//===-- WebAssemblyCFGStackify.cpp - CFG Stackification -------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
///
/// \file
/// This file implements a CFG stacking pass.
///
/// This pass inserts BLOCK, LOOP, and TRY markers to mark the start of scopes,
/// since scope boundaries serve as the labels for WebAssembly's control
/// transfers.
///
/// This is sufficient to convert arbitrary CFGs into a form that works on
/// WebAssembly, provided that all loops are single-entry.
///
/// In case we use exceptions, this pass also fixes mismatches in unwind
/// destinations created during transforming CFG into wasm structured format.
///
//===----------------------------------------------------------------------===//
#include "Utils/WebAssemblyTypeUtilities.h"
#include "Utils/WebAssemblyUtilities.h"
#include "WebAssembly.h"
#include "WebAssemblyExceptionInfo.h"
#include "WebAssemblyMachineFunctionInfo.h"
#include "WebAssemblySortRegion.h"
#include "WebAssemblySubtarget.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/WasmEHFuncInfo.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/Target/TargetMachine.h"
using namespace llvm;
using WebAssembly::SortRegionInfo;
#define DEBUG_TYPE "wasm-cfg-stackify"
STATISTIC(NumCallUnwindMismatches, "Number of call unwind mismatches found");
STATISTIC(NumCatchUnwindMismatches, "Number of catch unwind mismatches found");
namespace {
class WebAssemblyCFGStackify final : public MachineFunctionPass {
StringRef getPassName() const override { return "WebAssembly CFG Stackify"; }
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.addRequired<MachineDominatorTree>();
AU.addRequired<MachineLoopInfo>();
AU.addRequired<WebAssemblyExceptionInfo>();
MachineFunctionPass::getAnalysisUsage(AU);
}
bool runOnMachineFunction(MachineFunction &MF) override;
// For each block whose label represents the end of a scope, record the block
// which holds the beginning of the scope. This will allow us to quickly skip
// over scoped regions when walking blocks.
SmallVector<MachineBasicBlock *, 8> ScopeTops;
void updateScopeTops(MachineBasicBlock *Begin, MachineBasicBlock *End) {
int EndNo = End->getNumber();
if (!ScopeTops[EndNo] || ScopeTops[EndNo]->getNumber() > Begin->getNumber())
ScopeTops[EndNo] = Begin;
}
// Placing markers.
void placeMarkers(MachineFunction &MF);
void placeBlockMarker(MachineBasicBlock &MBB);
void placeLoopMarker(MachineBasicBlock &MBB);
void placeTryMarker(MachineBasicBlock &MBB);
// Exception handling related functions
bool fixCallUnwindMismatches(MachineFunction &MF);
bool fixCatchUnwindMismatches(MachineFunction &MF);
void addTryDelegate(MachineInstr *RangeBegin, MachineInstr *RangeEnd,
MachineBasicBlock *DelegateDest);
void recalculateScopeTops(MachineFunction &MF);
void removeUnnecessaryInstrs(MachineFunction &MF);
// Wrap-up
using EndMarkerInfo =
std::pair<const MachineBasicBlock *, const MachineInstr *>;
unsigned getBranchDepth(const SmallVectorImpl<EndMarkerInfo> &Stack,
const MachineBasicBlock *MBB);
unsigned getDelegateDepth(const SmallVectorImpl<EndMarkerInfo> &Stack,
const MachineBasicBlock *MBB);
unsigned
getRethrowDepth(const SmallVectorImpl<EndMarkerInfo> &Stack,
const SmallVectorImpl<const MachineBasicBlock *> &EHPadStack);
void rewriteDepthImmediates(MachineFunction &MF);
void fixEndsAtEndOfFunction(MachineFunction &MF);
void cleanupFunctionData(MachineFunction &MF);
// For each BLOCK|LOOP|TRY, the corresponding END_(BLOCK|LOOP|TRY) or DELEGATE
// (in case of TRY).
DenseMap<const MachineInstr *, MachineInstr *> BeginToEnd;
// For each END_(BLOCK|LOOP|TRY) or DELEGATE, the corresponding
// BLOCK|LOOP|TRY.
DenseMap<const MachineInstr *, MachineInstr *> EndToBegin;
// <TRY marker, EH pad> map
DenseMap<const MachineInstr *, MachineBasicBlock *> TryToEHPad;
// <EH pad, TRY marker> map
DenseMap<const MachineBasicBlock *, MachineInstr *> EHPadToTry;
// We need an appendix block to place 'end_loop' or 'end_try' marker when the
// loop / exception bottom block is the last block in a function
MachineBasicBlock *AppendixBB = nullptr;
MachineBasicBlock *getAppendixBlock(MachineFunction &MF) {
if (!AppendixBB) {
AppendixBB = MF.CreateMachineBasicBlock();
// Give it a fake predecessor so that AsmPrinter prints its label.
AppendixBB->addSuccessor(AppendixBB);
MF.push_back(AppendixBB);
}
return AppendixBB;
}
// Before running rewriteDepthImmediates function, 'delegate' has a BB as its
// destination operand. getFakeCallerBlock() returns a fake BB that will be
// used for the operand when 'delegate' needs to rethrow to the caller. This
// will be rewritten as an immediate value that is the number of block depths
// + 1 in rewriteDepthImmediates, and this fake BB will be removed at the end
// of the pass.
MachineBasicBlock *FakeCallerBB = nullptr;
MachineBasicBlock *getFakeCallerBlock(MachineFunction &MF) {
if (!FakeCallerBB)
FakeCallerBB = MF.CreateMachineBasicBlock();
return FakeCallerBB;
}
// Helper functions to register / unregister scope information created by
// marker instructions.
void registerScope(MachineInstr *Begin, MachineInstr *End);
void registerTryScope(MachineInstr *Begin, MachineInstr *End,
MachineBasicBlock *EHPad);
void unregisterScope(MachineInstr *Begin);
public:
static char ID; // Pass identification, replacement for typeid
WebAssemblyCFGStackify() : MachineFunctionPass(ID) {}
~WebAssemblyCFGStackify() override { releaseMemory(); }
void releaseMemory() override;
};
} // end anonymous namespace
char WebAssemblyCFGStackify::ID = 0;
INITIALIZE_PASS(WebAssemblyCFGStackify, DEBUG_TYPE,
"Insert BLOCK/LOOP/TRY markers for WebAssembly scopes", false,
false)
FunctionPass *llvm::createWebAssemblyCFGStackify() {
return new WebAssemblyCFGStackify();
}
/// Test whether Pred has any terminators explicitly branching to MBB, as
/// opposed to falling through. Note that it's possible (eg. in unoptimized
/// code) for a branch instruction to both branch to a block and fallthrough
/// to it, so we check the actual branch operands to see if there are any
/// explicit mentions.
static bool explicitlyBranchesTo(MachineBasicBlock *Pred,
MachineBasicBlock *MBB) {
for (MachineInstr &MI : Pred->terminators())
for (MachineOperand &MO : MI.explicit_operands())
if (MO.isMBB() && MO.getMBB() == MBB)
return true;
return false;
}
// Returns an iterator to the earliest position possible within the MBB,
// satisfying the restrictions given by BeforeSet and AfterSet. BeforeSet
// contains instructions that should go before the marker, and AfterSet contains
// ones that should go after the marker. In this function, AfterSet is only
// used for validation checking.
template <typename Container>
static MachineBasicBlock::iterator
getEarliestInsertPos(MachineBasicBlock *MBB, const Container &BeforeSet,
const Container &AfterSet) {
auto InsertPos = MBB->end();
while (InsertPos != MBB->begin()) {
if (BeforeSet.count(&*std::prev(InsertPos))) {
#ifndef NDEBUG
// Validation check
for (auto Pos = InsertPos, E = MBB->begin(); Pos != E; --Pos)
assert(!AfterSet.count(&*std::prev(Pos)));
#endif
break;
}
--InsertPos;
}
return InsertPos;
}
// Returns an iterator to the latest position possible within the MBB,
// satisfying the restrictions given by BeforeSet and AfterSet. BeforeSet
// contains instructions that should go before the marker, and AfterSet contains
// ones that should go after the marker. In this function, BeforeSet is only
// used for validation checking.
