ProgtoolsFeb 16
Aras Pranckevičius
“Meta-RFC: Long-term vision for improving build times” on LLVM dev list is good https://discourse.llvm.org/t/meta-rfc-long-term-vision-for-improving-build-times/89828
Meta-RFC: Long-term vision for improving build times

Introduction This RFC outlines a long-term vision and aggregated ideas for improving the build times. Generally this is more about mechanical and structural changes that can considerably improve build times in the long term. The short-term (<1 year) objective is to run a full build graph within the same, long-lived LLVM daemon process, moving beyond the traditional model of spawning a new OS process for each build action. This is motivated by the significant process creation and I/O overhead on...

LLVM Discussion Forums
Feb 16 at 6:45amWeb
Show threadChandler CarruthFeb 16

@aras

I'm a bit torn by a bunch of these directions....

On one hand, I do think there is a fundamentally more efficient architecture for builds. And I think this is at least (much) closer to that.

But on the other hand, I worry about the degree to which this (and many other efforts to improve Clang/LLVM compile performance) are baking in fundamentally inefficient architecture legacy from both C++ and LLVM... There is _so_ much efficiency on the table there, I can't help but suspect that if designed the build part once these are addressed, we'd end up with at least somewhat different architecture for the build as a whole.

Show threadAras PranckevičiusFeb 16

@chandlerc possibly. OTOH, when exactly llvm/clang will become much higher throughput than they are today?

A bunch of things Andrea talks about there are also orthogonal to llvm/clang: caching, process spawn overhead, etc.

Show threadJames WidmanFeb 17
@chandlerc @aras this is exactly the kind of topic (well, one of the topics) that i would want to see covered at SIGBUILD, if SIGBUILD existed
Show threadJames WidmanFeb 22

@chandlerc @aras er, on second thought...

since build strategies depend to some extent on language design... maybe papers/presentations on build systems belong under the SIGPLAN umbrella somewhere...? maybe at SPLASH?

Show threadStefan ReinalterFeb 16

@aras Alex has some really good ideas.

Now where's the AI agent army that will swoop in, learn the LLVM codebase and just build and refactor the code according to a few RFCs? Burn a few tokens, done in a week, ship it.

Show threadÉtienne ParmentierFeb 16
@aras I guess they are not considering switching to a more data oriented approach because the overhaul would be too big ? Zig has great compile Times with it’s DOD data structures.