//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "post-RA-sched"
+#include "ScheduleDAGInstrs.h"
#include "llvm/CodeGen/Passes.h"
-#include "llvm/CodeGen/ScheduleDAGInstrs.h"
#include "llvm/CodeGen/LatencyPriorityQueue.h"
#include "llvm/CodeGen/SchedulerRegistry.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
+#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/Compiler.h"
#include <map>
using namespace llvm;
+STATISTIC(NumNoops, "Number of noops inserted");
STATISTIC(NumStalls, "Number of pipeline stalls");
static cl::opt<bool>
cl::desc("Break post-RA scheduling anti-dependencies"),
cl::init(true), cl::Hidden);
+static cl::opt<bool>
+EnablePostRAHazardAvoidance("avoid-hazards",
+ cl::desc("Enable simple hazard-avoidance"),
+ cl::init(true), cl::Hidden);
+
namespace {
class VISIBILITY_HIDDEN PostRAScheduler : public MachineFunctionPass {
public:
/// Topo - A topological ordering for SUnits.
ScheduleDAGTopologicalSort Topo;
+ /// AllocatableSet - The set of allocatable registers.
+ /// We'll be ignoring anti-dependencies on non-allocatable registers,
+ /// because they may not be safe to break.
+ const BitVector AllocatableSet;
+
+ /// HazardRec - The hazard recognizer to use.
+ ScheduleHazardRecognizer *HazardRec;
+
public:
- SchedulePostRATDList(MachineBasicBlock *mbb, const TargetMachine &tm,
+ SchedulePostRATDList(MachineFunction &MF,
const MachineLoopInfo &MLI,
- const MachineDominatorTree &MDT)
- : ScheduleDAGInstrs(mbb, tm, MLI, MDT), Topo(SUnits) {}
+ const MachineDominatorTree &MDT,
+ ScheduleHazardRecognizer *HR)
+ : ScheduleDAGInstrs(MF, MLI, MDT), Topo(SUnits),
+ AllocatableSet(TRI->getAllocatableSet(MF)),
+ HazardRec(HR) {}
+
+ ~SchedulePostRATDList() {
+ delete HazardRec;
+ }
void Schedule();
void ListScheduleTopDown();
bool BreakAntiDependencies();
};
+
+ /// SimpleHazardRecognizer - A *very* simple hazard recognizer. It uses
+ /// a coarse classification and attempts to avoid that instructions of
+ /// a given class aren't grouped too densely together.
+ class SimpleHazardRecognizer : public ScheduleHazardRecognizer {
+ /// Class - A simple classification for SUnits.
+ enum Class {
+ Other, Load, Store
+ };
+
+ /// Window - The Class values of the most recently issued
+ /// instructions.
+ Class Window[8];
+
+ /// getClass - Classify the given SUnit.
+ Class getClass(const SUnit *SU) {
+ const MachineInstr *MI = SU->getInstr();
+ const TargetInstrDesc &TID = MI->getDesc();
+ if (TID.mayLoad())
+ return Load;
+ if (TID.mayStore())
+ return Store;
+ return Other;
+ }
+
+ /// Step - Rotate the existing entries in Window and insert the
+ /// given class value in position as the most recent.
+ void Step(Class C) {
+ std::copy(Window+1, array_endof(Window), Window);
+ Window[array_lengthof(Window)-1] = C;
+ }
+
+ public:
+ SimpleHazardRecognizer() : Window() {}
+
+ virtual HazardType getHazardType(SUnit *SU) {
+ Class C = getClass(SU);
+ if (C == Other)
+ return NoHazard;
+ unsigned Score = 0;
+ for (unsigned i = 0; i != array_lengthof(Window); ++i)
+ if (Window[i] == C)
+ Score += i + 1;
+ if (Score > array_lengthof(Window) * 2)
+ return Hazard;
+ return NoHazard;
+ }
+
+ virtual void EmitInstruction(SUnit *SU) {
+ Step(getClass(SU));
+ }
+
+ virtual void AdvanceCycle() {
+ Step(Other);
+ }
+ };
}
bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
const MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
const MachineDominatorTree &MDT = getAnalysis<MachineDominatorTree>();
+ ScheduleHazardRecognizer *HR = EnablePostRAHazardAvoidance ?
