glibc/sysdeps/x86_64/fpu/e_expl.S
Joseph Myers 5b0626b9c5 Fix x86 / x86_64 expl / expl10l wild results in directed rounding modes (bug 16356).
This patch fixes bug 16356, bad results from x86 / x86_64 expl /
exp10l in directed rounding modes, the most serious of the bugs shown
up by my patch expanding libm test coverage.  When I fixed bug 16293,
I thought it was only necessary to set round-to-nearest when using
frndint in expm1 functions, because in other cases the cancellation
error from having the resulting fractional part close to 1 or -1 would
not be significant.  However, in expl and exp10l, the way the final
fractional part gets computed (something more complicated than a
simple subtraction, because more precision is needed than you'd get
that way) can result in a value outside the range [-1, 1] when the
argument to frndint was very close to an integer and was rounded the
"wrong" way because of the rounding mode - and the f2xm1 instruction
has undefined results if its argument is outside [-1, 1], so resulting
in the large errors seen.  So this patch removes the USE_AS_EXPM1L
conditionals on the round-to-nearest settings, so all of expl, expm1l
and exp10l now get round-to-nearest used for frndint (meaning the
final fractional part can at most be slightly above 0.5 in
magnitude).  Associated tests of exp and exp10 are added and testing
of exp10 in directed rounding modes enabled.

Tested x86_64 and x86 and ulps updated accordingly.

	* sysdeps/i386/fpu/e_expl.S (IEEE754_EXPL): Also set
	round-to-nearest for [!USE_AS_EXPM1L].
	* sysdeps/x86_64/fpu/e_expl.S (IEEE754_EXPL): Likewise.
	* math/auto-libm-test-in: Do not expect cosh tests to fail.  Add
	more tests of exp and exp10.  Expect some exp10 tests to miss
	exceptions or fail in directed rounding modes.
	* math/auto-libm-test-out: Regenerated.
	* math/libm-test.inc (exp10_tonearest_test_data): New array.
	(exp10_test_tonearest): New function.
	(exp10_towardzero_test_data): New array.
	(exp10_test_towardzero): New function.
	(exp10_downward_test_data): New array.
	(exp10_test_downward): New function.
	(exp10_upward_test_data): New array.
	(exp10_test_upward): New function.
	(main): Call the new functions.
	* sysdeps/i386/fpu/libm-test-ulps: Update.
	* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
2013-12-21 13:07:16 +00:00

192 lines
5.0 KiB
ArmAsm

/*
* 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>
#ifdef USE_AS_EXP10L
# define IEEE754_EXPL __ieee754_exp10l
# define EXPL_FINITE __exp10l_finite
# define FLDLOG fldl2t
#elif defined USE_AS_EXPM1L
# define IEEE754_EXPL __expm1l
# undef EXPL_FINITE
# define FLDLOG fldl2e
#else
# define IEEE754_EXPL __ieee754_expl
# define EXPL_FINITE __expl_finite
# define FLDLOG fldl2e
#endif
.section .rodata.cst16,"aM",@progbits,16
.p2align 4
#ifdef USE_AS_EXP10L
.type c0,@object
c0: .byte 0, 0, 0, 0, 0, 0, 0x9a, 0xd4, 0x00, 0x40
.byte 0, 0, 0, 0, 0, 0
ASM_SIZE_DIRECTIVE(c0)
.type c1,@object
c1: .byte 0x58, 0x92, 0xfc, 0x15, 0x37, 0x9a, 0x97, 0xf0, 0xef, 0x3f
.byte 0, 0, 0, 0, 0, 0
ASM_SIZE_DIRECTIVE(c1)
#else
.type c0,@object
c0: .byte 0, 0, 0, 0, 0, 0, 0xaa, 0xb8, 0xff, 0x3f
.byte 0, 0, 0, 0, 0, 0
ASM_SIZE_DIRECTIVE(c0)
.type c1,@object
c1: .byte 0x20, 0xfa, 0xee, 0xc2, 0x5f, 0x70, 0xa5, 0xec, 0xed, 0x3f
.byte 0, 0, 0, 0, 0, 0
ASM_SIZE_DIRECTIVE(c1)
#endif
#ifndef USE_AS_EXPM1L
.type csat,@object
csat: .byte 0, 0, 0, 0, 0, 0, 0, 0x80, 0x0e, 0x40
.byte 0, 0, 0, 0, 0, 0
ASM_SIZE_DIRECTIVE(csat)
#endif
#ifdef PIC
# define MO(op) op##(%rip)
#else
# define MO(op) op
#endif
.text
ENTRY(IEEE754_EXPL)
#ifdef USE_AS_EXPM1L
movzwl 8+8(%rsp), %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
fldt 8(%rsp)
/* 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 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
movzwl 8+8(%rsp), %eax
andl $0x7fff, %eax
cmpl $0x400d, %eax
jle 3f
/* Overflow, underflow or infinity or NaN as argument. */
fstsw %ax
movb $0x45, %dh
andb %ah, %dh
cmpb $0x05, %dh
je 1f /* Is +-Inf, jump. */
cmpb $0x01, %dh
je 2f /* Is +-NaN, jump. */
/* Overflow or underflow; saturate. */
fstp %st
fldt MO(csat)
andb $2, %ah
jz 3f
fchs
#endif
3: FLDLOG /* 1 log2(base) */
fmul %st(1), %st /* 1 x log2(base) */
/* Set round-to-nearest temporarily. */
fstcw -4(%rsp)
movl $0xf3ff, %edx
andl -4(%rsp), %edx
movl %edx, -8(%rsp)
fldcw -8(%rsp)
frndint /* 1 i */
fld %st(1) /* 2 x */
frndint /* 2 xi */
fldcw -4(%rsp)
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 */
f2xm1 /* 3 2^(fract(x * log2(base))) - 1 */
#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
fld1 /* 4 1.0 */
faddp /* 3 2^(fract(x * log2(base))) */
fstp %st(1) /* 2 */
fscale /* 2 scale factor is st(1); base^x */
fstp %st(1) /* 1 */
#endif
fstp %st(1) /* 0 */
jmp 2f
1:
#ifdef USE_AS_EXPM1L
/* For expm1l, only negative sign gets here. */
fstp %st
fld1
fchs
#else
testl $0x200, %eax /* Test sign. */
jz 2f /* If positive, jump. */
fstp %st
fldz /* Set result to 0. */
#endif
2: ret
END(IEEE754_EXPL)
#ifdef USE_AS_EXPM1L
libm_hidden_def (__expm1l)
weak_alias (__expm1l, expm1l)
#else
strong_alias (IEEE754_EXPL, EXPL_FINITE)
#endif