Opened 7 years ago

Last modified 4 weeks ago

#989 new task

Windows "native" port

Reported by: simonmar Owned by:
Priority: normal Milestone:
Component: Compiler Version:
Keywords: Cc: ndmitchell@…, fryguybob@…, dterei, dagitj@…, idhameed@…
Operating System: Windows Architecture: x86
Type of failure: None/Unknown Difficulty: Difficult (2-5 days)
Test Case: N/A Blocked By:
Blocking: Related Tickets:

Description

Create a Windows "native" port of GHC, i.e. one that uses the Microsoft toolchain instead of GCC and the GNU binutils. The main tasks are:

  • Convert the pretty printer in the i386 NCG to generate MS/Intel syntax instead of AT&T syntax. We can then generate assembly code that MASM can grok.
  • Make the RTS compile with Microsoft's CL compiler. (this has been done in the past, so shouldn't be too much work).
  • Drop any dependencies on mingw32 functionality in the libraries.
  • Make appropriate modifications to the driver, build system etc. We'd still need Cygwin/MSYS for build tools (dropping dependency on GNU make is another task entirely...)

Change History (13)

comment:1 Changed 7 years ago by simonpj

Notes from Peter Tanski

The only caveat would be that the free version of MASM (an
add-on to Visual C++ Express) is licensed for non-commercial purposes
only. Commercial Windows developers would have to purchase a full
version of VC++. The most universal solution would be to restrict
the syntax from the pretty-printer to use assembler directives and
macros common to both MASM and IAS (Intel ASsembler), with some
special attention to linker differences between the two.

Here are a few notes on Task #989:

  • Task point 1: pretty printing MS/Intel asm syntax

Most of this seems to be concentrated in compiler/nativeGen/
PprMach?.hs, with a macro added to NCG.h.

Depending on how much you want to interoperate directly with .NET and
other non-C languages, the pretty-printer for assembly output may be
only part of the work: we might add special COFF sections to compiler/
nativeGen/AsmCodeGen.lhs.

  • Task point 2: compiling via Microsoft's CL compiler

Fun. Once we have gotten over point 4 (modifications to driver,
build system, etc.).

  • Task point 3: drop dependencies on mingw32 functionality in the libraries

Is there a Windows-version of readline, or a Windows-native version
of readline available? Even if I don't finish the replacement-GMP
library before this, if I recall correctly, GMP has a Windows-native
version (as a DLL, no less). Are there any other library
dependencies? Windows-native versions of Perl are available.

  • Task point 4: modifications to driver, build system, etc.

driver: DOS batch file or Perl (GHC users will need Perl installed to
run the mangler anyway); might also be a very small C program
(especially if you want to have a splash screen when GHC starts up).

build system: small but pervasive modifications for compiler
dependencies, etc.; as a general cleanup matter it might be nice to
organise these sorts of dependencies to add other compilers later.
The one problem I think people will run into is setting up the paths
in MSYS to use the Windows tools--not a GHC-support issue, really,
but many questions will be raised about it so we might head them off
with lots of build-documentation.

There are several related points to this project:

(1) for general Windows interoperability, GHC will need to produce
DLLs, at least: .NET or other Microsoft languages, such as VB, cannot
include foreign code from static libraries.

(2) once GHC may produce more than one type of assembler output,
would it be too much trouble to give it a cross-compiler option, say
"-arch i386" (compiler/nativeGen/PprMach.hs seems to use
"IF_ARCH_..." statements) or the GCC option, "-b i386v"? A cross-
compile option would probably require changing the compiler/utils/
Outputable.CodeStyle? data constructor AsmStyle? to contain several
different sub-styles (Intel, ATnT) and clutter up the code in
PprMach?.hs, at least for i386_TARGET_ARCH. Of course floatToBytes
and doubleToBytes (converting endianness to host byte order) would
also have to change. (Note: the reference in PprMach?.hs:2418 to
PprAbs? seems to refer to an old module, PprAbsC, which is no longer
used, correct?)

The easy, host-target-compiler-only solution might be to create a new
macro, IF_OS_windows(x,y), in NCG.h and a new Ppr section for the
MASM/Intel syntax in PprMach?.hs.

comment:2 Changed 7 years ago by igloo

  • Milestone set to _|_

comment:3 Changed 7 years ago by simonmar

A plausible plan seems to be to use YASM, which can generate Windows object files and understand GAS syntax. This way the changes to the NCG would be minimal, and hence we have a shorter path to getting something working.

comment:4 Changed 6 years ago by sharpe

  • Version 6.7 deleted

Win port of gmp at http://fp.gladman.plus.com/computing/gmp4win.htm, perhaps move gmp out of dist altogether and only link w/ .so or .dll?

Support utils (cgprof, hp2ps,.. etc.) need minor modifications to build and keep cl from choking on e.g. some exotic C macros (and simple ones such as macros suffixed with semi-colon).

RTS is more challenging as it contains a mix of C89, C99, C++ code which gcc happily compiles; minor inconsistencies (STATIC_INLINE here, static inline there), gcc extensions not supported by cl, inline assembly, void pointer arithmetic etc.

Adapting the build system is pretty terrible (esp. proper autoconf,libtool), custom make rules, custom config.h, custom mk files and some custom Makefile.win32 do the trick but are more difficult to maintain in the future (cmake?).

Used winsock is deprecated, also i will need to look into signal handling as this doesn't behave properly (mingw?).

Registered boot attempts fail with alignment errors, as of yet not resolved (->mix gcc/cl?); will continue on unregistered 6.4 series builds.

Haven't got to real low-level stuff, YASM seems ok and is easily integrated w/ VS2008.

prepping the environment is pretty easy once you've got around the unixy build system
(include vcvars variables PATH,LIBPATH,INCLUDE etc.), btw utils/pwd does not work in mingw using series 6.8.2 (easily resolved).

merging changes will be another fun item.

comment:5 Changed 6 years ago by NeilMitchell

  • Cc ndmitchell@… added

comment:6 Changed 4 years ago by simonmar

  • Difficulty changed from Difficult (1 week) to Difficult (2-5 days)

comment:7 Changed 3 years ago by fryguybob

  • Cc fryguybob@… added
  • Type of failure set to None/Unknown

comment:8 Changed 3 years ago by dterei

  • Cc dterei added

comment:9 Changed 3 years ago by dagit

  • Cc dagitj@… added

Being able to use MS compiler tool chain would be useful to me in the following two ways:

  • 64bit Haskell binaries (yay, 8gigs of ram!)
  • Directly use foo.lib files instead of using gendef/dlltool to (hopefully) construct a .a file. This comes up when trying to use OpenCL with Haskell on windows.

comment:10 Changed 3 years ago by simonmar

@dagit: using the MS compiler chain by itself wouldn't give us a 64-bit Windows port, though it would mean that we don't depend on a working 64-bit mingw port. See #1884.

comment:11 Changed 16 months ago by ihameed

  • Cc idhameed@… added

comment:12 Changed 15 months ago by morabbin

Is this still on anyone's radar?

comment:13 Changed 4 weeks ago by schyler

Just thought I would make a note that this is kind of blocking visual studio support. It's not easy to make a uniform visual studio extension for Haskell right now because it would be expected that it's able to link with other files in the same solution (.c, .cpp etc) however that's not the case since GHC uses gcc.

Note: See TracTickets for help on using tickets.