[LLVMdev] Proposal: bindings for the Go programming language
axwalk at gmail.com
Fri Oct 10 00:16:01 PDT 2014
On Thu, Oct 9, 2014 at 9:40 PM, Peter Collingbourne <peter at pcc.me.uk> wrote:
> On Thu, Oct 09, 2014 at 09:17:01AM +0200, Andrew Wilkins wrote:
>> > Importing this path would cause 'go get' to check out LLVM plus the bindings
>> > from SVN using the mechanism described in . We would check index.html
>> > files into the www repository to support this.
>> > There doesn't seem to be a good way to build complex C++ projects such as LLVM
>> > using 'go get', so there is a script (update_llvm.sh) that builds LLVM and
>> > configures the package tree with the appropriate compiler/linker flags. This
>> > is essentially the approach that I've taken with the existing gollvm project,
>> > but now the script uses the copy of LLVM in the parent directory instead of
>> > checking one out.
>> Some users will want to build/install the bindings against existing
>> LLVM libraries . For example, if the vNN URL were used, then I
>> think it would be nice to be able to fetch and build the bindings
>> against a released vNN LLVM without also fetching the LLVM tree via
>> "go get".
> I think the best way to solve that problem is to rely on distro-specific
> packaging. For example, on Debian derived distributions we could arrange so
> you could add /usr/share/gocode to your $GOPATH and get the packaged version
> of LLVM that way.
> But if we wanted to, we could support the scenario where the user manually
> checks out only the bindings, by letting people somehow run only the part
> of update_llvm.sh that generates llvm_config.go from a given llvm-config
> executable. I'd imagine this would be useful for distro packagers, as well.
>> What benefit does combining the checkout into the "go get"
>> bring, as opposed to deferring to update_llvm.sh as is done today?
> I think it is better to keep the management of the LLVM tree in $GOPATH in the
> hands of the developer as much as possible. In particular, I'm thinking of
> the scenario where someone does development on that copy of LLVM, and may
> want to check out arbitrary revisions, or use the LLVM git mirror (I can
> easily imagine doing this myself, for example). If we use separate checkouts
> for LLVM and the bindings, this becomes somewhat more cumbersome.
Fair enough. I made an invalid assumption that "go get" was going to
transform the standard LLVM tree so it would come under the Go package
>> One option for improving the "go get" experience for released versions
>> is to include pkg-config support in the OS distributions of LLVM. "go
>> get" is able to invoke pkg-config to determine include/library paths,
>> linker options, etc.
> I think I did try generating a pkg-config file for LLVM at one point, but it
> wouldn't work with the "go" tool because the format only supports a "cflags"
> attribute, and we need the current separation between cppflags and cxxflags
> because the compiler did not accept the -std=c++11 flag in C mode.
Well that is a pain. I wasn't aware that pkg-config's format was so restrictive.
I agree that packaging the bindings into distro packages is reasonably
neat, my only concern is that it is not the usual practice of a lot of
Go developers to use distro-provided Go packages. I'm not sure if any
distros other than Debian/Ubuntu even have any. Ideally I'd like to
avoid placing constraints on adoption of LLVM, but I think I'll have
to concede that this approach is the best we can do right now without
additional support from the "go" tool.
More information about the llvm-dev