https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92151
--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Alexander Monakov from comment #3)
>
> Here's a manually reduced variant that exhibits a similar issue at -O1:
>
> void foo(int num, int c) {
> asm("# %0" : "+r"(num));
> while (--c)
> asm goto("# %0" :: "r"(num) :: l2);
> l2:
> asm("# %0" :: "r"(num));
> }
Note the DEAD_REG is updated in IRA anyways so that is not the issue.
This reduced testcase seems to be fixed in GCC 11+, maybe r11-209
(https://gcc.gnu.org/pipermail/gcc-patches/2020-May/545442.html).
The original testcase did fix on mov issue.
But
__asm__("\t mov %q[w], (%q[d])\n" // write
: "=m" (*d) : [d] "r" (d), [w] "r" (w)) ;
__asm__("\t lea 8(%q[d]), %q[d]\n"
: [d] "+r" (d)) ;
Still causes an extra move to happen.
Note there was one improvement in GCC 11 in the original testcase too.
I think TWITCH difference is due to the way register variables are represented
in gimple ...
