llvm-project/llvm/lib/Analysis/LoopInfo.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

1227 lines
41 KiB
C++
Raw Normal View History

//===- LoopInfo.cpp - Natural Loop Calculator -----------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// This file defines the LoopInfo class that is used to identify natural loops
// and determine the loop depth of various nodes of the CFG. Note that the
// loops identified may actually be several natural loops that share the same
// header node... not just a single natural loop.
//
//===----------------------------------------------------------------------===//
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/ScopeExit.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/Analysis/IVDescriptors.h"
#include "llvm/Analysis/LoopInfoImpl.h"
#include "llvm/Analysis/LoopIterator.h"
#include "llvm/Analysis/LoopNestAnalysis.h"
#include "llvm/Analysis/MemorySSA.h"
#include "llvm/Analysis/MemorySSAUpdater.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Config/llvm-config.h"
#include "llvm/IR/CFG.h"
#include "llvm/IR/Constants.h"
Look for a loop's starting location in the llvm.loop metadata Getting accurate locations for loops is important, because those locations are used by the frontend to generate optimization remarks. Currently, optimization remarks for loops often appear on the wrong line, often the first line of the loop body instead of the loop itself. This is confusing because that line might itself be another loop, or might be somewhere else completely if the body was inlined function call. This happens because of the way we find the loop's starting location. First, we look for a preheader, and if we find one, and its terminator has a debug location, then we use that. Otherwise, we look for a location on an instruction in the loop header. The fallback heuristic is not bad, but will almost always find the beginning of the body, and not the loop statement itself. The preheader location search often fails because there's often not a preheader, and even when there is a preheader, depending on how it was formed, it sometimes carries the location of some preceeding code. I don't see any good theoretical way to fix this problem. On the other hand, this seems like a straightforward solution: Put the debug location in the loop's llvm.loop metadata. A companion Clang patch will cause Clang to insert llvm.loop metadata with appropriate locations when generating debugging information. With these changes, our loop remarks have much more accurate locations. Differential Revision: http://reviews.llvm.org/D19738 llvm-svn: 270771
2016-05-26 05:42:37 +08:00
#include "llvm/IR/DebugLoc.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/IRPrintingPasses.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Metadata.h"
#include "llvm/IR/PassManager.h"
#include "llvm/IR/PrintPasses.h"
#include "llvm/InitializePasses.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include <algorithm>
using namespace llvm;
// Explicitly instantiate methods in LoopInfoImpl.h for IR-level Loops.
template class llvm::LoopBase<BasicBlock, Loop>;
template class llvm::LoopInfoBase<BasicBlock, Loop>;
// Always verify loopinfo if expensive checking is enabled.
#ifdef EXPENSIVE_CHECKS
bool llvm::VerifyLoopInfo = true;
#else
bool llvm::VerifyLoopInfo = false;
#endif
static cl::opt<bool, true>
VerifyLoopInfoX("verify-loop-info", cl::location(VerifyLoopInfo),
cl::Hidden, cl::desc("Verify loop info (time consuming)"));
//===----------------------------------------------------------------------===//
// Loop implementation
//
bool Loop::isLoopInvariant(const Value *V) const {
if (const Instruction *I = dyn_cast<Instruction>(V))
return !contains(I);
return true; // All non-instructions are loop invariant
}
bool Loop::hasLoopInvariantOperands(const Instruction *I) const {
return all_of(I->operands(), [this](Value *V) { return isLoopInvariant(V); });
}
bool Loop::makeLoopInvariant(Value *V, bool &Changed, Instruction *InsertPt,
MemorySSAUpdater *MSSAU) const {
if (Instruction *I = dyn_cast<Instruction>(V))
return makeLoopInvariant(I, Changed, InsertPt, MSSAU);
return true; // All non-instructions are loop-invariant.
}
bool Loop::makeLoopInvariant(Instruction *I, bool &Changed,
Instruction *InsertPt,
MemorySSAUpdater *MSSAU) const {
// Test if the value is already loop-invariant.
if (isLoopInvariant(I))
return true;
if (!isSafeToSpeculativelyExecute(I))
return false;
if (I->mayReadFromMemory())
return false;
// EH block instructions are immobile.
if (I->isEHPad())
return false;
// Determine the insertion point, unless one was given.
if (!InsertPt) {
BasicBlock *Preheader = getLoopPreheader();
// Without a preheader, hoisting is not feasible.
if (!Preheader)
return false;
InsertPt = Preheader->getTerminator();
}
// Don't hoist instructions with loop-variant operands.
2016-01-15 08:08:10 +08:00
for (Value *Operand : I->operands())
if (!makeLoopInvariant(Operand, Changed, InsertPt, MSSAU))
return false;
2011-08-04 07:45:50 +08:00
// Hoist.
I->moveBefore(InsertPt);
if (MSSAU)
if (auto *MUD = MSSAU->getMemorySSA()->getMemoryAccess(I))
MSSAU->moveToPlace(MUD, InsertPt->getParent(),
MemorySSA::BeforeTerminator);
// There is possibility of hoisting this instruction above some arbitrary
// condition. Any metadata defined on it can be control dependent on this
// condition. Conservatively strip it here so that we don't give any wrong
// information to the optimizer.
I->dropUnknownNonDebugMetadata();
Changed = true;
return true;
}
bool Loop::getIncomingAndBackEdge(BasicBlock *&Incoming,
BasicBlock *&Backedge) const {
BasicBlock *H = getHeader();
Incoming = nullptr;
Backedge = nullptr;
pred_iterator PI = pred_begin(H);
assert(PI != pred_end(H) && "Loop must have at least one backedge!");
Backedge = *PI++;
if (PI == pred_end(H))
return false; // dead loop
Incoming = *PI++;
if (PI != pred_end(H))
return false; // multiple backedges?
if (contains(Incoming)) {
if (contains(Backedge))
return false;
std::swap(Incoming, Backedge);
} else if (!contains(Backedge))
return false;
assert(Incoming && Backedge && "expected non-null incoming and backedges");
return true;
}
PHINode *Loop::getCanonicalInductionVariable() const {
BasicBlock *H = getHeader();
BasicBlock *Incoming = nullptr, *Backedge = nullptr;
if (!getIncomingAndBackEdge(Incoming, Backedge))
return nullptr;
// Loop over all of the PHI nodes, looking for a canonical indvar.
for (BasicBlock::iterator I = H->begin(); isa<PHINode>(I); ++I) {
PHINode *PN = cast<PHINode>(I);
if (ConstantInt *CI =
dyn_cast<ConstantInt>(PN->getIncomingValueForBlock(Incoming)))
if (CI->isZero())
if (Instruction *Inc =
dyn_cast<Instruction>(PN->getIncomingValueForBlock(Backedge)))
if (Inc->getOpcode() == Instruction::Add && Inc->getOperand(0) == PN)
if (ConstantInt *CI = dyn_cast<ConstantInt>(Inc->getOperand(1)))
if (CI->isOne())
return PN;
}
return nullptr;
}
/// Get the latch condition instruction.
