bug#42147: 28.0.50; pure vs side-effect-free, missing optimizations?

[Mattias Engdegård]

3 juli 2020 kl. 01.16 skrev Paul Eggert <eggert@cs.ucla.edu>:

> As you can see, sometimes SSE2 is closer to the mathematically-correct answer
> and sometimes x87 is. In typical C math code, x87 is better; in Emacs I 
> imagine
> the reverse is true (for the reasons you mentioned), though I have not 
> attempted
> to measure this.

Thanks for the examples and these were indeed what I had in mind (there's also 
the effect from having a greater exponent range in the intermediate result); 
Monniaux [1] is a good reference.

In practice, the extra precision of x87 code is so unreliable and fickle 
(unless the 80-bit long double is used throughout) that it's almost never worth 
it. (Being much slower doesn't help either.)

Fortunately modern compilers generate SSE code by default, only passing return 
values on the x87 stack as per the x86 ABI (which causes no harm). This reduces 
an already tiny risk to nil. We could add an elaborate configure or run-time 
test and admonishments to the installation instructions but frankly we have 
better use of our time. I suggest we replace byte-opt--portable-numberp with 
numberp (or nothing at all, depending on where it occurs) and be done with it.

---
[1] 
https://hal.archives-ouvertes.fr/hal-00128124/file/floating-point-article.pdf