Use VEX encoding in inline math functions on x86-64 when possible

2 files changed, 142 insertions, 48 deletions

diff --git a/ChangeLog b/ChangeLog
index 203c6ec90e..1046feee0b 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,7 @@
+2011-10-25  Ulrich Drepper  <drepper@gmail.com>
+
+	* sysdeps/x86_64/fpu/math_private.h: Use VEX encoding when possible.
+
 2011-10-25  Andreas Schwab  <schwab@redhat.com>
 
 	* elf/dl-deps.c (_dl_map_object_deps): Remove always true
diff --git a/sysdeps/x86_64/fpu/math_private.h b/sysdeps/x86_64/fpu/math_private.h
index d3d84cfda4..7f52d5ee5f 100644
--- a/sysdeps/x86_64/fpu/math_private.h
+++ b/sysdeps/x86_64/fpu/math_private.h
@@ -1,59 +1,67 @@
 #ifndef _MATH_PRIVATE_H
 
 #define math_opt_barrier(x) \
-({ __typeof(x) __x;					\
-   if (sizeof (x) <= sizeof (double))			\
-     __asm ("" : "=x" (__x) : "0" (x));			\
-   else							\
-     __asm ("" : "=t" (__x) : "0" (x));			\
-   __x; })
+  ({ __typeof(x) __x;							      \
+     if (sizeof (x) <= sizeof (double))					      \
+       __asm ("" : "=x" (__x) : "0" (x));				      \
+     else								      \
+       __asm ("" : "=t" (__x) : "0" (x));				      \
+     __x; })
 #define math_force_eval(x) \
-do							\
-  {							\
-    if (sizeof (x) <= sizeof (double))			\
-      __asm __volatile ("" : : "x" (x));		\
-    else						\
-      __asm __volatile ("" : : "f" (x));		\
-  }							\
-while (0)
+  do {									      \
+    if (sizeof (x) <= sizeof (double))					      \
+      __asm __volatile ("" : : "x" (x));				      \
+    else								      \
+      __asm __volatile ("" : : "f" (x));				      \
+  } while (0)
 
 #include <math/math_private.h>
 
 /* We can do a few things better on x86-64.  */
 
+#ifdef __AVX__
+# define MOVD "vmovd"
+#else
+# define MOVD "movd"
+#endif
+
 /* Direct movement of float into integer register.  */
 #undef EXTRACT_WORDS64
-#define EXTRACT_WORDS64(i,d)					\
-do {								\
-  long int i_;							\
-  asm ("movd %1, %0" : "=rm" (i_) : "x" (d));			\
-  (i) = i_;							\
-} while (0)
+#define EXTRACT_WORDS64(i, d)						      \
+  do {									      \
+    long int i_;							      \
+    asm (MOVD " %1, %0" : "=rm" (i_) : "x" ((double) (d)));		      \
+    (i) = i_;								      \
+  } while (0)
 
 /* And the reverse.  */
 #undef INSERT_WORDS64
-#define INSERT_WORDS64(d,i) \
-do {								\
-  long int i_ = i;						\
-  asm ("movd %1, %0" : "=x" (d) : "rm" (i_));			\
-} while (0)
+#define INSERT_WORDS64(d, i) \
+  do {									      \
+    long int i_ = i;							      \
+    double d__;								      \
+    asm (MOVD " %1, %0" : "=x" (d__) : "rm" (i_));			      \
+    d = d__;								      \
+  } while (0)
 
 /* Direct movement of float into integer register.  */
 #undef GET_FLOAT_WORD
-#define GET_FLOAT_WORD(i,d) \
-do {								\
-  int i_;							\
-  asm ("movd %1, %0" : "=rm" (i_) : "x" (d));			\
-  (i) = i_;							\
-} while (0)
+#define GET_FLOAT_WORD(i, d) \
+  do {									      \
+    int i_;								      \
+    asm (MOVD " %1, %0" : "=rm" (i_) : "x" ((float) (d)));		      \
+    (i) = i_;								      \
+  } while (0)
 
 /* And the reverse.  */
 #undef SET_FLOAT_WORD
-#define SET_FLOAT_WORD(d,i) \
-do {								\
-  int i_ = i;							\
-  asm ("movd %1, %0" : "=x" (d) : "rm" (i_));			\
-} while (0)
+#define SET_FLOAT_WORD(f, i) \
+  do {									      \
+    int i_ = i;								      \
+    float f__;								      \
+    asm (MOVD " %1, %0" : "=x" (f__) : "rm" (i_));			      \
+    f = f__;								      \
+  } while (0)
 