template <typename Container>
static MachineBasicBlock::iterator
getLatestInsertPos(MachineBasicBlock *MBB, const Container &BeforeSet,
const Container &AfterSet) {
auto InsertPos = MBB->begin();
while (InsertPos != MBB->end()) {
if (AfterSet.count(&*InsertPos)) {
#ifndef NDEBUG
// Validation check
for (auto Pos = InsertPos, E = MBB->end(); Pos != E; ++Pos)
assert(!BeforeSet.count(&*Pos));
#endif
break;
}
++InsertPos;
}
return InsertPos;
}
void WebAssemblyCFGStackify::registerScope(MachineInstr *Begin,
MachineInstr *End) {
BeginToEnd[Begin] = End;
EndToBegin[End] = Begin;
}
// When 'End' is not an 'end_try' but 'delegate, EHPad is nullptr.
void WebAssemblyCFGStackify::registerTryScope(MachineInstr *Begin,
MachineInstr *End,
MachineBasicBlock *EHPad) {
registerScope(Begin, End);
TryToEHPad[Begin] = EHPad;
EHPadToTry[EHPad] = Begin;
}
void WebAssemblyCFGStackify::unregisterScope(MachineInstr *Begin) {
assert(BeginToEnd.count(Begin));
MachineInstr *End = BeginToEnd[Begin];
assert(EndToBegin.count(End));
BeginToEnd.erase(Begin);
EndToBegin.erase(End);
MachineBasicBlock *EHPad = TryToEHPad.lookup(Begin);
if (EHPad) {
assert(EHPadToTry.count(EHPad));
TryToEHPad.erase(Begin);
EHPadToTry.erase(EHPad);
}
}
/// Insert a BLOCK marker for branches to MBB (if needed).
// TODO Consider a more generalized way of handling block (and also loop and
// try) signatures when we implement the multi-value proposal later.
void WebAssemblyCFGStackify::placeBlockMarker(MachineBasicBlock &MBB) {
assert(!MBB.isEHPad());
MachineFunction &MF = *MBB.getParent();
auto &MDT = getAnalysis<MachineDominatorTree>();
const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
const auto &MFI = *MF.getInfo<WebAssemblyFunctionInfo>();
// First compute the nearest common dominator of all forward non-fallthrough
// predecessors so that we minimize the time that the BLOCK is on the stack,
// which reduces overall stack height.
MachineBasicBlock *Header = nullptr;
bool IsBranchedTo = false;
int MBBNumber = MBB.getNumber();
for (MachineBasicBlock *Pred : MBB.predecessors()) {
if (Pred->getNumber() < MBBNumber) {
Header = Header ? MDT.findNearestCommonDominator(Header, Pred) : Pred;
if (explicitlyBranchesTo(Pred, &MBB))
IsBranchedTo = true;
}
}
if (!Header)
return;
if (!IsBranchedTo)
return;
assert(&MBB != &MF.front() && "Header blocks shouldn't have predecessors");
MachineBasicBlock *LayoutPred = MBB.getPrevNode();
// If the nearest common dominator is inside a more deeply nested context,
// walk out to the nearest scope which isn't more deeply nested.
for (MachineFunction::iterator I(LayoutPred), E(Header); I != E; --I) {
if (MachineBasicBlock *ScopeTop = ScopeTops[I->getNumber()]) {
if (ScopeTop->getNumber() > Header->getNumber()) {
// Skip over an intervening scope.
I = std::next(ScopeTop->getIterator());
} else {
// We found a scope level at an appropriate depth.
Header = ScopeTop;
break;
}
}
}
// Decide where in Header to put the BLOCK.
// Instructions that should go before the BLOCK.
SmallPtrSet<const MachineInstr *, 4> BeforeSet;
// Instructions that should go after the BLOCK.
SmallPtrSet<const MachineInstr *, 4> AfterSet;
for (const auto &MI : *Header) {
// If there is a previously placed LOOP marker and the bottom block of the
// loop is above MBB, it should be after the BLOCK, because the loop is
// nested in this BLOCK. Otherwise it should be before the BLOCK.
if (MI.getOpcode() == WebAssembly::LOOP) {
auto *LoopBottom = BeginToEnd[&MI]->getParent()->getPrevNode();
if (MBB.getNumber() > LoopBottom->getNumber())
AfterSet.insert(&MI);
#ifndef NDEBUG
else
BeforeSet.insert(&MI);
#endif
}
// If there is a previously placed BLOCK/TRY marker and its corresponding
// END marker is before the current BLOCK's END marker, that should be
// placed after this BLOCK. Otherwise it should be placed before this BLOCK
// marker.
if (MI.getOpcode() == WebAssembly::BLOCK ||
MI.getOpcode() == WebAssembly::TRY) {
if (BeginToEnd[&MI]->getParent()->getNumber() <= MBB.getNumber())
AfterSet.insert(&MI);
#ifndef NDEBUG
else
BeforeSet.insert(&MI);
#endif
}
#ifndef NDEBUG
// All END_(BLOCK|LOOP|TRY) markers should be before the BLOCK.
if (MI.getOpcode() == WebAssembly::END_BLOCK ||
MI.getOpcode() == WebAssembly::END_LOOP ||
MI.getOpcode() == WebAssembly::END_TRY)
BeforeSet.insert(&MI);
#endif
// Terminators should go after the BLOCK.
if (MI.isTerminator())
AfterSet.insert(&MI);
}
// Local expression tree should go after the BLOCK.
for (auto I = Header->getFirstTerminator(), E = Header->begin(); I != E;
--I) {
if (std::prev(I)->isDebugInstr() || std::prev(I)->isPosition())
continue;
if (WebAssembly::isChild(*std::prev(I), MFI))
AfterSet.insert(&*std::prev(I));
else
break;
}
// Add the BLOCK.
WebAssembly::BlockType ReturnType = WebAssembly::BlockType::Void;
auto InsertPos = getLatestInsertPos(Header, BeforeSet, AfterSet);
MachineInstr *Begin =
BuildMI(*Header, InsertPos, Header->findDebugLoc(InsertPos),
TII.get(WebAssembly::BLOCK))
.addImm(int64_t(ReturnType));
// Decide where in Header to put the END_BLOCK.
BeforeSet.clear();
AfterSet.clear();
for (auto &MI : MBB) {
#ifndef NDEBUG
// END_BLOCK should precede existing LOOP and TRY markers.
if (MI.getOpcode() == WebAssembly::LOOP ||
MI.getOpcode() == WebAssembly::TRY)
AfterSet.insert(&MI);
#endif
// If there is a previously placed END_LOOP marker and the header of the
// loop is above this block's header, the END_LOOP should be placed after
// the BLOCK, because the loop contains this block. Otherwise the END_LOOP
// should be placed before the BLOCK. The same for END_TRY.
if (MI.getOpcode() == WebAssembly::END_LOOP ||
MI.getOpcode() == WebAssembly::END_TRY) {
if (EndToBegin[&MI]->getParent()->getNumber() >= Header->getNumber())
BeforeSet.insert(&MI);
#ifndef NDEBUG
else
AfterSet.insert(&MI);
#endif
}
}
// Mark the end of the block.
InsertPos = getEarliestInsertPos(&MBB, BeforeSet, AfterSet);
MachineInstr *End = BuildMI(MBB, InsertPos, MBB.findPrevDebugLoc(InsertPos),
TII.get(WebAssembly::END_BLOCK));
registerScope(Begin, End);
// Track the farthest-spanning scope that ends at this point.
updateScopeTops(Header, &MBB);
}
/// Insert a LOOP marker for a loop starting at MBB (if it's a loop header).
void WebAssemblyCFGStackify::placeLoopMarker(MachineBasicBlock &MBB) {
MachineFunction &MF = *MBB.getParent();
const auto &MLI = getAnalysis<MachineLoopInfo>();
const auto &WEI = getAnalysis<WebAssemblyExceptionInfo>();
SortRegionInfo SRI(MLI, WEI);
const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
MachineLoop *Loop = MLI.getLoopFor(&MBB);
if (!Loop || Loop->getHeader() != &MBB)
return;
// The operand of a LOOP is the first block after the loop. If the loop is the
// bottom of the function, insert a dummy block at the end.
MachineBasicBlock *Bottom = SRI.getBottom(Loop);
auto Iter = std::next(Bottom->getIterator());
if (Iter == MF.end()) {
getAppendixBlock(MF);
Iter = std::next(Bottom->getIterator());
}
MachineBasicBlock *AfterLoop = &*Iter;
// Decide where in Header to put the LOOP.
