[PATCH] Fix PR78305

[Marc Glisse]

On Wed, 16 Nov 2016, Richard Biener wrote:

> On Wed, 16 Nov 2016, Marc Glisse wrote:
>
>> On Wed, 16 Nov 2016, Richard Biener wrote:
>>
>>> I am testing the following to avoid undefined behavior when negating
>>> a multiplication (basically extending a previous fix to properly handle
>>> negative power of two).
>>>
>>> Bootstrap / regtest running on x86_64-unknown-linux-gnu.
>>>
>>> Richard.
>>>
>>> 2016-11-16  Richard Biener  <rguenther@suse.de>
>>>
>>> 	PR middle-end/78305
>>> 	* fold-const.c (negate_expr_p): Fix multiplication case.
>>>
>>> 	* gcc.dg/torture/pr78305.c: New testcase.
>>>
>>> Index: gcc/fold-const.c
>>> ===================================================================
>>> --- gcc/fold-const.c	(revision 242471)
>>> +++ gcc/fold-const.c	(working copy)
>>> @@ -450,13 +450,15 @@ negate_expr_p (tree t)
>>>       if (TYPE_UNSIGNED (type))
>>> 	break;
>>>       /* INT_MIN/n * n doesn't overflow while negating one operand it does
>>> -         if n is a power of two.  */
>>> +         if n is a power of (minus) two.  */
>>
>> if n is (minus) a power of two.
>> if n is a divisor of INT_MIN.
>
> n is a divisor of INT_MIN is correct.
>
>>>       if (INTEGRAL_TYPE_P (TREE_TYPE (t))
>>> 	  && ! TYPE_OVERFLOW_WRAPS (TREE_TYPE (t))
>>> 	  && ! ((TREE_CODE (TREE_OPERAND (t, 0)) == INTEGER_CST
>>> -		 && ! integer_pow2p (TREE_OPERAND (t, 0)))
>>> +		 && (wi::popcount (TREE_OPERAND (t, 0))
>>> +		     != 1 + wi::neg_p (TREE_OPERAND (t, 0), SIGNED)))
>>
>> Is that supposed to test for (possibly negated) powers of 2? I don't see it.
>> For -2, aka 0b11...110, popcount is 31 != 1 + 1.
>
> It's supposed to test for a power of two with the sign-bit ORed in.
> I believe those are which, when multiplied with some number can
> yield INT_MIN.  That is, we look for a test that ensures that there
> exists no number when multiplied with X yields INT_MIN.

The first sentence about ORing the sign bit sounds strange (except for a 
sign-magnitude representation). With 2's complement, INT_MIN is -2^31, the 
divisors are the 2^k and -(2^k). -2 * 2^30 yields INT_MIN, but your test 
misses -2 as a possible divisor. On the other hand, 0b100...001 (aka 
-INT_MAX) is not a divisor of INT_MIN but your test says the reverse.

-- 
Marc Glisse