2008-12-08 17:50:35 +00:00
|
|
|
//===---- ScheduleDAGInstrs.cpp - MachineInstr Rescheduling ---------------===//
|
2008-11-19 23:18:57 +00:00
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
2008-12-08 17:50:35 +00:00
|
|
|
// This implements the ScheduleDAGInstrs class, which implements re-scheduling
|
|
|
|
// of MachineInstrs.
|
2008-11-19 23:18:57 +00:00
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2012-12-03 16:50:05 +00:00
|
|
|
#include "llvm/CodeGen/ScheduleDAGInstrs.h"
|
|
|
|
#include "llvm/ADT/MapVector.h"
|
|
|
|
#include "llvm/ADT/SmallPtrSet.h"
|
|
|
|
#include "llvm/ADT/SmallSet.h"
|
2009-01-30 02:49:14 +00:00
|
|
|
#include "llvm/Analysis/AliasAnalysis.h"
|
2010-12-15 20:02:24 +00:00
|
|
|
#include "llvm/Analysis/ValueTracking.h"
|
2012-02-22 06:08:11 +00:00
|
|
|
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
|
2008-12-16 03:25:46 +00:00
|
|
|
#include "llvm/CodeGen/MachineFunctionPass.h"
|
2013-12-28 21:56:55 +00:00
|
|
|
#include "llvm/CodeGen/MachineInstrBuilder.h"
|
2009-09-25 20:36:54 +00:00
|
|
|
#include "llvm/CodeGen/MachineMemOperand.h"
|
2008-12-16 03:25:46 +00:00
|
|
|
#include "llvm/CodeGen/MachineRegisterInfo.h"
|
2008-12-04 01:35:46 +00:00
|
|
|
#include "llvm/CodeGen/PseudoSourceValue.h"
|
2012-06-06 19:47:35 +00:00
|
|
|
#include "llvm/CodeGen/RegisterPressure.h"
|
2012-11-28 05:13:24 +00:00
|
|
|
#include "llvm/CodeGen/ScheduleDFS.h"
|
2013-01-02 11:36:10 +00:00
|
|
|
#include "llvm/IR/Operator.h"
|
2011-06-29 01:14:12 +00:00
|
|
|
#include "llvm/MC/MCInstrItineraries.h"
|
2012-05-15 18:59:41 +00:00
|
|
|
#include "llvm/Support/CommandLine.h"
|
2008-11-19 23:18:57 +00:00
|
|
|
#include "llvm/Support/Debug.h"
|
2012-10-15 18:02:27 +00:00
|
|
|
#include "llvm/Support/Format.h"
|
2008-11-19 23:18:57 +00:00
|
|
|
#include "llvm/Support/raw_ostream.h"
|
2012-12-03 16:50:05 +00:00
|
|
|
#include "llvm/Target/TargetInstrInfo.h"
|
|
|
|
#include "llvm/Target/TargetMachine.h"
|
|
|
|
#include "llvm/Target/TargetRegisterInfo.h"
|
|
|
|
#include "llvm/Target/TargetSubtargetInfo.h"
|
2013-08-23 17:48:43 +00:00
|
|
|
#include <queue>
|
|
|
|
|
2008-11-19 23:18:57 +00:00
|
|
|
using namespace llvm;
|
|
|
|
|
2014-04-22 02:02:50 +00:00
|
|
|
#define DEBUG_TYPE "misched"
|
|
|
|
|
2012-05-15 18:59:41 +00:00
|
|
|
static cl::opt<bool> EnableAASchedMI("enable-aa-sched-mi", cl::Hidden,
|
|
|
|
cl::ZeroOrMore, cl::init(false),
|
|
|
|
cl::desc("Enable use of AA during MI GAD construction"));
|
|
|
|
|
2014-01-25 19:24:54 +00:00
|
|
|
static cl::opt<bool> UseTBAA("use-tbaa-in-sched-mi", cl::Hidden,
|
Reenable use of TBAA during CodeGen
We had disabled use of TBAA during CodeGen (even when otherwise using AA)
because the ptrtoint/inttoptr used by CGP for address sinking caused BasicAA to
miss basic type punning that it should catch (and, thus, we'd fail to override
TBAA when we should).
However, when AA is in use during CodeGen, CGP now uses normal GEPs and
bitcasts, instead of ptrtoint/inttoptr, when doing address sinking. As a
result, BasicAA should be able to make us do the right thing in the face of
type-punning, and it seems safe to enable use of TBAA again. self-hosting seems
fine on PPC64/Linux on the P7, with TBAA enabled and -misched=shuffle.
Note: We still don't update TBAA when merging stack slots, although because
BasicAA should now catch all such cases, this is no longer a blocking issue.
Nevertheless, I plan to commit code to deal with this properly in the near
future.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206093 91177308-0d34-0410-b5e6-96231b3b80d8
2014-04-12 01:26:00 +00:00
|
|
|
cl::init(true), cl::desc("Enable use of TBAA during MI GAD construction"));
|
2014-01-25 19:24:54 +00:00
|
|
|
|
2009-01-15 19:20:50 +00:00
|
|
|
ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
|
2008-12-16 03:25:46 +00:00
|
|
|
const MachineLoopInfo &mli,
|
2012-01-14 02:17:12 +00:00
|
|
|
const MachineDominatorTree &mdt,
|
2012-02-22 06:08:11 +00:00
|
|
|
bool IsPostRAFlag,
|
2013-12-28 21:56:55 +00:00
|
|
|
bool RemoveKillFlags,
|
2012-02-22 06:08:11 +00:00
|
|
|
LiveIntervals *lis)
|
2012-10-10 05:43:09 +00:00
|
|
|
: ScheduleDAG(mf), MLI(mli), MDT(mdt), MFI(mf.getFrameInfo()), LIS(lis),
|
2013-12-28 21:56:55 +00:00
|
|
|
IsPostRA(IsPostRAFlag), RemoveKillFlags(RemoveKillFlags),
|
2014-04-14 00:51:57 +00:00
|
|
|
CanHandleTerminators(false), FirstDbgValue(nullptr) {
|
2012-02-22 06:08:11 +00:00
|
|
|
assert((IsPostRA || LIS) && "PreRA scheduling requires LiveIntervals");
|
2011-06-02 20:07:12 +00:00
|
|
|
DbgValues.clear();
|
2012-02-22 06:08:13 +00:00
|
|
|
assert(!(IsPostRA && MRI.getNumVirtRegs()) &&
|
2012-02-21 04:51:23 +00:00
|
|
|
"Virtual registers must be removed prior to PostRA scheduling");
|
2012-09-18 18:20:00 +00:00
|
|
|
|
|
|
|
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
|
|
|
|
SchedModel.init(*ST.getSchedModel(), &ST, TII);
|
2009-10-18 19:58:47 +00:00
|
|
|
}
|
2008-11-19 23:18:57 +00:00
|
|
|
|
2009-01-30 02:49:14 +00:00
|
|
|
/// getUnderlyingObjectFromInt - This is the function that does the work of
|
|
|
|
/// looking through basic ptrtoint+arithmetic+inttoptr sequences.
|
|
|
|
static const Value *getUnderlyingObjectFromInt(const Value *V) {
|
|
|
|
do {
|
2009-07-17 20:58:59 +00:00
|
|
|
if (const Operator *U = dyn_cast<Operator>(V)) {
|
2009-01-30 02:49:14 +00:00
|
|
|
// If we find a ptrtoint, we can transfer control back to the
|
|
|
|
// regular getUnderlyingObjectFromInt.
|
2009-07-17 20:58:59 +00:00
|
|
|
if (U->getOpcode() == Instruction::PtrToInt)
|
2009-01-30 02:49:14 +00:00
|
|
|
return U->getOperand(0);
|
2012-11-28 03:42:49 +00:00
|
|
|
// If we find an add of a constant, a multiplied value, or a phi, it's
|
2009-01-30 02:49:14 +00:00
|
|
|
// likely that the other operand will lead us to the base
|
|
|
|
// object. We don't have to worry about the case where the
|
2009-08-07 01:26:06 +00:00
|
|
|
// object address is somehow being computed by the multiply,
|
2009-01-30 02:49:14 +00:00
|
|
|
// because our callers only care when the result is an
|
2012-10-26 04:27:49 +00:00
|
|
|
// identifiable object.
|
2009-07-17 20:58:59 +00:00
|
|
|
if (U->getOpcode() != Instruction::Add ||
|
2009-01-30 02:49:14 +00:00
|
|
|
(!isa<ConstantInt>(U->getOperand(1)) &&
|
2012-11-28 03:42:49 +00:00
|
|
|
Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
|
|
|
|
!isa<PHINode>(U->getOperand(1))))
|
2009-01-30 02:49:14 +00:00
|
|
|
return V;
|
|
|
|
V = U->getOperand(0);
|
|
|
|
} else {
|
|
|
|
return V;
|
|
|
|
}
|
2010-02-16 11:11:14 +00:00
|
|
|
assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
|
2009-01-30 02:49:14 +00:00
|
|
|
} while (1);
|
|
|
|
}
|
|
|
|
|
2012-12-10 18:49:16 +00:00
|
|
|
/// getUnderlyingObjects - This is a wrapper around GetUnderlyingObjects
|
2009-01-30 02:49:14 +00:00
|
|
|
/// and adds support for basic ptrtoint+arithmetic+inttoptr sequences.
|
2012-12-10 18:49:16 +00:00
|
|
|
static void getUnderlyingObjects(const Value *V,
|
|
|
|
SmallVectorImpl<Value *> &Objects) {
|
2014-04-15 07:22:52 +00:00
|
|
|
SmallPtrSet<const Value *, 16> Visited;
|
2012-12-10 18:49:16 +00:00
|
|
|
SmallVector<const Value *, 4> Working(1, V);
|
2009-01-30 02:49:14 +00:00
|
|
|
do {
|
2012-12-10 18:49:16 +00:00
|
|
|
V = Working.pop_back_val();
|
|
|
|
|
|
|
|
SmallVector<Value *, 4> Objs;
|
|
|
|
GetUnderlyingObjects(const_cast<Value *>(V), Objs);
|
|
|
|
|
2013-07-03 05:11:49 +00:00
|
|
|
for (SmallVectorImpl<Value *>::iterator I = Objs.begin(), IE = Objs.end();
|
2012-12-10 18:49:16 +00:00
|
|
|
I != IE; ++I) {
|
|
|
|
V = *I;
|
|
|
|
if (!Visited.insert(V))
|
|
|
|
continue;
|
|
|
|
if (Operator::getOpcode(V) == Instruction::IntToPtr) {
|
|
|
|
const Value *O =
|
|
|
|
getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
|
|
|
|
if (O->getType()->isPointerTy()) {
|
|
|
|
Working.push_back(O);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
Objects.push_back(const_cast<Value *>(V));
|
|
|
|
}
|
|
|
|
} while (!Working.empty());
|
2009-01-30 02:49:14 +00:00
|
|
|
}
|
|
|
|
|
2014-04-15 07:22:52 +00:00
|
|
|
typedef PointerUnion<const Value *, const PseudoSourceValue *> ValueType;
|
|
|
|
typedef SmallVector<PointerIntPair<ValueType, 1, bool>, 4>
|
2013-06-29 18:41:17 +00:00
|
|
|
UnderlyingObjectsVector;
|
|
|
|
|
2012-12-10 18:49:16 +00:00
|
|
|
/// getUnderlyingObjectsForInstr - If this machine instr has memory reference
|
2009-01-30 02:49:14 +00:00
|
|
|
/// information and it can be tracked to a normal reference to a known
|
2012-12-10 18:49:16 +00:00
|
|
|
/// object, return the Value for that object.
|
|
|
|
static void getUnderlyingObjectsForInstr(const MachineInstr *MI,
|
2013-06-29 18:41:17 +00:00
|
|
|
const MachineFrameInfo *MFI,
|
|
|
|
UnderlyingObjectsVector &Objects) {
|
2009-01-30 02:49:14 +00:00
|
|
|
if (!MI->hasOneMemOperand() ||
|
2014-04-15 07:22:52 +00:00
|
|
|
(!(*MI->memoperands_begin())->getValue() &&
|
|
|
|
!(*MI->memoperands_begin())->getPseudoValue()) ||
|
2009-09-25 20:36:54 +00:00
|
|
|
(*MI->memoperands_begin())->isVolatile())
|
2012-12-10 18:49:16 +00:00
|
|
|
return;
|
2009-01-30 02:49:14 +00:00
|
|
|
|
2014-04-15 07:22:52 +00:00
|
|
|
if (const PseudoSourceValue *PSV =
|
|
|
|
(*MI->memoperands_begin())->getPseudoValue()) {
|
2014-02-20 05:06:26 +00:00
|
|
|
// For now, ignore PseudoSourceValues which may alias LLVM IR values
|
|
|
|
// because the code that uses this function has no way to cope with
|
|
|
|
// such aliases.
