phobos commit, revision 2252


user: braddr

msg:
Integrate changes from the d1 branch, still not enough to pass the std/math.d unit tests.

http://www.dsource.org/projects/phobos/changeset/2252

paths changed:
U   trunk/phobos/std/math.d

On 2 January 2011 00:26, dsource.org <noreply at dsource.org> wrote:
> phobos commit, revision 2252
>
>
> user: braddr
>
> msg:
> Integrate changes from the d1 branch, still not enough to pass the std/math.d unit tests.
>
> http://www.dsource.org/projects/phobos/changeset/2252
>
> paths changed:
> U ? trunk/phobos/std/math.d

Which test is failing?

On 1/1/2011 10:42 PM, Don Clugston wrote:
> On 2 January 2011 00:26, dsource.org <noreply at dsource.org> wrote:
>> phobos commit, revision 2252
>>
>>
>> user: braddr
>>
>> msg:
>> Integrate changes from the d1 branch, still not enough to pass the std/math.d unit tests.
>>
>> http://www.dsource.org/projects/phobos/changeset/2252
>>
>> paths changed:
>> U   trunk/phobos/std/math.d
> 
> Which test is failing?

I'll need to spend time to reduce it.  It's segv'ing and the dwarf debug info make gdb decidedly unhappy still.  Some possibly very inaccurate info suggests that it might be the cos unit tests which all seem to be imaginary number tests.

I also think there's more than one segv inducing problems in there.  I commented out the cos block and another test was implicated.  The stacks are iffy enough that I'm not terribly confident in this data.

I'll spend the time tomorrow to investigate and create some smaller standalone test cases.

At this point, anyone with an x86-64 box ought to be able to build and test things pretty easily.  No changes to how you build dmd, and for druntime and phobos, just add MODEL=64 to your make commands.  If running dmd manually, just add -m64.  So, if you'd like to dig in yourself, please, feel encouraged to do so. :)

More later,
Brad

Thanks. Sounds like stack corruption. (I don't have a usable x86-64 box unfortunately. Yes, I live in the stone age).

On 2 January 2011 08:09, Brad Roberts <braddr at puremagic.com> wrote:
> On 1/1/2011 10:42 PM, Don Clugston wrote:
>> On 2 January 2011 00:26, dsource.org <noreply at dsource.org> wrote:
>>> phobos commit, revision 2252
>>>
>>>
>>> user: braddr
>>>
>>> msg:
>>> Integrate changes from the d1 branch, still not enough to pass the std/math.d unit tests.
>>>
>>> http://www.dsource.org/projects/phobos/changeset/2252
>>>
>>> paths changed:
>>> U ? trunk/phobos/std/math.d
>>
>> Which test is failing?
>
> I'll need to spend time to reduce it. ?It's segv'ing and the dwarf debug info make gdb decidedly unhappy still. ?Some possibly very inaccurate info suggests that it might be the cos unit tests which all seem to be imaginary number tests.
>
> I also think there's more than one segv inducing problems in there. ?I commented out the cos block and another test was implicated. ?The stacks are iffy enough that I'm not terribly confident in this data.
>
> I'll spend the time tomorrow to investigate and create some smaller standalone test cases.
>
> At this point, anyone with an x86-64 box ought to be able to build and test things pretty easily. ?No changes to how you build dmd, and for druntime and phobos, just add MODEL=64 to your make commands. ?If running dmd manually, just add -m64. ?So, if you'd like to dig in yourself, please, feel encouraged to do so. :)
>
> More later,
> Brad
> _______________________________________________
> phobos mailing list
> phobos at puremagic.com
> http://lists.puremagic.com/mailman/listinfo/phobos
>

On 2 January 2011 08:09, Brad Roberts <braddr at puremagic.com> wrote:
> On 1/1/2011 10:42 PM, Don Clugston wrote:
>> On 2 January 2011 00:26, dsource.org <noreply at dsource.org> wrote:
>>> phobos commit, revision 2252
>>>
>>>
>>> user: braddr
>>>
>>> msg:
>>> Integrate changes from the d1 branch, still not enough to pass the std/math.d unit tests.
>>>
>>> http://www.dsource.org/projects/phobos/changeset/2252
>>>
>>> paths changed:
>>> U ? trunk/phobos/std/math.d
>>
>> Which test is failing?
>
> I'll need to spend time to reduce it. ?It's segv'ing and the dwarf debug info make gdb decidedly unhappy still. ?Some possibly very inaccurate info suggests that it might be the cos unit tests which all seem to be imaginary number tests.

