RE: [PATCH]middle-end fix floating out of constants in conditionals

[Tamar Christina <Tamar.Christina@arm.com>]

> -----Original Message-----
> From: Richard Biener <rguenther@suse.de>
> Sent: Monday, September 26, 2022 11:28 AM
> To: Tamar Christina <Tamar.Christina@arm.com>
> Cc: gcc-patches@gcc.gnu.org; nd <nd@arm.com>; jeffreyalaw@gmail.com;
> ebotcazou@adacore.com
> Subject: Re: [PATCH]middle-end fix floating out of constants in conditionals
> 
> On Fri, 23 Sep 2022, Tamar Christina wrote:
> 
> > Hi All,
> >
> > The following testcase:
> >
> > int zoo1 (int a, int b, int c, int d)
> > {
> >    return (a > b ? c : d) & 1;
> > }
> >
> > gets de-optimized by the front-end since somewhere around GCC 4.x due
> > to a fix that was added to fold_binary_op_with_conditional_arg.
> >
> > The folding is supposed to succeed only if we have folded at least one
> > of the branches, however the check doesn't tests that all of the
> > values are non-constant.  So if one of the operators are a constant it
> accepts the folding.
> >
> > This ends up folding
> >
> >    return (a > b ? c : d) & 1;
> >
> > into
> >
> >    return (a > b ? c & 1 : d & 1);
> >
> > and thus performing the AND twice.
> >
> > change changes it to reject the folding if one of the arguments are a
> > constant and if the operations being performed are the same.
> >
> > Secondly it adds a new match.pd rule to now also fold the opposite
> > direction, so it now also folds:
> >
> >    return (a > b ? c & 1 : d & 1);
> >
> > into
> >
> >    return (a > b ? c : d) & 1;
> >
> > Bootstrapped Regtested on aarch64-none-linux-gnu, x86_64-pc-linux-gnu
> > and <on-goin> issues.
> >
> > Ok for master?
> >
> > Thanks,
> > Tamar
> >
> > gcc/ChangeLog:
> >
> > 	* fold-const.cc (fold_binary_op_with_conditional_arg): Add
> relaxation.
> > 	* match.pd: Add ternary constant fold rule.
> > 	* tree-cfg.cc (verify_gimple_assign_ternary): RHS1 of a COND_EXPR
> isn't
> > 	a value but an expression itself.
> >
> > gcc/testsuite/ChangeLog:
> >
> > 	* gcc.target/aarch64/if-compare_3.c: New test.
> >
> > --- inline copy of patch --
> > diff --git a/gcc/fold-const.cc b/gcc/fold-const.cc index
> >
> 4f4ec81c8d4b6937ade3141a14c695b67c874c35..0ee083f290d12104969f1b335d
> c3
> > 3917c97b4808 100644
> > --- a/gcc/fold-const.cc
> > +++ b/gcc/fold-const.cc
> > @@ -7212,7 +7212,9 @@ fold_binary_op_with_conditional_arg (location_t
> loc,
> >      }
> >
> >    /* Check that we have simplified at least one of the branches.  */
> > -  if (!TREE_CONSTANT (arg) && !TREE_CONSTANT (lhs) &&
> !TREE_CONSTANT
> > (rhs))
> > +  if ((!TREE_CONSTANT (arg) && !TREE_CONSTANT (lhs) &&
> !TREE_CONSTANT (rhs))
> > +      || (TREE_CONSTANT (arg) && TREE_CODE (lhs) == TREE_CODE (rhs)
> > +	  && !TREE_CONSTANT (lhs)))
> >      return NULL_TREE;
> 
> I think the better fix would be to only consider TREE_CONSTANT (arg) if it
> wasn't constant initially.  Because clearly "simplify" intends to be "constant"
> here.  In fact I wonder why we test !TREE_CONSTANT (arg) at all, we don't
> simplify 'arg' ...

The function allows this because even when !TREE_CONSTANT (arg) the
true_value or false_value can instead be constant.

Yes it's of limited use unless true_value and false_value are 0 or 1.

