysRegCopy(const SUnit *SU, bool isTop) {
|
|
|
|
const MachineInstr *MI = SU->getInstr();
|
|
|
|
if (!MI->isCopy())
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
unsigned ScheduledOper = isTop ? 1 : 0;
|
|
|
|
unsigned UnscheduledOper = isTop ? 0 : 1;
|
|
|
|
// If we have already scheduled the physreg produce/consumer, immediately
|
|
|
|
// schedule the copy.
|
|
|
|
if (TargetRegisterInfo::isPhysicalRegister(
|
|
|
|
MI->getOperand(ScheduledOper).getReg()))
|
|
|
|
return 1;
|
|
|
|
// If the physreg is at the boundary, defer it. Otherwise schedule it
|
|
|
|
// immediately to free the dependent. We can hoist the copy later.
|
|
|
|
bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft;
|
|
|
|
if (TargetRegisterInfo::isPhysicalRegister(
|
|
|
|
MI->getOperand(UnscheduledOper).getReg()))
|
|
|
|
return AtBoundary ? -1 : 1;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
void GenericScheduler::initCandidate(SchedCandidate &Cand, SUnit *SU,
|
|
|
|
bool AtTop,
|
|
|
|
const RegPressureTracker &RPTracker,
|
|
|
|
RegPressureTracker &TempTracker) {
|
|
|
|
Cand.SU = SU;
|
|
|
|
Cand.AtTop = AtTop;
|
2013-12-22 00:04:03 +00:00
|
|
|
if (DAG->isTrackingPressure()) {
|
2016-07-23 20:41:05 +00:00
|
|
|
if (AtTop) {
|
2013-12-22 00:04:03 +00:00
|
|
|
TempTracker.getMaxDownwardPressureDelta(
|
2016-07-23 20:41:05 +00:00
|
|
|
Cand.SU->getInstr(),
|
|
|
|
Cand.RPDelta,
|
2013-12-22 00:04:03 +00:00
|
|
|
DAG->getRegionCriticalPSets(),
|
|
|
|
DAG->getRegPressure().MaxSetPressure);
|
2016-07-23 20:41:05 +00:00
|
|
|
} else {
|
2013-12-22 00:04:03 +00:00
|
|
|
if (VerifyScheduling) {
|
|
|
|
TempTracker.getMaxUpwardPressureDelta(
|
2016-07-23 20:41:05 +00:00
|
|
|
Cand.SU->getInstr(),
|
|
|
|
&DAG->getPressureDiff(Cand.SU),
|
|
|
|
Cand.RPDelta,
|
2013-12-22 00:04:03 +00:00
|
|
|
DAG->getRegionCriticalPSets(),
|
|
|
|
DAG->getRegPressure().MaxSetPressure);
|
2016-07-23 20:41:05 +00:00
|
|
|
} else {
|
2013-12-22 00:04:03 +00:00
|
|
|
RPTracker.getUpwardPressureDelta(
|
2016-07-23 20:41:05 +00:00
|
|
|
Cand.SU->getInstr(),
|
|
|
|
DAG->getPressureDiff(Cand.SU),
|
|
|
|
Cand.RPDelta,
|
2013-12-22 00:04:03 +00:00
|
|
|
DAG->getRegionCriticalPSets(),
|
|
|
|
DAG->getRegPressure().MaxSetPressure);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2016-07-23 20:41:05 +00:00
|
|
|
DEBUG(if (Cand.RPDelta.Excess.isValid())
|
|
|
|
dbgs() << " Try SU(" << Cand.SU->NodeNum << ") "
|
|
|
|
<< TRI->getRegPressureSetName(Cand.RPDelta.Excess.getPSet())
|
|
|
|
<< ":" << Cand.RPDelta.Excess.getUnitInc() << "\n");
|
|
|
|
}
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2016-07-23 20:41:05 +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, or nullptr
|
|
|
|
// if Cand is from a different zone than TryCand.
|
|
|
|
void GenericScheduler::tryCandidate(SchedCandidate &Cand,
|
|
|
|
SchedCandidate &TryCand,
|
|
|
|
SchedBoundary *Zone) {
|
2012-12-02 13:10:19 +00:00
|
|
|
// Initialize the candidate if needed.
|
|
|
|
if (!Cand.isValid()) {
|
|
|
|
TryCand.Reason = NodeOrder;
|
|
|
|
return;
|
|
|
|
}
|
2013-06-10 20:36:52 +00:00
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
if (tryGreater(biasPhysRegCopy(TryCand.SU, TryCand.AtTop),
|
|
|
|
biasPhysRegCopy(Cand.SU, Cand.AtTop),
|
2013-06-10 20:36:52 +00:00
|
|
|
TryCand, Cand, PhysRegCopy))
|
|
|
|
return;
|
|
|
|
|
2015-05-27 18:44:32 +00:00
|
|
|
// Avoid exceeding the target's limit.
|
2013-12-22 00:04:03 +00:00
|
|
|
if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess,
|
|
|
|
Cand.RPDelta.Excess,
|
2015-12-30 11:46:15 +00:00
|
|
|
TryCand, Cand, RegExcess, TRI,
|
|
|
|
DAG->MF))
|
2012-12-02 13:10:19 +00:00
|
|
|
return;
|
|
|
|
|
|
|
|
// Avoid increasing the max critical pressure in the scheduled region.
