2012-05-17 22:37:09 +00:00
|
|
|
//===- MachineScheduler.cpp - Machine Instruction Scheduler ---------------===//
|
2012-01-13 06:30:30 +00:00
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// MachineScheduler schedules machine instructions after phi elimination. It
|
|
|
|
// preserves LiveIntervals so it can be invoked before register allocation.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2012-03-08 01:41:12 +00:00
|
|
|
#include "llvm/CodeGen/MachineScheduler.h"
|
2012-12-03 16:50:05 +00:00
|
|
|
#include "llvm/ADT/PriorityQueue.h"
|
|
|
|
#include "llvm/Analysis/AliasAnalysis.h"
|
|
|
|
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
|
2013-03-10 13:11:23 +00:00
|
|
|
#include "llvm/CodeGen/MachineDominators.h"
|
|
|
|
#include "llvm/CodeGen/MachineLoopInfo.h"
|
2013-06-21 18:32:58 +00:00
|
|
|
#include "llvm/CodeGen/MachineRegisterInfo.h"
|
2012-01-13 06:30:30 +00:00
|
|
|
#include "llvm/CodeGen/Passes.h"
|
2012-06-06 20:29:31 +00:00
|
|
|
#include "llvm/CodeGen/RegisterClassInfo.h"
|
2012-11-28 05:13:24 +00:00
|
|
|
#include "llvm/CodeGen/ScheduleDFS.h"
|
2012-05-24 22:11:09 +00:00
|
|
|
#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
|
2012-01-13 06:30:30 +00:00
|
|
|
#include "llvm/Support/CommandLine.h"
|
|
|
|
#include "llvm/Support/Debug.h"
|
|
|
|
#include "llvm/Support/ErrorHandling.h"
|
2013-01-25 07:45:29 +00:00
|
|
|
#include "llvm/Support/GraphWriter.h"
|
2012-01-13 06:30:30 +00:00
|
|
|
#include "llvm/Support/raw_ostream.h"
|
2013-06-14 00:00:13 +00:00
|
|
|
#include "llvm/Target/TargetInstrInfo.h"
|
2012-01-17 06:55:07 +00:00
|
|
|
#include <queue>
|
|
|
|
|
2012-01-13 06:30:30 +00:00
|
|
|
using namespace llvm;
|
|
|
|
|
2014-04-22 02:02:50 +00:00
|
|
|
#define DEBUG_TYPE "misched"
|
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
namespace llvm {
|
|
|
|
cl::opt<bool> ForceTopDown("misched-topdown", cl::Hidden,
|
|
|
|
cl::desc("Force top-down list scheduling"));
|
|
|
|
cl::opt<bool> ForceBottomUp("misched-bottomup", cl::Hidden,
|
|
|
|
cl::desc("Force bottom-up list scheduling"));
|
|
|
|
}
|
2012-03-14 04:00:41 +00:00
|
|
|
|
2012-03-07 00:18:25 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
static cl::opt<bool> ViewMISchedDAGs("view-misched-dags", cl::Hidden,
|
|
|
|
cl::desc("Pop up a window to show MISched dags after they are processed"));
|
2012-03-19 18:38:38 +00:00
|
|
|
|
|
|
|
static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
|
|
|
|
cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
|
2013-12-28 21:57:02 +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,
|
|
|
|
cl::desc("Only schedule this MBB#"));
|
2012-03-07 00:18:25 +00:00
|
|
|
#else
|
|
|
|
static bool ViewMISchedDAGs = false;
|
|
|
|
#endif // NDEBUG
|
|
|
|
|
2013-09-04 20:59:59 +00:00
|
|
|
static cl::opt<bool> EnableRegPressure("misched-regpressure", cl::Hidden,
|
|
|
|
cl::desc("Enable register pressure scheduling."), cl::init(true));
|
|
|
|
|
2013-08-23 17:48:43 +00:00
|
|
|
static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden,
|
2013-09-09 23:31:14 +00:00
|
|
|
cl::desc("Enable cyclic critical path analysis."), cl::init(true));
|
2013-08-23 17:48:43 +00:00
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
|
2012-11-13 08:47:29 +00:00
|
|
|
cl::desc("Enable load clustering."), cl::init(true));
|
2012-11-12 19:40:10 +00:00
|
|
|
|
2012-11-12 19:52:20 +00:00
|
|
|
// Experimental heuristics
|
|
|
|
static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
|
2012-11-13 08:47:29 +00:00
|
|
|
cl::desc("Enable scheduling for macro fusion."), cl::init(true));
|
2012-11-12 19:52:20 +00:00
|
|
|
|
2013-03-08 05:40:34 +00:00
|
|
|
static cl::opt<bool> VerifyScheduling("verify-misched", cl::Hidden,
|
|
|
|
cl::desc("Verify machine instrs before and after machine scheduling"));
|
|
|
|
|
2013-01-25 04:01:04 +00:00
|
|
|
// DAG subtrees must have at least this many nodes.
|
|
|
|
static const unsigned MinSubtreeSize = 8;
|
|
|
|
|
2013-11-19 00:57:56 +00:00
|
|
|
// Pin the vtables to this file.
|
|
|
|
void MachineSchedStrategy::anchor() {}
|
|
|
|
void ScheduleDAGMutation::anchor() {}
|
|
|
|
|
2012-01-14 02:17:06 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Machine Instruction Scheduling Pass and Registry
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2012-04-24 20:36:19 +00:00
|
|
|
MachineSchedContext::MachineSchedContext():
|
2014-04-14 00:51:57 +00:00
|
|
|
MF(nullptr), MLI(nullptr), MDT(nullptr), PassConfig(nullptr), AA(nullptr), LIS(nullptr) {
|
2012-04-24 20:36:19 +00:00
|
|
|
RegClassInfo = new RegisterClassInfo();
|
|
|
|
}
|
|
|
|
|
|
|
|
MachineSchedContext::~MachineSchedContext() {
|
|
|
|
delete RegClassInfo;
|
|
|
|
}
|
|
|
|
|
2012-01-13 06:30:30 +00:00
|
|
|
namespace {
|
2013-12-28 21:56:47 +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) {}
|
|
|
|
|
2014-04-14 00:51:57 +00:00
|
|
|
void print(raw_ostream &O, const Module* = nullptr) const override;
|
2013-12-28 21:56:47 +00:00
|
|
|
|
|
|
|
protected:
|
|
|
|
void scheduleRegions(ScheduleDAGInstrs &Scheduler);
|
|
|
|
};
|
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
/// MachineScheduler runs after coalescing and before register allocation.
|
2013-12-28 21:56:47 +00:00
|
|
|
class MachineScheduler : public MachineSchedulerBase {
|
2012-01-13 06:30:30 +00:00
|
|
|
public:
|
2012-01-17 06:55:03 +00:00
|
|
|
MachineScheduler();
|
2012-01-13 06:30:30 +00:00
|
|
|
|
2014-03-07 09:26:03 +00:00
|
|
|
void getAnalysisUsage(AnalysisUsage &AU) const override;
|
2012-01-13 06:30:30 +00:00
|
|
|
|
2014-03-07 09:26:03 +00:00
|
|
|
bool runOnMachineFunction(MachineFunction&) override;
|
2012-01-13 06:30:30 +00:00
|
|
|
|
|
|
|
static char ID; // Class identification, replacement for typeinfo
|
2013-09-20 05:14:41 +00:00
|
|
|
|
|
|
|
protected:
|
|
|
|
ScheduleDAGInstrs *createMachineScheduler();
|
2012-01-13 06:30:30 +00:00
|
|
|
};
|
2013-12-28 21:56:51 +00:00
|
|
|
|
|
|
|
/// PostMachineScheduler runs after shortly before code emission.
|
|
|
|
class PostMachineScheduler : public MachineSchedulerBase {
|
|
|
|
public:
|
|
|
|
PostMachineScheduler();
|
|
|
|
|
2014-03-07 09:26:03 +00:00
|
|
|
void getAnalysisUsage(AnalysisUsage &AU) const override;
|
2013-12-28 21:56:51 +00:00
|
|
|
|
2014-03-07 09:26:03 +00:00
|
|
|
bool runOnMachineFunction(MachineFunction&) override;
|
2013-12-28 21:56:51 +00:00
|
|
|
|
|
|
|
static char ID; // Class identification, replacement for typeinfo
|
|
|
|
|
|
|
|
protected:
|
|
|
|
ScheduleDAGInstrs *createPostMachineScheduler();
|
|
|
|
};
|
2012-01-13 06:30:30 +00:00
|
|
|
} // namespace
|
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
char MachineScheduler::ID = 0;
|
2012-01-13 06:30:30 +00:00
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
char &llvm::MachineSchedulerID = MachineScheduler::ID;
|
2012-01-13 06:30:30 +00:00
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
INITIALIZE_PASS_BEGIN(MachineScheduler, "misched",
|
2012-01-13 06:30:30 +00:00
|
|
|
"Machine Instruction Scheduler", false, false)
|
|
|
|
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
|
2012-01-17 06:55:03 +00:00
|
|
|
INITIALIZE_PASS_END(MachineScheduler, "misched",
|
2012-01-13 06:30:30 +00:00
|
|
|
"Machine Instruction Scheduler", false, false)
|
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
MachineScheduler::MachineScheduler()
|
2013-12-28 21:56:47 +00:00
|
|
|
: MachineSchedulerBase(ID) {
|
2012-01-17 06:55:03 +00:00
|
|
|
initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
|
2012-01-13 06:30:30 +00:00
|
|
|
}
|
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
|
2012-01-13 06:30:30 +00:00
|
|
|
AU.setPreservesCFG();
|
|
|
|
AU.addRequiredID(MachineDominatorsID);
|
|
|
|
AU.addRequired<MachineLoopInfo>();
|
|
|
|
AU.addRequired<AliasAnalysis>();
|
2012-03-09 00:52:20 +00:00
|
|
|
AU.addRequired<TargetPassConfig>();
|
2012-01-13 06:30:30 +00:00
|
|
|
AU.addRequired<SlotIndexes>();
|
|
|
|
AU.addPreserved<SlotIndexes>();
|
|
|
|
AU.addRequired<LiveIntervals>();
|
|
|
|
AU.addPreserved<LiveIntervals>();
|
|
|
|
MachineFunctionPass::getAnalysisUsage(AU);
|
|
|
|
}
|
|
|
|
|
2013-12-28 21:56:51 +00:00
|
|
|
char PostMachineScheduler::ID = 0;
|
|
|
|
|
|
|
|
char &llvm::PostMachineSchedulerID = PostMachineScheduler::ID;
|
|
|
|
|
|
|
|
INITIALIZE_PASS(PostMachineScheduler, "postmisched",
|
2013-12-28 22:47:55 +00:00
|
|
|
"PostRA Machine Instruction Scheduler", false, false)
|
2013-12-28 21:56:51 +00:00
|
|
|
|
|
|
|
PostMachineScheduler::PostMachineScheduler()
|
|
|
|
: MachineSchedulerBase(ID) {
|
|
|
|
initializePostMachineSchedulerPass(*PassRegistry::getPassRegistry());
|
|
|
|
}
|
|
|
|
|
|
|
|
void PostMachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
|
|
|
|
AU.setPreservesCFG();
|
|
|
|
AU.addRequiredID(MachineDominatorsID);
|
|
|
|
AU.addRequired<MachineLoopInfo>();
|
|
|
|
AU.addRequired<TargetPassConfig>();
|
|
|
|
MachineFunctionPass::getAnalysisUsage(AU);
|
|
|
|
}
|
|
|
|
|
2012-01-13 06:30:30 +00:00
|
|
|
MachinePassRegistry MachineSchedRegistry::Registry;
|
|
|
|
|
2012-03-09 00:52:20 +00:00
|
|
|
/// A dummy default scheduler factory indicates whether the scheduler
|
|
|
|
/// is overridden on the command line.
|
|
|
|
static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
|
2014-04-14 00:51:57 +00:00
|
|
|
return nullptr;
|
2012-03-09 00:52:20 +00:00
|
|
|
}
|
2012-01-13 06:30:30 +00:00
|
|
|
|
|
|
|
/// MachineSchedOpt allows command line selection of the scheduler.
|
|
|
|
static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
|
|
|
|
RegisterPassParser<MachineSchedRegistry> >
|
|
|
|
MachineSchedOpt("misched",
|
2012-03-09 00:52:20 +00:00
|
|
|
cl::init(&useDefaultMachineSched), cl::Hidden,
|
2012-01-13 06:30:30 +00:00
|
|
|
cl::desc("Machine instruction scheduler to use"));
|
|
|
|
|
2012-03-09 00:52:20 +00:00
|
|
|
static MachineSchedRegistry
|
2012-03-14 04:00:41 +00:00
|
|
|
DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
|
2012-03-09 00:52:20 +00:00
|
|
|
useDefaultMachineSched);
|
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
/// Forward declare the standard machine scheduler. This will be used as the
|
2012-03-09 00:52:20 +00:00
|
|
|
/// default scheduler if the target does not set a default.
|
2013-12-28 21:56:57 +00:00
|
|
|
static ScheduleDAGInstrs *createGenericSchedLive(MachineSchedContext *C);
|
|
|
|
static ScheduleDAGInstrs *createGenericSchedPostRA(MachineSchedContext *C);
|
2012-04-24 18:04:34 +00:00
|
|
|
|
|
|
|
/// Decrement this iterator until reaching the top or a non-debug instr.
|
2013-08-30 04:36:57 +00:00
|
|
|
static MachineBasicBlock::const_iterator
|
|
|
|
priorNonDebug(MachineBasicBlock::const_iterator I,
|
|
|
|
MachineBasicBlock::const_iterator Beg) {
|
2012-04-24 18:04:34 +00:00
|
|
|
assert(I != Beg && "reached the top of the region, cannot decrement");
|
|
|
|
while (--I != Beg) {
|
|
|
|
if (!I->isDebugValue())
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return I;
|
|
|
|
}
|
|
|
|
|
2013-08-30 04:36:57 +00:00
|
|
|
/// Non-const version.
|
|
|
|
static MachineBasicBlock::iterator
|
|
|
|
priorNonDebug(MachineBasicBlock::iterator I,
|
|
|
|
MachineBasicBlock::const_iterator Beg) {
|
|
|
|
return const_cast<MachineInstr*>(
|
|
|
|
&*priorNonDebug(MachineBasicBlock::const_iterator(I), Beg));
|
|
|
|
}
|
|
|
|
|
2012-04-24 18:04:34 +00:00
|
|
|
/// If this iterator is a debug value, increment until reaching the End or a
|
|
|
|
/// non-debug instruction.
|
2013-08-31 05:17:58 +00:00
|
|
|
static MachineBasicBlock::const_iterator
|
|
|
|
nextIfDebug(MachineBasicBlock::const_iterator I,
|
|
|
|
MachineBasicBlock::const_iterator End) {
|
2012-05-17 18:35:03 +00:00
|
|
|
for(; I != End; ++I) {
|
2012-04-24 18:04:34 +00:00
|
|
|
if (!I->isDebugValue())
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return I;
|
|
|
|
}
|
|
|
|
|
2013-08-31 05:17:58 +00:00
|
|
|
/// Non-const version.
|
|
|
|
static MachineBasicBlock::iterator
|
|
|
|
nextIfDebug(MachineBasicBlock::iterator I,
|
|
|
|
MachineBasicBlock::const_iterator End) {
|
|
|
|
// Cast the return value to nonconst MachineInstr, then cast to an
|
|
|
|
// instr_iterator, which does not check for null, finally return a
|
|
|
|
// bundle_iterator.
|
|
|
|
return MachineBasicBlock::instr_iterator(
|
|
|
|
const_cast<MachineInstr*>(
|
|
|
|
&*nextIfDebug(MachineBasicBlock::const_iterator(I), End)));
|
|
|
|
}
|
|
|
|
|
2013-09-24 17:11:19 +00:00
|
|
|
/// Instantiate a ScheduleDAGInstrs that will be owned by the caller.
|
2013-09-20 05:14:41 +00:00
|
|
|
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.
|
2013-12-28 21:56:57 +00:00
|
|
|
return createGenericSchedLive(this);
|
2013-09-20 05:14:41 +00:00
|
|
|
}
|
|
|
|
|
2013-12-28 21:56:51 +00:00
|
|
|
/// 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.
|
2013-12-28 21:56:57 +00:00
|
|
|
return createGenericSchedPostRA(this);
|
2013-12-28 21:56:51 +00:00
|
|
|
}
|
|
|
|
|
2012-03-14 04:00:38 +00:00
|
|
|
/// Top-level MachineScheduler pass driver.
|
|
|
|
///
|
|
|
|
/// Visit blocks in function order. Divide each block into scheduling regions
|
2012-03-14 04:00:41 +00:00
|
|
|
/// and visit them bottom-up. Visiting regions bottom-up is not required, but is
|
|
|
|
/// consistent with the DAG builder, which traverses the interior of the
|
|
|
|
/// scheduling regions bottom-up.
|
2012-03-14 04:00:38 +00:00
|
|
|
///
|
|
|
|
/// This design avoids exposing scheduling boundaries to the DAG builder,
|
2012-03-14 04:00:41 +00:00
|
|
|
/// simplifying the DAG builder's support for "special" target instructions.
|
|
|
|
/// At the same time the design allows target schedulers to operate across
|
2012-03-14 04:00:38 +00:00
|
|
|
/// scheduling boundaries, for example to bundle the boudary instructions
|
|
|
|
/// without reordering them. This creates complexity, because the target
|
|
|
|
/// scheduler must update the RegionBegin and RegionEnd positions cached by
|
|
|
|
/// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
|
|
|
|
/// design would be to split blocks at scheduling boundaries, but LLVM has a
|
|
|
|
/// general bias against block splitting purely for implementation simplicity.
|
2012-03-08 01:41:12 +00:00
|
|
|
bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
|
2012-05-10 21:06:21 +00:00
|
|
|
DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
|
|
|
|
|
2012-03-08 01:41:12 +00:00
|
|
|
// Initialize the context of the pass.
|
|
|
|
MF = &mf;
|
|
|
|
MLI = &getAnalysis<MachineLoopInfo>();
|
|
|
|
MDT = &getAnalysis<MachineDominatorTree>();
|
2012-03-09 00:52:20 +00:00
|
|
|
PassConfig = &getAnalysis<TargetPassConfig>();
|
2012-03-08 01:41:12 +00:00
|
|
|
AA = &getAnalysis<AliasAnalysis>();
|
|
|
|
|
|
|
|
LIS = &getAnalysis<LiveIntervals>();
|
|
|
|
|
2013-03-08 05:40:34 +00:00
|
|
|
if (VerifyScheduling) {
|
2013-07-25 07:26:26 +00:00
|
|
|
DEBUG(LIS->dump());
|
2013-03-08 05:40:34 +00:00
|
|
|
MF->verify(this, "Before machine scheduling.");
|
|
|
|
}
|
2012-04-24 20:36:19 +00:00
|
|
|
RegClassInfo->runOnMachineFunction(*MF);
|
2012-04-24 17:56:43 +00:00
|
|
|
|
2013-09-20 05:14:41 +00:00
|
|
|
// Instantiate the selected scheduler for this target, function, and
|
|
|
|
// optimization level.
|
2014-03-06 05:51:42 +00:00
|
|
|
std::unique_ptr<ScheduleDAGInstrs> Scheduler(createMachineScheduler());
|
2013-12-28 21:56:47 +00:00
|
|
|
scheduleRegions(*Scheduler);
|
|
|
|
|
|
|
|
DEBUG(LIS->dump());
|
|
|
|
if (VerifyScheduling)
|
|
|
|
MF->verify(this, "After machine scheduling.");
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2013-12-28 21:56:51 +00:00
|
|
|
bool PostMachineScheduler::runOnMachineFunction(MachineFunction &mf) {
|
2014-03-31 17:43:35 +00:00
|
|
|
if (skipOptnoneFunction(*mf.getFunction()))
|
|
|
|
return false;
|
|
|
|
|
2013-12-28 21:56:51 +00:00
|
|
|
DEBUG(dbgs() << "Before post-MI-sched:\n"; mf.print(dbgs()));
|
|
|
|
|
|
|
|
// Initialize the context of the pass.
