Don't use ProfPrologues's counters to estimate function invocations f…

…or hfsort

Summary:
That can be off because we often skip the prologue when entering JITed code.
Instead, use the counters for the translations that match the function's base
offset.

Reviewed By: mofarrell

Differential Revision: D15935666

fbshipit-source-id: 9b746357b5397871af31a9133273dabd8671b4eb

hphp/runtime/vm/jit/mcgen-order.cpp

@@ -55,14 +55,21 @@ createCallGraph(jit::hash_map<hfsort::TargetId, FuncId>& funcID) {
   FTRACE(3, "createCallGraph: maxFuncId = {}\n", maxFuncId);
   for (FuncId fid = 0; fid <= maxFuncId; fid++) {
     if (!Func::isFuncIdValid(fid) || !pd->profiling(fid)) continue;
+    const auto func = Func::fromFuncId(fid);
+    const auto baseOffset = func->base();
     const auto transIds = pd->funcProfTransIDs(fid);
     uint32_t size = 1; // avoid zero-sized functions
+    uint32_t profCount = 0;
     for (auto transId : transIds) {
       const auto trec = pd->transRec(transId);
       assertx(trec->kind() == TransKind::Profile);
       size += trec->asmSize();
+      if (trec->srcKey().offset() == baseOffset) {
+        profCount += pd->transCounter(transId);
+      }
     }
-    const auto targetId = cg.addTarget(size);
+    // NB: avoid division by 0
+    const auto targetId = cg.addTarget(size, profCount ? profCount : 1);
     targetID[fid] = targetId;
     funcID[targetId] = fid;
     FTRACE(3, "  - adding node FuncId = {} => TargetId = {}\n", fid, targetId);
@@ -105,7 +112,6 @@ createCallGraph(jit::hash_map<hfsort::TargetId, FuncId>& funcID) {
     const auto calleeTargetId = targetID[fid];
     const auto transIds = pd->funcProfTransIDs(fid);
     uint32_t totalCalls = 0;
-    uint32_t profCount  = 1; // avoid zero sample counts
     for (int nargs = 0; nargs <= func->numNonVariadicParams() + 1; nargs++) {
       auto transId = pd->proflogueTransId(func, nargs);
       if (transId == kInvalidTransID) continue;
@@ -119,9 +125,9 @@ createCallGraph(jit::hash_map<hfsort::TargetId, FuncId>& funcID) {
       }
       addCallersCount(calleeTargetId, trec->mainCallers(),  totalCalls);
       addCallersCount(calleeTargetId, trec->guardCallers(), totalCalls);
-      profCount += pd->transCounter(transId);
     }
-    cg.setSamples(calleeTargetId, std::max(totalCalls, profCount));
+    auto samples = cg.getSamples(calleeTargetId);
+    cg.setSamples(calleeTargetId, std::max(totalCalls, samples));
   }
   cg.normalizeArcWeights();
   return cg;

hphp/util/hfsort.cpp

@@ -65,6 +65,11 @@ void TargetGraph::setSamples(TargetId id, uint32_t samples) {
   targets[id].samples = samples;
 }
 
+uint32_t TargetGraph::getSamples(TargetId id) {
+  assertx(id < targets.size());
+  return targets[id].samples;
+}
+
 const Arc& TargetGraph::incArcWeight(TargetId src, TargetId dst, double w) {
   auto res = arcs.emplace(src, dst, w);
   if (!res.second) {

hphp/util/hfsort.h

@@ -70,6 +70,7 @@ struct Target {
 struct TargetGraph {
   TargetId addTarget(uint32_t size, uint32_t samples = 0);
   void setSamples(TargetId id, uint32_t samples);
+  uint32_t getSamples(TargetId id);
   const Arc& incArcWeight(TargetId src, TargetId dst, double w = 1.0);
   void normalizeArcWeights();