% Generated by roxygen2: do not edit by hand
% Please edit documentation in R/topic-nse.R
\name{topic-embrace-constants}
\alias{topic-embrace-constants}
\title{Why are strings and other constants enquosed in the empty environment?}
\description{
Function arguments are \link[=topic-defuse]{defused} into \link[=topic-quosure]{quosures} that keep track of the environment of the defused expression.
\if{html}{\out{
}}\preformatted{quo(1 + 1)
#>
#> expr: ^1 + 1
#> env: global
}\if{html}{\out{
}}
You might have noticed that when constants are supplied, the quosure tracks the empty environment instead of the current environmnent.
\if{html}{\out{}}\preformatted{quos("foo", 1, NULL)
#> >
#>
#> [[1]]
#>
#> expr: ^"foo"
#> env: empty
#>
#> [[2]]
#>
#> expr: ^1
#> env: empty
#>
#> [[3]]
#>
#> expr: ^NULL
#> env: empty
}\if{html}{\out{
}}
The reason for this has to do with compilation of R code which makes it impossible to consistently capture environments of constants from function arguments. Argument defusing relies on the \emph{promise} mechanism of R for lazy evaluation of arguments. When functions are compiled and R notices that an argument is constant, it avoids creating a promise since they slow down function evaluation. Instead, the function is directly supplied a naked constant instead of constant wrapped in a promise.
}
\section{Concrete case of promise unwrapping by compilation}{
We can observe this optimisation by calling into the C-level \code{findVar()} function to capture promises.
\if{html}{\out{}}\preformatted{# Return the object bound to `arg` without triggering evaluation of
# promises
f <- function(arg) \{
rlang:::find_var(current_env(), sym("arg"))
\}
# Call `f()` with a symbol or with a constant
g <- function(symbolic) \{
if (symbolic) \{
f(letters)
\} else \{
f("foo")
\}
\}
# Make sure these small functions are compiled
f <- compiler::cmpfun(f)
g <- compiler::cmpfun(g)
}\if{html}{\out{
}}
When \code{f()} is called with a symbolic argument, we get the promise object created by R.
\if{html}{\out{}}\preformatted{g(symbolic = TRUE)
#>
}\if{html}{\out{ }}
However, supplying a constant to \code{"f"} returns the constant directly.
\if{html}{\out{}}\preformatted{g(symbolic = FALSE)
#> [1] "foo"
}\if{html}{\out{
}}
Without a promise, there is no way to figure out the original environment of an argument.
}
\section{Do we need environments for constants?}{
Data-masking APIs in the tidyverse are intentionally designed so that they don't need an environment for constants.
\itemize{
\item Data-masking APIs should be able to interpret constants. These can arise from normal argument passing as we have seen, or by \link[=topic-inject]{injection} with \verb{!!}. There should be no difference between \code{dplyr::mutate(mtcars, var = cyl)} and \code{dplyr::mutate(mtcars, var = !!mtcars$cyl)}.
\item Data-masking is an \emph{evaluation} idiom, not an \emph{introspective} one. The behaviour of data-masking function should not depend on the calling environment when a constant (or a symbol evaluating to a given value) is supplied.
}
}
\keyword{internal}