|
2013-12-22 00:04:03 +00:00
|
|
|
if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax,
|
|
|
|
Cand.RPDelta.CriticalMax,
|
2015-12-30 11:46:15 +00:00
|
|
|
TryCand, Cand, RegCritical, TRI,
|
|
|
|
DAG->MF))
|
2013-12-22 00:04:03 +00:00
|
|
|
return;
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
// We only compare a subset of features when comparing nodes between
|
|
|
|
// Top and Bottom boundary. Some properties are simply incomparable, in many
|
|
|
|
// other instances we should only override the other boundary if something
|
|
|
|
// is a clear good pick on one boundary. Skip heuristics that are more
|
|
|
|
// "tie-breaking" in nature.
|
|
|
|
bool SameBoundary = Zone != nullptr;
|
|
|
|
if (SameBoundary) {
|
|
|
|
// For loops that are acyclic path limited, aggressively schedule for
|
2017-01-02 19:17:04 +00:00
|
|
|
// latency. Within an single cycle, whenever CurrMOps > 0, allow normal
|
|
|
|
// heuristics to take precedence.
|
2016-07-23 20:41:05 +00:00
|
|
|
if (Rem.IsAcyclicLatencyLimited && !Zone->getCurrMOps() &&
|
|
|
|
tryLatency(TryCand, Cand, *Zone))
|
|
|
|
return;
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
// Prioritize instructions that read unbuffered resources by stall cycles.
|
|
|
|
if (tryLess(Zone->getLatencyStallCycles(TryCand.SU),
|
|
|
|
Zone->getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall))
|
|
|
|
return;
|
|
|
|
}
|
2014-11-24 09:08:18 +00:00
|
|
|
|
2013-04-08 18:41:23 +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.
|
2016-07-23 20:41:05 +00:00
|
|
|
const SUnit *CandNextClusterSU =
|
|
|
|
Cand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
|
|
|
|
const SUnit *TryCandNextClusterSU =
|
|
|
|
TryCand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
|
|
|
|
if (tryGreater(TryCand.SU == TryCandNextClusterSU,
|
|
|
|
Cand.SU == CandNextClusterSU,
|
2013-04-08 18:41:23 +00:00
|
|
|
TryCand, Cand, Cluster))
|
|
|
|
return;
|
2013-06-10 20:36:52 +00:00
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
if (SameBoundary) {
|
|
|
|
// Weak edges are for clustering and other constraints.
|
|
|
|
if (tryLess(getWeakLeft(TryCand.SU, TryCand.AtTop),
|
|
|
|
getWeakLeft(Cand.SU, Cand.AtTop),
|
|
|
|
TryCand, Cand, Weak))
|
|
|
|
return;
|
2013-04-08 18:41:23 +00:00
|
|
|
}
|
2016-07-23 20:41:05 +00:00
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
// Avoid increasing the max pressure of the entire region.
|
|
|
|
if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax,
|
|
|
|
Cand.RPDelta.CurrentMax,
|
2015-12-30 11:46:15 +00:00
|
|
|
TryCand, Cand, RegMax, TRI,
|
|
|
|
DAG->MF))
|
2013-12-22 00:04:03 +00:00
|
|
|
return;
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
if (SameBoundary) {
|
|
|
|
// 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;
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
// Avoid serializing long latency dependence chains.
|
|
|
|
// For acyclic path limited loops, latency was already checked above.
|
|
|
|
if (!RegionPolicy.DisableLatencyHeuristic && TryCand.Policy.ReduceLatency &&
|
|
|
|
!Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, *Zone))
|
|
|
|
return;
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2016-07-23 20:41:05 +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;
|
|
|
|
}
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
/// Pick the best candidate from the queue.
|
2012-08-15 19:34:23 +00:00
|
|
|
///
|
|
|
|
/// 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.
|
2013-12-22 00:04:03 +00:00
|
|
|
void GenericScheduler::pickNodeFromQueue(SchedBoundary &Zone,
|
2016-07-23 20:41:05 +00:00
|
|
|
const CandPolicy &ZonePolicy,
|
2014-11-24 09:08:18 +00:00
|
|
|
const RegPressureTracker &RPTracker,
|
|
|
|
SchedCandidate &Cand) {
|
2012-08-15 19:34:23 +00:00
|
|
|
// getMaxPressureDelta temporarily modifies the tracker.
|
|
|
|
RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
|
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
ReadyQueue &Q = Zone.Available;
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SUnit *SU : Q) {
|
2012-08-15 19:34:23 +00:00
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
SchedCandidate TryCand(ZonePolicy);
|
2017-06-26 20:32:52 +00:00
|
|
|
initCandidate(TryCand, SU, Zone.isTop(), RPTracker, TempTracker);
|
2016-07-23 20:41:05 +00:00
|
|
|
// Pass SchedBoundary only when comparing nodes from the same boundary.
|
|
|
|
SchedBoundary *ZoneArg = Cand.AtTop == TryCand.AtTop ? &Zone : nullptr;
|
|
|
|
tryCandidate(Cand, TryCand, ZoneArg);
|
2012-12-02 13:10:19 +00:00
|
|
|
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);
|
2013-04-08 18:41:23 +00:00
|
|
|
DEBUG(traceCandidate(Cand));
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
/// Pick the best candidate node from either the top or bottom queue.
|
2013-12-22 00:04:03 +00:00
|
|
|
SUnit *GenericScheduler::pickNodeBidirectional(bool &IsTopNode) {
|
2012-08-15 19:34:23 +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.
