2012-05-05 21:34:31 +02:00
|
|
|
/*
|
|
|
|
* Written by J.T. Conklin <jtc@netbsd.org>.
|
|
|
|
* Public domain.
|
|
|
|
*
|
|
|
|
* Adapted for `long double' by Ulrich Drepper <drepper@cygnus.com>.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The 8087 method for the exponential function is to calculate
|
|
|
|
* exp(x) = 2^(x log2(e))
|
|
|
|
* after separating integer and fractional parts
|
|
|
|
* x log2(e) = i + f, |f| <= .5
|
|
|
|
* 2^i is immediate but f needs to be precise for long double accuracy.
|
|
|
|
* Suppress range reduction error in computing f by the following.
|
|
|
|
* Separate x into integer and fractional parts
|
|
|
|
* x = xi + xf, |xf| <= .5
|
|
|
|
* Separate log2(e) into the sum of an exact number c0 and small part c1.
|
|
|
|
* c0 + c1 = log2(e) to extra precision
|
|
|
|
* Then
|
|
|
|
* f = (c0 xi - i) + c0 xf + c1 x
|
|
|
|
* where c0 xi is exact and so also is (c0 xi - i).
|
|
|
|
* -- moshier@na-net.ornl.gov
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <machine/asm.h>
|
|
|
|
|
2012-05-06 20:23:44 +02:00
|
|
|
#ifdef USE_AS_EXP10L
|
|
|
|
# define IEEE754_EXPL __ieee754_exp10l
|
|
|
|
# define EXPL_FINITE __exp10l_finite
|
|
|
|
# define FLDLOG fldl2t
|
2012-05-07 21:13:08 +02:00
|
|
|
#elif defined USE_AS_EXPM1L
|
|
|
|
# define IEEE754_EXPL __expm1l
|
|
|
|
# undef EXPL_FINITE
|
|
|
|
# define FLDLOG fldl2e
|
2012-05-06 20:23:44 +02:00
|
|
|
#else
|
|
|
|
# define IEEE754_EXPL __ieee754_expl
|
|
|
|
# define EXPL_FINITE __expl_finite
|
|
|
|
# define FLDLOG fldl2e
|
|
|
|
#endif
|
|
|
|
|
2012-05-05 21:34:31 +02:00
|
|
|
.section .rodata.cst16,"aM",@progbits,16
|
|
|
|
|
|
|
|
.p2align 4
|
2012-05-06 20:23:44 +02:00
|
|
|
#ifdef USE_AS_EXP10L
|
2012-08-02 21:04:29 +02:00
|
|
|
.type c0,@object
|
2012-05-06 20:23:44 +02:00
|
|
|
c0: .byte 0, 0, 0, 0, 0, 0, 0x9a, 0xd4, 0x00, 0x40
|
|
|
|
.byte 0, 0, 0, 0, 0, 0
|
|
|
|
ASM_SIZE_DIRECTIVE(c0)
|
2012-08-02 21:04:29 +02:00
|
|
|
.type c1,@object
|
2012-05-06 20:23:44 +02:00
|
|
|
c1: .byte 0x58, 0x92, 0xfc, 0x15, 0x37, 0x9a, 0x97, 0xf0, 0xef, 0x3f
|
|
|
|
.byte 0, 0, 0, 0, 0, 0
|
|
|
|
ASM_SIZE_DIRECTIVE(c1)
|
|
|
|
#else
|
2012-08-02 21:04:29 +02:00
|
|
|
.type c0,@object
|
2012-05-05 21:34:31 +02:00
|
|
|
c0: .byte 0, 0, 0, 0, 0, 0, 0xaa, 0xb8, 0xff, 0x3f
|
|
|
|
.byte 0, 0, 0, 0, 0, 0
|
|
|
|
ASM_SIZE_DIRECTIVE(c0)
|
2012-08-02 21:04:29 +02:00
|
|
|
.type c1,@object
|
2012-05-05 21:34:31 +02:00
|
|
|
c1: .byte 0x20, 0xfa, 0xee, 0xc2, 0x5f, 0x70, 0xa5, 0xec, 0xed, 0x3f
