Subzero: Use cmov to improve lowering for the select instruction.

Subzero: Use cmov to improve lowering for the select instruction.

This is instead of explicit control flow which may interfere with branch prediction.  However, explicit control flow is still needed for types other than i16 and i32, due to cmov limitations.

The assembler for cmov is extended to allow the non-dest operand to be a memory operand.

The select lowering is getting large enough that it was in our best interest to combine the default lowering with the bool-folding optimization.

BUG= https://code.google.com/p/nativeclient/issues/detail?id=4095
R=jvoung@chromium.org

Committed: https://gerrit.chromium.org/gerrit/gitweb?p=native_client/pnacl-subzero.git;a=commit;h=537b5ba030eac501734276a39223e2b7a5465ad6

[Jim Stichnoth, 2015-05-18 22:37:05]

stichnot@chromium.org changed reviewers:
+ jvoung@chromium.org, kschimpf@google.com, mtrofin@google.com

[Jim Stichnoth, 2015-05-18 22:38:16]

This improves twolf by 13%, mcf by 5-6%, and a few others by lesser amounts, but
degrades some others by up to 3%.

[jvoung (off chromium), 2015-05-19 16:25:38]

LGTM

https://codereview.chromium.org/1125323004/diff/1/src/IceTargetLoweringX8632.cpp
File src/IceTargetLoweringX8632.cpp (right):

https://codereview.chromium.org/1125323004/diff/1/src/IceTargetLoweringX8632....
src/IceTargetLoweringX8632.cpp:4170: // XXX test select of i1 src operands, also
f32/f64.
Update XXX ?

https://codereview.chromium.org/1125323004/diff/1/src/IceTargetLoweringX8632....
src/IceTargetLoweringX8632.cpp:4184: // mov t, SrcT; cmov_!cond t, SrcF; mov
dest, t
Is it often that both are immediates (does that happen in the benchmarks that
fall behind a bit)?

[Jim Stichnoth, 2015-05-19 16:48:51]

Committed patchset #2 (id:20001) manually as
537b5ba030eac501734276a39223e2b7a5465ad6 (presubmit successful).

[Jim Stichnoth, 2015-05-19 20:41:51]

https://codereview.chromium.org/1125323004/diff/1/src/IceTargetLoweringX8632.cpp
File src/IceTargetLoweringX8632.cpp (right):

https://codereview.chromium.org/1125323004/diff/1/src/IceTargetLoweringX8632....
src/IceTargetLoweringX8632.cpp:4170: // XXX test select of i1 src operands, also
f32/f64.
On 2015/05/19 16:25:37, jvoung wrote:
> Update XXX ?

Done.  (forgot to cleanup my notes...)

https://codereview.chromium.org/1125323004/diff/1/src/IceTargetLoweringX8632....
src/IceTargetLoweringX8632.cpp:4184: // mov t, SrcT; cmov_!cond t, SrcF; mov
dest, t
On 2015/05/19 16:25:38, jvoung wrote:
> Is it often that both are immediates (does that happen in the benchmarks that
> fall behind a bit)?

I counted 1115 instances across all of spec2k:
  egrep 'select i.*, i.* [0-9]+, i.* [0-9]+$' *.ll | wc -l

This pattern accounts for 10-20% of select instructions in each benchmark, and
roughly 0.1% of all bitcode instructions.  Those are static numbers, I don't
know dynamic numbers.