|
2014-02-20 06:35:31 +00:00
|
|
|
if (!PSV->isAliased(MFI)) {
|
|
|
|
bool MayAlias = PSV->mayAlias(MFI);
|
2014-04-15 07:22:52 +00:00
|
|
|
Objects.push_back(UnderlyingObjectsVector::value_type(PSV, MayAlias));
|
2014-02-20 06:35:31 +00:00
|
|
|
}
|
2014-02-20 05:06:26 +00:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2014-04-15 07:22:52 +00:00
|
|
|
const Value *V = (*MI->memoperands_begin())->getValue();
|
|
|
|
if (!V)
|
|
|
|
return;
|
|
|
|
|
2012-12-10 18:49:16 +00:00
|
|
|
SmallVector<Value *, 4> Objs;
|
|
|
|
getUnderlyingObjects(V, Objs);
|
|
|
|
|
2013-07-03 05:11:49 +00:00
|
|
|
for (SmallVectorImpl<Value *>::iterator I = Objs.begin(), IE = Objs.end();
|
|
|
|
I != IE; ++I) {
|
2012-12-10 18:49:16 +00:00
|
|
|
V = *I;
|
|
|
|
|
2014-02-20 05:06:26 +00:00
|
|
|
if (!isIdentifiedObject(V)) {
|
2012-12-10 18:49:16 +00:00
|
|
|
Objects.clear();
|
|
|
|
return;
|
|
|
|
}
|
2009-10-18 18:16:27 +00:00
|
|
|
|
2014-02-20 05:06:26 +00:00
|
|
|
Objects.push_back(UnderlyingObjectsVector::value_type(V, true));
|
2012-12-10 18:49:16 +00:00
|
|
|
}
|
2009-01-30 02:49:14 +00:00
|
|
|
}
|
|
|
|
|
2012-04-20 20:05:21 +00:00
|
|
|
void ScheduleDAGInstrs::startBlock(MachineBasicBlock *bb) {
|
|
|
|
BB = bb;
|
2009-02-10 23:27:53 +00:00
|
|
|
}
|
|
|
|
|
2012-03-07 23:00:49 +00:00
|
|
|
void ScheduleDAGInstrs::finishBlock() {
|
2012-04-20 20:24:33 +00:00
|
|
|
// Subclasses should no longer refer to the old block.
|
2014-04-14 00:51:57 +00:00
|
|
|
BB = nullptr;
|
2012-03-07 05:21:52 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Initialize the DAG and common scheduler state for the current scheduling
|
|
|
|
/// region. This does not actually create the DAG, only clears it. The
|
|
|
|
/// scheduling driver may call BuildSchedGraph multiple times per scheduling
|
|
|
|
/// region.
|
|
|
|
void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
|
|
|
|
MachineBasicBlock::iterator begin,
|
|
|
|
MachineBasicBlock::iterator end,
|
2013-08-23 17:48:33 +00:00
|
|
|
unsigned regioninstrs) {
|
2012-04-20 20:05:21 +00:00
|
|
|
assert(bb == BB && "startBlock should set BB");
|
2012-03-09 04:29:02 +00:00
|
|
|
RegionBegin = begin;
|
|
|
|
RegionEnd = end;
|
2013-08-23 17:48:33 +00:00
|
|
|
NumRegionInstrs = regioninstrs;
|
2012-03-07 05:21:52 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Close the current scheduling region. Don't clear any state in case the
|
|
|
|
/// driver wants to refer to the previous scheduling region.
|
|
|
|
void ScheduleDAGInstrs::exitRegion() {
|
|
|
|
// Nothing to do.
|
|
|
|
}
|
|
|
|
|
2012-03-07 23:00:49 +00:00
|
|
|
/// addSchedBarrierDeps - Add dependencies from instructions in the current
|
2010-10-23 02:10:46 +00:00
|
|
|
/// list of instructions being scheduled to scheduling barrier by adding
|
|
|
|
/// the exit SU to the register defs and use list. This is because we want to
|
|
|
|
/// make sure instructions which define registers that are either used by
|
|
|
|
/// the terminator or are live-out are properly scheduled. This is
|
|
|
|
/// especially important when the definition latency of the return value(s)
|
|
|
|
/// are too high to be hidden by the branch or when the liveout registers
|
|
|
|
/// used by instructions in the fallthrough block.
|
2012-03-07 23:00:49 +00:00
|
|
|
void ScheduleDAGInstrs::addSchedBarrierDeps() {
|
2014-04-14 00:51:57 +00:00
|
|
|
MachineInstr *ExitMI = RegionEnd != BB->end() ? &*RegionEnd : nullptr;
|
2010-10-23 02:10:46 +00:00
|
|
|
ExitSU.setInstr(ExitMI);
|
|
|
|
bool AllDepKnown = ExitMI &&
|
2011-12-07 07:15:52 +00:00
|
|
|
(ExitMI->isCall() || ExitMI->isBarrier());
|
2010-10-23 02:10:46 +00:00
|
|
|
if (ExitMI && AllDepKnown) {
|
|
|
|
// If it's a call or a barrier, add dependencies on the defs and uses of
|
|
|
|
// instruction.
|
|
|
|
for (unsigned i = 0, e = ExitMI->getNumOperands(); i != e; ++i) {
|
|
|
|
const MachineOperand &MO = ExitMI->getOperand(i);
|
|
|
|
if (!MO.isReg() || MO.isDef()) continue;
|
|
|
|
unsigned Reg = MO.getReg();
|
|
|
|
if (Reg == 0) continue;
|
|
|
|
|
2012-01-14 02:17:18 +00:00
|
|
|
if (TRI->isPhysicalRegister(Reg))
|
2013-01-21 18:18:53 +00:00
|
|
|
Uses.insert(PhysRegSUOper(&ExitSU, -1, Reg));
|
2012-03-16 05:04:25 +00:00
|
|
|
else {
|
2012-01-14 02:17:18 +00:00
|
|
|
assert(!IsPostRA && "Virtual register encountered after regalloc.");
|
2012-12-01 01:22:44 +00:00
|
|
|
if (MO.readsReg()) // ignore undef operands
|
|
|
|
addVRegUseDeps(&ExitSU, i);
|
2012-03-16 05:04:25 +00:00
|
|
|
}
|
2010-10-23 02:10:46 +00:00
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// For others, e.g. fallthrough, conditional branch, assume the exit
|
2010-10-27 23:17:17 +00:00
|
|
|
// uses all the registers that are livein to the successor blocks.
|
2012-03-16 17:38:19 +00:00
|
|
|
assert(Uses.empty() && "Uses in set before adding deps?");
|
2010-10-27 23:17:17 +00:00
|
|
|
for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
|
|
|
|
SE = BB->succ_end(); SI != SE; ++SI)
|
|
|
|
for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
|
2011-05-05 19:24:06 +00:00
|
|
|
E = (*SI)->livein_end(); I != E; ++I) {
|
2010-10-27 23:17:17 +00:00
|
|
|
unsigned Reg = *I;
|
2012-03-16 17:38:19 +00:00
|
|
|
if (!Uses.contains(Reg))
|
2013-01-21 18:18:53 +00:00
|
|
|
Uses.insert(PhysRegSUOper(&ExitSU, -1, Reg));
|
2010-10-27 23:17:17 +00:00
|
|
|
}
|
2010-10-23 02:10:46 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-02-23 01:52:38 +00:00
|
|
|
/// MO is an operand of SU's instruction that defines a physical register. Add
|
|
|
|
/// data dependencies from SU to any uses of the physical register.
|
2012-08-23 00:39:43 +00:00
|
|
|
void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
|
|
|
|
const MachineOperand &MO = SU->getInstr()->getOperand(OperIdx);
|
2012-02-23 01:52:38 +00:00
|
|
|
assert(MO.isDef() && "expect physreg def");
|
2012-01-14 02:17:15 +00:00
|
|
|
|
|
|
|
// Ask the target if address-backscheduling is desirable, and if so how much.
|
|
|
|
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
|
|
|
|
|
2012-06-01 23:28:30 +00:00
|
|
|
for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
|
|
|
|
Alias.isValid(); ++Alias) {
|
2012-02-24 07:04:55 +00:00
|
|
|
if (!Uses.contains(*Alias))
|
2012-02-23 01:52:38 +00:00
|
|
|
continue;
|
2013-01-21 18:18:53 +00:00
|
|
|
for (Reg2SUnitsMap::iterator I = Uses.find(*Alias); I != Uses.end(); ++I) {
|
|
|
|
SUnit *UseSU = I->SU;
|
2012-01-14 02:17:15 +00:00
|
|
|
if (UseSU == SU)
|
|
|
|
continue;
|
2012-10-08 18:54:00 +00:00
|
|
|
|
|
|
|
// Adjust the dependence latency using operand def/use information,
|
|
|
|
// then allow the target to perform its own adjustments.
|
2013-01-21 18:18:53 +00:00
|
|
|
int UseOp = I->OpIdx;
|
2014-04-14 00:51:57 +00:00
|
|
|
MachineInstr *RegUse = nullptr;
|
2012-11-12 19:28:57 +00:00
|
|
|
SDep Dep;
|
|
|
|
if (UseOp < 0)
|
|
|
|
Dep = SDep(SU, SDep::Artificial);
|
|
|
|
else {
|
2013-04-13 06:07:40 +00:00
|
|
|
// Set the hasPhysRegDefs only for physreg defs that have a use within
|
|
|
|
// the scheduling region.
|
|
|
|
SU->hasPhysRegDefs = true;
|
2012-11-12 19:28:57 +00:00
|
|
|
Dep = SDep(SU, SDep::Data, *Alias);
|
|
|
|
RegUse = UseSU->getInstr();
|
|
|
|
}
|
|
|
|
Dep.setLatency(
|
2013-06-15 04:49:57 +00:00
|
|
|
SchedModel.computeOperandLatency(SU->getInstr(), OperIdx, RegUse,
|
|
|
|
UseOp));
|
2012-10-08 18:53:57 +00:00
|
|
|
|
2012-11-12 19:28:57 +00:00
|
|
|
ST.adjustSchedDependency(SU, UseSU, Dep);
|
|
|
|
UseSU->addPred(Dep);
|
2012-01-14 02:17:15 +00:00
|
|
|
}
|
2012-02-23 01:52:38 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// addPhysRegDeps - Add register dependencies (data, anti, and output) from
|
|
|
|
/// this SUnit to following instructions in the same scheduling region that
|
|
|
|
/// depend the physical register referenced at OperIdx.
|
|
|
|
void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
|
2013-12-28 21:56:55 +00:00
|
|
|
MachineInstr *MI = SU->getInstr();
|
|
|
|
MachineOperand &MO = MI->getOperand(OperIdx);
|
2012-02-23 01:52:38 +00:00
|
|
|
|
|
|
|
// Optionally add output and anti dependencies. For anti
|
|
|
|
// dependencies we use a latency of 0 because for a multi-issue
|
|
|
|
// target we want to allow the defining instruction to issue
|
|
|
|
// in the same cycle as the using instruction.
|
|
|
|
// TODO: Using a latency of 1 here for output dependencies assumes
|
|
|
|
// there's no cost for reusing registers.
|
|
|
|
SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
|
2012-06-01 23:28:30 +00:00
|
|
|
for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
|
|
|
|
Alias.isValid(); ++Alias) {
|
2012-02-24 07:04:55 +00:00
|
|
|
if (!Defs.contains(*Alias))
|
2012-02-23 01:52:38 +00:00
|
|
|
continue;
|
2013-01-21 18:18:53 +00:00
|
|
|
for (Reg2SUnitsMap::iterator I = Defs.find(*Alias); I != Defs.end(); ++I) {
|
|
|
|
SUnit *DefSU = I->SU;
|
2012-02-23 01:52:38 +00:00
|
|
|
if (DefSU == &ExitSU)
|
|
|
|
continue;
|
|
|
|
if (DefSU != SU &&
|
|
|
|
(Kind != SDep::Output || !MO.isDead() ||
|
|
|
|
!DefSU->getInstr()->registerDefIsDead(*Alias))) {
|
|
|
|
if (Kind == SDep::Anti)
|
2012-11-06 03:13:46 +00:00
|
|
|
DefSU->addPred(SDep(SU, Kind, /*Reg=*/*Alias));
|
2012-02-23 01:52:38 +00:00
|
|
|
else {
|
2012-11-06 03:13:46 +00:00
|
|
|
SDep Dep(SU, Kind, /*Reg=*/*Alias);
|
2013-06-15 04:49:57 +00:00
|
|
|
Dep.setLatency(
|
|
|
|
SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
|
2012-11-06 03:13:46 +00:00
|
|
|
DefSU->addPred(Dep);
|
2012-01-14 02:17:15 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2012-02-23 01:52:38 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!MO.isDef()) {
|
2013-04-13 06:07:40 +00:00
|
|
|
SU->hasPhysRegUses = true;
|
2012-02-23 01:52:38 +00:00
|
|
|
// Either insert a new Reg2SUnits entry with an empty SUnits list, or
|
|
|
|
// retrieve the existing SUnits list for this register's uses.
|
|
|
|
// Push this SUnit on the use list.
|
2013-01-21 18:18:53 +00:00
|
|
|
Uses.insert(PhysRegSUOper(SU, OperIdx, MO.getReg()));
|
2013-12-28 21:56:55 +00:00
|
|
|
if (RemoveKillFlags)
|
|
|
|
MO.setIsKill(false);
|
2012-02-23 01:52:38 +00:00
|
|
|
}
|
|
|
|
else {
|
2012-08-23 00:39:43 +00:00
|
|
|
addPhysRegDataDeps(SU, OperIdx);
|
2013-01-21 18:18:53 +00:00
|
|
|
unsigned Reg = MO.getReg();
|
2012-01-14 02:17:15 +00:00
|
|
|
|
2012-02-23 01:52:38 +00:00
|
|
|
// clear this register's use list
|
2013-01-21 18:18:53 +00:00
|
|
|
if (Uses.contains(Reg))
|
|
|
|
Uses.eraseAll(Reg);
|
|
|
|
|
|
|
|
if (!MO.isDead()) {
|
|
|
|
Defs.eraseAll(Reg);
|
|
|
|
} else if (SU->isCall) {
|
|
|
|
// Calls will not be reordered because of chain dependencies (see
|
|
|
|
// below). Since call operands are dead, calls may continue to be added
|
|
|
|
// to the DefList making dependence checking quadratic in the size of
|
|
|
|
// the block. Instead, we leave only one call at the back of the
|
|
|
|
// DefList.
|
|
|
|
Reg2SUnitsMap::RangePair P = Defs.equal_range(Reg);
|
|
|
|
Reg2SUnitsMap::iterator B = P.first;
|
|
|
|
Reg2SUnitsMap::iterator I = P.second;
|
|
|
|
for (bool isBegin = I == B; !isBegin; /* empty */) {
|
|
|
|
isBegin = (--I) == B;
|
|
|
|
if (!I->SU->isCall)
|
|
|
|
break;
|
|
|
|
I = Defs.erase(I);
|
|
|
|
}
|
2012-01-14 02:17:15 +00:00
|
|
|
}
|
2013-01-21 18:18:53 +00:00
|
|
|
|
2012-02-23 01:52:38 +00:00
|
|
|
// Defs are pushed in the order they are visited and never reordered.
