24e16d7 broke g++ build on FreeBSD-11.0 #15984
Comments
From @jkeenanSummary: A commit to Perl 5 blead in December 2016 made it impossible to build Preliminary remarks: 1. The default C-compiler on FreeBSD-11.0 is clang, not g++. When I 2. Hence, what I report below is an edge case that we should explore Bug report: At the point when perl-5.24.0 was released, perl could have been built Bisection indicates that the following is the first bad commit: ##### regexec.c: Make isFOO_lc() non-static The immediately preceding commit would be: ##### utf8.c: White space, comments only If I were to configure at ddb6593 as follows: ##### ... and then call 'make -j${TEST_JOBS} test_harness'. Perl would build If I now were to advance to commit ##### make: stopped in /usr/home/jkeenan/gitwork/perl2 make: stopped in /usr/home/jkeenan/gitwork/perl2 So this is a regression between 5.24 and 5.26. However, since it is a Thank you very much. |
From @jkeenanSummary of my perl5 (revision 5 version 25 subversion 9) configuration: Characteristics of this binary (from libperl): |
From @jkeenanext/re/re_exec.c:9753:Perl__is_grapheme(pTHX_ const U8 * strbeg, const U8 * s, const U8 * strend, const UV cp) |
From @craigberryOn Sat, May 20, 2017 at 10:54 PM, James E Keenan
Surely this was fixed by: commit 9637d2a Don't define Perl_isFOO_lc in the regex extension. Otherwise it ends up multiply defined, which annoys the VMS linker Can you you either do another bisect that starts from there or at |
|
The RT System itself - Status changed from 'new' to 'open' |
From @jkeenanOn Sun, 21 May 2017 13:11:38 GMT, craig.a.berry@gmail.com wrote:
I am attaching the log of 'make' (gzipped) that I got when compiling blead with g++ yesterday. Blead was probably at 2c4393c at that point. Thank you very much. |
From @jkeenan |
From @jkeenanOn Sun, 21 May 2017 13:11:38 GMT, craig.a.berry@gmail.com wrote:
Using this function to configure: ##### ... I tried to build perl at commit 9637d2a and got the same, "multiple definition of `Perl__is_grapheme'" build failure as previous posted. See attachment for gzipped output of 'make' at this revision. Thank you very much. |
From @craigberryOn Sun, May 21, 2017 at 11:24 AM, James E Keenan via RT
How do you mean "the same"? Previously it was Perl_isFOO_lc that was Inline Patchdiff --git a/regexec.c b/regexec.c
index 0b5c847f1c..ea52477781 100644
--- a/regexec.c
+++ b/regexec.c
@@ -9708,6 +9708,8 @@ S_to_byte_substr(pTHX_ regexp *prog)
return TRUE;
}
+#ifndef PERL_IN_XSUB_RE
+
bool
Perl__is_grapheme(pTHX_ const U8 * strbeg, const U8 * s, const U8 *
- /* |
From @jkeenanOn Mon, 22 May 2017 01:32:07 GMT, craig.a.berry@gmail.com wrote:
Your patch -- which I've pushed in the smoke-me/jkeenan/131337-craig branch -- does enable 'make' to complete when g++ is the C-compiler on FreeBSD-11.0. However, there is a persistent test failure in t/re/regexp_nonull.t (http://perl5.test-smoke.org/report/55820) -- a failure which that branch does not get when compiling with g++ on Linux (http://perl5.test-smoke.org/report/55823). I am attaching: * the gzipped log of 'make' for that build; * the ./perl -Ilib -V output for that build; * the output for "cd t; ./perl harness -v re/regexp_nonull.t; cd -" for that build * and by way of contrast, the output for "cd t; ./perl harness -v re/regexp_nonull.t; cd -" on FreeBSD-11.0 but using the regular configuration style with 'clang'. Thank you very much. -- |
From @jkeenanCan't locate XS/APItest.pm in @INC (you may need to install the XS::APItest module) (@INC contains: ../lib) at re/regexp_nonull.t line 10. Test Summary Report re/regexp_nonull.t (Wstat: 512 Tests: 0 Failed: 0) |
From @jkeenanSummary of my perl5 (revision 5 version 26 subversion 0) configuration: Characteristics of this binary (from libperl): |
From @jkeenanre/regexp_nonull.t .. |
From @jkeenanOn Mon, 22 May 2017 13:08:13 GMT, jkeenan wrote:
It appears that when, on FreeBSD-11.0, I configure with g++ in the smoke-me/jkeenan/131337-craig branch, the perl that gets built does not have XS::APItest. If, in a "regular_configure" of perl on this platform -- i.e., building with the default C-compiler clang -- at the end of 'make test_prep' I issue this command: ##### ... I get these results: ##### However, if I build with g++ instead, I only get these results: ##### I don't know why this is happening. Perhaps it is a different bug. Thank you very much. -- |
From @jkeenanOn Mon, 22 May 2017 14:11:58 GMT, jkeenan wrote:
