2017-09-11 23:00:48 +00:00
|
|
|
//===- PhiElimination.cpp - Eliminate PHI nodes by inserting copies -------===//
|
2005-04-21 22:36:52 +00:00
|
|
|
//
|
2019-01-19 08:50:56 +00:00
|
|
|
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
|
|
// See https://llvm.org/LICENSE.txt for license information.
|
|
|
|
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
2005-04-21 22:36:52 +00:00
|
|
|
//
|
2003-10-20 19:43:21 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
2003-01-13 20:01:16 +00:00
|
|
|
//
|
|
|
|
// This pass eliminates machine instruction PHI nodes by inserting copy
|
|
|
|
// instructions. This destroys SSA information, but is the desired input for
|
|
|
|
// some register allocators.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2024-07-17 11:26:56 +08:00
|
|
|
#include "llvm/CodeGen/PHIElimination.h"
|
2010-12-05 19:51:05 +00:00
|
|
|
#include "PHIEliminationUtils.h"
|
2017-09-11 23:00:48 +00:00
|
|
|
#include "llvm/ADT/DenseMap.h"
|
2012-12-03 16:50:05 +00:00
|
|
|
#include "llvm/ADT/SmallPtrSet.h"
|
|
|
|
#include "llvm/ADT/Statistic.h"
|
2017-09-11 23:00:48 +00:00
|
|
|
#include "llvm/Analysis/LoopInfo.h"
|
|
|
|
#include "llvm/CodeGen/LiveInterval.h"
|
2017-12-13 02:51:04 +00:00
|
|
|
#include "llvm/CodeGen/LiveIntervals.h"
|
2005-05-05 23:45:17 +00:00
|
|
|
#include "llvm/CodeGen/LiveVariables.h"
|
2017-09-11 23:00:48 +00:00
|
|
|
#include "llvm/CodeGen/MachineBasicBlock.h"
|
2009-11-14 00:38:06 +00:00
|
|
|
#include "llvm/CodeGen/MachineDominators.h"
|
2017-09-11 23:00:48 +00:00
|
|
|
#include "llvm/CodeGen/MachineFunction.h"
|
|
|
|
#include "llvm/CodeGen/MachineFunctionPass.h"
|
2003-01-13 20:01:16 +00:00
|
|
|
#include "llvm/CodeGen/MachineInstr.h"
|
2008-04-11 17:54:45 +00:00
|
|
|
#include "llvm/CodeGen/MachineInstrBuilder.h"
|
2010-08-17 01:20:36 +00:00
|
|
|
#include "llvm/CodeGen/MachineLoopInfo.h"
|
2017-09-11 23:00:48 +00:00
|
|
|
#include "llvm/CodeGen/MachineOperand.h"
|
2007-12-31 04:13:23 +00:00
|
|
|
#include "llvm/CodeGen/MachineRegisterInfo.h"
|
2017-09-11 23:00:48 +00:00
|
|
|
#include "llvm/CodeGen/SlotIndexes.h"
|
2017-11-08 01:01:31 +00:00
|
|
|
#include "llvm/CodeGen/TargetInstrInfo.h"
|
2017-11-17 01:07:10 +00:00
|
|
|
#include "llvm/CodeGen/TargetOpcodes.h"
|
|
|
|
#include "llvm/CodeGen/TargetRegisterInfo.h"
|
|
|
|
#include "llvm/CodeGen/TargetSubtargetInfo.h"
|
2017-09-11 23:00:48 +00:00
|
|
|
#include "llvm/Pass.h"
|
2011-03-10 05:59:17 +00:00
|
|
|
#include "llvm/Support/CommandLine.h"
|
2009-11-10 22:01:05 +00:00
|
|
|
#include "llvm/Support/Debug.h"
|
2015-03-23 19:32:43 +00:00
|
|
|
#include "llvm/Support/raw_ostream.h"
|
2017-09-11 23:00:48 +00:00
|
|
|
#include <cassert>
|
|
|
|
#include <iterator>
|
|
|
|
#include <utility>
|
|
|
|
|
2004-02-23 18:38:20 +00:00
|
|
|
using namespace llvm;
|
2003-11-11 22:41:34 +00:00
|
|
|
|
2017-05-10 23:13:26 +00:00
|
|
|
#define DEBUG_TYPE "phi-node-elimination"
|
2014-04-22 02:02:50 +00:00
|
|
|
|
2011-03-10 05:59:17 +00:00
|
|
|
static cl::opt<bool>
|
2024-07-10 17:13:02 +08:00
|
|
|
DisableEdgeSplitting("disable-phi-elim-edge-splitting", cl::init(false),
|
|
|
|
cl::Hidden,
|
|
|
|
cl::desc("Disable critical edge splitting "
|
|
|
|
"during PHI elimination"));
|
2011-03-10 05:59:17 +00:00
|
|
|
|
2013-02-12 03:49:25 +00:00
|
|
|
static cl::opt<bool>
|
2024-07-10 17:13:02 +08:00
|
|
|
SplitAllCriticalEdges("phi-elim-split-all-critical-edges", cl::init(false),
|
|
|
|
cl::Hidden,
|
|
|
|
cl::desc("Split all critical edges during "
|
|
|
|
"PHI elimination"));
|
2013-02-12 03:49:25 +00:00
|
|
|
|
2015-03-03 10:23:11 +00:00
|
|
|
static cl::opt<bool> NoPhiElimLiveOutEarlyExit(
|
|
|
|
"no-phi-elim-live-out-early-exit", cl::init(false), cl::Hidden,
|
|
|
|
cl::desc("Do not use an early exit if isLiveOutPastPHIs returns true."));
|
|
|
|
|
2010-12-05 21:39:42 +00:00
|
|
|
namespace {
|
2017-09-11 23:00:48 +00:00
|
|
|
|
2024-07-17 11:26:56 +08:00
|
|
|
class PHIEliminationImpl {
|
2024-07-10 17:13:02 +08:00
|
|
|
MachineRegisterInfo *MRI = nullptr; // Machine register information
|
|
|
|
LiveVariables *LV = nullptr;
|
|
|
|
LiveIntervals *LIS = nullptr;
|
2024-07-17 11:26:56 +08:00
|
|
|
MachineLoopInfo *MLI = nullptr;
|
|
|
|
MachineDominatorTree *MDT = nullptr;
|
2010-12-05 21:39:42 +00:00
|
|
|
|
2024-07-10 17:13:02 +08:00
|
|
|
/// EliminatePHINodes - Eliminate phi nodes by inserting copy instructions
|
|
|
|
/// in predecessor basic blocks.
|
|
|
|
bool EliminatePHINodes(MachineFunction &MF, MachineBasicBlock &MBB);
|
2017-09-11 23:00:48 +00:00
|
|
|
|
2024-07-10 17:13:02 +08:00
|
|
|
void LowerPHINode(MachineBasicBlock &MBB,
|
|
|
|
MachineBasicBlock::iterator LastPHIIt,
|
|
|
|
bool AllEdgesCritical);
|
2010-12-05 21:39:42 +00:00
|
|
|
|
2024-07-10 17:13:02 +08:00
|
|
|
/// analyzePHINodes - Gather information about the PHI nodes in
|
|
|
|
/// here. In particular, we want to map the number of uses of a virtual
|
|
|
|
/// register which is used in a PHI node. We map that to the BB the
|
|
|
|
/// vreg is coming from. This is used later to determine when the vreg
|
|
|
|
/// is killed in the BB.
|
|
|
|
void analyzePHINodes(const MachineFunction &MF);
|
2010-12-05 21:39:42 +00:00
|
|
|
|
2024-07-10 17:13:02 +08:00
|
|
|
/// Split critical edges where necessary for good coalescer performance.
|
|
|
|
bool SplitPHIEdges(MachineFunction &MF, MachineBasicBlock &MBB,
|
|
|
|
MachineLoopInfo *MLI,
|
|
|
|
std::vector<SparseBitVector<>> *LiveInSets);
|
2010-12-05 21:39:42 +00:00
|
|
|
|
2024-07-10 17:13:02 +08:00
|
|
|
// These functions are temporary abstractions around LiveVariables and
|
|
|
|
// LiveIntervals, so they can go away when LiveVariables does.
|
|
|
|
bool isLiveIn(Register Reg, const MachineBasicBlock *MBB);
|
|
|
|
bool isLiveOutPastPHIs(Register Reg, const MachineBasicBlock *MBB);
|
2013-02-10 23:29:49 +00:00
|
|
|
|
2024-07-10 17:13:02 +08:00
|
|
|
using BBVRegPair = std::pair<unsigned, Register>;
|
|
|
|
using VRegPHIUse = DenseMap<BBVRegPair, unsigned>;
|
2010-12-05 21:39:42 +00:00
|
|
|
|
2024-07-10 17:13:02 +08:00
|
|
|
// Count the number of non-undef PHI uses of each register in each BB.
|
|
|
|
VRegPHIUse VRegPHIUseCount;
|
2010-12-05 21:39:42 +00:00
|
|
|
|
2024-07-10 17:13:02 +08:00
|
|
|
// Defs of PHI sources which are implicit_def.
|
|
|
|
SmallPtrSet<MachineInstr *, 4> ImpDefs;
|
2010-12-05 21:39:42 +00:00
|
|
|
|
2024-07-10 17:13:02 +08:00
|
|
|
// Map reusable lowered PHI node -> incoming join register.