SmallPtrSet<const MachineInstr *, 4> BeforeSet;
SmallPtrSet<const MachineInstr *, 4> AfterSet;
for (const auto &MI : MBB) {
// LOOP marker should be after any existing loop that ends here. Otherwise
// we assume the instruction belongs to the loop.
if (MI.getOpcode() == WebAssembly::END_LOOP)
BeforeSet.insert(&MI);
#ifndef NDEBUG
else
AfterSet.insert(&MI);
#endif
}
// Mark the beginning of the loop.
auto InsertPos = getEarliestInsertPos(&MBB, BeforeSet, AfterSet);
MachineInstr *Begin = BuildMI(MBB, InsertPos, MBB.findDebugLoc(InsertPos),
TII.get(WebAssembly::LOOP))
.addImm(int64_t(WebAssembly::BlockType::Void));
// Decide where in Header to put the END_LOOP.
BeforeSet.clear();
AfterSet.clear();
#ifndef NDEBUG
for (const auto &MI : MBB)
// Existing END_LOOP markers belong to parent loops of this loop
if (MI.getOpcode() == WebAssembly::END_LOOP)
AfterSet.insert(&MI);
#endif
// Mark the end of the loop (using arbitrary debug location that branched to
// the loop end as its location).
InsertPos = getEarliestInsertPos(AfterLoop, BeforeSet, AfterSet);
DebugLoc EndDL = AfterLoop->pred_empty()
? DebugLoc()
: (*AfterLoop->pred_rbegin())->findBranchDebugLoc();
MachineInstr *End =
BuildMI(*AfterLoop, InsertPos, EndDL, TII.get(WebAssembly::END_LOOP));
registerScope(Begin, End);
assert((!ScopeTops[AfterLoop->getNumber()] ||
ScopeTops[AfterLoop->getNumber()]->getNumber() < MBB.getNumber()) &&
"With block sorting the outermost loop for a block should be first.");
updateScopeTops(&MBB, AfterLoop);
}
void WebAssemblyCFGStackify::placeTryMarker(MachineBasicBlock &MBB) {
assert(MBB.isEHPad());
MachineFunction &MF = *MBB.getParent();
auto &MDT = getAnalysis<MachineDominatorTree>();
const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
const auto &MLI = getAnalysis<MachineLoopInfo>();
const auto &WEI = getAnalysis<WebAssemblyExceptionInfo>();
SortRegionInfo SRI(MLI, WEI);
const auto &MFI = *MF.getInfo<WebAssemblyFunctionInfo>();
// Compute the nearest common dominator of all unwind predecessors
MachineBasicBlock *Header = nullptr;
int MBBNumber = MBB.getNumber();
for (auto *Pred : MBB.predecessors()) {
if (Pred->getNumber() < MBBNumber) {
Header = Header ? MDT.findNearestCommonDominator(Header, Pred) : Pred;
assert(!explicitlyBranchesTo(Pred, &MBB) &&
"Explicit branch to an EH pad!");
}
}
if (!Header)
return;
// If this try is at the bottom of the function, insert a dummy block at the
// end.
WebAssemblyException *WE = WEI.getExceptionFor(&MBB);
assert(WE);
MachineBasicBlock *Bottom = SRI.getBottom(WE);
auto Iter = std::next(Bottom->getIterator());
if (Iter == MF.end()) {
getAppendixBlock(MF);
Iter = std::next(Bottom->getIterator());
}
MachineBasicBlock *Cont = &*Iter;
assert(Cont != &MF.front());
MachineBasicBlock *LayoutPred = Cont->getPrevNode();
// If the nearest common dominator is inside a more deeply nested context,
// walk out to the nearest scope which isn't more deeply nested.
for (MachineFunction::iterator I(LayoutPred), E(Header); I != E; --I) {
if (MachineBasicBlock *ScopeTop = ScopeTops[I->getNumber()]) {
if (ScopeTop->getNumber() > Header->getNumber()) {
// Skip over an intervening scope.
I = std::next(ScopeTop->getIterator());
} else {
// We found a scope level at an appropriate depth.
Header = ScopeTop;
break;
}
}
}
// Decide where in Header to put the TRY.
// Instructions that should go before the TRY.
SmallPtrSet<const MachineInstr *, 4> BeforeSet;
// Instructions that should go after the TRY.
SmallPtrSet<const MachineInstr *, 4> AfterSet;
for (const auto &MI : *Header) {
// If there is a previously placed LOOP marker and the bottom block of the
// loop is above MBB, it should be after the TRY, because the loop is nested
// in this TRY. Otherwise it should be before the TRY.
if (MI.getOpcode() == WebAssembly::LOOP) {
auto *LoopBottom = BeginToEnd[&MI]->getParent()->getPrevNode();
if (MBB.getNumber() > LoopBottom->getNumber())
AfterSet.insert(&MI);
#ifndef NDEBUG
else
BeforeSet.insert(&MI);
#endif
}
// All previously inserted BLOCK/TRY markers should be after the TRY because
// they are all nested trys.
if (MI.getOpcode() == WebAssembly::BLOCK ||
MI.getOpcode() == WebAssembly::TRY)
AfterSet.insert(&MI);
#ifndef NDEBUG
// All END_(BLOCK/LOOP/TRY) markers should be before the TRY.
if (MI.getOpcode() == WebAssembly::END_BLOCK ||
MI.getOpcode() == WebAssembly::END_LOOP ||
MI.getOpcode() == WebAssembly::END_TRY)
BeforeSet.insert(&MI);
#endif
// Terminators should go after the TRY.
if (MI.isTerminator())
AfterSet.insert(&MI);
}
// If Header unwinds to MBB (= Header contains 'invoke'), the try block should
// contain the call within it. So the call should go after the TRY. The
// exception is when the header's terminator is a rethrow instruction, in
// which case that instruction, not a call instruction before it, is gonna
// throw.
MachineInstr *ThrowingCall = nullptr;
if (MBB.isPredecessor(Header)) {
auto TermPos = Header->getFirstTerminator();
if (TermPos == Header->end() ||
TermPos->getOpcode() != WebAssembly::RETHROW) {
for (auto &MI : reverse(*Header)) {
if (MI.isCall()) {
AfterSet.insert(&MI);
ThrowingCall = &MI;
// Possibly throwing calls are usually wrapped by EH_LABEL
// instructions. We don't want to split them and the call.
if (MI.getIterator() != Header->begin() &&
std::prev(MI.getIterator())->isEHLabel()) {
AfterSet.insert(&*std::prev(MI.getIterator()));
ThrowingCall = &*std::prev(MI.getIterator());
}
break;
}
}
}
}
// Local expression tree should go after the TRY.
// For BLOCK placement, we start the search from the previous instruction of a
// BB's terminator, but in TRY's case, we should start from the previous
// instruction of a call that can throw, or a EH_LABEL that precedes the call,
// because the return values of the call's previous instructions can be
// stackified and consumed by the throwing call.
auto SearchStartPt = ThrowingCall ? MachineBasicBlock::iterator(ThrowingCall)
: Header->getFirstTerminator();
for (auto I = SearchStartPt, E = Header->begin(); I != E; --I) {
if (std::prev(I)->isDebugInstr() || std::prev(I)->isPosition())
continue;
if (WebAssembly::isChild(*std::prev(I), MFI))
AfterSet.insert(&*std::prev(I));
else
break;
}
// Add the TRY.
auto InsertPos = getLatestInsertPos(Header, BeforeSet, AfterSet);
MachineInstr *Begin =
BuildMI(*Header, InsertPos, Header->findDebugLoc(InsertPos),
TII.get(WebAssembly::TRY))
.addImm(int64_t(WebAssembly::BlockType::Void));
// Decide where in Header to put the END_TRY.
BeforeSet.clear();
AfterSet.clear();
for (const auto &MI : *Cont) {
#ifndef NDEBUG
// END_TRY should precede existing LOOP and BLOCK markers.
if (MI.getOpcode() == WebAssembly::LOOP ||
MI.getOpcode() == WebAssembly::BLOCK)
AfterSet.insert(&MI);
// All END_TRY markers placed earlier belong to exceptions that contains
// this one.
if (MI.getOpcode() == WebAssembly::END_TRY)
AfterSet.insert(&MI);
#endif
// If there is a previously placed END_LOOP marker and its header is after
// where TRY marker is, this loop is contained within the 'catch' part, so
// the END_TRY marker should go after that. Otherwise, the whole try-catch
// is contained within this loop, so the END_TRY should go before that.
if (MI.getOpcode() == WebAssembly::END_LOOP) {
// For a LOOP to be after TRY, LOOP's BB should be after TRY's BB; if they
// are in the same BB, LOOP is always before TRY.
if (EndToBegin[&MI]->getParent()->getNumber() > Header->getNumber())
BeforeSet.insert(&MI);
#ifndef NDEBUG
else
AfterSet.insert(&MI);
#endif
}
// It is not possible for an END_BLOCK to be already in this block.
