Emulator

Please provide a memory card with a save just outside of the city. Confirming this bug - I've seen a lot of reports on the forums.

Could you try this build and tell me if the crashes still happen?

http://dl.dolphin-emu.org/wips/delroth-dolphin-master-3.5-135%2b-x64.7z

Ok, this is only the first step towards a fix though: the current patch removes an optimization in the JIT which causes that issue.

The issue, as I understand it:

• The game executes a loop for (i = 0; i < 0x1000; ++i) { ... } which does not modify i at all.
• The loop condition is translated to "addi , , 1; ...; cmpwi , 0x1000; ...; bne loop_again". Note the "...": there are instructions between the cmpwi and the bne.
• PPCAnalyst moves the cmpwi to just before the bne because this is the first instruction that depends on the flags or modify i.
• One of the instructions between cmpwi and bne raises an interrupt (GP write -> FIFO interrupt, I guess). In that case, the cmpwi was not executed because it was moved.
• When we come back to the loop after the interrupt occurred, we are in a new block starting after the cmpwi. That means the cmpwi will not be executed at all.

Knowing that, some ideas to fix this:

• Do not move a cmp after an instruction that might raise an interrupt. Might be hard to detect, especially when MMU is enabled and DSI can happen at every load. This will also make this optimization less effective.
• When we raise an interrupt, if there is a cmp pending (how to detect that?) execute it before interrupting.
• When we moved a PPC instruction, memorize that and insert the moved instruction to every other PPC block that might need it. This will force us to handle tricky cases due to possibly jumping inside a block (in this case we don't want a cmp from before our jump addr to be inserted).

Fix 1 is the easiest to implement. Fix 2 could maybe work but I'm not sure how easy it is to implement (maybe put that in CheckExceptions?). Fix 3 could work too, will keep all the performance wins of moving cmp but might be hard to implement due to tricky cases.

Cc-ing skid, I'd like to have your opinion on that issue.

A smarter fix seems very difficult to implement correctly for a very small performance improvement (I could not measure a performance decrease with my workaround). Moreover, this workaround fixed more games than expected: at least MMU games crashing at boot before this fix are now booting properly (yay!).

Closing as fixed.