[committed] Enable LRA on several ports

  Spurred by Segher's RFC, I went ahead and tested several ports with LRA 
enabled.  Not surprisingly, many failed, but a few built their full set 
of libraries successful and of those a few even ran their testsuites 
with no regressions.  In fact, enabling LRA fixes a small number of 
failures on the iq2000 port.

This patch converts the ports which built their libraries and have test 
results that are as good as or better than without LRA.    There may be 
minor code quality regressions or there may be minor code quality 
improvements -- I'm leaving that for the port maintainers to own going 
forward.

Committed to the trunk,

jeff
commit faf8bea79b62569af2891e7adc6f758141f738af
Author: Jeff Law <jlaw@ventanamicro>
Date:   Mon May 1 07:14:50 2023 -0600

    Enable LRA on several ports
    
    Spurred by Segher's RFC, I went ahead and tested several ports with LRA
    enabled.  Not surprisingly, many failed, but a few built their full set
    of libraries successful and of those a few even ran their testsuites
    with no regressions.  In fact, enabling LRA fixes a small number of
    failures on the iq2000 port.
    
    This patch converts the ports which built their libraries and have test
    results that are as good as or better than without LRA.    There may
    be minor code quality regressions or there may be minor code quality
    improvements -- I'm leaving that for the port maintainers to own going
    forward.
    
    gcc/
    
            * config/cris/cris.cc (TARGET_LRA_P): Remove.
            * config/epiphany/epiphany.cc (TARGET_LRA_P): Remove.
            * config/iq2000/iq2000.cc (TARGET_LRA_P): Remove.
            * config/m32r/m32r.cc (TARGET_LRA_P): Remove.
            * config/microblaze/microblaze.cc (TARGET_LRA_P): Remove.
            * config/mmix/mmix.cc (TARGET_LRA_P): Remove.
  

> Date: Mon, 1 May 2023 07:21:59 -0600
> From: Jeff Law <jlaw@ventanamicro.com>

> Spurred by Segher's RFC, I went ahead and tested several ports with LRA 
> enabled.  Not surprisingly, many failed, but a few built their full set 
> of libraries successful and of those a few even ran their testsuites 
> with no regressions.  In fact, enabling LRA fixes a small number of 
> failures on the iq2000 port.
> 
> This patch converts the ports which built their libraries and have test 
> results that are as good as or better than without LRA.

Yeah, I did fix test-suite errors for CRIS with LRA earlier
(see commit logs to the CRIS port this year).

>    There may be 
> minor code quality regressions or there may be minor code quality 
> improvements -- I'm leaving that for the port maintainers to own going 
> forward.

Right; I noticed performance regressions, and didn't want to
commit anything that knowingly degraded performance.  I did
follow the traces but fell into the rabbit-hole of
rtx_costs.  That's the main reason I didn't push a "-mlra"
option or remove the TARGET_LRA_P for CRIS.  (My story and I
stick to it.)

But thanks I guess, it saves me a commit, but (to all!)
please sync check_effective_target_lra for targets you
"convert".  ...oops, it's just CRIS and hppa there (wot, not
converted?)

brgds, H-P
  

On 5/1/23 21:24, Hans-Peter Nilsson via Gcc-patches wrote:

> 
>>     There may be
>> minor code quality regressions or there may be minor code quality
>> improvements -- I'm leaving that for the port maintainers to own going
>> forward.
> 
> Right; I noticed performance regressions, and didn't want to
> commit anything that knowingly degraded performance.  I did
> follow the traces but fell into the rabbit-hole of
> rtx_costs.  That's the main reason I didn't push a "-mlra"
> option or remove the TARGET_LRA_P for CRIS.  (My story and I
> stick to it.)
> 
> But thanks I guess, it saves me a commit, but (to all!)
> please sync check_effective_target_lra for targets you
> "convert".  ...oops, it's just CRIS and hppa there (wot, not
> converted?)
Well, I'd say that my plan would be to deprecate any target that is not 
converted by the end of this development cycle.  So the change keeps 
cris from falling into that bucket.

I wasn't aware of check_effective_target_lra.  Thanks.  I'll make sure 
to get that updated as we go forward.

jeff
  

On Tue, 2 May 2023, Jeff Law via Gcc-patches wrote:

> Well, I'd say that my plan would be to deprecate any target that is not
> converted by the end of this development cycle.  So the change keeps cris from
> falling into that bucket.

 As I noted in the other thread it is highly unlikely I will make it with 
the VAX target in this release cycle, owing to the catastrophic breakage 
of the exception unwinder, recently discovered, which I consider higher 
priority as a show-stopper for important software such as current GDB.  I 
will appreciate your taking this into consideration.

 That written the VAX target does build its target libraries with `-mlra', 
but there are ICE regressions in the test suite and overall code produced 
is brown paperbag quality.  And removing `-mno-lra' before that has been 
sorted will make making LRA match old reload quality much tougher.

  Maciej
  

On Fri, May 19, 2023 at 1:45 PM Maciej W. Rozycki <macro@orcam.me.uk> wrote:
>
> On Tue, 2 May 2023, Jeff Law via Gcc-patches wrote:
>
> > Well, I'd say that my plan would be to deprecate any target that is not
> > converted by the end of this development cycle.  So the change keeps cris from
> > falling into that bucket.
>
>  As I noted in the other thread it is highly unlikely I will make it with
> the VAX target in this release cycle, owing to the catastrophic breakage
> of the exception unwinder, recently discovered, which I consider higher
> priority as a show-stopper for important software such as current GDB.  I
> will appreciate your taking this into consideration.

You might end up with VAX working fine with reload for GCC 14 but
marked as deprecated.  You then have the full next cycle to GCC 15
to improve the code quality with LRA - note that reload is likely removed
early in the development cycle.

>  That written the VAX target does build its target libraries with `-mlra',
> but there are ICE regressions in the test suite and overall code produced
> is brown paperbag quality.  And removing `-mno-lra' before that has been
> sorted will make making LRA match old reload quality much tougher.

You can always compare to GCC 14 then or even work based off the
release branch.

Richard.

>   Maciej
  

On Mon, 1 May 2023, Jeff Law wrote:

> 
> Spurred by Segher's RFC, I went ahead and tested several ports with LRA
> enabled.  Not surprisingly, many failed, but a few built their full set of
> libraries successful and of those a few even ran their testsuites with no
> regressions.  In fact, enabling LRA fixes a small number of failures on the
> iq2000 port.
> 
> This patch converts the ports which built their libraries and have test
> results that are as good as or better than without LRA.    There may be minor
> code quality regressions or there may be minor code quality improvements --
> I'm leaving that for the port maintainers to own going forward.

How do you configure your builds?  Perhaps your cross-builds 
exclude C++?  I found that this (r14-383) broke MMIX building 
libstdc++-v3 from that commit up to and including r14-3180.
See commit r14-3187.

Thankfully there was just one single gotcha.  I temporarily 
reverted the LRA change for MMIX so that I can get honest 
repeatable baseline results.  There seems to have been one 
test-case regressing from the LRA switch (PR53948), thus I 
re-enabled LRA for MMIX again.  Sorry for the late reaction.

brgds, H-P