[wasm] Optimize jiterpreter type cast opcodes & add zero page optimizations

[kg]

This PR inlines a sizable portion of MINT_CASTCLASS/MINT_ISINST logic into jiterpreter traces to create fast paths for exact matches/null pointers that avoids running all the type check machinery we normally would run. Based on my instrumentation these fast paths will be used ~30% of the time, and measurements show a 1-3% improvement for the json section of browser-bench (this is pretty hard to measure, I may try to construct a decent synthetic benchmark for this).

This PR also applies zero page optimizations if available, by fusing the null check into the obj->vtable load. The helper code is all tweaked so that in the event a check fails due to a null ptr (null inputs are fairly uncommon in the instrumentation) it is converted into a success since MINT_CASTCLASS and MINT_ISINST both succeed for null instead of throwing.

It may be worth inlining the interface check entirely, but I haven't done the work to test that yet. The existence of special interfaces for arrays makes that harder to do.

This PR also fully inlines the implementation of MINT_UNBOX.

[ghost]

Tagging subscribers to 'arch-wasm': @lewing
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Opening as draft because it relies on #86403.

This PR inlines a sizable portion of CASTCLASS/ISINST logic into jiterpreter traces to create fast paths for exact matches/null pointers that avoids running all the type check machinery we normally would run. Based on my instrumentation these fast paths will be used ~30% of the time, and measurements show a 1-3% improvement for the json section of browser-bench (this is pretty hard to measure, I may try to construct a decent synthetic benchmark for this).

This PR also applies zero page optimizations if available, by fusing the null check into the obj->vtable load. The helper code is all tweaked so that in the event a check fails due to a null ptr (null inputs are fairly uncommon in the instrumentation) it is converted into a success since CASTCLASS and ISINST both succeed for null.

It may be worth inlining the interface check entirely, but I haven't done the work to test that yet. The existence of special interfaces for arrays makes that harder to do.

Author:	kg
Assignees:	-
Labels:	NO-MERGE, arch-wasm, area-Codegen-Jiterpreter-mono
Milestone:	-

[vargaz]

isinst doesn't succeed on nulls.

[kg]

isinst doesn't succeed on nulls.

Thanks, I'll revise it

[kg]

I was puzzled that none of the scenarios I expected to get faster were getting faster with these changes. It turned out that the scenarios I was testing were actually relying on MINT_UNBOX. So I wrote some toy BDN microbenchmarks to use and also optimized the unbox operation. Some comparison timings for main vs this PR:

scenario	main	this PR	relative %
(int)obj	1.285ms	1.159ms	90%
if (obj is int value)	1.62ms	1.218ms	75%