|
|
|
|
.byte 0, 0, 0, 0, 0, 0
|
|
|
|
ASM_SIZE_DIRECTIVE(c1)
|
2012-05-06 20:23:44 +02:00
|
|
|
#endif
|
2012-07-06 13:17:41 +02:00
|
|
|
#ifndef USE_AS_EXPM1L
|
2012-08-02 21:04:29 +02:00
|
|
|
.type csat,@object
|
2012-05-05 21:37:39 +02:00
|
|
|
csat: .byte 0, 0, 0, 0, 0, 0, 0, 0x80, 0x0e, 0x40
|
|
|
|
.byte 0, 0, 0, 0, 0, 0
|
|
|
|
ASM_SIZE_DIRECTIVE(csat)
|
2012-07-06 13:17:41 +02:00
|
|
|
#endif
|
2012-05-05 21:34:31 +02:00
|
|
|
|
|
|
|
#ifdef PIC
|
|
|
|
# define MO(op) op##@GOTOFF(%ecx)
|
|
|
|
#else
|
|
|
|
# define MO(op) op
|
|
|
|
#endif
|
|
|
|
|
|
|
|
.text
|
2012-05-06 20:23:44 +02:00
|
|
|
ENTRY(IEEE754_EXPL)
|
2012-05-07 21:13:08 +02:00
|
|
|
#ifdef USE_AS_EXPM1L
|
|
|
|
movzwl 4+8(%esp), %eax
|
|
|
|
xorb $0x80, %ah // invert sign bit (now 1 is "positive")
|
|
|
|
cmpl $0xc006, %eax // is num positive and exp >= 6 (number is >= 128.0)?
|
|
|
|
jae HIDDEN_JUMPTARGET (__expl) // (if num is denormal, it is at least >= 64.0)
|
|
|
|
#endif
|
2012-05-05 21:34:31 +02:00
|
|
|
fldt 4(%esp)
|
|
|
|
/* I added the following ugly construct because expl(+-Inf) resulted
|
|
|
|
in NaN. The ugliness results from the bright minds at Intel.
|
|
|
|
For the i686 the code can be written better.
|
|
|
|
-- drepper@cygnus.com. */
|
|
|
|
fxam /* Is NaN or +-Inf? */
|
|
|
|
#ifdef PIC
|
|
|
|
LOAD_PIC_REG (cx)
|
|
|
|
#endif
|
2012-07-06 13:17:41 +02:00
|
|
|
#ifdef USE_AS_EXPM1L
|
|
|
|
xorb $0x80, %ah
|
|
|
|
cmpl $0xc006, %eax
|
|
|
|
fstsw %ax
|
|
|
|
movb $0x45, %dh
|
|
|
|
jb 4f
|
|
|
|
|
|
|
|
/* Below -64.0 (may be -NaN or -Inf). */
|
|
|
|
andb %ah, %dh
|
|
|
|
cmpb $0x01, %dh
|
|
|
|
je 2f /* Is +-NaN, jump. */
|
|
|
|
jmp 1f /* -large, possibly -Inf. */
|
|
|
|
|
|
|
|
4: /* In range -64.0 to 64.0 (may be +-0 but not NaN or +-Inf). */
|
|
|
|
/* Test for +-0 as argument. */
|
|
|
|
andb %ah, %dh
|
|
|
|
cmpb $0x40, %dh
|
|
|
|
je 2f
|
|
|
|
#else
|
2012-05-05 21:37:39 +02:00
|
|
|
movzwl 4+8(%esp), %eax
|
|
|
|
andl $0x7fff, %eax
|
|
|
|
cmpl $0x400d, %eax
|
|
|
|
jle 3f
|
|
|
|
/* Overflow, underflow or infinity or NaN as argument. */
|
2012-05-05 21:34:31 +02:00
|
|
|
fstsw %ax
|
|
|
|
movb $0x45, %dh
|
|
|
|
andb %ah, %dh
|
|
|
|
cmpb $0x05, %dh
|
|
|
|
je 1f /* Is +-Inf, jump. */
|
2012-05-05 21:37:39 +02:00
|
|
|
cmpb $0x01, %dh
|
|
|
|
je 2f /* Is +-NaN, jump. */
|
|
|
|
/* Overflow or underflow; saturate. */
|
|
|
|
fstp %st
|
|
|
|
fldt MO(csat)
|
|
|
|
andb $2, %ah
|
|
|
|
jz 3f
|
|
|
|
fchs
