Re: RFA: one more version of the patch for PR61360

[Vladimir Makarov <vmakarov at redhat dot com>]

On 2014-10-02 7:31 AM, Markus Trippelsdorf wrote:
> On 2014.10.02 at 09:17 +0200, Markus Trippelsdorf wrote:
> > On 2014.09.26 at 16:31 -0400, Vladimir Makarov wrote:
> > > I guess we achieved the consensus about the following patch to fix PR61360
> > >
> > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61360
> > >
> > > The patch was successfully bootstrapped and tested (w/wo
> > > -march=amdfam10) on x86/x86-64.
> > >
> > > Is it ok to commit to trunk?
> >
> > I've tested your patch and unfortunately it doesn't work:
> >
> > In file included from /var/tmp/moz-build-dir/js/src/shell/Unified_cpp_js_src_shell0.cpp:15:0:
> > /var/tmp/mozilla-central/js/src/shell/js.cpp: In function ‘void Process(JSContext*, JSObject*, const char*, bool)’:
> > /var/tmp/mozilla-central/js/src/shell/js.cpp:592:1: internal compiler error: in lra_update_insn_recog_data, at lra.c:1221
> >   }
> >   ^
> > 0xa9d9ec lra_update_insn_recog_data(rtx_insn*)
> >          ../../gcc/gcc/lra.c:1220
> > 0xab450f eliminate_regs_in_insn
> >          ../../gcc/gcc/lra-eliminations.c:1077
> > 0xab450f process_insn_for_elimination
> >          ../../gcc/gcc/lra-eliminations.c:1344
> > 0xab450f lra_eliminate(bool, bool)
> >          ../../gcc/gcc/lra-eliminations.c:1408
> > 0xa9f2da lra(_IO_FILE*)
> >          ../../gcc/gcc/lra.c:2270
> > 0xa5d659 do_reload
> >          ../../gcc/gcc/ira.c:5311
> > 0xa5d659 execute
> >          ../../gcc/gcc/ira.c:5470
>
> Testcase is attached:
>
>   % g++ -c -march=amdfam10 -w -O2 js.ii
> js.ii: In function ‘void RunFile(C)’:
> js.ii:64:1: internal compiler error: in lra_update_insn_recog_data, at lra.c:1221

Thanks for reporting this, Marcus.  The problem now is in 
optimize_function_for_size_p.  It is false, when we define and cache 
enable attributes at early stages (instantation of virtual regs) and 
true later.

It is returning us to the same problem.  I believe that we should not 
have enable attributes depending on optimization options or on the 
current running pass if we don't want the current solution in the trunk 
(with recog_init).  Setting right value for optimize_function_for_size_p 
does not solve the problem as we can have different options for 
different functions in the same compilation file.

So minimal solution would be removing optimize_function_for_size_p from 
the attribute definition.  But I guess we could remove all condition. 
Unfortunately, Ganesh did not post is it really beneficial to switch off 
this alternative for AMD CPUs even if AMD optimization guide recommends it.