Ok, I've checked, and in -current, userland stack boundary alignment
is useless because the stack pointer is initially only aligned on a
word-boundary.
I've also verified that the correct alignment is indeed expected to
apply to frame pointers, i.e. stored frame pointers are at aligned
stack slots and the current frame pointer points to an aligned
address.
Alignment can be verified by doing something like this:
struct { int a[4]; } x __attribute__((aligned(16)));
printf("offset: %u\n", (unsigned int)&x & 15);
You have to run the program using different arguments and argument
lengths to ensure that what you get isn't coincidental. When things
are working properly, the above should always print 0.
Note that double-alignment vs. word-alignment can really have >30%
performance impact, at least on an Athlon and one meaningless floating
point microbenchmark (operations on small, fixed-sized
matrices...maybe it isn't even *that* meaningless).
Here's a patch against -current that fixes this problem (note that
this doesn't maintain alignment for constructors, and it isn't very
pretty because it has to be done all in asm...):
Index: crt1.c
===================================================================
RCS file: /m/cvs/freebsd/src/lib/csu/i386-elf/crt1.c,v
retrieving revision 1.4
diff -u -r1.4 crt1.c
--- crt1.c 1999/08/27 23:57:57 1.4
+++ crt1.c 1999/12/02 09:02:05
@@ -92,7 +92,17 @@
monstartup(&eprol, &etext);
#endif
_init();
- exit( main(argc, argv, env) );
+ asm volatile("andl $~15,%%esp;"
+ "addl $4,%%esp;"
+ "pushl %2;"
+ "pushl %1;"
+ "pushl %0;"
+ "call main;"
+ "movl %%eax,(%%esp);"
+ "call exit"
+ : : "rm" (argc), "rm" (argv), "rm" (env));
+ for (;;)
+ ;
}
#ifdef GCRT
To Unsubscribe: send mail to [EMAIL PROTECTED]
with "unsubscribe freebsd-current" in the body of the message
