Replying in both places because this mail is a pretty good summary of
issues and so it's worth replying to... my apologies for continuing in
Responding to LLVMDev - I for one had missed this part of the thread.
The only language aspect under discussion is integer division. Every
standardized language I'm aware of requires some form of check on integer
Either the language is C-based, in which case we need to provide compiler
options to achieve sanity:
I'm going to argue this one, this isn't part of language conformance,
unless you're optimizing your undefined behavior. I also haven't
seen patches to have clang emit this intrinsic anywhere.
Or the language requires a check to determine whether to throw an exception.
Like Java? Something else?
Java for one.
Or the language requires a check to determine whether the number needs to be
promoted to a different representation.
Guessing JS here then? You could just actually state the kind of
behavior you're trying to work with rather than handwaving it. How
about show me some IR and how we'd like this to work? Michael started
that, but I haven't seen something with a front end language and how
we'd like to map such a thing onto IR.
Most any language which uses full integer types rather than machine integers. (i.e. uses machine integers as an optimization) I don't have an example off the top of my head.
You have a good point about spelling out desired semantics. My particular semantics are (Java):
- INT_MIN/-1 = INT_MIN
- X/0 -> exception
- x/y = x/y (rounded towards zero as an integer)
Currently, we are using unoptimized checks for each of the corner cases emitted by our frontend. We haven't even started to optimize this yet. Which is why I'm interested when others seem to be pursuing the same goal.
Of course, these can all be implemented with explicit checks in the IR. And
at IR level, this is exactly the same situation as checking overflow on
add/sub/mul. I have argued in the past just as you are now that there is
already another way to represent the check. However, LLVM developers are now
very heavily using the llvm.arith.with.overflow intrinsics and making the
optimizer aware of them, I think we should be consistent. Anyway,
implementing explicit checks in the IR (before CodeGenPrepare) tends to make
the target independent IR heavily biased toward a particular architecture.
This is true, and we emit the IR with that instruction in the front
end so that we can do that reasonably effectively. I don't necessarily
know that the semantics of the divide instruction are as clear cut as
the overflow on add. Why is this the correct set of things that we
should be checking?
Because these are the corner cases which matter in practice. Or to say it differently, no one has spoken up to request others.
INT_MIN / -1 (or %) is definitely undefined
behavior in C11/C++11 and could trap on various processors, but is
apparently defined to ... something on ARM64?
You keep mentioning undefined behaviour in C++. I'm not compiling C++. Why is this relevant?