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

[Snehasish Kumar 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;
----------------
snehasish wrote:

This seems more like `firstN of 64K` rather than random sampling. I have a couple of concerns.
1. This seems like it could easily introduce bias.
2. We introduce yet another knob for tuning `sampled-instr-duration`. If it was randomized then we won't have to worry about this knob. A fast random implementation may not be too much increase in overhead compared to this approach.

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