glfw: Denormalize Errors

[hexops]

glfw: Denormalize Errors

[slimsag]

Fair enough, I think you've convinced me. Denormalization of all errors as you described originally SGTM!

[slimsag]

Very interesting arguments.

The one risk I see with error de-normalization is that we don't catch such an issue until someone's application makes its way into production, and they begin to experience a slew of unreachable panics on rarer system configurations (e.g. Wayland, obscure Linux distros/GPU drivers, etc.)

Looking at the current Error set and trying to group them into logical sets, we have:

• Can likely be eliminated entirely:
  • InvalidEnum
  • InvalidValue
• Graphics errors:
  • NoCurrentContext
  • APIUnavailable
  • VersionUnavailable
  • NoWindowContext
• Window creation and clipboard interaction:
  • FormatUnavailable
• Generic errors:
  • OutOfMemory
  • PlatformError

I think that I would feel comfortable with denormalizing all errors except those listed under "Generic errors" above. I am worried that basically any GLFW function can generate those two and we won't really know when/how unless we did a full audit of GLFW's sources

But I can also see your argument here about just denormalizing everything and improving GLFW docs when we run into anything inconsistent. I think overall I am OK with either of these approaches. After reading what I wrote above, what do you think the best path forward is?

[InKryption]

You make a strong point about denormalization being quite potentially harmful to production. Though, I think it then comes down to a trade-off: take the risk, causing inconvenience and some struggle, with the return being an overall net-positive in the long run, or avoid taking the risk, and simply continue with status-quo behaviour being, as you've described it, aspirational, as it pertains to documented behaviour, and actual behaviour.

After thinking through it for a bit, I would say I am still of the opinion that error denormalization is worth the benefits; whilst it does rock the boat, it's also a step in the same direction that pointed out the undefined behaviour that was present in GLFW for 6+ years.

In an extreme case, where this change causes too much havoc, we could re-normalize errors, and we'd come back to where we are now, but with additional information about the weaknesses in GLFW's error handling model, so still a net-positive.