This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: IF conversion bug with CC0
- From: Dan <dgisselq at verizon dot net>
- To: Jeff Law <law at redhat dot com>
- Cc: Eric Botcazou <ebotcazou at adacore dot com>, gcc at gcc dot gnu dot org, dgisselq at ieee dot org
- Date: Mon, 04 Apr 2016 18:49:09 -0400
- Subject: Re: IF conversion bug with CC0
- Authentication-results: sourceware.org; auth=none
- References: <1459801150 dot 1603 dot 51 dot camel at jericho> <1459804405 dot 1603 dot 55 dot camel at jericho> <5702DBDB dot 7080102 at redhat dot com> <1752525 dot tYkN1d13DP at polaris> <5702EB7C dot 2090404 at redhat dot com>
- Reply-to: dgisselq at ieee dot org
I like this position--with good kind natured folks arguing over the best
way to help me. ;P
Thank you both.
I took a quick look at Visium, and noticed arithmetic instructions in
the .md file doing a lot of clobbering of the condition codes register.
This doesn't seem very efficient, since it prevents the arithmetic
instructions from being able to set the CC register and have that value
be used.
Am I reading this right? Is this an oversight, an initial pass at
something more complicated, or the "right" way to do things?
Thanks,
Dan
P.S. My apologies if this is a second copy. I have my mailer
configured to automatically produce HTML ...
On Mon, 2016-04-04 at 16:32 -0600, Jeff Law wrote:
> On 04/04/2016 04:20 PM, Eric Botcazou wrote:
> >> From a 30 second view of your ISA, it appears that most
> >> arithmetic/logicals unconditionally set the condition codes.
> >>
> >> I would suggest modeling condition code handling similar to how it's
> >> done on the x86 port.
> >
> > No advertisement intended, but the Visium architecture is the typical 32-bit
> > RISC where every single arithmetic/logical unconditionally sets the condition
> > code. Moreover, the port was very recently converted from CC0 to CCmode (with
> > uniform post-reload splitters and define_substs tailored for the postreload
> > compare elimination pass), so it could be a good model.
> Visium might be an easier port to read/understand :-)
>
> jeff