|
2013-01-21 18:18:53 +00:00
|
|
|
Defs.insert(PhysRegSUOper(SU, OperIdx, Reg));
|
2012-01-14 02:17:15 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-01-14 02:17:18 +00:00
|
|
|
/// addVRegDefDeps - Add register output and data dependencies from this SUnit
|
|
|
|
/// to instructions that occur later in the same scheduling region if they read
|
|
|
|
/// from or write to the virtual register defined at OperIdx.
|
|
|
|
///
|
|
|
|
/// TODO: Hoist loop induction variable increments. This has to be
|
|
|
|
/// reevaluated. Generally, IV scheduling should be done before coalescing.
|
|
|
|
void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
|
|
|
|
const MachineInstr *MI = SU->getInstr();
|
|
|
|
unsigned Reg = MI->getOperand(OperIdx).getReg();
|
|
|
|
|
2012-07-28 01:48:15 +00:00
|
|
|
// Singly defined vregs do not have output/anti dependencies.
|
2012-02-22 18:34:49 +00:00
|
|
|
// The current operand is a def, so we have at least one.
|
2012-07-28 01:48:15 +00:00
|
|
|
// Check here if there are any others...
|
2012-07-30 23:48:17 +00:00
|
|
|
if (MRI.hasOneDef(Reg))
|
2012-07-28 01:48:15 +00:00
|
|
|
return;
|
2012-02-22 06:08:13 +00:00
|
|
|
|
2012-01-14 02:17:18 +00:00
|
|
|
// Add output dependence to the next nearest def of this vreg.
|
|
|
|
//
|
|
|
|
// Unless this definition is dead, the output dependence should be
|
|
|
|
// transitively redundant with antidependencies from this definition's
|
|
|
|
// uses. We're conservative for now until we have a way to guarantee the uses
|
|
|
|
// are not eliminated sometime during scheduling. The output dependence edge
|
|
|
|
// is also useful if output latency exceeds def-use latency.
|
2012-04-20 20:05:28 +00:00
|
|
|
VReg2SUnitMap::iterator DefI = VRegDefs.find(Reg);
|
2012-02-22 21:59:00 +00:00
|
|
|
if (DefI == VRegDefs.end())
|
|
|
|
VRegDefs.insert(VReg2SUnit(Reg, SU));
|
|
|
|
else {
|
|
|
|
SUnit *DefSU = DefI->SU;
|
|
|
|
if (DefSU != SU && DefSU != &ExitSU) {
|
2012-11-06 03:13:46 +00:00
|
|
|
SDep Dep(SU, SDep::Output, Reg);
|
2013-06-15 04:49:57 +00:00
|
|
|
Dep.setLatency(
|
|
|
|
SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
|
2012-11-06 03:13:46 +00:00
|
|
|
DefSU->addPred(Dep);
|
2012-02-22 21:59:00 +00:00
|
|
|
}
|
|
|
|
DefI->SU = SU;
|
2012-01-14 02:17:18 +00:00
|
|
|
}
|
2012-02-22 06:08:11 +00:00
|
|
|
}
|
2012-01-14 02:17:18 +00:00
|
|
|
|
2012-02-22 06:08:11 +00:00
|
|
|
/// addVRegUseDeps - Add a register data dependency if the instruction that
|
|
|
|
/// defines the virtual register used at OperIdx is mapped to an SUnit. Add a
|
|
|
|
/// register antidependency from this SUnit to instructions that occur later in
|
|
|
|
/// the same scheduling region if they write the virtual register.
|
|
|
|
///
|
|
|
|
/// TODO: Handle ExitSU "uses" properly.
|
|
|
|
void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
|
|
|
|
MachineInstr *MI = SU->getInstr();
|
|
|
|
unsigned Reg = MI->getOperand(OperIdx).getReg();
|
|
|
|
|
2013-08-23 17:48:39 +00:00
|
|
|
// Record this local VReg use.
|
2013-08-30 04:36:57 +00:00
|
|
|
VReg2UseMap::iterator UI = VRegUses.find(Reg);
|
|
|
|
for (; UI != VRegUses.end(); ++UI) {
|
|
|
|
if (UI->SU == SU)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (UI == VRegUses.end())
|
|
|
|
VRegUses.insert(VReg2SUnit(Reg, SU));
|
2013-08-23 17:48:39 +00:00
|
|
|
|
2012-02-22 06:08:11 +00:00
|
|
|
// Lookup this operand's reaching definition.
|
|
|
|
assert(LIS && "vreg dependencies requires LiveIntervals");
|
2013-10-10 21:28:52 +00:00
|
|
|
LiveQueryResult LRQ
|
|
|
|
= LIS->getInterval(Reg).Query(LIS->getInstructionIndex(MI));
|
2012-05-20 02:44:38 +00:00
|
|
|
VNInfo *VNI = LRQ.valueIn();
|
2012-04-24 18:04:41 +00:00
|
|
|
|
2012-02-23 03:16:24 +00:00
|
|
|
// VNI will be valid because MachineOperand::readsReg() is checked by caller.
|
2012-05-20 02:44:38 +00:00
|
|
|
assert(VNI && "No value to read by operand");
|
2012-02-22 06:08:11 +00:00
|
|
|
MachineInstr *Def = LIS->getInstructionFromIndex(VNI->def);
|
2012-02-23 03:16:24 +00:00
|
|
|
// Phis and other noninstructions (after coalescing) have a NULL Def.
|
2012-02-22 06:08:11 +00:00
|
|
|
if (Def) {
|
|
|
|
SUnit *DefSU = getSUnit(Def);
|
|
|
|
if (DefSU) {
|
|
|
|
// The reaching Def lives within this scheduling region.
|
|
|
|
// Create a data dependence.
|
2012-11-06 03:13:46 +00:00
|
|
|
SDep dep(DefSU, SDep::Data, Reg);
|
2012-10-08 18:53:57 +00:00
|
|
|
// Adjust the dependence latency using operand def/use information, then
|
|
|
|
// allow the target to perform its own adjustments.
|
|
|
|
int DefOp = Def->findRegisterDefOperandIdx(Reg);
|
2013-06-15 04:49:57 +00:00
|
|
|
dep.setLatency(SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx));
|
2012-10-08 18:53:57 +00:00
|
|
|
|
|
|
|
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
|
|
|
|
ST.adjustSchedDependency(DefSU, SU, const_cast<SDep &>(dep));
|
2012-02-22 06:08:11 +00:00
|
|
|
SU->addPred(dep);
|
2012-01-14 02:17:18 +00:00
|
|
|
}
|
|
|
|
}
|
2012-02-22 06:08:11 +00:00
|
|
|
|
|
|
|
// Add antidependence to the following def of the vreg it uses.
|
2012-04-20 20:05:28 +00:00
|
|
|
VReg2SUnitMap::iterator DefI = VRegDefs.find(Reg);
|
2012-02-22 21:59:00 +00:00
|
|
|
if (DefI != VRegDefs.end() && DefI->SU != SU)
|
2012-11-06 03:13:46 +00:00
|
|
|
DefI->SU->addPred(SDep(SU, SDep::Anti, Reg));
|
2012-01-14 02:17:18 +00:00
|
|
|
}
|
|
|
|
|
2012-05-15 18:59:41 +00:00
|
|
|
/// Return true if MI is an instruction we are unable to reason about
|
|
|
|
/// (like a call or something with unmodeled side effects).
|
|
|
|
static inline bool isGlobalMemoryObject(AliasAnalysis *AA, MachineInstr *MI) {
|
|
|
|
if (MI->isCall() || MI->hasUnmodeledSideEffects() ||
|
2012-08-29 21:19:21 +00:00
|
|
|
(MI->hasOrderedMemoryRef() &&
|
2012-05-15 18:59:41 +00:00
|
|
|
(!MI->mayLoad() || !MI->isInvariantLoad(AA))))
|
|
|
|
return true;
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// This MI might have either incomplete info, or known to be unsafe
|
|
|
|
// to deal with (i.e. volatile object).
|
|
|
|
static inline bool isUnsafeMemoryObject(MachineInstr *MI,
|
|
|
|
const MachineFrameInfo *MFI) {
|
|
|
|
if (!MI || MI->memoperands_empty())
|
|
|
|
return true;
|
|
|
|
// We purposefully do no check for hasOneMemOperand() here
|
|
|
|
// in hope to trigger an assert downstream in order to
|
|
|
|
// finish implementation.
|
|
|
|
if ((*MI->memoperands_begin())->isVolatile() ||
|
|
|
|
MI->hasUnmodeledSideEffects())
|
|
|
|
return true;
|
2014-04-15 07:22:52 +00:00
|
|
|
|
|
|
|
if ((*MI->memoperands_begin())->getPseudoValue()) {
|
|
|
|
// Similarly to getUnderlyingObjectForInstr:
|
|
|
|
// For now, ignore PseudoSourceValues which may alias LLVM IR values
|
|
|
|
// because the code that uses this function has no way to cope with
|
|
|
|
// such aliases.
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2012-05-15 18:59:41 +00:00
|
|
|
const Value *V = (*MI->memoperands_begin())->getValue();
|
|
|
|
if (!V)
|
|
|
|
return true;
|
|
|
|
|
2012-12-10 18:49:16 +00:00
|
|
|
SmallVector<Value *, 4> Objs;
|
|
|
|
getUnderlyingObjects(V, Objs);
|
2013-07-03 05:11:49 +00:00
|
|
|
for (SmallVectorImpl<Value *>::iterator I = Objs.begin(),
|
|
|
|
IE = Objs.end(); I != IE; ++I) {
|
2012-12-10 18:49:16 +00:00
|
|
|
// Does this pointer refer to a distinct and identifiable object?
|
2014-04-15 07:22:52 +00:00
|
|
|
if (!isIdentifiedObject(*I))
|
2012-05-15 18:59:41 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// This returns true if the two MIs need a chain edge betwee them.
|
|
|
|
/// If these are not even memory operations, we still may need
|
|
|
|
/// chain deps between them. The question really is - could
|
|
|
|
/// these two MIs be reordered during scheduling from memory dependency
|
|
|
|
/// point of view.
|
|
|
|
static bool MIsNeedChainEdge(AliasAnalysis *AA, const MachineFrameInfo *MFI,
|
|
|
|
MachineInstr *MIa,
|
|
|
|
MachineInstr *MIb) {
|
|
|
|
// Cover a trivial case - no edge is need to itself.
|
|
|
|
if (MIa == MIb)
|
|
|
|
return false;
|
|
|
|
|
2014-01-08 21:52:02 +00:00
|
|
|
// FIXME: Need to handle multiple memory operands to support all targets.
|
|
|
|
if (!MIa->hasOneMemOperand() || !MIb->hasOneMemOperand())
|
|
|
|
return true;
|
|
|
|
|
2012-05-15 18:59:41 +00:00
|
|
|
if (isUnsafeMemoryObject(MIa, MFI) || isUnsafeMemoryObject(MIb, MFI))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
// If we are dealing with two "normal" loads, we do not need an edge
|
|
|
|
// between them - they could be reordered.
|
|
|
|
if (!MIa->mayStore() && !MIb->mayStore())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// To this point analysis is generic. From here on we do need AA.
|
|
|
|
if (!AA)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
MachineMemOperand *MMOa = *MIa->memoperands_begin();
|
|
|
|
MachineMemOperand *MMOb = *MIb->memoperands_begin();
|
|
|
|
|
2014-04-15 07:22:52 +00:00
|
|
|
if (!MMOa->getValue() || !MMOb->getValue())
|
|
|
|
return true;
|
|
|
|
|
2012-05-15 18:59:41 +00:00
|
|
|
// The following interface to AA is fashioned after DAGCombiner::isAlias
|
|
|
|
// and operates with MachineMemOperand offset with some important
|
|
|
|
// assumptions:
|
|
|
|
// - LLVM fundamentally assumes flat address spaces.
|
|
|
|
// - MachineOperand offset can *only* result from legalization and
|
|
|
|
// cannot affect queries other than the trivial case of overlap
|
|
|
|
// checking.
|
|
|
|
// - These offsets never wrap and never step outside
|
|
|
|
// of allocated objects.
|
|
|
|
// - There should never be any negative offsets here.
|
|
|
|
//
|
|
|
|
// FIXME: Modify API to hide this math from "user"
|
|
|
|
// FIXME: Even before we go to AA we can reason locally about some
|
|
|
|
// memory objects. It can save compile time, and possibly catch some
|
|
|
|
// corner cases not currently covered.
|
|
|
|
|
|
|
|
assert ((MMOa->getOffset() >= 0) && "Negative MachineMemOperand offset");
|
|
|
|
assert ((MMOb->getOffset() >= 0) && "Negative MachineMemOperand offset");
|
|
|
|
|
|
|
|
int64_t MinOffset = std::min(MMOa->getOffset(), MMOb->getOffset());
|
|
|
|
int64_t Overlapa = MMOa->getSize() + MMOa->getOffset() - MinOffset;
|
|
|
|
int64_t Overlapb = MMOb->getSize() + MMOb->getOffset() - MinOffset;
|
|
|
|
|
|
|
|
AliasAnalysis::AliasResult AAResult = AA->alias(
|
2014-02-25 00:43:21 +00:00
|
|
|
AliasAnalysis::Location(MMOa->getValue(), Overlapa,
|
2014-04-14 00:51:57 +00:00
|
|
|
UseTBAA ? MMOa->getTBAAInfo() : nullptr),
|
2014-02-25 00:43:21 +00:00
|
|
|
AliasAnalysis::Location(MMOb->getValue(), Overlapb,
|
2014-04-14 00:51:57 +00:00
|
|
|
UseTBAA ? MMOb->getTBAAInfo() : nullptr));
|
2012-05-15 18:59:41 +00:00
|
|
|
|
|
|
|
return (AAResult != AliasAnalysis::NoAlias);
|
|
|
|
}
|
|
|
|
|
|
|
|
/// This recursive function iterates over chain deps of SUb looking for
|
|
|
|
/// "latest" node that needs a chain edge to SUa.