}
// Mark the end of the TRY.
InsertPos = getEarliestInsertPos(Cont, BeforeSet, AfterSet);
MachineInstr *End =
BuildMI(*Cont, InsertPos, Bottom->findBranchDebugLoc(),
TII.get(WebAssembly::END_TRY));
registerTryScope(Begin, End, &MBB);
// Track the farthest-spanning scope that ends at this point. We create two
// mappings: (BB with 'end_try' -> BB with 'try') and (BB with 'catch' -> BB
// with 'try'). We need to create 'catch' -> 'try' mapping here too because
// markers should not span across 'catch'. For example, this should not
// happen:
//
// try
// block --| (X)
// catch |
// end_block --|
// end_try
for (auto *End : {&MBB, Cont})
updateScopeTops(Header, End);
}
void WebAssemblyCFGStackify::removeUnnecessaryInstrs(MachineFunction &MF) {
const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
// When there is an unconditional branch right before a catch instruction and
// it branches to the end of end_try marker, we don't need the branch, because
// it there is no exception, the control flow transfers to that point anyway.
// bb0:
// try
// ...
// br bb2 <- Not necessary
// bb1 (ehpad):
// catch
// ...
// bb2: <- Continuation BB
// end
//
// A more involved case: When the BB where 'end' is located is an another EH
// pad, the Cont (= continuation) BB is that EH pad's 'end' BB. For example,
// bb0:
// try
// try
// ...
// br bb3 <- Not necessary
// bb1 (ehpad):
// catch
// bb2 (ehpad):
// end
// catch
// ...
// bb3: <- Continuation BB
// end
//
// When the EH pad at hand is bb1, its matching end_try is in bb2. But it is
// another EH pad, so bb0's continuation BB becomes bb3. So 'br bb3' in the
// code can be deleted. This is why we run 'while' until 'Cont' is not an EH
// pad.
for (auto &MBB : MF) {
if (!MBB.isEHPad())
continue;
MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
SmallVector<MachineOperand, 4> Cond;
MachineBasicBlock *EHPadLayoutPred = MBB.getPrevNode();
MachineBasicBlock *Cont = &MBB;
while (Cont->isEHPad()) {
MachineInstr *Try = EHPadToTry[Cont];
MachineInstr *EndTry = BeginToEnd[Try];
// We started from an EH pad, so the end marker cannot be a delegate
assert(EndTry->getOpcode() != WebAssembly::DELEGATE);
Cont = EndTry->getParent();
}
bool Analyzable = !TII.analyzeBranch(*EHPadLayoutPred, TBB, FBB, Cond);
// This condition means either
// 1. This BB ends with a single unconditional branch whose destinaion is
// Cont.
// 2. This BB ends with a conditional branch followed by an unconditional
// branch, and the unconditional branch's destination is Cont.
// In both cases, we want to remove the last (= unconditional) branch.
if (Analyzable && ((Cond.empty() && TBB && TBB == Cont) ||
(!Cond.empty() && FBB && FBB == Cont))) {
bool ErasedUncondBr = false;
(void)ErasedUncondBr;
for (auto I = EHPadLayoutPred->end(), E = EHPadLayoutPred->begin();
I != E; --I) {
auto PrevI = std::prev(I);
if (PrevI->isTerminator()) {
assert(PrevI->getOpcode() == WebAssembly::BR);
PrevI->eraseFromParent();
ErasedUncondBr = true;
break;
}
}
assert(ErasedUncondBr && "Unconditional branch not erased!");
}
}
// When there are block / end_block markers that overlap with try / end_try
// markers, and the block and try markers' return types are the same, the
// block /end_block markers are not necessary, because try / end_try markers
// also can serve as boundaries for branches.
// block <- Not necessary
// try
// ...
// catch
// ...
// end
// end <- Not necessary
SmallVector<MachineInstr *, 32> ToDelete;
for (auto &MBB : MF) {
for (auto &MI : MBB) {
if (MI.getOpcode() != WebAssembly::TRY)
continue;
MachineInstr *Try = &MI, *EndTry = BeginToEnd[Try];
if (EndTry->getOpcode() == WebAssembly::DELEGATE)
continue;
MachineBasicBlock *TryBB = Try->getParent();
MachineBasicBlock *Cont = EndTry->getParent();
int64_t RetType = Try->getOperand(0).getImm();
for (auto B = Try->getIterator(), E = std::next(EndTry->getIterator());
B != TryBB->begin() && E != Cont->end() &&
std::prev(B)->getOpcode() == WebAssembly::BLOCK &&
E->getOpcode() == WebAssembly::END_BLOCK &&
std::prev(B)->getOperand(0).getImm() == RetType;
--B, ++E) {
ToDelete.push_back(&*std::prev(B));
ToDelete.push_back(&*E);
}
}
}
for (auto *MI : ToDelete) {
if (MI->getOpcode() == WebAssembly::BLOCK)
unregisterScope(MI);
MI->eraseFromParent();
}
}
// When MBB is split into MBB and Split, we should unstackify defs in MBB that
// have their uses in Split.
static void unstackifyVRegsUsedInSplitBB(MachineBasicBlock &MBB,
MachineBasicBlock &Split) {
MachineFunction &MF = *MBB.getParent();
const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
auto &MFI = *MF.getInfo<WebAssemblyFunctionInfo>();
auto &MRI = MF.getRegInfo();
for (auto &MI : Split) {
for (auto &MO : MI.explicit_uses()) {
if (!MO.isReg() || Register::isPhysicalRegister(MO.getReg()))
continue;
if (MachineInstr *Def = MRI.getUniqueVRegDef(MO.getReg()))
if (Def->getParent() == &MBB)
MFI.unstackifyVReg(MO.getReg());
}
}
// In RegStackify, when a register definition is used multiple times,
// Reg = INST ...
// INST ..., Reg, ...
// INST ..., Reg, ...
// INST ..., Reg, ...
//
// we introduce a TEE, which has the following form:
// DefReg = INST ...
// TeeReg, Reg = TEE_... DefReg
// INST ..., TeeReg, ...
// INST ..., Reg, ...
// INST ..., Reg, ...
// with DefReg and TeeReg stackified but Reg not stackified.
//
// But the invariant that TeeReg should be stackified can be violated while we
// unstackify registers in the split BB above. In this case, we convert TEEs
// into two COPYs. This COPY will be eventually eliminated in ExplicitLocals.
// DefReg = INST ...
// TeeReg = COPY DefReg
// Reg = COPY DefReg
// INST ..., TeeReg, ...
// INST ..., Reg, ...
// INST ..., Reg, ...
for (MachineInstr &MI : llvm::make_early_inc_range(MBB)) {
if (!WebAssembly::isTee(MI.getOpcode()))
continue;
Register TeeReg = MI.getOperand(0).getReg();
Register Reg = MI.getOperand(1).getReg();
Register DefReg = MI.getOperand(2).getReg();
if (!MFI.isVRegStackified(TeeReg)) {
// Now we are not using TEE anymore, so unstackify DefReg too
MFI.unstackifyVReg(DefReg);
unsigned CopyOpc =
WebAssembly::getCopyOpcodeForRegClass(MRI.getRegClass(DefReg));
BuildMI(MBB, &MI, MI.getDebugLoc(), TII.get(CopyOpc), TeeReg)
.addReg(DefReg);
BuildMI(MBB, &MI, MI.getDebugLoc(), TII.get(CopyOpc), Reg).addReg(DefReg);
MI.eraseFromParent();
}
}
}
// Wrap the given range of instruction with try-delegate. RangeBegin and
// RangeEnd are inclusive.
void WebAssemblyCFGStackify::addTryDelegate(MachineInstr *RangeBegin,
MachineInstr *RangeEnd,
MachineBasicBlock *DelegateDest) {
auto *BeginBB = RangeBegin->getParent();
auto *EndBB = RangeEnd->getParent();
MachineFunction &MF = *BeginBB->getParent();
const auto &MFI = *MF.getInfo<WebAssemblyFunctionInfo>();
const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
// Local expression tree before the first call of this range should go
// after the nested TRY.
