.syntax unified
+#if HAVE_NEON
+ .arch armv7-a
+#elif HAVE_ARMV6T2
+ .arch armv6t2
+#elif HAVE_ARMV6
+ .arch armv6
+#endif
+
+.fpu neon
+
#ifdef PREFIX
# define EXTERN_ASM _
#else
#include "asm.S"
-.fpu neon
.align 2
// done in gas because .fpu neon overrides the refusal to assemble
#include "asm.S"
-.fpu neon
-
.section .rodata
.align 4
#include "asm.S"
-.fpu neon
-
.macro h264_loop_filter_start
ldr ip, [sp]
ldr ip, [ip]
#include "asm.S"
-.fpu neon
.text
// note: prefetch stuff assumes 64-byte cacheline, true for the Cortex-A8
#include "asm.S"
-.fpu neon
.section .rodata
.align 4
#include "asm.S"
-.fpu neon
-
.section .rodata
.align 4
#include "asm.S"
-.fpu neon
-
.section .rodata
.align 4
pmovmskb_byte: