From a47460b4c3582f416f84570ea3ea0c8e593c4076 Mon Sep 17 00:00:00 2001
From: Dmitry Babokin <babokin@gmail.com>
Date: Thu, 2 May 2013 00:12:16 +0200
Subject: [PATCH] Efficient library implementation of broadcast

---
 builtins/util.m4 | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/builtins/util.m4 b/builtins/util.m4
index 36ebdd2a..6dda67b9 100644
--- a/builtins/util.m4
+++ b/builtins/util.m4
@@ -569,10 +569,9 @@ divert`'dnl
 define(`shuffles', `
 define <WIDTH x $1> @__broadcast_$1(<WIDTH x $1>, i32) nounwind readnone alwaysinline {
   %v = extractelement <WIDTH x $1> %0, i32 %1
-  %r_0 = insertelement <WIDTH x $1> undef, $1 %v, i32 0
-forloop(i, 1, eval(WIDTH-1), `  %r_`'i = insertelement <WIDTH x $1> %r_`'eval(i-1), $1 %v, i32 i
-')
-  ret <WIDTH x $1> %r_`'eval(WIDTH-1)
+  %broadcast_init = insertelement <WIDTH x $1> undef, $1 %v, i32 0
+  %broadcast = shufflevector <WIDTH x $1> %broadcast_init, <WIDTH x $1> undef, <WIDTH x i32> zeroinitializer 
+  ret <WIDTH x $1> %broadcast
 }
 
 define <WIDTH x $1> @__rotate_$1(<WIDTH x $1>, i32) nounwind readnone alwaysinline {