evalOnce alternative

[mratsim]

Just FYI, I noticed that overloading via parameter constraints also work as an evalOnce alternative

nim-stew/stew/evalonce.nim

Lines 12 to 42 in 30dadcd

	macro evalOnceAs*(exp, alias: untyped): untyped =
	## Ensure that `exp` is evaluated only once unless it is a symbol in which
	## case it's used directly.
	##
	## A common case where this is useful is template parameters which, when
	## an expression is passed in, get evaluated multiple times.
	##
	## Based on a similar macro in std/sequtils
	expectKind(alias, nnkIdent)
	let
	body = nnkStmtList.newTree()
	val =
	if exp.kind == nnkSym:
	# The symbol can be used directly
	# TODO dot expressions? etc..
	exp
	else:
	let val = genSym(ident = "evalOnce_" & $alias)
	body.add newLetStmt(val, exp)
	val
	body.add(
	newProc(
	name = genSym(nskTemplate, $alias),
	params = [getType(untyped)],
	body = val,
	procType = nnkTemplateDef,
	)
	)
	body

vs

https://github.com/mratsim/tattletale/blob/5555745/workspace/libtorch/src/tensors.nim#L755-L766

  template `[]`*(t: Tensor{call}, args: varargs[untyped]): untyped =
    let tmp = t
    convertLibTorchExceptions:
      wrapTorchTensor:
        tmp.raw[args]

  template `[]`*(t: Tensor{`let`|`var`|`const`|lvalue|param}, args: varargs[untyped]): untyped =
    convertLibTorchExceptions:
      wrapTorchTensor:
        t.raw[args]

Though it's true that I didn't try with sym constraint, I'm not sure if it works in untyped templates.