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

for (int i = 0; i < 200; i++)
{
    Test(new Program());
    Thread.Sleep(10);
}


static void Test(object o)
{
    if (o is Program)
        Console.WriteLine();
}

; Assembly listing for method Program:Test(System.Object) (Tier1)
G_M3485_IG01:
       sub      rsp, 40
						;; size=4 bbWeight=1 PerfScore 0.25
G_M3485_IG02:
       mov      rax, rcx
       test     rax, rax
       je       SHORT G_M3485_IG04
						;; size=8 bbWeight=1 PerfScore 1.50
G_M3485_IG03:
       mov      rdx, 0xD1FFAB1E      ; Program
       cmp      qword ptr [rax], rdx
       jne      SHORT G_M3485_IG07
						;; size=15 bbWeight=0.50 PerfScore 2.12
G_M3485_IG04:
       test     rax, rax
       je       SHORT G_M3485_IG05
       call     [System.Console:WriteLine()]
						;; size=11 bbWeight=1 PerfScore 4.25
G_M3485_IG05:
       nop      
						;; size=1 bbWeight=1 PerfScore 0.25
G_M3485_IG06:
       add      rsp, 40
       ret      
						;; size=5 bbWeight=1 PerfScore 1.25
G_M3485_IG07:
       mov      rdx, rcx
       mov      rcx, 0xD1FFAB1E      ; Program
       call     CORINFO_HELP_ISINSTANCEOFCLASS
       jmp      SHORT G_M3485_IG04
						;; size=20 bbWeight=0 PerfScore 0.00

; Total bytes of code 64

@AndyAyersMS @dotnet/jit-contrib PTAL,

1. Change likelihood treshold from 75 to 40 (it's also defined as a in-place constant for GDV)
2. Save weights inside QMARK.

Small diffs due to more blocks marked as cold (due to profile) https://dev.azure.com/dnceng-public/public/_build/results?buildId=522520&view=ms.vss-build-web.run-extensions-tab (I suspect the real diffs could be bigger but that needs new collections)

In our class profile probes we ignore cases with null objects. I wonder if for casting we shouldn't ignore them, then you would know how often the null case might come up...?

We could make this change generally but would need to think through the repercussions on consumers; some of them are likely not expecting to see a null result.

For example:

HCIMPL2(void, JIT_ClassProfile32, Object *obj, ICorJitInfo::HandleHistogram32* classProfile)
{
    FCALL_CONTRACT;
    FC_GC_POLL_NOT_NEEDED();

    OBJECTREF objRef = ObjectToOBJECTREF(obj);
    VALIDATEOBJECTREF(objRef);

    size_t sampleIndex;
    if (!CheckSample(&classProfile->Count, &sampleIndex) || objRef == NULL)
    {
        return;
    }

    MethodTable* pMT = objRef->GetMethodTable();

could be

HCIMPL2(void, JIT_ClassProfile32, Object *obj, ICorJitInfo::HandleHistogram32* classProfile)
{
    FCALL_CONTRACT;
    FC_GC_POLL_NOT_NEEDED();

    OBJECTREF objRef = ObjectToOBJECTREF(obj);
    VALIDATEOBJECTREF(objRef);

    size_t sampleIndex;
    if (!CheckSample(&classProfile->Count, &sampleIndex))
    {
        return;
    }

    MethodTable* pMT = (objRef == NULL) ? NULL : objRef->GetMethodTable();

In our class profile probes we ignore cases with null objects. I wonder if for casting we shouldn't ignore them, then you would know how often the null case might come up...?

The part that changes weights definitely improves code densitiy and my microbenchmark. The changed threshold is purely speculative at this point. Although, there are CastingPerf benchmarks which test polymorphic casts. It's just that I noticed that a cast that has 50% Class1 and 50% Class2 is never expanded because the threshold was 75%. I think this all might be tunned because some casts are very expensinve while others are not.

In our class profile probes we ignore cases with null objects. I wonder if for casting we shouldn't ignore them, then you would know how often the null case might come up...?

That's a very good point - I was considering it too, I was just not sure how to trace null properly. I assume the empy table is basically all nulls, right? So how do I tell the difference between a table that is just almost empty (because we only had 1 hit) vs a very hot table that is half null because values are nulls?

Should I pre-init it with some magic value? Or it's not a bid deal to handle cold tables like hot ones?

I'll experiment with nulls separately

• [Perf] Linux/arm64: 25 Regressions on 1/3/2024 11:26:14 PM perf-autofiling-issues#27493
• [Perf] Windows/arm64: 4 Regressions on 1/11/2024 5:32:36 PM perf-autofiling-issues#27475