SmallPtrSet<const MachineInstr *, 4> AfterSet;
AfterSet.insert(RangeBegin);
for (auto I = MachineBasicBlock::iterator(RangeBegin), E = BeginBB->begin();
I != E; --I) {
if (std::prev(I)->isDebugInstr() || std::prev(I)->isPosition())
continue;
if (WebAssembly::isChild(*std::prev(I), MFI))
AfterSet.insert(&*std::prev(I));
else
break;
}
// Create the nested try instruction.
auto TryPos = getLatestInsertPos(
BeginBB, SmallPtrSet<const MachineInstr *, 4>(), AfterSet);
MachineInstr *Try = BuildMI(*BeginBB, TryPos, RangeBegin->getDebugLoc(),
TII.get(WebAssembly::TRY))
.addImm(int64_t(WebAssembly::BlockType::Void));
// Create a BB to insert the 'delegate' instruction.
MachineBasicBlock *DelegateBB = MF.CreateMachineBasicBlock();
// If the destination of 'delegate' is not the caller, adds the destination to
// the BB's successors.
if (DelegateDest != FakeCallerBB)
DelegateBB->addSuccessor(DelegateDest);
auto SplitPos = std::next(RangeEnd->getIterator());
if (SplitPos == EndBB->end()) {
// If the range's end instruction is at the end of the BB, insert the new
// delegate BB after the current BB.
MF.insert(std::next(EndBB->getIterator()), DelegateBB);
EndBB->addSuccessor(DelegateBB);
} else {
// When the split pos is in the middle of a BB, we split the BB into two and
// put the 'delegate' BB in between. We normally create a split BB and make
// it a successor of the original BB (PostSplit == true), but in case the BB
// is an EH pad and the split pos is before 'catch', we should preserve the
// BB's property, including that it is an EH pad, in the later part of the
// BB, where 'catch' is. In this case we set PostSplit to false.
bool PostSplit = true;
if (EndBB->isEHPad()) {
for (auto I = MachineBasicBlock::iterator(SplitPos), E = EndBB->end();
I != E; ++I) {
if (WebAssembly::isCatch(I->getOpcode())) {
PostSplit = false;
break;
}
}
}
MachineBasicBlock *PreBB = nullptr, *PostBB = nullptr;
if (PostSplit) {
// If the range's end instruction is in the middle of the BB, we split the
// BB into two and insert the delegate BB in between.
// - Before:
// bb:
// range_end
// other_insts
//
// - After:
// pre_bb: (previous 'bb')
// range_end
// delegate_bb: (new)
// delegate
// post_bb: (new)
// other_insts
PreBB = EndBB;
PostBB = MF.CreateMachineBasicBlock();
MF.insert(std::next(PreBB->getIterator()), PostBB);
MF.insert(std::next(PreBB->getIterator()), DelegateBB);
PostBB->splice(PostBB->end(), PreBB, SplitPos, PreBB->end());
PostBB->transferSuccessors(PreBB);
} else {
// - Before:
// ehpad:
// range_end
// catch
// ...
//
// - After:
// pre_bb: (new)
// range_end
// delegate_bb: (new)
// delegate
// post_bb: (previous 'ehpad')
// catch
// ...
assert(EndBB->isEHPad());
PreBB = MF.CreateMachineBasicBlock();
PostBB = EndBB;
MF.insert(PostBB->getIterator(), PreBB);
MF.insert(PostBB->getIterator(), DelegateBB);
PreBB->splice(PreBB->end(), PostBB, PostBB->begin(), SplitPos);
// We don't need to transfer predecessors of the EH pad to 'PreBB',
// because an EH pad's predecessors are all through unwind edges and they
// should still unwind to the EH pad, not PreBB.
}
unstackifyVRegsUsedInSplitBB(*PreBB, *PostBB);
PreBB->addSuccessor(DelegateBB);
PreBB->addSuccessor(PostBB);
}
// Add 'delegate' instruction in the delegate BB created above.
MachineInstr *Delegate = BuildMI(DelegateBB, RangeEnd->getDebugLoc(),
TII.get(WebAssembly::DELEGATE))
.addMBB(DelegateDest);
registerTryScope(Try, Delegate, nullptr);
}
bool WebAssemblyCFGStackify::fixCallUnwindMismatches(MachineFunction &MF) {
// Linearizing the control flow by placing TRY / END_TRY markers can create
// mismatches in unwind destinations for throwing instructions, such as calls.
//
// We use the 'delegate' instruction to fix the unwind mismatches. 'delegate'
// instruction delegates an exception to an outer 'catch'. It can target not
// only 'catch' but all block-like structures including another 'delegate',
// but with slightly different semantics than branches. When it targets a
// 'catch', it will delegate the exception to that catch. It is being
// discussed how to define the semantics when 'delegate''s target is a non-try
// block: it will either be a validation failure or it will target the next
// outer try-catch. But anyway our LLVM backend currently does not generate
// such code. The example below illustrates where the 'delegate' instruction
// in the middle will delegate the exception to, depending on the value of N.
// try
// try
// block
// try
// try
// call @foo
// delegate N ;; Where will this delegate to?
// catch ;; N == 0
// end
// end ;; N == 1 (invalid; will not be generated)
// delegate ;; N == 2
// catch ;; N == 3
// end
// ;; N == 4 (to caller)
// 1. When an instruction may throw, but the EH pad it will unwind to can be
// different from the original CFG.
//
// Example: we have the following CFG:
// bb0:
// call @foo ; if it throws, unwind to bb2
// bb1:
// call @bar ; if it throws, unwind to bb3
// bb2 (ehpad):
// catch
// ...
// bb3 (ehpad)
// catch
// ...
//
// And the CFG is sorted in this order. Then after placing TRY markers, it
// will look like: (BB markers are omitted)
// try
// try
// call @foo
// call @bar ;; if it throws, unwind to bb3
// catch ;; ehpad (bb2)
// ...
// end_try
// catch ;; ehpad (bb3)
// ...
// end_try
//
// Now if bar() throws, it is going to end up ip in bb2, not bb3, where it
// is supposed to end up. We solve this problem by wrapping the mismatching
// call with an inner try-delegate that rethrows the exception to the right
// 'catch'.
//
// try
// try
// call @foo
// try ;; (new)
// call @bar
// delegate 1 (bb3) ;; (new)
// catch ;; ehpad (bb2)
// ...
// end_try
// catch ;; ehpad (bb3)
// ...
// end_try
//
// ---
// 2. The same as 1, but in this case an instruction unwinds to a caller
// function and not another EH pad.
//
// Example: we have the following CFG:
// bb0:
// call @foo ; if it throws, unwind to bb2
// bb1:
// call @bar ; if it throws, unwind to caller
// bb2 (ehpad):
// catch
// ...
//
// And the CFG is sorted in this order. Then after placing TRY markers, it
// will look like:
// try
// call @foo
// call @bar ;; if it throws, unwind to caller
// catch ;; ehpad (bb2)
// ...
// end_try
//
// Now if bar() throws, it is going to end up ip in bb2, when it is supposed
// throw up to the caller. We solve this problem in the same way, but in this
// case 'delegate's immediate argument is the number of block depths + 1,
// which means it rethrows to the caller.
// try
// call @foo
// try ;; (new)
// call @bar
// delegate 1 (caller) ;; (new)
// catch ;; ehpad (bb2)
// ...
// end_try
//
// Before rewriteDepthImmediates, delegate's argument is a BB. In case of the
// caller, it will take a fake BB generated by getFakeCallerBlock(), which
// will be converted to a correct immediate argument later.
//
// In case there are multiple calls in a BB that may throw to the caller, they
// can be wrapped together in one nested try-delegate scope. (In 1, this
// couldn't happen, because may-throwing instruction there had an unwind
// destination, i.e., it was an invoke before, and there could be only one
// invoke within a BB.)
