mirror of
https://github.com/apple/swift.git
synced 2026-06-20 15:42:51 +02:00
097b0d3400
We cannot use spare bits or other overlapping storage layout tricks with fundamentally address-only enums, and we can take advantage of this to do borrowing switches or other in-place projections without copying the value. However, for resilient enums, the implementation may use spare bit packing, but the type must be handled address-only outside of its defining module, and we didn't have a way to express that with borrowing switch. Optimization passes have also been running into problems with the complexity that we were using `unchecked_take_enum_data_addr` sometimes as a pure operation. This patch splits the instruction into three: - `unchecked_inplace_enum_data_addr` represents a nondestructive in-place enum projection. It is only allowed for enums whose projection operation is nondestructive. - `unchecked_take_enum_data_addr` represents a destructive enum projection, invalidating the enum and leaving the payload to be further consumed. This matches the current instruction's semantics. - `unchecked_borrow_enum_data_addr` represents a borrowing enum projection. The instruction takes a second operand for "scratch" space, which the enum representation may be copied into in order to avoid invalidating the enum value, so the result is dependent on the lifetime of both the original enum and the scratch buffer. This allows for borrowing switches over resilient enums. `unchecked_borrow_enum_data_addr` is implemented by taking advantage of the "address-only enums can't do spare bit optimization" property at runtime. We inspect the operand type's bitwise-borrowability from its metadata. If the type is bitwise-borrowable, then we are allowed to bitwise-copy the enum to the scratch space and apply the projection to the scratch space, preserving the original value. If the type is not bitwise-borrowable, then we cannot use spare bit optimization in its layout, so we apply the projection in-place. Fixes rdar://174952822.
511 lines
18 KiB
C++
511 lines
18 KiB
C++
//===--- MemoryLocations.cpp ----------------------------------------------===//
|
|
//
|
|
// This source file is part of the Swift.org open source project
|
|
//
|
|
// Copyright (c) 2014 - 2021 Apple Inc. and the Swift project authors
|
|
// Licensed under Apache License v2.0 with Runtime Library Exception
|
|
//
|
|
// See https://swift.org/LICENSE.txt for license information
|
|
// See https://swift.org/CONTRIBUTORS.txt for the list of Swift project authors
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#define DEBUG_TYPE "sil-memory-locations"
|
|
#include "swift/Basic/Assertions.h"
|
|
#include "swift/SIL/MemoryLocations.h"
|
|
#include "swift/Basic/SmallBitVector.h"
|
|
#include "swift/SIL/ApplySite.h"
|
|
#include "swift/SIL/SILBasicBlock.h"
|
|
#include "swift/SIL/SILFunction.h"
|
|
#include "swift/SIL/SILModule.h"
|
|
#include "llvm/Support/raw_ostream.h"
|
|
|
|
using namespace swift;
|
|
|
|
namespace swift {
|
|
namespace {
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Utility functions
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
/// Enlarge the bitset if needed to set the bit with \p idx.
|
|
static void setBitAndResize(SmallBitVector &bits, unsigned idx) {
|
|
if (bits.size() <= idx)
|
|
bits.resize(idx + 1);
|
|
bits.set(idx);
|
|
}
|
|
|
|
static bool allUsesInSameBlock(AllocStackInst *ASI) {
|
|
SILBasicBlock *BB = ASI->getParent();
|
|
int numDeallocStacks = 0;
|
|
for (Operand *use : ASI->getUses()) {
|
|
SILInstruction *user = use->getUser();
|
|
if (isa<DeallocStackInst>(user)) {
|
|
++numDeallocStacks;
|
|
if (user->getParent() != BB)
|
|
return false;
|
|
}
|
|
}
|
|
// In case of an unreachable, the dealloc_stack can be missing. In this
|
|
// case we don't treat it as a single-block location.
