Add flag inlining_callee_size_threshold_with_constants (default 200) which causes more aggressive i…

Add flag inlining_callee_size_threshold_with_constants (default 200) which causes more aggressive inlining if constant arguments are passed to the callee.

R=vegorov@google.com

Committed: https://code.google.com/p/dart/source/detail?r=42184

[srdjan, 2014-12-08 19:03:27]

srdjan@google.com changed reviewers:
+ vegorov@google.com

[srdjan, 2014-12-08 19:03:28]

No differences in startup time measured using dart2js.
Significant speedup on SkeletalAnimationSIMD (2x)

[Vyacheslav Egorov (Google), 2014-12-08 19:11:30]

LGTM!

though the heuristic is a bit spooky because it increases the limit greatly but
there is now guarantee that huge method would suddenly shrink after inlining via
constant propagation.

We should check out the benchmark to see why it improves so muc

https://codereview.chromium.org/788673002/diff/1/runtime/vm/flow_graph_inline...
File runtime/vm/flow_graph_inliner.cc (right):

https://codereview.chromium.org/788673002/diff/1/runtime/vm/flow_graph_inline...
runtime/vm/flow_graph_inliner.cc:518: (instr_count <=
FLAG_inlining_constant_arguments_size_threshold)) {
maybe renamed this flag because now we have a positive threshold (one that
forces inlining  of small methods with constant arguments) and negative
threshold which are called in a similar way.

[srdjan, 2014-12-08 21:19:02]

On 2014/12/08 19:11:30, Vyacheslav Egorov (Google) wrote:
> LGTM!
> 
> though the heuristic is a bit spooky because it increases the limit greatly
but
> there is now guarantee that huge method would suddenly shrink after inlining
via
> constant propagation.
> 

IMO, it is not only about the size, but also about performance benefits vs
compilation slow down. Passing constant arguments can bring big performance
gains.

> We should check out the benchmark to see why it improves so muc

Matrix44SIMDOperations.multiply will get inlined with new inlining strategy.

> 
>
https://codereview.chromium.org/788673002/diff/1/runtime/vm/flow_graph_inline...
> File runtime/vm/flow_graph_inliner.cc (right):
> 
>
https://codereview.chromium.org/788673002/diff/1/runtime/vm/flow_graph_inline...
> runtime/vm/flow_graph_inliner.cc:518: (instr_count <=
> FLAG_inlining_constant_arguments_size_threshold)) {
> maybe renamed this flag because now we have a positive threshold (one that
> forces inlining  of small methods with constant arguments) and negative
> threshold which are called in a similar way.

[srdjan, 2014-12-08 21:19:09]

https://codereview.chromium.org/788673002/diff/1/runtime/vm/flow_graph_inline...
File runtime/vm/flow_graph_inliner.cc (right):

https://codereview.chromium.org/788673002/diff/1/runtime/vm/flow_graph_inline...
runtime/vm/flow_graph_inliner.cc:518: (instr_count <=
FLAG_inlining_constant_arguments_size_threshold)) {
On 2014/12/08 19:11:30, Vyacheslav Egorov (Google) wrote:
> maybe renamed this flag because now we have a positive threshold (one that
> forces inlining  of small methods with constant arguments) and negative
> threshold which are called in a similar way. 

Renamed flags to:
inlining_constant_arguments_max_size_threshold(above no inlining)
inlining_constant_arguments_min_size_threshold (below, always inline if
arg-count sufficient).

[srdjan, 2014-12-08 21:33:38]

Committed patchset #1 (id:1) manually as r42184 (presubmit successful).