Subzero: Use "pxor reg,reg" to load a floating-point scalar 0.0 value.

Subzero: Use "pxor reg,reg" to load a floating-point scalar 0.0 value.

BUG= none
R=jpp@chromium.org, sehr@chromium.org

Committed: https://gerrit.chromium.org/gerrit/gitweb?p=native_client/pnacl-subzero.git;a=commit;h=991656678c7c5d661093d9ca63d83e20afc9e468

[Jim Stichnoth, 2015-11-13 07:45:38]

Description was changed from

==========
Subzero: Use "pxor reg,reg" to load a floating-point scalar 0.0 value.

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

to

==========
Subzero: Use "pxor reg,reg" to load a floating-point scalar 0.0 value.

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

[Jim Stichnoth, 2015-11-13 07:45:39]

stichnot@chromium.org changed reviewers:
+ sehr@google.com

[Jim Stichnoth, 2015-11-13 07:48:07]

David - this is a prototype for the idea you had.  Small as it is, it could use
some cleanup.  Perf data is noisy as usual, but this suggests a 0.5% (or less)
geomean improvement in spec2k, with 5% on twolf and 3% on vortex.

[sehr (please use chromium), 2015-11-13 16:22:48]

LGTM.

Btw, did you ever find out the difference between pxor and xorp[sd]?  I intended
to look some more.

David

[Jim Stichnoth, 2015-11-13 21:42:17]

stichnot@chromium.org changed reviewers:
+ jpp@chromium.org, kschimpf@google.com, sehr@chromium.org
- sehr@google.com

[sehr, 2015-11-13 21:49:13]

One comment.  LGTM.

https://codereview.chromium.org/1439363002/diff/20001/tests_lit/llvm2ice_test...
File tests_lit/llvm2ice_tests/undef.ll (right):

https://codereview.chromium.org/1439363002/diff/20001/tests_lit/llvm2ice_test...
tests_lit/llvm2ice_tests/undef.ll:41: ; xCHECK: fld DWORD PTR {{.*}}
.L$float$00000000
Delete this.

[John, 2015-11-13 21:51:20]

lgtm

https://codereview.chromium.org/1439363002/diff/20001/src/IceTargetLoweringX8...
File src/IceTargetLoweringX86BaseImpl.h (right):

https://codereview.chromium.org/1439363002/diff/20001/src/IceTargetLoweringX8...
src/IceTargetLoweringX86BaseImpl.h:5327: _pxor(Reg, Reg);
What's the difference of pxor and xorp*?

https://codereview.chromium.org/1439363002/diff/20001/src/IceTargetLoweringX8...
src/IceTargetLoweringX86BaseImpl.h:5592: if (Value == 0 && !signbit(Value))
Optional: This is a good candidate for a little helper function

template <T> isPositiveZero(T Const) {
   ...
}

[Jim Stichnoth, 2015-11-13 22:20:24]

https://codereview.chromium.org/1439363002/diff/20001/src/IceTargetLoweringX8...
File src/IceTargetLoweringX86BaseImpl.h (right):

https://codereview.chromium.org/1439363002/diff/20001/src/IceTargetLoweringX8...
src/IceTargetLoweringX86BaseImpl.h:5327: _pxor(Reg, Reg);
On 2015/11/13 21:51:20, John wrote:
> What's the difference of pxor and xorp*?

We were trying to figure that out.  The Intel guys say that pxor may be better
when the result is to be used in a vector instruction, and xorps/xorpd may be
better for use in a scalar FP instruction, due to micro-architectural black
magic.

https://codereview.chromium.org/1439363002/diff/20001/src/IceTargetLoweringX8...
src/IceTargetLoweringX86BaseImpl.h:5592: if (Value == 0 && !signbit(Value))
On 2015/11/13 21:51:19, John wrote:
> Optional: This is a good candidate for a little helper function
> 
> template <T> isPositiveZero(T Const) {
>    ...
> }

Nice, done.

https://codereview.chromium.org/1439363002/diff/20001/tests_lit/llvm2ice_test...
File tests_lit/llvm2ice_tests/undef.ll (right):

https://codereview.chromium.org/1439363002/diff/20001/tests_lit/llvm2ice_test...
tests_lit/llvm2ice_tests/undef.ll:41: ; xCHECK: fld DWORD PTR {{.*}}
.L$float$00000000
On 2015/11/13 21:49:13, sehr (please use this account) wrote:
> Delete this.

Oops, done.

[Jim Stichnoth, 2015-11-13 22:20:45]

Committed patchset #3 (id:40001) manually as
991656678c7c5d661093d9ca63d83e20afc9e468 (presubmit successful).