Commit 792752af4e added infrastructure for using AVX-512 intrinsic
functions, and this commit uses that infrastructure to optimize
visibilitymap_count(). Specificially, a new pg_popcount_masked()
function is introduced that applies a bitmask to every byte in the
buffer prior to calculating the population count, which is used to
filter out the all-visible or all-frozen bits as needed. Platforms
without AVX-512 support should also see a nice speedup due to the
reduced number of calls to a function pointer.
Co-authored-by: Ants Aasma
Discussion: https://postgr.es/m/BL1PR11MB5304097DF7EA81D04C33F3D1DCA6A%40BL1PR11MB5304.namprd11.prod.outlook.com
Presently, pg_popcount() processes data in 32-bit or 64-bit chunks
when possible. Newer hardware that supports AVX-512 instructions
can use 512-bit chunks, which provides a nice speedup, especially
for larger buffers. This commit introduces the infrastructure
required to detect compiler and CPU support for the required
AVX-512 intrinsic functions, and it adds a new pg_popcount()
implementation that uses these functions. If CPU support for this
optimized implementation is detected at runtime, a function pointer
is updated so that it is used by subsequent calls to pg_popcount().
Most of the existing in-tree calls to pg_popcount() should benefit
from these instructions, and calls with smaller buffers should at
least not regress compared to v16. The new infrastructure
introduced by this commit can also be used to optimize
visibilitymap_count(), but that is left for a follow-up commit.
Co-authored-by: Paul Amonson, Ants Aasma
Reviewed-by: Matthias van de Meent, Tom Lane, Noah Misch, Akash Shankaran, Alvaro Herrera, Andres Freund, David Rowley
Discussion: https://postgr.es/m/BL1PR11MB5304097DF7EA81D04C33F3D1DCA6A%40BL1PR11MB5304.namprd11.prod.outlook.com
If there aren't many bytes to process, the function call overhead
of the optimized implementation isn't worth taking, so instead we
inline a loop that consults pg_number_of_ones in that case. If
there are many bytes to process, we accept the function call
overhead because the optimized versions are likely to be faster.
The threshold at which we use the optimized implementation is set
to the smallest amount of data required to use special popcount
instructions.
Reviewed-by: Alvaro Herrera, Tom Lane
Discussion: https://postgr.es/m/20240402155301.GA2750455%40nathanxps13
On some systems, calls to pg_popcount{32,64} are indirected through
a function pointer. This commit converts pg_popcount() to a
function pointer on those systems so that we can inline the
appropriate pg_popcount{32,64} implementations into each of the
pg_popcount() implementations. Since pg_popcount() may call
pg_popcount{32,64} several times, this can significantly improve
its performance.
Suggested-by: David Rowley
Reviewed-by: David Rowley
Discussion: https://postgr.es/m/CAApHDvrb7MJRB6JuKLDEY2x_LKdFHwVbogpjZBCX547i5%2BrXOQ%40mail.gmail.com
02a6a54ec added code to make use of the POPCNT instruction when available
for many of our common platforms. Here we do the same for MSVC for x86_64
machines.
MSVC's intrinsic functions for popcnt seem to differ from GCCs in that
they always appear to emit the popcnt instructions. In GCC the behavior
will depend on if the source file was compiled with -mpopcnt or not. For
this reason, the MSVC intrinsic function has been lumped into the
pg_popcount*_asm function, however doing that sort of invalidates the name
of that function, so let's rename it to pg_popcount*_fast().
Author: David Rowley
Reviewed-by: John Naylor
Discussion: https://postgr.es/m/CAApHDvqL3cbbK%3DGzNcwzsNR9Gi%2BaUvTudKkC4XgnQfXirJ_oRQ%40mail.gmail.com
This is still using the 2.0 version of pg_bsd_indent.
I thought it would be good to commit this separately,
so as to document the differences between 2.0 and 2.1 behavior.