|
|
|
|
MF = &mf;
|
|
|
|
PassConfig = &getAnalysis<TargetPassConfig>();
|
|
|
|
|
|
|
|
if (VerifyScheduling)
|
|
|
|
MF->verify(this, "Before post machine scheduling.");
|
|
|
|
|
|
|
|
// Instantiate the selected scheduler for this target, function, and
|
|
|
|
// optimization level.
|
2014-03-06 05:51:42 +00:00
|
|
|
std::unique_ptr<ScheduleDAGInstrs> Scheduler(createPostMachineScheduler());
|
2013-12-28 21:56:51 +00:00
|
|
|
scheduleRegions(*Scheduler);
|
|
|
|
|
|
|
|
if (VerifyScheduling)
|
|
|
|
MF->verify(this, "After post machine scheduling.");
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2013-12-28 21:56:57 +00:00
|
|
|
/// 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,
|
|
|
|
const TargetInstrInfo *TII,
|
|
|
|
bool IsPostRA) {
|
|
|
|
return MI->isCall() || TII->isSchedulingBoundary(MI, MBB, *MF);
|
|
|
|
}
|
|
|
|
|
2013-12-28 21:56:47 +00:00
|
|
|
/// Main driver for both MachineScheduler and PostMachineScheduler.
|
|
|
|
void MachineSchedulerBase::scheduleRegions(ScheduleDAGInstrs &Scheduler) {
|
|
|
|
const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
|
2013-12-28 21:56:57 +00:00
|
|
|
bool IsPostRA = Scheduler.isPostRA();
|
2012-03-08 01:41:12 +00:00
|
|
|
|
|
|
|
// Visit all machine basic blocks.
|
2012-04-24 17:56:43 +00:00
|
|
|
//
|
|
|
|
// TODO: Visit blocks in global postorder or postorder within the bottom-up
|
|
|
|
// loop tree. Then we can optionally compute global RegPressure.
|
2012-03-08 01:41:12 +00:00
|
|
|
for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
|
|
|
|
MBB != MBBEnd; ++MBB) {
|
|
|
|
|
2013-12-28 21:56:47 +00:00
|
|
|
Scheduler.startBlock(MBB);
|
2012-03-09 08:02:51 +00:00
|
|
|
|
2013-12-28 21:57:02 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
if (SchedOnlyFunc.getNumOccurrences() && SchedOnlyFunc != MF->getName())
|
|
|
|
continue;
|
|
|
|
if (SchedOnlyBlock.getNumOccurrences()
|
|
|
|
&& (int)SchedOnlyBlock != MBB->getNumber())
|
|
|
|
continue;
|
|
|
|
#endif
|
|
|
|
|
2012-03-08 01:41:12 +00:00
|
|
|
// Break the block into scheduling regions [I, RegionEnd), and schedule each
|
2012-07-23 08:51:15 +00:00
|
|
|
// region as soon as it is discovered. RegionEnd points the scheduling
|
2012-03-09 22:34:56 +00:00
|
|
|
// boundary at the bottom of the region. The DAG does not include RegionEnd,
|
|
|
|
// but the region does (i.e. the next RegionEnd is above the previous
|
|
|
|
// RegionBegin). If the current block has no terminator then RegionEnd ==
|
|
|
|
// MBB->end() for the bottom region.
|
|
|
|
//
|
|
|
|
// The Scheduler may insert instructions during either schedule() or
|
|
|
|
// exitRegion(), even for empty regions. So the local iterators 'I' and
|
|
|
|
// 'RegionEnd' are invalid across these calls.
|
2013-12-28 21:56:57 +00:00
|
|
|
//
|
|
|
|
// MBB::size() uses instr_iterator to count. Here we need a bundle to count
|
|
|
|
// as a single instruction.
|
|
|
|
unsigned RemainingInstrs = std::distance(MBB->begin(), MBB->end());
|
2012-03-09 03:46:39 +00:00
|
|
|
for(MachineBasicBlock::iterator RegionEnd = MBB->end();
|
2013-12-28 21:56:47 +00:00
|
|
|
RegionEnd != MBB->begin(); RegionEnd = Scheduler.begin()) {
|
2012-04-24 17:56:43 +00:00
|
|
|
|
2012-03-09 08:02:51 +00:00
|
|
|
// Avoid decrementing RegionEnd for blocks with no terminator.
|
2014-03-02 12:27:27 +00:00
|
|
|
if (RegionEnd != MBB->end() ||
|
|
|
|
isSchedBoundary(std::prev(RegionEnd), MBB, MF, TII, IsPostRA)) {
|
2012-03-09 08:02:51 +00:00
|
|
|
--RegionEnd;
|
|
|
|
// Count the boundary instruction.
|
2012-11-06 07:10:34 +00:00
|
|
|
--RemainingInstrs;
|
2012-03-09 08:02:51 +00:00
|
|
|
}
|
|
|
|
|
2012-03-08 01:41:12 +00:00
|
|
|
// The next region starts above the previous region. Look backward in the
|
|
|
|
// instruction stream until we find the nearest boundary.
|
2013-08-23 17:48:33 +00:00
|
|
|
unsigned NumRegionInstrs = 0;
|
2012-03-08 01:41:12 +00:00
|
|
|
MachineBasicBlock::iterator I = RegionEnd;
|
2013-08-23 17:48:33 +00:00
|
|
|
for(;I != MBB->begin(); --I, --RemainingInstrs, ++NumRegionInstrs) {
|
2014-03-02 12:27:27 +00:00
|
|
|
if (isSchedBoundary(std::prev(I), MBB, MF, TII, IsPostRA))
|
2012-03-08 01:41:12 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
// Notify the scheduler of the region, even if we may skip scheduling
|
|
|
|
// it. Perhaps it still needs to be bundled.
|
2013-12-28 21:56:47 +00:00
|
|
|
Scheduler.enterRegion(MBB, I, RegionEnd, NumRegionInstrs);
|
2012-03-08 01:41:12 +00:00
|
|
|
|
|
|
|
// Skip empty scheduling regions (0 or 1 schedulable instructions).
|
2014-03-02 12:27:27 +00:00
|
|
|
if (I == RegionEnd || I == std::prev(RegionEnd)) {
|
2012-03-08 01:41:12 +00:00
|
|
|
// Close the current region. Bundle the terminator if needed.
|
2012-03-09 22:34:56 +00:00
|
|
|
// This invalidates 'RegionEnd' and 'I'.
|
2013-12-28 21:56:47 +00:00
|
|
|
Scheduler.exitRegion();
|
2012-03-08 01:41:12 +00:00
|
|
|
continue;
|
|
|
|
}
|
2013-12-28 21:56:57 +00:00
|
|
|
DEBUG(dbgs() << "********** " << ((Scheduler.isPostRA()) ? "PostRA " : "")
|
|
|
|
<< "MI Scheduling **********\n");
|
2012-08-22 06:07:19 +00:00
|
|
|
DEBUG(dbgs() << MF->getName()
|
2013-01-25 07:45:31 +00:00
|
|
|
<< ":BB#" << MBB->getNumber() << " " << MBB->getName()
|
|
|
|
<< "\n From: " << *I << " To: ";
|
2012-03-08 01:41:12 +00:00
|
|
|
if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
|
|
|
|
else dbgs() << "End";
|
2013-08-23 17:48:33 +00:00
|
|
|
dbgs() << " RegionInstrs: " << NumRegionInstrs
|
|
|
|
<< " Remaining: " << RemainingInstrs << "\n");
|
2012-03-08 01:41:12 +00:00
|
|
|
|
2012-03-09 03:46:42 +00:00
|
|
|
// Schedule a region: possibly reorder instructions.
|
2012-03-09 22:34:56 +00:00
|
|
|
// This invalidates 'RegionEnd' and 'I'.
|
2013-12-28 21:56:47 +00:00
|
|
|
Scheduler.schedule();
|
2012-03-09 03:46:42 +00:00
|
|
|
|
|
|
|
// Close the current region.
|
2013-12-28 21:56:47 +00:00
|
|
|
Scheduler.exitRegion();
|
2012-03-08 01:41:12 +00:00
|
|
|
|
|
|
|
// Scheduling has invalidated the current iterator 'I'. Ask the
|
|
|
|
// scheduler for the top of it's scheduled region.
|
2013-12-28 21:56:47 +00:00
|
|
|
RegionEnd = Scheduler.begin();
|
2012-03-08 01:41:12 +00:00
|
|
|
}
|
2012-11-06 07:10:34 +00:00
|
|
|
assert(RemainingInstrs == 0 && "Instruction count mismatch!");
|
2013-12-28 21:56:47 +00:00
|
|
|
Scheduler.finishBlock();
|
2013-12-28 21:56:57 +00:00
|
|
|
if (Scheduler.isPostRA()) {
|
|
|
|
// FIXME: Ideally, no further passes should rely on kill flags. However,
|
|
|
|
// thumb2 size reduction is currently an exception.
|
|
|
|
Scheduler.fixupKills(MBB);
|
|
|
|
}
|
2012-03-08 01:41:12 +00:00
|
|
|
}
|
2013-12-28 21:56:47 +00:00
|
|
|
Scheduler.finalizeSchedule();
|
2012-03-08 01:41:12 +00:00
|
|
|
}
|
|
|
|
|
2013-12-28 21:56:47 +00:00
|
|
|
void MachineSchedulerBase::print(raw_ostream &O, const Module* m) const {
|
2012-03-08 01:41:12 +00:00
|
|
|
// unimplemented
|
|
|
|
}
|
|
|
|
|
2012-09-11 22:23:19 +00:00
|
|
|
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
|
2012-09-11 00:39:15 +00:00
|
|
|
void ReadyQueue::dump() {
|
2013-06-17 21:45:05 +00:00
|
|
|
dbgs() << Name << ": ";
|
2012-09-11 00:39:15 +00:00
|
|
|
for (unsigned i = 0, e = Queue.size(); i < e; ++i)
|
|
|
|
dbgs() << Queue[i]->NodeNum << " ";
|
|
|
|
dbgs() << "\n";
|
|
|
|
}
|
|
|
|
#endif
|
2012-03-14 04:00:41 +00:00
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
2013-12-28 21:56:47 +00:00
|
|
|
// ScheduleDAGMI - Basic machine instruction scheduling. This is
|
|
|
|
// independent of PreRA/PostRA scheduling and involves no extra book-keeping for
|
|
|
|
// virtual registers.
|
|
|
|
// ===----------------------------------------------------------------------===/
|
2012-01-14 02:17:06 +00:00
|
|
|
|
2014-04-21 20:32:32 +00:00
|
|
|
// Provide a vtable anchor.
|
2013-01-25 04:01:04 +00:00
|
|
|
ScheduleDAGMI::~ScheduleDAGMI() {
|
|
|
|
}
|
|
|
|
|
2013-04-24 15:54:43 +00:00
|
|
|
bool ScheduleDAGMI::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
|
|
|
|
return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
|
|
|
|
}
|
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
|
2012-11-12 19:52:20 +00:00
|
|
|
if (SuccSU != &ExitSU) {
|
|
|
|
// Do not use WillCreateCycle, it assumes SD scheduling.
|
|
|
|
// If Pred is reachable from Succ, then the edge creates a cycle.
|
|
|
|
if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
|
|
|
|
return false;
|
|
|
|
Topo.AddPred(SuccSU, PredDep.getSUnit());
|
|
|
|
}
|
2012-11-12 19:40:10 +00:00
|
|
|
SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
|
|
|
|
// Return true regardless of whether a new edge needed to be inserted.
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2012-01-17 06:55:07 +00:00
|
|
|
/// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
|
|
|
|
/// NumPredsLeft reaches zero, release the successor node.
|
2012-05-24 22:11:09 +00:00
|
|
|
///
|
|
|
|
/// FIXME: Adjust SuccSU height based on MinLatency.
|
2012-03-14 04:00:41 +00:00
|
|
|
void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
|
2012-01-17 06:55:07 +00:00
|
|
|
SUnit *SuccSU = SuccEdge->getSUnit();
|
|
|
|
|
2012-11-12 19:28:57 +00:00
|
|
|
if (SuccEdge->isWeak()) {
|
|
|
|
--SuccSU->WeakPredsLeft;
|
2012-11-12 19:40:10 +00:00
|
|
|
if (SuccEdge->isCluster())
|
|
|
|
NextClusterSucc = SuccSU;
|
2012-11-12 19:28:57 +00:00
|
|
|
return;
|
|
|
|
}
|
2012-01-17 06:55:07 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
if (SuccSU->NumPredsLeft == 0) {
|
|
|
|
dbgs() << "*** Scheduling failed! ***\n";
|
|
|
|
SuccSU->dump(this);
|
|
|
|
dbgs() << " has been released too many times!\n";
|
2014-04-14 00:51:57 +00:00
|
|
|
llvm_unreachable(nullptr);
|
2012-01-17 06:55:07 +00:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
--SuccSU->NumPredsLeft;
|
|
|
|
if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
|
2012-03-14 04:00:41 +00:00
|
|
|
SchedImpl->releaseTopNode(SuccSU);
|
2012-01-17 06:55:07 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// releaseSuccessors - Call releaseSucc on each of SU's successors.
|
2012-03-14 04:00:41 +00:00
|
|
|
void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
|
2012-01-17 06:55:07 +00:00
|
|
|
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
releaseSucc(SU, &*I);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
/// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
|
|
|
|
/// NumSuccsLeft reaches zero, release the predecessor node.
|
2012-05-24 22:11:09 +00:00
|
|
|
///
|
|
|
|
/// FIXME: Adjust PredSU height based on MinLatency.
|
2012-03-14 04:00:41 +00:00
|
|
|
void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
|
|
|
|
SUnit *PredSU = PredEdge->getSUnit();
|
|
|
|
|
2012-11-12 19:28:57 +00:00
|
|
|
if (PredEdge->isWeak()) {
|
|
|
|
--PredSU->WeakSuccsLeft;
|
2012-11-12 19:40:10 +00:00
|
|
|
if (PredEdge->isCluster())
|
|
|
|
NextClusterPred = PredSU;
|
2012-11-12 19:28:57 +00:00
|
|
|
return;
|
|
|
|
}
|
2012-03-14 04:00:41 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
if (PredSU->NumSuccsLeft == 0) {
|
|
|
|
dbgs() << "*** Scheduling failed! ***\n";
|
|
|
|
PredSU->dump(this);
|
|
|
|
dbgs() << " has been released too many times!\n";
|
2014-04-14 00:51:57 +00:00
|
|
|
llvm_unreachable(nullptr);
|
2012-03-14 04:00:41 +00:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
--PredSU->NumSuccsLeft;
|
|
|
|
if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
|
|
|
|
SchedImpl->releaseBottomNode(PredSU);
|
|
|
|
}
|
|
|
|
|
|
|
|
/// releasePredecessors - Call releasePred on each of SU's predecessors.
|
|
|
|
void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
|
|
|
|
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
releasePred(SU, &*I);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-12-28 21:56:47 +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);
|
|
|
|
}
|
|
|
|
|
2013-04-13 06:07:40 +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.
|
2013-12-28 21:56:47 +00:00
|
|
|
void ScheduleDAGMI::moveInstruction(
|
|
|
|
MachineInstr *MI, MachineBasicBlock::iterator InsertPos) {
|
2012-05-17 18:35:03 +00:00
|
|
|
// Advance RegionBegin if the first instruction moves down.
|
2012-03-21 04:12:10 +00:00
|
|
|
if (&*RegionBegin == MI)
|
2012-05-17 18:35:03 +00:00
|
|
|
++RegionBegin;
|
|
|
|
|
|
|
|
// Update the instruction stream.
|
2012-03-14 04:00:41 +00:00
|
|
|
BB->splice(InsertPos, BB, MI);
|
2012-05-17 18:35:03 +00:00
|
|
|
|
|
|
|
// Update LiveIntervals
|
2013-12-28 21:56:47 +00:00
|
|
|
if (LIS)
|
|
|
|
LIS->handleMove(MI, /*UpdateFlags=*/true);
|
2012-05-17 18:35:03 +00:00
|
|
|
|
|
|
|
// Recede RegionBegin if an instruction moves above the first.
|
2012-03-14 04:00:41 +00:00
|
|
|
if (RegionBegin == InsertPos)
|
|
|
|
RegionBegin = MI;
|
|
|
|
}
|
|
|
|
|
2012-03-21 04:12:07 +00:00
|
|
|
bool ScheduleDAGMI::checkSchedLimit() {
|
|
|
|
#ifndef NDEBUG
|
|
|
|
if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
|
|
|
|
CurrentTop = CurrentBottom;
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
++NumInstrsScheduled;
|
|
|
|
#endif
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2013-12-28 21:56:47 +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() {
|
|
|
|
// 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);
|
|
|
|
|
|
|
|
DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
|
|
|
|
SUnits[su].dumpAll(this));
|
|
|
|
if (ViewMISchedDAGs) viewGraph();
|
|
|
|
|
|
|
|
// Initialize ready queues now that the DAG and priority data are finalized.
|
|
|
|
initQueues(TopRoots, BotRoots);
|
|
|
|
|
|
|
|
bool IsTopNode = false;
|
|
|
|
while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
|
|
|
|
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);
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
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;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
updateQueues(SU, IsTopNode);
|
|
|
|
|
|
|
|
// Notify the scheduling strategy after updating the DAG.
|
|
|
|
SchedImpl->schedNode(SU, IsTopNode);
|
|
|
|
}
|
|
|
|
assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
|
|
|
|
|
|
|
|
placeDebugValues();
|
|
|
|
|
|
|
|
DEBUG({
|
|
|
|
unsigned BBNum = begin()->getParent()->getNumber();
|
|
|
|
dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
|
|
|
|
dumpSchedule();
|
|
|
|
dbgs() << '\n';
|
|
|
|
});
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Apply each ScheduleDAGMutation step in order.
|
|
|
|
void ScheduleDAGMI::postprocessDAG() {
|
|
|
|
for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
|
|
|
|
Mutations[i]->apply(this);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void ScheduleDAGMI::
|
|
|
|
findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
|
|
|
|
SmallVectorImpl<SUnit*> &BotRoots) {
|
|
|
|
for (std::vector<SUnit>::iterator
|
|
|
|
I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
|
|
|
|
SUnit *SU = &(*I);
|
|
|
|
assert(!SU->isBoundaryNode() && "Boundary node should not be in SUnits");
|
|
|
|
|
|
|
|
// Order predecessors so DFSResult follows the critical path.
|
|
|
|
SU->biasCriticalPath();
|
|
|
|
|
|
|
|
// A SUnit is ready to top schedule if it has no predecessors.
|
|
|
|
if (!I->NumPredsLeft)
|
|
|
|
TopRoots.push_back(SU);
|
|
|
|
// A SUnit is ready to bottom schedule if it has no successors.
|
|
|
|
if (!I->NumSuccsLeft)
|
|
|
|
BotRoots.push_back(SU);
|
|
|
|
}
|
|
|
|
ExitSU.biasCriticalPath();
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Identify DAG roots and setup scheduler queues.
|
|
|
|
void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
|
|
|
|
ArrayRef<SUnit*> BotRoots) {
|
2014-04-14 00:51:57 +00:00
|
|
|
NextClusterSucc = nullptr;
|
|
|
|
NextClusterPred = nullptr;
|
2013-12-28 21:56:47 +00:00
|
|
|
|
|
|
|
// 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.
|
|
|
|
for (SmallVectorImpl<SUnit*>::const_iterator
|
|
|
|
I = TopRoots.begin(), E = TopRoots.end(); I != E; ++I) {
|
|
|
|
SchedImpl->releaseTopNode(*I);
|
|
|
|
}
|
|
|
|
// 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;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
|
|
|
|
DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
|
2014-03-02 12:27:27 +00:00
|
|
|
std::pair<MachineInstr *, MachineInstr *> P = *std::prev(DI);
|
2013-12-28 21:56:47 +00:00
|
|
|
MachineInstr *DbgValue = P.first;
|
|
|
|
MachineBasicBlock::iterator OrigPrevMI = P.second;
|
|
|
|
if (&*RegionBegin == DbgValue)
|
|
|
|
++RegionBegin;
|
|
|
|
BB->splice(++OrigPrevMI, BB, DbgValue);
|
2014-03-02 12:27:27 +00:00
|
|
|
if (OrigPrevMI == std::prev(RegionEnd))
|
2013-12-28 21:56:47 +00:00
|
|
|
RegionEnd = DbgValue;
|
|
|
|
}
|
|
|
|
DbgValues.clear();
|
2014-04-14 00:51:57 +00:00
|
|
|
FirstDbgValue = nullptr;
|
2013-12-28 21:56:47 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
|
|
|
|
void ScheduleDAGMI::dumpSchedule() const {
|
|
|
|
for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
|
|
|
|
if (SUnit *SU = getSUnit(&(*MI)))
|
|
|
|
SU->dump(this);
|
|
|
|
else
|
|
|
|
dbgs() << "Missing SUnit\n";
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// ScheduleDAGMILive - Base class for MachineInstr scheduling with LiveIntervals
|
|
|
|
// preservation.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
ScheduleDAGMILive::~ScheduleDAGMILive() {
|
|
|
|
delete DFSResult;
|
|
|
|
}
|
|
|
|
|
2012-04-24 17:56:43 +00:00
|
|
|
/// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
|
|
|
|
/// crossing a scheduling boundary. [begin, end) includes all instructions in
|
|
|
|
/// the region, including the boundary itself and single-instruction regions
|
|
|
|
/// that don't get scheduled.