It's segfaulting in exp, which is not surprising -- I haven't made a
64 asm version of exp.
Disabling all of the tests which use exp(), which includes sinh(),
cosh(), and complex sin and cos,
it fails a couple of other tests:  sqrt(2) for ubyte, and
nextUp(double.min_normal) fails.

I think there might be a compiler bug with the floating point equality
involving implicit casts -- not sure.
But almost all of the tests are passing. Obviously we're very close.

On 1/2/2011 2:43 PM, Don Clugston wrote:
> On 2 January 2011 08:09, Brad Roberts <braddr at puremagic.com> wrote:
>> On 1/1/2011 10:42 PM, Don Clugston wrote:
>>> On 2 January 2011 00:26, dsource.org <noreply at dsource.org> wrote:
>>>> phobos commit, revision 2252
>>>>
>>>>
>>>> user: braddr
>>>>
>>>> msg:
>>>> Integrate changes from the d1 branch, still not enough to pass the std/math.d unit tests.
>>>>
>>>> http://www.dsource.org/projects/phobos/changeset/2252
>>>>
>>>> paths changed:
>>>> U   trunk/phobos/std/math.d
>>>
>>> Which test is failing?
>>
>> I'll need to spend time to reduce it.  It's segv'ing and the dwarf debug info make gdb decidedly unhappy still.  Some possibly very inaccurate info suggests that it might be the cos unit tests which all seem to be imaginary number tests.
> 
> It's segfaulting in exp, which is not surprising -- I haven't made a
> 64 asm version of exp.
> Disabling all of the tests which use exp(), which includes sinh(),
> cosh(), and complex sin and cos,
> it fails a couple of other tests:  sqrt(2) for ubyte, and
> nextUp(double.min_normal) fails.
> 
> I think there might be a compiler bug with the floating point equality
> involving implicit casts -- not sure.
> But almost all of the tests are passing. Obviously we're very close.

Yup.. closer every day too. :)

Let me know if you want me to reduce any of those failures, otherwise I'll continue with the other parts.

Thanks,
Brad

Yeah, I'll take of std.math.

On 3 January 2011 00:04, Brad Roberts <braddr at puremagic.com> wrote:
> On 1/2/2011 2:43 PM, Don Clugston wrote:
>> On 2 January 2011 08:09, Brad Roberts <braddr at puremagic.com> wrote:
>>> On 1/1/2011 10:42 PM, Don Clugston wrote:
>>>> On 2 January 2011 00:26, dsource.org <noreply at dsource.org> wrote:
>>>>> phobos commit, revision 2252
>>>>>
>>>>>
>>>>> user: braddr
>>>>>
>>>>> msg:
>>>>> Integrate changes from the d1 branch, still not enough to pass the std/math.d unit tests.
>>>>>
>>>>> http://www.dsource.org/projects/phobos/changeset/2252
>>>>>
>>>>> paths changed:
>>>>> U ? trunk/phobos/std/math.d
>>>>
>>>> Which test is failing?
>>>
>>> I'll need to spend time to reduce it. ?It's segv'ing and the dwarf debug info make gdb decidedly unhappy still. ?Some possibly very inaccurate info suggests that it might be the cos unit tests which all seem to be imaginary number tests.
>>
>> It's segfaulting in exp, which is not surprising -- I haven't made a
>> 64 asm version of exp.
>> Disabling all of the tests which use exp(), which includes sinh(),
>> cosh(), and complex sin and cos,
>> it fails a couple of other tests: ?sqrt(2) for ubyte, and
>> nextUp(double.min_normal) fails.
>>
>> I think there might be a compiler bug with the floating point equality
>> involving implicit casts -- not sure.
>> But almost all of the tests are passing. Obviously we're very close.
>
> Yup.. closer every day too. :)
>
> Let me know if you want me to reduce any of those failures, otherwise I'll continue with the other parts.
>
> Thanks,
> Brad
> _______________________________________________
> phobos mailing list
> phobos at puremagic.com
> http://lists.puremagic.com/mailman/listinfo/phobos
>