Check for MMXEXT and SSE at build-time if possible

include/vlc_cpu.h

@@ -32,8 +32,8 @@ VLC_API unsigned vlc_CPU(void);
 #  define HAVE_FPU 1
 #  define VLC_CPU_MMX    8
 #  define CPU_CAPABILITY_3DNOW   (1<<4)
-#  define CPU_CAPABILITY_MMXEXT  (1<<5)
-#  define CPU_CAPABILITY_SSE     (1<<6)
+#  define VLC_CPU_MMXEXT 32
+#  define VLC_CPU_SSE    64
 #  define CPU_CAPABILITY_SSE2    (1<<7)
 #  define CPU_CAPABILITY_SSE3    (1<<8)
 #  define CPU_CAPABILITY_SSSE3   (1<<9)
@@ -54,12 +54,18 @@ VLC_API unsigned vlc_CPU(void);
 # endif
 
 # if defined (__SSE__)
+#  define vlc_CPU_MMXEXT() (1)
+#  define vlc_CPU_SSE() (1)
 #  define VLC_SSE
-# elif VLC_GCC_VERSION(4, 4)
+# else
+#  define vlc_CPU_MMXEXT() ((vlc_CPU() & VLC_CPU_MMXEXT) != 0)
+#  define vlc_CPU_SSE() ((vlc_CPU() & VLC_CPU_SSE) != 0)
+#  if VLC_GCC_VERSION(4, 4)
 #   define VLC_SSE __attribute__ ((__target__ ("sse")))
 #  else
 #   define VLC_SSE VLC_SSE_is_not_implemented_on_this_compiler
 #  endif
+# endif
 
 # elif defined (__ppc__) || defined (__ppc64__) || defined (__powerpc__)
 #  define HAVE_FPU 1

modules/codec/avcodec/avcodec.c

@@ -334,11 +334,11 @@ static int OpenDecoder( vlc_object_t *p_this )
     unsigned i_cpu = vlc_CPU();
     if( !vlc_CPU_MMX() )
         p_context->dsp_mask |= AV_CPU_FLAG_MMX;
-    if( !(i_cpu & CPU_CAPABILITY_MMXEXT) )
+    if( !vlc_CPU_MMXEXT() )
         p_context->dsp_mask |= AV_CPU_FLAG_MMX2;
     if( !(i_cpu & CPU_CAPABILITY_3DNOW) )
         p_context->dsp_mask |= AV_CPU_FLAG_3DNOW;
-    if( !(i_cpu & CPU_CAPABILITY_SSE) )
+    if( !vlc_CPU_SSE() )
         p_context->dsp_mask |= AV_CPU_FLAG_SSE;
     if( !(i_cpu & CPU_CAPABILITY_SSE2) )
         p_context->dsp_mask |= AV_CPU_FLAG_SSE2;

modules/codec/avcodec/encoder.c

@@ -328,11 +328,11 @@ int OpenEncoder( vlc_object_t *p_this )
     unsigned i_cpu = vlc_CPU();
     if( !vlc_CPU_MMX() )
         p_context->dsp_mask |= AV_CPU_FLAG_MMX;
-    if( !(i_cpu & CPU_CAPABILITY_MMXEXT) )
+    if( !vlc_CPU_MMXEXT() )
         p_context->dsp_mask |= AV_CPU_FLAG_MMX2;
     if( !(i_cpu & CPU_CAPABILITY_3DNOW) )
         p_context->dsp_mask |= AV_CPU_FLAG_3DNOW;
-    if( !(i_cpu & CPU_CAPABILITY_SSE) )
+    if( !vlc_CPU_SSE() )
         p_context->dsp_mask |= AV_CPU_FLAG_SSE;
     if( !(i_cpu & CPU_CAPABILITY_SSE2) )
         p_context->dsp_mask |= AV_CPU_FLAG_SSE2;

modules/codec/libmpeg2.c

@@ -197,7 +197,7 @@ static int OpenDecoder( vlc_object_t *p_this )
         i_accel |= MPEG2_ACCEL_X86_MMX;
     if( cpu & CPU_CAPABILITY_3DNOW )
         i_accel |= MPEG2_ACCEL_X86_3DNOW;
-    if( cpu & CPU_CAPABILITY_MMXEXT )
+    if( vlc_CPU_MMXEXT() )
         i_accel |= MPEG2_ACCEL_X86_MMXEXT;
 #elif defined( __powerpc__ ) || defined( __ppc__ ) || defined( __ppc64__ )
     if( vlc_CPU_ALTIVEC() )

modules/codec/x264.c

@@ -1260,14 +1260,13 @@ static int  Open ( vlc_object_t *p_this )
     free( psz_val );
 