|
|
|
|
using LoweredPHIMap =
|
|
|
|
DenseMap<MachineInstr *, unsigned, MachineInstrExpressionTrait>;
|
|
|
|
LoweredPHIMap LoweredPHIs;
|
2024-07-17 11:26:56 +08:00
|
|
|
|
|
|
|
MachineFunctionPass *P = nullptr;
|
|
|
|
MachineFunctionAnalysisManager *MFAM = nullptr;
|
|
|
|
|
|
|
|
public:
|
|
|
|
PHIEliminationImpl(MachineFunctionPass *P) : P(P) {
|
|
|
|
auto *LVWrapper = P->getAnalysisIfAvailable<LiveVariablesWrapperPass>();
|
|
|
|
auto *LISWrapper = P->getAnalysisIfAvailable<LiveIntervalsWrapperPass>();
|
|
|
|
auto *MLIWrapper = P->getAnalysisIfAvailable<MachineLoopInfoWrapperPass>();
|
|
|
|
auto *MDTWrapper =
|
|
|
|
P->getAnalysisIfAvailable<MachineDominatorTreeWrapperPass>();
|
|
|
|
LV = LVWrapper ? &LVWrapper->getLV() : nullptr;
|
|
|
|
LIS = LISWrapper ? &LISWrapper->getLIS() : nullptr;
|
|
|
|
MLI = MLIWrapper ? &MLIWrapper->getLI() : nullptr;
|
|
|
|
MDT = MDTWrapper ? &MDTWrapper->getDomTree() : nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
PHIEliminationImpl(MachineFunction &MF, MachineFunctionAnalysisManager &AM)
|
|
|
|
: LV(AM.getCachedResult<LiveVariablesAnalysis>(MF)),
|
|
|
|
LIS(AM.getCachedResult<LiveIntervalsAnalysis>(MF)),
|
|
|
|
MLI(AM.getCachedResult<MachineLoopAnalysis>(MF)),
|
|
|
|
MDT(AM.getCachedResult<MachineDominatorTreeAnalysis>(MF)), MFAM(&AM) {}
|
|
|
|
|
|
|
|
bool run(MachineFunction &MF);
|
|
|
|
};
|
|
|
|
|
|
|
|
class PHIElimination : public MachineFunctionPass {
|
|
|
|
public:
|
|
|
|
static char ID; // Pass identification, replacement for typeid
|
|
|
|
|
|
|
|
PHIElimination() : MachineFunctionPass(ID) {
|
|
|
|
initializePHIEliminationPass(*PassRegistry::getPassRegistry());
|
|
|
|
}
|
|
|
|
|
|
|
|
bool runOnMachineFunction(MachineFunction &MF) override {
|
|
|
|
PHIEliminationImpl Impl(this);
|
|
|
|
return Impl.run(MF);
|
|
|
|
}
|
|
|
|
|
|
|
|
MachineFunctionProperties getSetProperties() const override {
|
|
|
|
return MachineFunctionProperties().set(
|
|
|
|
MachineFunctionProperties::Property::NoPHIs);
|
|
|
|
}
|
|
|
|
|
|
|
|
void getAnalysisUsage(AnalysisUsage &AU) const override;
|
2024-07-10 17:13:02 +08:00
|
|
|
};
|
2017-09-11 23:00:48 +00:00
|
|
|
|
|
|
|
} // end anonymous namespace
|
2010-12-05 21:39:42 +00:00
|
|
|
|
2024-07-17 11:26:56 +08:00
|
|
|
PreservedAnalyses
|
|
|
|
PHIEliminationPass::run(MachineFunction &MF,
|
|
|
|
MachineFunctionAnalysisManager &MFAM) {
|
|
|
|
PHIEliminationImpl Impl(MF, MFAM);
|
|
|
|
bool Changed = Impl.run(MF);
|
|
|
|
if (!Changed)
|
|
|
|
return PreservedAnalyses::all();
|
|
|
|
auto PA = getMachineFunctionPassPreservedAnalyses();
|
|
|
|
PA.preserve<LiveIntervalsAnalysis>();
|
|
|
|
PA.preserve<LiveVariablesAnalysis>();
|
|
|
|
PA.preserve<SlotIndexesAnalysis>();
|
|
|
|
PA.preserve<MachineDominatorTreeAnalysis>();
|
|
|
|
PA.preserve<MachineLoopAnalysis>();
|
|
|
|
return PA;
|
|
|
|
}
|
|
|
|
|
2013-02-10 06:42:32 +00:00
|
|
|
STATISTIC(NumLowered, "Number of phis lowered");
|
2011-02-14 02:09:11 +00:00
|
|
|
STATISTIC(NumCriticalEdgesSplit, "Number of critical edges split");
|
2009-12-16 18:55:53 +00:00
|
|
|
STATISTIC(NumReused, "Number of reused lowered phis");
|
2009-11-10 22:01:05 +00:00
|
|
|
|
2009-07-21 23:47:33 +00:00
|
|
|
char PHIElimination::ID = 0;
|
2017-09-11 23:00:48 +00:00
|
|
|
|
2024-07-10 17:13:02 +08:00
|
|
|
char &llvm::PHIEliminationID = PHIElimination::ID;
|
2003-01-13 20:01:16 +00:00
|
|
|
|
2017-05-25 21:26:32 +00:00
|
|
|
INITIALIZE_PASS_BEGIN(PHIElimination, DEBUG_TYPE,
|
2024-07-10 17:13:02 +08:00
|
|
|
"Eliminate PHI nodes for register allocation", false,
|
|
|
|
false)
|
2024-07-09 10:50:43 +08:00
|
|
|
INITIALIZE_PASS_DEPENDENCY(LiveVariablesWrapperPass)
|
2017-05-25 21:26:32 +00:00
|
|
|
INITIALIZE_PASS_END(PHIElimination, DEBUG_TYPE,
|
2012-02-10 04:10:36 +00:00
|
|
|
"Eliminate PHI nodes for register allocation", false, false)
|
|
|
|
|
2010-12-05 21:39:42 +00:00
|
|
|
void PHIElimination::getAnalysisUsage(AnalysisUsage &AU) const {
|
2024-07-09 10:50:43 +08:00
|
|
|
AU.addUsedIfAvailable<LiveVariablesWrapperPass>();
|
|
|
|
AU.addPreserved<LiveVariablesWrapperPass>();
|
2024-07-09 12:09:11 +08:00
|
|
|
AU.addPreserved<SlotIndexesWrapperPass>();
|
2024-07-10 19:34:48 +08:00
|
|
|
AU.addPreserved<LiveIntervalsWrapperPass>();
|
2024-06-11 21:27:14 +08:00
|
|
|
AU.addPreserved<MachineDominatorTreeWrapperPass>();
|
2024-07-09 09:11:18 +08:00
|
|
|
AU.addPreserved<MachineLoopInfoWrapperPass>();
|
2009-07-31 23:37:33 +00:00
|
|
|
MachineFunctionPass::getAnalysisUsage(AU);
|
|
|
|
}
|
2009-07-21 23:47:33 +00:00
|
|
|
|
2024-07-17 11:26:56 +08:00
|
|
|
bool PHIEliminationImpl::run(MachineFunction &MF) {
|
2010-05-04 17:12:26 +00:00
|
|
|
MRI = &MF.getRegInfo();
|
2008-04-03 16:38:20 +00:00
|
|
|
|
|
|
|
bool Changed = false;
|
|
|
|
|
2018-10-08 23:47:35 +00:00
|
|
|
// Split critical edges to help the coalescer.
|
2013-02-11 09:24:47 +00:00
|
|
|
if (!DisableEdgeSplitting && (LV || LIS)) {
|
2020-01-20 21:56:09 +01:00
|
|
|
// A set of live-in regs for each MBB which is used to update LV
|
|
|
|
// efficiently also with large functions.
|
|
|
|
std::vector<SparseBitVector<>> LiveInSets;
|
|
|
|
if (LV) {
|
|
|
|
LiveInSets.resize(MF.size());
|
|
|
|
for (unsigned Index = 0, e = MRI->getNumVirtRegs(); Index != e; ++Index) {
|
|
|
|
// Set the bit for this register for each MBB where it is
|
|
|
|
// live-through or live-in (killed).
|
2022-12-07 00:23:34 +00:00
|
|
|
Register VirtReg = Register::index2VirtReg(Index);
|
2020-01-20 21:56:09 +01:00
|
|
|
MachineInstr *DefMI = MRI->getVRegDef(VirtReg);
|
|
|
|
if (!DefMI)
|
|
|
|
continue;
|
|
|
|
LiveVariables::VarInfo &VI = LV->getVarInfo(VirtReg);
|
|
|
|
SparseBitVector<>::iterator AliveBlockItr = VI.AliveBlocks.begin();
|
|
|
|
SparseBitVector<>::iterator EndItr = VI.AliveBlocks.end();
|
|
|
|
while (AliveBlockItr != EndItr) {
|
|
|
|
unsigned BlockNum = *(AliveBlockItr++);
|
|
|
|
LiveInSets[BlockNum].set(Index);
|
|
|
|
}
|
|
|
|
// The register is live into an MBB in which it is killed but not
|
|
|
|
// defined. See comment for VarInfo in LiveVariables.h.
|
|
|
|
MachineBasicBlock *DefMBB = DefMI->getParent();
|
|
|
|
if (VI.Kills.size() > 1 ||
|
|
|
|
(!VI.Kills.empty() && VI.Kills.front()->getParent() != DefMBB))
|
|
|
|
for (auto *MI : VI.Kills)
|
|
|
|
LiveInSets[MI->getParent()->getNumber()].set(Index);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-06-11 07:45:05 +00:00
|
|
|
for (auto &MBB : MF)
|
2020-01-20 21:56:09 +01:00
|
|
|
Changed |= SplitPHIEdges(MF, MBB, MLI, (LV ? &LiveInSets : nullptr));
|
2010-08-17 21:00:37 +00:00
|
|
|
}
|
2009-11-11 19:31:31 +00:00
|
|
|
|
2020-01-20 21:56:09 +01:00
|
|
|
// This pass takes the function out of SSA form.
|
|
|
|
MRI->leaveSSA();
|
|
|
|
|
2009-11-11 19:31:31 +00:00
|
|
|
// Populate VRegPHIUseCount
|
[CodeGen] Only deduplicate PHIs on critical edges (#97064)
PHIElim deduplicates identical PHI nodes to reduce the number of copies
inserted. There are two cases:
1. Identical PHI nodes are in different blocks. That's the reason for
this optimization; this can't be avoided at SSA-level. A necessary
prerequisite for this is that the predecessors of all basic blocks
(where such a PHI node could occur) are the same. This implies that
all (>= 2) predecessors must have multiple successors, i.e. all edges
into the block are critical edges.
2. Identical PHI nodes are in the same block. CSE can remove these.
There are a few cases, however, where they still occur regardless:
- expand-large-div-rem creates PHI nodes with large integers, which
get lowered into one PHI per MVT. Later, some identical values
(zeroes) get folded, resulting in identical PHI nodes.
- peephole-opt occasionally inserts PHIs for the same value.
- Some pseudo instruction emitters create redundant PHI nodes (e.g.,
AVR's insertShift), merging the same values more than once.
In any case, this happens rarely and MachineCSE handles most cases
anyway, so that PHIElim only gets to see very few of such cases (see
changed test files).
Currently, all PHI nodes are inserted into a DenseMap that checks
equality not by pointer but by operands. This hash map is pretty
expensive (hashing itself and the hash map), but only really useful in
the first case.
Avoid this expensive hashing most of the time by restricting it to basic
blocks with only critical input edges. This improves performance for
code with many PHI nodes, especially at -O0. (Note that Clang often
doesn't generate PHI nodes and -O0 includes no mem2reg. Other
compilers always generate PHI nodes.)
2024-07-03 11:19:05 +02:00
|
|
|
if (LV || LIS)
|
|
|
|
analyzePHINodes(MF);
|
2009-11-11 19:31:31 +00:00
|
|
|
|
2008-04-03 16:38:20 +00:00
|
|
|
// Eliminate PHI instructions by inserting copies into predecessor blocks.
|
2015-06-11 07:45:05 +00:00
|
|
|
for (auto &MBB : MF)
|
|
|
|
Changed |= EliminatePHINodes(MF, MBB);
|
2008-04-03 16:38:20 +00:00
|
|
|
|
|
|
|
// Remove dead IMPLICIT_DEF instructions.
|
2014-08-24 23:23:06 +00:00
|
|
|
for (MachineInstr *DefMI : ImpDefs) {
|
Apply llvm-prefer-register-over-unsigned from clang-tidy to LLVM
Summary:
This clang-tidy check is looking for unsigned integer variables whose initializer
starts with an implicit cast from llvm::Register and changes the type of the
variable to llvm::Register (dropping the llvm:: where possible).
Partial reverts in:
X86FrameLowering.cpp - Some functions return unsigned and arguably should be MCRegister
X86FixupLEAs.cpp - Some functions return unsigned and arguably should be MCRegister
X86FrameLowering.cpp - Some functions return unsigned and arguably should be MCRegister
HexagonBitSimplify.cpp - Function takes BitTracker::RegisterRef which appears to be unsigned&
MachineVerifier.cpp - Ambiguous operator==() given MCRegister and const Register
PPCFastISel.cpp - No Register::operator-=()
PeepholeOptimizer.cpp - TargetInstrInfo::optimizeLoadInstr() takes an unsigned&
MachineTraceMetrics.cpp - MachineTraceMetrics lacks a suitable constructor
Manual fixups in:
ARMFastISel.cpp - ARMEmitLoad() now takes a Register& instead of unsigned&
HexagonSplitDouble.cpp - Ternary operator was ambiguous between unsigned/Register
HexagonConstExtenders.cpp - Has a local class named Register, used llvm::Register instead of Register.
