Increase cost of ARR_LENGTH node to match IND(ADD(..,CNS))

[EgorBo]

ARR_LENGTH was (2,2) while IND(ADD(addr, 8)) is (4,4) while both are effectively the same.

Diffs - seems to be a big PerfScore improvement. Surprisingly, it's also a size improvement on arm64.

[dotnet-policy-service]

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

[Copilot]

Pull Request Overview

This PR updates the evaluation cost of the GT_MDARR_LOWER_BOUND (array length) node so that it matches the cost of an IND(ADD(addr, smallCns)) sequence, improving both performance scoring and code size estimates.

• Revised the explanatory comment to reference the IND(ADD(...)) pattern.
• Increased costEx by 2 (from IND_COST_EX - 1 to IND_COST_EX + 1).
• Updated costSz from 2 to 4 (2 * 2).

[EgorBo]

@dotnet/jit-contrib @jakobbotsch opinions on this? diffs imply it's a perfscore improvement.

[jakobbotsch]

@dotnet/jit-contrib @jakobbotsch opinions on this? diffs imply it's a perfscore improvement.

Can you spot check some of the size regressions that are perfscore improvements? What happens in them and do they look beneficial?

I am fine with this change.

[EgorBo]

Can you spot check some of the size regressions that are perfscore improvements? What happens in them and do they look beneficial?

@jakobbotsch so my motivation to was to use more of LCL_VAR because those support constant folding with help of assertions, e.g.:

if (arr.Length == 10)
{
    memcpy(,, arr.Length); // I want to unroll this one.
}

in this case if it gets CSE'd then the assert-prop kicks in, if it's not - it doesn't. I actually tried to fix assertprop by just supporting ARR_LENGTH folding, but diffs were not nice #117505

Many size regressions seem to be caused by extra push+pop instructions for callee-saved registers used for extra CSEs. e.g.: