On Tue, Aug 9, 2011 at 12:40 AM, Uros Bizjak <ubiz...@gmail.com> wrote:
> On Mon, Aug 8, 2011 at 7:11 PM, Uros Bizjak <ubiz...@gmail.com> wrote:
>
>>> Moves are special as far as reload is concerned. If there is already
>>> a move instruction present *before* reload, it will get fixed up
>>> according to its constraints as any other instruction.
>>>
>>> However, reload will *introduce* new moves as part of its operation,
>>> and those will *not* themselves get reloaded. Instead, reload simply
>>> assumes that every plain move will just succeed without requiring
>>> any reload; if this is not true, the target *must* provide a
>>> secondary reload for this move.
>>>
>>> (Note that the secondary reload could also work by reloading the
>>> target address into a temporary; that's up to the target to
>>> implement.)
>>
>> Whoa, indeed.
>>
>> Using attached patch that reloads memory address instead of going
>> through XMM register, the code for the testcase improves from:
>
> Committed to mainline with following ChangeLog entry:
>
> 2011-08-09 Uros Bizjak <ubiz...@gmail.com>
>
> PR target/49781
> * config/i386/i386.md (reload_noff_load): New.
> (reload_noff_store): Ditto.
> * config/i386/i386.c (ix86_secondary_reload): Use
> CODE_FOR_reload_noff_load and CODE_FOR_reload_noff_store to handle
> double-word moves from/to non-offsetable addresses instead of
> generating XMM temporary.
>
> Re-bootstrapped and re-tested on x86_64-pc-linux-gnu {,-m32}.
>
No regressions on x32 with GCC, glibc and SPEC CPU 2K/2006.
Thanks.
--
H.J.
