// Place internal breakpoints to get out of the current function. This may place
// multiple internal breakpoints, and it may place them more than one frame up.
// Some instructions can cause PHP to be invoked without an explicit call. A set
// which causes a destructor to run, a iteration init which causes an object's
// next() method to run, a RetC which causes destructors to run, etc. This
// recgonizes such cases and ensures we have internal breakpoints to cover the
// destination(s) of such instructions.
void CmdFlowControl::setupStepOuts() {
// Existing step outs should be cleaned up before making new ones.
assert(!hasStepOuts());
auto fp = g_context->getFP();
if (!fp) return; // No place to step out to!
Offset returnOffset;
bool fromVMEntry;
while (!hasStepOuts()) {
fp = g_context->getPrevVMState(fp, &returnOffset, nullptr, &fromVMEntry);
// If we've run off the top of the stack, just return having setup no
// step outs. This will cause cmds like Next and Out to just let the program
// run, which is appropriate.
if (!fp) break;
Unit* returnUnit = fp->m_func->unit();
PC returnPC = returnUnit->at(returnOffset);
TRACE(2, "CmdFlowControl::setupStepOuts: at '%s' offset %d opcode %s\n",
fp->m_func->fullName()->data(), returnOffset,
opcodeToName(*reinterpret_cast<const Op*>(returnPC)));
// Don't step out to generated functions, keep looking.
if (fp->m_func->line1() == 0) continue;
if (fromVMEntry) {
TRACE(2, "CmdFlowControl::setupStepOuts: VM entry\n");
// We only execute this for opcodes which invoke more PHP, and that does
// not include switches. Thus, we'll have at most two destinations.
assert(!isSwitch(*reinterpret_cast<const Op*>(returnPC)) &&
(numSuccs(reinterpret_cast<const Op*>(returnPC)) <= 2));
// Set an internal breakpoint after the instruction if it can fall thru.
if (instrAllowsFallThru(*reinterpret_cast<const Op*>(returnPC))) {
Offset nextOffset = returnOffset + instrLen((Op*)returnPC);
TRACE(2, "CmdFlowControl: step out to '%s' offset %d (fall-thru)\n",
fp->m_func->fullName()->data(), nextOffset);
m_stepOut1 = StepDestination(returnUnit, nextOffset);
}
// Set an internal breakpoint at the target of a control flow instruction.
// A good example of a control flow op that invokes PHP is IterNext.
if (instrIsControlFlow(*reinterpret_cast<const Op*>(returnPC))) {
Offset target =
instrJumpTarget(reinterpret_cast<const Op*>(returnPC), 0);
if (target != InvalidAbsoluteOffset) {
Offset targetOffset = returnOffset + target;
TRACE(2, "CmdFlowControl: step out to '%s' offset %d (jump target)\n",
fp->m_func->fullName()->data(), targetOffset);
m_stepOut2 = StepDestination(returnUnit, targetOffset);
}
}
// If we have no place to step out to, then unwind another frame and try
// again. The most common case that leads here is Ret*, which does not
// fall-thru and has no encoded target.
} else {
TRACE(2, "CmdFlowControl: step out to '%s' offset %d\n",
fp->m_func->fullName()->data(), returnOffset);
m_stepOut1 = StepDestination(returnUnit, returnOffset);
}
}
}
void CmdNext::onBeginInterrupt(DebuggerProxy& proxy, CmdInterrupt& interrupt) {
TRACE(2, "CmdNext::onBeginInterrupt\n");
assertx(!m_complete); // Complete cmds should not be asked to do work.
ActRec *fp = vmfp();
if (!fp) {
// If we have no frame just wait for the next instruction to be interpreted.
m_needsVMInterrupt = true;
return;
}
PC pc = vmpc();
Unit* unit = fp->m_func->unit();
Offset offset = unit->offsetOf(pc);
TRACE(2, "CmdNext: pc %p, opcode %s at '%s' offset %d\n",
pc,
opcodeToName(peek_op(pc)),
fp->m_func->fullName()->data(),
offset);
int currentVMDepth = g_context->m_nesting;
int currentStackDepth = proxy.getStackDepth();
TRACE(2, "CmdNext: original depth %d:%d, current depth %d:%d\n",
m_vmDepth, m_stackDepth, currentVMDepth, currentStackDepth);
// Where are we on the stack now vs. when we started? Breaking the answer down
// into distinct variables helps the clarity of the algorithm below.
bool deeper = false;
bool originalDepth = false;
if ((currentVMDepth == m_vmDepth) && (currentStackDepth == m_stackDepth)) {
originalDepth = true;
} else if ((currentVMDepth > m_vmDepth) ||
((currentVMDepth == m_vmDepth) &&
(currentStackDepth > m_stackDepth))) {
deeper = true;
}
m_needsVMInterrupt = false; // Will be set again below if still needed.
// First consider if we've got internal breakpoints setup. These are used when
// we can make an accurate prediction of where execution should flow,
// eventually, and when we want to let the program run normally until we get
// there.
if (hasStepOuts() || hasStepResumable()) {
TRACE(2, "CmdNext: checking internal breakpoint(s)\n");
if (atStepOutOffset(unit, offset)) {
if (deeper) return; // Recursion
TRACE(2, "CmdNext: hit step-out\n");
} else if (atStepResumableOffset(unit, offset)) {
if (m_stepResumableId != getResumableId(fp)) return;
TRACE(2, "CmdNext: hit step-cont\n");
// We're in the resumable we expect. This may be at a
// different stack depth, though, especially if we've moved from
// the original function to the resumable. Update the depth
// accordingly.
if (!originalDepth) {
m_vmDepth = currentVMDepth;
m_stackDepth = currentStackDepth;
deeper = false;
originalDepth = true;
}
} else if (interrupt.getInterruptType() == ExceptionHandler) {
// Entering an exception handler may take us someplace we weren't
// expecting. Adjust internal breakpoints accordingly. First case is easy.
if (deeper) {
TRACE(2, "CmdNext: exception handler, deeper\n");
return;
}
// For step-conts, we ignore handlers at the original level if we're not
// in the original resumable. We don't care about exception handlers
// in resumables being driven at the same level.
if (hasStepResumable() && originalDepth &&
(m_stepResumableId != getResumableId(fp))) {
TRACE(2, "CmdNext: exception handler, original depth, wrong cont\n");
return;
}
// Sometimes we have handlers in generated code, i.e., Continuation::next.
// These just help propagate exceptions so ignore those.
if (fp->m_func->line1() == 0) {
TRACE(2, "CmdNext: exception handler, ignoring func with no source\n");
return;
}
if (fp->m_func->isBuiltin()) {
TRACE(2, "CmdNext: exception handler, ignoring builtin functions\n");
return;
}
TRACE(2, "CmdNext: exception handler altering expected flow\n");
} else {
// We have internal breakpoints setup, but we haven't hit one yet. Keep
// running until we reach one.
