LLVM: exit to dispatcher when mtmsr enables MSR[EE]

[dougchansan]

Summary

Pokemon Colosseum (GC6E01) rendered at 0.25 fps under the LLVM backend against 30 fps under the C backend, and could not cold-boot past its THP intro videos. Root cause: external-interrupt delivery starvation. This PR makes an mtmsr that enables MSR[EE] exit to the dispatcher, creating the delivery point the block-ending JITs get for free. The runtime half is ExpansionPak/RecompCore PR (delivers a pending external interrupt when a dispatch boundary shows EE=1); each half is safe alone, both are needed for the fix.

The mechanism

• Colosseum waits for VI retrace by yield-spinning: OSDisableInterrupts -> SelectThread(yield) -> OSRestoreInterrupts, check callback flag, check a 3-second bus-clock timeout, repeat. The other tested titles block on wait queues; this spin is why only Colosseum failed.
• Under the LLVM backend the whole critical section runs inside native calls within a single dispatch. The only burst-ending points are the budget guards, which sit at call sites -- in OS code, almost entirely inside interrupt-disabled windows. Burst boundaries phase-lock to EE=0.
• External interrupts are only deliverable when the host regains control with EE=1. Measured on the pinned scene: the VI interrupt was pending-but-blocked at 30% of delivery checks (C: 0.76%), and only 1.2% of raised interrupts were ever delivered. The post-retrace callback never ran; every frame escaped through the game's own 3-second timeout -- 1 frame / ~4 s = 0.2487 fps exactly.
• mtmsr was lowered as a plain state write. Dolphin's Jit64/JitArm64 end the block at mtmsr and check external exceptions; this change mirrors that: after the MSR store, a 0->1 transition of EE takes a side exit (sync, store pc+4, return). mtmsr is supervisor-only, so the exit is rare and cheap.

Measurements (Windows, pinned savestate, same runtime build)

	before	after
Colosseum LLVM	0.2487 fps	27.4-28.8 fps (4 runs)
Colosseum C (control)	29.98	29.99
MKDD LLVM (control)	48.65	51.0
Colosseum LLVM cold boot	hangs at THP videos	54.8-55.5 fps, boots

Interrupt counters on the pinned scene: raised 484 -> 31,777; delivered 462 -> 29,856; pending-but-blocked checks 30% -> 0.5% -- matching the healthy C profile.

Investigation history: #20.

[dougchansan]

Backend matrix on Colosseum, now that it runs

All five arms on the pinned Phenac City scene (part1-phenac-city.sav), Windows, same runtime build, two runs each. The mtmsr delivery exit was ported locally to the old and AOT emitters for this comparison (their DOLIR_OP_STATE_WRITE handling differs; local branches only) -- without it they still read 0.25 fps.

arm	run 1	run 2	vps
C backend	29.99	29.99	59.9
LLVM old + fix	25.4	27.3	50-54
LLVM new (this base) + fix	25.1-29.0 across runs		50-58
LLVM new + --state-in-memory + fix	29.97	29.97	59.9
LLVM AOT + fix	29.97	29.97	59.9

Reading the numbers honestly: Colosseum is retrace-locked at 30 fps -- the game's frame loop waits for the VI callback -- so fps here measures saturation, not backend speed, and no arm can beat C's 29.99. The signal is which arms hold the cap:

• state-in-memory and AOT hold 59.9 vps / 29.97 fps in every run.
• Plain new and old LLVM run 8-16% under cap. The overhead is measured, not guessed: each EE-enabling mtmsr side exit resumes at pc+4, which is mid-block, and the dispatcher covers the rest of the block (two instructions of OSRestoreInterrupts) through per-instruction interpreter fallback -- ~120M hook-fallback instructions per run, 59M of them the rlwinm/blr pair at 0x8009DF80/84. With state-in-memory the sync cost of those crossings collapses, which is why that arm saturates.

Two attempted optimizations are documented as not viable as-is: saturating the burst budget instead of exiting (deliver at the next guard), and marking pc+4 a region leader so the resume is native. Both build and both wedge Colosseum deterministically ~10-20 s into the scene, in an audio-handler interrupt loop -- delivery placement interacts with the AI/DSP handler's enable-interrupts-around-callback pattern in a way that needs its own investigation before optimizing this path. The shipped fix is the variant that has survived every run, the control titles, and a cold boot.

[dougchansan]

Incorporated upstream: emitStateWrite in 1bec355 implements this exit (msr_ee_exit at pc+4 on an EE 0->1 transition). Closing as superseded. Two follow-ups remain live: ExpansionPak/RecompCore#18 is the required runtime half -- without it the boundary exists but pending external interrupts still wait for slice granularity and Colosseum starves back to 0.25 fps -- and #24 tracks a large throughput regression measured on current main relative to this branch.