PPCFastISel.cpp - PPCEmitLoad() now takes a Register& instead of unsigned&
Depends on D65919
Reviewers: arsenm, bogner, craig.topper, RKSimon
Reviewed By: arsenm
Subscribers: RKSimon, craig.topper, lenary, aemerson, wuzish, jholewinski, MatzeB, qcolombet, dschuff, jyknight, dylanmckay, sdardis, nemanjai, jvesely, wdng, nhaehnle, sbc100, jgravelle-google, kristof.beyls, hiraditya, aheejin, kbarton, fedor.sergeev, javed.absar, asb, rbar, johnrusso, simoncook, apazos, sabuasal, niosHD, jrtc27, MaskRay, zzheng, edward-jones, atanasyan, rogfer01, MartinMosbeck, brucehoult, the_o, tpr, PkmX, jocewei, jsji, Petar.Avramovic, asbirlea, Jim, s.egerton, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65962
llvm-svn: 369041
2019-08-15 19:22:08 +00:00
|
|
|
Register DefReg = DefMI->getOperand(0).getReg();
|
2013-02-10 06:42:36 +00:00
|
|
|
if (MRI->use_nodbg_empty(DefReg)) {
|
|
|
|
if (LIS)
|
2016-02-27 06:40:41 +00:00
|
|
|
LIS->RemoveMachineInstrFromMaps(*DefMI);
|
2008-04-03 16:38:20 +00:00
|
|
|
DefMI->eraseFromParent();
|
2013-02-10 06:42:36 +00:00
|
|
|
}
|
2008-04-03 16:38:20 +00:00
|
|
|
}
|
|
|
|
|
2009-12-16 18:55:53 +00:00
|
|
|
// Clean up the lowered PHI instructions.
|
2016-02-27 06:40:41 +00:00
|
|
|
for (auto &I : LoweredPHIs) {
|
2013-02-12 05:48:56 +00:00
|
|
|
if (LIS)
|
2016-02-27 06:40:41 +00:00
|
|
|
LIS->RemoveMachineInstrFromMaps(*I.first);
|
2021-12-08 20:35:33 -08:00
|
|
|
MF.deleteMachineInstr(I.first);
|
2013-02-10 06:42:36 +00:00
|
|
|
}
|
2009-12-16 18:55:53 +00:00
|
|
|
|
2024-07-11 09:13:37 +02:00
|
|
|
// TODO: we should use the incremental DomTree updater here.
|
2024-07-17 11:26:56 +08:00
|
|
|
if (Changed && MDT)
|
|
|
|
MDT->getBase().recalculate(MF);
|
2024-07-11 09:13:37 +02:00
|
|
|
|
2009-12-17 23:42:32 +00:00
|
|
|
LoweredPHIs.clear();
|
2008-04-03 16:38:20 +00:00
|
|
|
ImpDefs.clear();
|
|
|
|
VRegPHIUseCount.clear();
|
2010-05-04 17:12:26 +00:00
|
|
|
|
2016-08-23 21:19:49 +00:00
|
|
|
MF.getProperties().set(MachineFunctionProperties::Property::NoPHIs);
|
|
|
|
|
2008-04-03 16:38:20 +00:00
|
|
|
return Changed;
|
|
|
|
}
|
|
|
|
|
2003-01-13 20:01:16 +00:00
|
|
|
/// EliminatePHINodes - Eliminate phi nodes by inserting copy instructions in
|
|
|
|
/// predecessor basic blocks.
|
2024-07-17 11:26:56 +08:00
|
|
|
bool PHIEliminationImpl::EliminatePHINodes(MachineFunction &MF,
|
|
|
|
MachineBasicBlock &MBB) {
|
2010-02-09 19:54:29 +00:00
|
|
|
if (MBB.empty() || !MBB.front().isPHI())
|
2024-07-10 17:13:02 +08:00
|
|
|
return false; // Quick exit for basic blocks without PHIs.
|
2003-01-13 20:01:16 +00:00
|
|
|
|
2018-09-30 17:23:21 +00:00
|
|
|
// Get an iterator to the last PHI node.
|
2013-07-01 19:42:46 +00:00
|
|
|
MachineBasicBlock::iterator LastPHIIt =
|
2024-07-10 17:13:02 +08:00
|
|
|
std::prev(MBB.SkipPHIsAndLabels(MBB.begin()));
|
2004-05-10 18:47:18 +00:00
|
|
|
|
[CodeGen] Only deduplicate PHIs on critical edges (#97064)
PHIElim deduplicates identical PHI nodes to reduce the number of copies
inserted. There are two cases:
1. Identical PHI nodes are in different blocks. That's the reason for
this optimization; this can't be avoided at SSA-level. A necessary
prerequisite for this is that the predecessors of all basic blocks
(where such a PHI node could occur) are the same. This implies that
all (>= 2) predecessors must have multiple successors, i.e. all edges
into the block are critical edges.
2. Identical PHI nodes are in the same block. CSE can remove these.
There are a few cases, however, where they still occur regardless:
- expand-large-div-rem creates PHI nodes with large integers, which
get lowered into one PHI per MVT. Later, some identical values
(zeroes) get folded, resulting in identical PHI nodes.
- peephole-opt occasionally inserts PHIs for the same value.
- Some pseudo instruction emitters create redundant PHI nodes (e.g.,
AVR's insertShift), merging the same values more than once.
In any case, this happens rarely and MachineCSE handles most cases
anyway, so that PHIElim only gets to see very few of such cases (see
changed test files).
Currently, all PHI nodes are inserted into a DenseMap that checks
equality not by pointer but by operands. This hash map is pretty
expensive (hashing itself and the hash map), but only really useful in
the first case.
Avoid this expensive hashing most of the time by restricting it to basic
blocks with only critical input edges. This improves performance for
code with many PHI nodes, especially at -O0. (Note that Clang often
doesn't generate PHI nodes and -O0 includes no mem2reg. Other
compilers always generate PHI nodes.)
2024-07-03 11:19:05 +02:00
|
|
|
// If all incoming edges are critical, we try to deduplicate identical PHIs so
|
|
|
|
// that we generate fewer copies. If at any edge is non-critical, we either
|
|
|
|
// have less than two predecessors (=> no PHIs) or a predecessor has only us
|
|
|
|
// as a successor (=> identical PHI node can't occur in different block).
|
|
|
|
bool AllEdgesCritical = MBB.pred_size() >= 2;
|
|
|
|
for (MachineBasicBlock *Pred : MBB.predecessors()) {
|
|
|
|
if (Pred->succ_size() < 2) {
|
|
|
|
AllEdgesCritical = false;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2010-02-09 19:54:29 +00:00
|
|
|
while (MBB.front().isPHI())
|
[CodeGen] Only deduplicate PHIs on critical edges (#97064)
PHIElim deduplicates identical PHI nodes to reduce the number of copies
inserted. There are two cases:
1. Identical PHI nodes are in different blocks. That's the reason for
this optimization; this can't be avoided at SSA-level. A necessary
prerequisite for this is that the predecessors of all basic blocks
(where such a PHI node could occur) are the same. This implies that
all (>= 2) predecessors must have multiple successors, i.e. all edges
into the block are critical edges.
2. Identical PHI nodes are in the same block. CSE can remove these.
There are a few cases, however, where they still occur regardless:
- expand-large-div-rem creates PHI nodes with large integers, which
get lowered into one PHI per MVT. Later, some identical values
(zeroes) get folded, resulting in identical PHI nodes.
- peephole-opt occasionally inserts PHIs for the same value.
- Some pseudo instruction emitters create redundant PHI nodes (e.g.,
AVR's insertShift), merging the same values more than once.
In any case, this happens rarely and MachineCSE handles most cases
anyway, so that PHIElim only gets to see very few of such cases (see
changed test files).
Currently, all PHI nodes are inserted into a DenseMap that checks
equality not by pointer but by operands. This hash map is pretty
expensive (hashing itself and the hash map), but only really useful in
the first case.
Avoid this expensive hashing most of the time by restricting it to basic
blocks with only critical input edges. This improves performance for
code with many PHI nodes, especially at -O0. (Note that Clang often
doesn't generate PHI nodes and -O0 includes no mem2reg. Other
compilers always generate PHI nodes.)
2024-07-03 11:19:05 +02:00
|
|
|
LowerPHINode(MBB, LastPHIIt, AllEdgesCritical);
|
2006-09-28 07:10:24 +00:00
|
|
|
|
2005-10-03 04:47:08 +00:00
|
|
|
return true;
|
|
|
|
}
|
2005-04-21 22:36:52 +00:00
|
|
|
|
2018-09-30 17:26:58 +00:00
|
|
|
/// Return true if all defs of VirtReg are implicit-defs.
|
2012-06-25 03:36:12 +00:00
|
|
|
/// This includes registers with no defs.
|
|
|
|
static bool isImplicitlyDefined(unsigned VirtReg,
|
2018-09-30 17:26:58 +00:00
|
|
|
const MachineRegisterInfo &MRI) {
|
|
|
|
for (MachineInstr &DI : MRI.def_instructions(VirtReg))
|
2014-03-17 19:36:09 +00:00
|
|
|
if (!DI.isImplicitDef())
|
2012-06-25 03:36:12 +00:00
|
|
|
return false;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2018-09-30 17:26:58 +00:00
|
|
|
/// Return true if all sources of the phi node are implicit_def's, or undef's.
|
|
|
|
static bool allPhiOperandsUndefined(const MachineInstr &MPhi,
|
|
|
|
const MachineRegisterInfo &MRI) {
|
|
|
|
for (unsigned I = 1, E = MPhi.getNumOperands(); I != E; I += 2) {
|
|
|
|
const MachineOperand &MO = MPhi.getOperand(I);
|
|
|
|
if (!isImplicitlyDefined(MO.getReg(), MRI) && !MO.isUndef())
|
2008-05-10 00:17:50 +00:00
|
|
|
return false;
|
2018-09-30 17:26:58 +00:00
|
|
|
}
|
2008-05-10 00:17:50 +00:00
|
|
|
return true;
|
2008-04-11 17:54:45 +00:00
|
|
|
}
|
2017-09-11 23:00:48 +00:00
|
|
|
/// LowerPHINode - Lower the PHI node at the top of the specified block.
|
2024-07-17 11:26:56 +08:00
|
|
|
void PHIEliminationImpl::LowerPHINode(MachineBasicBlock &MBB,
|
|
|
|
MachineBasicBlock::iterator LastPHIIt,
|
|
|
|
bool AllEdgesCritical) {
|
2013-02-10 06:42:32 +00:00
|
|
|
++NumLowered;
|
2013-07-01 19:42:46 +00:00
|
|
|
|
2014-03-02 12:27:27 +00:00
|
|
|
MachineBasicBlock::iterator AfterPHIsIt = std::next(LastPHIIt);
|
2013-07-01 19:42:46 +00:00
|
|
|
|
2005-10-03 04:47:08 +00:00
|
|
|
// Unlink the PHI node from the basic block, but don't delete the PHI yet.
|
2016-07-01 01:27:19 +00:00
|
|
|
MachineInstr *MPhi = MBB.remove(&*MBB.begin());
|
2005-10-03 04:47:08 +00:00
|
|
|
|
2008-04-11 17:54:45 +00:00
|
|
|
unsigned NumSrcs = (MPhi->getNumOperands() - 1) / 2;
|
Apply llvm-prefer-register-over-unsigned from clang-tidy to LLVM
Summary:
This clang-tidy check is looking for unsigned integer variables whose initializer
starts with an implicit cast from llvm::Register and changes the type of the
variable to llvm::Register (dropping the llvm:: where possible).
Partial reverts in:
X86FrameLowering.cpp - Some functions return unsigned and arguably should be MCRegister
X86FixupLEAs.cpp - Some functions return unsigned and arguably should be MCRegister
X86FrameLowering.cpp - Some functions return unsigned and arguably should be MCRegister
HexagonBitSimplify.cpp - Function takes BitTracker::RegisterRef which appears to be unsigned&
MachineVerifier.cpp - Ambiguous operator==() given MCRegister and const Register
PPCFastISel.cpp - No Register::operator-=()
PeepholeOptimizer.cpp - TargetInstrInfo::optimizeLoadInstr() takes an unsigned&
MachineTraceMetrics.cpp - MachineTraceMetrics lacks a suitable constructor
Manual fixups in:
ARMFastISel.cpp - ARMEmitLoad() now takes a Register& instead of unsigned&
HexagonSplitDouble.cpp - Ternary operator was ambiguous between unsigned/Register
HexagonConstExtenders.cpp - Has a local class named Register, used llvm::Register instead of Register.