 #if defined (__i386__) || defined (__x86_64__)
-    unsigned i_cpu = vlc_CPU();
     if( !vlc_CPU_MMX() )
         p_sys->param.cpu &= ~X264_CPU_MMX;
-    if( !(i_cpu & CPU_CAPABILITY_MMXEXT) )
+    if( !vlc_CPU_MMXEXT() )
         p_sys->param.cpu &= ~X264_CPU_MMXEXT;
-    if( !(i_cpu & CPU_CAPABILITY_SSE) )
+    if( !vlc_CPU_SSE() )
         p_sys->param.cpu &= ~X264_CPU_SSE;
-    if( !(i_cpu & CPU_CAPABILITY_SSE2) )
+    if( !(vlc_CPU() & CPU_CAPABILITY_SSE2) )
         p_sys->param.cpu &= ~X264_CPU_SSE2;
 #endif
 

modules/stream_out/switcher.c

@@ -383,11 +383,11 @@ static sout_stream_id_t *Add( sout_stream_t *p_stream, es_format_t *p_fmt )
         unsigned i_cpu = vlc_CPU();
         if( !vlc_CPU_MMX() )
             id->ff_enc_c->dsp_mask |= AV_CPU_FLAG_MMX;
-        if( !(i_cpu & CPU_CAPABILITY_MMXEXT) )
+        if( !vlc_CPU_MMXEXT() )
             id->ff_enc_c->dsp_mask |= AV_CPU_FLAG_MMX2;
         if( !(i_cpu & CPU_CAPABILITY_3DNOW) )
             id->ff_enc_c->dsp_mask |= AV_CPU_FLAG_3DNOW;
-        if( !(i_cpu & CPU_CAPABILITY_SSE) )
+        if( !vlc_CPU_SSE() )
             id->ff_enc_c->dsp_mask |= AV_CPU_FLAG_SSE;
         if( !(i_cpu & CPU_CAPABILITY_SSE2) )
             id->ff_enc_c->dsp_mask |= AV_CPU_FLAG_SSE2;
@@ -804,11 +804,11 @@ static mtime_t VideoCommand( sout_stream_t *p_stream, sout_stream_id_t *id )
         unsigned i_cpu = vlc_CPU();
         if( !vlc_CPU_MMX() )
             id->ff_enc_c->dsp_mask |= AV_CPU_FLAG_MMX;
-        if( !(i_cpu & CPU_CAPABILITY_MMXEXT) )
+        if( !vlc_CPU_MMXEXT() )
             id->ff_enc_c->dsp_mask |= AV_CPU_FLAG_MMX2;
         if( !(i_cpu & CPU_CAPABILITY_3DNOW) )
             id->ff_enc_c->dsp_mask |= AV_CPU_FLAG_3DNOW;
-        if( !(i_cpu & CPU_CAPABILITY_SSE) )
+        if( !vlc_CPU_SSE() )
             id->ff_enc_c->dsp_mask |= AV_CPU_FLAG_SSE;
         if( !(i_cpu & CPU_CAPABILITY_SSE2) )
             id->ff_enc_c->dsp_mask |= AV_CPU_FLAG_SSE2;

modules/video_filter/deinterlace/algo_phosphor.c

@@ -79,7 +79,9 @@ static void DarkenField( picture_t *p_dst,
     /* Bitwise ANDing with this clears the i_strength highest bits
        of each byte */
 #ifdef CAN_COMPILE_MMXEXT
-    unsigned u_cpu = vlc_CPU();
+# ifndef __SSE__
+    const unsigned u_cpu = vlc_CPU();
+# endif
     uint64_t i_strength_u64 = i_strength; /* for MMX version (needs to know
                                              number of bits) */
 #endif
@@ -113,7 +115,9 @@ static void DarkenField( picture_t *p_dst,
         int x = 0;
 
 #ifdef CAN_COMPILE_MMXEXT
-        if( u_cpu & CPU_CAPABILITY_MMXEXT )
+# ifndef __SSE__
+        if( u_cpu & VLC_CPU_MMXEXT )
+# endif
         {
             movq_m2r( i_strength_u64,  mm1 );
             movq_m2r( remove_high_u64, mm2 );
@@ -128,11 +132,11 @@ static void DarkenField( picture_t *p_dst,
             }
         }
         else
-        {
 #endif
+#if !defined (CAN_COMPILE_MMXEXT) || !defined (__SSE__)
+        {
             for( ; x < w8; x += 8, ++po )
                 (*po) = ( ((*po) >> i_strength) & remove_high_u64 );
-#ifdef CAN_COMPILE_MMXEXT
         }
 #endif
 
