From Bugzilla Helper: User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.4.1) Gecko/20031009 Description of problem: This version of libGL contains a buggy assembler routine for _mesa_test_os_sse_exception_support. The end of the assembler routine looks roughly like .... movups (%esp),%xmm1 addl $32,%esp divps %xmm0,%xmm1 ldmxcsr -4(%ebp) leave ret and that "addl $32,%esp" is buggy, since it de-allocates the stack. As a result, the exception that we are testing for that "divps" results in will _overwrite_ the stack location that "ldmxcsr" will restore the XCSR from. So first you get a (correct) SIGFPE from the divps, and when that returns you will get a (deadly) SIGSEGV from ldmxcsr and the program dies. Version-Release number of selected component (if applicable): XFree86-Mesa-libGL-4.3.0-37 How reproducible: Always Steps to Reproduce: 1. Run _just_ the right combination of kernel and stack offsets (which will depend on a lot of things, but 2.6.0-test8 generally shows it) 2. Run any program using libGL (eg "glxinfo") 3. See it segfault Actual Results: Program dies with SIGSEGV Expected Results: Program should just run Additional info: This is most trivially fixed by just _removing_ that one "addl $32,%esp" entirely from the _mesa_test_os_exception_support routine. The "leave" instruction will do the right thing anyway, and the "addl" instruction is simply buggy and has the wrong constant in it. This is also fixed in later versions of DRI CVS (and hopefully later versions of XFree86 too). Which is where I originally debugged this.
Ah yes, that makes a lot of sense. I must admit I haven't given the problem much thought until now, been too busy with other critical stuff for our release, however it's great to get this nasty one squashed. I've pulled the fix out of XFree86 CVS head by diffing the 4.3.0 vs. HEAD versions of that file. The addl is removed now, and should hopefully work. On a side note, what's odd to me is why ENTER/LEAVE is getting used, as that is God aweful slow instructions to use in modern code anywhere. I'm kindof surprised that is there at all. I've filed another bug report to fix that for the future as well. Thanks a bunch for pointing this out Linus. I wouldn't have had the time to allocate to investigate this prior to Fedora Core 1 due to other priorities, so our users will be very greatful to see this fixed. A lot less people will get hit while using 2.6.x. If only there were 36 hours in a day. ;o) Fix applied to 4.3.0-42 in rawhide as: XFree86-4.3.0-Mesa-SSE-fixes-from-MesaCVS-v2.patch
As to why the routine uses enter/leave at all: I suspect the original author really didn't know i386 assembly code, and had just created a C file that he compiled with some awful compile optimization options into assembler. And then he edited the assembler directly. The code is too ugly for words. But at least it should work now.
*** Bug 105581 has been marked as a duplicate of this bug. ***
An errata has been issued which should help the problem described in this bug report. This report is therefore being closed with a resolution of ERRATA. For more information on the solution and/or where to find the updated files, please follow the link below. You may reopen this bug report if the solution does not work for you. http://rhn.redhat.com/errata/RHBA-2003-406.html