[C++] Re: Free more of CFG in release_function_body

> > On Wed, Nov 25, 2020 at 3:11 PM Jan Hubicka <hubicka@ucw.cz> wrote:
> > >
> > > > On Tue, 24 Nov 2020, Jan Hubicka wrote:
> > > >
> > > > > Hi,
> > > > > at the end of processing function body we loop over basic blocks and
> > > > > free all edges while we do not free the rest.  I think this is leftover
> > > > > from time eges was not garbage collected and we was not using ggc_free.
> > > > > It makes more sense to free all associated structures (which is
> > > > > importnat for WPA memory footprint).
> > > > >
> > > > > Bootstrapped/regtested x86_64-linux, OK?
> > > >
> > > > OK.
> > >
> > > Unforutnately the patch does not surive LTO bootstrap.  The problem is
> > > that we keep DECL_INITIAL that points to blocks and blocks points to
> > > var_decls and these points to SSA_NAMES that points to statements and
> > > those points to basic blocks.
> > 
> > VAR_DECLs point to SSA_NAMEs?  It's the other way around.  We for sure
> > free SSA_NAMEs (well, maybe not explicitely with ggc_free).
> 
> I am going to debug this more carefully now.  I think it was VAR_DECL
> with variadic type pointing to SSA_NAME.  Should be easy to reduct with
> gcac compiler.

Hi,
it turns out that the pointers to statements leaks through saved scopes
in C++ FE.  Scopes seems to point to internal blocks of functions even
after we finish their compiling.

This patch adds code to free pointers.  I tried to clear saved_blocks
but it breaks since C++ finalization uses them, but it does not look
into previous class levels.

Patch lto-bootstraps/regtestes x86_64-linux with all languages. OK?

Honza

	* cfg.c (free_block): Call ggc_free on BB itself.

	* cp-tre.eh (cp_tree_c_finish_parsing): Declare.
	* semantics.c (finish_translation_unit): Call finish_parsing
	* tree.c (cp_tree_c_finish_parsing): New function.

On 11/27/20 6:26 AM, Jan Hubicka wrote:
>>> On Wed, Nov 25, 2020 at 3:11 PM Jan Hubicka <hubicka@ucw.cz> wrote:
>>>>> On Tue, 24 Nov 2020, Jan Hubicka wrote:
>>>>>
>>>>>> Hi,
>>>>>> at the end of processing function body we loop over basic blocks and
>>>>>> free all edges while we do not free the rest.  I think this is leftover
>>>>>> from time eges was not garbage collected and we was not using ggc_free.
>>>>>> It makes more sense to free all associated structures (which is
>>>>>> importnat for WPA memory footprint).
>>>>>>
>>>>>> Bootstrapped/regtested x86_64-linux, OK?
>>>>> OK.
>>>> Unforutnately the patch does not surive LTO bootstrap.  The problem is
>>>> that we keep DECL_INITIAL that points to blocks and blocks points to
>>>> var_decls and these points to SSA_NAMES that points to statements and
>>>> those points to basic blocks.
>>> VAR_DECLs point to SSA_NAMEs?  It's the other way around.  We for sure
>>> free SSA_NAMEs (well, maybe not explicitely with ggc_free).
>> I am going to debug this more carefully now.  I think it was VAR_DECL
>> with variadic type pointing to SSA_NAME.  Should be easy to reduct with
>> gcac compiler.
> Hi,
> it turns out that the pointers to statements leaks through saved scopes
> in C++ FE.  Scopes seems to point to internal blocks of functions even
> after we finish their compiling.
>
> This patch adds code to free pointers.  I tried to clear saved_blocks
> but it breaks since C++ finalization uses them, but it does not look
> into previous class levels.
>
> Patch lto-bootstraps/regtestes x86_64-linux with all languages. OK?
>
> Honza
>
> 	* cfg.c (free_block): Call ggc_free on BB itself.
I'll ACK this bit.  The C++ team should own the rest, while it looks
reasonable to me, there may be some reason I'm not aware of that they're
keeping that stuff around.

