progress

library/stdarch/examples/play.rs

@@ -1,26 +1,47 @@
+#![feature(target_feature)]
+
 extern crate stdsimd;
 
 use std::env;
 use stdsimd as s;
 
 #[inline(never)]
-fn foobar(a: s::f64x2, b: s::f64x2) -> bool {
-    s::_mm_ucomieq_sd(a, b)
+#[target_feature = "+sse4.2"]
+fn index(needle: &str, haystack: &str) -> usize {
+    assert!(needle.len() <= 16 && haystack.len() <= 16);
+
+    let (needle_len, hay_len) = (needle.len(), haystack.len());
+
+    let mut needle = needle.to_string().into_bytes();
+    needle.resize(16, 0);
+    let vneedle = s::__m128i::from(s::u8x16::load(&needle, 0));
+
+    let mut haystack = haystack.to_string().into_bytes();
+    haystack.resize(16, 0);
+    let vhaystack = s::__m128i::from(s::u8x16::load(&haystack, 0));
+
+    s::_mm_cmpestri(
+        vneedle, needle_len as i32, vhaystack, hay_len as i32,
+        s::_SIDD_CMP_EQUAL_ORDERED) as usize
 }
 
 fn main() {
-    let x0: f64 = env::args().nth(1).unwrap().parse().unwrap();
-    let x1: f64 = env::args().nth(2).unwrap().parse().unwrap();
-    let x2: f64 = env::args().nth(3).unwrap().parse().unwrap();
-    let x3: f64 = env::args().nth(4).unwrap().parse().unwrap();
+    // let x0: f64 = env::args().nth(1).unwrap().parse().unwrap();
+    // let x1: f64 = env::args().nth(2).unwrap().parse().unwrap();
+    // let x2: f64 = env::args().nth(3).unwrap().parse().unwrap();
+    // let x3: f64 = env::args().nth(4).unwrap().parse().unwrap();
     // let y0: i32 = env::args().nth(5).unwrap().parse().unwrap();
     // let y1: i32 = env::args().nth(6).unwrap().parse().unwrap();
     // let y2: i32 = env::args().nth(7).unwrap().parse().unwrap();
     // let y3: i32 = env::args().nth(8).unwrap().parse().unwrap();
 
-    let a = s::f64x2::new(x0, x1);
-    let b = s::f64x2::new(x2, x3);
+    // let a = s::f64x2::new(x0, x1);
+    // let b = s::f64x2::new(x2, x3);
     // let r = s::_mm_cmplt_sd(a, b);
-    let r = foobar(a, b);
-    println!("{:?}", r);
+    // let r = foobar(a, b);
+
+
+    let needle = env::args().nth(1).unwrap();
+    let haystack = env::args().nth(2).unwrap();
+    println!("{:?}", index(&needle, &haystack));
 }

library/stdarch/src/x86/mod.rs

@@ -1,6 +1,7 @@
 // pub use self::sse::*;
 pub use self::sse2::*;
 pub use self::ssse3::*;
+pub use self::sse42::*;
 
 #[allow(non_camel_case_types)]
 pub type __m128i = ::v128::i8x16;

library/stdarch/src/x86/sse42.rs

@@ -319,7 +319,7 @@ mod tests {
         let vb = __m128i::from(u8x16::load(b, 0));
         let i = sse42::_mm_cmpestri(
             va, 3, vb, 6,
-            sse42::_SIDD_CMP_EQUAL_ORDERED | sse42::_SIDD_MOST_SIGNIFICANT);
+            sse42::_SIDD_CMP_EQUAL_ORDERED);
         assert_eq!(3, i);
     }
 }