|
|
|
|
if (SUnit *SU = Bot.pickOnlyChoice()) {
|
|
|
|
IsTopNode = false;
|
2016-07-23 20:41:05 +00:00
|
|
|
tracePick(Only1, false);
|
2012-04-14 13:54:10 +00:00
|
|
|
return SU;
|
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
if (SUnit *SU = Top.pickOnlyChoice()) {
|
|
|
|
IsTopNode = true;
|
2016-07-23 20:41:05 +00:00
|
|
|
tracePick(Only1, true);
|
2012-08-15 19:34:23 +00:00
|
|
|
return SU;
|
|
|
|
}
|
2014-11-24 09:08:18 +00:00
|
|
|
// Set the bottom-up policy based on the state of the current bottom zone and
|
|
|
|
// the instructions outside the zone, including the top zone.
|
2016-07-23 20:41:05 +00:00
|
|
|
CandPolicy BotPolicy;
|
|
|
|
setPolicy(BotPolicy, /*IsPostRA=*/false, Bot, &Top);
|
2014-11-24 09:08:18 +00:00
|
|
|
// Set the top-down policy based on the state of the current top zone and
|
|
|
|
// the instructions outside the zone, including the bottom zone.
|
2016-07-23 20:41:05 +00:00
|
|
|
CandPolicy TopPolicy;
|
|
|
|
setPolicy(TopPolicy, /*IsPostRA=*/false, Top, &Bot);
|
|
|
|
|
|
|
|
// See if BotCand is still valid (because we previously scheduled from Top).
|
|
|
|
DEBUG(dbgs() << "Picking from Bot:\n");
|
|
|
|
if (!BotCand.isValid() || BotCand.SU->isScheduled ||
|
|
|
|
BotCand.Policy != BotPolicy) {
|
|
|
|
BotCand.reset(CandPolicy());
|
|
|
|
pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), BotCand);
|
|
|
|
assert(BotCand.Reason != NoCand && "failed to find the first candidate");
|
|
|
|
} else {
|
|
|
|
DEBUG(traceCandidate(BotCand));
|
|
|
|
#ifndef NDEBUG
|
|
|
|
if (VerifyScheduling) {
|
|
|
|
SchedCandidate TCand;
|
|
|
|
TCand.reset(CandPolicy());
|
|
|
|
pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), TCand);
|
|
|
|
assert(TCand.SU == BotCand.SU &&
|
|
|
|
"Last pick result should correspond to re-picking right now");
|
|
|
|
}
|
|
|
|
#endif
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
2016-07-23 20:41:05 +00:00
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
// Check if the top Q has a better candidate.
|
2016-07-23 20:41:05 +00:00
|
|
|
DEBUG(dbgs() << "Picking from Top:\n");
|
|
|
|
if (!TopCand.isValid() || TopCand.SU->isScheduled ||
|
|
|
|
TopCand.Policy != TopPolicy) {
|
|
|
|
TopCand.reset(CandPolicy());
|
|
|
|
pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TopCand);
|
|
|
|
assert(TopCand.Reason != NoCand && "failed to find the first candidate");
|
|
|
|
} else {
|
|
|
|
DEBUG(traceCandidate(TopCand));
|
|
|
|
#ifndef NDEBUG
|
|
|
|
if (VerifyScheduling) {
|
|
|
|
SchedCandidate TCand;
|
|
|
|
TCand.reset(CandPolicy());
|
|
|
|
pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TCand);
|
|
|
|
assert(TCand.SU == TopCand.SU &&
|
|
|
|
"Last pick result should correspond to re-picking right now");
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
|
2016-07-23 20:41:05 +00:00
|
|
|
// Pick best from BotCand and TopCand.
|
|
|
|
assert(BotCand.isValid());
|
|
|
|
assert(TopCand.isValid());
|
|
|
|
SchedCandidate Cand = BotCand;
|
|
|
|
TopCand.Reason = NoCand;
|
|
|
|
tryCandidate(Cand, TopCand, nullptr);
|
|
|
|
if (TopCand.Reason != NoCand) {
|
|
|
|
Cand.setBest(TopCand);
|
|
|
|
DEBUG(traceCandidate(Cand));
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
2016-07-23 20:41:05 +00:00
|
|
|
|
|
|
|
IsTopNode = Cand.AtTop;
|
|
|
|
tracePick(Cand);
|
|
|
|
return Cand.SU;
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
2012-04-14 13:54:10 +00:00
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
/// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
|
2013-12-22 00:04:03 +00:00
|
|
|
SUnit *GenericScheduler::pickNode(bool &IsTopNode) {
|
2012-08-15 19:34:23 +00:00
|
|
|
if (DAG->top() == DAG->bottom()) {
|
|
|
|
assert(Top.Available.empty() && Top.Pending.empty() &&
|
|
|
|
Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
|
2014-11-24 09:08:18 +00:00
|
|
|
return nullptr;
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2012-08-15 19:34:23 +00:00
|
|
|
SUnit *SU;
|
2012-12-02 13:10:19 +00:00
|
|
|
do {
|
2013-12-22 00:04:03 +00:00
|
|
|
if (RegionPolicy.OnlyTopDown) {
|
2012-12-02 13:10:19 +00:00
|
|
|
SU = Top.pickOnlyChoice();
|
|
|
|
if (!SU) {
|
|
|
|
CandPolicy NoPolicy;
|
2016-07-23 20:41:05 +00:00
|
|
|
TopCand.reset(NoPolicy);
|
|
|
|
pickNodeFromQueue(Top, NoPolicy, DAG->getTopRPTracker(), TopCand);
|
2013-12-22 00:04:03 +00:00
|
|
|
assert(TopCand.Reason != NoCand && "failed to find a candidate");