|
2012-05-07 21:13:08 +02:00
|
|
|
#endif
|
2012-07-06 13:17:41 +02:00
|
|
|
3: FLDLOG /* 1 log2(base) */
|
2012-05-06 20:23:44 +02:00
|
|
|
fmul %st(1), %st /* 1 x log2(base) */
|
2012-05-05 21:34:31 +02:00
|
|
|
frndint /* 1 i */
|
|
|
|
fld %st(1) /* 2 x */
|
|
|
|
frndint /* 2 xi */
|
|
|
|
fld %st(1) /* 3 i */
|
|
|
|
fldt MO(c0) /* 4 c0 */
|
|
|
|
fld %st(2) /* 5 xi */
|
|
|
|
fmul %st(1), %st /* 5 c0 xi */
|
|
|
|
fsubp %st, %st(2) /* 4 f = c0 xi - i */
|
|
|
|
fld %st(4) /* 5 x */
|
|
|
|
fsub %st(3), %st /* 5 xf = x - xi */
|
|
|
|
fmulp %st, %st(1) /* 4 c0 xf */
|
|
|
|
faddp %st, %st(1) /* 3 f = f + c0 xf */
|
|
|
|
fldt MO(c1) /* 4 */
|
|
|
|
fmul %st(4), %st /* 4 c1 * x */
|
|
|
|
faddp %st, %st(1) /* 3 f = f + c1 * x */
|
2012-05-06 20:23:44 +02:00
|
|
|
f2xm1 /* 3 2^(fract(x * log2(base))) - 1 */
|
2012-05-07 21:13:08 +02:00
|
|
|
#ifdef USE_AS_EXPM1L
|
|
|
|
fstp %st(1) /* 2 */
|
|
|
|
fscale /* 2 scale factor is st(1); base^x - 2^i */
|
|
|
|
fxch /* 2 i */
|
|
|
|
fld1 /* 3 1.0 */
|
|
|
|
fscale /* 3 2^i */
|
|
|
|
fld1 /* 4 1.0 */
|
|
|
|
fsubrp %st, %st(1) /* 3 2^i - 1.0 */
|
|
|
|
fstp %st(1) /* 2 */
|
|
|
|
faddp %st, %st(1) /* 1 base^x - 1.0 */
|
|
|
|
#else
|
2012-05-05 21:34:31 +02:00
|
|
|
fld1 /* 4 1.0 */
|
2012-05-06 20:23:44 +02:00
|
|
|
faddp /* 3 2^(fract(x * log2(base))) */
|
2012-05-05 21:34:31 +02:00
|
|
|
fstp %st(1) /* 2 */
|
2012-05-06 20:23:44 +02:00
|
|
|
fscale /* 2 scale factor is st(1); base^x */
|
2012-05-05 21:34:31 +02:00
|
|
|
fstp %st(1) /* 1 */
|
2012-05-07 21:13:08 +02:00
|
|
|
#endif
|
2012-05-05 21:34:31 +02:00
|
|
|
fstp %st(1) /* 0 */
|
|
|
|
jmp 2f
|
2012-07-06 13:17:41 +02:00
|
|
|
1:
|
2012-05-07 21:13:08 +02:00
|
|
|
#ifdef USE_AS_EXPM1L
|
2012-07-06 13:17:41 +02:00
|
|
|
/* For expm1l, only negative sign gets here. */
|
|
|
|
fstp %st
|
2012-05-07 21:13:08 +02:00
|
|
|
fld1
|
|
|
|
fchs
|
|
|
|
#else
|
2012-07-06 13:17:41 +02:00
|
|
|
testl $0x200, %eax /* Test sign. */
|
|
|
|
jz 2f /* If positive, jump. */
|
|
|
|
fstp %st
|
2012-05-05 21:34:31 +02:00
|
|
|
fldz /* Set result to 0. */
|
2012-05-07 21:13:08 +02:00
|
|
|
#endif
|
2012-05-05 21:34:31 +02:00
|
|
|
2: ret
|
2012-05-06 20:23:44 +02:00
|
|
|
END(IEEE754_EXPL)
|
2012-05-07 21:13:08 +02:00
|
|
|
#ifdef USE_AS_EXPM1L
|
|
|
|
libm_hidden_def (__expm1l)
|
|
|
|
weak_alias (__expm1l, expm1l)
|
|
|
|
#else
|
2012-05-06 20:23:44 +02:00
|
|
|
strong_alias (IEEE754_EXPL, EXPL_FINITE)
|
2012-05-07 21:13:08 +02:00
|
|
|
#endif
|