If there is an additional bug, it probably occurs much earlier -- in the configure stage. I am attaching the config.sh for blead built with my regular configuration command ... ##### ... and for the smoke-me/jkeenan/131337-craig branch built as follows: ##### If you grep for '_ext=', the differences are startling. regular_configure has a long list of libraries, including XS/APItest, as the value for 'dynamic_ext'. gpp_configure has an empty string for 'dynamic_ext'. In regular_configure, 'static_ext' has a value of a single whitespace. In gpp_configure, 'static_ext' has a long list of libraries which does not include XS/APItest. In regular_configure, 'XS/APItest' can be found in the values for 'dynamic_ext', 'extensions' and 'known_extensions'. In gpp_configure, 'XS/APItest' is found only in 'known_extensions'. Can anyone offer some insight into this? Thank you very much. -- |
From @LeontOn Mon, May 22, 2017 at 5:09 PM, James E Keenan via RT <
Yeah, the long list of .a files in the original post looked strange to me. Leon |
From @jkeenanOn Mon, 22 May 2017 17:03:43 GMT, LeonT wrote:
$ grep -n usedl *.config.sh So, 'undef' in the g++ case. |
From @jkeenanOn Mon, 22 May 2017 17:26:44 GMT, jkeenan wrote:
This is starting to look like a "Linux vs FreeBSD" case. If in my FreeBSD-10.3 VM (which is having its own problems with a g++ build; see RT #131336) I do a regular_configure in blead and then a gpp_configure in the smoke-me/jkeenan/131337-craig branch, I observe the same differences as I do in my FreeBSD-11.0 VM. In 10.3 in the branch, 'usedl' is 'undef', 'dynamic_ext' is '', 'static_ext' has many libraries but not XS/APItest. However, when on Linux I configure like this: ##### ... I get a "more plausible" config.sh: ##### Thank you very much. -- |
From @jkeenanOn Mon, 22 May 2017 17:49:51 GMT, jkeenan wrote:
After much experimentation, this is what I have got somewhat working. On FreeBSD-11.0, in the smoke-me/jkeenan/131337-craig branch, I configure as follows: ##### That populates config.sh as in attachment 131337.better.config.sh.txt. 'make' and 'make test_harness' complete successfully. But this begs the question: Why on FreeBSD --but not on Linux -- does a request to build with g++ require command-line switches 'usedl' and 'dlsrc' to be populated? I have stared at the relevant code in ./Configure but cannot figure out the answer. Thank you very much. |
From @jkeenan17:config_args='-des -Dusedevel -Duseithreads -Doptimize=-O2 -pipe -fstack-protector -fno-strict-aliasing -Dcc=g++ -Dusedl -Ddlsrc=dl_dlopen.xs' |
From @LeontOn Mon, May 22, 2017 at 9:59 PM, James E Keenan via RT <
The hints for linux mention that "If using g++, the Configure scan for Leon |
From @jkeenanOn Mon, 22 May 2017 21:07:39 GMT, LeonT wrote:
That looks promising. See https://perl5.git.perl.org/perl.git/commitdiff/48e836997 -- |
From @jkeenanOn Tue, 23 May 2017 02:15:19 GMT, jkeenan wrote:
Applied to blead in these commits: ##### When building with g++ on FreeBSD, explicitly set 'usedl' and 'dlsrc'. commit 66c5e3f Patch suggested by Craig Berry for RT 131337. See http://perl5.test-smoke.org/report/56009 Marking ticket Resolved. -- |
|
@jkeenan - Status changed from 'open' to 'resolved' |
From @doughera88On Mon, Jun 05, 2017 at 06:05:07AM -0700, James E Keenan via RT wrote:
I pushed a slightly simpler version that forces Configure to find commit 2c8efe4 Simpler hints fix for [perl #131337]. Having found dlopen(), the values of usedl=define and dlsrc=dl_dlopen.xs (I've also attached the patch here to record it in RT -- I meant to do The ticket should still stay Resolved. -- |
From @doughera88freebsd-hints.patchcommit 2c8efe4079b75c61cf34425054539a9c24913e9f
Author: Andy Dougherty <doughera@lafayette.edu>
Date: Mon Jun 12 08:02:10 2017 -0400
Simpler hints fix for [perl #131337].
The Configure scan fails to find dlopen() with g++. Explicitly making
it availble allows Configure to default to using dynamic loading, but
still allows the user to override and use static loading.
diff --git a/hints/freebsd.sh b/hints/freebsd.sh
index fc6d73f..b3422c9 100644
--- a/hints/freebsd.sh
+++ b/hints/freebsd.sh
@@ -322,10 +322,12 @@ d_printf_format_null='undef'
d_uselocale='undef'
# https://rt.perl.org/Ticket/Display.html?id=131337
+# Reported in 11.0-CURRENT with g++-4.8.5:
+# If using g++, the Configure scan for dlopen() fails.
+# Easier for now to just to forcibly set it.
case "$cc" in
*g++*)
- usedl='define'
- dlsrc='dl_dlopen.xs'
+ d_dlopen='define'
;;
esac
|
Migrated from rt.perl.org#131337 (status was 'resolved')
Searchable as RT131337$
The text was updated successfully, but these errors were encountered: