Implement _mm_shuffle_epi8

Author: bjorn3

example/std_example.rs

@@ -197,6 +197,7 @@ unsafe fn test_simd() {
 
     test_mm_extract_epi8();
     test_mm_insert_epi16();
+    test_mm_shuffle_epi8();
 
     test_mm256_shuffle_epi8();
     test_mm256_permute2x128_si256();
@@ -345,6 +346,26 @@ unsafe fn test_mm_insert_epi16() {
     assert_eq_m128i(r, e);
 }
 
+#[cfg(target_arch = "x86_64")]
+#[target_feature(enable = "ssse3")]
+unsafe fn test_mm_shuffle_epi8() {
+    #[rustfmt::skip]
+        let a = _mm_setr_epi8(
+            1, 2, 3, 4, 5, 6, 7, 8,
+            9, 10, 11, 12, 13, 14, 15, 16,
+        );
+    #[rustfmt::skip]
+        let b = _mm_setr_epi8(
+            4, 128_u8 as i8, 4, 3,
+            24, 12, 6, 19,
+            12, 5, 5, 10,
+            4, 1, 8, 0,
+        );
+    let expected = _mm_setr_epi8(5, 0, 5, 4, 9, 13, 7, 4, 13, 6, 6, 11, 5, 2, 9, 1);
+    let r = _mm_shuffle_epi8(a, b);
+    assert_eq_m128i(r, expected);
+}
+
 #[cfg(target_arch = "x86_64")]
 #[target_feature(enable = "avx2")]
 unsafe fn test_mm256_shuffle_epi8() {

src/intrinsics/llvm_x86.rs

@@ -222,7 +222,7 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
                 _ => fx.bcx.ins().iconst(types::I32, 0),
             });
         }
-        "llvm.x86.avx2.pshuf.b" => {
+        "llvm.x86.ssse3.pshuf.b.128" | "llvm.x86.avx2.pshuf.b" => {
             let (a, b) = match args {
                 [a, b] => (a, b),
                 _ => bug!("wrong number of args for intrinsic {intrinsic}"),
@@ -241,15 +241,18 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>(
                 let res = fx.bcx.ins().select(is_zero, zero, a_lane);
                 ret.place_lane(fx, i).to_ptr().store(fx, res, MemFlags::trusted());
             }
-            for i in 16..32 {
-                let b_lane = b.value_lane(fx, i).load_scalar(fx);
-                let is_zero = fx.bcx.ins().band_imm(b_lane, 0x80);
-                let b_lane_masked = fx.bcx.ins().band_imm(b_lane, 0xf);
-                let a_idx = fx.bcx.ins().iadd_imm(b_lane_masked, 16);
-                let a_idx = fx.bcx.ins().uextend(fx.pointer_type, a_idx);
-                let a_lane = a.value_lane_dyn(fx, a_idx).load_scalar(fx);
-                let res = fx.bcx.ins().select(is_zero, zero, a_lane);
-                ret.place_lane(fx, i).to_ptr().store(fx, res, MemFlags::trusted());
+
+            if intrinsic == "llvm.x86.avx2.pshuf.b" {
+                for i in 16..32 {
+                    let b_lane = b.value_lane(fx, i).load_scalar(fx);
+                    let is_zero = fx.bcx.ins().band_imm(b_lane, 0x80);
+                    let b_lane_masked = fx.bcx.ins().band_imm(b_lane, 0xf);
+                    let a_idx = fx.bcx.ins().iadd_imm(b_lane_masked, 16);
+                    let a_idx = fx.bcx.ins().uextend(fx.pointer_type, a_idx);
+                    let a_lane = a.value_lane_dyn(fx, a_idx).load_scalar(fx);
+                    let res = fx.bcx.ins().select(is_zero, zero, a_lane);
+                    ret.place_lane(fx, i).to_ptr().store(fx, res, MemFlags::trusted());
+                }
             }
         }
         "llvm.x86.avx2.vperm2i128" => {