Implement llvm.x86.avx2.pmovmskb llvm intrinsic

Author: bjorn3

example/std_example.rs

@@ -65,6 +65,8 @@ unsafe fn test_simd() {
     assert_eq!(std::mem::transmute::<_, [u16; 8]>(cmp_lt), [0, 0, 0, 0, 0, 0, 0, 0]);
 
     test_mm_slli_si128();
+    test_mm_movemask_epi8();
+    test_mm256_movemask_epi8();
 }
 
 #[target_feature(enable = "sse2")]
@@ -109,6 +111,31 @@ unsafe fn test_mm_slli_si128() {
     assert_eq_m128i(r, _mm_set1_epi8(0));
 }
 
+#[target_feature(enable = "sse2")]
+unsafe fn test_mm_movemask_epi8() {
+    use std::arch::x86_64::*;
+
+    #[rustfmt::skip]
+    let a = _mm_setr_epi8(
+        0b1000_0000u8 as i8, 0b0, 0b1000_0000u8 as i8, 0b01,
+        0b0101, 0b1111_0000u8 as i8, 0, 0,
+        0, 0, 0b1111_0000u8 as i8, 0b0101,
+        0b01, 0b1000_0000u8 as i8, 0b0, 0b1000_0000u8 as i8,
+    );
+    let r = _mm_movemask_epi8(a);
+    assert_eq!(r, 0b10100100_00100101);
+}
+
+#[target_feature(enable = "avx2")]
+unsafe fn test_mm256_movemask_epi8() {
+    use std::arch::x86_64::*;
+
+    let a = _mm256_set1_epi8(-1);
+    let r = _mm256_movemask_epi8(a);
+    let e = -1;
+    assert_eq!(r, e);
+}
+
 fn assert_eq_m128i(x: std::arch::x86_64::__m128i, y: std::arch::x86_64::__m128i) {
     unsafe {
         assert_eq!(std::mem::transmute::<_, [u8; 16]>(x), std::mem::transmute::<_, [u8; 16]>(x));

src/llvm_intrinsics.rs

@@ -33,15 +33,15 @@ pub fn codegen_llvm_intrinsic_call<'a, 'tcx: 'a>(
             crate::trap::trap_unimplemented(fx, intrinsic);
         };
 
-        // Used by _mm_movemask_epi8
-        llvm.x86.sse2.pmovmskb.128, (c a) {
+        // Used by `_mm_movemask_epi8` and `_mm256_movemask_epi8`
+        llvm.x86.sse2.pmovmskb.128 | llvm.x86.avx2.pmovmskb, (c a) {
             let (lane_layout, lane_count) = crate::intrinsics::lane_type_and_count(fx, a.layout(), intrinsic);
             assert_eq!(lane_layout.ty.sty, fx.tcx.types.i8.sty);
-            assert_eq!(lane_count, 16);
+            assert!(lane_count == 16 || lane_count == 32);
 
             let mut res = fx.bcx.ins().iconst(types::I32, 0);
 
-            for lane in 0..16 {
+            for lane in 0..lane_count {
                 let a_lane = a.value_field(fx, mir::Field::new(lane.try_into().unwrap())).load_scalar(fx);
                 let a_lane_sign = fx.bcx.ins().ushr_imm(a_lane, 7); // extract sign bit of 8bit int
                 let a_lane_sign = fx.bcx.ins().uextend(types::I32, a_lane_sign);
@@ -65,6 +65,5 @@ pub fn codegen_llvm_intrinsic_call<'a, 'tcx: 'a>(
 // llvm.x86.avx2.vperm2i128
 // llvm.x86.ssse3.pshuf.b.128
 // llvm.x86.avx2.pshuf.b
-// llvm.x86.avx2.pmovmskb
 // llvm.x86.avx2.psrli.w
 // llvm.x86.sse2.psrli.w