Files
swift-mirror/lib/SILOptimizer/Transforms/StackPromotion.cpp

497 lines
18 KiB
C++

//===------- StackPromotion.cpp - Promotes allocations to the stack -------===//
//
// This source file is part of the Swift.org open source project
//
// Copyright (c) 2014 - 2015 Apple Inc. and the Swift project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See http://swift.org/LICENSE.txt for license information
// See http://swift.org/CONTRIBUTORS.txt for the list of Swift project authors
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "stack-promotion"
#include "swift/SILOptimizer/PassManager/Passes.h"
#include "swift/SILOptimizer/PassManager/Transforms.h"
#include "swift/SILOptimizer/Analysis/EscapeAnalysis.h"
#include "swift/SILOptimizer/Analysis/DominanceAnalysis.h"
#include "swift/SIL/SILArgument.h"
#include "swift/SIL/SILBuilder.h"
#include "llvm/ADT/Statistic.h"
STATISTIC(NumStackPromoted, "Number of objects promoted to the stack");
using namespace swift;
/// Promotes heap allocated objects to the stack.
/// Following types of allocations are handled:
/// *) alloc_ref instructions of native swift classes: if promoted, the [stack]
/// attribute is set in the alloc_ref and a dealloc_ref [stack] is inserted
/// at the end of the object's lifetime.
/// *) Array buffers which are allocated by a call to swift_bufferAllocate: if
/// promoted the swift_bufferAllocate call is replaced by a call to
/// swift_bufferAllocateOnStack and a call to swift_bufferDeallocateFromStack
/// is inserted at the end of the buffer's lifetime.
/// Those calls are lowered by the LLVM SwiftStackPromotion pass.
/// TODO: This is a terrible hack, but necessary because we need constant
/// size and alignment for the final stack promotion decision. The arguments
/// to swift_bufferAllocate in SIL are not constant because they depend on
/// the not-yet-evaluatable sizeof and alignof builtins. Therefore we need
/// LLVM's constant propagation prior to deciding on stack promotion.
/// The solution to this problem is that we need native support for tail-
/// allocated arrays in SIL so that we can do the array buffer allocations
/// with alloc_ref instructions.
class StackPromoter {
// Some analysis we need.
SILFunction *F;
EscapeAnalysis::ConnectionGraph *ConGraph;
DominanceInfo *DT;
PostDominanceInfo *PDT;
EscapeAnalysis *EA;
// Pseudo-functions for (de-)allocating array buffers on the stack.
SILFunction *BufferAllocFunc = nullptr;
SILFunction *BufferDeallocFunc = nullptr;
bool ChangedInsts = false;
bool ChangedCalls = false;
/// Worklist for visiting all blocks.
class WorkListType {
/// The nesting depth of stack allocation instructions for each block.
/// A value of -1 means: not known yet.
llvm::DenseMap<SILBasicBlock *, int> Block2StackDepth;
/// The work list of not yet handled blocks.
llvm::SmallVector<SILBasicBlock *, 8> ToHandle;
public:
bool empty() const { return ToHandle.empty(); }
SILBasicBlock *pop_back_val() { return ToHandle.pop_back_val(); }
/// Insert a block into the worklist and set its stack depth.
void insert(SILBasicBlock *BB, int StackDepth) {
auto Iter = Block2StackDepth.find(BB);
if (Iter != Block2StackDepth.end()) {
// We already handled the block.
assert(StackDepth >= 0);
if (Iter->second < 0) {
// Update the stack depth if we didn't set it yet for the block.
Iter->second = StackDepth;
} else {
assert(Iter->second == StackDepth &&
"inconsistent stack depth at a CFG merge point");
}
} else {
Block2StackDepth[BB] = StackDepth;
ToHandle.push_back(BB);
}
}
int getStackDepth(SILBasicBlock *BB) {
assert(Block2StackDepth.find(BB) != Block2StackDepth.end());
int Depth = Block2StackDepth.lookup(BB);
assert(Depth >= 0 && "EndBlock not reachable from StartBlock");
return Depth;
}
};
/// Tries to promote the allocation \p AI.
void tryPromoteAlloc(SILInstruction *AI);
/// Creates the external declaration for swift_bufferAllocateOnStack.
SILFunction *getBufferAllocFunc(SILFunction *OrigFunc,
SILLocation Loc);
/// Creates the external declaration for swift_bufferDeallocateFromStack.
