300633 - patch to fix javascript math functions on linux/alpha

Status: RESOLVED WORKSFORME

(Core :: JavaScript Engine, defect)

Description

[Aron Griffis]

User-Agent:       Mozilla/5.0 (X11; U; Linux alpha; en-US; rv:1.7.8) Gecko/20050711 Firefox/1.0.4
Build Identifier: Mozilla/5.0 (X11; U; Linux alpha; en-US; rv:1.7.8) Gecko/20050711 Firefox/1.0.4

The jslibmath functions don't work correctly on alpha.  I tested building both
with and without -mieee.  In both cases, log() and pow() (and possibly others)
miscalculate their answers.

Looking into jslibmath.h, I found the following snippet:

#elif defined(linux)

#define fd_atan atan
#define fd_atan2 atan2
#define fd_ceil ceil
#define fd_cos cos
#define fd_fabs fabs
#define fd_floor floor
#define fd_fmod fmod
#define fd_sin sin
#define fd_sqrt sqrt
#define fd_tan tan
#define fd_copysign copysign

extern double fd_asin __P((double));
extern double fd_acos __P((double));
extern double fd_exp __P((double));
extern double fd_log __P((double));
extern double fd_pow __P((double, double));

I don't know the history that led to using the mozilla-defined versions of the
latter functions, but with modern glibc I can't imagine it's the best decision.
 For alpha, I am using the following patch in the Gentoo ebuild:

--- mozilla/js/src/jslibmath.h.agriffis 2005-07-11 14:58:35 -0400
+++ mozilla/js/src/jslibmath.h  2005-07-11 15:09:13 -0400
@@ -69,6 +69,9 @@
 #elif defined(HPUX)
 #define JS_USE_FDLIBM_MATH 1
 
+#elif defined(linux) && defined(__alpha__)
+#define JS_USE_FDLIBM_MATH 0
+
 #elif defined(linux)
 #define JS_USE_FDLIBM_MATH 1

I restricted the change to alpha, where it's necessary for correct operation, so
that I wouldn't accidentally break other architectures.  However my testing
indicates it would be safe for amd64 and x86 at least.  Frankly it would
probably be fine to disable JS_USE_FDLIBM_MATH for all Linux platforms.

Additionally I ran through the Javascript test suite from
http://www.mozilla.org/js/tests/library.html

Before the change, 362 of 1605 tests failed on alpha.  After the change, 300 of
1605 tests failed.  Additional checking shows that the 362 are a superset of the
300, so that's a full 62 tests fixed by this single change on alpha.

I ran through the same test suite on amd64; there were zero changes as a result
of disabling JS_USE_FDLIBM_MATH.

Reproducible: Always

Steps to Reproduce:
unit test:

1. go to http://bluweb.com/us/chouser/logcheck.html
2. note the answer is incorrect when run on alpha

practical test:

1. go to maps.google.com
2. click on the "satellite" button
3. the javascript will chug for a while, then finally announce that it's taking
a long time, would you like to terminate the script?

Comment 1

[Kevin Brosnan]

The file that is mentioned in this bug bears little resemblance to the patch in comment 0. Patches should be attached to bugs and set with a review flag. Unfortunately most developers don't have the bandwidth to poke through bugs and find patches in comments. If this is still an issue on an Alpha a new bug would be best.


https://developer.mozilla.org/En/Developer_Guide/How_to_Submit_a_Patch