ICmpInst *Loop::getLatchCmpInst() const {
if (BasicBlock *Latch = getLoopLatch())
if (BranchInst *BI = dyn_cast_or_null<BranchInst>(Latch->getTerminator()))
if (BI->isConditional())
return dyn_cast<ICmpInst>(BI->getCondition());
return nullptr;
}
/// Return the final value of the loop induction variable if found.
static Value *findFinalIVValue(const Loop &L, const PHINode &IndVar,
const Instruction &StepInst) {
ICmpInst *LatchCmpInst = L.getLatchCmpInst();
if (!LatchCmpInst)
return nullptr;
Value *Op0 = LatchCmpInst->getOperand(0);
Value *Op1 = LatchCmpInst->getOperand(1);
if (Op0 == &IndVar || Op0 == &StepInst)
return Op1;
if (Op1 == &IndVar || Op1 == &StepInst)
return Op0;
return nullptr;
}
Optional<Loop::LoopBounds> Loop::LoopBounds::getBounds(const Loop &L,
PHINode &IndVar,
ScalarEvolution &SE) {
InductionDescriptor IndDesc;
if (!InductionDescriptor::isInductionPHI(&IndVar, &L, &SE, IndDesc))
return None;
Value *InitialIVValue = IndDesc.getStartValue();
Instruction *StepInst = IndDesc.getInductionBinOp();
if (!InitialIVValue || !StepInst)
return None;
const SCEV *Step = IndDesc.getStep();
Value *StepInstOp1 = StepInst->getOperand(1);
Value *StepInstOp0 = StepInst->getOperand(0);
Value *StepValue = nullptr;
if (SE.getSCEV(StepInstOp1) == Step)
StepValue = StepInstOp1;
else if (SE.getSCEV(StepInstOp0) == Step)
StepValue = StepInstOp0;
Value *FinalIVValue = findFinalIVValue(L, IndVar, *StepInst);
if (!FinalIVValue)
return None;
return LoopBounds(L, *InitialIVValue, *StepInst, StepValue, *FinalIVValue,
SE);
}
using Direction = Loop::LoopBounds::Direction;
ICmpInst::Predicate Loop::LoopBounds::getCanonicalPredicate() const {
BasicBlock *Latch = L.getLoopLatch();
assert(Latch && "Expecting valid latch");
BranchInst *BI = dyn_cast_or_null<BranchInst>(Latch->getTerminator());
assert(BI && BI->isConditional() && "Expecting conditional latch branch");
ICmpInst *LatchCmpInst = dyn_cast<ICmpInst>(BI->getCondition());
assert(LatchCmpInst &&
"Expecting the latch compare instruction to be a CmpInst");
// Need to inverse the predicate when first successor is not the loop
// header
ICmpInst::Predicate Pred = (BI->getSuccessor(0) == L.getHeader())
? LatchCmpInst->getPredicate()
: LatchCmpInst->getInversePredicate();
if (LatchCmpInst->getOperand(0) == &getFinalIVValue())
Pred = ICmpInst::getSwappedPredicate(Pred);
// Need to flip strictness of the predicate when the latch compare instruction
// is not using StepInst
if (LatchCmpInst->getOperand(0) == &getStepInst() ||
LatchCmpInst->getOperand(1) == &getStepInst())
return Pred;
// Cannot flip strictness of NE and EQ
if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
return ICmpInst::getFlippedStrictnessPredicate(Pred);
Direction D = getDirection();
if (D == Direction::Increasing)
return ICmpInst::ICMP_SLT;
if (D == Direction::Decreasing)
return ICmpInst::ICMP_SGT;
// If cannot determine the direction, then unable to find the canonical
// predicate
return ICmpInst::BAD_ICMP_PREDICATE;
}
Direction Loop::LoopBounds::getDirection() const {
if (const SCEVAddRecExpr *StepAddRecExpr =
dyn_cast<SCEVAddRecExpr>(SE.getSCEV(&getStepInst())))
if (const SCEV *StepRecur = StepAddRecExpr->getStepRecurrence(SE)) {
if (SE.isKnownPositive(StepRecur))
return Direction::Increasing;
if (SE.isKnownNegative(StepRecur))
return Direction::Decreasing;
}
return Direction::Unknown;
}
Optional<Loop::LoopBounds> Loop::getBounds(ScalarEvolution &SE) const {
if (PHINode *IndVar = getInductionVariable(SE))
return LoopBounds::getBounds(*this, *IndVar, SE);
return None;
}
PHINode *Loop::getInductionVariable(ScalarEvolution &SE) const {
if (!isLoopSimplifyForm())
return nullptr;
BasicBlock *Header = getHeader();
assert(Header && "Expected a valid loop header");
ICmpInst *CmpInst = getLatchCmpInst();
if (!CmpInst)
return nullptr;
Instruction *LatchCmpOp0 = dyn_cast<Instruction>(CmpInst->getOperand(0));
Instruction *LatchCmpOp1 = dyn_cast<Instruction>(CmpInst->getOperand(1));
for (PHINode &IndVar : Header->phis()) {
InductionDescriptor IndDesc;
if (!InductionDescriptor::isInductionPHI(&IndVar, this, &SE, IndDesc))
continue;
Instruction *StepInst = IndDesc.getInductionBinOp();
// case 1:
// IndVar = phi[{InitialValue, preheader}, {StepInst, latch}]
// StepInst = IndVar + step
// cmp = StepInst < FinalValue
if (StepInst == LatchCmpOp0 || StepInst == LatchCmpOp1)
return &IndVar;
// case 2:
// IndVar = phi[{InitialValue, preheader}, {StepInst, latch}]
// StepInst = IndVar + step
// cmp = IndVar < FinalValue
if (&IndVar == LatchCmpOp0 || &IndVar == LatchCmpOp1)
return &IndVar;
}
return nullptr;
}
bool Loop::getInductionDescriptor(ScalarEvolution &SE,
InductionDescriptor &IndDesc) const {
if (PHINode *IndVar = getInductionVariable(SE))
return InductionDescriptor::isInductionPHI(IndVar, this, &SE, IndDesc);
return false;
}
bool Loop::isAuxiliaryInductionVariable(PHINode &AuxIndVar,
ScalarEvolution &SE) const {
// Located in the loop header
BasicBlock *Header = getHeader();
if (AuxIndVar.getParent() != Header)
return false;
// No uses outside of the loop
for (User *U : AuxIndVar.users())
if (const Instruction *I = dyn_cast<Instruction>(U))
if (!contains(I))
return false;
InductionDescriptor IndDesc;
if (!InductionDescriptor::isInductionPHI(&AuxIndVar, this, &SE, IndDesc))
return false;
// The step instruction opcode should be add or sub.
if (IndDesc.getInductionOpcode() != Instruction::Add &&
IndDesc.getInductionOpcode() != Instruction::Sub)
return false;
// Incremented by a loop invariant step for each loop iteration
return SE.isLoopInvariant(IndDesc.getStep(), this);
}
BranchInst *Loop::getLoopGuardBranch() const {
if (!isLoopSimplifyForm())
return nullptr;
BasicBlock *Preheader = getLoopPreheader();
assert(Preheader && getLoopLatch() &&
"Expecting a loop with valid preheader and latch");
// Loop should be in rotate form.
if (!isRotatedForm())
return nullptr;
// Disallow loops with more than one unique exit block, as we do not verify
// that GuardOtherSucc post dominates all exit blocks.
BasicBlock *ExitFromLatch = getUniqueExitBlock();
if (!ExitFromLatch)
return nullptr;
BasicBlock *GuardBB = Preheader->getUniquePredecessor();
if (!GuardBB)
return nullptr;
assert(GuardBB->getTerminator() && "Expecting valid guard terminator");
BranchInst *GuardBI = dyn_cast<BranchInst>(GuardBB->getTerminator());
if (!GuardBI || GuardBI->isUnconditional())
return nullptr;
BasicBlock *GuardOtherSucc = (GuardBI->getSuccessor(0) == Preheader)
? GuardBI->getSuccessor(1)
: GuardBI->getSuccessor(0);
// Check if ExitFromLatch (or any BasicBlock which is an empty unique
// successor of ExitFromLatch) is equal to GuardOtherSucc. If
// skipEmptyBlockUntil returns GuardOtherSucc, then the guard branch for the
// loop is GuardBI (return GuardBI), otherwise return nullptr.
if (&LoopNest::skipEmptyBlockUntil(ExitFromLatch, GuardOtherSucc,
/*CheckUniquePred=*/true) ==
GuardOtherSucc)
return GuardBI;
else
return nullptr;
}
bool Loop::isCanonical(ScalarEvolution &SE) const {
InductionDescriptor IndDesc;
if (!getInductionDescriptor(SE, IndDesc))
return false;
ConstantInt *Init = dyn_cast_or_null<ConstantInt>(IndDesc.getStartValue());
if (!Init || !Init->isZero())
return false;
if (IndDesc.getInductionOpcode() != Instruction::Add)
return false;
ConstantInt *Step = IndDesc.getConstIntStepValue();
if (!Step || !Step->isOne())
return false;
return true;
}
// Check that 'BB' doesn't have any uses outside of the 'L'
static bool isBlockInLCSSAForm(const Loop &L, const BasicBlock &BB,
const DominatorTree &DT) {
for (const Instruction &I : BB) {
// Tokens can't be used in PHI nodes and live-out tokens prevent loop
// optimizations, so for the purposes of considered LCSSA form, we
// can ignore them.