|
2013-12-28 21:56:47 +00:00
|
|
|
void ScheduleDAGMILive::enterRegion(MachineBasicBlock *bb,
|
2012-04-24 17:56:43 +00:00
|
|
|
MachineBasicBlock::iterator begin,
|
|
|
|
MachineBasicBlock::iterator end,
|
2013-08-23 17:48:33 +00:00
|
|
|
unsigned regioninstrs)
|
2012-04-24 17:56:43 +00:00
|
|
|
{
|
2013-12-28 21:56:47 +00:00
|
|
|
// ScheduleDAGMI initializes SchedImpl's per-region policy.
|
|
|
|
ScheduleDAGMI::enterRegion(bb, begin, end, regioninstrs);
|
2012-05-10 21:06:10 +00:00
|
|
|
|
|
|
|
// For convenience remember the end of the liveness region.
|
2014-03-02 12:27:27 +00:00
|
|
|
LiveRegionEnd = (RegionEnd == bb->end()) ? RegionEnd : std::next(RegionEnd);
|
2013-09-06 17:32:34 +00:00
|
|
|
|
2013-09-06 17:32:47 +00:00
|
|
|
SUPressureDiffs.clear();
|
|
|
|
|
2013-09-06 17:32:34 +00:00
|
|
|
ShouldTrackPressure = SchedImpl->shouldTrackPressure();
|
2012-05-10 21:06:10 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Setup the register pressure trackers for the top scheduled top and bottom
|
|
|
|
// scheduled regions.
|
2013-12-28 21:56:47 +00:00
|
|
|
void ScheduleDAGMILive::initRegPressure() {
|
2012-05-10 21:06:10 +00:00
|
|
|
TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
|
|
|
|
BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
|
|
|
|
|
|
|
|
// Close the RPTracker to finalize live ins.
|
|
|
|
RPTracker.closeRegion();
|
|
|
|
|
2013-07-30 19:59:12 +00:00
|
|
|
DEBUG(RPTracker.dump());
|
2012-05-24 22:11:14 +00:00
|
|
|
|
2012-05-10 21:06:10 +00:00
|
|
|
// Initialize the live ins and live outs.
|
|
|
|
TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
|
|
|
|
BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
|
|
|
|
|
|
|
|
// Close one end of the tracker so we can call
|
|
|
|
// getMaxUpward/DownwardPressureDelta before advancing across any
|
|
|
|
// instructions. This converts currently live regs into live ins/outs.
|
|
|
|
TopRPTracker.closeTop();
|
|
|
|
BotRPTracker.closeBottom();
|
|
|
|
|
2013-07-30 19:59:12 +00:00
|
|
|
BotRPTracker.initLiveThru(RPTracker);
|
|
|
|
if (!BotRPTracker.getLiveThru().empty()) {
|
|
|
|
TopRPTracker.initLiveThru(BotRPTracker.getLiveThru());
|
|
|
|
DEBUG(dbgs() << "Live Thru: ";
|
|
|
|
dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI));
|
|
|
|
};
|
|
|
|
|
2013-08-30 04:36:57 +00:00
|
|
|
// 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);
|
|
|
|
|
2012-05-10 21:06:10 +00:00
|
|
|
// Account for liveness generated by the region boundary.
|
2013-08-30 04:36:57 +00:00
|
|
|
if (LiveRegionEnd != RegionEnd) {
|
|
|
|
SmallVector<unsigned, 8> LiveUses;
|
|
|
|
BotRPTracker.recede(&LiveUses);
|
|
|
|
updatePressureDiffs(LiveUses);
|
|
|
|
}
|
2012-05-10 21:06:10 +00:00
|
|
|
|
|
|
|
assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
|
2012-05-17 18:35:10 +00:00
|
|
|
|
|
|
|
// Cache the list of excess pressure sets in this region. This will also track
|
|
|
|
// the max pressure in the scheduled code for these sets.
|
|
|
|
RegionCriticalPSets.clear();
|
2013-01-25 21:44:27 +00:00
|
|
|
const std::vector<unsigned> &RegionPressure =
|
|
|
|
RPTracker.getPressure().MaxSetPressure;
|
2012-05-17 18:35:10 +00:00
|
|
|
for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
|
2013-06-21 18:32:58 +00:00
|
|
|
unsigned Limit = RegClassInfo->getRegPressureSetLimit(i);
|
2013-06-21 18:33:01 +00:00
|
|
|
if (RegionPressure[i] > Limit) {
|
|
|
|
DEBUG(dbgs() << TRI->getRegPressureSetName(i)
|
|
|
|
<< " Limit " << Limit
|
|
|
|
<< " Actual " << RegionPressure[i] << "\n");
|
2013-08-30 03:49:48 +00:00
|
|
|
RegionCriticalPSets.push_back(PressureChange(i));
|
2013-06-21 18:33:01 +00:00
|
|
|
}
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
|
|
|
DEBUG(dbgs() << "Excess PSets: ";
|
|
|
|
for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
|
|
|
|
dbgs() << TRI->getRegPressureSetName(
|
2013-08-30 03:49:48 +00:00
|
|
|
RegionCriticalPSets[i].getPSet()) << " ";
|
2012-05-17 18:35:10 +00:00
|
|
|
dbgs() << "\n");
|
|
|
|
}
|
|
|
|
|
2013-12-28 21:56:47 +00:00
|
|
|
void ScheduleDAGMILive::
|
2013-09-06 17:32:47 +00:00
|
|
|
updateScheduledPressure(const SUnit *SU,
|
|
|
|
const std::vector<unsigned> &NewMaxPressure) {
|
|
|
|
const PressureDiff &PDiff = getPressureDiff(SU);
|
|
|
|
unsigned CritIdx = 0, CritEnd = RegionCriticalPSets.size();
|
|
|
|
for (PressureDiff::const_iterator I = PDiff.begin(), E = PDiff.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
if (!I->isValid())
|
|
|
|
break;
|
|
|
|
unsigned ID = I->getPSet();
|
|
|
|
while (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() < ID)
|
|
|
|
++CritIdx;
|
|
|
|
if (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() == ID) {
|
|
|
|
if ((int)NewMaxPressure[ID] > RegionCriticalPSets[CritIdx].getUnitInc()
|
|
|
|
&& NewMaxPressure[ID] <= INT16_MAX)
|
|
|
|
RegionCriticalPSets[CritIdx].setUnitInc(NewMaxPressure[ID]);
|
|
|
|
}
|
|
|
|
unsigned Limit = RegClassInfo->getRegPressureSetLimit(ID);
|
|
|
|
if (NewMaxPressure[ID] >= Limit - 2) {
|
|
|
|
DEBUG(dbgs() << " " << TRI->getRegPressureSetName(ID) << ": "
|
|
|
|
<< NewMaxPressure[ID] << " > " << Limit << "(+ "
|
|
|
|
<< BotRPTracker.getLiveThru()[ID] << " livethru)\n");
|
|
|
|
}
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
2012-04-24 17:56:43 +00:00
|
|
|
}
|
|
|
|
|
2013-08-30 04:36:57 +00:00
|
|
|
/// Update the PressureDiff array for liveness after scheduling this
|
|
|
|
/// instruction.
|
2013-12-28 21:56:47 +00:00
|
|
|
void ScheduleDAGMILive::updatePressureDiffs(ArrayRef<unsigned> LiveUses) {
|
2013-08-30 04:36:57 +00:00
|
|
|
for (unsigned LUIdx = 0, LUEnd = LiveUses.size(); LUIdx != LUEnd; ++LUIdx) {
|
|
|
|
/// FIXME: Currently assuming single-use physregs.
|
|
|
|
unsigned Reg = LiveUses[LUIdx];
|
2013-09-06 17:32:39 +00:00
|
|
|
DEBUG(dbgs() << " LiveReg: " << PrintVRegOrUnit(Reg, TRI) << "\n");
|
2013-08-30 04:36:57 +00:00
|
|
|
if (!TRI->isVirtualRegister(Reg))
|
|
|
|
continue;
|
2013-09-06 17:32:39 +00:00
|
|
|
|
2013-08-30 04:36:57 +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;
|
2013-08-31 05:17:58 +00:00
|
|
|
MachineBasicBlock::const_iterator I =
|
|
|
|
nextIfDebug(BotRPTracker.getPos(), BB->end());
|
|
|
|
if (I == BB->end())
|
2013-08-30 04:36:57 +00:00
|
|
|
VNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
|
|
|
|
else {
|
2013-10-10 21:28:52 +00:00
|
|
|
LiveQueryResult LRQ = LI.Query(LIS->getInstructionIndex(I));
|
2013-08-30 04:36:57 +00:00
|
|
|
VNI = LRQ.valueIn();
|
|
|
|
}
|
|
|
|
// RegisterPressureTracker guarantees that readsReg is true for LiveUses.
|
|
|
|
assert(VNI && "No live value at use.");
|
|
|
|
for (VReg2UseMap::iterator
|
|
|
|
UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
|
|
|
|
SUnit *SU = UI->SU;
|
2013-09-06 17:32:39 +00:00
|
|
|
DEBUG(dbgs() << " UpdateRegP: SU(" << SU->NodeNum << ") "
|
|
|
|
<< *SU->getInstr());
|
2013-08-30 04:36:57 +00:00
|
|
|
// If this use comes before the reaching def, it cannot be a last use, so
|
|
|
|
// descrease its pressure change.
|
|
|
|
if (!SU->isScheduled && SU != &ExitSU) {
|
2013-10-10 21:28:52 +00:00
|
|
|
LiveQueryResult LRQ
|
|
|
|
= LI.Query(LIS->getInstructionIndex(SU->getInstr()));
|
2013-08-30 04:36:57 +00:00
|
|
|
if (LRQ.valueIn() == VNI)
|
|
|
|
getPressureDiff(SU).addPressureChange(Reg, true, &MRI);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
/// schedule - Called back from MachineScheduler::runOnMachineFunction
|
2012-04-24 17:56:43 +00:00
|
|
|
/// after setting up the current scheduling region. [RegionBegin, RegionEnd)
|
|
|
|
/// only includes instructions that have DAG nodes, not scheduling boundaries.
|
2012-09-11 00:39:15 +00:00
|
|
|
///
|
|
|
|
/// This is a skeletal driver, with all the functionality pushed into helpers,
|
|
|
|
/// so that it can be easilly extended by experimental schedulers. Generally,
|
|
|
|
/// implementing MachineSchedStrategy should be sufficient to implement a new
|
|
|
|
/// scheduling algorithm. However, if a scheduler further subclasses
|
2013-12-28 21:56:47 +00:00
|
|
|
/// ScheduleDAGMILive then it will want to override this virtual method in order
|
|
|
|
/// to update any specialized state.
|
|
|
|
void ScheduleDAGMILive::schedule() {
|
2012-09-11 00:39:15 +00:00
|
|
|
buildDAGWithRegPressure();
|
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
Topo.InitDAGTopologicalSorting();
|
|
|
|
|
2012-09-14 17:22:42 +00:00
|
|
|
postprocessDAG();
|
|
|
|
|
2013-01-25 06:33:57 +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);
|
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
|
|
|
|
SUnits[su].dumpAll(this));
|
2013-01-25 06:33:57 +00:00
|
|
|
if (ViewMISchedDAGs) viewGraph();
|
2012-09-11 00:39:15 +00:00
|
|
|
|
2013-01-25 06:33:57 +00:00
|
|
|
// Initialize ready queues now that the DAG and priority data are finalized.
|
|
|
|
initQueues(TopRoots, BotRoots);
|
2012-09-11 00:39:15 +00:00
|
|
|
|
2013-12-28 21:56:47 +00:00
|
|
|
if (ShouldTrackPressure) {
|
|
|
|
assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
|
|
|
|
TopRPTracker.setPos(CurrentTop);
|
|
|
|
}
|
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
bool IsTopNode = false;
|
|
|
|
while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
|
2012-10-08 18:53:53 +00:00
|
|
|
assert(!SU->isScheduled && "Node already scheduled");
|
2012-09-11 00:39:15 +00:00
|
|
|
if (!checkSchedLimit())
|
|
|
|
break;
|
|
|
|
|
|
|
|
scheduleMI(SU, IsTopNode);
|
|
|
|
|
|
|
|
updateQueues(SU, IsTopNode);
|
2013-12-28 21:56:47 +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);
|
2012-09-11 00:39:15 +00:00
|
|
|
}
|
|
|
|
assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
|
|
|
|
|
|
|
|
placeDebugValues();
|
2012-11-07 07:05:09 +00:00
|
|
|
|
|
|
|
DEBUG({
|
2012-11-28 03:42:47 +00:00
|
|
|
unsigned BBNum = begin()->getParent()->getNumber();
|
2012-11-07 07:05:09 +00:00
|
|
|
dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
|
|
|
|
dumpSchedule();
|
|
|
|
dbgs() << '\n';
|
|
|
|
});
|
2012-09-11 00:39:15 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Build the DAG and setup three register pressure trackers.
|
2013-12-28 21:56:47 +00:00
|
|
|
void ScheduleDAGMILive::buildDAGWithRegPressure() {
|
2013-09-04 20:59:59 +00:00
|
|
|
if (!ShouldTrackPressure) {
|
|
|
|
RPTracker.reset();
|
|
|
|
RegionCriticalPSets.clear();
|
|
|
|
buildSchedGraph(AA);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2012-05-10 21:06:10 +00:00
|
|
|
// Initialize the register pressure tracker used by buildSchedGraph.
|
2013-07-30 19:59:12 +00:00
|
|
|
RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd,
|
|
|
|
/*TrackUntiedDefs=*/true);
|
2012-05-10 21:06:10 +00:00
|
|
|
|
|
|
|
// Account for liveness generate by the region boundary.
|
|
|
|
if (LiveRegionEnd != RegionEnd)
|
|
|
|
RPTracker.recede();
|
2012-04-24 17:56:43 +00:00
|
|
|
|
2012-05-10 21:06:10 +00:00
|
|
|
// Build the DAG, and compute current register pressure.
|
2013-08-30 03:49:48 +00:00
|
|
|
buildSchedGraph(AA, &RPTracker, &SUPressureDiffs);
|
2012-01-17 06:55:07 +00:00
|
|
|
|
2012-05-10 21:06:10 +00:00
|
|
|
// Initialize top/bottom trackers after computing region pressure.
|
|
|
|
initRegPressure();
|
2012-09-11 00:39:15 +00:00
|
|
|
}
|
2012-05-10 21:06:10 +00:00
|
|
|
|
2013-12-28 21:56:47 +00:00
|
|
|
void ScheduleDAGMILive::computeDFSResult() {
|
2013-01-25 04:01:04 +00:00
|
|
|
if (!DFSResult)
|
|
|
|
DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
|
|
|
|
DFSResult->clear();
|
|
|
|
ScheduledTrees.clear();
|
2013-01-25 06:33:57 +00:00
|
|
|
DFSResult->resize(SUnits.size());
|
|
|
|
DFSResult->compute(SUnits);
|
2013-01-25 04:01:04 +00:00
|
|
|
ScheduledTrees.resize(DFSResult->getNumSubtrees());
|
|
|
|
}
|
|
|
|
|
2013-08-29 18:04:49 +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
|
2013-08-30 02:02:12 +00:00
|
|
|
/// edge and considers the depth and height of the nodes. For example, consider
|
2013-08-29 18:04:49 +00:00
|
|
|
/// 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
|
2013-12-28 21:56:47 +00:00
|
|
|
///
|
|
|
|
/// This could be relevant to PostRA scheduling, but is currently implemented
|
|
|
|
/// assuming LiveIntervals.
|
|
|
|
unsigned ScheduleDAGMILive::computeCyclicCriticalPath() {
|
2013-08-29 18:04:49 +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.
|
|
|
|
ArrayRef<unsigned> LiveOuts = RPTracker.getPressure().LiveOutRegs;
|
|
|
|
for (ArrayRef<unsigned>::iterator RI = LiveOuts.begin(), RE = LiveOuts.end();
|
|
|
|
RI != RE; ++RI) {
|
|
|
|
unsigned Reg = *RI;
|
|
|
|
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.
|
|
|
|
for (VReg2UseMap::iterator
|
|
|
|
UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
|
|
|
|
if (UI->SU == &ExitSU)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Only consider uses of the phi.
|
2013-10-10 21:28:52 +00:00
|
|
|
LiveQueryResult LRQ =
|
|
|
|
LI.Query(LIS->getInstructionIndex(UI->SU->getInstr()));
|
2013-08-29 18:04:49 +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;
|
|
|
|
if (LiveOutDepth > UI->SU->getDepth())
|
|
|
|
CyclicLatency = LiveOutDepth - UI->SU->getDepth();
|
|
|
|
|
|
|
|
unsigned LiveInHeight = UI->SU->getHeight() + DefSU->Latency;
|
|
|
|
if (LiveInHeight > LiveOutHeight) {
|
|
|
|
if (LiveInHeight - LiveOutHeight < CyclicLatency)
|
|
|
|
CyclicLatency = LiveInHeight - LiveOutHeight;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
CyclicLatency = 0;
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "Cyclic Path: SU(" << DefSU->NodeNum << ") -> SU("
|
|
|
|
<< UI->SU->NodeNum << ") = " << CyclicLatency << "c\n");
|
|
|
|
if (CyclicLatency > MaxCyclicLatency)
|
|
|
|
MaxCyclicLatency = CyclicLatency;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "c\n");
|
|
|
|
return MaxCyclicLatency;
|
|
|
|
}
|
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
/// Move an instruction and update register pressure.
|
2013-12-28 21:56:47 +00:00
|
|
|
void ScheduleDAGMILive::scheduleMI(SUnit *SU, bool IsTopNode) {
|
2012-09-11 00:39:15 +00:00
|
|
|
// Move the instruction to its new location in the instruction stream.
|
|
|
|
MachineInstr *MI = SU->getInstr();
|
2012-05-10 21:06:10 +00:00
|
|
|
|
2012-09-11 00:39:15 +00:00
|
|
|
if (IsTopNode) {
|
|
|
|
assert(SU->isTopReady() && "node still has unscheduled dependencies");
|
|
|
|
if (&*CurrentTop == MI)
|
|
|
|
CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
|
|
|
|
else {
|
|
|
|
moveInstruction(MI, CurrentTop);
|
|
|
|
TopRPTracker.setPos(MI);
|
2012-03-14 04:00:41 +00:00
|
|
|
}
|
2012-09-11 00:39:15 +00:00
|
|
|
|
2013-09-04 20:59:59 +00:00
|
|
|
if (ShouldTrackPressure) {
|
|
|
|
// Update top scheduled pressure.
|
|
|
|
TopRPTracker.advance();
|
|
|
|
assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
|
2013-09-06 17:32:47 +00:00
|
|
|
updateScheduledPressure(SU, TopRPTracker.getPressure().MaxSetPressure);
|
2013-09-04 20:59:59 +00:00
|
|
|
}
|
2012-09-11 00:39:15 +00:00
|
|
|
}
|
|
|
|
else {
|
|
|
|
assert(SU->isBottomReady() && "node still has unscheduled dependencies");
|
|
|
|
MachineBasicBlock::iterator priorII =
|
|
|
|
priorNonDebug(CurrentBottom, CurrentTop);
|
|
|
|
if (&*priorII == MI)
|
|
|
|
CurrentBottom = priorII;
|
2012-01-17 06:55:07 +00:00
|
|
|
else {
|
2012-09-11 00:39:15 +00:00
|
|
|
if (&*CurrentTop == MI) {
|
|
|
|
CurrentTop = nextIfDebug(++CurrentTop, priorII);
|
|
|
|
TopRPTracker.setPos(CurrentTop);
|
2012-03-14 04:00:41 +00:00
|
|
|
}
|
2012-09-11 00:39:15 +00:00
|
|
|
moveInstruction(MI, CurrentBottom);
|
|
|
|
CurrentBottom = MI;
|
2012-01-17 06:55:07 +00:00
|
|
|
}
|
2013-09-04 20:59:59 +00:00
|
|
|
if (ShouldTrackPressure) {
|
|
|
|
// Update bottom scheduled pressure.