SmallVector<const MachineBasicBlock *, 8> EHPadStack;
// Range of intructions to be wrapped in a new nested try/catch. A range
// exists in a single BB and does not span multiple BBs.
using TryRange = std::pair<MachineInstr *, MachineInstr *>;
// In original CFG, <unwind destination BB, a vector of try ranges>
DenseMap<MachineBasicBlock *, SmallVector<TryRange, 4>> UnwindDestToTryRanges;
// Gather possibly throwing calls (i.e., previously invokes) whose current
// unwind destination is not the same as the original CFG. (Case 1)
for (auto &MBB : reverse(MF)) {
bool SeenThrowableInstInBB = false;
for (auto &MI : reverse(MBB)) {
if (MI.getOpcode() == WebAssembly::TRY)
EHPadStack.pop_back();
else if (WebAssembly::isCatch(MI.getOpcode()))
EHPadStack.push_back(MI.getParent());
// In this loop we only gather calls that have an EH pad to unwind. So
// there will be at most 1 such call (= invoke) in a BB, so after we've
// seen one, we can skip the rest of BB. Also if MBB has no EH pad
// successor or MI does not throw, this is not an invoke.
if (SeenThrowableInstInBB || !MBB.hasEHPadSuccessor() ||
!WebAssembly::mayThrow(MI))
continue;
SeenThrowableInstInBB = true;
// If the EH pad on the stack top is where this instruction should unwind
// next, we're good.
MachineBasicBlock *UnwindDest = getFakeCallerBlock(MF);
for (auto *Succ : MBB.successors()) {
// Even though semantically a BB can have multiple successors in case an
// exception is not caught by a catchpad, in our backend implementation
// it is guaranteed that a BB can have at most one EH pad successor. For
// details, refer to comments in findWasmUnwindDestinations function in
// SelectionDAGBuilder.cpp.
if (Succ->isEHPad()) {
UnwindDest = Succ;
break;
}
}
if (EHPadStack.back() == UnwindDest)
continue;
// Include EH_LABELs in the range before and afer the invoke
MachineInstr *RangeBegin = &MI, *RangeEnd = &MI;
if (RangeBegin->getIterator() != MBB.begin() &&
std::prev(RangeBegin->getIterator())->isEHLabel())
RangeBegin = &*std::prev(RangeBegin->getIterator());
if (std::next(RangeEnd->getIterator()) != MBB.end() &&
std::next(RangeEnd->getIterator())->isEHLabel())
RangeEnd = &*std::next(RangeEnd->getIterator());
// If not, record the range.
UnwindDestToTryRanges[UnwindDest].push_back(
TryRange(RangeBegin, RangeEnd));
LLVM_DEBUG(dbgs() << "- Call unwind mismatch: MBB = " << MBB.getName()
<< "\nCall = " << MI
<< "\nOriginal dest = " << UnwindDest->getName()
<< " Current dest = " << EHPadStack.back()->getName()
<< "\n\n");
}
}
assert(EHPadStack.empty());
// Gather possibly throwing calls that are supposed to unwind up to the caller
// if they throw, but currently unwind to an incorrect destination. Unlike the
// loop above, there can be multiple calls within a BB that unwind to the
// caller, which we should group together in a range. (Case 2)
MachineInstr *RangeBegin = nullptr, *RangeEnd = nullptr; // inclusive
// Record the range.
auto RecordCallerMismatchRange = [&](const MachineBasicBlock *CurrentDest) {
UnwindDestToTryRanges[getFakeCallerBlock(MF)].push_back(
TryRange(RangeBegin, RangeEnd));
LLVM_DEBUG(dbgs() << "- Call unwind mismatch: MBB = "
<< RangeBegin->getParent()->getName()
<< "\nRange begin = " << *RangeBegin
<< "Range end = " << *RangeEnd
<< "\nOriginal dest = caller Current dest = "
<< CurrentDest->getName() << "\n\n");
RangeBegin = RangeEnd = nullptr; // Reset range pointers
};
for (auto &MBB : reverse(MF)) {
bool SeenThrowableInstInBB = false;
for (auto &MI : reverse(MBB)) {
bool MayThrow = WebAssembly::mayThrow(MI);
// If MBB has an EH pad successor and this is the last instruction that
// may throw, this instruction unwinds to the EH pad and not to the
// caller.
if (MBB.hasEHPadSuccessor() && MayThrow && !SeenThrowableInstInBB)
SeenThrowableInstInBB = true;
// We wrap up the current range when we see a marker even if we haven't
// finished a BB.
else if (RangeEnd && WebAssembly::isMarker(MI.getOpcode()))
RecordCallerMismatchRange(EHPadStack.back());
// If EHPadStack is empty, that means it correctly unwinds to the caller
// if it throws, so we're good. If MI does not throw, we're good too.
else if (EHPadStack.empty() || !MayThrow) {
}
// We found an instruction that unwinds to the caller but currently has an
// incorrect unwind destination. Create a new range or increment the
// currently existing range.
else {
if (!RangeEnd)
RangeBegin = RangeEnd = &MI;
else
RangeBegin = &MI;
}
// Update EHPadStack.
if (MI.getOpcode() == WebAssembly::TRY)
EHPadStack.pop_back();
else if (WebAssembly::isCatch(MI.getOpcode()))
EHPadStack.push_back(MI.getParent());
}
if (RangeEnd)
RecordCallerMismatchRange(EHPadStack.back());
}
assert(EHPadStack.empty());
// We don't have any unwind destination mismatches to resolve.
if (UnwindDestToTryRanges.empty())
return false;
// Now we fix the mismatches by wrapping calls with inner try-delegates.
for (auto &P : UnwindDestToTryRanges) {
NumCallUnwindMismatches += P.second.size();
MachineBasicBlock *UnwindDest = P.first;
auto &TryRanges = P.second;
for (auto Range : TryRanges) {
MachineInstr *RangeBegin = nullptr, *RangeEnd = nullptr;
std::tie(RangeBegin, RangeEnd) = Range;
auto *MBB = RangeBegin->getParent();
// If this BB has an EH pad successor, i.e., ends with an 'invoke', now we
// are going to wrap the invoke with try-delegate, making the 'delegate'
// BB the new successor instead, so remove the EH pad succesor here. The
// BB may not have an EH pad successor if calls in this BB throw to the
// caller.
MachineBasicBlock *EHPad = nullptr;
for (auto *Succ : MBB->successors()) {
if (Succ->isEHPad()) {
EHPad = Succ;
break;
}
}
if (EHPad)
MBB->removeSuccessor(EHPad);
addTryDelegate(RangeBegin, RangeEnd, UnwindDest);
}
}
return true;
}
bool WebAssemblyCFGStackify::fixCatchUnwindMismatches(MachineFunction &MF) {
// There is another kind of unwind destination mismatches besides call unwind
// mismatches, which we will call "catch unwind mismatches". See this example
// after the marker placement:
// try
// try
// call @foo
// catch __cpp_exception ;; ehpad A (next unwind dest: caller)
// ...
// end_try
// catch_all ;; ehpad B
// ...
// end_try
//
// 'call @foo's unwind destination is the ehpad A. But suppose 'call @foo'
// throws a foreign exception that is not caught by ehpad A, and its next
// destination should be the caller. But after control flow linearization,
// another EH pad can be placed in between (e.g. ehpad B here), making the
// next unwind destination incorrect. In this case, the foreign exception
// will instead go to ehpad B and will be caught there instead. In this
// example the correct next unwind destination is the caller, but it can be
// another outer catch in other cases.
//
// There is no specific 'call' or 'throw' instruction to wrap with a
// try-delegate, so we wrap the whole try-catch-end with a try-delegate and
// make it rethrow to the right destination, as in the example below:
// try
// try ;; (new)
// try
// call @foo
// catch __cpp_exception ;; ehpad A (next unwind dest: caller)
// ...
// end_try
// delegate 1 (caller) ;; (new)
// catch_all ;; ehpad B
// ...
// end_try
const auto *EHInfo = MF.getWasmEHFuncInfo();
SmallVector<const MachineBasicBlock *, 8> EHPadStack;
// For EH pads that have catch unwind mismatches, a map of <EH pad, its
// correct unwind destination>.
