From: José Fonseca <jfonseca@vmware.com>
Date: Wed, 11 Aug 2010 17:44:17 +0000 (+0100)
Subject: gallivm: Fix and enable the extra Newton/Raphson step in lp_build_rcp().
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=b481a1237e00e1e1fb68ffca0653df3a96f21788;p=mesa.git

gallivm: Fix and enable the extra Newton/Raphson step in lp_build_rcp().

Thanks to Michal for spotting this.
---

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.c b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
index cf2feeb1633..816ee701194 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_arit.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
@@ -1274,7 +1274,7 @@ lp_build_rcp(struct lp_build_context *bld,
        * when we have a better system in place to track minimum precision.
        */
 
-#if 0
+#if 1
       /*
        * Do one Newton-Raphson step to improve precision:
        *
@@ -1291,7 +1291,7 @@ lp_build_rcp(struct lp_build_context *bld,
       res = LLVMBuildFSub(bld->builder, two, res, "");
       res = LLVMBuildFMul(bld->builder, res, rcp_a, "");
 
-      return rcp_a;
+      return res;
 #else
       return lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rcp.ps", lp_build_vec_type(type), a);
 #endif