PowerPC uses itineraries to describe processor pipelines (and dispatch-group
restrictions for P7/P8 cores). Unfortunately, the target-independent
implementation of TII.getInstrLatency calls ItinData->getStageLatency, and that
looks for the largest cycle count in the pipeline for any given instruction.
This, however, yields the wrong answer for the PPC itineraries, because we
don't encode the full pipeline. Because the functional units are fully
pipelined, we only model the initial stages (there are no relevant hazards in
the later stages to model), and so the technique employed by getStageLatency
does not really work. Instead, we should take the maximum output operand
latency, and that's what PPCInstrInfo::getInstrLatency now does.
This caused some test-case churn, including two unfortunate side effects.
First, the new arrangement of copies we get from function parameters now
sometimes blocks VSX FMA mutation (a FIXME has been added to the code and the
test cases), and we have one significant test-suite regression:
SingleSource/Benchmarks/BenchmarkGame/spectral-norm
56.4185% +/- 18.9398%
In this benchmark we have a loop with a vectorized FP divide, and it with the
new scheduling both divides end up in the same dispatch group (which in this
case seems to cause a problem, although why is not exactly clear). The grouping
structure is hard to predict from the bottom of the loop, and there may not be
much we can do to fix this.
Very few other test-suite performance effects were really significant, but
almost all weakly favor this change. However, in light of the issues
highlighted above, I've left the old behavior available via a
command-line flag.
llvm-svn: 242188
isel is actually a cracked instruction on the P7/P8, and must start a dispatch
group. The scheduling model should reflect this so that we don't bunch too many
of them together when possible.
Thanks to Bill Schmidt and Pat Haugen for helping to sort this out.
llvm-svn: 227758
Now that the way that the partial unrolling threshold for small loops is used
to compute the unrolling factor as been corrected, a slightly smaller threshold
is preferable. This is expected; other targets may need to re-tune as well.
llvm-svn: 225566
The P7 benefits from not have really-small loops so that we either have
multiple dispatch groups in the loop and/or the ability to form more-full
dispatch groups during scheduling. Setting the partial unrolling threshold to
44 seems good, empirically, for the P7. Compared to using no late partial
unrolling, this yields the following test-suite speedups:
SingleSource/Benchmarks/Adobe-C++/simple_types_constant_folding
-66.3253% +/- 24.1975%
SingleSource/Benchmarks/Misc-C++/oopack_v1p8
-44.0169% +/- 29.4881%
SingleSource/Benchmarks/Misc/pi
-27.8351% +/- 12.2712%
SingleSource/Benchmarks/Stanford/Bubblesort
-30.9898% +/- 22.4647%
I've speculatively added a similar setting for the P8. Also, I've noticed that
the unroller does not quite calculate the unrolling factor correctly for really
tiny loops because it neglects to account for the fact that not every loop body
replicant contains an ending branch and counter increment. I'll fix that later.
llvm-svn: 225522
The vector divide and sqrt instructions have high latencies, and the scalar
comparisons are like all of the others. On the P7, permutations take an extra
cycle over purely-simple vector ops.
llvm-svn: 205096
CR logicals (crand, crxor, etc.) on the P7 need to be in the first slot of each
dispatch group. The old itinerary entry was just wrong (but has not mattered
because we don't generate these instructions).
This will matter when, in an upcoming commit, we start generating these
instructions.
llvm-svn: 198359
Aside from a few minor latency corrections, the major change here is a new
hazard recognizer which focuses on better dispatch-group formation on the
POWER7. As with the PPC970's hazard recognizer, the most important thing it
does is avoid load-after-store hazards within the same dispatch group. It uses
the POWER7's special dispatch-group-terminating nop instruction (instead of
inserting multiple regular nop instructions). This new hazard recognizer makes
use of the scheduling dependency graph itself, built using AA information, to
robustly detect the possibility of load-after-store hazards.
