Feature #14104: Remove `trace` instructions

• Status: Closed
• Assignee: ko1 (Koichi Sasada)
• Target version: 2.5

[ko1 (Koichi Sasada)]

Abstract

Remove trace instructions from bytecode and it will improve performance, 10% faster.
Instead of trace instruction, we add event information for each instruction.
In other words, we unified trace instructions with the following instructions.

Backgroud

TracePoint and old set_trace_func method need to know which point they need to kick hooks at.
trace VM instruction is inserted to represent such hook points and kicks hooks if needed.
It is easy to introduce hook points into bytecode.
However, this technique introduces some overhead on "no-hooks" situations (and I believe 99% of workloads does not need hooks, especially on production environment).

For example, compiled code of the following method has several trace instructions.

def foo x 
  _y = x
end

Disasm output is here:

== disasm: #<ISeq:foo@<compiled>>=======================================
local table (size: 2, argc: 1 [opts: 0, rest: -1, post: 0, block: -1, kw: -1@-1, kwrest: -1])
[ 2] x<Arg>     [ 1] _y         
0000 trace            8                                               (   1)
0002 trace            1
0004 getlocal         x, 0
0007 dup              
0008 setlocal         _y, 0
0011 trace            16                                              (   2)
0013 leave                                                            (   1)

# disasm code is:
puts RubyVM::InstructionSequence.compile("def foo x
  _y = x
end", nil, nil, 0, trace_instruction: true, operands_unification: false).disasm

You can see 3 trace instructions on 0000, 0002 and 0011. They means hook points for :call, :line and :return events. It enables to support TracePoint, but introduces 3 dispatches for useless instructions.

We can remove trace instructions with the trace_instruction: true compile option, but we can't turn on TracePoint for such compiled ISeqs.

Proposal

Remove trace instruction and introduce event flags as information for each instructions.

With same code above, we can get the following disasm code.

== disasm: #<ISeq:foo@<compiled>>=======================================
local table (size: 2, argc: 1 [opts: 0, rest: -1, post: 0, block: -1, kw: -1@-1, kwrest: -1])
[ 2] x<Arg>     [ 1] _y
0000 getlocal         x, 0                                            (   1)[LiCa]
0003 dup
0004 setlocal         _y, 0
0007 leave            [Re]

0000 shows [LiCa], it means :line and :call events.
0007 shows [Re] which represents a :return event.

It is too heavy if we check these event flags on each instruction dispatch.
So we introduce new instructions: trace_ prefix instructions such as trace_getlocal and so on.
All normal instructions have corresponding trace_ prefix instructions (generated by the VM generator automatically).

If we need to enable TracePoint (or set_trace_func), then rewrite all of instruction sequences (ISeqs) in the heap from normal instructions to trace prefix instructions dynamically. trace_ prefix instructions check flags for an executing instruction and kick hooks if needed.

• Good:
  • a. Low overhead for normal execution because we don't need to dispatch useless trace instructions.
  • b. We don't have trace instructions, so that it is easy to turn on tail call optimization flags.
  • c. We can enable/disable traces on small units.
• Bad:
  • a. Big overhead to turn on/off TracePoint because we need to modify all of existing ISeqs.
  • b. Introducing a few incomatibilities for tracing.

For (Bad-a) I believe nobody turn on and trun off tracepoints many times, so that turning on/off overhead is not so big impact.

For (Good-c), I need to explain more.
Current TracePoint enables all of hooks for all methods, ISeqs. With this technique, however, we can select enable/disable the hooks in more small units, like files, classes, methods or a line. For example, we can enable the hooks only method named foo. Now we have no interface to specify it, but we can design later.

Evaluation

require 'benchmark'

def foo n
end

N = 100_000_000
Benchmark.bm(10){|x|
  x.report('trace off'){
    N.times{
      foo(10)
      foo(10)
      foo(10)
    }
  }
  x.report('trace on'){
    TracePoint.new{}.enable
    N.times{
      foo(10)
      foo(10)
      foo(10)
    }
  }
}

Currnt trunk:
                 user     system      total        real
trace off    9.090678   0.000000   9.090678 (  9.083007)
trace on   108.217320   0.005524 108.222844 (108.201941)

modified:
                 user     system      total        real
trace off    6.647247   0.000000   6.647247 (  6.641530)
trace on    93.405389   0.000000  93.405389 ( 93.404864)

Compatibility issues

With this change, we introduced a few incompatibilies:

• Line numbers on :return/:b_return events show the last executed lines, instead of end lines (without return statement).
• begin block doesn't show :line events.

