RPCSX/llvm
1
0
Fork 0
mirror of https://github.com/RPCSX/llvm.git synced 2024-11-25 04:39:44 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Break anti-dependence breaking out into its own class.

Browse Source 
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@85127 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
David Goodwin 2009-10-26 16:59:04 +00:00
parent ac16d44e58
commit 2e7be612d5
7 changed files with 737 additions and 545 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

56
include/llvm/CodeGen/AntiDepBreaker.h Normal file
View File

@ -0,0 +1,56 @@
//=- llvm/CodeGen/AntiDepBreaker.h - Anti-Dependence Breaking -*- C++ -*-=//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements the AntiDepBreaker class, which implements
// anti-dependence breaking heuristics for post-register-allocation scheduling.
//
//===----------------------------------------------------------------------===//
#ifndef LLVM_CODEGEN_ANTIDEPBREAKER_H
#define LLVM_CODEGEN_ANTIDEPBREAKER_H
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/Target/TargetRegisterInfo.h"
namespace llvm {
/// AntiDepBreaker - This class works into conjunction with the
/// post-RA scheduler to rename registers to break register
/// anti-dependencies.
class AntiDepBreaker {
public:
/// Start - Initialize anti-dep breaking for a new basic block.
virtual void StartBlock(MachineBasicBlock *BB) =0;
/// BreakAntiDependencies - Identifiy anti-dependencies within a
/// basic-block region and break them by renaming registers. Return
/// the number of anti-dependencies broken.
///
virtual unsigned BreakAntiDependencies(std::vector<SUnit>& SUnits,
MachineBasicBlock::iterator& Begin,
MachineBasicBlock::iterator& End,
unsigned InsertPosIndex) =0;
/// Observe - Update liveness information to account for the current
/// instruction, which will not be scheduled.
///
virtual void Observe(MachineInstr *MI, unsigned Count,
unsigned InsertPosIndex) =0;
/// Finish - Finish anti-dep breaking for a basic block.
virtual void FinishBlock() =0;
};
}
#endif

1
lib/CodeGen/CMakeLists.txt
View File

@ -1,6 +1,7 @@
add_llvm_library(LLVMCodeGen
BranchFolding.cpp
CodePlacementOpt.cpp
CriticalAntiDepBreaker.cpp
DeadMachineInstructionElim.cpp
DwarfEHPrepare.cpp
ELFCodeEmitter.cpp

