698407 - Performance degradation during string concatenation

Status: VERIFIED FIXED

(Tamarin Graveyard :: Virtual Machine, defect)

Description

[Brent Baker]

Attachment: simple testcase

Performance degradation during verification.

cd performance
./runtests.py language/string/typed/split.as

With the float patch queue this takes about 5 minutes to run, while the tamarin-redux repo runs in about 300ms.

Actually this is pretty interesting, the code reports that it only takes 300ms to actually run, but if you use >$ time $AVM test.abc you can see that it takes about 5 minutes to run. Is this unaccounted for time in the verifier or codegen?

tamarin:
>$ avmshell test.abc
305

float:
>$ time avmshell test.abc
308

real	5m11.250s
user	2m2.796s
sys	3m8.434s

Comment 1

[Brent Baker]

This also happens with -Dinterp on the float branch (OS X 64bit release )
>$ time avmshell -Dinterp test.abc 
286

real	5m10.118s
user	2m1.130s
sys	3m8.967s

Comment 2

[Brent Baker]

Issue is not during verification as -Dverifyonly runs quickly, same as tamarin-redux using release-debugger builds. Title of bug needs to be updated once somebody knows where all of the time is going.

Comment 3

[Brent Baker]

Attachment: simple testcase

testcase narrowed down to the string concatenation, I should have taken another second to look at the code the first time

Comment 4

[Brent Baker]

(In reply to Brent Baker from comment #3)
> Created attachment 570701 [details]
> simple testcase
> 
> testcase narrowed down to the string concatenation, I should have taken
> another second to look at the code the first time

GC cycles are completely different between tamarin and the float queue for this code:
Tamarin:
gccount=4 
incremental-marks=34
peak-occupancy: blocks-heap-allocated=7134 blocks-gc-allocated=6427 blocks-private=3665

Float:
gccount=9477
incremental-marks=9500
peak-occupancy: blocks-heap-allocated=38143 blocks-gc-allocated=24823 blocks-private=15095

Comment 5

[Brent Baker]

Compiling the shell with float disabled brings back the lost performance.

Recompile builtins (modify core/builtin.py to set VMCFG_FLOAT=false) and compile with --disable-swf16

Comment 6

[Virgil Palanciuc]

Note: interpreter issue (works fine with -Ojit in new VM; works slow with -Dinterp in new VM, but fast with -Dinterp in old VM). 
Note2: I can't for the life of me figure out why this would happen. It makes no sense. Oh well, maybe tomorrow...

Comment 7

[Virgil Palanciuc]

fixed in changeset 322:544cc0e3be9c

Comment 8

[Brent Baker]

Verified fixed in changeset 324:ba70538d70db