Collection of remaining Win64 issues (1)

[kinke]

Current master (+ https://github.com/kinke/dmd-testsuite), LLVM 3.7, Visual Studio 2015, Win8 x64.

• std.string-unittest-debug leads to an ICE with the -g switch; Debug info: Fixup value out of range #930 (probably an LLVM bug)
• core.thread-unittest fails in release in line https://github.com/ldc-developers/druntime/blob/ldc/src/core/thread.d#L5467, a Win64-specific test. Debug is fine.
• runnable\cppa.d complains about multiply defined C++ functions. The functions are implemented in the C++ file and in the D file, so maybe weak linkage is expected. The linker error can be worked around by using REQUIRED_ARGS: -L/FORCE:MULTIPLE in the test file. Edit: fixed by Linkage type and COMDAT #1367
• runnable\test42.d fails in line 1652: casting the result of cast(real) ulong.max back to ulong yields 0. Edit: fixed by Fix a test in runnable/test42.d failing for 64-bit reals dlang/dmd#5158, already in LDC master
• runnable\ldc_extern_weak.d fails in line 10. Apparently weak-linkage related too, but I haven't digged into it yet. Edit: disabled
• runnable\opover2.d leads to an ICE with -inline -g[c]; LLVM 3.7: dmd-testsuite-debug ICE wrt. debug infos and inlined functions #1032 Edit: fixed in master
• runnable\sdtor.d leads to an ICE Edit: fixed by Properly mark stack slot for struct return as lvalue #1109
• fail_compilation\diag4479.d fails with [...] Edit: fixed by Windows: replace forward slashes in cmdline source files by backslashes #1105

The dmd-testsuite issues are identical for both debug and release.
Needless to say, it'd be awesome if we could fix all but the first issue for the 0.16 release. :)

[redstar]

Are all math-related errors fixed? I asked because on Linux/PPC64le the only remaining druntime+Phobos test failures are all math precision related.

[dnadlinger]

Just disable ldc_extern_weak on Windows; it's mostly an internal hack for working around a Linux linker bug.

[kinke]

@redstar: Yes, all math related tests pass without any further hacks. The PPC tests may fail due to a too restrictive feqrel() (counting the most-significant identical mantissa bits of the actual and expected floats), and Walter likes to keep them really tight, not making it easier for non-x86 platforms.
@klickverbot: Will do.

[kinke]

Down to 6 now.

[kinke]

runnable\opover2.d fails on Linux x64 too when using LLVM 3.7 and -inline -gc. I don't know why Travis doesn't fail; the test specifies PERMUTE_ARGS: -inline -O -property, and dmd-testsuite-debug adds DFLAGS=-gc.

[dnadlinger]

The reason might be that we are running the test suite using quick (like the DMD autotester, IIRC), which does not try the different argument permutations.

[kinke]

Makes perfect sense - thx for the heads up.

[rainers]

core.thread-unittest fails in release in line https://github.com/ldc-developers/druntime/blob/ldc/src/core/thread.d#L5467, a Win64-specific test. Debug is fine.

I just stepped through the optimized build, it fails the testNonvolatileRegister!("RDI") test, because RDI is used by some inlined code dragged in from Fiber.callImpl. Is llvm expected to preserve registers used in asm blocks within the same function? I doubt the test is valid. It might be ok if inlining of zeroRegister.call() is prohibited.

[kinke]

Thx for digging, Rainer.

Is llvm expected to preserve registers used in asm blocks within the same function? I doubt the test is valid. It might be ok if inlining of zeroRegister.call() is prohibited.

I'd vote for simply disabling that test for LDC release.

Btw, this list of 3 issues was for D 2.067. 2.068 added more tests wrt. exception handling as you know, so the missing exception chaining support in ldc.eh.win64 now manifests itself in some tests. That's why your MSVC EH work is also critical for Win64. With LLVM 3.8 final out in a few weeks, you got the timing perfectly right. 👍

[kinke]

Updated. The 2 remaining issues are negligible now that we have exception chaining etc.

[kinke]

core.thread-unittest has been working for months...