[llvm] r212638 - AArch64: Better codegen for storing to __fp16.
grosbach at apple.com
Thu Jul 10 12:41:50 PDT 2014
> On Jul 10, 2014, at 12:37 PM, Tim Northover <t.p.northover at gmail.com> wrote:
>> Absolutely. Greedily fusing double -> float -> __fp16 is correct with higher frequency than the old state of affairs, given the current front-end.
> I don't think that makes it the right thing to do though. We've
> replaced 1 bug with 2 that just happen to cancel out most of the time.
> Feel free to assign the bug to me. I actually quite like
> IEEE-conformance stuff. Gives the pedant in me free reign, which is
> always refreshing. (And it'll give me a chance to fix InstCombine,
> which gets it wrong now).
Hehe. OK. You’d already convinced me to go dig into clang and I was just getting started on adding new intrinsics and such to do the to/from double, but I’m always glad to be able to hand off this sort of thing. :)
Excellent point about InstCombine. I hadn’t though about looking there, too.
More information about the llvm-commits