Harmony: optimize Math.clz32.

Harmony: optimize Math.clz32.

R=svenpanne@chromium.org
BUG=v8:2938
LOG=N

Committed: https://code.google.com/p/v8/source/detail?r=19487

[m.m.capewell, 2014-02-19 12:24:25]

https://codereview.chromium.org/172133003/diff/1/src/a64/lithium-codegen-a64.cc
File src/a64/lithium-codegen-a64.cc (right):

https://codereview.chromium.org/172133003/diff/1/src/a64/lithium-codegen-a64....
src/a64/lithium-codegen-a64.cc:3857: Register result =
ToRegister(instr->result());
The input and result should use ToRegister32, otherwise the number of leading
zeroes may be 32 more than expected.

[Yang, 2014-02-19 12:28:02]

On 2014/02/19 12:24:25, m.m.capewell wrote:
>
https://codereview.chromium.org/172133003/diff/1/src/a64/lithium-codegen-a64.cc
> File src/a64/lithium-codegen-a64.cc (right):
> 
>
https://codereview.chromium.org/172133003/diff/1/src/a64/lithium-codegen-a64....
> src/a64/lithium-codegen-a64.cc:3857: Register result =
> ToRegister(instr->result());
> The input and result should use ToRegister32, otherwise the number of leading
> zeroes may be 32 more than expected.

Thanks. Updated.

[Sven Panne, 2014-02-19 13:24:08]

https://codereview.chromium.org/172133003/diff/110001/src/a64/lithium-codegen...
File src/a64/lithium-codegen-a64.cc (right):

https://codereview.chromium.org/172133003/diff/110001/src/a64/lithium-codegen...
src/a64/lithium-codegen-a64.cc:3858: __ Clz(result, input);
Hmmm, can we mix W and X registers in a single instruction? The intention seems
to be "count in 32bits and silently expand the result to 64bit", no clue if this
is possible.

https://codereview.chromium.org/172133003/diff/110001/src/ia32/lithium-codege...
File src/ia32/lithium-codegen-ia32.cc (right):

https://codereview.chromium.org/172133003/diff/110001/src/ia32/lithium-codege...
src/ia32/lithium-codegen-ia32.cc:4134: __ Set(result, Immediate(63));  // 64^31
== 32
63^31=32

https://codereview.chromium.org/172133003/diff/110001/src/x64/lithium-codegen...
File src/x64/lithium-codegen-x64.cc (right):

https://codereview.chromium.org/172133003/diff/110001/src/x64/lithium-codegen...
src/x64/lithium-codegen-x64.cc:3739: __ Set(result, 63);  // 64^31 == 32
63^31=32

[Yang, 2014-02-19 13:39:47]

On 2014/02/19 13:24:08, Sven Panne wrote:
>
https://codereview.chromium.org/172133003/diff/110001/src/a64/lithium-codegen...
> File src/a64/lithium-codegen-a64.cc (right):
> 
>
https://codereview.chromium.org/172133003/diff/110001/src/a64/lithium-codegen...
> src/a64/lithium-codegen-a64.cc:3858: __ Clz(result, input);
> Hmmm, can we mix W and X registers in a single instruction? The intention
seems
> to be "count in 32bits and silently expand the result to 64bit", no clue if
this
> is possible.
> 
>
https://codereview.chromium.org/172133003/diff/110001/src/ia32/lithium-codege...
> File src/ia32/lithium-codegen-ia32.cc (right):
> 
>
https://codereview.chromium.org/172133003/diff/110001/src/ia32/lithium-codege...
> src/ia32/lithium-codegen-ia32.cc:4134: __ Set(result, Immediate(63));  //
64^31
> == 32
> 63^31=32
> 
>
https://codereview.chromium.org/172133003/diff/110001/src/x64/lithium-codegen...
> File src/x64/lithium-codegen-x64.cc (right):
> 
>
https://codereview.chromium.org/172133003/diff/110001/src/x64/lithium-codegen...
> src/x64/lithium-codegen-x64.cc:3739: __ Set(result, 63);  // 64^31 == 32
> 63^31=32

My bad. The debug build actually gets assertion failures for A64. Revised
version uploaded.

[Sven Panne, 2014-02-19 13:42:02]

lgtm

[m.m.capewell, 2014-02-19 13:45:30]

On 2014/02/19 13:24:08, Sven Panne wrote:
>
https://codereview.chromium.org/172133003/diff/110001/src/a64/lithium-codegen...
> src/a64/lithium-codegen-a64.cc:3858: __ Clz(result, input);
> Hmmm, can we mix W and X registers in a single instruction? The intention
seems
> to be "count in 32bits and silently expand the result to 64bit", no clue if
this
> is possible.

In most cases, you can't mix W and X registers in the same instruction. When you
can, you have to state how the W is interpreted as an X, ie. sxtw, uxtw.

[Yang, 2014-02-19 13:52:08]

Committed patchset #4 manually as r19487 (presubmit successful).