Re: [patch tree-optimization]: [2 of 3]: Boolify compares & more

[Kai Tietz <ktietz70@googlemail.com>]

2011/7/20 Richard Guenther <richard.guenther@gmail.com>:
> On Wed, Jul 20, 2011 at 3:31 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>> 2011/7/20 Richard Guenther <richard.guenther@gmail.com>:
>>> On Wed, Jul 20, 2011 at 3:05 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>>>> Hello,
>>>>
>>>> this is the revised version of the partial pre-approved patch for preserving
>>>> type-casts from/to boolean-types.  It fixes additionally the regression in
>>>> tree-ssa/builtin-expect-5.c testcase, which was caused by fold_builtin_expect.
>>>> Additionally there was a regression in gcc.dg/pr28685-1.c, which is fixed by
>>>> the change in tree-ssa-forwprop.c's function simplify_bitwise_binary.  This
>>>> is just temporary necessary.  As soon as we are boolifying comparisons in
>>>> gimplifier, the pattern-matching in tree-ssa-reassoc will match for 2
>>>> branched cases
>>>> again and we can remove the hunk from forward-propagation again.
>>>
>>> Hm, if we can't apply this pieces without regressions we shouldn't.  They
>>> can then wait for the boolification patch.
>>>
>>> Can you explain the fold_builtin_expect change?  I'm lost in the maze
>>> of inner/inner_arg0/arg0 renaming game.  It looks as if the patch only
>>> moves stuff - but that can't possibly be the case.  So, what's going on
>>> there?
>>
>> Well, the issue is here that fold_builtin_expect checks here for a
>> comparison.  If this comparison was created initially with a
>> boolean-type, the cast to 'long' will be in tree arg0 = (long)
>> CMP-with-boolean-type, as we are preserving here casts from
>> boolean-types (see the fold-const change). So we need to see through
>> this casts to match the compare and call cases. So I moved this "see
>> through" part before first pattern-match and introduced here a
>> helper-variable inner_arg0 to avoid double while-loop. The "inner"
>> variable might get invalid
>> ...
>>   if (COMPARISON_CLASS_P (inner)
>>       && TREE_CODE (TREE_OPERAND (inner, 1)) == INTEGER_CST)
>>     inner = TREE_OPERAND (inner, 0);
>> ...
>>
>> These are those "prefixed casts" you were asking in the other patch about.
>
> I see.  So, if the builtin.c parts bootstrap & test ok then they are ok
> to commit separately.
>
> Thanks,
> Richard.
>
>> Regards,
>> Kai
>>

Hello,

2011-07-20  Kai Tietz  <ktietz@redhat.com>

	* builtins.c (fold_builtin_expect): See through the cast
	from truthvalue_type_node to long.

Bootstrapped and regression tested for all languages (including Ada
and Obj-C++) on host x86_64-pc-linux-gnu.

Applied as pre-approved.

Regards,
Kai

Index: gcc-head/gcc/builtins.c
===================================================================
--- gcc-head.orig/gcc/builtins.c
+++ gcc-head/gcc/builtins.c
@@ -6264,13 +6264,22 @@ build_builtin_expect_predicate (location
 static tree
 fold_builtin_expect (location_t loc, tree arg0, tree arg1)
 {
-  tree inner, fndecl;
+  tree inner, fndecl, inner_arg0;
   enum tree_code code;

+  /* Distribute the expected value over short-circuiting operators.
+     See through the cast from truthvalue_type_node to long.  */
+  inner_arg0 = arg0;
+  while (TREE_CODE (inner_arg0) == NOP_EXPR
+	 && INTEGRAL_TYPE_P (TREE_TYPE (inner_arg0))
+	 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (inner_arg0, 0))))
+    inner_arg0 = TREE_OPERAND (inner_arg0, 0);
+
   /* If this is a builtin_expect within a builtin_expect keep the
      inner one.  See through a comparison against a constant.  It
      might have been added to create a thruthvalue.  */
-  inner = arg0;
+  inner = inner_arg0;
+
   if (COMPARISON_CLASS_P (inner)
       && TREE_CODE (TREE_OPERAND (inner, 1)) == INTEGER_CST)
     inner = TREE_OPERAND (inner, 0);
@@ -6281,14 +6290,7 @@ fold_builtin_expect (location_t loc, tre
       && DECL_FUNCTION_CODE (fndecl) == BUILT_IN_EXPECT)
     return arg0;

-  /* Distribute the expected value over short-circuiting operators.
-     See through the cast from truthvalue_type_node to long.  */
-  inner = arg0;
-  while (TREE_CODE (inner) == NOP_EXPR
-	 && INTEGRAL_TYPE_P (TREE_TYPE (inner))
-	 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (inner, 0))))
-    inner = TREE_OPERAND (inner, 0);
-
+  inner = inner_arg0;
   code = TREE_CODE (inner);
   if (code == TRUTH_ANDIF_EXPR || code == TRUTH_ORIF_EXPR)
     {
@@ -6303,13 +6305,13 @@ fold_builtin_expect (location_t loc, tre
     }

   /* If the argument isn't invariant then there's nothing else we can do.  */
-  if (!TREE_CONSTANT (arg0))
+  if (!TREE_CONSTANT (inner_arg0))
     return NULL_TREE;

   /* If we expect that a comparison against the argument will fold to
      a constant return the constant.  In practice, this means a true
      constant or the address of a non-weak symbol.  */
-  inner = arg0;
+  inner = inner_arg0;
   STRIP_NOPS (inner);
   if (TREE_CODE (inner) == ADDR_EXPR)
     {


2011-07-20  Kai Tietz  <ktietz@redhat.com>

	* fold-const.c (fold_unary_loc): Preserve
	non-boolean-typed casts.
	* tree-ssa.c (useless_type_conversion_p): Preserve incompatible
	casts from/to boolean,
	* builtins.c (fold_builtin_expect): See through the cast
	from truthvalue_type_node to long.
	* tree-ssa-forwprop.c (simplify_bitwise_binary): Add
	simplification for CMP bitwise-binary CMP.