Open ethanc8 opened 4 months ago
The original CAAnimation.m
gets CAAction compiled to:
$._OBJC_PROTOCOL_CAAction = comdat any
// a few hundred lines later
@52 = private unnamed_addr constant [9 x i8] c"CAAction\00", align 1
@._OBJC_PROTOCOL_CAAction = global { ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr } { ptr inttoptr (i32 4 to ptr), ptr @52, ptr @.objc_protocol_list.10, ptr @.objc_protocol_method_list.11, ptr @.objc_protocol_method_list.13, ptr @.objc_protocol_method_list.12, ptr @.objc_protocol_method_list.14, ptr null, ptr null, ptr null, ptr null }, section "__objc_protocols", comdat, align 8
// a few lines later
@.objc_protocol_list.20 = internal global { ptr, i64, [4 x ptr] } { ptr null, i64 4, [4 x ptr] [ptr @._OBJC_PROTOCOL_NSCoding, ptr @._OBJC_PROTOCOL_NSCopying, ptr @._OBJC_PROTOCOL_CAAction, ptr @._OBJC_PROTOCOL_CAMediaTiming] }, align 8
I've attached the full LLVM IR to https://gist.github.com/ethanc8/d02ef13441e1eb58d372cf650fa3f63b
That looks correct. The isa
pointer is that ptr inttoptr (i32 4 to ptr)
one: we're initialising it to 4 in the compiler, so it should be a GNUstepV2 protocol. So now the question is: where does that become 0? Can you put a watchpoint on the load address and see if there are any writes to it?
Can you put a watchpoint on the load address and see if there are any writes to it?
Which load address are you talking about?
Ok, I set a watchpoint at 0x7ffff6ab3a60
(the address of CAAction
).
Hardware watchpoint 1: *(int*)0x7ffff6ab3a60
Old value = 4
New value = 0
registerProtocol (proto=0x7ffff6ab3a60 <._OBJC_PROTOCOL_CAAction>) at /home/ethan/Projects/GNUstep/plaurent2/gnustep/libobjc2/protocol.c:608
608 if (protocol_for_name(proto->name) == NULL)
(gdb)
If I try to forcefully reset it back to 4, it does:
(gdb) set var proto->isa=0x4
(gdb) continue
Continuing.
Hardware watchpoint 1: *(int*)0x7ffff6ab3a60
Old value = 4
New value = -143207376
init_protocols (protocols=0x7ffff6aaa030 <objc_protocol_list>) at /home/ethan/Projects/GNUstep/plaurent2/gnustep/libobjc2/protocol.c:245
245 if (NULL != aProto->protocol_list)
(gdb)
and then fails with Unknown protocol version
.
Okay, this looks like a load-order problem. The line above the one in your gdb session is proto->isa = protocol_class_gsv2;
. If we haven’t yet loaded the protocol classes, this will be null, and that seems to be what you’re seeing.
Most of the paths that should be here are guarded on having loaded the runtime. Can you print a backtrace there? I suspect the simplest thing to do is treat the built-in classes as special and define them in C rather than relying on the normal load sequence.
Does the library defining this protocol explicitly link libobjc2? If it does, we should see the libobjc2 constructors run first (I think).
Here's the backtrace:
Hardware watchpoint 1: *(int*)0x7ffff6ab3a60
Old value = 4
New value = 0
registerProtocol (proto=0x7ffff6ab3a60 <._OBJC_PROTOCOL_CAAction>) at /home/ethan/Projects/GNUstep/plaurent2/gnustep/libobjc2/protocol.c:608
608 if (protocol_for_name(proto->name) == NULL)
(gdb) bt
#0 registerProtocol (proto=0x7ffff6ab3a60 <._OBJC_PROTOCOL_CAAction>) at /home/ethan/Projects/GNUstep/plaurent2/gnustep/libobjc2/protocol.c:608
#1 0x00007ffff774b1a5 in __objc_load (init=0x7ffff6aa9638 <objc_init>) at /home/ethan/Projects/GNUstep/plaurent2/gnustep/libobjc2/loader.c:241
#2 0x00007ffff7fc947e in call_init (l=<optimized out>, argc=argc@entry=1, argv=argv@entry=0x7fffffffd398, env=env@entry=0x7fffffffd3a8) at ./elf/dl-init.c:70
#3 0x00007ffff7fc9568 in call_init (env=0x7fffffffd3a8, argv=0x7fffffffd398, argc=1, l=<optimized out>) at ./elf/dl-init.c:33
#4 _dl_init (main_map=0x7ffff7ffe2e0, argc=1, argv=0x7fffffffd398, env=0x7fffffffd3a8) at ./elf/dl-init.c:117
#5 0x00007ffff7fe32ca in _dl_start_user () from /lib64/ld-linux-x86-64.so.2
#6 0x0000000000000001 in ?? ()
#7 0x00007fffffffd88b in ?? ()
#8 0x0000000000000000 in ?? ()
(gdb) p protocol_class_gsv2
$1 = (id) 0x0
All of the libraries and apps should be linking libobjc2:
$ gnustep-config --objc-libs -ldispatch -pthread -fexceptions -rdynamic -fobjc-runtime=gnustep-2.1 -fblocks -L/home/ethan/GNUstep/Library/Libraries -L/usr/GNUstep/Local/Library/Libraries -L/usr/GNUstep/System/Library/Libraries -lpthread -lobjc -lm
Please can you test #294 and see if this fixes it for you?
Please can you test #294 and see if this fixes it for you?
I will try to do so tonight, but I'm a bit busy today.
Please can you test #294 and see if this fixes it for you?
Yes, it seems to be working now. Thanks! Do you want me to confirm anything in the debugger for you? I have just noticed that it no longer aborts with "Unknown protocol version".
No, that’s great. Thanks.
We still need to figure out what caused the segfault in SparseArrayLookup.
Hello đź‘‹, I have encountered a segfault in
SparseArrayLookup
during the initialization of my port of the application GitY. Here is the backtrace from GDB:Interestingly, I have not encountered this when running any of the libs-opal tests, such as
images
. I have confirmed with GDB thatimages
calls+[CGImageDestinationTIFF load]
without segfaulting. Therefore, I don't know in which component this came from.The Objective-C components that are linked to by
GitY
are:It's possible that one of those has messed something up in their
+load
methods, but I'm not exactly sure. If needed, I can send you binaries of my GNUstep installation, compilelibobjc2
with different flags, peek around in GDB, or whatever else would be needed.