invert_tree_comparsion tweek for bettter inline predicates

invert_tree_comparsion tweek for bettter inline predicates

[Jan Hubicka]

Hi,
while looking on a testcase, i noticed that for simple code

  if (param > 6.0)
    BB1;
  else
    BB2;
the inline predicates currectly determine (param > 6.0) predicate
for BB1, but they give (true) predicate to BB2 unless -fno-trapping-math is
specified.  This is because invert_tree_comparison returns ERROR_MARK
when called with HONOR_NANS.

I see that this is most likely to preserve trappingness of the comparsion
- UNLE that is a logcial negative won't report unordered.  For my use it is
  however save to use (param UNLE 6.0) as predicate for BB2 as I do not really
care if the conditional will trap.  The inline predicates just need to be
conservative with respect to real program behaviour.
I can not pass HONOR_NANS=false because then the BB2 predicate would
be (param <= 6.0) and that would allow inliner to assume that for
param == UNORDERED both code paths are unreachable.

This patch adds extra paremter to invert_tree_comparsion to do what I want.
Does it seem sane?

Honza

	* fold-const.c (invert_tree_comparison): Add CONSIDER_TRAP.
	* fold-const.h (invert_tree_comparison): Update prototype.
	* ipa-inline-analysis.c (add_clause): Update use
	of invert_tree_comparsoin
	(set_cond_stmt_execution_predicate): Likewise.
Index: fold-const.c
===================================================================
--- fold-const.c	(revision 222052)
+++ fold-const.c	(working copy)
@@ -2436,13 +2436,17 @@ pedantic_non_lvalue_loc (location_t loc,
 /* Given a tree comparison code, return the code that is the logical inverse.
    It is generally not safe to do this for floating-point comparisons, except
    for EQ_EXPR, NE_EXPR, ORDERED_EXPR and UNORDERED_EXPR, so we return
-   ERROR_MARK in this case.  */
+   ERROR_MARK in this case.
+
+   if CONSDIER_TRAP is false, do not really care about the case when
+   conditional expression may trap.  */
 
 enum tree_code
-invert_tree_comparison (enum tree_code code, bool honor_nans)
+invert_tree_comparison (enum tree_code code, bool honor_nans,
+		        bool consider_trap)
 {
-  if (honor_nans && flag_trapping_math && code != EQ_EXPR && code != NE_EXPR
-      && code != ORDERED_EXPR && code != UNORDERED_EXPR)
+  if (honor_nans && consider_trap && flag_trapping_math && code != EQ_EXPR
+      && code != NE_EXPR && code != ORDERED_EXPR && code != UNORDERED_EXPR)
     return ERROR_MARK;
 
   switch (code)
Index: fold-const.h
===================================================================
--- fold-const.h	(revision 222052)
+++ fold-const.h	(working copy)
@@ -126,7 +126,8 @@ extern bool tree_swap_operands_p (const_
 extern enum tree_code swap_tree_comparison (enum tree_code);
 
 extern bool ptr_difference_const (tree, tree, HOST_WIDE_INT *);
-extern enum tree_code invert_tree_comparison (enum tree_code, bool);
+extern enum tree_code invert_tree_comparison (enum tree_code, bool,
+					      bool consider_trap = false);
 
 extern bool tree_unary_nonzero_warnv_p (enum tree_code, tree, tree, bool *);
 extern bool tree_binary_nonzero_warnv_p (enum tree_code, tree, tree, tree op1,
Index: ipa-inline-analysis.c
===================================================================
--- ipa-inline-analysis.c	(revision 222052)
+++ ipa-inline-analysis.c	(working copy)
@@ -381,7 +381,8 @@ add_clause (conditions conditions, struc
 		&& cc2->code != IS_NOT_CONSTANT
 		&& cc2->code != CHANGED
 		&& cc1->code == invert_tree_comparison (cc2->code,
-							HONOR_NANS (cc1->val)))
+							HONOR_NANS (cc1->val),
+							false))
 	      return;
 	  }
     }
@@ -1798,15 +1799,13 @@ set_cond_stmt_execution_predicate (struc
   if (unmodified_parm_or_parm_agg_item (info, last, op, &index, &aggpos))
     {
       code = gimple_cond_code (last);
-      inverted_code = invert_tree_comparison (code, HONOR_NANS (op));
+      inverted_code = invert_tree_comparison (code, HONOR_NANS (op), false);
 
