void Clusterizer::splitHotColdClusters() {
// compute the average weight of each cluster
jit::vector<uint64_t> clusterAvgWgt(m_clusters.size());
for (size_t c = 0; c < m_clusters.size(); c++) {
uint64_t totalWeight = 0;
uint64_t totalSize = 0;
for (auto b : m_clusters[c]) {
const auto numInsts = m_unit.blocks[b].code.size();
totalSize += numInsts;
totalWeight += numInsts * m_scale.weight(b);
}
clusterAvgWgt[c] = totalSize == 0 ? 0 : totalWeight / totalSize;
}
const auto entryAvgWgt = clusterAvgWgt[m_blockCluster[m_unit.entry]];
const uint64_t hotThreshold = entryAvgWgt *
RuntimeOption::EvalJitLayoutHotThreshold;
FTRACE(3, "splitHotColdClusters: entryAvgWgt = {} ; hotThreshold = {}\n",
entryAvgWgt, hotThreshold);
for (auto cid : m_clusterOrder) {
if (m_clusters[cid].size() == 0) continue;
const AreaIndex area = clusterAvgWgt[cid] >= hotThreshold ? AreaIndex::Main
: AreaIndex::Cold;
FTRACE(3, " -> C{}: {} (avg wgt = {}): ",
cid, area_names[unsigned(area)], clusterAvgWgt[cid]);
for (auto b : m_clusters[cid]) {
// don't reassign blocks that are in frozen
if (m_unit.blocks[b].area_idx == AreaIndex::Frozen) continue;
m_unit.blocks[b].area_idx = area;
FTRACE(3, "{}, ", b);
}
FTRACE(3, "\n");
}
}
/*
* Merge two state-stacks. The stacks must have the same depth. Returns
* whether any states changed.
*/
bool merge_into(jit::vector<FrameState>& dst, const jit::vector<FrameState>& src) {
always_assert(src.size() == dst.size());
auto changed = false;
for (auto idx = uint32_t{0}; idx < dst.size(); ++idx) {
changed |= merge_into(dst[idx], src[idx]);
}
return changed;
}
void Clusterizer::initClusters() {
m_clusters.resize(m_unit.blocks.size());
m_blockCluster.resize(m_unit.blocks.size());
for (auto b : m_blocks) {
m_clusters[b].push_back(b);
m_blockCluster[b] = b;
}
}
bool merge_memory_stack_into(jit::vector<StackState>& dst,
const jit::vector<StackState>& src) {
auto changed = false;
// We may need to merge different-sized memory stacks, because a predecessor
// may not touch some stack memory that another pred did. We just need to
// conservatively throw away slots that aren't tracked on all preds.
auto const result_size = std::min(dst.size(), src.size());
dst.resize(result_size);
for (auto i = uint32_t{0}; i < result_size; ++i) {
changed |= merge_into(dst[i], src[i]);
}
return changed;
}
void DFSSortClusters::dfs(uint32_t cid) {
if (m_visited.test(cid)) return;
m_visited.set(cid);
m_list.push_back(Vlabel(cid));
// find the best successor, which is the one to which cid has the
// highest weight among the ones that haven't been visited yet
int64_t maxWgt = 0;
uint32_t bestSucc = uint32_t(-1);
for (auto& sInfo : m_clusterSuccs[cid]) {
auto succId = sInfo.first;
if (m_visited.test(succId)) continue;
auto wgt = sInfo.second;
if (wgt >= maxWgt) {
maxWgt = wgt;
bestSucc = succId;
}
}
if (bestSucc == uint32_t(-1)) return;
// visit bestSucc first
dfs(bestSucc);
// now visit the remaining ones
for (auto& sInfo : m_clusterSuccs[cid]) {
if (sInfo.first != bestSucc) {
dfs(sInfo.first);
}
}
}
void natural_loop_dfs(jit::vector<Block*>& out, Visited& visited, Block* blk) {
if (visited.test(blk->id())) return;
visited.set(blk->id());
blk->forEachPred([&] (Block* pred) {
natural_loop_dfs(out, visited, pred);
});
out.push_back(blk);
}
void Vgen::emit(jcc& i) {
assertx(i.cc != CC_None);
if (i.targets[1] != i.targets[0]) {
if (next == i.targets[1]) {
// the taken branch is the fall-through block, invert the branch.
i = jcc{ccNegate(i.cc), i.sf, {i.targets[1], i.targets[0]}};
}
jccs.push_back({a->frontier(), i.targets[1]});
// B.cond range is +/- 1MB but this uses BR
backend.emitSmashableJump(*codeBlock, kEndOfTargetChain, i.cc);
}
emit(jmp{i.targets[0]});
}
void Vgen::emit(tbcc& i) {
assertx(i.cc == vixl::ne || i.cc == vixl::eq);
if (i.targets[1] != i.targets[0]) {
if (next == i.targets[1]) {
// the taken branch is the fall-through block, invert the branch.
