Introduce runtime function %OptimizeFunctionOnNextCall to manually trigger optimization.

Introduce runtime function %OptimizeFunctionOnNextCall to manually trigger optimization.


TEST=existing unit tests still pass

Committed: http://code.google.com/p/v8/source/detail?r=7572

[Jakob Kummerow, 2011-04-08 17:31:06]

Please review.

This fixes a bunch of tests that didn't trigger optimization before because
their loops didn't run long enough.
It also reduces the time needed for the tests run by "tools/test.py --arch=x64
-j16" on my machine from 38 to 36 seconds :-)

[Mads Ager (chromium), 2011-04-11 11:06:38]

Looks great.

One suggestion for replacing the contents of a test.

A number of explicit optimizations of something that should be inlined. I think
they can be removed.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/optimized...
File test/mjsunit/compiler/optimized-function-calls.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/optimized...
test/mjsunit/compiler/optimized-function-calls.js:52:
%OptimizeFunctionOnNextCall(object.f);
I'm not sure you should do this. If you optimize call_f, does o.f get inlined?
--trace-inlining will give you that info.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/property-...
File test/mjsunit/compiler/property-refs.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/property-...
test/mjsunit/compiler/property-refs.js:51: %OptimizeFunctionOnNextCall(StoreXY);
Similarly here. If you do not optimize this here, does StoreXY get inlined in
LoadXY?

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/recursive...
File test/mjsunit/compiler/recursive-deopt.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/recursive...
test/mjsunit/compiler/recursive-deopt.js:39: assertEquals(1 << 1, f(0));
I think we need to completely change this test. This will always deopt in the
base case here.

I think we should get rid of RunTests all together.

Use f(4) as our example. Call f(4) 5 times checking the result. Then optimize
and call f(4) checking the result. Then replace 'var one' and then just call
f(4) again checking the result. That will cause deopts at all the deopt places
in the function.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-l...
File test/mjsunit/compiler/regress-loadfield.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-l...
test/mjsunit/compiler/regress-loadfield.js:66: %OptimizeFunctionOnNextCall(bar);
Shouldn't be necessary to optimize bar here. It should be inlined into test I
think.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-m...
File test/mjsunit/compiler/regress-max.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-m...
test/mjsunit/compiler/regress-max.js:34: %OptimizeFunctionOnNextCall(Math.max);
This should get inlined if you just optimize f I think.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-s...
File test/mjsunit/compiler/regress-stacktrace-methods.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-s...
test/mjsunit/compiler/regress-stacktrace-methods.js:48:
%OptimizeFunctionOnNextCall(Hest.prototype.three);
Same here.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/switch-ba...
File test/mjsunit/compiler/switch-bailout.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/switch-ba...
test/mjsunit/compiler/switch-bailout.js:40: // [disabled optimization for: f /
7fb0972a2119] ?!
Remove?

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-11...
File test/mjsunit/regress/regress-1106.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-11...
test/mjsunit/regress/regress-1106.js:56: %OptimizeFunctionOnNextCall(x.gee);
Shouldn't be necessary. Should be inlined.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-12...
File test/mjsunit/regress/regress-1229.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-12...
test/mjsunit/regress/regress-1229.js:61: %OptimizeFunctionOnNextCall(foo);
Shouldn't be necessary.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-12...
test/mjsunit/regress/regress-1229.js:88: %OptimizeFunctionOnNextCall(baz);
Shouldn't be needed.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-12...
File test/mjsunit/regress/regress-1237.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-12...
test/mjsunit/regress/regress-1237.js:39: %OptimizeFunctionOnNextCall(observe);
Ditto.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-96...
File test/mjsunit/regress/regress-962.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-96...
test/mjsunit/regress/regress-962.js:57:
%OptimizeFunctionOnNextCall(L.prototype.c);
Ditto.

[Jakob Kummerow, 2011-04-11 12:55:51]

I agree with most of your comments, but a few of the methods do not get inlined.
Of course I can still remove their optimization triggers if you want me to.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/optimized...
File test/mjsunit/compiler/optimized-function-calls.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/optimized...
test/mjsunit/compiler/optimized-function-calls.js:52:
%OptimizeFunctionOnNextCall(object.f);
On 2011/04/11 11:06:38, Mads Ager wrote:
> I'm not sure you should do this. If you optimize call_f, does o.f get inlined?
> --trace-inlining will give you that info.