+ new SimpleHazardRecognizer :
+ new ScheduleHazardRecognizer();
+
+ SchedulePostRATDList Scheduler(Fn, MLI, MDT, HR);
// Loop over all of the basic blocks
for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
MBB != MBBe; ++MBB) {
+ // Schedule each sequence of instructions not interrupted by a label
+ // or anything else that effectively needs to shut down scheduling.
+ MachineBasicBlock::iterator Current = MBB->end(), Top = MBB->begin();
+ for (MachineBasicBlock::iterator I = Current; I != Top; ) {
+ MachineInstr *MI = --I;
+ if (MI->getDesc().isTerminator() || MI->isLabel()) {
+ Scheduler.Run(0, MBB, next(I), Current);
+ Scheduler.EmitSchedule();
+ Current = I;
+ }
+ }
- SchedulePostRATDList Scheduler(MBB, Fn.getTarget(), MLI, MDT);
-
- Scheduler.Run();
-
+ Scheduler.Run(0, MBB, Top, Current);
Scheduler.EmitSchedule();
}
/// instruction of the specified TargetInstrDesc.
static const TargetRegisterClass*
getInstrOperandRegClass(const TargetRegisterInfo *TRI,
- const TargetInstrInfo *TII, const TargetInstrDesc &II,
- unsigned Op) {
+ const TargetInstrDesc &II, unsigned Op) {
if (Op >= II.getNumOperands())
return NULL;
if (II.OpInfo[Op].isLookupPtrRegClass())
- return TII->getPointerRegClass();
+ return TRI->getPointerRegClass();
return TRI->getRegClass(II.OpInfo[Op].RegClass);
}
bool SchedulePostRATDList::BreakAntiDependencies() {
// The code below assumes that there is at least one instruction,
// so just duck out immediately if the block is empty.
- if (BB->empty()) return false;
+ if (SUnits.empty()) return false;
// Find the node at the bottom of the critical path.
SUnit *Max = 0;
}
DOUT << "Critical path has total latency "
- << (Max ? Max->getDepth() + Max->Latency : 0) << "\n";
-
- // We'll be ignoring anti-dependencies on non-allocatable registers, because
- // they may not be safe to break.
- const BitVector AllocatableSet = TRI->getAllocatableSet(*MF);
+ << (Max->getDepth() + Max->Latency) << "\n";
// Track progress along the critical path through the SUnit graph as we walk
// the instructions.
// instructions from the bottom up, tracking information about liveness
// as we go to help determine which registers are available.
bool Changed = false;
- unsigned Count = BB->size() - 1;
- for (MachineBasicBlock::reverse_iterator I = BB->rbegin(), E = BB->rend();
- I != E; ++I, --Count) {
- MachineInstr *MI = &*I;
+ unsigned Count = SUnits.size() - 1;
+ for (MachineBasicBlock::iterator I = End, E = Begin;
+ I != E; --Count) {
+ MachineInstr *MI = --I;
// After regalloc, IMPLICIT_DEF instructions aren't safe to treat as
// dependence-breaking. In the case of an INSERT_SUBREG, the IMPLICIT_DEF
AntiDepReg = Edge->getReg();
assert(AntiDepReg != 0 && "Anti-dependence on reg0?");
// Don't break anti-dependencies on non-allocatable registers.
