[Bug libstdc++/113811] New: std::rotate does 64-bit signed division

[Bug libstdc++/113811] New: std::rotate does 64-bit signed division

[terra at gnome dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113811

            Bug ID: 113811
           Summary: std::rotate does 64-bit signed division
           Product: gcc
           Version: 13.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: libstdc++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: terra at gnome dot org
  Target Milestone: ---

In stl_algo.h, function __rotate for RandomAccessIterator lines 1280-1362 for
me, there are two divisions of integers:

__n %= __k;

on lines 1332 and 1356.  They look harmless.

But in the common case on x86_64 where _Distance is, essentially, int64_t this
is a 64-bit signed division which is absurdly slow.

By my reading of https://www.agner.org/optimize/instruction_tables.pdf page
296:

64-bit signed: 57 cycles
64-bit unsigned: 36 cycles
smaller sizes: 10 cycles

(excluding the 64-to-128-bit sign extension needed too)

I believe the numbers involved are all positive, so at the very least the
division could be unsigned.  It might even make sense to check if __n is
smaller than 2^32 and do a 32-bit division instead.

Somewhat related: bug 102580

Note: I do not actually have benchmark results that show this matters in a
practical case.

[Bug libstdc++/113811] std::rotate does 64-bit signed division

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113811

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
In case __n is the minimum signed integer and __k is -1 the division would also
trap ;)  So yes, they should be unsigned.

[Bug libstdc++/113811] std::rotate does 64-bit signed division

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113811

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2024-02-08
     Ever confirmed|0                           |1

--- Comment #2 from Jonathan Wakely <redi at gcc dot gnu.org> ---
They're never negative unless the arguments to the algorithm are incorrect, in
which case maybe trapping is better than trashing the data!

I was going to suggest that we could just add an assertion and/or hint that
they're never negative, but I guess PR 102580 means that wouldn't help.

[Bug libstdc++/113811] std::rotate does 64-bit signed division

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113811

--- Comment #3 from Jonathan Wakely <redi at gcc dot gnu.org> ---
It seems fairly easy to do:

commit 12a028d76bbdf26d34d4d90a2ecdc39c6c0a4bd4 (HEAD -> master)
Author: Jonathan Wakely
Date:   Thu Feb 8 15:40:32 2024

    libstdc++: Use unsigned division in std::rotate [PR113811]

    Signed 64-bit division is much slower than unsigned, so cast the n and
    k values to unsigned before doing n %= k. We know this is safe because
    neither value can be negative.

    libstdc++-v3/ChangeLog:

            PR libstdc++/113811
            * include/bits/stl_algo.h (__rotate): Use unsigned values for
            division.

diff --git a/libstdc++-v3/include/bits/stl_algo.h
b/libstdc++-v3/include/bits/stl_algo.h
index 9496b53f887..7a0cf6b6737 100644
--- a/libstdc++-v3/include/bits/stl_algo.h
+++ b/libstdc++-v3/include/bits/stl_algo.h
@@ -1251,6 +1251,12 @@ _GLIBCXX_BEGIN_INLINE_ABI_NAMESPACE(_V2)
       typedef typename iterator_traits<_RandomAccessIterator>::value_type
        _ValueType;

+#if __cplusplus >= 201103L
+      typedef typename make_unsigned<_Distance>::type _UDistance;
+#else
+      typedef _Distance _UDistance;
+#endif
+
       _Distance __n = __last   - __first;
       _Distance __k = __middle - __first;

@@ -1281,7 +1287,7 @@ _GLIBCXX_BEGIN_INLINE_ABI_NAMESPACE(_V2)
                  ++__p;
                  ++__q;
                }
-             __n %= __k;
+             __n = static_cast<_UDistance>(__n) %
static_cast<_UDistance>(__k);
              if (__n == 0)
                return __ret;
              std::swap(__n, __k);
@@ -1305,7 +1311,7 @@ _GLIBCXX_BEGIN_INLINE_ABI_NAMESPACE(_V2)
                  --__q;
                  std::iter_swap(__p, __q);
                }
-             __n %= __k;
+             __n = static_cast<_UDistance>(__n) %
static_cast<_UDistance>(__k);
              if (__n == 0)
                return __ret;
              std::swap(__n, __k);


Conditionally using 32-bit types would be a bit trickier, as it needs runtime
branches, or making the type of __n and __k a template parameter, so we can
call __rotate_with<unsigned> to use a smaller type than
make_unsigned<_Distance> if max(n,k) < UINT_MAX.

[Bug libstdc++/113811] std::rotate does 64-bit signed division

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113811

--- Comment #4 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jonathan Wakely <redi@gcc.gnu.org>:

https://gcc.gnu.org/g:4d819db7f229a23cb15ef68f310e0bb51d201c45

commit r14-9001-g4d819db7f229a23cb15ef68f310e0bb51d201c45
Author: Jonathan Wakely <jwakely@redhat.com>
Date:   Thu Feb 8 15:40:32 2024 +0000

    libstdc++: Use unsigned division in std::rotate [PR113811]

    Signed 64-bit division is much slower than unsigned, so cast the n and
    k values to unsigned before doing n %= k. We know this is safe because
    neither value can be negative.

    libstdc++-v3/ChangeLog:

            PR libstdc++/113811
            * include/bits/stl_algo.h (__rotate): Use unsigned values for
            division.

[Bug libstdc++/113811] std::rotate does 64-bit signed division

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113811

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|NEW                         |RESOLVED
   Target Milestone|---                         |14.0

--- Comment #5 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Fixed for gcc-14, thanks for the suggestion.