519877 - NJ merge: lirasm --stress mode

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Graydon Hoare :graydon]

Or perhaps --self-benchmark. The idea here is to use some of the same machinery that'd be required to support --random mode in bug 519873, but have them run deterministically and at large volume, outputting some kinds of measurements of current NJ performance that could be compared to baseline numbers (code size and rdtsc-based cycle counting might be a good start).

In theory, just --random mode with a determinstic seed might suffice, assuming it's kept "legally executable", not meaningless jibberish. Though I imagine there's room for more interesting *targeted* stress tests too. Play around, be creative.

The idea with both bugs is really only to teach lirasm to test itself programmatically in "much more depth" than a bunch of individual unit tests would do. Unit tests are fixed and small. C++ code can generate more complex cases and run them rapidly.

Comment 1

[Nicholas Nethercote [inactive]]

--random with a large block size approximates this pretty well.  Eg. I can run 'lirasm --random 1000000' and it takes about 2 seconds.  It doesn't seed the random number generator so runs are repeatable.  (Perhaps I should allow the user to choose the seed.)

Comment 3

[Rick Reitmaier]

switching 503556 from blocking to a dependency.

Comment 4

[Nicholas Nethercote [inactive]]

I'm going to close this.  --random seems good enough.