Merge "Clarify GCC version check"

vp8/common/arm/neon/loopfilter_neon.c

@@ -10,6 +10,7 @@
 
 #include <arm_neon.h>
 #include "./vpx_config.h"
+#include "vpx_ports/arm.h"
 
 static INLINE void vp8_loop_filter_neon(
         uint8x16_t qblimit,  // flimit
@@ -253,23 +254,7 @@ void vp8_loop_filter_horizontal_edge_uv_neon(
 
 static INLINE void write_4x8(unsigned char *dst, int pitch,
                              const uint8x8x4_t result) {
-#if (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7))
-    vst4_lane_u8(dst, result, 0);
-    dst += pitch;
-    vst4_lane_u8(dst, result, 1);
-    dst += pitch;
-    vst4_lane_u8(dst, result, 2);
-    dst += pitch;
-    vst4_lane_u8(dst, result, 3);
-    dst += pitch;
-    vst4_lane_u8(dst, result, 4);
-    dst += pitch;
-    vst4_lane_u8(dst, result, 5);
-    dst += pitch;
-    vst4_lane_u8(dst, result, 6);
-    dst += pitch;
-    vst4_lane_u8(dst, result, 7);
-#else
+#ifdef VPX_INCOMPATIBLE_GCC
     /*
      * uint8x8x4_t result
     00 01 02 03 | 04 05 06 07
@@ -316,7 +301,23 @@ static INLINE void write_4x8(unsigned char *dst, int pitch,
     vst1_lane_u32((uint32_t *)dst, x_2_6, 1);
     dst += pitch;
     vst1_lane_u32((uint32_t *)dst, x_3_7, 1);
-#endif
+#else
+    vst4_lane_u8(dst, result, 0);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 1);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 2);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 3);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 4);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 5);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 6);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 7);
+#endif  // VPX_INCOMPATIBLE_GCC
 }
 
 void vp8_loop_filter_vertical_edge_y_neon(

vp8/common/arm/neon/loopfiltersimpleverticaledge_neon.c

@@ -10,45 +10,9 @@
 
 #include <arm_neon.h>
 #include "./vpx_config.h"
+#include "vpx_ports/arm.h"
 
-#if (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7))
-static INLINE void write_2x8(unsigned char *dst, int pitch,
-                             const uint8x8x2_t result,
-                             const uint8x8x2_t result2) {
-  vst2_lane_u8(dst, result, 0);
-  dst += pitch;
-  vst2_lane_u8(dst, result, 1);
-  dst += pitch;
-  vst2_lane_u8(dst, result, 2);
-  dst += pitch;
-  vst2_lane_u8(dst, result, 3);
-  dst += pitch;
-  vst2_lane_u8(dst, result, 4);
-  dst += pitch;
-  vst2_lane_u8(dst, result, 5);
-  dst += pitch;
-  vst2_lane_u8(dst, result, 6);
-  dst += pitch;
-  vst2_lane_u8(dst, result, 7);
-  dst += pitch;
-
-  vst2_lane_u8(dst, result2, 0);
-  dst += pitch;
-  vst2_lane_u8(dst, result2, 1);
-  dst += pitch;
-  vst2_lane_u8(dst, result2, 2);
-  dst += pitch;
-  vst2_lane_u8(dst, result2, 3);
-  dst += pitch;
-  vst2_lane_u8(dst, result2, 4);
-  dst += pitch;
-  vst2_lane_u8(dst, result2, 5);
-  dst += pitch;
-  vst2_lane_u8(dst, result2, 6);
-  dst += pitch;
-  vst2_lane_u8(dst, result2, 7);
-}
-#else
+#ifdef VPX_INCOMPATIBLE_GCC
 static INLINE void write_2x4(unsigned char *dst, int pitch,
                              const uint8x8x2_t result) {
     /*
@@ -88,30 +52,47 @@ static INLINE void write_2x8(unsigned char *dst, int pitch,
   dst += pitch * 8;
   write_2x4(dst, pitch, result2);
 }
-#endif
-
+#else
+static INLINE void write_2x8(unsigned char *dst, int pitch,
+                             const uint8x8x2_t result,
+                             const uint8x8x2_t result2) {
+  vst2_lane_u8(dst, result, 0);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 1);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 2);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 3);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 4);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 5);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 6);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 7);
+  dst += pitch;
 
-#if (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7))
-static INLINE
-uint8x8x4_t read_4x8(unsigned char *src, int pitch, uint8x8x4_t x) {
-    x = vld4_lane_u8(src, x, 0);
-    src += pitch;
-    x = vld4_lane_u8(src, x, 1);
-    src += pitch;
-    x = vld4_lane_u8(src, x, 2);
-    src += pitch;
-    x = vld4_lane_u8(src, x, 3);
-    src += pitch;
-    x = vld4_lane_u8(src, x, 4);
-    src += pitch;
-    x = vld4_lane_u8(src, x, 5);
-    src += pitch;
-    x = vld4_lane_u8(src, x, 6);
-    src += pitch;
-    x = vld4_lane_u8(src, x, 7);
-    return x;
+  vst2_lane_u8(dst, result2, 0);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 1);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 2);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 3);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 4);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 5);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 6);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 7);
 }
-#else
+#endif  // VPX_INCOMPATIBLE_GCC
+
+
+#ifdef VPX_INCOMPATIBLE_GCC
 static INLINE
 uint8x8x4_t read_4x8(unsigned char *src, int pitch, uint8x8x4_t x) {
     const uint8x8_t a = vld1_u8(src);
@@ -169,7 +150,27 @@ uint8x8x4_t read_4x8(unsigned char *src, int pitch, uint8x8x4_t x) {
 
     return x;
 }
-#endif
+#else
+static INLINE
+uint8x8x4_t read_4x8(unsigned char *src, int pitch, uint8x8x4_t x) {
+    x = vld4_lane_u8(src, x, 0);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 1);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 2);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 3);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 4);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 5);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 6);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 7);
+    return x;
+}
+#endif  // VPX_INCOMPATIBLE_GCC
 
 static INLINE void vp8_loop_filter_simple_vertical_edge_neon(
         unsigned char *s,

vpx_ports/arm.h

@@ -27,6 +27,12 @@ extern "C" {
 
 int arm_cpu_caps(void);
 
+// Earlier gcc compilers have issues with some neon intrinsics
+#if !defined(__clang__) && defined(__GNUC__) && \
+    __GNUC__ == 4 && __GNUC_MINOR__ <= 6
+#define VPX_INCOMPATIBLE_GCC
+#endif
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif