Project

General

Profile

Actions

Emulator Issues #917

closed

JIT compiler crashes

Added by Line524 over 15 years ago.

Status:
Fixed
Priority:
Normal
Assignee:
% Done:

0%

Operating system:
N/A
Issue type:
Bug
Milestone:
Regression:
No
Relates to usability:
No
Relates to performance:
No
Easy:
No
Relates to maintainability:
No
Regression start:
Fixed in:

Description

What steps will reproduce the problem?
1.Accessing inventory in Resident Evil 4 and trying to leave it
2.Crashes occasionally show up in Zelda Wind Waker in places like the
Tower of the Gods.
3.Additional info below (looks helpfull):

entry__13mDoExt_bpkAnmFP12J3DModelDataf
8000dff0: stwu sp, -0x0010 (sp)
8000dff4: mflr r0
8000dff8: stw r0, 0x0014 (sp)
8000dffc: addi r4, r4, 88
8000e000: bl ->0x8000E2F4
8000e004: lwz r0, 0x0014 (sp)
8000e008: mtlr r0
8000e00c: addi sp, sp, 16
8000e010: blr

It's a great function, really. But for some completely inexplicable
reason, by the time the game is running, it has turned into this:

entry__13mDoExt_bpkAnmFP12J3DModelDataf
8000dff0: stwu sp, -0x0010 (sp)
8000dff4: mflr r0
8000dff8: stw r0, 0x0014 (sp)
8000dffc: addi r4, r4, 88
8000e000: b ->0x80011F80
8000e004: JITblock 16256
8000e008: JITblock 934
8000e00c: addi sp, sp, 16
8000e010: blr

Note in particular the fact that a bl instruction to one address has
turned into a b instruction from an entirely different address.

In the original code, the hex values for the three modified instructions
are as follows:

480002f5
80010014
7c0803a6

I have absolutely no idea why this is happening, but I do know this: in
the corrupted code, the three instructions that have changed have the
following hex values:

48003f80
00003f80
000003a6

It would appear that 8 bytes of instructions have been overwritten with
3f8000003f800000.

No idea why this is, but it's not a coincidence, I know that much.
The exact same thing happens in Resident Evil 4 on a branch were the
result is always corrupted using JIT core, I haven't been able to find why
it fails though.
In Resident Evil 4 the crashes only usually occur when accessing the
inventory, map, treasure and file screens. In RE4 you'll be able to access
these screens only once, trying to access these a second time will crash
the game and emulator.

So is the Windwaker and RE4 issues similar, in that if one is fixed, it'll
fix the other. I don't get why RE4 only crashes when accessing these
screens but normal ingame play is fine.

What version of the product are you using? On what operating system?
Every Dolphin version gives this on any system. Deactivating JIT compiler
fix this, but playing on interpreter is too slow (2 or 3 fps.)


Related issues 1 (0 open1 closed)

Has duplicate Emulator - Emulator Issues #903: Memory corruption in Zelda: Wind WakerDuplicate

Actions
Actions #1

Updated by marcus over 15 years ago

  • Status changed from New to Accepted

I'll bet it'll fix issue 659, too.

Actions #2

Updated by BhaaL over 15 years ago

Some random thought about that:

"To me, it looks like either of the instructions before/after the corrupted block
operate on the wrong size of operands or something like that..."

Actions #3

Updated by marcus over 15 years ago

Issue 903 has been merged into this issue.

Actions #4

Updated by maroofsbp over 15 years ago

i think if this fix than it will fix most of the game

Actions #5

Updated by boulliastation over 15 years ago

This also happens for me in the Magmoor Caverns in Metroid Prime 1, in that room with
the single Magmoor and the morph ball rails, just before the elevator to Tallon
Overworld.

Actions #6

Updated by LinesPrower over 15 years ago

Is this happens in the IL version too?

Actions #7

Updated by sl1nk3.s over 15 years ago

It happens with the IL version too, here's where it crashed :

80a68b84: stwu sp, -0x0008 (sp)
80a68b88: mflr r0
80a68b8c: stw r0, 0x000C (sp)
80a68b90: li r0, 0
80a68b94: ori r0, r0, 0xFFFF
80a68b98: cmpw r4, r0
80a68b9c: bne- ->0x80A68BE4
80a68ba0: cmpwi r3, 0
80a68ba4: beq- ->0x80A68BC4
80a68ba8: JITblock 32969
80a68bac: subi r3, r3, 608
80a68bb0: bl ->0x8012AA94
80a68bb4: JITblock 32969
80a68bb8: subi r3, r3, 580
80a68bbc: bl ->0x8012ACB0
80a68bc0: b ->0x80A68BE4
80a68bc4: JITblock 32937
80a68bc8: lis r11, 0x80A9
80a68bcc: subi r9, r9, 24400
80a68bd0: subi r11, r11, 24472
80a68bd4: lis r10, 0x80C9
80a68bd8: lis r8, 0x80C9
80a68bdc: stw r9, -0x022C (r10)
80a68be0: stw r11, -0x0248 (r8)
80a68be4: lwz r0, 0x000C (sp)
80a68be8: mtlr r0
80a68bec: addi sp, sp, 8
80a68bf0: blr

still those JITblock and those branches :/

Actions #8

Updated by BhaaL over 15 years ago

Since it affects both JIT, but not the Interpreter, I'd guess that something in
JitCommon might be the problem (from the name, both JIT and JITIL use it, right?)

And as the original poster wrote, I'd say RE4 is a good game to test this bug.
According to him, it always crashes when opening the menu/inventory/whatever for the
second time, hence the first time will perform the JITting and thus break the code.
I dont have RE4, so I can't try it out.