DenseMap<MachineBasicBlock *, MachineBasicBlock *> EHPadToUnwindDest;
for (auto &MBB : reverse(MF)) {
for (auto &MI : reverse(MBB)) {
if (MI.getOpcode() == WebAssembly::TRY)
EHPadStack.pop_back();
else if (MI.getOpcode() == WebAssembly::DELEGATE)
EHPadStack.push_back(&MBB);
else if (WebAssembly::isCatch(MI.getOpcode())) {
auto *EHPad = &MBB;
// catch_all always catches an exception, so we don't need to do
// anything
if (MI.getOpcode() == WebAssembly::CATCH_ALL) {
}
// This can happen when the unwind dest was removed during the
// optimization, e.g. because it was unreachable.
else if (EHPadStack.empty() && EHInfo->hasUnwindDest(EHPad)) {
LLVM_DEBUG(dbgs() << "EHPad (" << EHPad->getName()
<< "'s unwind destination does not exist anymore"
<< "\n\n");
}
// The EHPad's next unwind destination is the caller, but we incorrectly
// unwind to another EH pad.
else if (!EHPadStack.empty() && !EHInfo->hasUnwindDest(EHPad)) {
EHPadToUnwindDest[EHPad] = getFakeCallerBlock(MF);
LLVM_DEBUG(dbgs()
<< "- Catch unwind mismatch:\nEHPad = " << EHPad->getName()
<< " Original dest = caller Current dest = "
<< EHPadStack.back()->getName() << "\n\n");
}
// The EHPad's next unwind destination is an EH pad, whereas we
// incorrectly unwind to another EH pad.
else if (!EHPadStack.empty() && EHInfo->hasUnwindDest(EHPad)) {
auto *UnwindDest = EHInfo->getUnwindDest(EHPad);
if (EHPadStack.back() != UnwindDest) {
EHPadToUnwindDest[EHPad] = UnwindDest;
LLVM_DEBUG(dbgs() << "- Catch unwind mismatch:\nEHPad = "
<< EHPad->getName() << " Original dest = "
<< UnwindDest->getName() << " Current dest = "
<< EHPadStack.back()->getName() << "\n\n");
}
}
EHPadStack.push_back(EHPad);
}
}
}
assert(EHPadStack.empty());
if (EHPadToUnwindDest.empty())
return false;
NumCatchUnwindMismatches += EHPadToUnwindDest.size();
SmallPtrSet<MachineBasicBlock *, 4> NewEndTryBBs;
for (auto &P : EHPadToUnwindDest) {
MachineBasicBlock *EHPad = P.first;
MachineBasicBlock *UnwindDest = P.second;
MachineInstr *Try = EHPadToTry[EHPad];
MachineInstr *EndTry = BeginToEnd[Try];
addTryDelegate(Try, EndTry, UnwindDest);
NewEndTryBBs.insert(EndTry->getParent());
}
// Adding a try-delegate wrapping an existing try-catch-end can make existing
// branch destination BBs invalid. For example,
//
// - Before:
// bb0:
// block
// br bb3
// bb1:
// try
// ...
// bb2: (ehpad)
// catch
// bb3:
// end_try
// end_block ;; 'br bb3' targets here
//
// Suppose this try-catch-end has a catch unwind mismatch, so we need to wrap
// this with a try-delegate. Then this becomes:
//
// - After:
// bb0:
// block
// br bb3 ;; invalid destination!
// bb1:
// try ;; (new instruction)
// try
// ...
// bb2: (ehpad)
// catch
// bb3:
// end_try ;; 'br bb3' still incorrectly targets here!
// delegate_bb: ;; (new BB)
// delegate ;; (new instruction)
// split_bb: ;; (new BB)
// end_block
//
// Now 'br bb3' incorrectly branches to an inner scope.
//
// As we can see in this case, when branches target a BB that has both
// 'end_try' and 'end_block' and the BB is split to insert a 'delegate', we
// have to remap existing branch destinations so that they target not the
// 'end_try' BB but the new 'end_block' BB. There can be multiple 'delegate's
// in between, so we try to find the next BB with 'end_block' instruction. In
// this example, the 'br bb3' instruction should be remapped to 'br split_bb'.
for (auto &MBB : MF) {
for (auto &MI : MBB) {
if (MI.isTerminator()) {
for (auto &MO : MI.operands()) {
if (MO.isMBB() && NewEndTryBBs.count(MO.getMBB())) {
auto *BrDest = MO.getMBB();
bool FoundEndBlock = false;
for (; std::next(BrDest->getIterator()) != MF.end();
BrDest = BrDest->getNextNode()) {
for (const auto &MI : *BrDest) {
if (MI.getOpcode() == WebAssembly::END_BLOCK) {
FoundEndBlock = true;
break;
}
}
if (FoundEndBlock)
break;
}
assert(FoundEndBlock);
MO.setMBB(BrDest);
}
}
}
}
}
return true;
}
void WebAssemblyCFGStackify::recalculateScopeTops(MachineFunction &MF) {
// Renumber BBs and recalculate ScopeTop info because new BBs might have been
// created and inserted during fixing unwind mismatches.
MF.RenumberBlocks();
ScopeTops.clear();
ScopeTops.resize(MF.getNumBlockIDs());
for (auto &MBB : reverse(MF)) {
for (auto &MI : reverse(MBB)) {
if (ScopeTops[MBB.getNumber()])
break;
switch (MI.getOpcode()) {
case WebAssembly::END_BLOCK:
case WebAssembly::END_LOOP:
case WebAssembly::END_TRY:
case WebAssembly::DELEGATE:
updateScopeTops(EndToBegin[&MI]->getParent(), &MBB);
break;
case WebAssembly::CATCH:
case WebAssembly::CATCH_ALL:
updateScopeTops(EHPadToTry[&MBB]->getParent(), &MBB);
break;
}
}
}
}
/// In normal assembly languages, when the end of a function is unreachable,
/// because the function ends in an infinite loop or a noreturn call or similar,
/// it isn't necessary to worry about the function return type at the end of
/// the function, because it's never reached. However, in WebAssembly, blocks
/// that end at the function end need to have a return type signature that
/// matches the function signature, even though it's unreachable. This function
/// checks for such cases and fixes up the signatures.
void WebAssemblyCFGStackify::fixEndsAtEndOfFunction(MachineFunction &MF) {
const auto &MFI = *MF.getInfo<WebAssemblyFunctionInfo>();
if (MFI.getResults().empty())
return;
// MCInstLower will add the proper types to multivalue signatures based on the
// function return type
WebAssembly::BlockType RetType =
MFI.getResults().size() > 1
? WebAssembly::BlockType::Multivalue
: WebAssembly::BlockType(
WebAssembly::toValType(MFI.getResults().front()));
SmallVector<MachineBasicBlock::reverse_iterator, 4> Worklist;
Worklist.push_back(MF.rbegin()->rbegin());
auto Process = [&](MachineBasicBlock::reverse_iterator It) {
auto *MBB = It->getParent();
while (It != MBB->rend()) {
MachineInstr &MI = *It++;
if (MI.isPosition() || MI.isDebugInstr())
continue;
switch (MI.getOpcode()) {
case WebAssembly::END_TRY: {
// If a 'try''s return type is fixed, both its try body and catch body
// should satisfy the return type, so we need to search 'end'
// instructions before its corresponding 'catch' too.
auto *EHPad = TryToEHPad.lookup(EndToBegin[&MI]);
assert(EHPad);
auto NextIt =
std::next(WebAssembly::findCatch(EHPad)->getReverseIterator());
if (NextIt != EHPad->rend())
Worklist.push_back(NextIt);
LLVM_FALLTHROUGH;
}
case WebAssembly::END_BLOCK:
case WebAssembly::END_LOOP:
case WebAssembly::DELEGATE:
EndToBegin[&MI]->getOperand(0).setImm(int32_t(RetType));
continue;
default:
// Something other than an `end`. We're done for this BB.
return;
}
}
// We've reached the beginning of a BB. Continue the search in the previous
// BB.
Worklist.push_back(MBB->getPrevNode()->rbegin());
};
while (!Worklist.empty())
Process(Worklist.pop_back_val());
}
// WebAssembly functions end with an end instruction, as if the function body
// were a block.
static void appendEndToFunction(MachineFunction &MF,
const WebAssemblyInstrInfo &TII) {
BuildMI(MF.back(), MF.back().end(),
MF.back().findPrevDebugLoc(MF.back().end()),
TII.get(WebAssembly::END_FUNCTION));
}
/// Insert LOOP/TRY/BLOCK markers at appropriate places.
void WebAssemblyCFGStackify::placeMarkers(MachineFunction &MF) {
// We allocate one more than the number of blocks in the function to
// accommodate for the possible fake block we may insert at the end.
