[Bug target/112573] Suboptimal code generation with `-fdata-sections` on aarch64

[wilco at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112573

Wilco <wilco at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2023-11-20
     Ever confirmed|0                           |1
                 CC|                            |wilco at gcc dot gnu.org
             Status|UNCONFIRMED                 |NEW

--- Comment #3 from Wilco <wilco at gcc dot gnu.org> ---
We should reassociate the immediate last for more optimal addressing like LLVM:

        adrp    x8, a
        add     x8, x8, :lo12:a
        lsr     w9, w0, #8
        add     x8, x8, w1, sxtw
        strb    w9, [x8, #1]
        lsr     w9, w0, #16
        strb    w0, [x8, #3]
        strb    w9, [x8, #2]
        lsr     w9, w0, #24
        strb    w9, [x8]
        ret

However GCC's reassociation is incorrect - it has been for many years and
things got much worse in GCC12...

As a result we may merge the immediate offset into the base address like in
'h'. Using -fdata-sections behaves like -fno-section-anchors, so it works as
expected (and 'extern' is the same as well). We could block merging offsets to
get more address CSEs if that ends up better overall.