This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Add code-hoisting to GIMPLE


On Fri, 8 Jul 2016, Rainer Orth wrote:

> Hi Richard,
> 
> > This is a final candidate patch to add code-hoisting to GIMPLE.
> >
> > I've already committed several patches fixing fallout and the following
> > one adds -fno-code-hoisting (I renamed the option) to a few testcases.
> > I filed PRs for the cases code-hoisting exposes missed optimization
> > opportunities in passes that I couldn't quickly fix (I fixed path
> > splitting and loop distribution but failed to grok SLSR).
> >
> > Bootstrapped and tested on x86_64-unknown-linux-gnu.
> >
> > I put the patch on the czerny tester for the weekend runs (x86_64 as 
> > well).
> >
> > Testing on other archs and comments are of course appreciated, if nothing
> > unusual happens I plan to commit this on Monday.
> 
> I've just bootstrapped the patch on sparc-sun-solaris2.12, which
> uncovered a couple of testsuite failures:
> 
> +FAIL: gcc.dg/tree-ssa/split-path-5.c scan-tree-dump-times split-paths "Duplicat
> ing join block" 2
> +FAIL: gcc.dg/tree-ssa/split-path-5.c scan-tree-dump-times split-paths "Duplicat
> ing join block" 2

Saw this on x86_64 as well and should have been fixed with

2016-07-05  Richard Biener  <rguenther@suse.de>                                 
                                                                                
        * gimple-ssa-split-paths.c (find_block_to_duplicate_for_splitting_pa):  
        Handle empty else block.                                                
        (is_feasible_trace): Likewise.                                          
        (split_paths): Likewise.            

> Message doesn't occur at all.
> 
> +FAIL: gfortran.dg/ldist-1.f90   -O   scan-tree-dump-not ldist "distributed: spl
> it to"
> 
> Likewise.

You mean it does occur (it's a scan-tree-dump-not).  I saw this on x86_64
as well and fixed it with

2016-07-05  Richard Biener  <rguenther@suse.de>

        * tree-loop-distribution.c (distribute_loop): Fix issue with
        the cost model loop.

maybe the fixes were not complete.  I'll have a second look with a
sparc-solaris cross on Monday.

Thanks for testing,
Richard.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]