|
|
|
|
SmallVector<unsigned, 8> LiveUses;
|
|
|
|
BotRPTracker.recede(&LiveUses);
|
|
|
|
assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
|
2013-09-06 17:32:47 +00:00
|
|
|
updateScheduledPressure(SU, BotRPTracker.getPressure().MaxSetPressure);
|
2013-09-04 20:59:59 +00:00
|
|
|
updatePressureDiffs(LiveUses);
|
|
|
|
}
|
2012-01-17 06:55:07 +00:00
|
|
|
}
|
2012-09-11 00:39:15 +00:00
|
|
|
}
|
2012-04-24 18:04:37 +00:00
|
|
|
|
2012-11-12 19:52:20 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// LoadClusterMutation - DAG post-processing to cluster loads.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
namespace {
|
|
|
|
/// \brief Post-process the DAG to create cluster edges between neighboring
|
|
|
|
/// loads.
|
|
|
|
class LoadClusterMutation : public ScheduleDAGMutation {
|
|
|
|
struct LoadInfo {
|
|
|
|
SUnit *SU;
|
|
|
|
unsigned BaseReg;
|
|
|
|
unsigned Offset;
|
|
|
|
LoadInfo(SUnit *su, unsigned reg, unsigned ofs)
|
|
|
|
: SU(su), BaseReg(reg), Offset(ofs) {}
|
2014-03-07 21:35:39 +00:00
|
|
|
|
|
|
|
bool operator<(const LoadInfo &RHS) const {
|
|
|
|
return std::tie(BaseReg, Offset) < std::tie(RHS.BaseReg, RHS.Offset);
|
|
|
|
}
|
2012-11-12 19:40:10 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
const TargetInstrInfo *TII;
|
|
|
|
const TargetRegisterInfo *TRI;
|
|
|
|
public:
|
|
|
|
LoadClusterMutation(const TargetInstrInfo *tii,
|
|
|
|
const TargetRegisterInfo *tri)
|
|
|
|
: TII(tii), TRI(tri) {}
|
|
|
|
|
2014-03-07 09:26:03 +00:00
|
|
|
void apply(ScheduleDAGMI *DAG) override;
|
2012-11-12 19:40:10 +00:00
|
|
|
protected:
|
|
|
|
void clusterNeighboringLoads(ArrayRef<SUnit*> Loads, ScheduleDAGMI *DAG);
|
|
|
|
};
|
|
|
|
} // anonymous
|
|
|
|
|
|
|
|
void LoadClusterMutation::clusterNeighboringLoads(ArrayRef<SUnit*> Loads,
|
|
|
|
ScheduleDAGMI *DAG) {
|
|
|
|
SmallVector<LoadClusterMutation::LoadInfo,32> LoadRecords;
|
|
|
|
for (unsigned Idx = 0, End = Loads.size(); Idx != End; ++Idx) {
|
|
|
|
SUnit *SU = Loads[Idx];
|
|
|
|
unsigned BaseReg;
|
|
|
|
unsigned Offset;
|
|
|
|
if (TII->getLdStBaseRegImmOfs(SU->getInstr(), BaseReg, Offset, TRI))
|
|
|
|
LoadRecords.push_back(LoadInfo(SU, BaseReg, Offset));
|
|
|
|
}
|
|
|
|
if (LoadRecords.size() < 2)
|
|
|
|
return;
|
2014-03-07 21:35:39 +00:00
|
|
|
std::sort(LoadRecords.begin(), LoadRecords.end());
|
2012-11-12 19:40:10 +00:00
|
|
|
unsigned ClusterLength = 1;
|
|
|
|
for (unsigned Idx = 0, End = LoadRecords.size(); Idx < (End - 1); ++Idx) {
|
|
|
|
if (LoadRecords[Idx].BaseReg != LoadRecords[Idx+1].BaseReg) {
|
|
|
|
ClusterLength = 1;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
SUnit *SUa = LoadRecords[Idx].SU;
|
|
|
|
SUnit *SUb = LoadRecords[Idx+1].SU;
|
2012-11-12 21:28:10 +00:00
|
|
|
if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
|
2012-11-12 19:40:10 +00:00
|
|
|
&& DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
|
|
|
|
<< SUb->NodeNum << ")\n");
|
|
|
|
// Copy successor edges from SUa to SUb. Interleaving computation
|
|
|
|
// dependent on SUa can prevent load combining due to register reuse.
|
|
|
|
// Predecessor edges do not need to be copied from SUb to SUa since nearby
|
|
|
|
// loads should have effectively the same inputs.
|
|
|
|
for (SUnit::const_succ_iterator
|
|
|
|
SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
|
|
|
|
if (SI->getSUnit() == SUb)
|
|
|
|
continue;
|
|
|
|
DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
|
|
|
|
DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
|
|
|
|
}
|
|
|
|
++ClusterLength;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
ClusterLength = 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// \brief Callback from DAG postProcessing to create cluster edges for loads.
|
|
|
|
void LoadClusterMutation::apply(ScheduleDAGMI *DAG) {
|
|
|
|
// Map DAG NodeNum to store chain ID.
|
|
|
|
DenseMap<unsigned, unsigned> StoreChainIDs;
|
|
|
|
// Map each store chain to a set of dependent loads.
|
|
|
|
SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
|
|
|
|
for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
|
|
|
|
SUnit *SU = &DAG->SUnits[Idx];
|
|
|
|
if (!SU->getInstr()->mayLoad())
|
|
|
|
continue;
|
|
|
|
unsigned ChainPredID = DAG->SUnits.size();
|
|
|
|
for (SUnit::const_pred_iterator
|
|
|
|
PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
|
|
|
|
if (PI->isCtrl()) {
|
|
|
|
ChainPredID = PI->getSUnit()->NodeNum;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Check if this chain-like pred has been seen
|
|
|
|
// before. ChainPredID==MaxNodeID for loads at the top of the schedule.
|
|
|
|
unsigned NumChains = StoreChainDependents.size();
|
|
|
|
std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
|
|
|
|
StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
|
|
|
|
if (Result.second)
|
|
|
|
StoreChainDependents.resize(NumChains + 1);
|
|
|
|
StoreChainDependents[Result.first->second].push_back(SU);
|
|
|
|
}
|
|
|
|
// Iterate over the store chains.
|
|
|
|
for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
|
|
|
|
clusterNeighboringLoads(StoreChainDependents[Idx], DAG);
|
|
|
|
}
|
|
|
|
|
2012-11-12 19:52:20 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// MacroFusion - DAG post-processing to encourage fusion of macro ops.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
/// \brief Post-process the DAG to create cluster edges between instructions
|
|
|
|
/// that may be fused by the processor into a single operation.
|
|
|
|
class MacroFusion : public ScheduleDAGMutation {
|
|
|
|
const TargetInstrInfo *TII;
|
|
|
|
public:
|
|
|
|
MacroFusion(const TargetInstrInfo *tii): TII(tii) {}
|
|
|
|
|
2014-03-07 09:26:03 +00:00
|
|
|
void apply(ScheduleDAGMI *DAG) override;
|
2012-11-12 19:52:20 +00:00
|
|
|
};
|
|
|
|
} // anonymous
|
|
|
|
|
|
|
|
/// \brief Callback from DAG postProcessing to create cluster edges to encourage
|
|
|
|
/// fused operations.
|
|
|
|
void MacroFusion::apply(ScheduleDAGMI *DAG) {
|
|
|
|
// For now, assume targets can only fuse with the branch.
|
|
|
|
MachineInstr *Branch = DAG->ExitSU.getInstr();
|
|
|
|
if (!Branch)
|
|
|
|
return;
|
|
|
|
|
|
|
|
for (unsigned Idx = DAG->SUnits.size(); Idx > 0;) {
|
|
|
|
SUnit *SU = &DAG->SUnits[--Idx];
|
|
|
|
if (!TII->shouldScheduleAdjacent(SU->getInstr(), Branch))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Create a single weak edge from SU to ExitSU. The only effect is to cause
|
|
|
|
// bottom-up scheduling to heavily prioritize the clustered SU. There is no
|
|
|
|
// need to copy predecessor edges from ExitSU to SU, since top-down
|
|
|
|
// scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
|
|
|
|
// of SU, we could create an artificial edge from the deepest root, but it
|
|
|
|
// hasn't been needed yet.
|
|
|
|
bool Success = DAG->addEdge(&DAG->ExitSU, SDep(SU, SDep::Cluster));
|
|
|
|
(void)Success;
|
|
|
|
assert(Success && "No DAG nodes should be reachable from ExitSU");
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "Macro Fuse SU(" << SU->NodeNum << ")\n");
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-04-24 15:54:43 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// CopyConstrain - DAG post-processing to encourage copy elimination.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
/// \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;
|
2013-04-24 23:19:56 +00:00
|
|
|
// RegionEndIdx is the slot index of the last non-debug instruction in the
|
|
|
|
// scheduling region. So we may have RegionBeginIdx == RegionEndIdx.
|
2013-04-24 15:54:43 +00:00
|
|
|
SlotIndex RegionEndIdx;
|
|
|
|
public:
|
|
|
|
CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {}
|
|
|
|
|
2014-03-07 09:26:03 +00:00
|
|
|
void apply(ScheduleDAGMI *DAG) override;
|
2013-04-24 15:54:43 +00:00
|
|
|
|
|
|
|
protected:
|
2013-12-28 21:56:47 +00:00
|
|
|
void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG);
|
2013-04-24 15:54:43 +00:00
|
|
|
};
|
|
|
|
} // anonymous
|
|
|
|
|
|
|
|
/// 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.
|
2013-12-28 21:56:47 +00:00
|
|
|
void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG) {
|
2013-04-24 15:54:43 +00:00
|
|
|
LiveIntervals *LIS = DAG->getLIS();
|
|
|
|
MachineInstr *Copy = CopySU->getInstr();
|
|
|
|
|
|
|
|
// Check for pure vreg copies.
|
|
|
|
unsigned SrcReg = Copy->getOperand(1).getReg();
|
|
|
|
if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
|
|
|
|
return;
|
|
|
|
|
|
|
|
unsigned DstReg = Copy->getOperand(0).getReg();
|
|
|
|
if (!TargetRegisterInfo::isVirtualRegister(DstReg))
|
|
|
|
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.
|
|
|
|
unsigned LocalReg = DstReg;
|
|
|
|
unsigned GlobalReg = SrcReg;
|
|
|
|
LiveInterval *LocalLI = &LIS->getInterval(LocalReg);
|
|
|
|
if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) {
|
|
|
|
LocalReg = SrcReg;
|
|
|
|
GlobalReg = DstReg;
|
|
|
|
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.
|
2014-03-02 12:27:27 +00:00
|
|
|
if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->end,
|
2013-04-24 15:54:43 +00:00
|
|
|
GlobalSegment->start)) {
|
|
|
|
return;
|
|
|
|
}
|
2013-07-30 19:59:08 +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.
|
2014-03-02 12:27:27 +00:00
|
|
|
if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->start,
|
2013-07-30 19:59:08 +00:00
|
|
|
LocalLI->beginIndex())) {
|
|
|
|
return;
|
|
|
|
}
|
2013-04-24 15:54:43 +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.
|
2014-03-02 12:27:27 +00:00
|
|
|
assert(std::prev(GlobalSegment)->start < LocalLI->beginIndex() &&
|
2013-04-24 15:54:43 +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);
|
|
|
|
for (SUnit::const_succ_iterator
|
|
|
|
I = LastLocalSU->Succs.begin(), E = LastLocalSU->Succs.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
if (I->getKind() != SDep::Data || I->getReg() != LocalReg)
|
|
|
|
continue;
|
|
|
|
if (I->getSUnit() == GlobalSU)
|
|
|
|
continue;
|
|
|
|
if (!DAG->canAddEdge(GlobalSU, I->getSUnit()))
|
|
|
|
return;
|
|
|
|
LocalUses.push_back(I->getSUnit());
|
|
|
|
}
|
|
|
|
// 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);
|
|
|
|
for (SUnit::const_pred_iterator
|
|
|
|
I = GlobalSU->Preds.begin(), E = GlobalSU->Preds.end(); I != E; ++I) {
|
|
|
|
if (I->getKind() != SDep::Anti || I->getReg() != GlobalReg)
|
|
|
|
continue;
|
|
|
|
if (I->getSUnit() == FirstLocalSU)
|
|
|
|
continue;
|
|
|
|
if (!DAG->canAddEdge(FirstLocalSU, I->getSUnit()))
|
|
|
|
return;
|
|
|
|
GlobalUses.push_back(I->getSUnit());
|
|
|
|
}
|
|
|
|
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.
|
|
|
|
void CopyConstrain::apply(ScheduleDAGMI *DAG) {
|
2013-12-28 21:56:47 +00:00
|
|
|
assert(DAG->hasVRegLiveness() && "Expect VRegs with LiveIntervals");
|
|
|
|
|
2013-04-24 23:19:56 +00:00
|
|
|
MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end());
|
|
|
|
if (FirstPos == DAG->end())
|
|
|
|
return;
|
|
|
|
RegionBeginIdx = DAG->getLIS()->getInstructionIndex(&*FirstPos);
|
2013-04-24 15:54:43 +00:00
|
|
|
RegionEndIdx = DAG->getLIS()->getInstructionIndex(
|
|
|
|
&*priorNonDebug(DAG->end(), DAG->begin()));
|
|
|
|
|
|
|
|
for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
|
|
|
|
SUnit *SU = &DAG->SUnits[Idx];
|
|
|
|
if (!SU->getInstr()->isCopy())
|
|
|
|
continue;
|
|
|
|
|
2013-12-28 21:56:47 +00:00
|
|
|
constrainLocalCopy(SU, static_cast<ScheduleDAGMILive*>(DAG));
|
2013-04-24 15:54:43 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-01-17 06:55:03 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
2013-12-07 05:59:44 +00:00
|
|
|
// MachineSchedStrategy helpers used by GenericScheduler, GenericPostScheduler
|
|
|
|
// and possibly other custom schedulers.
|
2013-12-28 21:56:57 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
2012-01-17 06:55:03 +00:00
|
|
|
|
2013-12-05 17:56:02 +00:00
|
|
|
static const unsigned InvalidCycle = ~0U;
|
|
|
|
|
2013-12-07 05:59:44 +00:00
|
|
|
SchedBoundary::~SchedBoundary() { delete HazardRec; }
|
2013-04-13 06:07:40 +00:00
|
|
|
|
2013-12-07 05:59:44 +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;
|
2014-04-14 00:51:57 +00:00
|
|
|
HazardRec = nullptr;
|
2013-12-07 05:59:44 +00:00
|
|
|
}
|
|
|
|
Available.clear();
|
|
|
|
Pending.clear();
|
|
|
|
CheckPending = false;
|
|
|
|
NextSUs.clear();
|
|
|
|
CurrCycle = 0;
|
|
|
|
CurrMOps = 0;
|
|
|
|
MinReadyCycle = UINT_MAX;
|
|
|
|
ExpectedLatency = 0;
|
|
|
|
DependentLatency = 0;
|
|
|
|
RetiredMOps = 0;
|
|
|
|
MaxExecutedResCount = 0;
|
|
|
|
ZoneCritResIdx = 0;
|
|
|
|
IsResourceLimited = false;
|
|
|
|
ReservedCycles.clear();
|
2012-05-10 21:06:19 +00:00
|
|
|
#ifndef NDEBUG
|
2013-12-28 21:56:57 +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).
|
2013-12-07 05:59:44 +00:00
|
|
|
MaxObservedLatency = 0;
|
2012-05-10 21:06:19 +00:00
|
|
|
#endif
|
2013-12-07 05:59:44 +00:00
|
|
|
// Reserve a zero-count for invalid CritResIdx.
|
|
|
|
ExecutedResCounts.resize(1);
|
|
|
|
assert(!ExecutedResCounts[0] && "nonzero count for bad resource");
|
|
|
|
}
|
2012-01-17 06:55:03 +00:00
|
|
|
|
2013-12-07 05:59:44 +00:00
|
|
|
void SchedRemainder::
|
2012-11-07 07:05:09 +00:00
|
|
|
init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
|
|
|
|
reset();
|
|
|
|
if (!SchedModel->hasInstrSchedModel())
|
|
|
|
return;
|
|
|
|
RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
|
|
|
|
for (std::vector<SUnit>::iterator
|
|
|
|
I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
|
|
|
|
const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
|
2013-06-15 05:39:19 +00:00
|
|
|
RemIssueCount += SchedModel->getNumMicroOps(I->getInstr(), SC)
|
|
|
|
* SchedModel->getMicroOpFactor();
|
2012-11-07 07:05:09 +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);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-12-07 05:59:44 +00:00
|
|
|
void SchedBoundary::
|
2012-11-07 07:05:09 +00:00
|
|
|
init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
|
|
|
|
reset();
|
|
|
|
DAG = dag;
|
|
|
|
SchedModel = smodel;
|
|
|
|
Rem = rem;
|
2013-12-05 17:56:02 +00:00
|
|
|
if (SchedModel->hasInstrSchedModel()) {
|
2013-06-15 05:39:19 +00:00
|
|
|
ExecutedResCounts.resize(SchedModel->getNumProcResourceKinds());
|
2013-12-05 17:56:02 +00:00
|
|
|
ReservedCycles.resize(SchedModel->getNumProcResourceKinds(), InvalidCycle);
|
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
|
|
|
|
2013-12-07 05:59:44 +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;
|
2013-09-04 21:00:11 +00:00
|
|
|
|
2013-12-07 05:59:44 +00:00
|
|
|
unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
|
|
|
|
if (ReadyCycle > CurrCycle)
|
|
|
|
return ReadyCycle - CurrCycle;
|
|
|
|
return 0;
|
|
|
|
}
|
2013-09-06 17:32:34 +00:00
|
|
|
|
2013-12-07 05:59:44 +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;
|
|
|
|
}
|
2013-12-05 17:56:02 +00:00
|
|
|
|
2012-06-29 03:23:24 +00:00
|
|
|
/// Does this SU have a hazard within the current instruction group.
|
|
|
|
///
|
|
|
|
/// The scheduler supports two modes of hazard recognition. The first is the
|
|
|
|
/// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
|
|
|
|
/// supports highly complicated in-order reservation tables
|
|
|
|
/// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
|
|
|
|
///
|
|
|
|
/// The second is a streamlined mechanism that checks for hazards based on
|
|
|
|
/// simple counters that the scheduler itself maintains. It explicitly checks
|
|
|
|
/// for instruction dispatch limitations, including the number of micro-ops that
|
|
|
|
/// can dispatch per cycle.
|
|
|
|
///
|
|
|
|
/// TODO: Also check whether the SU must start a new group.
|
2013-12-07 05:59:44 +00:00
|
|
|
bool SchedBoundary::checkHazard(SUnit *SU) {
|
2013-12-28 21:56:57 +00:00
|
|
|
if (HazardRec->isEnabled()
|
|
|
|
&& HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard) {
|
|
|
|
return true;
|
|
|
|
}
|
2012-10-10 05:43:09 +00:00
|
|
|
unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
|
2013-06-15 04:49:49 +00:00
|
|
|
if ((CurrMOps > 0) && (CurrMOps + uops > SchedModel->getIssueWidth())) {
|
2012-11-07 07:05:09 +00:00
|
|
|
DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
|
|
|
|
<< SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
|
2012-06-29 03:23:24 +00:00
|
|
|
return true;
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2013-12-05 17:56:02 +00:00
|
|
|
if (SchedModel->hasInstrSchedModel() && SU->hasReservedResource) {
|
|
|
|
const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
|
|
|
|
for (TargetSchedModel::ProcResIter
|
|
|
|
PI = SchedModel->getWriteProcResBegin(SC),
|
|
|
|
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
|
|
|
|
if (getNextResourceCycle(PI->ProcResourceIdx, PI->Cycles) > CurrCycle)
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
2012-06-29 03:23:24 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2013-06-15 05:39:19 +00:00
|
|
|
// Find the unscheduled node in ReadySUs with the highest latency.
