Re: [PATCH] Check the STRING_CSTs in varasm.c

[Richard Biener <rguenther@suse.de>]

On Fri, 17 Aug 2018, Bernd Edlinger wrote:

> Richard Biener wrote:
> > +embedded @code{NUL} characters.  However, the
> > +@code{TREE_STRING_LENGTH} always includes a trailing @code{NUL} that
> > +is not part of the language string literal but appended by the front end.
> > +If the string shall not be @code{NUL}-terminated the @code{TREE_TYPE}
> > +is one character shorter than @code{TREE_STRING_LENGTH}.
> > +Excess caracters other than one trailing @code{NUL} character are not

characters btw.

I read the above that the string literal for

char x[2] = "1";

is actually "1\0\0" - there's one NUL that is not part of the language
string literal.  The second sentence then suggests that both \0
are removed because 2 is less than 3?

As said, having this extra semantics of a STRING_CST tied to
another tree node (its TREE_TYPE) looks ugly.

> > +permitted.
> > 
> > I find this very confusing and oppose to that change.  Can we get
> > back to the drawing board please?  If we want an easy way to
> > see whether a string is "properly" terminated then maybe we can
> > simply use a flag that gets set by build_string?
> >
> 
> What I mean with that is the case like
> char x[2] = "123456";
> 
> which is build_string(7, "123456"), but with a type char[2],
> so varasm throws away "3456\0".

I think varasm throws away chars not because of the type of
the STRING_CST but because of the available storage in x.

> I want to say that this is not okay, the excess precision
> should only be used to strip the nul termination, in cases
> where it is intended to be a assembled as a not zero terminated
> string.  But maybe the wording could be improved?

ISTR we always assemble a NUL in .strings to get string merging
working.

Richard.