 #endif
 
@@ -78,14 +86,25 @@ do {								\
   ({ int __di; GET_FLOAT_WORD (__di, (float) d);			      \
      (__di & 0x7fffffff) < 0x7f800000; })
 
-#define __ieee754_sqrt(d) \
+#ifdef __AVX__
+# define __ieee754_sqrt(d) \
+  ({ double __res;							      \
+    asm ("vsqrtsd %1, %0, %0" : "=x" (__res) : "xm" ((double) (d)));	      \
+     __res; })
+# define __ieee754_sqrtf(d) \
+  ({ float __res;							      \
+    asm ("vsqrtss %1, %0, %0" : "=x" (__res) : "xm" ((float) (d)));	      \
+     __res; })
+#else
+# define __ieee754_sqrt(d) \
   ({ double __res;							      \
     asm ("sqrtsd %1, %0" : "=x" (__res) : "xm" ((double) (d)));		      \
      __res; })
-#define __ieee754_sqrtf(d) \
+# define __ieee754_sqrtf(d) \
   ({ float __res;							      \
     asm ("sqrtss %1, %0" : "=x" (__res) : "xm" ((float) (d)));		      \
      __res; })
+#endif
 #define __ieee754_sqrtl(d) \
   ({ long double __res;							      \
     asm ("fsqrt" : "=t" (__res) : "0" ((long double) (d)));		      \
@@ -93,29 +112,57 @@ do {								\
 
 #ifdef __SSE4_1__
 # ifndef __rint
-#  define __rint(d) \
+#  ifdef __AVX__
+#   define __rint(d) \
+  ({ double __res; \
+    asm ("vroundsd $4, %1, %0, %0" : "=x" (__res) : "xm" ((double) (d)));      \
+     __res; })
+#  else
+#   define __rint(d) \
   ({ double __res; \
     asm ("roundsd $4, %1, %0" : "=x" (__res) : "xm" ((double) (d)));	      \
      __res; })
+#  endif
 # endif
 # ifndef __rintf
-#  define __rintf(d) \
+#  ifdef __AVX__
+#   define __rintf(d) \
+  ({ float __res; \
+    asm ("vroundss $4, %1, %0, %0" : "=x" (__res) : "xm" ((float) (d)));      \
+     __res; })
+#  else
+#   define __rintf(d) \
   ({ float __res; \
     asm ("roundss $4, %1, %0" : "=x" (__res) : "xm" ((float) (d)));	      \
      __res; })
+#  endif
 # endif
 
 # ifndef __floor
-#  define __floor(d) \
+#  ifdef __AVX__
+#   define __floor(d) \
+  ({ double __res; \
+    asm ("vroundsd $1, %1, %0, %0" : "=x" (__res) : "xm" ((double) (d)));      \
+     __res; })
+#  else
+#   define __floor(d) \
   ({ double __res; \
     asm ("roundsd $1, %1, %0" : "=x" (__res) : "xm" ((double) (d)));	      \
      __res; })
+#  endif
 # endif
 # ifndef __floorf
-#  define __floorf(d) \
+#  ifdef __AVX__
+#   define __floorf(d) \
+  ({ float __res; \
+    asm ("vroundss $1, %1, %0, %0" : "=x" (__res) : "xm" ((float) (d)));      \
+     __res; })
+#  else
+#   define __floorf(d) \
   ({ float __res; \
     asm ("roundss $1, %1, %0" : "=x" (__res) : "xm" ((float) (d)));	      \
      __res; })
+#  endif
 # endif
 #endif
 
@@ -146,7 +193,17 @@ do {								\
 // #define libc_fesetroundl(r) (void) fesetround (r)
 
 #undef libc_feholdexcept
-#define libc_feholdexcept(e) \
+#ifdef __AVX__
+# define libc_feholdexcept(e) \
+  do {									      \
+     unsigned int mxcsr;						      \
+     asm ("vstmxcsr %0" : "=m" (*&mxcsr));				      \
+     (e)->__mxcsr = mxcsr;						      \
+     mxcsr = (mxcsr | 0x1f80) & ~0x3f;					      \
+     asm volatile ("vldmxcsr %0" : : "m" (*&mxcsr));			      \
+  } while (0)
+#else
+# define libc_feholdexcept(e) \
   do {									      \
      unsigned int mxcsr;						      \
      asm ("stmxcsr %0" : "=m" (*&mxcsr));				      \
@@ -154,12 +211,23 @@ do {								\
      mxcsr = (mxcsr | 0x1f80) & ~0x3f;					      \
      asm volatile ("ldmxcsr %0" : : "m" (*&mxcsr));			      \
   } while (0)
+#endif
 #undef libc_feholdexceptf
 #define libc_feholdexceptf(e) libc_feholdexcept (e)
 // #define libc_feholdexceptl(e) (void) feholdexcept (e)
 
