[Bug c/17913] [4.0 Regression] ICE jumping into statement expression

[jsm at polyomino dot org dot uk]

------- Additional Comments From jsm at polyomino dot org dot uk  2004-10-27 12:55 -------
Subject: Re:  [4.0 Regression] ICE jumping into statement
 expression

On Wed, 27 Oct 2004, steven at gcc dot gnu dot org wrote:

> Surely this is not valid?

The validity is the subject of bug 772 and the long thread linked from 
there.  This bug is for a particular ICE which is a regression; whether 
the code is valid or invalid depends on whether any resolution suggested 
in bug 772 or elsewhere is adopted.

I haven't seen a use for jumps into statement expressions and think we 
should disallow them.  This isn't as simple as making the labels local to 
the statement expression.  First, you need to check for switch statements 
jumping into the statement expression.  Second, if inside a statement 
expression you encounter a reference to an undeclared label, you don't 
know whether it refers to one later within that statement expression or 
one outside in the function (and in an address-of-label context, both 
could reasonably be valid) so don't know in what scope to declare the 
label (or, if there's one outside already seen with that name, whether the 
new one is a reference to that label or to a new local one).  For that 
reason I think it's best to keep the current scoping rules and have a 
separate check for any form of jumps into statement expressions, with 
jumps by computed goto explicitly runtime undefined.  (While jumps in from 
longjmp might be OK.)

Jumps out of statement expressions are or were used in real code 
<http://gcc.gnu.org/ml/gcc-patches/2003-05/msg00770.html>.  I don't think 
they are well-defined (except for jumps using longjmp), but I believe we 
need a deprecation period before removing them (e.g. deprecate in 4.0, 
remove in 4.1).



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17913