PPCFastISel.cpp - PPCEmitLoad() now takes a Register& instead of unsigned&
Depends on D65919
Reviewers: arsenm, bogner, craig.topper, RKSimon
Reviewed By: arsenm
Subscribers: RKSimon, craig.topper, lenary, aemerson, wuzish, jholewinski, MatzeB, qcolombet, dschuff, jyknight, dylanmckay, sdardis, nemanjai, jvesely, wdng, nhaehnle, sbc100, jgravelle-google, kristof.beyls, hiraditya, aheejin, kbarton, fedor.sergeev, javed.absar, asb, rbar, johnrusso, simoncook, apazos, sabuasal, niosHD, jrtc27, MaskRay, zzheng, edward-jones, atanasyan, rogfer01, MartinMosbeck, brucehoult, the_o, tpr, PkmX, jocewei, jsji, Petar.Avramovic, asbirlea, Jim, s.egerton, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65962
llvm-svn: 369041
2019-08-15 19:22:08 +00:00
|
|
|
Register DestReg = MPhi->getOperand(0).getReg();
|
2010-08-18 16:09:47 +00:00
|
|
|
assert(MPhi->getOperand(0).getSubReg() == 0 && "Can't handle sub-reg PHIs");
|
2008-07-03 09:09:37 +00:00
|
|
|
bool isDead = MPhi->getOperand(0).isDead();
|
2005-10-03 04:47:08 +00:00
|
|
|
|
2006-09-28 07:10:24 +00:00
|
|
|
// Create a new register for the incoming PHI arguments.
|
2005-10-03 04:47:08 +00:00
|
|
|
MachineFunction &MF = *MBB.getParent();
|
2008-07-03 09:09:37 +00:00
|
|
|
unsigned IncomingReg = 0;
|
2024-07-10 17:13:02 +08:00
|
|
|
bool EliminateNow = true; // delay elimination of nodes in LoweredPHIs
|
|
|
|
bool reusedIncoming = false; // Is IncomingReg reused from an earlier PHI?
|
2005-10-03 04:47:08 +00:00
|
|
|
|
2008-05-12 22:15:05 +00:00
|
|
|
// Insert a register to register copy at the top of the current block (but
|
2005-10-03 04:47:08 +00:00
|
|
|
// after any remaining phi nodes) which copies the new incoming register
|
|
|
|
// into the phi node destination.
|
2019-09-17 09:08:58 +00:00
|
|
|
MachineInstr *PHICopy = nullptr;
|
2014-08-05 02:39:49 +00:00
|
|
|
const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
|
2018-09-30 17:26:58 +00:00
|
|
|
if (allPhiOperandsUndefined(*MPhi, *MRI))
|
|
|
|
// If all sources of a PHI node are implicit_def or undef uses, just emit an
|
2008-07-03 09:09:37 +00:00
|
|
|
// implicit_def instead of a copy.
|
2019-09-17 09:08:58 +00:00
|
|
|
PHICopy = BuildMI(MBB, AfterPHIsIt, MPhi->getDebugLoc(),
|
2024-07-10 17:13:02 +08:00
|
|
|
TII->get(TargetOpcode::IMPLICIT_DEF), DestReg);
|
2008-07-03 09:09:37 +00:00
|
|
|
else {
|
2009-12-16 18:55:53 +00:00
|
|
|
// Can we reuse an earlier PHI node? This only happens for critical edges,
|
[CodeGen] Only deduplicate PHIs on critical edges (#97064)
PHIElim deduplicates identical PHI nodes to reduce the number of copies
inserted. There are two cases:
1. Identical PHI nodes are in different blocks. That's the reason for
this optimization; this can't be avoided at SSA-level. A necessary
prerequisite for this is that the predecessors of all basic blocks
(where such a PHI node could occur) are the same. This implies that
all (>= 2) predecessors must have multiple successors, i.e. all edges
into the block are critical edges.
2. Identical PHI nodes are in the same block. CSE can remove these.
There are a few cases, however, where they still occur regardless:
- expand-large-div-rem creates PHI nodes with large integers, which
get lowered into one PHI per MVT. Later, some identical values
(zeroes) get folded, resulting in identical PHI nodes.
- peephole-opt occasionally inserts PHIs for the same value.
- Some pseudo instruction emitters create redundant PHI nodes (e.g.,
AVR's insertShift), merging the same values more than once.
In any case, this happens rarely and MachineCSE handles most cases
anyway, so that PHIElim only gets to see very few of such cases (see
changed test files).
Currently, all PHI nodes are inserted into a DenseMap that checks
equality not by pointer but by operands. This hash map is pretty
expensive (hashing itself and the hash map), but only really useful in
the first case.
Avoid this expensive hashing most of the time by restricting it to basic
blocks with only critical input edges. This improves performance for
code with many PHI nodes, especially at -O0. (Note that Clang often
doesn't generate PHI nodes and -O0 includes no mem2reg. Other
compilers always generate PHI nodes.)
2024-07-03 11:19:05 +02:00
|
|
|
// typically those created by tail duplication. Typically, an identical PHI
|
|
|
|
// node can't occur, so avoid hashing/storing such PHIs, which is somewhat
|
|
|
|
// expensive.
|
|
|
|
unsigned *Entry = nullptr;
|
|
|
|
if (AllEdgesCritical)
|
|
|
|
Entry = &LoweredPHIs[MPhi];
|
|
|
|
if (Entry && *Entry) {
|
2009-12-16 18:55:53 +00:00
|
|
|
// An identical PHI node was already lowered. Reuse the incoming register.
|
[CodeGen] Only deduplicate PHIs on critical edges (#97064)
PHIElim deduplicates identical PHI nodes to reduce the number of copies
inserted. There are two cases:
1. Identical PHI nodes are in different blocks. That's the reason for
this optimization; this can't be avoided at SSA-level. A necessary
prerequisite for this is that the predecessors of all basic blocks
(where such a PHI node could occur) are the same. This implies that
all (>= 2) predecessors must have multiple successors, i.e. all edges
into the block are critical edges.
2. Identical PHI nodes are in the same block. CSE can remove these.
There are a few cases, however, where they still occur regardless:
- expand-large-div-rem creates PHI nodes with large integers, which
get lowered into one PHI per MVT. Later, some identical values
(zeroes) get folded, resulting in identical PHI nodes.
- peephole-opt occasionally inserts PHIs for the same value.
- Some pseudo instruction emitters create redundant PHI nodes (e.g.,
AVR's insertShift), merging the same values more than once.
In any case, this happens rarely and MachineCSE handles most cases
anyway, so that PHIElim only gets to see very few of such cases (see
changed test files).
Currently, all PHI nodes are inserted into a DenseMap that checks
equality not by pointer but by operands. This hash map is pretty
expensive (hashing itself and the hash map), but only really useful in
the first case.
Avoid this expensive hashing most of the time by restricting it to basic
blocks with only critical input edges. This improves performance for
code with many PHI nodes, especially at -O0. (Note that Clang often
doesn't generate PHI nodes and -O0 includes no mem2reg. Other
compilers always generate PHI nodes.)
2024-07-03 11:19:05 +02:00
|
|
|
IncomingReg = *Entry;
|
2009-12-16 18:55:53 +00:00
|
|
|
reusedIncoming = true;
|
|
|
|
++NumReused;
|
2018-05-14 12:53:11 +00:00
|
|
|
LLVM_DEBUG(dbgs() << "Reusing " << printReg(IncomingReg) << " for "
|
|
|
|
<< *MPhi);
|
2009-12-16 18:55:53 +00:00
|
|
|
} else {
|
2010-07-10 19:08:25 +00:00
|
|
|
const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(DestReg);
|
[CodeGen] Only deduplicate PHIs on critical edges (#97064)
PHIElim deduplicates identical PHI nodes to reduce the number of copies
inserted. There are two cases:
1. Identical PHI nodes are in different blocks. That's the reason for
this optimization; this can't be avoided at SSA-level. A necessary
prerequisite for this is that the predecessors of all basic blocks
(where such a PHI node could occur) are the same. This implies that
all (>= 2) predecessors must have multiple successors, i.e. all edges
into the block are critical edges.
2. Identical PHI nodes are in the same block. CSE can remove these.
There are a few cases, however, where they still occur regardless:
- expand-large-div-rem creates PHI nodes with large integers, which
get lowered into one PHI per MVT. Later, some identical values
(zeroes) get folded, resulting in identical PHI nodes.
- peephole-opt occasionally inserts PHIs for the same value.
- Some pseudo instruction emitters create redundant PHI nodes (e.g.,
AVR's insertShift), merging the same values more than once.
In any case, this happens rarely and MachineCSE handles most cases
anyway, so that PHIElim only gets to see very few of such cases (see
changed test files).
Currently, all PHI nodes are inserted into a DenseMap that checks
equality not by pointer but by operands. This hash map is pretty
expensive (hashing itself and the hash map), but only really useful in
the first case.
Avoid this expensive hashing most of the time by restricting it to basic
blocks with only critical input edges. This improves performance for
code with many PHI nodes, especially at -O0. (Note that Clang often
doesn't generate PHI nodes and -O0 includes no mem2reg. Other
compilers always generate PHI nodes.)
2024-07-03 11:19:05 +02:00
|
|
|
IncomingReg = MF.getRegInfo().createVirtualRegister(RC);
|
|
|
|
if (Entry) {
|
|
|
|
EliminateNow = false;
|
|
|
|
*Entry = IncomingReg;
|
|
|
|
}
|
2009-12-16 18:55:53 +00:00
|
|
|
}
|
[CodeGen] Only deduplicate PHIs on critical edges (#97064)
PHIElim deduplicates identical PHI nodes to reduce the number of copies
inserted. There are two cases:
1. Identical PHI nodes are in different blocks. That's the reason for
this optimization; this can't be avoided at SSA-level. A necessary
prerequisite for this is that the predecessors of all basic blocks
(where such a PHI node could occur) are the same. This implies that
all (>= 2) predecessors must have multiple successors, i.e. all edges
into the block are critical edges.
2. Identical PHI nodes are in the same block. CSE can remove these.
There are a few cases, however, where they still occur regardless:
- expand-large-div-rem creates PHI nodes with large integers, which
get lowered into one PHI per MVT. Later, some identical values
(zeroes) get folded, resulting in identical PHI nodes.
- peephole-opt occasionally inserts PHIs for the same value.
- Some pseudo instruction emitters create redundant PHI nodes (e.g.,
AVR's insertShift), merging the same values more than once.
In any case, this happens rarely and MachineCSE handles most cases
anyway, so that PHIElim only gets to see very few of such cases (see
changed test files).
Currently, all PHI nodes are inserted into a DenseMap that checks
equality not by pointer but by operands. This hash map is pretty
expensive (hashing itself and the hash map), but only really useful in
the first case.
Avoid this expensive hashing most of the time by restricting it to basic
blocks with only critical input edges. This improves performance for
code with many PHI nodes, especially at -O0. (Note that Clang often
doesn't generate PHI nodes and -O0 includes no mem2reg. Other
compilers always generate PHI nodes.)
2024-07-03 11:19:05 +02:00
|
|
|
|
2019-09-17 09:08:58 +00:00
|
|
|
// Give the target possiblity to handle special cases fallthrough otherwise
|
2024-07-10 17:13:02 +08:00
|
|
|
PHICopy = TII->createPHIDestinationCopy(
|
|
|
|
MBB, AfterPHIsIt, MPhi->getDebugLoc(), IncomingReg, DestReg);
|
2008-07-03 09:09:37 +00:00
|
|
|
}
|
2003-01-13 20:01:16 +00:00
|
|
|
|
[DebugInstrRef][1/3] Track PHI values through register allocation
This patch introduces "DBG_PHI" instructions, a marker of where a PHI
instruction used to be, before PHI elimination. Under the instruction
referencing model, we want to know where every value in the function is
defined -- and a PHI, even if implicit, is such a place.
Just like instruction numbers, we can use this to identify a value to be
used as a variable value, but we don't need to know what instruction
defines that value, for example:
bb1:
DBG_PHI $rax, 1
[... more insts ... ]
bb2:
DBG_INSTR_REF 1, 0, !1234, !DIExpression()
This specifies that on entry to bb1, whatever value is in $rax is known
as value number one -- and the later DBG_INSTR_REF marks the position
where variable !1234 should take on value number one.