|
2013-12-07 05:59:44 +00:00
|
|
|
unsigned SchedBoundary::
|
2013-06-15 05:39:19 +00:00
|
|
|
findMaxLatency(ArrayRef<SUnit*> ReadySUs) {
|
2014-04-14 00:51:57 +00:00
|
|
|
SUnit *LateSU = nullptr;
|
2013-06-15 05:39:19 +00:00
|
|
|
unsigned RemLatency = 0;
|
|
|
|
for (ArrayRef<SUnit*>::iterator I = ReadySUs.begin(), E = ReadySUs.end();
|
2012-12-18 20:52:56 +00:00
|
|
|
I != E; ++I) {
|
|
|
|
unsigned L = getUnscheduledLatency(*I);
|
2013-06-15 04:49:44 +00:00
|
|
|
if (L > RemLatency) {
|
2012-12-18 20:52:56 +00:00
|
|
|
RemLatency = L;
|
2013-06-15 05:39:19 +00:00
|
|
|
LateSU = *I;
|
2013-06-15 04:49:44 +00:00
|
|
|
}
|
2012-12-18 20:52:56 +00:00
|
|
|
}
|
2013-06-15 05:39:19 +00:00
|
|
|
if (LateSU) {
|
|
|
|
DEBUG(dbgs() << Available.getName() << " RemLatency SU("
|
|
|
|
<< LateSU->NodeNum << ") " << RemLatency << "c\n");
|
2012-12-18 20:52:56 +00:00
|
|
|
}
|
2013-06-15 05:39:19 +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.
|
2013-12-07 05:59:44 +00:00
|
|
|
unsigned SchedBoundary::
|
2013-06-15 05:39:19 +00:00
|
|
|
getOtherResourceCount(unsigned &OtherCritIdx) {
|
2013-07-19 08:55:18 +00:00
|
|
|
OtherCritIdx = 0;
|
2013-06-15 05:39:19 +00:00
|
|
|
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;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (OtherCritIdx) {
|
|
|
|
DEBUG(dbgs() << " " << Available.getName() << " + Remain CritRes: "
|
|
|
|
<< OtherCritCount / SchedModel->getResourceFactor(OtherCritIdx)
|
2013-12-07 05:59:44 +00:00
|
|
|
<< " " << SchedModel->getResourceName(OtherCritIdx) << "\n");
|
2013-06-15 05:39:19 +00:00
|
|
|
}
|
|
|
|
return OtherCritCount;
|
|
|
|
}
|
|
|
|
|
2013-12-07 05:59:44 +00:00
|
|
|
void SchedBoundary::releaseNode(SUnit *SU, unsigned ReadyCycle) {
|
2012-05-24 22:11:09 +00:00
|
|
|
if (ReadyCycle < MinReadyCycle)
|
|
|
|
MinReadyCycle = ReadyCycle;
|
|
|
|
|
|
|
|
// Check for interlocks first. For the purpose of other heuristics, an
|
|
|
|
// instruction that cannot issue appears as if it's not in the ReadyQueue.
|
2013-06-15 05:39:19 +00:00
|
|
|
bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
|
|
|
|
if ((!IsBuffered && ReadyCycle > CurrCycle) || checkHazard(SU))
|
2012-05-24 22:11:09 +00:00
|
|
|
Pending.push(SU);
|
|
|
|
else
|
|
|
|
Available.push(SU);
|
2012-11-07 07:05:09 +00:00
|
|
|
|
|
|
|
// Record this node as an immediate dependent of the scheduled node.
|
|
|
|
NextSUs.insert(SU);
|
2012-05-24 22:11:09 +00:00
|
|
|
}
|
|
|
|
|
2013-12-07 05:59:44 +00:00
|
|
|
void SchedBoundary::releaseTopNode(SUnit *SU) {
|
|
|
|
if (SU->isScheduled)
|
|
|
|
return;
|
|
|
|
|
|
|
|
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
if (I->isWeak())
|
|
|
|
continue;
|
|
|
|
unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
|
|
|
|
unsigned Latency = I->getLatency();
|
|
|
|
#ifndef NDEBUG
|
|
|
|
MaxObservedLatency = std::max(Latency, MaxObservedLatency);
|
|
|
|
#endif
|
|
|
|
if (SU->TopReadyCycle < PredReadyCycle + Latency)
|
|
|
|
SU->TopReadyCycle = PredReadyCycle + Latency;
|
|
|
|
}
|
|
|
|
releaseNode(SU, SU->TopReadyCycle);
|
|
|
|
}
|
|
|
|
|
|
|
|
void SchedBoundary::releaseBottomNode(SUnit *SU) {
|
|
|
|
if (SU->isScheduled)
|
|
|
|
return;
|
|
|
|
|
|
|
|
assert(SU->getInstr() && "Scheduled SUnit must have instr");
|
|
|
|
|
|
|
|
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
if (I->isWeak())
|
|
|
|
continue;
|
|
|
|
unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
|
|
|
|
unsigned Latency = I->getLatency();
|
|
|
|
#ifndef NDEBUG
|
|
|
|
MaxObservedLatency = std::max(Latency, MaxObservedLatency);
|
|
|
|
#endif
|
|
|
|
if (SU->BotReadyCycle < SuccReadyCycle + Latency)
|
|
|
|
SU->BotReadyCycle = SuccReadyCycle + Latency;
|
|
|
|
}
|
|
|
|
releaseNode(SU, SU->BotReadyCycle);
|
|
|
|
}
|
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
/// Move the boundary of scheduled code by one cycle.
|
2013-12-07 05:59:44 +00:00
|
|
|
void SchedBoundary::bumpCycle(unsigned NextCycle) {
|
2013-06-15 05:39:19 +00:00
|
|
|
if (SchedModel->getMicroOpBufferSize() == 0) {
|
|
|
|
assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
|
|
|
|
if (MinReadyCycle > NextCycle)
|
|
|
|
NextCycle = MinReadyCycle;
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2013-06-15 05:39:19 +00:00
|
|
|
// 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.
|
2013-06-15 04:49:44 +00:00
|
|
|
if ((NextCycle - CurrCycle) > DependentLatency)
|
|
|
|
DependentLatency = 0;
|
|
|
|
else
|
|
|
|
DependentLatency -= (NextCycle - CurrCycle);
|
2012-05-24 22:11:09 +00:00
|
|
|
|
|
|
|
if (!HazardRec->isEnabled()) {
|
2012-06-05 21:11:27 +00:00
|
|
|
// Bypass HazardRec virtual calls.
|
2012-05-24 22:11:09 +00:00
|
|
|
CurrCycle = NextCycle;
|
|
|
|
}
|
|
|
|
else {
|
2012-06-05 21:11:27 +00:00
|
|
|
// Bypass getHazardType calls in case of long latency.
|
2012-05-24 22:11:09 +00:00
|
|
|
for (; CurrCycle != NextCycle; ++CurrCycle) {
|
|
|
|
if (isTop())
|
|
|
|
HazardRec->AdvanceCycle();
|
|
|
|
else
|
|
|
|
HazardRec->RecedeCycle();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
CheckPending = true;
|
2013-06-15 05:39:19 +00:00
|
|
|
unsigned LFactor = SchedModel->getLatencyFactor();
|
|
|
|
IsResourceLimited =
|
|
|
|
(int)(getCriticalCount() - (getScheduledLatency() * LFactor))
|
|
|
|
> (int)LFactor;
|
2012-05-24 22:11:09 +00:00
|
|
|
|
2013-06-15 05:39:19 +00:00
|
|
|
DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() << '\n');
|
|
|
|
}
|
|
|
|
|
2013-12-07 05:59:44 +00:00
|
|
|
void SchedBoundary::incExecutedResources(unsigned PIdx, unsigned Count) {
|
2013-06-15 05:39:19 +00:00
|
|
|
ExecutedResCounts[PIdx] += Count;
|
|
|
|
if (ExecutedResCounts[PIdx] > MaxExecutedResCount)
|
|
|
|
MaxExecutedResCount = ExecutedResCounts[PIdx];
|
2012-05-24 22:11:09 +00:00
|
|
|
}
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
/// Add the given processor resource to this scheduled zone.
|
2013-06-15 05:39:19 +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.
|
2013-12-07 05:59:44 +00:00
|
|
|
unsigned SchedBoundary::
|
2013-12-05 17:56:02 +00:00
|
|
|
countResource(unsigned PIdx, unsigned Cycles, unsigned NextCycle) {
|
2012-11-07 07:05:09 +00:00
|
|
|
unsigned Factor = SchedModel->getResourceFactor(PIdx);
|
|
|
|
unsigned Count = Factor * Cycles;
|
2013-12-07 05:59:44 +00:00
|
|
|
DEBUG(dbgs() << " " << SchedModel->getResourceName(PIdx)
|
2013-06-15 05:39:19 +00:00
|
|
|
<< " +" << Cycles << "x" << Factor << "u\n");
|
|
|
|
|
|
|
|
// Update Executed resources counts.
|
|
|
|
incExecutedResources(PIdx, Count);
|
2012-11-07 07:05:09 +00:00
|
|
|
assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
|
|
|
|
Rem->RemainingCounts[PIdx] -= Count;
|
|
|
|
|
2013-07-19 00:20:07 +00:00
|
|
|
// Check if this resource exceeds the current critical resource. If so, it
|
|
|
|
// becomes the critical resource.
|
|
|
|
if (ZoneCritResIdx != PIdx && (getResourceCount(PIdx) > getCriticalCount())) {
|
2013-06-15 05:39:19 +00:00
|
|
|
ZoneCritResIdx = PIdx;
|
2012-11-07 07:05:09 +00:00
|
|
|
DEBUG(dbgs() << " *** Critical resource "
|
2013-12-07 05:59:44 +00:00
|
|
|
<< SchedModel->getResourceName(PIdx) << ": "
|
2013-06-15 05:39:19 +00:00
|
|
|
<< getResourceCount(PIdx) / SchedModel->getLatencyFactor() << "c\n");
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2013-12-05 17:56:02 +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;
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
|
|
|
|
2012-06-05 21:11:27 +00:00
|
|
|
/// Move the boundary of scheduled code by one SUnit.
|
2013-12-07 05:59:44 +00:00
|
|
|
void SchedBoundary::bumpNode(SUnit *SU) {
|
2012-06-05 21:11:27 +00:00
|
|
|
// Update the reservation table.
|
|
|
|
if (HazardRec->isEnabled()) {
|
|
|
|
if (!isTop() && SU->isCall) {
|
|
|
|
// Calls are scheduled with their preceding instructions. For bottom-up
|
|
|
|
// scheduling, clear the pipeline state before emitting.
|
|
|
|
HazardRec->Reset();
|
|
|
|
}
|
|
|
|
HazardRec->EmitInstruction(SU);
|
|
|
|
}
|
2013-12-05 17:56:02 +00:00
|
|
|
// checkHazard should prevent scheduling multiple instructions per cycle that
|
|
|
|
// exceed the issue width.
|
2013-06-15 05:39:19 +00:00
|
|
|
const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
|
|
|
|
unsigned IncMOps = SchedModel->getNumMicroOps(SU->getInstr());
|
2013-12-06 08:58:22 +00:00
|
|
|
assert(
|
|
|
|
(CurrMOps == 0 || (CurrMOps + IncMOps) <= SchedModel->getIssueWidth()) &&
|
2013-12-06 17:19:20 +00:00
|
|
|
"Cannot schedule this instruction's MicroOps in the current cycle.");
|
2013-12-05 17:56:02 +00:00
|
|
|
|
2013-06-15 05:39:19 +00:00
|
|
|
unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
|
|
|
|
DEBUG(dbgs() << " Ready @" << ReadyCycle << "c\n");
|
|
|
|
|
2013-12-05 17:56:02 +00:00
|
|
|
unsigned NextCycle = CurrCycle;
|
2013-06-15 05:39:19 +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
|
2013-12-05 17:55:58 +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;
|
2013-06-15 05:39:19 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
RetiredMOps += IncMOps;
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
// Update resource counts and critical resource.
|
|
|
|
if (SchedModel->hasInstrSchedModel()) {
|
2013-06-15 05:39:19 +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");
|
|
|
|
}
|
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
for (TargetSchedModel::ProcResIter
|
|
|
|
PI = SchedModel->getWriteProcResBegin(SC),
|
|
|
|
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
|
2013-06-15 05:39:19 +00:00
|
|
|
unsigned RCycle =
|
2013-12-05 17:56:02 +00:00
|
|
|
countResource(PI->ProcResourceIdx, PI->Cycles, NextCycle);
|
2013-06-15 05:39:19 +00:00
|
|
|
if (RCycle > NextCycle)
|
|
|
|
NextCycle = RCycle;
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2013-12-05 17:56:02 +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;
|
2013-12-28 21:56:57 +00:00
|
|
|
if (SchedModel->getProcResource(PIdx)->BufferSize == 0) {
|
2013-12-05 17:56:02 +00:00
|
|
|
ReservedCycles[PIdx] = isTop() ? NextCycle + PI->Cycles : NextCycle;
|
2013-12-28 21:56:57 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
MaxObservedLatency = std::max(PI->Cycles, MaxObservedLatency);
|
|
|
|
#endif
|
|
|
|
}
|
2013-12-05 17:56:02 +00:00
|
|
|
}
|
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2013-06-15 05:39:19 +00:00
|
|
|
// Update ExpectedLatency and DependentLatency.
|
2013-06-15 04:49:44 +00:00
|
|
|
unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency;
|
|
|
|
unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency;
|
2013-06-15 05:39:19 +00:00
|
|
|
if (SU->getDepth() > TopLatency) {
|
2013-06-15 04:49:44 +00:00
|
|
|
TopLatency = SU->getDepth();
|
2013-06-15 05:39:19 +00:00
|
|
|
DEBUG(dbgs() << " " << Available.getName()
|
|
|
|
<< " TopLatency SU(" << SU->NodeNum << ") " << TopLatency << "c\n");
|
|
|
|
}
|
|
|
|
if (SU->getHeight() > BotLatency) {
|
2013-06-15 04:49:44 +00:00
|
|
|
BotLatency = SU->getHeight();
|
2013-06-15 05:39:19 +00:00
|
|
|
DEBUG(dbgs() << " " << Available.getName()
|
|
|
|
<< " BotLatency SU(" << SU->NodeNum << ") " << BotLatency << "c\n");
|
2012-06-05 21:11:27 +00:00
|
|
|
}
|
2013-06-15 05:39:19 +00:00
|
|
|
// If we stall for any reason, bump the cycle.
|
|
|
|
if (NextCycle > CurrCycle) {
|
|
|
|
bumpCycle(NextCycle);
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
// After updating ZoneCritResIdx and ExpectedLatency, check if we're
|
2014-01-24 17:20:08 +00:00
|
|
|
// resource limited. If a stall occurred, bumpCycle does this.
|
2013-06-15 05:39:19 +00:00
|
|
|
unsigned LFactor = SchedModel->getLatencyFactor();
|
|
|
|
IsResourceLimited =
|
|
|
|
(int)(getCriticalCount() - (getScheduledLatency() * LFactor))
|
|
|
|
> (int)LFactor;
|
|
|
|
}
|
2013-12-05 17:56:02 +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;
|
|
|
|
while (CurrMOps >= SchedModel->getIssueWidth()) {
|
|
|
|
DEBUG(dbgs() << " *** Max MOps " << CurrMOps
|
|
|
|
<< " at cycle " << CurrCycle << '\n');
|
2013-12-28 21:56:57 +00:00
|
|
|
bumpCycle(++NextCycle);
|
2013-12-05 17:56:02 +00:00
|
|
|
}
|
2013-06-15 05:39:19 +00:00
|
|
|
DEBUG(dumpScheduledState());
|
2012-06-05 21:11:27 +00:00
|
|
|
}
|
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
/// Release pending ready nodes in to the available queue. This makes them
|
|
|
|
/// visible to heuristics.
|
2013-12-07 05:59:44 +00:00
|
|
|
void SchedBoundary::releasePending() {
|
2012-05-24 22:11:09 +00:00
|
|
|
// If the available queue is empty, it is safe to reset MinReadyCycle.
|
|
|
|
if (Available.empty())
|
|
|
|
MinReadyCycle = UINT_MAX;
|
|
|
|
|
|
|
|
// Check to see if any of the pending instructions are ready to issue. If
|
|
|
|
// so, add them to the available queue.
|
2013-06-15 05:39:19 +00:00
|
|
|
bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
|
2012-05-24 22:11:09 +00:00
|
|
|
for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
|
|
|
|
SUnit *SU = *(Pending.begin()+i);
|
2012-06-05 21:11:27 +00:00
|
|
|
unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
|
2012-05-24 22:11:09 +00:00
|
|
|
|
|
|
|
if (ReadyCycle < MinReadyCycle)
|
|
|
|
MinReadyCycle = ReadyCycle;
|
|
|
|
|
2013-06-15 05:39:19 +00:00
|
|
|
if (!IsBuffered && ReadyCycle > CurrCycle)
|
2012-05-24 22:11:09 +00:00
|
|
|
continue;
|
|
|
|
|
2012-06-29 03:23:24 +00:00
|
|
|
if (checkHazard(SU))
|
2012-05-24 22:11:09 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
Available.push(SU);
|
|
|
|
Pending.remove(Pending.begin()+i);
|
|
|
|
--i; --e;
|
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
DEBUG(if (!Pending.empty()) Pending.dump());
|
2012-05-24 22:11:09 +00:00
|
|
|
CheckPending = false;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Remove SU from the ready set for this boundary.
|
2013-12-07 05:59:44 +00:00
|
|
|
void SchedBoundary::removeReady(SUnit *SU) {
|
2012-05-24 22:11:09 +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,
|
2012-11-07 07:05:09 +00:00
|
|
|
/// defer any nodes that now hit a hazard, and advance the cycle until at least
|
|
|
|
/// one node is ready. If multiple instructions are ready, return NULL.
|
2013-12-07 05:59:44 +00:00
|
|
|
SUnit *SchedBoundary::pickOnlyChoice() {
|
2012-05-24 22:11:09 +00:00
|
|
|
if (CheckPending)
|
|
|
|
releasePending();
|
|
|
|
|
2013-06-15 04:49:49 +00:00
|
|
|
if (CurrMOps > 0) {
|
2012-11-07 07:05:09 +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;
|
|
|
|
}
|
|
|
|
}
|
2012-05-24 22:11:09 +00:00
|
|
|
for (unsigned i = 0; Available.empty(); ++i) {
|
2013-06-15 04:49:57 +00:00
|
|
|
assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedLatency) &&
|
2012-06-05 21:11:27 +00:00
|
|
|
"permanent hazard"); (void)i;
|
2013-06-15 05:39:19 +00:00
|
|
|
bumpCycle(CurrCycle + 1);
|
2012-05-24 22:11:09 +00:00
|
|
|
releasePending();
|
|
|
|
}
|
|
|
|
if (Available.size() == 1)
|
|
|
|
return *Available.begin();
|
2014-04-14 00:51:57 +00:00
|
|
|
return nullptr;
|
2012-05-24 22:11:09 +00:00
|
|
|
}
|
|
|
|
|
2013-06-15 05:46:47 +00:00
|
|
|
#ifndef NDEBUG
|
2013-06-15 05:39:19 +00:00
|
|
|
// This is useful information to dump after bumpNode.
|
|
|
|
// Note that the Queue contents are more useful before pickNodeFromQueue.
