make `proc nimQuery(setting: Setting): RootObj {.magic.}` => return a type that depends on the enum

[timotheecour]

@araq ok with following?

proposal

• deprecate compileOption, querySettingSeq and querySetting (or "soft" deprecate ie recommend the alternative here)
• introduce

  # in compilesettings
  proc nimQuery*(setting: Setting): RootObj {.compileTime, magic.}

  why RootObj ? because that's what used in iterator fields*[T: tuple|object](x: T): RootObj to allow returning a type that can depend on input argument

example

doAssert nimQuery(nimVerCT) == (1,3,7)
  # I need this in many places, after this we don't need to define new symbols in condsyms anymore since we can test for nimVerCT
  # i had originally introduced it in https://github.com/nim-lang/Nim/pull/14447#issuecomment-633725338 but moved it out of there
doAssert nimQuery(assertions) # can replace `compileOption("assertions")`
doAssert nimQuery(backend) == Backend.js
doAssert nimQuery(arguments) is seq[string]
doAssert Experimental.notnil in nimQuery(experimental)

note

compilesettings stays a "low-level" module in the sense it's usable by low level modules (eg doesn't depend on iterators, dollars etc)

implementation

no particular difficulty

[Araq]

Oh my, please ... no. I think it's a bad design, reminds me of sysctl. It's faking simplicity.

[timotheecour]

i was half expecting that, and I kind of agree. Then this?

proc nimVerCT*(): tuple(int,int,int)
proc getExperimental*(): seq[Experimental] # type Experimental = enum ...
proc getBackend(): Backend
when nimVerCT() < (1,3,5) : ...
when Experimental.notnil in getExperimental(): ...

[Araq]

What's wrong with the existing solutions?

[timotheecour]

• Hungarian-ish notation is redundant (querySetting, querySettingSeq, querySettingBool, ...)

• like you've said yourself, stringly typed API's ("everything is a string / seq[string]") aren't user friendly in a static language that makes things like enum etc trivial to use

  # we shouldn't need serialization + deserialization
  doAssert querySetting("nimverct").parseTo((int,int,int)) == (1,3,5)
  doAssert querySetting("assertions").parseTo(bool) # or reach for `system.compileOption` ... but why

  and the compiler can't always help you if you misuse:

  doAssert "forLoopmacros" in querySetting("experimental")  # bug (wrong style => you'll silently get wrong answer)
  doAssert "forLoopMacros".normalize in querySetting("experimental") # still buggy if you mistype

This is more user friendly, self documenting and impossible to misuse or you'll get CT error:

# either the "sysctl"-like way i gave in top post, or as individual procs, I'm agnostic:
doAssert nimVerCT() == (1,3,5)
doAssert Experimental.forLoopMacros in getExperimental()
doAssert assertions() # or maybe `nimQuery(assertions)` still in this case

• a 2ndary goal is to consolidate these in compilesettings so you don't have some arbitrary set of flags in system.compileOption and some other arbitrary set of flags in compilesettings.{querySetting,querySettingSeq}

[Araq]

Well as long as it doesn't bloat system.nim further, I'm fine with more type-safe additions.