Add support for running on Windows #233
Conversation
| @@ -0,0 +1,108 @@ | |||
| # Makefile for "nmake", part of Microsoft Visual Studio Compiler (MSVC) on Windows | |||
There was a problem hiding this comment.
I wonder if we should go ahead and switch to a platform-independent build script to avoid makefile drift. Since we already include Ruby, maybe it could be a set of rake tasks?
There was a problem hiding this comment.
Fair point re: Makefile drift. I wonder if there is a way to maintain the list of tests (and examples) another way, or just discover all of them in the test/examples directories, to avoid having to add new tests to both.
Generating Makefiles could work, but I worry that the indirection (i.e. having to go through a generation script first) isn't obvious - i.e. I'd prefer if we could find a way to avoid that extra step.
There was a problem hiding this comment.
What about something like CMake / Meson, which generates Makefiles on *nix platforms and NMake specific Makefiles on Windows?
There was a problem hiding this comment.
Postgres itself started using Meson for this purpose, but that of course then introduces a build-time dependency (which I'm 50/50 on whether its okay to add).
I think for now I'll not adjust this since this is already a massive PR, but I think we could come back to this at a future time.
One approach to solving the duplication for tests/examples specifically could be to have a Makefile in the tests and examples folders that can run on either platform. The main motivation for not doing that in the root folder Makefile is that there are a bunch of features not supported by nmake that make it hard to write a portable Makefile that can do the code generation/etc conditionally.
…units If we place them together with regular source to be compiled, we always need to maintain a special list of files not to compile directly. Instead, treat them as include files in the src/postgres folder.
This avoids having any .c files in the top-level src/ folder that can't be directly compiled, and thus lets us simplify the logic for defining which source units are to be compiled. In passing drop the src/postgres/guc-file.c stub which is no longer used.
Previously there used to be a recurring bug where the defines suddenly became empty when regenerating sources, due to an ordering issue. Fix by always using the versions defined on top of the Makefile as the source. Additionally, rewrite PG_VERSION_STR to avoid the string containing information from the environment the source was extracted in (which caused unnecessary diffs between different machines).
There is no reason to keep using the old version, so we might as well upgrade, since we're about to add another invocation of this.
This removes an unnecessary whitespace that was added between the "__thread" keyword and the type, as well improves the regexp for cases where variables are both extern and PGDLLIMPORT.
All systems we support have it (HAVE_DECL_STRNLEN is always 1), and it might unnecessarily conflict with built-ins in some cases.
This adds guards for HAVE_PTHREAD around any logic that requires pthread
directly (used for automatically cleaning up memory on thread exit),
removes an unnecessary include for libgen.h, and ensures structs are
initialized with "{0}", not "{}", to support C89 compilers.
…ectly This moves the decision which includes we copy, on top of those we discover to be needed, from the Makefile to the extract_source script, to make it easier to manage the list. In passing this also broadens the patterns of which atomics includes are kept (previously we only kept x86/ARM/PPC, but it doesn't hurt to keep all of them).
In the pg_query context these do not need to do the work they usually do (we always output to stderr, and we never output to the client), and mocking them avoids pulling in postmaster.c amongst other benefits. It also avoids problems with Windows-specific code in a follow-up commit.
This uses a few steps to workaround the problem that we're not actually running on Windows during code extraction: - Define EXEC_BACKEND for all code extraction, since this is safe to do always and mainly allows us to discover some code that can be removed - Define WIN32 when processing elog.c and mbutils.c - this causes the extraction script to throw a bunch of errors, but we still get an analysis result that detects unused code guarded by a "#ifdef WIN32" - Explicitly add a few methods to be resolve that otherwise wouldn't be needed (but are needed when compiling code guarded by WIN32 defines) - Workaround our lack of extracting port/win32 by adding some mocks for port/win32/signal.c directly at the end of elog.c instead
This uses the defines gathered from running Postgres' own ./configure script on Windows under either MSVC or mingw64. Additionally it also defines "__thread" to be "__declspec( thread )" when compiling with MSVC.
We previously had the hardcoded macOS ("darwin") contents for this file,
which is unnecessary and incorrect. Instead keep this file empty in all
cases except on Windows, where we include port/win32.h (just like Postgres
itself does), since it is actually necessary for our use case.
Different from Postgres itself we undef PGDLLIMPORT and PGDLLEXPORT, since
they are not needed for pg_query, and PGDLLIMPORT conflicts with the
thread local attribute.
Additionally, we workaround a Windows-specific problem with sigsetjmp
defines when using clang whilst __MINGW64__ is defined (which can happen
on MSYS2-based environments).
Git automatically converts simple LFs to CRLFs on Windows when checking out code, which appears to break the deparser tests. For now simply turn off this behaviour for any *.sql files in the repository.
lfittl
force-pushed
the
windows-support
branch
from
31ee266
to
59586e4
Compare
| @@ -0,0 +1,108 @@ | |||
| # Makefile for "nmake", part of Microsoft Visual Studio Compiler (MSVC) on Windows | |||
There was a problem hiding this comment.
What about something like CMake / Meson, which generates Makefiles on *nix platforms and NMake specific Makefiles on Windows?
Fixes #44