[Bug target/86673] [8/9 regression] inline asm sometimes ignores 'register asm("reg")' declarations

[thopre01 at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86673

--- Comment #8 from Thomas Preud'homme <thopre01 at gcc dot gnu.org> ---
(In reply to Thomas Preud'homme from comment #7)
> (In reply to Thomas Preud'homme from comment #6)
> > The following simpler testcase also shows the problem:
> > 
> > void fn1() {
> >   register const int h asm("r2") = 1;
> >   __asm__(".ifnc %0,r2; .err; .endif\n\t"
> >           "bl   __put_user_4" :: "r"(h));
> > }
> 
> The register label gets optimized during gimple stages.

Dump for original already shows propagation of the constant has happened which
later lead to the removal of the register declaration:

;; Function fn1 (null)
;; enabled by -tree-original


{
  register const int h __asm__ (*r2) = 1;

    register const int h __asm__ (*r2) = 1;
  __asm__ __volatile__(".ifnc %0,r2; .err; .endif\n\tbl\t__put_user_4"::"r" 1);
}