https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87246
--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Jakub Jelinek from comment #3)
> Started with r233107.
>
> Slightly adjusted testcase:
>
> __int128 a;
> int b;
>
> void
> bar (__int128 *x)
> {
> if (*x != 0)
> {
> a = 0;
> b = b <= *x;
> }
> }
>
> void
> foo (unsigned int x)
> {
> bar (x + 1);
> }
>
> This is on
> (mem:TI (zero_extend:DI (plus:SI (reg/v:SI 91 [ x ]) (const_int 1 [0x1]))))
> which matches the predicate of the *movti_internal instruction -
> general_operand, but the selected alternative needs offsettable memory.
> LRA attempts:
> (mem:DI (plus:DI (zero_extend:DI (plus:SI (reg/v:SI 91 [ x ])
> (const_int 1 [0x1])))
> (const_int 8 [0x8])) [1 *_3+8 S8 A64])
> but that isn't a valid offsettable address, it should have instead forced
> the zero_extend into a register.
>
> Or should some reload hook in the backend tell LRA that it should do that if
> it wants an offsettable address out of that?
This should be handled in ix86_secondary_reload TARGET_SECONDARY_RELOAD hook,
which has the code to handle
/* Double-word spills from general registers to non-offsettable memory
references (zero-extended addresses) require special handling. */
via reload_noff_{load,store} patterns.