TRACE(2, "CmdNext: waiting to hit internal breakpoint...\n");
return;
}
// We've hit one internal breakpoint at a useful place, or decided we don't,
// need them, so we can remove them all now.
cleanupStepOuts();
cleanupStepResumable();
}
if (interrupt.getInterruptType() == ExceptionHandler) {
// If we're about to enter an exception handler we turn interrupts on to
// ensure we stop when control reaches the handler. The normal logic below
// will decide if we're done at that point or not.
TRACE(2, "CmdNext: exception handler\n");
removeLocationFilter();
m_needsVMInterrupt = true;
return;
}
if (m_skippingAwait) {
m_skippingAwait = false;
stepAfterAwait();
return;
}
if (deeper) {
TRACE(2, "CmdNext: deeper, setup step out to get back to original line\n");
setupStepOuts();
// We can nuke the entire location filter here since we'll re-install it
// when we get back to the old level. Keeping it installed may be more
// efficient if we were on a large line, but there is a penalty for every
// opcode executed while it's installed and that's bad if there's a lot of
// code called from that line.
removeLocationFilter();
return;
}
if (originalDepth && (m_loc == interrupt.getFileLine())) {
TRACE(2, "CmdNext: not complete, still on same line\n");
stepCurrentLine(interrupt, fp, pc);
return;
}
TRACE(2, "CmdNext: operation complete.\n");
m_complete = (decCount() == 0);
if (!m_complete) {
TRACE(2, "CmdNext: repeat count > 0, start fresh.\n");
onSetup(proxy, interrupt);
}
}
RegionDescPtr selectHotTrace(HotTransContext& ctx,
TransIDSet& selectedSet,
TransIDVec* selectedVec /* = nullptr */) {
auto region = std::make_shared<RegionDesc>();
TransID tid = ctx.tid;
TransID prevId = kInvalidTransID;
selectedSet.clear();
if (selectedVec) selectedVec->clear();
TypedLocations accumPostConds;
// Maps from BlockIds to accumulated post conditions for that block.
// Used to determine if we can add branch-over edges by checking the
// pre-conditions of the successor block.
hphp_hash_map<RegionDesc::BlockId, TypedLocations> blockPostConds;
auto numBCInstrs = ctx.maxBCInstrs;
FTRACE(1, "selectHotTrace: starting with maxBCInstrs = {}\n", numBCInstrs);
while (!selectedSet.count(tid)) {
auto rec = ctx.profData->transRec(tid);
auto blockRegion = rec->region();
if (blockRegion == nullptr) break;
// Break if region would be larger than the specified limit.
if (blockRegion->instrSize() > numBCInstrs) {
FTRACE(2, "selectHotTrace: breaking region at Translation {} because "
"size would exceed of maximum translation limit\n", tid);
break;
}
// If the debugger is attached, only allow single-block regions.
if (prevId != kInvalidTransID && isDebuggerAttachedProcess()) {
FTRACE(2, "selectHotTrace: breaking region at Translation {} "
"because of debugger is attached\n", tid);
break;
}
// Break if block is not the first and it corresponds to the main
// function body entry. This is to prevent creating multiple
// large regions containing the function body (starting at various
// DV funclets).
if (prevId != kInvalidTransID) {
auto const func = rec->func();
auto const bcOffset = rec->startBcOff();
if (func->base() == bcOffset) {
FTRACE(2, "selectHotTrace: breaking region because reached the main "
"function body entry at Translation {} (BC offset {})\n",
tid, bcOffset);
break;
}
}
if (prevId != kInvalidTransID) {
auto sk = rec->srcKey();
if (ctx.profData->optimized(sk)) {
FTRACE(2, "selectHotTrace: breaking region because next sk already "
"optimized, for Translation {}\n", tid);
break;
}
}
bool hasPredBlock = !region->empty();
RegionDesc::BlockId predBlockId = (hasPredBlock ?
region->blocks().back().get()->id() : 0);
auto const& newFirstBlock = blockRegion->entry();
auto newFirstBlockId = newFirstBlock->id();
// Add blockRegion's blocks and arcs to region.
region->append(*blockRegion);
numBCInstrs -= blockRegion->instrSize();
assertx(numBCInstrs >= 0);
if (hasPredBlock) {
region->addArc(predBlockId, newFirstBlockId);
}
selectedSet.insert(tid);
if (selectedVec) selectedVec->push_back(tid);
const auto lastSk = rec->lastSrcKey();
if (breaksRegion(lastSk)) {
FTRACE(2, "selectHotTrace: breaking region because of last instruction "
"in Translation {}: {}\n", tid, opcodeToName(lastSk.op()));
break;
}
auto outArcs = ctx.cfg->outArcs(tid);
if (outArcs.size() == 0) {
FTRACE(2, "selectHotTrace: breaking region because there's no successor "
"for Translation {}\n", tid);
break;
}
auto newLastBlock = blockRegion->blocks().back();
discardPoppedTypes(accumPostConds,
blockRegion->entry()->initialSpOffset());
mergePostConds(accumPostConds, newLastBlock->postConds());
blockPostConds[newLastBlock->id()] = accumPostConds;
TransCFG::ArcPtrVec possibleOutArcs;
for (auto arc : outArcs) {
auto dstRec = ctx.profData->transRec(arc->dst());
auto possibleNext = dstRec->region()->entry();
if (preCondsAreSatisfied(possibleNext, accumPostConds)) {
possibleOutArcs.emplace_back(arc);
}
}
if (possibleOutArcs.size() == 0) {
FTRACE(2, "selectHotTrace: breaking region because postcondition check "
"pruned all successors of Translation {}\n", tid);
break;
}
auto maxWeight = std::numeric_limits<int64_t>::min();
TransCFG::Arc* maxArc = nullptr;
for (auto arc : possibleOutArcs) {
if (arc->weight() >= maxWeight) {
maxWeight = arc->weight();
maxArc = arc;
}
}
assertx(maxArc != nullptr);
prevId = tid;
tid = maxArc->dst();
}
FTRACE(3, "selectHotTrace: before chainRetransBlocks:\n{}\n", show(*region));
region->chainRetransBlocks();
FTRACE(3, "selectHotTrace: after chainRetransBlocks:\n{}\n", show(*region));
return region;
}
bool InliningDecider::shouldInline(SrcKey callerSk,
const Func* callee,
const RegionDesc& region,
uint32_t maxTotalCost) {
auto sk = region.empty() ? SrcKey() : region.start();
assertx(callee);
assertx(sk.func() == callee);
// Tracing return lambdas.
auto refuse = [&] (const char* why) {
FTRACE(2, "shouldInline: rejecting callee region: {}", show(region));
return traceRefusal(m_topFunc, callee, why);
};
auto accept = [&, this] (const char* kind) {
FTRACE(2, "InliningDecider: inlining {}() <- {}()\t<reason: {}>\n",
m_topFunc->fullName()->data(), callee->fullName()->data(), kind);
return true;
};
if (m_stackDepth + callee->maxStackCells() >= kStackCheckLeafPadding) {
return refuse("inlining stack depth limit exceeded");
}
// Even if the func contains NativeImpl we may have broken the trace before
// we hit it.
auto containsNativeImpl = [&] {
for (auto block : region.blocks()) {
if (!block->empty() && block->last().op() == OpNativeImpl) return true;
}
return false;
};
// Try to inline CPP builtin functions.