SILFunction *getBufferDeallocFunc(SILFunction *OrigFunc,
SILLocation Loc);
/// Returns true if the allocation \p AI can be promoted.
/// In this case it sets the \a DeallocInsertionPoint to the instruction
/// where the deallocation must be inserted.
/// It optionally also sets \a AllocInsertionPoint in case the allocation
/// instruction must be moved to another place.
bool canPromoteAlloc(SILInstruction *AI,
SILInstruction *&AllocInsertionPoint,
SILInstruction *&DeallocInsertionPoint);
bool strictlyDominates(SILBasicBlock *A, SILBasicBlock *B) {
return A != B && DT->dominates(A, B);
}
bool strictlyPostDominates(SILBasicBlock *A, SILBasicBlock *B) {
return A != B && PDT->dominates(A, B);
}
SILBasicBlock *getImmediatePostDom(SILBasicBlock *BB) {
auto *Node = PDT->getNode(BB);
if (!Node)
return nullptr;
auto *IDomNode = Node->getIDom();
if (!IDomNode)
return nullptr;
return IDomNode->getBlock();
}
public:
StackPromoter(SILFunction *F, EscapeAnalysis::ConnectionGraph *ConGraph,
DominanceInfo *DT, PostDominanceInfo *PDT,
EscapeAnalysis *EA) :
F(F), ConGraph(ConGraph), DT(DT), PDT(PDT), EA(EA) { }
/// What did the optimization change?
enum class ChangeState {
None,
Insts,
Calls
};
/// The main entry point for the optimization.
ChangeState promote();
};
/// Returns true if instruction \p I is an allocation we can handle.
static bool isPromotableAllocInst(SILInstruction *I) {
// Check for swift object allocation.
if (auto *ARI = dyn_cast<AllocRefInst>(I)) {
if (!ARI->isObjC())
return true;
return false;
}
// Check for array buffer allocation.
auto *AI = dyn_cast<ApplyInst>(I);
if (AI && AI->getNumArguments() == 3) {
if (auto *Callee = AI->getCalleeFunction()) {
if (Callee->getName() == "swift_bufferAllocate")
return true;
}
return false;
}
return false;
}
StackPromoter::ChangeState StackPromoter::promote() {
// Search the whole function for stack promotable allocations.
for (SILBasicBlock &BB : *F) {
for (auto Iter = BB.begin(); Iter != BB.end();) {
// The allocation instruction may be moved, so increment Iter prior to
// doing the optimization.
SILInstruction *I = &*Iter++;
if (isPromotableAllocInst(I)) {
tryPromoteAlloc(I);
}
}
}
if (ChangedCalls)
return ChangeState::Calls;
if (ChangedInsts)
return ChangeState::Insts;
return ChangeState::None;
}
void StackPromoter::tryPromoteAlloc(SILInstruction *I) {
SILInstruction *AllocInsertionPoint = nullptr;
SILInstruction *DeallocInsertionPoint = nullptr;
if (!canPromoteAlloc(I, AllocInsertionPoint, DeallocInsertionPoint))
return;
DEBUG(llvm::dbgs() << "Promoted " << *I);
DEBUG(llvm::dbgs() << " in " << I->getFunction()->getName() << '\n');
NumStackPromoted++;
SILBuilder B(DeallocInsertionPoint);
if (auto *ARI = dyn_cast<AllocRefInst>(I)) {
// It's an object allocation. We set the [stack] attribute in the alloc_ref.
ARI->setStackAllocatable();
if (AllocInsertionPoint)
ARI->moveBefore(AllocInsertionPoint);
/// And create a dealloc_ref [stack] at the end of the object's lifetime.
B.createDeallocRef(I->getLoc(), I, true);
ChangedInsts = true;
return;
}
if (auto *AI = dyn_cast<ApplyInst>(I)) {
assert(!AllocInsertionPoint && "can't move call to swift_bufferAlloc");
// It's an array buffer allocation.
auto *OldFRI = cast<FunctionRefInst>(AI->getCallee());
SILFunction *OldF = OldFRI->getReferencedFunction();
SILLocation loc = (OldF->hasLocation() ? OldF->getLocation() : AI->getLoc());
SILFunction *DeallocFun = getBufferDeallocFunc(OldF, loc);
// We insert a swift_bufferDeallocateFromStack at the end of the buffer's
// lifetime.
auto *DeallocFRI = B.createFunctionRef(OldFRI->getLoc(), DeallocFun);
B.createApply(loc, DeallocFRI, { AI }, false);
// And replace the call to swift_bufferAllocate with a call to
// swift_bufferAllocateOnStack.