if (I.getType()->isTokenTy())
continue;
for (const Use &U : I.uses()) {
const Instruction *UI = cast<Instruction>(U.getUser());
const BasicBlock *UserBB = UI->getParent();
// For practical purposes, we consider that the use in a PHI
// occurs in the respective predecessor block. For more info,
// see the `phi` doc in LangRef and the LCSSA doc.
if (const PHINode *P = dyn_cast<PHINode>(UI))
UserBB = P->getIncomingBlock(U);
// Check the current block, as a fast-path, before checking whether
// the use is anywhere in the loop. Most values are used in the same
// block they are defined in. Also, blocks not reachable from the
// entry are special; uses in them don't need to go through PHIs.
if (UserBB != &BB && !L.contains(UserBB) &&
DT.isReachableFromEntry(UserBB))
return false;
}
}
return true;
}
bool Loop::isLCSSAForm(const DominatorTree &DT) const {
// For each block we check that it doesn't have any uses outside of this loop.
return all_of(this->blocks(), [&](const BasicBlock *BB) {
return isBlockInLCSSAForm(*this, *BB, DT);
});
}
bool Loop::isRecursivelyLCSSAForm(const DominatorTree &DT,
const LoopInfo &LI) const {
// For each block we check that it doesn't have any uses outside of its
// innermost loop. This process will transitively guarantee that the current
// loop and all of the nested loops are in LCSSA form.
return all_of(this->blocks(), [&](const BasicBlock *BB) {
return isBlockInLCSSAForm(*LI.getLoopFor(BB), *BB, DT);
});
}
bool Loop::isLoopSimplifyForm() const {
// Normal-form loops have a preheader, a single backedge, and all of their
// exits have all their predecessors inside the loop.
return getLoopPreheader() && getLoopLatch() && hasDedicatedExits();
}
// Routines that reform the loop CFG and split edges often fail on indirectbr.
bool Loop::isSafeToClone() const {
// Return false if any loop blocks contain indirectbrs, or there are any calls
// to noduplicate functions.
// FIXME: it should be ok to clone CallBrInst's if we correctly update the
// operand list to reflect the newly cloned labels.
2016-01-15 08:08:10 +08:00
for (BasicBlock *BB : this->blocks()) {
if (isa<IndirectBrInst>(BB->getTerminator()) ||
isa<CallBrInst>(BB->getTerminator()))
return false;
for (Instruction &I : *BB)
if (auto *CB = dyn_cast<CallBase>(&I))
if (CB->cannotDuplicate())
return false;
}
return true;
}
MDNode *Loop::getLoopID() const {
MDNode *LoopID = nullptr;
// Go through the latch blocks and check the terminator for the metadata.
SmallVector<BasicBlock *, 4> LatchesBlocks;
getLoopLatches(LatchesBlocks);
for (BasicBlock *BB : LatchesBlocks) {
Instruction *TI = BB->getTerminator();
MDNode *MD = TI->getMetadata(LLVMContext::MD_loop);
if (!MD)
return nullptr;
if (!LoopID)
LoopID = MD;
else if (MD != LoopID)
return nullptr;
}
if (!LoopID || LoopID->getNumOperands() == 0 ||
LoopID->getOperand(0) != LoopID)
return nullptr;
return LoopID;
}
void Loop::setLoopID(MDNode *LoopID) const {
[Unroll/UnrollAndJam/Vectorizer/Distribute] Add followup loop attributes. When multiple loop transformation are defined in a loop's metadata, their order of execution is defined by the order of their respective passes in the pass pipeline. For instance, e.g. #pragma clang loop unroll_and_jam(enable) #pragma clang loop distribute(enable) is the same as #pragma clang loop distribute(enable) #pragma clang loop unroll_and_jam(enable) and will try to loop-distribute before Unroll-And-Jam because the LoopDistribute pass is scheduled after UnrollAndJam pass. UnrollAndJamPass only supports one inner loop, i.e. it will necessarily fail after loop distribution. It is not possible to specify another execution order. Also,t the order of passes in the pipeline is subject to change between versions of LLVM, optimization options and which pass manager is used. This patch adds 'followup' attributes to various loop transformation passes. These attributes define which attributes the resulting loop of a transformation should have. For instance, !0 = !{!0, !1, !2} !1 = !{!"llvm.loop.unroll_and_jam.enable"} !2 = !{!"llvm.loop.unroll_and_jam.followup_inner", !3} !3 = !{!"llvm.loop.distribute.enable"} defines a loop ID (!0) to be unrolled-and-jammed (!1) and then the attribute !3 to be added to the jammed inner loop, which contains the instruction to distribute the inner loop. Currently, in both pass managers, pass execution is in a fixed order and UnrollAndJamPass will not execute again after LoopDistribute. We hope to fix this in the future by allowing pass managers to run passes until a fixpoint is reached, use Polly to perform these transformations, or add a loop transformation pass which takes the order issue into account. For mandatory/forced transformations (e.g. by having been declared by #pragma omp simd), the user must be notified when a transformation could not be performed. It is not possible that the responsible pass emits such a warning because the transformation might be 'hidden' in a followup attribute when it is executed, or it is not present in the pipeline at all. For this reason, this patche introduces a WarnMissedTransformations pass, to warn about orphaned transformations. Since this changes the user-visible diagnostic message when a transformation is applied, two test cases in the clang repository need to be updated. To ensure that no other transformation is executed before the intended one, the attribute `llvm.loop.disable_nonforced` can be added which should disable transformation heuristics before the intended transformation is applied. E.g. it would be surprising if a loop is distributed before a #pragma unroll_and_jam is applied. With more supported code transformations (loop fusion, interchange, stripmining, offloading, etc.), transformations can be used as building blocks for more complex transformations (e.g. stripmining+stripmining+interchange -> tiling). Reviewed By: hfinkel, dmgreen Differential Revision: https://reviews.llvm.org/D49281 Differential Revision: https://reviews.llvm.org/D55288 llvm-svn: 348944
2018-12-13 01:32:52 +08:00
assert((!LoopID || LoopID->getNumOperands() > 0) &&
"Loop ID needs at least one operand");
assert((!LoopID || LoopID->getOperand(0) == LoopID) &&
"Loop ID should refer to itself");
SmallVector<BasicBlock *, 4> LoopLatches;
getLoopLatches(LoopLatches);
for (BasicBlock *BB : LoopLatches)
BB->getTerminator()->setMetadata(LLVMContext::MD_loop, LoopID);
}
void Loop::setLoopAlreadyUnrolled() {
LLVMContext &Context = getHeader()->getContext();
MDNode *DisableUnrollMD =
MDNode::get(Context, MDString::get(Context, "llvm.loop.unroll.disable"));
MDNode *LoopID = getLoopID();
MDNode *NewLoopID = makePostTransformationMetadata(
Context, LoopID, {"llvm.loop.unroll."}, {DisableUnrollMD});
setLoopID(NewLoopID);
}
void Loop::setLoopMustProgress() {
LLVMContext &Context = getHeader()->getContext();
MDNode *MustProgress = findOptionMDForLoop(this, "llvm.loop.mustprogress");
if (MustProgress)
return;
MDNode *MustProgressMD =
MDNode::get(Context, MDString::get(Context, "llvm.loop.mustprogress"));
MDNode *LoopID = getLoopID();
MDNode *NewLoopID =
makePostTransformationMetadata(Context, LoopID, {}, {MustProgressMD});
setLoopID(NewLoopID);
}
bool Loop::isAnnotatedParallel() const {
MDNode *DesiredLoopIdMetadata = getLoopID();
if (!DesiredLoopIdMetadata)
return false;
Introduce llvm.loop.parallel_accesses and llvm.access.group metadata. The current llvm.mem.parallel_loop_access metadata has a problem in that it uses LoopIDs. LoopID unfortunately is not loop identifier. It is neither unique (there's even a regression test assigning the some LoopID to multiple loops; can otherwise happen if passes such as LoopVersioning make copies of entire loops) nor persistent (every time a property is removed/added from a LoopID's MDNode, it will also receive a new LoopID; this happens e.g. when calling Loop::setLoopAlreadyUnrolled()). Since most loop transformation passes change the loop attributes (even if it just to mark that a loop should not be processed again as llvm.loop.isvectorized does, for the versioned and unversioned loop), the parallel access information is lost for any subsequent pass. This patch unlinks LoopIDs and parallel accesses. llvm.mem.parallel_loop_access metadata on instruction is replaced by llvm.access.group metadata. llvm.access.group points to a distinct MDNode with no operands (avoiding the problem to ever need to add/remove operands), called "access group". Alternatively, it can point to a list of access groups. The LoopID then has an attribute llvm.loop.parallel_accesses with all the access groups that are parallel (no dependencies carries by this loop). This intentionally avoid any kind of "ID". Loops that are clones/have their attributes modifies retain the llvm.loop.parallel_accesses attribute. Access instructions that a cloned point to the same access group. It is not necessary for each access to have it's own "ID" MDNode, but those memory access instructions with the same behavior can be grouped together. The behavior of llvm.mem.parallel_loop_access is not changed by this patch, but should be considered deprecated. Differential Revision: https://reviews.llvm.org/D52116 llvm-svn: 349725
2018-12-20 12:58:07 +08:00
MDNode *ParallelAccesses =
findOptionMDForLoop(this, "llvm.loop.parallel_accesses");
SmallPtrSet<MDNode *, 4>
ParallelAccessGroups; // For scalable 'contains' check.