significant test-suite performance changes (the error bars are 99.5% confidence
intervals based on 5 test-suite runs both with and without the change --
speedups are negative):
speedups:
MultiSource/Benchmarks/FreeBench/pcompress2/pcompress2
-0.55171% +/- 0.333168%
MultiSource/Benchmarks/TSVC/CrossingThresholds-dbl/CrossingThresholds-dbl
-17.5576% +/- 14.598%
MultiSource/Benchmarks/TSVC/Reductions-dbl/Reductions-dbl
-29.5708% +/- 7.09058%
MultiSource/Benchmarks/TSVC/Reductions-flt/Reductions-flt
-34.9471% +/- 11.4391%
SingleSource/Benchmarks/BenchmarkGame/puzzle
-25.1347% +/- 11.0104%
SingleSource/Benchmarks/Misc/flops-8
-17.7297% +/- 9.79061%
SingleSource/Benchmarks/Shootout-C++/ary3
-35.5018% +/- 23.9458%
SingleSource/Regression/C/uint64_to_float
-56.3165% +/- 25.4234%
SingleSource/UnitTests/Vectorizer/gcc-loops
-18.5309% +/- 6.8496%
regressions:
MultiSource/Benchmarks/ASCI_Purple/SMG2000/smg2000
18.351% +/- 12.156%
SingleSource/Benchmarks/Shootout-C++/methcall
27.3086% +/- 14.4733%
llvm-svn: 197099
This adds a scheduling model for the POWER7 (P7) core, and enables the
machine-instruction scheduler when targeting the P7. Scheduling for the P7,
like earlier ooo PPC cores, requires considering both dispatch group hazards,
and functional unit resources and latencies. These are both modeled in a
combined itinerary. Dispatch group formation is still handled by the post-RA
scheduler (which still needs to be updated for the P7, but nevertheless does a
pretty good job).
One interesting aspect of this change is that I've also enabled to use of AA
duing CodeGen for the P7 (just as it is for the embedded cores). The benchmark
results seem to support this decision (see below), and while this is normally
useful for in-order cores, and not for ooo cores like the P7, I think that the
dispatch slot hazards are enough like in-order resources to make the AA useful.
Test suite significant performance differences (where negative is a speedup,
and positive is a regression) vs. the current situation:
MultiSource/Benchmarks/BitBench/drop3/drop3
with AA: N/A
without AA: -28.7614% +/- 19.8356%
(significantly against AA)
MultiSource/Benchmarks/FreeBench/neural/neural
with AA: -17.7406% +/- 11.2712%
without AA: N/A
(significantly in favor of AA)
MultiSource/Benchmarks/SciMark2-C/scimark2
with AA: -11.2079% +/- 1.80543%
without AA: -11.3263% +/- 2.79651%
MultiSource/Benchmarks/TSVC/Symbolics-flt/Symbolics-flt
with AA: -41.8649% +/- 17.0053%
without AA: -34.5256% +/- 23.7072%
MultiSource/Benchmarks/mafft/pairlocalalign
with AA: 25.3016% +/- 17.8614%
without AA: 38.6629% +/- 14.9391%
(significantly in favor of AA)
MultiSource/Benchmarks/sim/sim
with AA: N/A
without AA: 13.4844% +/- 7.18195%
(significantly in favor of AA)
SingleSource/Benchmarks/BenchmarkGame/Large/fasta
with AA: 15.0664% +/- 6.70216%
without AA: 12.7747% +/- 8.43043%
SingleSource/Benchmarks/BenchmarkGame/puzzle
with AA: 82.2713% +/- 26.3567%
without AA: 75.7525% +/- 41.1842%
SingleSource/Benchmarks/Misc/flops-2
with AA: -37.1621% +/- 20.7964%
without AA: -35.2342% +/- 20.2999%
(significantly in favor of AA)
These are 99.5% confidence intervals from 5 runs per configuration. Regarding
the choice to turn on AA during CodeGen, of these results, four seem
significantly in favor of using AA, and one seems significantly against. I'm
not making this decision based on these numbers alone, but these results
seem consistent with results I have from other tests, and so I think that, on
balance, using AA is a win.
llvm-svn: 195981
Hal Finkel