From fab76b1c4fa21f16d009d756225d6876df260c40 Mon Sep 17 00:00:00 2001
From: Jonathan Moore Liles <j.liles@unix.net>
Date: Fri, 9 Aug 2013 17:41:11 -0700
Subject: [PATCH] Set FTZ and DAZ flags by default for all JACK client threads.

---
 nonlib/JACK/Client.C |  9 +++++-
 wscript              | 77 ++++++++++++++++++++++++--------------------
 2 files changed, 50 insertions(+), 36 deletions(-)

diff --git a/nonlib/JACK/Client.C b/nonlib/JACK/Client.C
index 380d1c9..489f439 100644
--- a/nonlib/JACK/Client.C
+++ b/nonlib/JACK/Client.C
@@ -26,7 +26,9 @@
 
 #include "debug.h"
 
-
+#ifdef __SSE2_MATH__
+#include <xmmintrin.h>
+#endif
 
 namespace JACK
 {
@@ -113,6 +115,11 @@ namespace JACK
     void
     Client::thread_init ( void *arg )
     {
+#if __SSE2_MATH__
+    /* set FTZ and DAZ flags */
+    _mm_setcsr(_mm_getcsr() | 0x8040);
+#endif
+
         ((Client*)arg)->thread_init();
     }
 
diff --git a/wscript b/wscript
index 4535b5f..8ec772b 100644
--- a/wscript
+++ b/wscript
@@ -19,6 +19,8 @@ def options(opt):
     #                help="Link to system-installed shared NTK instead of bundled version")    
     opt.add_option('--enable-debug', action='store_true', default=False, dest='debug',
                     help='Build for debugging')
+    opt.add_option('--disable-sse', action='store_false', default=True, dest='sse',
+                    help='Disable SSE optimization')
     opt.add_option('--project', action='store', default=False, dest='project',
                     help='Limit build to a single project (' + ', '.join( projects ) + ')')
 
@@ -33,24 +35,59 @@ def configure(conf):
     conf.load('gccdeps')
     conf.line_just = 52
 
+    conf.env['LIB_PTHREAD'] = ['pthread']
+    conf.env['LIB_DL'] = ['dl']
+    conf.env['LIB_M'] = ['m']
+
+    # NTK_EXTRA_FLAGS=''
+    # if not Options.options.use_system_ntk:
+    #     print 'Using bundled NTK'
+    #     os.environ['PKG_CONFIG_PATH'] = 'lib/ntk/build/:' + os.environ.get('PKG_CONFIG_PATH','')
+    #    NTK_EXTRA_FLAGS='--static'
+    #     PWD = os.environ.get('PWD','')
+    #     os.environ['PATH'] = PWD + '/lib/ntk/build/fluid:' + os.environ.get('PATH','')
+
+    conf.check_cfg(package='ntk', uselib_store='NTK', args='--cflags --libs',
+                   atleast_version='1.3.0', mandatory=True)
+    conf.check_cfg(package='ntk_images', uselib_store='NTK_IMAGES', args=' --cflags --libs',
+                   atleast_version='1.3.0', mandatory=True)
+
+    conf.find_program('ntk-fluid', var='NTK_FLUID')
+
+    conf.check_cfg(package='jack', uselib_store='JACK', args="--cflags --libs",
+                   atleast_version='0.103.0', mandatory=True)
+
+    conf.check_cfg(package='liblo', uselib_store='LIBLO',args="--cflags --libs",
+                      atleast_version='0.26', mandatory=True)
+
+###
+
+    for i in common:
+        conf.recurse(i)
+
     optimization_flags = [
             "-O3",
             "-fomit-frame-pointer",
             "-ffast-math",
-            "-msse",
-            "-mfpmath=sse",
-            "-ftree-vectorize",
-#            "-fstrength-reduce",
             "-pipe"
             ]
+    
+    if Options.options.sse:
+        print('Using SSE optimization')
+        optimization_flags.extend( [
+            "-msse2",
+            "-mfpmath=sse",
+            "-ftree-vectorize" ] )
+        conf.define( 'USE_SSE', 1 )
 
     debug_flags = [ '-O0', '-g3' ]
 
-
     if Options.options.debug:
+        print('Building for debugging')
         conf.env.append_value('CFLAGS', debug_flags )
         conf.env.append_value('CXXFLAGS', debug_flags )
     else:
+        print('Building for performance')
         conf.env.append_value('CFLAGS', optimization_flags )
         conf.env.append_value('CXXFLAGS', optimization_flags )
         conf.define( 'NDEBUG', 1 )
@@ -70,36 +107,6 @@ def configure(conf):
     conf.env.append_value('CFLAGS', global_flags )
     conf.env.append_value('CXXFLAGS', global_flags )
 
-    conf.env['LIB_PTHREAD'] = ['pthread']
-    conf.env['LIB_DL'] = ['dl']
-    conf.env['LIB_M'] = ['m']
-
-    # NTK_EXTRA_FLAGS=''
-    # if not Options.options.use_system_ntk:
-    #     print 'Using bundled NTK'
-    #     os.environ['PKG_CONFIG_PATH'] = 'lib/ntk/build/:' + os.environ.get('PKG_CONFIG_PATH','')
-    #    NTK_EXTRA_FLAGS='--static'
-    #     PWD = os.environ.get('PWD','')
-    #     os.environ['PATH'] = PWD + '/lib/ntk/build/fluid:' + os.environ.get('PATH','')
-
-    conf.check_cfg(package='ntk', uselib_store='NTK', args='--cflags --libs',
-                   atleast_version='1.3.0', mandatory=True)
-    conf.check_cfg(package='ntk_images', uselib_store='NTK_IMAGES', args=' --cflags --libs',
-                   atleast_version='1.3.0', mandatory=True)
-
-    conf.find_program('ntk-fluid', var='NTK_FLUID')
-
-    conf.check_cfg(package='jack', uselib_store='JACK', args="--cflags --libs",
-                   atleast_version='0.103.0', mandatory=True)
-
-    conf.check_cfg(package='liblo', uselib_store='LIBLO',args="--cflags --libs",
-                      atleast_version='0.26', mandatory=True)
-
-
-###
-
-    for i in common:
-        conf.recurse(i)
 
     conf.env.PROJECT = conf.options.project