rs6000: Add load density heuristic
We noticed that SPEC2017 503.bwaves_r run time degrades by
about 8% on P8 and P9 if we enabled vectorization at O2
fast-math (with cheap vect cost model). Comparing to Ofast,
compiler doesn't do the loop interchange on the innermost
loop, it's not profitable to vectorize it then.
As Richi's comments [1], this follows the similar idea to
over price the vector construction fed by VMAT_ELEMENTWISE
or VMAT_STRIDED_SLP. Instead of adding the extra cost on
vector construction costing immediately, it firstly records
how many loads and vectorized statements in the given loop,
later in rs6000_density_test (called by finish_cost) it
computes the load density ratio against all vectorized
statements, and check with the corresponding thresholds
DENSITY_LOAD_NUM_THRESHOLD and DENSITY_LOAD_PCT_THRESHOLD,
do the actual extra pricing if both thresholds are exceeded.
Note that this new load density heuristic check is based on
some fields in target cost which are updated as needed when
scanning each add_stmt_cost entry, it's independent of the
current function rs6000_density_test which requires to scan
non_vect stmts. Since it's checking the load stmts count
vs. all vectorized stmts, it's kind of density, so I put
it in function rs6000_density_test. With the same reason to
keep it independent, I didn't put it as an else arm of the
current existing density threshold check hunk or before this
hunk.
In the investigation of -1.04% degradation from 526.blender_r
on Power8, I noticed that the extra penalized cost 320 on one
single vector construction for mode V16QI is much exaggerated,
which makes the final body cost unreliable, so this patch adds
one maximum bound for the extra penalized cost for each vector
construction statement.
Full SPEC2017 performance evaluation on Power8/Power9 with
option combinations:
* -O2 -ftree-vectorize {,-fvect-cost-model=very-cheap}
{,-ffast-math}
* {-O3, -Ofast} {,-funroll-loops}
bwaves_r degradations on P8/P9 have been fixed, nothing else
remarkable was observed. Power10 -Ofast -funroll-loops run
shows it's neutral, while -O2 -ftree-vectorize run shows the
bwaves_r degradation is fixed expectedly.
[1] https://gcc.gnu.org/pipermail/gcc-patches/2021-May/570076.html
gcc/ChangeLog:
* config/rs6000/rs6000.c (struct rs6000_cost_data): New members
nstmts, nloads and extra_ctor_cost.
(rs6000_density_test): Add load density related heuristics. Do
extra costing on vector construction statements if need.
(rs6000_init_cost): Init new members.
(rs6000_update_target_cost_per_stmt): New function.
(rs6000_add_stmt_cost): Factor vect_nonmem hunk out to function
rs6000_update_target_cost_per_stmt and call it.