------- Comment #19 from dominiq at lps dot ens dot fr 2008-04-03 15:04 -------
(In reply to comment #17)
> I tried the patch referenced in comment #16 on powerpc64-linux. It
> allows tests pr11832.c and pr33009 to pass, but adding
> -frtl-abstract-sequences to all tests causes several tests to timeout on
> compilation, and gcc.dg/pr32912-1.c fails at execution for -m32. I built
> the C tests from SPEC CPU2000 with "-O2 -frtl-abstract-sequences" using
> the patched compiler and ran them with the short test input. A couple
> took a very long time to compile (try, for example, the file toke.c in
> test perlbmk) and four failed at execution time: vpr and crafty for -m32,
> gcc and gap for -m64.
I am not sure to understand: does it mean that the failures were
introduced by the patch? i.e., did you do the same tests for without the
patch? I don't know how '-frtl-abstract-sequences' is working, but I won't
be surprised if its cost is quadratic in the size of the problem.
The gcc manual says:
...
-O3
Optimize yet more. -O3 turns on all optimizations specified by -O2 and
also turns on the -finline-functions, -funswitch-loops,
-fpredictive-commoning, -fgcse-after-reload and -ftree-vectorize options.
...
-frtl-abstract-sequences
It is a size optimization method. This option is to find identical
sequences of code, which can be turned into pseudo-procedures and then
replace all occurrences with calls to the newly created subroutine. It is
kind of an opposite of -finline-functions. This optimization runs at RTL
level.
Although I have no doubt that examples can be crafted that show a speed
increase with '-O3 -frtl-abstract-sequences', the net result of this
combination will probably a (much) longer compilation time and a poor
overall result. It seems to me that '-frtl-abstract-sequences' is more
intended to complement the '-Os' option, so I did the following comparison
on the polyhedron benchmark (I don't have access to SPEC). Three tests
failed to compile, two shown a less than 10% decrease in the executable
size at the expense of more than a factor 3 in the compile time.
Core2Duo 2.16Ghz, i686-apple-darwin9, gcc version 4.4.0 20080403 (experimental)
(GCC)
+ patches (including the patch referenced in comment #16):
================================================================================
Polyhedron Benchmark Validator
Copyright (C) Polyhedron Software Ltd - 2004 - All rights reserved
================================================================================
-Os -Os -frtl-... -m64 -O3
...
Benchmark Compile Executable Compile Executable Compile
Executable
Name (secs) (bytes) (secs) (bytes) (secs)
(bytes)
--------- ------- ----------- ------- --------- -------
----------
ac 0.92 25748 1.82 25748 4.32
46616
aermod* 80.06 980996 298.68 931844 106.97
1225304
air 4.24 56048 8.41 56048 6.97
73200
capacita 1.86 35232 1.91 35232 3.78
64520
channel 0.85 25888 1.91 0 2.61
42752
doduc 8.47 117396 9.11 117396 15.18
183600
fatigue* 2.94 55556 18.51 51460 6.18
76696
gas_dyn 2.20 646272 2.91 642176 6.39
700392
induct 8.50 131152 29.73 0 13.41
160672
linpk 0.78 17508 0.66 17508 1.58
38400
mdbx 2.52 47760 2.64 47760 3.92
68856
nf 0.90 21804 0.93 21804 24.34
153240
protein 3.46 56228 6.49 0 11.10
118240
rnflow 4.07 55936 4.79 55936 11.31
167240
test_fpu 3.29 46996 3.89 46996 10.35
154176
tfft 0.62 17704 0.62 17704 1.42
26392
Benchmark Ave Run Number Estim Ave Run Number Estim Ave Run
Number Estim
Name (secs) Repeats Err % (secs) Repeats Err % (secs)
Repeats Err %
--------- ------- ------- ------- ------- ------- ----- -------
------- ------
ac 23.40 2 0.1197 23.45 2 0.0149 12.59
2 0.0318
aermod 44.30 2 0.0508 44.76 2 0.1251 29.77
2 0.1411
air 14.43 2 0.0139 15.06 5 1.2305 8.57
3 0.1683
capacita 82.95 2 0.0645 82.25 2 0.1088 55.79
2 0.1434
channel 7.34 5 0.1510 -1.00 2 0.1088 2.41
5 1.6186
doduc 61.77 2 0.0559 61.66 2 0.0592 42.85
2 0.0198
fatigue 18.19 2 0.0825 18.09 4 0.1745 10.65
2 0.1502
gas_dyn 29.75 3 0.1419 29.90 3 0.1623 10.21
2 0.1322
induct 94.22 2 0.0085 -1.00 3 0.1623 61.04
2 0.0549
linpk 28.37 2 0.0599 31.07 3 0.1955 28.23
2 0.0797
mdbx 17.47 2 0.0544 17.54 2 0.0171 15.14
2 0.0429
nf 35.04 2 0.0357 35.03 5 0.0733 32.39
2 0.0232
protein 57.37 2 0.0148 -1.00 5 0.0733 45.70
2 0.0088
rnflow 54.89 2 0.0173 54.87 2 0.0802 37.10
2 0.0526
test_fpu 22.33 2 0.0224 22.40 2 0.1407 12.73
2 0.1688
tfft 3.30 2 0.0909 3.31 2 0.0905 2.93
5 0.0907
Geom Mean Time = 27.63s
17.86s
where '-frtl-abstract-...' stands for '-frtl-abstract-sequences' and '-m64 -O3
...'