Discussion: https://postgr.es/m/16296.1558103386@sss.pgh.pa.us
Test for the compiler builtins __builtin_clz, __builtin_ctz, and
__builtin_popcount, and make use of these in preference to
handwritten C code if they're available. Create src/port
infrastructure for "leftmost one", "rightmost one", and "popcount"
so as to centralize these decisions.
On x86_64, __builtin_popcount generally won't make use of the POPCNT
opcode because that's not universally supported yet. Provide code
that checks CPUID and then calls POPCNT via asm() if available.
This requires indirecting through a function pointer, which is
an annoying amount of overhead for a one-instruction operation,
but it's probably not worth working harder than this for our
current use-cases.
I'm not sure we've found all the existing places that could profit
from this new infrastructure; but we at least touched all the
ones that used copied-and-pasted versions of the bitmapset.c code,
and got rid of multiple copies of the associated constant arrays.
While at it, replace c-compiler.m4's one-per-builtin-function
macros with a single one that can handle all the cases we need
to worry about so far. Also, because I'm paranoid, make those
checks into AC_LINK checks rather than just AC_COMPILE; the
former coding failed to verify that libgcc has support for the
builtin, in cases where it's not inline code.
David Rowley, Thomas Munro, Alvaro Herrera, Tom Lane
Discussion: https://postgr.es/m/CAKJS1f9WTAGG1tPeJnD18hiQW5gAk59fQ6WK-vfdAKEHyRg2RA@mail.gmail.com
This reverts commits fc6c72747a, 109de05cbb, d0b4663c23 and
711bab1e4d.
Somebody will have to try harder before submitting this patch again.
I've spent entirely too much time on it already, and the #ifdef maze yet
to be written in order for it to build at all got on my nerves. The
amount of work needed to get a platform-specific performance improvement
that's barely above the noise level is not worth it.
Split out these new functions in three parts: one in a new file that
uses the compiler builtin and gets compiled with the -mpopcnt compiler
option if it exists; another one that uses the compiler builtin but not
the compiler option; and finally the fallback with open-coded
algorithms.
Split out the configure logic: in the original commit, it was selecting
to use the -mpopcnt compiler switch together with deciding whether to
use the compiler builtin, but those two things are really separate.
Split them out. Also, expose whether the builtin exists to
Makefile.global, so that src/port's Makefile can decide whether to
compile the hw-optimized file.
Remove CPUID test for CTZ/CLZ. Make pg_{right,left}most_ones use either
the compiler intrinsic or open-coded algo; trying to use the
HW-optimized version is a waste of time. Make them static inline
functions.
Discussion: https://postgr.es/m/20190213221719.GA15976@alvherre.pgsql
We were using uint64 function arguments as "long int" arguments to
compiler builtins, which fails on machines where long ints are 32 bits:
the upper half of the uint64 was being ignored. Fix by using the "ll"
builtin variants instead, which on those machines take 64 bit arguments.
Also, remove configure tests for __builtin_popcountl() (as well as
"long" variants for ctz and clz): the theory here is that any compiler
version will provide all widths or none, so one test suffices. Were
this theory to be wrong, we'd have to add tests for
__builtin_popcountll() and friends, which would be tedious.
Per failures in buildfarm member lapwing and ensuing discussion.
These opcodes have been around in the AMD world since 2007, and 2008 in
the case of intel. They're supported in GCC and Clang via some __builtin
macros. The opcodes may be unavailable during runtime, in which case we
fall back on a C-based implementation of the code. In order to get the
POPCNT instruction we must pass the -mpopcnt option to the compiler. We
do this only for the pg_bitutils.c file.
David Rowley (with fragments taken from a patch by Thomas Munro)
Discussion: https://postgr.es/m/CAKJS1f9WTAGG1tPeJnD18hiQW5gAk59fQ6WK-vfdAKEHyRg2RA@mail.gmail.com
Nathan Bossart