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

[via llvm-commits]

================
@@ -635,33 +664,161 @@ PreservedAnalyses InstrProfilingLoweringPass::run(Module &M,
   return PreservedAnalyses::none();
 }
 
+//
+// Perform instrumentation sampling.
+//
+// There are 3 favors of sampling:
+// (1) Full burst sampling: We transform:
+//   Increment_Instruction;
+// to:
+//   if (__llvm_profile_sampling__ < SampledInstrBurstDuration) {
+//     Increment_Instruction;
+//   }
+//   __llvm_profile_sampling__ += 1;
+//   if (__llvm_profile_sampling__ >= SampledInstrPeriod) {
+//     __llvm_profile_sampling__ = 0;
+//   }
+//
+// "__llvm_profile_sampling__" is a thread-local global shared by all PGO
+// counters (value-instrumentation and edge instrumentation).
+//
+// (2) Fast burst sampling:
+// The value is an unsigned type, meaning it will wrap around to zero when
----------------
xur-llvm wrote:

ack.

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