Implement __bswap[sdt]i2 intrinsics

These can be emitted by gcc, at least if requested specifically via __builtin_bswap{32,64,128}.

Author: tea

README.md

@@ -154,6 +154,9 @@ rely on CI.
 - [ ] arm/unordsf2vfp.S
 - [x] ashldi3.c
 - [x] ashrdi3.c
+- [x] bswapdi2.c
+- [x] bswapsi2.c
+- [x] bswapti2.c
 - [x] comparedf2.c
 - [x] comparesf2.c
 - [x] divdf3.c

build.rs

@@ -161,6 +161,9 @@ fn configure_check_cfg() {
         "__ashlsi3",
         "__ashrdi3",
         "__ashrsi3",
+        "__bswapsi2",
+        "__bswapdi2",
+        "__bswapti2",
         "__clzsi2",
         "__divdi3",
         "__divsi3",

src/int/bswap.rs

@@ -0,0 +1,32 @@
+use crate::int::Int;
+
+fn bswap<I: Int + From<u8>>(x: I) -> I {
+    let mut ret = I::ZERO;
+    for (i, byte) in x.to_be_bytes().into_iter().enumerate() {
+        ret |= I::from(byte) << (i as u32 * 8);
+    }
+    ret
+}
+
+intrinsics! {
+    #[maybe_use_optimized_c_shim]
+    #[avr_skip]
+    /// Swaps bytes in 32-bit number
+    pub extern "C" fn __bswapsi2(x: u32) -> u32 {
+        bswap(x)
+    }
+
+    #[maybe_use_optimized_c_shim]
+    #[avr_skip]
+    /// Swaps bytes in 64-bit number
+    pub extern "C" fn __bswapdi2(x: u64) -> u64 {
+        bswap(x)
+    }
+
+    #[maybe_use_optimized_c_shim]
+    #[avr_skip]
+    /// Swaps bytes in 128-bit number
+    pub extern "C" fn __bswapti2(x: u128) -> u128 {
+        bswap(x)
+    }
+}

src/int/mod.rs

@@ -4,6 +4,7 @@ mod specialized_div_rem;
 
 pub mod addsub;
 mod big;
+pub mod bswap;
 pub mod leading_zeros;
 pub mod mul;
 pub mod sdiv;
@@ -62,6 +63,8 @@ pub(crate) trait Int: MinInt
     + ops::BitXor<Output = Self>
     + ops::BitAnd<Output = Self>
 {
+    type Bytes: IntoIterator<Item = u8>;
+
     /// LUT used for maximizing the space covered and minimizing the computational cost of fuzzing
     /// in `testcrate`. For example, Self = u128 produces [0,1,2,7,8,15,16,31,32,63,64,95,96,111,
     /// 112,119,120,125,126,127].
@@ -103,6 +106,8 @@ pub(crate) trait Int: MinInt
     fn overflowing_add(self, other: Self) -> (Self, bool);
     fn leading_zeros(self) -> u32;
     fn ilog2(self) -> u32;
+
+    fn to_be_bytes(self) -> Self::Bytes;
 }
 }
 
@@ -154,6 +159,8 @@ pub(crate) const fn make_fuzz_lengths(bits: u32) -> [u8; 20] {
 
 macro_rules! int_impl_common {
     ($ty:ty) => {
+        type Bytes = [u8; Self::BITS as usize / 8];
+
         fn from_bool(b: bool) -> Self {
             b as $ty
         }
@@ -205,6 +212,10 @@ macro_rules! int_impl_common {
         fn ilog2(self) -> u32 {
             <Self>::ilog2(self)
         }
+
+        fn to_be_bytes(self) -> Self::Bytes {
+            self.to_be_bytes()
+        }
     };
 }
 

testcrate/tests/misc.rs

@@ -92,6 +92,40 @@ fn leading_zeros() {
     })
 }
 
+#[test]
+#[cfg(not(target_arch = "avr"))]
+fn bswap() {
+    use compiler_builtins::int::bswap::{__bswapdi2, __bswapsi2};
+    fuzz(N, |x: u32| {
+        assert_eq!(x.swap_bytes(), __bswapsi2(x));
+    });
+    fuzz(N, |x: u64| {
+        assert_eq!(x.swap_bytes(), __bswapdi2(x));
+    });
+
+    assert_eq!(__bswapsi2(0x12345678u32), 0x78563412u32);
+    assert_eq!(__bswapsi2(0x00000001u32), 0x01000000u32);
+    assert_eq!(__bswapdi2(0x123456789ABCDEF0u64), 0xF0DEBC9A78563412u64);
+    assert_eq!(__bswapdi2(0x0200000001000000u64), 0x0000000100000002u64);
+
+    #[cfg(any(target_pointer_width = "32", target_pointer_width = "64"))]
+    {
+        use compiler_builtins::int::bswap::__bswapti2;
+        fuzz(N, |x: u128| {
+            assert_eq!(x.swap_bytes(), __bswapti2(x));
+        });
+
+        assert_eq!(
+            __bswapti2(0x123456789ABCDEF013579BDF02468ACEu128),
+            0xCE8A4602DF9B5713F0DEBC9A78563412u128
+        );
+        assert_eq!(
+            __bswapti2(0x04000000030000000200000001000000u128),
+            0x00000001000000020000000300000004u128
+        );
+    }
+}
+
 // This is approximate because of issues related to
 // https://github.com/rust-lang/rust/issues/73920.
 // TODO how do we resolve this indeterminacy?