|
|
|
|
static unsigned
|
|
|
|
iterateChainSucc(AliasAnalysis *AA, const MachineFrameInfo *MFI,
|
|
|
|
SUnit *SUa, SUnit *SUb, SUnit *ExitSU, unsigned *Depth,
|
|
|
|
SmallPtrSet<const SUnit*, 16> &Visited) {
|
|
|
|
if (!SUa || !SUb || SUb == ExitSU)
|
|
|
|
return *Depth;
|
|
|
|
|
|
|
|
// Remember visited nodes.
|
|
|
|
if (!Visited.insert(SUb))
|
|
|
|
return *Depth;
|
|
|
|
// If there is _some_ dependency already in place, do not
|
|
|
|
// descend any further.
|
|
|
|
// TODO: Need to make sure that if that dependency got eliminated or ignored
|
|
|
|
// for any reason in the future, we would not violate DAG topology.
|
|
|
|
// Currently it does not happen, but makes an implicit assumption about
|
|
|
|
// future implementation.
|
|
|
|
//
|
|
|
|
// Independently, if we encounter node that is some sort of global
|
|
|
|
// object (like a call) we already have full set of dependencies to it
|
|
|
|
// and we can stop descending.
|
|
|
|
if (SUa->isSucc(SUb) ||
|
|
|
|
isGlobalMemoryObject(AA, SUb->getInstr()))
|
|
|
|
return *Depth;
|
|
|
|
|
|
|
|
// If we do need an edge, or we have exceeded depth budget,
|
|
|
|
// add that edge to the predecessors chain of SUb,
|
|
|
|
// and stop descending.
|
|
|
|
if (*Depth > 200 ||
|
|
|
|
MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr())) {
|
2012-11-06 03:13:46 +00:00
|
|
|
SUb->addPred(SDep(SUa, SDep::MayAliasMem));
|
2012-05-15 18:59:41 +00:00
|
|
|
return *Depth;
|
|
|
|
}
|
|
|
|
// Track current depth.
|
|
|
|
(*Depth)++;
|
|
|
|
// Iterate over chain dependencies only.
|
|
|
|
for (SUnit::const_succ_iterator I = SUb->Succs.begin(), E = SUb->Succs.end();
|
|
|
|
I != E; ++I)
|
|
|
|
if (I->isCtrl())
|
|
|
|
iterateChainSucc (AA, MFI, SUa, I->getSUnit(), ExitSU, Depth, Visited);
|
|
|
|
return *Depth;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// This function assumes that "downward" from SU there exist
|
|
|
|
/// tail/leaf of already constructed DAG. It iterates downward and
|
|
|
|
/// checks whether SU can be aliasing any node dominated
|
|
|
|
/// by it.
|
|
|
|
static void adjustChainDeps(AliasAnalysis *AA, const MachineFrameInfo *MFI,
|
2012-06-13 02:39:03 +00:00
|
|
|
SUnit *SU, SUnit *ExitSU, std::set<SUnit *> &CheckList,
|
|
|
|
unsigned LatencyToLoad) {
|
2012-05-15 18:59:41 +00:00
|
|
|
if (!SU)
|
|
|
|
return;
|
|
|
|
|
|
|
|
SmallPtrSet<const SUnit*, 16> Visited;
|
|
|
|
unsigned Depth = 0;
|
|
|
|
|
|
|
|
for (std::set<SUnit *>::iterator I = CheckList.begin(), IE = CheckList.end();
|
|
|
|
I != IE; ++I) {
|
|
|
|
if (SU == *I)
|
|
|
|
continue;
|
2012-06-13 02:39:03 +00:00
|
|
|
if (MIsNeedChainEdge(AA, MFI, SU->getInstr(), (*I)->getInstr())) {
|
2012-11-06 03:13:46 +00:00
|
|
|
SDep Dep(SU, SDep::MayAliasMem);
|
|
|
|
Dep.setLatency(((*I)->getInstr()->mayLoad()) ? LatencyToLoad : 0);
|
|
|
|
(*I)->addPred(Dep);
|
2012-06-13 02:39:03 +00:00
|
|
|
}
|
2012-05-15 18:59:41 +00:00
|
|
|
// Now go through all the chain successors and iterate from them.
|
|
|
|
// Keep track of visited nodes.
|
|
|
|
for (SUnit::const_succ_iterator J = (*I)->Succs.begin(),
|
|
|
|
JE = (*I)->Succs.end(); J != JE; ++J)
|
|
|
|
if (J->isCtrl())
|
|
|
|
iterateChainSucc (AA, MFI, SU, J->getSUnit(),
|
|
|
|
ExitSU, &Depth, Visited);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Check whether two objects need a chain edge, if so, add it
|
|
|
|
/// otherwise remember the rejected SU.
|
|
|
|
static inline
|
|
|
|
void addChainDependency (AliasAnalysis *AA, const MachineFrameInfo *MFI,
|
|
|
|
SUnit *SUa, SUnit *SUb,
|
|
|
|
std::set<SUnit *> &RejectList,
|
|
|
|
unsigned TrueMemOrderLatency = 0,
|
|
|
|
bool isNormalMemory = false) {
|
|
|
|
// If this is a false dependency,
|
|
|
|
// do not add the edge, but rememeber the rejected node.
|
2013-08-29 03:25:05 +00:00
|
|
|
if (!AA || MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr())) {
|
2012-11-06 03:13:46 +00:00
|
|
|
SDep Dep(SUa, isNormalMemory ? SDep::MayAliasMem : SDep::Barrier);
|
|
|
|
Dep.setLatency(TrueMemOrderLatency);
|
|
|
|
SUb->addPred(Dep);
|
|
|
|
}
|
2012-05-15 18:59:41 +00:00
|
|
|
else {
|
|
|
|
// Duplicate entries should be ignored.
|
|
|
|
RejectList.insert(SUb);
|
|
|
|
DEBUG(dbgs() << "\tReject chain dep between SU("
|
|
|
|
<< SUa->NodeNum << ") and SU("
|
|
|
|
<< SUb->NodeNum << ")\n");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-02-22 06:08:11 +00:00
|
|
|
/// Create an SUnit for each real instruction, numbered in top-down toplological
|
|
|
|
/// order. The instruction order A < B, implies that no edge exists from B to A.
|
|
|
|
///
|
|
|
|
/// Map each real instruction to its SUnit.
|
|
|
|
///
|
2012-03-14 04:00:41 +00:00
|
|
|
/// After initSUnits, the SUnits vector cannot be resized and the scheduler may
|
|
|
|
/// hang onto SUnit pointers. We may relax this in the future by using SUnit IDs
|
|
|
|
/// instead of pointers.
|
|
|
|
///
|
|
|
|
/// MachineScheduler relies on initSUnits numbering the nodes by their order in
|
|
|
|
/// the original instruction list.
|
2012-02-22 06:08:11 +00:00
|
|
|
void ScheduleDAGInstrs::initSUnits() {
|
|
|
|
// We'll be allocating one SUnit for each real instruction in the region,
|
|
|
|
// which is contained within a basic block.
|
2013-08-23 17:48:33 +00:00
|
|
|
SUnits.reserve(NumRegionInstrs);
|
2012-01-14 02:17:18 +00:00
|
|
|
|
2012-03-09 04:29:02 +00:00
|
|
|
for (MachineBasicBlock::iterator I = RegionBegin; I != RegionEnd; ++I) {
|
2012-02-22 06:08:11 +00:00
|
|
|
MachineInstr *MI = I;
|
|
|
|
if (MI->isDebugValue())
|
|
|
|
continue;
|
2012-01-14 02:17:18 +00:00
|
|
|
|
2012-03-07 23:00:49 +00:00
|
|
|
SUnit *SU = newSUnit(MI);
|
2012-02-22 06:08:11 +00:00
|
|
|
MISUnitMap[MI] = SU;
|
|
|
|
|
|
|
|
SU->isCall = MI->isCall();
|
|
|
|
SU->isCommutable = MI->isCommutable();
|
|
|
|
|
|
|
|
// Assign the Latency field of SU using target-provided information.
|
2012-10-10 05:43:09 +00:00
|
|
|
SU->Latency = SchedModel.computeInstrLatency(SU->getInstr());
|
2013-12-05 17:55:58 +00:00
|
|
|
|
2014-04-18 17:35:08 +00:00
|
|
|
// If this SUnit uses a reserved or unbuffered resource, mark it as such.
|
|
|
|
//
|
2014-05-15 01:52:21 +00:00
|
|
|
// Reserved resources block an instruction from issuing and stall the
|
2014-04-18 17:35:08 +00:00
|
|
|
// entire pipeline. These are identified by BufferSize=0.
|
|
|
|
//
|
2014-05-15 01:52:21 +00:00
|
|
|
// Unbuffered resources prevent execution of subsequent instructions that
|
2014-04-18 17:35:08 +00:00
|
|
|
// require the same resources. This is used for in-order execution pipelines
|
|
|
|
// within an out-of-order core. These are identified by BufferSize=1.
|
2013-12-05 17:55:58 +00:00
|
|
|
if (SchedModel.hasInstrSchedModel()) {
|
|
|
|
const MCSchedClassDesc *SC = getSchedClass(SU);
|
|
|
|
for (TargetSchedModel::ProcResIter
|
|
|
|
PI = SchedModel.getWriteProcResBegin(SC),
|
|
|
|
PE = SchedModel.getWriteProcResEnd(SC); PI != PE; ++PI) {
|
2013-12-05 17:56:02 +00:00
|
|
|
switch (SchedModel.getProcResource(PI->ProcResourceIdx)->BufferSize) {
|
|
|
|
case 0:
|
|
|
|
SU->hasReservedResource = true;
|
|
|
|
break;
|
|
|
|
case 1:
|
2013-12-05 17:55:58 +00:00
|
|
|
SU->isUnbuffered = true;
|
|
|
|
break;
|
2013-12-05 17:56:02 +00:00
|
|
|
default:
|
|
|
|
break;
|
2013-12-05 17:55:58 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2012-02-22 06:08:11 +00:00
|
|
|
}
|
2012-01-14 02:17:15 +00:00
|
|
|
}
|
|
|
|
|
2013-12-05 05:44:44 +00:00
|
|
|
/// If RegPressure is non-null, compute register pressure as a side effect. The
|
2012-04-24 17:56:43 +00:00
|
|
|
/// DAG builder is an efficient place to do it because it already visits
|
|
|
|
/// operands.
|
|
|
|
void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
|
2013-08-30 03:49:48 +00:00
|
|
|
RegPressureTracker *RPTracker,
|
|
|
|
PressureDiffs *PDiffs) {
|
2013-08-29 03:25:05 +00:00
|
|
|
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
|
|
|
|
bool UseAA = EnableAASchedMI.getNumOccurrences() > 0 ? EnableAASchedMI
|
|
|
|
: ST.useAA();
|
2014-04-14 00:51:57 +00:00
|
|
|
AliasAnalysis *AAForDep = UseAA ? AA : nullptr;
|
2013-08-29 03:25:05 +00:00
|
|
|
|
2013-09-04 21:00:02 +00:00
|
|
|
MISUnitMap.clear();
|
|
|
|
ScheduleDAG::clearDAG();
|
|
|
|
|
2012-02-22 06:08:11 +00:00
|
|
|
// Create an SUnit for each real instruction.
|
|
|
|
initSUnits();
|
2008-11-19 23:18:57 +00:00
|
|
|
|
2013-08-30 03:49:48 +00:00
|
|
|
if (PDiffs)
|
|
|
|
PDiffs->init(SUnits.size());
|
|
|
|
|
2008-12-04 01:35:46 +00:00
|
|
|
// We build scheduling units by walking a block's instruction list from bottom
|
|
|
|
// to top.
|
|
|
|
|
2009-11-09 19:22:17 +00:00
|
|
|
// Remember where a generic side-effecting instruction is as we procede.
|
2014-04-14 00:51:57 +00:00
|
|
|
SUnit *BarrierChain = nullptr, *AliasChain = nullptr;
|
2008-12-04 01:35:46 +00:00
|
|
|
|
2009-11-09 19:22:17 +00:00
|
|
|
// Memory references to specific known memory locations are tracked
|
|
|
|
// so that they can be given more precise dependencies. We track
|
|
|
|
// separately the known memory locations that may alias and those
|
|
|
|
// that are known not to alias
|
2014-04-15 07:22:52 +00:00
|
|
|
MapVector<ValueType, std::vector<SUnit *> > AliasMemDefs, NonAliasMemDefs;
|
|
|
|
MapVector<ValueType, std::vector<SUnit *> > AliasMemUses, NonAliasMemUses;
|
2012-05-15 18:59:41 +00:00
|
|
|
std::set<SUnit*> RejectMemNodes;
|
2008-12-04 01:35:46 +00:00
|
|
|
|
2010-03-10 22:13:47 +00:00
|
|
|
// Remove any stale debug info; sometimes BuildSchedGraph is called again
|
|
|
|
// without emitting the info from the previous call.
