Add c++ ABI tests

[gchatelet]

@yebblies Here is a first simple version of C++ abi tests.
It fails for real numbers on Linux.

[gchatelet]

So real type compiled with -O seems to produce an invalid value when passing the value 0
See failing test.

[dnadlinger]

Your comparison includes the padding bytes, which are not guaranteed to be set to any particular value. (At least not by C/C++.)

[gchatelet]

@klickverbot I see, so for real values I should check only for the first 80 bits ?

[dnadlinger]

real is not always 80 bit once you leave DMD-land, but yeah, that's the idea. @ibuclaw can probably tell you exactly how he'd like to see this implemented. Can't you just use is, though?

[gchatelet]

is or == doesn't work with NaN since NaN==NaN is always false.
I think a better strategy is to use == expect when value is NaN.

[ibuclaw]

is will use memcmp at the correct precision for you (so you don't need to have logic in your own tests for that).

[dnadlinger]

@gchatelet: I'm pretty sure nan is nan.

[jmdavis]

@gchatelet: I'm pretty sure nan is nan.

Except that there are different variants of NaN, so two NaNs may not be the same via is. e.g. https://github.com/D-Programming-Language/phobos/blob/master/std/bitmanip.d#L2288

[yebblies]

Checking float equality with is is the correct thing to do, but be wary of https://issues.dlang.org/show_bug.cgi?id=3632

We really do want to be checking that the floats match bit for bit.

[gchatelet]

So as @jmdavis pointed out two NaNs might not have the same representation. Especially since the value is copied in C++ passthrough functions. The C++ code might choose a different representation for NaN.

Options are:

1. For robustness and also valid semantic I'll use std.math.isIdentical for floating point and is otherwise.
2. Do not use value NaN or 0 for floating points since they can have different binary representations.
3. Special case floating points on T.mant_dig as in std.math

I went for 1. right now. What do you think ?

[gchatelet]

BTW the Windows tests are failing, I suspect a name mangling issue.

[yebblies]

Phobos should not be used in the dmd test suite.

[gchatelet]

Please take another look. There is now no more dependencies on phobos.
Parameters are now passed: by value, by pointer and by reference.

[yebblies]

Looks good. Please squash the commits together.

[gchatelet]

Done. Thx for the review.

[dnadlinger]

Auto-merge toggled on

[dnadlinger]

The funny pointer type formatting is not my cup of tea, but LGTM.

[gchatelet]

thx @klickverbot . What do you mean by funny pointer ? Having the star next to the type instead of next to the variable ? I don't mind providing a fix if you want.

[dnadlinger]

It's not worth the effort, and Walter has been doing this from time to time too, but I'm really not a fan of having the start anywhere but next to the type in D.

[yebblies]

Having the star next to the type instead of next to the variable ?

Yes, that's correct C++ style but incorrect D style, since the association is different.

[dnadlinger]

This seems to have triggered a spurious failure in the "stage 3" GDC DDMD tests on Travis: https://travis-ci.org/D-Programming-Language/dmd/builds/80190427 ?!

[gchatelet]

Hmm why do you think this is linked to this PR, it looks like it's linked to druntime, not dmd tests.

make[2]: Entering directory `/home/travis/build/D-Programming-Language/druntime/test/profile'
Testing profile
make[2]: execvp: ./obj/linux/64/profile: Permission denied
make[2]: *** [obj/linux/64/profile.done] Error 127

[dnadlinger]

It seems improbable that the failure would be related to this. I've invalidated the test run; let's see whether it comes up again.