1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
use crate::simd::{cmp::SimdPartialEq, LaneCount, Simd, SimdElement, SupportedLaneCount};
use core::ops::{Add, Mul};
use core::ops::{BitAnd, BitOr, BitXor};
use core::ops::{Div, Rem, Sub};
use core::ops::{Shl, Shr};

mod assign;
mod deref;
mod shift_scalar;
mod unary;

impl<I, T, const N: usize> core::ops::Index<I> for Simd<T, N>
where
    T: SimdElement,
    LaneCount<N>: SupportedLaneCount,
    I: core::slice::SliceIndex<[T]>,
{
    type Output = I::Output;
    #[inline]
    fn index(&self, index: I) -> &Self::Output {
        &self.as_array()[index]
    }
}

impl<I, T, const N: usize> core::ops::IndexMut<I> for Simd<T, N>
where
    T: SimdElement,
    LaneCount<N>: SupportedLaneCount,
    I: core::slice::SliceIndex<[T]>,
{
    #[inline]
    fn index_mut(&mut self, index: I) -> &mut Self::Output {
        &mut self.as_mut_array()[index]
    }
}

macro_rules! unsafe_base {
    ($lhs:ident, $rhs:ident, {$simd_call:ident}, $($_:tt)*) => {
        // Safety: $lhs and $rhs are vectors
        unsafe { core::intrinsics::simd::$simd_call($lhs, $rhs) }
    };
}

/// SAFETY: This macro should not be used for anything except Shl or Shr, and passed the appropriate shift intrinsic.
/// It handles performing a bitand in addition to calling the shift operator, so that the result
/// is well-defined: LLVM can return a poison value if you shl, lshr, or ashr if `rhs >= <Int>::BITS`
/// At worst, this will maybe add another instruction and cycle,
/// at best, it may open up more optimization opportunities,
/// or simply be elided entirely, especially for SIMD ISAs which default to this.
///
// FIXME: Consider implementing this in cg_llvm instead?
// cg_clif defaults to this, and scalar MIR shifts also default to wrapping
macro_rules! wrap_bitshift {
    ($lhs:ident, $rhs:ident, {$simd_call:ident}, $int:ident) => {
        #[allow(clippy::suspicious_arithmetic_impl)]
        // Safety: $lhs and the bitand result are vectors
        unsafe {
            core::intrinsics::simd::$simd_call(
                $lhs,
                $rhs.bitand(Simd::splat(<$int>::BITS as $int - 1)),
            )
        }
    };
}

/// SAFETY: This macro must only be used to impl Div or Rem and given the matching intrinsic.
/// It guards against LLVM's UB conditions for integer div or rem using masks and selects,
/// thus guaranteeing a Rust value returns instead.
///
/// |                  | LLVM | Rust
/// | :--------------: | :--- | :----------
/// | N {/,%} 0        | UB   | panic!()
/// | <$int>::MIN / -1 | UB   | <$int>::MIN
/// | <$int>::MIN % -1 | UB   | 0
///
macro_rules! int_divrem_guard {
    (   $lhs:ident,
        $rhs:ident,
        {   const PANIC_ZERO: &'static str = $zero:literal;
            $simd_call:ident, $op:tt
        },
        $int:ident ) => {
        if $rhs.simd_eq(Simd::splat(0 as _)).any() {
            panic!($zero);
        } else {
            // Prevent otherwise-UB overflow on the MIN / -1 case.
            let rhs = if <$int>::MIN != 0 {
                // This should, at worst, optimize to a few branchless logical ops
                // Ideally, this entire conditional should evaporate
                // Fire LLVM and implement those manually if it doesn't get the hint
                ($lhs.simd_eq(Simd::splat(<$int>::MIN))
                // type inference can break here, so cut an SInt to size
                & $rhs.simd_eq(Simd::splat(-1i64 as _)))
                .select(Simd::splat(1 as _), $rhs)
            } else {
                // Nice base case to make it easy to const-fold away the other branch.
                $rhs
            };

            // aarch64 div fails for arbitrary `v % 0`, mod fails when rhs is MIN, for non-powers-of-two
            // these operations aren't vectorized on aarch64 anyway
            #[cfg(target_arch = "aarch64")]
            {
                let mut out = Simd::splat(0 as _);
                for i in 0..Self::LEN {
                    out[i] = $lhs[i] $op rhs[i];
                }
                out
            }

