r225244 - Sema: analyze I,J,K,M,N,O constraints

[Joerg Sonnenberger]

On Thu, Jan 15, 2015 at 11:31:19AM -0800, Chandler Carruth wrote:
> On Thu, Jan 15, 2015 at 11:06 AM, Joerg Sonnenberger <
> joerg at britannica.bec.de> wrote:
> 
> > On Thu, Jan 15, 2015 at 10:41:51AM -0800, Chandler Carruth wrote:
> > > On Thu, Jan 15, 2015 at 10:29 AM, Joerg Sonnenberger <
> > > joerg at britannica.bec.de> wrote:
> > >
> > > > See the inb example. All fixes are pessimations for the code. People
> > > > complained enough about the (in)ability of __builtin_constant_p to get
> > > > it delayed to the backend, this is IMO just the same.
> > > >
> > >
> > > I agree it is the same, however I have heard extremely little complaining
> > > about __builtin_constant_p. I've also not heard any complaints other than
> > > yours about this change -- most of the complaints I've heard have been
> > > about the terribly brittle code it uncovered and is now getting fixed.
> >
> > Given that I am often the first person to see breakage in random
> > software, I'm not too surprised.
> 
> 
> I'm not sure this is true. We had already hit and fixed the pixman issue
> locally when you reported it. However, it doesn't really matter who hits it
> first...

pixman is not the only place. tomcrypt also hit it in pkgsrc for AMD64,
can't tell about i386 or ARM for now.

> > So the real question would be -- why is
> > this not just a warning? I don't think anyone (including me) would
> > object to that, but it would make it possible to still use the feature
> > when necessary / useful.
> 
> 
> Because as a fundamental matter of design, Clang rejects inputs which it
> cannot validate will succeed in the code generator. This is not new, and an
> inline assembly constraint is not a compelling reason (to me) to undo this
> design principle.

Assembler validation in the frontend will always go only so far. The inb
example is *always* compilable, so rejecting it is certainly bogus.
While that is not true for all inline asm use, we should not artifically
constraint it either. I also don't believe that "never fail in codegen"
is a design principle of clang, otherwise intrinsincs on a lot of
platforms have been fundamentally flawed for a long time.

Joerg