[turbofan] Reduce memory consumption of graph building

[turbofan] Reduce memory consumption of graph building

Allow reservation of additional input capacity when creating nodes to prevent switching to deque representation when adding well-known additional inputs.

Also ensure that only a single temporary buffer is used to create temporary input arrays before allocating nodes.

R=mstarzinger@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=24896

[danno, 2014-10-22 13:53:22]

danno@chromium.org changed reviewers:
+ mstarzinger@chromium.org

[danno, 2014-10-22 13:53:36]

PTAL

[Benedikt Meurer, 2014-10-27 09:29:12]

bmeurer@chromium.org changed reviewers:
+ bmeurer@chromium.org

[Benedikt Meurer, 2014-10-27 09:29:14]

https://codereview.chromium.org/644083003/diff/160001/test/unittests/compiler...
File test/unittests/compiler/value-numbering-reducer-unittest.cc (right):

https://codereview.chromium.org/644083003/diff/160001/test/unittests/compiler...
test/unittests/compiler/value-numbering-reducer-unittest.cc:71:
static_cast<Operator::Opcode>(IrOpcode::kLast - i);
What is the purpose of this change?

https://codereview.chromium.org/644083003/diff/160001/test/unittests/compiler...
test/unittests/compiler/value-numbering-reducer-unittest.cc:99:
static_cast<Operator::Opcode>(IrOpcode::kLast - i);
What is the purpose of this change?

[Michael Starzinger, 2014-10-27 09:34:20]

LGTM with comments.

https://codereview.chromium.org/644083003/diff/160001/src/compiler/graph-buil...
File src/compiler/graph-builder.cc (right):

https://codereview.chromium.org/644083003/diff/160001/src/compiler/graph-buil...
src/compiler/graph-builder.cc:143: graph()->NewNode(common()->Merge(1),
arraysize(inputs), inputs, true);
suggestion: Having the "incomplete" parameter (with default value) on the
one-operand NewNode method might slightly increase readability.

https://codereview.chromium.org/644083003/diff/160001/src/compiler/graph-buil...
src/compiler/graph-builder.cc:216: control = graph()->NewNode(op,
arraysize(inputs), inputs, true);
suggestion: Having the "incomplete" parameter (with default value) on the
two-operand NewNode method might slightly increase readability.

https://codereview.chromium.org/644083003/diff/160001/src/compiler/graph.cc
File src/compiler/graph.cc (right):

https://codereview.chromium.org/644083003/diff/160001/src/compiler/graph.cc#n...
src/compiler/graph.cc:38: if (!incomplete &&
!IrOpcode::IsControlOpcode(result->opcode())) {
Is there a particular reason not to decorate control nodes?

https://codereview.chromium.org/644083003/diff/160001/test/unittests/compiler...
File test/unittests/compiler/value-numbering-reducer-unittest.cc (right):

https://codereview.chromium.org/644083003/diff/160001/test/unittests/compiler...
test/unittests/compiler/value-numbering-reducer-unittest.cc:71:
static_cast<Operator::Opcode>(IrOpcode::kLast - i);
Is there any particular reason for this change? I actually would like to get rid
of IrOpcode::kLast completely if we can, it's only used in one assertion AFAICT.

[danno, 2014-10-27 10:12:32]

Committed patchset #9 (id:160001) manually as 24896 (presubmit successful).