Work around slow _mm_packus_epi16 in Rust 1.96 through 1.98, inclusive

See rust-lang/stdarch#2208

Author: hsivonen

Cargo.toml

@@ -5,14 +5,15 @@ version = "0.8.35"
 edition = '2024'
 authors = ["Henri Sivonen <hsivonen@hsivonen.fi>"]
 license = "(Apache-2.0 OR MIT) AND BSD-3-Clause"
-include = ["src/**/*.rs", "/data", "Cargo.toml", "COPYRIGHT", "LICENSE*", "README.md"]
+include = ["build.rs", "src/**/*.rs", "/data", "Cargo.toml", "COPYRIGHT", "LICENSE*", "README.md"]
 readme = "README.md"
 documentation = "https://docs.rs/encoding_rs/"
 homepage = "https://docs.rs/encoding_rs/"
 repository = "https://github.com/hsivonen/encoding_rs"
 keywords = ["encoding", "web", "unicode", "charset"]
 categories = ["text-processing", "encoding", "web-programming", "internationalization"]
 rust-version = "1.88"
+build = "build.rs"
 
 [features]
 default = ["alloc"]
@@ -58,6 +59,9 @@ serde_derive = "1.0"
 bincode = "1.0"
 serde_json = "1.0"
 
+[build-dependencies]
+rustversion = "1.0.19"
+
 [profile.release]
 lto = true
 

build.rs

@@ -0,0 +1,6 @@
+fn main() {
+    println!("cargo::rustc-check-cfg=cfg(slow_mm_packus_epi16)");
+    if rustversion::cfg!(all(since(1.96), before(1.99))) {
+        println!("cargo:rustc-cfg=slow_mm_packus_epi16");
+    }
+}

src/lib.rs

@@ -728,6 +728,22 @@
     all(feature = "simd-accel", target_endian = "little"),
     feature(portable_simd)
 )]
+// These are for working around
+// https://github.com/rust-lang/stdarch/issues/2208
+// https://github.com/rust-lang/rust/issues/159464
+// https://github.com/rust-lang/rust/pull/161558
+#![cfg_attr(
+    all(slow_mm_packus_epi16, feature = "simd-accel", target_feature = "sse2"),
+    feature(link_llvm_intrinsics)
+)]
+#![cfg_attr(
+    all(slow_mm_packus_epi16, feature = "simd-accel", target_feature = "sse2"),
+    feature(abi_unadjusted)
+)]
+#![cfg_attr(
+    all(slow_mm_packus_epi16, feature = "simd-accel", target_feature = "sse2"),
+    feature(simd_ffi)
+)]
 
 #[cfg(feature = "alloc")]
 #[cfg_attr(test, macro_use)]

src/simd_funcs.rs

@@ -54,17 +54,43 @@ cfg_if! {
         #[allow(unused)]
         use core::arch::x86_64::__m128i;
         use core::arch::x86_64::_mm_movemask_epi8;
+        #[rustversion::any(before(1.96), since(1.99))]
         use core::arch::x86_64::_mm_packus_epi16;
     } else if #[cfg(all(target_feature = "sse2", target_arch = "x86"))] {
         #[allow(unused)]
         use core::arch::x86::__m128i;
         use core::arch::x86::_mm_movemask_epi8;
+        #[rustversion::any(before(1.96), since(1.99))]
         use core::arch::x86::_mm_packus_epi16;
     } else {
 
     }
 }
 
+cfg_if! {
+    if #[cfg(target_feature = "sse2")] {
+        // See:
+        //
+        // https://github.com/rust-lang/stdarch/issues/2208
+        // https://github.com/rust-lang/rust/issues/159464
+        // https://github.com/rust-lang/rust/pull/161558
+        #[rustversion::all(since(1.96), before(1.99))]
+        unsafe fn _mm_packus_epi16(a: __m128i, b: __m128i) -> __m128i {
+            unsafe { packus(core::simd::i16x8::from(a), core::simd::i16x8::from(b)).into() }
+        }
+
+        #[rustversion::all(since(1.96), before(1.99))]
+        #[allow(improper_ctypes)]
+        unsafe extern "unadjusted" {
+            #[link_name = "llvm.x86.sse2.packuswb.128"]
+            fn packus(a: core::simd::i16x8, b: core::simd::i16x8) -> u8x16;
+        }
+
+    } else {
+
+    }
+}
+
 // #[inline(always)]
 // fn simd_byte_swap_u8(s: u8x16) -> u8x16 {
 //     unsafe {