Forwarded conversation
Subject: [developer] SUNWspro compiler bug
------------------------
From: Matthew Ahrens <mahr...@delphix.com>
Date: 11 September 2012 06:35
To: develo...@lists.illumos.org
Is there a good place to report a SUNWspro compiler bug?
I've found that functions that use alloca() (or c99 variable-length
arrays) and are compiled with "-m64 -xO1 -Wu,-save_args" (as the
kernel is) restore callee-saved registers incorrectly.
Note that extant code works correctly because the kernel doesn't call
alloca(), and 64-bit userland is not compiled with save_args.
The incorrect assembly looks like:
func: pushq %rbp
func+1: movq %rsp,%rbp
# save arguments:
func+4: subq $0x18,%rsp
func+8: movq %rdi,-0x8(%rbp)
func+0xc: movq %rsi,-0x10(%rbp)
func+0x10: movq %rdx,-0x18(%rbp)
# callee-saved registers which this function needs:
func+0x14: pushq %r12
func+0x16: pushq %r13
func+0x18: pushq %r14
func+0x1a: pushq %r15
# alloca(1):
func+0x1c: subq $0x8,%rsp
func+0x20: decq %rsp
func+0x23: andq $0xfffffffffffffff0,%rsp
... some computation that uses %r12, %r13, %r14, %r15
# incorrectly change %rsp to what it was before the alloca()
# should be -0x38 to include saved arguments
func+0x187: leaq -0x20(%rbp),%rsp
# restore callee-saved registers:
func+0x18b: popq %r15
func+0x18d: popq %r14
func+0x18f: popq %r13
func+0x191: popq %r12
func+0x193: leave
func+0x194: ret
For an example, see the attached test program.
--matt
illumos-developer | Archives | Modify Your Subscription
----------
From: Richard Lowe <richl...@richlowe.net>
Date: 11 September 2012 06:40
To: Matthew Ahrens <mahr...@delphix.com>, illumos-dev
<develo...@lists.illumos.org>
On Tue, Sep 11, 2012 at 12:35 AM, Matthew Ahrens <mahr...@delphix.com> wrote:
>
> Is there a good place to report a SUNWspro compiler bug?
>
> I've found that functions that use alloca() (or c99 variable-length arrays)
> and are compiled with "-m64 -xO1 -Wu,-save_args" (as the kernel is) restore
> callee-saved registers incorrectly.
I hit this same issue when concocting the libproc changes I just sent
out for review. I think you can file Studio bugs at bugs.sun.com
still, but it won't be any good for illumos, since it'll only be fixed
in a version of the compiler we can't really use.
-- Rich
illumos-developer | Archives | Modify Your Subscription
Ced
--
Cedric Blancher <cedric.blanc...@googlemail.com>
Institute Pasteur
/*
* fails when compile with:
* /opt/SUNWspro/bin/cc -m64 -xO1 -Wu,-save_args bad.c
* succeeds when compiled with:
* /usr/sfw/bin/gcc -m64 -O2 -msave-args bad.c
*/
#include <sys/types.h>
#include <stdlib.h>
#include <alloca.h>
#include <stdio.h>
int
func(void *a, void *b, void *c)
{
/* fails with either of these */
#if 0
static const int buflen = 1;
char buf[buflen];
#else
alloca(1);
#endif
/*
* This computation will require several registers, thus forcing
* some registers to be saved when the function is entered.
*/
uint64_t x = (uint64_t)a + 1;
uint64_t y = (uint64_t)b + 1;
uint64_t z = (uint64_t)c + 1;
uint64_t tmp =
((((x*y + y*z) * (x*x + y*y)) /
(1+(z*z + x*z) * (z*x + y*x))) %
(1+((x*x + y*y) * (z*z + x*y)) /
(1+(z*y + x*x) * (z*z + y*z)))) +
(1+(((x*x + y*y) * (y*x + z*y)) /
(1+(z*x + x*x) * (z*z + y*z))) %
(1+((x*y + x*y) * (x*z + y*y)) /
(1+(y*y + x*y) * (z*x + z*z))))
;
snprintf(NULL, 0, "%llx\n", tmp);
printf("%p %p %p\n", a, b, c);
return (0);
}
int main(int argc, char *argv[])
{
printf("%p %p %p\n", argv[0], argv[1], argv[2]);
func(argv[0], argv[1], argv[2]);
printf("%p %p %p\n", argv[0], argv[1], argv[2]);
return (0);
}
_______________________________________________
opensolaris-discuss mailing list
opensolaris-discuss@opensolaris.org
