Segmentation fault with CXXFLAGS="-O3" or CXXFLAGS="-O1 -ftree-loop-vectorize"

[Arfrever]

Segmentation fault in marisa-test occurs if Marisa is configured with CXXFLAGS="-O3".
Problem does not occur for me with additional -march=westmere or -march=native (e.g. CXXFLAGS="-O3 -march=westmere").
I use GCC 10.2.0.

$ CXXFLAGS="-g -O3" ./configure
...
$ make
...
$ make check
...
make  check-TESTS
make[2]: Entering directory '/tmp/marisa-trie/tests'
make[3]: Entering directory '/tmp/marisa-trie/tests'
PASS: base-test
PASS: io-test
PASS: vector-test
PASS: trie-test
../test-driver: line 109: 14027 Segmentation fault      "$@" > $log_file 2>&1
FAIL: marisa-test
============================================================================
Testsuite summary for marisa 0.2.6
============================================================================
# TOTAL: 5
# PASS:  4
# SKIP:  0
# XFAIL: 0
# FAIL:  1
# XPASS: 0
# ERROR: 0
============================================================================
See tests/test-suite.log
Please report to susumu.yata@gmail.com
============================================================================
make[3]: *** [Makefile:715: test-suite.log] Error 1
make[3]: Leaving directory '/tmp/marisa-trie/tests'
make[2]: *** [Makefile:823: check-TESTS] Error 2
make[2]: Leaving directory '/tmp/marisa-trie/tests'
make[1]: *** [Makefile:925: check-am] Error 2
make[1]: Leaving directory '/tmp/marisa-trie/tests'
make: *** [Makefile:456: check-recursive] Error 1
$ LD_LIBRARY_PATH="$(pwd)/lib/marisa/.libs" gdb tests/.libs/marisa-test
...
(gdb) r
Starting program: /tmp/marisa-trie/tests/.libs/marisa-test 
marisa-test.cc:13: TestEmptyTrie(): ok
marisa-test.cc:87: TestTinyTrie(): ok

Program received signal SIGSEGV, Segmentation fault.
0x00007ffff7fab3c5 in marisa::grimoire::algorithm::details::get_label<marisa::grimoire::trie::Entry> (depth=1, unit=...)
    at ../../../../lib/marisa/grimoire/algorithm/sort.h:19
19        return (depth < unit.length()) ? (int)(UInt8)unit[depth] : -1;
(gdb) bt
#0  0x00007ffff7fab3c5 in marisa::grimoire::algorithm::details::get_label<marisa::grimoire::trie::Entry> (depth=1, unit=...)
    at ../../../../lib/marisa/grimoire/algorithm/sort.h:19
#1  marisa::grimoire::algorithm::details::sort<marisa::grimoire::trie::Entry*> (l=l@entry=0x55555557c110, r=r@entry=0x55555557c660, depth=depth@entry=1)
    at ../../../../lib/marisa/grimoire/algorithm/sort.h:103
#2  0x00007ffff7fab4ea in marisa::grimoire::algorithm::details::sort<marisa::grimoire::trie::Entry*> (l=0x55555557aed0, r=0x55555557dce0, 
    depth=depth@entry=0) at ../../../../lib/marisa/grimoire/algorithm/sort.h:132
#3  0x00007ffff7fac2c3 in marisa::grimoire::algorithm::sort<marisa::grimoire::trie::Entry*> (end=<optimized out>, begin=<optimized out>)
    at ../../../../lib/marisa/grimoire/algorithm/sort.h:189
#4  marisa::grimoire::Algorithm::sort<marisa::grimoire::trie::Entry*> (end=<optimized out>, begin=<optimized out>, this=<optimized out>)
    at ../../../../lib/marisa/grimoire/algorithm.h:15
#5  marisa::grimoire::trie::Tail::build_ (this=0x7fffffffd130, entries=..., offsets=0x7fffffffd310, mode=MARISA_TEXT_TAIL) at tail.cc:161
#6  0x00007ffff7facde7 in marisa::grimoire::trie::Tail::build (this=this@entry=0x7fffffffd918, entries=..., offsets=offsets@entry=0x7fffffffd310, 
    mode=MARISA_TEXT_TAIL) at tail.cc:41
#7  0x00007ffff7fb569b in marisa::grimoire::trie::LoudsTrie::build_next_trie<marisa::grimoire::trie::Key> (this=0x7fffffffd630, keys=..., 
    terminals=0x7fffffffd310, config=..., trie_id=1) at ../../../../lib/marisa/grimoire/trie/config.h:35
#8  0x00007ffff7fbea57 in marisa::grimoire::trie::LoudsTrie::build_trie<marisa::grimoire::trie::Key> (this=this@entry=0x7fffffffd630, keys=..., 
    terminals=terminals@entry=0x7fffffffd4c0, config=..., trie_id=trie_id@entry=1) at louds-trie.cc:300
#9  0x00007ffff7fb5940 in marisa::grimoire::trie::LoudsTrie::build_ (this=0x7fffffffd630, keyset=..., config=...) at louds-trie.cc:257
#10 0x00007ffff7fb69e3 in marisa::grimoire::trie::LoudsTrie::build (this=this@entry=0x555555576bd0, keyset=..., flags=flags@entry=135169)
    at louds-trie.cc:27
#11 0x00007ffff7fa975a in marisa::Trie::build (this=this@entry=0x7fffffffdb38, keyset=..., config_flags=config_flags@entry=135169) at trie.cc:16
#12 0x000055555555ada4 in (anonymous namespace)::TestTrie (num_tries=1, tail_mode=MARISA_TEXT_TAIL, node_order=MARISA_WEIGHT_ORDER, keyset=...)
    at marisa-test.cc:268
#13 0x000055555555bd0a in (anonymous namespace)::TestTrie (tail_mode=MARISA_TEXT_TAIL, node_order=MARISA_WEIGHT_ORDER, keyset=...) at marisa-test.cc:357
#14 0x000055555555be2e in (anonymous namespace)::TestTrie (tail_mode=MARISA_TEXT_TAIL) at marisa-test.cc:367
#15 0x0000555555559b7e in (anonymous namespace)::TestTrie () at marisa-test.cc:372
#16 main () at marisa-test.cc:383
(gdb) 

[Arfrever]

More precisely, segmentation fault occurs with -ftree-loop-vectorize and at least -O1 (so either CXXFLAGS="-O1 -ftree-loop-vectorize" or CXXFLAGS="-O2 -ftree-loop-vectorize").
Segmentation fault occurs with GCC 10.2.0, and does not occur with GCC 8.4.0 or 9.3.0.

[Arfrever]

Reported as GCC bug 99375.

[Arfrever]

Problem was fixed in GCC 10 branch:
GCC commit 1ab88985631dd2c5a5e3b5c0dce47cf8b6ed2f82
Fix will be released in GCC 10.3.0.

(I will close this report in Marisa after release of GCC 10.3.0.)

[Arfrever]

GCC 10.3.0 was released on 2021-04-08.