[llvm-commits] CVS: llvm/test/Feature/varargs_new.ll

Chris Lattner sabre at nondot.org
Sun Jun 19 06:36:11 PDT 2005


On Sun, 19 Jun 2005, Andrew Lenharth wrote:

> On Sat, 2005-06-18 at 22:17 -0500, Chris Lattner wrote:
>> Reid: I think this patch should be reverted.  This regression test failing
>> indicates a bug in the auto-upgrader for .ll files.  Imagine if an old
>> LLVM front-end generated this file: now llvm-as would crash on it.
>
> No front end would generate this file.  This is not a bug in the upgrade
> code but a matter of what to do on illegal/invalid .ll files.  The
> problem here is using oldstyle vaarg with new style va_start (new style
> is va_arg, and must be to disambiguate certain legal constructs).

Ok, makes sense.

>> Not only is this not acceptable from a backwards compatibility
>> perspective, llvm-as should not crash on *any* input.
>
> It already does for similar problems.  try defining llvm.va_start as
> taking 5 args.  Verify will hit an assertion.  This is no different than
> hitting an arguement number check assertion earlier.

That's a bug and should be fixed then :).  It should use ThrowException to 
report the error, which will tack on a line number to the error report.

Sorry for the misunderstanding!

-Chris

-- 
http://nondot.org/sabre/
http://llvm.cs.uiuc.edu/




More information about the llvm-commits mailing list