PHI locations are stored in MachineFunction for the duration of the
regalloc phase in the DebugPHIPositions map. The map is populated by
PHIElimination, and then flushed back into the instruction stream by
virtregrewriter. A small amount of maintenence is needed in
LiveDebugVariables to account for registers being split, but only for
individual positions, not for entire ranges of blocks.
Differential Revision: https://reviews.llvm.org/D86812
2021-05-26 19:53:33 +01:00
|
|
|
if (MPhi->peekDebugInstrNum()) {
|
|
|
|
// If referred to by debug-info, store where this PHI was.
|
|
|
|
MachineFunction *MF = MBB.getParent();
|
|
|
|
unsigned ID = MPhi->peekDebugInstrNum();
|
|
|
|
auto P = MachineFunction::DebugPHIRegallocPos(&MBB, IncomingReg, 0);
|
|
|
|
auto Res = MF->DebugPHIPositions.insert({ID, P});
|
|
|
|
assert(Res.second);
|
|
|
|
(void)Res;
|
|
|
|
}
|
|
|
|
|
2008-05-12 22:15:05 +00:00
|
|
|
// Update live variable information if there is any.
|
2005-10-03 04:47:08 +00:00
|
|
|
if (LV) {
|
2008-07-03 09:09:37 +00:00
|
|
|
if (IncomingReg) {
|
2009-12-16 18:55:53 +00:00
|
|
|
LiveVariables::VarInfo &VI = LV->getVarInfo(IncomingReg);
|
|
|
|
|
2020-07-30 08:18:50 +00:00
|
|
|
MachineInstr *OldKill = nullptr;
|
|
|
|
bool IsPHICopyAfterOldKill = false;
|
|
|
|
|
|
|
|
if (reusedIncoming && (OldKill = VI.findKill(&MBB))) {
|
|
|
|
// Calculate whether the PHICopy is after the OldKill.
|
|
|
|
// In general, the PHICopy is inserted as the first non-phi instruction
|
|
|
|
// by default, so it's before the OldKill. But some Target hooks for
|
|
|
|
// createPHIDestinationCopy() may modify the default insert position of
|
|
|
|
// PHICopy.
|
2024-07-10 17:13:02 +08:00
|
|
|
for (auto I = MBB.SkipPHIsAndLabels(MBB.begin()), E = MBB.end(); I != E;
|
|
|
|
++I) {
|
2020-07-30 08:18:50 +00:00
|
|
|
if (I == PHICopy)
|
|
|
|
break;
|
|
|
|
|
|
|
|
if (I == OldKill) {
|
|
|
|
IsPHICopyAfterOldKill = true;
|
|
|
|
break;
|
|
|
|
}
|
2009-12-16 18:55:53 +00:00
|
|
|
}
|
2020-07-30 08:18:50 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// When we are reusing the incoming register and it has been marked killed
|
|
|
|
// by OldKill, if the PHICopy is after the OldKill, we should remove the
|
|
|
|
// killed flag from OldKill.
|
|
|
|
if (IsPHICopyAfterOldKill) {
|
|
|
|
LLVM_DEBUG(dbgs() << "Remove old kill from " << *OldKill);
|
|
|
|
LV->removeVirtualRegisterKilled(IncomingReg, *OldKill);
|
|
|
|
LLVM_DEBUG(MBB.dump());
|
|
|
|
}
|
2008-07-03 09:09:37 +00:00
|
|
|
|
2020-07-30 08:18:50 +00:00
|
|
|
// Add information to LiveVariables to know that the first used incoming
|
|
|
|
// value or the resued incoming value whose PHICopy is after the OldKIll
|
|
|
|
// is killed. Note that because the value is defined in several places
|
|
|
|
// (once each for each incoming block), the "def" block and instruction
|
|
|
|
// fields for the VarInfo is not filled in.
|
|
|
|
if (!OldKill || IsPHICopyAfterOldKill)
|
|
|
|
LV->addVirtualRegisterKilled(IncomingReg, *PHICopy);
|
2008-07-03 09:09:37 +00:00
|
|
|
}
|
2003-01-13 20:01:16 +00:00
|
|
|
|
2008-05-12 22:15:05 +00:00
|
|
|
// Since we are going to be deleting the PHI node, if it is the last use of
|
|
|
|
// any registers, or if the value itself is dead, we need to move this
|
2005-10-03 04:47:08 +00:00
|
|
|
// information over to the new copy we just inserted.
|
2016-07-01 01:51:32 +00:00
|
|
|
LV->removeVirtualRegistersKilled(*MPhi);
|
2005-10-03 04:47:08 +00:00
|
|
|
|
2005-10-03 07:22:07 +00:00
|
|
|
// If the result is dead, update LV.
|
2008-07-03 09:09:37 +00:00
|
|
|
if (isDead) {
|
2019-09-17 09:08:58 +00:00
|
|
|
LV->addVirtualRegisterDead(DestReg, *PHICopy);
|
2016-07-01 01:51:32 +00:00
|
|
|
LV->removeVirtualRegisterDead(DestReg, *MPhi);
|
2003-05-12 03:55:21 +00:00
|
|
|
}
|
2005-10-03 04:47:08 +00:00
|
|
|
}
|
|
|
|
|
2013-02-10 06:42:36 +00:00
|
|
|
// Update LiveIntervals for the new copy or implicit def.
|
|
|
|
if (LIS) {
|
2019-09-17 09:08:58 +00:00
|
|
|
SlotIndex DestCopyIndex = LIS->InsertMachineInstrInMaps(*PHICopy);
|
2013-02-10 06:42:36 +00:00
|
|
|
|
|
|
|
SlotIndex MBBStartIndex = LIS->getMBBStartIdx(&MBB);
|
|
|
|
if (IncomingReg) {
|
|
|
|
// Add the region from the beginning of MBB to the copy instruction to
|
|
|
|
// IncomingReg's live interval.
|
2023-11-13 12:16:26 +09:00
|
|
|
LiveInterval &IncomingLI = LIS->getOrCreateEmptyInterval(IncomingReg);
|
2013-02-10 06:42:36 +00:00
|
|
|
VNInfo *IncomingVNI = IncomingLI.getVNInfoAt(MBBStartIndex);
|
|
|
|
if (!IncomingVNI)
|
2024-07-10 17:13:02 +08:00
|
|
|
IncomingVNI =
|
|
|
|
IncomingLI.getNextValue(MBBStartIndex, LIS->getVNInfoAllocator());
|
|
|
|
IncomingLI.addSegment(LiveInterval::Segment(
|
|
|
|
MBBStartIndex, DestCopyIndex.getRegSlot(), IncomingVNI));
|
2013-02-10 06:42:36 +00:00
|
|
|
}
|
|
|
|
|
2013-02-21 08:51:55 +00:00
|
|
|
LiveInterval &DestLI = LIS->getInterval(DestReg);
|
2023-11-13 12:16:26 +09:00
|
|
|
assert(!DestLI.empty() && "PHIs should have non-empty LiveIntervals.");
|
|
|
|
|
|
|
|
SlotIndex NewStart = DestCopyIndex.getRegSlot();
|
|
|
|
|
|
|
|
SmallVector<LiveRange *> ToUpdate({&DestLI});
|
|
|
|
for (auto &SR : DestLI.subranges())
|
|
|
|
ToUpdate.push_back(&SR);
|
|
|
|
|
|
|
|
for (auto LR : ToUpdate) {
|
|
|
|
auto DestSegment = LR->find(MBBStartIndex);
|
|
|
|
assert(DestSegment != LR->end() &&
|
|
|
|
"PHI destination must be live in block");
|
|
|
|
|
|
|
|
if (LR->endIndex().isDead()) {
|
|
|
|
// A dead PHI's live range begins and ends at the start of the MBB, but
|
|
|
|
// the lowered copy, which will still be dead, needs to begin and end at
|
|
|
|
// the copy instruction.
|
|
|
|
VNInfo *OrigDestVNI = LR->getVNInfoAt(DestSegment->start);
|
|
|
|
assert(OrigDestVNI && "PHI destination should be live at block entry.");
|
|
|
|
LR->removeSegment(DestSegment->start, DestSegment->start.getDeadSlot());
|
|
|
|
LR->createDeadDef(NewStart, LIS->getVNInfoAllocator());
|
|
|
|
LR->removeValNo(OrigDestVNI);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Destination copies are not inserted in the same order as the PHI nodes
|
|
|
|
// they replace. Hence the start of the live range may need to be adjusted
|
|
|
|
// to match the actual slot index of the copy.
|
|
|
|
if (DestSegment->start > NewStart) {
|
|
|
|
VNInfo *VNI = LR->getVNInfoAt(DestSegment->start);
|
|
|
|
assert(VNI && "value should be defined for known segment");
|
|
|
|
LR->addSegment(
|
|
|
|
LiveInterval::Segment(NewStart, DestSegment->start, VNI));
|
|
|
|
} else if (DestSegment->start < NewStart) {
|
|
|
|
assert(DestSegment->start >= MBBStartIndex);
|
|
|
|
assert(DestSegment->end >= DestCopyIndex.getRegSlot());
|
|
|
|
LR->removeSegment(DestSegment->start, NewStart);
|
|
|
|
}
|
|
|
|
VNInfo *DestVNI = LR->getVNInfoAt(NewStart);
|
2013-02-10 06:42:36 +00:00
|
|
|
assert(DestVNI && "PHI destination should be live at its definition.");
|
2023-11-13 12:16:26 +09:00
|
|
|
DestVNI->def = NewStart;
|
2013-02-10 06:42:36 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-05-12 22:15:05 +00:00
|
|
|
// Adjust the VRegPHIUseCount map to account for the removal of this PHI node.
|
[CodeGen] Only deduplicate PHIs on critical edges (#97064)
PHIElim deduplicates identical PHI nodes to reduce the number of copies
inserted. There are two cases:
1. Identical PHI nodes are in different blocks. That's the reason for
this optimization; this can't be avoided at SSA-level. A necessary
prerequisite for this is that the predecessors of all basic blocks
(where such a PHI node could occur) are the same. This implies that
all (>= 2) predecessors must have multiple successors, i.e. all edges
into the block are critical edges.
2. Identical PHI nodes are in the same block. CSE can remove these.
There are a few cases, however, where they still occur regardless:
- expand-large-div-rem creates PHI nodes with large integers, which
get lowered into one PHI per MVT. Later, some identical values
(zeroes) get folded, resulting in identical PHI nodes.
- peephole-opt occasionally inserts PHIs for the same value.
- Some pseudo instruction emitters create redundant PHI nodes (e.g.,
AVR's insertShift), merging the same values more than once.
In any case, this happens rarely and MachineCSE handles most cases
anyway, so that PHIElim only gets to see very few of such cases (see
changed test files).
Currently, all PHI nodes are inserted into a DenseMap that checks
equality not by pointer but by operands. This hash map is pretty
expensive (hashing itself and the hash map), but only really useful in
the first case.
Avoid this expensive hashing most of the time by restricting it to basic
blocks with only critical input edges. This improves performance for
code with many PHI nodes, especially at -O0. (Note that Clang often
doesn't generate PHI nodes and -O0 includes no mem2reg. Other
compilers always generate PHI nodes.)
2024-07-03 11:19:05 +02:00
|
|
|
if (LV || LIS) {
|
|
|
|
for (unsigned i = 1; i != MPhi->getNumOperands(); i += 2) {
|
|
|
|
if (!MPhi->getOperand(i).isUndef()) {
|
|
|
|
--VRegPHIUseCount[BBVRegPair(
|
|
|
|
MPhi->getOperand(i + 1).getMBB()->getNumber(),
|
|
|
|
MPhi->getOperand(i).getReg())];
|
|
|
|
}
|
2021-10-11 16:02:40 +01:00
|
|
|
}
|
|
|
|
}
|
2005-10-03 04:47:08 +00:00
|
|
|
|
2008-05-12 22:15:05 +00:00
|
|
|
// Now loop over all of the incoming arguments, changing them to copy into the
|
|
|
|
// IncomingReg register in the corresponding predecessor basic block.
|
2024-07-10 17:13:02 +08:00
|
|
|
SmallPtrSet<MachineBasicBlock *, 8> MBBsInsertedInto;
|
2008-04-11 17:54:45 +00:00
|
|
|
for (int i = NumSrcs - 1; i >= 0; --i) {
|
Apply llvm-prefer-register-over-unsigned from clang-tidy to LLVM
Summary:
This clang-tidy check is looking for unsigned integer variables whose initializer
starts with an implicit cast from llvm::Register and changes the type of the
variable to llvm::Register (dropping the llvm:: where possible).
