[EgorBo]

Use up to three compatible profiled runtime types when expanding isinst/castclass, rather than only the first candidate. Guards sharing a result block must agree on returning the object or null, and multi-check expansions load the method table once. The general cast helper remains the fallback for unprofiled types.

Add DOTNET_JitCastMaxTypeChecks (default 3) to control cast candidates independently of virtual-call GDV. NativeAOT's getExactClasses enumeration uses the same budget without relaxing its existing exactness checks; size-optimized NativeAOT casts consequently default to three candidates instead of one.

Clamp normalized conditional edge likelihoods to avoid floating-point rounding assertions when candidate probabilities sum to 100%.

[azure-pipelines]

Azure Pipelines:
Successfully started running 5 pipeline(s).
11 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[EgorBo]

@EgorBot -arm -amd

using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

BenchmarkSwitcher.FromAssembly(typeof(CastBench).Assembly).Run(args);

public class CastBench {
    private readonly object[] _objects = [new B(), new C()];

    [Benchmark(OperationsPerInvoke = 2)]
    public int IsA() {
        int count = 0;
        foreach (object obj in _objects) {
            if (obj is A) {
                count++;
            }
        }
        return count;
    }

    private class A { }
    private sealed class B : A { }
    private sealed class C : A { }
}

Note

This benchmark comment was composed with GitHub Copilot.

[EgorBo]

PTAL @dotnet/jit-contrib @adamperlin

A few improvements to cast expansion, enable multiple guessed by default.
Example of a benchmark - improved by 1.5x

[jkotas]

size-optimized NativeAOT casts consequently default to three candidates instead of one.

What is the size optimization that this is talking about? Checking three candidates inline does not sound like a size optimization to me - it is more code than a call to a cast helper.

[EgorBo]

size-optimized NativeAOT casts consequently default to three candidates instead of one.

What is the size optimization that this is talking about? Checking three candidates inline does not sound like a size optimization to me - it is more code than a call to a cast helper.

@jkotas oops, that was not intentional, although, e.g. 2 exact checks (no fallback) is not much bigger than just helper-call: https://godbolt.org/z/bcfqc6v5W (in case of NAOT we mostly talk about the non-fallbac get-exact-classes case only).

Fixed.