Introduce 'top-of-stack caching' to the toplevel code generator by...

Introduce 'top-of-stack caching' to the toplevel code generator by
allowing the value of the rightmost subexpressions to be evaluated
into an accumulator register (eg, eax, rax, or r0) rather than onto
the stack.


Committed: http://code.google.com/p/v8/source/detail?r=3630

[Kevin Millikin (Chromium), 2010-01-13 17:15:55]

This is just the ia32 version.  I'll do a straightforward port to ARM and amd64
after I've got your comments.

[fschneider, 2010-01-14 00:04:50]

LGTM so far.

http://codereview.chromium.org/541047/diff/1004/1007
File src/fast-codegen.cc (right):

http://codereview.chromium.org/541047/diff/1004/1007#newcode104
src/fast-codegen.cc:104: __ pop(result_register());
As already discussed the push/pop around TestAndBranch could be optimized in
case we don't have to call the ToBooleanStub in TestAndBranch.

http://codereview.chromium.org/541047/diff/1004/1008
File test/mjsunit/compiler/short-circuit.js (right):

http://codereview.chromium.org/541047/diff/1004/1008#newcode1
test/mjsunit/compiler/short-circuit.js:1: // Copyright 2009 the V8 project
authors. All rights reserved.
2010?

[William Hesse, 2010-01-14 12:25:59]

LGTM.

http://codereview.chromium.org/541047/diff/1004/1006
File src/ia32/fast-codegen-ia32.cc (right):

http://codereview.chromium.org/541047/diff/1004/1006#newcode1062
src/ia32/fast-codegen-ia32.cc:1062: __ RecordWrite(ecx, offset, edx, ebx);
Are we using result_register() instead of eax for flexibility?  If so, shouldn't
we assert that ecx, edx, and ebx are not result_register()?

[Lasse Reichstein, 2010-01-14 12:50:49]

Drive-by comments (just the usual :).

http://codereview.chromium.org/541047/diff/1004/1006
File src/ia32/fast-codegen-ia32.cc (right):

http://codereview.chromium.org/541047/diff/1004/1006#newcode200
src/ia32/fast-codegen-ia32.cc:200: switch (context) {
Could this function be in fast-codegen.cc? It seems entirely platform
independent.

http://codereview.chromium.org/541047/diff/1004/1006#newcode247
src/ia32/fast-codegen-ia32.cc:247: Apply(context, eax);
If eax is converted to result_register(), this function looks generic too.

http://codereview.chromium.org/541047/diff/1004/1006#newcode286
src/ia32/fast-codegen-ia32.cc:286: __ mov(eax, Operand(esp, 0));
If we hae a macro to load the top of stack into a register (or possibly any
stack-pointer indexed word), this function could be generic too.

http://codereview.chromium.org/541047/diff/1004/1006#newcode483
src/ia32/fast-codegen-ia32.cc:483: __ mov(Operand(ebp, SlotOffset(slot)),
result_register());
Could this use StoreToFrameField?

http://codereview.chromium.org/541047/diff/1004/1006#newcode507
src/ia32/fast-codegen-ia32.cc:507: result_register());
Could we have an opposite of LoadContextField to use here?

http://codereview.chromium.org/541047/diff/1004/1006#newcode1022
src/ia32/fast-codegen-ia32.cc:1022: TestAndBranch(eax, true_label_, &discard);
Maybe use result_register() here for consistency?

[Kevin Millikin (Chromium), 2010-01-18 11:30:14]

Hi Florian, I've ported this to arm and x64, which required some more
refactoring.  Can you take another look?

[fschneider, 2010-01-18 13:56:59]

LGTM.

http://codereview.chromium.org/541047/diff/3005/2002
File src/fast-codegen.h (right):

http://codereview.chromium.org/541047/diff/3005/2002#newcode232
src/fast-codegen.h:232: void DropAndApply(int count, Expression::Context
context, Register reg);
Extra space.

http://codereview.chromium.org/541047/diff/3005/2003
File src/ia32/fast-codegen-ia32.cc (right):

http://codereview.chromium.org/541047/diff/3005/2003#newcode506
src/ia32/fast-codegen-ia32.cc:506: __ j(not_zero, if_true);
&discard as well.

http://codereview.chromium.org/541047/diff/3005/2006
File src/x64/fast-codegen-x64.cc (right):

http://codereview.chromium.org/541047/diff/3005/2006#newcode512
src/x64/fast-codegen-x64.cc:512: __ j(not_zero, if_true);
branch to &discard as well here?

http://codereview.chromium.org/541047/diff/3005/2006#newcode1043
src/x64/fast-codegen-x64.cc:1043: // Reciever is now under the key and value.
Typo -> "receiver"