[llvm] [PGO] Sampled instrumentation in PGO to speed up instrumentation binary (PR #69535)

[Hongtao Yu via llvm-commits]

================
@@ -412,30 +424,91 @@ PreservedAnalyses InstrProfiling::run(Module &M, ModuleAnalysisManager &AM) {
   return PreservedAnalyses::none();
 }
 
+// Perform instrumentation sampling.
+// We transform:
+//   Increment_Instruction;
+// to:
+//   if (__llvm_profile_sampling__ <= SampleDuration) {
+//     Increment_Instruction;
+//   }
+//   __llvm_profile_sampling__ += 1;
----------------
htyu wrote:

Since `__llvm_profile_sampling__` is shared by all PGO counters, I'm wondering if biasing could happen to code always shadowed in the range `[__llvm_profile_sampling__, 65536]`. E.g, assume two loops in the program, they count up to 65536 iterations in all, and the second loop will unlikely be counted?

https://github.com/llvm/llvm-project/pull/69535