From 7f9585ca725029bcc7bad389f677071dcbf6669b Mon Sep 17 00:00:00 2001
From: Recep Aslantas <info@recp.me>
Date: Fri, 7 May 2021 01:46:24 +0300
Subject: [PATCH] arrm, neon: impove hadd performance

---
 include/cglm/simd/arm.h | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/include/cglm/simd/arm.h b/include/cglm/simd/arm.h
index d0c4bc3..50cec46 100644
--- a/include/cglm/simd/arm.h
+++ b/include/cglm/simd/arm.h
@@ -47,8 +47,13 @@ glmm_abs(float32x4_t v) {
 static inline
 float32x4_t
 glmm_vhadd(float32x4_t v) {
-  v = vaddq_f32(v, vrev64q_f32(v));
-  return vaddq_f32(v, vcombine_f32(vget_high_f32(v), vget_low_f32(v)));
+  return vaddq_f32(vaddq_f32(glmm_splat_x(v), glmm_splat_y(v)),
+                   vaddq_f32(glmm_splat_z(v), glmm_splat_w(v)));
+  /*
+   this seems slower:
+   v = vaddq_f32(v, vrev64q_f32(v));
+   return vaddq_f32(v, vcombine_f32(vget_high_f32(v), vget_low_f32(v)));
+   */
 }
 
 static inline