Subzero: Build a better bitcast.

Subzero: Build a better bitcast.

The x86 lowering of bitcast between integers and floats forced the transfer through a stack slot (the original implementer *cough* *cough* wasn't aware of the movd instruction).  This requires excess instructions, but also a store to memory followed immediately by a load from that location is very slow.

This fixes the problem by using the movd instruction instead.

BUG= none
R=jpp@chromium.org

Committed: https://gerrit.chromium.org/gerrit/gitweb?p=native_client/pnacl-subzero.git;a=commit;h=a1410df95992d58e06ccdacd62896dd2a7bffb93

[Jim Stichnoth, 2016-06-16 04:26:25]

Description was changed from

==========
Subzero: Build a better bitcast.

BUG= none
==========

to

==========
Subzero: Build a better bitcast.

The x86 lowering of bitcast between integers and floats forced the transfer
through a stack slot (the original implementer *cough* *cough* wasn't aware of
the movd instruction).  This requires excess instructions, but also a store to
memory followed immediately by a load from that location is very slow.

This fixes the problem by using the movd instruction instead.

BUG= none
==========

[Jim Stichnoth, 2016-06-16 04:47:24]

stichnot@chromium.org changed reviewers:
+ eholk@chromium.org, jpp@chromium.org, kschimpf@google.com, manasijm@google.com

[Jim Stichnoth, 2016-06-16 04:49:03]

This gives ~3% improvement on 177.mesa.  The pattern doesn't really affect any
hot blocks in the other spec2k components.

[John, 2016-06-16 13:20:35]

lgtm

[Jim Stichnoth, 2016-06-16 17:02:55]

Description was changed from

==========
Subzero: Build a better bitcast.

The x86 lowering of bitcast between integers and floats forced the transfer
through a stack slot (the original implementer *cough* *cough* wasn't aware of
the movd instruction).  This requires excess instructions, but also a store to
memory followed immediately by a load from that location is very slow.

This fixes the problem by using the movd instruction instead.

BUG= none
==========

to

==========
Subzero: Build a better bitcast.

The x86 lowering of bitcast between integers and floats forced the transfer
through a stack slot (the original implementer *cough* *cough* wasn't aware of
the movd instruction).  This requires excess instructions, but also a store to
memory followed immediately by a load from that location is very slow.

This fixes the problem by using the movd instruction instead.

BUG= none
R=jpp@chromium.org

Committed:
https://gerrit.chromium.org/gerrit/gitweb?p=native_client/pnacl-subzero.git;a...
==========

[Jim Stichnoth, 2016-06-16 17:02:56]

Committed patchset #1 (id:1) manually as
a1410df95992d58e06ccdacd62896dd2a7bffb93 (presubmit successful).