Should we clarify the behavior of 'nsz'?

From LangRef:
"No Signed Zeros - Allow optimizations to treat the sign of a zero argument or result as insignificant."

If we make that like "nnan" or "ninf":
"Allow optimizations to assume the arguments and result are not -0.0. If an argument is -0.0 or the result would be -0.0, it produces a poison value instead."

I'm not sure what the implications of that change would be.

If we keep the current definition that uses "insignificant", is there a way to say that -0.0 is an alias bit-value for +0.0 in Alive2?

Should we clarify the behavior of 'nsz'?

From LangRef:
"No Signed Zeros - Allow optimizations to treat the sign of a zero argument
or result as insignificant."

If we keep the current definition that uses "insignificant", is there a way
to say that -0.0 is an alias bit-value for +0.0 in Alive2?

LangRef is a bit vague indeed. What Alive2 implements is that if the input or output is -0 or +0, then it is changed to be a non-deterministic choice of +/- 0.
This is the closest I can think of +/- 0 bit patterns being an alias. But that doesn't make the example in this report correct, though.
The problem is that the source returns the value of a non-nsz instruction, while the target returns from an nsz instruction. We can't accept an optimization as correct that changes the bit-pattern of the result unless it has some fast-math flag.
(the example would pass if the fadd was nsz)

If we make that like "nnan" or "ninf":
"Allow optimizations to assume the arguments and result are not -0.0. If an
argument is -0.0 or the result would be -0.0, it produces a poison value
instead."

I'm not sure what the implications of that change would be.

Uhm, not sure we could use that nsz for clang's -ffast-math then. It feels a bit extreme to give poison for e.g. 1.0 * -0.0.

The problem is that the source returns the value of a non-nsz instruction,
while the target returns from an nsz instruction. We can't accept an
optimization as correct that changes the bit-pattern of the result unless it
has some fast-math flag.
(the example would pass if the fadd was nsz)

Right - just wanted to decide where to make the logic fix. Proposal:
https://reviews.llvm.org/D79422

https://reviews.llvm.org/rGa954b8a363ad