core/stdarch/crates/core_arch/src/x86_64/
sse2.rs

1//! `x86_64`'s Streaming SIMD Extensions 2 (SSE2)
2
3use crate::core_arch::x86::*;
4
5#[cfg(test)]
6use stdarch_test::assert_instr;
7
8#[allow(improper_ctypes)]
9unsafe extern "C" {
10    #[link_name = "llvm.x86.sse2.cvtsd2si64"]
11    fn cvtsd2si64(a: __m128d) -> i64;
12    #[link_name = "llvm.x86.sse2.cvttsd2si64"]
13    fn cvttsd2si64(a: __m128d) -> i64;
14}
15
16/// Converts the lower double-precision (64-bit) floating-point element in a to
17/// a 64-bit integer.
18///
19/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvtsd_si64)
20#[inline]
21#[target_feature(enable = "sse2")]
22#[cfg_attr(test, assert_instr(cvtsd2si))]
23#[stable(feature = "simd_x86", since = "1.27.0")]
24pub fn _mm_cvtsd_si64(a: __m128d) -> i64 {
25    unsafe { cvtsd2si64(a) }
26}
27
28/// Alias for `_mm_cvtsd_si64`
29///
30/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvtsd_si64x)
31#[inline]
32#[target_feature(enable = "sse2")]
33#[cfg_attr(test, assert_instr(cvtsd2si))]
34#[stable(feature = "simd_x86", since = "1.27.0")]
35pub fn _mm_cvtsd_si64x(a: __m128d) -> i64 {
36    _mm_cvtsd_si64(a)
37}
38
39/// Converts the lower double-precision (64-bit) floating-point element in `a`
40/// to a 64-bit integer with truncation.
41///
42/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvttsd_si64)
43#[inline]
44#[target_feature(enable = "sse2")]
45#[cfg_attr(test, assert_instr(cvttsd2si))]
46#[stable(feature = "simd_x86", since = "1.27.0")]
47pub fn _mm_cvttsd_si64(a: __m128d) -> i64 {
48    unsafe { cvttsd2si64(a) }
49}
50
51/// Alias for `_mm_cvttsd_si64`
52///
53/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvttsd_si64x)
54#[inline]
55#[target_feature(enable = "sse2")]
56#[cfg_attr(test, assert_instr(cvttsd2si))]
57#[stable(feature = "simd_x86", since = "1.27.0")]
58pub fn _mm_cvttsd_si64x(a: __m128d) -> i64 {
59    _mm_cvttsd_si64(a)
60}
61
62/// Stores a 64-bit integer value in the specified memory location.
63/// To minimize caching, the data is flagged as non-temporal (unlikely to be
64/// used again soon).
65///
66/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_stream_si64)
67///
68/// # Safety of non-temporal stores
69///
70/// After using this intrinsic, but before any other access to the memory that this intrinsic
71/// mutates, a call to [`_mm_sfence`] must be performed by the thread that used the intrinsic. In
72/// particular, functions that call this intrinsic should generally call `_mm_sfence` before they
73/// return.
74///
75/// See [`_mm_sfence`] for details.
76#[inline]
77#[target_feature(enable = "sse2")]
78#[cfg_attr(test, assert_instr(movnti))]
79#[stable(feature = "simd_x86", since = "1.27.0")]
80pub unsafe fn _mm_stream_si64(mem_addr: *mut i64, a: i64) {
81    // see #1541, we should use inline asm to be sure, because LangRef isn't clear enough
82    crate::arch::asm!(
83        vps!("movnti", ",{a}"),
84        p = in(reg) mem_addr,
85        a = in(reg) a,
86        options(nostack, preserves_flags),
87    );
88}
89
90/// Returns a vector whose lowest element is `a` and all higher elements are
91/// `0`.
92///
93/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvtsi64_si128)
94#[inline]
95#[target_feature(enable = "sse2")]
96#[cfg_attr(test, assert_instr(movq))]
97#[stable(feature = "simd_x86", since = "1.27.0")]
98#[rustc_const_unstable(feature = "stdarch_const_x86", issue = "149298")]
99pub const fn _mm_cvtsi64_si128(a: i64) -> __m128i {
100    _mm_set_epi64x(0, a)
101}
102
103/// Returns a vector whose lowest element is `a` and all higher elements are
104/// `0`.
105///
106/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvtsi64x_si128)
107#[inline]
108#[target_feature(enable = "sse2")]
109#[cfg_attr(test, assert_instr(movq))]
110#[stable(feature = "simd_x86", since = "1.27.0")]
111#[rustc_const_unstable(feature = "stdarch_const_x86", issue = "149298")]
112pub const fn _mm_cvtsi64x_si128(a: i64) -> __m128i {
113    _mm_cvtsi64_si128(a)
114}
115
116/// Returns the lowest element of `a`.
