fix digits(n::Unsigned) with neg base for n > typemax(n)÷2
#29205
Conversation
StefanKarpinski
force-pushed
the
sk/digits-unsigned
branch
2 times, most recently
from
731f384
to
0a9bc41
Compare
| # manually peel one loop iteration for type stability | ||
| n, d = fldmod(n, -base) | ||
| a[firstindex(a)] = d | ||
| n = -signed(n) |
There was a problem hiding this comment.
This is type unstable, but maybe the compiler is good enough to handle that, or that the the cost of the divisions makes this negligible?
EDIT: after some very basic benchmarking, this doesn't seem indeed to have any significant performance impact.
There was a problem hiding this comment.
This used to be type unstable but since the introduction of SSAIR the compiler doesn't care about the reuse of variable names at all (that's basically what the SS part means: single static assignment so reassigning amounts to just assigning a different variable). It only matters if the same "slot" in a basic block can have different types, e.g. in the loop body below. Even then the union splitting can deal with it fairly often.
There was a problem hiding this comment.
Thanks for this explanation, that's very good to know, it allows not bothering with this kind of gymnastic!
StefanKarpinski
force-pushed
the
sk/digits-unsigned
branch
from
0a9bc41
to
4fb103a
Compare
Alternative based on #29187 Tests from rforquet's PR linked above.
StefanKarpinski
force-pushed
the
sk/digits-unsigned
branch
from
4fb103a
to
15ff960
Compare
StefanKarpinski
added
backport pending 1.0
bugfix
Alternative based on #29187