|
2011-06-02 20:07:12 +00:00
|
|
|
DbgValues.clear();
|
2014-04-14 00:51:57 +00:00
|
|
|
FirstDbgValue = nullptr;
|
2010-03-10 22:13:47 +00:00
|
|
|
|
2012-02-23 01:52:38 +00:00
|
|
|
assert(Defs.empty() && Uses.empty() &&
|
|
|
|
"Only BuildGraph should update Defs/Uses");
|
2013-01-21 18:18:53 +00:00
|
|
|
Defs.setUniverse(TRI->getNumRegs());
|
|
|
|
Uses.setUniverse(TRI->getNumRegs());
|
2011-05-06 21:52:52 +00:00
|
|
|
|
2012-02-22 21:59:00 +00:00
|
|
|
assert(VRegDefs.empty() && "Only BuildSchedGraph may access VRegDefs");
|
2013-08-23 17:48:39 +00:00
|
|
|
VRegUses.clear();
|
2012-02-22 21:59:00 +00:00
|
|
|
VRegDefs.setUniverse(MRI.getNumVirtRegs());
|
2013-08-23 17:48:39 +00:00
|
|
|
VRegUses.setUniverse(MRI.getNumVirtRegs());
|
2012-01-14 02:17:18 +00:00
|
|
|
|
2012-02-23 01:52:38 +00:00
|
|
|
// Model data dependencies between instructions being scheduled and the
|
|
|
|
// ExitSU.
|
2012-03-07 23:00:49 +00:00
|
|
|
addSchedBarrierDeps();
|
2012-02-23 01:52:38 +00:00
|
|
|
|
2009-02-10 23:27:53 +00:00
|
|
|
// Walk the list of instructions, from bottom moving up.
|
2014-04-14 00:51:57 +00:00
|
|
|
MachineInstr *DbgMI = nullptr;
|
2012-03-09 04:29:02 +00:00
|
|
|
for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
|
2008-11-19 23:18:57 +00:00
|
|
|
MII != MIE; --MII) {
|
2014-03-02 12:27:27 +00:00
|
|
|
MachineInstr *MI = std::prev(MII);
|
2012-12-01 01:22:49 +00:00
|
|
|
if (MI && DbgMI) {
|
|
|
|
DbgValues.push_back(std::make_pair(DbgMI, MI));
|
2014-04-14 00:51:57 +00:00
|
|
|
DbgMI = nullptr;
|
2011-06-02 20:07:12 +00:00
|
|
|
}
|
|
|
|
|
2010-03-10 22:13:47 +00:00
|
|
|
if (MI->isDebugValue()) {
|
2012-12-01 01:22:49 +00:00
|
|
|
DbgMI = MI;
|
2010-03-10 22:13:47 +00:00
|
|
|
continue;
|
|
|
|
}
|
2013-08-30 03:49:48 +00:00
|
|
|
SUnit *SU = MISUnitMap[MI];
|
|
|
|
assert(SU && "No SUnit mapped to this MI");
|
|
|
|
|
2012-04-24 17:56:43 +00:00
|
|
|
if (RPTracker) {
|
2014-04-14 00:51:57 +00:00
|
|
|
PressureDiff *PDiff = PDiffs ? &(*PDiffs)[SU->NodeNum] : nullptr;
|
|
|
|
RPTracker->recede(/*LiveUses=*/nullptr, PDiff);
|
2014-03-02 12:27:27 +00:00
|
|
|
assert(RPTracker->getPos() == std::prev(MII) &&
|
|
|
|
"RPTracker can't find MI");
|
2012-04-24 17:56:43 +00:00
|
|
|
}
|
2011-06-02 20:07:12 +00:00
|
|
|
|
2014-03-07 06:08:31 +00:00
|
|
|
assert(
|
|
|
|
(CanHandleTerminators || (!MI->isTerminator() && !MI->isPosition())) &&
|
|
|
|
"Cannot schedule terminators or labels!");
|
2008-11-19 23:18:57 +00:00
|
|
|
|
2008-12-04 01:35:46 +00:00
|
|
|
// Add register-based dependencies (data, anti, and output).
|
2012-12-18 20:53:01 +00:00
|
|
|
bool HasVRegDef = false;
|
2008-11-19 23:18:57 +00:00
|
|
|
for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
|
|
|
|
const MachineOperand &MO = MI->getOperand(j);
|
|
|
|
if (!MO.isReg()) continue;
|
|
|
|
unsigned Reg = MO.getReg();
|
|
|
|
if (Reg == 0) continue;
|
|
|
|
|
2012-01-14 02:17:15 +00:00
|
|
|
if (TRI->isPhysicalRegister(Reg))
|
|
|
|
addPhysRegDeps(SU, j);
|
|
|
|
else {
|
|
|
|
assert(!IsPostRA && "Virtual register encountered!");
|
2012-12-18 20:53:01 +00:00
|
|
|
if (MO.isDef()) {
|
|
|
|
HasVRegDef = true;
|
2012-01-14 02:17:18 +00:00
|
|
|
addVRegDefDeps(SU, j);
|
2012-12-18 20:53:01 +00:00
|
|
|
}
|
2012-02-23 03:16:24 +00:00
|
|
|
else if (MO.readsReg()) // ignore undef operands
|
2012-01-14 02:17:18 +00:00
|
|
|
addVRegUseDeps(SU, j);
|
2008-11-19 23:18:57 +00:00
|
|
|
}
|
|
|
|
}
|
2012-12-18 20:53:01 +00:00
|
|
|
// If we haven't seen any uses in this scheduling region, create a
|
|
|
|
// dependence edge to ExitSU to model the live-out latency. This is required
|
|
|
|
// for vreg defs with no in-region use, and prefetches with no vreg def.
|
|
|
|
//
|
|
|
|
// FIXME: NumDataSuccs would be more precise than NumSuccs here. This
|
|
|
|
// check currently relies on being called before adding chain deps.
|
|
|
|
if (SU->NumSuccs == 0 && SU->Latency > 1
|
|
|
|
&& (HasVRegDef || MI->mayLoad())) {
|
|
|
|
SDep Dep(SU, SDep::Artificial);
|
|
|
|
Dep.setLatency(SU->Latency - 1);
|
|
|
|
ExitSU.addPred(Dep);
|
|
|
|
}
|
2008-12-04 01:35:46 +00:00
|
|
|
|
|
|
|
// Add chain dependencies.
|
2009-11-02 17:06:28 +00:00
|
|
|
// Chain dependencies used to enforce memory order should have
|
|
|
|
// latency of 0 (except for true dependency of Store followed by
|
|
|
|
// aliased Load... we estimate that with a single cycle of latency
|
|
|
|
// assuming the hardware will bypass)
|
2008-12-04 01:35:46 +00:00
|
|
|
// Note that isStoreToStackSlot and isLoadFromStackSLot are not usable
|
|
|
|
// after stack slots are lowered to actual addresses.
|
|
|
|
// TODO: Use an AliasAnalysis and do real alias-analysis queries, and
|
|
|
|
// produce more precise dependence information.
|
2012-06-13 02:39:03 +00:00
|
|
|
unsigned TrueMemOrderLatency = MI->mayStore() ? 1 : 0;
|
2012-05-15 18:59:41 +00:00
|
|
|
if (isGlobalMemoryObject(AA, MI)) {
|
2009-11-09 19:22:17 +00:00
|
|
|
// Be conservative with these and add dependencies on all memory
|
|
|
|
// references, even those that are known to not alias.
|
2014-04-15 07:22:52 +00:00
|
|
|
for (MapVector<ValueType, std::vector<SUnit *> >::iterator I =
|
2009-11-09 19:22:17 +00:00
|
|
|
NonAliasMemDefs.begin(), E = NonAliasMemDefs.end(); I != E; ++I) {
|
Track multiple stores per object when using AA in ScheduleDAGInstrs
When using AA to break false chain dependencies, we need to track multiple
stores per object in ScheduleDAGInstrs. Historically, we tracked potential alias
chains at the object level, and so all loads of an object would retain
dependencies on any store to that object. With AA, however, this is not
sufficient: non-overlapping stores and loads to the same object all need to be
tested for dependencies separately, we cannot only test all loads to an object
against only the last store (see PR18497 for an explicit example).
To mitigate any unwelcome compile-time impact when not using AA, only one store
is kept in the list per object when not using AA.
This, along with a stack coloring change to come shortly, will provide a test
case, fix PR18497 (and allow LLVM to compile itself using -enable-aa-sched-mi
on x86-64).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@199657 91177308-0d34-0410-b5e6-96231b3b80d8
2014-01-20 14:03:02 +00:00
|
|
|
for (unsigned i = 0, e = I->second.size(); i != e; ++i) {
|
|
|
|
I->second[i]->addPred(SDep(SU, SDep::Barrier));
|
|
|
|
}
|
2009-11-09 19:22:17 +00:00
|
|
|
}
|
2014-04-15 07:22:52 +00:00
|
|
|
for (MapVector<ValueType, std::vector<SUnit *> >::iterator I =
|
2009-11-09 19:22:17 +00:00
|
|
|
NonAliasMemUses.begin(), E = NonAliasMemUses.end(); I != E; ++I) {
|
2012-11-06 03:13:46 +00:00
|
|
|
for (unsigned i = 0, e = I->second.size(); i != e; ++i) {
|
|
|
|
SDep Dep(SU, SDep::Barrier);
|
|
|
|
Dep.setLatency(TrueMemOrderLatency);
|
|
|
|
I->second[i]->addPred(Dep);
|
|
|
|
}
|
2009-11-09 19:22:17 +00:00
|
|
|
}
|
|
|
|
// Add SU to the barrier chain.
|
|
|
|
if (BarrierChain)
|
2012-11-06 03:13:46 +00:00
|
|
|
BarrierChain->addPred(SDep(SU, SDep::Barrier));
|
2009-11-09 19:22:17 +00:00
|
|
|
BarrierChain = SU;
|
2012-05-15 18:59:41 +00:00
|
|
|
// This is a barrier event that acts as a pivotal node in the DAG,
|
|
|
|
// so it is safe to clear list of exposed nodes.
|
2012-06-13 02:39:03 +00:00
|
|
|
adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
|
|
|
|
TrueMemOrderLatency);
|
2012-05-15 18:59:41 +00:00
|
|
|
RejectMemNodes.clear();
|
|
|
|
NonAliasMemDefs.clear();
|
|
|
|
NonAliasMemUses.clear();
|
2009-11-09 19:22:17 +00:00
|
|
|
|
|
|
|
// fall-through
|
|
|
|
new_alias_chain:
|
|
|
|
// Chain all possibly aliasing memory references though SU.
|
2012-06-13 02:39:03 +00:00
|
|
|
if (AliasChain) {
|
|
|
|
unsigned ChainLatency = 0;
|
|
|
|
if (AliasChain->getInstr()->mayLoad())
|
|
|
|
ChainLatency = TrueMemOrderLatency;
|
2013-08-29 03:25:05 +00:00
|
|
|
addChainDependency(AAForDep, MFI, SU, AliasChain, RejectMemNodes,
|
2012-06-13 02:39:03 +00:00
|
|
|
ChainLatency);
|
|
|
|
}
|
2009-11-09 19:22:17 +00:00
|
|
|
AliasChain = SU;
|
2008-11-19 23:18:57 +00:00
|
|
|
for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
|
2013-08-29 03:25:05 +00:00
|
|
|
addChainDependency(AAForDep, MFI, SU, PendingLoads[k], RejectMemNodes,
|
2012-05-15 18:59:41 +00:00
|
|
|
TrueMemOrderLatency);
|
2014-04-15 07:22:52 +00:00
|
|
|
for (MapVector<ValueType, std::vector<SUnit *> >::iterator I =
|
Track multiple stores per object when using AA in ScheduleDAGInstrs
When using AA to break false chain dependencies, we need to track multiple
stores per object in ScheduleDAGInstrs. Historically, we tracked potential alias
chains at the object level, and so all loads of an object would retain
dependencies on any store to that object. With AA, however, this is not
sufficient: non-overlapping stores and loads to the same object all need to be
tested for dependencies separately, we cannot only test all loads to an object
against only the last store (see PR18497 for an explicit example).
To mitigate any unwelcome compile-time impact when not using AA, only one store
is kept in the list per object when not using AA.
This, along with a stack coloring change to come shortly, will provide a test
case, fix PR18497 (and allow LLVM to compile itself using -enable-aa-sched-mi
on x86-64).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@199657 91177308-0d34-0410-b5e6-96231b3b80d8
2014-01-20 14:03:02 +00:00
|
|
|
AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I) {
|
|
|
|
for (unsigned i = 0, e = I->second.size(); i != e; ++i)
|
|
|
|
addChainDependency(AAForDep, MFI, SU, I->second[i], RejectMemNodes);
|
|
|
|
}
|
2014-04-15 07:22:52 +00:00
|
|
|
for (MapVector<ValueType, std::vector<SUnit *> >::iterator I =
|
2009-11-09 19:22:17 +00:00
|
|
|
AliasMemUses.begin(), E = AliasMemUses.end(); I != E; ++I) {
|
2008-12-04 01:35:46 +00:00
|
|
|
for (unsigned i = 0, e = I->second.size(); i != e; ++i)
|
2013-08-29 03:25:05 +00:00
|
|
|
addChainDependency(AAForDep, MFI, SU, I->second[i], RejectMemNodes,
|
2012-05-15 18:59:41 +00:00
|
|
|
TrueMemOrderLatency);
|
2008-12-04 01:35:46 +00:00
|
|
|
}
|
2012-06-13 02:39:03 +00:00
|
|
|
adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
|
|
|
|
TrueMemOrderLatency);
|
2009-11-09 19:22:17 +00:00
|
|
|
PendingLoads.clear();
|
|
|
|
AliasMemDefs.clear();
|
|
|
|
AliasMemUses.clear();
|
2011-12-07 07:15:52 +00:00
|
|
|
} else if (MI->mayStore()) {
|
2013-06-29 18:41:17 +00:00
|
|
|
UnderlyingObjectsVector Objs;
|
2012-12-10 18:49:16 +00:00
|
|
|
getUnderlyingObjectsForInstr(MI, MFI, Objs);
|
|
|
|
|
|
|
|
if (Objs.empty()) {
|
|
|
|
// Treat all other stores conservatively.
|
|
|
|
goto new_alias_chain;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool MayAlias = false;
|
2013-06-29 18:41:17 +00:00
|
|
|
for (UnderlyingObjectsVector::iterator K = Objs.begin(), KE = Objs.end();
|
|
|
|
K != KE; ++K) {
|
2014-04-15 07:22:52 +00:00
|
|
|
ValueType V = K->getPointer();
|
2013-06-29 18:41:17 +00:00
|
|
|
bool ThisMayAlias = K->getInt();
|
2012-12-10 18:49:16 +00:00
|
|
|
if (ThisMayAlias)
|
|
|
|
MayAlias = true;
|
|
|
|
|
2008-12-04 01:35:46 +00:00
|
|
|
// A store to a specific PseudoSourceValue. Add precise dependencies.
|
2009-11-09 19:22:17 +00:00
|
|
|
// Record the def in MemDefs, first adding a dep if there is
|
|
|
|
// an existing def.
|
2014-04-15 07:22:52 +00:00
|
|
|
MapVector<ValueType, std::vector<SUnit *> >::iterator I =
|
2012-12-10 18:49:16 +00:00
|
|
|
((ThisMayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
|
2014-04-15 07:22:52 +00:00
|
|
|
MapVector<ValueType, std::vector<SUnit *> >::iterator IE =
|
2012-12-10 18:49:16 +00:00
|
|
|
((ThisMayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
|
2009-11-09 19:22:17 +00:00
|
|
|
if (I != IE) {
|
Track multiple stores per object when using AA in ScheduleDAGInstrs
When using AA to break false chain dependencies, we need to track multiple
stores per object in ScheduleDAGInstrs. Historically, we tracked potential alias
chains at the object level, and so all loads of an object would retain
dependencies on any store to that object. With AA, however, this is not
sufficient: non-overlapping stores and loads to the same object all need to be
tested for dependencies separately, we cannot only test all loads to an object
against only the last store (see PR18497 for an explicit example).