i = tbcc{i.cc == vixl::ne ? vixl::eq : vixl::ne, i.bit, i.s,
{i.targets[1], i.targets[0]}};
}
bccs.push_back({a->frontier(), i.targets[1]});
// offset range +/- 32KB
if (i.cc == vixl::ne) {
a->tbnz(X(i.s), i.bit, 0);
} else {
a->tbz(X(i.s), i.bit, 0);
}
}
emit(jmp{i.targets[0]});
}
void Vgen::emit(jmp i) {
if (next == i.target) return;
jmps.push_back({a->frontier(), i.target});
// B range is +/- 128MB but this uses BR
backend.emitSmashableJump(*codeBlock, kEndOfTargetChain, CC_None);
}
void Vgen::emit(hcunwind& i) {
catches.push_back({points[i.call], i.targets[1]});
emit(jmp{i.targets[0]});
}
// overall emitter
void Vgen::emit(jit::vector<Vlabel>& labels) {
// Some structures here track where we put things just for debug printing.
struct Snippet {
const IRInstruction* origin;
TcaRange range;
};
struct BlockInfo {
jit::vector<Snippet> snippets;
};
// This is under the printir tracemod because it mostly shows you IR and
// machine code, not vasm and machine code (not implemented).
bool shouldUpdateAsmInfo = !!m_asmInfo
&& Trace::moduleEnabledRelease(HPHP::Trace::printir, kCodeGenLevel);
std::vector<TransBCMapping>* bcmap = nullptr;
if (mcg->tx().isTransDBEnabled() || RuntimeOption::EvalJitUseVtuneAPI) {
bcmap = &mcg->cgFixups().m_bcMap;
}
jit::vector<jit::vector<BlockInfo>> areaToBlockInfos;
if (shouldUpdateAsmInfo) {
areaToBlockInfos.resize(areas.size());
for (auto& r : areaToBlockInfos) {
r.resize(unit.blocks.size());
}
}
for (int i = 0, n = labels.size(); i < n; ++i) {
assertx(checkBlockEnd(unit, labels[i]));
auto b = labels[i];
auto& block = unit.blocks[b];
codeBlock = &area(block.area).code;
vixl::MacroAssembler as { *codeBlock };
a = &as;
auto blockStart = a->frontier();
addrs[b] = blockStart;
{
// Compute the next block we will emit into the current area.
auto cur_start = start(labels[i]);
auto j = i + 1;
while (j < labels.size() && cur_start != start(labels[j])) {
j++;
}
next = j < labels.size() ? labels[j] : Vlabel(unit.blocks.size());
}
const IRInstruction* currentOrigin = nullptr;
auto blockInfo = shouldUpdateAsmInfo
? &areaToBlockInfos[unsigned(block.area)][b]
: nullptr;
auto start_snippet = [&](Vinstr& inst) {
if (!shouldUpdateAsmInfo) return;
blockInfo->snippets.push_back(
Snippet { inst.origin, TcaRange { codeBlock->frontier(), nullptr } }
);
};
auto finish_snippet = [&] {
if (!shouldUpdateAsmInfo) return;
if (!blockInfo->snippets.empty()) {
auto& snip = blockInfo->snippets.back();
snip.range = TcaRange { snip.range.start(), codeBlock->frontier() };
}
};
for (auto& inst : block.code) {
if (currentOrigin != inst.origin) {
finish_snippet();
start_snippet(inst);
currentOrigin = inst.origin;
}
if (bcmap && inst.origin) {
auto sk = inst.origin->marker().sk();
if (bcmap->empty() ||
bcmap->back().md5 != sk.unit()->md5() ||
bcmap->back().bcStart != sk.offset()) {
bcmap->push_back(TransBCMapping{sk.unit()->md5(), sk.offset(),
main().frontier(), cold().frontier(),
frozen().frontier()});
}
}
switch (inst.op) {
#define O(name, imms, uses, defs) \
case Vinstr::name: emit(inst.name##_); break;
VASM_OPCODES
#undef O
}
}
finish_snippet();
}
for (auto& p : jccs) {
assertx(addrs[p.target]);
backend.smashJcc(p.instr, addrs[p.target]);
}
for (auto& p : bccs) {
assertx(addrs[p.target]);
auto link = (Instruction*) p.instr;
link->SetImmPCOffsetTarget(Instruction::Cast(addrs[p.target]));
}
for (auto& p : jmps) {
assertx(addrs[p.target]);
backend.smashJmp(p.instr, addrs[p.target]);
}
for (auto& p : catches) {
mcg->registerCatchBlock(p.instr, addrs[p.target]);
}
for (auto& p : ldpoints) {
CodeCursor cc(main(), p.instr);
MacroAssembler a{main()};
a.Mov(X(p.d), points[p.pos]);
}
if (!shouldUpdateAsmInfo) {
return;
}
for (auto i = 0; i < areas.size(); ++i) {
const IRInstruction* currentOrigin = nullptr;
auto& blockInfos = areaToBlockInfos[i];
for (auto const blockID : labels) {
auto const& blockInfo = blockInfos[static_cast<size_t>(blockID)];
if (blockInfo.snippets.empty()) continue;
for (auto const& snip : blockInfo.snippets) {
if (currentOrigin != snip.origin && snip.origin) {
currentOrigin = snip.origin;
}
m_asmInfo->updateForInstruction(
currentOrigin,
static_cast<AreaIndex>(i),
snip.range.start(),
snip.range.end());
}
}
}
}
