diff mbox

Fix inlining checks wrt optimize attribute

Message ID 54C0CF92.6080001@st.com
State New
Headers show

Commit Message

Christian Bruel Jan. 22, 2015, 10:23 a.m. UTC 
Hi Richard,

I thought one of my current issue would be solved by this patch, but it 
is not : I have some inlining failures with the attribute target on ARM. 
(e.g inline-3.c) where obvious early inline fails with because we fail 
into the last can_inline_edge_p case:

opt_for_fn (callee->decl, optimize)
		   >= opt_for_fn (caller->decl, optimize)))

when callee and caller are both -O2 and 
targetm.target_option.can_inline_p was true, they should be inlined as 
in the general case (no DECL_FUNCTION_SPECIFIC_OPTIMIZATION)

I'm currently testing this additional change:


Since this is a hot topic for you, I though you would have useful 
comments on this before I ask for a commit (when stage 4 close) ?

Cheers

Christian





On 01/22/2015 10:24 AM, Richard Biener wrote:
>
> As said in the other thread - this makes sure we don't perform inlining
> that might end up generating invalid code.  It also preserves
> user-provided optimize attributes more properly.
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.
>
> Richard.
>
> 2015-01-22  Richard Biener  <rguenther@suse.de>
>
> 	* ipa-inline.c (can_inline_edge_p): Disable inlining of edges
> 	with IL incompatible options.  Properly honor user optimize
> 	attributes.
>
> Index: gcc/ipa-inline.c
> ===================================================================
> --- gcc/ipa-inline.c	(revision 219929)
> +++ gcc/ipa-inline.c	(working copy)
> @@ -404,17 +404,56 @@ can_inline_edge_p (struct cgraph_edge *e
>        optimization attribute.  */
>     else if (caller_tree != callee_tree)
>       {
> -      /* gcc.dg/pr43564.c.  Look at forced inline even in -O0.  */
> -      if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
> +      /* There are some options that change IL semantics which means
> +         we cannot inline in these cases for correctness reason.
> +	 Not even for always_inline declared functions.  */
> +      /* Strictly speaking only when the callee contains signed integer
> +         math where overflow is undefined.  */
> +      if ((opt_for_fn (e->caller->decl, flag_strict_overflow)
> +	   != opt_for_fn (e->caller->decl, flag_strict_overflow))
> +	  || (opt_for_fn (e->caller->decl, flag_wrapv)
> +	      != opt_for_fn (e->caller->decl, flag_wrapv))
> +	  || (opt_for_fn (e->caller->decl, flag_trapv)
> +	      != opt_for_fn (e->caller->decl, flag_trapv))
> +	  /* Strictly speaking only when the callee contains memory
> +	     accesses that are not using alias-set zero anyway.  */
> +	  || (opt_for_fn (e->caller->decl, flag_strict_aliasing)
> +	      != opt_for_fn (e->caller->decl, flag_strict_aliasing))
> +	  /* Strictly speaking only when the callee uses FP math.  */
> +	  || (opt_for_fn (e->caller->decl, flag_rounding_math)
> +	      != opt_for_fn (e->caller->decl, flag_rounding_math))
> +	  || (opt_for_fn (e->caller->decl, flag_trapping_math)
> +	      != opt_for_fn (e->caller->decl, flag_trapping_math))
> +	  || (opt_for_fn (e->caller->decl, flag_unsafe_math_optimizations)
> +	      != opt_for_fn (e->caller->decl, flag_unsafe_math_optimizations))
> +	  || (opt_for_fn (e->caller->decl, flag_finite_math_only)
> +	      != opt_for_fn (e->caller->decl, flag_finite_math_only))
> +	  || (opt_for_fn (e->caller->decl, flag_signaling_nans)
> +	      != opt_for_fn (e->caller->decl, flag_signaling_nans))
> +	  || (opt_for_fn (e->caller->decl, flag_cx_limited_range)
> +	      != opt_for_fn (e->caller->decl, flag_cx_limited_range))
> +	  || (opt_for_fn (e->caller->decl, flag_signed_zeros)
> +	      != opt_for_fn (e->caller->decl, flag_signed_zeros))
> +	  || (opt_for_fn (e->caller->decl, flag_associative_math)
> +	      != opt_for_fn (e->caller->decl, flag_associative_math))
> +	  || (opt_for_fn (e->caller->decl, flag_reciprocal_math)
> +	      != opt_for_fn (e->caller->decl, flag_reciprocal_math))
> +	  /* Strictly speaking only when the callee contains function
> +	     calls that may end up setting errno.  */
> +	  || (opt_for_fn (e->caller->decl, flag_errno_math)
> +	      != opt_for_fn (e->caller->decl, flag_errno_math)))
> +	{
> +	  e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
> +	  inlinable = false;
> +	}
> +      /* gcc.dg/pr43564.c.  Apply user-forced inline even at -O0.  */
> +      else if (DECL_DISREGARD_INLINE_LIMITS (callee->decl)
> +	       && lookup_attribute ("always_inline",
> +				    DECL_ATTRIBUTES (callee->decl)))
>   	;
> -      /* When user added an attribute, honnor it.  */
> -      else if ((lookup_attribute ("optimize", DECL_ATTRIBUTES (caller->decl))
> -		|| lookup_attribute ("optimize",
> -				     DECL_ATTRIBUTES (callee->decl)))
> -	       && ((opt_for_fn (caller->decl, optimize)
> -		   > opt_for_fn (callee->decl, optimize))
> -		   || (opt_for_fn (caller->decl, optimize_size)
> -		       != opt_for_fn (callee->decl, optimize_size))))
> +      /* When user added an attribute to the callee honor it.  */
> +      else if (lookup_attribute ("optimize", DECL_ATTRIBUTES (callee->decl))
> +	       && opts_for_fn (caller->decl) != opts_for_fn (callee->decl))
>   	{
>   	  e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
>   	  inlinable = false;
>