To mitigate any unwelcome compile-time impact when not using AA, only one store
is kept in the list per object when not using AA.
This, along with a stack coloring change to come shortly, will provide a test
case, fix PR18497 (and allow LLVM to compile itself using -enable-aa-sched-mi
on x86-64).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@199657 91177308-0d34-0410-b5e6-96231b3b80d8
2014-01-20 14:03:02 +00:00
|
|
|
for (unsigned i = 0, e = I->second.size(); i != e; ++i)
|
|
|
|
addChainDependency(AAForDep, MFI, SU, I->second[i], RejectMemNodes,
|
|
|
|
0, true);
|
|
|
|
|
|
|
|
// If we're not using AA, then we only need one store per object.
|
|
|
|
if (!AAForDep)
|
|
|
|
I->second.clear();
|
|
|
|
I->second.push_back(SU);
|
2008-12-04 01:35:46 +00:00
|
|
|
} else {
|
Track multiple stores per object when using AA in ScheduleDAGInstrs
When using AA to break false chain dependencies, we need to track multiple
stores per object in ScheduleDAGInstrs. Historically, we tracked potential alias
chains at the object level, and so all loads of an object would retain
dependencies on any store to that object. With AA, however, this is not
sufficient: non-overlapping stores and loads to the same object all need to be
tested for dependencies separately, we cannot only test all loads to an object
against only the last store (see PR18497 for an explicit example).
To mitigate any unwelcome compile-time impact when not using AA, only one store
is kept in the list per object when not using AA.
This, along with a stack coloring change to come shortly, will provide a test
case, fix PR18497 (and allow LLVM to compile itself using -enable-aa-sched-mi
on x86-64).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@199657 91177308-0d34-0410-b5e6-96231b3b80d8
2014-01-20 14:03:02 +00:00
|
|
|
if (ThisMayAlias) {
|
|
|
|
if (!AAForDep)
|
|
|
|
AliasMemDefs[V].clear();
|
|
|
|
AliasMemDefs[V].push_back(SU);
|
|
|
|
} else {
|
|
|
|
if (!AAForDep)
|
|
|
|
NonAliasMemDefs[V].clear();
|
|
|
|
NonAliasMemDefs[V].push_back(SU);
|
|
|
|
}
|
2008-12-04 01:35:46 +00:00
|
|
|
}
|
|
|
|
// Handle the uses in MemUses, if there are any.
|
2014-04-15 07:22:52 +00:00
|
|
|
MapVector<ValueType, std::vector<SUnit *> >::iterator J =
|
2012-12-10 18:49:16 +00:00
|
|
|
((ThisMayAlias) ? AliasMemUses.find(V) : NonAliasMemUses.find(V));
|
2014-04-15 07:22:52 +00:00
|
|
|
MapVector<ValueType, std::vector<SUnit *> >::iterator JE =
|
2012-12-10 18:49:16 +00:00
|
|
|
((ThisMayAlias) ? AliasMemUses.end() : NonAliasMemUses.end());
|
2009-11-09 19:22:17 +00:00
|
|
|
if (J != JE) {
|
2008-12-04 01:35:46 +00:00
|
|
|
for (unsigned i = 0, e = J->second.size(); i != e; ++i)
|
2013-08-29 03:25:05 +00:00
|
|
|
addChainDependency(AAForDep, MFI, SU, J->second[i], RejectMemNodes,
|
2012-05-15 18:59:41 +00:00
|
|
|
TrueMemOrderLatency, true);
|
2008-12-04 01:35:46 +00:00
|
|
|
J->second.clear();
|
|
|
|
}
|
2009-11-02 17:06:28 +00:00
|
|
|
}
|
2012-12-10 18:49:16 +00:00
|
|
|
if (MayAlias) {
|
|
|
|
// Add dependencies from all the PendingLoads, i.e. loads
|
|
|
|
// with no underlying object.
|
|
|
|
for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
|
2013-08-29 03:25:05 +00:00
|
|
|
addChainDependency(AAForDep, MFI, SU, PendingLoads[k], RejectMemNodes,
|
2012-12-10 18:49:16 +00:00
|
|
|
TrueMemOrderLatency);
|
|
|
|
// Add dependence on alias chain, if needed.
|
|
|
|
if (AliasChain)
|
2013-08-29 03:25:05 +00:00
|
|
|
addChainDependency(AAForDep, MFI, SU, AliasChain, RejectMemNodes);
|
2012-12-10 18:49:16 +00:00
|
|
|
// But we also should check dependent instructions for the
|
|
|
|
// SU in question.
|
|
|
|
adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
|
|
|
|
TrueMemOrderLatency);
|
|
|
|
}
|
|
|
|
// Add dependence on barrier chain, if needed.
|
|
|
|
// There is no point to check aliasing on barrier event. Even if
|
|
|
|
// SU and barrier _could_ be reordered, they should not. In addition,
|
|
|
|
// we have lost all RejectMemNodes below barrier.
|
|
|
|
if (BarrierChain)
|
|
|
|
BarrierChain->addPred(SDep(SU, SDep::Barrier));
|
2011-12-07 07:15:52 +00:00
|
|
|
} else if (MI->mayLoad()) {
|
2009-11-03 20:15:00 +00:00
|
|
|
bool MayAlias = true;
|
2009-10-09 23:27:56 +00:00
|
|
|
if (MI->isInvariantLoad(AA)) {
|
2008-12-04 01:35:46 +00:00
|
|
|
// Invariant load, no chain dependencies needed!
|
2009-11-05 00:16:44 +00:00
|
|
|
} else {
|
2013-06-29 18:41:17 +00:00
|
|
|
UnderlyingObjectsVector Objs;
|
2012-12-10 18:49:16 +00:00
|
|
|
getUnderlyingObjectsForInstr(MI, MFI, Objs);
|
|
|
|
|
|
|
|
if (Objs.empty()) {
|
2009-11-09 19:22:17 +00:00
|
|
|
// A load with no underlying object. Depend on all
|
|
|
|
// potentially aliasing stores.
|
2014-04-15 07:22:52 +00:00
|
|
|
for (MapVector<ValueType, std::vector<SUnit *> >::iterator I =
|
2009-11-09 19:22:17 +00:00
|
|
|
AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I)
|
Track multiple stores per object when using AA in ScheduleDAGInstrs
When using AA to break false chain dependencies, we need to track multiple
stores per object in ScheduleDAGInstrs. Historically, we tracked potential alias
chains at the object level, and so all loads of an object would retain
dependencies on any store to that object. With AA, however, this is not
sufficient: non-overlapping stores and loads to the same object all need to be
tested for dependencies separately, we cannot only test all loads to an object
against only the last store (see PR18497 for an explicit example).
To mitigate any unwelcome compile-time impact when not using AA, only one store
is kept in the list per object when not using AA.
This, along with a stack coloring change to come shortly, will provide a test
case, fix PR18497 (and allow LLVM to compile itself using -enable-aa-sched-mi
on x86-64).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@199657 91177308-0d34-0410-b5e6-96231b3b80d8
2014-01-20 14:03:02 +00:00
|
|
|
for (unsigned i = 0, e = I->second.size(); i != e; ++i)
|
|
|
|
addChainDependency(AAForDep, MFI, SU, I->second[i],
|
|
|
|
RejectMemNodes);
|
2011-05-05 19:24:06 +00:00
|
|
|
|
2009-11-09 19:22:17 +00:00
|
|
|
PendingLoads.push_back(SU);
|
|
|
|
MayAlias = true;
|
2012-12-10 18:49:16 +00:00
|
|
|
} else {
|
|
|
|
MayAlias = false;
|
|
|
|
}
|
|
|
|
|
2013-06-29 18:41:17 +00:00
|
|
|
for (UnderlyingObjectsVector::iterator
|
2012-12-10 18:49:16 +00:00
|
|
|
J = Objs.begin(), JE = Objs.end(); J != JE; ++J) {
|
2014-04-15 07:22:52 +00:00
|
|
|
ValueType V = J->getPointer();
|
2013-06-29 18:41:17 +00:00
|
|
|
bool ThisMayAlias = J->getInt();
|
2012-12-10 18:49:16 +00:00
|
|
|
|
|
|
|
if (ThisMayAlias)
|
|
|
|
MayAlias = true;
|
|
|
|
|
|
|
|
// A load from a specific PseudoSourceValue. Add precise dependencies.
|
2014-04-15 07:22:52 +00:00
|
|
|
MapVector<ValueType, std::vector<SUnit *> >::iterator I =
|
2012-12-10 18:49:16 +00:00
|
|
|
((ThisMayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
|
2014-04-15 07:22:52 +00:00
|
|
|
MapVector<ValueType, std::vector<SUnit *> >::iterator IE =
|
2012-12-10 18:49:16 +00:00
|
|
|
((ThisMayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
|
|
|
|
if (I != IE)
|
Track multiple stores per object when using AA in ScheduleDAGInstrs
When using AA to break false chain dependencies, we need to track multiple
stores per object in ScheduleDAGInstrs. Historically, we tracked potential alias
chains at the object level, and so all loads of an object would retain
dependencies on any store to that object. With AA, however, this is not
sufficient: non-overlapping stores and loads to the same object all need to be
tested for dependencies separately, we cannot only test all loads to an object
against only the last store (see PR18497 for an explicit example).
To mitigate any unwelcome compile-time impact when not using AA, only one store
is kept in the list per object when not using AA.
This, along with a stack coloring change to come shortly, will provide a test
case, fix PR18497 (and allow LLVM to compile itself using -enable-aa-sched-mi
on x86-64).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@199657 91177308-0d34-0410-b5e6-96231b3b80d8
2014-01-20 14:03:02 +00:00
|
|
|
for (unsigned i = 0, e = I->second.size(); i != e; ++i)
|
|
|
|
addChainDependency(AAForDep, MFI, SU, I->second[i],
|
|
|
|
RejectMemNodes, 0, true);
|
2012-12-10 18:49:16 +00:00
|
|
|
if (ThisMayAlias)
|
|
|
|
AliasMemUses[V].push_back(SU);
|
|
|
|
else
|
|
|
|
NonAliasMemUses[V].push_back(SU);
|
2009-11-03 20:15:00 +00:00
|
|
|
}
|
2012-05-15 18:59:41 +00:00
|
|
|
if (MayAlias)
|
2012-06-13 02:39:03 +00:00
|
|
|
adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes, /*Latency=*/0);
|
2009-11-09 19:22:17 +00:00
|
|
|
// Add dependencies on alias and barrier chains, if needed.
|
|
|
|
if (MayAlias && AliasChain)
|
2013-08-29 03:25:05 +00:00
|
|
|
addChainDependency(AAForDep, MFI, SU, AliasChain, RejectMemNodes);
|
2009-11-09 19:22:17 +00:00
|
|
|
if (BarrierChain)
|
2012-11-06 03:13:46 +00:00
|
|
|
BarrierChain->addPred(SDep(SU, SDep::Barrier));
|
2011-05-05 19:24:06 +00:00
|
|
|
}
|
2008-11-19 23:18:57 +00:00
|
|
|
}
|
|
|
|
}
|
2012-12-01 01:22:49 +00:00
|
|
|
if (DbgMI)
|
|
|
|
FirstDbgValue = DbgMI;
|
2009-01-15 19:20:50 +00:00
|
|
|
|
2012-02-23 01:52:38 +00:00
|
|
|
Defs.clear();
|
|
|
|
Uses.clear();
|
2012-01-14 02:17:18 +00:00
|
|
|
VRegDefs.clear();
|
2009-01-15 19:20:50 +00:00
|
|
|
PendingLoads.clear();
|
2008-11-19 23:18:57 +00:00
|
|
|
}
|
|
|
|
|
2013-12-28 21:56:55 +00:00
|
|
|
/// \brief Initialize register live-range state for updating kills.
|
|
|
|
void ScheduleDAGInstrs::startBlockForKills(MachineBasicBlock *BB) {
|
|
|
|
// Start with no live registers.
|
|
|
|
LiveRegs.reset();
|
|
|
|
|
|
|
|
// 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;
|
|
|
|
// Repeat, for reg and all subregs.
