Re: [PATCH, C++] Make Canonical ICE instead of just warn when mismatching

[Brooks Moses <brooks.moses@codesourcery.com>]

Mike Stump wrote:
> On Jan 4, 2007, at 3:42 PM, Andrew_Pinski@PlayStation.Sony.Com wrote:
>> This is just like any other internal check inside GCC we have.
> 
> It could be.  It need not be.  In our design for this one, we choose  
> to use software fault tolerance.  99.9% of the other checks, don't.   
> If you want to argue that software is no place to practice fault  
> tolerance, fine, do that.  Feel free to cite references if you want.   
> We don't have to, but, for the transition hump, I do think it is  
> better to offer fault tolerance.

There's one thing that I'm not sure I'm understanding here, in trying to 
follow this discussion.

Normally, on a random compiler ICE, the presence of the condition that 
caused the ICE is an indication that something has happened that does 
not fit our understanding of how the code works, and thus we are in a 
situation where we have no reason to expect that the generated results 
will be correct -- and, in fact, chances are pretty good that it will be 
incorrect and buggy.

Thus, in all of those cases, there _is_ no fault tolerance, regardless 
of whether the ICE is an error or a warning -- in either case, usable 
results are not produced.  The distinction is in how we tell the user of 
this failure, and how clear we make it that the results are unusable. 
An ICE and no output makes this blantantly clear; a warning and 
probably-buggy output (which masquerades as a valid executable) makes it 
far less clear, and it's possible that some users will overlook the 
failure entirely.

In that case, an ICE and error-out is almost certainly the correct answer.

My impression from your argument is that this case is not like that; 
that here we know that a failure to pass this particular internal check 
will never result in invalid code.  And thus a warning is appropriate 
here, because the user will still get valid usable code if they ignore 
the warning.

Is this a correct impression?  I haven't seen any mention of this 
distinction, but it seems to me that it's the only way that the argument 
for a warning instead of a hard error would make sense.

- Brooks