[LLVMdev] [RFC v4] BPF backend
alexei.starovoitov at gmail.com
Fri Dec 19 10:20:12 PST 2014
On Fri, Dec 19, 2014 at 7:22 AM, Tom Stellard <tom at stellard.net> wrote:
> On Thu, Dec 18, 2014 at 07:31:07PM -0800, Alexei Starovoitov wrote:
>> On Thu, Dec 18, 2014 at 7:25 PM, Alexei Starovoitov
>> <alexei.starovoitov at gmail.com> wrote:
>> > V3->V4:
>> > - added CODE_OWNERS entry for BPF backend
>> > V2->V3:
>> > - fix metadata in tests
>> > V1->V2:
>> > - addressed feedback from Tom and Matt
>> > - removed top level change to configure (now everything via 'experimental-backend')
>> > - reworked error reporting via DiagnosticInfo (similar to R600)
>> > - added few more tests
>> > - added cmake build
>> > - added Triple::bpf
>> > - tested on linux and darwin
>> > http://reviews.llvm.org/D6494
>> cc-ing llvmdev to get feedback.
>> this patch is not the end of development,
>> rather are solid base that folks can start using now.
>> I'm still working on better error handling, more tests,
>> fixes for some corner cases, debug gen, few optimizations, etc.
>> I can keep posting next versions, but the patch will
>> only get bigger which is against 'incremental changes' policy.
>> I would prefer to get this base in and keep working
>> in the tree. What is the way forward here?
>> Since it's not build by default and doesn't touch core,
>> it shouldn't cause any breakage, but
>> may be someone can trigger buildbot run for this patch?
> As far as I know it's up to Chris Lattner to approve new backends, so
> you should start cc'ing him on these emails. Also, due to the holidays
> and the recent mailing list issues reviews / responses may be delayed.
Chris, what's your take?
The links to isa description, linux kernel bits, motivation are in:
It seems phab didn't pick up the earlier discussion,
but it can be seen in email archives:
> Personally, I think this should be added as an experimental target.
> I have reviewed the code and it looks pretty good. I also think it
> benefits the project to have a simple backend like this in tree, because
> it's much easier to use a reference then some of the more complex targets.
Though the holidays and Jan 14 cutoff of 3.6 are looming,
and if Chris is ok with this, I'll be on watch out for any fallout :)
More information about the llvm-dev