library/portable-simd/crates/test_helpers/src/subnormals.rs

   1 pub trait FlushSubnormals: Sized {
   2     fn flush(self) -> Self {
   3         self
   4     }
   5 }
   6
   7 impl<T> FlushSubnormals for *const T {}
   8 impl<T> FlushSubnormals for *mut T {}
   9
  10 macro_rules! impl_float {
  11     { $($ty:ty),* } => {
  12         $(
  13         impl FlushSubnormals for $ty {
  14             fn flush(self) -> Self {
  15                 let is_f32 = core::mem::size_of::<Self>() == 4;
  16                 let ppc_flush = is_f32 && cfg!(all(
  17                     any(target_arch = "powerpc", all(target_arch = "powerpc64", target_endian = "big")),
  18                     target_feature = "altivec",
  19                     not(target_feature = "vsx"),
  20                 ));
  21                 let arm_flush = is_f32 && cfg!(all(target_arch = "arm", target_feature = "neon"));
  22                 let flush = ppc_flush || arm_flush;
  23                 if flush && self.is_subnormal() {
  24                     <$ty>::copysign(0., self)
  25                 } else {
  26                     self
  27                 }
  28             }
  29         }
  30         )*
  31     }
  32 }
  33
  34 macro_rules! impl_else {
  35     { $($ty:ty),* } => {
  36         $(
  37         impl FlushSubnormals for $ty {}
  38         )*
  39     }
  40 }
  41
  42 impl_float! { f32, f64 }
  43 impl_else! { i8, i16, i32, i64, isize, u8, u16, u32, u64, usize }
  44
  45 /// AltiVec should flush subnormal inputs to zero, but QEMU seems to only flush outputs.
  46 /// https://gitlab.com/qemu-project/qemu/-/issues/1779
  47 #[cfg(all(
  48     any(target_arch = "powerpc", target_arch = "powerpc64"),
  49     target_feature = "altivec"
  50 ))]
  51 fn in_buggy_qemu() -> bool {
  52     use std::sync::OnceLock;
  53     static BUGGY: OnceLock<bool> = OnceLock::new();
  54
  55     fn add(x: f32, y: f32) -> f32 {
  56         #[cfg(target_arch = "powerpc")]
  57         use core::arch::powerpc::*;
  58         #[cfg(target_arch = "powerpc64")]
  59         use core::arch::powerpc64::*;
  60
  61         let array: [f32; 4] =
  62             unsafe { core::mem::transmute(vec_add(vec_splats(x), vec_splats(y))) };
  63         array[0]
  64     }
  65
  66     *BUGGY.get_or_init(|| add(-1.0857398e-38, 0.).is_sign_negative())
  67 }
  68
  69 #[cfg(all(
  70     any(target_arch = "powerpc", target_arch = "powerpc64"),
  71     target_feature = "altivec"
  72 ))]
  73 pub fn flush_in<T: FlushSubnormals>(x: T) -> T {
  74     if in_buggy_qemu() {
  75         x
  76     } else {
  77         x.flush()
  78     }
  79 }
  80
  81 #[cfg(not(all(
  82     any(target_arch = "powerpc", target_arch = "powerpc64"),
  83     target_feature = "altivec"
  84 )))]
  85 pub fn flush_in<T: FlushSubnormals>(x: T) -> T {
  86     x.flush()
  87 }
  88
  89 pub fn flush<T: FlushSubnormals>(x: T) -> T {
  90     x.flush()
  91 }