|
2013-12-07 05:59:44 +00:00
|
|
|
void SchedBoundary::dumpScheduledState() {
|
2013-06-15 05:39:19 +00:00
|
|
|
unsigned ResFactor;
|
|
|
|
unsigned ResCount;
|
|
|
|
if (ZoneCritResIdx) {
|
|
|
|
ResFactor = SchedModel->getResourceFactor(ZoneCritResIdx);
|
|
|
|
ResCount = getResourceCount(ZoneCritResIdx);
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2013-06-15 05:39:19 +00:00
|
|
|
else {
|
|
|
|
ResFactor = SchedModel->getMicroOpFactor();
|
|
|
|
ResCount = RetiredMOps * SchedModel->getMicroOpFactor();
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2013-06-15 05:39:19 +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, "
|
2013-12-07 05:59:44 +00:00
|
|
|
<< ResCount / ResFactor << " "
|
|
|
|
<< SchedModel->getResourceName(ZoneCritResIdx)
|
2013-06-15 05:39:19 +00:00
|
|
|
<< "\n ExpectedLatency: " << ExpectedLatency << "c\n"
|
|
|
|
<< (IsResourceLimited ? " - Resource" : " - Latency")
|
|
|
|
<< " limited.\n";
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
2013-06-15 05:46:47 +00:00
|
|
|
#endif
|
2012-11-07 07:05:09 +00:00
|
|
|
|
2013-12-07 05:59:44 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
2013-12-28 21:56:57 +00:00
|
|
|
// GenericScheduler - Generic implementation of MachineSchedStrategy.
|
2013-12-07 05:59:44 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
2012-11-07 07:05:09 +00:00
|
|
|
|
2013-12-07 05:59:44 +00:00
|
|
|
namespace {
|
2013-12-28 21:56:57 +00:00
|
|
|
/// Base class for GenericScheduler. This class maintains information about
|
|
|
|
/// scheduling candidates based on TargetSchedModel making it easy to implement
|
|
|
|
/// heuristics for either preRA or postRA scheduling.
|
|
|
|
class GenericSchedulerBase : public MachineSchedStrategy {
|
2013-12-07 05:59:44 +00:00
|
|
|
public:
|
|
|
|
/// Represent the type of SchedCandidate found within a single queue.
|
|
|
|
/// pickNodeBidirectional depends on these listed by decreasing priority.
|
|
|
|
enum CandReason {
|
|
|
|
NoCand, PhysRegCopy, RegExcess, RegCritical, Stall, Cluster, Weak, RegMax,
|
|
|
|
ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
|
|
|
|
TopDepthReduce, TopPathReduce, NextDefUse, NodeOrder};
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
2013-12-28 21:56:57 +00:00
|
|
|
static const char *getReasonStr(GenericSchedulerBase::CandReason Reason);
|
2013-12-07 05:59:44 +00:00
|
|
|
#endif
|
|
|
|
|
|
|
|
/// Policy for scheduling the next instruction in the candidate's zone.
|
|
|
|
struct CandPolicy {
|
|
|
|
bool ReduceLatency;
|
|
|
|
unsigned ReduceResIdx;
|
|
|
|
unsigned DemandResIdx;
|
|
|
|
|
|
|
|
CandPolicy(): ReduceLatency(false), ReduceResIdx(0), DemandResIdx(0) {}
|
|
|
|
};
|
|
|
|
|
|
|
|
/// Status of an instruction's critical resource consumption.
|
|
|
|
struct SchedResourceDelta {
|
|
|
|
// Count critical resources in the scheduled region required by SU.
|
|
|
|
unsigned CritResources;
|
|
|
|
|
|
|
|
// Count critical resources from another region consumed by SU.
|
|
|
|
unsigned DemandedResources;
|
|
|
|
|
|
|
|
SchedResourceDelta(): CritResources(0), DemandedResources(0) {}
|
|
|
|
|
|
|
|
bool operator==(const SchedResourceDelta &RHS) const {
|
|
|
|
return CritResources == RHS.CritResources
|
|
|
|
&& DemandedResources == RHS.DemandedResources;
|
|
|
|
}
|
|
|
|
bool operator!=(const SchedResourceDelta &RHS) const {
|
|
|
|
return !operator==(RHS);
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
/// Store the state used by GenericScheduler heuristics, required for the
|
|
|
|
/// lifetime of one invocation of pickNode().
|
|
|
|
struct SchedCandidate {
|
|
|
|
CandPolicy Policy;
|
|
|
|
|
|
|
|
// The best SUnit candidate.
|
|
|
|
SUnit *SU;
|
|
|
|
|
|
|
|
// The reason for this candidate.
|
|
|
|
CandReason Reason;
|
|
|
|
|
|
|
|
// Set of reasons that apply to multiple candidates.
|
|
|
|
uint32_t RepeatReasonSet;
|
|
|
|
|
|
|
|
// Register pressure values for the best candidate.
|
|
|
|
RegPressureDelta RPDelta;
|
|
|
|
|
|
|
|
// Critical resource consumption of the best candidate.
|
|
|
|
SchedResourceDelta ResDelta;
|
|
|
|
|
|
|
|
SchedCandidate(const CandPolicy &policy)
|
2014-04-14 00:51:57 +00:00
|
|
|
: Policy(policy), SU(nullptr), Reason(NoCand), RepeatReasonSet(0) {}
|
2013-12-07 05:59:44 +00:00
|
|
|
|
|
|
|
bool isValid() const { return SU; }
|
|
|
|
|
|
|
|
// Copy the status of another candidate without changing policy.
|
|
|
|
void setBest(SchedCandidate &Best) {
|
|
|
|
assert(Best.Reason != NoCand && "uninitialized Sched candidate");
|
|
|
|
SU = Best.SU;
|
|
|
|
Reason = Best.Reason;
|
|
|
|
RPDelta = Best.RPDelta;
|
|
|
|
ResDelta = Best.ResDelta;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool isRepeat(CandReason R) { return RepeatReasonSet & (1 << R); }
|
|
|
|
void setRepeat(CandReason R) { RepeatReasonSet |= (1 << R); }
|
2012-11-07 07:05:09 +00:00
|
|
|
|
2013-12-28 21:56:57 +00:00
|
|
|
void initResourceDelta(const ScheduleDAGMI *DAG,
|
2013-12-07 05:59:44 +00:00
|
|
|
const TargetSchedModel *SchedModel);
|
|
|
|
};
|
|
|
|
|
2013-12-28 21:56:57 +00:00
|
|
|
protected:
|
2013-12-07 05:59:44 +00:00
|
|
|
const MachineSchedContext *Context;
|
|
|
|
const TargetSchedModel *SchedModel;
|
|
|
|
const TargetRegisterInfo *TRI;
|
|
|
|
|
|
|
|
SchedRemainder Rem;
|
2013-12-28 21:56:57 +00:00
|
|
|
protected:
|
|
|
|
GenericSchedulerBase(const MachineSchedContext *C):
|
2014-04-14 00:51:57 +00:00
|
|
|
Context(C), SchedModel(nullptr), TRI(nullptr) {}
|
2013-12-28 21:56:57 +00:00
|
|
|
|
|
|
|
void setPolicy(CandPolicy &Policy, bool IsPostRA, SchedBoundary &CurrZone,
|
|
|
|
SchedBoundary *OtherZone);
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
void traceCandidate(const SchedCandidate &Cand);
|
|
|
|
#endif
|
|
|
|
};
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
void GenericSchedulerBase::SchedCandidate::
|
|
|
|
initResourceDelta(const ScheduleDAGMI *DAG,
|
|
|
|
const TargetSchedModel *SchedModel) {
|
|
|
|
if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
|
|
|
|
return;
|
|
|
|
|
|
|
|
const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
|
|
|
|
for (TargetSchedModel::ProcResIter
|
|
|
|
PI = SchedModel->getWriteProcResBegin(SC),
|
|
|
|
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
|
|
|
|
if (PI->ProcResourceIdx == Policy.ReduceResIdx)
|
|
|
|
ResDelta.CritResources += PI->Cycles;
|
|
|
|
if (PI->ProcResourceIdx == Policy.DemandResIdx)
|
|
|
|
ResDelta.DemandedResources += PI->Cycles;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Set the CandPolicy given a scheduling zone given the current resources and
|
|
|
|
/// latencies inside and outside the zone.
|
|
|
|
void GenericSchedulerBase::setPolicy(CandPolicy &Policy,
|
|
|
|
bool IsPostRA,
|
|
|
|
SchedBoundary &CurrZone,
|
|
|
|
SchedBoundary *OtherZone) {
|
|
|
|
// Apply preemptive heuristics based on the the total latency and resources
|
|
|
|
// 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.
|
2013-12-28 22:25:57 +00:00
|
|
|
unsigned OtherCritIdx = 0;
|
2013-12-28 21:56:57 +00:00
|
|
|
unsigned OtherCount =
|
|
|
|
OtherZone ? OtherZone->getOtherResourceCount(OtherCritIdx) : 0;
|
|
|
|
|
|
|
|
bool OtherResLimited = false;
|
|
|
|
if (SchedModel->hasInstrSchedModel()) {
|
|
|
|
unsigned LFactor = SchedModel->getLatencyFactor();
|
|
|
|
OtherResLimited = (int)(OtherCount - (RemLatency * LFactor)) > (int)LFactor;
|
|
|
|
}
|
|
|
|
// 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 ";
|
|
|
|
case PhysRegCopy: return "PREG-COPY";
|
|
|
|
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;
|
|
|
|
}
|
|
|
|
dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
|
|
|
|
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;
|
|
|
|
}
|
|
|
|
Cand.setRepeat(Reason);
|
|
|
|
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;
|
|
|
|
}
|
|
|
|
Cand.setRepeat(Reason);
|
|
|
|
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;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
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;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void tracePick(const GenericSchedulerBase::SchedCandidate &Cand,
|
|
|
|
bool IsTop) {
|
|
|
|
DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ")
|
|
|
|
<< GenericSchedulerBase::getReasonStr(Cand.Reason) << '\n');
|
|
|
|
}
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
/// GenericScheduler shrinks the unscheduled zone using heuristics to balance
|
|
|
|
/// the schedule.
|
|
|
|
class GenericScheduler : public GenericSchedulerBase {
|
|
|
|
ScheduleDAGMILive *DAG;
|
|
|
|
|
|
|
|
// State of the top and bottom scheduled instruction boundaries.
|
2013-12-07 05:59:44 +00:00
|
|
|
SchedBoundary Top;
|
|
|
|
SchedBoundary Bot;
|
|
|
|
|
|
|
|
MachineSchedPolicy RegionPolicy;
|
|
|
|
public:
|
|
|
|
GenericScheduler(const MachineSchedContext *C):
|
2014-04-14 00:51:57 +00:00
|
|
|
GenericSchedulerBase(C), DAG(nullptr), Top(SchedBoundary::TopQID, "TopQ"),
|
2013-12-28 21:56:57 +00:00
|
|
|
Bot(SchedBoundary::BotQID, "BotQ") {}
|
2013-12-07 05:59:44 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void initPolicy(MachineBasicBlock::iterator Begin,
|
|
|
|
MachineBasicBlock::iterator End,
|
|
|
|
unsigned NumRegionInstrs) override;
|
2013-12-07 05:59:44 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
bool shouldTrackPressure() const override {
|
2013-12-28 21:56:57 +00:00
|
|
|
return RegionPolicy.ShouldTrackPressure;
|
|
|
|
}
|
2013-12-07 05:59:44 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void initialize(ScheduleDAGMI *dag) override;
|
2013-12-07 05:59:44 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
SUnit *pickNode(bool &IsTopNode) override;
|
2013-12-07 05:59:44 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void schedNode(SUnit *SU, bool IsTopNode) override;
|
2013-12-07 05:59:44 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void releaseTopNode(SUnit *SU) override {
|
2013-12-28 21:56:57 +00:00
|
|
|
Top.releaseTopNode(SU);
|
|
|
|
}
|
2013-12-07 05:59:44 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void releaseBottomNode(SUnit *SU) override {
|
2013-12-28 21:56:57 +00:00
|
|
|
Bot.releaseBottomNode(SU);
|
|
|
|
}
|
2013-12-07 05:59:44 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void registerRoots() override;
|
2013-12-07 05:59:44 +00:00
|
|
|
|
|
|
|
protected:
|
|
|
|
void checkAcyclicLatency();
|
|
|
|
|
|
|
|
void tryCandidate(SchedCandidate &Cand,
|
|
|
|
SchedCandidate &TryCand,
|
|
|
|
SchedBoundary &Zone,
|
|
|
|
const RegPressureTracker &RPTracker,
|
|
|
|
RegPressureTracker &TempTracker);
|
|
|
|
|
|
|
|
SUnit *pickNodeBidirectional(bool &IsTopNode);
|
|
|
|
|
|
|
|
void pickNodeFromQueue(SchedBoundary &Zone,
|
|
|
|
const RegPressureTracker &RPTracker,
|
|
|
|
SchedCandidate &Candidate);
|
|
|
|
|
|
|
|
void reschedulePhysRegCopies(SUnit *SU, bool isTop);
|
|
|
|
};
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
void GenericScheduler::initialize(ScheduleDAGMI *dag) {
|
2013-12-28 21:56:47 +00:00
|
|
|
assert(dag->hasVRegLiveness() &&
|
|
|
|
"(PreRA)GenericScheduler needs vreg liveness");
|
|
|
|
DAG = static_cast<ScheduleDAGMILive*>(dag);
|
2013-12-07 05:59:44 +00:00
|
|
|
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();
|
|
|
|
const TargetMachine &TM = DAG->MF.getTarget();
|
|
|
|
if (!Top.HazardRec) {
|
|
|
|
Top.HazardRec =
|
|
|
|
TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
|
|
|
|
}
|
|
|
|
if (!Bot.HazardRec) {
|
|
|
|
Bot.HazardRec =
|
|
|
|
TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Initialize the per-region scheduling policy.
|
|
|
|
void GenericScheduler::initPolicy(MachineBasicBlock::iterator Begin,
|
|
|
|
MachineBasicBlock::iterator End,
|
|
|
|
unsigned NumRegionInstrs) {
|
|
|
|
const TargetMachine &TM = Context->MF->getTarget();
|
2014-01-22 03:38:55 +00:00
|
|
|
const TargetLowering *TLI = TM.getTargetLowering();
|
2013-12-07 05:59:44 +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.
|
2014-01-21 21:27:37 +00:00
|
|
|
RegionPolicy.ShouldTrackPressure = true;
|
2014-01-22 03:38:55 +00:00
|
|
|
for (unsigned VT = MVT::i32; VT > (unsigned)MVT::i1; --VT) {
|
|
|
|
MVT::SimpleValueType LegalIntVT = (MVT::SimpleValueType)VT;
|
|
|
|
if (TLI->isTypeLegal(LegalIntVT)) {
|
2014-01-21 21:27:37 +00:00
|
|
|
unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs(
|
2014-01-22 03:38:55 +00:00
|
|
|
TLI->getRegClassFor(LegalIntVT));
|
2014-01-21 21:27:37 +00:00
|
|
|
RegionPolicy.ShouldTrackPressure = NumRegionInstrs > (NIntRegs / 2);
|
|
|
|
}
|
|
|
|
}
|
2013-12-07 05:59:44 +00:00
|
|
|
|
|
|
|
// 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.
|
|
|
|
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
|
|
|
|
ST.overrideSchedPolicy(RegionPolicy, Begin, End, NumRegionInstrs);
|
|
|
|
|
|
|
|
// 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;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// 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;
|
|
|
|
|
|
|
|
// 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");
|
|
|
|
}
|
|
|
|
|
|
|
|
void GenericScheduler::registerRoots() {
|
|
|
|
Rem.CriticalPath = DAG->ExitSU.getDepth();
|
|
|
|
|
|
|
|
// Some roots may not feed into ExitSU. Check all of them in case.
|
|
|
|
for (std::vector<SUnit*>::const_iterator
|
|
|
|
I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
|
|
|
|
if ((*I)->getDepth() > Rem.CriticalPath)
|
|
|
|
Rem.CriticalPath = (*I)->getDepth();
|
|
|
|
}
|
|
|
|
DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
|
|
|
|
|
|
|
|
if (EnableCyclicPath) {
|
|
|
|
Rem.CyclicCritPath = DAG->computeCyclicCriticalPath();
|
|
|
|
checkAcyclicLatency();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-08-30 03:49:48 +00:00
|
|
|
static bool tryPressure(const PressureChange &TryP,
|
|
|
|
const PressureChange &CandP,
|
2013-12-28 21:56:57 +00:00
|
|
|
GenericSchedulerBase::SchedCandidate &TryCand,
|
|
|
|
GenericSchedulerBase::SchedCandidate &Cand,
|
|
|
|
GenericSchedulerBase::CandReason Reason) {
|
2013-08-30 04:27:29 +00:00
|
|
|
int TryRank = TryP.getPSetOrMax();
|
|
|
|
int CandRank = CandP.getPSetOrMax();
|
|
|
|
// If both candidates affect the same set, go with the smallest increase.
|
|
|
|
if (TryRank == CandRank) {
|
|
|
|
return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand,
|
|
|
|
Reason);
|
|
|
|
}
|
|
|
|
// If one candidate decreases and the other increases, go with it.
|
|
|
|
// Invalid candidates have UnitInc==0.
|
|
|
|
if (tryLess(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand,
|
|
|
|
Reason)) {
|
|
|
|
return true;
|
2013-07-25 07:26:35 +00:00
|
|
|
}
|
|
|
|
// If the candidates are decreasing pressure, reverse priority.
|
2013-08-30 03:49:48 +00:00
|
|
|
if (TryP.getUnitInc() < 0)
|
2013-07-25 07:26:35 +00:00
|
|
|
std::swap(TryRank, CandRank);
|
|
|
|
return tryGreater(TryRank, CandRank, TryCand, Cand, Reason);
|
|
|
|
}
|
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
|
|
|
|
return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
|
|
|
|
}
|
|
|
|
|
2013-04-13 06:07:40 +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;
|
|
|
|
}
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
/// Apply a set of heursitics to a new candidate. Heuristics are currently
|
|
|
|
/// hierarchical. This may be more efficient than a graduated cost model because
|
|
|
|
/// we don't need to evaluate all aspects of the model for each node in the
|
|
|
|
/// queue. But it's really done to make the heuristics easier to debug and
|
|
|
|
/// statistically analyze.
|
|
|
|
///
|
|
|
|
/// \param Cand provides the policy and current best candidate.
|
|
|
|
/// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
|
|
|
|
/// \param Zone describes the scheduled zone that we are extending.
|
|
|
|
/// \param RPTracker describes reg pressure within the scheduled zone.
|
|
|
|
/// \param TempTracker is a scratch pressure tracker to reuse in queries.
|
2013-09-19 23:10:59 +00:00
|
|
|
void GenericScheduler::tryCandidate(SchedCandidate &Cand,
|
2013-12-05 17:55:47 +00:00
|
|
|
SchedCandidate &TryCand,
|
|
|
|
SchedBoundary &Zone,
|
|
|
|
const RegPressureTracker &RPTracker,
|
|
|
|
RegPressureTracker &TempTracker) {
|
2012-11-07 07:05:09 +00:00
|
|
|
|
2013-09-04 21:00:11 +00:00
|
|
|
if (DAG->isTrackingPressure()) {
|
2013-09-04 21:00:02 +00:00
|
|
|
// Always initialize TryCand's RPDelta.
|
|
|
|
if (Zone.isTop()) {
|
|
|
|
TempTracker.getMaxDownwardPressureDelta(
|
2013-08-30 03:49:48 +00:00
|
|
|
TryCand.SU->getInstr(),
|
|
|
|
TryCand.RPDelta,
|
|
|
|
DAG->getRegionCriticalPSets(),
|
|
|
|
DAG->getRegPressure().MaxSetPressure);
|
|
|
|
}
|
|
|
|
else {
|
2013-09-04 21:00:02 +00:00
|
|
|
if (VerifyScheduling) {
|
|
|
|
TempTracker.getMaxUpwardPressureDelta(
|
|
|
|
TryCand.SU->getInstr(),
|
|
|
|
&DAG->getPressureDiff(TryCand.SU),
|
|
|
|
TryCand.RPDelta,
|
|
|
|
DAG->getRegionCriticalPSets(),
|
|
|
|
DAG->getRegPressure().MaxSetPressure);
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
RPTracker.getUpwardPressureDelta(
|
|
|
|
TryCand.SU->getInstr(),
|
|
|
|
DAG->getPressureDiff(TryCand.SU),
|
|
|
|
TryCand.RPDelta,
|
|
|
|
DAG->getRegionCriticalPSets(),
|
|
|
|
DAG->getRegPressure().MaxSetPressure);
|
|
|
|
}
|
2013-08-30 03:49:48 +00:00
|
|
|
}
|
|
|
|
}
|
2013-09-06 17:32:44 +00:00
|
|
|
DEBUG(if (TryCand.RPDelta.Excess.isValid())
|
|
|
|
dbgs() << " SU(" << TryCand.SU->NodeNum << ") "
|
|
|
|
<< TRI->getRegPressureSetName(TryCand.RPDelta.Excess.getPSet())
|
|
|
|
<< ":" << TryCand.RPDelta.Excess.getUnitInc() << "\n");
|
2012-11-07 07:05:09 +00:00
|
|
|
|
|
|
|
// Initialize the candidate if needed.
