Fix PR 22018

[Diego Novillo]

On Wed, Jun 15, 2005 at 05:53:05PM +0200, Eric Botcazou wrote:

> Sorry for being picky
>
No problem.  Thanks for the feedback.  It's easy for me to miss
documenting all the silent assumptions the code makes.

> , but this is still not symmetric: "VAL1 + VAL2 grows if 
> VAL2 is >= 0".  Surely VAL1 and VAL2 can be swapped and the conclusion should 
> stay the same.  This works because VAL1 + VAL2 can overflow only if VAL1 and 
> VAL2 have the same sign, so either (sign1>0 && sign2>0) or (sign1<0 && 
> sign2<0) is true inside the TREE_OVERFLOW (res) block for PLUS_EXPR.  So you 
> are allowed to only test one of the 2 signs, but IMHO that should be 
> explained.
> 
> Same for MINUS: "VAL1 - VAL2 grows if VAL2 is < 0" is not antisymmetric.  VAL1 
> and VAL2 can be swapped and the conclusion must be swapped too.  Again this 
> works because VAL1 - VAL2 can overflow only if VAL1 and VAL2 doesn't have the 
> same sign, so either (sign1>0 && sign2<0) or (sign1<0 && sign2>0) is true 
> inside the TREE_OVERFLOW (res) block for MINUS_EXPR.  So again you are 
> allowed to only test one of the 2 signs.
> 
Sure.  Mind patching the comment with these added details?


Thanks.  Diego.