[ruby-core:105800] [Ruby master Bug#18269] trace_opt_not and trace_opt_regexpmatch2 insns are indistinguishable

Issue #18269 has been reported by mame (Yusuke Endoh).

----------------------------------------
Bug #18269: trace_opt_not and trace_opt_regexpmatch2 insns are indistinguishable
https://bugs.ruby-lang.org/issues/18269

* Author: mame (Yusuke Endoh)
* Status: Open
* Priority: Normal
* Backport: 2.6: UNKNOWN, 2.7: UNKNOWN, 3.0: UNKNOWN
----------------------------------------
On some platforms and compiler tool chain, `RubyVM::ISeq#to_a` shows a wrong instruction.

```
set_trace_func(Proc.new { })
set_trace_func(nil)

iseq = RubyVM::InstructionSequence.compile("/true/ =~ 'true'")
pp iseq.to_a.last
# expected:
#=> [1,
#    :RUBY_EVENT_LINE,
#    [:putobject, /true/],
#    [:putstring, "true"],
#    [:opt_regexpmatch2, {:mid=>:=~, :flag=>16, :orig_argc=>1}],
#    [:leave]]

# actual:
#=> [1,
#    :RUBY_EVENT_LINE,
#    [:putobject, /true/],
#    [:putstring, "true"],
#    [:opt_not, {:mid=>:=~, :flag=>16, :orig_argc=>1}],
#    [:leave]]
```

In this case, `opt_regexpmatch2` is a correct insn, but actually `opt_not` is printed.

This is caused by the fix of #14870. commit:48c8df9e0eb295af06d593ce37ce1933c0ee1d90 made almost all traced and optimized instructions (`trace_opt_*` insns) delegate to `opt_send_without_block`. Under some conditions, GCC seems to merge them to one code block because the bodies of `trace_opt_*` instructions are all identical. I'm unsure why, but in this case, only the pair of `trace_opt_not` and `trace_opt_regexpmatch2` are merged, and their addresses in direct-threaded code are now indistinguishable.

Fortunately, there seems to be no path to revert `trace_*` insn to its original (non-trace) version, except `RubyVM::ISeq#to_a`. (YJIT might be also affected because it uses `rb_vm_insn_addr2opcode`, but I'm unsure.)
So, AFAIK, this is not a significant problem, at the present time. However, this will bring trouble if we attempt to revert `trace_*` insn in future.

Some tests of MJIT uses `RubyVM::ISeq#to_a` to check if the insn that the test targets is correctly tested. This leads to false positive warnings:

http://rubyci.s3.amazonaws.com/graviton2/ruby-master/log/20211026T020530Z.log.html.gz

```
[20397/21172] TestJIT#test_compile_insn_opt_regexpmatch2
/home/chkbuild/chkbuild/tmp/build/20211026T020530Z/ruby/test/ruby/test_jit.rb:599: warning: 'opt_regexpmatch2' insn is not included in the script. Actual insns are: putobject putstring opt_not leave

/home/chkbuild/chkbuild/tmp/build/20211026T020530Z/ruby/test/ruby/test_jit.rb:600: warning: 'opt_regexpmatch2' insn is not included in the script. Actual insns are: putstring putobject opt_not leave
 = 0.58 s
```

@jeremyevans0 @k0kubun Do you have any idea to fix this issue?



-- 
https://bugs.ruby-lang.org/

Unsubscribe: <mailto:[e-mail]?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>