[PATCH] reassoc: Fix up (ab) handling in eliminate_redundant_comparison [PR108783]

[PATCH] reassoc: Fix up (ab) handling in eliminate_redundant_comparison [PR108783]

[Jakub Jelinek]

Hi!

The following testcase ICEs because eliminate_redundant_comparison sees
redundant comparisons in &&/|| where the comparison has (ab) SSA_NAME,
maybe_fold_{and,or}_comparisons optimizes them into a single comparison
and build_and_add_sum emits a new comparison close to the definition
operands, which in this case is before a returns_twice call (which is
invalid).  Generally reassoc just punts on (ab) SSA_NAMEs, declares them
non-reassociable etc., so the second half of this patch does that.

Though we can do better in this case; the function has special code
when maybe_fold_{and,or}_comparisons returns INTEGER_CST (false/true)
or when what it returns is the same as curr->op (the first of the
comparisons we are considering) - in that case we just remove the
second one and keep the first one.  The reason it doesn't match is that
curr->op is a SSA_NAME whose SSA_NAME_DEF_STMT is checked to be a
comparison, in this case _42 = a_1(ab) != 0 and the other comparison
is also like that.  maybe_fold_{and,or}_comparisons looks through the
definitions though and so returns a_1(ab) != 0 as tree.
So the first part of the patch checks whether that returned comparison
isn't the same as the curr->op comparison and if yes, it just overrides
t back to curr->op so that its SSA_NAME is reused.  In that case we can
handle even (ab) in {,new}op{1,2} because we don't create a new comparison
of that, just keep using the existing one.  And t can't be (ab) because
otherwise it wouldn't be considered a reassociable operand.

The (ab) checks are needed say when we have a_1(ab) == 42 || a_1(ab) > 42
kind of comparisons where maybe_fold_{and,or}_comparisons returns a new
comparison not existing in the IL yet.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2023-02-15  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/108783
	* tree-ssa-reassoc.cc (eliminate_redundant_comparison): If lcode
	is equal to TREE_CODE (t), op1 to newop1 and op2 to newop2, set
	t to curr->op.  Otherwise, punt if either newop1 or newop2 are
	SSA_NAME_OCCURS_IN_ABNORMAL_PHI SSA_NAMEs.

	* gcc.c-torture/compile/pr108783.c: New test.

--- gcc/tree-ssa-reassoc.cc.jj	2023-01-12 21:04:08.726238049 +0100
+++ gcc/tree-ssa-reassoc.cc	2023-02-15 13:28:04.987278895 +0100
@@ -2272,6 +2272,15 @@ eliminate_redundant_comparison (enum tre
 	  STRIP_USELESS_TYPE_CONVERSION (newop2);
 	  if (!is_gimple_val (newop1) || !is_gimple_val (newop2))
 	    continue;
+	  if (lcode == TREE_CODE (t)
+	      && operand_equal_p (op1, newop1, 0)
+	      && operand_equal_p (op2, newop2, 0))
+	    t = curr->op;
+	  else if ((TREE_CODE (newop1) == SSA_NAME
+		    && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (newop1))
+		   || (TREE_CODE (newop2) == SSA_NAME
+		       && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (newop2)))
+	    continue;
 	}
 
       if (dump_file && (dump_flags & TDF_DETAILS))
--- gcc/testsuite/gcc.c-torture/compile/pr108783.c.jj	2023-02-15 12:42:46.244340524 +0100
+++ gcc/testsuite/gcc.c-torture/compile/pr108783.c	2023-02-15 13:24:47.515187118 +0100
@@ -0,0 +1,42 @@
+/* PR tree-optimization/108783 */
+
+__attribute__((returns_twice)) int baz (int, int);
+
+int
+bar (int x)
+{
+  return x;
+}
+
+int
+foo (int x, int y)
+{
+  int a;
+
+  a = bar (x);
+  baz (x, y);
+
+  return y && a && a;
+}
+
+int
+qux (int x, int y)
+{
+  int a;
+
+  a = bar (x);
+  baz (x, y);
+
+  return y && a != 42 && a >= 42;
+}
+
+int
+corge (int x, int y)
+{
+  int a;
+
+  a = bar (x);
+  baz (x, y);
+
+  return y || a == 42 || a > 42;
+}

	Jakub

Re: [PATCH] reassoc: Fix up (ab) handling in eliminate_redundant_comparison [PR108783]

[Richard Biener]

On Thu, 16 Feb 2023, Jakub Jelinek wrote:

