[PATCH] Move X - (X / Y) * Y folding to match.pd

[PATCH] Move X - (X / Y) * Y folding to match.pd

[Marek Polacek]

This is an attempt to move one pattern from fold-const.c to match.pd.
It ought to be 1:1, but is not, e.g. with this patch we won't fold e.g.

int
f (int a, int b)
{
  return a - (unsigned) ((a / b) * b)
}

anymore, but we're able to fold

int
ff (int a, unsigned int b)
{
  return a - ((a / b) * b); 
}

and fold-const.c is not.  I played around with converts, but didn't find
anything that would work well.  Any suggestions how to make this pattern
better?

More to come...

Bootstrapped/regtested on x86_64-linux, ok for trunk?

2015-06-26  Marek Polacek  <polacek@redhat.com>

	* fold-const.c (fold_binary_loc): Move X - (X / Y) * Y -> X % Y to ...
	* match.pd: ... pattern here.

diff --git gcc/fold-const.c gcc/fold-const.c
index 6f12dd0..01e3983 100644
--- gcc/fold-const.c
+++ gcc/fold-const.c
@@ -10509,19 +10509,6 @@ fold_binary_loc (location_t loc,
 			    fold_convert_loc (loc, type,
 					      TREE_OPERAND (arg0, 0)));
 
-      /* X - (X / Y) * Y is X % Y.  */
-      if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
-	  && TREE_CODE (arg1) == MULT_EXPR
-	  && TREE_CODE (TREE_OPERAND (arg1, 0)) == TRUNC_DIV_EXPR
-	  && operand_equal_p (arg0,
-			      TREE_OPERAND (TREE_OPERAND (arg1, 0), 0), 0)
-	  && operand_equal_p (TREE_OPERAND (TREE_OPERAND (arg1, 0), 1),
-			      TREE_OPERAND (arg1, 1), 0))
-	return
-	  fold_convert_loc (loc, type,
-			    fold_build2_loc (loc, TRUNC_MOD_EXPR, TREE_TYPE (arg0),
-					 arg0, TREE_OPERAND (arg1, 1)));
-
       if (! FLOAT_TYPE_P (type))
 	{
 	  /* Fold A - (A & B) into ~B & A.  */
diff --git gcc/match.pd gcc/match.pd
index b2f8429..2bc158b 100644
--- gcc/match.pd
+++ gcc/match.pd
@@ -238,6 +238,12 @@ along with GCC; see the file COPYING3.  If not see
       && tree_nop_conversion_p (type, TREE_TYPE (@1)))
   (trunc_mod @0 (convert @1))))
 
+/* X - (X / Y) * Y is the same as X % Y.  */
+(simplify
+ (minus @0 (mult (trunc_div @0 @1) @1))
+ (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
+  (trunc_mod @0 @1)))
+
 /* Optimize TRUNC_MOD_EXPR by a power of two into a BIT_AND_EXPR,
    i.e. "X % C" into "X & (C - 1)", if X and C are positive.
    Also optimize A % (C << N)  where C is a power of 2,

	Marek

Re: [PATCH] Move X - (X / Y) * Y folding to match.pd

[Marc Glisse]

On Fri, 26 Jun 2015, Marek Polacek wrote:

> This is an attempt to move one pattern from fold-const.c to match.pd.
> It ought to be 1:1, but is not, e.g. with this patch we won't fold e.g.
>
> int
> f (int a, int b)
> {
>  return a - (unsigned) ((a / b) * b)
> }
>
> anymore, but we're able to fold
>
> int
> ff (int a, unsigned int b)
> {
>  return a - ((a / b) * b);
> }
>
> and fold-const.c is not.  I played around with converts, but didn't find
> anything that would work well.  Any suggestions how to make this pattern
> better?

Anything wrong with this?

+/* X - (X / Y) * Y is the same as X % Y.  */
+(simplify
+ (minus (convert? @0) (convert? (mult (trunc_div @0 @1) @1)))
+ (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
+  (convert (trunc_mod @0 @1))))

(the other div/mod pairs could benefit from the same transformation as 
long as there are no conversions, but the conversion seems easier to 
handle with trunc_)

