[wasm] Fix bugs related to jiterpreter heuristic

[kg]

While investigating some benchmark regressions I noticed a couple bugs related to the jiterpreter.

I did 3 runs of browser-bench for before/after and the difference is within noise for me, so I don't know whether this makes anything faster or slower. The existing instruction counter is broken in an unpredictable way though, and this correct implementation of it is at least possible to tune.

[ghost]

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

Issue Details

---

While investigating some benchmark regressions I noticed a couple bugs related to the jiterpreter.

I did 3 runs of browser-bench for before/after and the difference is within noise for me, so I don't know whether this makes anything faster or slower. The existing instruction counter is broken in an unpredictable way though, and this correct implementation of it is at least possible to tune.

Author:	kg
Assignees:	-
Labels:	arch-wasm, area-Codegen-Interpreter-mono
Milestone:	-

[kg]

Blocked by

[21:37:54] info: Starting:    System.Text.RegularExpressions.Tests.dll
[21:38:40] info: long superbranch detected with opcode 259 (beq.i4.imm.sp) in method <PcreTestData_Cases>d__1.MoveNext
[21:38:40] fail: [MONO] * Assertion at /__w/1/s/src/mono/mono/mini/interp/transform.c:8558, condition `<disabled>' not met
[22:07:50] fail: Tests timed out. Killing driver service pid 80

Which is caused by a 2900-line generator function, the generated state machine seems to have a chance to end up with complex long branches in it depending on how the opcodes end up arranged, and fixing the jiterpreter heuristic permutes the opcodes just so.

We could hide this problem by splitting the generator function in half, but for now we're going to figure out whether there's a better solution.