ScopeTops.resize(MF.getNumBlockIDs() + 1);
// Place the LOOP for MBB if MBB is the header of a loop.
for (auto &MBB : MF)
placeLoopMarker(MBB);
const MCAsmInfo *MCAI = MF.getTarget().getMCAsmInfo();
for (auto &MBB : MF) {
if (MBB.isEHPad()) {
// Place the TRY for MBB if MBB is the EH pad of an exception.
if (MCAI->getExceptionHandlingType() == ExceptionHandling::Wasm &&
MF.getFunction().hasPersonalityFn())
placeTryMarker(MBB);
} else {
// Place the BLOCK for MBB if MBB is branched to from above.
placeBlockMarker(MBB);
}
}
// Fix mismatches in unwind destinations induced by linearizing the code.
if (MCAI->getExceptionHandlingType() == ExceptionHandling::Wasm &&
MF.getFunction().hasPersonalityFn()) {
bool Changed = fixCallUnwindMismatches(MF);
Changed |= fixCatchUnwindMismatches(MF);
if (Changed)
recalculateScopeTops(MF);
}
}
unsigned WebAssemblyCFGStackify::getBranchDepth(
const SmallVectorImpl<EndMarkerInfo> &Stack, const MachineBasicBlock *MBB) {
unsigned Depth = 0;
for (auto X : reverse(Stack)) {
if (X.first == MBB)
break;
++Depth;
}
assert(Depth < Stack.size() && "Branch destination should be in scope");
return Depth;
}
unsigned WebAssemblyCFGStackify::getDelegateDepth(
const SmallVectorImpl<EndMarkerInfo> &Stack, const MachineBasicBlock *MBB) {
if (MBB == FakeCallerBB)
return Stack.size();
// Delegate's destination is either a catch or a another delegate BB. When the
// destination is another delegate, we can compute the argument in the same
// way as branches, because the target delegate BB only contains the single
// delegate instruction.
if (!MBB->isEHPad()) // Target is a delegate BB
return getBranchDepth(Stack, MBB);
// When the delegate's destination is a catch BB, we need to use its
// corresponding try's end_try BB because Stack contains each marker's end BB.
// Also we need to check if the end marker instruction matches, because a
// single BB can contain multiple end markers, like this:
// bb:
// END_BLOCK
// END_TRY
// END_BLOCK
// END_TRY
// ...
//
// In case of branches getting the immediate that targets any of these is
// fine, but delegate has to exactly target the correct try.
unsigned Depth = 0;
const MachineInstr *EndTry = BeginToEnd[EHPadToTry[MBB]];
for (auto X : reverse(Stack)) {
if (X.first == EndTry->getParent() && X.second == EndTry)
break;
++Depth;
}
assert(Depth < Stack.size() && "Delegate destination should be in scope");
return Depth;
}
unsigned WebAssemblyCFGStackify::getRethrowDepth(
const SmallVectorImpl<EndMarkerInfo> &Stack,
const SmallVectorImpl<const MachineBasicBlock *> &EHPadStack) {
unsigned Depth = 0;
// In our current implementation, rethrows always rethrow the exception caught
// by the innermost enclosing catch. This means while traversing Stack in the
// reverse direction, when we encounter END_TRY, we should check if the
// END_TRY corresponds to the current innermost EH pad. For example:
// try
// ...
// catch ;; (a)
// try
// rethrow 1 ;; (b)
// catch ;; (c)
// rethrow 0 ;; (d)
// end ;; (e)
// end ;; (f)
//
// When we are at 'rethrow' (d), while reversely traversing Stack the first
// 'end' we encounter is the 'end' (e), which corresponds to the 'catch' (c).
// And 'rethrow' (d) rethrows the exception caught by 'catch' (c), so we stop
// there and the depth should be 0. But when we are at 'rethrow' (b), it
// rethrows the exception caught by 'catch' (a), so when traversing Stack
// reversely, we should skip the 'end' (e) and choose 'end' (f), which
// corresponds to 'catch' (a).
for (auto X : reverse(Stack)) {
const MachineInstr *End = X.second;
if (End->getOpcode() == WebAssembly::END_TRY) {
auto *EHPad = TryToEHPad[EndToBegin[End]];
if (EHPadStack.back() == EHPad)
break;
}
++Depth;
}
assert(Depth < Stack.size() && "Rethrow destination should be in scope");
return Depth;
}
void WebAssemblyCFGStackify::rewriteDepthImmediates(MachineFunction &MF) {
// Now rewrite references to basic blocks to be depth immediates.
SmallVector<EndMarkerInfo, 8> Stack;
SmallVector<const MachineBasicBlock *, 8> EHPadStack;
for (auto &MBB : reverse(MF)) {
for (MachineInstr &MI : llvm::reverse(MBB)) {
switch (MI.getOpcode()) {
case WebAssembly::BLOCK:
case WebAssembly::TRY:
assert(ScopeTops[Stack.back().first->getNumber()]->getNumber() <=
MBB.getNumber() &&
"Block/try marker should be balanced");
Stack.pop_back();
break;
case WebAssembly::LOOP:
assert(Stack.back().first == &MBB && "Loop top should be balanced");
Stack.pop_back();
break;
case WebAssembly::END_BLOCK:
Stack.push_back(std::make_pair(&MBB, &MI));
break;
case WebAssembly::END_TRY: {
// We handle DELEGATE in the default level, because DELEGATE has
// immediate operands to rewrite.
Stack.push_back(std::make_pair(&MBB, &MI));
auto *EHPad = TryToEHPad[EndToBegin[&MI]];
EHPadStack.push_back(EHPad);
break;
}
case WebAssembly::END_LOOP:
Stack.push_back(std::make_pair(EndToBegin[&MI]->getParent(), &MI));
break;
case WebAssembly::CATCH:
case WebAssembly::CATCH_ALL:
EHPadStack.pop_back();
break;
case WebAssembly::RETHROW:
MI.getOperand(0).setImm(getRethrowDepth(Stack, EHPadStack));
break;
default:
if (MI.isTerminator()) {
// Rewrite MBB operands to be depth immediates.
SmallVector<MachineOperand, 4> Ops(MI.operands());
while (MI.getNumOperands() > 0)
MI.removeOperand(MI.getNumOperands() - 1);
for (auto MO : Ops) {
if (MO.isMBB()) {
if (MI.getOpcode() == WebAssembly::DELEGATE)
MO = MachineOperand::CreateImm(
getDelegateDepth(Stack, MO.getMBB()));
else
MO = MachineOperand::CreateImm(
getBranchDepth(Stack, MO.getMBB()));
}
MI.addOperand(MF, MO);
}
}
if (MI.getOpcode() == WebAssembly::DELEGATE)
Stack.push_back(std::make_pair(&MBB, &MI));
break;
}
}
}
assert(Stack.empty() && "Control flow should be balanced");
}
void WebAssemblyCFGStackify::cleanupFunctionData(MachineFunction &MF) {
if (FakeCallerBB)
MF.deleteMachineBasicBlock(FakeCallerBB);
AppendixBB = FakeCallerBB = nullptr;
}
void WebAssemblyCFGStackify::releaseMemory() {
ScopeTops.clear();
BeginToEnd.clear();
EndToBegin.clear();
TryToEHPad.clear();
EHPadToTry.clear();
}
bool WebAssemblyCFGStackify::runOnMachineFunction(MachineFunction &MF) {
LLVM_DEBUG(dbgs() << "********** CFG Stackifying **********\n"
"********** Function: "
<< MF.getName() << '\n');
const MCAsmInfo *MCAI = MF.getTarget().getMCAsmInfo();
releaseMemory();
// Liveness is not tracked for VALUE_STACK physreg.
MF.getRegInfo().invalidateLiveness();
// Place the BLOCK/LOOP/TRY markers to indicate the beginnings of scopes.
placeMarkers(MF);
// Remove unnecessary instructions possibly introduced by try/end_trys.
if (MCAI->getExceptionHandlingType() == ExceptionHandling::Wasm &&
MF.getFunction().hasPersonalityFn())
removeUnnecessaryInstrs(MF);
// Convert MBB operands in terminators to relative depth immediates.
rewriteDepthImmediates(MF);
// Fix up block/loop/try signatures at the end of the function to conform to
// WebAssembly's rules.
fixEndsAtEndOfFunction(MF);
// Add an end instruction at the end of the function body.
const auto &TII = *MF.getSubtarget<WebAssemblySubtarget>().getInstrInfo();
if (!MF.getSubtarget<WebAssemblySubtarget>()
.getTargetTriple()
.isOSBinFormatELF())
appendEndToFunction(MF, TII);
cleanupFunctionData(MF);
MF.getInfo<WebAssemblyFunctionInfo>()->setCFGStackified();
return true;
}