[llvm] [MemProf] Extend MemProfUse pass to make use of data access profiles to partition data (PR #151238)

[Mingming Liu via llvm-commits]

================
@@ -752,3 +766,67 @@ PreservedAnalyses MemProfUsePass::run(Module &M, ModuleAnalysisManager &AM) {
 
   return PreservedAnalyses::none();
 }
+
+bool MemProfUsePass::annotateGlobalVariables(
+    Module &M, const memprof::DataAccessProfData *DataAccessProf) {
+  if (!AnnotateStaticDataSectionPrefix || M.globals().empty())
+    return false;
+
+  if (!DataAccessProf) {
+    M.getContext().diagnose(DiagnosticInfoPGOProfile(
+        MemoryProfileFileName.data(),
+        StringRef("Data access profiles not found in memprof. Ignore "
+                  "-memprof-annotate-static-data-prefix."),
+        DS_Warning));
+    return false;
+  }
+
+  bool Changed = false;
+  // Iterate all global variables in the module and annotate them based on
+  // data access profiles. Note it's up to the linker to decide how to map input
+  // sections to output sections, and one conservative practice is to map
+  // unlikely-prefixed ones to unlikely output section, and map the rest
+  // (hot-prefixed or prefix-less) to the canonical output section.
+  for (GlobalVariable &GVar : M.globals()) {
+    assert(!GVar.getSectionPrefix().has_value() &&
+           "GVar shouldn't have section prefix yet");
+    if (GVar.isDeclarationForLinker())
+      continue;
+
+    StringRef Name = GVar.getName();
+    // Skip string literals as their mangled names don't stay stable across
+    // binary binary releases.
+    // TODO: Track string content hash in the profiles and compute it inside the
+    // compiler to categeorize the hotness string literals.
+    if (Name.starts_with(".str")) {
+      LLVM_DEBUG(dbgs() << "Skip annotating string literal " << Name << "\n");
+      continue;
+    }
+
+    // DataAccessProfRecord's get* methods will canonicalize the name under the
+    // hood before looking it up, so optimizer doesn't need to do it.
+    std::optional<DataAccessProfRecord> Record =
+        DataAccessProf->getProfileRecord(Name);
+    // Annotate a global variable as hot if it has non-zero sampled count, and
+    // annotate it as cold if it's seen in the profiled binary
+    // file but doesn't have any access sample.
+    // For logging, optimization remark emitter requires a llvm::Function, but
+    // it's not well defined how to associate a global variable with a function.
+    // So we just print out the static data section prefix in LLVM_DEBUG.
----------------
mingmingl-llvm wrote:

I'm onboard with adding counters so implemented it; though if I remember correctly stats are not always enabled in a release build compiler, and there is a `DLLVM_ENABLE_STATS` flag.

https://github.com/llvm/llvm-project/pull/151238