for '-m64 -O3 -ffast-math -funroll-loops -ftree-loop-linear
-fomit-frame-pointer
-finline-limit=600 --param min-vect-loop-bound=2'
[ibook-dhum] lin/test% gfc -Os -frtl-abstract-sequences channel.f90
/var/folders/iU/iUj3xngxGYe3MPCc0TZUcE+++TI/-Tmp-//ccM5Aq1n.s:531:non-relocatable
subtraction expression, "L43" minus "L00000000002$pb"
/var/folders/iU/iUj3xngxGYe3MPCc0TZUcE+++TI/-Tmp-//ccM5Aq1n.s:531:symbol:
"L00000000002$pb" can't be undefined in a subtraction expression
/var/folders/iU/iUj3xngxGYe3MPCc0TZUcE+++TI/-Tmp-//ccM5Aq1n.s:509:non-relocatable
subtraction expression, "L42" minus "L00000000002$pb"
/var/folders/iU/iUj3xngxGYe3MPCc0TZUcE+++TI/-Tmp-//ccM5Aq1n.s:509:symbol:
"L00000000002$pb" can't be undefined in a subtraction expression
/var/folders/iU/iUj3xngxGYe3MPCc0TZUcE+++TI/-Tmp-//ccM5Aq1n.s:unknown:Undefined
local symbol L00000000002$pb
[ibook-dhum] lin/test% gfc -Os -frtl-abstract-sequences induct.f90
induct.f90: In function 'gen_resq_mesh':
induct.f90:3704: internal compiler error: in compensate_edge, at
reg-stack.c:2759
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.
[ibook-dhum] lin/test% gfc -Os -frtl-abstract-sequences protein.f90
/var/folders/iU/iUj3xngxGYe3MPCc0TZUcE+++TI/-Tmp-//ccs13LFN.s:7048:non-relocatable
subtraction expression, "L729" minus "L00000000010$pb"
/var/folders/iU/iUj3xngxGYe3MPCc0TZUcE+++TI/-Tmp-//ccs13LFN.s:7048:symbol:
"L00000000010$pb" can't be undefined in a subtraction expression
/var/folders/iU/iUj3xngxGYe3MPCc0TZUcE+++TI/-Tmp-//ccs13LFN.s:6755:non-relocatable
subtraction expression, "L730" minus "L00000000010$pb"
/var/folders/iU/iUj3xngxGYe3MPCc0TZUcE+++TI/-Tmp-//ccs13LFN.s:6755:symbol:
"L00000000010$pb" can't be undefined in a subtraction expression
/var/folders/iU/iUj3xngxGYe3MPCc0TZUcE+++TI/-Tmp-//ccs13LFN.s:unknown:Undefined
local symbol L00000000010$pb
On a G5 1.8Ghz, powerpc-apple-darwin9, gcc version 4.4.0 20080403
(experimental) (GCC)
+ patches (including the patch referenced in comment #16), channel.f90 and
protein.f90 compile, and induct gives an ICE:
[karma] lin/test% gfc -Os -frtl-abstract-sequences induct.f90
induct.f90: In function 'convert_lower_case':
induct.f90:719: error: unrecognizable insn:
(jump_insn 45 19 40 3 (return) -1 (nil))
induct.f90:719: internal compiler error: in extract_insn, at recog.c:1983
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.
Note that all the tests fail on both platforms with gfortran 4.3.0 and
'-Os -frtl-abstract-sequences':
[ibook-dhum] lin/test% gfortran -w -Os -frtl-abstract-sequences ac.f90
ac.f90: In function 'suscep':
ac.f90:761: error: unrecognizable insn:
(insn 114 0 0 (set (reg:SI 0 ax)
(symbol_ref:SI ("*L8") [flags 0x2])) -1 (nil))
ac.f90:761: internal compiler error: in insn_default_length, at
insn-attrtab.c:1339
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.
...
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33642