// The NativeImpl opcode may appear later in the function because of Asserts
// generated in hhbbc
if (callee->isCPPBuiltin() && containsNativeImpl()) {
if (isInlinableCPPBuiltin(callee)) {
return accept("inlinable CPP builtin");
}
return refuse("non-inlinable CPP builtin");
}
// If the function may use a VarEnv (which is stored in the ActRec) or may be
// variadic, we restrict inlined callees to certain whitelisted instructions
// which we know won't actually require these features.
const bool needsCheckVVSafe = callee->attrs() & AttrMayUseVV;
bool hasRet = false;
// Iterate through the region, checking its suitability for inlining.
for (auto const& block : region.blocks()) {
sk = block->start();
for (auto i = 0, n = block->length(); i < n; ++i, sk.advance()) {
auto op = sk.op();
// We don't allow inlined functions in the region. The client is
// expected to disable inlining for the region it gives us to peek.
if (sk.func() != callee) {
return refuse("got region with inlined calls");
}
// Restrict to VV-safe opcodes if necessary.
if (needsCheckVVSafe && !isInliningVVSafe(op)) {
return refuse(folly::format("{} may use dynamic environment",
opcodeToName(op)).str().c_str());
}
// Count the returns.
if (isReturnish(op)) {
hasRet = true;
}
// We can't inline FCallArray. XXX: Why?
if (op == Op::FCallArray) {
return refuse("can't inline FCallArray");
}
}
}
if (!hasRet) {
return refuse("region has no returns");
}
// Refuse if the cost exceeds our thresholds.
// We measure the cost of inlining each callstack and stop when it exceeds a
// certain threshold. (Note that we do not measure the total cost of all the
// inlined calls for a given caller---just the cost of each nested stack.)
const int maxCost = maxTotalCost - m_cost;
const int cost = computeTranslationCost(callerSk, region);
if (cost > maxCost) {
return refuse("too expensive");
}
return accept("small region with return");
}
RegionDescPtr selectHotTrace(TransID triggerId,
const ProfData* profData,
TransCFG& cfg,
TransIDSet& selectedSet,
TransIDVec* selectedVec) {
auto region = std::make_shared<RegionDesc>();
TransID tid = triggerId;
TransID prevId = kInvalidTransID;
selectedSet.clear();
if (selectedVec) selectedVec->clear();
PostConditions accumPostConds;
// Maps BlockIds to the set of BC offsets for its successor blocks.
// Used to prevent multiple successors with the same SrcKey for now.
// This can go away once task #4157613 is done.
hphp_hash_map<RegionDesc::BlockId, SrcKeySet> succSKSet;
// Maps from BlockIds to accumulated post conditions for that block.
// Used to determine if we can add branch-over edges by checking the
// pre-conditions of the successor block.
hphp_hash_map<RegionDesc::BlockId, PostConditions> blockPostConds;
while (!selectedSet.count(tid)) {
RegionDescPtr blockRegion = profData->transRegion(tid);
if (blockRegion == nullptr) break;
// If the debugger is attached, only allow single-block regions.
if (prevId != kInvalidTransID && isDebuggerAttachedProcess()) {
FTRACE(2, "selectHotTrace: breaking region at Translation {} "
"because of debugger is attached\n", tid);
break;
}
// Break if block is not the first and requires reffiness checks.
// Task #2589970: fix translateRegion to support mid-region reffiness checks
if (prevId != kInvalidTransID) {
auto nRefDeps = blockRegion->blocks[0]->reffinessPreds().size();
if (nRefDeps > 0) {
FTRACE(2, "selectHotTrace: breaking region because of refDeps ({}) at "
"Translation {}\n", nRefDeps, tid);
break;
}
}
// Break if block is not the first and it corresponds to the main
// function body entry. This is to prevent creating multiple
// large regions containing the function body (starting at various
// DV funclets).
if (prevId != kInvalidTransID) {
const Func* func = profData->transFunc(tid);
Offset bcOffset = profData->transStartBcOff(tid);
if (func->base() == bcOffset) {
FTRACE(2, "selectHotTrace: breaking region because reached the main "
"function body entry at Translation {} (BC offset {})\n",
tid, bcOffset);
break;
}
}
if (prevId != kInvalidTransID) {
auto sk = profData->transSrcKey(tid);
if (profData->optimized(sk)) {
FTRACE(2, "selectHotTrace: breaking region because next sk already "
"optimized, for Translation {}\n", tid);
break;
}
}
// Break trace if translation tid cannot follow the execution of
// the entire translation prevId. This can only happen if the
// execution of prevId takes a side exit that leads to the
// execution of tid.
if (prevId != kInvalidTransID) {
Op* lastInstr = profData->transLastInstr(prevId);
const Unit* unit = profData->transFunc(prevId)->unit();
OffsetSet succOffs = findSuccOffsets(lastInstr, unit);
if (!succOffs.count(profData->transSrcKey(tid).offset())) {
if (HPHP::Trace::moduleEnabled(HPHP::Trace::pgo, 2)) {
FTRACE(2, "selectHotTrace: WARNING: Breaking region @: {}\n",
show(*region));
FTRACE(2, "selectHotTrace: next translation selected: tid = {}\n{}\n",
tid, show(*blockRegion));
FTRACE(2, "\nsuccOffs = {}\n", folly::join(", ", succOffs));
}
break;
}
}
if (region->blocks.size() > 0) {
auto& newBlock = blockRegion->blocks.front();
auto newBlockId = newBlock->id();
auto predBlockId = region->blocks.back().get()->id();
if (!RuntimeOption::EvalHHIRBytecodeControlFlow) {
region->addArc(predBlockId, newBlockId);
} else {
// With bytecode control-flow, we add all forward arcs in the TransCFG
// that are induced by the blocks in the region, as a simple way
// to expose control-flow for now.
