Bug #9382: [patch] add opt_aref_str and opt_aset_str

• Status: Closed
• Assignee: tmm1 (Aman Karmani)
• Target version: 2.2.0
• ruby -v: trunk
• Backport: 1.9.3: DONTNEED, 2.0.0: DONTNEED, 2.1: DONTNEED

[tmm1 (Aman Karmani)]

@charliesome, @normalperson (Eric Wong) and I have been working on a patch to add two new instructions:

https://bugs.ruby-lang.org/issues/9188#note-9

These optimize Hash#[] and Hash#[]=, when used with a string literal key.

Before:

0010 putstring "str"
0012 opt_aref <callinfo!mid:[], argc:1, ARGS_SKIP>

After:

0010 opt_aref_str <callinfo!mid:[], argc:1, ARGS_SKIP>, "str"

The opt_(aref|aset)_str instructions avoid str_duplicate, resulting in reduced allocation overhead. For example, this loop requires zero allocations after the patch:

hash={}; loop{ hash["key"] += 1 }

In our rails app, 2.5% of iseqs can benefit from these optimizations:

$ ruby -rconfig/environment -e' p ObjectSpace.each_object(RubyVM::InstructionSequence).count '
63314

$ ruby -rconfig/environment -e' p ObjectSpace.each_object(RubyVM::InstructionSequence).select{ |i| i.disasm.match(/ opt_(aref|aset)_str /) }.size '
1464

@headius (Charles Nutter) is also adding a similar optimization to jruby: https://bugs.ruby-lang.org/issues/8998#note-15

I plan to commit the patch above, so if you have any feedback on the implementation please share.

[tmm1 (Aman Karmani)]

With the patch, "allocations per request" in our app is reduced by ~3%.

[nobu (Nobuyoshi Nakada)]

• Backport changed from 1.9.3: UNKNOWN, 2.0.0: UNKNOWN, 2.1: UNKNOWN to 1.9.3: DONTNEED, 2.0.0: DONTNEED, 2.1: DONTNEED
• Status changed from Open to Closed

[shyouhei (Shyouhei Urabe)]

• Related to Bug #12855: Inconsistent keys identity in compare_by_identity Hash when using literals added