@@ -174,7 +178,9 @@ static void DarkenField( picture_t *p_dst,
 
 #ifdef CAN_COMPILE_MMXEXT
                 /* See also easy-to-read C version below. */
-                if( u_cpu & CPU_CAPABILITY_MMXEXT )
+# ifndef __SSE__
+                if( u_cpu & VLC_CPU_MMXEXT )
+# endif
                 {
                     static const mmx_t b128 = { .uq = 0x8080808080808080ULL };
                     movq_m2r( b128, mm5 );
@@ -216,7 +222,9 @@ static void DarkenField( picture_t *p_dst,
     } /* if process_chroma */
 
 #ifdef CAN_COMPILE_MMXEXT
-    if( u_cpu & CPU_CAPABILITY_MMXEXT )
+# ifndef __SSE__
+    if( u_cpu & VLC_CPU_MMXEXT )
+# endif
         emms();
 #endif
 }

modules/video_filter/deinterlace/algo_x.c

@@ -535,8 +535,8 @@ static inline void XDeintBand8x8MMXEXT( uint8_t *dst, int i_dst,
 void RenderX( picture_t *p_outpic, picture_t *p_pic )
 {
     int i_plane;
-#ifdef CAN_COMPILE_MMXEXT
-    unsigned u_cpu = vlc_CPU();
+#if defined (CAN_COMPILE_MMXEXT) && !defined(__SSE__)
+    const unsigned u_cpu = vlc_CPU();
 #endif
 
     /* Copy image and skip lines */
@@ -559,11 +559,15 @@ void RenderX( picture_t *p_outpic, picture_t *p_pic )
             uint8_t *src = &p_pic->p[i_plane].p_pixels[8*y*i_src];
 
 #ifdef CAN_COMPILE_MMXEXT
-            if( u_cpu & CPU_CAPABILITY_MMXEXT )
+# ifndef __SSE__
+            if( u_cpu & VLC_CPU_MMXEXT )
+# endif
                 XDeintBand8x8MMXEXT( dst, i_dst, src, i_src, i_mbx, i_modx );
             else
 #endif
+#ifndef __SSE__
                 XDeintBand8x8C( dst, i_dst, src, i_src, i_mbx, i_modx );
+#endif
         }
 
         /* Last line (C only)*/
@@ -586,7 +590,9 @@ void RenderX( picture_t *p_outpic, picture_t *p_pic )
     }
 
 #ifdef CAN_COMPILE_MMXEXT
-    if( u_cpu & CPU_CAPABILITY_MMXEXT )
+# ifndef __SSE__
+    if( u_cpu & VLC_CPU_MMXEXT )
+# endif
         emms();
 #endif
 }

modules/video_filter/deinterlace/deinterlace.c

@@ -641,7 +641,7 @@ int Open( vlc_object_t *p_this )
     else
 #endif
 #if defined(CAN_COMPILE_MMXEXT)
-    if( chroma->pixel_size == 1 && (vlc_CPU() & CPU_CAPABILITY_MMXEXT) )
+    if( chroma->pixel_size == 1 && vlc_CPU_MMXEXT() )
     {
         p_sys->pf_merge = MergeMMXEXT;
         p_sys->pf_end_merge = EndMMX;

modules/video_filter/deinterlace/helpers.c

@@ -113,7 +113,7 @@ static void FieldFromPlane( plane_t *p_dst, const plane_t *p_src, int i_field )
  * @param[in] p_pix_c Base pointer to the same block in current picture
  * @param i_pitch_prev i_pitch of previous picture
  * @param i_pitch_curr i_pitch of current picture
- * @param b_mmx (vlc_CPU() & CPU_CAPABILITY_MMXEXT) or false.
+ * @param b_mmx (vlc_CPU() & VLC_CPU_MMXEXT) or false.
  * @param[out] pi_top 1 if top field of the block had motion, 0 if no
  * @param[out] pi_bot 1 if bottom field of the block had motion, 0 if no
  * @return 1 if the block had motion, 0 if no
@@ -388,9 +388,9 @@ int EstimateNumBlocksWithMotion( const picture_t* p_prev,
 
     /* We must tell our inline helper whether to use MMX acceleration. */
 #ifdef CAN_COMPILE_MMXEXT
-    bool b_mmx = ( vlc_CPU() & CPU_CAPABILITY_MMXEXT );
+    const bool b_mmx = vlc_CPU_MMXEXT();
 #else
-    bool b_mmx = false;
+    const bool b_mmx = false;
 #endif
 
     int i_score = 0;
@@ -466,9 +466,11 @@ int CalculateInterlaceScore( const picture_t* p_pic_top,
     int32_t i_score_c   = 0; /* this counts as-is (used for non-MMX parts) */
 