|
2016-07-23 20:41:05 +00:00
|
|
|
tracePick(TopCand);
|
2012-12-02 13:10:19 +00:00
|
|
|
SU = TopCand.SU;
|
|
|
|
}
|
|
|
|
IsTopNode = true;
|
2016-07-23 20:41:05 +00:00
|
|
|
} else if (RegionPolicy.OnlyBottomUp) {
|
2012-12-02 13:10:19 +00:00
|
|
|
SU = Bot.pickOnlyChoice();
|
|
|
|
if (!SU) {
|
|
|
|
CandPolicy NoPolicy;
|
2016-07-23 20:41:05 +00:00
|
|
|
BotCand.reset(NoPolicy);
|
|
|
|
pickNodeFromQueue(Bot, NoPolicy, DAG->getBotRPTracker(), BotCand);
|
2013-12-22 00:04:03 +00:00
|
|
|
assert(BotCand.Reason != NoCand && "failed to find a candidate");
|
2016-07-23 20:41:05 +00:00
|
|
|
tracePick(BotCand);
|
2012-12-02 13:10:19 +00:00
|
|
|
SU = BotCand.SU;
|
|
|
|
}
|
|
|
|
IsTopNode = false;
|
2016-07-23 20:41:05 +00:00
|
|
|
} else {
|
2012-12-02 13:10:19 +00:00
|
|
|
SU = pickNodeBidirectional(IsTopNode);
|
|
|
|
}
|
|
|
|
} while (SU->isScheduled);
|
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
if (SU->isTopReady())
|
|
|
|
Top.removeReady(SU);
|
|
|
|
if (SU->isBottomReady())
|
|
|
|
Bot.removeReady(SU);
|
|
|
|
|
2013-06-10 20:36:52 +00:00
|
|
|
DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
|
2012-08-15 19:34:23 +00:00
|
|
|
return SU;
|
|
|
|
}
|
|
|
|
|
2013-12-22 00:04:03 +00:00
|
|
|
void GenericScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) {
|
2013-06-10 20:36:52 +00:00
|
|
|
MachineBasicBlock::iterator InsertPos = SU->getInstr();
|
|
|
|
if (!isTop)
|
|
|
|
++InsertPos;
|
|
|
|
SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs;
|
|
|
|
|
|
|
|
// Find already scheduled copies with a single physreg dependence and move
|
|
|
|
// them just above the scheduled instruction.
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SDep &Dep : Deps) {
|
|
|
|
if (Dep.getKind() != SDep::Data || !TRI->isPhysicalRegister(Dep.getReg()))
|
2013-06-10 20:36:52 +00:00
|
|
|
continue;
|
2017-06-26 20:32:52 +00:00
|
|
|
SUnit *DepSU = Dep.getSUnit();
|
2013-06-10 20:36:52 +00:00
|
|
|
if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1)
|
|
|
|
continue;
|
|
|
|
MachineInstr *Copy = DepSU->getInstr();
|
|
|
|
if (!Copy->isCopy())
|
|
|
|
continue;
|
|
|
|
DEBUG(dbgs() << " Rescheduling physreg copy ";
|
2017-06-26 20:32:52 +00:00
|
|
|
Dep.getSUnit()->dump(DAG));
|
2013-06-10 20:36:52 +00:00
|
|
|
DAG->moveInstruction(Copy, InsertPos);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-08-15 19:34:23 +00:00
|
|
|
/// Update the scheduler's state after scheduling a node. This is the same node
|
2014-11-24 09:08:18 +00:00
|
|
|
/// that was just returned by pickNode(). However, ScheduleDAGMILive needs to
|
|
|
|
/// update it's state based on the current cycle before MachineSchedStrategy
|
|
|
|
/// does.
|
2013-06-10 20:36:52 +00:00
|
|
|
///
|
|
|
|
/// FIXME: Eventually, we may bundle physreg copies rather than rescheduling
|
|
|
|
/// them here. See comments in biasPhysRegCopy.
|
2013-12-22 00:04:03 +00:00
|
|
|
void GenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
|
2012-08-15 19:34:23 +00:00
|
|
|
if (IsTopNode) {
|
2014-11-24 09:08:18 +00:00
|
|
|
SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
|
2012-08-15 19:34:23 +00:00
|
|
|
Top.bumpNode(SU);
|
2013-06-10 20:36:52 +00:00
|
|
|
if (SU->hasPhysRegUses)
|
|
|
|
reschedulePhysRegCopies(SU, true);
|
2016-07-23 20:41:05 +00:00
|
|
|
} else {
|
2014-11-24 09:08:18 +00:00
|
|
|
SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.getCurrCycle());
|
2012-08-15 19:34:23 +00:00
|
|
|
Bot.bumpNode(SU);
|
2013-06-10 20:36:52 +00:00
|
|
|
if (SU->hasPhysRegDefs)
|
|
|
|
reschedulePhysRegCopies(SU, false);
|
2012-08-15 19:34:23 +00:00
|
|
|
}
|
|
|
|
}
|
2012-04-14 13:54:10 +00:00
|
|
|
|
|
|
|
/// Create the standard converging machine scheduler. This will be used as the
|
|
|
|
/// default scheduler if the target does not set a default.
|
2017-01-02 19:17:04 +00:00
|
|
|
ScheduleDAGMILive *llvm::createGenericSchedLive(MachineSchedContext *C) {
|
2017-04-16 16:01:22 +00:00
|
|
|
ScheduleDAGMILive *DAG =
|
|
|
|
new ScheduleDAGMILive(C, llvm::make_unique<GenericScheduler>(C));
|
2013-04-08 18:41:23 +00:00
|
|
|
// Register DAG post-processors.
|
2013-06-10 20:36:52 +00:00
|
|
|
//
|
|
|
|
// FIXME: extend the mutation API to allow earlier mutations to instantiate
|
|
|
|
// data and pass it to later mutations. Have a single mutation that gathers
|
|
|
|
// the interesting nodes in one pass.
|
2017-01-02 19:17:04 +00:00
|
|
|
DAG->addMutation(createCopyConstrainDAGMutation(DAG->TII, DAG->TRI));
|
2013-04-08 18:41:23 +00:00
|
|
|
return DAG;
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2014-11-24 09:08:18 +00:00
|
|
|
|
2017-01-02 19:17:04 +00:00
|
|
|
static ScheduleDAGInstrs *createConveringSched(MachineSchedContext *C) {
|
|
|
|
return createGenericSchedLive(C);
|
|
|
|
}
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
static MachineSchedRegistry
|
2013-12-22 00:04:03 +00:00
|
|
|
GenericSchedRegistry("converge", "Standard converging scheduler.",
|
2017-01-02 19:17:04 +00:00
|
|
|
createConveringSched);
|
2014-11-24 09:08:18 +00:00
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// PostGenericScheduler - Generic PostRA implementation of MachineSchedStrategy.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
void PostGenericScheduler::initialize(ScheduleDAGMI *Dag) {
|
|
|
|
DAG = Dag;
|
|
|
|
SchedModel = DAG->getSchedModel();
|
|
|
|
TRI = DAG->TRI;
|
|
|
|
|
|
|
|
Rem.init(DAG, SchedModel);
|
|
|
|
Top.init(DAG, SchedModel, &Rem);
|
|
|
|
BotRoots.clear();
|
|
|
|
|
|
|
|
// Initialize the HazardRecognizers. If itineraries don't exist, are empty,
|
|
|
|
// or are disabled, then these HazardRecs will be disabled.
|
|
|
|
const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
|
|
|
|
if (!Top.HazardRec) {
|
|
|
|
Top.HazardRec =
|
2015-01-18 16:17:27 +00:00
|
|
|
DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer(
|
|
|
|
Itin, DAG);
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void PostGenericScheduler::registerRoots() {
|
|
|
|
Rem.CriticalPath = DAG->ExitSU.getDepth();
|
|
|
|
|
|
|
|
// Some roots may not feed into ExitSU. Check all of them in case.
|
2017-06-26 20:32:52 +00:00
|
|
|
for (const SUnit *SU : BotRoots) {
|
|
|
|
if (SU->getDepth() > Rem.CriticalPath)
|
|
|
|
Rem.CriticalPath = SU->getDepth();
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
2015-01-18 16:17:27 +00:00
|
|
|
DEBUG(dbgs() << "Critical Path: (PGS-RR) " << Rem.CriticalPath << '\n');
|
|
|
|
if (DumpCriticalPathLength) {
|
|
|
|
errs() << "Critical Path(PGS-RR ): " << Rem.CriticalPath << " \n";
|
|
|
|
}
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Apply a set of heursitics to a new candidate for PostRA scheduling.
|
|
|
|
///
|
|
|
|
/// \param Cand provides the policy and current best candidate.
|
|
|
|
/// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
|
|
|
|
void PostGenericScheduler::tryCandidate(SchedCandidate &Cand,
|
|
|
|
SchedCandidate &TryCand) {
|
|
|
|
// Initialize the candidate if needed.
|
|
|
|
if (!Cand.isValid()) {
|
|
|
|
TryCand.Reason = NodeOrder;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Prioritize instructions that read unbuffered resources by stall cycles.
|
|
|
|
if (tryLess(Top.getLatencyStallCycles(TryCand.SU),
|
|
|
|
Top.getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall))
|
|
|
|
return;
|
2017-05-29 16:25:25 +00:00
|
|
|
|
|
|
|
// Keep clustered nodes together.
|
|
|
|
if (tryGreater(TryCand.SU == DAG->getNextClusterSucc(),
|
|
|
|
Cand.SU == DAG->getNextClusterSucc(),
|
|
|
|
TryCand, Cand, Cluster))
|
|
|
|
return;
|
2014-11-24 09:08:18 +00:00
|
|
|
|
|
|
|
// Avoid critical resource consumption and balance the schedule.
|
|
|
|
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 && tryLatency(TryCand, Cand, Top)) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Fall through to original instruction order.
|
|
|
|
if (TryCand.SU->NodeNum < Cand.SU->NodeNum)
|
|
|
|
TryCand.Reason = NodeOrder;
|
|
|
|
}
|
|
|
|
|
|
|
|
void PostGenericScheduler::pickNodeFromQueue(SchedCandidate &Cand) {
|
|
|
|
ReadyQueue &Q = Top.Available;
|
2017-06-26 20:32:52 +00:00
|
|
|
for (SUnit *SU : Q) {
|
2014-11-24 09:08:18 +00:00
|
|
|
SchedCandidate TryCand(Cand.Policy);
|
2017-06-26 20:32:52 +00:00
|
|
|
TryCand.SU = SU;
|
2016-07-23 20:41:05 +00:00
|
|
|
TryCand.AtTop = true;
|
2014-11-24 09:08:18 +00:00
|
|
|
TryCand.initResourceDelta(DAG, SchedModel);
|
|
|
|
tryCandidate(Cand, TryCand);
|
|
|
|
if (TryCand.Reason != NoCand) {
|
|
|
|
Cand.setBest(TryCand);
|
|
|
|
DEBUG(traceCandidate(Cand));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Pick the next node to schedule.