if (ParallelAccesses) {
for (const MDOperand &MD : drop_begin(ParallelAccesses->operands())) {
Introduce llvm.loop.parallel_accesses and llvm.access.group metadata. The current llvm.mem.parallel_loop_access metadata has a problem in that it uses LoopIDs. LoopID unfortunately is not loop identifier. It is neither unique (there's even a regression test assigning the some LoopID to multiple loops; can otherwise happen if passes such as LoopVersioning make copies of entire loops) nor persistent (every time a property is removed/added from a LoopID's MDNode, it will also receive a new LoopID; this happens e.g. when calling Loop::setLoopAlreadyUnrolled()). Since most loop transformation passes change the loop attributes (even if it just to mark that a loop should not be processed again as llvm.loop.isvectorized does, for the versioned and unversioned loop), the parallel access information is lost for any subsequent pass. This patch unlinks LoopIDs and parallel accesses. llvm.mem.parallel_loop_access metadata on instruction is replaced by llvm.access.group metadata. llvm.access.group points to a distinct MDNode with no operands (avoiding the problem to ever need to add/remove operands), called "access group". Alternatively, it can point to a list of access groups. The LoopID then has an attribute llvm.loop.parallel_accesses with all the access groups that are parallel (no dependencies carries by this loop). This intentionally avoid any kind of "ID". Loops that are clones/have their attributes modifies retain the llvm.loop.parallel_accesses attribute. Access instructions that a cloned point to the same access group. It is not necessary for each access to have it's own "ID" MDNode, but those memory access instructions with the same behavior can be grouped together. The behavior of llvm.mem.parallel_loop_access is not changed by this patch, but should be considered deprecated. Differential Revision: https://reviews.llvm.org/D52116 llvm-svn: 349725
2018-12-20 12:58:07 +08:00
MDNode *AccGroup = cast<MDNode>(MD.get());
assert(isValidAsAccessGroup(AccGroup) &&
"List item must be an access group");
ParallelAccessGroups.insert(AccGroup);
}
}
// The loop branch contains the parallel loop metadata. In order to ensure
// that any parallel-loop-unaware optimization pass hasn't added loop-carried
// dependencies (thus converted the loop back to a sequential loop), check
Introduce llvm.loop.parallel_accesses and llvm.access.group metadata. The current llvm.mem.parallel_loop_access metadata has a problem in that it uses LoopIDs. LoopID unfortunately is not loop identifier. It is neither unique (there's even a regression test assigning the some LoopID to multiple loops; can otherwise happen if passes such as LoopVersioning make copies of entire loops) nor persistent (every time a property is removed/added from a LoopID's MDNode, it will also receive a new LoopID; this happens e.g. when calling Loop::setLoopAlreadyUnrolled()). Since most loop transformation passes change the loop attributes (even if it just to mark that a loop should not be processed again as llvm.loop.isvectorized does, for the versioned and unversioned loop), the parallel access information is lost for any subsequent pass. This patch unlinks LoopIDs and parallel accesses. llvm.mem.parallel_loop_access metadata on instruction is replaced by llvm.access.group metadata. llvm.access.group points to a distinct MDNode with no operands (avoiding the problem to ever need to add/remove operands), called "access group". Alternatively, it can point to a list of access groups. The LoopID then has an attribute llvm.loop.parallel_accesses with all the access groups that are parallel (no dependencies carries by this loop). This intentionally avoid any kind of "ID". Loops that are clones/have their attributes modifies retain the llvm.loop.parallel_accesses attribute. Access instructions that a cloned point to the same access group. It is not necessary for each access to have it's own "ID" MDNode, but those memory access instructions with the same behavior can be grouped together. The behavior of llvm.mem.parallel_loop_access is not changed by this patch, but should be considered deprecated. Differential Revision: https://reviews.llvm.org/D52116 llvm-svn: 349725
2018-12-20 12:58:07 +08:00
// that all the memory instructions in the loop belong to an access group that
// is parallel to this loop.
2016-01-15 08:08:10 +08:00
for (BasicBlock *BB : this->blocks()) {
for (Instruction &I : *BB) {
if (!I.mayReadOrWriteMemory())
continue;
Introduce llvm.loop.parallel_accesses and llvm.access.group metadata. The current llvm.mem.parallel_loop_access metadata has a problem in that it uses LoopIDs. LoopID unfortunately is not loop identifier. It is neither unique (there's even a regression test assigning the some LoopID to multiple loops; can otherwise happen if passes such as LoopVersioning make copies of entire loops) nor persistent (every time a property is removed/added from a LoopID's MDNode, it will also receive a new LoopID; this happens e.g. when calling Loop::setLoopAlreadyUnrolled()). Since most loop transformation passes change the loop attributes (even if it just to mark that a loop should not be processed again as llvm.loop.isvectorized does, for the versioned and unversioned loop), the parallel access information is lost for any subsequent pass. This patch unlinks LoopIDs and parallel accesses. llvm.mem.parallel_loop_access metadata on instruction is replaced by llvm.access.group metadata. llvm.access.group points to a distinct MDNode with no operands (avoiding the problem to ever need to add/remove operands), called "access group". Alternatively, it can point to a list of access groups. The LoopID then has an attribute llvm.loop.parallel_accesses with all the access groups that are parallel (no dependencies carries by this loop). This intentionally avoid any kind of "ID". Loops that are clones/have their attributes modifies retain the llvm.loop.parallel_accesses attribute. Access instructions that a cloned point to the same access group. It is not necessary for each access to have it's own "ID" MDNode, but those memory access instructions with the same behavior can be grouped together. The behavior of llvm.mem.parallel_loop_access is not changed by this patch, but should be considered deprecated. Differential Revision: https://reviews.llvm.org/D52116 llvm-svn: 349725
2018-12-20 12:58:07 +08:00
if (MDNode *AccessGroup = I.getMetadata(LLVMContext::MD_access_group)) {
auto ContainsAccessGroup = [&ParallelAccessGroups](MDNode *AG) -> bool {
if (AG->getNumOperands() == 0) {
assert(isValidAsAccessGroup(AG) && "Item must be an access group");
return ParallelAccessGroups.count(AG);
}
for (const MDOperand &AccessListItem : AG->operands()) {
MDNode *AccGroup = cast<MDNode>(AccessListItem.get());
assert(isValidAsAccessGroup(AccGroup) &&
"List item must be an access group");
if (ParallelAccessGroups.count(AccGroup))
return true;
}
return false;
};
if (ContainsAccessGroup(AccessGroup))
continue;
}
// The memory instruction can refer to the loop identifier metadata
// directly or indirectly through another list metadata (in case of
// nested parallel loops). The loop identifier metadata refers to
// itself so we can check both cases with the same routine.
