diff --git a/src/builder.rs b/src/builder.rs
index ee9983830ff8..adcd6235b709 100644
--- a/src/builder.rs
+++ b/src/builder.rs
@@ -1597,7 +1597,6 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         unimplemented!();
     }
 
-
     pub fn vector_select(&mut self, cond: RValue<'gcc>, then_val: RValue<'gcc>, else_val: RValue<'gcc>) -> RValue<'gcc> {
         // cond is a vector of integers, not of bools.
         let cond_type = cond.get_type();
@@ -1607,10 +1606,12 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         let zeros = vec![self.context.new_rvalue_zero(element_type); num_units];
         let zeros = self.context.new_rvalue_from_vector(None, cond_type, &zeros);
 
+        let result_type = then_val.get_type();
+
         let masks = self.context.new_comparison(None, ComparisonOp::NotEquals, cond, zeros);
         // NOTE: masks is a vector of integers, but the values can be vectors of floats, so use bitcast to make
         // the & operation work.
-        let masks = self.bitcast_if_needed(masks, then_val.get_type());
+        let then_val = self.bitcast_if_needed(then_val, masks.get_type());
         let then_vals = masks & then_val;
 
         let minus_ones = vec![self.context.new_rvalue_from_int(element_type, -1); num_units];
@@ -1623,7 +1624,8 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         let else_val = self.context.new_bitcast(None, else_val, then_val.get_type());
         let else_vals = inverted_masks & else_val;
 
-        then_vals | else_vals
+        let res = then_vals | else_vals;
+        self.bitcast_if_needed(res, result_type)
     }
 }
 
diff --git a/src/intrinsic/simd.rs b/src/intrinsic/simd.rs
index 7d7890098265..fbfcebe46a13 100644
--- a/src/intrinsic/simd.rs
+++ b/src/intrinsic/simd.rs
@@ -93,14 +93,19 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>,
         let arg1_vector_type = arg1_type.unqualified().dyncast_vector().expect("vector type");
         let arg1_element_type = arg1_vector_type.get_element_type();
 
+        // NOTE: since the arguments can be vectors of floats, make sure the mask is a vector of
+        // integer.
+        let mask_element_type = bx.type_ix(arg1_element_type.get_size() as u64 * 8);
+        let vector_mask_type = bx.context.new_vector_type(mask_element_type, arg1_vector_type.get_num_units() as u64);
+
         let mut elements = vec![];
         let one = bx.context.new_rvalue_one(mask.get_type());
         for _ in 0..len {
-            let element = bx.context.new_cast(None, mask & one, arg1_element_type);
+            let element = bx.context.new_cast(None, mask & one, mask_element_type);
             elements.push(element);
             mask = mask >> one;
         }
-        let vector_mask = bx.context.new_rvalue_from_vector(None, arg1_type, &elements);
+        let vector_mask = bx.context.new_rvalue_from_vector(None, vector_mask_type, &elements);
 
         return Ok(bx.vector_select(vector_mask, arg1, args[2].immediate()));
     }