Make std.math.isIdentical work in CTFE

[n8sh]

Works because we are in CTFE and there is no way in CTFE to set more bits of NaN payload than can fit in a double, and since 2.087 changed real.init to be non-signaling I think there is no way in CTFE for a real to be a signaling NaN unless real and double have the same representation so real's bits can be manipulated directly.

If the above is not true this should not be approved.

[dlang-bot]

Thanks for your pull request, @n8sh!

Bugzilla references

Auto-close	Bugzilla	Severity	Description
✓	20197	enhancement	Make std.math.isIdentical work in CTFE

Testing this PR locally

If you don't have a local development environment setup, you can use Digger to test this PR:

dub run digger -- build "master + phobos#7515"

[kinke]

In general, it seems the only reason for that function is that DMD includes the padding bytes for real is real. LDC handles this (only comparing the 10 payload bytes for x87 real), so this would work:

bool isIdentical(real a, real b) {
    return a is b;
}

static assert(isIdentical(2, 0x1p1));
static assert(!isIdentical(0.0L, -0.0L));

The comment is misleading as well, as equal but different representations (think 0x2p0, 0x1p1) aren't (bitwise) identical, so it's not just about sign for zero and NaN payloads.

Edit: GDC apparently also just compares the 10 bytes: https://godbolt.org/z/3z3W36

[n8sh]

In general, it seems the only reason for that function is that DMD includes the padding bytes for real is real.

I'd give another reason: is considers all NaNs equivalent regardless of representation.

[kinke]

If you look at the asm, LDC and GDC definitely don't - they use memcmp.

[n8sh]

If you look at the asm, LDC and GDC definitely don't - they use memcmp.

Really? Because assert(real.nan is -real.nan) runs fine with LDC on run.dlang.io and on my computer.

[kinke]

Yes, really, as easily shown in IR or ASM output. When looking at the AST output, you'll see that the frontend 'optimizes' your assert to an assert(true).

[kinke]

The culprit almost certainly being https://github.com/dlang/dmd/blob/2efa5242753e75f88c2f559c5be8adb4584d5e5c/src/dmd/expression.d#L566-L569.

[berni44]

Shouldn't that be x == y?

[n8sh]

Yes

[n8sh]

Ah, now I remember. It doesn't matter: the only circumstance where x is y but not x == y is if both x and y are NaN. So checking x == x is the same as checking x == y (it's a NaN check).

[n8sh]

By the current specification

For struct objects and floating point values, identity is defined as the bits in the operands being identical

so this whole function could in principle be rewritten as return x is y. However as of DMD 2.095 that still isn't true for floating point numbers during CTFE. Possibly related open PR: dlang/dmd#7568

[LightBender]

Sending this to @thewilsonator to see if he can Salvage it.