Speed up APFloat division by using short division for small divisors. #44051
Conversation
|
r? @arielb1 (rust_highfive has picked a reviewer for you, use r? to override) |
|
@bors p=7 This PR should be tested outside of a rollup for perf.rust-lang.org. |
eddyb
force-pushed
the
apfloat-faster-div
branch
from
757a146 to
b9c69ec
Compare
aidanhs
added
the
S-waiting-on-review
|
@alexcrichton Do you want to review this? It think @arielb1 is unavailable for a few days. |
|
I could give this a rubber stamp but unfortunately I don't think I'm equipped to give this an actual review |
|
r? @nagisa |
| /// One, not zero, based LSB. That is, returns 0 for a zeroed significand. | ||
| pub(super) fn olsb(limbs: &[Limb]) -> usize { | ||
| for i in 0..limbs.len() { | ||
| if limbs[i] != 0 { |
There was a problem hiding this comment.
Why not just iterate the limbs slice directly (with enumerate adapter)?
There was a problem hiding this comment.
With enumerate? That's a transformation that can probably be done in a few places but I didn't do it in the original port.
|
@bors r+ |
|
📌 Commit b9c69ec has been approved by |
|
Oh wow the "rollup" in my previous message confused bors... |
|
@bors rollup- |
Speed up APFloat division by using short division for small divisors. Fixes #43828 (hopefully), by not doing long division bit-by-bit for small divisors. When parsing the ~200,000 decimal float literals in the `tuple-stress` benchmark, this change brings roughly a 5x speed increase (from `0.6s` to `0.12s`), and the hottest instructions are native `div`s.
|
☀️ Test successful - status-appveyor, status-travis |
Fixes #43828 (hopefully), by not doing long division bit-by-bit for small divisors.
When parsing the ~200,000 decimal float literals in the
tuple-stressbenchmark, this change brings roughly a 5x speed increase (from0.6sto0.12s), and the hottest instructions are nativedivs.