|
|
|
|
SUnit *PostGenericScheduler::pickNode(bool &IsTopNode) {
|
|
|
|
if (DAG->top() == DAG->bottom()) {
|
|
|
|
assert(Top.Available.empty() && Top.Pending.empty() && "ReadyQ garbage");
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
SUnit *SU;
|
|
|
|
do {
|
|
|
|
SU = Top.pickOnlyChoice();
|
2016-07-23 20:41:05 +00:00
|
|
|
if (SU) {
|
|
|
|
tracePick(Only1, true);
|
|
|
|
} else {
|
2014-11-24 09:08:18 +00:00
|
|
|
CandPolicy NoPolicy;
|
|
|
|
SchedCandidate TopCand(NoPolicy);
|
|
|
|
// Set the top-down policy based on the state of the current top zone and
|
|
|
|
// the instructions outside the zone, including the bottom zone.
|
|
|
|
setPolicy(TopCand.Policy, /*IsPostRA=*/true, Top, nullptr);
|
|
|
|
pickNodeFromQueue(TopCand);
|
|
|
|
assert(TopCand.Reason != NoCand && "failed to find a candidate");
|
2016-07-23 20:41:05 +00:00
|
|
|
tracePick(TopCand);
|
2014-11-24 09:08:18 +00:00
|
|
|
SU = TopCand.SU;
|
|
|
|
}
|
|
|
|
} while (SU->isScheduled);
|
|
|
|
|
|
|
|
IsTopNode = true;
|
|
|
|
Top.removeReady(SU);
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
|
|
|
|
return SU;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Called after ScheduleDAGMI has scheduled an instruction and updated
|
|
|
|
/// scheduled/remaining flags in the DAG nodes.
|
|
|
|
void PostGenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
|
|
|
|
SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
|
|
|
|
Top.bumpNode(SU);
|
|
|
|
}
|
|
|
|
|
2017-01-02 19:17:04 +00:00
|
|
|
ScheduleDAGMI *llvm::createGenericSchedPostRA(MachineSchedContext *C) {
|
2017-04-16 16:01:22 +00:00
|
|
|
return new ScheduleDAGMI(C, llvm::make_unique<PostGenericScheduler>(C),
|
2017-01-02 19:17:04 +00:00
|
|
|
/*RemoveKillFlags=*/true);
|
2014-11-24 09:08:18 +00:00
|
|
|
}
|
2012-04-14 13:54:10 +00:00
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// ILP Scheduler. Currently for experimental analysis of heuristics.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
/// \brief Order nodes by the ILP metric.
|
|
|
|
struct ILPOrder {
|
2017-04-16 16:01:22 +00:00
|
|
|
const SchedDFSResult *DFSResult = nullptr;
|
|
|
|
const BitVector *ScheduledTrees = nullptr;
|
2012-12-02 13:10:19 +00:00
|
|
|
bool MaximizeILP;
|
|
|
|
|
2017-04-16 16:01:22 +00:00
|
|
|
ILPOrder(bool MaxILP) : MaximizeILP(MaxILP) {}
|
2012-12-02 13:10:19 +00:00
|
|
|
|
|
|
|
/// \brief Apply a less-than relation on node priority.
|
2013-04-08 18:41:23 +00:00
|
|
|
///
|
|
|
|
/// (Return true if A comes after B in the Q.)
|
2012-12-02 13:10:19 +00:00
|
|
|
bool operator()(const SUnit *A, const SUnit *B) const {
|
2013-04-08 18:41:23 +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);
|
|
|
|
}
|
|
|
|
}
|
2012-12-02 13:10:19 +00:00
|
|
|
if (MaximizeILP)
|
2013-04-08 18:41:23 +00:00
|
|
|
return DFSResult->getILP(A) < DFSResult->getILP(B);
|
2012-12-02 13:10:19 +00:00
|
|
|
else
|
2013-04-08 18:41:23 +00:00
|
|
|
return DFSResult->getILP(A) > DFSResult->getILP(B);
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
/// \brief Schedule based on the ILP metric.
|
|
|
|
class ILPScheduler : public MachineSchedStrategy {
|
2017-04-16 16:01:22 +00:00
|
|
|
ScheduleDAGMILive *DAG = nullptr;
|
2012-12-02 13:10:19 +00:00
|
|
|
ILPOrder Cmp;
|
|
|
|
|
|
|
|
std::vector<SUnit*> ReadyQ;
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
public:
|
2017-04-16 16:01:22 +00:00
|
|
|
ILPScheduler(bool MaximizeILP) : Cmp(MaximizeILP) {}
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void initialize(ScheduleDAGMI *dag) override {
|
|
|
|
assert(dag->hasVRegLiveness() && "ILPScheduler needs vreg liveness");
|
|
|
|
DAG = static_cast<ScheduleDAGMILive*>(dag);
|
2013-04-08 18:41:23 +00:00
|
|
|
DAG->computeDFSResult();
|
|
|
|
Cmp.DFSResult = DAG->getDFSResult();
|
|
|
|
Cmp.ScheduledTrees = &DAG->getScheduledTrees();
|
2012-12-02 13:10:19 +00:00
|
|
|
ReadyQ.clear();
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void registerRoots() override {
|
2013-04-08 18:41:23 +00:00
|
|
|
// Restore the heap in ReadyQ with the updated DFS results.
|
|
|
|
std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Implement MachineSchedStrategy interface.
|
|
|
|
/// -----------------------------------------
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
/// Callback to select the highest priority node from the ready Q.