Partial reverts in:
X86FrameLowering.cpp - Some functions return unsigned and arguably should be MCRegister
X86FixupLEAs.cpp - Some functions return unsigned and arguably should be MCRegister
X86FrameLowering.cpp - Some functions return unsigned and arguably should be MCRegister
HexagonBitSimplify.cpp - Function takes BitTracker::RegisterRef which appears to be unsigned&
MachineVerifier.cpp - Ambiguous operator==() given MCRegister and const Register
PPCFastISel.cpp - No Register::operator-=()
PeepholeOptimizer.cpp - TargetInstrInfo::optimizeLoadInstr() takes an unsigned&
MachineTraceMetrics.cpp - MachineTraceMetrics lacks a suitable constructor
Manual fixups in:
ARMFastISel.cpp - ARMEmitLoad() now takes a Register& instead of unsigned&
HexagonSplitDouble.cpp - Ternary operator was ambiguous between unsigned/Register
HexagonConstExtenders.cpp - Has a local class named Register, used llvm::Register instead of Register.
PPCFastISel.cpp - PPCEmitLoad() now takes a Register& instead of unsigned&
Depends on D65919
Reviewers: arsenm, bogner, craig.topper, RKSimon
Reviewed By: arsenm
Subscribers: RKSimon, craig.topper, lenary, aemerson, wuzish, jholewinski, MatzeB, qcolombet, dschuff, jyknight, dylanmckay, sdardis, nemanjai, jvesely, wdng, nhaehnle, sbc100, jgravelle-google, kristof.beyls, hiraditya, aheejin, kbarton, fedor.sergeev, javed.absar, asb, rbar, johnrusso, simoncook, apazos, sabuasal, niosHD, jrtc27, MaskRay, zzheng, edward-jones, atanasyan, rogfer01, MartinMosbeck, brucehoult, the_o, tpr, PkmX, jocewei, jsji, Petar.Avramovic, asbirlea, Jim, s.egerton, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65962
llvm-svn: 369041
2019-08-15 19:22:08 +00:00
|
|
|
Register SrcReg = MPhi->getOperand(i * 2 + 1).getReg();
|
2024-07-10 17:13:02 +08:00
|
|
|
unsigned SrcSubReg = MPhi->getOperand(i * 2 + 1).getSubReg();
|
|
|
|
bool SrcUndef = MPhi->getOperand(i * 2 + 1).isUndef() ||
|
|
|
|
isImplicitlyDefined(SrcReg, *MRI);
|
2023-01-13 14:38:08 -08:00
|
|
|
assert(SrcReg.isVirtual() &&
|
2005-10-03 07:22:07 +00:00
|
|
|
"Machine PHI Operands must all be virtual registers!");
|
2005-10-03 04:47:08 +00:00
|
|
|
|
2009-07-23 04:34:03 +00:00
|
|
|
// Get the MachineBasicBlock equivalent of the BasicBlock that is the source
|
|
|
|
// path the PHI.
|
2024-07-10 17:13:02 +08:00
|
|
|
MachineBasicBlock &opBlock = *MPhi->getOperand(i * 2 + 2).getMBB();
|
2009-07-23 04:34:03 +00:00
|
|
|
|
2005-10-03 04:47:08 +00:00
|
|
|
// Check to make sure we haven't already emitted the copy for this block.
|
2008-05-12 22:15:05 +00:00
|
|
|
// This can happen because PHI nodes may have multiple entries for the same
|
|
|
|
// basic block.
|
2014-11-19 07:49:26 +00:00
|
|
|
if (!MBBsInsertedInto.insert(&opBlock).second)
|
2024-07-10 17:13:02 +08:00
|
|
|
continue; // If the copy has already been emitted, we're done.
|
2009-11-10 22:00:56 +00:00
|
|
|
|
2020-12-10 12:14:23 +00:00
|
|
|
MachineInstr *SrcRegDef = MRI->getVRegDef(SrcReg);
|
|
|
|
if (SrcRegDef && TII->isUnspillableTerminator(SrcRegDef)) {
|
|
|
|
assert(SrcRegDef->getOperand(0).isReg() &&
|
|
|
|
SrcRegDef->getOperand(0).isDef() &&
|
|
|
|
"Expected operand 0 to be a reg def!");
|
|
|
|
// Now that the PHI's use has been removed (as the instruction was
|
|
|
|
// removed) there should be no other uses of the SrcReg.
|
|
|
|
assert(MRI->use_empty(SrcReg) &&
|
|
|
|
"Expected a single use from UnspillableTerminator");
|
|
|
|
SrcRegDef->getOperand(0).setReg(IncomingReg);
|
2021-10-01 15:39:38 +01:00
|
|
|
|
|
|
|
// Update LiveVariables.
|
|
|
|
if (LV) {
|
|
|
|
LiveVariables::VarInfo &SrcVI = LV->getVarInfo(SrcReg);
|
|
|
|
LiveVariables::VarInfo &IncomingVI = LV->getVarInfo(IncomingReg);
|
|
|
|
IncomingVI.AliveBlocks = std::move(SrcVI.AliveBlocks);
|
|
|
|
SrcVI.AliveBlocks.clear();
|
|
|
|
}
|
|
|
|
|
2020-12-10 12:14:23 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2008-05-12 22:15:05 +00:00
|
|
|
// Find a safe location to insert the copy, this may be the first terminator
|
|
|
|
// in the block (or end()).
|
2009-11-13 21:56:15 +00:00
|
|
|
MachineBasicBlock::iterator InsertPos =
|
2024-07-10 17:13:02 +08:00
|
|
|
findPHICopyInsertPoint(&opBlock, &MBB, SrcReg);
|
2009-03-13 22:59:14 +00:00
|
|
|
|
2005-10-03 07:22:07 +00:00
|
|
|
// Insert the copy.
|
2014-04-14 00:51:57 +00:00
|
|
|
MachineInstr *NewSrcInstr = nullptr;
|
2012-06-25 03:36:12 +00:00
|
|
|
if (!reusedIncoming && IncomingReg) {
|
|
|
|
if (SrcUndef) {
|
|
|
|
// The source register is undefined, so there is no need for a real
|
|
|
|
// COPY, but we still need to ensure joint dominance by defs.
|
|
|
|
// Insert an IMPLICIT_DEF instruction.
|
2024-07-10 17:13:02 +08:00
|
|
|
NewSrcInstr =
|
|
|
|
BuildMI(opBlock, InsertPos, MPhi->getDebugLoc(),
|
|
|
|
TII->get(TargetOpcode::IMPLICIT_DEF), IncomingReg);
|
2012-06-25 03:36:12 +00:00
|
|
|
|
|
|
|
// Clean up the old implicit-def, if there even was one.
|
|
|
|
if (MachineInstr *DefMI = MRI->getVRegDef(SrcReg))
|
|
|
|
if (DefMI->isImplicitDef())
|
|
|
|
ImpDefs.insert(DefMI);
|
|
|
|
} else {
|
2021-04-20 13:26:40 -07:00
|
|
|
// Delete the debug location, since the copy is inserted into a
|
|
|
|
// different basic block.
|
|
|
|
NewSrcInstr = TII->createPHISourceCopy(opBlock, InsertPos, nullptr,
|
|
|
|
SrcReg, SrcSubReg, IncomingReg);
|
2012-06-25 03:36:12 +00:00
|
|
|
}
|
|
|
|
}
|
2005-10-03 07:22:07 +00:00
|
|
|
|
2013-02-10 06:42:36 +00:00
|
|
|
// We only need to update the LiveVariables kill of SrcReg if this was the
|
|
|
|
// last PHI use of SrcReg to be lowered on this CFG edge and it is not live
|
|
|
|
// out of the predecessor. We can also ignore undef sources.
|
|
|
|
if (LV && !SrcUndef &&
|
|
|
|
!VRegPHIUseCount[BBVRegPair(opBlock.getNumber(), SrcReg)] &&
|
|
|
|
!LV->isLiveOut(SrcReg, opBlock)) {
|
|
|
|
// We want to be able to insert a kill of the register if this PHI (aka,
|
|
|
|
// the copy we just inserted) is the last use of the source value. Live
|
|
|
|
// variable analysis conservatively handles this by saying that the value
|
|
|
|
// is live until the end of the block the PHI entry lives in. If the value
|
|
|
|
// really is dead at the PHI copy, there will be no successor blocks which
|
|
|
|
// have the value live-in.
|
|
|
|
|
|
|
|
// Okay, if we now know that the value is not live out of the block, we
|
|
|
|
// can add a kill marker in this block saying that it kills the incoming
|
|
|
|
// value!
|
2005-10-03 07:22:07 +00:00
|
|
|
|
2006-01-04 07:12:21 +00:00
|
|
|
// In our final twist, we have to decide which instruction kills the
|
2012-07-04 19:52:05 +00:00
|
|
|
// register. In most cases this is the copy, however, terminator
|
|
|
|
// instructions at the end of the block may also use the value. In this
|
|
|
|
// case, we should mark the last such terminator as being the killing
|
|
|
|
// block, not the copy.
|
|
|
|
MachineBasicBlock::iterator KillInst = opBlock.end();
|
2021-09-30 15:35:43 +01:00
|
|
|
for (MachineBasicBlock::iterator Term = InsertPos; Term != opBlock.end();
|
|
|
|
++Term) {
|
[CodeGen] Make the parameter TRI required in some functions. (#85968)
Fixes #82659
There are some functions, such as `findRegisterDefOperandIdx` and `findRegisterDefOperand`, that have too many default parameters. As a result, we have encountered some issues due to the lack of TRI parameters, as shown in issue #82411.
Following @RKSimon 's suggestion, this patch refactors 9 functions, including `{reads, kills, defines, modifies}Register`, `registerDefIsDead`, and `findRegister{UseOperandIdx, UseOperand, DefOperandIdx, DefOperand}`, adjusting the order of the TRI parameter and making it required. In addition, all the places that call these functions have also been updated correctly to ensure no additional impact.
After this, the caller of these functions should explicitly know whether to pass the `TargetRegisterInfo` or just a `nullptr`.
2024-04-24 21:24:14 +08:00
|
|
|
if (Term->readsRegister(SrcReg, /*TRI=*/nullptr))
|
2012-07-04 19:52:05 +00:00
|
|
|
KillInst = Term;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (KillInst == opBlock.end()) {
|
|
|
|
// No terminator uses the register.
|
|
|
|
|
|
|
|
if (reusedIncoming || !IncomingReg) {
|
|
|
|
// We may have to rewind a bit if we didn't insert a copy this time.
|
2021-09-30 15:35:43 +01:00
|
|
|
KillInst = InsertPos;
|
2012-07-04 19:52:05 +00:00
|
|
|
while (KillInst != opBlock.begin()) {
|
|
|
|
--KillInst;
|
2018-05-09 02:42:00 +00:00
|
|
|
if (KillInst->isDebugInstr())
|
2012-07-04 19:52:05 +00:00
|
|
|
continue;
|
[CodeGen] Make the parameter TRI required in some functions. (#85968)
Fixes #82659
There are some functions, such as `findRegisterDefOperandIdx` and `findRegisterDefOperand`, that have too many default parameters. As a result, we have encountered some issues due to the lack of TRI parameters, as shown in issue #82411.
