Google Git
Sign in
llvm/llvm-project/beffef686fde396bee59bf31ce44fffa0e6fc2ca/./llvm/lib/Transforms/Utils/LoopUnroll.cpp
blob: d00de52e1cb0b37dc24564a5ddcbdaa1731c5aac [file]
//===-- UnrollLoop.cpp - Loop unrolling utilities -------------------------===//
//
// 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 implements some loop unrolling utilities. It does not define any
// actual pass or policy, but provides a single function to perform loop
// unrolling.
//
// The process of unrolling can produce extraneous basic blocks linked with
// unconditional branches. This will be corrected in the future.
//
//===----------------------------------------------------------------------===//
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/MapVector.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/ScopedHashTable.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/ADT/Twine.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/DomTreeUpdater.h"
#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/LoopIterator.h"
#include "llvm/Analysis/MemorySSA.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/CFG.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Metadata.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/IR/Use.h"
#include "llvm/IR/User.h"
#include "llvm/IR/ValueHandle.h"
#include "llvm/IR/ValueMap.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/GenericDomTree.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Cloning.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Transforms/Utils/LoopSimplify.h"
#include "llvm/Transforms/Utils/LoopUtils.h"
#include "llvm/Transforms/Utils/SimplifyIndVar.h"
#include "llvm/Transforms/Utils/UnrollLoop.h"
#include "llvm/Transforms/Utils/ValueMapper.h"
#include <assert.h>
#include <cmath>
#include <numeric>
#include <vector>
namespace llvm {
class DataLayout;
class Value;
} // namespace llvm
using namespace llvm;
#define DEBUG_TYPE "loop-unroll"
// TODO: Should these be here or in LoopUnroll?
STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
STATISTIC(NumUnrolledNotLatch, "Number of loops unrolled without a conditional "
"latch (completely or otherwise)");
static cl::opt<bool>
UnrollRuntimeEpilog("unroll-runtime-epilog", cl::init(false), cl::Hidden,
cl::desc("Allow runtime unrolled loops to be unrolled "
"with epilog instead of prolog."));
static cl::opt<bool>
UnrollVerifyDomtree("unroll-verify-domtree", cl::Hidden,
cl::desc("Verify domtree after unrolling"),
#ifdef EXPENSIVE_CHECKS
cl::init(true)
#else
cl::init(false)
#endif
);
static cl::opt<bool>
UnrollVerifyLoopInfo("unroll-verify-loopinfo", cl::Hidden,
cl::desc("Verify loopinfo after unrolling"),
#ifdef EXPENSIVE_CHECKS
cl::init(true)
#else
cl::init(false)
#endif
);
static cl::opt<bool> UnrollAddParallelReductions(
"unroll-add-parallel-reductions", cl::init(false), cl::Hidden,
cl::desc("Allow unrolling to add parallel reduction phis."));
/// Check if unrolling created a situation where we need to insert phi nodes to
/// preserve LCSSA form.
/// \param Blocks is a vector of basic blocks representing unrolled loop.
/// \param L is the outer loop.
/// It's possible that some of the blocks are in L, and some are not. In this
/// case, if there is a use is outside L, and definition is inside L, we need to
/// insert a phi-node, otherwise LCSSA will be broken.
/// The function is just a helper function for llvm::UnrollLoop that returns
/// true if this situation occurs, indicating that LCSSA needs to be fixed.
static bool needToInsertPhisForLCSSA(Loop *L,
const std::vector<BasicBlock *> &Blocks,
LoopInfo *LI) {
for (BasicBlock *BB : Blocks) {
if (LI->getLoopFor(BB) == L)
continue;
for (Instruction &I : *BB) {
for (Use &U : I.operands()) {
if (const auto *Def = dyn_cast<Instruction>(U)) {
Loop *DefLoop = LI->getLoopFor(Def->getParent());
if (!DefLoop)
continue;
if (DefLoop->contains(L))
return true;
}
}
}
}
return false;
}
/// Adds ClonedBB to LoopInfo, creates a new loop for ClonedBB if necessary
/// and adds a mapping from the original loop to the new loop to NewLoops.
/// Returns nullptr if no new loop was created and a pointer to the
/// original loop OriginalBB was part of otherwise.
const Loop* llvm::addClonedBlockToLoopInfo(BasicBlock *OriginalBB,
BasicBlock *ClonedBB, LoopInfo *LI,
NewLoopsMap &NewLoops) {
// Figure out which loop New is in.
const Loop *OldLoop = LI->getLoopFor(OriginalBB);
assert(OldLoop && "Should (at least) be in the loop being unrolled!");
Loop *&NewLoop = NewLoops[OldLoop];
if (!NewLoop) {
// Found a new sub-loop.
assert(OriginalBB == OldLoop->getHeader() &&
"Header should be first in RPO");
NewLoop = LI->AllocateLoop();
Loop *NewLoopParent = NewLoops.lookup(OldLoop->getParentLoop());
if (NewLoopParent)
NewLoopParent->addChildLoop(NewLoop);
else
LI->addTopLevelLoop(NewLoop);
NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
return OldLoop;
} else {
NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
return nullptr;
}
}
/// The function chooses which type of unroll (epilog or prolog) is more
/// profitabale.
/// Epilog unroll is more profitable when there is PHI that starts from
/// constant. In this case epilog will leave PHI start from constant,
/// but prolog will convert it to non-constant.
///
/// loop:
/// PN = PHI [I, Latch], [CI, PreHeader]
/// I = foo(PN)
/// ...
///
/// Epilog unroll case.
/// loop:
/// PN = PHI [I2, Latch], [CI, PreHeader]
/// I1 = foo(PN)
/// I2 = foo(I1)
/// ...
/// Prolog unroll case.
/// NewPN = PHI [PrologI, Prolog], [CI, PreHeader]
/// loop:
/// PN = PHI [I2, Latch], [NewPN, PreHeader]
/// I1 = foo(PN)
/// I2 = foo(I1)
/// ...
///
static bool isEpilogProfitable(Loop *L) {
BasicBlock *PreHeader = L->getLoopPreheader();
BasicBlock *Header = L->getHeader();
assert(PreHeader && Header);
for (const PHINode &PN : Header->phis()) {
if (isa<ConstantInt>(PN.getIncomingValueForBlock(PreHeader)))
return true;
}
return false;
}
struct LoadValue {
Instruction *DefI = nullptr;
unsigned Generation = 0;
LoadValue() = default;
LoadValue(Instruction *Inst, unsigned Generation)
: DefI(Inst), Generation(Generation) {}
};
class StackNode {
ScopedHashTable<const SCEV *, LoadValue>::ScopeTy LoadScope;
unsigned CurrentGeneration;
unsigned ChildGeneration;
DomTreeNode *Node;
DomTreeNode::const_iterator ChildIter;
DomTreeNode::const_iterator EndIter;
bool Processed = false;
public:
StackNode(ScopedHashTable<const SCEV *, LoadValue> &AvailableLoads,
unsigned cg, DomTreeNode *N, DomTreeNode::const_iterator Child,
DomTreeNode::const_iterator End)
: LoadScope(AvailableLoads), CurrentGeneration(cg), ChildGeneration(cg),
Node(N), ChildIter(Child), EndIter(End) {}
// Accessors.
unsigned currentGeneration() const { return CurrentGeneration; }
unsigned childGeneration() const { return ChildGeneration; }
void childGeneration(unsigned generation) { ChildGeneration = generation; }
DomTreeNode *node() { return Node; }
DomTreeNode::const_iterator childIter() const { return ChildIter; }
DomTreeNode *nextChild() {
DomTreeNode *Child = *ChildIter;
++ChildIter;
return Child;
}
DomTreeNode::const_iterator end() const { return EndIter; }
bool isProcessed() const { return Processed; }
void process() { Processed = true; }
};
Value *getMatchingValue(LoadValue LV, LoadInst *LI, unsigned CurrentGeneration,
BatchAAResults &BAA,
function_ref<MemorySSA *()> GetMSSA) {
if (!LV.DefI)
return nullptr;
if (LV.DefI->getType() != LI->getType())
return nullptr;
if (LV.Generation != CurrentGeneration) {
MemorySSA *MSSA = GetMSSA();
if (!MSSA)
return nullptr;
auto *EarlierMA = MSSA->getMemoryAccess(LV.DefI);
MemoryAccess *LaterDef =
MSSA->getWalker()->getClobberingMemoryAccess(LI, BAA);
if (!MSSA->dominates(LaterDef, EarlierMA))
return nullptr;
}
return LV.DefI;
}
void loadCSE(Loop *L, DominatorTree &DT, ScalarEvolution &SE, LoopInfo &LI,
BatchAAResults &BAA, function_ref<MemorySSA *()> GetMSSA) {
ScopedHashTable<const SCEV *, LoadValue> AvailableLoads;
SmallVector<std::unique_ptr<StackNode>> NodesToProcess;
DomTreeNode *HeaderD = DT.getNode(L->getHeader());
NodesToProcess.emplace_back(new StackNode(AvailableLoads, 0, HeaderD,
HeaderD->begin(), HeaderD->end()));
unsigned CurrentGeneration = 0;
while (!NodesToProcess.empty()) {
StackNode *NodeToProcess = &*NodesToProcess.back();
CurrentGeneration = NodeToProcess->currentGeneration();
if (!NodeToProcess->isProcessed()) {
// Process the node.
