Re: [PATCH] tree-optimization/106912 - IPA profile and pure/const

[Jakub Jelinek <jakub@redhat.com>]

On Fri, Mar 17, 2023 at 08:09:17PM +0100, Jan Hubicka wrote:
> > 
> > I have in the meantime briefly tested following.
> > 
> > But if you want to the above way, then at least the testcase could be
> > useful.  Though, not sure if the above is all that is needed.  Shouldn't
> > set_const_flag_1 upon TREE_READONLY (node->decl) = 0; also adjust
> > TREE_TYPE on the function to
> > 	  tree fntype = TREE_TYPE (node->decl);
> > 	  TREE_TYPE (node->decl)
> > 	    = build_qualified_type (fntype,
> > 				    TYPE_QUALS (fntype) & ~TYPE_QUAL_CONST);
> > too (perhaps guarded on TREE_READONLY (fntype))?
> > 
> > 2023-03-16  Jakub Jelinek  <jakub@redhat.com>
> > 
> > 	PR tree-optimization/106912
> > gcc/
> > 	* calls.h (ignore_const_fntype): Declare.
> > 	* calls.cc (flags_from_decl_or_type): Ignore TYPE_READONLY
> > 	on types if ignore_const_fntype.
> > 	* tree-profile.cc: Include calls.h.
> > 	(tree_profiling): Set ignore_const_fntype if profile_arc_flag
> > 	or flag_test_coverage.
> > gcc/lto/
> > 	* lto-lang.cc (lto_post_options): Set ignore_const_fntype if
> > 	profile_arc_flag or flag_test_coverage and not flag_auto_profile.
> > gcc/testsuite/
> > 	* gcc.dg/pr106912.c: New test.
> > 
> > --- gcc/calls.h.jj	2023-01-02 09:32:51.252868185 +0100
> > +++ gcc/calls.h	2023-03-16 12:23:51.632460586 +0100
> > @@ -134,5 +134,6 @@ extern void maybe_complain_about_tail_ca
> >  
> >  extern rtx rtx_for_static_chain (const_tree, bool);
> >  extern bool cxx17_empty_base_field_p (const_tree);
> > +extern bool ignore_const_fntype;
> >  
> >  #endif // GCC_CALLS_H
> > --- gcc/calls.cc.jj	2023-02-21 11:44:48.460464845 +0100
> > +++ gcc/calls.cc	2023-03-16 12:27:45.427032110 +0100
> > @@ -800,6 +800,13 @@ is_tm_builtin (const_tree fndecl)
> >    return false;
> >  }
> >  
> > +/* Set if flags_from_decl_or_type should ignore TYPE_READONLY of function
> > +   types.  This is used when tree-profile.cc instruments const calls,
> > +   clears TREE_READONLY on FUNCTION_DECLs which have been instrumented, but
> > +   for function types or indirect calls we don't know if the callees have been
> > +   instrumented or not.  */
> > +bool ignore_const_fntype;
> > +
> >  /* Detect flags (function attributes) from the function decl or type node.  */
> >  
> >  int
> > @@ -849,7 +856,7 @@ flags_from_decl_or_type (const_tree exp)
> >      }
> >    else if (TYPE_P (exp))
> >      {
> > -      if (TYPE_READONLY (exp))
> > +      if (TYPE_READONLY (exp) && !ignore_const_fntype)
> 
> I think we want to ignore PURE flag too: if callee modifies counters
> seen by caller, we need to take that into account when optimizing it.
> 
> It is not very pretty to have global flag for prifled state, but I can't
> think of scenario where this would break, so perhaps it is good way to
> go?

Note, Richi posted what I think is much cleaner patch in
https://gcc.gnu.org/pipermail/gcc-patches/2023-March/614074.html

	Jakub