Date: Wed, 4 Jul 2012 23:33:16 +0900 From: Taku YAMAMOTO <taku@tackymt.homeip.net> To: freebsd-current@freebsd.org Subject: FYI: SIGBUS with world built by clang Message-ID: <20120704233316.70ec8654.taku@tackymt.homeip.net>
next in thread | raw e-mail | index | archive | help
For people having SIGBUS with clang-build world + gcc-build binaries,
In short words, for any libraries (and never forget about rtld-elf!)
which are potentially called from arbitrary binaries,
compile them with either -mstackrealign or -mstack-alignment=8!
The detail is as follows.
I've observed that clang carelessly expects the stack being aligned at
16 byte boundary.
For example, the following code:
#include <stdarg.h>
#include <stdio.h>
int
foo(const char *format,...)
{
int ret;
va_list ap;
FILE f = {};
va_start(ap, format);
ret = vfprintf(&f, format, ap);
va_end(ap);
return ret;
}
which turns into:
pushl %ebp
movl %esp, %ebp
subl $264, %esp # imm = 0x108
xorps %xmm0, %xmm0
movaps %xmm0, -40(%ebp)
movaps %xmm0, -56(%ebp)
(snip; lots of movaps insns follows)
which results in SIGBUS if %esp - 4 is not at 16 byte boundary.
(Note: movaps expects the address aligned to 16 bytes!)
This problem becomes visible when such functions get called from binaries
compiled with other compilers which don't care about stack alignment.
If the above code is compiled by clang with -mstackrealign:
pushl %ebp
movl %esp, %ebp
andl $-16, %esp
subl $272, %esp # imm = 0x110
xorps %xmm0, %xmm0
movaps %xmm0, 224(%esp)
movaps %xmm0, 208(%esp)
(snip; lots of movaps insns follows)
it tries to align the stack prior to allocating local variables thus no
problem.
If the above code is compiled by clang with -mstack-alignment=8:
pushl %ebp
movl %esp, %ebp
pushl %esi
subl $252, %esp
leal -240(%ebp), %esi
movl %esi, (%esp)
movl $232, 8(%esp)
movl $0, 4(%esp)
calll memset
(snip)
it calls memset instead of a bunch of movaps to clear the storage thus no
problem, of course.
# I don't know why clang doesn't utilize rep stosl, though.
Pros and cons:
-mstackrealign
Pros: no function calls to memset
probably faster because of SSE
Cons: use of SSE means the need of saving FP registers
potentially more stack consumption
-mstack-alignment=#
Pros: normal and predictive stack consumption
don't spill SSE registers; no extra overhead on context switching
Cons: depends on memset
--
-|-__ YAMAMOTO, Taku
| __ < <taku@tackymt.homeip.net>
- A chicken is an egg's way of producing more eggs. -
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20120704233316.70ec8654.taku>