This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Better handling for multi-word values?
- To: Joern Rennecke <amylaar at cygnus dot co dot uk>
- Subject: Re: Better handling for multi-word values?
- From: Bernd Schmidt <crux at pool dot informatik dot rwth-aachen dot de>
- Date: Mon, 7 Sep 1998 17:27:02 +0200 (MET DST)
- cc: egcs at cygnus dot com
>
> An alternate solution woul be to introduce a target macro that takes
> a register class and a mode, and tells if a register group actually needs
> to use consecutive registers.
Actually, in my current patch, the machine description gets to decide whether
a given mode is allocated as a MULTIWORD, or as a normal single pseudo reg.
The requirement might differ between different instructions: consider the
i386 mulsidi3 pattern, which does need consecutive registers, although none
of the other DImode operations do. So there needs to be a way for reload to
decide per operand whether a group is needed or not, the macro is only a hint.
Are there any machine descriptions which actually require multi-word integer
values to be in consecutive hard regs (except for special cases like the i386
mulsidi3)?
Bernd