diff --git a/library/stdarch/src/x86/sse3.rs b/library/stdarch/src/x86/sse3.rs
index 0354f83c1eb7..32231e557e97 100644
--- a/library/stdarch/src/x86/sse3.rs
+++ b/library/stdarch/src/x86/sse3.rs
@@ -40,6 +40,15 @@ pub unsafe fn _mm_hadd_ps(a: f32x4, b: f32x4) -> f32x4 {
     haddps(a, b)
 }
 
+/// Horizontally subtract adjacent pairs of double-precision (64-bit)
+/// floating-point elements in `a` and `b`, and pack the results.
+#[inline(always)]
+#[target_feature = "+sse3"]
+#[cfg_attr(test, assert_instr(hsub))]
+pub unsafe fn _mm_hsub_pd(a: f64x2, b: f64x2) -> f64x2 {
+    hsubpd(a, b)
+}
+
 /// Load 128-bits of integer data from unaligned memory.
 /// This intrinsic may perform better than `_mm_loadu_si128`
 /// when the data crosses a cache line boundary.
@@ -60,6 +69,8 @@ extern {
     fn haddpd(a: f64x2, b: f64x2) -> f64x2;
     #[link_name = "llvm.x86.sse3.hadd.ps"]
     fn haddps(a: f32x4, b: f32x4) -> f32x4;
+    #[link_name = "llvm.x86.sse3.hsub.pd"]
+    fn hsubpd(a: f64x2, b: f64x2) -> f64x2;
     #[link_name = "llvm.x86.sse3.ldu.dq"]
     fn lddqu(mem_addr: *const i8) -> __m128i;
 }
@@ -104,6 +115,14 @@ mod tests {
         assert_eq!(r, f32x4::new(4.0, -10.0, -80.0, -5.0));
     }
 
+    #[simd_test = "sse3"]
+    unsafe fn _mm_hsub_pd() {
+        let a = f64x2::new(-1.0, 5.0);
+        let b = f64x2::new(-100.0, 20.0);
+        let r = sse3::_mm_hsub_pd(a, b);
+        assert_eq!(r, f64x2::new(-6.0, -120.0));
+    }
+
     #[simd_test = "sse3"]
     unsafe fn _mm_lddqu_si128() {
         let a = i8x16::new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16);