Do you have numbers how much time was spend there? :)

Do you have numbers how much time was spend there? :)

It's important to note that the time vcpkg does something is very small. Calculating abis takes around 75 ms or 10% of the total time vcpkg runs (excluding sub processes) when installing one port. I'm not sure if it's worth parallelizing all of it on a per-port level or parallelizing certain tasks like parsing cmake and finding script helpers. The goal is to find out what's most efficient and the code is split up to allow easy combination of the required steps.

Isn't the main bottleneck just I/O? Also the parallelization is not allow to break existing abi hashes.

Isn't the main bottleneck just I/O?

Yes! That's why I pointed at script hashes which is something I'll definitely consider parallelizing. I'd also take a look at saving the abi file (and maybe also that weird sbom file) on disk.

My profiling also suggests that parsing the cmake contents of a port for the use of helper functions takes some time - we loop through the cmake contents >100 times!

Also the parallelization is not allow to break existing abi hashes.

I'm not sure what you mean by that. ABI can't be broken by calculating hashes of one port in different order because the vector is sorted in the end. I'm also aware that we can't parallelize everything without iterating multiple times over install plan actions.

The top 3 hottest paths are get_compiler_info, get_cmake_script_hashes, and get_toolset. get_cmake_script_hashes is executed exactly once and the other two functions only a few times at most. These are 2 different operations (get_compiler_info needs get_toolset). Therefore, we could do them asynchronously.

If we could get the compiler info for host and target triplet in parallel we could be even faster, but I don't think this is possible?

#802 adds a new PortAbiCache to speed up port abi creation. But this is only helpful if you have multiple ABIs for the same port, which you normally only have if you have host-triplet != target-triplet. But #802 builds a port multiple times with different abis and needed this (resulted in a multiple minutes speedup when using --all).
https://github.com/microsoft/vcpkg-tool/pull/802/files#diff-a93efc9f4d36565c420ffe731eb4aafb23b3f5f300db52302f898c4a64884a94

#802 adds a new PortAbiCache to speed up port abi creation. But this is only helpful if you have multiple ABIs for the same port, which you normally only have if you have host-triplet != target-triplet. But #802 builds a port multiple times with different abis and needed this (resulted in a multiple minutes speedup when using --all). https://github.com/microsoft/vcpkg-tool/pull/802/files#diff-a93efc9f4d36565c420ffe731eb4aafb23b3f5f300db52302f898c4a64884a94

I don't think it's worth optimizing for host-triplet != target-triplet because it applies only to few ports. However, I think I might end up parallelizing make_abi_tag().