MIPS: Implement sqrt in inline assembly.

MIPS: Implement sqrt in inline assembly.

Port r18506 (c7b2885)

Original commit message:
Call VSQRT directly to avoid the tiniest (1ulp) precision
error that occurs in the system-supplied sqrt on QNX/ARM.

All precision tests in SunSpider are now passing on this platform.

BUG=

[Paul Lind, 2014-01-09 17:46:44]

Please see the inline comments.

https://codereview.chromium.org/131263003/diff/1/src/mips/codegen-mips.cc
File src/mips/codegen-mips.cc (right):

https://codereview.chromium.org/131263003/diff/1/src/mips/codegen-mips.cc#new...
src/mips/codegen-mips.cc:538: __ SetCallCDoubleArguments(f12);
This is incorrect. For the hard float case it works, since the ABI will pass the
param in f12. But for the SoftFloatABI case, we need to move from a0, a1 (GPR
pair) to f12. 

This function does the opposite since it is meant to prepare for Calling a C
function, and this code is being called FROM a C function with the soft-float
ABI. The ARM code just swapped Result and Arguments calls, since they are
symmetric. But on MIPS, they are not symmetric, so we either need to just code
this inline here, or generate a couple new helper functions. A possible name
might be GetFromCDoubleArguments().

https://codereview.chromium.org/131263003/diff/1/src/mips/codegen-mips.cc#new...
src/mips/codegen-mips.cc:540: __ GetCFunctionDoubleResult(f0);
As above, in SoftFloatABI mode this also does the wrong move for C code calling
this routine, we need to move from f0 to v0,v1 GPR pair. Maybe
SetForCDoubleResult() for the new helper routine? Better names are welcome!

[kilvadyb, 2014-01-09 18:46:09]

https://codereview.chromium.org/131263003/diff/1/src/mips/codegen-mips.cc
File src/mips/codegen-mips.cc (right):

https://codereview.chromium.org/131263003/diff/1/src/mips/codegen-mips.cc#new...
src/mips/codegen-mips.cc:538: __ SetCallCDoubleArguments(f12);
On 2014/01/09 17:46:45, Paul Lind wrote:
> This is incorrect. For the hard float case it works, since the ABI will pass
the
> param in f12. But for the SoftFloatABI case, we need to move from a0, a1 (GPR
> pair) to f12. 
> 
> This function does the opposite since it is meant to prepare for Calling a C
> function, and this code is being called FROM a C function with the soft-float
> ABI. The ARM code just swapped Result and Arguments calls, since they are
> symmetric. But on MIPS, they are not symmetric, so we either need to just code
> this inline here, or generate a couple new helper functions. A possible name
> might be GetFromCDoubleArguments(). 

Done.

https://codereview.chromium.org/131263003/diff/1/src/mips/codegen-mips.cc#new...
src/mips/codegen-mips.cc:540: __ GetCFunctionDoubleResult(f0);
On 2014/01/09 17:46:45, Paul Lind wrote:
> As above, in SoftFloatABI mode this also does the wrong move for C code
calling
> this routine, we need to move from f0 to v0,v1 GPR pair. Maybe
> SetForCDoubleResult() for the new helper routine? Better names are welcome!

Done.

[Paul Lind, 2014-01-09 18:51:15]

LGTM, thanks.

[Paul Lind, 2014-01-09 19:01:14]

Committed as r18523.