https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117251
--- Comment #4 from Michael Meissner <meissner at gcc dot gnu.org> ---
I tracked down the commit that first made the slowdown visible:
commit 3a61ca1b9256535e1bfb19b2d46cde21f3908a5d (HEAD)
Author: Jan Hubicka <j...@suse.cz>
Date: Thu Jul 6 18:56:22 2023 +0200
Improve profile updates after loop-ch and cunroll
Extend loop-ch and loop unrolling to fix profile in case the loop is
known to not iterate at all (or iterate few times) while profile claims it
iterates more. While this is kind of symptomatic fix, it is best we can do
incase profile was originally esitmated incorrectly.
In the testcase the problematic loop is produced by vectorizer and I think
vectorizer should know and account into its costs that vectorizer loop
and/or
epilogue is not going to loop after the transformation. So it would be
nice
to fix it on that side, too.
The patch avoids about half of profile mismatches caused by cunroll.
Pass dump id and name |static mismatcdynamic mismatch
|in count |in count
107t cunrolli | 3 +3| 17251 +17251
115t threadfull | 3 | 14376 -2875
116t vrp | 5 +2| 30908 +16532
117t dse | 5 | 30908
118t dce | 3 -2| 17251 -13657
127t ch | 13 +10| 17251
131t dom | 39 +26| 17251
133t isolate-paths | 47 +8| 17251
134t reassoc | 49 +2| 17251
136t forwprop | 53 +4| 202501 +185250
159t cddce | 61 +8| 216211 +13710
161t ldist | 62 +1| 216211
172t ifcvt | 66 +4| 373711 +157500
173t vect | 143 +77| 9802097 +9428386
176t cunroll | 221 +78| 15639591 +5837494
183t loopdone | 218 -3| 15577640 -61951
195t fre | 214 -4| 15577640
197t dom | 213 -1| 16671606 +1093966
199t threadfull | 215 +2| 16879581 +207975
200t vrp | 217 +2| 17077750 +198169
204t dce | 215 -2| 17004486 -73264
206t sink | 213 -2| 17004486
211t cddce | 219 +6| 17005926 +1440
255t optimized | 217 -2| 17005926
256r expand | 210 -7| 19571573 +2565647
258r into_cfglayout | 208 -2| 19571573
275r loop2_unroll | 212 +4| 22992432 +3420859
291r ce2 | 210 -2| 23011838
312r pro_and_epilogue | 230 +20| 23073776 +61938
315r jump2 | 236 +6| 27110534 +4036758
323r bbro | 229 -7| 21826835 -5283699
W/o the patch cunroll does:
176t cunroll | 294 +151|126548439 +116746342
and we end up with 291 mismatches at bbro.
Bootstrapped/regtested x86_64-linux. Plan to commit it after the
scale_loop_frequency patch.
gcc/ChangeLog:
PR middle-end/25623
* tree-ssa-loop-ch.cc (ch_base::copy_headers): Scale loop frequency
to maximal number
of iterations determined.
* tree-ssa-loop-ivcanon.cc (try_unroll_loop_completely): Likewise.
gcc/testsuite/ChangeLog:
PR middle-end/25623
* gfortran.dg/pr25623-2.f90: New test.
However, I backed that particular patch back out of the trunk sources, and it
shows similar regressions.
Here is the scale loop patch which was mentioned above, and is the adjacent
patch. At present, I have not tried backing out this patch:
commit d4c2e34deef8cbd81ba2ef3389fdbaf95c70e225
Author: Jan Hubicka <j...@suse.cz>
Date: Thu Jul 6 18:51:02 2023 +0200
Improve scale_loop_profile
Original scale_loop_profile was implemented to only handle very simple
loops
produced by vectorizer at that time (basically loops with only one exit and
no
subloops). It also has not been updated to new profile-count API very
carefully.
The function does two thigs
1) scales down the loop profile by a given probability.
This is useful, for example, to scale down profile after peeling when
loop
body is executed less often than before
2) update profile to cap iteration count by ITERATION_BOUND parameter.
I changed ITERATION_BOUND to be actual bound on number of iterations as
used elsewhere (i.e. number of executions of latch edge) rather then
number of iterations + 1 as it was before.
To do 2) one needs to do the following
a) scale own loop profile so frquency o header is at most
the sum of in-edge counts * (iteration_bound + 1)
b) update loop exit probabilities so their count is the same
as before scaling.
c) reduce frequencies of basic blocks after loop exit
old code did b) by setting probability to 1 / iteration_bound which is
correctly only of the basic block containing exit executes precisely one
per
iteration (it is not insie other conditional or inner loop). This is fixed
now by using set_edge_probability_and_rescale_others
aldo c) was implemented only for special case when the exit was just before
latch bacis block. I now use dominance info to get right some of addional
case.
I still did not try to do anything for multiple exit loops, though the
implementatoin could be generalized.
Bootstrapped/regtested x86_64-linux. Plan to cmmit it tonight if there
are no complains.
gcc/ChangeLog:
* cfgloopmanip.cc (scale_loop_profile): Rewrite exit edge
probability update to be safe on loops with subloops.
Make bound parameter to be iteration bound.
* tree-ssa-loop-ivcanon.cc (try_peel_loop): Update call
of scale_loop_profile.
* tree-vect-loop-manip.cc (vect_do_peeling): Likewise.
