Re: useless cast blocking some optimization in gcc 4.7.3

[Marc Glisse <marc dot glisse at inria dot fr>]

Hello,

On Mon, 8 Apr 2013, Laurent Alfonsi wrote:

> I have identified a big performance regression between 4.6 and 4.7. (I have 
> enclosed a pathological test).
>
> After investigation, it is because of the += statement applied on 2 signed 
> chars.
>  - It is now type-promoted to "int" when it is written "result += foo()". 
> (since 4.7)
>  - it is type promoted to "unsigned char" when it is written "result = 
> result + foo()".
>
> The "char->int->char" cast is blocking some optimizations in later phases.

Which ones?

> Anyway, this doesn't look wrong, so I extended fold optimization in order to 
> catch this case. (patch enclosed)
> The patch basically transforms :
>    (TypeA)  ( (TypeB) a1 + (TypeB) a2 )                /* with a1 and a2 of 
> the signed type TypeA */
> into :
>    a1 + a2
>
> I believe this is legal for any licit a1/a2 input values (no overflow on 
> signed char).

I don't think this is ok, please refer to the discussion around the 
PR and patch that added this conversion, it was done on purpose.

According to this (4th item)
http://gcc.gnu.org/onlinedocs/gcc/Integers-implementation.html

char a=100;
a+=a;

is perfectly defined and a is -56 (assuming a signed 8 bit char and a 
strictly larger int). However, your transformation turns it into undefined 
behavior: an addition that overflows in a type with 
TYPE_OVERFLOW_UNDEFINED.

--
Marc Glisse