|
|
|
|
for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
|
|
|
|
SubRegs.isValid(); ++SubRegs)
|
|
|
|
LiveRegs.set(*SubRegs);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
bool ScheduleDAGInstrs::toggleKillFlag(MachineInstr *MI, MachineOperand &MO) {
|
|
|
|
// Setting kill flag...
|
|
|
|
if (!MO.isKill()) {
|
|
|
|
MO.setIsKill(true);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// If MO itself is live, clear the kill flag...
|
|
|
|
if (LiveRegs.test(MO.getReg())) {
|
|
|
|
MO.setIsKill(false);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// If any subreg of MO is live, then create an imp-def for that
|
|
|
|
// subreg and keep MO marked as killed.
|
|
|
|
MO.setIsKill(false);
|
|
|
|
bool AllDead = true;
|
|
|
|
const unsigned SuperReg = MO.getReg();
|
|
|
|
MachineInstrBuilder MIB(MF, MI);
|
|
|
|
for (MCSubRegIterator SubRegs(SuperReg, TRI); SubRegs.isValid(); ++SubRegs) {
|
|
|
|
if (LiveRegs.test(*SubRegs)) {
|
|
|
|
MIB.addReg(*SubRegs, RegState::ImplicitDefine);
|
|
|
|
AllDead = false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if(AllDead)
|
|
|
|
MO.setIsKill(true);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// FIXME: Reuse the LivePhysRegs utility for this.
|
|
|
|
void ScheduleDAGInstrs::fixupKills(MachineBasicBlock *MBB) {
|
|
|
|
DEBUG(dbgs() << "Fixup kills for BB#" << MBB->getNumber() << '\n');
|
|
|
|
|
|
|
|
LiveRegs.resize(TRI->getNumRegs());
|
|
|
|
BitVector killedRegs(TRI->getNumRegs());
|
|
|
|
|
|
|
|
startBlockForKills(MBB);
|
|
|
|
|
|
|
|
// Examine block from end to start...
|
|
|
|
unsigned Count = MBB->size();
|
|
|
|
for (MachineBasicBlock::iterator I = MBB->end(), E = MBB->begin();
|
|
|
|
I != E; --Count) {
|
|
|
|
MachineInstr *MI = --I;
|
|
|
|
if (MI->isDebugValue())
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Update liveness. Registers that are defed but not used in this
|
|
|
|
// instruction are now dead. Mark register and all subregs as they
|
|
|
|
// are completely defined.
|
|
|
|
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
|
|
|
|
MachineOperand &MO = MI->getOperand(i);
|
|
|
|
if (MO.isRegMask())
|
|
|
|
LiveRegs.clearBitsNotInMask(MO.getRegMask());
|
|
|
|
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;
|
|
|
|
|
|
|
|
// Repeat for reg and all subregs.
|
|
|
|
for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
|
|
|
|
SubRegs.isValid(); ++SubRegs)
|
|
|
|
LiveRegs.reset(*SubRegs);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Examine all used registers and set/clear kill flag. When a
|
|
|
|
// register is used multiple times we only set the kill flag on
|
|
|
|
// the first use. Don't set kill flags on undef operands.
|
|
|
|
killedRegs.reset();
|
|
|
|
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
|
|
|
|
MachineOperand &MO = MI->getOperand(i);
|
|
|
|
if (!MO.isReg() || !MO.isUse() || MO.isUndef()) continue;
|
|
|
|
unsigned Reg = MO.getReg();
|
|
|
|
if ((Reg == 0) || MRI.isReserved(Reg)) continue;
|
|
|
|
|
|
|
|
bool kill = false;
|
|
|
|
if (!killedRegs.test(Reg)) {
|
|
|
|
kill = true;
|
|
|
|
// A register is not killed if any subregs are live...
|
|
|
|
for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) {
|
|
|
|
if (LiveRegs.test(*SubRegs)) {
|
|
|
|
kill = false;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// If subreg is not live, then register is killed if it became
|
|
|
|
// live in this instruction
|
|
|
|
if (kill)
|
|
|
|
kill = !LiveRegs.test(Reg);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (MO.isKill() != kill) {
|
|
|
|
DEBUG(dbgs() << "Fixing " << MO << " in ");
|
|
|
|
// Warning: toggleKillFlag may invalidate MO.
|
|
|
|
toggleKillFlag(MI, MO);
|
|
|
|
DEBUG(MI->dump());
|
|
|
|
}
|
|
|
|
|
|
|
|
killedRegs.set(Reg);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Mark any used register (that is not using undef) and subregs as
|
|
|
|
// now live...
|
|
|
|
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
|
|
|
|
MachineOperand &MO = MI->getOperand(i);
|
|
|
|
if (!MO.isReg() || !MO.isUse() || MO.isUndef()) continue;
|
|
|
|
unsigned Reg = MO.getReg();
|
|
|
|
if ((Reg == 0) || MRI.isReserved(Reg)) continue;
|
|
|
|
|
|
|
|
for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
|
|
|
|
SubRegs.isValid(); ++SubRegs)
|
|
|
|
LiveRegs.set(*SubRegs);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-11-19 23:18:57 +00:00
|
|
|
void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
|
2012-09-11 22:23:19 +00:00
|
|
|
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
|
2008-11-19 23:18:57 +00:00
|
|
|
SU->getInstr()->dump();
|
2012-09-06 19:06:06 +00:00
|
|
|
#endif
|
2008-11-19 23:18:57 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
|
|
|
|
std::string s;
|
|
|
|
raw_string_ostream oss(s);
|
2009-02-10 23:27:53 +00:00
|
|
|
if (SU == &EntrySU)
|
|
|
|
oss << "<entry>";
|
|
|
|
else if (SU == &ExitSU)
|
|
|
|
oss << "<exit>";
|
|
|
|
else
|
2013-01-25 07:45:25 +00:00
|
|
|
SU->getInstr()->print(oss, &TM, /*SkipOpers=*/true);
|
2008-11-19 23:18:57 +00:00
|
|
|
return oss.str();
|
|
|
|
}
|
|
|
|
|
2012-03-07 00:18:22 +00:00
|
|
|
/// Return the basic block label. It is not necessarilly unique because a block
|
|
|
|
/// contains multiple scheduling regions. But it is fine for visualization.
|
|
|
|
std::string ScheduleDAGInstrs::getDAGName() const {
|
|
|
|
return "dag." + BB->getFullName();
|
|
|
|
}
|
2012-10-15 18:02:27 +00:00
|
|
|
|
2012-11-28 05:13:28 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// SchedDFSResult Implementation
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace llvm {
|
|
|
|
/// \brief Internal state used to compute SchedDFSResult.
|
|
|
|
class SchedDFSImpl {
|
|
|
|
SchedDFSResult &R;
|
|
|
|
|
|
|
|
/// Join DAG nodes into equivalence classes by their subtree.
|
|
|
|
IntEqClasses SubtreeClasses;
|
|
|
|
/// List PredSU, SuccSU pairs that represent data edges between subtrees.
|
|
|
|
std::vector<std::pair<const SUnit*, const SUnit*> > ConnectionPairs;
|
|
|
|
|
2013-01-25 06:52:27 +00:00
|
|
|
struct RootData {
|
|
|
|
unsigned NodeID;
|
|
|
|
unsigned ParentNodeID; // Parent node (member of the parent subtree).
|
|
|
|
unsigned SubInstrCount; // Instr count in this tree only, not children.
|
|
|
|
|
|
|
|
RootData(unsigned id): NodeID(id),
|
|
|
|
ParentNodeID(SchedDFSResult::InvalidSubtreeID),
|
|
|
|
SubInstrCount(0) {}
|
|
|
|
|
|
|
|
unsigned getSparseSetIndex() const { return NodeID; }
|
|
|
|
};
|
|
|
|
|
|
|
|
SparseSet<RootData> RootSet;
|
|
|
|
|
2012-11-28 05:13:28 +00:00
|
|
|
public:
|
2013-01-25 06:52:27 +00:00
|
|
|
SchedDFSImpl(SchedDFSResult &r): R(r), SubtreeClasses(R.DFSNodeData.size()) {
|
|
|
|
RootSet.setUniverse(R.DFSNodeData.size());
|
|
|
|
}
|
2012-11-28 05:13:28 +00:00
|
|
|
|
2013-01-25 06:02:44 +00:00
|
|
|
/// Return true if this node been visited by the DFS traversal.
|
|
|
|
///
|
|
|
|
/// During visitPostorderNode the Node's SubtreeID is assigned to the Node
|
|
|
|
/// ID. Later, SubtreeID is updated but remains valid.
|
2012-11-28 05:13:28 +00:00
|
|
|
bool isVisited(const SUnit *SU) const {
|
2013-01-25 06:52:27 +00:00
|
|
|
return R.DFSNodeData[SU->NodeNum].SubtreeID
|
|
|
|
!= SchedDFSResult::InvalidSubtreeID;
|
2012-11-28 05:13:28 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Initialize this node's instruction count. We don't need to flag the node
|
|
|
|
/// visited until visitPostorder because the DAG cannot have cycles.
|
|
|
|
void visitPreorder(const SUnit *SU) {
|
2013-01-25 06:52:27 +00:00
|
|
|
R.DFSNodeData[SU->NodeNum].InstrCount =
|
|
|
|
SU->getInstr()->isTransient() ? 0 : 1;
|
2013-01-25 06:02:44 +00:00
|
|
|
}
|
2012-11-28 05:13:28 +00:00
|
|
|
|
2013-01-25 06:02:44 +00:00
|
|
|
/// Called once for each node after all predecessors are visited. Revisit this
|
|
|
|
/// node's predecessors and potentially join them now that we know the ILP of
|
|
|
|
/// the other predecessors.
|
|
|
|
void visitPostorderNode(const SUnit *SU) {
|
|
|
|
// Mark this node as the root of a subtree. It may be joined with its
|
|
|
|
// successors later.
|
2013-01-25 06:52:27 +00:00
|
|
|
R.DFSNodeData[SU->NodeNum].SubtreeID = SU->NodeNum;
|
|
|
|
RootData RData(SU->NodeNum);
|
|
|
|
RData.SubInstrCount = SU->getInstr()->isTransient() ? 0 : 1;
|
2012-11-28 05:13:28 +00:00
|
|
|
|
2013-01-25 06:02:44 +00:00
|
|
|
// If any predecessors are still in their own subtree, they either cannot be
|
|
|
|
// joined or are large enough to remain separate. If this parent node's
|
|
|
|
// total instruction count is not greater than a child subtree by at least
|
|
|
|
// the subtree limit, then try to join it now since splitting subtrees is
|
|
|
|
// only useful if multiple high-pressure paths are possible.
|
2013-01-25 06:52:27 +00:00
|
|
|
unsigned InstrCount = R.DFSNodeData[SU->NodeNum].InstrCount;
|
2013-01-25 06:02:44 +00:00
|
|
|
for (SUnit::const_pred_iterator
|
|
|
|
PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
|
|
|
|
if (PI->getKind() != SDep::Data)
|
|
|
|
continue;
|
|
|
|
unsigned PredNum = PI->getSUnit()->NodeNum;
|
2013-01-25 06:52:27 +00:00
|
|
|
if ((InstrCount - R.DFSNodeData[PredNum].InstrCount) < R.SubtreeLimit)
|
2013-01-25 06:02:44 +00:00
|
|
|
joinPredSubtree(*PI, SU, /*CheckLimit=*/false);
|
2013-01-25 06:52:27 +00:00
|
|
|
|
|
|
|
// Either link or merge the TreeData entry from the child to the parent.
|
2013-01-25 06:52:30 +00:00
|
|
|
if (R.DFSNodeData[PredNum].SubtreeID == PredNum) {
|
|
|
|
// If the predecessor's parent is invalid, this is a tree edge and the
|
|
|
|
// current node is the parent.
|
|
|
|
if (RootSet[PredNum].ParentNodeID == SchedDFSResult::InvalidSubtreeID)
|
|
|
|
RootSet[PredNum].ParentNodeID = SU->NodeNum;
|
|
|
|
}
|
|
|
|
else if (RootSet.count(PredNum)) {
|
|
|
|
// The predecessor is not a root, but is still in the root set. This
|
|
|
|
// must be the new parent that it was just joined to. Note that
|
|
|
|
// RootSet[PredNum].ParentNodeID may either be invalid or may still be
|
|
|
|
// set to the original parent.
|
2013-01-25 06:52:27 +00:00
|
|
|
RData.SubInstrCount += RootSet[PredNum].SubInstrCount;
|
|
|
|
RootSet.erase(PredNum);
|
|
|
|
}
|
2013-01-25 06:02:44 +00:00
|
|
|
}
|
2013-01-25 06:52:27 +00:00
|
|
|
RootSet[SU->NodeNum] = RData;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Called once for each tree edge after calling visitPostOrderNode on the
|
|
|
|
/// predecessor. Increment the parent node's instruction count and
|
|
|
|
/// preemptively join this subtree to its parent's if it is small enough.
|
|
|
|
void visitPostorderEdge(const SDep &PredDep, const SUnit *Succ) {
|
|
|
|
R.DFSNodeData[Succ->NodeNum].InstrCount
|
|
|
|
+= R.DFSNodeData[PredDep.getSUnit()->NodeNum].InstrCount;
|
|
|
|
joinPredSubtree(PredDep, Succ);
|
2012-11-28 05:13:28 +00:00
|
|
|
}
|
|
|
|
|
2013-01-25 06:02:44 +00:00
|
|
|
/// Add a connection for cross edges.
|
|
|
|
void visitCrossEdge(const SDep &PredDep, const SUnit *Succ) {
|
2012-11-28 05:13:28 +00:00
|
|
|
ConnectionPairs.push_back(std::make_pair(PredDep.getSUnit(), Succ));
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Set each node's subtree ID to the representative ID and record connections
|
|
|
|
/// between trees.