|
|
assert(numDeallocStacks <= 1 &&
|
|
"A single-block stack location cannot have multiple deallocations");
|
|
return numDeallocStacks == 1;
|
|
}
|
|
|
|
} // anonymous namespace
|
|
} // namespace swift
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// MemoryLocations members
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
MemoryLocations::Location::Location(SILValue val, unsigned index, int parentIdx) :
|
|
representativeValue(val),
|
|
parentIdx(parentIdx) {
|
|
assert(((parentIdx >= 0) ==
|
|
(isa<StructElementAddrInst>(val) || isa<TupleElementAddrInst>(val) ||
|
|
isa<InitEnumDataAddrInst>(val) ||
|
|
isa<UncheckedEnumDataAddrInstBase>(val) ||
|
|
isa<InitExistentialAddrInst>(val) ||
|
|
isa<OpenExistentialAddrInst>(val) || isa<ApplyInst>(val) ||
|
|
isa<StoreBorrowInst>(val))) &&
|
|
"sub-locations can only be introduced with "
|
|
"struct/tuple/enum/store_borrow/borrow accessor "
|
|
"projections");
|
|
setBitAndResize(subLocations, index);
|
|
setBitAndResize(selfAndParents, index);
|
|
}
|
|
|
|
void MemoryLocations::Location::updateFieldCounters(SILType ty, int increment) {
|
|
SILFunction *function = representativeValue->getFunction();
|
|
if (!ty.isEmpty(*function)) {
|
|
numFieldsNotCoveredBySubfields += increment;
|
|
if (!ty.isTrivial(*function))
|
|
numNonTrivialFieldsNotCovered += increment;
|
|
}
|
|
}
|
|
|
|
static SILValue getBaseValue(SILValue addr) {
|
|
while (true) {
|
|
switch (addr->getKind()) {
|
|
case ValueKind::BeginAccessInst:
|
|
addr = cast<BeginAccessInst>(addr)->getOperand();
|
|
break;
|
|
case ValueKind::MarkDependenceInst:
|
|
addr = cast<MarkDependenceInst>(addr)->getValue();
|
|
break;
|
|
default:
|
|
return addr;
|
|
}
|
|
}
|
|
}
|
|
|
|
int MemoryLocations::getLocationIdx(SILValue addr) const {
|
|
auto iter = addr2LocIdx.find(getBaseValue(addr));
|
|
if (iter == addr2LocIdx.end())
|
|
return -1;
|
|
return iter->second;
|
|
}
|
|
|
|
const MemoryLocations::Location *
|
|
MemoryLocations::getRootLocation(unsigned index) const {
|
|
while (true) {
|
|
const Location &loc = locations[index];
|
|
if (loc.parentIdx < 0)
|
|
return &loc;
|
|
index = loc.parentIdx;
|
|
}
|
|
}
|
|
|
|
static bool canHandleAllocStack(AllocStackInst *asi) {
|
|
assert(asi);
|
|
|
|
// An alloc_stack with dynamic lifetime set has a lifetime that relies on
|
|
// unrelated conditional control flow for correctness. This means that we may
|
|
// statically leak along paths that were known by the emitter to never be
|
|
// taken if the value is live. So bail since we can't verify this.
|
|
if (asi->hasDynamicLifetime())
|
|
return false;
|
|
|
|
// Otherwise we can optimize!
|
|
return true;
|
|
}
|
|
|
|
void MemoryLocations::analyzeLocations(SILFunction *function) {
|
|
// As we have to limit the set of handled locations to memory, which is
|
|
// guaranteed to be not aliased, we currently only handle indirect function
|
|
// arguments and alloc_stack locations.
|
|
for (SILArgument *arg : function->getArguments()) {
|
|
SILFunctionArgument *funcArg = cast<SILFunctionArgument>(arg);
|
|
switch (funcArg->getArgumentConvention()) {
|
|
case SILArgumentConvention::Indirect_In:
|
|
case SILArgumentConvention::Indirect_In_Guaranteed:
|
|
case SILArgumentConvention::Indirect_Out:
|
|
// These are not SIL addresses under -enable-sil-opaque-values
|
|
if (!function->getConventions().useLoweredAddresses())
|
|
break;
|
|
|
|
LLVM_FALLTHROUGH;
|
|
case SILArgumentConvention::Indirect_Inout:
|
|
analyzeLocation(funcArg);
|
|
break;
|
|
default:
|
|
break;
|
|
}
|
|
}
|
|
for (SILBasicBlock &BB : *function) {
|
|
for (SILInstruction &I : BB) {
|
|
if (auto *ASI = dyn_cast<AllocStackInst>(&I)) {
|
|
if (canHandleAllocStack(ASI)) {
|
|
if (allUsesInSameBlock(ASI)) {
|
|
singleBlockLocations.push_back(ASI);
|
|
} else {
|
|
analyzeLocation(ASI);
|
|
}
|
|
}
|
|
}
|
|
if (auto *BAI = dyn_cast<BeginApplyInst>(&I)) {
|
|
auto convention = BAI->getSubstCalleeConv();
|
|
auto yields = convention.getYields();
|
|
auto yieldedValues = BAI->getYieldedValues();
|
|
for (auto index : indices(yields)) {
|
|
if (convention.isSILIndirect(yields[index])) {
|
|
analyzeLocation(yieldedValues[index]);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
void MemoryLocations::analyzeLocation(SILValue loc) {
|
|
SILFunction *function = loc->getFunction();
|
|
assert(function && "cannot analyze a SILValue which is not in a function");
|
|
|
|
// Ignore trivial types to keep the number of locations small. Trivial types
|
|
// are not interesting anyway, because such memory locations are not
|
|
// destroyed.
|
|
if (!handleTrivialLocations && loc->getType().isTrivial(*function))
|
|
return;
|
|
|
|
/// We don't handle empty tuples and empty structs.
|
|
///
|
|
/// Locations with empty types don't even need a store to count as
|
|
/// "initialized". We don't handle such cases.
|
|
if (loc->getType().isEmpty(*function))
|
|
return;
|
|
|
|
unsigned currentLocIdx = locations.size();
|
|
locations.push_back(Location(loc, currentLocIdx));
|
|
SmallVector<SILValue, 8> collectedVals;
|
|
SubLocationMap subLocationMap;
|
|
if (!analyzeLocationUsesRecursively(loc, currentLocIdx, collectedVals,
|
|
subLocationMap)) {
|
|
locations.truncate(currentLocIdx);
|
|
for (SILValue V : collectedVals) {
|
|
addr2LocIdx.erase(V);
|
|
}
|
|
return;
|
|
}
|
|
addr2LocIdx[loc] = currentLocIdx;
|
|
}
|
|
|
|
void MemoryLocations::handleSingleBlockLocations(
|
|
std::function<void (SILBasicBlock *block)> handlerFunc) {
|
|
SILBasicBlock *currentBlock = nullptr;
|
|
clear();
|
|
|
|
// Walk over all collected single-block locations.
|
|
for (SingleValueInstruction *SVI : singleBlockLocations) {
|
|
// Whenever the parent block changes, process the block's locations.
|
|
if (currentBlock && SVI->getParent() != currentBlock) {
|
|
handlerFunc(currentBlock);
|
|
clear();
|
|
}
|
|
currentBlock = SVI->getParent();
|
|
analyzeLocation(SVI);
|
|
}
|
|
// Process the last block's locations.
|
|
if (currentBlock)
|
|
handlerFunc(currentBlock);
|
|
clear();
|
|
}
|
|
|
|
const MemoryLocations::Bits &MemoryLocations::getNonTrivialLocations() {
|
|
if (nonTrivialLocations.empty()) {
|
|
// Compute the bitset lazily.
