*
* Developed at SunPro, a Sun Microsystems, Inc. business.
* Permission to use, copy, modify, and distribute this
- * software is freely granted, provided that this notice
+ * software is freely granted, provided that this notice
* is preserved.
* ====================================================
*/
-/*
- * from: @(#)fdlibm.h 5.1 93/09/24
- * $Id: math_private.h,v 1.2 2002/06/26 09:10:51 andersen Exp $
- */
-
#ifndef _MATH_PRIVATE_H_
#define _MATH_PRIVATE_H_
ints. */
/*
- * Math on arm is little endian except for the FP word order which is
- * big endian.
+ * Math on arm is special (read: stupid):
+ * For FPA, float words are always big-endian.
+ * For VFP, float words follow the memory system mode.
+ * For Maverick, float words are always little-endian.
*/
-#if (__BYTE_ORDER == __BIG_ENDIAN) || defined(__arm__)
+#if !defined(__MAVERICK__) && ((__BYTE_ORDER == __BIG_ENDIAN) || \
+ (!defined(__VFP_FP__) && (defined(__arm__) || defined(__thumb__))))
-typedef union
+typedef union
{
double value;
- struct
+ struct
{
u_int32_t msw;
u_int32_t lsw;
} parts;
} ieee_double_shape_type;
-#endif
-
-#if (__BYTE_ORDER == __LITTLE_ENDIAN) && !defined(__arm__)
+#else
-typedef union
+typedef union
{
double value;
- struct
+ struct
{
u_int32_t lsw;
u_int32_t msw;
} while (0)
/* ieee style elementary functions */
-extern double __ieee754_sqrt __P((double));
-extern double __ieee754_acos __P((double));
-extern double __ieee754_acosh __P((double));
-extern double __ieee754_log __P((double));
-extern double __ieee754_atanh __P((double));
-extern double __ieee754_asin __P((double));
-extern double __ieee754_atan2 __P((double,double));
-extern double __ieee754_exp __P((double));
-extern double __ieee754_cosh __P((double));
-extern double __ieee754_fmod __P((double,double));
-extern double __ieee754_pow __P((double,double));
-extern double __ieee754_lgamma_r __P((double,int *));
-extern double __ieee754_gamma_r __P((double,int *));
-extern double __ieee754_lgamma __P((double));
-extern double __ieee754_gamma __P((double));
-extern double __ieee754_log10 __P((double));
-extern double __ieee754_sinh __P((double));
-extern double __ieee754_hypot __P((double,double));
-extern double __ieee754_j0 __P((double));
-extern double __ieee754_j1 __P((double));
-extern double __ieee754_y0 __P((double));
-extern double __ieee754_y1 __P((double));
-extern double __ieee754_jn __P((int,double));
-extern double __ieee754_yn __P((int,double));
-extern double __ieee754_remainder __P((double,double));
-extern int __ieee754_rem_pio2 __P((double,double*));
-#if defined(_SCALB_INT)
-extern double __ieee754_scalb __P((double,int));
-#else
-extern double __ieee754_scalb __P((double,double));
-#endif
+extern double __ieee754_sqrt (double) attribute_hidden;
+extern double __ieee754_acos (double) attribute_hidden;
+extern double __ieee754_acosh (double) attribute_hidden;
+extern double __ieee754_log (double) attribute_hidden;
+extern double __ieee754_log2 (double) attribute_hidden;
+extern double __ieee754_atanh (double) attribute_hidden;
+extern double __ieee754_asin (double) attribute_hidden;
+extern double __ieee754_atan2 (double,double) attribute_hidden;
+extern double __ieee754_exp (double) attribute_hidden;
+extern double __ieee754_cosh (double) attribute_hidden;
+extern double __ieee754_fmod (double,double) attribute_hidden;
+extern double __ieee754_pow (double,double) attribute_hidden;
+extern double __ieee754_lgamma_r (double,int *) attribute_hidden;
+/*extern double __ieee754_gamma_r (double,int *) attribute_hidden;*/
+extern double __ieee754_lgamma (double) attribute_hidden;
+/*extern double __ieee754_gamma (double) attribute_hidden;*/
+extern double __ieee754_log10 (double) attribute_hidden;
+extern double __ieee754_sinh (double) attribute_hidden;
+extern double __ieee754_hypot (double,double) attribute_hidden;
+extern double __ieee754_j0 (double) attribute_hidden;
+extern double __ieee754_j1 (double) attribute_hidden;
+extern double __ieee754_y0 (double) attribute_hidden;
+extern double __ieee754_y1 (double) attribute_hidden;
+extern double __ieee754_jn (int,double) attribute_hidden;
+extern double __ieee754_yn (int,double) attribute_hidden;
+extern double __ieee754_remainder (double,double) attribute_hidden;
+extern int __ieee754_rem_pio2 (double,double*) attribute_hidden;
+extern double __ieee754_scalb (double,double) attribute_hidden;
/* fdlibm kernel function */
-extern double __kernel_standard __P((double,double,int));
-extern double __kernel_sin __P((double,double,int));
-extern double __kernel_cos __P((double,double));