            #[cfg(not(target_arch = "aarch64"))]
            {
                // Safety: $lhs and rhs are vectors
                unsafe { core::intrinsics::simd::$simd_call($lhs, rhs) }
            }
        }
    };
}

macro_rules! for_base_types {
    (   T = ($($scalar:ident),*);
        type Lhs = Simd<T, N>;
        type Rhs = Simd<T, N>;
        type Output = $out:ty;

        impl $op:ident::$call:ident {
            $macro_impl:ident $inner:tt
        }) => {
            $(
                impl<const N: usize> $op<Self> for Simd<$scalar, N>
                where
                    $scalar: SimdElement,
                    LaneCount<N>: SupportedLaneCount,
                {
                    type Output = $out;

                    #[inline]
                    #[must_use = "operator returns a new vector without mutating the inputs"]
                    // TODO: only useful for int Div::div, but we hope that this
                    // will essentially always always get inlined anyway.
                    #[track_caller]
                    fn $call(self, rhs: Self) -> Self::Output {
                        $macro_impl!(self, rhs, $inner, $scalar)
                    }
                }
            )*
    }
}

// A "TokenTree muncher": takes a set of scalar types `T = {};`
// type parameters for the ops it implements, `Op::fn` names,
// and a macro that expands into an expr, substituting in an intrinsic.
// It passes that to for_base_types, which expands an impl for the types,
// using the expanded expr in the function, and recurses with itself.
//
// tl;dr impls a set of ops::{Traits} for a set of types
macro_rules! for_base_ops {
    (
        T = $types:tt;
        type Lhs = Simd<T, N>;
        type Rhs = Simd<T, N>;
        type Output = $out:ident;
        impl $op:ident::$call:ident
            $inner:tt
        $($rest:tt)*
    ) => {
        for_base_types! {
            T = $types;
            type Lhs = Simd<T, N>;
            type Rhs = Simd<T, N>;
            type Output = $out;
            impl $op::$call
                $inner
        }
        for_base_ops! {
            T = $types;
            type Lhs = Simd<T, N>;
            type Rhs = Simd<T, N>;
            type Output = $out;
            $($rest)*
        }
    };
    ($($done:tt)*) => {
        // Done.
    }
}

// Integers can always accept add, mul, sub, bitand, bitor, and bitxor.
// For all of these operations, simd_* intrinsics apply wrapping logic.
for_base_ops! {
    T = (i8, i16, i32, i64, isize, u8, u16, u32, u64, usize);
    type Lhs = Simd<T, N>;
    type Rhs = Simd<T, N>;
    type Output = Self;

    impl Add::add {
        unsafe_base { simd_add }
    }

    impl Mul::mul {
        unsafe_base { simd_mul }
    }

    impl Sub::sub {
        unsafe_base { simd_sub }
    }

    impl BitAnd::bitand {
        unsafe_base { simd_and }
    }

    impl BitOr::bitor {
        unsafe_base { simd_or }
    }

    impl BitXor::bitxor {
        unsafe_base { simd_xor }
    }

    impl Div::div {
        int_divrem_guard {
            const PANIC_ZERO: &'static str = "attempt to divide by zero";
            simd_div, /
        }
    }

    impl Rem::rem {
        int_divrem_guard {
            const PANIC_ZERO: &'static str = "attempt to calculate the remainder with a divisor of zero";
            simd_rem, %
        }
    }

    // The only question is how to handle shifts >= <Int>::BITS?
    // Our current solution uses wrapping logic.
    impl Shl::shl {
        wrap_bitshift { simd_shl }
    }

    impl Shr::shr {
        wrap_bitshift {
            // This automatically monomorphizes to lshr or ashr, depending,
            // so it's fine to use it for both UInts and SInts.
            simd_shr
        }
    }
}

// We don't need any special precautions here:
// Floats always accept arithmetic ops, but may become NaN.
for_base_ops! {
    T = (f32, f64);
    type Lhs = Simd<T, N>;
    type Rhs = Simd<T, N>;
    type Output = Self;

    impl Add::add {
        unsafe_base { simd_add }
    }

    impl Mul::mul {
        unsafe_base { simd_mul }
    }

    impl Sub::sub {
        unsafe_base { simd_sub }
    }

    impl Div::div {
        unsafe_base { simd_div }
    }

    impl Rem::rem {
        unsafe_base { simd_rem }
    }
}