https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79506
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, so it's IVOPTs becoming relatively slower fastest (with --param
inline-min-speedup=8 it surpasses IRA as the most expensive pass). 34% of IVO
is
spent in simple_iv (and siblings), the other 34% are spent in
determine_group_iv_cost.
Overall it seems like just the regular "larger functions compile slower".
inline change (look at the weird -0.0 badness?!)
--- tfft2.f90.078i.inline.10 2017-02-14 12:06:39.638906303 +0100
+++ tfft2.f90.078i.inline.8 2017-02-14 12:06:49.199057817 +0100
@@ -4940,12 +4940,62 @@
Processing frequency rfft
Called by tfft that is executed once
Called in loop
- not inlinable: tfft/3 -> rfft/1, --param max-inline-insns-auto limit reached
+ decreasing badness tfft/3 -> rfft/1, -0.000000 to -0.000000
Estimating body: tfft/3
Known to be false: not inlined
size:590 time:24774
Inlined into tfft which now has time 24776 and size 593, net change of +462.
-Unit growth for small function inlining: 945->1407 (48%)
+
+Considering rfft/1 with 535 size
+ to be inlined into tfft/3 in tfft2.f90:36
+ Estimated badness is -0.000000, frequency 5.67.
+ Badness calculation for tfft/3 -> rfft/1
+ size growth 462, time 4083 big_speedup
+ -0.000000: guessed profile. frequency 5.670000, count 0 caller count 0
time w/o inlining 52479.620026, time w/ inlining 47841.559998 overall growth
462 (current) 924 (original) 66204 (compensated)
+ Adjusted by hints -0.000000
+ Scaling time by probability:0.000000
+ Accounting size:2.50, time:0.00 on predicate:(true)
+ Scaling time by probability:0.000000
+ Accounting size:2.00, time:0.00 on predicate:(true)
+ Accounting size:13.00, time:261.35 on predicate:(true)
+ Scaling time by probability:0.000000
+ Accounting size:4.00, time:0.00 on predicate:(true)
+ Scaling time by probability:0.000000
+ Accounting size:2.50, time:0.00 on predicate:(true)
+ Accounting size:6.00, time:35.31 on predicate:(true)
+ Scaling time by probability:0.000000
+ Accounting size:2.00, time:0.00 on predicate:(true)
+ Accounting size:28.00, time:524.72 on predicate:(true)
+ Scaling time by probability:0.000000
+ Accounting size:4.00, time:0.00 on predicate:(true)
+ Scaling time by probability:0.000000
+ Accounting size:12.00, time:0.00 on predicate:(true)
+ Accounting size:14.00, time:4.84 on predicate:(true)
+ Scaling time by probability:0.000000
+ Accounting size:2.00, time:0.00 on predicate:(true)
+ Accounting size:41.00, time:54.74 on predicate:(true)
+ Scaling time by probability:0.000000
+ Accounting size:2.00, time:0.00 on predicate:(true)
+ Accounting size:20.50, time:65.23 on predicate:(true)
+ Accounting size:65.00, time:2192.07 on predicate:(true)
+ Scaling time by probability:0.000000
+ Accounting size:1.50, time:0.00 on predicate:(true)
+ Accounting size:79.00, time:7097.12 on predicate:(true)
+ Accounting size:79.00, time:12617.33 on predicate:(true)
+ Accounting size:46.00, time:131.93 on predicate:(true)
+ Scaling time by probability:0.000000
+ Accounting size:2.50, time:0.00 on predicate:(true)
+ Accounting size:4.00, time:2.95 on predicate:(true)
+ Accounting size:4.00, time:6.58 on predicate:(true)
+ Accounting size:6.00, time:0.00 on predicate:(true)
+Processing frequency rfft
+ Called by tfft that is executed once
+ Called in loop
+ Estimating body: tfft/3
+ Known to be false: not inlined
+ size:1051 time:47842
+ Inlined into tfft which now has time 47844 and size 1054, net change of +461.
+Unit growth for small function inlining: 945->1868 (97%)
so we inline the main computation routine into the "main" try-this-N-times
loop. Not sure why we get any estimated speedup here... (maybe ipa-cp
propagated that constant call arg from both call sites, but inlining above
doesn't suggest that).
