From 59fd3b651e0a3ab8459ef784463f3cd8a3815f2a Mon Sep 17 00:00:00 2001
From: "K. Noel Belcourt" <kbelco@sandia.gov>
Date: Fri, 16 May 2014 23:06:32 -0600
Subject: [PATCH] Mods to get Cray building cleanly, target native hardware on
 compute nodes.

---
 src/tools/cray.jam | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/src/tools/cray.jam b/src/tools/cray.jam
index a7d3b51c3..1fa1ddc01 100644
--- a/src/tools/cray.jam
+++ b/src/tools/cray.jam
@@ -50,8 +50,10 @@ generators.register-c-compiler cray.compile.c : C : OBJ : <toolset>cray ;
 
 # unlike most compliers, Cray defaults to static linking.
 # flags cxx LINKFLAGS <runtime-link>static : -bstatic ;
-flags cray.compile OPTIONS <debug-symbols>on : -G0 ;
-flags cray.link OPTIONS <debug-symbols>on : -G0 ;
+# flags cray.compile OPTIONS <debug-symbols>on : -G0 ;
+flags cray.compile OPTIONS <debug-symbols>on : -g ;
+# flags cray.link OPTIONS <debug-symbols>on : -G0 ;
+flags cray.link OPTIONS <debug-symbols>on : -g ;
 
 flags cray.compile OPTIONS <optimization>off : -O0 ;
 flags cray.compile OPTIONS <optimization>speed : -O2 ;
@@ -67,11 +69,14 @@ flags cray.compile DEFINES <define> ;
 flags cray.compile INCLUDES <include> ;
 flags cray.link OPTIONS <linkflags> ;
 
-flags cray.compile OPTIONS : -hgnu -fPIC -h system_alloc -h tolerant -h ipa0 ;
+# flags cray.compile OPTIONS : -hgnu -fPIC -h system_alloc -h tolerant -h ipa0 ;
+flags cray.compile OPTIONS : -march=bdver1 -mfpmath=sse -mfma4 -mavx -funroll-all-loops -mprefer-avx128 -fprefetch-loop-arrays --param prefetch-latency=300 -minline-all-stringops -ffast-math -fno-finite-math-only ;
 flags cray.compile OPTIONS <link>shared : -dynamic ;
 flags cray.compile OPTIONS <link>static : -static ;
-flags cray.link OPTIONS <link>static : -static ;
-flags cray.link OPTIONS <link>shared : ;
+# flags cray.link OPTIONS <link>static : -static ;
+# flags cray.link OPTIONS <link>shared ;
+flags cray.link OPTIONS <link>static : -static -march=bdver1 -mfpmath=sse -mfma4 -mavx -funroll-all-loops -mprefer-avx128 -fprefetch-loop-arrays --param prefetch-latency=300 -minline-all-stringops -ffast-math -fno-finite-math-only -Xlinker --allow-multiple-definition ;
+flags cray.link OPTIONS <link>shared : -march=bdver1 -mfpmath=sse -mfma4 -mavx -funroll-all-loops -mprefer-avx128 -fprefetch-loop-arrays --param prefetch-latency=300 -minline-all-stringops -ffast-math -fno-finite-math-only -Xlinker --allow-multiple-definition ;
 flags cray.link LOPTIONS <link>shared : -dynamic ;
 
 flags cray.link LIBPATH <library-path> ;