       FOR_EACH_EDGE (e, ei, bb->succs)
 	{
 	  enum tree_code this_code = (e->flags & EDGE_TRUE_VALUE
 				      ? code : inverted_code);
-	  /* invert_tree_comparison will return ERROR_MARK on FP
-	     comparsions that are not EQ/NE instead of returning proper
-	     unordered one.  Be sure it is not confused with NON_CONSTANT.  */
+	  /*  Be sure to not confuse ERROR_MARK with NON_CONSTANT.  */
 	  if (this_code != ERROR_MARK)
 	    {
 	      struct predicate p = add_condition (summary, index, &aggpos,

Re: invert_tree_comparsion tweek for bettter inline predicates

[Richard Biener]

On Tue, 14 Apr 2015, Jan Hubicka wrote:

> Hi,
> while looking on a testcase, i noticed that for simple code
> 
>   if (param > 6.0)
>     BB1;
>   else
>     BB2;
> the inline predicates currectly determine (param > 6.0) predicate
> for BB1, but they give (true) predicate to BB2 unless -fno-trapping-math is
> specified.  This is because invert_tree_comparison returns ERROR_MARK
> when called with HONOR_NANS.
> 
> I see that this is most likely to preserve trappingness of the comparsion
> - UNLE that is a logcial negative won't report unordered.  For my use it is
>   however save to use (param UNLE 6.0) as predicate for BB2 as I do not really
> care if the conditional will trap.  The inline predicates just need to be
> conservative with respect to real program behaviour.
> I can not pass HONOR_NANS=false because then the BB2 predicate would
> be (param <= 6.0) and that would allow inliner to assume that for
> param == UNORDERED both code paths are unreachable.
> 
> This patch adds extra paremter to invert_tree_comparsion to do what I want.
> Does it seem sane?

You could also temporarily un-set flag_trapping_math around the call.

That said, invert_tree_comparison should get flag_trapping_math
as argument, not another that ignores it.  That would be a cleaner
interface.

You can add an inline overload

inline enum tree_code
invert_tree_comparison (enum tree_code code, bool honor_nans)
{
  return invert_tree_comparison (code, honor_nans, flag_trapping_math);
}

to make old code work without change.

Richard.

