8 files changed, 491 insertions, 486 deletions

diff --git a/lib/Headers/emmintrin.h b/lib/Headers/emmintrin.h
index b09a62731c..6b9dd2aba5 100644
--- a/lib/Headers/emmintrin.h
+++ b/lib/Headers/emmintrin.h
@@ -36,417 +36,417 @@ typedef long long __m128i __attribute__((__vector_size__(16)));
 typedef short __v8hi __attribute__((__vector_size__(16)));
 typedef char __v16qi __attribute__((__vector_size__(16)));
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_add_sd(__m128d a, __m128d b)
 {
   a[0] += b[0];
   return a;
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_add_pd(__m128d a, __m128d b)
 {
   return a + b;
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_sub_sd(__m128d a, __m128d b)
 {
   a[0] -= b[0];
   return a;
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_sub_pd(__m128d a, __m128d b)
 {
   return a - b;
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_mul_sd(__m128d a, __m128d b)
 {
   a[0] *= b[0];
   return a;
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_mul_pd(__m128d a, __m128d b)
 {
   return a * b;
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_div_sd(__m128d a, __m128d b)
 {
   a[0] /= b[0];
   return a;
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_div_pd(__m128d a, __m128d b)
 {
   return a / b;
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_sqrt_sd(__m128d a, __m128d b)
 {
   __m128d c = __builtin_ia32_sqrtsd(b);
   return (__m128d) { c[0], a[1] };
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_sqrt_pd(__m128d a)
 {
   return __builtin_ia32_sqrtpd(a);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_min_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_minsd(a, b);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_min_pd(__m128d a, __m128d b)
 {
   return __builtin_ia32_minpd(a, b);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_max_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_maxsd(a, b);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_max_pd(__m128d a, __m128d b)
 {
   return __builtin_ia32_maxpd(a, b);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_and_pd(__m128d a, __m128d b)
 {
   return (__m128d)((__v4si)a & (__v4si)b);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_andnot_pd(__m128d a, __m128d b)
 {
   return (__m128d)(~(__v4si)a & (__v4si)b);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_or_pd(__m128d a, __m128d b)
 {
   return (__m128d)((__v4si)a | (__v4si)b);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_xor_pd(__m128d a, __m128d b)
 {
   return (__m128d)((__v4si)a ^ (__v4si)b);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpeq_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(a, b, 0);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmplt_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(a, b, 1);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmple_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(a, b, 2);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpgt_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(b, a, 1);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpge_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(b, a, 2);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpord_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(a, b, 7);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpunord_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(a, b, 3);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpneq_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(a, b, 4);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnlt_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(a, b, 5);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnle_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(a, b, 6);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpngt_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(b, a, 5);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnge_pd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmppd(b, a, 6);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpeq_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(a, b, 0);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmplt_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(a, b, 1);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmple_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(a, b, 2);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpgt_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(b, a, 1);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpge_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(b, a, 2);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpord_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(a, b, 7);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpunord_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(a, b, 3);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpneq_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(a, b, 4);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnlt_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(a, b, 5);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnle_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(a, b, 6);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpngt_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(b, a, 5);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cmpnge_sd(__m128d a, __m128d b)
 {
   return (__m128d)__builtin_ia32_cmpsd(b, a, 6);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_comieq_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_comisdeq(a, b);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_comilt_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_comisdlt(a, b);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_comile_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_comisdle(a, b);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_comigt_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_comisdgt(a, b);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_comineq_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_comisdneq(a, b);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_ucomieq_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_ucomisdeq(a, b);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_ucomilt_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_ucomisdlt(a, b);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_ucomile_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_ucomisdle(a, b);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_ucomigt_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_ucomisdgt(a, b);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_ucomineq_sd(__m128d a, __m128d b)
 {
   return __builtin_ia32_ucomisdneq(a, b);
 }
 
