[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;
----------------
WenleiHe wrote:

> xur@'s sampling scheme is bursty style which seems less likely to introduce bias compared with every N hits scheme

Why? In general, sampling needs to be evenly distributed to avoid bias. HW sampling doesn't do burst either.  Do you have data and example to illustrate the benefit of bursty sampling? 

The only reason I can see to choose bursty sample is to open up the opportunity for coalescing increments, but that is not done here. 

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