Merging upstream version 1.76.0+dfsg1.

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

4 files changed, 250 insertions, 7 deletions

diff --git a/library/portable-simd/crates/test_helpers/Cargo.toml b/library/portable-simd/crates/test_helpers/Cargo.toml
index 1d2bc8b51..23dae7c93 100644
--- a/library/portable-simd/crates/test_helpers/Cargo.toml
+++ b/library/portable-simd/crates/test_helpers/Cargo.toml
@@ -4,10 +4,8 @@ version = "0.1.0"
 edition = "2021"
 publish = false
 
-[dependencies.proptest]
-version = "0.10"
-default-features = false
-features = ["alloc"]
+[dependencies]
+proptest = { version = "0.10", default-features = false, features = ["alloc"] }
 
 [features]
 all_lane_counts = []
diff --git a/library/portable-simd/crates/test_helpers/src/biteq.rs b/library/portable-simd/crates/test_helpers/src/biteq.rs
index 7d91260d8..cbc20cda0 100644
--- a/library/portable-simd/crates/test_helpers/src/biteq.rs
+++ b/library/portable-simd/crates/test_helpers/src/biteq.rs
@@ -113,6 +113,27 @@ impl<T: BitEq> core::fmt::Debug for BitEqWrapper<'_, T> {
     }
 }
 
+#[doc(hidden)]
+pub struct BitEqEitherWrapper<'a, T>(pub &'a T, pub &'a T);
+
+impl<T: BitEq> PartialEq<BitEqEitherWrapper<'_, T>> for BitEqWrapper<'_, T> {
+    fn eq(&self, other: &BitEqEitherWrapper<'_, T>) -> bool {
+        self.0.biteq(other.0) || self.0.biteq(other.1)
+    }
+}
+
+impl<T: BitEq> core::fmt::Debug for BitEqEitherWrapper<'_, T> {
+    fn fmt(&self, f: &mut core::fmt::Formatter) -> core::fmt::Result {
+        if self.0.biteq(self.1) {
+            self.0.fmt(f)
+        } else {
+            self.0.fmt(f)?;
+            write!(f, " or ")?;
+            self.1.fmt(f)
+        }
+    }
+}
+
 #[macro_export]
 macro_rules! prop_assert_biteq {
     { $a:expr, $b:expr $(,)? } => {
@@ -122,5 +143,14 @@ macro_rules! prop_assert_biteq {
             let b = $b;
             proptest::prop_assert_eq!(BitEqWrapper(&a), BitEqWrapper(&b));
         }
-    }
+    };
+    { $a:expr, $b:expr, $c:expr $(,)? } => {
+        {
+            use $crate::biteq::{BitEqWrapper, BitEqEitherWrapper};
+            let a = $a;
+            let b = $b;
+            let c = $c;
+            proptest::prop_assert_eq!(BitEqWrapper(&a), BitEqEitherWrapper(&b, &c));
+        }
+    };
 }
diff --git a/library/portable-simd/crates/test_helpers/src/lib.rs b/library/portable-simd/crates/test_helpers/src/lib.rs
index b26cdc311..b80c745aa 100644
--- a/library/portable-simd/crates/test_helpers/src/lib.rs
+++ b/library/portable-simd/crates/test_helpers/src/lib.rs
@@ -1,3 +1,5 @@
+#![feature(stdsimd, powerpc_target_feature)]
+
 pub mod array;
 
 #[cfg(target_arch = "wasm32")]
@@ -6,6 +8,9 @@ pub mod wasm;
 #[macro_use]
 pub mod biteq;
 
+pub mod subnormals;
+use subnormals::FlushSubnormals;
+
 /// Specifies the default strategy for testing a type.
 ///
 /// This strategy should be what "makes sense" to test.
@@ -151,7 +156,6 @@ pub fn test_3<
 }
 
 /// Test a unary vector function against a unary scalar function, applied elementwise.
-#[inline(never)]
 pub fn test_unary_elementwise<Scalar, ScalarResult, Vector, VectorResult, const LANES: usize>(
     fv: &dyn Fn(Vector) -> VectorResult,
     fs: &dyn Fn(Scalar) -> ScalarResult,
@@ -178,6 +182,48 @@ pub fn test_unary_elementwise<Scalar, ScalarResult, Vector, VectorResult, const
 }
 
