|
| 1 | +// This test check that even if we mixup target feature of function with homogenous floats, |
| 2 | +// the abi is sound and still produce the right answer. |
| 3 | +// |
| 4 | +// This is basically the same test as src/test/ui/simd/target-feature-mixup.rs but for floats and |
| 5 | +// without #[repr(simd)] |
| 6 | + |
| 7 | +// run-pass |
| 8 | +// ignore-emscripten |
| 9 | +// ignore-sgx no processes |
| 10 | + |
| 11 | +#![feature(avx512_target_feature)] |
| 12 | + |
| 13 | +#![allow(overflowing_literals)] |
| 14 | +#![allow(unused_variables)] |
| 15 | + |
| 16 | +use std::process::{Command, ExitStatus}; |
| 17 | +use std::env; |
| 18 | + |
| 19 | +fn main() { |
| 20 | + if let Some(level) = env::args().nth(1) { |
| 21 | + return test::main(&level) |
| 22 | + } |
| 23 | + |
| 24 | + match std::env::var("TARGET") { |
| 25 | + Ok(s) => { |
| 26 | + // Skip this tests on i586-unknown-linux-gnu where sse2 is disabled |
| 27 | + if s.contains("i586") { |
| 28 | + return |
| 29 | + } |
| 30 | + } |
| 31 | + Err(_) => return, |
| 32 | + } |
| 33 | + |
| 34 | + let me = env::current_exe().unwrap(); |
| 35 | + for level in ["sse", "avx", "avx512"].iter() { |
| 36 | + let status = Command::new(&me).arg(level).status().unwrap(); |
| 37 | + if status.success() { |
| 38 | + println!("success with {}", level); |
| 39 | + continue |
| 40 | + } |
| 41 | + |
| 42 | + // We don't actually know if our computer has the requisite target features |
| 43 | + // for the test below. Testing for that will get added to libstd later so |
| 44 | + // for now just assume sigill means this is a machine that can't run this test. |
| 45 | + if is_sigill(status) { |
| 46 | + println!("sigill with {}, assuming spurious", level); |
| 47 | + continue |
| 48 | + } |
| 49 | + panic!("invalid status at {}: {}", level, status); |
| 50 | + } |
| 51 | +} |
| 52 | + |
| 53 | +#[cfg(unix)] |
| 54 | +fn is_sigill(status: ExitStatus) -> bool { |
| 55 | + use std::os::unix::prelude::*; |
| 56 | + status.signal() == Some(4) |
| 57 | +} |
| 58 | + |
| 59 | +#[cfg(windows)] |
| 60 | +fn is_sigill(status: ExitStatus) -> bool { |
| 61 | + status.code() == Some(0xc000001d) |
| 62 | +} |
| 63 | + |
| 64 | +#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 65 | +#[allow(nonstandard_style)] |
| 66 | +mod test { |
| 67 | + #[derive(PartialEq, Debug, Clone, Copy)] |
| 68 | + struct f32x2(f32, f32); |
| 69 | + |
| 70 | + #[derive(PartialEq, Debug, Clone, Copy)] |
| 71 | + struct f32x4(f32, f32, f32, f32); |
| 72 | + |
| 73 | + #[derive(PartialEq, Debug, Clone, Copy)] |
| 74 | + struct f32x8(f32, f32, f32, f32, f32, f32, f32, f32); |
| 75 | + |
| 76 | + pub fn main(level: &str) { |
| 77 | + unsafe { |
| 78 | + main_normal(level); |
| 79 | + main_sse(level); |
| 80 | + if level == "sse" { |
| 81 | + return |
| 82 | + } |
| 83 | + main_avx(level); |
| 84 | + if level == "avx" { |
| 85 | + return |
| 86 | + } |
| 87 | + main_avx512(level); |
| 88 | + } |
| 89 | + } |
| 90 | + |
| 91 | + macro_rules! mains { |
| 92 | + ($( |
| 93 | + $(#[$attr:meta])* |
| 94 | + unsafe fn $main:ident(level: &str) { |
| 95 | + ... |
| 96 | + } |
| 97 | + )*) => ($( |
| 98 | + $(#[$attr])* |
| 99 | + unsafe fn $main(level: &str) { |
