compute deps fingerprint only

[soulomoon]

Just follow implementation from #4443 (comment) @guibou very good bisect and findings

[guibou]

Thank you for working on that topic.

I applied the patch and did not observe much improvement in my work codebase (e.g. any added import triggers a long, 30s-1minute rebuild).

I will have a look a bit later during the day with the "simple" examples that I had provided in #4443, maybe this actually fix / improves them and maybe there is another issue with my codebase.

[guibou]

@soulomoon Thank you for this code.

I've tested it with the replication depicted in #4443 (comment)

(after rebasing the replication code on the commit just before this MR).

What I can observe:

With commit before this MR:

• Opening Lib2.hs: diagnostics are available after 10s
• Adding/removing the module: diagnostics are there immediately
• Then, in the same editor session, opening MyLib.hs: rebuilds take 10s
• Adding/removing the module: diagnostics takes 10s to come back

With this commit from this MR:

• Opening Lib2.hs: diagnostics are available immediately! I'm actually kinda surprised by that. Is there some sort of cache that HLS may exploit when restarting? I would have assumed that in order to provide diagnostics for Lib2.hs, the initial startup would have required to typecheck MyLib.hs at least once. Anyway, that's great!
• Adding/removing the module: diagnostics are there immediately (as before)
• Then, in the same editor session, opening MyLib.hs: rebuild for MyLib.hs takes 10s. This is expected and interesting, it means that to typecheck / diag Lib2.hs (which depends on MyLib.hs) it does not need to fully typecheck MyLib.hs (somethig is in cache, or typecheck is partial). But in cannot use "the cache" when typechecking MyLib.hs. Anyway, I'm ok with this result.
• Adding/removing the module in Lib2.hs: diagnostics are back immediately.

This is an awesome success!

I unfortunately do not understand the implementation so I cannot comment more than that, but this is awesome and should be merged ASAP ;) @fendor maybe you can have a look.

EDIT: I've cleaned my .cache/ghcide AND when reopening Lib2.hs it initially took like 15s to typecheck. However after closing and reopening, it is back to instant. So it means that there is indeed a cache and that this cache is working great and that your fix actually fix the cache usage too when starting hls. I'm really excited because on all project I had worked with HLS, I had never ever imagined that there was a cache at startup (I always observe startup time of something like many minutes) and now I'm starting to dream about super fast HLS startup time.

[soulomoon]

I forget to change this rule previously.

[soulomoon]

Thank you so much for your detailed feedback. I am so happy that the fix worked !

For this PR, all the compilation rules that use GetModuleGraph rule switched to use GetFileModuleGraph file for each file, now they only depend on ModSummary fingerprints for the file and its depencies. (Originally they depend on the whole build graph's ModSummaries' fingerprints)

your fix actually fix the cache usage too when starting hls

Interesting, it seems to fix some thing more than I expected.

I just discovered that I forget to update the rule needsCompilationRule to use the new GetFileModuleGraph in my previous commits , now I changed that one too. Perhaps it would help with your @guibou codebase.

• One concern for this PR is that we have GetFileModuleGraph for each file, it would increase the memory usage since their results are all cached in shake. One improvement could be, do not return DependencyInformation directly, but use a IOref or something to share DependencyInformation for all the GetFileModuleGraph's result. But we might want to measure how much more mem the current approach would take to see if it worth changing it to IOref.
• The other concern is that ReverseModuleDeps might not be up to date. Possible fix, single out a new rule to compute ReverseModuleDeps for a file, so our rule system can keep it up to date.

[guibou]

@soulomoon regarding your concern about memory usage, I've discussed that a bit in #4598 and I'm sure we can shrink the memory usage of the dependency graph by a factor 10 at a minimum.

I'll try to work on that asap.

Interesting, it seems to fix some thing more than I expected.

It does make sense, I suppose (and hope) that modules with template haskell do have a ModSummary which is always invalidated. So your change does not consider irrelevant changes in these modules.