-static inline __m128 __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cvtpd_ps(__m128d a)
 {
   return __builtin_ia32_cvtpd2ps(a);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cvtps_pd(__m128 a)
 {
   return __builtin_ia32_cvtps2pd(a);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cvtepi32_pd(__m128i a)
 {
   return __builtin_ia32_cvtdq2pd((__v4si)a);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_cvtpd_epi32(__m128d a)
 {
   return __builtin_ia32_cvtpd2dq(a);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_cvtsd_si32(__m128d a)
 {
   return __builtin_ia32_cvtsd2si(a);
 }
 
-static inline __m128 __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_cvtsd_ss(__m128 a, __m128d b)
 {
   a[0] = b[0];
   return a;
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cvtsi32_sd(__m128d a, int b)
 {
   a[0] = b;
   return a;
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cvtss_sd(__m128d a, __m128 b)
 {
   a[0] = b[0];
   return a;
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_cvttpd_epi32(__m128d a)
 {
   return (__m128i)__builtin_ia32_cvttpd2dq(a);
 }
 
-static inline int __attribute__((__always_inline__, __nodebug__))
+static __inline__ int __attribute__((__always_inline__, __nodebug__))
 _mm_cvttsd_si32(__m128d a)
 {
   return a[0];
 }
 
-static inline __m64 __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
 _mm_cvtpd_pi32(__m128d a)
 {
   return (__m64)__builtin_ia32_cvtpd2pi(a);
 }
 
-static inline __m64 __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
 _mm_cvttpd_pi32(__m128d a)
 {
   return (__m64)__builtin_ia32_cvttpd2pi(a);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_cvtpi32_pd(__m64 a)
 {
   return __builtin_ia32_cvtpi2pd((__v2si)a);
 }
 
-static inline double __attribute__((__always_inline__, __nodebug__))
+static __inline__ double __attribute__((__always_inline__, __nodebug__))
 _mm_cvtsd_f64(__m128d a)
 {
   return a[0];
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_load_pd(double const *dp)
 {
   return *(__m128d*)dp;
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_load1_pd(double const *dp)
 {
   return (__m128d){ dp[0], dp[0] };
@@ -454,542 +454,542 @@ _mm_load1_pd(double const *dp)
 
 #define        _mm_load_pd1(dp)        _mm_load1_pd(dp)
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_loadr_pd(double const *dp)
 {
   return (__m128d){ dp[1], dp[0] };
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_loadu_pd(double const *dp)
 {
   return __builtin_ia32_loadupd(dp);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_load_sd(double const *dp)
 {
   return (__m128d){ *dp, 0.0 };
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_loadh_pd(__m128d a, double const *dp)
 {
   return __builtin_shufflevector(a, *(__m128d *)dp, 0, 2);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_loadl_pd(__m128d a, double const *dp)
 {
   return __builtin_shufflevector(a, *(__m128d *)dp, 2, 1);
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_set_sd(double w)
 {
   return (__m128d){ w, 0 };
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_set1_pd(double w)
 {
   return (__m128d){ w, w };
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_set_pd(double w, double x)
 {
   return (__m128d){ x, w };
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_setr_pd(double w, double x)
 {
   return (__m128d){ w, x };
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_setzero_pd(void)
 {
   return (__m128d){ 0, 0 };
 }
 
-static inline __m128d __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_move_sd(__m128d a, __m128d b)
 {
   return (__m128d){ b[0], a[1] };
 }
 
-static inline void __attribute__((__always_inline__, __nodebug__))
+static __inline__ void __attribute__((__always_inline__, __nodebug__))
 _mm_store_sd(double *dp, __m128d a)
 {
   dp[0] = a[0];
 }
 
-static inline void __attribute__((__always_inline__, __nodebug__))
+static __inline__ void __attribute__((__always_inline__, __nodebug__))
 _mm_store1_pd(double *dp, __m128d a)
 {
   dp[0] = a[0];
   dp[1] = a[0];
 }
 