// If this block has a single predecessor, then the predecessor is the
// parent
// of the domtree node and all of the live out memory values are still
// current in this block. If this block has multiple predecessors, then
// they could have invalidated the live-out memory values of our parent
// value. For now, just be conservative and invalidate memory if this
// block has multiple predecessors.
if (!NodeToProcess->node()->getBlock()->getSinglePredecessor())
++CurrentGeneration;
for (auto &I : make_early_inc_range(*NodeToProcess->node()->getBlock())) {
auto *Load = dyn_cast<LoadInst>(&I);
if (!Load || !Load->isSimple()) {
if (I.mayWriteToMemory())
CurrentGeneration++;
continue;
}
const SCEV *PtrSCEV = SE.getSCEV(Load->getPointerOperand());
LoadValue LV = AvailableLoads.lookup(PtrSCEV);
if (Value *M =
getMatchingValue(LV, Load, CurrentGeneration, BAA, GetMSSA)) {
if (LI.replacementPreservesLCSSAForm(Load, M)) {
Load->replaceAllUsesWith(M);
Load->eraseFromParent();
}
} else {
AvailableLoads.insert(PtrSCEV, LoadValue(Load, CurrentGeneration));
}
}
NodeToProcess->childGeneration(CurrentGeneration);
NodeToProcess->process();
} else if (NodeToProcess->childIter() != NodeToProcess->end()) {
// Push the next child onto the stack.
DomTreeNode *Child = NodeToProcess->nextChild();
if (!L->contains(Child->getBlock()))
continue;
NodesToProcess.emplace_back(
new StackNode(AvailableLoads, NodeToProcess->childGeneration(), Child,
Child->begin(), Child->end()));
} else {
// It has been processed, and there are no more children to process,
// so delete it and pop it off the stack.
NodesToProcess.pop_back();
}
}
}
/// Perform some cleanup and simplifications on loops after unrolling. It is
/// useful to simplify the IV's in the new loop, as well as do a quick
/// simplify/dce pass of the instructions.
void llvm::simplifyLoopAfterUnroll(Loop *L, bool SimplifyIVs, LoopInfo *LI,
ScalarEvolution *SE, DominatorTree *DT,
AssumptionCache *AC,
const TargetTransformInfo *TTI,
ArrayRef<BasicBlock *> Blocks,
AAResults *AA) {
using namespace llvm::PatternMatch;
// Simplify any new induction variables in the partially unrolled loop.
if (SE && SimplifyIVs) {
SmallVector<WeakTrackingVH, 16> DeadInsts;
simplifyLoopIVs(L, SE, DT, LI, TTI, DeadInsts);
// Aggressively clean up dead instructions that simplifyLoopIVs already
// identified. Any remaining should be cleaned up below.
while (!DeadInsts.empty()) {
Value *V = DeadInsts.pop_back_val();
if (Instruction *Inst = dyn_cast_or_null<Instruction>(V))
RecursivelyDeleteTriviallyDeadInstructions(Inst);
}
if (AA) {
std::unique_ptr<MemorySSA> MSSA = nullptr;
BatchAAResults BAA(*AA);
loadCSE(L, *DT, *SE, *LI, BAA, [L, AA, DT, &MSSA]() -> MemorySSA * {
if (!MSSA)
MSSA.reset(new MemorySSA(*L, AA, DT));
return &*MSSA;
});
}
}
// At this point, the code is well formed. Perform constprop, instsimplify,
// and dce.
SmallVector<WeakTrackingVH, 16> DeadInsts;
for (BasicBlock *BB : Blocks) {
// Remove repeated debug instructions after loop unrolling.
if (BB->getParent()->getSubprogram())
RemoveRedundantDbgInstrs(BB);
for (Instruction &Inst : llvm::make_early_inc_range(*BB)) {
if (Value *V = simplifyInstruction(
&Inst, {BB->getDataLayout(), nullptr, DT, AC}))
if (LI->replacementPreservesLCSSAForm(&Inst, V))
Inst.replaceAllUsesWith(V);
if (isInstructionTriviallyDead(&Inst))
DeadInsts.emplace_back(&Inst);
// Fold ((add X, C1), C2) to (add X, C1+C2). This is very common in
// unrolled loops, and handling this early allows following code to
// identify the IV as a "simple recurrence" without first folding away
// a long chain of adds.
{
Value *X;
const APInt *C1, *C2;
if (match(&Inst, m_Add(m_Add(m_Value(X), m_APInt(C1)), m_APInt(C2)))) {
auto *InnerI = dyn_cast<Instruction>(Inst.getOperand(0));
auto *InnerOBO = cast<OverflowingBinaryOperator>(Inst.getOperand(0));
bool SignedOverflow;
APInt NewC = C1->sadd_ov(*C2, SignedOverflow);
Inst.setOperand(0, X);
Inst.setOperand(1, ConstantInt::get(Inst.getType(), NewC));
Inst.setHasNoUnsignedWrap(Inst.hasNoUnsignedWrap() &&
InnerOBO->hasNoUnsignedWrap());
Inst.setHasNoSignedWrap(Inst.hasNoSignedWrap() &&
InnerOBO->hasNoSignedWrap() &&
!SignedOverflow);
if (InnerI && isInstructionTriviallyDead(InnerI))
DeadInsts.emplace_back(InnerI);
}
}
}
// We can't do recursive deletion until we're done iterating, as we might
// have a phi which (potentially indirectly) uses instructions later in
// the block we're iterating through.
RecursivelyDeleteTriviallyDeadInstructions(DeadInsts);
}
}
// Loops containing convergent instructions that are uncontrolled or controlled
// from outside the loop must have a count that divides their TripMultiple.
LLVM_ATTRIBUTE_USED
static bool canHaveUnrollRemainder(const Loop *L) {
if (getLoopConvergenceHeart(L))
return false;
// Check for uncontrolled convergent operations.
for (auto &BB : L->blocks()) {
for (auto &I : *BB) {
if (isa<ConvergenceControlInst>(I))
return true;
if (auto *CB = dyn_cast<CallBase>(&I))
if (CB->isConvergent())
return CB->getConvergenceControlToken();
}
}
return true;
}
// If LoopUnroll has proven OriginalLoopProb is incorrect for some iterations
// of the original loop, adjust latch probabilities in the unrolled loop to
// maintain the original total frequency of the original loop body.
//
// OriginalLoopProb is practical but imprecise
// -------------------------------------------
//
// The latch branch weights that LLVM originally adds to a loop encode one latch
// probability, OriginalLoopProb, applied uniformly across the loop's infinite
// set of theoretically possible iterations. While this uniform latch
// probability serves as a practical statistic summarizing the trip counts
// observed during profiling, it is imprecise. Specifically, unless it is zero,
// it is impossible for it to be the actual probability observed at every
// individual iteration. To see why, consider that the only way to actually
// observe at run time that the latch probability remains non-zero is to profile
// at least one loop execution that has an infinite number of iterations. I do
// not know how to profile an infinite number of loop iterations, and most loops
// I work with are always finite.
//
// LoopUnroll proves OriginalLoopProb is incorrect
// ------------------------------------------------
//
// LoopUnroll reorganizes the original loop so that loop iterations are no
// longer all implemented by the same code, and then it analyzes some of those
// loop iteration implementations independently of others. In particular, it
// converts some of their conditional latches to unconditional. That is, by
// examining code structure without any profile data, LoopUnroll proves that the
// actual latch probability at the end of such an iteration is either 1 or 0.
// When an individual iteration's actual latch probability is 1 or 0, that means
// it always behaves the same, so it is impossible to observe it as having any
// other probability. The original uniform latch probability is rarely 1 or 0
// because, when applied to all possible iterations, that would yield an
// estimated trip count of infinity or 1, respectively.
//
// Thus, the new probabilities of 1 or 0 are proven corrections to
// OriginalLoopProb for individual iterations in the original loop. However,
// LoopUnroll often is able to perform these corrections for only some
// iterations, leaving other iterations with OriginalLoopProb, and thus
// corrupting the aggregate effect on the total frequency of the original loop
// body.
//
// Adjusting latch probabilities
// -----------------------------
//
// This function ensures that the total frequency of the original loop body,
// summed across all its occurrences in the unrolled loop after the
// aforementioned latch conversions, is the same as in the original loop. To do
// so, it adjusts probabilities on the remaining conditional latches. However,
// it cannot derive the new probabilities directly from the original uniform
// latch probability because the latter has been proven incorrect for some
// original loop iterations.
