Actions
Feature #15902
closed
Add a specialized instruction for `.nil?`
Status:
Closed
Assignee:
-
Target version:
-
Description
I'd like to add a specialized instruction for .nil?. We have specialized instructions for .length and .empty?, and surprisingly our application also calls .nil? a lot:
[aaron@TC ~/g/github (gc-boot-stats)]$ git grep '.empty?' | wc -l
2553
[aaron@TC ~/g/github (gc-boot-stats)]$ git grep '.length' | wc -l
3975
[aaron@TC ~/g/github (gc-boot-stats)]$ git grep '.nil?' | wc -l
3117
I'm not sure how hot any of the .nil? callsites are, but I think this instruction will speed up most of them.
I tried two benchmark runners:
Benchmark/ips¶
require "benchmark/ips"
class Niller
def nil?; true; end
end
not_nil = Object.new
xnil = nil
niller = Niller.new
Benchmark.ips do |x|
x.report("nil?") { xnil.nil? }
x.report("not nil") { not_nil.nil? }
x.report("niller") { niller.nil? }
end
Results¶
On Ruby master:
[aaron@TC ~/g/ruby (master)]$ ./ruby compil.rb
Warming up --------------------------------------
nil? 429.195k i/100ms
not nil 437.889k i/100ms
niller 437.935k i/100ms
Calculating -------------------------------------
nil? 20.166M (± 8.1%) i/s - 100.002M in 5.002794s
not nil 20.046M (± 7.6%) i/s - 99.839M in 5.020086s
niller 22.467M (± 6.1%) i/s - 112.111M in 5.013817s
[aaron@TC ~/g/ruby (master)]$ ./ruby compil.rb
Warming up --------------------------------------
nil? 449.660k i/100ms
not nil 433.836k i/100ms
niller 443.073k i/100ms
Calculating -------------------------------------
nil? 19.997M (± 8.8%) i/s - 99.375M in 5.020458s
not nil 20.529M (± 7.0%) i/s - 102.385M in 5.020689s
niller 21.796M (± 8.0%) i/s - 108.110M in 5.002300s
[aaron@TC ~/g/ruby (master)]$ ./ruby compil.rb
Warming up --------------------------------------
nil? 402.119k i/100ms
not nil 438.968k i/100ms
niller 398.226k i/100ms
Calculating -------------------------------------
nil? 20.050M (±12.2%) i/s - 98.519M in 5.008817s
not nil 20.614M (± 8.0%) i/s - 102.280M in 5.004531s
niller 22.223M (± 8.8%) i/s - 110.309M in 5.013106s
On this patch:
[aaron@TC ~/g/ruby (specialized-nilp)]$ ./ruby compil.rb
Warming up --------------------------------------
nil? 468.371k i/100ms
not nil 456.517k i/100ms
niller 454.981k i/100ms
Calculating -------------------------------------
nil? 27.849M (± 7.8%) i/s - 138.169M in 5.001730s
not nil 26.417M (± 8.7%) i/s - 131.020M in 5.011674s
niller 21.561M (± 7.5%) i/s - 107.376M in 5.018113s
[aaron@TC ~/g/ruby (specialized-nilp)]$ ./ruby compil.rb
Warming up --------------------------------------
nil? 477.259k i/100ms
not nil 428.712k i/100ms
niller 446.109k i/100ms
Calculating -------------------------------------
nil? 28.071M (± 7.3%) i/s - 139.837M in 5.016590s
not nil 25.789M (±12.9%) i/s - 126.470M in 5.011144s
niller 20.002M (±12.2%) i/s - 98.144M in 5.001737s
[aaron@TC ~/g/ruby (specialized-nilp)]$ ./ruby compil.rb
Warming up --------------------------------------
nil? 467.676k i/100ms
not nil 445.791k i/100ms
niller 415.024k i/100ms
Calculating -------------------------------------
nil? 26.907M (± 8.0%) i/s - 133.755M in 5.013915s
not nil 25.319M (± 7.9%) i/s - 125.713M in 5.007758s
niller 19.569M (±11.8%) i/s - 96.286M in 5.008533s
According to benchmark/ips, it's about 27% faster when the object is nil or a regular object. When it's an object that implements .nil?, I think it might be slower but it's hard to tell.
