Trust single-edge synthetic profile #116054
Conversation
github-actions
bot
added
the
area-CodeGen-coreclr
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
|
@EgorBot -amd -arm using System.Collections.Generic;
using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;
public class Bench
{
private static Dictionary<string, int> Dictionary =
new()
{
{ "1", 1 },
{ "222", 2 },
{ "3333", 3 },
{ "44444", 4 },
};
[Benchmark]
[Arguments("44444")]
public int Lookup(string key) => Dictionary[key];
} |
Co-authored-by: Copilot <[email protected]>
|
PTAL @AndyAyersMS @dotnet/jit-contrib @AndyAyersMS I remember you recently made a fix for a similar case, but in your case it was not the root inlinee. E.g. the benchmark ^ is significantly improved as |
There was a problem hiding this comment.
I thought about doing something lilke this but wondered if it might be a bit too benchmarkish. I had been thinking of modifying the "wrapper method" heuristic but your approach seems simpler and more general, since trust/sufficiency plays out more broadly.
It seems quite possible there is user code written this way where for whatever reason the trivial root method can't be inlined.
This will show an enormous amount of benchmark wins on Windows/Viper when the inlinee is fairly trivial (100's of benchmarks) due to elimination of CET overhead.
Normally, we don't instrument single-edge methods, it means if such a method is an inline root (it happens in Micro-benchmarks as normally single-edge methods are inlined somewhere), all callees don't receive an inlining boost.