Interpreter_Branch looks a bit easier and shorter to me than the JIT_Branch ones of
both JIT and JITIL. As earlier mentioned, it looks like something in there operates
on the wrong operand size (most probably the bl) and corrupts the memory afterwards.
Hard to say from just looking at the code, not running anything, and (besides) my
limited knowledge of the PPC Instruction set - just my 2ct here.

Actions #9

Updated by Line524 over 15 years ago

Yeah Jack, you right ! So it will be possible to fix if you'll get RE4 game ?:)
I think RE4 is the best (or one of the best) game for Dolphin. PC version was badly
ported from PS2 version :( And PS2 version have reduced gfx detail level... So GC
version is original (and BEST !).

Actions #10

Updated by sl1nk3.s over 15 years ago

btw, i tried to change all the JIT_Branch functions to use the interpreter, and it
crashed too, not at the exact same moment though.
I mean originally it crashes when you open the inventory for the second time, when
using Interpreter_Branch it crashed for the first time i opened it.
So, my guess is that JIT_Branch could be actually fine and that the issue is elsewhere :(

Actions #11

Updated by BhaaL over 15 years ago

Well, as said, its just a vague idea of where it could be, but since sl1nk3 has
proven me wrong... ;)

Gonna see if I can borrow RE4 somewhere (I'd bet my ass that one of my colleagues got
it), but I'll probably have a hard time dumping it without my Wii.
Besides, I never really liked/cared about RE, so chances of that might still be low :P

Another possible thing would be Wind Waker with the Tower of Gods (where I
experienced the crash a lot of times, pretty much reliably), if I can find my save file.

But since sl1nk3 pointed out that switching to Interpreter also crashes, it might be
a general problem with the PPC instructions (as earlier mentioned, wrong oeprand
size, just with all of the implementations?)
I'll try that out if I can pile up some spare time, painting my room and stuff the
next few days.

Actions #12

Updated by sl1nk3.s over 15 years ago

err, well completly switching to interpreter works, it won't crash, that's just i
though it was the JIT_Branch but as switching to Interpreter_Branch still crash, it
may not be the branch code (don't know if that's what you said ? o.o)

So err, no idea how to debug this actually, but if you want to give it a try and
reproduce the crash, RE4 is still the easiest one to try, just open the inventory two
times, i don't really know about Wind Waker though :)

Actions #13

Updated by BhaaL over 15 years ago

Odd, tho it was just a random guess of mine at all.
Wasnt it too much of a coincidence that the bl stays in-place partially (becoming a
b) and corrupting the stuff right after it?
The same for your examples, sl1nk3, the JITblocks occur right after a branch operation.
Btw, do you think you could grab that corrupted snippet of yours and post the initial
state?
Debugging it is one thing, but when you don't have a clue about where it happens, its
pretty much a dead end.

As for Wind Waker, it might be easier than RE4 at all.
Starting a savegame seems to perform the stuff that breaks the function at 8000dff0
as shown in the first post.
Gonna dig around that a bit.

Actions #14

Updated by BhaaL over 15 years ago

Hm, odd. I used some PanicAlerts to find the places where that location (0x8000e000)
is used, yet the Panics seemed to hit false positives.
Instructions and addresses werent the same as if used in the (positively triggered)
condition, and what-not.

Figured I might switch to SC mode, then this happened:
Heap corruption detected at 0CBBD0F8
Heap corruption detected at 0CBBD0F8
HEAP[DolphinDF.exe]: HEAP: Free Heap block cbbd0f0 modified at cbbd100 after it was freed
Windows has triggered a breakpoint in DolphinDF.exe.

Either way, could anyone point me to the place where the opcodes are stored that are
executed and shown in the debugger? I couldn't seem to pinpoint the location where
that branch gets corrupted.

Actions #15

Updated by malo.sasha over 15 years ago

Did you try to use HLE SOUND Plugin with audio backend = null . I had Zelda WW
crashing in some places until i choose that option for sound

Actions #16

Updated by Line524 over 15 years ago

So sound plugin could really be the reason of this problem ?? I would like to help
Jack.Frost, but I'm not a coder :( And I would test Null sound mode, but someone cut
my power supply =(

Actions #17

Updated by BhaaL over 15 years ago

I'm using HLE with all sounds disabled.
I don't think its sound-related, altho it is as much of a possibility as all the
other plugins are.

Still stuck at the point about finding out where the operations are stored; i'm
planning to put a memory breakpoint on the 0x8000e000 one, and see where it hits.

Actions #18

Updated by BhaaL over 15 years ago

Hm, interresting, the interpreter mode doesnt seem to fix this either.

  1. Run Zelda:WW until the load savegame screen.
  2. Pause and switch to interpreter mode, resume.
  3. Load the game.
    The memory around 0x8000e000 still gets corrupted...
Actions #19

Updated by BhaaL over 15 years ago

Seems only to occur when the savegame is after the tower of gods, see attached GCI
for a working (breaking?) savegame.

Actions #20

Updated by hrydgard over 15 years ago

The code gets overwritten by this guy:
8029b00c stfs f1, 0 (r3)

For some reason, r3 is 0x0000e002. So, boom.

Have fun debugging! :p

Actions #21

Updated by kairi.streetfighter over 15 years ago

This issue occurs with Capcom vs. SNK 2 EO as well. The boss battle in Arcade mode or
a simple matchup in VS or Training mode, anyone of these can't be done twice because
of this crash.

Actions #22

Updated by xatnys over 15 years ago

Has progress been made regarding this issue?

Actions #23

Updated by LinesPrower about 15 years ago

  • Status changed from Accepted to Work started
Actions #24

Updated by LinesPrower about 15 years ago

  • Status changed from Work started to Fixed

This issue was closed by revision r4357. Please verify

Actions

Also available in: Atom PDF