MDNode *LoopIdMD =
2016-01-15 08:08:10 +08:00
I.getMetadata(LLVMContext::MD_mem_parallel_loop_access);
if (!LoopIdMD)
return false;
2021-02-15 00:36:20 +08:00
if (!llvm::is_contained(LoopIdMD->operands(), DesiredLoopIdMetadata))
return false;
}
}
return true;
}
DebugLoc Loop::getStartLoc() const { return getLocRange().getStart(); }
Loop::LocRange Loop::getLocRange() const {
Look for a loop's starting location in the llvm.loop metadata Getting accurate locations for loops is important, because those locations are used by the frontend to generate optimization remarks. Currently, optimization remarks for loops often appear on the wrong line, often the first line of the loop body instead of the loop itself. This is confusing because that line might itself be another loop, or might be somewhere else completely if the body was inlined function call. This happens because of the way we find the loop's starting location. First, we look for a preheader, and if we find one, and its terminator has a debug location, then we use that. Otherwise, we look for a location on an instruction in the loop header. The fallback heuristic is not bad, but will almost always find the beginning of the body, and not the loop statement itself. The preheader location search often fails because there's often not a preheader, and even when there is a preheader, depending on how it was formed, it sometimes carries the location of some preceeding code. I don't see any good theoretical way to fix this problem. On the other hand, this seems like a straightforward solution: Put the debug location in the loop's llvm.loop metadata. A companion Clang patch will cause Clang to insert llvm.loop metadata with appropriate locations when generating debugging information. With these changes, our loop remarks have much more accurate locations. Differential Revision: http://reviews.llvm.org/D19738 llvm-svn: 270771
2016-05-26 05:42:37 +08:00
// If we have a debug location in the loop ID, then use it.
if (MDNode *LoopID = getLoopID()) {
DebugLoc Start;
// We use the first DebugLoc in the header as the start location of the loop
// and if there is a second DebugLoc in the header we use it as end location
// of the loop.
for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
if (DILocation *L = dyn_cast<DILocation>(LoopID->getOperand(i))) {
if (!Start)
Start = DebugLoc(L);
else
return LocRange(Start, DebugLoc(L));
}
}
if (Start)
return LocRange(Start);
}
Look for a loop's starting location in the llvm.loop metadata Getting accurate locations for loops is important, because those locations are used by the frontend to generate optimization remarks. Currently, optimization remarks for loops often appear on the wrong line, often the first line of the loop body instead of the loop itself. This is confusing because that line might itself be another loop, or might be somewhere else completely if the body was inlined function call. This happens because of the way we find the loop's starting location. First, we look for a preheader, and if we find one, and its terminator has a debug location, then we use that. Otherwise, we look for a location on an instruction in the loop header. The fallback heuristic is not bad, but will almost always find the beginning of the body, and not the loop statement itself. The preheader location search often fails because there's often not a preheader, and even when there is a preheader, depending on how it was formed, it sometimes carries the location of some preceeding code. I don't see any good theoretical way to fix this problem. On the other hand, this seems like a straightforward solution: Put the debug location in the loop's llvm.loop metadata. A companion Clang patch will cause Clang to insert llvm.loop metadata with appropriate locations when generating debugging information. With these changes, our loop remarks have much more accurate locations. Differential Revision: http://reviews.llvm.org/D19738 llvm-svn: 270771
2016-05-26 05:42:37 +08:00
// Try the pre-header first.
if (BasicBlock *PHeadBB = getLoopPreheader())
if (DebugLoc DL = PHeadBB->getTerminator()->getDebugLoc())
return LocRange(DL);
Look for a loop's starting location in the llvm.loop metadata Getting accurate locations for loops is important, because those locations are used by the frontend to generate optimization remarks. Currently, optimization remarks for loops often appear on the wrong line, often the first line of the loop body instead of the loop itself. This is confusing because that line might itself be another loop, or might be somewhere else completely if the body was inlined function call. This happens because of the way we find the loop's starting location. First, we look for a preheader, and if we find one, and its terminator has a debug location, then we use that. Otherwise, we look for a location on an instruction in the loop header. The fallback heuristic is not bad, but will almost always find the beginning of the body, and not the loop statement itself. The preheader location search often fails because there's often not a preheader, and even when there is a preheader, depending on how it was formed, it sometimes carries the location of some preceeding code. I don't see any good theoretical way to fix this problem. On the other hand, this seems like a straightforward solution: Put the debug location in the loop's llvm.loop metadata. A companion Clang patch will cause Clang to insert llvm.loop metadata with appropriate locations when generating debugging information. With these changes, our loop remarks have much more accurate locations. Differential Revision: http://reviews.llvm.org/D19738 llvm-svn: 270771
2016-05-26 05:42:37 +08:00
// If we have no pre-header or there are no instructions with debug
// info in it, try the header.
if (BasicBlock *HeadBB = getHeader())
return LocRange(HeadBB->getTerminator()->getDebugLoc());
Look for a loop's starting location in the llvm.loop metadata Getting accurate locations for loops is important, because those locations are used by the frontend to generate optimization remarks. Currently, optimization remarks for loops often appear on the wrong line, often the first line of the loop body instead of the loop itself. This is confusing because that line might itself be another loop, or might be somewhere else completely if the body was inlined function call. This happens because of the way we find the loop's starting location. First, we look for a preheader, and if we find one, and its terminator has a debug location, then we use that. Otherwise, we look for a location on an instruction in the loop header. The fallback heuristic is not bad, but will almost always find the beginning of the body, and not the loop statement itself. The preheader location search often fails because there's often not a preheader, and even when there is a preheader, depending on how it was formed, it sometimes carries the location of some preceeding code. I don't see any good theoretical way to fix this problem. On the other hand, this seems like a straightforward solution: Put the debug location in the loop's llvm.loop metadata. A companion Clang patch will cause Clang to insert llvm.loop metadata with appropriate locations when generating debugging information. With these changes, our loop remarks have much more accurate locations. Differential Revision: http://reviews.llvm.org/D19738 llvm-svn: 270771
2016-05-26 05:42:37 +08:00
return LocRange();
Look for a loop's starting location in the llvm.loop metadata Getting accurate locations for loops is important, because those locations are used by the frontend to generate optimization remarks. Currently, optimization remarks for loops often appear on the wrong line, often the first line of the loop body instead of the loop itself. This is confusing because that line might itself be another loop, or might be somewhere else completely if the body was inlined function call. This happens because of the way we find the loop's starting location. First, we look for a preheader, and if we find one, and its terminator has a debug location, then we use that. Otherwise, we look for a location on an instruction in the loop header. The fallback heuristic is not bad, but will almost always find the beginning of the body, and not the loop statement itself. The preheader location search often fails because there's often not a preheader, and even when there is a preheader, depending on how it was formed, it sometimes carries the location of some preceeding code. I don't see any good theoretical way to fix this problem. On the other hand, this seems like a straightforward solution: Put the debug location in the loop's llvm.loop metadata. A companion Clang patch will cause Clang to insert llvm.loop metadata with appropriate locations when generating debugging information. With these changes, our loop remarks have much more accurate locations. Differential Revision: http://reviews.llvm.org/D19738 llvm-svn: 270771
2016-05-26 05:42:37 +08:00
}
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
LLVM_DUMP_METHOD void Loop::dump() const { print(dbgs()); }
LLVM_DUMP_METHOD void Loop::dumpVerbose() const {
print(dbgs(), /*Verbose=*/true);
}
#endif
//===----------------------------------------------------------------------===//
// UnloopUpdater implementation
//
namespace {
/// Find the new parent loop for all blocks within the "unloop" whose last
/// backedges has just been removed.
class UnloopUpdater {
Loop &Unloop;
LoopInfo *LI;
LoopBlocksDFS DFS;
// Map unloop's immediate subloops to their nearest reachable parents. Nested
// loops within these subloops will not change parents. However, an immediate
// subloop's new parent will be the nearest loop reachable from either its own
// exits *or* any of its nested loop's exits.