-extern double __kernel_tan __P((double,double,int));
-extern int __kernel_rem_pio2 __P((double*,double*,int,int,int,const int*));
-
-
-/* ieee style elementary float functions */
-extern float __ieee754_sqrtf __P((float));
-extern float __ieee754_acosf __P((float));
-extern float __ieee754_acoshf __P((float));
-extern float __ieee754_logf __P((float));
-extern float __ieee754_atanhf __P((float));
-extern float __ieee754_asinf __P((float));
-extern float __ieee754_atan2f __P((float,float));
-extern float __ieee754_expf __P((float));
-extern float __ieee754_coshf __P((float));
-extern float __ieee754_fmodf __P((float,float));
-extern float __ieee754_powf __P((float,float));
-extern float __ieee754_lgammaf_r __P((float,int *));
-extern float __ieee754_gammaf_r __P((float,int *));
-extern float __ieee754_lgammaf __P((float));
-extern float __ieee754_gammaf __P((float));
-extern float __ieee754_log10f __P((float));
-extern float __ieee754_sinhf __P((float));
-extern float __ieee754_hypotf __P((float,float));
-extern float __ieee754_j0f __P((float));
-extern float __ieee754_j1f __P((float));
-extern float __ieee754_y0f __P((float));
-extern float __ieee754_y1f __P((float));
-extern float __ieee754_jnf __P((int,float));
-extern float __ieee754_ynf __P((int,float));
-extern float __ieee754_remainderf __P((float,float));
-extern int __ieee754_rem_pio2f __P((float,float*));
-extern float __ieee754_scalbf __P((float,float));
-
-/* float versions of fdlibm kernel functions */
-extern float __kernel_sinf __P((float,float,int));
-extern float __kernel_cosf __P((float,float));
-extern float __kernel_tanf __P((float,float,int));
-extern int __kernel_rem_pio2f __P((float*,float*,int,int,int,const int*));
+#ifndef _IEEE_LIBM
+extern double __kernel_standard (double,double,int) attribute_hidden;
+#endif
+extern double __kernel_sin (double,double,int) attribute_hidden;
+extern double __kernel_cos (double,double) attribute_hidden;
+extern double __kernel_tan (double,double,int) attribute_hidden;
+extern int __kernel_rem_pio2 (double*,double*,int,int,int,const int*) attribute_hidden;
+
+/*
+ * math_opt_barrier(x): safely load x, even if it was manipulated
+ * by non-floationg point operations. This macro returns the value of x.
+ * This ensures compiler does not (ab)use its knowledge about x value
+ * and don't optimize future operations. Example:
+ * float x;
+ * SET_FLOAT_WORD(x, 0x80000001); // sets a bit pattern
+ * y = math_opt_barrier(x); // "compiler, do not cheat!"
+ * y = y * y; // compiler can't optimize, must use real multiply insn
+ *
+ * math_force_eval(x): force expression x to be evaluated.
+ * Useful if otherwise compiler may eliminate the expression
+ * as unused. This macro returns no value.
+ * Example: "void fn(float f) { f = f * f; }"
+ * versus "void fn(float f) { f = f * f; math_force_eval(f); }"
+ *
+ * Currently, math_force_eval(x) stores x into
+ * a floating point register or memory *of the appropriate size*.
+ * There is no guarantee this will not change.
+ */
+#if defined(__i386__)
+#define math_opt_barrier(x) ({ \
+ __typeof(x) __x = (x); \
+ /* "t": load x into top-of-stack fpreg */ \
+ __asm__ ("" : "=t" (__x) : "0" (__x)); \
+ __x; \
+})
+#define math_force_eval(x) do { \
+ __typeof(x) __x = (x); \
+ if (sizeof(__x) <= sizeof(double)) \
+ /* "m": store x into a memory location */ \
+ __asm__ __volatile__ ("" : : "m" (__x)); \
+ else /* long double */ \
+ /* "f": load x into (any) fpreg */ \
+ __asm__ __volatile__ ("" : : "f" (__x)); \
+} while (0)
+#endif
+
+#if defined(__x86_64__)
+#define math_opt_barrier(x) ({ \
+ __typeof(x) __x = (x); \
+ if (sizeof(__x) <= sizeof(double)) \
+ /* "x": load into XMM SSE register */ \
+ __asm__ ("" : "=x" (__x) : "0" (__x)); \
+ else /* long double */ \
+ /* "t": load x into top-of-stack fpreg */ \
+ __asm__ ("" : "=t" (__x) : "0" (__x)); \
+ __x; \
+})
+#define math_force_eval(x) do { \
+ __typeof(x) __x = (x); \
+ if (sizeof(__x) <= sizeof(double)) \
+ /* "x": load into XMM SSE register */ \
+ __asm__ __volatile__ ("" : : "x" (__x)); \
+ else /* long double */ \
+ /* "f": load x into (any) fpreg */ \
+ __asm__ __volatile__ ("" : : "f" (__x)); \
+} while (0)
+#endif
+
+/* Default implementations force store to a memory location */
+#ifndef math_opt_barrier
+#define math_opt_barrier(x) ({ __typeof(x) __x = (x); __asm__ ("" : "+m" (__x)); __x; })
+#endif
+#ifndef math_force_eval
+#define math_force_eval(x) do { __typeof(x) __x = (x); __asm__ __volatile__ ("" : : "m" (__x)); } while (0)
+#endif
+
#endif /* _MATH_PRIVATE_H_ */
OSDN Git Service
