Switch to a new CPU flag infrastructure

This opens a possibility to selectively disable
particular implementations at runtime.

5 files changed, 74 insertions, 61 deletions

diff --git a/libass/ass_bitmap_engine.c b/libass/ass_bitmap_engine.c
index 54bb5c7..ef3f6b3 100644
--- a/libass/ass_bitmap_engine.c
+++ b/libass/ass_bitmap_engine.c
@@ -19,7 +19,11 @@
 #include "config.h"
 #include "ass_compat.h"
 
+#include <stdbool.h>
+
 #include "ass_bitmap_engine.h"
+#include "x86/cpuid.h"
+
 
 #define RASTERIZER_PROTOTYPES(tile_size, suffix) \
     FillSolidTileFunc     ass_fill_solid_tile     ## tile_size ## _ ## suffix; \
@@ -99,3 +103,60 @@ BITMAP_ENGINE(5, 4, 16, avx2)
 #endif
 
 #endif
+
+
+unsigned ass_get_cpu_flags(unsigned mask)
+{
+    unsigned flags = ASS_CPU_FLAG_NONE;
+
+#if CONFIG_ASM && ARCH_X86
+
+    if (!ass_has_cpuid())
+        return flags & mask;
+
+    uint32_t eax = 0, ebx, ecx, edx;
+    ass_get_cpuid(&eax, &ebx, &ecx, &edx);
+    uint32_t max_leaf = eax;
+
+    bool avx = false;
+    if (max_leaf >= 1) {
+        eax = 1;
+        ass_get_cpuid(&eax, &ebx, &ecx, &edx);
+        if (edx & (1 << 26))  // SSE2
+            flags |= ASS_CPU_FLAG_X86_SSE2;
+
+        if (ecx & (1 << 27) &&  // OSXSAVE
+            ecx & (1 << 28)) {  // AVX
+            uint32_t xcr0l, xcr0h;
+            ass_get_xgetbv(0, &xcr0l, &xcr0h);
+            if (xcr0l & (1 << 1) &&  // XSAVE for XMM
+                xcr0l & (1 << 2))    // XSAVE for YMM
+                    avx = true;
+        }
+    }
+
+    if (max_leaf >= 7) {
+        eax = 7;
+        ass_get_cpuid(&eax, &ebx, &ecx, &edx);
+        if (avx && ebx & (1 << 5))  // AVX2
+            flags |= ASS_CPU_FLAG_X86_AVX2;
+    }
+
+#endif
+
+    return flags & mask;
+}
+
+const BitmapEngine *ass_bitmap_engine_init(unsigned mask)
+{
+#if CONFIG_ASM
+    unsigned flags = ass_get_cpu_flags(mask);
+#if ARCH_X86
+    if (flags & ASS_CPU_FLAG_X86_AVX2)
+        return &ass_bitmap_engine_avx2;
+    if (flags & ASS_CPU_FLAG_X86_SSE2)
+        return &ass_bitmap_engine_sse2;
+#endif
+#endif
+    return &ass_bitmap_engine_c;
+}
diff --git a/libass/ass_bitmap_engine.h b/libass/ass_bitmap_engine.h
index 8628bd9..7175031 100644
--- a/libass/ass_bitmap_engine.h
+++ b/libass/ass_bitmap_engine.h
@@ -80,8 +80,17 @@ typedef struct {
     ParamFilterFunc *blur_horz[5], *blur_vert[5];
 } BitmapEngine;
 
-extern const BitmapEngine ass_bitmap_engine_c;
-extern const BitmapEngine ass_bitmap_engine_sse2;
-extern const BitmapEngine ass_bitmap_engine_avx2;
+enum {
+    ASS_CPU_FLAG_NONE          = 0x0000,
+#if ARCH_X86
+    ASS_CPU_FLAG_X86_SSE2      = 0x0001,
+    ASS_CPU_FLAG_X86_AVX2      = 0x0002,
+#endif
+    ASS_CPU_FLAG_ALL           = 0x0FFF,
+};
+
+unsigned ass_get_cpu_flags(unsigned mask);
+
+const BitmapEngine *ass_bitmap_engine_init(unsigned mask);
 
 #endif /* LIBASS_BITMAP_ENGINE_H */
diff --git a/libass/ass_render.c b/libass/ass_render.c
index d0e1447..29a4374 100644
--- a/libass/ass_render.c
+++ b/libass/ass_render.c
@@ -106,18 +106,7 @@ ASS_Renderer *ass_renderer_init(ASS_Library *library)
     priv->ftlibrary = ft;
     // images_root and related stuff is zero-filled in calloc
 
-#if CONFIG_ASM && ARCH_X86
-    bool sse2, avx2;
-    ass_cpu_capabilities(&sse2, &avx2);
-    if (avx2)
-        priv->engine = &ass_bitmap_engine_avx2;
-    else if (sse2)
-        priv->engine = &ass_bitmap_engine_sse2;
-    else
-        priv->engine = &ass_bitmap_engine_c;
-#else
-    priv->engine = &ass_bitmap_engine_c;
-#endif
+    priv->engine = ass_bitmap_engine_init(ASS_CPU_FLAG_ALL);
 
     if (!ass_rasterizer_init(priv->engine, &priv->state.rasterizer, RASTERIZER_PRECISION))
         goto fail;
diff --git a/libass/ass_utils.c b/libass/ass_utils.c
index 9075e20..2c7923d 100644
--- a/libass/ass_utils.c
+++ b/libass/ass_utils.c
@@ -30,48 +30,6 @@
 #include "ass_utils.h"
 #include "ass_string.h"
 
-#if CONFIG_ASM && ARCH_X86
-
-#include "x86/cpuid.h"
-
-void ass_cpu_capabilities(bool *sse2, bool *avx2)
-{
-    *sse2 = false;
-    *avx2 = false;
-
-    if (!ass_has_cpuid())
-        return;
-
-    uint32_t eax = 0, ebx, ecx, edx;
-    ass_get_cpuid(&eax, &ebx, &ecx, &edx);
-    uint32_t max_leaf = eax;
-    bool avx = false;
-
-    if (max_leaf >= 1) {
-        eax = 1;
-        ass_get_cpuid(&eax, &ebx, &ecx, &edx);
-        if (edx & (1 << 26))  // SSE2
-            *sse2 = true;
-
-        if (ecx & (1 << 27) &&  // OSXSAVE
-            ecx & (1 << 28)) {  // AVX
-            uint32_t xcr0l, xcr0h;
-            ass_get_xgetbv(0, &xcr0l, &xcr0h);
-            if (xcr0l & (1 << 1) &&  // XSAVE for XMM
-                xcr0l & (1 << 2))    // XSAVE for YMM
-                    avx = true;
-        }
-    }
-
-    if (max_leaf >= 7) {
-        eax = 7;
-        ass_get_cpuid(&eax, &ebx, &ecx, &edx);
-        if (avx && ebx & (1 << 5))  // AVX2
-            *avx2 = true;
-    }
-}
-#endif // ASM
-
 // Fallbacks
 #ifndef HAVE_STRDUP
 char *ass_strdup_fallback(const char *str)
diff --git a/libass/ass_utils.h b/libass/ass_utils.h
index 7d68326..a1beb76 100644
--- a/libass/ass_utils.h
+++ b/libass/ass_utils.h
@@ -51,10 +51,6 @@
 
 #define FEATURE_MASK(feat) (((uint32_t) 1) << (feat))
 
-#if CONFIG_ASM && ARCH_X86
-void ass_cpu_capabilities(bool *sse2, bool *avx2);
-#endif
-
 typedef struct {
     const char *str;
     size_t len;