-static inline void __attribute__((__always_inline__, __nodebug__))
+static __inline__ void __attribute__((__always_inline__, __nodebug__))
 _mm_store_pd(double *dp, __m128d a)
 {
   *(__m128d *)dp = a;
 }
 
-static inline void __attribute__((__always_inline__, __nodebug__))
+static __inline__ void __attribute__((__always_inline__, __nodebug__))
 _mm_storeu_pd(double *dp, __m128d a)
 {
   __builtin_ia32_storeupd(dp, a);
 }
 
-static inline void __attribute__((__always_inline__, __nodebug__))
+static __inline__ void __attribute__((__always_inline__, __nodebug__))
 _mm_storer_pd(double *dp, __m128d a)
 {
   dp[0] = a[1];
   dp[1] = a[0];
 }
 
-static inline void __attribute__((__always_inline__, __nodebug__))
+static __inline__ void __attribute__((__always_inline__, __nodebug__))
 _mm_storeh_pd(double *dp, __m128d a)
 {
   dp[0] = a[1];
 }
 
-static inline void __attribute__((__always_inline__, __nodebug__))
+static __inline__ void __attribute__((__always_inline__, __nodebug__))
 _mm_storel_pd(double *dp, __m128d a)
 {
   dp[0] = a[0];
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_add_epi8(__m128i a, __m128i b)
 {
   return (__m128i)((__v16qi)a + (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_add_epi16(__m128i a, __m128i b)
 {
   return (__m128i)((__v8hi)a + (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_add_epi32(__m128i a, __m128i b)
 {
   return (__m128i)((__v4si)a + (__v4si)b);
 }
 
-static inline __m64 __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
 _mm_add_si64(__m64 a, __m64 b)
 {
   return a + b;
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_add_epi64(__m128i a, __m128i b)
 {
   return a + b;
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_adds_epi8(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_paddsb128((__v16qi)a, (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_adds_epi16(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_paddsw128((__v8hi)a, (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_adds_epu8(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_paddusb128((__v16qi)a, (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_adds_epu16(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_paddusw128((__v8hi)a, (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_avg_epu8(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_pavgb128((__v16qi)a, (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_avg_epu16(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_pavgw128((__v8hi)a, (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_madd_epi16(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_pmaddwd128((__v8hi)a, (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_max_epi16(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_pmaxsw128((__v8hi)a, (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_max_epu8(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_pmaxub128((__v16qi)a, (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_min_epi16(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_pminsw128((__v8hi)a, (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_min_epu8(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_pminub128((__v16qi)a, (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_mulhi_epi16(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_pmulhw128((__v8hi)a, (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_mulhi_epu16(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)a, (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_mullo_epi16(__m128i a, __m128i b)
 {
   return (__m128i)((__v8hi)a * (__v8hi)b);
 }
 
-static inline __m64 __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
 _mm_mul_su32(__m64 a, __m64 b)
 {
   return __builtin_ia32_pmuludq((__v2si)a, (__v2si)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_mul_epu32(__m128i a, __m128i b)
 {
   return __builtin_ia32_pmuludq128((__v4si)a, (__v4si)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_sad_epu8(__m128i a, __m128i b)
 {
   return __builtin_ia32_psadbw128((__v16qi)a, (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_sub_epi8(__m128i a, __m128i b)
 {
   return (__m128i)((__v16qi)a - (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_sub_epi16(__m128i a, __m128i b)
 {
   return (__m128i)((__v8hi)a - (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_sub_epi32(__m128i a, __m128i b)
 {
   return (__m128i)((__v4si)a - (__v4si)b);
 }
 
