Bug 178312 – [JSC][LLInt] Implement Polymorphic Inline Caches on op_get_by_id

• Status: ASSIGNED
• Component: JavaScriptCore

[Caio Lima]

Polymorphic Inline Caching is being used by Baseline+ tiers in JSC for a long time. However, It is not implemented on LLInt, which means that any CLoop build won't be able to get this optimization. My idea is to introduce a "Pseudo-PIC" on LLInt, to get performance progressions in some benchmarks.

[Caio Lima]

Created attachment 323816 [details]
WIP - It starts

This Patch starts the implementation. I'm interested in collecting feedback from you to keep working on it. My idea is to cache Structures using FIFO policy, to be able to catch different call-sites of a function and optimize for them (Probably I'm going to create a case for that). IMHO, this policy seems very simple, but even with it, I could measure some improvements on benchmarks:

PS.: I collected following results running benchmarks with "JSC_useJIT=false"

Benchmark report for V8Spider, Octane, and Kraken on Caios-MacBook-Pro (MacBookPro12,1).

VMs tested:
"baseline" at /Users/caiolima/open_projects/WebKit/WebKitBaselineBuild/Release/jsc
"changes" at /Users/caiolima/open_projects/WebKit/WebKitBuild/Release/jsc

Collected 6 samples per benchmark/VM, with 6 VM invocations per benchmark. Emitted a call to gc() between sample
measurements. Used 1 benchmark iteration per VM invocation for warm-up. Used the jsc-specific preciseTime() function to
get microsecond-level timing. Reporting benchmark execution times with 95% confidence intervals in milliseconds.

                                                 baseline                  changes                                      
V8Spider:
   crypto                                   683.1670+-19.2621    ?    835.8115+-431.9966      ? might be 1.2234x slower
   deltablue                               1362.3902+-27.2429    ?   1370.7209+-18.6387       ?
   earley-boyer                             659.4617+-399.6603        490.7356+-11.2124         might be 1.3438x faster
   raytrace                                 264.5184+-19.7043    ?    266.9626+-16.2422       ?
   regexp                                   148.3394+-3.1761          148.2460+-8.9221        
   richards                                1456.1012+-30.0906        1454.5450+-40.0814       
   splay                                    231.0293+-8.9334     ^    194.3757+-2.3684        ^ definitely 1.1886x faster

   <geometric>                              496.4563+-34.7263         480.6935+-38.1704         might be 1.0328x faster

                                                 baseline                  changes                                      
Octane:
   encrypt                                   4.29210+-0.08702    ?     4.32454+-0.08857       ?
   decrypt                                  77.00315+-1.79384    ?    79.30127+-1.12004       ? might be 1.0298x slower
   deltablue                        x2       9.31920+-0.18430          9.13937+-0.25287         might be 1.0197x faster
   earley                                    7.89351+-2.62695          6.85874+-0.15562         might be 1.1509x faster
   boyer                                   117.29007+-1.82930        115.10083+-0.65218         might be 1.0190x faster
   navier-stokes                    x2      90.61067+-17.35486        80.68529+-2.79175         might be 1.1230x faster
   raytrace                         x2      45.02483+-0.48965         43.93484+-0.63733         might be 1.0248x faster
   richards                         x2       4.28244+-0.06543          4.25291+-0.08199       
   splay                            x2       3.04093+-0.05215    ^     2.71874+-0.07910       ^ definitely 1.1185x faster
   regexp                           x2      82.33913+-16.75664        71.77397+-5.44225         might be 1.1472x faster
   pdfjs                            x2     217.93367+-28.88964       207.78765+-3.79388         might be 1.0488x faster
   mandreel                         x2    2951.07356+-61.71802      2905.05659+-50.70337        might be 1.0158x faster
   gbemu                            x2     623.67314+-30.73659   ?   656.00748+-23.37269      ? might be 1.0518x slower
   closure                                   0.73940+-0.01035    ?     0.74472+-0.00569       ?
   jquery                                   10.41921+-1.39356          9.88610+-0.63410         might be 1.0539x faster
   box2d                            x2     163.96675+-8.26118        161.95834+-3.38487         might be 1.0124x faster
   zlib                             x2   13847.71583+-227.59617  ^ 12545.11145+-496.99654     ^ definitely 1.1038x faster
   typescript                       x2    3462.74837+-54.67934      3430.36820+-66.78332      

   <geometric>                              89.44004+-1.44483    ^    86.00409+-0.88517       ^ definitely 1.0400x faster

Since it isn't possible to re-patch code on Interpreter, I'm using a loop and a new "op_get_by_id_polymorphic" to handle cached cases and avoid slow paths on property lookups.

[Build Bot]

Attachment 323816 [details] did not pass style-queue:


ERROR: Source/JavaScriptCore/bytecode/StructureProfile.h:46:  Missing space inside { }.  [whitespace/braces] [5]
ERROR: Source/JavaScriptCore/bytecode/CodeBlock.cpp:1311:  Weird number of spaces at line-start.  Are you using a 4-space indent?  [whitespace/indent] [3]
ERROR: Source/JavaScriptCore/bytecode/CodeBlock.cpp:1312:  Weird number of spaces at line-start.  Are you using a 4-space indent?  [whitespace/indent] [3]
ERROR: Source/JavaScriptCore/bytecode/CodeBlock.cpp:1313:  Weird number of spaces at line-start.  Are you using a 4-space indent?  [whitespace/indent] [3]
ERROR: Source/JavaScriptCore/bytecode/CodeBlock.cpp:1314:  Weird number of spaces at line-start.  Are you using a 4-space indent?  [whitespace/indent] [3]
ERROR: Source/JavaScriptCore/bytecode/CodeBlock.cpp:1315:  Weird number of spaces at line-start.  Are you using a 4-space indent?  [whitespace/indent] [3]
ERROR: Source/JavaScriptCore/bytecode/CodeBlock.cpp:1316:  Weird number of spaces at line-start.  Are you using a 4-space indent?  [whitespace/indent] [3]
ERROR: Source/JavaScriptCore/bytecode/CodeBlock.cpp:1317:  Weird number of spaces at line-start.  Are you using a 4-space indent?  [whitespace/indent] [3]
ERROR: Source/JavaScriptCore/bytecode/BytecodeDumper.cpp:445:  One line control clauses should not use braces.  [whitespace/braces] [4]
Total errors found: 9 in 15 files


If any of these errors are false positives, please file a bug against check-webkit-style.

[Caio Lima]

Created attachment 323977 [details]
Patch

[Caio Lima]

Created attachment 323981 [details]
Patch

[Build Bot]

Comment on attachment 323981 [details]
Patch

Attachment 323981 [details] did not pass mac-debug-ews (mac):
Output: http://webkit-queues.webkit.org/results/4878634

Number of test failures exceeded the failure limit.

[Build Bot]

Created attachment 323994 [details]
Archive of layout-test-results from ews113 for mac-elcapitan

The attached test failures were seen while running run-webkit-tests on the mac-debug-ews.
Bot: ews113  Port: mac-elcapitan  Platform: Mac OS X 10.11.6

[Caio Lima]

Created attachment 324177 [details]
Proposed Patch

Patch fixing crashes on Bot

[Caio Lima]

Ping review

[Caio Lima]

Created attachment 329167 [details]
WIP - Patch RFC

It's a WIP project. I'm testing some new approaches to avoid some regressions I've found into Octane benchmarks.

[Caio Lima]

Created attachment 329170 [details]
Benchmark

This is the report of running with JSC_useJIT=false. I could notice some regressions that need further investigation.