Uh oh!
There was an error while loading. Please reload this page.
bpo-29782: Consolidate _Py_Bit_Length() - #20739
Conversation
vstinner
left a comment
There was a problem hiding this comment.
Maybe add test_bit_length() to _testinternalcapi.c: see test_popcount().
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
vstinner
commented
Jun 8, 2020
Previous rejected attempt: PR #594. |
Uh oh!
There was an error while loading. Please reload this page.
vstinner
commented
Jun 8, 2020
cc @mdickinson |
Uh oh!
There was an error while loading. Please reload this page.
vstinner
commented
Jun 8, 2020
The Windows build looks broken: |
In pythonGH-2866, _Py_Bit_Length() was added to pymath.h for lack of a better location. pythonGH-20518 added a more appropriate header file for bit utilities. It also shows how to properly use intrinsics. This allows reconsidering bpo-29782. * Move the function to the new header. * Changed return type to match __builtin_clzl and reviewed usage. * Use intrinsics where available. * Pick a (mostly theoretical) fallback implementation suitable for inlining.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Co-authored-by: Victor Stinner <vstinner@python.org>
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
niklasf
commented
Jun 8, 2020
Thanks for reviewing. Also added unit tests as suggested (in addition to the existing test coverage). I briefly disabled the branch with clz (8e3ce66) to get a CI run with the fallback implementation. The current fallback is identical to the original (except for inlining). |
vstinner
left a comment
There was a problem hiding this comment.
Thanks for the many updates! The code now looks better! New review.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Co-authored-by: Victor Stinner <vstinner@python.org>
Co-authored-by: Victor Stinner <vstinner@python.org>
vstinner
commented
Jun 8, 2020
Using gcc 10 -O3 -flto, I get the following machine code in test_bit_length(): x86 BSR (Bit Scan Reverse) instruction is used as expected. |
vstinner
left a comment
There was a problem hiding this comment.
LGTM!
@mdickinson: Do you want to review this PR as well?
mdickinson
commented
Jun 9, 2020
I'd very much like to take a look, but am short on available time. Can you give me until the weekend? If I've failed to review by the start of next week, go ahead and merge. |
vstinner
commented
Jun 9, 2020
@mdickinson: this PR can wait one week :-) |
| static int | ||
| bit_length_digit(digit x) | ||
| { | ||
| Py_BUILD_ASSERT(PyLong_SHIFT <= sizeof(unsigned long) * 8); |
There was a problem hiding this comment.
It would be nice if C provided some standard way to get the width (in the sense of C99 §6.2.6.2p6) of an integer type, so that we didn't have to use sizeof(unsigned long) * 8, which could fail in the presence of padding bits. But as far as I know it doesn't. Related: https://stackoverflow.com/q/3957252/270986
vstinner
commented
Jun 15, 2020
Thanks @niklasf, I merged your PR. I removed "(mostly theoretical)" from your commit message: We support other C compilers, like XLC. |
In GH-2866, _Py_Bit_Length() was added to pymath.h for lack of a better
location. GH-20518 added a more appropriate header file for bit utilities. It
also shows how to properly use intrinsics. This allows reconsidering bpo-29782.
inlining.
https://bugs.python.org/issue29782