On 1/1/20 7:45 AM, Stefan Monnier wrote:
> We could/should do hash-consing of bignums, tho.  It won't affect code
> that doesn't use bignums, and should have a fairly minor performance
> cost for bignums while making their semantics more clean.

I did that, and surprise! it sped up 'make compile-always' by about 7% on the
two platforms I tried it on (Fedora 31 and Ubuntu 18.04.3, both x86-64).

Although hash-consing of bignums slows down computations that do nothing but
generate new bignums (a slowdown of 38% on my microbenchmark) such computations
are reasonably rare, and the benefit having 'eq' be more-compatible with
traditional Emacs (as well as the abovementioned speedup in more-typical code)
seems to make this tradeoff well worthwhile.

Code patch against Emacs master attached. Given the performance boost, at least
this should go into master; not so sure about emacs-27, though (pro: eq is more
compatible; con: it's a last-minute change to emacs-27).

Two or three issues:

* Should we document that eq == eql on bignums, or continue to leave this stuff
unspecified?

* Should we try hash-consing floats too? Maybe it wouldn't be as slow as we
thought, for typical computations anyway....

* The attached patch could probably be sped up a bit by supporting sets as well
as mappings at the low level, since bignum_map is really just a set of bignums.
Not sure it's worth the effort, though.