g++ 6.3 build failure

[amitdo]

Environment

• Tesseract Version: 5.0 alpha
• Commit Number: e3fd938
• Platform: Linux debian 4.9.0-14-amd64 defect issue #1 SMP Debian 4.9.246-2 (2020-12-17) x86_64 GNU/Linux

gcc version 6.3.0 20170516 (Debian 6.3.0-18+deb9u1)

g++ 6.3 fails to compile commit e3fd938.

config.log

make.log

[stweil]

Building with g++-6 fails since commit 4fc467a according to my tests.

So Tesseract no longer compiles with Debian stretch (old stable).

[stweil]

The problems are indeed related to C++11. Those related to the STRING class can easily be fixed. As far as I see the only remaining code which needs a newer compiler is this:

src/ccutil/unicharset.cpp: In member function ‘void tesseract::UNICHARSET::unichar_insert(const char*, tesseract::OldUncleanUnichars)’:
../../../src/ccutil/unicharset.cpp:618:37: error: forming reference to void
     auto &u = unichars.emplace_back();

In C++-11 emplace_back is a void function, so that needs different code if we want to support C++-11.

[stweil]

The modifications which I mentioned above are not sufficient, because training has more dependencies on newer compilers. In addition, I get unresolved symbols when linking.

So I am afraid that supporting g++-6 would require more work. There are currently other, more important, problems.

[stweil]

Builds of git master currently work with -std=c++14, but 3 code locations like the following one then show warnings because they use C++17 code:

src/training/mftraining.cpp: In function ‘int main(int, char**)’:
src/training/mftraining.cpp:207:8: warning: structured bindings only available with ‘-std=c++17’ or ‘-std=gnu++17’
  207 |   auto [trainer,shape_table] = tesseract::LoadTrainingData(argc, argv,
      |        ^

[amitdo]

It's not important to me to have gcc-6 support, but I do want to know what are the minimum GCC, Clang and MSVC versions that can compile the current 5.0 code.

[amitdo]

Travis CI seems to use gcc 6 and was able to build (with cmake) commit e3fd938.

[Shreeshrii]

Travis CI seems to use gcc 6 and was able to build (with cmake) commit e3fd938.

Travis build for MacOS does not build the lstm related training tools. See log

$ sudo make install
[ 88%] Built target libtesseract
[ 89%] Built target tesseract
[ 93%] Built target common_training
[ 93%] Built target ambiguous_words
[ 94%] Built target wordlist2dawg
[ 95%] Built target cntraining
[ 96%] Built target dawg2wordlist
[ 97%] Built target shapeclustering
[ 98%] Built target combine_tessdata
[ 99%] Built target mftraining
[100%] Built target classifier_tester
Install the project...
-- Install configuration: "Release"

Linux builds the following. See log

$ sudo make install
[ 78%] Built target libtesseract
[ 78%] Built target tesseract
[ 82%] Built target common_training
[ 83%] Built target shapeclustering
[ 84%] Built target mftraining
[ 89%] Built target unicharset_training
[ 91%] Built target pango_training
[ 92%] Built target text2image
[ 93%] Built target wordlist2dawg
[ 93%] Built target ambiguous_words
[ 94%] Built target classifier_tester
[ 95%] Built target combine_tessdata
[ 95%] Built target combine_lang_model
[ 96%] Built target lstmeval
[ 96%] Built target lstmtraining
[ 96%] Built target dawg2wordlist
[ 97%] Built target cntraining
[ 98%] Built target merge_unicharsets
[ 99%] Built target set_unicharset_properties
[100%] Built target unicharset_extractor```

[amitdo]

Travis CI seems to use gcc 6 and was able to build (with cmake) commit e3fd938.

My mistake.

tesseract/.travis.yml

Line 49 in a377b4e

#- if [[ $LINUX && "$CXX" = "g++" ]]; then export CXX="g++-6" CC="gcc-6"; fi

I didn't notice this line was commented out.

My report didn't include training tools.

I don't think @stweil submmited his partial fix.