//
// There are often many sets of latch probabilities that can produce the
// original total loop body frequency. If there are many remaining conditional
// latches, this function just quickly hacks a few of their probabilities to
// restore the original total loop body frequency. Otherwise, it determines
// less arbitrary probabilities.
static void fixProbContradiction(Loop *L, UnrollLoopOptions ULO,
OptimizationRemarkEmitter *ORE,
BranchProbability OriginalLoopProb,
bool CompletelyUnroll,
std::vector<unsigned> &IterCounts,
const std::vector<BasicBlock *> &CondLatches,
std::vector<BasicBlock *> &CondLatchNexts) {
// Runtime unrolling is handled later in LoopUnroll not here.
//
// There are two scenarios in which LoopUnroll sets ProbUpdateRequired to true
// because it needs to update probabilities that were originally
// OriginalLoopProb, but only in one scenario has LoopUnroll proven
// OriginalLoopProb incorrect for iterations within the original loop:
// - If ULO.Runtime, LoopUnroll adds new guards that enforce new reaching
// conditions for new loop iteration implementations (e.g., one unrolled
// loop iteration executes only if at least ULO.Count original loop
// iterations remain). Those reaching conditions dictate how conditional
// latches can be converted to unconditional (e.g., within an unrolled loop
// iteration, there is no need to recheck the number of remaining original
// loop iterations). None of this reorganization alters the set of possible
// original loop iteration counts or proves OriginalLoopProb incorrect for
// any of the original loop iterations. Thus, LoopUnroll derives
// probabilities for the new guards and latches directly from
// OriginalLoopProb based on the probabilities that their reaching
// conditions would occur in the original loop. Doing so maintains the
// total frequency of the original loop body.
// - If !ULO.Runtime, LoopUnroll initially adds new loop iteration
// implementations, which have the same latch probabilities as in the
// original loop because there are no new guards that change their reaching
// conditions. Sometimes, LoopUnroll is then done, and so does not set
// ProbUpdateRequired to true. Other times, LoopUnroll then proves that
// some latches are unconditional, directly contradicting OriginalLoopProb
// for the corresponding original loop iterations. That reduces the set of
// possible original loop iteration counts, possibly producing a finite set
// if it manages to eliminate the backedge. LoopUnroll has to choose a new
// set of latch probabilities that produce the same total loop body
// frequency.
//
// This function addresses the second scenario only.
if (ULO.Runtime)
return;
// If CondLatches.empty(), there are no latch branches with probabilities we
// can adjust. That should mean that the actual trip count is always exactly
// the number of remaining unrolled iterations, and so OriginalLoopProb should
// have yielded that trip count as the original loop body frequency. Of
// course, OriginalLoopProb could be based on inaccurate profile data, but
// there is nothing we can do about that here.
if (CondLatches.empty())
return;
// If the original latch probability is 1, the original frequency is infinity.
// Leaving all remaining probabilities set to 1 might or might not get us
// there (e.g., a completely unrolled loop cannot be infinite), but it is the
// closest we can come.
assert(!OriginalLoopProb.isUnknown() &&
"Expected to have loop probability to fix");
if (OriginalLoopProb.isOne())
return;
// FreqDesired is the frequency implied by the original loop probability.
double FreqDesired = 1 / (1 - OriginalLoopProb.toDouble());
// Get the probability at CondLatches[I].
auto GetProb = [&](unsigned I) {
CondBrInst *B = cast<CondBrInst>(CondLatches[I]->getTerminator());
bool FirstTargetIsNext = B->getSuccessor(0) == CondLatchNexts[I];
return getBranchProbability(B, FirstTargetIsNext).toDouble();
};
// Set the probability at CondLatches[I] to Prob.
auto SetProb = [&](unsigned I, double Prob) {
CondBrInst *B = cast<CondBrInst>(CondLatches[I]->getTerminator());
bool FirstTargetIsNext = B->getSuccessor(0) == CondLatchNexts[I];
setBranchProbability(B, BranchProbability::getBranchProbability(Prob),
FirstTargetIsNext);
};
// Set all probabilities in CondLatches to Prob.
auto SetAllProbs = [&](double Prob) {
for (unsigned I = 0, E = CondLatches.size(); I < E; ++I)
SetProb(I, Prob);
};
// If n <= 2, we choose the simplest probability model we can think of: every
// remaining conditional branch instruction has the same probability, Prob,
// of continuing to the next iteration. This model has several helpful
// properties:
// - We have no reason to think one latch branch's probability should be
// higher or lower than another, and so this model makes them all the same.
// In the worst cases, we thus avoid setting just some probabilities to 0 or
// 1, which can unrealistically make some code appear unreachable. There
// are cases where they *all* must become 0 or 1 to achieve the total
// frequency of original loop body, and our model does permit that.
// - The frequency, FreqOne, of the original loop body in a single iteration
// of the unrolled loop is computed by a simple polynomial, where p=Prob,
// n=CondLatches.size(), and c_i=IterCounts[i]:
//
// FreqOne = Sum(i=0..n)(c_i * p^i)
//
// - If the backedge has been eliminated, FreqOne is the total frequency of
// the original loop body in the unrolled loop.
// - If the backedge remains, Sum(i=0..inf)(FreqOne * p^(n*i)) =
// FreqOne / (1 - p^n) is the total frequency of the original loop body in
// the unrolled loop, regardless of whether the backedge is conditional or
// unconditional.
// - For n <= 2, we can use simple formulas to solve the above polynomial
// equations exactly for p without performing a search.
// When iterating for a solution, we stop early if we find probabilities
// that produce a Freq whose difference from FreqDesired is small
// (FreqPrec). Otherwise, we expect to compute a solution at least that
// accurate (but surely far more accurate).
const double FreqPrec = 1e-6;
// Compute the probability that, used at CondLaches[0] where
// CondLatches.size() == 1, gets as close as possible to FreqDesired.
auto ComputeProbForLinear = [&]() {
// The polynomial is linear (0 = A*p + B), so just solve it.
double A = IterCounts[1] + (CompletelyUnroll ? 0 : FreqDesired);
double B = IterCounts[0] - FreqDesired;
assert(A > 0 && "Expected iterations after last conditional latch");
double Prob = -B / A;
Prob = std::max(Prob, 0.);
Prob = std::min(Prob, 1.);
return Prob;
};
// Compute the probability that, used throughout CondLatches where
// CondLatches.size() == 2, gets as close as possible to FreqDesired.
auto ComputeProbForQuadratic = [&]() {
// The polynomial is quadratic (0 = A*p^2 + B*p + C), so just solve it.
double A = IterCounts[2] + (CompletelyUnroll ? 0 : FreqDesired);
double B = IterCounts[1];
double C = IterCounts[0] - FreqDesired;
assert(A > 0 && "Expected iterations after last conditional latch");
double Prob = (-B + sqrt(B * B - 4 * A * C)) / (2 * A);
Prob = std::max(Prob, 0.);
Prob = std::min(Prob, 1.);
return Prob;
};
// Adjust the probability at CondLatches[ComputeIdx] to get as close as
// possible to FreqDesired without replacing probabilities elsewhere in
// CondLatches. Return the new total frequency.
//
// Given a CondLatches index I, then for a single unrolled loop iteration:
// - ProbBefore or ProbAfter is the probability that control flow can pass
// through every CondLatches[J] for J < I or J > I, respectively.
// - FreqBefore or FreqAfter is the total frequency accumulated before or
// after CondLatches[I], respectively, while the probability at
// CondLatches[I] is treated as 1.
//
// If ComputeIdx == 0, then ComputeProb will set those values for I == 0 and
// ignore the current values. If ComputeIdx > 0, then it expects those values
// to already be set for I == ComputeIdx - 1, and it will set them for I ==
// ComputeIdx.
auto AdjustProb = [&](unsigned ComputeIdx, double &ProbBefore,
double &ProbAfter, double &FreqBefore,
double &FreqAfter) {
assert(ComputeIdx < CondLatches.size() &&
"Expected valid CondLatches index");
// Compute or update ProbBefore, ProbAfter, FreqBefore, and FreqAfter.
auto ComputeAfter = [&]() {
ProbAfter = 1;
FreqAfter = IterCounts[ComputeIdx + 1];
for (unsigned I = ComputeIdx + 1, E = CondLatches.size(); I < E; ++I) {
double Prob = GetProb(I);
ProbAfter *= Prob;
// After Prob == 0, ProbAfter and FreqAfter won't change, so save time.
if (Prob == 0)
break;
FreqAfter += IterCounts[I + 1] * ProbAfter;
}
};
if (ComputeIdx == 0) {
ProbBefore = 1;
FreqBefore = IterCounts[0];
ComputeAfter();
} else {
// Rather than iterating all of CondLatches again, we fix up the
// previously computed values.
double ProbOld = GetProb(ComputeIdx);
if (ProbOld > 0) {
FreqAfter -= IterCounts[ComputeIdx] * ProbBefore;
ProbAfter /= ProbOld;
FreqAfter /= ProbOld;
} else {
// We cannot divide out the old zero probability. We short-circuited
// the iteration at that zero in the previous ComputeAfter call, so now
// we pick up where we left off.