DenseMap<Loop *, Loop *> SubloopParents;
// Flag the presence of an irreducible backedge whose destination is a block
// directly contained by the original unloop.
bool FoundIB;
public:
UnloopUpdater(Loop *UL, LoopInfo *LInfo)
: Unloop(*UL), LI(LInfo), DFS(UL), FoundIB(false) {}
void updateBlockParents();
void removeBlocksFromAncestors();
void updateSubloopParents();
protected:
Loop *getNearestLoop(BasicBlock *BB, Loop *BBLoop);
};
} // end anonymous namespace
/// Update the parent loop for all blocks that are directly contained within the
/// original "unloop".
void UnloopUpdater::updateBlockParents() {
if (Unloop.getNumBlocks()) {
// Perform a post order CFG traversal of all blocks within this loop,
// propagating the nearest loop from successors to predecessors.
LoopBlocksTraversal Traversal(DFS, LI);
for (BasicBlock *POI : Traversal) {
Loop *L = LI->getLoopFor(POI);
Loop *NL = getNearestLoop(POI, L);
if (NL != L) {
// For reducible loops, NL is now an ancestor of Unloop.
assert((NL != &Unloop && (!NL || NL->contains(&Unloop))) &&
"uninitialized successor");
LI->changeLoopFor(POI, NL);
} else {
// Or the current block is part of a subloop, in which case its parent
// is unchanged.
assert((FoundIB || Unloop.contains(L)) && "uninitialized successor");
}
}
}
// Each irreducible loop within the unloop induces a round of iteration using
// the DFS result cached by Traversal.
bool Changed = FoundIB;
for (unsigned NIters = 0; Changed; ++NIters) {
assert(NIters < Unloop.getNumBlocks() && "runaway iterative algorithm");
// Iterate over the postorder list of blocks, propagating the nearest loop
// from successors to predecessors as before.
Changed = false;
for (LoopBlocksDFS::POIterator POI = DFS.beginPostorder(),
POE = DFS.endPostorder();
POI != POE; ++POI) {
Loop *L = LI->getLoopFor(*POI);
Loop *NL = getNearestLoop(*POI, L);
if (NL != L) {
assert(NL != &Unloop && (!NL || NL->contains(&Unloop)) &&
"uninitialized successor");
LI->changeLoopFor(*POI, NL);
Changed = true;
}
}
}
}
/// Remove unloop's blocks from all ancestors below their new parents.
void UnloopUpdater::removeBlocksFromAncestors() {
// Remove all unloop's blocks (including those in nested subloops) from
// ancestors below the new parent loop.
for (BasicBlock *BB : Unloop.blocks()) {
Loop *OuterParent = LI->getLoopFor(BB);
if (Unloop.contains(OuterParent)) {
while (OuterParent->getParentLoop() != &Unloop)
OuterParent = OuterParent->getParentLoop();
OuterParent = SubloopParents[OuterParent];
}
// Remove blocks from former Ancestors except Unloop itself which will be
// deleted.
for (Loop *OldParent = Unloop.getParentLoop(); OldParent != OuterParent;
OldParent = OldParent->getParentLoop()) {
assert(OldParent && "new loop is not an ancestor of the original");
OldParent->removeBlockFromLoop(BB);
}
}
}
/// Update the parent loop for all subloops directly nested within unloop.
void UnloopUpdater::updateSubloopParents() {
while (!Unloop.isInnermost()) {
Loop *Subloop = *std::prev(Unloop.end());
Unloop.removeChildLoop(std::prev(Unloop.end()));
assert(SubloopParents.count(Subloop) && "DFS failed to visit subloop");
if (Loop *Parent = SubloopParents[Subloop])
Parent->addChildLoop(Subloop);
else
LI->addTopLevelLoop(Subloop);
}
}
/// Return the nearest parent loop among this block's successors. If a successor
/// is a subloop header, consider its parent to be the nearest parent of the
/// subloop's exits.
///
/// For subloop blocks, simply update SubloopParents and return NULL.
Loop *UnloopUpdater::getNearestLoop(BasicBlock *BB, Loop *BBLoop) {
// Initially for blocks directly contained by Unloop, NearLoop == Unloop and
// is considered uninitialized.
Loop *NearLoop = BBLoop;
Loop *Subloop = nullptr;
if (NearLoop != &Unloop && Unloop.contains(NearLoop)) {
Subloop = NearLoop;
// Find the subloop ancestor that is directly contained within Unloop.
while (Subloop->getParentLoop() != &Unloop) {
Subloop = Subloop->getParentLoop();
assert(Subloop && "subloop is not an ancestor of the original loop");
}
// Get the current nearest parent of the Subloop exits, initially Unloop.
NearLoop = SubloopParents.insert({Subloop, &Unloop}).first->second;
}
succ_iterator I = succ_begin(BB), E = succ_end(BB);
if (I == E) {
assert(!Subloop && "subloop blocks must have a successor");
NearLoop = nullptr; // unloop blocks may now exit the function.
}
for (; I != E; ++I) {
if (*I == BB)
continue; // self loops are uninteresting
Loop *L = LI->getLoopFor(*I);
if (L == &Unloop) {
// This successor has not been processed. This path must lead to an
// irreducible backedge.
assert((FoundIB || !DFS.hasPostorder(*I)) && "should have seen IB");
FoundIB = true;
}
if (L != &Unloop && Unloop.contains(L)) {
// Successor is in a subloop.
if (Subloop)
continue; // Branching within subloops. Ignore it.
// BB branches from the original into a subloop header.
assert(L->getParentLoop() == &Unloop && "cannot skip into nested loops");
// Get the current nearest parent of the Subloop's exits.
L = SubloopParents[L];
// L could be Unloop if the only exit was an irreducible backedge.
}
if (L == &Unloop) {
continue;
}
// Handle critical edges from Unloop into a sibling loop.
if (L && !L->contains(&Unloop)) {
L = L->getParentLoop();
}
// Remember the nearest parent loop among successors or subloop exits.
if (NearLoop == &Unloop || !NearLoop || NearLoop->contains(L))
NearLoop = L;
}
if (Subloop) {
SubloopParents[Subloop] = NearLoop;
return BBLoop;
}
return NearLoop;
}
LoopInfo::LoopInfo(const DomTreeBase<BasicBlock> &DomTree) { analyze(DomTree); }
bool LoopInfo::invalidate(Function &F, const PreservedAnalyses &PA,
FunctionAnalysisManager::Invalidator &) {
// Check whether the analysis, all analyses on functions, or the function's
// CFG have been preserved.
auto PAC = PA.getChecker<LoopAnalysis>();
return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>() ||
PAC.preservedSet<CFGAnalyses>());
}
void LoopInfo::erase(Loop *Unloop) {
assert(!Unloop->isInvalid() && "Loop has already been erased!");
auto InvalidateOnExit = make_scope_exit([&]() { destroy(Unloop); });
// First handle the special case of no parent loop to simplify the algorithm.
if (Unloop->isOutermost()) {
// Since BBLoop had no parent, Unloop blocks are no longer in a loop.
for (BasicBlock *BB : Unloop->blocks()) {
// Don't reparent blocks in subloops.
if (getLoopFor(BB) != Unloop)
continue;
// Blocks no longer have a parent but are still referenced by Unloop until
// the Unloop object is deleted.
changeLoopFor(BB, nullptr);
}
// Remove the loop from the top-level LoopInfo object.
for (iterator I = begin();; ++I) {
assert(I != end() && "Couldn't find loop");
if (*I == Unloop) {
removeLoop(I);
break;
}
}
// Move all of the subloops to the top-level.
while (!Unloop->isInnermost())
addTopLevelLoop(Unloop->removeChildLoop(std::prev(Unloop->end())));
return;
}
// Update the parent loop for all blocks within the loop. Blocks within
// subloops will not change parents.
UnloopUpdater Updater(Unloop, this);
Updater.updateBlockParents();
// Remove blocks from former ancestor loops.
Updater.removeBlocksFromAncestors();
// Add direct subloops as children in their new parent loop.
