Support for ARM64 Windows

Summary

I was wondering whether it would be possible to build an ARM64 Windows GHC. One thing I'm not immediately understanding is how do we bootstrap such a compiler? Do I need to build a cross compiler targeting arm64 windows to then build a native arm64 windows compiler?

This came up as I was trying to debug a windows issue with virtualization on my aarch64 M2 machine, and was using an aarch64 virtualized windows for performance... Emulating a x64 windows on arm is much more expensive unfortunately.

added needs triage label

changed the description

AArch64/Windows is currently not supported. It's not impossible to support, but it also is both non-trivial and not a high priority. The steps involved are essentially:

• Teach ghc-toolchain, configure, and the compiler about the existence of the platform
• Add STG register assignments to MachRegs
• Optional: Upstream the STG register assignment to LLVM
• Teach the AArch64 NCG about the calling convention
• Add PE/AArch64 support to the RTS linker

Sady I don't see progress being made here unless someone steps up to take ownership of the platform.

removed needs triage label

added Tbug Windows aarch64 labels

added Tfeature request label and removed Tbug label

It seems Windows is starting to push Arm64 a bit harder. For example they are offering Arm-based surface notebooks now: https://www.theverge.com/2024/5/20/24160498/microsoft-arm-surface-laptop-6-qualcomm-snapdragon-x-elite

So we might want to prioritize this somewhat higher as those become more common.

changed milestone to %9.14.1

https://github.com/mstorsjo/llvm-mingw provides aarch64-w64-mingw32- prefixed llvm toolchain that can compile to aarch64 windows native binaries on linux. So a good first step would be trying to get a Haskell hello world working via unreg backend, then trying to bootstrap a non-threaded unreg GHC native build. The native build can use clangarm64 packages provided by msys2 and later developments can happen on windows natively.

Do you think this could reasonably be done by a beginner? Maybe it'd make for an interesting task at ZuriHac

Only for the work that leads to a hello world but not anything beyond that. Also if you want to get people at zurihac working for this, do make sure they have a arm laptop with windows vm, because wine support for arm64 is still in its infancy.

OK, maybe I won't

It seems arm on windows is documented here: https://learn.microsoft.com/en-us/windows/arm/overview

Update: after applying this patch, I've managed to cross-compile the ghc api hello world to aarch64 executable and actually run it on aarch64 windows:

For anyone motivated enough to give it a try:

LLVM_MINGW_PATH=/workspace/llvm-mingw-20240518-ucrt-ubuntu-20.04-x86_64
export AR=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-llvm-ar
export CC=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-clang
export CXX=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-clang++
export NM=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-nm
export OBJCOPY=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-objcopy
export OBJDUMP=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-objdump
export RANLIB=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-llvm-ranlib
export SIZE=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-size
export STRINGS=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-strings
export STRIP=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-strip
export WindresCmd=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-windres
export CONFIGURE_ARGS="--host=x86_64-unknown-linux --target=aarch64-unknown-mingw32 --enable-unregisterised --with-intree-gmp"

hadrian/build --bignum=native

Just signalling of successful reproduction!

I faced with:

d.lld: error: undefined symbol: ___chkstk_ms

I was needed to remove the following line:

       RTS_WIN64_ONLY(SymI_NeedsProto(___chkstk_ms))      \

at rts/RtsSymbols.c.

It compiled for ARM64:

serge@DESKTOP-VSP1KUB CLANGARM64 /c/Users/serge/Workspace
$ file ./T22455.exe
./T22455.exe: PE32+ executable for MS Windows 6.00 (console), ARM64, 15 sections

Also I was needed to disable windows security to let it run because it had been treated falsy positive as a backdoor ))

My diff:

diff --git a/libraries/base/src/System/CPUTime/Windows.hsc b/libraries/base/src/System/CPUTime/Windows.hsc
index 547e7fa480..4bae075d46 100644
--- a/libraries/base/src/System/CPUTime/Windows.hsc
+++ b/libraries/base/src/System/CPUTime/Windows.hsc
@@ -60,7 +60,7 @@ type HANDLE = ()
 #if defined(i386_HOST_ARCH)
 foreign import stdcall unsafe "GetCurrentProcess" getCurrentProcess :: IO (Ptr HANDLE)
 foreign import stdcall unsafe "GetProcessTimes" getProcessTimes :: Ptr HANDLE -> Ptr FILETIME -> Ptr FILETIME -> Ptr FILETIME -> Ptr FILETIME -> IO CInt
-#elif defined(x86_64_HOST_ARCH)
+#elif defined(x86_64_HOST_ARCH) || defined(aarch64_HOST_ARCH)
 foreign import ccall unsafe "GetCurrentProcess" getCurrentProcess :: IO (Ptr HANDLE)
 foreign import ccall unsafe "GetProcessTimes" getProcessTimes :: Ptr HANDLE -> Ptr FILETIME -> Ptr FILETIME -> Ptr FILETIME -> Ptr FILETIME -> IO CInt
 #else
diff --git a/rts/RtsSymbols.c b/rts/RtsSymbols.c
index d4ce9acb94..56e8c58faa 100644
--- a/rts/RtsSymbols.c
+++ b/rts/RtsSymbols.c
@@ -84,7 +84,7 @@ extern char **environ;
 #if defined(mingw32_HOST_OS)
 #define RTS_POSIX_ONLY_SYMBOLS  /**/
 
-#if defined(x86_64_HOST_ARCH)
+#if defined(x86_64_HOST_ARCH) || defined(aarch64_HOST_ARCH)
 #define RTS_WIN64_ONLY(X) X
 #else
 #define RTS_WIN64_ONLY(X) /**/
@@ -164,7 +164,6 @@ extern char **environ;
       SymI_HasProto(rts_IOManagerIsWin32Native)          \
       SymI_HasProto(rts_ConsoleHandlerDone)              \
       SymI_NeedsProto(__mingw_module_is_dll)             \
-      RTS_WIN64_ONLY(SymI_NeedsProto(___chkstk_ms))      \
       SymI_HasProto(__stdio_common_vswprintf_s)          \
       SymI_HasProto(__stdio_common_vswprintf)            \
       SymI_HasProto(_errno)                              \
@@ -1020,7 +1019,7 @@ extern char **environ;
 #define SymI_NeedsDataProto(vvv)  extern StgWord vvv[];
 #if defined(COMPILING_WINDOWS_DLL)
 #define SymE_HasProto(vvv)    SymE_HasProto(vvv);
-#  if defined(x86_64_HOST_ARCH)
+#  if defined(x86_64_HOST_ARCH) || defined(aarch64_HOST_ARCH)
 #    define SymE_NeedsProto(vvv)    extern void __imp_ ## vvv (void);
 #    define SymE_NeedsDataProto(vvv) SymE_NeedsProto(vvv)
 #  else

My commands:

$ ./configure --target=aarch64-unknown-mingw32 CC=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-clang AR=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-llvm-ar CXX=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-clang++ NM=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-nm OBJCOPY=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-objcopy OBJDUMP=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-objdump RANLIB=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-llvm-ranlib SIZE=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-size STRINGS=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-strings STRIP=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-strip WindresCmd=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-windres LLVMAS=$LLVM_MINGW_PATH/bin/aarch64-w64-mingw32-clang --enable-unregisterised --with-intree-gmp

$ hadrian/build --bignum=native --flavour=default+no_profiled_libs --docs=none -j

I used the tag: ghc-9.12.1-release from ghc repo.

Windows 11 ARM GHC build.

@TerrorJack Merry Christmas!

1. Download 7z compiled by https://github.com/mstorsjo/llvm-mingw/releases/download/20241217/llvm-mingw-20241217-ucrt-macos-universal.tar.xz
2. Add mingw folder with content from here.
3. Change your settings to setup windres and c compiler paths.
4. Use stage0/bin *.exe files. NOT stage1. By some reason it contains MacOS but libraries should be OK.

windows-11-from-darwin-cross-compile-daf659b6e3.diff

remote-hsc2hs.sh

cc @angerman who might be interested in porting the AArch64 NCG to Windows :-)

mentioned in issue #24883 (closed)

mentioned in commit 812c9cea

mentioned in commit b056e326

Next steps I have in mind:

• Land !12767 (closed)
• Land required changes in submodules
• Set up a special aarch64 deb12 image with llvm-mingw toolchain and wine; I've verified wine shipped by deb12 works for the example above. It's also possible to run it in x86_64 using qemu user mode but we might as well avoid an extra layer of emulation.
• Set up a CI job (could be nightly only) to verify the minimum example works, similar to how the existing x86_64->aarch64 cross job is tested.

The above steps would give us a nice solid foundation to iterate on more complex stuff.

Land required changes in submodules

1. Cosmetic one https://github.com/haskell/terminfo/pull/58
2. https://github.com/haskell/cabal/pull/10705
3. https://github.com/haskell/win32/pull/238
4. https://github.com/haskell/directory/pull/196
5. https://github.com/haskell/haskeline/pull/193
6. https://github.com/haskell/hsc2hs/pull/98
7. https://github.com/haskell/process/pull/333
8. It is more about #25619 but noted only when Windows 11 ARM is used as a host platform https://github.com/haskell/unix/pull/334 with MSYS2 clangarm compilers and https://github.com/mstorsjo/llvm-mingw. May be it is related to usage LLVM instead of GCC and could be reproduced at Windows x86 as well.

The legend:

• - received at least one response from maintainers, communication is in progress.
• - changes has been applied.
• No mark - no information yet

@bgamari could you possibly assist with packages under your control? terminfo / haskeline / hsc2hs / process.

Thanks for the ping, @Bodigrim. Yes, I'll have a look.

I believe these should all now be merged. Thanks @gulin.serge !

Added one more checks are green

Thanks!

One more please!

Thanks!

mentioned in merge request !12767 (closed)

mentioned in merge request !12787 (closed)

mentioned in commit ce8fb51e