[foreign] RFR: Windows support

[Maurizio Cimadamore]

Hi Jorn,
Thanks for the hard work; we are in the process of reviewing the 
changes. Here's what we'll do:

1) we will create issues for each of the patches (Sundar is doing that 
as we speak)
2) we will try to get easier changes in sooner rather  than later (that 
is, patches 1, 3, 4)
3) in parallel, I'll be doing an in-depth review of your binder changes 
and perhaps start a separate thread on that

Thanks
Maurizio


On 09/01/2019 21:17, Jorn Vernee wrote:

> Hi,
>
> I have fixed the last few failing tests on Windows so this is an RFR 
> for adding Windows support to Panama FFI. There's a lot of changes so 
> I've split it into 4 patches. We should probably focus on one of these 
> one at a time, but I wanted to include all the changes needed to pass 
> the tests at least. It's probably best to start with [2], since that 
> implements the core functionality.
>
> Default lib fixes [1] (needed for the default library to work on Windows)
> Hotspot, binder tweaks [2] (Changes to the stub generation and shared 
> binder code. Includes the Windows implementation of SystemABI)
> jextract tweaks [3] (minor jextract portability changes)
> test tweaks [4] (changes needed to run, and pass the tests on Windows)
>
> There are some tests that are disabled:
>   - EmptyStructTest; Empty structs are a GCC extension.
>   - BadBitfieldTest; this is testing for a problem the MSVC ABI 
> doesn't seem to have.
>   - StdioTest; relies on standard system header path, which on Windows 
> is more tricky to find.
>   - ComplexTest; Windows complex types work very differently in C 
> source, so the current test doesn't compile.
>   - LongDoubleTest; long double is only 64 bits on Windows. This might 
> need a replacement test still, but it should work the same as `double` 
> on Windows.
>   - Getpid; No getpid on Windows in the C standard.
>   - UnixSystem; Windows is not unix.
>
> (Some tests specifically for Windows could be added, but I'm leaving 
> that for a different RFR)
>
> Changes for [1]:
>   - replace default library handle with a findEntryInProcess since 
> Windows doesn't have an RTLD_DEFAULT equivalent.
>
> Changes for [2]:
>   - changed upcall/downcall stub generation to follow the MS ABI on 
> windows (using compiler switches)
>   - added SKIPs to shuffle recipe generation when e.g. the first float 
> is passed in the second float register (which can be the case on 
> windows).
>   - moved boxValue and unboxValue into SystemABI, since they turned 
> out to be ABI specific, and made tweaks to the windows version for 
> handeling structs being passed by value.
>   - moved upcall stub natives into shared ABI package.
>   - Let each ABI have it's own VarargsInvoker implementation, since we 
> need to propagate more info about which arguments are variadic on 
> windows.
>   - CallingSequence::storageOffset is no longer relying on 
> Storage::getStorageIndex, since that is not the same as what's 
> actually needed; being the index of the binding. (the 2 can differ on 
> windows, but not on sysv).
>   - Added Windows ABI implementation classes based on the SysV ones.
>   - Let UniversalUpcallHandler write the in-memory-return pointer to 
> RAX since that is required on Windows (and I think this change is 
> portable).
>
> Changes for [3]:
>   - instead of hardcoding the size of `long` `unsigned long` and `long 
> double`, delegate to libclang to tell us the size, since these types 
> have different sizes on the 2 ABIs. (This is pretty ad-hoc, so 
> probably needs to be improved. But, I have little experience with the 
> jextract code, so I'm open to suggestion of how to better fix this)
>
> Changes for [4]:
>   - added dll exports for native symbols using compiler switches.
>   - added test for small (fits-in-register) struct by-value passing, 
> since windows has special handling for those.
>   - removed depencies on POSIX functions from some tests where it 
> wasn't really needed, so they could run on windows.
>   - propagate clang library name to TestJextractFFI, since on Windows 
> we need to load "libclang".
>   - let Runner test use windows specific files to compare with for 
> bitfields.h and simple.h
>   - use `long long` instead of `long` in some cases since `long long` 
> is 64 bits on SysV and MSx64
>
> Thanks,
> Jorn
>
> [1] : 
> http://cr.openjdk.java.net/~jvernee/panama/webrevs/8211060/webrev.01/
> [2] : 
> http://cr.openjdk.java.net/~jvernee/panama/webrevs/windows/webrev.03/
> [3] : 
> http://cr.openjdk.java.net/~jvernee/panama/webrevs/windows_jextract/webrev.00/
> [4] : 
> http://cr.openjdk.java.net/~jvernee/panama/webrevs/windows_tests/webrev.02/