Fix PPC_FP on non-sse4.1 code paths.

[phire]

The Invalid bit on the x87 fpu is sticky, so once a single NaN goes
through the old code on CPUs without sse4.1 all future floats are
mutilated.

Patch to emulate PTEST by Fiora.

Should fix issue 7237 and issue 7510.

[FioraAeterna]

The other possible solution that phire suggested was to clear the x87 FPU flags on the NaN path (so if a NAN appears, the flags get set back accordingly). This has a latency of 22 cycles on Haswell, so it might stall the FPU for quite a while, but if that path is rare it might not matter.

[lioncash]

Could this not be abstracted out into a function called PTEST or something equivalent(considering it is emulating it, may as well toss a concrete name onto it)? This is sort of duplicating code.

[FioraAeterna]

It's only a partial emulation; ptest sets both the C and Z flags while this only partially replicates the functionality (just, only the part this code needs). I mean, it could still be a separate function, it's just not quite generic enough to be a PTEST replacement in full (like the MOVDDUP replacement and similar things in the x64emitter), I think?

[lioncash]

Making it its own function would be sort of nice, because then you can move the m_zero to be a local.

[FioraAeterna]

Also note the only reason I have the m_zero there is because I ran out of registers (I'd need three, I think, unless I'm missing something, to use pxor).

[phire]

Comments fixed to c++ style.

[FioraAeterna]

The "AX could get trashed" comment for ConvertSingleToDouble and such should probably be changed to "EAX could get trashed", I think

[JMC47]

This is a serious issue for users without SSE4.1, is there anything I can do to help test it on my computers?

[pauldacheez]

@JMC47 What I'd do is checkout the branch, find any files with #ifdef _M_SSE4.1 or whatever and put an #undef _M_SSE4.1 at the top of the file before compiling. That's a pretty derpy way to do it, though – you can probably just pass a compiler flag that does the same job better, but I don't know enough about MSVC to tell you what it is.

[FioraAeterna]

You could delete the SSE4 side of the branch in those two functions to force it to run the pre-SSE4 code, maybe?

[Tilka]

lgtm once the comment in the header file is fixed. @JMC47 If you want to test this, #undef _M_SSE4.1 won't change anything, you need to delete the branch.

[phire]

@JMC47 I set the bSSE4_1 variable to false to test the whole jit as these users would experience it. x64CPUDetect.cpp:160

The comment in the header has been fixed. This can be merged if everyone is happy.

[lioncash]

You still haven't answered the code de-duplication suggestion. You can encapsulate a member into a local that way.

EDIT: Never mind, looks good to me.

[phire]

oh, I was hoping everyone had forgotten about that :D

[phire]

Code isn't common, notice one block uses movsd and the other uses movss.