> 
> Eric added this test (previosuly we'd just always done the folding), but I think
> not enough?
> 
> >
> >    return fold_build3_loc (loc, cond_code, type, test, lhs, rhs); diff
> > --git a/gcc/match.pd b/gcc/match.pd index
> >
> b225d36dc758f1581502c8d03761544bfd499c01..b61ed70e69b881a49177f10f20
> c1
> > f92712bb8665 100644
> > --- a/gcc/match.pd
> > +++ b/gcc/match.pd
> > @@ -4318,6 +4318,16 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> >    (op @3 (vec_cond:s @0 @1 @2))
> >    (vec_cond @0 (op! @3 @1) (op! @3 @2))))
> >
> > +/* Float out binary operations from branches if they can't be folded.
> > +   Fold (a ? (b op c) : (d op c)) --> (op (a ? b : d) c).  */ (for op
> > +(plus mult min max bit_and bit_ior bit_xor minus lshift rshift rdiv
> > +	 trunc_div ceil_div floor_div round_div trunc_mod ceil_mod
> floor_mod
> > +	 round_mod)
> > + (simplify
> > +  (cond @0 (op @1 @2) (op @3 @2))
> > +   (if (!FLOAT_TYPE_P (type) || !(HONOR_NANS (@1) &&
> flag_trapping_math))
> > +    (op (cond @0 @1 @3) @2))))
> 
> Ick.  Adding a reverse tranform is going to be prone to recursing :/
> 
> Why do you need to care about NANs or FP exceptions?

Because the function in fold-const.cc has a check in the start:

  /* Do not move possibly trapping operations into the conditional as this
     pessimizes code and causes gimplification issues when applied late.  */

And so I stayed conservative and just didn't want to touch them.

> How do you know if they can't be folded?

Because otherwise they would have been reduced in fold-const.cc.
The new conditions long with the rest in fold-const.cc require:

1. at least one of the operands to be constant
2. if the operands are equal, at least one of them must have been reduced to a constant.

These two means that (cond @0 (op @1 @2) (op @3 @2)) can't match if it's something
that fold-const.cc can handle.

> Since match.pd cannot handle arbitrary operations it
> isn't a good fit for match.pd patterns, instead this would be a forwprop
> pattern or, in case you want to catch GENERIC, a fold-const.cc one?

I'll try to move it to fold-const.cc then.

Tamar

> 
> Thanks and sorry for the late reply - hope Jeffs approval didn't make you
> apply it yet.
> 
> Richard.
> 
> > +
> >  #if GIMPLE
> >  (match (nop_atomic_bit_test_and_p @0 @1 @4)
> >   (bit_and (convert?@4 (ATOMIC_FETCH_OR_XOR_N @2 INTEGER_CST@0
> @3))
> > diff --git a/gcc/testsuite/gcc.target/aarch64/if-compare_3.c
> > b/gcc/testsuite/gcc.target/aarch64/if-compare_3.c
> > new file mode 100644
> > index
> >
> 0000000000000000000000000000000000000000..1d97da5c0d6454175881c2199
> 274
> > 71a567a6f0c7
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/aarch64/if-compare_3.c
> > @@ -0,0 +1,27 @@
> > +/* { dg-do compile } */
> > +/* { dg-additional-options "-O3 -std=c99" } */
> > +/* { dg-final { check-function-bodies "**" "" "" { target { le } } }
> > +} */
> > +
> > +/*
> > +**zoo:
> > +**	cmp	w0, w1
> > +**	csel	w0, w3, w2, le
> > +**	ret
> > +*/
> > +int zoo (int a, int b, int c, int d)
> > +{
> > +   return a > b ? c : d;
> > +}
> > +
> > +/*
> > +**zoo1:
> > +**	cmp	w0, w1
> > +**	csel	w0, w3, w2, le
> > +**	and	w0, w0, 1
> > +**	ret
> > +*/
> > +int zoo1 (int a, int b, int c, int d) {
> > +   return (a > b ? c : d) & 1;
> > +}
> > +
> > diff --git a/gcc/tree-cfg.cc b/gcc/tree-cfg.cc index
> >
> b19710392940cf469de52d006603ae1e3deb6b76..aaf1b29da5c598add25dad2c
> 38b8
> > 28eaa89c49ce 100644
> > --- a/gcc/tree-cfg.cc
> > +++ b/gcc/tree-cfg.cc
> > @@ -4244,7 +4244,9 @@ verify_gimple_assign_ternary (gassign *stmt)
> >        return true;
> >      }
> >
> > -  if (!is_gimple_val (rhs1)
> > +  /* In a COND_EXPR the rhs1 is the condition and thus isn't
> > +     a gimple value by definition.  */  if ((!is_gimple_val (rhs1) &&
> > + rhs_code != COND_EXPR)
> >        || !is_gimple_val (rhs2)
> >        || !is_gimple_val (rhs3))
> >      {
> >
> >
> >
> >
> >
> 
> --
> Richard Biener <rguenther@suse.de>
> SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461
> Nuernberg, Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald,
> Boudien Moerman; HRB 36809 (AG Nuernberg)