Yorick Peterse (Oct 28 2022 at 14:52): Yorick Peterse (Oct 28 2022 at 14:52):Does Cranelift have anything to make it easier to maintain stacktraces that map back to your source files? Right now in my interpreter I essentially maintain a mapping of instruction indexes to locations in a separate data structure. From what I gather this would require fiddling with dwarf/libunwind? The other option is to insert a bunch of code in places where necessary. This is sort of what Nim does, but apparently this comes with quite a bit of overhead (https://github.com/nim-lang/Nim/issues/12702).
Yorick Peterse (Oct 28 2022 at 14:53):Basically how my setup works now is that every frame stores the method it belongs to, and the method has a mapping of instruction index => location. Building a stacktrace is then a matter of walking those frames and finding an entry for the current index in each frame. Mind you I only need traces in the event of a panic in my language, they're not used for e.g. exceptions, so maximum performance isn't my main interest
Yorick Peterse (Oct 28 2022 at 15:14):Alternatively perhaps backtrace-rs is an option :thinking:
Chris Fallin (Oct 28 2022 at 17:23):We have support for emitting DWARF info for frames; Wasmtime used to use this, before it grew its own backtracer
Chris Fallin (Oct 28 2022 at 17:23):Right now what we do in Wasmtime is turn on the Cranelift option to always use frame pointers; then we can trace the linked list of (fp, return addr) tuples very quickly, and look up the addresses
Chris Fallin (Oct 28 2022 at 17:24):separately, srclocs get mapped to offsets in the output code and so you can correlate PCs back to whatever source-level info your srclocs represent
Yorick Peterse (Oct 28 2022 at 18:43):Ah thanks, I'll dig around to see if I could do something similar :smile:
Last updated: Dec 23 2024 at 12:05 UTC