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");
}
/*
* Checks if the given region is well-formed, which entails the
* following properties:
*
* 1) The region has at least one block.
*
* 2) Each block in the region has a different id.
*
* 3) All arcs involve blocks within the region.
*
* 4) For each arc, the bytecode offset of the dst block must
* possibly follow the execution of the src block.
*
* 5) Each block contains at most one successor corresponding to a
* given SrcKey.
*
* 6) The region doesn't contain any loops, unless JitLoops is
* enabled.
*
* 7) All blocks are reachable from the entry block.
*
* 8) For each block, there must be a path from the entry to it that
* includes only earlier blocks in the region.
*
* 9) The region is topologically sorted unless loops are enabled.
*
* 10) The block-retranslation chains cannot have cycles.
*
*/
bool check(const RegionDesc& region, std::string& error) {
auto bad = [&](const std::string& errorMsg) {
error = errorMsg;
return false;
};
// 1) The region has at least one block.
if (region.empty()) return bad("empty region");
RegionDesc::BlockIdSet blockSet;
for (auto b : region.blocks()) {
auto bid = b->id();
// 2) Each block in the region has a different id.
if (blockSet.count(bid)) {
return bad(folly::sformat("many blocks with id {}", bid));
}
blockSet.insert(bid);
}
for (auto b : region.blocks()) {
auto bid = b->id();
SrcKey lastSk = region.block(bid)->last();
OffsetSet validSuccOffsets = lastSk.succOffsets();
OffsetSet succOffsets;
for (auto succ : region.succs(bid)) {
SrcKey succSk = region.block(succ)->start();
Offset succOffset = succSk.offset();
// 3) All arcs involve blocks within the region.
if (blockSet.count(succ) == 0) {
return bad(folly::sformat("arc with dst not in the region: {} -> {}",
bid, succ));
}
// Checks 4) and 5) below don't make sense for arcs corresponding
// to inlined calls and returns, so skip them in such cases.
// This won't be possible once task #4076399 is done.
if (lastSk.func() != succSk.func()) continue;
// 4) For each arc, the bytecode offset of the dst block must
// possibly follow the execution of the src block.
if (validSuccOffsets.count(succOffset) == 0) {
return bad(folly::sformat("arc with impossible control flow: {} -> {}",
bid, succ));
}
// 5) Each block contains at most one successor corresponding to a
// given SrcKey.
if (succOffsets.count(succOffset) > 0) {
return bad(folly::sformat("block {} has multiple successors with SK {}",
bid, show(succSk)));
}
succOffsets.insert(succOffset);
}
for (auto pred : region.preds(bid)) {
if (blockSet.count(pred) == 0) {
return bad(folly::sformat("arc with src not in the region: {} -> {}",
pred, bid));
}
}
}
// 6) is checked by dfsCheck.
DFSChecker dfsCheck(region);
if (!dfsCheck.check(region.entry()->id())) {
return bad("region is cyclic");
}
// 7) All blocks are reachable from the entry (first) block.
if (dfsCheck.numVisited() != blockSet.size()) {
return bad("region has unreachable blocks");
}
// 8) and 9) are checked below.
RegionDesc::BlockIdSet visited;
auto& blocks = region.blocks();
for (unsigned i = 0; i < blocks.size(); i++) {
auto bid = blocks[i]->id();
unsigned nVisited = 0;
for (auto pred : region.preds(bid)) {
nVisited += visited.count(pred);
}
// 8) For each block, there must be a path from the entry to it that
// includes only earlier blocks in the region.
if (nVisited == 0 && i != 0) {
return bad(folly::sformat("block {} appears before all its predecessors",
bid));
}
// 9) The region is topologically sorted unless loops are enabled.
if (!RuntimeOption::EvalJitLoops && nVisited != region.preds(bid).size()) {
return bad(folly::sformat("non-topological order (bid: {})", bid));
}
visited.insert(bid);
}
// 10) The block-retranslation chains cannot have cycles.
for (auto b : blocks) {
auto bid = b->id();
RegionDesc::BlockIdSet chainSet;
chainSet.insert(bid);
while (auto next = region.nextRetrans(bid)) {
auto nextId = next.value();
if (chainSet.count(nextId)) {
return bad(folly::sformat("cyclic retranslation chain for block {}",
bid));
}
chainSet.insert(nextId);
bid = nextId;
}
}
return true;
}
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");
}