[libc++] Make the benchmarking data pseudo-random

[philnik777]

Having truly random data makes it quite difficult to compare benchmark results, since a significant amount of difference can simply be due to different input data. Making the data determinisic helps a lot when trying to figure out whether specific changes indeed improve the performance.

[llvmbot]

@llvm/pr-subscribers-libcxx

Author: Nikolas Klauser (philnik777)

Changes

Having truly random data makes it quite difficult to compare benchmark results, since a significant amount of difference can simply be due to different input data. Making the data determinisic helps a lot when trying to figure out whether specific changes indeed improve the performance.

---

Full diff: https://github.com/llvm/llvm-project/pull/159525.diff

1 Files Affected:

• (modified) libcxx/test/benchmarks/GenerateInput.h (+1-1)

diff --git a/libcxx/test/benchmarks/GenerateInput.h b/libcxx/test/benchmarks/GenerateInput.h
index 06387852f76a6..d07cb8857579d 100644
--- a/libcxx/test/benchmarks/GenerateInput.h
+++ b/libcxx/test/benchmarks/GenerateInput.h
@@ -25,7 +25,7 @@ static const char Letters[] = {
 static const std::size_t LettersSize = sizeof(Letters);
 
 inline std::default_random_engine& getRandomEngine() {
-  static std::default_random_engine RandEngine(std::random_device{}());
+  static std::default_random_engine RandEngine(123456);
   return RandEngine;
 }
 

[ldionne]

Can you confirm that you don't see any wild differences when running the microbenchmarks before and after this patch?

[philnik777]

Can you confirm that you don't see any wild differences when running the microbenchmarks before and after this patch?

I've checked with vector.bench.cpp and string.bench.cpp and didn't see anything out of the ordinary.