summaryrefslogtreecommitdiff
path: root/libc/sysdeps/linux/i386/bits/endian.h
diff options
context:
space:
mode:
authorEric Andersen <andersen@codepoet.org>2001-10-05 11:31:48 +0000
committerEric Andersen <andersen@codepoet.org>2001-10-05 11:31:48 +0000
commit8a309c2fde98f9fcba538fcada54248eafdd34ad (patch)
tree7edbb9f5153b509f0e1604abc266ac83f738d85f /libc/sysdeps/linux/i386/bits/endian.h
parentab79ee6808a768bc72cd1158f93ba8a50a0615e5 (diff)
Wohoo! David McCullough found the bug! His comments follow:
I had a look at it and you won't believe it was always broken. I'll try and explain it, let me know if it doesn't make sense. * ash calls setjmp, which messes with the stack to look like it has two args instead of one and then jmps (actually falls) into sigsetjmp. BUG * sigsetjmp then saves the registers and "jumps" to __sigset_save, a C function. BUG1 - because the caller pops its args off the stack, a program that changes it's number of args is broken because the caller will not pop the correct number of args. I think that jumping from the sigsetjmp asm to the 'C' code is unsafe but I can't think of an example. Anyway, I have attached what I think is a working fix. The reason this worked without -fomit-frame-pointer is that the _sigset_save 'C' code would restore the stack pointer from %ebp (the frame pointer) and because none of the asm had moded it, when we returned from __sigset_save the stack was back to it's correct position for 1 argument despite our best attempts to stuff it up ;-)
Diffstat (limited to 'libc/sysdeps/linux/i386/bits/endian.h')
0 files changed, 0 insertions, 0 deletions