539
lib/CodeGen/CriticalAntiDepBreaker.cpp Normal file
View File

@ -0,0 +1,539 @@
//===----- CriticalAntiDepBreaker.cpp - Anti-dep breaker -------- ---------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements the CriticalAntiDepBreaker class, which
// implements register anti-dependence breaking along a blocks
// critical path during post-RA scheduler.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "critical-antidep"
#include "CriticalAntiDepBreaker.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
CriticalAntiDepBreaker::
CriticalAntiDepBreaker(MachineFunction& MFi) :
AntiDepBreaker(), MF(MFi),
MRI(MF.getRegInfo()),
TRI(MF.getTarget().getRegisterInfo()),
AllocatableSet(TRI->getAllocatableSet(MF))
{
}
CriticalAntiDepBreaker::~CriticalAntiDepBreaker() {
}
void CriticalAntiDepBreaker::StartBlock(MachineBasicBlock *BB) {
// Clear out the register class data.
std::fill(Classes, array_endof(Classes),
static_cast<const TargetRegisterClass *>(0));
// Initialize the indices to indicate that no registers are live.
std::fill(KillIndices, array_endof(KillIndices), ~0u);
std::fill(DefIndices, array_endof(DefIndices), BB->size());
// Clear "do not change" set.
KeepRegs.clear();
bool IsReturnBlock = (!BB->empty() && BB->back().getDesc().isReturn());
// Determine the live-out physregs for this block.
if (IsReturnBlock) {
// In a return block, examine the function live-out regs.
for (MachineRegisterInfo::liveout_iterator I = MRI.liveout_begin(),
E = MRI.liveout_end(); I != E; ++I) {
unsigned Reg = *I;
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[Reg] = BB->size();
DefIndices[Reg] = ~0u;
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[AliasReg] = BB->size();
DefIndices[AliasReg] = ~0u;
}
}
} else {
// In a non-return block, examine the live-in regs of all successors.
for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
SE = BB->succ_end(); SI != SE; ++SI)
for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
E = (*SI)->livein_end(); I != E; ++I) {
unsigned Reg = *I;
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[Reg] = BB->size();
DefIndices[Reg] = ~0u;
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[AliasReg] = BB->size();
DefIndices[AliasReg] = ~0u;
}
}
}
// Mark live-out callee-saved registers. In a return block this is
// all callee-saved registers. In non-return this is any
// callee-saved register that is not saved in the prolog.
const MachineFrameInfo *MFI = MF.getFrameInfo();
BitVector Pristine = MFI->getPristineRegs(BB);
for (const unsigned *I = TRI->getCalleeSavedRegs(); *I; ++I) {
unsigned Reg = *I;
if (!IsReturnBlock && !Pristine.test(Reg)) continue;
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[Reg] = BB->size();
DefIndices[Reg] = ~0u;
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[AliasReg] = BB->size();
DefIndices[AliasReg] = ~0u;
}
}
}
void CriticalAntiDepBreaker::FinishBlock() {
RegRefs.clear();
KeepRegs.clear();
}
void CriticalAntiDepBreaker::Observe(MachineInstr *MI, unsigned Count,
unsigned InsertPosIndex) {
assert(Count < InsertPosIndex && "Instruction index out of expected range!");
// Any register which was defined within the previous scheduling region
// may have been rescheduled and its lifetime may overlap with registers
// in ways not reflected in our current liveness state. For each such
// register, adjust the liveness state to be conservatively correct.
for (unsigned Reg = 0; Reg != TargetRegisterInfo::FirstVirtualRegister; ++Reg)
if (DefIndices[Reg] < InsertPosIndex && DefIndices[Reg] >= Count) {
assert(KillIndices[Reg] == ~0u && "Clobbered register is live!");
// Mark this register to be non-renamable.
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
// Move the def index to the end of the previous region, to reflect
// that the def could theoretically have been scheduled at the end.
DefIndices[Reg] = InsertPosIndex;
}
PrescanInstruction(MI);
ScanInstruction(MI, Count);
}
/// CriticalPathStep - Return the next SUnit after SU on the bottom-up
/// critical path.
static SDep *CriticalPathStep(SUnit *SU) {
SDep *Next = 0;
unsigned NextDepth = 0;
// Find the predecessor edge with the greatest depth.
for (SUnit::pred_iterator P = SU->Preds.begin(), PE = SU->Preds.end();
P != PE; ++P) {
SUnit *PredSU = P->getSUnit();
unsigned PredLatency = P->getLatency();
unsigned PredTotalLatency = PredSU->getDepth() + PredLatency;
// In the case of a latency tie, prefer an anti-dependency edge over
// other types of edges.
if (NextDepth < PredTotalLatency ||
(NextDepth == PredTotalLatency && P->getKind() == SDep::Anti)) {
NextDepth = PredTotalLatency;
Next = &*P;
}
}
return Next;
}
void CriticalAntiDepBreaker::PrescanInstruction(MachineInstr *MI) {
// Scan the register operands for this instruction and update
// Classes and RegRefs.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
const TargetRegisterClass *NewRC = 0;
if (i < MI->getDesc().getNumOperands())
NewRC = MI->getDesc().OpInfo[i].getRegClass(TRI);
// For now, only allow the register to be changed if its register
// class is consistent across all uses.
if (!Classes[Reg] && NewRC)
Classes[Reg] = NewRC;
else if (!NewRC || Classes[Reg] != NewRC)
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
// Now check for aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
// If an alias of the reg is used during the live range, give up.
// Note that this allows us to skip checking if AntiDepReg
// overlaps with any of the aliases, among other things.
unsigned AliasReg = *Alias;
if (Classes[AliasReg]) {
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
}
}
// If we're still willing to consider this register, note the reference.
if (Classes[Reg] != reinterpret_cast<TargetRegisterClass *>(-1))
RegRefs.insert(std::make_pair(Reg, &MO));
// It's not safe to change register allocation for source operands of
// that have special allocation requirements.
if (MO.isUse() && MI->getDesc().hasExtraSrcRegAllocReq()) {
if (KeepRegs.insert(Reg)) {
for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
*Subreg; ++Subreg)
KeepRegs.insert(*Subreg);
}
}
}
}
void CriticalAntiDepBreaker::ScanInstruction(MachineInstr *MI,
unsigned Count) {
// Update liveness.
// Proceding upwards, registers that are defed but not used in this
// instruction are now dead.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
if (!MO.isDef()) continue;
// Ignore two-addr defs.
if (MI->isRegTiedToUseOperand(i)) continue;
DefIndices[Reg] = Count;
KillIndices[Reg] = ~0u;
assert(((KillIndices[Reg] == ~0u) !=
(DefIndices[Reg] == ~0u)) &&
"Kill and Def maps aren't consistent for Reg!");
KeepRegs.erase(Reg);
Classes[Reg] = 0;
RegRefs.erase(Reg);
// Repeat, for all subregs.
for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
*Subreg; ++Subreg) {
unsigned SubregReg = *Subreg;
DefIndices[SubregReg] = Count;
KillIndices[SubregReg] = ~0u;
KeepRegs.erase(SubregReg);
Classes[SubregReg] = 0;
RegRefs.erase(SubregReg);
}
// Conservatively mark super-registers as unusable.
for (const unsigned *Super = TRI->getSuperRegisters(Reg);
*Super; ++Super) {
unsigned SuperReg = *Super;
Classes[SuperReg] = reinterpret_cast<TargetRegisterClass *>(-1);
}
}
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
if (!MO.isUse()) continue;
const TargetRegisterClass *NewRC = 0;
if (i < MI->getDesc().getNumOperands())
NewRC = MI->getDesc().OpInfo[i].getRegClass(TRI);
// For now, only allow the register to be changed if its register
// class is consistent across all uses.
if (!Classes[Reg] && NewRC)
Classes[Reg] = NewRC;
else if (!NewRC || Classes[Reg] != NewRC)
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
RegRefs.insert(std::make_pair(Reg, &MO));
// It wasn't previously live but now it is, this is a kill.
if (KillIndices[Reg] == ~0u) {
KillIndices[Reg] = Count;
DefIndices[Reg] = ~0u;
assert(((KillIndices[Reg] == ~0u) !=
(DefIndices[Reg] == ~0u)) &&
"Kill and Def maps aren't consistent for Reg!");
}
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
if (KillIndices[AliasReg] == ~0u) {
KillIndices[AliasReg] = Count;
DefIndices[AliasReg] = ~0u;
}
}
}
}
unsigned