> diff --git gcc/match.pd gcc/match.pd
> index b2f8429..2bc158b 100644
> --- gcc/match.pd
> +++ gcc/match.pd
> @@ -238,6 +238,12 @@ along with GCC; see the file COPYING3.  If not see
>       && tree_nop_conversion_p (type, TREE_TYPE (@1)))
>   (trunc_mod @0 (convert @1))))
>
> +/* X - (X / Y) * Y is the same as X % Y.  */
> +(simplify
> + (minus @0 (mult (trunc_div @0 @1) @1))
> + (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
> +  (trunc_mod @0 @1)))
> +
> /* Optimize TRUNC_MOD_EXPR by a power of two into a BIT_AND_EXPR,
>    i.e. "X % C" into "X & (C - 1)", if X and C are positive.
>    Also optimize A % (C << N)  where C is a power of 2,
>
> 	Marek
>

-- 
Marc Glisse

Re: [PATCH] Move X - (X / Y) * Y folding to match.pd

[Richard Biener]

On Sat, 27 Jun 2015, Marc Glisse wrote:

> On Fri, 26 Jun 2015, Marek Polacek wrote:
> 
> > This is an attempt to move one pattern from fold-const.c to match.pd.
> > It ought to be 1:1, but is not, e.g. with this patch we won't fold e.g.
> > 
> > int
> > f (int a, int b)
> > {
> >  return a - (unsigned) ((a / b) * b)
> > }
> > 
> > anymore, but we're able to fold
> > 
> > int
> > ff (int a, unsigned int b)
> > {
> >  return a - ((a / b) * b);
> > }
> > 
> > and fold-const.c is not.

Interesting.

> >  I played around with converts, but didn't find
> > anything that would work well.  Any suggestions how to make this pattern
> > better?
> 
> Anything wrong with this?
> 
> +/* X - (X / Y) * Y is the same as X % Y.  */
> +(simplify
> + (minus (convert? @0) (convert? (mult (trunc_div @0 @1) @1)))
> + (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
> +  (convert (trunc_mod @0 @1))))

Yes.  Eventually even (convert? (mult (convert1? (trunc_div ...)?
Of course with matching @0 between the two operands of the minus
you constrain types quite a bit.

I'd say just single-step through fold and see what types it get
present when folding a - (unsigned) ((a / b) * b).

Thanks,
Richard.

> (the other div/mod pairs could benefit from the same transformation as long as
> there are no conversions, but the conversion seems easier to handle with
> trunc_)
> 
> > diff --git gcc/match.pd gcc/match.pd
> > index b2f8429..2bc158b 100644
> > --- gcc/match.pd
> > +++ gcc/match.pd
> > @@ -238,6 +238,12 @@ along with GCC; see the file COPYING3.  If not see
> >       && tree_nop_conversion_p (type, TREE_TYPE (@1)))
> >   (trunc_mod @0 (convert @1))))
> > 
> > +/* X - (X / Y) * Y is the same as X % Y.  */
> > +(simplify
> > + (minus @0 (mult (trunc_div @0 @1) @1))
> > + (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
> > +  (trunc_mod @0 @1)))
> > +
> > /* Optimize TRUNC_MOD_EXPR by a power of two into a BIT_AND_EXPR,
> >    i.e. "X % C" into "X & (C - 1)", if X and C are positive.
> >    Also optimize A % (C << N)  where C is a power of 2,
> > 
> > 	Marek
> > 
> 
> 

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

Re: [PATCH] Move X - (X / Y) * Y folding to match.pd

[Marek Polacek]

On Mon, Jun 29, 2015 at 09:36:59AM +0200, Richard Biener wrote:
> > Anything wrong with this?
> > 
> > +/* X - (X / Y) * Y is the same as X % Y.  */
> > +(simplify
> > + (minus (convert? @0) (convert? (mult (trunc_div @0 @1) @1)))
> > + (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
> > +  (convert (trunc_mod @0 @1))))
 
That looks awfully similar to a variant I also tried (but I remember
having convert1? and convert2? in it).  Not sure what was wrong with
that one; certainly yours seems to work fine.

Not sure whether we need some tree_nop_conversion_p in it.  Perhaps not.

> Yes.  Eventually even (convert? (mult (convert1? (trunc_div ...)?
> Of course with matching @0 between the two operands of the minus
> you constrain types quite a bit.

I'm starting to dislike this whole convert business ;).

> I'd say just single-step through fold and see what types it get
> present when folding a - (unsigned) ((a / b) * b).

I did that.  The whole expression has type "unsigned int", arg0
is "a" of type int and arg1 is "(a / b) * b" of type int.

The following version is what Marc suggests.

Bootstrapped/regtested on x86_64-linux, ok for trunk?