Following @RKSimon 's suggestion, this patch refactors 9 functions, including `{reads, kills, defines, modifies}Register`, `registerDefIsDead`, and `findRegister{UseOperandIdx, UseOperand, DefOperandIdx, DefOperand}`, adjusting the order of the TRI parameter and making it required. In addition, all the places that call these functions have also been updated correctly to ensure no additional impact.
After this, the caller of these functions should explicitly know whether to pass the `TargetRegisterInfo` or just a `nullptr`.
2024-04-24 21:24:14 +08:00
|
|
|
if (KillInst->readsRegister(SrcReg, /*TRI=*/nullptr))
|
2012-07-04 19:52:05 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// We just inserted this copy.
|
2019-09-17 09:08:58 +00:00
|
|
|
KillInst = NewSrcInstr;
|
2011-01-14 06:33:45 +00:00
|
|
|
}
|
2006-01-04 07:12:21 +00:00
|
|
|
}
|
[CodeGen] Make the parameter TRI required in some functions. (#85968)
Fixes #82659
There are some functions, such as `findRegisterDefOperandIdx` and `findRegisterDefOperand`, that have too many default parameters. As a result, we have encountered some issues due to the lack of TRI parameters, as shown in issue #82411.
Following @RKSimon 's suggestion, this patch refactors 9 functions, including `{reads, kills, defines, modifies}Register`, `registerDefIsDead`, and `findRegister{UseOperandIdx, UseOperand, DefOperandIdx, DefOperand}`, adjusting the order of the TRI parameter and making it required. In addition, all the places that call these functions have also been updated correctly to ensure no additional impact.
After this, the caller of these functions should explicitly know whether to pass the `TargetRegisterInfo` or just a `nullptr`.
2024-04-24 21:24:14 +08:00
|
|
|
assert(KillInst->readsRegister(SrcReg, /*TRI=*/nullptr) &&
|
|
|
|
"Cannot find kill instruction");
|
2009-11-10 22:00:56 +00:00
|
|
|
|
2006-01-04 07:12:21 +00:00
|
|
|
// Finally, mark it killed.
|
2016-07-01 01:51:32 +00:00
|
|
|
LV->addVirtualRegisterKilled(SrcReg, *KillInst);
|
2005-10-03 07:22:07 +00:00
|
|
|
|
|
|
|
// This vreg no longer lives all of the way through opBlock.
|
|
|
|
unsigned opBlockNum = opBlock.getNumber();
|
2009-11-10 22:00:56 +00:00
|
|
|
LV->getVarInfo(SrcReg).AliveBlocks.reset(opBlockNum);
|
2003-01-13 20:01:16 +00:00
|
|
|
}
|
2013-02-10 06:42:36 +00:00
|
|
|
|
|
|
|
if (LIS) {
|
|
|
|
if (NewSrcInstr) {
|
2016-02-27 06:40:41 +00:00
|
|
|
LIS->InsertMachineInstrInMaps(*NewSrcInstr);
|
|
|
|
LIS->addSegmentToEndOfBlock(IncomingReg, *NewSrcInstr);
|
2013-02-10 06:42:36 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!SrcUndef &&
|
|
|
|
!VRegPHIUseCount[BBVRegPair(opBlock.getNumber(), SrcReg)]) {
|
|
|
|
LiveInterval &SrcLI = LIS->getInterval(SrcReg);
|
|
|
|
|
|
|
|
bool isLiveOut = false;
|
2021-02-19 22:44:14 -08:00
|
|
|
for (MachineBasicBlock *Succ : opBlock.successors()) {
|
|
|
|
SlotIndex startIdx = LIS->getMBBStartIdx(Succ);
|
2013-02-12 05:48:58 +00:00
|
|
|
VNInfo *VNI = SrcLI.getVNInfoAt(startIdx);
|
|
|
|
|
|
|
|
// Definitions by other PHIs are not truly live-in for our purposes.
|
|
|
|
if (VNI && VNI->def != startIdx) {
|
2013-02-10 06:42:36 +00:00
|
|
|
isLiveOut = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!isLiveOut) {
|
|
|
|
MachineBasicBlock::iterator KillInst = opBlock.end();
|
2021-09-30 15:35:43 +01:00
|
|
|
for (MachineBasicBlock::iterator Term = InsertPos;
|
|
|
|
Term != opBlock.end(); ++Term) {
|
[CodeGen] Make the parameter TRI required in some functions. (#85968)
Fixes #82659
There are some functions, such as `findRegisterDefOperandIdx` and `findRegisterDefOperand`, that have too many default parameters. As a result, we have encountered some issues due to the lack of TRI parameters, as shown in issue #82411.
Following @RKSimon 's suggestion, this patch refactors 9 functions, including `{reads, kills, defines, modifies}Register`, `registerDefIsDead`, and `findRegister{UseOperandIdx, UseOperand, DefOperandIdx, DefOperand}`, adjusting the order of the TRI parameter and making it required. In addition, all the places that call these functions have also been updated correctly to ensure no additional impact.
After this, the caller of these functions should explicitly know whether to pass the `TargetRegisterInfo` or just a `nullptr`.
2024-04-24 21:24:14 +08:00
|
|
|
if (Term->readsRegister(SrcReg, /*TRI=*/nullptr))
|
2013-02-10 06:42:36 +00:00
|
|
|
KillInst = Term;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (KillInst == opBlock.end()) {
|
|
|
|
// No terminator uses the register.
|
|
|
|
|
|
|
|
if (reusedIncoming || !IncomingReg) {
|
|
|
|
// We may have to rewind a bit if we didn't just insert a copy.
|
2021-09-30 15:35:43 +01:00
|
|
|
KillInst = InsertPos;
|
2013-02-10 06:42:36 +00:00
|
|
|
while (KillInst != opBlock.begin()) {
|
|
|
|
--KillInst;
|
2018-05-09 02:42:00 +00:00
|
|
|
if (KillInst->isDebugInstr())
|
2013-02-10 06:42:36 +00:00
|
|
|
continue;
|
[CodeGen] Make the parameter TRI required in some functions. (#85968)
Fixes #82659
There are some functions, such as `findRegisterDefOperandIdx` and `findRegisterDefOperand`, that have too many default parameters. As a result, we have encountered some issues due to the lack of TRI parameters, as shown in issue #82411.
Following @RKSimon 's suggestion, this patch refactors 9 functions, including `{reads, kills, defines, modifies}Register`, `registerDefIsDead`, and `findRegister{UseOperandIdx, UseOperand, DefOperandIdx, DefOperand}`, adjusting the order of the TRI parameter and making it required. In addition, all the places that call these functions have also been updated correctly to ensure no additional impact.
After this, the caller of these functions should explicitly know whether to pass the `TargetRegisterInfo` or just a `nullptr`.
2024-04-24 21:24:14 +08:00
|
|
|
if (KillInst->readsRegister(SrcReg, /*TRI=*/nullptr))
|
2013-02-10 06:42:36 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// We just inserted this copy.
|
2014-03-02 12:27:27 +00:00
|
|
|
KillInst = std::prev(InsertPos);
|
2013-02-10 06:42:36 +00:00
|
|
|
}
|
|
|
|
}
|
[CodeGen] Make the parameter TRI required in some functions. (#85968)
Fixes #82659
There are some functions, such as `findRegisterDefOperandIdx` and `findRegisterDefOperand`, that have too many default parameters. As a result, we have encountered some issues due to the lack of TRI parameters, as shown in issue #82411.
Following @RKSimon 's suggestion, this patch refactors 9 functions, including `{reads, kills, defines, modifies}Register`, `registerDefIsDead`, and `findRegister{UseOperandIdx, UseOperand, DefOperandIdx, DefOperand}`, adjusting the order of the TRI parameter and making it required. In addition, all the places that call these functions have also been updated correctly to ensure no additional impact.
After this, the caller of these functions should explicitly know whether to pass the `TargetRegisterInfo` or just a `nullptr`.
2024-04-24 21:24:14 +08:00
|
|
|
assert(KillInst->readsRegister(SrcReg, /*TRI=*/nullptr) &&
|
2013-02-10 06:42:36 +00:00
|
|
|
"Cannot find kill instruction");
|
|
|
|
|
2016-02-27 06:40:41 +00:00
|
|
|
SlotIndex LastUseIndex = LIS->getInstructionIndex(*KillInst);
|
2013-10-10 21:28:43 +00:00
|
|
|
SrcLI.removeSegment(LastUseIndex.getRegSlot(),
|
|
|
|
LIS->getMBBEndIdx(&opBlock));
|
2023-11-13 12:16:26 +09:00
|
|
|
for (auto &SR : SrcLI.subranges()) {
|
|
|
|
SR.removeSegment(LastUseIndex.getRegSlot(),
|
|
|
|
LIS->getMBBEndIdx(&opBlock));
|
|
|
|
}
|
2013-02-10 06:42:36 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2003-01-13 20:01:16 +00:00
|
|
|
}
|
2009-11-10 22:00:56 +00:00
|
|
|
|
2009-12-16 18:55:53 +00:00
|
|
|
// Really delete the PHI instruction now, if it is not in the LoweredPHIs map.
|
[CodeGen] Only deduplicate PHIs on critical edges (#97064)
PHIElim deduplicates identical PHI nodes to reduce the number of copies
inserted. There are two cases:
1. Identical PHI nodes are in different blocks. That's the reason for
this optimization; this can't be avoided at SSA-level. A necessary
prerequisite for this is that the predecessors of all basic blocks
(where such a PHI node could occur) are the same. This implies that
all (>= 2) predecessors must have multiple successors, i.e. all edges
into the block are critical edges.
2. Identical PHI nodes are in the same block. CSE can remove these.
There are a few cases, however, where they still occur regardless:
- expand-large-div-rem creates PHI nodes with large integers, which
get lowered into one PHI per MVT. Later, some identical values
(zeroes) get folded, resulting in identical PHI nodes.
- peephole-opt occasionally inserts PHIs for the same value.
- Some pseudo instruction emitters create redundant PHI nodes (e.g.,
AVR's insertShift), merging the same values more than once.
In any case, this happens rarely and MachineCSE handles most cases
anyway, so that PHIElim only gets to see very few of such cases (see
changed test files).
Currently, all PHI nodes are inserted into a DenseMap that checks
equality not by pointer but by operands. This hash map is pretty
expensive (hashing itself and the hash map), but only really useful in
the first case.
Avoid this expensive hashing most of the time by restricting it to basic
blocks with only critical input edges. This improves performance for
code with many PHI nodes, especially at -O0. (Note that Clang often
doesn't generate PHI nodes and -O0 includes no mem2reg. Other
compilers always generate PHI nodes.)