Updater.updateSubloopParents();
// Remove unloop from its parent loop.
Loop *ParentLoop = Unloop->getParentLoop();
for (Loop::iterator I = ParentLoop->begin();; ++I) {
assert(I != ParentLoop->end() && "Couldn't find loop");
if (*I == Unloop) {
ParentLoop->removeChildLoop(I);
break;
}
}
}
bool
LoopInfo::wouldBeOutOfLoopUseRequiringLCSSA(const Value *V,
const BasicBlock *ExitBB) const {
if (V->getType()->isTokenTy())
// We can't form PHIs of token type, so the definition of LCSSA excludes
// values of that type.
return false;
const Instruction *I = dyn_cast<Instruction>(V);
if (!I)
return false;
const Loop *L = getLoopFor(I->getParent());
if (!L)
return false;
if (L->contains(ExitBB))
// Could be an exit bb of a subloop and contained in defining loop
return false;
// We found a (new) out-of-loop use location, for a value defined in-loop.
// (Note that because of LCSSA, we don't have to account for values defined
// in sibling loops. Such values will have LCSSA phis of their own in the
// common parent loop.)
return true;
}
[PM] Change the static object whose address is used to uniquely identify analyses to have a common type which is enforced rather than using a char object and a `void *` type when used as an identifier. This has a number of advantages. First, it at least helps some of the confusion raised in Justin Lebar's code review of why `void *` was being used everywhere by having a stronger type that connects to documentation about this. However, perhaps more importantly, it addresses a serious issue where the alignment of these pointer-like identifiers was unknown. This made it hard to use them in pointer-like data structures. We were already dodging this in dangerous ways to create the "all analyses" entry. In a subsequent patch I attempted to use these with TinyPtrVector and things fell apart in a very bad way. And it isn't just a compile time or type system issue. Worse than that, the actual alignment of these pointer-like opaque identifiers wasn't guaranteed to be a useful alignment as they were just characters. This change introduces a type to use as the "key" object whose address forms the opaque identifier. This both forces the objects to have proper alignment, and provides type checking that we get it right everywhere. It also makes the types somewhat less mysterious than `void *`. We could go one step further and introduce a truly opaque pointer-like type to return from the `ID()` static function rather than returning `AnalysisKey *`, but that didn't seem to be a clear win so this is just the initial change to get to a reliably typed and aligned object serving is a key for all the analyses. Thanks to Richard Smith and Justin Lebar for helping pick plausible names and avoid making this refactoring many times. =] And thanks to Sean for the super fast review! While here, I've tried to move away from the "PassID" nomenclature entirely as it wasn't really helping and is overloaded with old pass manager constructs. Now we have IDs for analyses, and key objects whose address can be used as IDs. Where possible and clear I've shortened this to just "ID". In a few places I kept "AnalysisID" to make it clear what was being identified. Differential Revision: https://reviews.llvm.org/D27031 llvm-svn: 287783
2016-11-24 01:53:26 +08:00
AnalysisKey LoopAnalysis::Key;
LoopInfo LoopAnalysis::run(Function &F, FunctionAnalysisManager &AM) {
// FIXME: Currently we create a LoopInfo from scratch for every function.
// This may prove to be too wasteful due to deallocating and re-allocating
// memory each time for the underlying map and vector datastructures. At some
// point it may prove worthwhile to use a freelist and recycle LoopInfo
// objects. I don't want to add that kind of complexity until the scope of
// the problem is better understood.
LoopInfo LI;
LI.analyze(AM.getResult<DominatorTreeAnalysis>(F));
return LI;
}
PreservedAnalyses LoopPrinterPass::run(Function &F,
FunctionAnalysisManager &AM) {
AM.getResult<LoopAnalysis>(F).print(OS);
return PreservedAnalyses::all();
}
[PM] Rewrite the loop pass manager to use a worklist and augmented run arguments much like the CGSCC pass manager. This is a major redesign following the pattern establish for the CGSCC layer to support updates to the set of loops during the traversal of the loop nest and to support invalidation of analyses. An additional significant burden in the loop PM is that so many passes require access to a large number of function analyses. Manually ensuring these are cached, available, and preserved has been a long-standing burden in LLVM even with the help of the automatic scheduling in the old pass manager. And it made the new pass manager extremely unweildy. With this design, we can package the common analyses up while in a function pass and make them immediately available to all the loop passes. While in some cases this is unnecessary, I think the simplicity afforded is worth it. This does not (yet) address loop simplified form or LCSSA form, but those are the next things on my radar and I have a clear plan for them. While the patch is very large, most of it is either mechanically updating loop passes to the new API or the new testing for the loop PM. The code for it is reasonably compact. I have not yet updated all of the loop passes to correctly leverage the update mechanisms demonstrated in the unittests. I'll do that in follow-up patches along with improved FileCheck tests for those passes that ensure things work in more realistic scenarios. In many cases, there isn't much we can do with these until the loop simplified form and LCSSA form are in place. Differential Revision: https://reviews.llvm.org/D28292 llvm-svn: 291651
2017-01-11 14:23:21 +08:00
void llvm::printLoop(Loop &L, raw_ostream &OS, const std::string &Banner) {
if (forcePrintModuleIR()) {
// handling -print-module-scope
OS << Banner << " (loop: ";
L.getHeader()->printAsOperand(OS, false);
OS << ")\n";
// printing whole module
OS << *L.getHeader()->getModule();
return;
}
OS << Banner;
auto *PreHeader = L.getLoopPreheader();
if (PreHeader) {
OS << "\n; Preheader:";
PreHeader->print(OS);
OS << "\n; Loop:";
}
for (auto *Block : L.blocks())
if (Block)
Block->print(OS);
else
OS << "Printing <null> block";
SmallVector<BasicBlock *, 8> ExitBlocks;
L.getExitBlocks(ExitBlocks);
if (!ExitBlocks.empty()) {
OS << "\n; Exit blocks";
for (auto *Block : ExitBlocks)
if (Block)
Block->print(OS);
else
OS << "Printing <null> block";
}
}
Introduce llvm.loop.parallel_accesses and llvm.access.group metadata. The current llvm.mem.parallel_loop_access metadata has a problem in that it uses LoopIDs. LoopID unfortunately is not loop identifier. It is neither unique (there's even a regression test assigning the some LoopID to multiple loops; can otherwise happen if passes such as LoopVersioning make copies of entire loops) nor persistent (every time a property is removed/added from a LoopID's MDNode, it will also receive a new LoopID; this happens e.g. when calling Loop::setLoopAlreadyUnrolled()). Since most loop transformation passes change the loop attributes (even if it just to mark that a loop should not be processed again as llvm.loop.isvectorized does, for the versioned and unversioned loop), the parallel access information is lost for any subsequent pass. This patch unlinks LoopIDs and parallel accesses. llvm.mem.parallel_loop_access metadata on instruction is replaced by llvm.access.group metadata. llvm.access.group points to a distinct MDNode with no operands (avoiding the problem to ever need to add/remove operands), called "access group". Alternatively, it can point to a list of access groups. The LoopID then has an attribute llvm.loop.parallel_accesses with all the access groups that are parallel (no dependencies carries by this loop). This intentionally avoid any kind of "ID". Loops that are clones/have their attributes modifies retain the llvm.loop.parallel_accesses attribute. Access instructions that a cloned point to the same access group. It is not necessary for each access to have it's own "ID" MDNode, but those memory access instructions with the same behavior can be grouped together. The behavior of llvm.mem.parallel_loop_access is not changed by this patch, but should be considered deprecated. Differential Revision: https://reviews.llvm.org/D52116 llvm-svn: 349725
2018-12-20 12:58:07 +08:00
MDNode *llvm::findOptionMDForLoopID(MDNode *LoopID, StringRef Name) {
// No loop metadata node, no loop properties.
if (!LoopID)
return nullptr;
// First operand should refer to the metadata node itself, for legacy reasons.
assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
// Iterate over the metdata node operands and look for MDString metadata.
for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
if (!MD || MD->getNumOperands() < 1)
continue;
MDString *S = dyn_cast<MDString>(MD->getOperand(0));
if (!S)
continue;
// Return the operand node if MDString holds expected metadata.
if (Name.equals(S->getString()))
return MD;
}
// Loop property not found.
return nullptr;
}
MDNode *llvm::findOptionMDForLoop(const Loop *TheLoop, StringRef Name) {
return findOptionMDForLoopID(TheLoop->getLoopID(), Name);
}
/// Find string metadata for loop
///
/// If it has a value (e.g. {"llvm.distribute", 1} return the value as an
/// operand or null otherwise. If the string metadata is not found return
/// Optional's not-a-value.