CriticalAntiDepBreaker::findSuitableFreeRegister(unsigned AntiDepReg,
unsigned LastNewReg,
const TargetRegisterClass *RC) {
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;
// Don't replace a register with one that was recently used to repair
// an anti-dependence with this AntiDepReg, because that would
// re-introduce that anti-dependence.
if (NewReg == LastNewReg) continue;
// If NewReg is dead and NewReg's most recent def is not before
// AntiDepReg's kill, it's safe to replace AntiDepReg with NewReg.
assert(((KillIndices[AntiDepReg] == ~0u) != (DefIndices[AntiDepReg] == ~0u)) &&
"Kill and Def maps aren't consistent for AntiDepReg!");
assert(((KillIndices[NewReg] == ~0u) != (DefIndices[NewReg] == ~0u)) &&
"Kill and Def maps aren't consistent for NewReg!");
if (KillIndices[NewReg] != ~0u ||
Classes[NewReg] == reinterpret_cast<TargetRegisterClass *>(-1) ||
KillIndices[AntiDepReg] > DefIndices[NewReg])
continue;
return NewReg;
}
// No registers are free and available!
return 0;
}
unsigned CriticalAntiDepBreaker::
BreakAntiDependencies(std::vector<SUnit>& SUnits,
MachineBasicBlock::iterator& Begin,
MachineBasicBlock::iterator& End,
unsigned InsertPosIndex) {
// The code below assumes that there is at least one instruction,
// so just duck out immediately if the block is empty.
if (SUnits.empty()) return 0;
// Find the node at the bottom of the critical path.
SUnit *Max = 0;
for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
SUnit *SU = &SUnits[i];
if (!Max || SU->getDepth() + SU->Latency > Max->getDepth() + Max->Latency)
Max = SU;
}
#ifndef NDEBUG
{
DEBUG(errs() << "Critical path has total latency "
<< (Max->getDepth() + Max->Latency) << "\n");
DEBUG(errs() << "Available regs:");
for (unsigned Reg = 0; Reg < TRI->getNumRegs(); ++Reg) {
if (KillIndices[Reg] == ~0u)
DEBUG(errs() << " " << TRI->getName(Reg));
}
DEBUG(errs() << '\n');
}
#endif
// Track progress along the critical path through the SUnit graph as we walk
// the instructions.
SUnit *CriticalPathSU = Max;
MachineInstr *CriticalPathMI = CriticalPathSU->getInstr();
// Consider this pattern:
// A = ...
// ... = A
// A = ...
// ... = A
// A = ...
// ... = A
// A = ...
// ... = A
// There are three anti-dependencies here, and without special care,
// we'd break all of them using the same register:
// A = ...
// ... = A
// B = ...
// ... = B
// B = ...
// ... = B
// B = ...
// ... = B
// because at each anti-dependence, B is the first register that
// isn't A which is free. This re-introduces anti-dependencies
// at all but one of the original anti-dependencies that we were
// trying to break. To avoid this, keep track of the most recent
// register that each register was replaced with, avoid
// using it to repair an anti-dependence on the same register.
// This lets us produce this:
// A = ...
// ... = A
// B = ...
// ... = B
// C = ...
// ... = C
// B = ...
// ... = B
// This still has an anti-dependence on B, but at least it isn't on the
// original critical path.
//
// TODO: If we tracked more than one register here, we could potentially
// fix that remaining critical edge too. This is a little more involved,
// because unlike the most recent register, less recent registers should
// still be considered, though only if no other registers are available.
unsigned LastNewReg[TargetRegisterInfo::FirstVirtualRegister] = {};
// Attempt to break anti-dependence edges on the critical path. Walk the
// instructions from the bottom up, tracking information about liveness
// as we go to help determine which registers are available.
unsigned Broken = 0;
unsigned Count = InsertPosIndex - 1;
for (MachineBasicBlock::iterator I = End, E = Begin;
I != E; --Count) {
MachineInstr *MI = --I;
// Check if this instruction has a dependence on the critical path that
// is an anti-dependence that we may be able to break. If it is, set
// AntiDepReg to the non-zero register associated with the anti-dependence.
//
// We limit our attention to the critical path as a heuristic to avoid
// breaking anti-dependence edges that aren't going to significantly
// impact the overall schedule. There are a limited number of registers
// and we want to save them for the important edges.
//
// TODO: Instructions with multiple defs could have multiple
// anti-dependencies. The current code here only knows how to break one
// edge per instruction. Note that we'd have to be able to break all of
// the anti-dependencies in an instruction in order to be effective.
unsigned AntiDepReg = 0;
if (MI == CriticalPathMI) {
if (SDep *Edge = CriticalPathStep(CriticalPathSU)) {
SUnit *NextSU = Edge->getSUnit();
// Only consider anti-dependence edges.
if (Edge->getKind() == SDep::Anti) {
AntiDepReg = Edge->getReg();
assert(AntiDepReg != 0 && "Anti-dependence on reg0?");
if (!AllocatableSet.test(AntiDepReg))
// Don't break anti-dependencies on non-allocatable registers.
AntiDepReg = 0;
else if (KeepRegs.count(AntiDepReg))
// Don't break anti-dependencies if an use down below requires
// this exact register.
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
// scheduled past each other regardless.
//
// Also, if there are dependencies on other SUnits with the
// same register as the anti-dependency, don't attempt to
// break it.
for (SUnit::pred_iterator P = CriticalPathSU->Preds.begin(),
PE = CriticalPathSU->Preds.end(); P != PE; ++P)
if (P->getSUnit() == NextSU ?
(P->getKind() != SDep::Anti || P->getReg() != AntiDepReg) :
(P->getKind() == SDep::Data && P->getReg() == AntiDepReg)) {
AntiDepReg = 0;
break;
}
}
}
CriticalPathSU = NextSU;
CriticalPathMI = CriticalPathSU->getInstr();
} else {
// We've reached the end of the critical path.
CriticalPathSU = 0;
CriticalPathMI = 0;
}
}
PrescanInstruction(MI);
if (MI->getDesc().hasExtraDefRegAllocReq())
// If this instruction's defs have special allocation requirement, don't
// break this anti-dependency.
AntiDepReg = 0;
else if (AntiDepReg) {
// If this instruction has a use of AntiDepReg, breaking it
// is invalid.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
if (MO.isUse() && AntiDepReg == Reg) {
AntiDepReg = 0;
break;
}
}
}
// Determine AntiDepReg's register class, if it is live and is
// consistently used within a single class.
const TargetRegisterClass *RC = AntiDepReg != 0 ? Classes[AntiDepReg] : 0;
assert((AntiDepReg == 0 || RC != NULL) &&
"Register should be live if it's causing an anti-dependence!");
if (RC == reinterpret_cast<TargetRegisterClass *>(-1))
AntiDepReg = 0;
// Look for a suitable register to use to break the anti-depenence.
//
// TODO: Instead of picking the first free register, consider which might
// be the best.
if (AntiDepReg != 0) {
if (unsigned NewReg = findSuitableFreeRegister(AntiDepReg,
LastNewReg[AntiDepReg],
RC)) {
DEBUG(errs() << "Breaking anti-dependence edge on "
<< TRI->getName(AntiDepReg)
<< " with " << RegRefs.count(AntiDepReg) << " references"
<< " using " << TRI->getName(NewReg) << "!\n");
// Update the references to the old register to refer to the new
// register.
std::pair<std::multimap<unsigned, MachineOperand *>::iterator,
std::multimap<unsigned, MachineOperand *>::iterator>
Range = RegRefs.equal_range(AntiDepReg);
for (std::multimap<unsigned, MachineOperand *>::iterator
Q = Range.first, QE = Range.second; Q != QE; ++Q)
Q->second->setReg(NewReg);
// We just went back in time and modified history; the
// liveness information for the anti-depenence reg is now
// inconsistent. Set the state as if it were dead.
Classes[NewReg] = Classes[AntiDepReg];
DefIndices[NewReg] = DefIndices[AntiDepReg];
KillIndices[NewReg] = KillIndices[AntiDepReg];
assert(((KillIndices[NewReg] == ~0u) !=
(DefIndices[NewReg] == ~0u)) &&
"Kill and Def maps aren't consistent for NewReg!");
Classes[AntiDepReg] = 0;
DefIndices[AntiDepReg] = KillIndices[AntiDepReg];
KillIndices[AntiDepReg] = ~0u;
assert(((KillIndices[AntiDepReg] == ~0u) !=
(DefIndices[AntiDepReg] == ~0u)) &&
"Kill and Def maps aren't consistent for AntiDepReg!");
RegRefs.erase(AntiDepReg);
LastNewReg[AntiDepReg] = NewReg;
++Broken;
}
}
ScanInstruction(MI, Count);
}
return Broken;
}

