Midnight Club 3 - JIT cache getting cleared repeatedly

[hrydgard]

About every four seconds When a rewind snapshot is taken and rounding modes are used, the jit cache gets cleared. Looking at the log, it looks really silly:

04-15 23:37:57.059 15248 15492 W PPSSPP  : [JIT] Detected rounding mode usage, rebuilding jit with checks
04-15 23:37:57.059 15248 15492 I PPSSPP  : [JIT] ARM64Jit: Clearing the cache!
04-15 23:38:01.081 15248 15492 W PPSSPP  : [JIT] Detected rounding mode usage, rebuilding jit with checks
04-15 23:38:01.081 15248 15492 I PPSSPP  : [JIT] ARM64Jit: Clearing the cache!
04-15 23:38:05.374 15248 15492 W PPSSPP  : [JIT] Detected rounding mode usage, rebuilding jit with checks
04-15 23:38:05.374 15248 15492 I PPSSPP  : [JIT] ARM64Jit: Clearing the cache!
04-15 23:38:09.111 15248 15492 W PPSSPP  : [JIT] Detected rounding mode usage, rebuilding jit with checks
04-15 23:38:09.111 15248 15492 I PPSSPP  : [JIT] ARM64Jit: Clearing the cache!
04-15 23:38:13.133 15248 15492 W PPSSPP  : [JIT] Detected rounding mode usage, rebuilding jit with checks
04-15 23:38:13.133 15248 15492 I PPSSPP  : [JIT] ARM64Jit: Clearing the cache!

• We should probably remember the rounding mode state across jit clears within a game, right?

[hrydgard]

Actually this is caused by rewind snapshots, we must be doing something bad.

[unknownbrackets]

This doesn't look right:

	Do(p, js.startDefaultPrefix);
	if (s >= 2) {
		Do(p, js.hasSetRounding);
		js.lastSetRounding = 0;
	} else {
		js.hasSetRounding = 1;
	}

-[Unknown]

[hrydgard]

Ah, we had already fixed that on x86 it seems, assuming this is right:

	Do(p, js.startDefaultPrefix);
	if (s >= 2) {
		Do(p, js.hasSetRounding);
		if (p.mode == PointerWrap::MODE_READ) {
			js.lastSetRounding = 0;
		}
	} else {
		js.hasSetRounding = 1;
	}

[unknownbrackets]

I guess that was only fixed in 718cb9e.

-[Unknown]

[hrydgard]

Yup, oversight by me.

[hrydgard]

Maybe should create a jitcommon base class, instead of just having those pure interfaces - this function can be the same in all of them. But I don't think I wanna do that right now, so I'll just submit a PR that fixes them all individually.