This is part of my GSoC'22 project "Automatic system images based on a specific user workflow," under the supervision of @ianatol. The project's main idea is to introduce a new library that would obtain precompile statements during regular usage of Julia. Then the system image is customized for a specific user. So, it's crucial to speed up building the system image (sysimage) because it's pretty slow now (empty >2 min). So, this PR rebases #40414 (initially proposed by @Keno) for Julia v1.9-dev. But there are some modifications:

• It stores some extra information in precompile field of CodeInstance.
  • precompile & 2 - CodeInstance is loaded from a current sysimage.
  • precompile & 4 - CodeInstance should be emitted into the incremental sysimage.
  • These bits are only temporary and not serialized into sysimage when produced.
• It speeds up building system image by emitting code only for the tagged instances (by precompile & 4).
  • jl_precompiles_for_sysimage function is introduced to tag only user-specific precompile statements. Precompiles generated before this function is called will be ignored.
• New detection if CodeInstance is already in the sysimage. Originally, it was done only by the naming convention. Now, the instance is not included if contained in ReverseLocalSymbolTable checked by lookup_sysimage_fname, which seems a bit more robust and systematic way to do so.

Given these changes, it's unnecessary to emit binary code for all precompile() statements defined in loaded libraries, resulting in relatively fast compilation times.

In the future, I'll will investigate why there is a greater overhead for Plots (not yet clear to me).

All times are in seconds, N - normal run, S - using chained sysimage, SysImg - create chained sysimage, load - call using, work - time to first work (TTFX). Sources for the comparison are in BenchmarkIncrementalSysimage.

There are probably some edge cases, but it seems to work fine for now. In the future, I'd like to try to disable precompile() entirely for the included libraries; that would probably save some loading time because the sysimage would contain less inferred CodeInstances.

This PR also introduces a CI test for building an almost empty sysimage. It seems possible to make the test multi-platform by using LLVM_full_jll library that contains all the necessary utils.

Solved issues

• It may produce multiple definitions (and thus fails if ld --allow-multiple-definition is not used). Solved by initializing globalUniqueGeneratedNames, see 431b6ba.
• Archive sys-o.a (207MB) is not currently shipped with julia, but it can be re-created easily together with sys.so. This is necessary only once.
• 32-bit version seems fixed.
• Most of the CI passes (only for Linux)

Known issues (to be solved) / TODOs

• It interfere with Base.compilecache. Precompilation do not work when building incremental sysimage, so the pakages need to be already precompiled.
• Chained sysimages for Windows are not yet supported.

Open questions and ideas for and improvement

• Would it be possible to copy compiled code directly from the loaded sysimage when outputting the archive?
• Is it worthy to support chains of multiple sysimages (to enable incrementally adding precompile statements)?