On Fri, Aug 5, 2016 at 10:02 PM, Jim Meyering <j...@meyering.net> wrote:
> On Fri, Aug 5, 2016 at 7:32 PM, Norihiro Tanaka <nori...@kcn.ne.jp> wrote:
>>
>> On Fri, 5 Aug 2016 13:29:43 -0700
>> Jim Meyering <j...@meyering.net> wrote:
>>
>>> On Fri, Aug 5, 2016 at 4:30 AM, Norihiro Tanaka <nori...@kcn.ne.jp> wrote:
>>> > dfaoptimize() is not set fast flag even if it is success, but it is wrong.
>>> > If success, dfa matcher uses algorithm for single byte, and it is so fast.
>>> >
>>> > I think this bug does not affect for grep, but it will affect with the
>>> > patch that I just sent to gawk.
>>>
>>> Thank you for the patch.
>>> I was going to push it with the attached slightly updated log message.
>>> Note however that grep does use that -> fast member via dfasearch.c's
>>> use of the dfaisfast function.
>>> But then I realized I should at least verify with "make check", and
>>> found that this makes grep's dfa-match test fail.
>>> Thus, I will not be pushing it as-is.
>>
>> Thanks for review and adjustment. I re-ran all tests including dfa-match,
>> and they were passwd again in my machine. Next, I will re-run them on
>> Fedora24, as my machine is RHEL 6.8 and GCC 4.4.7 which is too old.
>>
>> However, I do not know why dfa-match test fails on your machine.
>> dfa-match test does not use grep. It directly calls dfa functions through
>> dfa-match-aux executable in order to test codes of dfa which grep does
>> not use. dfa-match-aux does not referer to the ->fast member.
>
> I have examined the logs, which suggest it was a false positive in a
> parallelized "make check" run, due to that test's 3-second timeout. I
> have tried repeatedly to reproduce that failure, so far without
> success, but in coreutils development, with parallelized tests, we
> fixed many hard-to-reproduce tests with small timeout limits like this
> -- most of them now use 10 seconds as the limit, so I will change this
> one, too (and several others) with the attached patch.
>
> I have pushed your patch.
While trying to reproduce that, I ran some tests on OS X and saw
frequent failure of one of the tests, so wrote the attached to work
around what I assume is an aggressive write-to-/dev/null optimization:
From 428a433ca0a19628922554f10e55a5d91540c15f Mon Sep 17 00:00:00 2001
From: Jim Meyering <meyer...@fb.com>
Date: Fri, 5 Aug 2016 21:38:30 -0700
Subject: [PATCH] tests: avoid occasional false positive on OS X
* tests/backref-multibyte-slow: When redirecting the "fast"
LC_ALL=C run's output to /dev/null, we got an artificially low
timing (of 0), because of how some systems ignore writes to
/dev/null. With an initial timing of 0 on that first run, the
derived timeout for the UTF-8 run (which redirects to a file)
would be a mere 1 second. The fix: also redirect that first run's
output to a file, not to /dev/null.
---
tests/backref-multibyte-slow | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tests/backref-multibyte-slow b/tests/backref-multibyte-slow
index d447a4a..05bb62e 100755
--- a/tests/backref-multibyte-slow
+++ b/tests/backref-multibyte-slow
@@ -14,7 +14,7 @@ $AWK 'BEGIN {for (i=0; i<13000; i++) print "aba"}' /dev/null
> in || fail=1
# when running in en_US.UTF-8. Round up to whole seconds, since timeout
# can't deal with fractional seconds.
max_seconds=$(LC_ALL=C perl -le 'use Time::HiRes qw(time); my $s = time();
- system q,grep -E '\''^([a-z]).\1$'\'' in > /dev/null,;
+ system q,grep -E '\''^([a-z]).\1$'\'' in > junk,;
my $elapsed = time() - $s; print int (1 + 10 * $elapsed)') \
|| { max_seconds=5;
warn_ "$ME_: warning: no perl? using default of 5s timeout"; }
--
2.8.0-rc2