|
|
|
|
if (!Cand.isValid()) {
|
|
|
|
TryCand.Reason = NodeOrder;
|
|
|
|
return;
|
|
|
|
}
|
2013-04-13 06:07:40 +00:00
|
|
|
|
|
|
|
if (tryGreater(biasPhysRegCopy(TryCand.SU, Zone.isTop()),
|
|
|
|
biasPhysRegCopy(Cand.SU, Zone.isTop()),
|
|
|
|
TryCand, Cand, PhysRegCopy))
|
|
|
|
return;
|
|
|
|
|
2013-07-25 07:26:35 +00:00
|
|
|
// Avoid exceeding the target's limit. If signed PSetID is negative, it is
|
|
|
|
// invalid; convert it to INT_MAX to give it lowest priority.
|
2013-09-04 21:00:11 +00:00
|
|
|
if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess,
|
|
|
|
Cand.RPDelta.Excess,
|
|
|
|
TryCand, Cand, RegExcess))
|
2012-11-07 07:05:09 +00:00
|
|
|
return;
|
|
|
|
|
|
|
|
// Avoid increasing the max critical pressure in the scheduled region.
|
2013-09-04 21:00:11 +00:00
|
|
|
if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax,
|
|
|
|
Cand.RPDelta.CriticalMax,
|
|
|
|
TryCand, Cand, RegCritical))
|
2012-11-07 07:05:09 +00:00
|
|
|
return;
|
|
|
|
|
2013-09-06 17:32:36 +00:00
|
|
|
// For loops that are acyclic path limited, aggressively schedule for latency.
|
2013-09-09 22:28:08 +00:00
|
|
|
// This can result in very long dependence chains scheduled in sequence, so
|
|
|
|
// once every cycle (when CurrMOps == 0), switch to normal heuristics.
|
2013-12-07 05:59:44 +00:00
|
|
|
if (Rem.IsAcyclicLatencyLimited && !Zone.getCurrMOps()
|
2013-09-09 22:28:08 +00:00
|
|
|
&& tryLatency(TryCand, Cand, Zone))
|
2013-09-06 17:32:36 +00:00
|
|
|
return;
|
|
|
|
|
2013-12-05 17:55:58 +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;
|
|
|
|
|
2012-11-12 19:40:10 +00:00
|
|
|
// Keep clustered nodes together to encourage downstream peephole
|
|
|
|
// optimizations which may reduce resource requirements.
|
|
|
|
//
|
|
|
|
// This is a best effort to set things up for a post-RA pass. Optimizations
|
|
|
|
// like generating loads of multiple registers should ideally be done within
|
|
|
|
// the scheduler pass by combining the loads during DAG postprocessing.
|
|
|
|
const SUnit *NextClusterSU =
|
|
|
|
Zone.isTop() ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
|
|
|
|
if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
|
|
|
|
TryCand, Cand, Cluster))
|
|
|
|
return;
|
2013-04-24 15:54:43 +00:00
|
|
|
|
|
|
|
// Weak edges are for clustering and other constraints.
|
2012-11-12 19:40:10 +00:00
|
|
|
if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
|
|
|
|
getWeakLeft(Cand.SU, Zone.isTop()),
|
2013-04-24 15:54:43 +00:00
|
|
|
TryCand, Cand, Weak)) {
|
2012-11-12 19:40:10 +00:00
|
|
|
return;
|
|
|
|
}
|
2013-06-17 21:45:13 +00:00
|
|
|
// Avoid increasing the max pressure of the entire region.
|
2013-09-04 21:00:11 +00:00
|
|
|
if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax,
|
|
|
|
Cand.RPDelta.CurrentMax,
|
|
|
|
TryCand, Cand, RegMax))
|
2013-06-17 21:45:13 +00:00
|
|
|
return;
|
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
// Avoid critical resource consumption and balance the schedule.
|
|
|
|
TryCand.initResourceDelta(DAG, SchedModel);
|
|
|
|
if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
|
|
|
|
TryCand, Cand, ResourceReduce))
|
|
|
|
return;
|
|
|
|
if (tryGreater(TryCand.ResDelta.DemandedResources,
|
|
|
|
Cand.ResDelta.DemandedResources,
|
|
|
|
TryCand, Cand, ResourceDemand))
|
|
|
|
return;
|
|
|
|
|
|
|
|
// Avoid serializing long latency dependence chains.
|
2013-08-23 17:48:43 +00:00
|
|
|
// For acyclic path limited loops, latency was already checked above.
|
|
|
|
if (Cand.Policy.ReduceLatency && !Rem.IsAcyclicLatencyLimited
|
|
|
|
&& tryLatency(TryCand, Cand, Zone)) {
|
|
|
|
return;
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Prefer immediate defs/users of the last scheduled instruction. This is a
|
2013-06-15 05:39:19 +00:00
|
|
|
// local pressure avoidance strategy that also makes the machine code
|
|
|
|
// readable.
|
2013-12-07 05:59:44 +00:00
|
|
|
if (tryGreater(Zone.isNextSU(TryCand.SU), Zone.isNextSU(Cand.SU),
|
2012-11-12 19:40:10 +00:00
|
|
|
TryCand, Cand, NextDefUse))
|
2012-11-07 07:05:09 +00:00
|
|
|
return;
|
2012-11-12 19:40:10 +00:00
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
// Fall through to original instruction order.
|
|
|
|
if ((Zone.isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
|
|
|
|
|| (!Zone.isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
|
|
|
|
TryCand.Reason = NodeOrder;
|
|
|
|
}
|
2012-05-10 21:06:19 +00:00
|
|
|
}
|
|
|
|
|
2013-09-06 17:32:44 +00:00
|
|
|
/// Pick the best candidate from the queue.
|
2012-05-10 21:06:16 +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.
|
2013-09-19 23:10:59 +00:00
|
|
|
void GenericScheduler::pickNodeFromQueue(SchedBoundary &Zone,
|
2013-12-05 17:55:47 +00:00
|
|
|
const RegPressureTracker &RPTracker,
|
|
|
|
SchedCandidate &Cand) {
|
2012-11-07 07:05:09 +00:00
|
|
|
ReadyQueue &Q = Zone.Available;
|
|
|
|
|
2012-05-24 22:11:12 +00:00
|
|
|
DEBUG(Q.dump());
|
2012-05-17 18:35:10 +00:00
|
|
|
|
2012-05-10 21:06:16 +00:00
|
|
|
// getMaxPressureDelta temporarily modifies the tracker.
|
|
|
|
RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
|
|
|
|
|
2012-05-24 22:11:03 +00:00
|
|
|
for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
|
2012-05-10 21:06:16 +00:00
|
|
|
|
2012-11-07 07:05:09 +00:00
|
|
|
SchedCandidate TryCand(Cand.Policy);
|
|
|
|
TryCand.SU = *I;
|
|
|
|
tryCandidate(Cand, TryCand, Zone, RPTracker, TempTracker);
|
|
|
|
if (TryCand.Reason != NoCand) {
|
|
|
|
// Initialize resource delta if needed in case future heuristics query it.
|
|
|
|
if (TryCand.ResDelta == SchedResourceDelta())
|
|
|
|
TryCand.initResourceDelta(DAG, SchedModel);
|
|
|
|
Cand.setBest(TryCand);
|
2013-04-05 00:31:29 +00:00
|
|
|
DEBUG(traceCandidate(Cand));
|
2012-05-10 21:06:16 +00:00
|
|
|
}
|
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
}
|
|
|
|
|
2012-05-17 18:35:10 +00:00
|
|
|
/// Pick the best candidate node from either the top or bottom queue.
|
2013-09-19 23:10:59 +00:00
|
|
|
SUnit *GenericScheduler::pickNodeBidirectional(bool &IsTopNode) {
|
2012-05-17 18:35:10 +00:00
|
|
|
// Schedule as far as possible in the direction of no choice. This is most
|
|
|
|
// efficient, but also provides the best heuristics for CriticalPSets.
|
2012-05-24 22:11:09 +00:00
|
|
|
if (SUnit *SU = Bot.pickOnlyChoice()) {
|
2012-05-17 18:35:10 +00:00
|
|
|
IsTopNode = false;
|
2013-06-15 05:39:19 +00:00
|
|
|
DEBUG(dbgs() << "Pick Bot NOCAND\n");
|
2012-05-24 22:11:09 +00:00
|
|
|
return SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
2012-05-24 22:11:09 +00:00
|
|
|
if (SUnit *SU = Top.pickOnlyChoice()) {
|
2012-05-17 18:35:10 +00:00
|
|
|
IsTopNode = true;
|
2013-06-15 05:39:19 +00:00
|
|
|
DEBUG(dbgs() << "Pick Top NOCAND\n");
|
2012-05-24 22:11:09 +00:00
|
|
|
return SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
2012-11-07 07:05:09 +00:00
|
|
|
CandPolicy NoPolicy;
|
|
|
|
SchedCandidate BotCand(NoPolicy);
|
|
|
|
SchedCandidate TopCand(NoPolicy);
|
2013-12-07 05:59:44 +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.
|
2013-12-28 21:56:57 +00:00
|
|
|
setPolicy(BotCand.Policy, /*IsPostRA=*/false, Bot, &Top);
|
2013-12-07 05:59:44 +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.
|
2013-12-28 21:56:57 +00:00
|
|
|
setPolicy(TopCand.Policy, /*IsPostRA=*/false, Top, &Bot);
|
2012-11-07 07:05:09 +00:00
|
|
|
|
2012-05-17 18:35:10 +00:00
|
|
|
// Prefer bottom scheduling when heuristics are silent.
|
2012-11-07 07:05:09 +00:00
|
|
|
pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
|
|
|
|
assert(BotCand.Reason != NoCand && "failed to find the first candidate");
|
2012-05-17 18:35:10 +00:00
|
|
|
|
|
|
|
// If either Q has a single candidate that provides the least increase in
|
|
|
|
// Excess pressure, we can immediately schedule from that Q.
|
|
|
|
//
|
|
|
|
// RegionCriticalPSets summarizes the pressure within the scheduled region and
|
|
|
|
// affects picking from either Q. If scheduling in one direction must
|
|
|
|
// increase pressure for one of the excess PSets, then schedule in that
|
|
|
|
// direction first to provide more freedom in the other direction.
|
2013-06-17 21:45:05 +00:00
|
|
|
if ((BotCand.Reason == RegExcess && !BotCand.isRepeat(RegExcess))
|
|
|
|
|| (BotCand.Reason == RegCritical
|
|
|
|
&& !BotCand.isRepeat(RegCritical)))
|
|
|
|
{
|
2012-05-17 18:35:10 +00:00
|
|
|
IsTopNode = false;
|
2012-11-07 07:05:09 +00:00
|
|
|
tracePick(BotCand, IsTopNode);
|
2012-05-24 22:11:09 +00:00
|
|
|
return BotCand.SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
|
|
|
// Check if the top Q has a better candidate.
|
2012-11-07 07:05:09 +00:00
|
|
|
pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
|
|
|
|
assert(TopCand.Reason != NoCand && "failed to find the first candidate");
|
2012-05-17 18:35:10 +00:00
|
|
|
|
2013-06-17 21:45:05 +00:00
|
|
|
// Choose the queue with the most important (lowest enum) reason.
|
2012-11-07 07:05:09 +00:00
|
|
|
if (TopCand.Reason < BotCand.Reason) {
|
2012-05-17 18:35:10 +00:00
|
|
|
IsTopNode = true;
|
2012-11-07 07:05:09 +00:00
|
|
|
tracePick(TopCand, IsTopNode);
|
2012-05-24 22:11:09 +00:00
|
|
|
return TopCand.SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
2013-06-17 21:45:05 +00:00
|
|
|
// Otherwise prefer the bottom candidate, in node order if all else failed.
|
2012-05-17 18:35:10 +00:00
|
|
|
IsTopNode = false;
|
2012-11-07 07:05:09 +00:00
|
|
|
tracePick(BotCand, IsTopNode);
|
2012-05-24 22:11:09 +00:00
|
|
|
return BotCand.SU;
|
2012-05-17 18:35:10 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
|
2013-09-19 23:10:59 +00:00
|
|
|
SUnit *GenericScheduler::pickNode(bool &IsTopNode) {
|
2012-05-10 21:06:16 +00:00
|
|
|
if (DAG->top() == DAG->bottom()) {
|
2012-05-24 22:11:09 +00:00
|
|
|
assert(Top.Available.empty() && Top.Pending.empty() &&
|
|
|
|
Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
|
2014-04-14 00:51:57 +00:00
|
|
|
return nullptr;
|
2012-05-10 21:06:16 +00:00
|
|
|
}
|
|
|
|
SUnit *SU;
|
2012-10-08 18:53:53 +00:00
|
|
|
do {
|
2013-09-06 17:32:34 +00:00
|
|
|
if (RegionPolicy.OnlyTopDown) {
|
2012-10-08 18:53:53 +00:00
|
|
|
SU = Top.pickOnlyChoice();
|
|
|
|
if (!SU) {
|
2012-11-07 07:05:09 +00:00
|
|
|
CandPolicy NoPolicy;
|
|
|
|
SchedCandidate TopCand(NoPolicy);
|
|
|
|
pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
|
2013-09-04 21:00:13 +00:00
|
|
|
assert(TopCand.Reason != NoCand && "failed to find a candidate");
|
2013-09-04 21:00:16 +00:00
|
|
|
tracePick(TopCand, true);
|
2012-10-08 18:53:53 +00:00
|
|
|
SU = TopCand.SU;
|
|
|
|
}
|
|
|
|
IsTopNode = true;
|
2012-05-24 23:11:17 +00:00
|
|
|
}
|
2013-09-06 17:32:34 +00:00
|
|
|
else if (RegionPolicy.OnlyBottomUp) {
|
2012-10-08 18:53:53 +00:00
|
|
|
SU = Bot.pickOnlyChoice();
|
|
|
|
if (!SU) {
|
2012-11-07 07:05:09 +00:00
|
|
|
CandPolicy NoPolicy;
|
|
|
|
SchedCandidate BotCand(NoPolicy);
|
|
|
|
pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
|
2013-09-04 21:00:13 +00:00
|
|
|
assert(BotCand.Reason != NoCand && "failed to find a candidate");
|
2013-09-04 21:00:16 +00:00
|
|
|
tracePick(BotCand, false);
|
2012-10-08 18:53:53 +00:00
|
|
|
SU = BotCand.SU;
|
|
|
|
}
|
|
|
|
IsTopNode = false;
|
2012-05-24 23:11:17 +00:00
|
|
|
}
|
2012-10-08 18:53:53 +00:00
|
|
|
else {
|
2012-11-07 07:05:09 +00:00
|
|
|
SU = pickNodeBidirectional(IsTopNode);
|
2012-10-08 18:53:53 +00:00
|
|
|
}
|
|
|
|
} while (SU->isScheduled);
|
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
if (SU->isTopReady())
|
|
|
|
Top.removeReady(SU);
|
|
|
|
if (SU->isBottomReady())
|
|
|
|
Bot.removeReady(SU);
|
2012-05-25 02:02:39 +00:00
|
|
|
|
2013-04-13 06:07:49 +00:00
|
|
|
DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
|
2012-05-24 22:11:09 +00:00
|
|
|
return SU;
|
|
|
|
}
|
|
|
|
|
2013-09-19 23:10:59 +00:00
|
|
|
void GenericScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) {
|
2013-04-13 06:07:40 +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.
|
|
|
|
for (SmallVectorImpl<SDep>::iterator I = Deps.begin(), E = Deps.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
if (I->getKind() != SDep::Data || !TRI->isPhysicalRegister(I->getReg()))
|
|
|
|
continue;
|
|
|
|
SUnit *DepSU = I->getSUnit();
|
|
|
|
if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1)
|
|
|
|
continue;
|
|
|
|
MachineInstr *Copy = DepSU->getInstr();
|
|
|
|
if (!Copy->isCopy())
|
|
|
|
continue;
|
|
|
|
DEBUG(dbgs() << " Rescheduling physreg copy ";
|
|
|
|
I->getSUnit()->dump(DAG));
|
|
|
|
DAG->moveInstruction(Copy, InsertPos);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
/// Update the scheduler's state after scheduling a node. This is the same node
|
2013-12-28 21:56:57 +00:00
|
|
|
/// that was just returned by pickNode(). However, ScheduleDAGMILive needs to
|
|
|
|
/// update it's state based on the current cycle before MachineSchedStrategy
|
|
|
|
/// does.
|
2013-04-13 06:07:40 +00:00
|
|
|
///
|
|
|
|
/// FIXME: Eventually, we may bundle physreg copies rather than rescheduling
|
|
|
|
/// them here. See comments in biasPhysRegCopy.
|
2013-09-19 23:10:59 +00:00
|
|
|
void GenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
|
2012-06-05 21:11:27 +00:00
|
|
|
if (IsTopNode) {
|
2013-12-07 05:59:44 +00:00
|
|
|
SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
|
2012-06-29 03:23:22 +00:00
|
|
|
Top.bumpNode(SU);
|
2013-04-13 06:07:40 +00:00
|
|
|
if (SU->hasPhysRegUses)
|
|
|
|
reschedulePhysRegCopies(SU, true);
|
2012-05-10 21:06:16 +00:00
|
|
|
}
|
2012-06-05 21:11:27 +00:00
|
|
|
else {
|
2013-12-07 05:59:44 +00:00
|
|
|
SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.getCurrCycle());
|
2012-06-29 03:23:22 +00:00
|
|
|
Bot.bumpNode(SU);
|
2013-04-13 06:07:40 +00:00
|
|
|
if (SU->hasPhysRegDefs)
|
|
|
|
reschedulePhysRegCopies(SU, false);
|
2012-05-10 21:06:16 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
/// Create the standard converging machine scheduler. This will be used as the
|
|
|
|
/// default scheduler if the target does not set a default.
|
2013-12-28 21:56:57 +00:00
|
|
|
static ScheduleDAGInstrs *createGenericSchedLive(MachineSchedContext *C) {
|
2014-04-21 20:32:32 +00:00
|
|
|
ScheduleDAGMILive *DAG = new ScheduleDAGMILive(C, make_unique<GenericScheduler>(C));
|
2012-11-12 19:40:10 +00:00
|
|
|
// Register DAG post-processors.
|
2013-04-24 15:54:43 +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.
|
2014-04-21 20:32:32 +00:00
|
|
|
DAG->addMutation(make_unique<CopyConstrain>(DAG->TII, DAG->TRI));
|
2013-09-04 21:00:08 +00:00
|
|
|
if (EnableLoadCluster && DAG->TII->enableClusterLoads())
|
2014-04-21 20:32:32 +00:00
|
|
|
DAG->addMutation(make_unique<LoadClusterMutation>(DAG->TII, DAG->TRI));
|
2012-11-12 19:52:20 +00:00
|
|
|
if (EnableMacroFusion)
|
2014-04-21 20:32:32 +00:00
|
|
|
DAG->addMutation(make_unique<MacroFusion>(DAG->TII));
|
2012-11-12 19:40:10 +00:00
|
|
|
return DAG;
|
2012-01-17 06:55:03 +00:00
|
|
|
}
|
2013-12-28 21:56:57 +00:00
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
static MachineSchedRegistry
|
2013-09-19 23:10:59 +00:00
|
|
|
GenericSchedRegistry("converge", "Standard converging scheduler.",
|
2013-12-28 21:56:57 +00:00
|
|
|
createGenericSchedLive);
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// PostGenericScheduler - Generic PostRA implementation of MachineSchedStrategy.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
/// PostGenericScheduler - Interface to the scheduling algorithm used by
|
|
|
|
/// ScheduleDAGMI.
|
|
|
|
///
|
|
|
|
/// Callbacks from ScheduleDAGMI:
|
|
|
|
/// initPolicy -> initialize(DAG) -> registerRoots -> pickNode ...