117///
118/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvtsi128_si64)
119#[inline]
120#[target_feature(enable = "sse2")]
121#[cfg_attr(test, assert_instr(movq))]
122#[stable(feature = "simd_x86", since = "1.27.0")]
123#[rustc_const_unstable(feature = "stdarch_const_x86", issue = "149298")]
124pub const fn _mm_cvtsi128_si64(a: __m128i) -> i64 {
125    unsafe { simd_extract!(a.as_i64x2(), 0) }
126}
127
128/// Returns the lowest element of `a`.
129///
130/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvtsi128_si64x)
131#[inline]
132#[target_feature(enable = "sse2")]
133#[cfg_attr(test, assert_instr(movq))]
134#[stable(feature = "simd_x86", since = "1.27.0")]
135#[rustc_const_unstable(feature = "stdarch_const_x86", issue = "149298")]
136pub const fn _mm_cvtsi128_si64x(a: __m128i) -> i64 {
137    _mm_cvtsi128_si64(a)
138}
139
140/// Returns `a` with its lower element replaced by `b` after converting it to
141/// an `f64`.
142///
143/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvtsi64_sd)
144#[inline]
145#[target_feature(enable = "sse2")]
146#[cfg_attr(test, assert_instr(cvtsi2sd))]
147#[stable(feature = "simd_x86", since = "1.27.0")]
148#[rustc_const_unstable(feature = "stdarch_const_x86", issue = "149298")]
149pub const fn _mm_cvtsi64_sd(a: __m128d, b: i64) -> __m128d {
150    unsafe { simd_insert!(a, 0, b as f64) }
151}
152
153/// Returns `a` with its lower element replaced by `b` after converting it to
154/// an `f64`.
155///
156/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvtsi64x_sd)
157#[inline]
158#[target_feature(enable = "sse2")]
159#[cfg_attr(test, assert_instr(cvtsi2sd))]
160#[stable(feature = "simd_x86", since = "1.27.0")]
161#[rustc_const_unstable(feature = "stdarch_const_x86", issue = "149298")]
162pub const fn _mm_cvtsi64x_sd(a: __m128d, b: i64) -> __m128d {
163    _mm_cvtsi64_sd(a, b)
164}
165
166#[cfg(test)]
167mod tests {
168    use crate::core_arch::arch::x86_64::*;
169    use crate::core_arch::assert_eq_const as assert_eq;
170    use std::boxed;
171    use std::ptr;
172    use stdarch_test::simd_test;
173
174    #[simd_test(enable = "sse2")]
175    fn test_mm_cvtsd_si64() {
176        let r = _mm_cvtsd_si64(_mm_setr_pd(-2.0, 5.0));
177        assert_eq!(r, -2_i64);
178
179        let r = _mm_cvtsd_si64(_mm_setr_pd(f64::MAX, f64::MIN));
180        assert_eq!(r, i64::MIN);
181    }
182
183    #[simd_test(enable = "sse2")]
184    fn test_mm_cvtsd_si64x() {
185        let r = _mm_cvtsd_si64x(_mm_setr_pd(f64::NAN, f64::NAN));
186        assert_eq!(r, i64::MIN);
187    }
188
189    #[simd_test(enable = "sse2")]
190    fn test_mm_cvttsd_si64() {
191        let a = _mm_setr_pd(-1.1, 2.2);
192        let r = _mm_cvttsd_si64(a);
193        assert_eq!(r, -1_i64);
194    }
195
196    #[simd_test(enable = "sse2")]
197    fn test_mm_cvttsd_si64x() {
198        let a = _mm_setr_pd(f64::NEG_INFINITY, f64::NAN);
199        let r = _mm_cvttsd_si64x(a);
200        assert_eq!(r, i64::MIN);
201    }
202
203    #[simd_test(enable = "sse2")]
204    // Miri cannot support this until it is clear how it fits in the Rust memory model
205    // (non-temporal store)
206    #[cfg_attr(miri, ignore)]
207    fn test_mm_stream_si64() {
208        let a: i64 = 7;
209        let mut mem = boxed::Box::<i64>::new(-1);
210        unsafe {
211            _mm_stream_si64(ptr::addr_of_mut!(*mem), a);
212        }
213        _mm_sfence();
214        assert_eq!(a, *mem);
215    }
216
217    #[simd_test(enable = "sse2")]
218    const fn test_mm_cvtsi64_si128() {
219        let r = _mm_cvtsi64_si128(5);
220        assert_eq_m128i(r, _mm_setr_epi64x(5, 0));
221    }
222
223    #[simd_test(enable = "sse2")]
224    const fn test_mm_cvtsi128_si64() {
225        let r = _mm_cvtsi128_si64(_mm_setr_epi64x(5, 0));
226        assert_eq!(r, 5);
227    }
228
229    #[simd_test(enable = "sse2")]
230    const fn test_mm_cvtsi64_sd() {
231        let a = _mm_set1_pd(3.5);
232        let r = _mm_cvtsi64_sd(a, 5);
233        assert_eq_m128d(r, _mm_setr_pd(5.0, 3.5));
234    }
235}
core/stdarch/crates/core_arch/src/x86_64/sse2.rs

core/stdarch/crates/core_arch/src/x86_64/
sse2.rs
