Re: Invalid compilation without -fno-strict-aliasing

[torvalds@transmeta.com (Linus Torvalds)]

In article <20030226172213.O3910@devserv.devel.redhat.com>,
Jakub Jelinek  <jakub@redhat.com> wrote:
>
>To fix that, __constant_memcpy would have to access the data through
>union,

Which is impossible, since memcpy _fundamentally_ cannot know what the
different types are..

> or you could as well forget about __constant_memcpy and use
>__builtin_memcpy where gcc will take care about the constant copying.

Which is impossible because (a) historically __builtin_memcpy does a bad
job and (b) it doesn't solve the generic case anyway, ie for other
non-memcpy things.

The fact is, for type-based alias analysis gcc needs a way to tell it
"this can alias", which it doesn't have.  Unions are _not_ useful,
_regardless_ of what silly language lawyers say, since they are not a
generic method.  Unions only work for trivial and largely uninteresting
cases, and it doesn't _matter_ what C99 says about the issue, since that
nasty thing called "real life" interferes.

Until we get some non-union way to say "this can alias", that
-fno-strict-alias has to stay because gcc is too broken to allow us
doing interesting stuff in-line without it. 

My personal opinion is (and was several years ago when this started
coming up) that a cast (any cast) should do it. But I don't are _what_
it is, as long as it is syntactically sane and isn't limited to special
cases like unions.

		Linus