// This can go away once Task #4075822 is done.
auto newBlockSrcKey = blockRegion->blocks.front().get()->start();
if (succSKSet[predBlockId].count(newBlockSrcKey)) break;
region->addArc(predBlockId, newBlockId);
succSKSet[predBlockId].insert(newBlockSrcKey);
assert(hasTransId(newBlockId));
auto newTransId = getTransId(newBlockId);
for (auto iOther = 0; iOther < region->blocks.size(); iOther++) {
auto other = region->blocks[iOther];
auto otherBlockId = other.get()->id();
if (!hasTransId(otherBlockId)) continue;
auto otherTransId = getTransId(otherBlockId);
auto otherBlockSrcKey = other.get()->start();
if (cfg.hasArc(otherTransId, newTransId) &&
!other.get()->inlinedCallee() &&
// Task #4157613 will allow the following check to go away
!succSKSet[otherBlockId].count(newBlockSrcKey) &&
preCondsAreSatisfied(newBlock, blockPostConds[otherBlockId])) {
region->addArc(otherBlockId, newBlockId);
succSKSet[otherBlockId].insert(newBlockSrcKey);
}
// When Eval.JitLoops is set, insert back-edges in the
// region if they exist in the TransCFG.
if (RuntimeOption::EvalJitLoops &&
cfg.hasArc(newTransId, otherTransId) &&
// Task #4157613 will allow the following check to go away
!succSKSet[newBlockId].count(otherBlockSrcKey)) {
region->addArc(newBlockId, otherBlockId);
succSKSet[newBlockId].insert(otherBlockSrcKey);
}
}
}
}
region->blocks.insert(region->blocks.end(), blockRegion->blocks.begin(),
blockRegion->blocks.end());
region->arcs.insert(region->arcs.end(), blockRegion->arcs.begin(),
blockRegion->arcs.end());
if (cfg.outArcs(tid).size() > 1) {
region->setSideExitingBlock(blockRegion->blocks.front()->id());
}
selectedSet.insert(tid);
if (selectedVec) selectedVec->push_back(tid);
Op lastOp = *(profData->transLastInstr(tid));
if (breaksRegion(lastOp)) {
FTRACE(2, "selectHotTrace: breaking region because of last instruction "
"in Translation {}: {}\n", tid, opcodeToName(lastOp));
break;
}
auto outArcs = cfg.outArcs(tid);
if (outArcs.size() == 0) {
FTRACE(2, "selectHotTrace: breaking region because there's no successor "
"for Translation {}\n", tid);
break;
}
auto lastNewBlock = blockRegion->blocks.back();
discardPoppedTypes(accumPostConds,
blockRegion->blocks[0]->initialSpOffset());
mergePostConds(accumPostConds, lastNewBlock->postConds());
blockPostConds[lastNewBlock->id()] = accumPostConds;
TransCFG::ArcPtrVec possibleOutArcs;
for (auto arc : outArcs) {
RegionDesc::BlockPtr possibleNext =
profData->transRegion(arc->dst())->blocks[0];
if (preCondsAreSatisfied(possibleNext, accumPostConds)) {
possibleOutArcs.emplace_back(arc);
}
}
if (possibleOutArcs.size() == 0) {
FTRACE(2, "selectHotTrace: breaking region because postcondition check "
"pruned all successors of Translation {}\n", tid);
break;
}
auto maxWeight = std::numeric_limits<int64_t>::min();
TransCFG::Arc* maxArc = nullptr;
for (auto arc : possibleOutArcs) {
if (arc->weight() >= maxWeight) {
maxWeight = arc->weight();
maxArc = arc;
}
}
assert(maxArc != nullptr);
prevId = tid;
tid = maxArc->dst();
}
return region;
}
RegionDescPtr selectHotTrace(TransID triggerId,
const ProfData* profData,
TransCFG& cfg,
TransIDSet& selectedSet,
TransIDVec* selectedVec) {
auto region = std::make_shared<RegionDesc>();
TransID tid = triggerId;
TransID prevId = kInvalidTransID;
selectedSet.clear();
if (selectedVec) selectedVec->clear();
PostConditions accumPostConds;
// Maps from BlockIds to accumulated post conditions for that block.
// Used to determine if we can add branch-over edges by checking the
// pre-conditions of the successor block.
hphp_hash_map<RegionDesc::BlockId, PostConditions> blockPostConds;
uint32_t numBCInstrs = 0;
while (!selectedSet.count(tid)) {
RegionDescPtr blockRegion = profData->transRegion(tid);
if (blockRegion == nullptr) break;
// Break if region would be larger than the specified limit.
auto newInstrSize = numBCInstrs + blockRegion->instrSize();
if (newInstrSize > RuntimeOption::EvalJitMaxRegionInstrs) {
FTRACE(2, "selectHotTrace: breaking region at Translation {} because "
"size ({}) would exceed of maximum translation limit\n",
tid, newInstrSize);
break;
}
// If the debugger is attached, only allow single-block regions.
if (prevId != kInvalidTransID && isDebuggerAttachedProcess()) {
FTRACE(2, "selectHotTrace: breaking region at Translation {} "
"because of debugger is attached\n", tid);
break;
}
// Break if block is not the first and it corresponds to the main
// function body entry. This is to prevent creating multiple
// large regions containing the function body (starting at various
// DV funclets).
if (prevId != kInvalidTransID) {
const Func* func = profData->transFunc(tid);
Offset bcOffset = profData->transStartBcOff(tid);
if (func->base() == bcOffset) {
FTRACE(2, "selectHotTrace: breaking region because reached the main "
"function body entry at Translation {} (BC offset {})\n",
tid, bcOffset);
break;
}
}
if (prevId != kInvalidTransID) {
auto sk = profData->transSrcKey(tid);
if (profData->optimized(sk)) {
FTRACE(2, "selectHotTrace: breaking region because next sk already "
"optimized, for Translation {}\n", tid);
break;
}
}
bool hasPredBlock = !region->empty();
RegionDesc::BlockId predBlockId = (hasPredBlock ?
region->blocks().back().get()->id() : 0);
auto const& newFirstBlock = blockRegion->entry();
auto newFirstBlockId = newFirstBlock->id();
auto newLastBlockId = blockRegion->blocks().back()->id();
// Add blockRegion's blocks and arcs to region.
region->append(*blockRegion);
numBCInstrs += blockRegion->instrSize();
if (hasPredBlock) {
region->addArc(predBlockId, newFirstBlockId);
}
// When Eval.JitLoops is set, insert back-edges in the region if
// they exist in the TransCFG.
if (RuntimeOption::EvalJitLoops) {
assertx(hasTransId(newFirstBlockId));
auto newTransId = getTransId(newFirstBlockId);
// Don't add the arc if the last opcode in the source block ends
// the region.
