Switch to special quiet NaN as init value for all FP types

[kinke]

Update:

By letting LLVM truncate compile-time real_t values to float/double IR constants. We previously relied on the host, which turns signalling real_t NaNs to quiet float/double NaNs.

Signalling NaNs are characterized by their most-significant mantissa bit being 0 (and at least for D the 2nd-most-significant mantissa bit being set to distinguish them from infinity). The host truncation led to the MSB
getting set while keeping the rest of the mantissa (i.e., both MSBs set).

Here's the resulting bitpattern table in hex:

	float	double	x87 real
init old	7fe00000	7ffc000000000000	7fffa000000000000000
init new	7fa00000	7ff4000000000000	7fffa000000000000000
qnan	7fc00000	7ff8000000000000	7fffc000000000000000

DMD 2.074.0 produces wrongly quiet float/double init values. It's known: https://forum.dlang.org/post/nsp1ql$ivu$1@digitalmars.com

[thewilsonator]

In codegen/vector_init.d on x64 the NaNs in the __initZ symbol seem to initialised to

0x7FF4000000000000

instead of

0x7FFC000000000000

Not sure why, considering it passes on x86. It is the only failure though.

Should we be implementing Target::RealProperties::[s]nan() in C++ as well?

[kinke]

The new value 0x7FF4000000000000 is a signalling NaN, whereas the tested one is a quiet NaN. The inconsistency between 32 and 64 bits is especially ugly (and must be caused by DMD [2.074.0], as that's the first compiler I used to generate the 1.3 LDC bootstrap compiler which then builds 1.3 final).

[kinke]

[Changed the scope of the PR after figuring out what the actual issue is.]

[dnadlinger]

Why did you remove isSNaN? Seems like a gratuitous frontend change at first glance.

[kinke]

To make clear we don't need that buggy version assuming x87 real_t if size != 8.

[dnadlinger]

Maybe at least add a comment saying // buggy; assumes x87 for real.sizeof != 8 or something like that to the version line?

[kinke]

Win32 unexpectedly still fails with quiet double.init.

[kinke]

To be more precise: the Win32 LDC compiler generates quiet ones; the 64-bit build with -m32 signalling ones. Possibly something to do with SSE vs. x87.

[kinke]

In the long term, using quiet FP inits (with the second-most significant mantissa bit set too, to distinguish them from normal quiet NaNs) for all 3 types may make more sense.

[kinke]

Well, why wait => switched to quiet {float,double,real}.init.

[kinke]

Any objections to a consistent special quiet NaN for all FP types?

[JohanEngelen]

can't help judging here, maybe ask on the forums if you're unsure.

[kinke]

No feedback in the forum - merging.