Benchmark-driver¶
I added a benchmark driver file:
prelude: |
class Niller; def nil?; true; end; end
xnil, notnil = nil, Object.new
niller = Niller.new
benchmark:
- xnil.nil?
- notnil.nil?
- niller.nil?
loop_count: 10000000
Results (tested against master @ c9b74f9fd95113df903fc34cc1d6ec3fb3160c85 )¶
[aaron@TC ~/g/ruby (specialized-nilp)]$ make benchmark ARGS=benchmark/nil_p.yml
./revision.h unchanged
/Users/aaron/.rbenv/shims/ruby --disable=gems -rrubygems -I./benchmark/lib ./benchmark/benchmark-driver/exe/benchmark-driver \
--executables="compare-ruby::/Users/aaron/.rbenv/shims/ruby --disable=gems -I.ext/common --disable-gem" \
--executables="built-ruby::./miniruby -I./lib -I. -I.ext/common ./tool/runruby.rb --extout=.ext -- --disable-gems --disable-gem" \
benchmark/nil_p.yml
Calculating -------------------------------------
compare-ruby built-ruby
xnil.nil? 68.825M 405.121M i/s - 10.000M times in 0.145296s 0.024684s
notnil.nil? 66.357M 267.874M i/s - 10.000M times in 0.150700s 0.037331s
niller.nil? 110.273M 123.089M i/s - 10.000M times in 0.090684s 0.081242s
Comparison:
xnil.nil?
built-ruby: 405120725.0 i/s
compare-ruby: 68825019.2 i/s - 5.89x slower
notnil.nil?
built-ruby: 267873885.2 i/s
compare-ruby: 66357000.6 i/s - 4.04x slower
niller.nil?
built-ruby: 123089042.6 i/s
compare-ruby: 110273035.8 i/s - 1.12x slower
[aaron@TC ~/g/ruby (specialized-nilp)]$ make benchmark ARGS=benchmark/nil_p.yml
./revision.h unchanged
/Users/aaron/.rbenv/shims/ruby --disable=gems -rrubygems -I./benchmark/lib ./benchmark/benchmark-driver/exe/benchmark-driver \
--executables="compare-ruby::/Users/aaron/.rbenv/shims/ruby --disable=gems -I.ext/common --disable-gem" \
--executables="built-ruby::./miniruby -I./lib -I. -I.ext/common ./tool/runruby.rb --extout=.ext -- --disable-gems --disable-gem" \
benchmark/nil_p.yml
Calculating -------------------------------------
compare-ruby built-ruby
xnil.nil? 45.083M 360.998M i/s - 10.000M times in 0.221811s 0.027701s
notnil.nil? 69.558M 271.054M i/s - 10.000M times in 0.143765s 0.036893s
niller.nil? 115.423M 79.667M i/s - 10.000M times in 0.086638s 0.125523s
Comparison:
xnil.nil?
built-ruby: 360997801.1 i/s
compare-ruby: 45083426.9 i/s - 8.01x slower
notnil.nil?
built-ruby: 271054130.3 i/s
compare-ruby: 69557959.1 i/s - 3.90x slower
niller.nil?
compare-ruby: 115422793.6 i/s
built-ruby: 79666674.5 i/s - 1.45x slower
I think there is too much noise for the third case.
I'm not happy about making rb_false non-static, but I'm not sure how else to do this patch.
What do you think?
Files
Updated by Eregon (Benoit Daloze) over 5 years ago
FWIW, we already inline nil? in the AST for TruffleRuby:
https://github.com/oracle/truffleruby/commit/57628008#diff-09677670ed37ee212c217374c6468718
Makes sense since it's such a small operation and it's frequently used.
