Merge pull request #645 from tea/ctz

Implement __ctz*i2 intrinsics

library/compiler-builtins/README.md

@@ -171,6 +171,9 @@ rely on CI.
 - [x] clzti2.c
 - [x] comparedf2.c
 - [x] comparesf2.c
+- [x] ctzdi2.c
+- [x] ctzsi2.c
+- [x] ctzti2.c
 - [x] divdf3.c
 - [x] divdi3.c
 - [x] divmoddi4.c
@@ -395,9 +398,6 @@ These builtins are never called by LLVM.
 - ~~arm/switchu8.S~~
 - ~~cmpdi2.c~~
 - ~~cmpti2.c~~
-- ~~ctzdi2.c~~
-- ~~ctzsi2.c~~
-- ~~ctzti2.c~~
 - ~~ffssi2.c~~
 - ~~ffsdi2.c~~ - this is [called by gcc][jemalloc-fail] though!
 - ~~ffsti2.c~~

library/compiler-builtins/build.rs

@@ -345,8 +345,6 @@ mod c {
             ("__addvdi3", "addvdi3.c"),
             ("__addvsi3", "addvsi3.c"),
             ("__cmpdi2", "cmpdi2.c"),
-            ("__ctzdi2", "ctzdi2.c"),
-            ("__ctzsi2", "ctzsi2.c"),
             ("__int_util", "int_util.c"),
             ("__mulvdi3", "mulvdi3.c"),
             ("__mulvsi3", "mulvsi3.c"),
@@ -380,7 +378,6 @@ mod c {
                 ("__absvti2", "absvti2.c"),
                 ("__addvti3", "addvti3.c"),
                 ("__cmpti2", "cmpti2.c"),
-                ("__ctzti2", "ctzti2.c"),
                 ("__ffsti2", "ffsti2.c"),
                 ("__mulvti3", "mulvti3.c"),
                 ("__negti2", "negti2.c"),

library/compiler-builtins/src/int/mod.rs

@@ -9,6 +9,7 @@ pub mod leading_zeros;
 pub mod mul;
 pub mod sdiv;
 pub mod shift;
+pub mod trailing_zeros;
 pub mod udiv;
 
 pub use big::{i256, u256};

library/compiler-builtins/src/int/trailing_zeros.rs

@@ -0,0 +1,64 @@
+use crate::int::{CastInto, Int};
+
+public_test_dep! {
+/// Returns number of trailing binary zeros in `x`.
+#[allow(dead_code)]
+pub(crate) fn trailing_zeros<T: Int + CastInto<u32> + CastInto<u16> + CastInto<u8>>(x: T) -> usize {
+    let mut x = x;
+    let mut r: u32 = 0;
+    let mut t: u32;
+
+    const { assert!(T::BITS <= 64) };
+    if T::BITS >= 64 {
+        r += ((CastInto::<u32>::cast(x) == 0) as u32) << 5; // if (x has no 32 small bits) t = 32 else 0
+        x >>= r; // remove 32 zero bits
+    }
+
+    if T::BITS >= 32 {
+        t = ((CastInto::<u16>::cast(x) == 0) as u32) << 4; // if (x has no 16 small bits) t = 16 else 0
+        r += t;
+        x >>= t;         // x = [0 - 0xFFFF] + higher garbage bits
+    }
+
+    const { assert!(T::BITS >= 16) };
+    t = ((CastInto::<u8>::cast(x) == 0) as u32) << 3;
+    x >>= t; // x = [0 - 0xFF] + higher garbage bits
+    r += t;
+
+    let mut x: u8 = x.cast();
+
+    t = (((x & 0x0F) == 0) as u32) << 2;
+    x >>= t; // x = [0 - 0xF] + higher garbage bits
+    r += t;
+
+    t = (((x & 0x3) == 0) as u32) << 1;
+    x >>= t;  // x = [0 - 0x3] + higher garbage bits
+    r += t;
+
+    x &= 3;
+
+    r as usize + ((2 - (x >> 1) as usize) & (((x & 1) == 0) as usize).wrapping_neg())
+}
+}
+
+intrinsics! {
+    /// Returns the number of trailing binary zeros in `x` (32 bit version).
+    pub extern "C" fn __ctzsi2(x: u32) -> usize {
+        trailing_zeros(x)
+    }
+
+    /// Returns the number of trailing binary zeros in `x` (64 bit version).
+    pub extern "C" fn __ctzdi2(x: u64) -> usize {
+        trailing_zeros(x)
+    }
+
+    /// Returns the number of trailing binary zeros in `x` (128 bit version).
+    pub extern "C" fn __ctzti2(x: u128) -> usize {
+        let lo = x as u64;
+        if lo == 0 {
+            64 + __ctzdi2((x >> 64) as u64)
+        } else {
+            __ctzdi2(lo)
+        }
+    }
+}

library/compiler-builtins/testcrate/tests/misc.rs

@@ -131,6 +131,49 @@ fn leading_zeros() {
     }
 }
 
+#[test]
+fn trailing_zeros() {
+    use compiler_builtins::int::trailing_zeros::{__ctzdi2, __ctzsi2, __ctzti2, trailing_zeros};
+    fuzz(N, |x: u32| {
+        if x == 0 {
+            return; // undefined value for an intrinsic
+        }
+        let tz = x.trailing_zeros() as usize;
+        let tz0 = __ctzsi2(x);
+        let tz1 = trailing_zeros(x);
+        if tz0 != tz {
+            panic!("__ctzsi2({}): std: {}, builtins: {}", x, tz, tz0);
+        }
+        if tz1 != tz {
+            panic!("trailing_zeros({}): std: {}, builtins: {}", x, tz, tz1);
+        }
+    });
+    fuzz(N, |x: u64| {
+        if x == 0 {
+            return; // undefined value for an intrinsic
+        }
+        let tz = x.trailing_zeros() as usize;
+        let tz0 = __ctzdi2(x);
+        let tz1 = trailing_zeros(x);
+        if tz0 != tz {
+            panic!("__ctzdi2({}): std: {}, builtins: {}", x, tz, tz0);
+        }
+        if tz1 != tz {
+            panic!("trailing_zeros({}): std: {}, builtins: {}", x, tz, tz1);
+        }
+    });
+    fuzz(N, |x: u128| {
+        if x == 0 {
+            return; // undefined value for an intrinsic
+        }
+        let tz = x.trailing_zeros() as usize;
+        let tz0 = __ctzti2(x);
+        if tz0 != tz {
+            panic!("__ctzti2({}): std: {}, builtins: {}", x, tz, tz0);
+        }
+    });
+}
+
 #[test]
 #[cfg(not(target_arch = "avr"))]
 fn bswap() {