[shevegen (Robert A. Heiler)]

                 user     system      total        real
trace off    9.090678   0.000000   9.090678 (  9.083007)
trace on   108.217320   0.005524 108.222844 (108.201941)

                 user     system      total        real
trace off    6.647247   0.000000   6.647247 (  6.641530)
trace on    93.405389   0.000000  93.405389 ( 93.404864)

\o/

[ko1 (Koichi Sasada)]

• Description updated (diff)

[ko1 (Koichi Sasada)]

Here is a result of benchmark result with

ruby ~/ruby/src/trunk/benchmark/driver.rb -e ruby_2_4::~/ruby/install/ruby_2_4/bin/ruby -e trunk_oct::~/ruby/install/trunk_r60079/bin/ruby -e trunk::~/ruby/install/trunk/bin/ruby -e modified::~/ruby/install/gitruby/bin/ruby  -v -r 5

https://gist.github.com/anonymous/64b7e73e5972befc33a9acde09612f90

[ko1 (Koichi Sasada)]

• Status changed from Open to Closed

Applied in changeset trunk|r60763.

---

remove trace instruction. [Feature #14104]

• tool/instruction.rb: create trace_ prefix instructions.

• compile.c (ADD_TRACE): do not add trace instructions but add
  TRACE link elements. TRACE elements will be unified with a next
  instruction as instruction information.

• vm_trace.c (update_global_event_hook): modify all ISeqs when
  hooks are enabled.

• iseq.c (rb_iseq_trace_set): added to toggle trace_ instructions.

• vm_insnhelper.c (vm_trace): added.
  This function is a body of trace_ prefix instructions.

• vm_insnhelper.h (JUMP): save PC to a control frame.

• insns.def (trace): removed.

• vm_exec.h (INSN_ENTRY_SIG): add debug output (disabled).

[vmakarov (Vladimir Makarov)]

ko1 (Koichi Sasada) wrote:

Abstract

Remove trace instructions from bytecode and it will improve performance, 10% faster.
Instead of trace instruction, we add event information for each instruction.
In other words, we unified trace instructions with the following instructions.

Hi, Koichi. I am very glad that you addressed this issue. The current implementation of tracing is just wasting CPU time. I thought about this too but my approach would be a different because I thought that I had to maintain the full compatibility of trace behaviour.

So I considered to remove and insert trace insns when the tracing is off or on. It requires to change branch offset, catch table offsets, and line tables. To speed up this, I would keep the log of pairs (<place in iseq code, catch table, or line table>, ) and log of pairs (, ).

I also thought that such general log mechanism would be useful in possible speculative iseq optimizations on MRI insns level.

But if the full trace behaviour compatibility is not so important, your solution is much much simpler and practically has the same code locality as you keep codes of trace and usual insns in disjoint parts of the interpreter switch-stmt.

I think rebuilding iseq is not a problem. Sorry if I missed the code change but I don't see a check of version of loaded iseq on compatibility. So what will happen if somebody tries to load the old iseq with trace insns into a new version of MRI?

[ko1 (Koichi Sasada)]

vmakarov (Vladimir Makarov) wrote:

So I considered to remove and insert trace insns when the tracing is off or on. It requires to change branch offset, catch table offsets, and line tables. To speed up this, I would keep the log of pairs (<place in iseq code, catch table, or line table>, ) and log of pairs (, ).

I also thought that such general log mechanism would be useful in possible speculative iseq optimizations on MRI insns level.

I completely agree your approach is more compatible and flexible for other optimization techniques. But it is some tough work to implement it because PC info are located variety of places. I also believe (sometime) we need to implement similar technique you explained.

But if the full trace behaviour compatibility is not so important, your solution is much much simpler and practically has the same code locality as you keep codes of trace and usual insns in disjoint parts of the interpreter switch-stmt.

Yes, simplicity is the reason why I use this technique.
Only few days I needed to make.

I think rebuilding iseq is not a problem. Sorry if I missed the code change but I don't see a check of version of loaded iseq on compatibility. So what will happen if somebody tries to load the old iseq with trace insns into a new version of MRI?

We don't guarantee the cross-version dump/load compatibility (this is why we embed version info into dumped results). So I believe there are no problem on this change.

[mame (Yusuke Endoh)]

• Related to Bug #14582: Unable to use `method__entry` and `method_return` tracing probes since 2.5 added