This is an extension of a transform that tries to produce positive floating-point constants to improve canonicalization (and hopefully lead to more reassociation and CSE).
But as the test diffs show, these were limited to basic patterns by walking from an instruction to its single user rather than recursively moving up the def-use sequence. No fast-math is required here because we're only rearranging implicit FP negations in intermediate ops.
A motivating bug is:
...but unfortunately the transform in that example gets inverted in the backend, so we end up with the same code. I'll investigate that case as a follow-up, but I think the changes here can be viewed as an independent improvement.