|
|
nonTrivialLocations.resize(getNumLocations());
|
|
nonTrivialLocations.reset();
|
|
unsigned idx = 0;
|
|
for (Location &loc : locations) {
|
|
initFieldsCounter(loc);
|
|
if (loc.numNonTrivialFieldsNotCovered != 0)
|
|
nonTrivialLocations.set(idx);
|
|
++idx;
|
|
}
|
|
}
|
|
return nonTrivialLocations;
|
|
}
|
|
|
|
void MemoryLocations::dump() const {
|
|
unsigned idx = 0;
|
|
for (const Location &loc : locations) {
|
|
llvm::dbgs() << "location #" << idx << ": sublocs=" << loc.subLocations
|
|
<< ", parent=" << loc.parentIdx
|
|
<< ", parentbits=" << loc.selfAndParents
|
|
<< ", #f=" << loc.numFieldsNotCoveredBySubfields
|
|
<< ", #ntf=" << loc.numNonTrivialFieldsNotCovered
|
|
<< ": " << loc.representativeValue;
|
|
++idx;
|
|
}
|
|
}
|
|
|
|
void MemoryLocations::clear() {
|
|
locations.clear();
|
|
addr2LocIdx.clear();
|
|
nonTrivialLocations.clear();
|
|
}
|
|
|
|
static bool hasInoutArgument(ApplySite AS) {
|
|
for (Operand &op : AS.getArgumentOperands()) {
|
|
switch (AS.getArgumentConvention(op)) {
|
|
case SILArgumentConvention::Indirect_Inout:
|
|
case SILArgumentConvention::Indirect_InoutAliasable:
|
|
return true;
|
|
default:
|
|
break;
|
|
}
|
|
}
|
|
return false;
|
|
}
|
|
|
|
bool MemoryLocations::analyzeLocationUsesRecursively(SILValue V, unsigned locIdx,
|
|
SmallVectorImpl<SILValue> &collectedVals,
|
|
SubLocationMap &subLocationMap) {
|
|
for (Operand *use : V->getUses()) {
|
|
// We can safely ignore type dependent operands, because the lifetime of a
|
|
// type is decoupled from the lifetime of its value. For example, even if
|
|
// the result of an open_existential_addr is destroyed its type is still
|
|
// valid.
|
|
if (use->isTypeDependent())
|
|
continue;
|
|
|
|
SILInstruction *user = use->getUser();
|
|
|
|
// We only handle addr-instructions which are planned to be used with
|
|
// opaque values. We can still consider to support other addr-instructions
|
|
// like addr-cast instructions. This somehow depends how opaque values will
|
|
// look like.
|
|
switch (user->getKind()) {
|
|
case SILInstructionKind::StructElementAddrInst: {
|
|
auto SEAI = cast<StructElementAddrInst>(user);
|
|
if (!analyzeAddrProjection(SEAI, locIdx, SEAI->getFieldIndex(),
|
|
collectedVals, subLocationMap))
|
|
return false;
|
|
break;
|
|
}
|
|
case SILInstructionKind::TupleElementAddrInst: {
|
|
auto *TEAI = cast<TupleElementAddrInst>(user);
|
|
if (!analyzeAddrProjection(TEAI, locIdx, TEAI->getFieldIndex(),
|
|
collectedVals, subLocationMap))
|
|
return false;
|
|
break;
|
|
}
|
|
case SILInstructionKind::BeginAccessInst:
|
|
if (!analyzeLocationUsesRecursively(cast<BeginAccessInst>(user), locIdx,
|
|
collectedVals, subLocationMap))
|
|
return false;
|
|
break;
|
|
case SILInstructionKind::UncheckedBorrowEnumDataAddrInst:
|
|
// borrow acts as a projection of its enum operand, but not of its
|
|
// scratch space.
|
|
if (use->getOperandNumber() != 0) {
|
|
break;
|
|
}
|
|
LLVM_FALLTHROUGH;
|
|
case SILInstructionKind::InitExistentialAddrInst:
|
|
case SILInstructionKind::OpenExistentialAddrInst:
|
|
case SILInstructionKind::InitEnumDataAddrInst:
|
|
case SILInstructionKind::UncheckedTakeEnumDataAddrInst:
|
|
case SILInstructionKind::UncheckedInPlaceEnumDataAddrInst:
|
|
if (!handleNonTrivialProjections)
|
|
return false;
|
|
// The payload is represented as a single sub-location of the enum.
|
|
if (!analyzeAddrProjection(cast<SingleValueInstruction>(user), locIdx,
|
|
/*fieldNr*/ 0, collectedVals, subLocationMap))
|
|
return false;
|
|
break;
|
|
case SILInstructionKind::PartialApplyInst:
|
|
// inout/inout_aliasable conventions means that the argument "escapes".
|
|
// This is okay for memory verification, but cannot handled by other
|
|
// optimizations, like DestroyHoisting.
|
|
if (!handleNonTrivialProjections && hasInoutArgument(ApplySite(user)))
|
|
return false;
|
|
break;
|
|
case SILInstructionKind::LoadBorrowInst:
|
|
// Reborrows are not handled
|
|
if (!cast<LoadBorrowInst>(user)->getUsersOfType<BranchInst>().empty())
|
|
return false;
|
|
break;
|
|
case SILInstructionKind::MarkDependenceInst: {
|
|
auto *mdi = cast<MarkDependenceInst>(user);
|
|
if (use == &mdi->getAllOperands()[MarkDependenceInst::Dependent]) {
|
|
if (!analyzeLocationUsesRecursively(mdi, locIdx, collectedVals, subLocationMap))
|
|
return false;
|
|
}
|
|
break;
|
|
}
|
|
case SILInstructionKind::DebugValueInst:
|
|
if (cast<DebugValueInst>(user)->hasAddrVal())
|
|
break;
|
|
return false;
|
|
case SILInstructionKind::ApplyInst: {
|
|
auto *apply = cast<ApplyInst>(user);
|
|
if (apply->hasAddressResult()) {
|
|
if (!analyzeAddrProjection(apply, locIdx, 0, collectedVals,
|
|
subLocationMap))
|
|
return false;
|
|
}
|
|
break;
|
|
}
|
|
case SILInstructionKind::StoreBorrowInst: {
|
|
if (!analyzeAddrProjection(cast<StoreBorrowInst>(user), locIdx, 0,
|
|
collectedVals, subLocationMap))
|
|
return false;
|
|
break;
|
|
}
|
|
case SILInstructionKind::InjectEnumAddrInst:
|
|
case SILInstructionKind::SelectEnumAddrInst:
|
|
case SILInstructionKind::ExistentialMetatypeInst:
|
|
case SILInstructionKind::ValueMetatypeInst:
|
|
case SILInstructionKind::IsUniqueInst:
|
|
case SILInstructionKind::FixLifetimeInst:
|
|
case SILInstructionKind::LoadInst:
|
|
case SILInstructionKind::StoreInst:
|
|
case SILInstructionKind::EndAccessInst:
|
|
case SILInstructionKind::DestroyAddrInst:
|
|
case SILInstructionKind::CheckedCastAddrBranchInst:
|
|
case SILInstructionKind::UncheckedRefCastAddrInst:
|
|
case SILInstructionKind::UnconditionalCheckedCastAddrInst:
|
|
case SILInstructionKind::TryApplyInst:
|
|
case SILInstructionKind::BeginApplyInst:
|
|
case SILInstructionKind::CopyAddrInst:
|
|
case SILInstructionKind::YieldInst:
|
|
case SILInstructionKind::DeallocStackInst:
|
|
case SILInstructionKind::SwitchEnumAddrInst:
|
|
case SILInstructionKind::WitnessMethodInst:
|
|
case SILInstructionKind::EndBorrowInst:
|
|
break;
|
|
case SILInstructionKind::MarkUnresolvedMoveAddrInst:
|
|
// We do not want the memory lifetime verifier to verify move_addr inst
|
|
// since the MarkUnresolvedMoveAddrChecker will validate that its uses
|
|
// are correct.