 #ifdef CAN_COMPILE_MMXEXT
-    unsigned u_cpu = vlc_CPU();
+# ifndef __SSE__
+    const unsigned u_cpu = vlc_CPU();
 
-    if( u_cpu & CPU_CAPABILITY_MMXEXT )
+    if( u_cpu & VLC_CPU_MMXEXT )
+# endif
         pxor_r2r( mm7, mm7 ); /* we will keep score in mm7 */
 #endif
 
@@ -512,7 +514,9 @@ int CalculateInterlaceScore( const picture_t* p_pic_top,
                             # of pixels < (2^32)/255
                Note: calculates score * 255
             */
-            if( u_cpu & CPU_CAPABILITY_MMXEXT )
+# ifndef __SSE__
+            if( u_cpu & VLC_CPU_MMXEXT )
+# endif
             {
                 static const mmx_t b0   = { .uq = 0x0000000000000000ULL };
                 static const mmx_t b128 = { .uq = 0x8080808080808080ULL };
@@ -591,7 +595,9 @@ int CalculateInterlaceScore( const picture_t* p_pic_top,
     }
 
 #ifdef CAN_COMPILE_MMXEXT
-    if( u_cpu & CPU_CAPABILITY_MMXEXT )
+# ifndef __SSE__
+    if( u_cpu & VLC_CPU_MMXEXT )
+# endif
     {
         movd_r2m( mm7, i_score_mmx );
         emms();

modules/video_filter/gradfun.c

@@ -133,21 +133,24 @@ static int Open(vlc_object_t *object)
     cfg->thresh      = 0.0;
     cfg->radius      = 0;
     cfg->buf         = NULL;
-    cfg->filter_line = filter_line_c;
-    cfg->blur_line   = blur_line_c;
 
 #if HAVE_SSE2 && HAVE_6REGS
     if (vlc_CPU() & CPU_CAPABILITY_SSE2)
         cfg->blur_line = blur_line_sse2;
+    else
 #endif
-#if HAVE_MMX2
-    if (vlc_CPU() & CPU_CAPABILITY_MMXEXT)
-        cfg->filter_line = filter_line_mmx2;
-#endif
+        cfg->blur_line   = blur_line_c;
 #if HAVE_SSSE3
     if (vlc_CPU() & CPU_CAPABILITY_SSSE3)
         cfg->filter_line = filter_line_ssse3;
+    else
+#endif
+#if HAVE_MMX2
+    if (vlc_CPU_MMXEXT())
+        cfg->filter_line = filter_line_mmx2;
+    else
 #endif
+        cfg->filter_line = filter_line_c;
 
     filter->p_sys           = sys;
     filter->pf_video_filter = Filter;

modules/video_filter/postproc.c

@@ -134,12 +134,11 @@ static int OpenPostproc( vlc_object_t *p_this )
 
     /* Set CPU capabilities */
 #if defined(__i386__) || defined(__x86_64__)
-    unsigned i_cpu = vlc_CPU();
     if( vlc_CPU_MMX() )
         i_flags |= PP_CPU_CAPS_MMX;
-    if( i_cpu & CPU_CAPABILITY_MMXEXT )
+    if( vlc_CPU_MMXEXT() )
         i_flags |= PP_CPU_CAPS_MMX2;
-    if( i_cpu & CPU_CAPABILITY_3DNOW )
+    if( vlc_CPU() & CPU_CAPABILITY_3DNOW )
         i_flags |= PP_CPU_CAPS_3DNOW;
 #elif defined(__ppc__) || defined(__ppc64__) || defined(__powerpc__)
     if( vlc_CPU_ALTIVEC() )

modules/video_filter/swscale.c

@@ -232,14 +232,13 @@ static int GetSwsCpuMask(void)
     int i_sws_cpu = 0;
 
 #if defined(__i386__) || defined(__x86_64__)
-    const unsigned int i_cpu = vlc_CPU();
     if( vlc_CPU_MMX() )
         i_sws_cpu |= SWS_CPU_CAPS_MMX;
 #if (LIBSWSCALE_VERSION_INT >= ((0<<16)+(5<<8)+0))
-    if( i_cpu & CPU_CAPABILITY_MMXEXT )
+    if( vlc_CPU_MMXEXT() )
         i_sws_cpu |= SWS_CPU_CAPS_MMX2;
 #endif
-    if( i_cpu & CPU_CAPABILITY_3DNOW )
+    if( vlc_CPU() & CPU_CAPABILITY_3DNOW )
         i_sws_cpu |= SWS_CPU_CAPS_3DNOW;
 #elif defined(__ppc__) || defined(__ppc64__) || defined(__powerpc__)
     if( vlc_CPU_ALTIVEC() )

src/misc/cpu.c

@@ -223,19 +223,14 @@ void vlc_CPU_init (void)
 # endif
     i_capabilities |= VLC_CPU_MMX;
 