> Hi!
> 
> The following testcase ICEs because eliminate_redundant_comparison sees
> redundant comparisons in &&/|| where the comparison has (ab) SSA_NAME,
> maybe_fold_{and,or}_comparisons optimizes them into a single comparison
> and build_and_add_sum emits a new comparison close to the definition
> operands, which in this case is before a returns_twice call (which is
> invalid).  Generally reassoc just punts on (ab) SSA_NAMEs, declares them
> non-reassociable etc., so the second half of this patch does that.
> 
> Though we can do better in this case; the function has special code
> when maybe_fold_{and,or}_comparisons returns INTEGER_CST (false/true)
> or when what it returns is the same as curr->op (the first of the
> comparisons we are considering) - in that case we just remove the
> second one and keep the first one.  The reason it doesn't match is that
> curr->op is a SSA_NAME whose SSA_NAME_DEF_STMT is checked to be a
> comparison, in this case _42 = a_1(ab) != 0 and the other comparison
> is also like that.  maybe_fold_{and,or}_comparisons looks through the
> definitions though and so returns a_1(ab) != 0 as tree.
> So the first part of the patch checks whether that returned comparison
> isn't the same as the curr->op comparison and if yes, it just overrides
> t back to curr->op so that its SSA_NAME is reused.  In that case we can
> handle even (ab) in {,new}op{1,2} because we don't create a new comparison
> of that, just keep using the existing one.  And t can't be (ab) because
> otherwise it wouldn't be considered a reassociable operand.
> 
> The (ab) checks are needed say when we have a_1(ab) == 42 || a_1(ab) > 42
> kind of comparisons where maybe_fold_{and,or}_comparisons returns a new
> comparison not existing in the IL yet.
> 
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

OK.

Thanks,
Richard.

> 2023-02-15  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR tree-optimization/108783
> 	* tree-ssa-reassoc.cc (eliminate_redundant_comparison): If lcode
> 	is equal to TREE_CODE (t), op1 to newop1 and op2 to newop2, set
> 	t to curr->op.  Otherwise, punt if either newop1 or newop2 are
> 	SSA_NAME_OCCURS_IN_ABNORMAL_PHI SSA_NAMEs.
> 
> 	* gcc.c-torture/compile/pr108783.c: New test.
> 
> --- gcc/tree-ssa-reassoc.cc.jj	2023-01-12 21:04:08.726238049 +0100
> +++ gcc/tree-ssa-reassoc.cc	2023-02-15 13:28:04.987278895 +0100
> @@ -2272,6 +2272,15 @@ eliminate_redundant_comparison (enum tre
>  	  STRIP_USELESS_TYPE_CONVERSION (newop2);
>  	  if (!is_gimple_val (newop1) || !is_gimple_val (newop2))
>  	    continue;
> +	  if (lcode == TREE_CODE (t)
> +	      && operand_equal_p (op1, newop1, 0)
> +	      && operand_equal_p (op2, newop2, 0))
> +	    t = curr->op;
> +	  else if ((TREE_CODE (newop1) == SSA_NAME
> +		    && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (newop1))
> +		   || (TREE_CODE (newop2) == SSA_NAME
> +		       && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (newop2)))
> +	    continue;
>  	}
>  
>        if (dump_file && (dump_flags & TDF_DETAILS))
> --- gcc/testsuite/gcc.c-torture/compile/pr108783.c.jj	2023-02-15 12:42:46.244340524 +0100
> +++ gcc/testsuite/gcc.c-torture/compile/pr108783.c	2023-02-15 13:24:47.515187118 +0100
> @@ -0,0 +1,42 @@
> +/* PR tree-optimization/108783 */
> +
> +__attribute__((returns_twice)) int baz (int, int);
> +
> +int
> +bar (int x)
> +{
> +  return x;
> +}
> +
> +int
> +foo (int x, int y)
> +{
> +  int a;
> +
> +  a = bar (x);
> +  baz (x, y);
> +
> +  return y && a && a;
> +}
> +
> +int
> +qux (int x, int y)
> +{
> +  int a;
> +
> +  a = bar (x);
> +  baz (x, y);
> +
> +  return y && a != 42 && a >= 42;
> +}
> +
> +int
> +corge (int x, int y)
> +{
> +  int a;
> +
> +  a = bar (x);
> +  baz (x, y);
> +
> +  return y || a == 42 || a > 42;
> +}
> 
> 	Jakub
> 
> 

-- 
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)