465649 - TM: Make nanojit reverse pipeline customizable

Status: RESOLVED INVALID

(Core :: JavaScript Engine, enhancement)

Comment 1

[Andreas Gal :gal]

Attachment: patch

Comment 2

[Andreas Gal :gal]

Attachment: enable dead code filter for regexp, seems to be required (very slow otherwise)

Comment 3

[David Mandelin [:dmandelin]]

We wanted to do this experiment to see if it would improve compilation time for regexps, because they don't need all the filters. The results:

- DeadCodeFilter is needed, without it we get a 12ms regression, presumably because of LIR immediate instructions that are not eliminated.
- We couldn't detect any compiler speedup on regexp-dna or crypto-aes.
- Shark seems to indicate that for regexps, DeadCodeFilter is somewhat costly (perhaps about 1/8 of total compilation time) but the other backward filters are cheap.

So we concluded that for now this won't help.

Comment 4

[Edwin Smith]

The Assembler cases for each instruction actually assume the instruction isn't dead, i.e. the instruction will already have a register assignment and/or spill location assignment when it is reached.  So DCF is really required.

A reason it would account for a large % of time is it can be busy skipping instructions, potentially lots of instructions.  you mentioned immediates.  One thing to look at is whether the LIR generation pipeline can produce less instructions that become dead.  CseFilter will reuse immediates effectively, but it's expensive.  a cheap immediate-only filter, with a fast value-lookup-table, might do the trick.

Comment 5

[Nicholas Nethercote [inactive]]

This bug is older than the hills and barely relevant.  Closing.