Comments

Richard Biener Jan. 22, 2015, 10:33 a.m. UTC | #1 
On Thu, 22 Jan 2015, Christian Bruel wrote:

> Hi Richard,
> 
> I thought one of my current issue would be solved by this patch, but it is not
> : I have some inlining failures with the attribute target on ARM. (e.g
> inline-3.c) where obvious early inline fails with because we fail into the
> last can_inline_edge_p case:
> 
> opt_for_fn (callee->decl, optimize)
> 		   >= opt_for_fn (caller->decl, optimize)))
> 
> when callee and caller are both -O2 and targetm.target_option.can_inline_p was
> true, they should be inlined as in the general case (no
> DECL_FUNCTION_SPECIFIC_OPTIMIZATION)
> 
> I'm currently testing this additional change:
> 
> Index: ipa-inline.c
> ===================================================================
> --- ipa-inline.c	(revision 219989)
> +++ ipa-inline.c	(working copy)
> @@ -489,7 +489,7 @@
>        else if (opt_for_fn (callee->decl, optimize_size)
>  	       < opt_for_fn (caller->decl, optimize_size)
>  	       || (opt_for_fn (callee->decl, optimize)
> -		   >= opt_for_fn (caller->decl, optimize)))
> +		   > opt_for_fn (caller->decl, optimize)))
>  	{
>  	  if (estimate_edge_time (e)
>  	      >= 20 + inline_edge_summary (e)->call_stmt_time)
> 
> Since this is a hot topic for you, I though you would have useful comments on
> this before I ask for a commit (when stage 4 close) ?

Yeah - the above looks like an obvious change to me.  Thus,
approved if it passes bootstrap/regtest.

Thanks,
Richard.

> Cheers
> 
> Christian
> 
> 
> 
> 
> 
> On 01/22/2015 10:24 AM, Richard Biener wrote:
> > 
> > As said in the other thread - this makes sure we don't perform inlining
> > that might end up generating invalid code.  It also preserves
> > user-provided optimize attributes more properly.
> > 
> > Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.
> > 
> > Richard.
> > 
> > 2015-01-22  Richard Biener  <rguenther@suse.de>
> > 
> > 	* ipa-inline.c (can_inline_edge_p): Disable inlining of edges
> > 	with IL incompatible options.  Properly honor user optimize
> > 	attributes.
> > 
> > Index: gcc/ipa-inline.c
> > ===================================================================
> > --- gcc/ipa-inline.c	(revision 219929)
> > +++ gcc/ipa-inline.c	(working copy)
> > @@ -404,17 +404,56 @@ can_inline_edge_p (struct cgraph_edge *e
> >        optimization attribute.  */
> >     else if (caller_tree != callee_tree)
> >       {
> > -      /* gcc.dg/pr43564.c.  Look at forced inline even in -O0.  */
> > -      if (DECL_DISREGARD_INLINE_LIMITS (callee->decl))
> > +      /* There are some options that change IL semantics which means
> > +         we cannot inline in these cases for correctness reason.
> > +	 Not even for always_inline declared functions.  */
> > +      /* Strictly speaking only when the callee contains signed integer
> > +         math where overflow is undefined.  */
> > +      if ((opt_for_fn (e->caller->decl, flag_strict_overflow)
> > +	   != opt_for_fn (e->caller->decl, flag_strict_overflow))
> > +	  || (opt_for_fn (e->caller->decl, flag_wrapv)
> > +	      != opt_for_fn (e->caller->decl, flag_wrapv))
> > +	  || (opt_for_fn (e->caller->decl, flag_trapv)
> > +	      != opt_for_fn (e->caller->decl, flag_trapv))
> > +	  /* Strictly speaking only when the callee contains memory
> > +	     accesses that are not using alias-set zero anyway.  */
> > +	  || (opt_for_fn (e->caller->decl, flag_strict_aliasing)
> > +	      != opt_for_fn (e->caller->decl, flag_strict_aliasing))
> > +	  /* Strictly speaking only when the callee uses FP math.  */
> > +	  || (opt_for_fn (e->caller->decl, flag_rounding_math)
> > +	      != opt_for_fn (e->caller->decl, flag_rounding_math))
> > +	  || (opt_for_fn (e->caller->decl, flag_trapping_math)
> > +	      != opt_for_fn (e->caller->decl, flag_trapping_math))
> > +	  || (opt_for_fn (e->caller->decl, flag_unsafe_math_optimizations)
> > +	      != opt_for_fn (e->caller->decl, flag_unsafe_math_optimizations))
> > +	  || (opt_for_fn (e->caller->decl, flag_finite_math_only)
> > +	      != opt_for_fn (e->caller->decl, flag_finite_math_only))
> > +	  || (opt_for_fn (e->caller->decl, flag_signaling_nans)
> > +	      != opt_for_fn (e->caller->decl, flag_signaling_nans))
> > +	  || (opt_for_fn (e->caller->decl, flag_cx_limited_range)
> > +	      != opt_for_fn (e->caller->decl, flag_cx_limited_range))
> > +	  || (opt_for_fn (e->caller->decl, flag_signed_zeros)
> > +	      != opt_for_fn (e->caller->decl, flag_signed_zeros))
> > +	  || (opt_for_fn (e->caller->decl, flag_associative_math)
> > +	      != opt_for_fn (e->caller->decl, flag_associative_math))
> > +	  || (opt_for_fn (e->caller->decl, flag_reciprocal_math)
> > +	      != opt_for_fn (e->caller->decl, flag_reciprocal_math))
> > +	  /* Strictly speaking only when the callee contains function
> > +	     calls that may end up setting errno.  */
> > +	  || (opt_for_fn (e->caller->decl, flag_errno_math)
> > +	      != opt_for_fn (e->caller->decl, flag_errno_math)))
> > +	{
> > +	  e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
> > +	  inlinable = false;
> > +	}
