compiler_builtins/int/
leading_zeros.rs

1// Note: these functions happen to produce the correct `usize::leading_zeros(0)` value
2// without a explicit zero check. Zero is probably common enough that it could warrant
3// adding a zero check at the beginning, but `__clzsi2` has a precondition that `x != 0`.
4// Compilers will insert the check for zero in cases where it is needed.
5
6#[cfg(feature = "unstable-public-internals")]
7pub use implementation::{leading_zeros_default, leading_zeros_riscv};
8#[cfg(not(feature = "unstable-public-internals"))]
9pub(crate) use implementation::{leading_zeros_default, leading_zeros_riscv};
10
11mod implementation {
12    use crate::int::{CastInto, Int};
13
14    /// Returns the number of leading binary zeros in `x`.
15    #[allow(dead_code)]
16    pub fn leading_zeros_default<T: Int + CastInto<usize>>(x: T) -> usize {
17        // The basic idea is to test if the higher bits of `x` are zero and bisect the number
18        // of leading zeros. It is possible for all branches of the bisection to use the same
19        // code path by conditionally shifting the higher parts down to let the next bisection
20        // step work on the higher or lower parts of `x`. Instead of starting with `z == 0`
21        // and adding to the number of zeros, it is slightly faster to start with
22        // `z == usize::MAX.count_ones()` and subtract from the potential number of zeros,
23        // because it simplifies the final bisection step.
24        let mut x = x;
25        // the number of potential leading zeros
26        let mut z = T::BITS as usize;
27        // a temporary
28        let mut t: T;
29
30        const { assert!(T::BITS <= 64) };
31        if T::BITS >= 64 {
32            t = x >> 32;
33            if t != T::ZERO {
34                z -= 32;
35                x = t;
36            }
37        }
38        if T::BITS >= 32 {
39            t = x >> 16;
40            if t != T::ZERO {
41                z -= 16;
42                x = t;
43            }
44        }
45        const { assert!(T::BITS >= 16) };
46        t = x >> 8;
47        if t != T::ZERO {
48            z -= 8;
49            x = t;
50        }
51        t = x >> 4;
52        if t != T::ZERO {
53            z -= 4;
54            x = t;
55        }
56        t = x >> 2;
57        if t != T::ZERO {
58            z -= 2;
59            x = t;
60        }
61        // the last two bisections are combined into one conditional
62        t = x >> 1;
63        if t != T::ZERO { z - 2 } else { z - x.cast() }
64
65        // We could potentially save a few cycles by using the LUT trick from
66        // "https://embeddedgurus.com/state-space/2014/09/
67        // fast-deterministic-and-portable-counting-leading-zeros/".
68        // However, 256 bytes for a LUT is too large for embedded use cases. We could remove
69        // the last 3 bisections  and use this 16 byte LUT for the rest of the work:
70        //const LUT: [u8; 16] = [0, 1, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4];
71        //z -= LUT[x] as usize;
72        //z
73        // However, it ends up generating about the same number of instructions. When benchmarked
74        // on x86_64, it is slightly faster to use the LUT, but this is probably because of OOO
75        // execution effects. Changing to using a LUT and branching is risky for smaller cores.
76    }
77
78    // The above method does not compile well on RISC-V (because of the lack of predicated
79    // instructions), producing code with many branches or using an excessively long
80    // branchless solution. This method takes advantage of the set-if-less-than instruction on
81    // RISC-V that allows `(x >= power-of-two) as usize` to be branchless.
82
83    /// Returns the number of leading binary zeros in `x`.
84    #[allow(dead_code)]
85    pub fn leading_zeros_riscv<T: Int + CastInto<usize>>(x: T) -> usize {
86        let mut x = x;
87        // the number of potential leading zeros
88        let mut z = T::BITS;
89        // a temporary
90        let mut t: u32;
91
92        // RISC-V does not have a set-if-greater-than-or-equal instruction and
93        // `(x >= power-of-two) as usize` will get compiled into two instructions, but this is
94        // still the most optimal method. A conditional set can only be turned into a single
95        // immediate instruction if `x` is compared with an immediate `imm` (that can fit into
96        // 12 bits) like `x < imm` but not `imm < x` (because the immediate is always on the
97        // right). If we try to save an instruction by using `x < imm` for each bisection, we
98        // have to shift `x` left and compare with powers of two approaching `usize::MAX + 1`,
99        // but the immediate will never fit into 12 bits and never save an instruction.
100        const { assert!(T::BITS <= 64) };
101        if T::BITS >= 64 {
102            // If the upper 32 bits of `x` are not all 0, `t` is set to `1 << 5`, otherwise
103            // `t` is set to 0.
104            t = ((x >= (T::ONE << 32)) as u32) << 5;
105            // If `t` was set to `1 << 5`, then the upper 32 bits are shifted down for the
106            // next step to process.
107            x >>= t;
108            // If `t` was set to `1 << 5`, then we subtract 32 from the number of potential
109            // leading zeros
110            z -= t;
111        }
112        if T::BITS >= 32 {
113            t = ((x >= (T::ONE << 16)) as u32) << 4;
114            x >>= t;
115            z -= t;
116        }
117        const { assert!(T::BITS >= 16) };
118        t = ((x >= (T::ONE << 8)) as u32) << 3;
119        x >>= t;
120        z -= t;
121        t = ((x >= (T::ONE << 4)) as u32) << 2;
122        x >>= t;
123        z -= t;
124        t = ((x >= (T::ONE << 2)) as u32) << 1;
125        x >>= t;
126        z -= t;
127        t = (x >= (T::ONE << 1)) as u32;
128        x >>= t;
129        z -= t;
130        // All bits except the LSB are guaranteed to be zero for this final bisection step.
131        // If `x != 0` then `x == 1` and subtracts one potential zero from `z`.
132        z as usize - x.cast()
133    }
134}
135
136intrinsics! {
137    /// Returns the number of leading binary zeros in `x`
138    pub extern "C" fn __clzsi2(x: u32) -> usize {
139        if cfg!(any(target_arch = "riscv32", target_arch = "riscv64")) {
140            leading_zeros_riscv(x)
141        } else {
142            leading_zeros_default(x)
143        }
144    }
145
146    /// Returns the number of leading binary zeros in `x`
147    pub extern "C" fn __clzdi2(x: u64) -> usize {
148        if cfg!(any(target_arch = "riscv32", target_arch = "riscv64")) {
149            leading_zeros_riscv(x)
150        } else {
151            leading_zeros_default(x)
152        }
153    }
154
155    /// Returns the number of leading binary zeros in `x`
156    pub extern "C" fn __clzti2(x: u128) -> usize {
157        let hi = (x >> 64) as u64;
158        if hi == 0 {
159            64 + __clzdi2(x as u64)
160        } else {
161            __clzdi2(hi)
162        }
163    }
164}