Bug #14892: r63763 breaks -DMJIT_FORCE_ENABLE=1

• Status: Closed
• Assignee: k0kubun (Takashi Kokubun)
• Target version: -
• Backport: 2.3: UNKNOWN, 2.4: UNKNOWN, 2.5: UNKNOWN

[normalperson (Eric Wong)]

r63763 ("give up insn attr handles_frame") breaks -DMJIT_FORCE_ENABLE=1

(commit 6b534134a78e3e43c344682c3585e1abab015216 in git)
Related [Bug #14809] [Bug #14834]

Is this being worked on? I mentioned it in [ruby-core:87674],
"make test" fails in test_insns.rb for me on Debian 9 (x86 and amd64).
Not sure what's wrong.

Reverting that commit allows even "make exam" to pass for me.

[k0kubun (Takashi Kokubun)]

• Assignee set to k0kubun (Takashi Kokubun)

I recognize that. I'm currently taking time for RubyBench project but planning to take a look after that. Please ignore the -DMJIT_FORCE_ENABLE=1 until I fix it.

[wanabe (_ wanabe)]

A mini reproductive command is here.

ruby --jit-wait --disable=gems -e '1000.times { a, b = nil }'

[wanabe (_ wanabe)]

And another one.

ruby --jit-wait --disable-gems -e '1000.times { break if /a/ =~ "ab" && !$~[0] }'

[wanabe (_ wanabe)]

Many cases (including a, b = nil) go well after r63863.
Now I found that there remain 2 issues at least.

1. pc is set to next_pos when insn is opt_xxx and body->catch_except_p is falsy.
   • ex. 1000.times { break if /a/ =~ "ab" && !$~[0] }
2. opt_aref_with / opt_aset_with push rb_str_resurrect(key) when MJIT execution is cancelled, but it is unexpected for VM's opt_aref_with / opt_aset_with.
   • ex. 1000.times { Thread.current["a"] = {} }

[k0kubun (Takashi Kokubun)]

• Status changed from Open to Closed

Applied in changeset trunk|r63875.

---

insns.def: stop pushing unnecessary keys for MJIT

[Bug #14892]