python rounding errors fix

[its-pointless]

fixes #2236

[tomty89]

Is there any reason for you not to patch pyport.h directly (add #undef PY_NO_SHORT_FLOAT_REPR after the two lines where it could be defined)?

From your PR I assume only the first one is of our concern:
https://github.com/python/cpython/blob/3.6/Include/pyport.h#L476
For that there are three possible cases that could avoid the case, what makes you think the little endian one is appropriate for all possible platforms (especially for arm and arm64)?

There's a reason that none of the three is defined for cross-compiling, because tests are performed in configure when built on site, which can't be done for the target platform when cross-compile.

You should patch configure.ac instead even if this is necessary and appropriate:
https://github.com/python/cpython/blob/3.6/configure.ac#L4272

[tomty89]

Or configure if we don't regenerate it:
https://github.com/python/cpython/blob/3.6/configure#L13978

[tomty89]

Also python2?

[its-pointless]

1 python2 doesn't even attempt it. seriously 1/10 is 0. Even on normal linux pc. luckly this has nothing to do with the actual calculations going on.
2. i was under the impression that DOUBLE_IS_LITTLE_ENDIAN_IEEE754 on android. ...

[tomty89]

https://github.com/python/cpython/blob/2.7/Include/pyport.h#L611

Also:

$ python2
Python 2.7.14 (default, Jan  7 2018, 04:00:23)
[GCC 4.2.1 Compatible Android Clang 5.0.300080 ] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> 1/10
0
>>> 1.0/10.0
0.10000000000000001
>>> 1/10.0
0.10000000000000001
>>> 1.0/10
0.10000000000000001
>>>

You need to make at least one of the numbers a float if you want to perform a float calculation in python2. For the record:

$ python
Python 3.6.4 (default, Jan  7 2018, 03:52:16)
[GCC 4.2.1 Compatible Android Clang 5.0.300080 ] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> 8/4
2.0
>>>
$ python2
Python 2.7.14 (default, Jan  7 2018, 04:00:23)
[GCC 4.2.1 Compatible Android Clang 5.0.300080 ] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> 8/4
2
>>>

It has nothing to do with the OS but CPU I think.

DOUBLE_IS_LITTLE_ENDIAN_IEEE754 should be fine on most cases, it's just I wonder if there could be any corner ones that Termux might be serving, and whether the is a necessity to define it instead of enforcing short repr directly.

[its-pointless]

Would it differ vs gcc?
it would have to do with the os if the vendor of the os stipulates behavior for vendors to adhere to. We use the ndk as its a standard. I wonder how many other packages assume a standard from cpus?

[tomty89]

I don't think the compiler has anything to do with it.

And I don't know. The thing is the question stands anyway: we don't we just undefine PY_NO_SHORT_FLOAT_REPR directly instead? It will be less confusing as well. Does it not work?

And if that's the case, we will go for defining DOUBLE_IS_LITTLE_ENDIAN_IEEE754 by patching configure. Simple, isn't it?

[fornwall]

So, it seems that ac_cv_little_endian_double=yes is enough? This is applied in 8dc5d78 which is applied in the latest package versions (3.6.5-1 for python, and 2.7.14-3 for python2). Seems to work fine for all arches, right?

[its-pointless]

it should since java uses ieee754 and from jni.h
/* Primitive types that match up with Java equivalents. /
typedef uint8_t jboolean; / unsigned 8 bits /
typedef int8_t jbyte; / signed 8 bits /
typedef uint16_t jchar; / unsigned 16 bits /
typedef int16_t jshort; / signed 16 bits /
typedef int32_t jint; / signed 32 bits /
typedef int64_t jlong; / signed 64 bits /
typedef float jfloat; / 32-bit IEEE 754 /
typedef double jdouble; / 64-bit IEEE 754 */

[tomty89]

IEEE 754 does not specify endianness, that's why there are three DOUBLE_IS_..._ENDIAN_IEEE754.

I think the question is more like, whether assuming support of any of the variants has effect other than preventing PY_NO_SHORT_FLOAT_REPR to be defined.

That's why if we want to avoid unexpected side effect, patching pyport.h should be our first approach, and only if that doesn't work, we assume little endian double, which should work fine on most platforms.

But whatever the patch is up and it's probably the easiest way anyway, so who cares, right?

[its-pointless]

ieee754 doesn;t specify endianess but android does. its always little endian.

[tomty89]

If you say so.

[its-pointless]

No i like tomty89 being around even if disagreements occur as long as they are about the what is best technically its all good. Disagreement is fine if the end point is better. So far that has been the case.

[tomty89]

Should I feel lucky that I didn't also delete my reply to your quoted reply:
SDRausty/TermuxArch#54 (comment)
Coz apparently others deleting their comments could drive you mad. How does it matter that I deleted a few of my broken ideas?

I don't have broad knowledge but yes I am picky even in that case, and I do think that that pickiness can be a form of contribution.

And since when I do personal attacks? (Does attacking someone's grammar count? Either way that's not what I did) I mean, can I even do that? Like I pretty much know nothing about any of you guys other than your code.

Can anyone explain what is technically correct about my reply to a non-existent comment?

Yes I can. Because GitHub provides a delete button to whoever made specific comments, so yes, it's technically (and logically) correct that they could vanish at any point. Now would you just stop harassing others just because you think I have an "attitude"? File an issue to my parents or primary school teachers, if you can.

[Grimler91]

@SDRausty enough already. This pull request is about rounding errors in python and I'm not really not sure what you are trying to achieve here.

I'll delete additional aggressive posts (from any source).

[ghost]

I am not happy with these actions either

@SDRausty No one is happy there when you posting aggressive offtopic comments, especially if they are posted in pull request closed 2 weeks ago.

what do you do when your projects are attacked?

Github has a special button that allows to block user. More info about this: Github Help.