585433 - Skip deadvars_analyze in non-looping methods.

Status: RESOLVED FIXED

(Tamarin Graveyard :: Baseline JIT (CodegenLIR), defect, P2)

Description

[Edwin Smith]

Methods that have no back edges should not require iteration in deadvars.  We
could do a single backwards sweep (deadvars_kill only) to save time. 
Furthermore this sweep could be concurrent with code generation to avoid
traversing LIR one extra time.  (*very* similar to nanojit::StackFilter).

Stats from a selection of 175 SWF's:

methods:    48645
cyclic:      5111 (11%)
total deadvars passes: 104745 (2.15 per method)

If we skip deadvars_analyze in the 43534 acyclic methods, its a 42% reduction in the # of passes, down to 61211 (1.26 per method).  

Deadvars is 14% of the overall jit+verify time in the same run; 42% off would reduce overall jit+verify time by 6%, with deadvars contributing 9%.

Comment 1

[Edwin Smith]

Attachment: (v1) proof of concept

This cuts pure JIT time of my as3-big-100 swfs from 27sec to 26sec, about 3% speedup.  (that's about 700MB of compressed swf content)

Comment 2

[Edwin Smith]

Attachment: (v2) Skip deadvars_analyze in non-looping methods.

Only minor changes since v1:
* renamed haveLoops to hasBackedes
* set hasBackedges when we emit actual LIR backedges, which should be more robust than setting the flag based on indirect conditions.

Comment 3

[Rick Reitmaier]

Comment on attachment 465509 [details] [diff] [review]
(v2) Skip deadvars_analyze in non-looping methods.

Looks good; small nit, typedef HashMap<LIns*,nanojit::BitSet*> to something would improve readability.

Comment 4

[Edwin Smith]

TR: http://hg.mozilla.org/tamarin-redux/rev/5c5db9106e55

Comment 5

[Edwin Smith]

patch includes:

    typedef HashMap<LIns*, nanojit::BitSet*> LabelBitSet;