> Honza
> 
> 	* fold-const.c (invert_tree_comparison): Add CONSIDER_TRAP.
> 	* fold-const.h (invert_tree_comparison): Update prototype.
> 	* ipa-inline-analysis.c (add_clause): Update use
> 	of invert_tree_comparsoin
> 	(set_cond_stmt_execution_predicate): Likewise.
> Index: fold-const.c
> ===================================================================
> --- fold-const.c	(revision 222052)
> +++ fold-const.c	(working copy)
> @@ -2436,13 +2436,17 @@ pedantic_non_lvalue_loc (location_t loc,
>  /* Given a tree comparison code, return the code that is the logical inverse.
>     It is generally not safe to do this for floating-point comparisons, except
>     for EQ_EXPR, NE_EXPR, ORDERED_EXPR and UNORDERED_EXPR, so we return
> -   ERROR_MARK in this case.  */
> +   ERROR_MARK in this case.
> +
> +   if CONSDIER_TRAP is false, do not really care about the case when
> +   conditional expression may trap.  */
>  
>  enum tree_code
> -invert_tree_comparison (enum tree_code code, bool honor_nans)
> +invert_tree_comparison (enum tree_code code, bool honor_nans,
> +		        bool consider_trap)
>  {
> -  if (honor_nans && flag_trapping_math && code != EQ_EXPR && code != NE_EXPR
> -      && code != ORDERED_EXPR && code != UNORDERED_EXPR)
> +  if (honor_nans && consider_trap && flag_trapping_math && code != EQ_EXPR
> +      && code != NE_EXPR && code != ORDERED_EXPR && code != UNORDERED_EXPR)
>      return ERROR_MARK;
>  
>    switch (code)
> Index: fold-const.h
> ===================================================================
> --- fold-const.h	(revision 222052)
> +++ fold-const.h	(working copy)
> @@ -126,7 +126,8 @@ extern bool tree_swap_operands_p (const_
>  extern enum tree_code swap_tree_comparison (enum tree_code);
>  
>  extern bool ptr_difference_const (tree, tree, HOST_WIDE_INT *);
> -extern enum tree_code invert_tree_comparison (enum tree_code, bool);
> +extern enum tree_code invert_tree_comparison (enum tree_code, bool,
> +					      bool consider_trap = false);
>  
>  extern bool tree_unary_nonzero_warnv_p (enum tree_code, tree, tree, bool *);
>  extern bool tree_binary_nonzero_warnv_p (enum tree_code, tree, tree, tree op1,
> Index: ipa-inline-analysis.c
> ===================================================================
> --- ipa-inline-analysis.c	(revision 222052)
> +++ ipa-inline-analysis.c	(working copy)
> @@ -381,7 +381,8 @@ add_clause (conditions conditions, struc
>  		&& cc2->code != IS_NOT_CONSTANT
>  		&& cc2->code != CHANGED
>  		&& cc1->code == invert_tree_comparison (cc2->code,
> -							HONOR_NANS (cc1->val)))
> +							HONOR_NANS (cc1->val),
> +							false))
>  	      return;
>  	  }
>      }
> @@ -1798,15 +1799,13 @@ set_cond_stmt_execution_predicate (struc
>    if (unmodified_parm_or_parm_agg_item (info, last, op, &index, &aggpos))
>      {
>        code = gimple_cond_code (last);
> -      inverted_code = invert_tree_comparison (code, HONOR_NANS (op));
> +      inverted_code = invert_tree_comparison (code, HONOR_NANS (op), false);
>  
>        FOR_EACH_EDGE (e, ei, bb->succs)
>  	{
>  	  enum tree_code this_code = (e->flags & EDGE_TRUE_VALUE
>  				      ? code : inverted_code);
> -	  /* invert_tree_comparison will return ERROR_MARK on FP
> -	     comparsions that are not EQ/NE instead of returning proper
> -	     unordered one.  Be sure it is not confused with NON_CONSTANT.  */
> +	  /*  Be sure to not confuse ERROR_MARK with NON_CONSTANT.  */
>  	  if (this_code != ERROR_MARK)
>  	    {
>  	      struct predicate p = add_condition (summary, index, &aggpos,
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Jennifer Guild,
Dilip Upmanyu, Graham Norton HRB 21284 (AG Nuernberg)

Re: invert_tree_comparsion tweek for bettter inline predicates

[Bernhard Reutner-Fischer]

On April 14, 2015 1:36:14 AM GMT+02:00, Jan Hubicka <hubicka@ucw.cz> wrote:
>Hi,
>while looking on a testcase, i noticed that for simple code
>
>  if (param > 6.0)
>    BB1;
>  else
>    BB2;
>the inline predicates currectly determine (param > 6.0) predicate
>for BB1, but they give (true) predicate to BB2 unless
>-fno-trapping-math is
>specified.  This is because invert_tree_comparison returns ERROR_MARK
>when called with HONOR_NANS.
>
>I see that this is most likely to preserve trappingness of the
>comparsion
>- UNLE that is a logcial negative won't report unordered.  For my use
>it is
>however save to use (param UNLE 6.0) as predicate for BB2 as I do not
>really
>care if the conditional will trap.  The inline predicates just need to
>be
>conservative with respect to real program behaviour.
>I can not pass HONOR_NANS=false because then the BB2 predicate would
>be (param <= 6.0) and that would allow inliner to assume that for
>param == UNORDERED both code paths are unreachable.
>
>This patch adds extra paremter to invert_tree_comparsion to do what I
>want.
>Does it seem sane?

Don't you have to default CONSIDER_TRAP to true to retain previous behaviour for e.g. fold_truth_not_expr and other users that pass down HONOR_NANS to honour_nans?