2015-06-29  Marek Polacek  <polacek@redhat.com>
	    Marc Glisse  <marc.glisse@inria.fr>

	* fold-const.c (fold_binary_loc): Move X - (X / Y) * Y -> X % Y to ...
	* match.pd: ... pattern here.

diff --git gcc/fold-const.c gcc/fold-const.c
index 6f12dd0..01e3983 100644
--- gcc/fold-const.c
+++ gcc/fold-const.c
@@ -10509,19 +10509,6 @@ fold_binary_loc (location_t loc,
 			    fold_convert_loc (loc, type,
 					      TREE_OPERAND (arg0, 0)));
 
-      /* X - (X / Y) * Y is X % Y.  */
-      if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
-	  && TREE_CODE (arg1) == MULT_EXPR
-	  && TREE_CODE (TREE_OPERAND (arg1, 0)) == TRUNC_DIV_EXPR
-	  && operand_equal_p (arg0,
-			      TREE_OPERAND (TREE_OPERAND (arg1, 0), 0), 0)
-	  && operand_equal_p (TREE_OPERAND (TREE_OPERAND (arg1, 0), 1),
-			      TREE_OPERAND (arg1, 1), 0))
-	return
-	  fold_convert_loc (loc, type,
-			    fold_build2_loc (loc, TRUNC_MOD_EXPR, TREE_TYPE (arg0),
-					 arg0, TREE_OPERAND (arg1, 1)));
-
       if (! FLOAT_TYPE_P (type))
 	{
 	  /* Fold A - (A & B) into ~B & A.  */
diff --git gcc/match.pd gcc/match.pd
index b2f8429..2bc158b 100644
--- gcc/match.pd
+++ gcc/match.pd
@@ -238,6 +238,12 @@ along with GCC; see the file COPYING3.  If not see
       && tree_nop_conversion_p (type, TREE_TYPE (@1)))
   (trunc_mod @0 (convert @1))))
 
+/* X - (X / Y) * Y is the same as X % Y.  */
+(simplify
+ (minus (convert? @0) (convert? (mult (trunc_div @0 @1) @1)))
+ (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
+  (convert (trunc_mod @0 @1))))
+
 /* Optimize TRUNC_MOD_EXPR by a power of two into a BIT_AND_EXPR,
    i.e. "X % C" into "X & (C - 1)", if X and C are positive.
    Also optimize A % (C << N)  where C is a power of 2,

	Marek

Re: [PATCH] Move X - (X / Y) * Y folding to match.pd

[Richard Biener]

On Mon, 29 Jun 2015, Marek Polacek wrote:

> On Mon, Jun 29, 2015 at 09:36:59AM +0200, Richard Biener wrote:
> > > Anything wrong with this?
> > > 
> > > +/* X - (X / Y) * Y is the same as X % Y.  */
> > > +(simplify
> > > + (minus (convert? @0) (convert? (mult (trunc_div @0 @1) @1)))
> > > + (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
> > > +  (convert (trunc_mod @0 @1))))
>  
> That looks awfully similar to a variant I also tried (but I remember
> having convert1? and convert2? in it).  Not sure what was wrong with
> that one; certainly yours seems to work fine.
> 
> Not sure whether we need some tree_nop_conversion_p in it.  Perhaps not.
> 
> > Yes.  Eventually even (convert? (mult (convert1? (trunc_div ...)?
> > Of course with matching @0 between the two operands of the minus
> > you constrain types quite a bit.
> 
> I'm starting to dislike this whole convert business ;).

;)  fold-const.c STRIP_NOPS certainly was both convenient and
error-prone at the same time.

> > I'd say just single-step through fold and see what types it get
> > present when folding a - (unsigned) ((a / b) * b).
> 
> I did that.  The whole expression has type "unsigned int", arg0
> is "a" of type int and arg1 is "(a / b) * b" of type int.
> 
> The following version is what Marc suggests.
> 
> Bootstrapped/regtested on x86_64-linux, ok for trunk?

Ok.

Thanks,
Richard.

