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");
}
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");
}