|
2014-11-24 09:08:18 +00:00
|
|
|
SUnit *pickNode(bool &IsTopNode) override {
|
|
|
|
if (ReadyQ.empty()) return nullptr;
|
2013-04-08 18:41:23 +00:00
|
|
|
std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
2012-12-02 13:10:19 +00:00
|
|
|
SUnit *SU = ReadyQ.back();
|
|
|
|
ReadyQ.pop_back();
|
|
|
|
IsTopNode = false;
|
2013-06-10 20:36:52 +00:00
|
|
|
DEBUG(dbgs() << "Pick node " << "SU(" << SU->NodeNum << ") "
|
2013-04-08 18:41:23 +00:00
|
|
|
<< " ILP: " << DAG->getDFSResult()->getILP(SU)
|
|
|
|
<< " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
|
|
|
|
<< DAG->getDFSResult()->getSubtreeLevel(
|
2013-06-10 20:36:52 +00:00
|
|
|
DAG->getDFSResult()->getSubtreeID(SU)) << '\n'
|
|
|
|
<< "Scheduling " << *SU->getInstr());
|
2012-12-02 13:10:19 +00:00
|
|
|
return SU;
|
|
|
|
}
|
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
/// \brief Scheduler callback to notify that a new subtree is scheduled.
|
2014-11-24 09:08:18 +00:00
|
|
|
void scheduleTree(unsigned SubtreeID) override {
|
2013-04-08 18:41:23 +00:00
|
|
|
std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Callback after a node is scheduled. Mark a newly scheduled tree, notify
|
|
|
|
/// DFSResults, and resort the priority Q.
|
2014-11-24 09:08:18 +00:00
|
|
|
void schedNode(SUnit *SU, bool IsTopNode) override {
|
2013-04-08 18:41:23 +00:00
|
|
|
assert(!IsTopNode && "SchedDFSResult needs bottom-up");
|
|
|
|
}
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void releaseTopNode(SUnit *) override { /*only called for top roots*/ }
|
2012-12-02 13:10:19 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void releaseBottomNode(SUnit *SU) override {
|
2012-12-02 13:10:19 +00:00
|
|
|
ReadyQ.push_back(SU);
|
|
|
|
std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
|
|
|
}
|
|
|
|
};
|
2017-04-16 16:01:22 +00:00
|
|
|
|
|
|
|
} // end anonymous namespace
|
2012-12-02 13:10:19 +00:00
|
|
|
|
|
|
|
static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
|
2017-04-16 16:01:22 +00:00
|
|
|
return new ScheduleDAGMILive(C, llvm::make_unique<ILPScheduler>(true));
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
|
|
|
static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
|
2017-04-16 16:01:22 +00:00
|
|
|
return new ScheduleDAGMILive(C, llvm::make_unique<ILPScheduler>(false));
|
2012-12-02 13:10:19 +00:00
|
|
|
}
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2012-12-02 13:10:19 +00:00
|
|
|
static MachineSchedRegistry ILPMaxRegistry(
|
|
|
|
"ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
|
|
|
|
static MachineSchedRegistry ILPMinRegistry(
|
|
|
|
"ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
|
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Machine Instruction Shuffler for Correctness Testing
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
namespace {
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2012-04-14 13:54:10 +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 {
|
|
|
|
bool operator()(SUnit *A, SUnit *B) const {
|
|
|
|
if (IsReverse)
|
|
|
|
return A->NodeNum > B->NodeNum;
|
|
|
|
else
|
|
|
|
return A->NodeNum < B->NodeNum;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
/// Reorder instructions as much as possible.
|
|
|
|
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.
|
2017-04-16 16:01:22 +00:00
|
|
|
PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false>>
|
2012-04-14 13:54:10 +00:00
|
|
|
TopQ;
|
2017-12-18 20:10:56 +00:00
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
// When scheduling bottom-up, use greater-than as the queue priority.
|
2017-04-16 16:01:22 +00:00
|
|
|
PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true>>
|
2012-04-14 13:54:10 +00:00
|
|
|
BottomQ;
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
public:
|
|
|
|
InstructionShuffler(bool alternate, bool topdown)
|
|
|
|
: IsAlternating(alternate), IsTopDown(topdown) {}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void initialize(ScheduleDAGMI*) override {
|
2012-04-14 13:54:10 +00:00
|
|
|
TopQ.clear();
|
|
|
|
BottomQ.clear();
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Implement MachineSchedStrategy interface.
|
|
|
|
/// -----------------------------------------
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
SUnit *pickNode(bool &IsTopNode) override {
|
2012-04-14 13:54:10 +00:00
|
|
|
SUnit *SU;
|
|
|
|
if (IsTopDown) {
|
|
|
|
do {
|
2014-11-24 09:08:18 +00:00
|
|
|
if (TopQ.empty()) return nullptr;
|
2012-04-14 13:54:10 +00:00
|
|
|
SU = TopQ.top();
|
|
|
|
TopQ.pop();
|
|
|
|
} while (SU->isScheduled);
|
|
|
|
IsTopNode = true;
|
2016-07-23 20:41:05 +00:00
|
|
|
} else {
|
2012-04-14 13:54:10 +00:00
|
|
|
do {
|
2014-11-24 09:08:18 +00:00
|
|
|
if (BottomQ.empty()) return nullptr;
|
2012-04-14 13:54:10 +00:00
|
|
|
SU = BottomQ.top();
|
|
|
|
BottomQ.pop();
|
|
|
|
} while (SU->isScheduled);
|
|
|
|
IsTopNode = false;
|
|
|
|
}
|
|
|
|
if (IsAlternating)
|
|
|
|
IsTopDown = !IsTopDown;
|
|
|
|
return SU;
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void schedNode(SUnit *SU, bool IsTopNode) override {}
|
2012-08-15 19:34:23 +00:00
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
void releaseTopNode(SUnit *SU) override {
|
2012-04-14 13:54:10 +00:00
|
|
|
TopQ.push(SU);
|
|
|
|
}
|
2014-11-24 09:08:18 +00:00
|
|
|
void releaseBottomNode(SUnit *SU) override {
|
2012-04-14 13:54:10 +00:00
|
|
|
BottomQ.push(SU);
|
|
|
|
}
|
|
|
|
};
|
2017-04-16 16:01:22 +00:00
|
|
|
|
|
|
|
} // end anonymous namespace
|
2012-04-14 13:54:10 +00:00
|
|
|
|
|
|
|
static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
|
|
|
|
bool Alternate = !ForceTopDown && !ForceBottomUp;
|
|
|
|
bool TopDown = !ForceBottomUp;
|
|
|
|
assert((TopDown || !ForceTopDown) &&
|
|
|
|
"-misched-topdown incompatible with -misched-bottomup");
|
2017-04-16 16:01:22 +00:00
|
|
|
return new ScheduleDAGMILive(
|
|
|
|
C, llvm::make_unique<InstructionShuffler>(Alternate, TopDown));
|
2012-04-14 13:54:10 +00:00
|
|
|
}
|
2017-04-16 16:01:22 +00:00
|
|
|
|
2012-04-14 13:54:10 +00:00
|
|
|
static MachineSchedRegistry ShufflerRegistry(
|
|
|
|
"shuffle", "Shuffle machine instructions alternating directions",
|
|
|
|
createInstructionShuffler);
|
|
|
|
#endif // !NDEBUG
|
2013-04-08 18:41:23 +00:00
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
2014-11-24 09:08:18 +00:00
|
|
|
// GraphWriter support for ScheduleDAGMILive.
