Closed
Bug 605196
Opened 14 years ago
Closed 13 years ago
Need switch to enable/disable specializations and inlining at JIT-time
Categories
(Tamarin Graveyard :: Baseline JIT (CodegenLIR), defect, P2)
Tamarin Graveyard
Baseline JIT (CodegenLIR)
Tracking
(Not tracked)
RESOLVED
FIXED
Q3 11 - Serrano
People
(Reporter: wmaddox, Assigned: wmaddox)
References
Details
(Whiteboard: fixed-in-spicier)
Attachments
(1 file)
(deleted),
patch
|
rreitmai
:
review+
|
Details | Diff | Splinter Review |
Ongoing optimization work has lead to more code paths and greater subtlety in how a given opcode or builtin is translated. In order to rapidly isolate bugs arising from this new machinery, it should be possible to quickly disable it.
- At a minimum, there should be a way to turn off all the "fancy stuff" from the command-line to aid in bug triage.
- It would be helpful to allow fine-grained control of which optimizations are performed to facilitate performance studies.
- Inlining increases code size and puts pressure on analysis algorithms (e.g., deadvars) that are sensitive to the number of branches and labels in the code. In some cases, we may need to fall back to a less aggressive space-for-speed tradeoff in order to control code size and/or compilation time.
Assignee | ||
Updated•14 years ago
|
Summary: Need ability to enable/disable specializations and inlining → Need switch to enable/disable specializations and inlining at runtime
Assignee | ||
Updated•14 years ago
|
Summary: Need switch to enable/disable specializations and inlining at runtime → Need switch to enable/disable specializations and inlining at JIT-time
Assignee | ||
Updated•14 years ago
|
Assignee: nobody → wmaddox
Updated•14 years ago
|
Target Milestone: --- → flash10.2.x-Spicy
Updated•14 years ago
|
Priority: -- → P2
Comment 1•14 years ago
|
||
Arguably we also need a mm.cfg switch for Flash.
Comment 2•14 years ago
|
||
(In reply to comment #1)
> Arguably we also need a mm.cfg switch for Flash.
mms.cfg, actually
Assignee | ||
Comment 3•14 years ago
|
||
Command line switches -osr and -noinline were included in tr-spicy as part of the following patches:
https://bugzilla.mozilla.org/attachment.cgi?id=489290&action=diff
https://bugzilla.mozilla.org/attachment.cgi?id=491050&action=diff
Flash mms.cfg support has also been added for these in Spicier.
Leaving the bug open until these make it into TR, and fine-grained control issues are addressed. Retargeting for Serrano.
Status: NEW → ASSIGNED
Whiteboard: fixed-in-spicier
Target Milestone: flash10.2.x-Spicy → flash10.x-Serrano
Updated•14 years ago
|
Flags: flashplayer-bug-
Comment 4•13 years ago
|
||
Bill: is this part of the current work?
Assignee | ||
Comment 5•13 years ago
|
||
(In reply to comment #4)
> Bill: is this part of the current work?
The -noinline switch in tr-spicy and mentioned in comment 3 has landed in TR as part of the port of OSR from tr-spicy:
http://hg.mozilla.org/tamarin-redux/rev/ff7c627718e8
The FP work that will be needed once TR is next integrated into FRMain is covered by WE 2793437.
There is no longer any particular urgency to this bug, but it remains desirable that have more fine-grained control for performance studies.
Assignee | ||
Comment 6•13 years ago
|
||
TR has added an additional inlining optimization since the -Dnoinline switch and the internal inlineFastpath flag were added to tr-spicy.
The patch merely wraps an 'if (inlineFastpath)' around a couple of blocks of code that were already guarded by compile-time conditionalization -- the rest is just indentation.
Attachment #527604 -
Flags: review?(rreitmai)
Comment 7•13 years ago
|
||
Comment on attachment 527604 [details] [diff] [review]
Make inlining of fused add and vector index conditional on inlineFastpath flag
r+
Attachment #527604 -
Flags: review?(rreitmai) → review+
Comment 8•13 years ago
|
||
changeset: 6219:f432ac087cde
user: William Maddox <wmaddox@adobe.com>
summary: Bug 605196 - followup: allow inlining enable/disable mechanism to control fused add + vector-index (r=rreitmai)
http://hg.mozilla.org/tamarin-redux/rev/f432ac087cde
Assignee | ||
Updated•13 years ago
|
Status: ASSIGNED → RESOLVED
Closed: 13 years ago
Resolution: --- → FIXED
You need to log in
before you can comment on or make changes to this bug.
Description
•