[Bug target/71695] m68k: long long multiplication broken

[martin at netbsd dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71695

Martin Husemann <martin at netbsd dot org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |REOPENED
         Resolution|INVALID                     |---

--- Comment #3 from Martin Husemann <martin at netbsd dot org> ---
Chuck Silvers has looked a bit closer and found that while it does not happen
with an unmodified gcc 5.4 configured for --target=m68k-linux, it is
reproducable when configuring for --target=m68k-netbsdelf.

One of the differences in the config is the override to BIGGEST_ALIGNEMENT in
netbsd-elf.h:

    For m68k/SVR4, some types (doubles for example) are aligned on 8 byte
    boundaries */

 #undef BIGGEST_ALIGNMENT
 #define BIGGEST_ALIGNMENT 64

while Linux uses:

    Most published ABIs say that ints should be aligned on 16-bit
    boundaries, but CPUs with 32-bit busses get better performance
    aligned on 32-bit boundaries.  */
 #define BIGGEST_ALIGNMENT (TARGET_ALIGN_INT ? 32 : 16)


This seems to have been there since prehistoric times. Removing this override
makes gcc generate the same code as shown by Andreas.

More interesting though is that adding BIGGEST_ALIGNMENT 64 to the linux
configuration does NOT make gcc emit the bad code, so there must be another
factor involved. Any hints appreciated!