> > +      /* gcc.dg/pr43564.c.  Apply user-forced inline even at -O0.  */
> > +      else if (DECL_DISREGARD_INLINE_LIMITS (callee->decl)
> > +	       && lookup_attribute ("always_inline",
> > +				    DECL_ATTRIBUTES (callee->decl)))
> >   	;
> > -      /* When user added an attribute, honnor it.  */
> > -      else if ((lookup_attribute ("optimize", DECL_ATTRIBUTES
> > (caller->decl))
> > -		|| lookup_attribute ("optimize",
> > -				     DECL_ATTRIBUTES (callee->decl)))
> > -	       && ((opt_for_fn (caller->decl, optimize)
> > -		   > opt_for_fn (callee->decl, optimize))
> > -		   || (opt_for_fn (caller->decl, optimize_size)
> > -		       != opt_for_fn (callee->decl, optimize_size))))
> > +      /* When user added an attribute to the callee honor it.  */
> > +      else if (lookup_attribute ("optimize", DECL_ATTRIBUTES
> > (callee->decl))
> > +	       && opts_for_fn (caller->decl) != opts_for_fn (callee->decl))
> >   	{
> >   	  e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
> >   	  inlinable = false;
> > 
> 
>
Christian Bruel April 17, 2015, 6:51 a.m. UTC | #2 
On 01/22/2015 11:33 AM, Richard Biener wrote:
> On Thu, 22 Jan 2015, Christian Bruel wrote:
> 
>> Hi Richard,
>>
>> I thought one of my current issue would be solved by this patch, but it is not
>> : I have some inlining failures with the attribute target on ARM. (e.g
>> inline-3.c) where obvious early inline fails with because we fail into the
>> last can_inline_edge_p case:
>>
>> opt_for_fn (callee->decl, optimize)
>> 		   >= opt_for_fn (caller->decl, optimize)))
>>
>> when callee and caller are both -O2 and targetm.target_option.can_inline_p was
>> true, they should be inlined as in the general case (no
>> DECL_FUNCTION_SPECIFIC_OPTIMIZATION)
>>
>> I'm currently testing this additional change:
>>
>> Index: ipa-inline.c
>> ===================================================================
>> --- ipa-inline.c	(revision 219989)
>> +++ ipa-inline.c	(working copy)
>> @@ -489,7 +489,7 @@
>>        else if (opt_for_fn (callee->decl, optimize_size)
>>  	       < opt_for_fn (caller->decl, optimize_size)
>>  	       || (opt_for_fn (callee->decl, optimize)
>> -		   >= opt_for_fn (caller->decl, optimize)))
>> +		   > opt_for_fn (caller->decl, optimize)))
>>  	{
>>  	  if (estimate_edge_time (e)
>>  	      >= 20 + inline_edge_summary (e)->call_stmt_time)
>>
>> Since this is a hot topic for you, I though you would have useful comments on
>> this before I ask for a commit (when stage 4 close) ?
> 
> Yeah - the above looks like an obvious change to me.  Thus,
> approved if it passes bootstrap/regtest.
> 
> Thanks,
> Richard.
> 

thanks, sorry for the delay (stage1 blocked)

committed with the Changelog:

* ipa-inline.c (can_inline_edge_p): Allow inlining of functions with
same attributes.

This frees the road for

 - [PATCH, x86] [PR target/64835] Add
TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE hook
 https://gcc.gnu.org/ml/gcc-patches/2015-04/msg00594.html

 and

 - [PATCH, ARM] attribute target (thumb,arm) [0-6]
  https://gcc.gnu.org/ml/gcc-patches/2015-04/msg00706.html

with new regressions tests.
diff mbox

Patch

Index: ipa-inline.c
===================================================================
--- ipa-inline.c	(revision 219989)
+++ ipa-inline.c	(working copy)
@@ -489,7 +489,7 @@ 
        else if (opt_for_fn (callee->decl, optimize_size)
  	       < opt_for_fn (caller->decl, optimize_size)
  	       || (opt_for_fn (callee->decl, optimize)
-		   >= opt_for_fn (caller->decl, optimize)))
+		   > opt_for_fn (caller->decl, optimize)))
  	{
  	  if (estimate_edge_time (e)
  	      >= 20 + inline_edge_summary (e)->call_stmt_time)