Faster to number conversion

Avoid tagged values for Instructions that truncate the operands with ToNumber.

I case the ToNumber is applied to a non numeric value but its not observable (some constants and oddballs) we should already do it in hydrogen...

BUG=
R=verwaest@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=15818

[oliv, 2013-07-19 12:20:17]

This part of the split up of https://codereview.chromium.org/19271008/

[Sven Panne, 2013-07-22 07:06:36]

DBC

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions.h
File src/hydrogen-instructions.h (right):

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions...
src/hydrogen-instructions.h:411: if (IsTaggedPrimitive() || IsSmi() ||
IsHeapNumber() ||
o_O o_O o_O

return IsTaggedPrimitive() || IsSmi() || IsHeapNumber() || IsBoolean() ||
IsJSArray();

[oliv, 2013-07-22 07:34:30]

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions.h
File src/hydrogen-instructions.h (right):

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions...
src/hydrogen-instructions.h:411: if (IsTaggedPrimitive() || IsSmi() ||
IsHeapNumber() ||
On 2013/07/22 07:06:36, Sven Panne wrote:
> o_O o_O o_O
> 
> return IsTaggedPrimitive() || IsSmi() || IsHeapNumber() || IsBoolean() ||
> IsJSArray();

lol, that's embarrassing

[Jakob Kummerow, 2013-07-22 07:43:47]

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions.h
File src/hydrogen-instructions.h (right):

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions...
src/hydrogen-instructions.h:3362: HConstant* CopyToTruncatedNumber(Zone* zone);
This sounds more generic than it is. How about "CopyOddballToNumber"?

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen.cc
File src/hydrogen.cc (right):

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen.cc#newcode7857
src/hydrogen.cc:7857: Push(Add<HConstant>(OS::nan_value(),
Representation::Double()));
Stack height mismatches at basic block joins are a pattern that we need to
eliminate/avoid. Suggestion:

HValue copied_constant = Add<HConstant>(...);
IfBuilder if_nan(this);
if_nan.If<...>(...);
if_nan.Then();
if_nan.ElseDeopt();
if_nan.End();
return copied_constant;

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen.h
File src/hydrogen.h (right):

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen.h#newcode1121
src/hydrogen.h:1121: void BuildHydrogenBinaryOperation(BinaryOperation* expr);
I don't see this defined or used anywhere.

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen.h#newcode1123
src/hydrogen.h:1123: HValue* ConstantToNumber(HValue* value);
I don't see this anywhere either.

[Toon Verwaest, 2013-07-22 08:30:20]

Looking good. Some comments.

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions.h
File src/hydrogen-instructions.h (right):

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions...
src/hydrogen-instructions.h:410: bool IsNoString() const {
Can we name this IsNonString, in line with IsNonPrimitive?

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions...
src/hydrogen-instructions.h:3361: HConstant* CopyToTruncatedInt32(Zone* zone);
I guess you tried hard to keep this const? I'd prefer to keep it; but if it's
too hard... meh :)

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions...
src/hydrogen-instructions.h:3362: HConstant* CopyToTruncatedNumber(Zone* zone);
I think the original name is fine. CopyToTruncatedInt32 doesn't handle all cases
either, and will return NULL if not possible. We could rename both to
TryCopyToTruncated* though.

On 2013/07/22 07:43:47, Jakob wrote:
> This sounds more generic than it is. How about "CopyOddballToNumber"?

[Toon Verwaest, 2013-07-22 08:31:44]

One more comment.

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions.cc
File src/hydrogen-instructions.cc (right):

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions...
src/hydrogen-instructions.cc:2331: HConstant*
HConstant::CopyToTruncatedNumber(Zone* zone) {
Shouldn't we also return "this" for values that are already numbers?

[oliv, 2013-07-22 13:56:43]

PTAL

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions.cc
File src/hydrogen-instructions.cc (right):

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions...
src/hydrogen-instructions.cc:2331: HConstant*
HConstant::CopyToTruncatedNumber(Zone* zone) {
On 2013/07/22 08:31:44, Toon Verwaest wrote:
> Shouldn't we also return "this" for values that are already numbers?

there is no point to copying the constant if we are already a number (since
unlike the int32 case we do not change representation).

returning 'this' is afaik not a good idea, since then the client does not know,
if the instruction was already added or not...

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions.h
File src/hydrogen-instructions.h (right):

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions...
src/hydrogen-instructions.h:410: bool IsNoString() const {
On 2013/07/22 08:30:21, Toon Verwaest wrote:
> Can we name this IsNonString, in line with IsNonPrimitive?

Done.

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions...
src/hydrogen-instructions.h:3361: HConstant* CopyToTruncatedInt32(Zone* zone);
On 2013/07/22 08:30:21, Toon Verwaest wrote:
> I guess you tried hard to keep this const? I'd prefer to keep 
> it; but if it's
> too hard... meh :)

the problem is deferred handle dereference in TruncateToNumber

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen-instructions...
src/hydrogen-instructions.h:3362: HConstant* CopyToTruncatedNumber(Zone* zone);
On 2013/07/22 08:30:21, Toon Verwaest wrote:
> I think the original name is fine. CopyToTruncatedInt32 doesn't handle all
cases
> either, and will return NULL if not possible. We could rename both to
> TryCopyToTruncated* though.
> 
> On 2013/07/22 07:43:47, Jakob wrote:
> > This sounds more generic than it is. How about "CopyOddballToNumber"?
> 

How about returning Maybe<>, now it is clear, that the conversion is partial

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen.cc
File src/hydrogen.cc (right):

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen.cc#newcode7857
src/hydrogen.cc:7857: Push(Add<HConstant>(OS::nan_value(),
Representation::Double()));
On 2013/07/22 07:43:47, Jakob wrote:
> Stack height mismatches at basic block joins are a pattern that we need to
> eliminate/avoid. Suggestion:
> 
> HValue copied_constant = Add<HConstant>(...);
> IfBuilder if_nan(this);
> if_nan.If<...>(...);
> if_nan.Then();
> if_nan.ElseDeopt();
> if_nan.End();
> return copied_constant;

Done.

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen.h
File src/hydrogen.h (right):

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen.h#newcode1121
src/hydrogen.h:1121: void BuildHydrogenBinaryOperation(BinaryOperation* expr);
On 2013/07/22 07:43:47, Jakob wrote:
> I don't see this defined or used anywhere.

Done.

https://codereview.chromium.org/19798002/diff/16001/src/hydrogen.h#newcode1123
src/hydrogen.h:1123: HValue* ConstantToNumber(HValue* value);
On 2013/07/22 07:43:47, Jakob wrote:
> I don't see this anywhere either.

Done.

[Toon Verwaest, 2013-07-22 16:46:45]

lgtm with nit

https://codereview.chromium.org/19798002/diff/41001/src/hydrogen-instructions.h
File src/hydrogen-instructions.h (right):

https://codereview.chromium.org/19798002/diff/41001/src/hydrogen-instructions...
src/hydrogen-instructions.h:411: return (IsTaggedPrimitive() || IsSmi() ||
IsHeapNumber() ||
Nit: remove () around the expression.

[oliv, 2013-07-23 09:14:05]

Committed patchset #5 manually as r15818 (presubmit successful).