95
lib/CodeGen/CriticalAntiDepBreaker.h Normal file
View File

@ -0,0 +1,95 @@
//=- llvm/CodeGen/CriticalAntiDepBreaker.h - Anti-Dep Support -*- C++ -*-=//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements the CriticalAntiDepBreaker class, which
// implements register anti-dependence breaking along a blocks
// critical path during post-RA scheduler.
//
//===----------------------------------------------------------------------===//
#ifndef LLVM_CODEGEN_CRITICALANTIDEPBREAKER_H
#define LLVM_CODEGEN_CRITICALANTIDEPBREAKER_H
#include "llvm/CodeGen/AntiDepBreaker.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/SmallSet.h"
namespace llvm {
class CriticalAntiDepBreaker : public AntiDepBreaker {
MachineFunction& MF;
MachineRegisterInfo &MRI;
const TargetRegisterInfo *TRI;
/// 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;
/// Classes - For live regs that are only used in one register class in a
/// live range, the register class. If the register is not live, the
/// corresponding value is null. If the register is live but used in
/// multiple register classes, the corresponding value is -1 casted to a
/// pointer.
const TargetRegisterClass *
Classes[TargetRegisterInfo::FirstVirtualRegister];
/// RegRegs - Map registers to all their references within a live range.
std::multimap<unsigned, MachineOperand *> RegRefs;
/// KillIndices - The index of the most recent kill (proceding bottom-up),
/// or ~0u if the register is not live.
unsigned KillIndices[TargetRegisterInfo::FirstVirtualRegister];
/// DefIndices - The index of the most recent complete def (proceding bottom
/// up), or ~0u if the register is live.
unsigned DefIndices[TargetRegisterInfo::FirstVirtualRegister];
/// KeepRegs - A set of registers which are live and cannot be changed to
/// break anti-dependencies.
SmallSet<unsigned, 4> KeepRegs;
public:
CriticalAntiDepBreaker(MachineFunction& MFi);
~CriticalAntiDepBreaker();
/// Start - Initialize anti-dep breaking for a new basic block.
void StartBlock(MachineBasicBlock *BB);
/// BreakAntiDependencies - Identifiy anti-dependencies along the critical path
/// of the ScheduleDAG and break them by renaming registers.
///
unsigned BreakAntiDependencies(std::vector<SUnit>& SUnits,
MachineBasicBlock::iterator& Begin,
MachineBasicBlock::iterator& End,
unsigned InsertPosIndex);
/// Observe - Update liveness information to account for the current
/// instruction, which will not be scheduled.
///
void Observe(MachineInstr *MI, unsigned Count, unsigned InsertPosIndex);
/// Finish - Finish anti-dep breaking for a basic block.
void FinishBlock();
private:
void PrescanInstruction(MachineInstr *MI);
void ScanInstruction(MachineInstr *MI, unsigned Count);
unsigned findSuitableFreeRegister(unsigned AntiDepReg,
unsigned LastNewReg,
const TargetRegisterClass *);
};
}
#endif

585
lib/CodeGen/PostRASchedulerList.cpp
View File

@ -19,6 +19,7 @@
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "post-RA-sched"
#include "CriticalAntiDepBreaker.h"
#include "ExactHazardRecognizer.h"
#include "SimpleHazardRecognizer.h"
#include "ScheduleDAGInstrs.h"
@ -40,6 +41,7 @@
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/Statistic.h"
#include <map>
#include <set>
@ -47,6 +49,7 @@ using namespace llvm;
STATISTIC(NumNoops, "Number of noops inserted");
STATISTIC(NumStalls, "Number of pipeline stalls");
STATISTIC(NumFixedAnti, "Number of fixed anti-dependencies");
// Post-RA scheduling is enabled with
// TargetSubtarget.enablePostRAScheduler(). This flag can be used to
@ -55,10 +58,11 @@ static cl::opt<bool>
EnablePostRAScheduler("post-RA-scheduler",
cl::desc("Enable scheduling after register allocation"),
cl::init(false), cl::Hidden);
static cl::opt<bool>
static cl::opt<std::string>
EnableAntiDepBreaking("break-anti-dependencies",
cl::desc("Break post-RA scheduling anti-dependencies"),
cl::init(true), cl::Hidden);
cl::desc("Break post-RA scheduling anti-dependencies: "
"\"critical\", \"all\", or \"none\""),
cl::init("none"), cl::Hidden);
static cl::opt<bool>
EnablePostRAHazardAvoidance("avoid-hazards",
cl::desc("Enable exact hazard avoidance"),
@ -116,56 +120,30 @@ namespace {
/// 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;
/// AntiDepBreak - Anti-dependence breaking object, or NULL if none
AntiDepBreaker *AntiDepBreak;
/// AA - AliasAnalysis for making memory reference queries.
AliasAnalysis *AA;
/// AntiDepMode - Anti-dependence breaking mode
TargetSubtarget::AntiDepBreakMode AntiDepMode;
/// Classes - For live regs that are only used in one register class in a
/// live range, the register class. If the register is not live, the
/// corresponding value is null. If the register is live but used in
/// multiple register classes, the corresponding value is -1 casted to a
/// pointer.
const TargetRegisterClass *
Classes[TargetRegisterInfo::FirstVirtualRegister];
/// RegRegs - Map registers to all their references within a live range.
std::multimap<unsigned, MachineOperand *> RegRefs;
/// KillIndices - The index of the most recent kill (proceding bottom-up),
/// or ~0u if the register is not live.
unsigned KillIndices[TargetRegisterInfo::FirstVirtualRegister];
/// DefIndices - The index of the most recent complete def (proceding bottom
/// up), or ~0u if the register is live.
unsigned DefIndices[TargetRegisterInfo::FirstVirtualRegister];
/// KeepRegs - A set of registers which are live and cannot be changed to
/// break anti-dependencies.
SmallSet<unsigned, 4> KeepRegs;
public:
SchedulePostRATDList(MachineFunction &MF,
const MachineLoopInfo &MLI,
const MachineDominatorTree &MDT,
ScheduleHazardRecognizer *HR,
AliasAnalysis *aa,
TargetSubtarget::AntiDepBreakMode adm)
AntiDepBreaker *ADB,
AliasAnalysis *aa)
: ScheduleDAGInstrs(MF, MLI, MDT), Topo(SUnits),
AllocatableSet(TRI->getAllocatableSet(MF)),
HazardRec(HR), AA(aa), AntiDepMode(adm) {}
HazardRec(HR), AntiDepBreak(ADB), AA(aa) {}
~SchedulePostRATDList() {
delete HazardRec;
}
/// StartBlock - Initialize register live-range state for scheduling in
@ -177,11 +155,6 @@ namespace {
///
void Schedule();
/// FixupKills - Fix register kill flags that have been made
/// invalid due to scheduling
///
void FixupKills(MachineBasicBlock *MBB);
/// Observe - Update liveness information to account for the current
/// instruction, which will not be scheduled.
///
@ -191,17 +164,16 @@ namespace {
///
void FinishBlock();
/// FixupKills - Fix register kill flags that have been made
/// invalid due to scheduling
///
void FixupKills(MachineBasicBlock *MBB);
private:
void PrescanInstruction(MachineInstr *MI);
void ScanInstruction(MachineInstr *MI, unsigned Count);
void ReleaseSucc(SUnit *SU, SDep *SuccEdge);
void ReleaseSuccessors(SUnit *SU);
void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
void ListScheduleTopDown();
bool BreakAntiDependencies();
unsigned findSuitableFreeRegister(unsigned AntiDepReg,
unsigned LastNewReg,
const TargetRegisterClass *);
void StartBlockForKills(MachineBasicBlock *BB);
// ToggleKillFlag - Toggle a register operand kill flag. Other
@ -250,8 +222,9 @@ bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
// Check for antidep breaking override...
if (EnableAntiDepBreaking.getPosition() > 0) {
AntiDepMode = (EnableAntiDepBreaking) ?
TargetSubtarget::ANTIDEP_CRITICAL : TargetSubtarget::ANTIDEP_NONE;
AntiDepMode = (EnableAntiDepBreaking == "all") ? TargetSubtarget::ANTIDEP_ALL :
(EnableAntiDepBreaking == "critical") ? TargetSubtarget::ANTIDEP_CRITICAL :
TargetSubtarget::ANTIDEP_NONE;
}
DEBUG(errs() << "PostRAScheduler\n");
@ -262,8 +235,12 @@ bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
ScheduleHazardRecognizer *HR = EnablePostRAHazardAvoidance ?
(ScheduleHazardRecognizer *)new ExactHazardRecognizer(InstrItins) :
(ScheduleHazardRecognizer *)new SimpleHazardRecognizer();
AntiDepBreaker *ADB =
(AntiDepMode == TargetSubtarget::ANTIDEP_ALL) ? NULL /* FIXME */ :
(AntiDepMode == TargetSubtarget::ANTIDEP_CRITICAL) ?
new CriticalAntiDepBreaker(Fn) : NULL;
SchedulePostRATDList Scheduler(Fn, MLI, MDT, HR, AA, AntiDepMode);
SchedulePostRATDList Scheduler(Fn, MLI, MDT, HR, ADB, AA);
// Loop over all of the basic blocks
for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
@ -311,6 +288,9 @@ bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
Scheduler.FixupKills(MBB);
}
delete HR;
delete ADB;
return true;
}
@ -321,78 +301,10 @@ void SchedulePostRATDList::StartBlock(MachineBasicBlock *BB) {
// Call the superclass.
ScheduleDAGInstrs::StartBlock(BB);
// Reset the hazard recognizer.
// Reset the hazard recognizer and anti-dep breaker.
HazardRec->Reset();
