DBC: Fixes typed data bugs. Adds unboxed int32 instructions

DBC: Fixes typed data bugs. Adds unboxed int32 instructions

Typed data indexing was all wrong. It was also wrong to clobber the index register in LoadIndexed.

RSA benchmark: 88s -> 0.7s

R=fschneider@google.com

Committed: https://github.com/dart-lang/sdk/commit/08ff3c039e2f5fe7f15d21368e9a0a72c2b5bbb2

[zra, 2016-08-19 20:20:39]

Description was changed from

==========
DBC: Unboxed (u)int32s
==========

to

==========
DBC: Fixes typed data bugs. Adds unboxed int32 instructions

Typed data indexing was all wrong. It was also wrong to clobber the index
register in LoadIndexed.
==========

[zra, 2016-08-19 20:24:02]

Description was changed from

==========
DBC: Fixes typed data bugs. Adds unboxed int32 instructions

Typed data indexing was all wrong. It was also wrong to clobber the index
register in LoadIndexed.
==========

to

==========
DBC: Fixes typed data bugs. Adds unboxed int32 instructions

Typed data indexing was all wrong. It was also wrong to clobber the index
register in LoadIndexed.

RSA benchmark: 88s -> 0.7s
==========

[zra, 2016-08-19 20:24:19]

zra@google.com changed reviewers:
+ fschneider@google.com, rmacnak@google.com

[Florian Schneider, 2016-08-19 21:39:06]

Lgtm.

https://codereview.chromium.org/2258493004/diff/40001/runtime/vm/simulator_db...
File runtime/vm/simulator_dbc.cc (right):

https://codereview.chromium.org/2258493004/diff/40001/runtime/vm/simulator_db...
runtime/vm/simulator_dbc.cc:1983: FP[rA] = reinterpret_cast<RawObject*>((value
<< 32) >> (32 - kSmiTagSize));
Why is shifting left by 32 and right by 31 necessary? Is the upper half not
always zeroed?

https://codereview.chromium.org/2258493004/diff/40001/runtime/vm/simulator_db...
runtime/vm/simulator_dbc.cc:1990: FP[rA] = reinterpret_cast<RawObject*>((value
<< 32) >> (32 - kSmiTagSize));
Maybe add a local helper SmiTagInt32(intptr_t value)?

[zra, 2016-08-19 22:16:04]

https://codereview.chromium.org/2258493004/diff/40001/runtime/vm/simulator_db...
File runtime/vm/simulator_dbc.cc (right):

https://codereview.chromium.org/2258493004/diff/40001/runtime/vm/simulator_db...
runtime/vm/simulator_dbc.cc:1983: FP[rA] = reinterpret_cast<RawObject*>((value
<< 32) >> (32 - kSmiTagSize));
On 2016/08/19 21:39:06, Florian Schneider wrote:
> Why is shifting left by 32 and right by 31 necessary? Is the upper half not
> always zeroed?
> 

Both the x64 and arm64 implementations zero or sign-extend the upper bits as
appropriate. It looks like on x64, arithmetic could result in non-zero bits in
the upper half, though here and on arm64 the needed IL instructions aren't
implemented. I'll leave this as-is in case me or someone else implements those
IL instructions.

https://codereview.chromium.org/2258493004/diff/40001/runtime/vm/simulator_db...
runtime/vm/simulator_dbc.cc:1990: FP[rA] = reinterpret_cast<RawObject*>((value
<< 32) >> (32 - kSmiTagSize));
On 2016/08/19 21:39:05, Florian Schneider wrote:
> Maybe add a local helper SmiTagInt32(intptr_t value)?

Decided to do casting instead of shifting.

[zra, 2016-08-19 22:23:18]

Description was changed from

==========
DBC: Fixes typed data bugs. Adds unboxed int32 instructions

Typed data indexing was all wrong. It was also wrong to clobber the index
register in LoadIndexed.

RSA benchmark: 88s -> 0.7s
==========

to

==========
DBC: Fixes typed data bugs. Adds unboxed int32 instructions

Typed data indexing was all wrong. It was also wrong to clobber the index
register in LoadIndexed.

RSA benchmark: 88s -> 0.7s

R=fschneider@google.com

Committed:
https://github.com/dart-lang/sdk/commit/08ff3c039e2f5fe7f15d21368e9a0a72c2b5bbb2
==========

[zra, 2016-08-19 22:23:20]

Committed patchset #4 (id:60001) manually as
08ff3c039e2f5fe7f15d21368e9a0a72c2b5bbb2 (presubmit successful).