ComputeAfter();
}
ProbBefore *= GetProb(ComputeIdx - 1);
FreqBefore += IterCounts[ComputeIdx] * ProbBefore;
}
// Compute the required probability, and limit it to a valid probability (0
// <= p <= 1). See the FreqCompute formula below for how to derive the
// ProbCompute formula.
double ProbReachingBackedge = CompletelyUnroll ? 0 : ProbBefore * ProbAfter;
double ProbComputeNumerator = FreqDesired - FreqBefore;
double ProbComputeDenominator =
FreqAfter + FreqDesired * ProbReachingBackedge;
double ProbCompute = -1; // Init expected to be unused.
if (ProbComputeNumerator <= 0) {
// FreqBefore has already reached or surpassed FreqDesired, so add no more
// frequency. It is possible that ProbComputeDenominator == 0 here
// because some latch probability (maybe the original) was set to zero, so
// this check avoids setting ProbCompute=1 (in the else if below) and
// division by zero where the numerator <= 0 (in the else below).
ProbCompute = 0;
} else if (ProbComputeDenominator == 0) {
// Analytically, this case seems impossible. It would occur if either:
// - Both FreqAfter and FreqDesired are zero. But the latter would cause
// ProbComputeNumerator < 0, which we catch above, and FreqDesired
// should always be >= 1 anyway.
// - There are no iterations after CondLatches[ComputeIdx], not even via
// a backedge, so that both FreqAfter and ProbReachingBackedge are zero.
// But iterations should exist after even the last conditional latch.
// - Some latch probability (maybe the original) was set to zero so that
// both FreqAfter and ProbReachingBackedge are zero. But that should
// not have happened because, according to the above
// ProbComputeNumerator check, we have not yet reached FreqDesired
// (which, if the original latch probability is zero, is just 1 and thus
// always reached or surpassed).
//
// Numerically, perhaps this case is possible. We interpret it to mean we
// need more frequency (ProbComputeNumerator > 0) but have no way to get
// any (ProbComputeDenominator is analytically too small to distinguish it
// from 0 in floating point), suggesting infinite probability is needed,
// but 1 is the maximum valid probability and thus the best we can do.
//
// TODO: Cover this case in the test suite if you can.
ProbCompute = 1;
} else {
ProbCompute = ProbComputeNumerator / ProbComputeDenominator;
ProbCompute = std::max(ProbCompute, 0.);
ProbCompute = std::min(ProbCompute, 1.);
}
SetProb(ComputeIdx, ProbCompute);
// Compute the resulting total frequency.
double FreqCompute = -1; // Init expected to be unused.
if (ProbReachingBackedge * ProbCompute == 1) {
// Analytically, this case seems impossible. It requires that there is a
// backedge and that FreqDesired == infinity so that every conditional
// latch's probability had to be set to 1. But FreqDesired == infinity
// means OriginalLoopProb.isOne(), which we guarded against earlier.
//
// Numerically, perhaps this case is possible. We interpret it to mean
// that analytically the probability has to be so near 1 that, in floating
// point, the frequency is computed as infinite.
//
// TODO: Cover this case in the test suite if you can.
FreqCompute = std::numeric_limits<double>::infinity();
if (ORE) {
ORE->emit([&]() {
return OptimizationRemark(DEBUG_TYPE, "InfiniteFrequency",
L->getStartLoc(), L->getHeader());
});
}
} else {
assert(FreqBefore > 0 &&
"Expected at least one iteration before first latch");
// In this equation, if we replace the left-hand side with FreqDesired and
// then solve for ProbCompute, we get the ProbCompute formula above.
FreqCompute = (FreqBefore + FreqAfter * ProbCompute) /
(1 - ProbReachingBackedge * ProbCompute);
}
assert(FreqCompute > 0 && "Expected valid frequency");
return FreqCompute;
};
// Determine and set branch weights.
if (CondLatches.size() == 1) {
SetAllProbs(ComputeProbForLinear());
} else if (CondLatches.size() == 2) {
SetAllProbs(ComputeProbForQuadratic());
} else {
// The polynomial is too complex for a simple formula, so the quick and
// dirty fix has been selected. Adjust probabilities starting from the
// first latch, which has the most influence on the total frequency, so
// starting there should minimize the number of latches that have to be
// visited. We do have to iterate because the first latch alone might not
// be enough. For example, we might need to set all probabilities to 1 if
// the frequency is the unroll factor.
double ProbBefore = -1, ProbAfter = -1; // Inits expected to be unused.
double FreqBefore = -1, FreqAfter = -1; // Inits expected to be unused.
for (unsigned I = 0; I != CondLatches.size(); ++I) {
double Freq = AdjustProb(I, ProbBefore, ProbAfter, FreqBefore, FreqAfter);
if (fabs(Freq - FreqDesired) < FreqPrec)
break;
}
}
// FIXME: We have not considered non-latch loop exits:
// - Their original probabilities are not considered in our calculation of
// FreqDesired.
// - Their probabilities are not considered in our probability model used to
// determine new probabilities for remaining conditional branches.
// - If they are conditional and LoopUnroll converts them to unconditional,
// LoopUnroll has proven their original probabilities are incorrect for some
// original loop iterations, but that does not cause ProbUpdateRequired to
// be set to true.
//
// To adjust FreqDesired and our probability model correctly for a non-latch
// loop exit, we would need to compute the original probability that the exit
// is reached from the loop header (in contrast, we currently assume that
// probability is 1 in the case of a latch exit) and the probability that the
// exit is taken if it is conditional (use the branch's old or new weights for
// FreqDesired or the probability model, respectively). Does computing the
// reaching probability require a CFG traversal, or is there some existing
// library that can do it? Prior discussions suggest some such libraries are
// difficult to use within LoopUnroll:
// <https://github.com/llvm/llvm-project/pull/164799#issuecomment-3438681519>.
// For now, we just let our corrected probabilities be less accurate in that
// scenario. Alternatively, we could refuse to correct probabilities at all
// in that scenario, but that seems worse.
}
/// Unroll the given loop by Count. The loop must be in LCSSA form. Unrolling
/// can only fail when the loop's latch block is not terminated by a conditional
/// branch instruction. However, if the trip count (and multiple) are not known,
/// loop unrolling will mostly produce more code that is no faster.
///
/// If Runtime is true then UnrollLoop will try to insert a prologue or
/// epilogue that ensures the latch has a trip multiple of Count. UnrollLoop
/// will not runtime-unroll the loop if computing the run-time trip count will
/// be expensive and AllowExpensiveTripCount is false.
///
/// The LoopInfo Analysis that is passed will be kept consistent.
///
/// This utility preserves LoopInfo. It will also preserve ScalarEvolution and
/// DominatorTree if they are non-null.
///
/// If RemainderLoop is non-null, it will receive the remainder loop (if
/// required and not fully unrolled).
LoopUnrollResult
llvm::UnrollLoop(Loop *L, UnrollLoopOptions ULO, LoopInfo *LI,
ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC,
const TargetTransformInfo *TTI, OptimizationRemarkEmitter *ORE,
bool PreserveLCSSA, Loop **RemainderLoop, AAResults *AA) {
assert(DT && "DomTree is required");
if (!L->getLoopPreheader()) {
LLVM_DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
return LoopUnrollResult::Unmodified;
}
if (!L->getLoopLatch()) {
LLVM_DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
return LoopUnrollResult::Unmodified;
}
// Loops with indirectbr cannot be cloned.
if (!L->isSafeToClone()) {
LLVM_DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n");
return LoopUnrollResult::Unmodified;
}
if (L->getHeader()->hasAddressTaken()) {
// The loop-rotate pass can be helpful to avoid this in many cases.
LLVM_DEBUG(
dbgs() << " Won't unroll loop: address of header block is taken.\n");
return LoopUnrollResult::Unmodified;
}
assert(ULO.Count > 0);
// All these values should be taken only after peeling because they might have
// changed.