Jeff

On 11/27/20 8:26 AM, Jan Hubicka wrote:
>>> On Wed, Nov 25, 2020 at 3:11 PM Jan Hubicka <hubicka@ucw.cz> wrote:
>>>>
>>>>> On Tue, 24 Nov 2020, Jan Hubicka wrote:
>>>>>
>>>>>> Hi,
>>>>>> at the end of processing function body we loop over basic blocks and
>>>>>> free all edges while we do not free the rest.  I think this is leftover
>>>>>> from time eges was not garbage collected and we was not using ggc_free.
>>>>>> It makes more sense to free all associated structures (which is
>>>>>> importnat for WPA memory footprint).
>>>>>>
>>>>>> Bootstrapped/regtested x86_64-linux, OK?
>>>>>
>>>>> OK.
>>>>
>>>> Unforutnately the patch does not surive LTO bootstrap.  The problem is
>>>> that we keep DECL_INITIAL that points to blocks and blocks points to
>>>> var_decls and these points to SSA_NAMES that points to statements and
>>>> those points to basic blocks.
>>>
>>> VAR_DECLs point to SSA_NAMEs?  It's the other way around.  We for sure
>>> free SSA_NAMEs (well, maybe not explicitely with ggc_free).
>>
>> I am going to debug this more carefully now.  I think it was VAR_DECL
>> with variadic type pointing to SSA_NAME.  Should be easy to reduct with
>> gcac compiler.
> 
> Hi,
> it turns out that the pointers to statements leaks through saved scopes
> in C++ FE.  Scopes seems to point to internal blocks of functions even
> after we finish their compiling.
> 
> This patch adds code to free pointers.  I tried to clear saved_blocks
> but it breaks since C++ finalization uses them, but it does not look
> into previous class levels.
> 
> Patch lto-bootstraps/regtestes x86_64-linux with all languages. OK?
> 
> Honza
> 
> 	* cfg.c (free_block): Call ggc_free on BB itself.
> 
> 	* cp-tre.eh (cp_tree_c_finish_parsing): Declare.
> 	* semantics.c (finish_translation_unit): Call finish_parsing
> 	* tree.c (cp_tree_c_finish_parsing): New function.
> diff --git a/gcc/cfg.c b/gcc/cfg.c
> index 529b6ed2105..e8bd1456c9f 100644
> --- a/gcc/cfg.c
> +++ b/gcc/cfg.c
> @@ -102,8 +102,7 @@ free_block (basic_block bb)
>      bb->succs = NULL;
>      vec_free (bb->preds);
>      bb->preds = NULL;
> -   /* Do not free BB itself yet since we leak pointers to dead statements
> -      that points to dead basic blocks.  */
> +   ggc_free (bb);
>   }
>   
>   /* Free the memory associated with the CFG in FN.  */
> diff --git a/gcc/cp/cp-tree.h b/gcc/cp/cp-tree.h
> index 021de76e142..665d171d9b0 100644
> --- a/gcc/cp/cp-tree.h
> +++ b/gcc/cp/cp-tree.h
> @@ -7986,6 +7986,8 @@ struct uid_sensitive_constexpr_evaluation_checker
>     bool evaluation_restricted_p () const;
>   };
>   
> +void cp_tree_c_finish_parsing ();
> +
>   /* In cp-ubsan.c */
>   extern void cp_ubsan_maybe_instrument_member_call (tree);
>   extern void cp_ubsan_instrument_member_accesses (tree *);
> diff --git a/gcc/cp/semantics.c b/gcc/cp/semantics.c
> index 5ff70ff4844..e9d17c21985 100644
> --- a/gcc/cp/semantics.c
> +++ b/gcc/cp/semantics.c
> @@ -3094,6 +3094,7 @@ finish_translation_unit (void)
>   	       "%<#pragma omp end declare target%>");
>         scope_chain->omp_declare_target_attribute = 0;
>       }
> +  cp_tree_c_finish_parsing ();

This is too soon for this call; it should be from c_parse_final_cleanups 
by the call to fini_constexpr, so that it follows template instantiation 
at EOF.

>   }
>   
>   /* Finish a template type parameter, specified as AGGR IDENTIFIER.
> diff --git a/gcc/cp/tree.c b/gcc/cp/tree.c
> index 28e591086b3..e63d383c0a3 100644
> --- a/gcc/cp/tree.c
> +++ b/gcc/cp/tree.c
> @@ -5844,6 +5844,19 @@ maybe_warn_zero_as_null_pointer_constant (tree expr, location_t loc)
>     return false;
>   }
>   
> +/* Release memory we no longer need after parsing.  */
> +void
> +cp_tree_c_finish_parsing ()
> +{
> +  saved_scope *chain = scope_chain;
> +  while (chain)
> +    {
> +      chain->x_previous_class_level = NULL;
> +      chain = chain->prev;
> +    }

This can just be invalidate_class_lookup_cache ().  scope_chain->prev 
will always be NULL at this point.

OK with those changes.

> +  deleted_copy_types = NULL;
> +}
> +
>   #if defined ENABLE_TREE_CHECKING && (GCC_VERSION >= 2007)
>   /* Complain that some language-specific thing hanging off a tree
>      node has been accessed improperly.  */