- if (AllocatableSet.test(AntiDepReg)) {
+ if (!AllocatableSet.test(AntiDepReg))
+ AntiDepReg = 0;
+ else {
// If the SUnit has other dependencies on the SUnit that it
// anti-depends on, don't bother breaking the anti-dependency
// since those edges would prevent such units from being
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
const TargetRegisterClass *NewRC =
- getInstrOperandRegClass(TRI, TII, MI->getDesc(), i);
+ getInstrOperandRegClass(TRI, MI->getDesc(), i);
// If this instruction has a use of AntiDepReg, breaking it
// is invalid.
// TODO: Instead of picking the first free register, consider which might
// be the best.
if (AntiDepReg != 0) {
- for (TargetRegisterClass::iterator R = RC->allocation_order_begin(*MF),
- RE = RC->allocation_order_end(*MF); R != RE; ++R) {
+ for (TargetRegisterClass::iterator R = RC->allocation_order_begin(MF),
+ RE = RC->allocation_order_end(MF); R != RE; ++R) {
unsigned NewReg = *R;
// Don't replace a register with itself.
if (NewReg == AntiDepReg) continue;
if (!MO.isUse()) continue;
const TargetRegisterClass *NewRC =
- getInstrOperandRegClass(TRI, TII, MI->getDesc(), i);
+ getInstrOperandRegClass(TRI, MI->getDesc(), i);
// For now, only allow the register to be changed if its register
// class is consistent across all uses.
// While Available queue is not empty, grab the node with the highest
// priority. If it is not ready put it back. Schedule the node.
+ std::vector<SUnit*> NotReady;
Sequence.reserve(SUnits.size());
while (!AvailableQueue.empty() || !PendingQueue.empty()) {
// Check to see if any of the pending instructions are ready to issue. If
MinDepth = PendingQueue[i]->getDepth();
}
- // If there are no instructions available, don't try to issue anything.
+ // If there are no instructions available, don't try to issue anything, and
+ // don't advance the hazard recognizer.
if (AvailableQueue.empty()) {
CurCycle = MinDepth != ~0u ? MinDepth : CurCycle + 1;
continue;
}
- SUnit *FoundSUnit = AvailableQueue.pop();
-
+ SUnit *FoundSUnit = 0;
+
+ bool HasNoopHazards = false;
+ while (!AvailableQueue.empty()) {
+ SUnit *CurSUnit = AvailableQueue.pop();
+
+ ScheduleHazardRecognizer::HazardType HT =
+ HazardRec->getHazardType(CurSUnit);
+ if (HT == ScheduleHazardRecognizer::NoHazard) {
+ FoundSUnit = CurSUnit;
+ break;
+ }
+
+ // Remember if this is a noop hazard.
+ HasNoopHazards |= HT == ScheduleHazardRecognizer::NoopHazard;
+
+ NotReady.push_back(CurSUnit);
+ }
+
+ // Add the nodes that aren't ready back onto the available list.
+ if (!NotReady.empty()) {
+ AvailableQueue.push_all(NotReady);
+ NotReady.clear();
+ }
+
// If we found a node to schedule, do it now.
if (FoundSUnit) {
ScheduleNodeTopDown(FoundSUnit, CurCycle);
+ HazardRec->EmitInstruction(FoundSUnit);
// If this is a pseudo-op node, we don't want to increment the current
// cycle.
if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
- ++CurCycle;
- } else {
+ ++CurCycle;
+ } else if (!HasNoopHazards) {
// Otherwise, we have a pipeline stall, but no other problem, just advance
// the current cycle and try again.
DOUT << "*** Advancing cycle, no work to do\n";
+ HazardRec->AdvanceCycle();
++NumStalls;
++CurCycle;
+ } else {
+ // Otherwise, we have no instructions to issue and we have instructions
+ // that will fault if we don't do this right. This is the case for
+ // processors without pipeline interlocks and other cases.
+ DOUT << "*** Emitting noop\n";
+ HazardRec->EmitNoop();
+ Sequence.push_back(0); // NULL here means noop
+ ++NumNoops;
+ ++CurCycle;
}
}