> 2015-06-29  Marek Polacek  <polacek@redhat.com>
> 	    Marc Glisse  <marc.glisse@inria.fr>
> 
> 	* fold-const.c (fold_binary_loc): Move X - (X / Y) * Y -> X % Y to ...
> 	* match.pd: ... pattern here.
> 
> diff --git gcc/fold-const.c gcc/fold-const.c
> index 6f12dd0..01e3983 100644
> --- gcc/fold-const.c
> +++ gcc/fold-const.c
> @@ -10509,19 +10509,6 @@ fold_binary_loc (location_t loc,
>  			    fold_convert_loc (loc, type,
>  					      TREE_OPERAND (arg0, 0)));
>  
> -      /* X - (X / Y) * Y is X % Y.  */
> -      if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
> -	  && TREE_CODE (arg1) == MULT_EXPR
> -	  && TREE_CODE (TREE_OPERAND (arg1, 0)) == TRUNC_DIV_EXPR
> -	  && operand_equal_p (arg0,
> -			      TREE_OPERAND (TREE_OPERAND (arg1, 0), 0), 0)
> -	  && operand_equal_p (TREE_OPERAND (TREE_OPERAND (arg1, 0), 1),
> -			      TREE_OPERAND (arg1, 1), 0))
> -	return
> -	  fold_convert_loc (loc, type,
> -			    fold_build2_loc (loc, TRUNC_MOD_EXPR, TREE_TYPE (arg0),
> -					 arg0, TREE_OPERAND (arg1, 1)));
> -
>        if (! FLOAT_TYPE_P (type))
>  	{
>  	  /* Fold A - (A & B) into ~B & A.  */
> diff --git gcc/match.pd gcc/match.pd
> index b2f8429..2bc158b 100644
> --- gcc/match.pd
> +++ gcc/match.pd
> @@ -238,6 +238,12 @@ along with GCC; see the file COPYING3.  If not see
>        && tree_nop_conversion_p (type, TREE_TYPE (@1)))
>    (trunc_mod @0 (convert @1))))
>  
> +/* X - (X / Y) * Y is the same as X % Y.  */
> +(simplify
> + (minus (convert? @0) (convert? (mult (trunc_div @0 @1) @1)))
> + (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
> +  (convert (trunc_mod @0 @1))))
> +
>  /* Optimize TRUNC_MOD_EXPR by a power of two into a BIT_AND_EXPR,
>     i.e. "X % C" into "X & (C - 1)", if X and C are positive.
>     Also optimize A % (C << N)  where C is a power of 2,
> 
> 	Marek
> 
> 

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

Re: [PATCH] Move X - (X / Y) * Y folding to match.pd

[Marc Glisse]

On Mon, 29 Jun 2015, Marek Polacek wrote:

> On Mon, Jun 29, 2015 at 09:36:59AM +0200, Richard Biener wrote:
>>> Anything wrong with this?
>>>
>>> +/* X - (X / Y) * Y is the same as X % Y.  */
>>> +(simplify
>>> + (minus (convert? @0) (convert? (mult (trunc_div @0 @1) @1)))
>>> + (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
>>> +  (convert (trunc_mod @0 @1))))
>
> That looks awfully similar to a variant I also tried (but I remember
> having convert1? and convert2? in it).  Not sure what was wrong with
> that one; certainly yours seems to work fine.

Afterwards I thought of a limitation. Nothing bad, but it highlights a 
trap I regularly fall into: several @0 in the same pattern may have 
different types (for INTEGER_CST, operand_equal_p mostly ignores the 
type). So for an int x, 42L-42/x*x should fail to match, while using 
convert1? and convert2? should match.

-- 
Marc Glisse

Re: [PATCH] Move X - (X / Y) * Y folding to match.pd

[Richard Biener]

On Tue, 30 Jun 2015, Marc Glisse wrote:

> On Mon, 29 Jun 2015, Marek Polacek wrote:
> 
> > On Mon, Jun 29, 2015 at 09:36:59AM +0200, Richard Biener wrote:
> > > > Anything wrong with this?
> > > > 
> > > > +/* X - (X / Y) * Y is the same as X % Y.  */
> > > > +(simplify
> > > > + (minus (convert? @0) (convert? (mult (trunc_div @0 @1) @1)))
> > > > + (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
> > > > +  (convert (trunc_mod @0 @1))))
> > 
> > That looks awfully similar to a variant I also tried (but I remember
> > having convert1? and convert2? in it).  Not sure what was wrong with
> > that one; certainly yours seems to work fine.
> 
> Afterwards I thought of a limitation. Nothing bad, but it highlights a trap I
> regularly fall into: several @0 in the same pattern may have different types
> (for INTEGER_CST, operand_equal_p mostly ignores the type). So for an int x,
> 42L-42/x*x should fail to match, while using convert1? and convert2? should
> match.

Indeed that's a subtle issue with using operand_equal_p for matching
operands.  Note that 42L-42/x*x will appear as
42L-(long)(42/x*x) in the IL just in case that wasn't obvious.

Thus ok to adjust the pattern to convert1? / convert2? if you add such
a testcase (maybe also add the one that the variants you tried on
originally failed to match).

Richard.