517121 - avoid loading Toplevel* in jit code

Status: VERIFIED FIXED

(Tamarin Graveyard :: Virtual Machine, defect, P3)

Description

[Edwin Smith]

instead, pass MethodEnv* to helpers that currently need Toplevel*, and load it within the helpers.  for all code, this reduces the size of jit code and slightly reduces jit time.  (less instructions).

for hot code, its a wash, except that when we also can inline the fast path inside the helper, we usually can avoid loading Toplevel* completely, which is a win.  plus, inlining the fast path in this manner is a win too.

individual patches and/or dependent bugs will be tracked here.

combined, the changes currently net about ~10% speedup on x86 and 5-20% speedup on arm, on v8.  other tests TBD.

Comment 1

[Edwin Smith]

Attachment: OP_applytype: delay loading Toplevel* until we need it in the error path.

Comment 2

[Edwin Smith]

Attachment: load Toplevel* in the newarray helper instead of inline jit code

Comment 3

[Edwin Smith]

Attachment: load Toplevel* in astype helper instead of inline jit code

Comment 4

[Edwin Smith]

Attachment: load Toplevel* in instanceof helper

Comment 5

[Edwin Smith]

Attachment: load Toplevel* in "in" operator

Comment 6

[Tom Harwood]

Comment on attachment 401805 [details] [diff] [review]
OP_applytype: delay loading Toplevel* until we need it in the error path.

Worth doing a sandbox build here?  I'm not sure of the xplt implications of taking the address of an ostensibly inlined function.

Comment 7

[Edwin Smith]

Comment on attachment 401805 [details] [diff] [review]
OP_applytype: delay loading Toplevel* until we need it in the error path.

applytype pushed http://hg.mozilla.org/tamarin-redux/rev/66a4cb7df0aa

Comment 8

[Edwin Smith]

Comment on attachment 401806 [details] [diff] [review]
load Toplevel* in the newarray helper instead of inline jit code

newarray pushed http://hg.mozilla.org/tamarin-redux/rev/740ec102ce34

Comment 9

[Edwin Smith]

Comment on attachment 401808 [details] [diff] [review]
load Toplevel* in astype helper instead of inline jit code

astype pushed http://hg.mozilla.org/tamarin-redux/rev/e945a38bd5f5

Comment 10

[Edwin Smith]

Comment on attachment 401810 [details] [diff] [review]
load Toplevel* in instanceof helper

instanceof pushed
http://hg.mozilla.org/tamarin-redux/rev/4499c517b563

Comment 11

[Edwin Smith]

Comment on attachment 401822 [details] [diff] [review]
load Toplevel* in "in" operator

in pushed 
http://hg.mozilla.org/tamarin-redux/rev/0cf78fb25264

Comment 12

[Edwin Smith]

Attachment: Removes boilerplate around callproperty slow path

less code, no significant perf change now that inline caching is in place.

Comment 13

[Edwin Smith]

Attachment: clean up boilerplate around construct slow paths

less jit code, no significant perf change.  (applies on top of the previous patch for callproperty cleanup)

Comment 14

[Edwin Smith]

Comment on attachment 404049 [details] [diff] [review]
Removes boilerplate around callproperty slow path

pushed 
http://hg.mozilla.org/tamarin-redux/rev/f11a8d8ca82c

Comment 15

[Edwin Smith]

Comment on attachment 404050 [details] [diff] [review]
clean up boilerplate around construct slow paths

pushed http://hg.mozilla.org/tamarin-redux/rev/70726a525ca0

Comment 16

[Edwin Smith]

Attachment: eliminate boilerplate in jit around calls to coerce()

This patch also moves the coerce() code to instr.h, too.  R- if you want me to split this into two patches to make the review easier.

Comment 17

[Edwin Smith]

Comment on attachment 405098 [details] [diff] [review]
eliminate boilerplate in jit around calls to coerce()

http://hg.mozilla.org/tamarin-redux/rev/172691e389c8

Comment 18

[Chris Peyer]

Engineering work item.  Marking as verified.