CMake: builds always use -march=native

[bmwiedemann]

Current Behavior

While working on reproducible builds for openSUSE, I found that
when building packages, machine code depends on which build worker is selected from our build pool.

Expected Behavior

It should be possible to build without -march=native to get reproducible binaries.

Suggested Fix

There is a fix in #3864

tesseract -v

5.3.0

Operating System

No response

Other Operating System

openSUSE-Tumbleweed 20230315

uname -a

6.2.4

Compiler

gcc-12

CPU

various

Virtualization / Containers

No response

Other Information

No response

[stweil]

Linux distributions should use the automake build which fixed that in commit 5884036. Does openSUSE use the cmake build? Why?

[bmwiedemann]

https://code.opensuse.org/package/tesseract-ocr/blob/master/f/tesseract-ocr.spec#_79 says we use cmake since we moved from automake 2 months ago.

[amitdo]

Does openSUSE use the cmake build? Why?

People see that we support both Autotools and CMake, and they choose the tool they prefer.

[stweil]

That's okay for individual users and for those who have special needs like using IDEs for Windows or macOS. But for Linux distributions I suggest to move back to automake. That minimized their risk and makes it easier for us.

[bmwiedemann]

Even outside of Linux, #3864 would be useful, because not everyone runs their binaries on the same type of machine as their build machine.

[mimi1vx]

@stweil .. usually is cmake faster, you can use different builder than make like ninja, and it can be pretty easy debug and modified in comparsion with autotools ?

[zdenop]

Please check current code. Now the default setting should be the same as at autotools.

[bmwiedemann]

commit f779c43 looks good, except that it does not mention this issue.
Test passed with it, too.

[stweil]

@stweil .. usually is cmake faster

Autotools by default builds shared and static libraries and builds executables which use the shared tesseract library.

CMake by default seems to build only the static library, so the executables don't use a shared tesseract library and are much larger. @bmwiedemann, Linux distributions typically prefer shared linking. Do you use special flags when running CMake?

A single threaded build with CMake takes around 1:46 on my MacBook and uses 99 % cpu.
A single threaded build with Autotools and --disable-shared takes around 2:26 and uses 91 % cpu. So yes, it is a bit slower because it obviously does not use the full cpu power.

[amitdo]

Autotools by default builds shared and static libraries

I don't think it's a good default. what do you think?

[stweil]

As far as I know that's the normal default for autotools. And typically a distribution requires both, shared libraries for the binaries and additional static libraries for development packages. So for distributions that default is fine.

[zdenop]

@stweil : with -DBUILD_SHARED_LIBS=ON you can build shared libs with cmake