Hacker News Reader

Comment by saagarjha | Hacker News Reader

Since floating point operations are not associative, I wonder if it is legal for compilers to rearrange them. Maybe that’s what’s happening here?

View on HN

Even if it is legal to rearrange them, the compiler should not unless it is guaranteed to get the same result.

As far as I know, floating point add and multiply are commutative (but not associative) in IEEE 754. And I believe C and C++ allow you to detect if IEEE 754 is used by the hardware and if so, execute the floating point expressions as written.

Note: That is of course no longer true under -ffast-math.

Oops, I meant associative. Sorry about that; I fixed it above.

As far as I know, floating point add and multiply are commutative (but not associative) in IEEE 754.

Usually, but not always. (+0) + (-0) yields (+0), while (-0) + (+0) yields (-0).

Can you cite any references about this? I don’t recall anything like this at all unless perhaps it’s specific to FENV_ACCESS and non-default rounding modes.

In a sister comment, I quoted IEEE 754-2008 Section 6.3 which shows that 0 is commutative in the newer standard. The older revision has similar language.

I can't reproduce this: https://godbolt.org/z/PK4gvZ

It works if the compiler can't do constant folding, like so: https://godbolt.org/z/nQDrSW

Your first line of main is adding -0.0 and -0.0, which is not what I think you intended to show.

Oops, you're right. https://godbolt.org/z/pAwqUA shows it better. Adding a -0 and +0 together in either order gives +0, using the actual x86-64 addsd instruction.

In IEEE 754-1985 and IEEE 754-2008 this is not valid behavior. I have mostly worked on the 2008 revision, so I didn't know if 0 was commutative in the original standard. Thanks for making me look it up.

IEEE 754-2008

> Section 6.3 The sign bit

> When the sum of two operands with opposite signs (or the difference of two operands with like signs) is exactly zero, the sign of that sum (or difference) shall be +0 in all rounding-direction attributes except roundTowardNegative; under that attribute, the sign of an exact zero sum (or difference) shall be −0. However, x + x = x − (−x) retains the same sign as x even when x is zero.

No quote for IEEE 754-1985 because I didn't want to type it out and it is mostly the same.

Huh, interesting. I've definitely seen systems where the sum of two zeroes took the sign of the first value. I assumed they were inheriting 754 behaviour but I guess not.

Reply on news.ycombinator.com