|
|
|
|
void finalize() {
|
|
|
|
SubtreeClasses.compress();
|
2013-01-25 06:52:27 +00:00
|
|
|
R.DFSTreeData.resize(SubtreeClasses.getNumClasses());
|
|
|
|
assert(SubtreeClasses.getNumClasses() == RootSet.size()
|
|
|
|
&& "number of roots should match trees");
|
|
|
|
for (SparseSet<RootData>::const_iterator
|
|
|
|
RI = RootSet.begin(), RE = RootSet.end(); RI != RE; ++RI) {
|
|
|
|
unsigned TreeID = SubtreeClasses[RI->NodeID];
|
|
|
|
if (RI->ParentNodeID != SchedDFSResult::InvalidSubtreeID)
|
|
|
|
R.DFSTreeData[TreeID].ParentTreeID = SubtreeClasses[RI->ParentNodeID];
|
|
|
|
R.DFSTreeData[TreeID].SubInstrCount = RI->SubInstrCount;
|
2013-01-25 06:52:30 +00:00
|
|
|
// Note that SubInstrCount may be greater than InstrCount if we joined
|
|
|
|
// subtrees across a cross edge. InstrCount will be attributed to the
|
|
|
|
// original parent, while SubInstrCount will be attributed to the joined
|
|
|
|
// parent.
|
2013-01-25 06:52:27 +00:00
|
|
|
}
|
2012-11-28 05:13:28 +00:00
|
|
|
R.SubtreeConnections.resize(SubtreeClasses.getNumClasses());
|
|
|
|
R.SubtreeConnectLevels.resize(SubtreeClasses.getNumClasses());
|
|
|
|
DEBUG(dbgs() << R.getNumSubtrees() << " subtrees:\n");
|
2013-01-25 06:52:27 +00:00
|
|
|
for (unsigned Idx = 0, End = R.DFSNodeData.size(); Idx != End; ++Idx) {
|
|
|
|
R.DFSNodeData[Idx].SubtreeID = SubtreeClasses[Idx];
|
2012-11-28 05:13:28 +00:00
|
|
|
DEBUG(dbgs() << " SU(" << Idx << ") in tree "
|
2013-01-25 06:52:27 +00:00
|
|
|
<< R.DFSNodeData[Idx].SubtreeID << '\n');
|
2012-11-28 05:13:28 +00:00
|
|
|
}
|
|
|
|
for (std::vector<std::pair<const SUnit*, const SUnit*> >::const_iterator
|
|
|
|
I = ConnectionPairs.begin(), E = ConnectionPairs.end();
|
|
|
|
I != E; ++I) {
|
|
|
|
unsigned PredTree = SubtreeClasses[I->first->NodeNum];
|
|
|
|
unsigned SuccTree = SubtreeClasses[I->second->NodeNum];
|
|
|
|
if (PredTree == SuccTree)
|
|
|
|
continue;
|
|
|
|
unsigned Depth = I->first->getDepth();
|
|
|
|
addConnection(PredTree, SuccTree, Depth);
|
|
|
|
addConnection(SuccTree, PredTree, Depth);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
protected:
|
2013-01-25 06:02:44 +00:00
|
|
|
/// Join the predecessor subtree with the successor that is its DFS
|
|
|
|
/// parent. Apply some heuristics before joining.
|
|
|
|
bool joinPredSubtree(const SDep &PredDep, const SUnit *Succ,
|
|
|
|
bool CheckLimit = true) {
|
|
|
|
assert(PredDep.getKind() == SDep::Data && "Subtrees are for data edges");
|
|
|
|
|
|
|
|
// Check if the predecessor is already joined.
|
|
|
|
const SUnit *PredSU = PredDep.getSUnit();
|
|
|
|
unsigned PredNum = PredSU->NodeNum;
|
2013-01-25 06:52:27 +00:00
|
|
|
if (R.DFSNodeData[PredNum].SubtreeID != PredNum)
|
2013-01-25 06:02:44 +00:00
|
|
|
return false;
|
2013-01-25 00:12:57 +00:00
|
|
|
|
|
|
|
// Four is the magic number of successors before a node is considered a
|
|
|
|
// pinch point.
|
|
|
|
unsigned NumDataSucs = 0;
|
|
|
|
for (SUnit::const_succ_iterator SI = PredSU->Succs.begin(),
|
|
|
|
SE = PredSU->Succs.end(); SI != SE; ++SI) {
|
|
|
|
if (SI->getKind() == SDep::Data) {
|
|
|
|
if (++NumDataSucs >= 4)
|
2013-01-25 06:02:44 +00:00
|
|
|
return false;
|
2013-01-25 00:12:57 +00:00
|
|
|
}
|
|
|
|
}
|
2013-01-25 06:52:27 +00:00
|
|
|
if (CheckLimit && R.DFSNodeData[PredNum].InstrCount > R.SubtreeLimit)
|
2013-01-25 06:02:44 +00:00
|
|
|
return false;
|
2013-01-25 06:52:27 +00:00
|
|
|
R.DFSNodeData[PredNum].SubtreeID = Succ->NodeNum;
|
2013-01-25 06:02:44 +00:00
|
|
|
SubtreeClasses.join(Succ->NodeNum, PredNum);
|
|
|
|
return true;
|
2013-01-25 00:12:57 +00:00
|
|
|
}
|
|
|
|
|
2012-11-28 05:13:28 +00:00
|
|
|
/// Called by finalize() to record a connection between trees.
|
|
|
|
void addConnection(unsigned FromTree, unsigned ToTree, unsigned Depth) {
|
|
|
|
if (!Depth)
|
|
|
|
return;
|
|
|
|
|
2013-01-25 06:52:27 +00:00
|
|
|
do {
|
|
|
|
SmallVectorImpl<SchedDFSResult::Connection> &Connections =
|
|
|
|
R.SubtreeConnections[FromTree];
|
|
|
|
for (SmallVectorImpl<SchedDFSResult::Connection>::iterator
|
|
|
|
I = Connections.begin(), E = Connections.end(); I != E; ++I) {
|
|
|
|
if (I->TreeID == ToTree) {
|
|
|
|
I->Level = std::max(I->Level, Depth);
|
|
|
|
return;
|
|
|
|
}
|
2012-11-28 05:13:28 +00:00
|
|
|
}
|
2013-01-25 06:52:27 +00:00
|
|
|
Connections.push_back(SchedDFSResult::Connection(ToTree, Depth));
|
|
|
|
FromTree = R.DFSTreeData[FromTree].ParentTreeID;
|
|
|
|
} while (FromTree != SchedDFSResult::InvalidSubtreeID);
|
2012-11-28 05:13:28 +00:00
|
|
|
}
|
|
|
|
};
|
|
|
|
} // namespace llvm
|
|
|
|
|
2012-10-15 18:02:27 +00:00
|
|
|
namespace {
|
|
|
|
/// \brief Manage the stack used by a reverse depth-first search over the DAG.
|
|
|
|
class SchedDAGReverseDFS {
|
|
|
|
std::vector<std::pair<const SUnit*, SUnit::const_pred_iterator> > DFSStack;
|
|
|
|
public:
|
|
|
|
bool isComplete() const { return DFSStack.empty(); }
|
|
|
|
|
|
|
|
void follow(const SUnit *SU) {
|
|
|
|
DFSStack.push_back(std::make_pair(SU, SU->Preds.begin()));
|
|
|
|
}
|
|
|
|
void advance() { ++DFSStack.back().second; }
|
|
|
|
|
2012-11-28 05:13:28 +00:00
|
|
|
const SDep *backtrack() {
|
|
|
|
DFSStack.pop_back();
|
2014-04-14 00:51:57 +00:00
|
|
|
return DFSStack.empty() ? nullptr : std::prev(DFSStack.back().second);
|
2012-11-28 05:13:28 +00:00
|
|
|
}
|
2012-10-15 18:02:27 +00:00
|
|
|
|
|
|
|
const SUnit *getCurr() const { return DFSStack.back().first; }
|
|
|
|
|
|
|
|
SUnit::const_pred_iterator getPred() const { return DFSStack.back().second; }
|
|
|
|
|
|
|
|
SUnit::const_pred_iterator getPredEnd() const {
|
|
|
|
return getCurr()->Preds.end();
|
|
|
|
}
|
|
|
|
};
|
|
|
|
} // anonymous
|
|
|
|
|
2013-01-25 06:02:44 +00:00
|
|
|
static bool hasDataSucc(const SUnit *SU) {
|
|
|
|
for (SUnit::const_succ_iterator
|
|
|
|
SI = SU->Succs.begin(), SE = SU->Succs.end(); SI != SE; ++SI) {
|
2013-01-25 06:52:30 +00:00
|
|
|
if (SI->getKind() == SDep::Data && !SI->getSUnit()->isBoundaryNode())
|
2013-01-25 06:02:44 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2012-10-15 18:02:27 +00:00
|
|
|
/// Compute an ILP metric for all nodes in the subDAG reachable via depth-first
|
|
|
|
/// search from this root.
|
2013-01-25 06:33:57 +00:00
|
|
|
void SchedDFSResult::compute(ArrayRef<SUnit> SUnits) {
|
2012-10-15 18:02:27 +00:00
|
|
|
if (!IsBottomUp)
|
|
|
|
llvm_unreachable("Top-down ILP metric is unimplemnted");
|
|
|
|
|
2012-11-28 05:13:28 +00:00
|
|
|
SchedDFSImpl Impl(*this);
|
2013-01-25 06:33:57 +00:00
|
|
|
for (ArrayRef<SUnit>::const_iterator
|
|
|
|
SI = SUnits.begin(), SE = SUnits.end(); SI != SE; ++SI) {
|
|
|
|
const SUnit *SU = &*SI;
|
|
|
|
if (Impl.isVisited(SU) || hasDataSucc(SU))
|
|
|
|
continue;
|
|
|
|
|
2012-11-28 05:13:28 +00:00
|
|
|
SchedDAGReverseDFS DFS;
|
2013-01-25 06:33:57 +00:00
|
|
|
Impl.visitPreorder(SU);
|
|
|
|
DFS.follow(SU);
|
2012-11-28 05:13:28 +00:00
|
|
|
for (;;) {
|
|
|
|
// Traverse the leftmost path as far as possible.
|
|
|
|
while (DFS.getPred() != DFS.getPredEnd()) {
|
|
|
|
const SDep &PredDep = *DFS.getPred();
|
|
|
|
DFS.advance();
|
2013-01-25 06:02:44 +00:00
|
|
|
// Ignore non-data edges.
|
2013-01-25 06:52:30 +00:00
|
|
|
if (PredDep.getKind() != SDep::Data
|
|
|
|
|| PredDep.getSUnit()->isBoundaryNode()) {
|
2013-01-25 06:02:44 +00:00
|
|
|
continue;
|
2013-01-25 06:52:30 +00:00
|
|
|
}
|
2013-01-25 06:02:44 +00:00
|
|
|
// An already visited edge is a cross edge, assuming an acyclic DAG.
|
2012-11-28 05:13:28 +00:00
|
|
|
if (Impl.isVisited(PredDep.getSUnit())) {
|
2013-01-25 06:02:44 +00:00
|
|
|
Impl.visitCrossEdge(PredDep, DFS.getCurr());
|
2012-11-28 05:13:28 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
Impl.visitPreorder(PredDep.getSUnit());
|
|
|
|
DFS.follow(PredDep.getSUnit());
|
|
|
|
}
|
|
|
|
// Visit the top of the stack in postorder and backtrack.
|
|
|
|
const SUnit *Child = DFS.getCurr();
|
|
|
|
const SDep *PredDep = DFS.backtrack();
|
2013-01-25 06:02:44 +00:00
|
|
|
Impl.visitPostorderNode(Child);
|
|
|
|
if (PredDep)
|
|
|
|
Impl.visitPostorderEdge(*PredDep, DFS.getCurr());
|
2012-11-28 05:13:28 +00:00
|
|
|
if (DFS.isComplete())
|
|
|
|
break;
|
2012-10-15 18:02:27 +00:00
|
|
|
}
|
2012-11-28 05:13:28 +00:00
|
|
|
}
|
|
|
|
Impl.finalize();
|
|
|
|
}
|
|
|
|
|
|
|
|
/// The root of the given SubtreeID was just scheduled. For all subtrees
|
|
|
|
/// connected to this tree, record the depth of the connection so that the
|
|
|
|
/// nearest connected subtrees can be prioritized.
|
|
|
|
void SchedDFSResult::scheduleTree(unsigned SubtreeID) {
|
|
|
|
for (SmallVectorImpl<Connection>::const_iterator
|
|
|
|
I = SubtreeConnections[SubtreeID].begin(),
|
|
|
|
E = SubtreeConnections[SubtreeID].end(); I != E; ++I) {
|
|
|
|
SubtreeConnectLevels[I->TreeID] =
|
|
|
|
std::max(SubtreeConnectLevels[I->TreeID], I->Level);
|
|
|
|
DEBUG(dbgs() << " Tree: " << I->TreeID
|
|
|
|
<< " @" << SubtreeConnectLevels[I->TreeID] << '\n');
|
2012-10-15 18:02:27 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
|
|
|
|
void ILPValue::print(raw_ostream &OS) const {
|
2012-11-28 05:13:28 +00:00
|
|
|
OS << InstrCount << " / " << Length << " = ";
|
|
|
|
if (!Length)
|
2012-10-15 18:02:27 +00:00
|
|
|
OS << "BADILP";
|
2012-11-28 05:13:28 +00:00
|
|
|
else
|
|
|
|
OS << format("%g", ((double)InstrCount / Length));
|
2012-10-15 18:02:27 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void ILPValue::dump() const {
|
|
|
|
dbgs() << *this << '\n';
|
|
|
|
}
|
|
|
|
|
|
|
|
namespace llvm {
|
|
|
|
|
|
|
|
raw_ostream &operator<<(raw_ostream &OS, const ILPValue &Val) {
|
|
|
|
Val.print(OS);
|
|
|
|
return OS;
|
|
|
|
}
|
|
|
|
|
|
|
|
} // namespace llvm
|
|
|
|
#endif // !NDEBUG || LLVM_ENABLE_DUMP
|