Fix Win64 ABI wrt. passing structs > 64 bit

[kinke]

This fixes #754.

[redstar]

Is there a reason to not call isPassedWithByvalSemantics()? The sole effect is that the byval attribute is set (which has no effect because the argument is rewritten). The other ABI implementations use the same approach.

[kinke]

No particular reason, I just wasn't sure whether byval would really be ignored if the argument is rewritten.

[kinke]

Now we're down to 63 errors in druntime/phobos (release and debug) on Win64:

core.thread (SEGFAULT)
core.time (SEGFAULT)
rt.lifetime (SEGFAULT)
rt.minfo (SEGFAULT)
std.algorithm (SEGFAULT)
std.array (SEGFAULT)
std.complex (Failed)
std.concurrency (SEGFAULT)
std.conv (SEGFAULT)
std.csv (Failed)
std.datetime (SEGFAULT)
std.exception (SEGFAULT)
std.getopt (SEGFAULT)
std.math (Failed)
std.numeric (Failed)
std.outbuffer (Failed)
std.parallelism (SEGFAULT)
std.path (SEGFAULT)
std.process (SEGFAULT)
std.regex (SEGFAULT)
std.socket (SEGFAULT)
std.stdio (SEGFAULT)
std.string (SEGFAULT)
std.uni (SEGFAULT)
std.uri (SEGFAULT)
std.variant (Failed)
std.xml (SEGFAULT)
std.zlib (SEGFAULT)
std.container.array (SEGFAULT)
std.net.isemail (SEGFAULT)
std.internal.math.errorfunction (Failed)
std.internal.math.gammafunction (Failed)
core.time-debug (SEGFAULT)
rt.lifetime-debug (SEGFAULT)
rt.minfo-debug (SEGFAULT)
std.algorithm-debug (SEGFAULT)
std.array-debug (SEGFAULT)
std.complex-debug (Failed)
std.concurrency-debug (SEGFAULT)
std.conv-debug (SEGFAULT)
std.csv-debug (Failed)
std.datetime-debug (SEGFAULT)
std.exception-debug (SEGFAULT)
std.getopt-debug (SEGFAULT)
std.json-debug (Failed)
std.math-debug (Failed)
std.numeric-debug (Failed)
std.outbuffer-debug (Failed)
std.parallelism-debug (SEGFAULT)
std.path-debug (Failed)
std.process-debug (SEGFAULT)
std.regex-debug (SEGFAULT)
std.socket-debug (SEGFAULT)
std.stdio-debug (SEGFAULT)
std.uni-debug (SEGFAULT)
std.uri-debug (SEGFAULT)
std.variant-debug (SEGFAULT)
std.xml-debug (SEGFAULT)
std.zlib-debug (SEGFAULT)
std.container.array-debug (SEGFAULT)
std.net.isemail-debug (SEGFAULT)
std.internal.math.errorfunction-debug (Failed)
std.internal.math.gammafunction-debug (Failed)

I.e., this fixes 15-20 errors. :)

[redstar]

Thanks!

[Trass3r]

Did you build with assertions enabled?

[kinke]

Yes I did, getting the same problem now. It's due to the byval attribute, see the outdated line diff comments. Simply returning false in passByVal() works.

[Trass3r]

Hmm ok first I thought it's cause of the llvm update but returning false makes the assertions go away indeed.

[kinke]

Hotfix: #757

[kinke]

@Trass3r Note that you'll run into ldc-developers/druntime@2c2c8ca#commitcomment-8248875 as well. Simply remove the INTRINSICS_FROM_302 version guard for the moment.

[Trass3r]

Do you also get empty "application error" message boxes with the crashing unittests?

[kinke]

Hmm, I get the normal app-crashed Windows dialog on segfault (with the option to close or debug it) and 2-3 message boxes from the MSVCRT ("abort() called" iirc). Is that what you mean?

[Trass3r]

With the former assertions in ldc I get the usual close or debug dialog. But the test runner crashes suddenly (?) only produce a messagebox with title Application Error, an icon, no description and no text on the ok button. Maybe some /MT weirdness.

[kinke]

Good guess, I've seen some warnings about /MT overriding /MTd or something like that. I guess your weird dialog should be the MSVCRT dialog about abort().
/edit: Although I've seen dialogs such as these in case one runs out of memory. Note that the 6 parallel LDC2 processes using ninja (on my system: 4 cores + 2 extra tasks) can eat up quite huge amounts, iirc, std.algorithm-unittest alone takes ~3.5 gigs.
And the build time for the tests has increased considerably (+60% or so, especially for phobos-unittest in release mode) since real == double (I don't recall since when exactly). I guess it may have to do with CTFE which was previously blocked by inline assembly in std.math.

[Trass3r]

Yeah one problem was I modified the CXX flags for static C runtime, but not the C flags.
The other one was the curl library.

[Trass3r]

Also see #758