Optimize FastDtoa, in particular Grisu3. In addition to making
a microbenchmark, there are a number of smaller and larger
changes here:
- Replace divisions by power-of-ten with multiplications by
their inverses, using an algorithm very similar to the one
in libdivide.
- For DiyFp::Times(), use 128-bit hardware multiplication
if available (which it generally is on 64-bit platforms).
- Where possible, send around a pointer to the end of the string,
instead of a pointer and a length, reducing register pressure
(especially for Intel). Where not (easily) possible, add
a local variable to make the compiler understand that length
and decimal_point cannot alias.
- Change some ints to unsigneds where it helps us avoid sign
extensions.
- Some minor changes to reduce instruction dependency chains.
- Inline BiggestPowerTen().
Actual performance gain is wildly different between platforms.
On my 3990X workstation (Zen 2), gains are about 21%. On a M1
Mac Mini, they are about 17%. But on my i7-10610U laptop
(Comet Lake, so Skylake microarchitecture), the function is
78% faster. This is probably because large divisions
(divisor over 255) seem to hurt a lot on Skylake, but I haven't
gone through it in detail.
Change-Id: I5b67c257d788a3f7d1be7065d055456852451d68
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/4110741
Commit-Queue: Steinar H Gunderson <sesse@chromium.org>
Reviewed-by: Michael Lippautz <mlippautz@chromium.org>
Cr-Commit-Position: refs/heads/main@{#84906}
6da6e45099