ea74a3f548
duplicate_and_interleave is the main fallback way of loading
a repeating sequence of elements into variable-length vectors.
The code handles cases in which the number of elements in the
sequence is potentially several times greater than the number
of elements in a vector.
Let:
- NE be the (compile-time) number of elements in the sequence
- NR be the (compile-time) number of vector results and
- VE be the (run-time) number of elements in each vector
The basic approach is to duplicate each element into a
separate vector, giving NE vectors in total, then use
log2(NE) rows of NE permutes to generate NE results.
In the worst case — when VE has no known compile-time factor
and NR >= NE — all of these permutes are necessary. However,
if VE is known to be a multiple of 2**F, then each of the
first F permute rows produces duplicate results; specifically,
the high permute for a given pair is the same as the low permute.
The code dealt with this by reusing the low result for the
high result. This part was OK.
However, having duplicate results from one row meant that the
next row did duplicate work. The redundancies would be optimised
away by later passes, but the code tried to avoid generating them
in the first place. This is the part that went wrong.
Specifically, NR is typically less than NE when some permutes are
redundant, so the code tried to use NR to reduce the amount of work
performed. The problem was that, although it correctly calculated
a conservative bound on how many results were needed in each row,
it chose the wrong results for anything other than the final row.
This doesn't usually matter for fully-packed SVE vectors. We first
try to coalesce smaller elements into larger ones, so normally
VE ends up being 2**VQ (where VQ is the number of 128-bit blocks
in an SVE vector). In that situation we'd only apply the faulty
optimisation to the final row, i.e. the case it handled correctly.
E.g. for things like:
void
f (long *x)
{
for (int i = 0; i < 100; i += 8)
{
x[i] += 1;
x[i + 1] += 2;
x[i + 2] += 3;
x[i + 3] += 4;
x[i + 4] += 5;
x[i + 5] += 6;
x[i + 6] += 7;
x[i + 7] += 8;
}
}
(already tested by the testsuite), we'd have 3 rows of permutes
producing 4 vector results. The schemne produced:
1st row: 8 results from 4 permutes, highs duplicates of lows
2nd row: 8 results from 8 permutes (half of which are actually redundant)
3rd row: 4 results from 4 permutes
However, coalescing elements is trickier for unpacked vectors,
and at the moment we don't try to do it (see the GET_MODE_SIZE
check in can_duplicate_and_interleave_p). Unpacked vectors
therefore stress the code in ways that packed vectors didn't.
The patch fixes this by removing the redundancies from each row,
rather than trying to work around them later. This also removes
the redundant work in the second row of the example above.
gcc/
PR tree-optimization/98535
* tree-vect-slp.c (duplicate_and_interleave): Use quick_grow_cleared.
If the high and low permutes are the same, remove the high permutes
from the working set and only continue with the low ones.
|
||
|---|---|---|
| c++tools | ||
| config | ||
| contrib | ||
| fixincludes | ||
| gcc | ||
| gnattools | ||
| gotools | ||
| include | ||
| INSTALL | ||
| intl | ||
| libada | ||
| libatomic | ||
| libbacktrace | ||
| libcc1 | ||
| libcody | ||
| libcpp | ||
| libdecnumber | ||
| libffi | ||
| libgcc | ||
| libgfortran | ||
| libgo | ||
| libgomp | ||
| libhsail-rt | ||
| libiberty | ||
| libitm | ||
| libobjc | ||
| liboffloadmic | ||
| libphobos | ||
| libquadmath | ||
| libsanitizer | ||
| libssp | ||
| libstdc++-v3 | ||
| libvtv | ||
| lto-plugin | ||
| maintainer-scripts | ||
| zlib | ||
| .dir-locals.el | ||
| .gitattributes | ||
| .gitignore | ||
| ABOUT-NLS | ||
| ar-lib | ||
| ChangeLog | ||
| ChangeLog.jit | ||
| ChangeLog.tree-ssa | ||
| compile | ||
| config-ml.in | ||
| config.guess | ||
| config.rpath | ||
| config.sub | ||
| configure | ||
| configure.ac | ||
| COPYING | ||
| COPYING3 | ||
| COPYING3.LIB | ||
| COPYING.LIB | ||
| COPYING.RUNTIME | ||
| depcomp | ||
| install-sh | ||
| libtool-ldflags | ||
| libtool.m4 | ||
| lt~obsolete.m4 | ||
| ltgcc.m4 | ||
| ltmain.sh | ||
| ltoptions.m4 | ||
| ltsugar.m4 | ||
| ltversion.m4 | ||
| MAINTAINERS | ||
| Makefile.def | ||
| Makefile.in | ||
| Makefile.tpl | ||
| missing | ||
| mkdep | ||
| mkinstalldirs | ||
| move-if-change | ||
| multilib.am | ||
| README | ||
| symlink-tree | ||
| test-driver | ||
| ylwrap | ||
This directory contains the GNU Compiler Collection (GCC). The GNU Compiler Collection is free software. See the files whose names start with COPYING for copying permission. The manuals, and some of the runtime libraries, are under different terms; see the individual source files for details. The directory INSTALL contains copies of the installation information as HTML and plain text. The source of this information is gcc/doc/install.texi. The installation information includes details of what is included in the GCC sources and what files GCC installs. See the file gcc/doc/gcc.texi (together with other files that it includes) for usage and porting information. An online readable version of the manual is in the files gcc/doc/gcc.info*. See http://gcc.gnu.org/bugs/ for how to report bugs usefully. Copyright years on GCC source files may be listed using range notation, e.g., 1987-2012, indicating that every year in the range, inclusive, is a copyrightable year that could otherwise be listed individually.