On 30-Oct-08, at 5:55 AM, Christian Seiler wrote:
1) Internal FPU precision on x86.
Do you have any test cases show the error?
2) Specification problem (which rounding mode should actually be
used?)
3) Dividing/multiplying by >= 10^23 is not exact.
4) round (1.255, 2) should give 1.26 but gives 1.25. The FUZZ stuff
tries to resolve this issue (but not the other three) but I've
come to the conclusion that the FUZZ is actually the wrong solution
for the problem.
The current implementation does return 1.26
The proposed solution seems fairly complex and wide-reaching to me, I
am also concerned with the overheads it introduces as that was a
problem with many in-C implementation libc folks have tried and
rejected.
Since I didn't get any reaction on the RFC on internals, I pinged
Lukas
and Johannes (as they are RMs for PHP 5.3) in private in order to ask
whether it was possible to include my proposal in 5.3 (I don't have
ZE2
Karma and my patch also needs to change zend_strtod).
Lukas and Johannes were concerned about the interopability of my
solution of problem (1). So I did some tests on different platforms
and
operating systems and Lukas and Johannes asked around for other people
to do tests, too. I've summarized results of these tests and proposed
solution for correct cross-platform floating point arithmetics here:
http://www.christian-seiler.de/projekte/fpmath/
I've mailed patches for PHP 5.3 and HEAD to Johannes and Lukas for
ZendEngine2 that only address the above issue (1) and do the
following:
1) Define some macros for math-related functions that will ensure the
function itself always uses double precision. Add configure checks for
these macros.
2) Modified zend_strtod and the add/sub/div/mul functions in
zend_operators.c to make use of those macros. This ensures that PHP
will
always use double precision for calculations and math is thus
portable.
3) Added a test that determines if the calculations precision is
actually correct.
The patches for 5.3 and HEAD are found here:
http://www.christian-seiler.de/temp/php/2008-10-28-fpu/php-float-precision-5.3.patch
http://www.christian-seiler.de/temp/php/2008-10-28-fpu/php-float-precision-6.patch
My next step (as discussed with Johannes and Lukas) would have been to
apply the Non-ZE2-part of my patch to ext/standard/math.c in 5.3 and
HEAD (I don't have separate patches for that yet but they are quite
trivial to adapt to the new macros).
Now the question is: Where do we go from here? Your commit does not
solve all the problems of PHP's round but is at least better than the
previous implementation since at least the platform-dependency on
whether or not to use the FUZZ is removed. The other problems,
however,
remain.
Christian
Ilia Alshanetsky
--
PHP Internals - PHP Runtime Development Mailing List
To unsubscribe, visit: http://www.php.net/unsub.php
