Floating-point constant folding in Emacs byte compiler

[Paul Eggert]

While preparing and installing the attached patch to Emacs master, I 
noticed that the byte optimizer assumes that floating-point arithmetic 
behaves the same at compile-time that it does at runtime. For example, 
on my x86-64 platform the byte compiler optimizes (/ 0 0.0) to -NaN, (- 
1 1.0000000000000001) to 0.0, and (< 1 1.0000000000000001) to nil, even 
though these expressions will evaluate to different values on (say) an 
IBM mainframe that uses IBM floating point, and the first expression 
yields +NaN on some IEEE platforms (e.g., ARM).

These discrepancies mean that .elc files containing floating-point 
constants might not be platform-independent, in that byte-compiling a 
file on one machine X and running it on another machine Y can yield 
different results than byte-compiling and running the same file on Y. Is 
this sort of discrepancy intended? If so, should it be documented in the 
Emacs Lisp manual? On the one hand, I doubt whether this sort of 
optimization buys us much performance; on the other, I doubt whether 
many users care about the discrepancies.

0001-Fix-byte-opt-lists-of-pure-functions-etc.txt
Description: Text document