 #undef libc_feholdexcept_setround
-#define libc_feholdexcept_setround(e, r) \
+#ifdef __AVX__
+# define libc_feholdexcept_setround(e, r) \
+  do {									      \
+     unsigned int mxcsr;						      \
+     asm ("vstmxcsr %0" : "=m" (*&mxcsr));				      \
+     (e)->__mxcsr = mxcsr;						      \
+     mxcsr = ((mxcsr | 0x1f80) & ~0x603f) | ((r) << 3);			      \
+     asm volatile ("vldmxcsr %0" : : "m" (*&mxcsr));			      \
+  } while (0)
+#else
+# define libc_feholdexcept_setround(e, r) \
   do {									      \
      unsigned int mxcsr;						      \
      asm ("stmxcsr %0" : "=m" (*&mxcsr));				      \
@@ -167,33 +235,55 @@ do {								\
      mxcsr = ((mxcsr | 0x1f80) & ~0x603f) | ((r) << 3);			      \
      asm volatile ("ldmxcsr %0" : : "m" (*&mxcsr));			      \
   } while (0)
+#endif
 #undef libc_feholdexcept_setroundf
 #define libc_feholdexcept_setroundf(e, r) libc_feholdexcept_setround (e, r)
 // #define libc_feholdexcept_setroundl(e, r) ...
 
 #undef libc_fetestexcept
-#define libc_fetestexcept(e) \
-  ({ unsigned int mxcsr; asm volatile ("stmxcsr %0" : "=m" (*&mxcsr)); \
+#ifdef __AVX__
+# define libc_fetestexcept(e) \
+  ({ unsigned int mxcsr; asm volatile ("vstmxcsr %0" : "=m" (*&mxcsr));	      \
      mxcsr & (e) & FE_ALL_EXCEPT; })
+#else
+# define libc_fetestexcept(e) \
+  ({ unsigned int mxcsr; asm volatile ("stmxcsr %0" : "=m" (*&mxcsr));	      \
+     mxcsr & (e) & FE_ALL_EXCEPT; })
+#endif
 #undef libc_fetestexceptf
 #define libc_fetestexceptf(e) libc_fetestexcept (e)
 // #define libc_fetestexceptl(e) fetestexcept (e)
 
 #undef libc_fesetenv
-#define libc_fesetenv(e) \
+#ifdef __AVX__
+# define libc_fesetenv(e) \
+  asm volatile ("vldmxcsr %0" : : "m" ((e)->__mxcsr))
+#else
+# define libc_fesetenv(e) \
   asm volatile ("ldmxcsr %0" : : "m" ((e)->__mxcsr))
+#endif
 #undef libc_fesetenvf
 #define libc_fesetenvf(e) libc_fesetenv (e)
 // #define libc_fesetenvl(e) (void) fesetenv (e)
 
 #undef libc_feupdateenv
-#define libc_feupdateenv(e) \
+#ifdef __AVX__
+# define libc_feupdateenv(e) \
+  do {									      \
+    unsigned int mxcsr;							      \
+    asm volatile ("vstmxcsr %0" : "=m" (*&mxcsr));			      \
+    asm volatile ("vldmxcsr %0" : : "m" ((e)->__mxcsr));		      \
+    __feraiseexcept (mxcsr & FE_ALL_EXCEPT);				      \
+  } while (0)
+#else
+# define libc_feupdateenv(e) \
   do {									      \
     unsigned int mxcsr;							      \
     asm volatile ("stmxcsr %0" : "=m" (*&mxcsr));			      \
     asm volatile ("ldmxcsr %0" : : "m" ((e)->__mxcsr));			      \
     __feraiseexcept (mxcsr & FE_ALL_EXCEPT);				      \
   } while (0)
+#endif
 #undef libc_feupdateenvf
 #define libc_feupdateenvf(e) libc_feupdateenv (e)
 // #define libc_feupdateenvl(e) (void) feupdateenv (e)