BasicBlock *Preheader = L->getLoopPreheader();
BasicBlock *Header = L->getHeader();
BasicBlock *LatchBlock = L->getLoopLatch();
SmallVector<BasicBlock *, 4> ExitBlocks;
L->getExitBlocks(ExitBlocks);
std::vector<BasicBlock *> OriginalLoopBlocks = L->getBlocks();
const unsigned MaxTripCount = SE->getSmallConstantMaxTripCount(L);
const bool MaxOrZero = SE->isBackedgeTakenCountMaxOrZero(L);
std::optional<unsigned> OriginalTripCount =
llvm::getLoopEstimatedTripCount(L);
BranchProbability OriginalLoopProb = llvm::getLoopProbability(L);
// Effectively "DCE" unrolled iterations that are beyond the max tripcount
// and will never be executed.
if (MaxTripCount && ULO.Count > MaxTripCount)
ULO.Count = MaxTripCount;
struct ExitInfo {
unsigned TripCount;
unsigned TripMultiple;
unsigned BreakoutTrip;
bool ExitOnTrue;
BasicBlock *FirstExitingBlock = nullptr;
SmallVector<BasicBlock *> ExitingBlocks;
};
MapVector<BasicBlock *, ExitInfo> ExitInfos;
SmallVector<BasicBlock *, 4> ExitingBlocks;
L->getExitingBlocks(ExitingBlocks);
for (auto *ExitingBlock : ExitingBlocks) {
// The folding code is not prepared to deal with non-branch instructions
// right now.
auto *BI = dyn_cast<CondBrInst>(ExitingBlock->getTerminator());
if (!BI)
continue;
ExitInfo &Info = ExitInfos[ExitingBlock];
Info.TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
Info.TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
if (Info.TripCount != 0) {
Info.BreakoutTrip = Info.TripCount % ULO.Count;
Info.TripMultiple = 0;
} else {
Info.BreakoutTrip = Info.TripMultiple =
(unsigned)std::gcd(ULO.Count, Info.TripMultiple);
}
Info.ExitOnTrue = !L->contains(BI->getSuccessor(0));
Info.ExitingBlocks.push_back(ExitingBlock);
LLVM_DEBUG(dbgs() << " Exiting block %" << ExitingBlock->getName()
<< ": TripCount=" << Info.TripCount
<< ", TripMultiple=" << Info.TripMultiple
<< ", BreakoutTrip=" << Info.BreakoutTrip << "\n");
}
// Are we eliminating the loop control altogether? Note that we can know
// we're eliminating the backedge without knowing exactly which iteration
// of the unrolled body exits.
const bool CompletelyUnroll = ULO.Count == MaxTripCount;
const bool PreserveOnlyFirst = CompletelyUnroll && MaxOrZero;
// There's no point in performing runtime unrolling if this unroll count
// results in a full unroll.
if (CompletelyUnroll)
ULO.Runtime = false;
// Go through all exits of L and see if there are any phi-nodes there. We just
// conservatively assume that they're inserted to preserve LCSSA form, which
// means that complete unrolling might break this form. We need to either fix
// it in-place after the transformation, or entirely rebuild LCSSA. TODO: For
// now we just recompute LCSSA for the outer loop, but it should be possible
// to fix it in-place.
bool NeedToFixLCSSA =
PreserveLCSSA && CompletelyUnroll &&
any_of(ExitBlocks,
[](const BasicBlock *BB) { return isa<PHINode>(BB->begin()); });
// The current loop unroll pass can unroll loops that have
// (1) single latch; and
// (2a) latch is unconditional; or
// (2b) latch is conditional and is an exiting block
// FIXME: The implementation can be extended to work with more complicated
// cases, e.g. loops with multiple latches.
Instruction *LatchTerm = LatchBlock->getTerminator();
// A conditional branch which exits the loop, which can be optimized to an
// unconditional branch in the unrolled loop in some cases.
bool LatchIsExiting = L->isLoopExiting(LatchBlock);
if (!isa<UncondBrInst>(LatchTerm) &&
!(isa<CondBrInst>(LatchTerm) && LatchIsExiting)) {
LLVM_DEBUG(
dbgs() << "Can't unroll; a conditional latch must exit the loop");
return LoopUnrollResult::Unmodified;
}
bool EpilogProfitability =
UnrollRuntimeEpilog.getNumOccurrences() ? UnrollRuntimeEpilog
: isEpilogProfitable(L);
if (ULO.Runtime &&
!UnrollRuntimeLoopRemainder(
L, ULO.Count, ULO.AllowExpensiveTripCount, EpilogProfitability,
ULO.UnrollRemainder, ULO.ForgetAllSCEV, LI, SE, DT, AC, TTI,
PreserveLCSSA, ULO.SCEVExpansionBudget, ULO.RuntimeUnrollMultiExit,
RemainderLoop, OriginalTripCount, OriginalLoopProb)) {
if (ULO.Force)
ULO.Runtime = false;
else {
LLVM_DEBUG(dbgs() << "Won't unroll; remainder loop could not be "
"generated when assuming runtime trip count\n");
return LoopUnrollResult::Unmodified;
}
}
using namespace ore;
// Determine whether this loop originated from the vectorizer so we can
// produce more informative remarks.
StringRef LoopKind = getLoopVectorizeKindPrefix(L);
// Report the unrolling decision.
if (CompletelyUnroll) {
LLVM_DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
<< " with trip count " << ULO.Count << "!\n");
if (ORE)
ORE->emit([&]() {
return OptimizationRemark(DEBUG_TYPE, "FullyUnrolled", L->getStartLoc(),
L->getHeader())
<< "completely unrolled " + LoopKind.str() + "loop with "
<< NV("UnrollCount", ULO.Count) << " iterations";
});
} else {
LLVM_DEBUG({
dbgs() << "UNROLLING loop %" << Header->getName() << " by " << ULO.Count;
if (ULO.Runtime) {
dbgs() << " with run-time trip count";
if (ULO.UnrollRemainder)
dbgs() << " (remainder unrolled)";
}
dbgs() << "!\n";
});
if (ORE)
ORE->emit([&]() {
OptimizationRemark Diag(DEBUG_TYPE, "PartialUnrolled", L->getStartLoc(),
L->getHeader());
Diag << "unrolled " + LoopKind.str() + "loop by a factor of "
<< NV("UnrollCount", ULO.Count);
if (ULO.Runtime)
Diag << " with run-time trip count"
<< (ULO.UnrollRemainder ? " (remainder unrolled)" : "");
return Diag;
});
}
// We are going to make changes to this loop. SCEV may be keeping cached info
// about it, in particular about backedge taken count. The changes we make
// are guaranteed to invalidate this information for our loop. It is tempting
// to only invalidate the loop being unrolled, but it is incorrect as long as
// all exiting branches from all inner loops have impact on the outer loops,
// and if something changes inside them then any of outer loops may also
// change. When we forget outermost loop, we also forget all contained loops
// and this is what we need here.
if (SE) {
if (ULO.ForgetAllSCEV)
SE->forgetAllLoops();
else {
SE->forgetTopmostLoop(L);
SE->forgetBlockAndLoopDispositions();
}
}
if (!LatchIsExiting)
++NumUnrolledNotLatch;
// For the first iteration of the loop, we should use the precloned values for
// PHI nodes. Insert associations now.
ValueToValueMapTy LastValueMap;
std::vector<PHINode*> OrigPHINode;
for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
OrigPHINode.push_back(cast<PHINode>(I));
}
// Collect phi nodes for reductions for which we can introduce multiple
// parallel reduction phis and compute the final reduction result after the
// loop. This requires a single exit block after unrolling. This is ensured by
// restricting to single-block loops where the unrolled iterations are known
// to not exit.
DenseMap<PHINode *, RecurrenceDescriptor> Reductions;
bool CanAddAdditionalAccumulators =
(UnrollAddParallelReductions.getNumOccurrences() > 0
? UnrollAddParallelReductions
: ULO.AddAdditionalAccumulators) &&
!CompletelyUnroll && L->getNumBlocks() == 1 &&
(ULO.Runtime ||
(ExitInfos.contains(Header) && ((ExitInfos[Header].TripCount != 0 &&
ExitInfos[Header].BreakoutTrip == 0))));
// Limit parallelizing reductions to unroll counts of 4 or less for now.
// TODO: The number of parallel reductions should depend on the number of
// execution units. We also don't have to add a parallel reduction phi per
// unrolled iteration, but could for example add a parallel phi for every 2
// unrolled iterations.
if (CanAddAdditionalAccumulators && ULO.Count <= 4) {
for (PHINode &Phi : Header->phis()) {
auto RdxDesc = canParallelizeReductionWhenUnrolling(Phi, L, SE);
if (!RdxDesc)
continue;
// Only handle duplicate phis for a single reduction for now.
// TODO: Handle any number of reductions
if (!Reductions.empty())
continue;
Reductions[&Phi] = *RdxDesc;
}
}
std::vector<BasicBlock *> Headers;
std::vector<BasicBlock *> Latches;
Headers.push_back(Header);
Latches.push_back(LatchBlock);
// The current on-the-fly SSA update requires blocks to be processed in
// reverse postorder so that LastValueMap contains the correct value at each
// exit.
LoopBlocksDFS DFS(L);
DFS.perform(LI);
// Stash the DFS iterators before adding blocks to the loop.
LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
std::vector<BasicBlock*> UnrolledLoopBlocks = L->getBlocks();
// Loop Unrolling might create new loops. While we do preserve LoopInfo, we
// might break loop-simplified form for these loops (as they, e.g., would
// share the same exit blocks). We'll keep track of loops for which we can
// break this so that later we can re-simplify them.
SmallSetVector<Loop *, 4> LoopsToSimplify;
LoopsToSimplify.insert_range(*L);
// When a FSDiscriminator is enabled, we don't need to add the multiply
// factors to the discriminators.
if (Header->getParent()->shouldEmitDebugInfoForProfiling() &&
!EnableFSDiscriminator)
for (BasicBlock *BB : L->getBlocks())
for (Instruction &I : *BB)
if (!I.isDebugOrPseudoInst())
if (const DILocation *DIL = I.getDebugLoc()) {
auto NewDIL = DIL->cloneByMultiplyingDuplicationFactor(ULO.Count);
if (NewDIL)
I.setDebugLoc(*NewDIL);
else
LLVM_DEBUG(dbgs()
<< "Failed to create new discriminator: "
<< DIL->getFilename() << " Line: " << DIL->getLine());
}
// Identify what noalias metadata is inside the loop: if it is inside the
// loop, the associated metadata must be cloned for each iteration.
SmallVector<MDNode *, 6> LoopLocalNoAliasDeclScopes;
identifyNoAliasScopesToClone(L->getBlocks(), LoopLocalNoAliasDeclScopes);
// We place the unrolled iterations immediately after the original loop
// latch. This is a reasonable default placement if we don't have block
// frequencies, and if we do, well the layout will be adjusted later.
auto BlockInsertPt = std::next(LatchBlock->getIterator());
SmallVector<Instruction *> PartialReductions;
for (unsigned It = 1; It != ULO.Count; ++It) {
SmallVector<BasicBlock *, 8> NewBlocks;
SmallDenseMap<const Loop *, Loop *, 4> NewLoops;
NewLoops[L] = L;
for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
ValueToValueMapTy VMap;
BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
Header->getParent()->insert(BlockInsertPt, New);
assert((*BB != Header || LI->getLoopFor(*BB) == L) &&
"Header should not be in a sub-loop");
// Tell LI about New.
const Loop *OldLoop = addClonedBlockToLoopInfo(*BB, New, LI, NewLoops);
if (OldLoop)
LoopsToSimplify.insert(NewLoops[OldLoop]);
if (*BB == Header) {
// Loop over all of the PHI nodes in the block, changing them to use
// the incoming values from the previous block.
for (PHINode *OrigPHI : OrigPHINode) {
PHINode *NewPHI = cast<PHINode>(VMap[OrigPHI]);
Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
// Use cloned phis as parallel phis for partial reductions, which will
// get combined to the final reduction result after the loop.
if (Reductions.contains(OrigPHI)) {
// Collect partial reduction results.
if (PartialReductions.empty())
PartialReductions.push_back(cast<Instruction>(InVal));
PartialReductions.push_back(cast<Instruction>(VMap[InVal]));
// Update the start value for the cloned phis to use the identity
// value for the reduction.
const RecurrenceDescriptor &RdxDesc = Reductions[OrigPHI];
NewPHI->setIncomingValueForBlock(
L->getLoopPreheader(),
getRecurrenceIdentity(RdxDesc.getRecurrenceKind(),
OrigPHI->getType(),
RdxDesc.getFastMathFlags()));
// Update NewPHI to use the cloned value for the iteration and move
// to header.
NewPHI->replaceUsesOfWith(InVal, VMap[InVal]);
NewPHI->moveBefore(OrigPHI->getIterator());
continue;
}
if (Instruction *InValI = dyn_cast<Instruction>(InVal))
if (It > 1 && L->contains(InValI))
InVal = LastValueMap[InValI];
VMap[OrigPHI] = InVal;
NewPHI->eraseFromParent();
}
// Eliminate copies of the loop heart intrinsic, if any.
if (ULO.Heart) {
auto it = VMap.find(ULO.Heart);
assert(it != VMap.end());
Instruction *heartCopy = cast<Instruction>(it->second);
heartCopy->eraseFromParent();
VMap.erase(it);
}
}
// Remap source location atom instance. Do this now, rather than
// when we remap instructions, because remap is called once we've
// cloned all blocks (all the clones would get the same atom
// number).
if (!VMap.AtomMap.empty())
for (Instruction &I : *New)
RemapSourceAtom(&I, VMap);
// Update our running map of newest clones
LastValueMap[*BB] = New;
for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
VI != VE; ++VI)
LastValueMap[VI->first] = VI->second;
// Add phi entries for newly created values to all exit blocks.
for (BasicBlock *Succ : successors(*BB)) {
if (L->contains(Succ))
continue;
for (PHINode &PHI : Succ->phis()) {
Value *Incoming = PHI.getIncomingValueForBlock(*BB);
ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
if (It != LastValueMap.end())
Incoming = It->second;
PHI.addIncoming(Incoming, New);
SE->forgetLcssaPhiWithNewPredecessor(L, &PHI);
}
}
// Keep track of new headers and latches as we create them, so that
// we can insert the proper branches later.
if (*BB == Header)
Headers.push_back(New);
if (*BB == LatchBlock)
Latches.push_back(New);
// Keep track of the exiting block and its successor block contained in
// the loop for the current iteration.
auto ExitInfoIt = ExitInfos.find(*BB);
if (ExitInfoIt != ExitInfos.end())
ExitInfoIt->second.ExitingBlocks.push_back(New);
NewBlocks.push_back(New);
UnrolledLoopBlocks.push_back(New);
// Update DomTree: since we just copy the loop body, and each copy has a
// dedicated entry block (copy of the header block), this header's copy
// dominates all copied blocks. That means, dominance relations in the
// copied body are the same as in the original body.
if (*BB == Header)
DT->addNewBlock(New, Latches[It - 1]);
else {
auto BBDomNode = DT->getNode(*BB);
auto BBIDom = BBDomNode->getIDom();
BasicBlock *OriginalBBIDom = BBIDom->getBlock();
DT->addNewBlock(
New, cast<BasicBlock>(LastValueMap[cast<Value>(OriginalBBIDom)]));
}
}
// Remap all instructions in the most recent iteration.
// Key Instructions: Nothing to do - we've already remapped the atoms.
remapInstructionsInBlocks(NewBlocks, LastValueMap);
for (BasicBlock *NewBlock : NewBlocks)
for (Instruction &I : *NewBlock)
if (auto *II = dyn_cast<AssumeInst>(&I))
AC->registerAssumption(II);
{
// Identify what other metadata depends on the cloned version. After
// cloning, replace the metadata with the corrected version for both
// memory instructions and noalias intrinsics.
std::string ext = (Twine("It") + Twine(It)).str();
cloneAndAdaptNoAliasScopes(LoopLocalNoAliasDeclScopes, NewBlocks,
Header->getContext(), ext);
}
}
// Loop over the PHI nodes in the original block, setting incoming values.
for (PHINode *PN : OrigPHINode) {
if (CompletelyUnroll) {
PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
PN->eraseFromParent();
} else if (ULO.Count > 1) {
if (Reductions.contains(PN))
continue;
Value *InVal = PN->removeIncomingValue(LatchBlock, false);
// If this value was defined in the loop, take the value defined by the
// last iteration of the loop.
if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
if (L->contains(InValI))
InVal = LastValueMap[InVal];
}
assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
PN->addIncoming(InVal, Latches.back());
}
}
// Connect latches of the unrolled iterations to the headers of the next
// iteration. Currently they point to the header of the same iteration.
for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
unsigned j = (i + 1) % e;
Latches[i]->getTerminator()->replaceSuccessorWith(Headers[i], Headers[j]);
}
// Remove loop metadata copied from the original loop latch to branches that
// are no longer latches.
for (unsigned I = 0, E = Latches.size() - (CompletelyUnroll ? 0 : 1); I < E;
++I)
Latches[I]->getTerminator()->setMetadata(LLVMContext::MD_loop, nullptr);
// Update dominators of blocks we might reach through exits.
// Immediate dominator of such block might change, because we add more
// routes which can lead to the exit: we can now reach it from the copied
// iterations too.
if (ULO.Count > 1) {
for (auto *BB : OriginalLoopBlocks) {
auto *BBDomNode = DT->getNode(BB);
SmallVector<BasicBlock *, 16> ChildrenToUpdate;
for (auto *ChildDomNode : BBDomNode->children()) {
auto *ChildBB = ChildDomNode->getBlock();
if (!L->contains(ChildBB))
ChildrenToUpdate.push_back(ChildBB);
}
// The new idom of the block will be the nearest common dominator
// of all copies of the previous idom. This is equivalent to the
// nearest common dominator of the previous idom and the first latch,
// which dominates all copies of the previous idom.
