me: "i like it how the entire LLVM project is one monorepo"
her: "when are you going to compile the entirety of LLVM to WebAssembly?"
me: "yes."
her: "... I said that as a joke"
if i recall, linking with -flto=thin is nearly as bad since basically all time is spent in the instruction selector, and linking without lto is still pretty bad, but it's not quite "twenty minutes each time" bad
@whitequark This genuinely might be the most blessed thing I’ve seen all month. Only a lunatic would attempt this, let alone complete it. Congratulations!
@whitequark Does this mean I could have an online C compiler test environment entirely in my browser? Hell, could it be ported so that the Compiler Explorer runs entirely in the browser?
@whitequark First thing I'd want to do is a client-side Rust playground.
I also have hopes to build a Python build backend that would run Python and a C compiler in WASM to have deterministic, safe, builds, but that is a bit of a deep hole :)
afaict the compiler driver calls wasm-opt unconditionally if:
it is on PATH
you pass -On with n>0 to the linker
so to make it not spend 10 minutes of real time in wasm-opt you need to pass -O0 to the linker. i only figured this out by reading the sources of the compiler driver.
nobody should be figuring out compiler options by reading the sources of the compiler driver. nobody!
so now that i've generated half a gig of debug information i can just feed it into wasmtime, have it generate DWARF for me, and be done with it, right?
wrong.
thread 'main' panicked at /home/whitequark/.cargo/registry/src/index.crates.io-6f17d22bba15001f/wasmtime-cranelift-19.0.1/src/compiler.rs:565:13:
do you think the debug information was valid? nope! wasm-opt unconditionally launched by wasm-ld (which performs link time optimization despite lack of -flto) also corrupted the DWARF data
did adding -O0 turn it off? yes! did that make the DWARF data seemingly well formed? yes! did it give me line numbers? of course not, it just gave me function names
i have solved my problem by disassembling the wasm module (i went through three different disassemblers until i found one that is able to finish this in finite time) and correlating it manually with the source code
... i'm pretty sure that wasmtime isn't reporting the trap correctly T__T
there should be one more frame on top of the frame 0
@whitequark Oooh, I admire your patience and determination! Does whitequark refer to quantum physics or something, or even have a meaning beyond surface level (although I personally need an explanation even for the surface level)?
@whitequark ...do you need more than the ability to run processes synchronously? because if you just need "processes as function calls" we don't even need process management.
@SoniEx2 yep, that's the plan. I want it to be fully synchronous because SAB is kind of a pain, though I recently discovered that you could use service workers to add CORS headers client-side apparently?
@whitequark it may be NP-hard to prove whether advanced optimizers terminate? I’m pretty sure there are only subset proofs. many optimizers just „hammer things until nothing changes“ - without any guarantee of local/global optimality.
Add comment