Fix FXTRACT for 0.0 and -0.0

[pmatos]

Fixes fxtract by returning the correct values for 0.0 and -0.0. We moved the split of fxtract into _sig and _exp, to the opcode dispatcher, to ease some comparisons.

Also removed the IR node F80XTRACTStack which is not needed anymore.

Depends on landing #4092 (commit d62dd50 included here but to be merged separately)

[pmatos]

For context the spec mentions that for fscale:

The FSCALE instruction can also be used to reverse the action of the FXTRACT instruction, as shown in the following example:

FXTRACT;
FSCALE;
FSTP ST(1);

When testing this, through FScaleFXtract.asm unit test, I noticed that for 0.0 and -0.0, fxtract was generating incorrect values. I fixed these there.

[pmatos]

All good here now - instcounci failure coming from #4092 that needs to land before this one.

[pmatos]

pinging here for review!
One thing I would like another pair of eyes on is the IR definition of NZCVSelectV. Do we actually need to specify a ResultSize? Can't we just say that the ResultSize is the same size as the TrueVal (or FalseVal). Also maybe we need to validate that these have the same size?

[Sonicadvance1]

Seems reasonable to me, but I'll leave it to Alyssa for the IR change.

[pmatos]

ping @alyssarosenzweig

[alyssarosenzweig]

instead of And+SubNZCV, just use TestNZ to save an instruction

[alyssarosenzweig]

CondClassType {.Val = FEXCore::IR::COND_EQ} can maybe just be {COND_EQ}

[alyssarosenzweig]

+  Ref ExpNZ = _And(OpSize::i64Bit, Gpr, _Constant(0x7ff0000000000000LL));
+  ExpNZ = _Lshr(OpSize::i64Bit, ExpNZ, _Constant(52));

can you combine these two into a _Bfe instruction? (start=52, count=11 or something)

[alyssarosenzweig]

+  Ref SigNZ = _And(OpSize::i64Bit, Gpr, _Constant(0x800fffffffffffffLL));
+  SigNZ = _Or(OpSize::i64Bit, SigNZ, _Constant(0x3ff0000000000000LL));

reexpress as x |( ~(1 << 63)) followed by x & ~(1 << 63). This lets you use constants that will inline, whereas these ones don't, thereby saving a few moves

[JosJuice]

0x800f'ffff'ffff'ffff and 0x3ff0'0000'0000'0000 do successfully inline for logical op instructions. You can see it in the InstCountCI output.

[alyssarosenzweig]

0x800f'ffff'ffff'ffff and 0x3ff0'0000'0000'0000 do successfully inline for logical op instructions. You can see it in the InstCountCI output.

Oops, right. disregard that part!

[alyssarosenzweig]

as commented

[pmatos]

as commented

Thanks for the comments - I think I have addressed them now.