[compiler-rt] [compiler-rt][ctx_profile] Add the instrumented contextual profiling APIs (PR #89838)

Teresa Johnson via llvm-commits llvm-commits at lists.llvm.org
Mon May 6 08:46:44 PDT 2024


================
@@ -38,3 +128,177 @@ void Arena::freeArenaList(Arena *&A) {
   }
   A = nullptr;
 }
+
+inline ContextNode *ContextNode::alloc(char *Place, GUID Guid,
+                                       uint32_t NrCounters,
+                                       uint32_t NrCallsites,
+                                       ContextNode *Next) {
+  assert(reinterpret_cast<uint64_t>(Place) % sizeof(void *) == 0);
+  return new (Place) ContextNode(Guid, NrCounters, NrCallsites, Next);
+}
+
+void ContextNode::reset() {
+  for (uint32_t I = 0; I < NrCounters; ++I)
+    counters()[I] = 0;
+  for (uint32_t I = 0; I < NrCallsites; ++I)
+    for (auto *Next = subContexts()[I]; Next; Next = Next->Next)
+      Next->reset();
+}
+
+// If this is the first time we hit a callsite with this (Guid) particular
+// callee, we need to allocate.
+ContextNode *getCallsiteSlow(uint64_t Guid, ContextNode **InsertionPoint,
+                             uint32_t NrCounters, uint32_t NrCallsites) {
+  auto AllocSize = ContextNode::getAllocSize(NrCounters, NrCallsites);
+  auto *Mem = __llvm_ctx_profile_current_context_root->CurrentMem;
+  char *AllocPlace = Mem->tryBumpAllocate(AllocSize);
+  if (!AllocPlace) {
+    // if we failed to allocate on the current arena, allocate a new arena,
+    // and place it on __llvm_ctx_profile_current_context_root->CurrentMem so we
+    // find it from now on for other cases when we need to getCallsiteSlow.
+    // Note that allocateNewArena will link the allocated memory in the list of
+    // Arenas.
+    __llvm_ctx_profile_current_context_root->CurrentMem = Mem =
+        Mem->allocateNewArena(getArenaAllocSize(AllocSize), Mem);
+  }
+  auto *Ret = ContextNode::alloc(AllocPlace, Guid, NrCounters, NrCallsites,
+                                 *InsertionPoint);
+  *InsertionPoint = Ret;
+  return Ret;
+}
+
+ContextNode *__llvm_ctx_profile_get_context(void *Callee, GUID Guid,
+                                            uint32_t NrCounters,
+                                            uint32_t NrCallsites) {
+  // fast "out" if we're not even doing contextual collection.
+  if (!__llvm_ctx_profile_current_context_root)
+    return TheScratchContext;
+
+  // also fast "out" if the caller is scratch.
+  auto **CallsiteContext = consume(__llvm_ctx_profile_callsite[0]);
+  if (!CallsiteContext || isScratch(*CallsiteContext))
+    return TheScratchContext;
+
+  // if the callee isn't the expected one, return scratch.
+  // Signal handler(s) could have been invoked at any point in the execution.
+  // Should that have happened, and had it (the handler) be built with
+  // instrumentation, its __llvm_ctx_profile_get_context would have failed here.
+  // Its sub call graph would have then populated
+  // __llvm_ctx_profile_{expected_callee | callsite} at index 1.
+  // The normal call graph may be impacted in that, if the signal handler
+  // happened somewhere before we read the TLS here, we'd see the TLS reset and
+  // we'd also fail here. That would just mean we would loose counter values for
+  // the normal subgraph, this time around. That should be very unlikely, but if
+  // it happens too frequently, we should be able to detect discrepancies in
+  // entry counts (caller-callee). At the moment, the design goes on the
+  // assumption that is so unfrequent, though, that it's not worth doing more
+  // for that case.
+  auto *ExpectedCallee = consume(__llvm_ctx_profile_expected_callee[0]);
+  if (ExpectedCallee != Callee)
+    return TheScratchContext;
+
+  auto *Callsite = *CallsiteContext;
+  while (Callsite && Callsite->guid() != Guid) {
----------------
teresajohnson wrote:

What is this loop doing?

https://github.com/llvm/llvm-project/pull/89838


More information about the llvm-commits mailing list