|
|
return false;
|
|
default:
|
|
return false;
|
|
}
|
|
}
|
|
return true;
|
|
}
|
|
|
|
bool MemoryLocations::analyzeAddrProjection(
|
|
SingleValueInstruction *projection, unsigned parentLocIdx,unsigned fieldNr,
|
|
SmallVectorImpl<SILValue> &collectedVals, SubLocationMap &subLocationMap) {
|
|
|
|
if (projection->getType().isEmpty(*projection->getFunction()))
|
|
return false;
|
|
|
|
auto key = std::make_pair(parentLocIdx, fieldNr);
|
|
unsigned subLocIdx = subLocationMap[key];
|
|
if (subLocIdx == 0) {
|
|
subLocIdx = locations.size();
|
|
assert(subLocIdx > 0);
|
|
subLocationMap[key] = subLocIdx;
|
|
locations.push_back(Location(projection, subLocIdx, parentLocIdx));
|
|
|
|
Location &parentLoc = locations[parentLocIdx];
|
|
locations.back().selfAndParents |= parentLoc.selfAndParents;
|
|
|
|
int idx = (int)parentLocIdx;
|
|
do {
|
|
Location &loc = locations[idx];
|
|
setBitAndResize(loc.subLocations, subLocIdx);
|
|
idx = loc.parentIdx;
|
|
} while (idx >= 0);
|
|
|
|
initFieldsCounter(parentLoc);
|
|
assert(parentLoc.numFieldsNotCoveredBySubfields >= 1);
|
|
parentLoc.updateFieldCounters(projection->getType(), -1);
|
|
|
|
if (parentLoc.numFieldsNotCoveredBySubfields == 0) {
|
|
int idx = (int)parentLocIdx;
|
|
do {
|
|
Location &loc = locations[idx];
|
|
loc.subLocations.reset(parentLocIdx);
|
|
idx = loc.parentIdx;
|
|
} while (idx >= 0);
|
|
}
|
|
} else if (!isa<OpenExistentialAddrInst>(projection)) {
|
|
Location *loc = &locations[subLocIdx];
|
|
if (loc->representativeValue->getType() != projection->getType()) {
|
|
assert(isa<InitEnumDataAddrInst>(projection) ||
|
|
isa<UncheckedEnumDataAddrInstBase>(projection) ||
|
|
isa<InitExistentialAddrInst>(projection));
|
|
|
|
// We can only handle a single enum payload type for a location or or a
|
|
// single concrete existential type. Mismatching types can have a different
|
|
// number of (non-trivial) sub-locations and we cannot handle this.
|
|
// But we ignore opened existential types, because those cannot have
|
|
// sub-locations (there cannot be an address projection on an
|
|
// open_existential_addr).
|
|
if (!isa<OpenExistentialAddrInst>(loc->representativeValue))
|
|
return false;
|
|
assert(loc->representativeValue->getType().is<ExistentialArchetypeType>());
|
|
loc->representativeValue = projection;
|
|
}
|
|
}
|
|
|
|
if (!analyzeLocationUsesRecursively(projection, subLocIdx, collectedVals,
|
|
subLocationMap)) {
|
|
return false;
|
|
}
|
|
registerProjection(projection, subLocIdx);
|
|
collectedVals.push_back(projection);
|
|
return true;
|
|
}
|
|
|
|
void MemoryLocations::initFieldsCounter(Location &loc) {
|
|
if (loc.numFieldsNotCoveredBySubfields >= 0)
|
|
return;
|
|
|
|
assert(loc.numNonTrivialFieldsNotCovered < 0);
|
|
|
|
loc.numFieldsNotCoveredBySubfields = 0;
|
|
loc.numNonTrivialFieldsNotCovered = 0;
|
|
SILFunction *function = loc.representativeValue->getFunction();
|
|
SILType ty = loc.representativeValue->getType();
|
|
if (StructDecl *decl = ty.getStructOrBoundGenericStruct()) {
|
|
if (decl->isResilient(function->getModule().getSwiftModule(),
|
|
function->getResilienceExpansion())) {
|
|
loc.numFieldsNotCoveredBySubfields = INT_MAX;
|
|
return;
|
|
}
|
|
SILModule &module = function->getModule();
|
|
for (VarDecl *field : decl->getStoredProperties()) {
|
|
loc.updateFieldCounters(
|
|
ty.getFieldType(field, module, TypeExpansionContext(*function)), +1);
|
|
}
|
|
return;
|
|
}
|
|
if (auto tupleTy = ty.getAs<TupleType>()) {
|
|
for (unsigned idx = 0, end = tupleTy->getNumElements(); idx < end; ++idx) {
|
|
loc.updateFieldCounters(ty.getTupleElementType(idx), +1);
|
|
}
|
|
return;
|
|
}
|
|
loc.updateFieldCounters(ty, +1);
|
|
}
|