Update to LLVM head and mark various ptrs as nonnull. #14306
Conversation
|
These also need to be marked in trans/callee.rs for indirect calls. It seems marking the out pointer as |
|
Does this help with #11751 at all? (I'd guess not, since the pointers are all nested there.) |
|
I think it might help if LLVM learned to leveraged the |
|
(actually it's not going to help either way) |
|
I do think that there are many more venues of optimization possible here, but this should hopefully be a clear win in at least some situations. For example, this doesn't look like this handles structs which contain non-null pointers. Things like trait objects, tuples, slices, etc, will all benefit from that, but just attaching to a parameter may not be sufficient. This is merged upstream, however, and this definitely isn't detrimental, so let's push forward and see how far we can get with these optimizations! |
|
cc #9546 |
| ty::ty_rptr(_, mt) if mt.mutbl == ast::MutMutable => { | ||
| attrs.push((idx, lib::llvm::NoAliasAttribute as u64)); | ||
| attrs.push((idx, lib::llvm::NonNullAttribute as u64)); | ||
| } |
There was a problem hiding this comment.
This case doesn't get nocapture, but sometimes it may be able to?
|
Nice cleanup! |
|
r=me with a squash of the last three commits (they all seem very closely related) |
…uments and returns.
We can now mark arguments and returns as `nonnull` in LLVM. It's still somewhat limited by the fact that LLVM loses this information after inlining but it might help in certain cases.
fix block with no termination in or patterns fix rust-lang#14298
We can now mark arguments and returns as
nonnullin LLVM. It's still somewhat limited by the fact that LLVM loses this information after inlining but it might help in certain cases.