-# if defined (__SSE__)
-    i_capabilities |= CPU_CAPABILITY_MMXEXT | CPU_CAPABILITY_SSE;
-# else
     if( i_edx & 0x02000000 )
     {
-        i_capabilities |= CPU_CAPABILITY_MMXEXT;
-
+        i_capabilities |= VLC_CPU_MMXEXT;
 #   ifdef CAN_COMPILE_SSE
         if (vlc_CPU_check ("SSE", SSE_test))
-            i_capabilities |= CPU_CAPABILITY_SSE;
+            i_capabilities |= VLC_CPU_SSE;
 #   endif
     }
-# endif
 
 # if defined (__SSE2__)
     i_capabilities |= CPU_CAPABILITY_SSE2;
@@ -289,9 +284,7 @@ void vlc_CPU_init (void)
 # endif
 
     if( b_amd && ( i_edx & 0x00400000 ) )
-    {
-        i_capabilities |= CPU_CAPABILITY_MMXEXT;
-    }
+        i_capabilities |= VLC_CPU_MMXEXT;
 out:
 
 #elif defined( __powerpc__ ) || defined( __ppc__ ) || defined( __powerpc64__ ) \
@@ -353,15 +346,15 @@ void vlc_CPU_dump (vlc_object_t *obj)
         p += sprintf (p, "%s ", (string) )
 
     if (vlc_CPU_MMX()) p += sprintf (p, "MMX ");
-    PRINT_CAPABILITY(CPU_CAPABILITY_3DNOW, "3DNow!");
-    PRINT_CAPABILITY(CPU_CAPABILITY_MMXEXT, "MMXEXT");
-    PRINT_CAPABILITY(CPU_CAPABILITY_SSE, "SSE");
+    if (vlc_CPU_MMXEXT()) p += sprintf (p, "MMXEXT ");
+    if (vlc_CPU_SSE()) p += sprintf (p, "SSE ");;
     PRINT_CAPABILITY(CPU_CAPABILITY_SSE2, "SSE2");
     PRINT_CAPABILITY(CPU_CAPABILITY_SSE3, "SSE3");
     PRINT_CAPABILITY(CPU_CAPABILITY_SSSE3, "SSSE3");
     PRINT_CAPABILITY(CPU_CAPABILITY_SSE4_1, "SSE4.1");
     PRINT_CAPABILITY(CPU_CAPABILITY_SSE4_2, "SSE4.2");
     PRINT_CAPABILITY(CPU_CAPABILITY_SSE4A,  "SSE4A");
+    PRINT_CAPABILITY(CPU_CAPABILITY_3DNOW, "3DNow!");
 
 #elif defined (__powerpc__) || defined (__ppc__) || defined (__ppc64__)
     if (vlc_CPU_ALTIVEC())  p += sprintf (p, "AltiVec");

src/posix/linux_cpu.c

@@ -69,12 +69,10 @@ static void vlc_CPU_init (void)
 #elif defined (__i386__) || defined (__x86_64__)
             if (!strcmp (cap, "mmx"))
                 core_caps |= VLC_CPU_MMX;
-# ifndef __SSE__
             if (!strcmp (cap, "sse"))
-                core_caps |= CPU_CAPABILITY_SSE | CPU_CAPABILITY_MMXEXT;
+                core_caps |= VLC_CPU_SSE | VLC_CPU_MMXEXT;
             if (!strcmp (cap, "mmxext"))
-                core_caps |= CPU_CAPABILITY_MMXEXT;
-# endif
+                core_caps |= VLC_CPU_MMXEXT;
 # ifndef __SSE2__
             if (!strcmp (cap, "sse2"))
                 core_caps |= CPU_CAPABILITY_SSE2;
@@ -119,9 +117,6 @@ static void vlc_CPU_init (void)
 
     /* Always enable capabilities that were forced during compilation */
 #if defined (__i386__) || defined (__x86_64__)
-# ifdef __SSE__
-    all_caps |= CPU_CAPABILITY_SSE | CPU_CAPABILITY_MMXEXT;
-# endif
 # ifdef __SSE2__
     all_caps |= CPU_CAPABILITY_SSE2;
 # endif