Done. (Removed this line.)

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/property-...
File test/mjsunit/compiler/property-refs.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/property-...
test/mjsunit/compiler/property-refs.js:51: %OptimizeFunctionOnNextCall(StoreXY);
On 2011/04/11 11:06:38, Mads Ager wrote:
> Similarly here. If you do not optimize this here, does StoreXY get inlined in
> LoadXY?

Yes, StoreXY does get inlined if only LoadXY is optimized explicitly. Removed
the two unnecessary lines.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/recursive...
File test/mjsunit/compiler/recursive-deopt.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/recursive...
test/mjsunit/compiler/recursive-deopt.js:39: assertEquals(1 << 1, f(0));
On 2011/04/11 11:06:38, Mads Ager wrote:
> I think we need to completely change this test. This will always deopt in the
> base case here.
> 
> I think we should get rid of RunTests all together.
> 
> Use f(4) as our example. Call f(4) 5 times checking the result. Then optimize
> and call f(4) checking the result. Then replace 'var one' and then just call
> f(4) again checking the result. That will cause deopts at all the deopt places
> in the function.

Done.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-l...
File test/mjsunit/compiler/regress-loadfield.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-l...
test/mjsunit/compiler/regress-loadfield.js:66: %OptimizeFunctionOnNextCall(bar);
On 2011/04/11 11:06:38, Mads Ager wrote:
> Shouldn't be necessary to optimize bar here. It should be inlined into test I
> think.

Done.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-m...
File test/mjsunit/compiler/regress-max.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-m...
test/mjsunit/compiler/regress-max.js:34: %OptimizeFunctionOnNextCall(Math.max);
On 2011/04/11 11:06:38, Mads Ager wrote:
> This should get inlined if you just optimize f I think.

No: "Did not inline max called from f (target not inlineable).", and it doesn't
get optimized either without the explicit statement.
Want me to remove %OptimizeFunctionOnNextCall(f); instead? I guess that's not
really required for this test.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-s...
File test/mjsunit/compiler/regress-stacktrace-methods.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-s...
test/mjsunit/compiler/regress-stacktrace-methods.js:48:
%OptimizeFunctionOnNextCall(Hest.prototype.three);
On 2011/04/11 11:06:38, Mads Ager wrote:
> Same here.

Hest.prototype.three can't be inlined (because it throws), and doesn't get
optimized unless explicitly requested.
(Note that I don't request optimization of Svin.prototype.two, because that
method does get inlined.)

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/switch-ba...
File test/mjsunit/compiler/switch-bailout.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/switch-ba...
test/mjsunit/compiler/switch-bailout.js:40: // [disabled optimization for: f /
7fb0972a2119] ?!
On 2011/04/11 11:06:38, Mads Ager wrote:
> Remove?

Done.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-11...
File test/mjsunit/regress/regress-1106.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-11...
test/mjsunit/regress/regress-1106.js:56: %OptimizeFunctionOnNextCall(x.gee);
On 2011/04/11 11:06:38, Mads Ager wrote:
> Shouldn't be necessary. Should be inlined.

Not according to --trace-inlining. I therefore left the line in there.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-12...
File test/mjsunit/regress/regress-1229.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-12...
test/mjsunit/regress/regress-1229.js:61: %OptimizeFunctionOnNextCall(foo);
On 2011/04/11 11:06:38, Mads Ager wrote:
> Shouldn't be necessary.

Done. This means foo no longer gets optimized (and doesn't get inlined either),
but that should be OK for this test.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-12...
test/mjsunit/regress/regress-1229.js:88: %OptimizeFunctionOnNextCall(baz);
On 2011/04/11 11:06:38, Mads Ager wrote:
> Shouldn't be needed.

Done.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-12...
File test/mjsunit/regress/regress-1237.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-12...
test/mjsunit/regress/regress-1237.js:39: %OptimizeFunctionOnNextCall(observe);
On 2011/04/11 11:06:38, Mads Ager wrote:
> Ditto.

