[PATCH] D23064: AMDGPU: Custom lower FP_TO_FP16 for f64

[Jan Vesely via llvm-commits]

jvesely added a comment.

In https://reviews.llvm.org/D23064#504311, @arsenm wrote:

> This isn't a correct lowering, the correctly rounded conversion is complex and needs to be implemented


feel free to click 'reject', although a more detailed answer would be appreciated.
The OpenCL specs say that mantissa is correctly rounded using default rounding mode [rtne]. my understanding is that using trunc should be enough to not destroy the rounding information.


Repository:
  rL LLVM

https://reviews.llvm.org/D23064