2024-07-03 11:19:05 +02:00
|
|
|
if (EliminateNow) {
|
2013-02-10 06:42:36 +00:00
|
|
|
if (LIS)
|
2016-02-27 06:40:41 +00:00
|
|
|
LIS->RemoveMachineInstrFromMaps(*MPhi);
|
2021-12-08 20:35:33 -08:00
|
|
|
MF.deleteMachineInstr(MPhi);
|
2013-02-10 06:42:36 +00:00
|
|
|
}
|
2003-01-13 20:01:16 +00:00
|
|
|
}
|
2006-09-28 07:10:24 +00:00
|
|
|
|
|
|
|
/// analyzePHINodes - Gather information about the PHI nodes in here. In
|
|
|
|
/// particular, we want to map the number of uses of a virtual register which is
|
|
|
|
/// used in a PHI node. We map that to the BB the vreg is coming from. This is
|
|
|
|
/// used later to determine when the vreg is killed in the BB.
|
2024-07-17 11:26:56 +08:00
|
|
|
void PHIEliminationImpl::analyzePHINodes(const MachineFunction &MF) {
|
2021-10-11 16:02:40 +01:00
|
|
|
for (const auto &MBB : MF) {
|
2014-04-30 22:17:38 +00:00
|
|
|
for (const auto &BBI : MBB) {
|
|
|
|
if (!BBI.isPHI())
|
|
|
|
break;
|
2021-10-11 16:02:40 +01:00
|
|
|
for (unsigned i = 1, e = BBI.getNumOperands(); i != e; i += 2) {
|
|
|
|
if (!BBI.getOperand(i).isUndef()) {
|
|
|
|
++VRegPHIUseCount[BBVRegPair(
|
|
|
|
BBI.getOperand(i + 1).getMBB()->getNumber(),
|
|
|
|
BBI.getOperand(i).getReg())];
|
|
|
|
}
|
|
|
|
}
|
2014-04-30 22:17:38 +00:00
|
|
|
}
|
2021-10-11 16:02:40 +01:00
|
|
|
}
|
2006-09-28 07:10:24 +00:00
|
|
|
}
|
2009-11-10 22:00:56 +00:00
|
|
|
|
2024-07-17 11:26:56 +08:00
|
|
|
bool PHIEliminationImpl::SplitPHIEdges(
|
|
|
|
MachineFunction &MF, MachineBasicBlock &MBB, MachineLoopInfo *MLI,
|
|
|
|
std::vector<SparseBitVector<>> *LiveInSets) {
|
2015-08-27 23:27:47 +00:00
|
|
|
if (MBB.empty() || !MBB.front().isPHI() || MBB.isEHPad())
|
2024-07-10 17:13:02 +08:00
|
|
|
return false; // Quick exit for basic blocks without PHIs.
|
2009-11-18 18:01:35 +00:00
|
|
|
|
2014-04-14 00:51:57 +00:00
|
|
|
const MachineLoop *CurLoop = MLI ? MLI->getLoopFor(&MBB) : nullptr;
|
2012-07-20 20:49:53 +00:00
|
|
|
bool IsLoopHeader = CurLoop && &MBB == CurLoop->getHeader();
|
|
|
|
|
2010-08-17 01:20:36 +00:00
|
|
|
bool Changed = false;
|
2011-12-06 22:12:01 +00:00
|
|
|
for (MachineBasicBlock::iterator BBI = MBB.begin(), BBE = MBB.end();
|
2010-02-09 19:54:29 +00:00
|
|
|
BBI != BBE && BBI->isPHI(); ++BBI) {
|
2009-11-10 22:01:05 +00:00
|
|
|
for (unsigned i = 1, e = BBI->getNumOperands(); i != e; i += 2) {
|
Apply llvm-prefer-register-over-unsigned from clang-tidy to LLVM
Summary:
This clang-tidy check is looking for unsigned integer variables whose initializer
starts with an implicit cast from llvm::Register and changes the type of the
variable to llvm::Register (dropping the llvm:: where possible).
Partial reverts in:
X86FrameLowering.cpp - Some functions return unsigned and arguably should be MCRegister
X86FixupLEAs.cpp - Some functions return unsigned and arguably should be MCRegister
X86FrameLowering.cpp - Some functions return unsigned and arguably should be MCRegister
HexagonBitSimplify.cpp - Function takes BitTracker::RegisterRef which appears to be unsigned&
MachineVerifier.cpp - Ambiguous operator==() given MCRegister and const Register
PPCFastISel.cpp - No Register::operator-=()
PeepholeOptimizer.cpp - TargetInstrInfo::optimizeLoadInstr() takes an unsigned&
MachineTraceMetrics.cpp - MachineTraceMetrics lacks a suitable constructor
Manual fixups in:
ARMFastISel.cpp - ARMEmitLoad() now takes a Register& instead of unsigned&
HexagonSplitDouble.cpp - Ternary operator was ambiguous between unsigned/Register
HexagonConstExtenders.cpp - Has a local class named Register, used llvm::Register instead of Register.
PPCFastISel.cpp - PPCEmitLoad() now takes a Register& instead of unsigned&
Depends on D65919
Reviewers: arsenm, bogner, craig.topper, RKSimon
Reviewed By: arsenm
Subscribers: RKSimon, craig.topper, lenary, aemerson, wuzish, jholewinski, MatzeB, qcolombet, dschuff, jyknight, dylanmckay, sdardis, nemanjai, jvesely, wdng, nhaehnle, sbc100, jgravelle-google, kristof.beyls, hiraditya, aheejin, kbarton, fedor.sergeev, javed.absar, asb, rbar, johnrusso, simoncook, apazos, sabuasal, niosHD, jrtc27, MaskRay, zzheng, edward-jones, atanasyan, rogfer01, MartinMosbeck, brucehoult, the_o, tpr, PkmX, jocewei, jsji, Petar.Avramovic, asbirlea, Jim, s.egerton, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65962
llvm-svn: 369041
2019-08-15 19:22:08 +00:00
|
|
|
Register Reg = BBI->getOperand(i).getReg();
|
2024-07-10 17:13:02 +08:00
|
|
|
MachineBasicBlock *PreMBB = BBI->getOperand(i + 1).getMBB();
|
2012-07-20 20:49:53 +00:00
|
|
|
// Is there a critical edge from PreMBB to MBB?
|
|
|
|
if (PreMBB->succ_size() == 1)
|
|
|
|
continue;
|
|
|
|
|
2010-08-17 17:43:50 +00:00
|
|
|
// Avoid splitting backedges of loops. It would introduce small
|
|
|
|
// out-of-line blocks into the loop which is very bad for code placement.
|
2013-02-12 03:49:25 +00:00
|
|
|
if (PreMBB == &MBB && !SplitAllCriticalEdges)
|
2012-07-20 20:49:53 +00:00
|
|
|
continue;
|
2014-04-14 00:51:57 +00:00
|
|
|
const MachineLoop *PreLoop = MLI ? MLI->getLoopFor(PreMBB) : nullptr;
|
2013-02-12 03:49:25 +00:00
|
|
|
if (IsLoopHeader && PreLoop == CurLoop && !SplitAllCriticalEdges)
|
2012-07-20 20:49:53 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
// LV doesn't consider a phi use live-out, so isLiveOut only returns true
|
|
|
|
// when the source register is live-out for some other reason than a phi
|
|
|
|
// use. That means the copy we will insert in PreMBB won't be a kill, and
|
|
|
|
// there is a risk it may not be coalesced away.
|
|
|
|
//
|
|
|
|
// If the copy would be a kill, there is no need to split the edge.
|
2015-03-03 10:23:11 +00:00
|
|
|
bool ShouldSplit = isLiveOutPastPHIs(Reg, PreMBB);
|
|
|
|
if (!ShouldSplit && !NoPhiElimLiveOutEarlyExit)
|
2012-07-20 20:49:53 +00:00
|
|
|
continue;
|
2015-03-03 10:23:11 +00:00
|
|
|
if (ShouldSplit) {
|
2018-05-14 12:53:11 +00:00
|
|
|
LLVM_DEBUG(dbgs() << printReg(Reg) << " live-out before critical edge "
|
|
|
|
<< printMBBReference(*PreMBB) << " -> "
|
|
|
|
<< printMBBReference(MBB) << ": " << *BBI);
|
2015-03-03 10:23:11 +00:00
|
|
|
}
|
2012-07-20 20:49:53 +00:00
|
|
|
|
|
|
|
// If Reg is not live-in to MBB, it means it must be live-in to some
|
|
|
|
// other PreMBB successor, and we can avoid the interference by splitting
|
|
|
|
// the edge.
|
|
|
|
//
|
|
|
|
// If Reg *is* live-in to MBB, the interference is inevitable and a copy
|
|
|
|
// is likely to be left after coalescing. If we are looking at a loop
|
|
|
|
// exiting edge, split it so we won't insert code in the loop, otherwise
|
|
|
|
// don't bother.
|
2015-03-03 10:23:11 +00:00
|
|
|
ShouldSplit = ShouldSplit && !isLiveIn(Reg, &MBB);
|
2012-07-20 20:49:53 +00:00
|
|
|
|
|
|
|
// Check for a loop exiting edge.
|
|
|
|
if (!ShouldSplit && CurLoop != PreLoop) {
|
2018-05-14 12:53:11 +00:00
|
|
|
LLVM_DEBUG({
|
2012-07-20 20:49:53 +00:00
|
|
|
dbgs() << "Split wouldn't help, maybe avoid loop copies?\n";
|
2018-05-14 12:53:11 +00:00
|
|
|
if (PreLoop)
|
|
|
|
dbgs() << "PreLoop: " << *PreLoop;
|
|
|
|
if (CurLoop)
|
|
|
|
dbgs() << "CurLoop: " << *CurLoop;
|
2012-07-20 20:49:53 +00:00
|
|
|
});
|
|
|
|
// This edge could be entering a loop, exiting a loop, or it could be
|
|
|
|
// both: Jumping directly form one loop to the header of a sibling
|
|
|
|
// loop.
|
|
|
|
// Split unless this edge is entering CurLoop from an outer loop.
|
|
|
|
ShouldSplit = PreLoop && !PreLoop->contains(CurLoop);
|
|
|
|
}
|
2015-03-03 10:23:11 +00:00
|
|
|
if (!ShouldSplit && !SplitAllCriticalEdges)
|
2012-07-20 20:49:53 +00:00
|
|
|
continue;
|
2024-07-17 11:26:56 +08:00
|
|
|
if (!(P ? PreMBB->SplitCriticalEdge(&MBB, *P, LiveInSets)
|
|
|
|
: PreMBB->SplitCriticalEdge(&MBB, *MFAM, LiveInSets))) {
|
2018-05-14 12:53:11 +00:00
|
|
|
LLVM_DEBUG(dbgs() << "Failed to split critical edge.\n");
|
2012-07-20 20:49:53 +00:00
|
|
|
continue;
|
2010-08-17 17:43:50 +00:00
|
|
|
}
|
2012-07-20 20:49:53 +00:00
|
|
|
Changed = true;
|
|
|
|
++NumCriticalEdgesSplit;
|
2009-11-10 22:01:05 +00:00
|
|
|
}
|
|
|
|
}
|
2011-02-17 06:13:43 +00:00
|
|
|
return Changed;
|
2009-11-10 22:01:05 +00:00
|
|
|
}
|
2013-02-10 23:29:49 +00:00
|
|
|
|
2024-07-17 11:26:56 +08:00
|
|
|
bool PHIEliminationImpl::isLiveIn(Register Reg, const MachineBasicBlock *MBB) {
|
2013-02-10 23:29:49 +00:00
|
|
|
assert((LV || LIS) &&
|
|
|
|
"isLiveIn() requires either LiveVariables or LiveIntervals");
|
|
|
|
if (LIS)
|
|
|
|
return LIS->isLiveInToMBB(LIS->getInterval(Reg), MBB);
|
|
|
|
else
|
|
|
|
return LV->isLiveIn(Reg, *MBB);
|
|
|
|
}
|
|
|
|
|
2024-07-17 11:26:56 +08:00
|
|
|
bool PHIEliminationImpl::isLiveOutPastPHIs(Register Reg,
|
|
|
|
const MachineBasicBlock *MBB) {
|
2013-02-10 23:29:49 +00:00
|
|
|
assert((LV || LIS) &&
|
|
|
|
"isLiveOutPastPHIs() requires either LiveVariables or LiveIntervals");
|
|
|
|
// LiveVariables considers uses in PHIs to be in the predecessor basic block,
|
|
|
|
// so that a register used only in a PHI is not live out of the block. In
|
2024-07-10 17:13:02 +08:00
|
|
|
// contrast, LiveIntervals considers uses in PHIs to be on the edge rather
|
|
|
|
// than in the predecessor basic block, so that a register used only in a PHI
|
|
|
|
// is live out of the block.
|
2013-02-10 23:29:49 +00:00
|
|
|
if (LIS) {
|
|
|
|
const LiveInterval &LI = LIS->getInterval(Reg);
|
2015-06-11 07:45:05 +00:00
|
|
|
for (const MachineBasicBlock *SI : MBB->successors())
|
|
|
|
if (LI.liveAt(LIS->getMBBStartIdx(SI)))
|
2013-02-10 23:29:49 +00:00
|
|
|
return true;
|
|
|
|
return false;
|
|
|
|
} else {
|
|
|
|
return LV->isLiveOut(Reg, *MBB);
|
|
|
|
}
|
|
|
|
}
|