On 1 June 2013 14:06, Jan Hubicka <hubi...@ucw.cz> wrote:
> Hi,
> this patch cleanups way we handle aliases. The main point is to merge code
> that was previously done separately for variables and functions.
>
Hello, This patch appears to break both arm and aarch64. I don't
fully understand the mechanism. The issue shows up when building
res_libc.o over in glibc. The following fragment demonstrates the
issue:
/* aarch64-none-linux-gnueabi-gcc -O2 -c -std=gnu99 -fgnu89-inline r.c */
struct __res_state {
int retrans;
};
extern __thread struct __res_state *__libc_resp __attribute__
((tls_model ("initial-exec")));
struct __res_state _res __attribute__((section (".bss")));
__thread struct __res_state *__resp = &_res;
extern __thread struct __res_state *__libc_resp __attribute__ ((alias
("__resp"))) ;
int
__res_init(void) {
(*__libc_resp).retrans = 5;
}
We seem to lose track of the alias between _resp and __llbc_resp
resulting in the following output:
.size __res_init, .-__res_init
.global __resp
.global __libc_resp
__libc_resp = __resp
.global _res
.section .bss,"aw",%nobits
.align 3
.type _res, %object
.size _res, 4
_res:
.zero 4
.section .tbss,"awT",%nobits
.align 3
.LANCHOR0 = . + 0
.type __libc_resp, %object
.size __libc_resp, 8
__libc_resp:
.zero 8
.section .tdata,"awT",%progbits
.align 3
.type __resp, %object
.size __resp, 8
__resp:
.xword _res
Note the definition of both __libc_resp and __resp, where previously
we would have defined only __resp.
I suspect the section anchor appeared because
varasm.c:use_blocks_for_decl_p test for an 'alias' failed.
/Marcus