B.setInsertionPoint(AI);
auto *AllocFRI = B.createFunctionRef(OldFRI->getLoc(),
getBufferAllocFunc(OldF, loc));
AI->setOperand(0, AllocFRI);
ChangedCalls = true;
return;
}
llvm_unreachable("unhandled allocation instruction");
}
SILFunction *StackPromoter::getBufferAllocFunc(SILFunction *OrigFunc,
SILLocation Loc) {
if (!BufferAllocFunc) {
BufferAllocFunc = OrigFunc->getModule().getOrCreateFunction(
Loc,
"swift_bufferAllocateOnStack",
OrigFunc->getLinkage(),
OrigFunc->getLoweredFunctionType(),
OrigFunc->isBare(), IsNotTransparent,
OrigFunc->isFragile());
}
return BufferAllocFunc;
}
SILFunction *StackPromoter::getBufferDeallocFunc(SILFunction *OrigFunc,
SILLocation Loc) {
if (!BufferDeallocFunc) {
SILModule &M = OrigFunc->getModule();
const ASTContext &Ctx = OrigFunc->getModule().getASTContext();
CanSILFunctionType OrigTy = OrigFunc->getLoweredFunctionType();
CanType ObjectTy = OrigTy->getResult().getType();
// The function type for swift_bufferDeallocateFromStack.
CanSILFunctionType FunTy = SILFunctionType::get(
OrigTy->getGenericSignature(),
OrigTy->getExtInfo(),
OrigTy->getCalleeConvention(),
{ SILParameterInfo(ObjectTy, ParameterConvention::Direct_Guaranteed) },
SILResultInfo(TupleType::getEmpty(Ctx), ResultConvention::Owned),
OrigTy->getOptionalErrorResult(),
M.getASTContext());
BufferDeallocFunc = M.getOrCreateFunction(
Loc,
"swift_bufferDeallocateFromStack",
OrigFunc->getLinkage(),
FunTy,
OrigFunc->isBare(), IsNotTransparent, OrigFunc->isFragile());
}
return BufferDeallocFunc;
}
bool StackPromoter::canPromoteAlloc(SILInstruction *AI,
SILInstruction *&AllocInsertionPoint,
SILInstruction *&DeallocInsertionPoint) {
AllocInsertionPoint = nullptr;
DeallocInsertionPoint = nullptr;
auto *Node = ConGraph->getNodeOrNull(AI, EA);
if (!Node)
return false;
// The most important check: does the object escape the current function?
if (Node->escapes())
return false;
// Now we have to determine the lifetime of the allocated object in its
// function.
// Get all interesting uses of the object (e.g. release instructions). This
// includes uses of objects where the allocation is stored to.
int NumUsePointsToFind = ConGraph->getNumUsePoints(Node);
if (NumUsePointsToFind == 0) {
// There should always be at least one release for an allocated object.
// But in case all pathes from this block end in unreachable then the
// final release of the object may be optimized away. We bail out in this
// case.
return false;
}
// In the following we check two requirements for stack promotion:
// 1) Are all uses in the same control region as the alloc? E.g. if the
// allocation is in a loop then there may not be any uses of the object
// outside the loop.
// 2) We need to find an insertion place for the deallocation so that it
// preserves a properly nested stack allocation-deallocation structure.
SILBasicBlock *StartBlock = AI->getParent();
// The block where we assume we can insert the deallocation.
SILBasicBlock *EndBlock = StartBlock;
// We visit all instructions starting at the allocation instruction.
WorkListType WorkList;
// It's important that the EndBlock is at the head of the WorkList so that
// we handle it after all other blocks.
WorkList.insert(EndBlock, -1);
WorkList.insert(StartBlock, 0);
for (;;) {
SILBasicBlock *BB = WorkList.pop_back_val();
int StackDepth = 0;
SILBasicBlock::iterator Iter;
if (BB == StartBlock) {
// In the first block we start at the allocation instruction and not at
// the begin of the block.
Iter = AI->getIterator();
} else {
// Track all uses in the block arguments.
for (SILArgument *BBArg : BB->getBBArgs()) {
if (ConGraph->isUsePoint(BBArg, Node))
NumUsePointsToFind--;
}
// Make sure that the EndBlock is not inside a loop (which does not
// contain the StartBlock).
