AddCallProtocol unconditionally forces frame

[headius]

I see that push/pop of binding are omitted when local variables are not needed across calls, but push/pop of frame appear to be emitted all the time:

For code "def foo(a); p a; end" we get the following instrs:

  0 push_frame
  1 check_arity(1, 0, -1)
  2 %t_a_1 = recv_pre_reqd_arg(0)
  3 thread_poll
  4 line_num(0)
  5 %v_0 = call_1o(FUNCTIONAL, 'p', %self, [%t_a_1]){1O}
  6 pop_frame
  7 return(%v_0)
  8 [DEAD]%v_3 = recv_jruby_exc
  9 [DEAD]pop_frame
  10    [DEAD]throw(%v_3)

I guess this reflects a lack of awareness about frame fields in IR right now.

/cc @subbuss @enebo

[subbuss]

That is correct. IRScope needs to add a needsFrame flag and set it based on whatever heuristics are currently used in AST JIT.

[enebo]

This is silly but:

%frame = push_frame

pop_frame %frame

Could be made to work naturally without needing to add a frame but that
might be a bit too much work to not add another flag to IRScope

-Tom

On Tue, Jan 28, 2014 at 9:43 AM, Charles Oliver Nutter <
notifications@github.com> wrote:

I see that push/pop of binding are omitted when local variables are not
needed across calls, but push/pop of frame appear to be emitted all the
time:

For code "def foo(a); p a; end" we get the following instrs:

0 push_frame
1 check_arity(1, 0, -1)
2 %t_a_1 = recv_pre_reqd_arg(0)
3 thread_poll
4 line_num(0)
5 %v_0 = call_1o(FUNCTIONAL, 'p', %self, [%t_a_1]){1O}
6 pop_frame
7 return(%v_0)
8 [DEAD]%v_3 = recv_jruby_exc
9 [DEAD]pop_frame
10 [DEAD]throw(%v_3)

I guess this reflects a lack of awareness about frame fields in IR right
now.

/cc @subbuss https://github.com/subbuss @enebohttps://github.com/enebo

Reply to this email directly or view it on GitHubhttps://github.com//issues/1456
.

blog: http://blog.enebo.com twitter: tom_enebo
mail: tom.enebo@gmail.com

[enebo]

I should have replied to this instead of an email thread on jruby mailing list...Just compile all dead instructions until this is resolved since compiling dead instructions will not affect programs correctness. It might also stress the JIT more anyways as well.

[enebo]

Fixed a long time ago.