if (!breaksRegion(*profData->transLastInstr(newTransId))) {
auto& blocks = region->blocks();
for (auto iOther = 0; iOther < blocks.size(); iOther++) {
auto other = blocks[iOther];
auto otherFirstBlockId = other.get()->id();
if (!hasTransId(otherFirstBlockId)) continue;
auto otherTransId = getTransId(otherFirstBlockId);
if (cfg.hasArc(newTransId, otherTransId)) {
region->addArc(newLastBlockId, otherFirstBlockId);
}
}
}
}
if (cfg.outArcs(tid).size() > 1) {
region->setSideExitingBlock(blockRegion->entry()->id());
}
selectedSet.insert(tid);
if (selectedVec) selectedVec->push_back(tid);
Op lastOp = *(profData->transLastInstr(tid));
if (breaksRegion(lastOp)) {
FTRACE(2, "selectHotTrace: breaking region because of last instruction "
"in Translation {}: {}\n", tid, opcodeToName(lastOp));
break;
}
auto outArcs = cfg.outArcs(tid);
if (outArcs.size() == 0) {
FTRACE(2, "selectHotTrace: breaking region because there's no successor "
"for Translation {}\n", tid);
break;
}
auto newLastBlock = blockRegion->blocks().back();
discardPoppedTypes(accumPostConds,
blockRegion->entry()->initialSpOffset());
mergePostConds(accumPostConds, newLastBlock->postConds());
blockPostConds[newLastBlock->id()] = accumPostConds;
TransCFG::ArcPtrVec possibleOutArcs;
for (auto arc : outArcs) {
RegionDesc::BlockPtr possibleNext =
profData->transRegion(arc->dst())->entry();
if (preCondsAreSatisfied(possibleNext, accumPostConds)) {
possibleOutArcs.emplace_back(arc);
}
}
if (possibleOutArcs.size() == 0) {
FTRACE(2, "selectHotTrace: breaking region because postcondition check "
"pruned all successors of Translation {}\n", tid);
break;
}
auto maxWeight = std::numeric_limits<int64_t>::min();
TransCFG::Arc* maxArc = nullptr;
for (auto arc : possibleOutArcs) {
if (arc->weight() >= maxWeight) {
maxWeight = arc->weight();
maxArc = arc;
}
}
assertx(maxArc != nullptr);
prevId = tid;
tid = maxArc->dst();
}
FTRACE(3, "selectHotTrace: before chainRetransBlocks:\n{}\n", show(*region));
region->chainRetransBlocks();
FTRACE(3, "selectHotTrace: after chainRetransBlocks:\n{}\n", show(*region));
return region;
}
/*
* Get location metadata for the inputs of `ni'.
*/
InputInfoVec getInputs(NormalizedInstruction& ni, FPInvOffset bcSPOff) {
InputInfoVec inputs;
if (isAlwaysNop(ni.op())) return inputs;
always_assert_flog(
instrInfo.count(ni.op()),
"Invalid opcode in getInputsImpl: {}\n",
opcodeToName(ni.op())
);
UNUSED auto const sk = ni.source;
auto const& info = instrInfo[ni.op()];
auto const flags = info.in;
auto stackOff = bcSPOff;
if (flags & FStack) {
stackOff -= ni.imm[0].u_IVA; // arguments consumed
stackOff -= kNumActRecCells; // ActRec is torn down as well
}
if (flags & FuncdRef) inputs.needsRefCheck = true;
if (flags & IgnoreInnerType) ni.ignoreInnerType = true;
if (flags & Stack1) {
SKTRACE(1, sk, "getInputs: Stack1 %d\n", stackOff.offset);
inputs.emplace_back(Location::Stack { stackOff-- });
if (flags & DontGuardStack1) inputs.back().dontGuard = true;
if (flags & Stack2) {
SKTRACE(1, sk, "getInputs: Stack2 %d\n", stackOff.offset);
inputs.emplace_back(Location::Stack { stackOff-- });
if (flags & Stack3) {
SKTRACE(1, sk, "getInputs: Stack3 %d\n", stackOff.offset);
inputs.emplace_back(Location::Stack { stackOff-- });
}
}
}
if (flags & StackI) {
inputs.emplace_back(Location::Stack {
BCSPRelOffset{ni.imm[0].u_IVA}.to<FPInvOffset>(bcSPOff)
});
}
if (flags & StackN) {
int numArgs = (ni.op() == Op::NewPackedArray ||
ni.op() == Op::NewVecArray ||
ni.op() == Op::ConcatN)
? ni.imm[0].u_IVA
: ni.immVec.numStackValues();
SKTRACE(1, sk, "getInputs: StackN %d %d\n", stackOff.offset, numArgs);
for (int i = 0; i < numArgs; i++) {
inputs.emplace_back(Location::Stack { stackOff-- });
inputs.back().dontGuard = true;
inputs.back().dontBreak = true;
}
}
if (flags & BStackN) {
int numArgs = ni.imm[0].u_IVA;
SKTRACE(1, sk, "getInputs: BStackN %d %d\n", stackOff.offset, numArgs);
for (int i = 0; i < numArgs; i++) {
inputs.emplace_back(Location::Stack { stackOff-- });
}
}
if (flags & Local) {
// (Almost) all instructions that take a Local have its index at their
// first immediate.
auto const loc = ni.imm[localImmIdx(ni.op())].u_IVA;
SKTRACE(1, sk, "getInputs: local %d\n", loc);
inputs.emplace_back(Location::Local { uint32_t(loc) });
}
if (flags & AllLocals) ni.ignoreInnerType = true;
if (flags & MKey) {
auto mk = ni.imm[memberKeyImmIdx(ni.op())].u_KA;
switch (mk.mcode) {
case MEL:
case MPL:
inputs.emplace_back(Location::Local { uint32_t(mk.iva) });
break;
case MEC:
case MPC:
inputs.emplace_back(Location::Stack {
BCSPRelOffset{int32_t(mk.iva)}.to<FPInvOffset>(bcSPOff)
});
break;
case MW:
case MEI:
case MET:
case MPT:
case MQT:
// The inputs vector is only used for deciding when to break the
// tracelet, which can never happen for these cases.
break;
}
}
SKTRACE(1, sk, "stack args: virtual sfo now %d\n", stackOff.offset);
TRACE(1, "%s\n", Trace::prettyNode("Inputs", inputs).c_str());
if ((flags & DontGuardAny) || dontGuardAnyInputs(ni.op())) {
for (auto& info : inputs) info.dontGuard = true;
}
return inputs;
}
RegionDescPtr selectHotTrace(TransID triggerId,
const ProfData* profData,
TransCFG& cfg,
TransIDSet& selectedSet,
TransIDVec* selectedVec) {
auto region = std::make_shared<RegionDesc>();
TransID tid = triggerId;
TransID prevId = InvalidID;
selectedSet.clear();
if (selectedVec) selectedVec->clear();
PostConditions accumPostConds;
while (!selectedSet.count(tid)) {
RegionDescPtr blockRegion = profData->transRegion(tid);
if (blockRegion == nullptr) break;
// If the debugger is attached, only allow single-block regions.
if (prevId != InvalidID && isDebuggerAttachedProcess()) {
FTRACE(2, "selectHotTrace: breaking region at Translation {} "
"because of debugger is attached\n", tid);
break;
}
// Break if block is not the first and requires reffiness checks.