BasicBlock *NewIDom = DT->findNearestCommonDominator(BB, LatchBlock);
for (auto *ChildBB : ChildrenToUpdate)
DT->changeImmediateDominator(ChildBB, NewIDom);
}
}
assert(!UnrollVerifyDomtree ||
DT->verify(DominatorTree::VerificationLevel::Fast));
SmallVector<DominatorTree::UpdateType> DTUpdates;
auto SetDest = [&](BasicBlock *Src, bool WillExit, bool ExitOnTrue) {
auto *Term = cast<CondBrInst>(Src->getTerminator());
const unsigned Idx = ExitOnTrue ^ WillExit;
BasicBlock *Dest = Term->getSuccessor(Idx);
BasicBlock *DeadSucc = Term->getSuccessor(1-Idx);
// Remove predecessors from all non-Dest successors.
DeadSucc->removePredecessor(Src, /* KeepOneInputPHIs */ true);
// Replace the conditional branch with an unconditional one.
auto *BI = UncondBrInst::Create(Dest, Term->getIterator());
BI->setDebugLoc(Term->getDebugLoc());
Term->eraseFromParent();
DTUpdates.emplace_back(DominatorTree::Delete, Src, DeadSucc);
};
auto WillExit = [&](const ExitInfo &Info, unsigned i, unsigned j,
bool IsLatch) -> std::optional<bool> {
if (CompletelyUnroll) {
if (PreserveOnlyFirst) {
if (i == 0)
return std::nullopt;
return j == 0;
}
// Complete (but possibly inexact) unrolling
if (j == 0)
return true;
if (Info.TripCount && j != Info.TripCount)
return false;
return std::nullopt;
}
if (ULO.Runtime) {
// If runtime unrolling inserts a prologue, information about non-latch
// exits may be stale.
if (IsLatch && j != 0)
return false;
return std::nullopt;
}
if (j != Info.BreakoutTrip &&
(Info.TripMultiple == 0 || j % Info.TripMultiple != 0)) {
// If we know the trip count or a multiple of it, we can safely use an
// unconditional branch for some iterations.
return false;
}
return std::nullopt;
};
// Fold branches for iterations where we know that they will exit or not
// exit. In the case of an iteration's latch, if we thus find
// *OriginalLoopProb is incorrect, set ProbUpdateRequired to true.
bool ProbUpdateRequired = false;
for (auto &Pair : ExitInfos) {
ExitInfo &Info = Pair.second;
for (unsigned i = 0, e = Info.ExitingBlocks.size(); i != e; ++i) {
// The branch destination.
unsigned j = (i + 1) % e;
bool IsLatch = Pair.first == LatchBlock;
std::optional<bool> KnownWillExit = WillExit(Info, i, j, IsLatch);
if (!KnownWillExit) {
if (!Info.FirstExitingBlock)
Info.FirstExitingBlock = Info.ExitingBlocks[i];
continue;
}
// We don't fold known-exiting branches for non-latch exits here,
// because this ensures that both all loop blocks and all exit blocks
// remain reachable in the CFG.
// TODO: We could fold these branches, but it would require much more
// sophisticated updates to LoopInfo.
if (*KnownWillExit && !IsLatch) {
if (!Info.FirstExitingBlock)
Info.FirstExitingBlock = Info.ExitingBlocks[i];
continue;
}
// For a latch, record any OriginalLoopProb contradiction.
if (!OriginalLoopProb.isUnknown() && IsLatch) {
BranchProbability ActualProb = *KnownWillExit
? BranchProbability::getZero()
: BranchProbability::getOne();
ProbUpdateRequired |= OriginalLoopProb != ActualProb;
}
SetDest(Info.ExitingBlocks[i], *KnownWillExit, Info.ExitOnTrue);
}
}
DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
DomTreeUpdater *DTUToUse = &DTU;
if (ExitingBlocks.size() == 1 && ExitInfos.size() == 1) {
// Manually update the DT if there's a single exiting node. In that case
// there's a single exit node and it is sufficient to update the nodes
// immediately dominated by the original exiting block. They will become
// dominated by the first exiting block that leaves the loop after
// unrolling. Note that the CFG inside the loop does not change, so there's
// no need to update the DT inside the unrolled loop.
DTUToUse = nullptr;
auto &[OriginalExit, Info] = *ExitInfos.begin();
if (!Info.FirstExitingBlock)
Info.FirstExitingBlock = Info.ExitingBlocks.back();
for (auto *C : to_vector(DT->getNode(OriginalExit)->children())) {
if (L->contains(C->getBlock()))
continue;
C->setIDom(DT->getNode(Info.FirstExitingBlock));
}
} else {
DTU.applyUpdates(DTUpdates);
}
// When completely unrolling, the last latch becomes unreachable.
if (!LatchIsExiting && CompletelyUnroll) {
// There is no need to update the DT here, because there must be a unique
// latch. Hence if the latch is not exiting it must directly branch back to
// the original loop header and does not dominate any nodes.
assert(LatchBlock->getSingleSuccessor() && "Loop with multiple latches?");
changeToUnreachable(Latches.back()->getTerminator(), PreserveLCSSA);
}
// After merging adjacent blocks in Latches below:
// - CondLatches will list the blocks from Latches that are still terminated
// with conditional branches.
// - For 1 <= I < CondLatches.size(), IterCounts[I] will store the number of
// the original loop iterations through which control flows from
// CondLatches[I-1] to CondLatches[I].
// - For I == 0 or I == CondLatches.size(), IterCounts[I] will store the
// number of the original loop iterations through which control can flow
// before CondLatches.front() or after CondLatches.back(), respectively,
// without taking the unrolled loop's backedge, if any.
// - CondLatchNexts[I] will store the CondLatches[I] branch target for the
// next of the original loop's iterations (as opposed to the exit target).
assert(ULO.Count == Latches.size() &&
"Expected one latch block per unrolled iteration");
std::vector<unsigned> IterCounts(1, 0);
std::vector<BasicBlock *> CondLatches;
std::vector<BasicBlock *> CondLatchNexts;
IterCounts.reserve(Latches.size() + 1);
CondLatches.reserve(Latches.size());
CondLatchNexts.reserve(Latches.size());
// Merge adjacent basic blocks, if possible.
for (auto [I, Latch] : enumerate(Latches)) {
++IterCounts.back();
assert((isa<UncondBrInst, CondBrInst>(Latch->getTerminator()) ||
(CompletelyUnroll && !LatchIsExiting && Latch == Latches.back())) &&
"Need a branch as terminator, except when fully unrolling with "
"unconditional latch");
if (auto *Term = dyn_cast<UncondBrInst>(Latch->getTerminator())) {
BasicBlock *Dest = Term->getSuccessor();
BasicBlock *Fold = Dest->getUniquePredecessor();
if (MergeBlockIntoPredecessor(Dest, /*DTU=*/DTUToUse, LI,
/*MSSAU=*/nullptr, /*MemDep=*/nullptr,
/*PredecessorWithTwoSuccessors=*/false,
DTUToUse ? nullptr : DT)) {
// Dest has been folded into Fold. Update our worklists accordingly.
llvm::replace(Latches, Dest, Fold);
llvm::erase(UnrolledLoopBlocks, Dest);
}
} else if (isa<CondBrInst>(Latch->getTerminator())) {
IterCounts.push_back(0);
CondLatches.push_back(Latch);
CondLatchNexts.push_back(Headers[(I + 1) % Latches.size()]);
}
}
// Fix probabilities we contradicted above.
if (ProbUpdateRequired) {
fixProbContradiction(L, ULO, ORE, OriginalLoopProb, CompletelyUnroll,
IterCounts, CondLatches, CondLatchNexts);
}
// If there are partial reductions, create code in the exit block to compute
// the final result and update users of the final result.
if (!PartialReductions.empty()) {
BasicBlock *ExitBlock = L->getExitBlock();
assert(ExitBlock &&
"Can only introduce parallel reduction phis with single exit block");
assert(Reductions.size() == 1 &&
"currently only a single reduction is supported");
Value *FinalRdxValue = PartialReductions.back();
Value *RdxResult = nullptr;
for (PHINode &Phi : ExitBlock->phis()) {
if (Phi.getIncomingValueForBlock(L->getLoopLatch()) != FinalRdxValue)
continue;
if (!RdxResult) {
RdxResult = PartialReductions.front();
IRBuilder Builder(ExitBlock, ExitBlock->getFirstNonPHIIt());
Builder.setFastMathFlags(Reductions.begin()->second.getFastMathFlags());
RecurKind RK = Reductions.begin()->second.getRecurrenceKind();
for (Instruction *RdxPart : drop_begin(PartialReductions)) {
if (RecurrenceDescriptor::isMinMaxRecurrenceKind(RK))
RdxResult = createMinMaxOp(Builder, RK, RdxResult, RdxPart);
else
RdxResult = Builder.CreateBinOp(
(Instruction::BinaryOps)RecurrenceDescriptor::getOpcode(RK),
RdxPart, RdxResult, "bin.rdx");
}
NeedToFixLCSSA = true;
for (Instruction *RdxPart : PartialReductions)
RdxPart->dropPoisonGeneratingFlags();
}
Phi.replaceAllUsesWith(RdxResult);
}
}
if (DTUToUse) {
// Apply updates to the DomTree.