|
|
|
|
class PostGenericScheduler : public GenericSchedulerBase {
|
|
|
|
ScheduleDAGMI *DAG;
|
|
|
|
SchedBoundary Top;
|
|
|
|
SmallVector<SUnit*, 8> BotRoots;
|
|
|
|
public:
|
|
|
|
PostGenericScheduler(const MachineSchedContext *C):
|
|
|
|
GenericSchedulerBase(C), Top(SchedBoundary::TopQID, "TopQ") {}
|
|
|
|
|
|
|
|
virtual ~PostGenericScheduler() {}
|
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void initPolicy(MachineBasicBlock::iterator Begin,
|
|
|
|
MachineBasicBlock::iterator End,
|
|
|
|
unsigned NumRegionInstrs) override {
|
2013-12-28 21:56:57 +00:00
|
|
|
/* no configurable policy */
|
|
|
|
};
|
|
|
|
|
|
|
|
/// PostRA scheduling does not track pressure.
|
2014-03-10 05:29:18 +00:00
|
|
|
bool shouldTrackPressure() const override { return false; }
|
2013-12-28 21:56:57 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void initialize(ScheduleDAGMI *Dag) override {
|
2013-12-28 21:56:57 +00:00
|
|
|
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();
|
|
|
|
const TargetMachine &TM = DAG->MF.getTarget();
|
|
|
|
if (!Top.HazardRec) {
|
|
|
|
Top.HazardRec =
|
|
|
|
TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void registerRoots() override;
|
2013-12-28 21:56:57 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
SUnit *pickNode(bool &IsTopNode) override;
|
2013-12-28 21:56:57 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void scheduleTree(unsigned SubtreeID) override {
|
2013-12-28 21:56:57 +00:00
|
|
|
llvm_unreachable("PostRA scheduler does not support subtree analysis.");
|
|
|
|
}
|
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void schedNode(SUnit *SU, bool IsTopNode) override;
|
2013-12-28 21:56:57 +00:00
|
|
|
|
2014-03-10 05:29:18 +00:00
|
|
|
void releaseTopNode(SUnit *SU) override {
|
2013-12-28 21:56:57 +00:00
|
|
|
Top.releaseTopNode(SU);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Only called for roots.
|
2014-03-10 05:29:18 +00:00
|
|
|
void releaseBottomNode(SUnit *SU) override {
|
2013-12-28 21:56:57 +00:00
|
|
|
BotRoots.push_back(SU);
|
|
|
|
}
|
|
|
|
|
|
|
|
protected:
|
|
|
|
void tryCandidate(SchedCandidate &Cand, SchedCandidate &TryCand);
|
|
|
|
|
|
|
|
void pickNodeFromQueue(SchedCandidate &Cand);
|
|
|
|
};
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
void PostGenericScheduler::registerRoots() {
|
|
|
|
Rem.CriticalPath = DAG->ExitSU.getDepth();
|
|
|
|
|
|
|
|
// Some roots may not feed into ExitSU. Check all of them in case.
|
|
|
|
for (SmallVectorImpl<SUnit*>::const_iterator
|
|
|
|
I = BotRoots.begin(), E = BotRoots.end(); I != E; ++I) {
|
|
|
|
if ((*I)->getDepth() > Rem.CriticalPath)
|
|
|
|
Rem.CriticalPath = (*I)->getDepth();
|
|
|
|
}
|
|
|
|
DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
|
|
|
|
}
|
|
|
|
|
|
|
|
/// 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;
|
|
|
|
|
|
|
|
// 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;
|
|
|
|
|
|
|
|
DEBUG(Q.dump());
|
|
|
|
|
|
|
|
for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
|
|
|
|
SchedCandidate TryCand(Cand.Policy);
|
|
|
|
TryCand.SU = *I;
|
|
|
|
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");
|
2014-04-14 00:51:57 +00:00
|
|
|
return nullptr;
|
2013-12-28 21:56:57 +00:00
|
|
|
}
|
|
|
|
SUnit *SU;
|
|
|
|
do {
|
|
|
|
SU = Top.pickOnlyChoice();
|
|
|
|
if (!SU) {
|
|
|
|
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.
|
2014-04-14 00:51:57 +00:00
|
|
|
setPolicy(TopCand.Policy, /*IsPostRA=*/true, Top, nullptr);
|
2013-12-28 21:56:57 +00:00
|
|
|
pickNodeFromQueue(TopCand);
|
|
|
|
assert(TopCand.Reason != NoCand && "failed to find a candidate");
|
|
|
|
tracePick(TopCand, true);
|
|
|
|
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);
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Create a generic scheduler with no vreg liveness or DAG mutation passes.
|
|
|
|
static ScheduleDAGInstrs *createGenericSchedPostRA(MachineSchedContext *C) {
|
2014-04-21 20:32:32 +00:00
|
|
|
return new ScheduleDAGMI(C, make_unique<PostGenericScheduler>(C), /*IsPostRA=*/true);
|
2013-12-28 21:56:57 +00:00
|
|
|
}
|
2012-01-17 06:55:03 +00:00
|
|
|
|
2012-10-15 18:02:27 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// ILP Scheduler. Currently for experimental analysis of heuristics.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
/// \brief Order nodes by the ILP metric.
|
|
|
|
struct ILPOrder {
|
2013-01-25 04:01:04 +00:00
|
|
|
const SchedDFSResult *DFSResult;
|
|
|
|
const BitVector *ScheduledTrees;
|
2012-10-15 18:02:27 +00:00
|
|
|
bool MaximizeILP;
|
|
|
|
|
2014-04-14 00:51:57 +00:00
|
|
|
ILPOrder(bool MaxILP)
|
|
|
|
: DFSResult(nullptr), ScheduledTrees(nullptr), MaximizeILP(MaxILP) {}
|
2012-10-15 18:02:27 +00:00
|
|
|
|
|
|
|
/// \brief Apply a less-than relation on node priority.
|
2012-11-28 05:13:28 +00:00
|
|
|
///
|
|
|
|
/// (Return true if A comes after B in the Q.)
|
2012-10-15 18:02:27 +00:00
|
|
|
bool operator()(const SUnit *A, const SUnit *B) const {
|
2012-11-28 05:13:28 +00:00
|
|
|
unsigned SchedTreeA = DFSResult->getSubtreeID(A);
|
|
|
|
unsigned SchedTreeB = DFSResult->getSubtreeID(B);
|
|
|
|
if (SchedTreeA != SchedTreeB) {
|
|
|
|
// Unscheduled trees have lower priority.
|
|
|
|
if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
|
|
|
|
return ScheduledTrees->test(SchedTreeB);
|
|
|
|
|
|
|
|
// Trees with shallower connections have have lower priority.
|
|
|
|
if (DFSResult->getSubtreeLevel(SchedTreeA)
|
|
|
|
!= DFSResult->getSubtreeLevel(SchedTreeB)) {
|
|
|
|
return DFSResult->getSubtreeLevel(SchedTreeA)
|
|
|
|
< DFSResult->getSubtreeLevel(SchedTreeB);
|
|
|
|
}
|
|
|
|
}
|
2012-10-15 18:02:27 +00:00
|
|
|
if (MaximizeILP)
|
2012-11-28 05:13:28 +00:00
|
|
|
return DFSResult->getILP(A) < DFSResult->getILP(B);
|
2012-10-15 18:02:27 +00:00
|
|
|
else
|
2012-11-28 05:13:28 +00:00
|
|
|
return DFSResult->getILP(A) > DFSResult->getILP(B);
|
2012-10-15 18:02:27 +00:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
/// \brief Schedule based on the ILP metric.
|
|
|
|
class ILPScheduler : public MachineSchedStrategy {
|
2013-12-28 21:56:47 +00:00
|
|
|
ScheduleDAGMILive *DAG;
|
2012-10-15 18:02:27 +00:00
|
|
|
ILPOrder Cmp;
|
|
|
|
|
|
|
|
std::vector<SUnit*> ReadyQ;
|
|
|
|
public:
|
2014-04-14 00:51:57 +00:00
|
|
|
ILPScheduler(bool MaximizeILP): DAG(nullptr), Cmp(MaximizeILP) {}
|
2012-10-15 18:02:27 +00:00
|
|
|
|
2014-03-07 09:26:03 +00:00
|
|
|
void initialize(ScheduleDAGMI *dag) override {
|
2013-12-28 21:56:47 +00:00
|
|
|
assert(dag->hasVRegLiveness() && "ILPScheduler needs vreg liveness");
|
|
|
|
DAG = static_cast<ScheduleDAGMILive*>(dag);
|
2013-01-25 06:33:57 +00:00
|
|
|
DAG->computeDFSResult();
|
2013-01-25 04:01:04 +00:00
|
|
|
Cmp.DFSResult = DAG->getDFSResult();
|
|
|
|
Cmp.ScheduledTrees = &DAG->getScheduledTrees();
|
2012-10-15 18:02:27 +00:00
|
|
|
ReadyQ.clear();
|
|
|
|
}
|
|
|
|
|
2014-03-07 09:26:03 +00:00
|
|
|
void registerRoots() override {
|
2012-11-29 14:36:26 +00:00
|
|
|
// Restore the heap in ReadyQ with the updated DFS results.
|
|
|
|
std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
2012-10-15 18:02:27 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Implement MachineSchedStrategy interface.
|
|
|
|
/// -----------------------------------------
|
|
|
|
|
2012-11-28 05:13:28 +00:00
|
|
|
/// Callback to select the highest priority node from the ready Q.
|
2014-03-07 09:26:03 +00:00
|
|
|
SUnit *pickNode(bool &IsTopNode) override {
|
2014-04-14 00:51:57 +00:00
|
|
|
if (ReadyQ.empty()) return nullptr;
|
2013-03-21 00:57:21 +00:00
|
|
|
std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
2012-10-15 18:02:27 +00:00
|
|
|
SUnit *SU = ReadyQ.back();
|
|
|
|
ReadyQ.pop_back();
|
|
|
|
IsTopNode = false;
|
2013-04-13 06:07:49 +00:00
|
|
|
DEBUG(dbgs() << "Pick node " << "SU(" << SU->NodeNum << ") "
|
2013-01-25 04:01:04 +00:00
|
|
|
<< " ILP: " << DAG->getDFSResult()->getILP(SU)
|
|
|
|
<< " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
|
|
|
|
<< DAG->getDFSResult()->getSubtreeLevel(
|
2013-04-13 06:07:49 +00:00
|
|
|
DAG->getDFSResult()->getSubtreeID(SU)) << '\n'
|
|
|
|
<< "Scheduling " << *SU->getInstr());
|
2012-10-15 18:02:27 +00:00
|
|
|
return SU;
|
|
|
|
}
|
|
|
|
|
2013-01-25 04:01:04 +00:00
|
|
|
/// \brief Scheduler callback to notify that a new subtree is scheduled.
|
2014-03-07 09:26:03 +00:00
|
|
|
void scheduleTree(unsigned SubtreeID) override {
|
2013-01-25 04:01:04 +00:00
|
|
|
std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
|
|
|
}
|
|
|
|
|
2012-11-28 05:13:28 +00:00
|
|
|
/// Callback after a node is scheduled. Mark a newly scheduled tree, notify
|
|
|
|
/// DFSResults, and resort the priority Q.
|
2014-03-07 09:26:03 +00:00
|
|
|
void schedNode(SUnit *SU, bool IsTopNode) override {
|
2012-11-28 05:13:28 +00:00
|
|
|
assert(!IsTopNode && "SchedDFSResult needs bottom-up");
|
|
|
|
}
|
2012-10-15 18:02:27 +00:00
|
|
|
|
2014-03-07 09:26:03 +00:00
|
|
|
void releaseTopNode(SUnit *) override { /*only called for top roots*/ }
|
2012-10-15 18:02:27 +00:00
|
|
|
|
2014-03-07 09:26:03 +00:00
|
|
|
void releaseBottomNode(SUnit *SU) override {
|
2012-10-15 18:02:27 +00:00
|
|
|
ReadyQ.push_back(SU);
|
|
|
|
std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
|
|
|
|
}
|
|
|
|
};
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
|
2014-04-21 20:32:32 +00:00
|
|
|
return new ScheduleDAGMILive(C, make_unique<ILPScheduler>(true));
|
2012-10-15 18:02:27 +00:00
|
|
|
}
|
|
|
|
static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
|
2014-04-21 20:32:32 +00:00
|
|
|
return new ScheduleDAGMILive(C, make_unique<ILPScheduler>(false));
|
2012-10-15 18:02:27 +00:00
|
|
|
}
|
|
|
|
static MachineSchedRegistry ILPMaxRegistry(
|
|
|
|
"ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
|
|
|
|
static MachineSchedRegistry ILPMinRegistry(
|
|
|
|
"ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
|
|
|
|
|
2012-01-14 02:17:06 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Machine Instruction Shuffler for Correctness Testing
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2012-01-13 06:30:30 +00:00
|
|
|
#ifndef NDEBUG
|
|
|
|
namespace {
|
2012-03-14 04:00:41 +00:00
|
|
|
/// Apply a less-than relation on the node order, which corresponds to the
|
|
|
|
/// instruction order prior to scheduling. IsReverse implements greater-than.
|
|
|
|
template<bool IsReverse>
|
|
|
|
struct SUnitOrder {
|
2012-01-17 06:55:07 +00:00
|
|
|
bool operator()(SUnit *A, SUnit *B) const {
|
2012-03-14 04:00:41 +00:00
|
|
|
if (IsReverse)
|
|
|
|
return A->NodeNum > B->NodeNum;
|
|
|
|
else
|
|
|
|
return A->NodeNum < B->NodeNum;
|
2012-01-17 06:55:07 +00:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
2012-01-13 06:30:30 +00:00
|
|
|
/// Reorder instructions as much as possible.
|
2012-03-14 04:00:41 +00:00
|
|
|
class InstructionShuffler : public MachineSchedStrategy {
|
|
|
|
bool IsAlternating;
|
|
|
|
bool IsTopDown;
|
|
|
|
|
|
|
|
// Using a less-than relation (SUnitOrder<false>) for the TopQ priority
|
|
|
|
// gives nodes with a higher number higher priority causing the latest
|
|
|
|
// instructions to be scheduled first.
|
|
|
|
PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
|
|
|
|
TopQ;
|
|
|
|
// When scheduling bottom-up, use greater-than as the queue priority.
|
|
|
|
PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
|
|
|
|
BottomQ;
|
2012-01-13 06:30:30 +00:00
|
|
|
public:
|
2012-03-14 04:00:41 +00:00
|
|
|
InstructionShuffler(bool alternate, bool topdown)
|
|
|
|
: IsAlternating(alternate), IsTopDown(topdown) {}
|
2012-01-13 06:30:30 +00:00
|
|
|
|
2013-12-28 21:56:47 +00:00
|
|
|
virtual void initialize(ScheduleDAGMI*) {
|
2012-03-14 04:00:41 +00:00
|
|
|
TopQ.clear();
|
|
|
|
BottomQ.clear();
|
|
|
|
}
|
2012-01-17 06:55:07 +00:00
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
/// Implement MachineSchedStrategy interface.
|
|
|
|
/// -----------------------------------------
|
|
|
|
|
|
|
|
virtual SUnit *pickNode(bool &IsTopNode) {
|
|
|
|
SUnit *SU;
|
|
|
|
if (IsTopDown) {
|
|
|
|
do {
|
2014-04-14 00:51:57 +00:00
|
|
|
if (TopQ.empty()) return nullptr;
|
2012-03-14 04:00:41 +00:00
|
|
|
SU = TopQ.top();
|
|
|
|
TopQ.pop();
|
|
|
|
} while (SU->isScheduled);
|
|
|
|
IsTopNode = true;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
do {
|
2014-04-14 00:51:57 +00:00
|
|
|
if (BottomQ.empty()) return nullptr;
|
2012-03-14 04:00:41 +00:00
|
|
|
SU = BottomQ.top();
|
|
|
|
BottomQ.pop();
|
|
|
|
} while (SU->isScheduled);
|
|
|
|
IsTopNode = false;
|
|
|
|
}
|
|
|
|
if (IsAlternating)
|
|
|
|
IsTopDown = !IsTopDown;
|
2012-01-17 06:55:07 +00:00
|
|
|
return SU;
|
|
|
|
}
|
|
|
|
|
2012-05-24 22:11:09 +00:00
|
|
|
virtual void schedNode(SUnit *SU, bool IsTopNode) {}
|
|
|
|
|
2012-03-14 04:00:41 +00:00
|
|
|
virtual void releaseTopNode(SUnit *SU) {
|
|
|
|
TopQ.push(SU);
|
|
|
|
}
|
|
|
|
virtual void releaseBottomNode(SUnit *SU) {
|
|
|
|
BottomQ.push(SU);
|
2012-01-13 06:30:30 +00:00
|
|
|
}
|
|
|
|
};
|
|
|
|
} // namespace
|
|
|
|
|
2012-03-08 01:41:12 +00:00
|
|
|
static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
|
2012-03-14 04:00:41 +00:00
|
|
|
bool Alternate = !ForceTopDown && !ForceBottomUp;
|
|
|
|
bool TopDown = !ForceBottomUp;
|
2012-03-14 11:26:37 +00:00
|
|
|
assert((TopDown || !ForceTopDown) &&
|
2012-03-14 04:00:41 +00:00
|
|
|
"-misched-topdown incompatible with -misched-bottomup");
|
2014-04-21 20:32:32 +00:00
|
|
|
return new ScheduleDAGMILive(C, make_unique<InstructionShuffler>(Alternate, TopDown));
|
2012-01-13 06:30:30 +00:00
|
|
|
}
|
2012-03-14 04:00:41 +00:00
|
|
|
static MachineSchedRegistry ShufflerRegistry(
|
|
|
|
"shuffle", "Shuffle machine instructions alternating directions",
|
|
|
|
createInstructionShuffler);
|
2012-01-13 06:30:30 +00:00
|
|
|
#endif // !NDEBUG
|
2013-01-25 07:45:29 +00:00
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
2013-12-28 21:56:47 +00:00
|
|
|
// GraphWriter support for ScheduleDAGMILive.
|
2013-01-25 07:45:29 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
namespace llvm {
|
|
|
|
|
|
|
|
template<> struct GraphTraits<
|
|
|
|
ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
|
|
|
|
|
|
|
|
template<>
|
|
|
|
struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
|
|
|
|
|
|
|
|
DOTGraphTraits (bool isSimple=false) : DefaultDOTGraphTraits(isSimple) {}
|
|
|
|
|
|
|
|
static std::string getGraphName(const ScheduleDAG *G) {
|
|
|
|
return G->MF.getName();
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool renderGraphFromBottomUp() {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool isNodeHidden(const SUnit *Node) {
|
2013-09-04 21:00:18 +00:00
|
|
|
return (Node->Preds.size() > 10 || Node->Succs.size() > 10);
|
2013-01-25 07:45:29 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static bool hasNodeAddressLabel(const SUnit *Node,
|
|
|
|
const ScheduleDAG *Graph) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// 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);
|
2013-12-28 21:56:47 +00:00
|
|
|
const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
|
|
|
|
const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
|
2014-04-14 00:51:57 +00:00
|
|
|
static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr;
|
2013-09-06 17:32:42 +00:00
|
|
|
SS << "SU:" << SU->NodeNum;
|
|
|
|
if (DFS)
|
|
|
|
SS << " I:" << DFS->getNumInstrs(SU);
|
2013-01-25 07:45:29 +00:00
|
|
|
return SS.str();
|
|
|
|
}
|
|
|
|
static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
|
|
|
|
return G->getGraphNodeLabel(SU);
|
|
|
|
}
|
|
|
|
|
2013-12-28 21:56:47 +00:00
|
|
|
static std::string getNodeAttributes(const SUnit *N, const ScheduleDAG *G) {
|
2013-01-25 07:45:29 +00:00
|
|
|
std::string Str("shape=Mrecord");
|
2013-12-28 21:56:47 +00:00
|
|
|
const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
|
|
|
|
const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
|
2014-04-14 00:51:57 +00:00
|
|
|
static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr;
|
2013-01-25 07:45:29 +00:00
|
|
|
if (DFS) {
|
|
|
|
Str += ",style=filled,fillcolor=\"#";
|
|
|
|
Str += DOT::getColorString(DFS->getSubtreeID(N));
|
|
|
|
Str += '"';
|
|
|
|
}
|
|
|
|
return Str;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
} // namespace llvm
|
|
|
|
#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());
|
|
|
|
}
|