StaticRecomp: skip guest busy-wait loops the analyzer can identify

[dougchansan]

The dispatcher already returns the rest of its slice to CoreTiming when the guest reaches MAIN_STATICRECOMP_IDLE_PC, but that only covers loops configured ahead of time. Every other idle loop is spun through in real time.

A dispatch that returns to the address it entered at is a self-loop. Running PPCAnalyst over that address tells us whether it is an idle loop (a branch with branchIsIdleLoop back to the entry); if it is, there is nothing to gain from executing it until the next event.

• result cached per address, invalidated wherever guest code can change — ClearCache and SMC re-verification
• skipped entirely when a debugger or branch watch is active, since both want every block to actually execute
• 44 lines across 4 files, no new dependencies (analyzer, code_block, m_code_buffer all come from JitBase)

Measurements

x86-64, static recompilation active, frames advanced in a fixed 20s window. The same binary is used for both arms, with detection toggled at runtime, so nothing but this behaviour differs.

title / scene	detection on	off	ratio	t	95% CI
Skyward Sword — gameplay	796.2 (n=6)	571.8	1.392	10.47	1.317–1.467
Luigi's Mansion — foyer	751.9 (n=8)	672.1	1.119	3.58	1.052–1.185
Mario Kart DD — bench	1033.5 (n=6)	965.3	1.071	2.64	1.017–1.124
Pokémon Colosseum — 2 scenes	—	—	0.98–1.01	—	neutral

No title regressed. On Skyward Sword the arms do not overlap at all: the worst detection-on run (730 frames) still beats the best off run (606).

Why these numbers can be trusted

The noise floor was measured rather than assumed, by supplying the same module as both control and arm so the true ratio is exactly 1.0000:

n=13 null comparisons across 4 titles / 6 scenes
12 of 13 within +/-1.72%
 1 of 13 at +14.97%  (one scene; 3 repeats of it then gave 1.0164 / 0.9972 / 1.0019)

Single runs therefore carry a rare ~15% excursion, and every figure above is a mean over repeated runs rather than one measurement.

Three confounds had to be controlled before the effect was measurable at all, each of which produced plausible-looking but wrong numbers:

• host input reaching the guest. With no device bound and BackgroundInput = False the detection-on arm spreads 651–810 frames; without that pinning the same arm spread 829–7157, an 8.6x range.
• Wii extension config. A Wii title reads its extension from WiimoteNew.ini, not from the savestate, so omitting Extension = Nunchuk leaves Skyward Sword sitting on a "connect the Nunchuk" prompt. That measured +119% and was an artifact of the dialog, not of this change.
• one emulator at a time. A second concurrent instance compressed the same Skyward Sword effect from +39% to +30%.

Honest characterisation

The win is opportunistic: detection only fires when a dispatch happens to re-enter at the loop head. That shows up as higher variance in the detection-on arm (sd 46–60, against 18–25 with it off). It is a consistent gain, not a uniform one, and Colosseum shows it can also be simply neutral.

[dougchansan]

Note on the measurement environment, since it came up while working on ModernGekko #34.

Every number in this PR was taken with MODERNGEKKO_CACHE_AFFINITY unset, which on a 9950X3D leaves the emulated CPU thread free to migrate off the V-Cache die. Enabling it is worth about +25% in absolute speed on that machine.

The ratios here are unaffected. I checked this directly rather than assuming it: pinning lifts the C and LLVM arms by the same amount (+13.2% each on Colosseum, n=6 per cell), and the arm/control ratio moves +0.6%, which is noise. Both arms carried the same handicap, so the comparisons stand exactly as reported.

Two things worth knowing for anyone reproducing these figures:

• Absolute speeds understate the machine. Re-running with pinning on will produce higher numbers in both arms, not a different conclusion.
• Pinning also cuts run-to-run spread sharply — one Colosseum scene went from ±0.297 to ±0.015 standard error over three samples. That makes replication cheaper, so I would recommend it for any re-measurement even though it does not change the answer.