// Task #2589970: fix translateRegion to support mid-region reffiness checks
if (prevId != InvalidID) {
auto nRefDeps = blockRegion->blocks[0]->reffinessPreds().size();
if (nRefDeps > 0) {
FTRACE(2, "selectHotTrace: breaking region because of refDeps ({}) at "
"Translation {}\n", nRefDeps, tid);
break;
}
}
// Break if block is not the first and it corresponds to the main
// function body entry. This is to prevent creating multiple
// large regions containing the function body (starting at various
// DV funclets).
if (prevId != InvalidID) {
const Func* func = profData->transFunc(tid);
Offset bcOffset = profData->transStartBcOff(tid);
if (func->base() == bcOffset) {
FTRACE(2, "selectHotTrace: breaking region because reached the main "
"function body entry at Translation {} (BC offset {})\n",
tid, bcOffset);
break;
}
}
if (prevId != InvalidID) {
auto sk = profData->transSrcKey(tid);
if (profData->optimized(sk)) {
FTRACE(2, "selectHotTrace: breaking region because next sk already "
"optimized, for Translation {}\n", tid);
break;
}
}
// Break trace if translation tid cannot follow the execution of
// the entire translation prevId. This can only happen if the
// execution of prevId takes a side exit that leads to the
// execution of tid.
if (prevId != InvalidID) {
Op* lastInstr = profData->transLastInstr(prevId);
const Unit* unit = profData->transFunc(prevId)->unit();
OffsetSet succOffs = findSuccOffsets(lastInstr, unit);
if (!succOffs.count(profData->transSrcKey(tid).offset())) {
if (HPHP::Trace::moduleEnabled(HPHP::Trace::pgo, 2)) {
FTRACE(2, "selectHotTrace: WARNING: Breaking region @: {}\n",
JIT::show(*region));
FTRACE(2, "selectHotTrace: next translation selected: tid = {}\n{}\n",
tid, JIT::show(*blockRegion));
FTRACE(2, "\nsuccOffs = {}\n", folly::join(", ", succOffs));
}
break;
}
}
region->blocks.insert(region->blocks.end(), blockRegion->blocks.begin(),
blockRegion->blocks.end());
selectedSet.insert(tid);
if (selectedVec) selectedVec->push_back(tid);
Op lastOp = *(profData->transLastInstr(tid));
if (breaksRegion(lastOp)) {
FTRACE(2, "selectHotTrace: breaking region because of last instruction "
"in Translation {}: {}\n", tid, opcodeToName(lastOp));
break;
}
auto outArcs = cfg.outArcs(tid);
if (outArcs.size() == 0) {
FTRACE(2, "selectHotTrace: breaking region because there's no successor "
"for Translation {}\n", tid);
break;
}
auto lastNewBlock = blockRegion->blocks.back();
discardPoppedTypes(accumPostConds,
blockRegion->blocks[0]->initialSpOffset());
mergePostConds(accumPostConds, lastNewBlock->postConds());
TransCFG::ArcPtrVec possibleOutArcs;
for (auto arc : outArcs) {
RegionDesc::BlockPtr possibleNext =
profData->transRegion(arc->dst())->blocks[0];
if (preCondsAreSatisfied(possibleNext, accumPostConds)) {
possibleOutArcs.emplace_back(arc);
}
}
if (possibleOutArcs.size() == 0) {
FTRACE(2, "selectHotTrace: breaking region because postcondition check "
"pruned all successors of Translation {}\n", tid);
break;
}
auto maxWeight = std::numeric_limits<int64_t>::min();
TransCFG::Arc* maxArc = nullptr;
for (auto arc : possibleOutArcs) {
if (arc->weight() >= maxWeight) {
maxWeight = arc->weight();
maxArc = arc;
}
}
assert(maxArc != nullptr);
prevId = tid;
tid = maxArc->dst();
}
return region;
}
RegionDescPtr selectHotTrace(TransID triggerId,
const ProfData* profData,
TransCFG& cfg,
TransIDSet& selectedSet,
TransIDVec* selectedVec) {
auto region = std::make_shared<RegionDesc>();
TransID tid = triggerId;
TransID prevId = kInvalidTransID;
selectedSet.clear();
if (selectedVec) selectedVec->clear();
PostConditions accumPostConds;
// Maps BlockIds to the set of BC offsets for its successor blocks.
// Used to prevent multiple successors with the same SrcKey for now.
// This can go away once task #4157613 is done.
hphp_hash_map<RegionDesc::BlockId, SrcKeySet> succSKSet;
// Maps from BlockIds to accumulated post conditions for that block.
// Used to determine if we can add branch-over edges by checking the
// pre-conditions of the successor block.
hphp_hash_map<RegionDesc::BlockId, PostConditions> blockPostConds;
while (!selectedSet.count(tid)) {
RegionDescPtr blockRegion = profData->transRegion(tid);
if (blockRegion == nullptr) break;
// If the debugger is attached, only allow single-block regions.
if (prevId != kInvalidTransID && isDebuggerAttachedProcess()) {
FTRACE(2, "selectHotTrace: breaking region at Translation {} "
"because of debugger is attached\n", tid);
break;
}
// Break if block is not the first and it corresponds to the main
// function body entry. This is to prevent creating multiple
// large regions containing the function body (starting at various
// DV funclets).
if (prevId != kInvalidTransID) {
const Func* func = profData->transFunc(tid);
Offset bcOffset = profData->transStartBcOff(tid);
if (func->base() == bcOffset) {
FTRACE(2, "selectHotTrace: breaking region because reached the main "
"function body entry at Translation {} (BC offset {})\n",
tid, bcOffset);
break;
}
}
if (prevId != kInvalidTransID) {
auto sk = profData->transSrcKey(tid);
if (profData->optimized(sk)) {
FTRACE(2, "selectHotTrace: breaking region because next sk already "
"optimized, for Translation {}\n", tid);
break;
}
}
bool hasPredBlock = !region->empty();
RegionDesc::BlockId predBlockId = (hasPredBlock ?
region->blocks().back().get()->id() : 0);
auto const& newFirstBlock = blockRegion->entry();
auto newFirstBlockId = newFirstBlock->id();
auto newFirstBlockSk = newFirstBlock->start();
auto newLastBlockId = blockRegion->blocks().back()->id();
// Make sure we don't end up with multiple successors for the same
// SrcKey. Task #4157613 will allow the following check to go away.
// This needs to be done before we insert blockRegion into region,
// to avoid creating unreachable blocks.
if (RuntimeOption::EvalHHIRBytecodeControlFlow && hasPredBlock &&
succSKSet[predBlockId].count(newFirstBlockSk)) {
break;
}
// Add blockRegion's blocks and arcs to region.
region->append(*blockRegion);
if (hasPredBlock) {
if (RuntimeOption::EvalHHIRBytecodeControlFlow) {
// This is checked above.
assert(succSKSet[predBlockId].count(newFirstBlockSk) == 0);
succSKSet[predBlockId].insert(newFirstBlockSk);
}
region->addArc(predBlockId, newFirstBlockId);
}
// With bytecode control-flow, we add all forward arcs in the TransCFG
// that are induced by the blocks in the region, as a simple way
// to expose control-flow for now.
