RFR: 8375285: Port fdlibm asinh to Java [v6]
Raffaello Giulietti
rgiulietti at openjdk.org
Tue Feb 3 14:45:38 UTC 2026
On Thu, 29 Jan 2026 12:01:46 GMT, Anton Artemov <aartemov at openjdk.org> wrote:
>> Hi, please consider the following changes:
>>
>> This is a port of FDLIBM asinh method.
>>
>> Original C vs transliteration port:
>>
>>
>> $ diff -w fdlib_asinh.c Asinh.translit.java
>> 1c1,3
>> < /* asinh(x)
>> ---
>>> /**
>>> * Return the Inverse Hyperbolic Sine of x
>>> *
>> 2a5
>>> *
>> 10a14,17
>>> private static final class Asinh {
>>> private static final double one = 1.0;
>>> private static final double ln2 = 6.93147180559945286227e-01;
>>> private static final double huge = 1.0e300;
>> 12,29c19
>> < #include "fdlibm.h"
>> <
>> < #ifdef __STDC__
>> < static const double
>> < #else
>> < static double
>> < #endif
>> < one = 1.00000000000000000000e+00, /* 0x3FF00000, 0x00000000 */
>> < ln2 = 6.93147180559945286227e-01, /* 0x3FE62E42, 0xFEFA39EF */
>> < huge= 1.00000000000000000000e+300;
>> <
>> < #ifdef __STDC__
>> < double asinh(double x)
>> < #else
>> < double asinh(x)
>> < double x;
>> < #endif
>> < {
>> ---
>>> static double compute(double x) {
>> 39c29
>> < w = __ieee754_log(fabs(x))+ln2;
>> ---
>>> w = log(Math.abs(x))+ln2;
>> 41,42c31,32
>> < t = fabs(x);
>> < w = __ieee754_log(2.0*t+one/(sqrt(x*x+one)+t));
>> ---
>>> t = Math.abs(x);
>>> w = log(2.0*t+one/(sqrt(x*x+one)+t));
>> 45c35
>> < w =log1p(fabs(x)+t/(one+sqrt(one+t)));
>> ---
>>> w =log1p(Math.abs(x)+t/(one+sqrt(one+t)));
>> 47a38
>>> }
>>
>> Transliteration port vs more idiomatic port:
>>
>>
>> $ diff -w Asinh.translit.java Asinh.fdlibm.java
>> 6,9c6,12
>> < * Based on
>> < * asinh(x) = sign(x) * log [ |x| + sqrt(x*x+1) ]
>> < * we have
>> < * asinh(x) := x if 1+x*x=1,
>> ---
>>> *
>>> * asinh(x) is defined so that asinh(sinh(alpha)) = alpha -INF < alpha < < INF
>>> * and sinh(asinh(x)) = x, -INF < x < INF.
>>> * It can be written as asinh(x) = ln(x + sqrt(x^2 + 1)), -INF < x < INF.
>>> * 1. Replace x by |x| as the function is odd.
>>> * 2.
>>> * asinh(x) := x, if 1+x^2 = 1,
>> 12a16,22
>>> *
>>> *
>>> *
>>> * Special cases:
>>> * only asinh(0)=0 is exact for finite x.
>>> * asinh(NaN) is NaN
>>> * asinh(INF) is INF
>> 14,15c24
>> < private static final class Asinh {
>> < private static final double one = 1.0;
>> ---
>>> static final class Asinh {
>> 24c33,35
>> < if(ix>=0x7ff00000) return x+x; /* x is inf or NaN */
>> ---
>>> if(ix >= 0x7ff00000) {
>>> return x...
>
> Anton Artemov has updated the pull request incrementally with one additional commit since the last revision:
>
> 8375285: Added missed link in documentation.
Will review the test code in another round.
src/java.base/share/classes/java/lang/FdLibm.java line 3518:
> 3516: *
> 3517: *
> 3518: * asinh(x) is defined so that asinh(sinh(alpha)) = alpha, -INF < alpha < < INF
Suggestion:
* asinh(x) is defined so that asinh(sinh(alpha)) = alpha, -∞ < alpha < ∞
∞ looks better
< avoids confusing an HTML parser (not sure if it really needed, though)
src/java.base/share/classes/java/lang/FdLibm.java line 3519:
> 3517: *
> 3518: * asinh(x) is defined so that asinh(sinh(alpha)) = alpha, -INF < alpha < < INF
> 3519: * and sinh(asinh(x)) = x, -INF < x < INF.
Similarly
src/java.base/share/classes/java/lang/FdLibm.java line 3520:
> 3518: * asinh(x) is defined so that asinh(sinh(alpha)) = alpha, -INF < alpha < < INF
> 3519: * and sinh(asinh(x)) = x, -INF < x < INF.
> 3520: * It can be written as asinh(x) = ln(x + sqrt(x^2 + 1)), -INF < x < INF.
This is not the formulation in the original fdlibm code.
src/java.base/share/classes/java/lang/FdLibm.java line 3531:
> 3529: *
> 3530: * Special cases:
> 3531: * only asinh(0)=0 is exact for finite x.
Suggestion:
* only asinh(±0) = ±0 is exact for finite x.
src/java.base/share/classes/java/lang/FdLibm.java line 3533:
> 3531: * only asinh(0)=0 is exact for finite x.
> 3532: * asinh(NaN) is NaN
> 3533: * asinh(INF) is INF
Suggestion:
* asinh(±∞) = ±∞
src/java.base/share/classes/java/lang/FdLibm.java line 3543:
> 3541: int hx, ix;
> 3542: hx = __HI(x);
> 3543: ix = hx & 0x7fffffff;
Suggestion:
ix = hx & 0x7fff_ffff;
as well as other hexadecimal literals below. It's easier for counting whether the number of hex digits is correct.
src/java.base/share/classes/java/lang/Math.java line 2765:
> 2763: * The inverse hyperbolic sine of <i>x</i> is defined to be a function such that
> 2764: * asinh({@linkplain Math#sinh sinh(<i>x</i>)}) = <i>x</i> for any <i>x</i>.
> 2765: * Note that range of the exact asinh is not limited.
Suggestion:
* Note that domain of the exact asinh is unrestricted.
src/java.base/share/classes/java/lang/Math.java line 2777:
> 2775: * <li>If the argument is negative infinity, then the result is
> 2776: * negative infinity.
> 2777: *
It is possible to collapse both items into one, as is done for the zero case.
src/java.base/share/classes/java/lang/StrictMath.java line 2197:
> 2195: * @return The inverse hyperbolic sine of {@code x}.
> 2196: * @since 27
> 2197: */
See the comment for Math
-------------
PR Review: https://git.openjdk.org/jdk/pull/29273#pullrequestreview-3745650028
PR Review Comment: https://git.openjdk.org/jdk/pull/29273#discussion_r2759404605
PR Review Comment: https://git.openjdk.org/jdk/pull/29273#discussion_r2759404841
PR Review Comment: https://git.openjdk.org/jdk/pull/29273#discussion_r2759405044
PR Review Comment: https://git.openjdk.org/jdk/pull/29273#discussion_r2759405227
PR Review Comment: https://git.openjdk.org/jdk/pull/29273#discussion_r2759405470
PR Review Comment: https://git.openjdk.org/jdk/pull/29273#discussion_r2759405711
PR Review Comment: https://git.openjdk.org/jdk/pull/29273#discussion_r2759415326
PR Review Comment: https://git.openjdk.org/jdk/pull/29273#discussion_r2759415580
PR Review Comment: https://git.openjdk.org/jdk/pull/29273#discussion_r2759419099
More information about the core-libs-dev
mailing list