Thanks,
>
>Honza
>
>	* fold-const.c (invert_tree_comparison): Add CONSIDER_TRAP.
>	* fold-const.h (invert_tree_comparison): Update prototype.
>	* ipa-inline-analysis.c (add_clause): Update use
>	of invert_tree_comparsoin
>	(set_cond_stmt_execution_predicate): Likewise.
>Index: fold-const.c
>===================================================================
>--- fold-const.c	(revision 222052)
>+++ fold-const.c	(working copy)
>@@ -2436,13 +2436,17 @@ pedantic_non_lvalue_loc (location_t loc,
>/* Given a tree comparison code, return the code that is the logical
>inverse.
>It is generally not safe to do this for floating-point comparisons,
>except
>    for EQ_EXPR, NE_EXPR, ORDERED_EXPR and UNORDERED_EXPR, so we return
>-   ERROR_MARK in this case.  */
>+   ERROR_MARK in this case.
>+
>+   if CONSDIER_TRAP is false, do not really care about the case when
>+   conditional expression may trap.  */
> 
> enum tree_code
>-invert_tree_comparison (enum tree_code code, bool honor_nans)
>+invert_tree_comparison (enum tree_code code, bool honor_nans,
>+		        bool consider_trap)
> {
>-  if (honor_nans && flag_trapping_math && code != EQ_EXPR && code !=
>NE_EXPR
>-      && code != ORDERED_EXPR && code != UNORDERED_EXPR)
>+  if (honor_nans && consider_trap && flag_trapping_math && code !=
>EQ_EXPR
>+      && code != NE_EXPR && code != ORDERED_EXPR && code !=
>UNORDERED_EXPR)
>     return ERROR_MARK;
> 
>   switch (code)
>Index: fold-const.h
>===================================================================
>--- fold-const.h	(revision 222052)
>+++ fold-const.h	(working copy)
>@@ -126,7 +126,8 @@ extern bool tree_swap_operands_p (const_
> extern enum tree_code swap_tree_comparison (enum tree_code);
> 
> extern bool ptr_difference_const (tree, tree, HOST_WIDE_INT *);
>-extern enum tree_code invert_tree_comparison (enum tree_code, bool);
>+extern enum tree_code invert_tree_comparison (enum tree_code, bool,
>+					      bool consider_trap = false);
> 
>extern bool tree_unary_nonzero_warnv_p (enum tree_code, tree, tree,
>bool *);
>extern bool tree_binary_nonzero_warnv_p (enum tree_code, tree, tree,
>tree op1,
>Index: ipa-inline-analysis.c
>===================================================================
>--- ipa-inline-analysis.c	(revision 222052)
>+++ ipa-inline-analysis.c	(working copy)
>@@ -381,7 +381,8 @@ add_clause (conditions conditions, struc
> 		&& cc2->code != IS_NOT_CONSTANT
> 		&& cc2->code != CHANGED
> 		&& cc1->code == invert_tree_comparison (cc2->code,
>-							HONOR_NANS (cc1->val)))
>+							HONOR_NANS (cc1->val),
>+							false))
> 	      return;
> 	  }
>     }
>@@ -1798,15 +1799,13 @@ set_cond_stmt_execution_predicate (struc
>if (unmodified_parm_or_parm_agg_item (info, last, op, &index, &aggpos))
>     {
>       code = gimple_cond_code (last);
>-      inverted_code = invert_tree_comparison (code, HONOR_NANS (op));
>+      inverted_code = invert_tree_comparison (code, HONOR_NANS (op),
>false);
> 
>       FOR_EACH_EDGE (e, ei, bb->succs)
> 	{
> 	  enum tree_code this_code = (e->flags & EDGE_TRUE_VALUE
> 				      ? code : inverted_code);
>-	  /* invert_tree_comparison will return ERROR_MARK on FP
>-	     comparsions that are not EQ/NE instead of returning proper
>-	     unordered one.  Be sure it is not confused with NON_CONSTANT. 
>*/
>+	  /*  Be sure to not confuse ERROR_MARK with NON_CONSTANT.  */
> 	  if (this_code != ERROR_MARK)
> 	    {
> 	      struct predicate p = add_condition (summary, index, &aggpos,