Updated by tenderlovemaking (Aaron Patterson) over 5 years ago
I applied this optimization and tested against our application while logging instructions:
> insn_info[order(insn_info$Count),]
Instruction Count
94 reverse 1
95 opt_newarray_max 49
93 setblockparam 54
92 opt_call_c_function 60
22 opt_aset_with 302
78 setglobal 417
39 setclassvariable 441
87 opt_div 597
88 once 1263
91 branchnil 1357
52 opt_regexpmatch1 1386
53 duphash 1513
76 getglobal 1538
58 topn 1607
56 opt_and 2129
89 opt_or 2344
86 opt_mult 2859
46 toregexp 11460
28 definesmethod 11880
20 setconstant 12930
81 intern 19087
5 defineclass 32953
55 opt_mod 33719
24 opt_aref_with 36994
80 throw 44954
35 duparray 53448
96 setlocal 58770
30 getspecial 63765
69 opt_length 74235
77 newrange 77459
71 opt_size 98244
15 definemethod 104818
42 tostring 124320
70 opt_gt 130296
74 concatarray 133235
48 getclassvariable 143407
34 opt_str_freeze 192432
68 opt_regexpmatch2 272036
90 checkkeyword 279475
44 freezestring 290999
14 putspecialobject 338706
43 concatstrings 344539
18 putstring 435151
45 getblockparam 455764
64 expandarray 490262
41 checktype 541355
21 newhash 698166
63 opt_neq 764259
57 swap 1146326
66 opt_ge 1283820
73 opt_le 1345993
51 opt_nil_p 1536827
72 opt_minus 1598463
67 invokeblock 1614973
79 getblockparamproxy 1675137
82 getlocal 1799493
38 opt_ltlt 1888635
65 splatarray 1971596
83 opt_case_dispatch 2049897
85 setlocal_WC_1 2062346
50 invokesuper 2153631
84 opt_lt 2229708
3 opt_setinlinecache 2422345
54 opt_plus 2803688
60 opt_empty_p 2820474
27 opt_aset 3204087
47 putobject_INT2FIX_0_ 3790567
75 adjuststack 3924393
31 opt_not 3986585
2 getconstant 4722342
62 jump 4781094
49 dupn 4854287
37 setinstancevariable 5385254
61 checkmatch 5448507
36 newarray 5539621
26 setn 6319773
59 putobject_INT2FIX_1_ 9551850
16 defined 10071630
32 getlocal_WC_1 15007475
29 opt_eq 16058544
10 send 16372975
9 putobject 20392124
4 putnil 26095091
19 branchif 30425098
33 opt_aref 31976718
1 opt_getinlinecache 32302934
40 getinstancevariable 34616623
13 pop 36429507
7 setlocal_WC_0 39844133
23 dup 41106221
11 nop 51714288
8 putself 56476479
17 branchunless 62428811
12 leave 86843816
6 opt_send_without_block 147141336
25 getlocal_WC_0 213198244
Our application boot process executes this instruction 1.5mil times:
> insn_info[which(insn_info$Instruction == "opt_nil_p"),]
Instruction Count
51 opt_nil_p 1536827
>
I think this is a useful instruction, so I'd like to apply the patch. :)
Updated by ko1 (Koichi Sasada) over 5 years ago
All right, go ahead.
Updated by tenderlovemaking (Aaron Patterson) over 5 years ago
- Status changed from Open to Closed
I pushed this as 9faef3113fb4331524b81ba73005ba13fa0ef6c6
We found it decreased our app boot time by ~1.8 seconds (which is not large percentage wise, but definitely measurable). I tested by booting our app in production mode 50 times with and without the patch.
Here is a graph of the results:
Again with Y at 0:
Updated by ned (Ned Hadzo) about 5 years ago
This was reverted, right? https://github.com/ruby/ruby/commit/a0980f2446c0db735b8ffeb37e241370c458a626
Updated by mame (Yusuke Endoh) about 5 years ago
ned (Ned Hadzo) wrote:
This was reverted, right? https://github.com/ruby/ruby/commit/a0980f2446c0db735b8ffeb37e241370c458a626
It has been re-introduced. The Mojave issue was also fixed.
Actions
Like0
Like0Like0Like0Like0Like0Like0