-static inline __m64 __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
 _mm_sub_si64(__m64 a, __m64 b)
 {
   return a - b;
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_sub_epi64(__m128i a, __m128i b)
 {
   return a - b;
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_subs_epi8(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_psubsb128((__v16qi)a, (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_subs_epi16(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_psubsw128((__v8hi)a, (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_subs_epu8(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_psubusb128((__v16qi)a, (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_subs_epu16(__m128i a, __m128i b)
 {
   return (__m128i)__builtin_ia32_psubusw128((__v8hi)a, (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_and_si128(__m128i a, __m128i b)
 {
   return a & b;
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_andnot_si128(__m128i a, __m128i b)
 {
   return ~a & b;
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_or_si128(__m128i a, __m128i b)
 {
   return a | b;
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_xor_si128(__m128i a, __m128i b)
 {
   return a ^ b;
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_slli_si128(__m128i a, int imm)
 {
   return __builtin_ia32_pslldqi128(a, imm * 8);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_slli_epi16(__m128i a, int count)
 {
   return (__m128i)__builtin_ia32_psllwi128((__v8hi)a, count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_sll_epi16(__m128i a, __m128i count)
 {
   return (__m128i)__builtin_ia32_psllw128((__v8hi)a, (__v8hi)count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_slli_epi32(__m128i a, int count)
 {
   return (__m128i)__builtin_ia32_pslldi128((__v4si)a, count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_sll_epi32(__m128i a, __m128i count)
 {
   return (__m128i)__builtin_ia32_pslld128((__v4si)a, (__v4si)count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_slli_epi64(__m128i a, int count)
 {
   return __builtin_ia32_psllqi128(a, count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_sll_epi64(__m128i a, __m128i count)
 {
   return __builtin_ia32_psllq128(a, count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_srai_epi16(__m128i a, int count)
 {
   return (__m128i)__builtin_ia32_psrawi128((__v8hi)a, count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_sra_epi16(__m128i a, __m128i count)
 {
   return (__m128i)__builtin_ia32_psraw128((__v8hi)a, (__v8hi)count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_srai_epi32(__m128i a, int count)
 {
   return (__m128i)__builtin_ia32_psradi128((__v4si)a, count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_sra_epi32(__m128i a, __m128i count)
 {
   return (__m128i)__builtin_ia32_psrad128((__v4si)a, (__v4si)count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_srli_si128(__m128i a, int imm)
 {
   return __builtin_ia32_psrldqi128(a, imm * 8);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_srli_epi16(__m128i a, int count)
 {
   return (__m128i)__builtin_ia32_psrlwi128((__v8hi)a, count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_srl_epi16(__m128i a, __m128i count)
 {
   return (__m128i)__builtin_ia32_psrlw128((__v8hi)a, (__v8hi)count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_srli_epi32(__m128i a, int count)
 {
   return (__m128i)__builtin_ia32_psrldi128((__v4si)a, count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_srl_epi32(__m128i a, __m128i count)
 {
   return (__m128i)__builtin_ia32_psrld128((__v4si)a, (__v4si)count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_srli_epi64(__m128i a, int count)
 {
   return __builtin_ia32_psrlqi128(a, count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_srl_epi64(__m128i a, __m128i count)
 {
   return __builtin_ia32_psrlq128(a, count);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_cmpeq_epi8(__m128i a, __m128i b)
 {
   return (__m128i)((__v16qi)a == (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_cmpeq_epi16(__m128i a, __m128i b)
 {
   return (__m128i)((__v8hi)a == (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_cmpeq_epi32(__m128i a, __m128i b)
 {
   return (__m128i)((__v4si)a == (__v4si)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_cmpgt_epi8(__m128i a, __m128i b)
 {
   return (__m128i)((__v16qi)a > (__v16qi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_cmpgt_epi16(__m128i a, __m128i b)
 {
   return (__m128i)((__v8hi)a > (__v8hi)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_cmpgt_epi32(__m128i a, __m128i b)
 {
   return (__m128i)((__v4si)a > (__v4si)b);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_cmplt_epi8(__m128i a, __m128i b)
 {
   return _mm_cmpgt_epi8(b,a);
 }
 
-static inline __m128i __attribute__((__always_inline__, __nodebug__))
+static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_cmplt_epi16(__m128i a, __m128i b)
 {
   return _mm_cmpgt_epi16(b,a);
 }
 <