[x64] simply tweak materialization of float/double constants

[x64] simply tweak materialization of float/double constants

port 24485
Fixed a bug of "psllq" instruction in x64

R=bmeurer@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=24673

[Weiliang, 2014-10-16 00:52:37]

weiliang.lin@intel.com changed reviewers:
+ bmeurer@chromium.org, danno@chromium.org, dcarney@chromium.org,
titzer@chromium.org

[Weiliang, 2014-10-16 00:52:38]

PTAL

[Benedikt Meurer, 2014-10-16 05:35:21]

How about porting the MacroAssembler::Move(XMMRegister, intX_t) functions to X64
instead?

[Weiliang, 2014-10-16 08:01:09]

On 2014/10/16 05:35:21, Benedikt Meurer wrote:
> How about porting the MacroAssembler::Move(XMMRegister, intX_t) functions to
X64
> instead?

How about the patch set2? 

I did not port the case of (nlz + cnt + ntz == 32 or nlz + cnt + ntz == 64)
eagerly, because X64 has a dedicated scratch register.

A better solution maybe place all double/float constants of a function at a
constant pool, so that ia32/x64 could load them into SSE register by only one
movsd(xmm, mem) without any temporary register.

[Benedikt Meurer, 2014-10-16 08:20:28]

Yeah, long-term plan is to use constant pool. Still, isn't it faster to
materialize inside SSE register?

[Weiliang, 2014-10-17 04:02:08]

On 2014/10/16 08:20:28, Benedikt Meurer wrote:
> Yeah, long-term plan is to use constant pool. Still, isn't it faster to
> materialize inside SSE register?

OK, i did it. 

Please see the patch set 3 and help land it.

[Benedikt Meurer, 2014-10-17 04:18:22]

LGTM, will land, thanks.

[Benedikt Meurer, 2014-10-17 04:18:52]

Committed patchset #3 (id:40001) manually as 24673 (presubmit successful).