Optional<const MDOperand *> llvm::findStringMetadataForLoop(const Loop *TheLoop,
StringRef Name) {
MDNode *MD = findOptionMDForLoop(TheLoop, Name);
if (!MD)
return None;
switch (MD->getNumOperands()) {
case 1:
return nullptr;
case 2:
return &MD->getOperand(1);
default:
llvm_unreachable("loop metadata has 0 or 1 operand");
}
}
Optional<bool> llvm::getOptionalBoolLoopAttribute(const Loop *TheLoop,
StringRef Name) {
MDNode *MD = findOptionMDForLoop(TheLoop, Name);
if (!MD)
return None;
switch (MD->getNumOperands()) {
case 1:
// When the value is absent it is interpreted as 'attribute set'.
return true;
case 2:
if (ConstantInt *IntMD =
mdconst::extract_or_null<ConstantInt>(MD->getOperand(1).get()))
return IntMD->getZExtValue();
return true;
}
llvm_unreachable("unexpected number of options");
}
bool llvm::getBooleanLoopAttribute(const Loop *TheLoop, StringRef Name) {
return getOptionalBoolLoopAttribute(TheLoop, Name).getValueOr(false);
}
llvm::Optional<int> llvm::getOptionalIntLoopAttribute(const Loop *TheLoop,
StringRef Name) {
const MDOperand *AttrMD =
findStringMetadataForLoop(TheLoop, Name).getValueOr(nullptr);
if (!AttrMD)
return None;
ConstantInt *IntMD = mdconst::extract_or_null<ConstantInt>(AttrMD->get());
if (!IntMD)
return None;
return IntMD->getSExtValue();
}
int llvm::getIntLoopAttribute(const Loop *TheLoop, StringRef Name,
int Default) {
return getOptionalIntLoopAttribute(TheLoop, Name).getValueOr(Default);
}
static const char *LLVMLoopMustProgress = "llvm.loop.mustprogress";
bool llvm::hasMustProgress(const Loop *L) {
return getBooleanLoopAttribute(L, LLVMLoopMustProgress);
}
bool llvm::isMustProgress(const Loop *L) {
return L->getHeader()->getParent()->mustProgress() || hasMustProgress(L);
}
Introduce llvm.loop.parallel_accesses and llvm.access.group metadata. The current llvm.mem.parallel_loop_access metadata has a problem in that it uses LoopIDs. LoopID unfortunately is not loop identifier. It is neither unique (there's even a regression test assigning the some LoopID to multiple loops; can otherwise happen if passes such as LoopVersioning make copies of entire loops) nor persistent (every time a property is removed/added from a LoopID's MDNode, it will also receive a new LoopID; this happens e.g. when calling Loop::setLoopAlreadyUnrolled()). Since most loop transformation passes change the loop attributes (even if it just to mark that a loop should not be processed again as llvm.loop.isvectorized does, for the versioned and unversioned loop), the parallel access information is lost for any subsequent pass. This patch unlinks LoopIDs and parallel accesses. llvm.mem.parallel_loop_access metadata on instruction is replaced by llvm.access.group metadata. llvm.access.group points to a distinct MDNode with no operands (avoiding the problem to ever need to add/remove operands), called "access group". Alternatively, it can point to a list of access groups. The LoopID then has an attribute llvm.loop.parallel_accesses with all the access groups that are parallel (no dependencies carries by this loop). This intentionally avoid any kind of "ID". Loops that are clones/have their attributes modifies retain the llvm.loop.parallel_accesses attribute. Access instructions that a cloned point to the same access group. It is not necessary for each access to have it's own "ID" MDNode, but those memory access instructions with the same behavior can be grouped together. The behavior of llvm.mem.parallel_loop_access is not changed by this patch, but should be considered deprecated. Differential Revision: https://reviews.llvm.org/D52116 llvm-svn: 349725
2018-12-20 12:58:07 +08:00
bool llvm::isValidAsAccessGroup(MDNode *Node) {
return Node->getNumOperands() == 0 && Node->isDistinct();
}
MDNode *llvm::makePostTransformationMetadata(LLVMContext &Context,
MDNode *OrigLoopID,
ArrayRef<StringRef> RemovePrefixes,
ArrayRef<MDNode *> AddAttrs) {
// First remove any existing loop metadata related to this transformation.
SmallVector<Metadata *, 4> MDs;
// Reserve first location for self reference to the LoopID metadata node.
MDs.push_back(nullptr);
// Remove metadata for the transformation that has been applied or that became
// outdated.
if (OrigLoopID) {
for (unsigned i = 1, ie = OrigLoopID->getNumOperands(); i < ie; ++i) {
bool IsVectorMetadata = false;
Metadata *Op = OrigLoopID->getOperand(i);
if (MDNode *MD = dyn_cast<MDNode>(Op)) {
const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
if (S)
IsVectorMetadata =
llvm::any_of(RemovePrefixes, [S](StringRef Prefix) -> bool {
return S->getString().startswith(Prefix);
});
}
if (!IsVectorMetadata)
MDs.push_back(Op);
}
}
// Add metadata to avoid reapplying a transformation, such as
// llvm.loop.unroll.disable and llvm.loop.isvectorized.
MDs.append(AddAttrs.begin(), AddAttrs.end());
MDNode *NewLoopID = MDNode::getDistinct(Context, MDs);
// Replace the temporary node with a self-reference.
NewLoopID->replaceOperandWith(0, NewLoopID);
return NewLoopID;
}
//===----------------------------------------------------------------------===//
// LoopInfo implementation
//
LoopInfoWrapperPass::LoopInfoWrapperPass() : FunctionPass(ID) {
initializeLoopInfoWrapperPassPass(*PassRegistry::getPassRegistry());
}
char LoopInfoWrapperPass::ID = 0;
INITIALIZE_PASS_BEGIN(LoopInfoWrapperPass, "loops", "Natural Loop Information",
true, true)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_END(LoopInfoWrapperPass, "loops", "Natural Loop Information",
true, true)
bool LoopInfoWrapperPass::runOnFunction(Function &) {
releaseMemory();
LI.analyze(getAnalysis<DominatorTreeWrapperPass>().getDomTree());
return false;
}
void LoopInfoWrapperPass::verifyAnalysis() const {
// LoopInfoWrapperPass is a FunctionPass, but verifying every loop in the
// function each time verifyAnalysis is called is very expensive. The
// -verify-loop-info option can enable this. In order to perform some
// checking by default, LoopPass has been taught to call verifyLoop manually
// during loop pass sequences.
if (VerifyLoopInfo) {
auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
LI.verify(DT);
}
}
void LoopInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
AU.addRequiredTransitive<DominatorTreeWrapperPass>();
}
void LoopInfoWrapperPass::print(raw_ostream &OS, const Module *) const {
LI.print(OS);
}
PreservedAnalyses LoopVerifierPass::run(Function &F,
FunctionAnalysisManager &AM) {
LoopInfo &LI = AM.getResult<LoopAnalysis>(F);
auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
LI.verify(DT);
return PreservedAnalyses::all();
}
//===----------------------------------------------------------------------===//
// LoopBlocksDFS implementation
//
/// Traverse the loop blocks and store the DFS result.
/// Useful for clients that just want the final DFS result and don't need to
/// visit blocks during the initial traversal.
void LoopBlocksDFS::perform(LoopInfo *LI) {
LoopBlocksTraversal Traversal(*this, LI);
for (LoopBlocksTraversal::POTIterator POI = Traversal.begin(),
POE = Traversal.end();
POI != POE; ++POI)
;
}