// This can go away once Task #4075822 is done.
if (RuntimeOption::EvalHHIRBytecodeControlFlow) {
assert(hasTransId(newFirstBlockId));
auto newTransId = getTransId(newFirstBlockId);
auto& blocks = region->blocks();
for (auto iOther = 0; iOther < blocks.size(); iOther++) {
auto other = blocks[iOther];
auto otherFirstBlockId = other.get()->id();
if (!hasTransId(otherFirstBlockId)) continue;
auto otherTransId = getTransId(otherFirstBlockId);
auto otherFirstBlockSk = other.get()->start();
auto otherRegion = profData->transRegion(otherTransId);
auto otherLastBlockId = otherRegion->blocks().back()->id();
// When loops are off, stop once we hit the newTransId we just inserted.
if (!RuntimeOption::EvalJitLoops && otherTransId == newTransId) break;
if (cfg.hasArc(otherTransId, newTransId) &&
// Task #4157613 will allow the following check to go away
!succSKSet[otherLastBlockId].count(newFirstBlockSk) &&
preCondsAreSatisfied(newFirstBlock,
blockPostConds[otherLastBlockId])) {
region->addArc(otherLastBlockId, newFirstBlockId);
succSKSet[otherLastBlockId].insert(newFirstBlockSk);
}
// When Eval.JitLoops is set, insert back-edges in the
// region if they exist in the TransCFG.
if (RuntimeOption::EvalJitLoops &&
cfg.hasArc(newTransId, otherTransId) &&
// Task #4157613 will allow the following check to go away
!succSKSet[newLastBlockId].count(otherFirstBlockSk)) {
region->addArc(newLastBlockId, otherFirstBlockId);
succSKSet[newLastBlockId].insert(otherFirstBlockSk);
}
}
}
if (cfg.outArcs(tid).size() > 1) {
region->setSideExitingBlock(blockRegion->entry()->id());
}
selectedSet.insert(tid);
if (selectedVec) selectedVec->push_back(tid);
Op lastOp = *(profData->transLastInstr(tid));
if (breaksRegion(lastOp)) {
FTRACE(2, "selectHotTrace: breaking region because of last instruction "
"in Translation {}: {}\n", tid, opcodeToName(lastOp));
break;
}
auto outArcs = cfg.outArcs(tid);
if (outArcs.size() == 0) {
FTRACE(2, "selectHotTrace: breaking region because there's no successor "
"for Translation {}\n", tid);
break;
}
auto newLastBlock = blockRegion->blocks().back();
discardPoppedTypes(accumPostConds,
blockRegion->entry()->initialSpOffset());
mergePostConds(accumPostConds, newLastBlock->postConds());
blockPostConds[newLastBlock->id()] = accumPostConds;
TransCFG::ArcPtrVec possibleOutArcs;
for (auto arc : outArcs) {
RegionDesc::BlockPtr possibleNext =
profData->transRegion(arc->dst())->entry();
if (preCondsAreSatisfied(possibleNext, accumPostConds)) {
possibleOutArcs.emplace_back(arc);
}
}
if (possibleOutArcs.size() == 0) {
FTRACE(2, "selectHotTrace: breaking region because postcondition check "
"pruned all successors of Translation {}\n", tid);
break;
}
auto maxWeight = std::numeric_limits<int64_t>::min();
TransCFG::Arc* maxArc = nullptr;
for (auto arc : possibleOutArcs) {
if (arc->weight() >= maxWeight) {
maxWeight = arc->weight();
maxArc = arc;
}
}
assert(maxArc != nullptr);
prevId = tid;
tid = maxArc->dst();
}
return region;
}
void SialOpsParallel::log_statement(opcode_t type, int line){
SIP_LOG(
std::cout<< "W " << sip_mpi_attr_.global_rank()
<< " : Line "<<line << ", type: " << opcodeToName(type)<<std::endl);
}
/*
* getInputs --
* Returns locations for this instruction's inputs.
*/
InputInfoVec getInputs(NormalizedInstruction& ni) {
InputInfoVec inputs;
auto UNUSED sk = ni.source;
if (isAlwaysNop(ni.op())) return inputs;
assertx(inputs.empty());
always_assert_flog(
instrInfo.count(ni.op()),
"Invalid opcode in getInputsImpl: {}\n",
opcodeToName(ni.op())
);
const InstrInfo& info = instrInfo[ni.op()];
Operands input = info.in;
BCSPOffset spOff{0};
if (input & FuncdRef) {
inputs.needsRefCheck = true;
}
if (input & Iter) {
inputs.emplace_back(Location(Location::Iter, ni.imm[0].u_IVA));
}
if (input & FStack) {
spOff += ni.imm[0].u_IVA; // arguments consumed
spOff += kNumActRecCells; // ActRec is torn down as well
}
if (input & IgnoreInnerType) ni.ignoreInnerType = true;
if (input & Stack1) {
SKTRACE(1, sk, "getInputs: stack1 %d\n", spOff.offset);
inputs.emplace_back(Location(spOff++));
if (input & DontGuardStack1) inputs.back().dontGuard = true;
if (input & Stack2) {
SKTRACE(1, sk, "getInputs: stack2 %d\n", spOff.offset);
inputs.emplace_back(Location(spOff++));
if (input & Stack3) {
SKTRACE(1, sk, "getInputs: stack3 %d\n", spOff.offset);
inputs.emplace_back(Location(spOff++));
}
}
}
if (input & StackI) {
inputs.emplace_back(Location(BCSPOffset{ni.imm[0].u_IVA}));
}
if (input & StackN) {
int numArgs = (ni.op() == Op::NewPackedArray ||
ni.op() == Op::ConcatN)
? ni.imm[0].u_IVA
: ni.immVec.numStackValues();
SKTRACE(1, sk, "getInputs: stackN %d %d\n", spOff.offset, numArgs);
for (int i = 0; i < numArgs; i++) {
inputs.emplace_back(Location(spOff++));
inputs.back().dontGuard = true;
inputs.back().dontBreak = true;
}
}
if (input & BStackN) {
int numArgs = ni.imm[0].u_IVA;
SKTRACE(1, sk, "getInputs: BStackN %d %d\n", spOff.offset, numArgs);
for (int i = 0; i < numArgs; i++) {
inputs.emplace_back(Location(spOff++));
}
}
if (input & Local) {
// (Almost) all instructions that take a Local have its index at
// their first immediate.
auto const loc = ni.imm[localImmIdx(ni.op())].u_IVA;
SKTRACE(1, sk, "getInputs: local %d\n", loc);
inputs.emplace_back(Location(Location::Local, loc));
}
if (input & MKey) {
auto mk = ni.imm[memberKeyImmIdx(ni.op())].u_KA;
switch (mk.mcode) {
case MEL: case MPL:
inputs.emplace_back(Location(Location::Local, mk.iva));
break;
case MEC: case MPC:
inputs.emplace_back(Location(BCSPOffset{int32_t(mk.iva)}));
break;
case MW: case MEI: case MET: case MPT: case MQT:
// The inputs vector is only used for deciding when to break the
// tracelet, which can never happen for these cases.
break;
}
}
if (input & AllLocals) {
ni.ignoreInnerType = true;
}
SKTRACE(1, sk, "stack args: virtual sfo now %d\n", spOff.offset);
TRACE(1, "%s\n", Trace::prettyNode("Inputs", inputs).c_str());
if (inputs.size() &&
((input & DontGuardAny) || dontGuardAnyInputs(ni.op()))) {
for (int i = inputs.size(); i--; ) {
inputs[i].dontGuard = true;
}
}
if (input & This) {
inputs.emplace_back(Location(Location::This));
}
return inputs;
}
RegionDescPtr selectHotTrace(TransID triggerId,
const ProfData* profData,
TransCFG& cfg,
TransIDSet& selectedSet) {
JIT::RegionDescPtr region = smart::make_unique<JIT::RegionDesc>();
TransID tid = triggerId;
TransID prevId = InvalidID;
selectedSet.clear();
while (!setContains(selectedSet, tid)) {
RegionDesc::BlockPtr block = profData->transBlock(tid);
if (block == nullptr) break;
// If the debugger is attached, only allow single-block regions.
if (prevId != InvalidID && isDebuggerAttachedProcess()) {
FTRACE(5, "selectHotRegion: breaking region at Translation {} "
"because of debugger is attached\n", tid);
break;
}
// Break if block is not the first and requires reffiness checks.
