[<prev] [next>] [day] [month] [year] [list]
Message-ID: <7582613-e06a-eeed-a6be-c6dd6159e5d@esi.com.au>
Date: Sun, 1 Feb 2026 10:55:08 +1100 (AEDT)
From: Damian McGuckin <damianm@....com.au>
To: MUSL <musl@...ts.openwall.com>
Subject: Slightly faster 'scalbn(x,i) = x*2**i'
In the course of helping a colleage debug a problem with ldexp/scalbn in
Apple's LIBM for ARM, I ported a custom version of scalbn() written in
Chapel to C. The resulting C code was pretty much like MUSL for 2**i being
a normal number (or a number greater than the largest floating point
number). But, for 2**i being subnormal (or effectively even smaller) and
|x| being normal, it avoids the multiplications seen in MUSL.
The X86-64 assembler code size from both GCC11 and CLANG19 is much the
same. Just for fun, I timed it against MUSL's routine and it is about 30%
faster over a broad range of input.
As MUSL's scalbn() is not a performance hog and it easy to read, I see no
reason to suggest changing it.
Let me know if anybody thinks otherwise.
Thanks - Damian
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
