[META][Win64] build and test issues on MinGW-w64

[rubenvb]

Bugzilla Link	9072
Resolution	FIXED
Resolved on	Nov 17, 2015 11:48
Version	trunk
OS	Windows NT
Depends On	#9205 #9220 #9222 llvm/llvm-bugzilla-archive#9099
Blocks	llvm/llvm-bugzilla-archive#9100
Attachments	patch for the issues listed in the report
CC	@rnk

Extended Description

When compiling LLVM with MinGW-w64, several easy to fix/workaround issues popped up:

1. _WIN32_WINNT redefined
2. lib/support/Windows/DynamicLibrary.inc: bad ifdef for mingw-w64 which needs the function ifdef'ed out
3. explicit symbol declaration for functions not meant to be present on win64 MinGW platforms.
4. tblgen.exe crashes on CellSPU *.inc files

I have attached a patch for these issues (except 3). I would be much obliged if applied to trunk.

I do not have a clue about number 4, and already opened a bug report about this here: http://llvm.org/bugs/show_bug.cgi?id=8850
Disabling the target in question lets the build continue and finish happily.

I will update the bug report with "make check" results.

[llvmbot]

Ruben,

• Please split issues as possible.
• Don't leave a duplicated issue bug 8848.
• You may post patches to llvm-commits when you have both matters and solutions.

1. _WIN32_WINNT redefined

I prefer as below;

#ifdef _WIN32_WINNT
#if _WIN32_WINNT < 0x0500
#error
#endif
#else
#define _WIN32_WINNT 0x0500
...

How do you think? :)

2. lib/support/Windows/DynamicLibrary.inc: bad ifdef for mingw-w64 which needs
   the function ifdef'ed out

Would you like to refer to threads below?
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20110124/115720.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20110124/115726.html

As simply you hide a matter, I guess you would get test/ExeutionEngine failed.

I will work on this.

I will update the bug report with "make check" results.

How are your tests?

[rubenvb]

"mingw32-make check" results
I have made this the "main" bug for all important issues (ie disasterous evil crashes and miscompiles) with mingw-w64 (and 64-bit Windows as far as anything is encountered).

I hope this counts as "splitting and not leaving duplicate issues".

If not, I can reorganize my mess ;)

I understand that you (@nakamura) will try to sort the 32-bit symbols? I can't really help with anything code-related on that issue, but I can point you to mingw-w64-public@lists.sourceforge.net where (a lot) more knowledgeable people are. They are interested in getting LLVM/Clang working.

I have also attached the first test results, which seem rather awful :s. I did notice several output difference problems (like 0,00e+17 expected 0,000e+017 etc...) which might be due to incompatibility with the MSVC runtime.

[rubenvb]

"mingw32-make check" results
This time with "cat" and "uniq" in PATH. 164 failures.

[llvmbot]

I have made this the "main" bug for all important issues (ie disasterous evil
crashes and miscompiles) with mingw-w64 (and 64-bit Windows as far as anything
is encountered).

I hope this counts as "splitting and not leaving duplicate issues".

Good. I added a few testing issues.

FYI, on my host (x86_64-w64-mingw32(20101129), msys autoconf, python-2.7) with my local patches;

-- Testing: 5789 tests, 8 threads --
Testing: 0 .. 10.. 20.. 30.. 40.. 50.. 60.. 70.. 80.. 90..
Testing Time: 109.72s
Expected Passes : 5182
Expected Failures : 43
Unsupported Tests : 564

[rubenvb]

"mingw32-make check" LLVM results
Attached are new test results from LLVM r126497 built and tested with mingw-w64 4.5.2 toolchain built by sezero. Many tests now pass, but there are still 44 failures left. I noticed some are due to the difference in "printf" functionality (amount of digits in scientific notation output between the tested GNU libc output and msvcrt output).

[rubenvb]

"mingw32-make check" tools/clang results
Results from Clang tests, also under the same GCC 4.5.2 mingw-w64 compiler. Only 20 failures.

[rubenvb]

"mingw32-make check" results
Expected Passes : 8160
Expected Failures : 69
Unsupported Tests : 556
Unexpected Failures: 28

These are results for both LLVM (11 failures) and Clang (17 failures).

[rubenvb]

output from "mingw32-make check-all"
New "mingw32-make check-all" results. LLVM passes fully, Clang has some failures, a lot of them due to pointer->int conversions, which don't work well for Win64.

[rubenvb]

*** Bug llvm/llvm-bugzilla-archive#10729 has been marked as a duplicate of this bug. ***

[rnk]

Is this fixed?

[rubenvb]

[Test report 28https://user-images.githubusercontent.com/92601424/143742614-d16ccf35-87a3-4c05-a45d-8bea721881b0.gz)
I ran the tests on a fresh build of development tip with MSYS2's MinGW-w64 64-bit GCC 4.9.2 on Windows 8.1 x64.

97 failed tests for LLVM+Clang. See attachment for the output log.

[rnk]

I believe this is fixed. We've built with mingw64 for a long time. This bot provides proof that it works:
http://bb.pgr.jp/builders/ninja-clang-x64-mingw64-RA/

I don't think we need a meta bug for this now that it works. Any breakage can be filed separately.

[llvmbot]

mentioned in issue llvm/llvm-bugzilla-archive#9099

[llvmbot]

mentioned in issue llvm/llvm-bugzilla-archive#9100