Hi Daniel, Thank you for the extensive information. I have looked at the ordering of pass short_help strings and I find them to be ordered by iterating over std::map<std::string, Pass*> pass_register; which clearly is defined regardless of locale since std::map uses std::less for comparison which uses operator< which for std::string is given as std::lexicographical_compare which is described <https://en.cppreference.com/w/cpp/algorithm/lexicographical_compare> as "If one range is a prefix of another, the shorter range is lexicographically *less* than the other."
Therefore synth_intel ought to always come before synth_intel_alm regardless of stability or locale. Would you mind sharing information about which locale was used and the commit this was tested at? Cheers Emil On Mon, Jul 8, 2024 at 6:16 PM Daniel Gröber <d...@darkboxed.org> wrote: > Hi Emil, > Hi all (CC'ing rb-general to fact check my recomendation and incentives > analysis below), > > On Mon, Jul 08, 2024 at 05:34:19PM +0200, Emil Jiří Tywoniak wrote: > > Thanks for the reply. I saw that PR > > <https://github.com/YosysHQ/yosys/pull/3951> - so the blocker is > related to > > how Debian uses faketime in chroot, and triggered by yosys using faketime > > on pdfs? Do you have some related links? > > That turned out to be a false-positive. Faketime was never really the > problem, my assertion that graphviz supports SOURCE_DATE_EPOCH was also > wrong see https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1054170. > > If you look at > > https://tests.reproducible-builds.org/debian/rb-pkg/unstable/amd64/yosys.html > > this still shows the old repro problems with CreationDate differing. > > <key>CreationDate</key> > - <value><string·size="22">D:20240629125837-12'00</string></value> > + <value><string·size="22">D:20260905043016+14'00</string></value> > > IIRC I was able to suppress that problem in the latest experimental > upload. It turned out to be a makefile variable override fuckup on my part. > > Unfortunately the experimental package currently fails to build so the > diffoscope output is not on the CI server anymore. Last I built it what I > was seeing was (suspected) sort() reordering of help output depending on > locale. This is incorporated into the docs build hence the repro problem. > > Looked something like this according to my notes: > > -synthesis for Intel (Altera) FPGAs. > -synth_intel > synthesis for ALM-based Intel (Altera) FPGAs. > synth_intel_alm > +synthesis for Intel (Altera) FPGAs. > +synth_intel > synthesis for Lattice FPGAs > synth_lattice > <bram_type> > > The problem comes from me (quite intentionally) only overriding LC_*/TZ for > the docs build, not the main yosys binary build. So my guess is the help > texts are hardcoded into the binary at build time somehow and this depends > on locale. This is where you it would help if you know or could ask your > collegues if that hypothesis makes any sense or what :) > > I did an audit of all the sort()ish calls in yosys during mDebConf Berlin > to find the culprit but alas I forgot to write down my suspect :] Note: > keep in mind this could also be an unstable ordering rather than a local > dependent sort() I have not yet proved the sort() hypothesis. > > An easy fix would be to override LC_* in Debian but we don't like to do > that because ideally your upstream version shouldn't have this > reproducibility hazard in the first place. This way other distros can > benefit too. > > BTW: The reason overriding in our packaging is OK for latex/graphviz IMO is > that the bugs are problems with those packages not yours so we'll either > fix them upstream or in the corresponding Debian packages. If > reproducibility is a concern you care about too you could opt to do this > override in your build system but I would perhaps recommend against doing > it in the released yosys Makefile as it would lessen the incentives for the > root cause upstreams to fix their problems. > > The help text reordering does represent a problem in yosys though so it > should be fixed there. > > Thanks, > --Daniel >