DT = &DTU.getDomTree();
}
assert(!UnrollVerifyDomtree ||
DT->verify(DominatorTree::VerificationLevel::Fast));
Loop *OuterL = L->getParentLoop();
std::vector<BasicBlock *> Blocks;
// Update LoopInfo if the loop is completely removed.
if (CompletelyUnroll) {
Blocks = L->getBlocks();
LI->erase(L);
// We shouldn't try to use `L` anymore.
L = nullptr;
}
// At this point, the code is well formed. We now simplify the unrolled loop,
// doing constant propagation and dead code elimination as we go.
simplifyLoopAfterUnroll(
L, !CompletelyUnroll && ULO.Count > 1, LI, SE, DT, AC, TTI,
CompletelyUnroll ? ArrayRef<BasicBlock *>(Blocks) : L->getBlocks(), AA);
NumCompletelyUnrolled += CompletelyUnroll;
++NumUnrolled;
if (!CompletelyUnroll) {
// Update metadata for the loop's branch weights and estimated trip count:
// - If ULO.Runtime, UnrollRuntimeLoopRemainder sets the guard branch
// weights, latch branch weights, and estimated trip count of the
// remainder loop it creates. It also sets the branch weights for the
// unrolled loop guard it creates. The branch weights for the unrolled
// loop latch are adjusted below. FIXME: Handle prologue loops.
// - Otherwise, if unrolled loop iteration latches become unconditional,
// branch weights are adjusted by the fixProbContradiction call above.
// - Otherwise, the original loop's branch weights are correct for the
// unrolled loop, so do not adjust them.
// - In all cases, the unrolled loop's estimated trip count is set below.
//
// As an example of the last case, consider what happens if the unroll count
// is 4 for a loop with an estimated trip count of 10 when we do not create
// a remainder loop and all iterations' latches remain conditional. Each
// unrolled iteration's latch still has the same probability of exiting the
// loop as it did when in the original loop, and thus it should still have
// the same branch weights. Each unrolled iteration's non-zero probability
// of exiting already appropriately reduces the probability of reaching the
// remaining iterations just as it did in the original loop. Trying to also
// adjust the branch weights of the final unrolled iteration's latch (i.e.,
// the backedge for the unrolled loop as a whole) to reflect its new trip
// count of 3 will erroneously further reduce its block frequencies.
// However, in case an analysis later needs to estimate the trip count of
// the unrolled loop as a whole without considering the branch weights for
// each unrolled iteration's latch within it, we store the new trip count as
// separate metadata.
if (!OriginalLoopProb.isUnknown() && ULO.Runtime && EpilogProfitability) {
assert((CondLatches.size() == 1 &&
(ProbUpdateRequired || OriginalLoopProb.isOne())) &&
"Expected ULO.Runtime to give unrolled loop 1 conditional latch, "
"the backedge, requiring a probability update unless infinite");
// Where p is always the probability of executing at least 1 more
// iteration, the probability for at least n more iterations is p^n.
setLoopProbability(L, OriginalLoopProb.pow(ULO.Count));
}
if (OriginalTripCount) {
unsigned NewTripCount = *OriginalTripCount / ULO.Count;
if (!ULO.Runtime && *OriginalTripCount % ULO.Count)
++NewTripCount;
setLoopEstimatedTripCount(L, NewTripCount);
}
}
// LoopInfo should not be valid, confirm that.
if (UnrollVerifyLoopInfo)
LI->verify(*DT);
// After complete unrolling most of the blocks should be contained in OuterL.
// However, some of them might happen to be out of OuterL (e.g. if they
// precede a loop exit). In this case we might need to insert PHI nodes in
// order to preserve LCSSA form.
// We don't need to check this if we already know that we need to fix LCSSA
// form.
// TODO: For now we just recompute LCSSA for the outer loop in this case, but
// it should be possible to fix it in-place.
if (PreserveLCSSA && OuterL && CompletelyUnroll && !NeedToFixLCSSA)
NeedToFixLCSSA |= ::needToInsertPhisForLCSSA(OuterL, UnrolledLoopBlocks, LI);
// Make sure that loop-simplify form is preserved. We want to simplify
// at least one layer outside of the loop that was unrolled so that any
// changes to the parent loop exposed by the unrolling are considered.
if (OuterL) {
// OuterL includes all loops for which we can break loop-simplify, so
// it's sufficient to simplify only it (it'll recursively simplify inner
// loops too).
if (NeedToFixLCSSA) {
// LCSSA must be performed on the outermost affected loop. The unrolled
// loop's last loop latch is guaranteed to be in the outermost loop
// after LoopInfo's been updated by LoopInfo::erase.
Loop *LatchLoop = LI->getLoopFor(Latches.back());
Loop *FixLCSSALoop = OuterL;
if (!FixLCSSALoop->contains(LatchLoop))
while (FixLCSSALoop->getParentLoop() != LatchLoop)
FixLCSSALoop = FixLCSSALoop->getParentLoop();
formLCSSARecursively(*FixLCSSALoop, *DT, LI, SE);
} else if (PreserveLCSSA) {
assert(OuterL->isLCSSAForm(*DT) &&
"Loops should be in LCSSA form after loop-unroll.");
}
// TODO: That potentially might be compile-time expensive. We should try
// to fix the loop-simplified form incrementally.
simplifyLoop(OuterL, DT, LI, SE, AC, nullptr, PreserveLCSSA);
} else {
// Simplify loops for which we might've broken loop-simplify form.
for (Loop *SubLoop : LoopsToSimplify)
simplifyLoop(SubLoop, DT, LI, SE, AC, nullptr, PreserveLCSSA);
}
return CompletelyUnroll ? LoopUnrollResult::FullyUnrolled
: LoopUnrollResult::PartiallyUnrolled;
}
/// Given an llvm.loop loop id metadata node, returns the loop hint metadata
/// node with the given name (for example, "llvm.loop.unroll.count"). If no
/// such metadata node exists, then nullptr is returned.
MDNode *llvm::GetUnrollMetadata(MDNode *LoopID, StringRef Name) {
// First operand should refer to the loop id itself.
assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
for (const MDOperand &MDO : llvm::drop_begin(LoopID->operands())) {
MDNode *MD = dyn_cast<MDNode>(MDO);
if (!MD)
continue;
MDString *S = dyn_cast<MDString>(MD->getOperand(0));
if (!S)
continue;
if (Name == S->getString())
return MD;
}
return nullptr;
}
// Returns the loop hint metadata node with the given name (for example,
// "llvm.loop.unroll.count"). If no such metadata node exists, then nullptr is
// returned.
MDNode *llvm::getUnrollMetadataForLoop(const Loop *L, StringRef Name) {
if (MDNode *LoopID = L->getLoopID())
return GetUnrollMetadata(LoopID, Name);
return nullptr;
}
std::optional<RecurrenceDescriptor>
llvm::canParallelizeReductionWhenUnrolling(PHINode &Phi, Loop *L,
ScalarEvolution *SE) {
RecurrenceDescriptor RdxDesc;
if (!RecurrenceDescriptor::isReductionPHI(&Phi, L, RdxDesc,
/*DemandedBits=*/nullptr,
/*AC=*/nullptr, /*DT=*/nullptr, SE))
return std::nullopt;
if (RdxDesc.hasUsesOutsideReductionChain())
return std::nullopt;
RecurKind RK = RdxDesc.getRecurrenceKind();
// Skip unsupported reductions.
// TODO: Handle any-of and find-last reductions.
if (RecurrenceDescriptor::isAnyOfRecurrenceKind(RK) ||
RecurrenceDescriptor::isFindRecurrenceKind(RK))
return std::nullopt;
if (RdxDesc.hasExactFPMath())
return std::nullopt;
if (RdxDesc.IntermediateStore)
return std::nullopt;
BasicBlock *Latch = L->getLoopLatch();
if (!Latch)
return std::nullopt;
Instruction *LatchInst =
cast<Instruction>(Phi.getIncomingValueForBlock(Latch));
// Don't unroll reductions with constant ops; those can be folded to a
// single induction update. For calls (e.g. fmuladd or min/max
// intrinsics), the called function is itself a Constant operand and is
// not a reduction operand, so restrict the check to the argument list.
auto Ops = isa<CallBase>(LatchInst) ? cast<CallBase>(LatchInst)->args()
: LatchInst->operands();
if (any_of(Ops, IsaPred<Constant>))
return std::nullopt;
if (!is_contained(LatchInst->operands(), &Phi))
return std::nullopt;
return RdxDesc;
}