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

[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;
+//
+// "__llvm_profile_sampling__" is a thread-local global shared by all PGO
+// instrumentation variables (value-instrumentation and edge instrumentation).
+// It has a unsigned short type and will wrapper around when overflow.
+//
+// Note that, the code snippet after the transformation can still be
+// counter promoted. But I don't see a reason for that because the
+// counter updated should be sparse. That's the reason we disable
+// counter promotion by default when sampling is enabled.
----------------
xur-llvm wrote:

OK. I can expand a bit more there: The downside of counter promotion is that we can get incomplete profile if we dump the counter in the middle of the loop. The benefit is improve the instrumentation speed. With this patch, the benefit is very small and won't overweight the potential downside.

As I said in the comment, these two techniques can work together without any issue. I actually added a test case for them to work together.

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