Use TryGetRangeFromAssertions to fold more relops in global assert prop

[EgorBo]

This PR does:

1. Folds more various "implied relops" in optAssertionPropGlobal_RelOp with help of RangeCheck::TryGetRangeFromAssertions (it's not too expensive). For now, it only handles "Op1 relop IntCns". For that, I introduced EvalRelop that can answer questions about two ranges (e.g. range1 is >= range2, etc)
2. Improved MergeEdgeAssertions a bit:
   • if VN is actually ArrayLen, then for bounds-check NO_THROW assertions we can also check if our vn equals lenVN - this gives us a hint that lenVN is >= indexVN (in case of constant)
   • for X >= CNS we now produce "[CNS...INT32_MAX]" instead of "[CNS...UNKNOWN]" - this helps DoesOverflow to be less conservative for full constant ranges

One note that RangeCheck::TryGetRangeFromAssertions is more or less cheap - it relies only on assertions. There is also RangeCheck::TryGetRange that is SSA-based - it finds a lot more places to optimize, with it, I see -250k diffs. Unfortunately, we have to pay with up to +1% TP for them. I want to invest some time in improving TP in RangeCheck to see if I can get 80% of th e diffs for a small price

Example of an improved pattern:

void Test(Span<int> src)
{
    src[5] = 42; // creates an assertion
    Consume(src.Slice(4)); // fold checks here based on the assertion ^
}

Diffs

[EgorBo]

@MihuBot

[Copilot]

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

[EgorBo]

/azp run runtime-coreclr libraries-pgo, runtime-coreclr jitstress, runtime-coreclr libraries-jitstress, Fuzzlyn

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).

[EgorBo]

/azp run runtime-coreclr libraries-pgo, runtime-coreclr jitstress, runtime-coreclr libraries-jitstress, Fuzzlyn

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).

[EgorBo]

PTAL @jakobbotsch @dotnet/jit-contrib diffs

I ended up removing SSA-based for now for two reasons:

1. it's too TP heavy
2. it has a small issue in DoesOverflow here (it should be properly handled)

I'll try again separately, SSA-based one increases diffs significantly (up to -250k for win-x64)

Fuzzlyn's failures repro on Main too, although, looks like some of them are my fault (previous PRs) - I'm looking now

[AndyAyersMS]

Can you update the description to describe what this does now?

[AndyAyersMS]

Could this use op2's VN instead relying on it being a literal constant?

[EgorBo]

Sure, I just assumed that we always propagate constants by value before global assertprop so while we generally should rely on VNs, for constants it's more or less fine 🙂 but in the sake of consistency I guess I can

[EgorBo]

@AndyAyersMS I've addressed your feedback in 877112c and updated the issue description