x86: Add asm for mbtree fixed point conversion

[x264] / common / cpu.h

diff --git a/common/cpu.h b/common/cpu.h
index 7f76e435b050e6b357b1fda0acb569f1f6ee4869..138141946fc4ff6929a5aa7eed29bb95643478be 100644 (file)
--- a/common/cpu.h
+++ b/common/cpu.h
@@ -1,7 +1,7 @@
 /*****************************************************************************
  * cpu.h: cpu detection
  *****************************************************************************
- * Copyright (C) 2004-2013 x264 project
+ * Copyright (C) 2004-2016 x264 project
  *
  * Authors: Loren Merritt <lorenm@u.washington.edu>
  *
@@ -45,19 +45,19 @@ void     x264_cpu_sfence( void );
 #define x264_emms()
 #endif
 #define x264_sfence x264_cpu_sfence
-void     x264_cpu_mask_misalign_sse( void );
-void     x264_safe_intel_cpu_indicator_init( void );
 
-/* kluge:
+/* kludge:
  * gcc can't give variables any greater alignment than the stack frame has.
- * We need 16 byte alignment for SSE2, so here we make sure that the stack is
- * aligned to 16 bytes.
+ * We need 32 byte alignment for AVX2, so here we make sure that the stack is
+ * aligned to 32 bytes.
  * gcc 4.2 introduced __attribute__((force_align_arg_pointer)) to fix this
  * problem, but I don't want to require such a new version.
- * This applies only to x86_32, since other architectures that need alignment
- * either have ABIs that ensure aligned stack, or don't support it at all. */
-#if ARCH_X86 && HAVE_MMX
-int x264_stack_align( void (*func)(), ... );
+ * aligning to 32 bytes only works if the compiler supports keeping that
+ * alignment between functions (osdep.h handles manual alignment of arrays
+ * if it doesn't).
+ */
+#if (ARCH_X86 || STACK_ALIGNMENT > 16) && HAVE_MMX
+intptr_t x264_stack_align( void (*func)(), ... );
 #define x264_stack_align(func,...) x264_stack_align((void (*)())func, __VA_ARGS__)
 #else
 #define x264_stack_align(func,...) func(__VA_ARGS__)