 /// Test a unary vector function against a unary scalar function, applied elementwise.
+///
+/// Where subnormals are flushed, use approximate equality.
+pub fn test_unary_elementwise_flush_subnormals<
+    Scalar,
+    ScalarResult,
+    Vector,
+    VectorResult,
+    const LANES: usize,
+>(
+    fv: &dyn Fn(Vector) -> VectorResult,
+    fs: &dyn Fn(Scalar) -> ScalarResult,
+    check: &dyn Fn([Scalar; LANES]) -> bool,
+) where
+    Scalar: Copy + core::fmt::Debug + DefaultStrategy + FlushSubnormals,
+    ScalarResult: Copy + biteq::BitEq + core::fmt::Debug + DefaultStrategy + FlushSubnormals,
+    Vector: Into<[Scalar; LANES]> + From<[Scalar; LANES]> + Copy,
+    VectorResult: Into<[ScalarResult; LANES]> + From<[ScalarResult; LANES]> + Copy,
+{
+    let flush = |x: Scalar| subnormals::flush(fs(subnormals::flush_in(x)));
+    test_1(&|x: [Scalar; LANES]| {
+        proptest::prop_assume!(check(x));
+        let result_v: [ScalarResult; LANES] = fv(x.into()).into();
+        let result_s: [ScalarResult; LANES] = x
+            .iter()
+            .copied()
+            .map(fs)
+            .collect::<Vec<_>>()
+            .try_into()
+            .unwrap();
+        let result_sf: [ScalarResult; LANES] = x
+            .iter()
+            .copied()
+            .map(flush)
+            .collect::<Vec<_>>()
+            .try_into()
+            .unwrap();
+        crate::prop_assert_biteq!(result_v, result_s, result_sf);
+        Ok(())
+    });
+}
+
+/// Test a unary vector function against a unary scalar function, applied elementwise.
 #[inline(never)]
 pub fn test_unary_mask_elementwise<Scalar, Vector, Mask, const LANES: usize>(
     fv: &dyn Fn(Vector) -> Mask,
@@ -204,7 +250,6 @@ pub fn test_unary_mask_elementwise<Scalar, Vector, Mask, const LANES: usize>(
 }
 
 /// Test a binary vector function against a binary scalar function, applied elementwise.
-#[inline(never)]
 pub fn test_binary_elementwise<
     Scalar1,
     Scalar2,
@@ -241,6 +286,85 @@ pub fn test_binary_elementwise<
     });
 }
 