| 100 | + let m128 = f32x2(1., 2.); |
| 101 | + let m256 = f32x4(3., 4., 5., 6.); |
| 102 | + let m512 = f32x8(7., 8., 9., 10., 11., 12., 13., 14.); |
| 103 | + assert_eq!(id_sse_128(m128), m128); |
| 104 | + assert_eq!(id_sse_256(m256), m256); |
| 105 | + assert_eq!(id_sse_512(m512), m512); |
| 106 | + |
| 107 | + if level == "sse" { |
| 108 | + return |
| 109 | + } |
| 110 | + assert_eq!(id_avx_128(m128), m128); |
| 111 | + assert_eq!(id_avx_256(m256), m256); |
| 112 | + assert_eq!(id_avx_512(m512), m512); |
| 113 | + |
| 114 | + if level == "avx" { |
| 115 | + return |
| 116 | + } |
| 117 | + assert_eq!(id_avx512_128(m128), m128); |
| 118 | + assert_eq!(id_avx512_256(m256), m256); |
| 119 | + assert_eq!(id_avx512_512(m512), m512); |
| 120 | + } |
| 121 | + )*) |
| 122 | + } |
| 123 | + |
| 124 | + mains! { |
| 125 | + unsafe fn main_normal(level: &str) { ... } |
| 126 | + #[target_feature(enable = "sse2")] |
| 127 | + unsafe fn main_sse(level: &str) { ... } |
| 128 | + #[target_feature(enable = "avx")] |
| 129 | + unsafe fn main_avx(level: &str) { ... } |
| 130 | + #[target_feature(enable = "avx512bw")] |
| 131 | + unsafe fn main_avx512(level: &str) { ... } |
| 132 | + } |
| 133 | + |
| 134 | + #[target_feature(enable = "sse2")] |
| 135 | + unsafe fn id_sse_128(a: f32x2) -> f32x2 { |
| 136 | + assert_eq!(a, f32x2(1., 2.)); |
| 137 | + a.clone() |
| 138 | + } |
| 139 | + |
| 140 | + #[target_feature(enable = "sse2")] |
| 141 | + unsafe fn id_sse_256(a: f32x4) -> f32x4 { |
| 142 | + assert_eq!(a, f32x4(3., 4., 5., 6.)); |
| 143 | + a.clone() |
| 144 | + } |
| 145 | + |
| 146 | + #[target_feature(enable = "sse2")] |
| 147 | + unsafe fn id_sse_512(a: f32x8) -> f32x8 { |
| 148 | + assert_eq!(a, f32x8(7., 8., 9., 10., 11., 12., 13., 14.)); |
| 149 | + a.clone() |
| 150 | + } |
| 151 | + |
| 152 | + #[target_feature(enable = "avx")] |
| 153 | + unsafe fn id_avx_128(a: f32x2) -> f32x2 { |
| 154 | + assert_eq!(a, f32x2(1., 2.)); |
| 155 | + a.clone() |
| 156 | + } |
| 157 | + |
| 158 | + #[target_feature(enable = "avx")] |
| 159 | + unsafe fn id_avx_256(a: f32x4) -> f32x4 { |
| 160 | + assert_eq!(a, f32x4(3., 4., 5., 6.)); |
| 161 | + a.clone() |
| 162 | + } |
| 163 | + |
| 164 | + #[target_feature(enable = "avx")] |
| 165 | + unsafe fn id_avx_512(a: f32x8) -> f32x8 { |
| 166 | + assert_eq!(a, f32x8(7., 8., 9., 10., 11., 12., 13., 14.)); |
| 167 | + a.clone() |
| 168 | + } |
| 169 | + |
| 170 | + #[target_feature(enable = "avx512bw")] |
| 171 | + unsafe fn id_avx512_128(a: f32x2) -> f32x2 { |
| 172 | + assert_eq!(a, f32x2(1., 2.)); |
| 173 | + a.clone() |
| 174 | + } |
| 175 | + |
| 176 | + #[target_feature(enable = "avx512bw")] |
| 177 | + unsafe fn id_avx512_256(a: f32x4) -> f32x4 { |
| 178 | + assert_eq!(a, f32x4(3., 4., 5., 6.)); |
| 179 | + a.clone() |
| 180 | + } |
| 181 | + |
| 182 | + #[target_feature(enable = "avx512bw")] |
| 183 | + unsafe fn id_avx512_512(a: f32x8) -> f32x8 { |
| 184 | + assert_eq!(a, f32x8(7., 8., 9., 10., 11., 12., 13., 14.)); |
| 185 | + a.clone() |
| 186 | + } |
| 187 | +} |
| 188 | + |
| 189 | +#[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))] |
| 190 | +mod test { |
| 191 | + pub fn main(level: &str) {} |
| 192 | +} |
0 commit comments