diff --git a/tests/reg-encodings.R b/tests/reg-encodings.R
index de66237f184..9d35a9ddec5 100644
--- a/tests/reg-encodings.R
+++ b/tests/reg-encodings.R
@@ -336,3 +336,22 @@ source(tf, encoding = c("UTF-8", "latin1"))
 ## in R 4.2.{0,1} gave Warning (that would now be an error):
 ##   'length(x) = 2 > 1' in coercion to 'logical(1)'
 if (UTF8) stopifnot(identical(Encoding(x), "UTF-8"))
+
+## Check that UTF-16 with BOM can be read from a connection.  This tests a
+## work-around in R for a bug in libiconv-86 on macOS (at least since
+## libiconv-107).
+words <- c(0xfeff, 0x30+c(1:9,0,1:9,0), 0x0a) # bom + 12345678901234567890 + newline
+hi <- as.raw(words %/% 0x100)
+low <- as.raw(words %% 0x100)
+be <- c(rbind(hi, low))
+befile <- tempfile("be_", fileext=".txt")
+writeBin(be, befile)
+becon <- file(befile, encoding = "UTF-16", open="r")
+stopifnot(identical(readLines(becon), "12345678901234567890"))
+close(becon)
+le <- c(rbind(low, hi))
+lefile <- tempfile("le_", fileext=".txt")
+writeBin(le, lefile)
+lecon <- file(lefile, encoding = "UTF-16", open="r")
+stopifnot(identical(readLines(lecon), "12345678901234567890"))
+close(lecon)