+/// Test a binary vector function against a binary scalar function, applied elementwise.
+///
+/// Where subnormals are flushed, use approximate equality.
+pub fn test_binary_elementwise_flush_subnormals<
+    Scalar1,
+    Scalar2,
+    ScalarResult,
+    Vector1,
+    Vector2,
+    VectorResult,
+    const LANES: usize,
+>(
+    fv: &dyn Fn(Vector1, Vector2) -> VectorResult,
+    fs: &dyn Fn(Scalar1, Scalar2) -> ScalarResult,
+    check: &dyn Fn([Scalar1; LANES], [Scalar2; LANES]) -> bool,
+) where
+    Scalar1: Copy + core::fmt::Debug + DefaultStrategy + FlushSubnormals,
+    Scalar2: Copy + core::fmt::Debug + DefaultStrategy + FlushSubnormals,
+    ScalarResult: Copy + biteq::BitEq + core::fmt::Debug + DefaultStrategy + FlushSubnormals,
+    Vector1: Into<[Scalar1; LANES]> + From<[Scalar1; LANES]> + Copy,
+    Vector2: Into<[Scalar2; LANES]> + From<[Scalar2; LANES]> + Copy,
+    VectorResult: Into<[ScalarResult; LANES]> + From<[ScalarResult; LANES]> + Copy,
+{
+    let flush = |x: Scalar1, y: Scalar2| {
+        subnormals::flush(fs(subnormals::flush_in(x), subnormals::flush_in(y)))
+    };
+    test_2(&|x: [Scalar1; LANES], y: [Scalar2; LANES]| {
+        proptest::prop_assume!(check(x, y));
+        let result_v: [ScalarResult; LANES] = fv(x.into(), y.into()).into();
+        let result_s: [ScalarResult; LANES] = x
+            .iter()
+            .copied()
+            .zip(y.iter().copied())
+            .map(|(x, y)| fs(x, y))
+            .collect::<Vec<_>>()
+            .try_into()
+            .unwrap();
+        let result_sf: [ScalarResult; LANES] = x
+            .iter()
+            .copied()
+            .zip(y.iter().copied())
+            .map(|(x, y)| flush(x, y))
+            .collect::<Vec<_>>()
+            .try_into()
+            .unwrap();
+        crate::prop_assert_biteq!(result_v, result_s, result_sf);
+        Ok(())
+    });
+}
+
+/// Test a unary vector function against a unary scalar function, applied elementwise.
+#[inline(never)]
+pub fn test_binary_mask_elementwise<Scalar1, Scalar2, Vector1, Vector2, Mask, const LANES: usize>(
+    fv: &dyn Fn(Vector1, Vector2) -> Mask,
+    fs: &dyn Fn(Scalar1, Scalar2) -> bool,
+    check: &dyn Fn([Scalar1; LANES], [Scalar2; LANES]) -> bool,
+) where
+    Scalar1: Copy + core::fmt::Debug + DefaultStrategy,
+    Scalar2: Copy + core::fmt::Debug + DefaultStrategy,
+    Vector1: Into<[Scalar1; LANES]> + From<[Scalar1; LANES]> + Copy,
+    Vector2: Into<[Scalar2; LANES]> + From<[Scalar2; LANES]> + Copy,
+    Mask: Into<[bool; LANES]> + From<[bool; LANES]> + Copy,
+{
+    test_2(&|x: [Scalar1; LANES], y: [Scalar2; LANES]| {
+        proptest::prop_assume!(check(x, y));
+        let result_v: [bool; LANES] = fv(x.into(), y.into()).into();
+        let result_s: [bool; LANES] = x
+            .iter()
+            .copied()
+            .zip(y.iter().copied())
+            .map(|(x, y)| fs(x, y))
+            .collect::<Vec<_>>()
+            .try_into()
+            .unwrap();
+        crate::prop_assert_biteq!(result_v, result_s);
+        Ok(())
+    });
+}
+
 /// Test a binary vector-scalar function against a binary scalar function, applied elementwise.
 #[inline(never)]
 pub fn test_binary_scalar_rhs_elementwise<
diff --git a/library/portable-simd/crates/test_helpers/src/subnormals.rs b/library/portable-simd/crates/test_helpers/src/subnormals.rs
new file mode 100644
index 000000000..ec0f1fb24
--- /dev/null
+++ b/library/portable-simd/crates/test_helpers/src/subnormals.rs
@@ -0,0 +1,91 @@
+pub trait FlushSubnormals: Sized {
+    fn flush(self) -> Self {
+        self
+    }
+}
+
+impl<T> FlushSubnormals for *const T {}
+impl<T> FlushSubnormals for *mut T {}
+
+macro_rules! impl_float {
+    { $($ty:ty),* } => {
+        $(
+        impl FlushSubnormals for $ty {
+            fn flush(self) -> Self {
+                let is_f32 = core::mem::size_of::<Self>() == 4;
+                let ppc_flush = is_f32 && cfg!(all(
+                    any(target_arch = "powerpc", all(target_arch = "powerpc64", target_endian = "big")),
+                    target_feature = "altivec",
+                    not(target_feature = "vsx"),
+                ));
+                let arm_flush = is_f32 && cfg!(all(target_arch = "arm", target_feature = "neon"));
+                let flush = ppc_flush || arm_flush;
+                if flush && self.is_subnormal() {
+                    <$ty>::copysign(0., self)
+                } else {
+                    self
+                }
+            }
+        }
+        )*
+    }
+}
+
+macro_rules! impl_else {
+    { $($ty:ty),* } => {
+        $(
+        impl FlushSubnormals for $ty {}
+        )*
+    }
+}
+
+impl_float! { f32, f64 }
+impl_else! { i8, i16, i32, i64, isize, u8, u16, u32, u64, usize }
+
+/// AltiVec should flush subnormal inputs to zero, but QEMU seems to only flush outputs.
+/// https://gitlab.com/qemu-project/qemu/-/issues/1779
+#[cfg(all(
+    any(target_arch = "powerpc", target_arch = "powerpc64"),
+    target_feature = "altivec"
+))]
+fn in_buggy_qemu() -> bool {
+    use std::sync::OnceLock;
+    static BUGGY: OnceLock<bool> = OnceLock::new();
+
+    fn add(x: f32, y: f32) -> f32 {
+        #[cfg(target_arch = "powerpc")]
+        use core::arch::powerpc::*;
+        #[cfg(target_arch = "powerpc64")]
+        use core::arch::powerpc64::*;
+
+        let array: [f32; 4] =
+            unsafe { core::mem::transmute(vec_add(vec_splats(x), vec_splats(y))) };
+        array[0]
+    }
+
+    *BUGGY.get_or_init(|| add(-1.0857398e-38, 0.).is_sign_negative())
+}
+
+#[cfg(all(
+    any(target_arch = "powerpc", target_arch = "powerpc64"),
+    target_feature = "altivec"
+))]
+pub fn flush_in<T: FlushSubnormals>(x: T) -> T {
+    if in_buggy_qemu() {
+        x
+    } else {
+        x.flush()
+    }
+}
+
+#[cfg(not(all(
+    any(target_arch = "powerpc", target_arch = "powerpc64"),
+    target_feature = "altivec"
+)))]
+pub fn flush_in<T: FlushSubnormals>(x: T) -> T {
+    x.flush()
+}
+
+pub fn flush<T: FlushSubnormals>(x: T) -> T {
+    x.flush()
+}