// E.g.:
// %obj = alloc_ref // the allocation
// br loop
// loop:
// the_only_use_of_obj(%obj)
// cond_br ..., loop, exit
// exit:
// ... // this is the new EndBlock
for (SILBasicBlock *Pred : BB->getPreds()) {
// Extend the lifetime region until the EndBlock post dominates the
// StartBlock.
while (!strictlyPostDominates(EndBlock, Pred)) {
EndBlock = getImmediatePostDom(EndBlock);
if (!EndBlock)
return false;
}
}
Iter = BB->begin();
StackDepth = WorkList.getStackDepth(BB);
}
// Visit all instructions of the current block.
while (Iter != BB->end()) {
SILInstruction &I = *Iter++;
if (BB == EndBlock && StackDepth == 0 && NumUsePointsToFind == 0) {
// We found a place to insert the stack deallocation.
DeallocInsertionPoint = &I;
return true;
}
if (I.isAllocatingStack()) {
StackDepth++;
} else if (I.isDeallocatingStack()) {
if (StackDepth == 0) {
// The allocation is inside a stack alloc-dealloc region and we are
// now leaving this region without having found a place for the
// deallocation. E.g.
// E.g.:
// %1 = alloc_stack
// %obj = alloc_ref // the allocation
// dealloc_stack %1
// use_of_obj(%obj)
//
// In this case we can move the alloc_ref before the alloc_stack
// to fix the nesting.
if (!isa<AllocRefInst>(AI))
return false;
auto *Alloc = dyn_cast<SILInstruction>(I.getOperand(0).getDef());
if (!Alloc)
return false;
// This should always be the case, but let's be on the safe side.
if (!PDT->dominates(StartBlock, Alloc->getParent()))
return false;
AllocInsertionPoint = Alloc;
StackDepth++;
}
StackDepth--;
}
// Track a use.
if (ConGraph->isUsePoint(&I, Node) != 0)
NumUsePointsToFind--;
}
if (WorkList.empty()) {
if (EndBlock == BB) {
// We reached the EndBlock but didn't find a place for the deallocation
// so far (because we didn't find all uses yet or we entered a another
// stack alloc-dealloc region). Let's extend our lifetime region.
// E.g.:
// %obj = alloc_ref // the allocation
// %1 = alloc_stack
// use_of_obj(%obj) // can't insert the deallocation in this block
// cond_br ..., bb1, bb2
// bb1:
// ...
// br bb2
// bb2:
// dealloc_stack %1 // this is the new EndBlock
EndBlock = getImmediatePostDom(EndBlock);
if (!EndBlock)
return false;
}
// Again, it's important that the EndBlock is the first in the WorkList.
WorkList.insert(EndBlock, -1);
}
// Push the successor blocks to the WorkList.
for (SILBasicBlock *Succ : BB->getSuccessors()) {
if (!strictlyDominates(StartBlock, Succ)) {
// The StartBlock is inside a loop but we couldn't find a deallocation
// place in this loop, e.g. because there are uses outside the loop.
// E.g.:
// %container = alloc_ref
// br loop
// loop:
// %obj = alloc_ref // the allocation
// store %obj to %some_field_in_container
// cond_br ..., loop, exit
// exit:
// use(%container)
return false;
}
WorkList.insert(Succ, StackDepth);
}
}
}
//===----------------------------------------------------------------------===//
// Top Level Driver
//===----------------------------------------------------------------------===//
namespace {
class StackPromotion : public SILFunctionTransform {
public:
StackPromotion() {}
private:
/// The entry point to the transformation.
void run() override {
DEBUG(llvm::dbgs() << "** StackPromotion **\n");
auto *EA = PM->getAnalysis<EscapeAnalysis>();
auto *DA = PM->getAnalysis<DominanceAnalysis>();
auto *PDA = PM->getAnalysis<PostDominanceAnalysis>();
SILFunction *F = getFunction();
if (auto *ConGraph = EA->getConnectionGraph(F)) {
StackPromoter promoter(F, ConGraph, DA->get(F), PDA->get(F), EA);
switch (promoter.promote()) {
case StackPromoter::ChangeState::None:
break;
case StackPromoter::ChangeState::Insts:
invalidateAnalysis(SILAnalysis::InvalidationKind::Instructions);
break;
case StackPromoter::ChangeState::Calls: {
invalidateAnalysis(SILAnalysis::InvalidationKind::CallsAndInstructions);
}
break;
}
}
}
StringRef getName() override { return "StackPromotion"; }
};
} // end anonymous namespace
SILTransform *swift::createStackPromotion() {
return new StackPromotion();
}