Done.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-96...
File test/mjsunit/regress/regress-962.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-96...
test/mjsunit/regress/regress-962.js:57:
%OptimizeFunctionOnNextCall(L.prototype.c);
On 2011/04/11 11:06:38, Mads Ager wrote:
> Ditto.

Done. Again, L.prototype.c no longer gets optimized, but it's pretty
uninteresting anyway.

[Mads Ager (chromium), 2011-04-11 13:03:44]

LGTM

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-m...
File test/mjsunit/compiler/regress-max.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-m...
test/mjsunit/compiler/regress-max.js:34: %OptimizeFunctionOnNextCall(Math.max);
On 2011/04/11 12:55:51, jkummerow wrote:
> On 2011/04/11 11:06:38, Mads Ager wrote:
> > This should get inlined if you just optimize f I think.
> 
> No: "Did not inline max called from f (target not inlineable).", and it
doesn't
> get optimized either without the explicit statement.
> Want me to remove %OptimizeFunctionOnNextCall(f); instead? I guess that's not
> really required for this test.

I agree with you. What should matter here is that Math.max gets optimize.
Introducing f just to give Math.max a shorter name seems unnecessary. Replace
all calls to f with calls to Math.max, optimize that and check the results?

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-s...
File test/mjsunit/compiler/regress-stacktrace-methods.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-s...
test/mjsunit/compiler/regress-stacktrace-methods.js:48:
%OptimizeFunctionOnNextCall(Hest.prototype.three);
On 2011/04/11 12:55:51, jkummerow wrote:
> On 2011/04/11 11:06:38, Mads Ager wrote:
> > Same here.
> 
> Hest.prototype.three can't be inlined (because it throws), and doesn't get
> optimized unless explicitly requested.
> (Note that I don't request optimization of Svin.prototype.two, because that
> method does get inlined.)

Sounds good. Thanks.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-11...
File test/mjsunit/regress/regress-1106.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-11...
test/mjsunit/regress/regress-1106.js:56: %OptimizeFunctionOnNextCall(x.gee);
On 2011/04/11 12:55:51, jkummerow wrote:
> On 2011/04/11 11:06:38, Mads Ager wrote:
> > Shouldn't be necessary. Should be inlined.
> 
> Not according to --trace-inlining. I therefore left the line in there.

For this test it doesn't matter whether or not gee is optimized. I'd remove the
call in this case.

[Jakob Kummerow, 2011-04-11 13:27:04]

...and landed.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-m...
File test/mjsunit/compiler/regress-max.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/compiler/regress-m...
test/mjsunit/compiler/regress-max.js:34: %OptimizeFunctionOnNextCall(Math.max);
On 2011/04/11 13:03:44, Mads Ager wrote:
> On 2011/04/11 12:55:51, jkummerow wrote:
> > On 2011/04/11 11:06:38, Mads Ager wrote:
> > > This should get inlined if you just optimize f I think.
> > 
> > No: "Did not inline max called from f (target not inlineable).", and it
> doesn't
> > get optimized either without the explicit statement.
> > Want me to remove %OptimizeFunctionOnNextCall(f); instead? I guess that's
not
> > really required for this test.
> 
> I agree with you. What should matter here is that Math.max gets optimize.
> Introducing f just to give Math.max a shorter name seems unnecessary. Replace
> all calls to f with calls to Math.max, optimize that and check the results? 

Done.

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-11...
File test/mjsunit/regress/regress-1106.js (right):

http://codereview.chromium.org/6821009/diff/1/test/mjsunit/regress/regress-11...
test/mjsunit/regress/regress-1106.js:56: %OptimizeFunctionOnNextCall(x.gee);
On 2011/04/11 13:03:44, Mads Ager wrote:
> On 2011/04/11 12:55:51, jkummerow wrote:
> > On 2011/04/11 11:06:38, Mads Ager wrote:
> > > Shouldn't be necessary. Should be inlined.
> > 
> > Not according to --trace-inlining. I therefore left the line in there.
> 
> For this test it doesn't matter whether or not gee is optimized. I'd remove
the
> call in this case.

Done.