Re: [PATCH] predict: Adjust optimize_function_for_size_p [PR105818]

["Kewen.Lin" <linkw@linux.ibm.com>]

on 2022/7/11 11:42, Kewen.Lin wrote:
> on 2022/6/15 14:20, Kewen.Lin wrote:
>> Hi Honza,
>>
>> Thanks for the comments!  Some replies are inlined below.
>>
>> on 2022/6/14 19:37, Jan Hubicka wrote:
>>>> Hi,
>>>>
>>>> Function optimize_function_for_size_p returns OPTIMIZE_SIZE_NO
>>>> if func->decl is not null but no cgraph node is available for it.
>>>> As PR105818 shows, this could give unexpected result.  For the
>>>> case in PR105818, when parsing bar decl in function foo, the cfun
>>>> is a function structure for foo, for which there is none cgraph
>>>> node, so it returns OPTIMIZE_SIZE_NO.  But it's incorrect since
>>>> the context is to optimize for size, the flag optimize_size is
>>>> true.
>>>>
>>>> The patch is to make optimize_function_for_size_p to check
>>>> optimize_size as what it does when func->decl is unavailable.
>>>>
>>>> One regression failure got exposed on aarch64-linux-gnu:
>>>>
>>>> PASS->FAIL: gcc.dg/guality/pr54693-2.c   -Os \
>>>> 	    -DPREVENT_OPTIMIZATION  line 21 x == 10 - i
>>>>
>>>> The difference comes from the macro LOGICAL_OP_NON_SHORT_CIRCUIT
>>>> used in function fold_range_test during c parsing, it uses
>>>> optimize_function_for_speed_p which is equal to the invertion
>>>> of optimize_function_for_size_p.  At that time cfun->decl is valid
>>>> but no cgraph node for it, w/o this patch function
>>>> optimize_function_for_speed_p returns true eventually, while it
>>>> returns false with this patch.  Since the command line option -Os
>>>> is specified, there is no reason to interpret it as "for speed".
>>>> I think this failure is expected and adjust the test case
>>>> accordingly.
>>>>
>>>> Is it ok for trunk?
>>>>
>>>> BR,
>>>> Kewen
>>>> -----
>>>>
>>>> 	PR target/105818
>>>>
>>>> gcc/ChangeLog:
>>>>
>>>> 	* predict.cc (optimize_function_for_size_p): Check optimize_size when
>>>> 	func->decl is valid but its cgraph node is unavailable.
>>>>
>>>> gcc/testsuite/ChangeLog:
>>>>
>>>> 	* gcc.target/powerpc/pr105818.c: New test.
>>>> 	* gcc.dg/guality/pr54693-2.c: Adjust for aarch64.
>>>> ---
>>>>  gcc/predict.cc                              | 2 +-
>>>>  gcc/testsuite/gcc.dg/guality/pr54693-2.c    | 2 +-
>>>>  gcc/testsuite/gcc.target/powerpc/pr105818.c | 9 +++++++++
>>>>  3 files changed, 11 insertions(+), 2 deletions(-)
>>>>  create mode 100644 gcc/testsuite/gcc.target/powerpc/pr105818.c
>>>>
>>>> diff --git a/gcc/predict.cc b/gcc/predict.cc
>>>> index 5734e4c8516..6c60a973236 100644
>>>> --- a/gcc/predict.cc
>>>> +++ b/gcc/predict.cc
>>>> @@ -268,7 +268,7 @@ optimize_function_for_size_p (struct function *fun)
>>>>    cgraph_node *n = cgraph_node::get (fun->decl);
>>>>    if (n)
>>>>      return n->optimize_for_size_p ();
>>>> -  return OPTIMIZE_SIZE_NO;
>>>> +  return optimize_size ? OPTIMIZE_SIZE_MAX : OPTIMIZE_SIZE_NO;
>>>
>>> We could also do (opt_for_fn (cfun->decl, optimize_size) that is
>>> probably better since one can change optimize_size with optimization
>>> attribute.
>>
>> Good point, agree!
>>
>>> However I think in most cases we check for optimize_size early I think
>>> we are doing something wrong, since at that time htere is no profile
>>> available.  Why exactly PR105818 hits the flag change issue?
>>
>> For PR105818, the reason why the flag changs is that:
>>
>> Firstly, the inconsistent flag is OPTION_MASK_SAVE_TOC_INDIRECT bit
>> of rs6000_isa_flags_explicit, it's set as below:
>>
>> /* If we can shrink-wrap the TOC register save separately, then use
>>    -msave-toc-indirect unless explicitly disabled.  */
>> if ((rs6000_isa_flags_explicit & OPTION_MASK_SAVE_TOC_INDIRECT) == 0
>>     && flag_shrink_wrap_separate
>>     && optimize_function_for_speed_p (cfun))
>>   rs6000_isa_flags |= OPTION_MASK_SAVE_TOC_INDIRECT;
>>
>> Initially, rs6000 initialize target_option_default_node with
>> OPTION_MASK_SAVE_TOC_INDIRECT unset, at that time cfun is NULL
>> and optimize_size is true.
>>
>> Later, when c parser handling function foo, it builds target
>> option node as target_option_default_node in function
>> handle_optimize_attribute, it does global option saving and
>> verifying there as well, at that time the cfun is NULL, no
>> issue is found.  And function store_parm_decls allocates
>> struct_function for foo then, cfun becomes function struct
>> for foo, when c parser continues to handle the decl bar in
>> foo, function handle_optimize_attribute works as before,
>> tries to restore the target options at the end, it calls
>> targetm.target_option.restore (rs6000_function_specific_restore)
>> which calls function rs6000_option_override_internal again,
>> at this time the cfun is not NULL while there is no cgraph
>> node for its decl, optimize_function_for_speed_p returns true
>> and gets the OPTION_MASK_SAVE_TOC_INDIRECT bit of flag
>> rs6000_isa_flags set unexpectedly.  It becomes inconsistent
>> as the one saved previously.
>>
>> IMHO, both contexts of global and function decl foo here hold
>> optimize_size, function optimize_function_for_speed_p should
>> not return true anyway.
>>
>> btw, the aarch64 failed case also gets the unexpected
>> result for optimize_function_for_speed_p during c parsing
>> (fold_range_test <- ... <- c_parser_condition).
>>
>> IIUC, in parsing time we don't have the profile information
>> available.
>>
> 
> Hi Honza,
> 
> Does the above explanation sound reasonable to you?
> 

Hi Honza,

Gentle ping ...

BR,
Kewen