Relational comparisons for unboxed mints.

Relational comparisons for unboxed mints.
Committed: https://code.google.com/p/dart/source/detail?r=14343

[Florian Schneider, 2012-10-29 12:11:44]

Added vegorov@ as reviewer.

[Vyacheslav Egorov (Google), 2012-10-29 12:46:40]

lgtm

https://codereview.chromium.org/11344011/diff/1/runtime/vm/intermediate_langu...
File runtime/vm/intermediate_language_ia32.cc (right):

https://codereview.chromium.org/11344011/diff/1/runtime/vm/intermediate_langu...
runtime/vm/intermediate_language_ia32.cc:707: __ cmpl(lo, Address(ESP, 0 *
kWordSize));
can you benchmark another pattern: 

extract higher
extract higher 
compare
extract lower
extract lower
compare

?

[Vyacheslav Egorov (Google), 2012-10-29 12:49:10]

I retract LGTM.

There are issues with signed comparison here.

[Florian Schneider, 2012-10-29 13:07:32]

Right, I have to do a subtract-with-carry instead.

On 2012/10/29 12:49:10, Vyacheslav Egorov (Google) wrote:
> I retract LGTM.
> 
> There are issues with signed comparison here.

[Florian Schneider, 2012-10-29 15:32:38]

New version uploaded. Please have another look.

[srdjan, 2012-10-29 19:30:25]

https://codereview.chromium.org/11344011/diff/18001/runtime/vm/flow_graph_opt...
File runtime/vm/flow_graph_optimizer.cc (right):

https://codereview.chromium.org/11344011/diff/18001/runtime/vm/flow_graph_opt...
runtime/vm/flow_graph_optimizer.cc:1278:
comp->set_operands_class_id(kNumberCid);
What do we need the kNumberCid-s for (here and below)?

https://codereview.chromium.org/11344011/diff/18001/runtime/vm/intermediate_l...
File runtime/vm/intermediate_language_ia32.cc (right):

https://codereview.chromium.org/11344011/diff/18001/runtime/vm/intermediate_l...
runtime/vm/intermediate_language_ia32.cc:692: BranchInstr* branch) {
Isn't it slower to unbox Mint's for comparison, vs using 4 CPU registers to
compare, especially if we must unbox Mint and Smi when comparing? Please measure
vs current solution where we call out to method with intrinsified prolog.

[Florian Schneider, 2012-10-30 12:09:03]

https://codereview.chromium.org/11344011/diff/18001/runtime/vm/flow_graph_opt...
File runtime/vm/flow_graph_optimizer.cc (right):

https://codereview.chromium.org/11344011/diff/18001/runtime/vm/flow_graph_opt...
runtime/vm/flow_graph_optimizer.cc:1278:
comp->set_operands_class_id(kNumberCid);
On 2012/10/29 19:30:25, srdjan wrote:
> What do we need the kNumberCid-s for (here and below)?

Removed. Actually not needed anymore.

https://codereview.chromium.org/11344011/diff/18001/runtime/vm/intermediate_l...
File runtime/vm/intermediate_language_ia32.cc (right):

https://codereview.chromium.org/11344011/diff/18001/runtime/vm/intermediate_l...
runtime/vm/intermediate_language_ia32.cc:692: BranchInstr* branch) {
On 2012/10/29 19:30:25, srdjan wrote:
> Isn't it slower to unbox Mint's for comparison, vs using 4 CPU registers to
> compare, especially if we must unbox Mint and Smi when comparing? Please
measure
> vs current solution where we call out to method with intrinsified prolog.

Yes, in the worst cast where we have two boxed mints the inlined mint code is
~2x faster. I measured the following function executed in a loop 10^8 times:

foo(a, b) => a > b;

Current VM:

real	0m3.938s
user	0m3.920s
sys	0m0.000s

With my change:

real	0m1.829s
user	0m1.810s
sys	0m0.010s

[Florian Schneider, 2012-10-31 12:23:08]

any further comments? Otherwise I'll land this now.

[srdjan, 2012-10-31 14:39:22]

LGTM