[llvm-dev] ORC JIT Weekly #27 -- ORC runtime status update

Lang Hames via llvm-dev llvm-dev at lists.llvm.org
Mon Dec 7 04:29:57 PST 2020


Hi All,

Apologies for the lack of updates recently -- things have been busy over
the holidays.

Work on the Orc runtime has continued, however I decided to tackle the hard
part first so working demo code is a little way off. In the meantime anyone
who's particularly curious can check out the rough work taking shape in [1].

The "hard part" being tackled is the initializer problem: How should static
initializers be run in the executor process? Historically we've run
initializers imperatively, with the JIT ordering the executor to run a
particular sequence of functions, There are two substantial problems with
current initializer support though:

(1) We've never offered in-tree support for running initializers
out-of-process. This *could* be fixed now using the TargetProcessControl
API, but...

(2) This isn't how initializers are run in ahead-of-time compiled code:
They're run by the dynamic loader during process launch and dlopen calls,
and on the thread that made the call. This becomes significant if the
initializers spawn threads, take locks, or trigger new rounds of
initialization (e.g. by calling dlopen).

Now that we have a runtime we can tackle the initializer problem by
introducing a jit version dlopen: Executors call the jit_dlopen function in
the Orc runtime, passing the name of a JITDylib as the argument. The
jit_dlopen implementation calls over to the JIT (via __orc_rt_jit_dispatch,
which insulates the runtime from any RPC / IPC specifics) to obtain the
list of initializers to be run, then executes them. Initializers can then
be run on the same thread that they would have run on in a statically
compiled program.

That's the theory anyway. The implementation is still a work in progress.
In [1] I have started adding the plumbing for __orc_rt_jit_dispatch, and
taught llvm-jitlink to load the runtime into executors and use the
following sequence to run the JIT'd program:

  auto *H = jit_dlopen("Main");
  auto *Main = ((int)(*)(int, char*[]))jit_dlsym(H, "_main");
  int Result = Main(argc, argv);
  jit_dlclose(H);
  return Result;

This doesn't actually work yet as the runtime functions are empty, but I
hope to fill them in over the next few days. Hopefully by next week I'll
have a working demo for you.

Once that's done the next thing on the list will be (re)implementing
eh-frame registration and atexit handling as runtime functions, and then
implementing macOS TLS support.

-- Lang.

[1]
https://github.com/lhames/llvm-project/commit/8c683c549fbf53455b3a4cf26957d4ce489fb90a
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20201207/827bd502/attachment.html>


More information about the llvm-dev mailing list