Bsb build order is seemingly non-deterministic (wrong-deterministic?)

[glennsl]

When doing a refactor that spans multiple files, bsb doesn't "stick" to reporting errors in the file you are currently working on. This leads to having to jump back and forth between different files to make one error fix attempt per file at a time, which then causes you to lose track of what you were doing in one file when bsb temporarily moves on to some other file.

Bsb should perhaps instead prioritize the last modified file, if that's possible.

[bobzhang]

@glennsl we can do it this way, suppose we know file_a is changed and file watcher is correct (sort of by Node), we call

ninja path/to/file_a.cm{j,i} all

(I need check if ninja respect the order though)

If it does, it should work

[glennsl]

I've looked around a bit, and I've not found anything to indicate that target order is respected, but I have found evidence to the contrary: ninja-build/ninja#1055

I've also found this exact issue discussed in several issues dating as far back as 2011, with no resolution in sight:

• Build edges that previously failed asap ninja-build/ninja#60
• explicit control of the build order ninja-build/ninja#232
• Use better logic for picking which node to build first ninja-build/ninja#376

The only possible solution I've found is to set up explicit order-only dependencies in the .ninja file, which would at least make it deterministic. But introducing additional constraints will also remove opportunities for parallelism, which might have a significant effect on performance, especially for full builds.

[bobzhang]

@glennsl actually, we can do this from bsb side without bother ninja. so you get a file changed, you just recompile that file (assume that we compute the compiler flags from bsb) and run ninja later(no matter the former succeeds or not).

[glennsl]

Ok, so when bsb_watcher calls back to bsb it should pass the name of the file that's changed (if avialable), bsb picks that up, execs bsc for just that file, then proceed to run ninja as usual?

[bobzhang]

Yes, it requires some refactoring though

[bobzhang]

@glennsl This will help reduce the latency even for very large projects, I think it is the right direction