| Summary: | gtk-demo crash in "Color Selector" + Solution | ||
|---|---|---|---|
| Product: | pixman | Reporter: | Ronald Wahl <rwahl> |
| Component: | pixman | Assignee: | Carl Worth <cworth> |
| Status: | RESOLVED FIXED | QA Contact: | cairo-bugs mailing list <cairo-bugs> |
| Severity: | normal | ||
| Priority: | high | CC: | dberkholz |
| Version: | 0.1.3 | ||
| Hardware: | x86 (IA32) | ||
| OS: | Linux (All) | ||
| Whiteboard: | |||
| i915 platform: | i915 features: | ||
| Attachments: |
Fix for segfault
Fix segfault (updated - missed %esi in clobber list) Patch applied to libpixman |
||
|
Description
Ronald Wahl
2005-08-27 08:37:01 UTC
Created attachment 3061 [details] [review] Fix for segfault Created attachment 3062 [details] [review] Fix segfault (updated - missed %esi in clobber list) I had trouble understanding the description provided, but looking at
the patch, I think the idea of it is:
While generally gcc will try to avoid reusing %ebx (since it is
the "pic" register and used as the base offset for accessing global
data), with some combinations of of compiler flags it's possible
that it *will* be used for one of the destination arguments, so
doing:
"mov %%ebx, %1\n"
"mov %%edx, %2\n"
"mov %%ecx, %3\n"
[...]
"mov %%edx, %0\n"
doesn't work. The obvious fix would be to mark %ebx as clobbered,
but that doesn't work either since because %ebx is the pic register,
it's not allowed to be clobbered. So, the patch is careful to restore
ebx at the points where the destination registeres are accessed.
The patch looks OK to me, though it seems to me that %eax could have been
used rather than adding %esi as another clobber. (While it's unlikely
here, if you have too many clobbers, it's conceivable that the compiler
will be unable to satisfy the constraints.)
If you have no frame pointer in ebp then the compiler uses esp to access local variables. Since esp is modified behind compilers back by pushing ebx on the stack the compiler will fail to access the right storage place of the local variable(s). Using of eax instead of esi is ok. Created attachment 3069 [details] [review] Patch applied to libpixman I'm applying the patch (with a comment added and the esi => eax change) to libpixman. Billy - can you take care of making sure the fix is propagated as appropriate in case someone wants to try and compile their X server with -fomit-frmae-pointer? Fix applied to xserver, I'll add it to the list that should go to Xorg. |
Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.