// Task #2589970: fix translateRegion to support mid-region reffiness checks
if (prevId != InvalidID) {
auto nRefDeps = block->reffinessPreds().size();
if (nRefDeps > 0) {
FTRACE(5, "selectHotRegion: breaking region because of refDeps ({}) at "
"Translation {}\n", nRefDeps, tid);
break;
}
}
// Break trace if translation tid cannot follow the execution of
// the entire translation prevTd. This can only happen if the
// execution of prevId takes a side exit that leads to the
// execution of tid.
if (prevId != InvalidID) {
Op* lastInstr = profData->transLastInstr(prevId);
const Unit* unit = profData->transFunc(prevId)->unit();
OffsetSet succOffs = findSuccOffsets(lastInstr, unit);
if (!setContains(succOffs, profData->transSrcKey(tid).offset())) {
if (HPHP::Trace::moduleEnabled(HPHP::Trace::pgo, 5)) {
FTRACE(5, "selectHotTrace: WARNING: Breaking region @: {}\n",
JIT::show(*region));
FTRACE(5, "selectHotTrace: next translation selected: tid = {}\n{}\n",
tid, JIT::show(*block));
std::string succStr("succOffs = ");
for (auto succ : succOffs) {
succStr += lexical_cast<std::string>(succ);
}
FTRACE(5, "\n{}\n", succStr);
}
break;
}
}
region->blocks.emplace_back(block);
selectedSet.insert(tid);
Op lastOp = *(profData->transLastInstr(tid));
if (breaksRegion(lastOp)) {
FTRACE(5, "selectHotTrace: breaking region because of last instruction "
"in Translation {}: {}\n", tid, opcodeToName(lastOp));
break;
}
auto outArcs = cfg.outArcs(tid);
if (outArcs.size() == 0) {
FTRACE(5, "selectHotTrace: breaking region because there's no successor "
"for Translation {}\n", tid);
break;
}
auto maxWeight = std::numeric_limits<int64_t>::min();
TransCFG::Arc* maxArc = nullptr;
for (auto arc : outArcs) {
if (arc->weight() >= maxWeight) {
maxWeight = arc->weight();
maxArc = arc;
}
}
assert(maxArc != nullptr);
prevId = tid;
tid = maxArc->dst();
}
return region;
}
bool InliningDecider::shouldInline(const Func* callee,
const RegionDesc& region) {
auto sk = region.empty() ? SrcKey() : region.start();
assertx(callee);
assertx(sk.func() == callee);
int cost = 0;
// Tracing return lambdas.
auto refuse = [&] (const char* why) {
return traceRefusal(m_topFunc, callee, why);
};
auto accept = [&, this] (const char* kind) {
FTRACE(1, "InliningDecider: inlining {}() <- {}()\t<reason: {}>\n",
m_topFunc->fullName()->data(), callee->fullName()->data(), kind);
// Update our context.
m_costStack.push_back(cost);
m_cost += cost;
m_callDepth += 1;
m_stackDepth += callee->maxStackCells();
return true;
};
// Check inlining depths.
if (m_callDepth + 1 >= RuntimeOption::EvalHHIRInliningMaxDepth) {
return refuse("inlining call depth limit exceeded");
}
if (m_stackDepth + callee->maxStackCells() >= kStackCheckLeafPadding) {
return refuse("inlining stack depth limit exceeded");
}
// Even if the func contains NativeImpl we may have broken the trace before
// we hit it.
auto containsNativeImpl = [&] {
for (auto block : region.blocks()) {
if (!block->empty() && block->last().op() == OpNativeImpl) return true;
}
return false;
};
// Try to inline CPP builtin functions.
// The NativeImpl opcode may appear later in the function because of Asserts
// generated in hhbbc
if (callee->isCPPBuiltin() && containsNativeImpl()) {
if (isInlinableCPPBuiltin(callee)) {
return accept("inlinable CPP builtin");
}
return refuse("non-inlinable CPP builtin");
}
// If the function may use a VarEnv (which is stored in the ActRec) or may be
// variadic, we restrict inlined callees to certain whitelisted instructions
// which we know won't actually require these features.
const bool needsCheckVVSafe = callee->attrs() & AttrMayUseVV;
// We measure the cost of inlining each callstack and stop when it exceeds a
// certain threshold. (Note that we do not measure the total cost of all the
// inlined calls for a given caller---just the cost of each nested stack.)
const int maxCost = RuntimeOption::EvalHHIRInliningMaxCost - m_cost;
// We only inline callee regions that have exactly one return.
//
// NOTE: Currently, the tracelet selector uses the first Ret in the child's
// region to determine when to stop inlining. However, the safety of this
// behavior should not be considered guaranteed by InliningDecider; the
// "right" way to decide when inlining ends is to inline all of `region'.
int numRets = 0;
// Iterate through the region, checking its suitability for inlining.
for (auto const& block : region.blocks()) {
sk = block->start();
for (auto i = 0, n = block->length(); i < n; ++i, sk.advance()) {
auto op = sk.op();
// We don't allow inlined functions in the region. The client is
// expected to disable inlining for the region it gives us to peek.
if (sk.func() != callee) {
return refuse("got region with inlined calls");
}
// Restrict to VV-safe opcodes if necessary.
if (needsCheckVVSafe && !isInliningVVSafe(op)) {
return refuse(folly::format("{} may use dynamic environment",
opcodeToName(op)).str().c_str());
}
// Count the returns.
if (isRet(op) || op == Op::NativeImpl) {
if (++numRets > 1) {
return refuse("region has too many returns");
}
continue;
}
// We can't inline FCallArray. XXX: Why?
if (op == Op::FCallArray) {
return refuse("can't inline FCallArray");
}
// Assert opcodes don't contribute to the inlining cost.
if (op == Op::AssertRATL || op == Op::AssertRATStk) continue;
cost += 1;
// Add the size of immediate vectors to the cost.
auto const pc = reinterpret_cast<const Op*>(sk.pc());
if (hasMVector(op)) {
cost += getMVector(pc).size();
} else if (hasImmVector(op)) {
cost += getImmVector(pc).size();
}
// Refuse if the cost exceeds our thresholds.
if (cost > maxCost) {
return refuse("too expensive");
}
}
}
if (numRets != 1) {
return refuse("region has no returns");
}
return accept("small region with single return");
}
