diff --git a/simd/src/arm/mod.rs b/simd/src/arm/mod.rs
index c39380b2..43d410ec 100644
--- a/simd/src/arm/mod.rs
+++ b/simd/src/arm/mod.rs
@@ -79,12 +79,12 @@ impl F32x2 {
 
     #[inline]
     pub fn min(self, other: F32x2) -> F32x2 {
-        unsafe { F32x2(simd_fmin(self.0, other.0)) }
+        unsafe { F32x2(simd_minimum_number_nsz(self.0, other.0)) }
     }
 
     #[inline]
     pub fn max(self, other: F32x2) -> F32x2 {
-        unsafe { F32x2(simd_fmax(self.0, other.0)) }
+        unsafe { F32x2(simd_maximum_number_nsz(self.0, other.0)) }
     }
 
     #[inline]
@@ -268,12 +268,12 @@ impl F32x4 {
 
     #[inline]
     pub fn min(self, other: F32x4) -> F32x4 {
-        unsafe { F32x4(simd_fmin(self.0, other.0)) }
+        unsafe { F32x4(simd_minimum_number_nsz(self.0, other.0)) }
     }
 
     #[inline]
     pub fn max(self, other: F32x4) -> F32x4 {
-        unsafe { F32x4(simd_fmax(self.0, other.0)) }
+        unsafe { F32x4(simd_maximum_number_nsz(self.0, other.0)) }
     }
 
     #[inline]
@@ -604,12 +604,12 @@ impl I32x4 {
 
     #[inline]
     pub fn max(self, other: I32x4) -> I32x4 {
-        unsafe { I32x4(simd_cast(simd_fmax(self.to_f32x4().0, other.to_f32x4().0))) }
+        unsafe { I32x4(simd_cast(simd_maximum_number_nsz(self.to_f32x4().0, other.to_f32x4().0))) }
     }
 
     #[inline]
     pub fn min(self, other: I32x4) -> I32x4 {
-        unsafe { I32x4(simd_cast(simd_fmin(self.to_f32x4().0, other.to_f32x4().0))) }
+        unsafe { I32x4(simd_cast(simd_minimum_number_nsz(self.to_f32x4().0, other.to_f32x4().0))) }
     }
 
     // Packed comparisons
diff --git a/simd/src/test.rs b/simd/src/test.rs
index 07d45513..0f287525 100644
--- a/simd/src/test.rs
+++ b/simd/src/test.rs
@@ -37,7 +37,17 @@ fn test_f32x4_accessors_and_mutators() {
 fn test_f32x4_basic_ops() {
     let a = F32x4::new(1.0, 3.0, 5.0, 7.0);
     let b = F32x4::new(2.0, 2.0, 6.0, 6.0);
-    assert_eq!(a.approx_recip(), F32x4::new(0.99975586, 0.333313, 0.19995117, 0.14282227));
+    let approx_recip = a.approx_recip();
+
+    // The greatest diff we observe between the correct answer to the reciprocal and the answer that
+    // llvm gives us is 1.0 vs 0.9980469, so we need to have the allowable diff be 0.002 so that it
+    // passes. It's approximate, so it's fine to be a bit off.
+    const ALLOWABLE_RECIP_DIFF: f32 = 0.002;
+    assert!((approx_recip[0] - (1. / 1.)).abs() < ALLOWABLE_RECIP_DIFF);
+    assert!((approx_recip[1] - (1. / 3.)).abs() < ALLOWABLE_RECIP_DIFF);
+    assert!((approx_recip[2] - (1. / 5.)).abs() < ALLOWABLE_RECIP_DIFF);
+    assert!((approx_recip[3] - (1. / 7.)).abs() < ALLOWABLE_RECIP_DIFF);
+
     assert_eq!(a.min(b), F32x4::new(1.0, 2.0, 5.0, 6.0));
     assert_eq!(a.max(b), F32x4::new(2.0, 3.0, 6.0, 7.0));
     let c = F32x4::new(-1.0, 1.3, -20.0, 3.6);