Lower the fcmp instruction for <4 x float> operands.

Lower the fcmp instruction for <4 x float> operands.

Most fcmp conditions map directly to single x86 instructions. For
these, the lowering is table driven.

BUG=none
R=jvoung@chromium.org, stichnot@chromium.org

Committed: https://gerrit.chromium.org/gerrit/gitweb?p=native_client/pnacl-subzero.git;a=commit;h=ce0ca8f

[Jim Stichnoth, 2014-07-24 16:16:20]

https://codereview.chromium.org/413053002/diff/40001/src/IceTargetLoweringX86...
File src/IceTargetLoweringX8632.def (right):

https://codereview.chromium.org/413053002/diff/40001/src/IceTargetLoweringX86...
src/IceTargetLoweringX8632.def:38: #define VECTOR_FCMPX8632_TABLE      \
Can you just add swap and cmpps_predicate columns to the scalar x-macros table,
instead of defining a new table?

You could also consider combining the two concrete tables in
IceTargetLoweringX8632.cpp, but that's more a matter of preference.

[wala, 2014-07-24 17:13:24]

https://codereview.chromium.org/413053002/diff/40001/src/IceTargetLoweringX86...
File src/IceTargetLoweringX8632.def (right):

https://codereview.chromium.org/413053002/diff/40001/src/IceTargetLoweringX86...
src/IceTargetLoweringX8632.def:38: #define VECTOR_FCMPX8632_TABLE      \
On 2014/07/24 16:16:19, stichnot wrote:
> Can you just add swap and cmpps_predicate columns to the scalar x-macros
table,
> instead of defining a new table?
> 
> You could also consider combining the two concrete tables in
> IceTargetLoweringX8632.cpp, but that's more a matter of preference.

Done.

Also used a single table for IceTargetLoweringX8632.cpp.

[Jim Stichnoth, 2014-07-24 17:29:36]

LGTM with bikeshedding.

https://codereview.chromium.org/413053002/diff/80001/src/IceTargetLoweringX86...
File src/IceTargetLoweringX8632.def (right):

https://codereview.chromium.org/413053002/diff/80001/src/IceTargetLoweringX86...
src/IceTargetLoweringX8632.def:19: /*       scalar comparison             vector
comparison */  \
In the pursuit of beautiful ASCII art, you could do something like:

/* <---- scalar comparison ----> ... */

https://codereview.chromium.org/413053002/diff/80001/src/IceTargetLoweringX86...
src/IceTargetLoweringX8632.def:37: //#define X(val, dflt, swap, C1, C2, swap2,
pred)
Instead of swap and swap2, how about swapS and swapV?

This would also apply to the comment above, and wherever FCMPX8632_TABLE is
used.

[jvoung (off chromium), 2014-07-24 18:39:24]

lgtm

[wala, 2014-07-24 19:33:34]

https://codereview.chromium.org/413053002/diff/80001/src/IceTargetLoweringX86...
File src/IceTargetLoweringX8632.def (right):

https://codereview.chromium.org/413053002/diff/80001/src/IceTargetLoweringX86...
src/IceTargetLoweringX8632.def:19: /*       scalar comparison             vector
comparison */  \
On 2014/07/24 17:29:36, stichnot wrote:
> In the pursuit of beautiful ASCII art, you could do something like:
> 
> /* <---- scalar comparison ----> ... */

Done.

https://codereview.chromium.org/413053002/diff/80001/src/IceTargetLoweringX86...
src/IceTargetLoweringX8632.def:37: //#define X(val, dflt, swap, C1, C2, swap2,
pred)
On 2014/07/24 17:29:36, stichnot wrote:
> Instead of swap and swap2, how about swapS and swapV?
> 
> This would also apply to the comment above, and wherever FCMPX8632_TABLE is
> used.

Done.

[wala, 2014-07-24 19:34:24]

Committed patchset #6 manually as rce0ca8f (presubmit successful).