Reduce number of sreal operator* calls

Message ID alpine.LSU.2.20.1811071244260.1827@zhemvz.fhfr.qr
State New
Headers show
Series
  • Reduce number of sreal operator* calls
Related show

Commit Message

Richard Biener Nov. 7, 2018, 12:11 p.m.
This reduces the number of $subject calls by computing big_speedup_p
lazily.  This caller accounts for roughly a quarter of all operator*
calls for PR38474 and operator* is top of the profile of the whole
compilation.

Next offenders (callers) are compute_inlined_call_time and
edge_badness.  profile_count::to_sreal_scale is also quite
bad in performance btw (probably due to the sreal division).

Bootstrap/regtest in progress.

OK?

Thanks,
Richard.

2018-11-07  Richard Biener  <rguenther@suse.de>

	* ipa-inline.c (want_inline_small_function_p): Compute
	big_speedup_p lazily and last.

Comments

Jan Hubicka Nov. 7, 2018, 12:24 p.m. | #1
> 
> This reduces the number of $subject calls by computing big_speedup_p
> lazily.  This caller accounts for roughly a quarter of all operator*
> calls for PR38474 and operator* is top of the profile of the whole
> compilation.
> 
> Next offenders (callers) are compute_inlined_call_time and
> edge_badness.  profile_count::to_sreal_scale is also quite
> bad in performance btw (probably due to the sreal division).
> 
> Bootstrap/regtest in progress.
> 
> OK?

OK, thanks!
Honza
> 
> Thanks,
> Richard.
> 
> 2018-11-07  Richard Biener  <rguenther@suse.de>
> 
> 	* ipa-inline.c (want_inline_small_function_p): Compute
> 	big_speedup_p lazily and last.
> 
> Index: gcc/ipa-inline.c
> ===================================================================
> --- gcc/ipa-inline.c	(revision 265860)
> +++ gcc/ipa-inline.c	(working copy)
> @@ -779,7 +779,7 @@ want_inline_small_function_p (struct cgr
>      {
>        int growth = estimate_edge_growth (e);
>        ipa_hints hints = estimate_edge_hints (e);
> -      bool big_speedup = big_speedup_p (e);
> +      int big_speedup = -1; /* compute this lazily */
>  
>        if (growth <= 0)
>  	;
> @@ -787,13 +787,13 @@ want_inline_small_function_p (struct cgr
>  	 hints suggests that inlining given function is very profitable.  */
>        else if (DECL_DECLARED_INLINE_P (callee->decl)
>  	       && growth >= MAX_INLINE_INSNS_SINGLE
> -	       && ((!big_speedup
> -		    && !(hints & (INLINE_HINT_indirect_call
> +	       && (growth >= MAX_INLINE_INSNS_SINGLE * 16
> +		   || (!(hints & (INLINE_HINT_indirect_call
>  				  | INLINE_HINT_known_hot
>  				  | INLINE_HINT_loop_iterations
>  				  | INLINE_HINT_array_index
> -				  | INLINE_HINT_loop_stride)))
> -		   || growth >= MAX_INLINE_INSNS_SINGLE * 16))
> +				  | INLINE_HINT_loop_stride))
> +		       && !(big_speedup = big_speedup_p (e)))))
>  	{
>            e->inline_failed = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT;
>  	  want_inline = false;
> @@ -813,7 +813,6 @@ want_inline_small_function_p (struct cgr
>  	 Upgrade it to MAX_INLINE_INSNS_SINGLE when hints suggests that
>  	 inlining given function is very profitable.  */
>        else if (!DECL_DECLARED_INLINE_P (callee->decl)
> -	       && !big_speedup
>  	       && !(hints & INLINE_HINT_known_hot)
>  	       && growth >= ((hints & (INLINE_HINT_indirect_call
>  				       | INLINE_HINT_loop_iterations
> @@ -821,7 +820,8 @@ want_inline_small_function_p (struct cgr
>  				       | INLINE_HINT_loop_stride))
>  			     ? MAX (MAX_INLINE_INSNS_AUTO,
>  				    MAX_INLINE_INSNS_SINGLE)
> -			     : MAX_INLINE_INSNS_AUTO))
> +			     : MAX_INLINE_INSNS_AUTO)
> +	       && !(big_speedup == -1 ? big_speedup_p (e) : big_speedup))
>  	{
>  	  /* growth_likely_positive is expensive, always test it last.  */
>            if (growth >= MAX_INLINE_INSNS_SINGLE

Patch

Index: gcc/ipa-inline.c
===================================================================
--- gcc/ipa-inline.c	(revision 265860)
+++ gcc/ipa-inline.c	(working copy)
@@ -779,7 +779,7 @@  want_inline_small_function_p (struct cgr
     {
       int growth = estimate_edge_growth (e);
       ipa_hints hints = estimate_edge_hints (e);
-      bool big_speedup = big_speedup_p (e);
+      int big_speedup = -1; /* compute this lazily */
 
       if (growth <= 0)
 	;
@@ -787,13 +787,13 @@  want_inline_small_function_p (struct cgr
 	 hints suggests that inlining given function is very profitable.  */
       else if (DECL_DECLARED_INLINE_P (callee->decl)
 	       && growth >= MAX_INLINE_INSNS_SINGLE
-	       && ((!big_speedup
-		    && !(hints & (INLINE_HINT_indirect_call
+	       && (growth >= MAX_INLINE_INSNS_SINGLE * 16
+		   || (!(hints & (INLINE_HINT_indirect_call
 				  | INLINE_HINT_known_hot
 				  | INLINE_HINT_loop_iterations
 				  | INLINE_HINT_array_index
-				  | INLINE_HINT_loop_stride)))
-		   || growth >= MAX_INLINE_INSNS_SINGLE * 16))
+				  | INLINE_HINT_loop_stride))
+		       && !(big_speedup = big_speedup_p (e)))))
 	{
           e->inline_failed = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT;
 	  want_inline = false;
@@ -813,7 +813,6 @@  want_inline_small_function_p (struct cgr
 	 Upgrade it to MAX_INLINE_INSNS_SINGLE when hints suggests that
 	 inlining given function is very profitable.  */
       else if (!DECL_DECLARED_INLINE_P (callee->decl)
-	       && !big_speedup
 	       && !(hints & INLINE_HINT_known_hot)
 	       && growth >= ((hints & (INLINE_HINT_indirect_call
 				       | INLINE_HINT_loop_iterations
@@ -821,7 +820,8 @@  want_inline_small_function_p (struct cgr
 				       | INLINE_HINT_loop_stride))
 			     ? MAX (MAX_INLINE_INSNS_AUTO,
 				    MAX_INLINE_INSNS_SINGLE)
-			     : MAX_INLINE_INSNS_AUTO))
+			     : MAX_INLINE_INSNS_AUTO)
+	       && !(big_speedup == -1 ? big_speedup_p (e) : big_speedup))
 	{
 	  /* growth_likely_positive is expensive, always test it last.  */
           if (growth >= MAX_INLINE_INSNS_SINGLE