Use [u]int64 -> FloatingPoint conversions even on 32b targets #70541
Conversation
…tforms. This means that we'll end up going int32 -> int64 -> float/double sometiems, but LLVM knows how to optimize away the intermediate conversion so we end up with just a normal 32b->float conversion as desired, and we get much, much better performance on oddball platforms like arm64_32.
|
@swift-ci test |
|
@swift-ci please benchmark |
|
@glessard the CI bencharks only test targets where there's no difference at all from this change, sadly. |
|
With integer conversions, there's no chance of a difference in rounding, correct? I remember that we had a bunch of platform-specific problems around NSNumber bridging when changing one of these initializers, but I'm pretty sure that was a float-to-float conversion. |
|
@swift-ci smoke test linux |
The only way we'd see a change here is if the generic conversion routine had a bug that we'd now avoid, which could be a hassle but also a change we'd want to make. We're reasonably confident that this isn't the case for standard library types, but the input space is too large to test exhaustively, so we can't be certain. |
…tforms. (swiftlang#70541) This means that we'll end up going int32 -> int64 -> float/double sometiems, but LLVM knows how to optimize away the intermediate conversion so we end up with just a normal 32b->float conversion as desired, and we get much, much better performance on oddball platforms like arm64_32.
…tforms. (swiftlang#70541) This means that we'll end up going int32 -> int64 -> float/double sometiems, but LLVM knows how to optimize away the intermediate conversion so we end up with just a normal 32b->float conversion as desired, and we get much, much better performance on oddball platforms like arm64_32.
This means that we'll end up going int32 -> int64 -> float/double sometimes, but LLVM knows how to optimize away the intermediate conversion so we end up with just a normal 32b->float conversion as desired, and we get much, much better performance on oddball platforms like arm64_32.