On 03/23/2018 01:24 AM, Eli Zaretskii wrote:
> I think we should strive to leave the byte code platform-independent.

That's my thought as well. The byte-compiler was taking the Fortran 
attitude that if an optimization would be allowed on an 
infinite-precision machine, then do it even if it results in different 
answers on real machines due to rounding error or overflow. This loses 
predictability and means that .elc files are not machine-independent. 
Althopugh performance trumps predictability for Fortran, predictability 
is more important for Emacs, as its users will not notice any 
performance loss but they might be affected by the differing answers.

I took a quick look through the byte optimizer and removed the 
Fortran-style floating-point optimizations that I found, by installing 
the attached. This doesn't seem to affect performance significantly on 
Emacs itself.