// Clear out the register class data.
std::fill(Classes, array_endof(Classes),
static_cast<const TargetRegisterClass *>(0));
// Initialize the indices to indicate that no registers are live.
std::fill(KillIndices, array_endof(KillIndices), ~0u);
std::fill(DefIndices, array_endof(DefIndices), BB->size());
// Clear "do not change" set.
KeepRegs.clear();
bool IsReturnBlock = (!BB->empty() && BB->back().getDesc().isReturn());
// Determine the live-out physregs for this block.
if (IsReturnBlock) {
// In a return block, examine the function live-out regs.
for (MachineRegisterInfo::liveout_iterator I = MRI.liveout_begin(),
E = MRI.liveout_end(); I != E; ++I) {
unsigned Reg = *I;
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[Reg] = BB->size();
DefIndices[Reg] = ~0u;
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[AliasReg] = BB->size();
DefIndices[AliasReg] = ~0u;
}
}
} else {
// In a non-return block, examine the live-in regs of all successors.
for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
SE = BB->succ_end(); SI != SE; ++SI)
for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
E = (*SI)->livein_end(); I != E; ++I) {
unsigned Reg = *I;
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[Reg] = BB->size();
DefIndices[Reg] = ~0u;
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[AliasReg] = BB->size();
DefIndices[AliasReg] = ~0u;
}
}
}
// Mark live-out callee-saved registers. In a return block this is
// all callee-saved registers. In non-return this is any
// callee-saved register that is not saved in the prolog.
const MachineFrameInfo *MFI = MF.getFrameInfo();
BitVector Pristine = MFI->getPristineRegs(BB);
for (const unsigned *I = TRI->getCalleeSavedRegs(); *I; ++I) {
unsigned Reg = *I;
if (!IsReturnBlock && !Pristine.test(Reg)) continue;
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[Reg] = BB->size();
DefIndices[Reg] = ~0u;
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[AliasReg] = BB->size();
DefIndices[AliasReg] = ~0u;
}
}
if (AntiDepBreak != NULL)
AntiDepBreak->StartBlock(BB);
}
/// Schedule - Schedule the instruction range using list scheduling.
@ -403,8 +315,11 @@ void SchedulePostRATDList::Schedule() {
// Build the scheduling graph.
BuildSchedGraph(AA);
if (AntiDepMode != TargetSubtarget::ANTIDEP_NONE) {
if (BreakAntiDependencies()) {
if (AntiDepBreak != NULL) {
unsigned Broken =
AntiDepBreak->BreakAntiDependencies(SUnits, Begin, InsertPos,
InsertPosIndex);
if (Broken > 0) {
// We made changes. Update the dependency graph.
// Theoretically we could update the graph in place:
// When a live range is changed to use a different register, remove
@ -415,6 +330,8 @@ void SchedulePostRATDList::Schedule() {
EntrySU = SUnit();
ExitSU = SUnit();
BuildSchedGraph(AA);
NumFixedAnti += Broken;
}
}
@ -432,436 +349,20 @@ void SchedulePostRATDList::Schedule() {
/// instruction, which will not be scheduled.
///
void SchedulePostRATDList::Observe(MachineInstr *MI, unsigned Count) {
assert(Count < InsertPosIndex && "Instruction index out of expected range!");
// Any register which was defined within the previous scheduling region
// may have been rescheduled and its lifetime may overlap with registers
// in ways not reflected in our current liveness state. For each such
// register, adjust the liveness state to be conservatively correct.
for (unsigned Reg = 0; Reg != TargetRegisterInfo::FirstVirtualRegister; ++Reg)
if (DefIndices[Reg] < InsertPosIndex && DefIndices[Reg] >= Count) {
assert(KillIndices[Reg] == ~0u && "Clobbered register is live!");
// Mark this register to be non-renamable.
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
// Move the def index to the end of the previous region, to reflect
// that the def could theoretically have been scheduled at the end.
DefIndices[Reg] = InsertPosIndex;
}
PrescanInstruction(MI);
ScanInstruction(MI, Count);
if (AntiDepBreak != NULL)
AntiDepBreak->Observe(MI, Count, InsertPosIndex);
}
/// FinishBlock - Clean up register live-range state.
///
void SchedulePostRATDList::FinishBlock() {
RegRefs.clear();
if (AntiDepBreak != NULL)
AntiDepBreak->FinishBlock();
// Call the superclass.
ScheduleDAGInstrs::FinishBlock();
}
/// CriticalPathStep - Return the next SUnit after SU on the bottom-up
/// critical path.
static SDep *CriticalPathStep(SUnit *SU) {
SDep *Next = 0;
unsigned NextDepth = 0;
// Find the predecessor edge with the greatest depth.
for (SUnit::pred_iterator P = SU->Preds.begin(), PE = SU->Preds.end();
P != PE; ++P) {
SUnit *PredSU = P->getSUnit();
unsigned PredLatency = P->getLatency();
unsigned PredTotalLatency = PredSU->getDepth() + PredLatency;
// In the case of a latency tie, prefer an anti-dependency edge over
// other types of edges.
if (NextDepth < PredTotalLatency ||
(NextDepth == PredTotalLatency && P->getKind() == SDep::Anti)) {
NextDepth = PredTotalLatency;
Next = &*P;
}
}
return Next;
}
void SchedulePostRATDList::PrescanInstruction(MachineInstr *MI) {
// Scan the register operands for this instruction and update
// Classes and RegRefs.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
const TargetRegisterClass *NewRC = 0;
if (i < MI->getDesc().getNumOperands())
NewRC = MI->getDesc().OpInfo[i].getRegClass(TRI);
// For now, only allow the register to be changed if its register
// class is consistent across all uses.
if (!Classes[Reg] && NewRC)
Classes[Reg] = NewRC;
else if (!NewRC || Classes[Reg] != NewRC)
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
// Now check for aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
// If an alias of the reg is used during the live range, give up.
// Note that this allows us to skip checking if AntiDepReg
// overlaps with any of the aliases, among other things.
unsigned AliasReg = *Alias;
if (Classes[AliasReg]) {