|
2013-04-08 18:41:23 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
namespace llvm {
|
|
|
|
|
|
|
|
template<> struct GraphTraits<
|
|
|
|
ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
|
|
|
|
|
|
|
|
template<>
|
|
|
|
struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
|
2017-04-16 16:01:22 +00:00
|
|
|
DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {}
|
2013-04-08 18:41:23 +00:00
|
|
|
|
|
|
|
static std::string getGraphName(const ScheduleDAG *G) {
|
|
|
|
return G->MF.getName();
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool renderGraphFromBottomUp() {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool isNodeHidden(const SUnit *Node) {
|
2015-12-30 11:46:15 +00:00
|
|
|
if (ViewMISchedCutoff == 0)
|
|
|
|
return false;
|
|
|
|
return (Node->Preds.size() > ViewMISchedCutoff
|
|
|
|
|| Node->Succs.size() > ViewMISchedCutoff);
|
2013-04-08 18:41:23 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// If you want to override the dot attributes printed for a particular
|
|
|
|
/// edge, override this method.
|
|
|
|
static std::string getEdgeAttributes(const SUnit *Node,
|
|
|
|
SUnitIterator EI,
|
|
|
|
const ScheduleDAG *Graph) {
|
|
|
|
if (EI.isArtificialDep())
|
|
|
|
return "color=cyan,style=dashed";
|
|
|
|
if (EI.isCtrlDep())
|
|
|
|
return "color=blue,style=dashed";
|
|
|
|
return "";
|
|
|
|
}
|
|
|
|
|
|
|
|
static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
|
|
|
|
std::string Str;
|
|
|
|
raw_string_ostream SS(Str);
|
2014-11-24 09:08:18 +00:00
|
|
|
const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
|
|
|
|
const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
|
|
|
|
static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr;
|
2013-12-22 00:04:03 +00:00
|
|
|
SS << "SU:" << SU->NodeNum;
|
|
|
|
if (DFS)
|
|
|
|
SS << " I:" << DFS->getNumInstrs(SU);
|
2013-04-08 18:41:23 +00:00
|
|
|
return SS.str();
|
|
|
|
}
|
2017-12-18 20:10:56 +00:00
|
|
|
|
2013-04-08 18:41:23 +00:00
|
|
|
static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
|
|
|
|
return G->getGraphNodeLabel(SU);
|
|
|
|
}
|
|
|
|
|
2014-11-24 09:08:18 +00:00
|
|
|
static std::string getNodeAttributes(const SUnit *N, const ScheduleDAG *G) {
|
2013-04-08 18:41:23 +00:00
|
|
|
std::string Str("shape=Mrecord");
|
2014-11-24 09:08:18 +00:00
|
|
|
const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
|
|
|
|
const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
|
|
|
|
static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr;
|
2013-04-08 18:41:23 +00:00
|
|
|
if (DFS) {
|
|
|
|
Str += ",style=filled,fillcolor=\"#";
|
|
|
|
Str += DOT::getColorString(DFS->getSubtreeID(N));
|
|
|
|
Str += '"';
|
|
|
|
}
|
|
|
|
return Str;
|
|
|
|
}
|
|
|
|
};
|
2017-04-16 16:01:22 +00:00
|
|
|
|
|
|
|
} // end namespace llvm
|
2013-04-08 18:41:23 +00:00
|
|
|
#endif // NDEBUG
|
|
|
|
|
|
|
|
/// viewGraph - Pop up a ghostview window with the reachable parts of the DAG
|
|
|
|
/// rendered using 'dot'.
|
|
|
|
void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) {
|
|
|
|
#ifndef NDEBUG
|
|
|
|
ViewGraph(this, Name, false, Title);
|
|
|
|
#else
|
|
|
|
errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on "
|
|
|
|
<< "systems with Graphviz or gv!\n";
|
|
|
|
#endif // NDEBUG
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Out-of-line implementation with no arguments is handy for gdb.
|
|
|
|
void ScheduleDAGMI::viewGraph() {
|
|
|
|
viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName());
|
|
|
|
}
|