Re: [PATCH] Fix PR60911

[Richard Biener <rguenther at suse dot de>]

On Fri, 25 Apr 2014, Jan Hubicka wrote:

> > On Thu, 24 Apr 2014, Jan Hubicka wrote:
> > 
> > > > 
> > > > Simple IPA passes are supposed to see function bodies with IPA transforms 
> > > > applied - this is what the code in execute_one_pass tries to ensure.
> > > > But that doesn't work anymore with on-demand function-body loading.
> > > > The following addresses this in the least intrusive way - inlining
> > > > do_per_function (apply_ipa_transforms) and adjusting it accordingly.
> > > > 
> > > > This IMHO is definitely the solution for the 4.9 branch (and for
> > > > the time being on trunk).
> > > > 
> > > > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
> > > > 
> > > > Ok for trunk and branch?
> > > 
> > > I think it is fine for both 4.9 and mainline. I will try to make better version
> > > for mainline as explained in PR hortly.
> > > 
> > > Can you, please, double check that it won't load all bodies prior late 
> > > optimization by default? Looking at gate of pass_omp_simd_clone, perhaps 
> > 
> > Well, first of all it will only load bodies with IPA transforms to apply
> > (yeah, that includes inlining, right?).  Then it's only executed if
> > a small IPA pass actually executes, but ...
> 
> All functions have transforms to apply. They are added to all bodies of
> functions that are around when pass is run. Not only to those pass wants to do
> something. This is how first cut got implemented in 2004 and it is still the
> same. We may want to change this, since the per-function transform lists tends
> to be memory consuming these days. The reason why the lists are per-function
> was that I expected IPA passes to introduce new functions in the middle of
> optimization processing (it now happens for sample with ctor merging) and those
> sould not be processed by transforms of IPA passes that never see them.
> > 
> > > it actually kills late loading of bodies and perhaps we need to mark in 
> > > cgraph node whether the given node needs clonning and page the gate 
> > > return false if partition has no such unit? bool 
> > > pass_omp_simd_clone::gate (function *) {
> > >   return ((flag_openmp || flag_openmp_simd
> > >            || flag_cilkplus
> > >            || (in_lto_p && !flag_wpa))
> > >           && (targetm.simd_clone.compute_vecsize_and_simdlen != NULL));
> > > }
> > > 
> > > I did not see there the in_lto_p previously.
> > 
> > ... this is IIRC because you can't rely on -fopenmp/-fopenmp-simd/-fcilk+
> > to be present on the LTO commandline.
> 
> Yep, together with your change it will kill lazy loading, because the pass is
> practicaly unconditoinal with LTO and all functions will have transforms on them.
> Adding simple flag if function needs openmp_simd processing would solve it.

Not really - you'd also have to defer body loading to the simple IPA
passes (well, and IPA PTA really needs all bodies anyway).  Here
OMP SIMD isn't really an "IPA" pass (other than it creating new
cgraph nodes).

Richard.

> Honza
> > 
> > Richard.
> > 
> > > Honza
> > > > 
> > > > Thanks,
> > > > Richard.
> > > > 
> > > > 2014-04-24  Richard Biener  <rguenther@suse.de>
> > > > 
> > > > 	PR ipa/60911
> > > > 	* passes.c (apply_ipa_transforms): Inline into only caller ...
> > > > 	(execute_one_pass): ... here.  Properly bring in function
> > > > 	bodies for nodes we want to apply IPA transforms to.
> > > > 
> > > > 	* gcc.dg/lto/pr60911_0.c: New testcase.
> > > > 
> > > > Index: gcc/passes.c
> > > > ===================================================================
> > > > --- gcc/passes.c	(revision 209742)
> > > > +++ gcc/passes.c	(working copy)
> > > > @@ -2109,20 +2109,6 @@ execute_all_ipa_transforms (void)
> > > >      }
> > > >  }
> > > >  
> > > > -/* Callback for do_per_function to apply all IPA transforms.  */
> > > > -
> > > > -static void
> > > > -apply_ipa_transforms (void *data)
> > > > -{
> > > > -  struct cgraph_node *node = cgraph_get_node (current_function_decl);
> > > > -  if (!node->global.inlined_to && node->ipa_transforms_to_apply.exists ())
> > > > -    {
> > > > -      *(bool *)data = true;
> > > > -      execute_all_ipa_transforms ();
> > > > -      rebuild_cgraph_edges ();
> > > > -    }
> > > > -}
> > > > -
> > > >  /* Check if PASS is explicitly disabled or enabled and return
> > > >     the gate status.  FUNC is the function to be processed, and
> > > >     GATE_STATUS is the gate status determined by pass manager by
> > > > @@ -2194,8 +2180,26 @@ execute_one_pass (opt_pass *pass)
> > > >       Apply all trnasforms first.  */
> > > >    if (pass->type == SIMPLE_IPA_PASS)
> > > >      {
> > > > +      struct cgraph_node *node;
> > > >        bool applied = false;
> > > > -      do_per_function (apply_ipa_transforms, (void *)&applied);
> > > > +      FOR_EACH_DEFINED_FUNCTION (node)
> > > > +	if (node->analyzed
> > > > +	    && cgraph_function_with_gimple_body_p (node)
> > > > +	    && (!node->clone_of || node->decl != node->clone_of->decl))
> > > > +	  {
> > > > +	    if (!node->global.inlined_to
> > > > +		&& node->ipa_transforms_to_apply.exists ())
> > > > +	      {
> > > > +		cgraph_get_body (node);
> > > > +		push_cfun (DECL_STRUCT_FUNCTION (node->decl));
> > > > +		execute_all_ipa_transforms ();
> > > > +		rebuild_cgraph_edges ();
> > > > +		free_dominance_info (CDI_DOMINATORS);
> > > > +		free_dominance_info (CDI_POST_DOMINATORS);
> > > > +		pop_cfun ();
> > > > +		applied = true;
> > > > +	      }
> > > > +	  }
> > > >        if (applied)
> > > >          symtab_remove_unreachable_nodes (true, dump_file);
> > > >        /* Restore current_pass.  */
> > > > Index: gcc/testsuite/gcc.dg/lto/pr60911_0.c
> > > > ===================================================================
> > > > --- gcc/testsuite/gcc.dg/lto/pr60911_0.c	(revision 0)
> > > > +++ gcc/testsuite/gcc.dg/lto/pr60911_0.c	(working copy)
> > > > @@ -0,0 +1,21 @@
> > > > +// { dg-lto-do run }
> > > > +// { dg-lto-options { { -O2 -flto -fipa-pta } } }
> > > > +
> > > > +int __attribute__ ((__noinline__)) f (unsigned *p, int *x)
> > > > +{
> > > > +  int y = *p++ & 0xfff;
> > > > +  *x++ = y;
> > > > +  *x = *p;
> > > > +  return y;
> > > > +}
> > > > +
> > > > +int
> > > > +main ()
> > > > +{
> > > > +  unsigned u[2] = { 0x3aad, 0x5ad1 };
> > > > +  int x[2] = { 17689, 23456 };
> > > > +
> > > > +  if (f (u, x) != 0xaad || x[0] != 0xaad || x[1] != 0x5ad1)
> > > > +    __builtin_abort ();
> > > > +  return 0;
> > > > +}
> > > 
> > > 
> > 
> > -- 
> > Richard Biener <rguenther@suse.de>
> > SUSE / SUSE Labs
> > SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
> > GF: Jeff Hawn, Jennifer Guild, Felix Imend"orffer
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE / SUSE Labs
SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
GF: Jeff Hawn, Jennifer Guild, Felix Imend"orffer