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
}
}
// If we're still willing to consider this register, note the reference.
if (Classes[Reg] != reinterpret_cast<TargetRegisterClass *>(-1))
RegRefs.insert(std::make_pair(Reg, &MO));
// It's not safe to change register allocation for source operands of
// that have special allocation requirements.
if (MO.isUse() && MI->getDesc().hasExtraSrcRegAllocReq()) {
if (KeepRegs.insert(Reg)) {
for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
*Subreg; ++Subreg)
KeepRegs.insert(*Subreg);
}
}
}
}
void SchedulePostRATDList::ScanInstruction(MachineInstr *MI,
unsigned Count) {
// Update liveness.
// Proceding upwards, registers that are defed but not used in this
// instruction are now dead.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
if (!MO.isDef()) continue;
// Ignore two-addr defs.
if (MI->isRegTiedToUseOperand(i)) continue;
DefIndices[Reg] = Count;
KillIndices[Reg] = ~0u;
assert(((KillIndices[Reg] == ~0u) !=
(DefIndices[Reg] == ~0u)) &&
"Kill and Def maps aren't consistent for Reg!");
KeepRegs.erase(Reg);
Classes[Reg] = 0;
RegRefs.erase(Reg);
// Repeat, for all subregs.
for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
*Subreg; ++Subreg) {
unsigned SubregReg = *Subreg;
DefIndices[SubregReg] = Count;
KillIndices[SubregReg] = ~0u;
KeepRegs.erase(SubregReg);
Classes[SubregReg] = 0;
RegRefs.erase(SubregReg);
}
// Conservatively mark super-registers as unusable.
for (const unsigned *Super = TRI->getSuperRegisters(Reg);
*Super; ++Super) {
unsigned SuperReg = *Super;
Classes[SuperReg] = reinterpret_cast<TargetRegisterClass *>(-1);
}
}
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
if (!MO.isUse()) continue;
const TargetRegisterClass *NewRC = 0;
if (i < MI->getDesc().getNumOperands())
NewRC = MI->getDesc().OpInfo[i].getRegClass(TRI);
// For now, only allow the register to be changed if its register
// class is consistent across all uses.
if (!Classes[Reg] && NewRC)
Classes[Reg] = NewRC;
else if (!NewRC || Classes[Reg] != NewRC)
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
RegRefs.insert(std::make_pair(Reg, &MO));
// It wasn't previously live but now it is, this is a kill.
if (KillIndices[Reg] == ~0u) {
KillIndices[Reg] = Count;
DefIndices[Reg] = ~0u;
assert(((KillIndices[Reg] == ~0u) !=
(DefIndices[Reg] == ~0u)) &&
"Kill and Def maps aren't consistent for Reg!");
}
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
if (KillIndices[AliasReg] == ~0u) {
KillIndices[AliasReg] = Count;
DefIndices[AliasReg] = ~0u;
}
}
}
}
unsigned
SchedulePostRATDList::findSuitableFreeRegister(unsigned AntiDepReg,
unsigned LastNewReg,
const TargetRegisterClass *RC) {
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;
// Don't replace a register with one that was recently used to repair
// an anti-dependence with this AntiDepReg, because that would
// re-introduce that anti-dependence.
if (NewReg == LastNewReg) continue;
// If NewReg is dead and NewReg's most recent def is not before
// AntiDepReg's kill, it's safe to replace AntiDepReg with NewReg.
assert(((KillIndices[AntiDepReg] == ~0u) != (DefIndices[AntiDepReg] == ~0u)) &&
"Kill and Def maps aren't consistent for AntiDepReg!");
assert(((KillIndices[NewReg] == ~0u) != (DefIndices[NewReg] == ~0u)) &&
"Kill and Def maps aren't consistent for NewReg!");
if (KillIndices[NewReg] != ~0u ||
Classes[NewReg] == reinterpret_cast<TargetRegisterClass *>(-1) ||
KillIndices[AntiDepReg] > DefIndices[NewReg])
continue;
return NewReg;
}
// No registers are free and available!
return 0;
}
/// BreakAntiDependencies - Identifiy anti-dependencies along the critical path
/// of the ScheduleDAG and break them by renaming registers.
///
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 (SUnits.empty()) return false;
// Find the node at the bottom of the critical path.
SUnit *Max = 0;
for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
SUnit *SU = &SUnits[i];
if (!Max || SU->getDepth() + SU->Latency > Max->getDepth() + Max->Latency)
Max = SU;
}
#ifndef NDEBUG
{
DEBUG(errs() << "Critical path has total latency "
<< (Max->getDepth() + Max->Latency) << "\n");
DEBUG(errs() << "Available regs:");
for (unsigned Reg = 0; Reg < TRI->getNumRegs(); ++Reg) {
if (KillIndices[Reg] == ~0u)
DEBUG(errs() << " " << TRI->getName(Reg));
}
DEBUG(errs() << '\n');
}
#endif
// Track progress along the critical path through the SUnit graph as we walk
// the instructions.
SUnit *CriticalPathSU = Max;
MachineInstr *CriticalPathMI = CriticalPathSU->getInstr();
// Consider this pattern:
// A = ...
// ... = A
// A = ...
// ... = A
// A = ...
// ... = A
// A = ...
// ... = A
// There are three anti-dependencies here, and without special care,
// we'd break all of them using the same register:
// A = ...
// ... = A
// B = ...
// ... = B
// B = ...
// ... = B
// B = ...
// ... = B
// because at each anti-dependence, B is the first register that
// isn't A which is free. This re-introduces anti-dependencies
// at all but one of the original anti-dependencies that we were
// trying to break. To avoid this, keep track of the most recent
// register that each register was replaced with, avoid
// using it to repair an anti-dependence on the same register.
// This lets us produce this:
// A = ...
// ... = A
// B = ...
// ... = B
// C = ...
// ... = C
// B = ...
// ... = B
// This still has an anti-dependence on B, but at least it isn't on the
// original critical path.
//
// TODO: If we tracked more than one register here, we could potentially
// fix that remaining critical edge too. This is a little more involved,
// because unlike the most recent register, less recent registers should
// still be considered, though only if no other registers are available.
unsigned LastNewReg[TargetRegisterInfo::FirstVirtualRegister] = {};
