enh-google wrote:

> > > Typically we don't bother to implement these unless libc starts using 
> > > these, or needs them. Making builtins for these functions without the 
> > > library needs them is kinda silly.
> > > DO we have a request for these coming from the libc maintainers?
> > 
> > 
> > I think they're important for two reasons: 1) we're going to want constexpr 
> > support for these for the same reason we want constexpr support for 
> > `strlen` in C so defining them as recognized library builtins is the way we 
> > do that, and 2) I think libc is going to want to have full support for 
> > bit-precise integer types and that's easier to support from a builtin 
> > currently. That said, CC @michaelrj-google for additional opinions
> 
> From the libc side having these builtins would be handy for both of the 
> reasons Aaron mentioned. For optimization it would be helpful if the compiler 
> could replace the libcall with a builtin, since things like `leading_zeros` 
> can sometimes be reduced to a single instruction. Actually calling these bit 
> functions as functions is unlikely to be optimal. CC: @enh-google

as i said in android's <stdbit.h>:
```
 * Ideally we'd have clang builtins for all of these, like gcc has.
 * If/when clang catches up, we can replace all of this with direct calls to
 * the relevant builtins.
```
so, yeah, "please subscribe me to your newsletter" --- i'll switch android's 
libc over to these as soon as they're available in the platform's clang :-)

see 
https://cs.android.com/android/platform/superproject/+/android-latest-release:bionic/libc/include/stdbit.h;l=2?q=file:stdbit.h
 for our current implementation, rewriting these in terms of existing builtins, 
which is mostly (but not always) fairly straightforward.

https://github.com/llvm/llvm-project/pull/185978
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to