build: Add scaffolding for SSE3

It's currently unused, but is useful to have it around when comparing
different vectorization strategies.

https://bugzilla.gnome.org/show_bug.cgi?id=795686

diff --git a/configure.ac b/configure.ac
index 1dfdccdebbcd0a9f7ab6c64a69f0652fed5943ae..8aa37c5e1f2e6a4293208b0ae21759b9c3e89eed 100644 (file)
--- a/configure.ac
+++ b/configure.ac
@@ -324,6 +324,10 @@ AC_ARG_ENABLE(sse2,
   [  --enable-sse2            enable SSE2 support (default=auto)],,
   enable_sse2=$enable_sse)
 
+AC_ARG_ENABLE(sse3,
+  [  --enable-sse3            enable SSE3 support (default=auto)],,
+  enable_sse3=$enable_sse2)
+
 AC_ARG_ENABLE(sse4_1,
   [  --enable-sse4_1            enable SSE4_1 support (default=auto)],,
   enable_sse4_1=$enable_sse)
@@ -388,22 +392,40 @@ if test "x$enable_mmx" = xyes; then
           AC_MSG_WARN([The assembler does not support the SSE2 command set.])
         )
 
-        if test "x$enable_sse4_1" = xyes; then
-          BABL_DETECT_CFLAGS(sse4_1_flag, '-msse4.1')
-          SSE4_1_EXTRA_CFLAGS="$SSE_EXTRA_CFLAGS $sse4_1_flag"
+        if test "x$enable_sse3" = xyes; then
+          BABL_DETECT_CFLAGS(sse3_flag, '-msse3')
+          SSE3_EXTRA_CFLAGS="$SSE2_EXTRA_CFLAGS $sse3_flag"
 
-          AC_MSG_CHECKING(whether we can compile SSE4_1 code)
+          AC_MSG_CHECKING(whether we can compile SSE3 code)
 
-          CFLAGS="$CFLAGS $sse4_1_flag"
+          CFLAGS="$CFLAGS $sse3_flag"
 
-          AC_COMPILE_IFELSE([AC_LANG_PROGRAM(,[asm ("pmovzxbd %xmm0,%xmm1");])],
-            AC_DEFINE(USE_SSE4_1, 1, [Define to 1 if SSE4_1 assembly is available.])
+          AC_COMPILE_IFELSE([AC_LANG_PROGRAM(,[asm ("addsubpd %xmm0,%xmm1");])],
+            AC_DEFINE(USE_SSE3, 1, [Define to 1 if SSE3 assembly is available.])
             AC_MSG_RESULT(yes)
           ,
-            enable_sse4_1=no
+            enable_sse3=no
             AC_MSG_RESULT(no)
-            AC_MSG_WARN([The assembler does not support the SSE4_1 command set.])
+            AC_MSG_WARN([The assembler does not support the SSE3 command set.])
           )
+
+          if test "x$enable_sse4_1" = xyes; then
+            BABL_DETECT_CFLAGS(sse4_1_flag, '-msse4.1')
+            SSE4_1_EXTRA_CFLAGS="$SSE_EXTRA_CFLAGS $sse4_1_flag"
+
+            AC_MSG_CHECKING(whether we can compile SSE4_1 code)
+
+            CFLAGS="$CFLAGS $sse4_1_flag"
+
+            AC_COMPILE_IFELSE([AC_LANG_PROGRAM(,[asm ("pmovzxbd %xmm0,%xmm1");])],
+              AC_DEFINE(USE_SSE4_1, 1, [Define to 1 if SSE4_1 assembly is available.])
+              AC_MSG_RESULT(yes)
+            ,
+              enable_sse4_1=no
+              AC_MSG_RESULT(no)
+              AC_MSG_WARN([The assembler does not support the SSE4_1 command set.])
+            )
+          fi
         fi
       fi
 
@@ -439,6 +461,7 @@ if test "x$enable_mmx" = xyes; then
   AC_SUBST(MMX_EXTRA_CFLAGS)
   AC_SUBST(SSE_EXTRA_CFLAGS)
   AC_SUBST(SSE2_EXTRA_CFLAGS)
+  AC_SUBST(SSE3_EXTRA_CFLAGS)
   AC_SUBST(SSE4_1_EXTRA_CFLAGS)
   AC_SUBST(F16C_EXTRA_CFLAGS)
 fi

diff --git a/meson.build b/meson.build
index 5795c9645bb5a9b99b583b466bf27f600e0e1bed..41880193459c030c5819e08804001da0a8e0a5ce 100644 (file)
--- a/meson.build
+++ b/meson.build
@@ -163,6 +163,12 @@ if has_sse2
     language: 'c')
 endif
 
+has_sse3 = cc.has_argument('-msse3') and get_option('enable-sse3')
+if has_sse3
+  add_project_arguments(   '-msse3',
+    language: 'c')
+endif
+
 has_sse41= cc.has_argument('-msse4.1') and get_option('enable-sse4_1')
 if has_sse41
   add_project_arguments(   '-msse4.1',

diff --git a/meson_options.txt b/meson_options.txt
index 51e5fdc711579d4aaf1df9c19d9677339e28da89..febd3067988bbce4bc390f6dbc617dce6ce941de 100644 (file)
--- a/meson_options.txt
+++ b/meson_options.txt
@@ -1,6 +1,7 @@
 option('enable-mmx',    type: 'boolean', value: true, description: 'enable MMX support')
 option('enable-sse',    type: 'boolean', value: true, description: 'enable SSE support')
 option('enable-sse2',   type: 'boolean', value: true, description: 'enable SSE2 support')
+option('enable-sse3',   type: 'boolean', value: true, description: 'enable SSE3 support')
 option('enable-sse4_1', type: 'boolean', value: true, description: 'enable SSE4.1 support')
 option('enable-f16c',   type: 'boolean', value: true, description: 'enable hardware half-float support')
 option('with-docs',     type: 'boolean', value: true)