// Attempt to break anti-dependence edges on the critical path. Walk the
// instructions from the bottom up, tracking information about liveness
// as we go to help determine which registers are available.
bool Changed = false;
unsigned Count = InsertPosIndex - 1;
for (MachineBasicBlock::iterator I = InsertPos, E = Begin;
I != E; --Count) {
MachineInstr *MI = --I;
// Check if this instruction has a dependence on the critical path that
// is an anti-dependence that we may be able to break. If it is, set
// AntiDepReg to the non-zero register associated with the anti-dependence.
//
// We limit our attention to the critical path as a heuristic to avoid
// breaking anti-dependence edges that aren't going to significantly
// impact the overall schedule. There are a limited number of registers
// and we want to save them for the important edges.
//
// TODO: Instructions with multiple defs could have multiple
// anti-dependencies. The current code here only knows how to break one
// edge per instruction. Note that we'd have to be able to break all of
// the anti-dependencies in an instruction in order to be effective.
unsigned AntiDepReg = 0;
if (MI == CriticalPathMI) {
if (SDep *Edge = CriticalPathStep(CriticalPathSU)) {
SUnit *NextSU = Edge->getSUnit();
// Only consider anti-dependence edges.
if (Edge->getKind() == SDep::Anti) {
AntiDepReg = Edge->getReg();
assert(AntiDepReg != 0 && "Anti-dependence on reg0?");
if (!AllocatableSet.test(AntiDepReg))
// Don't break anti-dependencies on non-allocatable registers.
AntiDepReg = 0;
else if (KeepRegs.count(AntiDepReg))
// Don't break anti-dependencies if an use down below requires
// this exact register.
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
// scheduled past each other regardless.
//
// Also, if there are dependencies on other SUnits with the
// same register as the anti-dependency, don't attempt to
// break it.
for (SUnit::pred_iterator P = CriticalPathSU->Preds.begin(),
PE = CriticalPathSU->Preds.end(); P != PE; ++P)
if (P->getSUnit() == NextSU ?
(P->getKind() != SDep::Anti || P->getReg() != AntiDepReg) :
(P->getKind() == SDep::Data && P->getReg() == AntiDepReg)) {
AntiDepReg = 0;
break;
}
}
}
CriticalPathSU = NextSU;
CriticalPathMI = CriticalPathSU->getInstr();
} else {
// We've reached the end of the critical path.
CriticalPathSU = 0;
CriticalPathMI = 0;
}
}
PrescanInstruction(MI);
if (MI->getDesc().hasExtraDefRegAllocReq())
// If this instruction's defs have special allocation requirement, don't
// break this anti-dependency.
AntiDepReg = 0;
else if (AntiDepReg) {
// If this instruction has a use of AntiDepReg, breaking it
// is invalid.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
if (MO.isUse() && AntiDepReg == Reg) {
AntiDepReg = 0;
break;
}
}
}
// Determine AntiDepReg's register class, if it is live and is
// consistently used within a single class.
const TargetRegisterClass *RC = AntiDepReg != 0 ? Classes[AntiDepReg] : 0;
assert((AntiDepReg == 0 || RC != NULL) &&
"Register should be live if it's causing an anti-dependence!");
if (RC == reinterpret_cast<TargetRegisterClass *>(-1))
AntiDepReg = 0;
// Look for a suitable register to use to break the anti-depenence.
//
// TODO: Instead of picking the first free register, consider which might
// be the best.
if (AntiDepReg != 0) {
if (unsigned NewReg = findSuitableFreeRegister(AntiDepReg,
LastNewReg[AntiDepReg],
RC)) {
DEBUG(errs() << "Breaking anti-dependence edge on "
<< TRI->getName(AntiDepReg)
<< " with " << RegRefs.count(AntiDepReg) << " references"
<< " using " << TRI->getName(NewReg) << "!\n");
// Update the references to the old register to refer to the new
// register.
std::pair<std::multimap<unsigned, MachineOperand *>::iterator,
std::multimap<unsigned, MachineOperand *>::iterator>
Range = RegRefs.equal_range(AntiDepReg);
for (std::multimap<unsigned, MachineOperand *>::iterator
Q = Range.first, QE = Range.second; Q != QE; ++Q)
Q->second->setReg(NewReg);
// We just went back in time and modified history; the
// liveness information for the anti-depenence reg is now
// inconsistent. Set the state as if it were dead.
Classes[NewReg] = Classes[AntiDepReg];
DefIndices[NewReg] = DefIndices[AntiDepReg];
KillIndices[NewReg] = KillIndices[AntiDepReg];
assert(((KillIndices[NewReg] == ~0u) !=
(DefIndices[NewReg] == ~0u)) &&
"Kill and Def maps aren't consistent for NewReg!");
Classes[AntiDepReg] = 0;
DefIndices[AntiDepReg] = KillIndices[AntiDepReg];
KillIndices[AntiDepReg] = ~0u;
assert(((KillIndices[AntiDepReg] == ~0u) !=
(DefIndices[AntiDepReg] == ~0u)) &&
"Kill and Def maps aren't consistent for AntiDepReg!");
RegRefs.erase(AntiDepReg);
Changed = true;
LastNewReg[AntiDepReg] = NewReg;
}
}
ScanInstruction(MI, Count);
}
return Changed;
}
/// StartBlockForKills - Initialize register live-range state for updating kills
///
void SchedulePostRATDList::StartBlockForKills(MachineBasicBlock *BB) {

2
lib/Target/ARM/ARMSubtarget.h
View File

@ -130,7 +130,7 @@ protected:
/// for Thumb1.
bool enablePostRAScheduler(CodeGenOpt::Level OptLevel,
TargetSubtarget::AntiDepBreakMode& mode) const {
mode = TargetSubtarget::ANTIDEP_NONE;
mode = TargetSubtarget::ANTIDEP_CRITICAL;
return PostRAScheduler && OptLevel >= CodeGenOpt::Default;
}

4
test/CodeGen/X86/break-anti-dependencies.ll
View File

@ -1,7 +1,7 @@
; RUN: llc < %s -march=x86-64 -post-RA-scheduler -break-anti-dependencies=false > %t
; RUN: llc < %s -march=x86-64 -post-RA-scheduler -break-anti-dependencies=none > %t
; RUN: grep {%xmm0} %t | count 14
; RUN: not grep {%xmm1} %t
; RUN: llc < %s -march=x86-64 -post-RA-scheduler -break-anti-dependencies > %t
; RUN: llc < %s -march=x86-64 -post-RA-scheduler -break-anti-dependencies=critical > %t
; RUN: grep {%xmm0} %t | count 7
; RUN: grep {%xmm1} %t | count 7