Keep the exit status when the CLI cannot print an error to stderr

lib/elixir/lib/kernel/cli.ex

@@ -185,7 +185,15 @@ defmodule Kernel.CLI do
 
   defp print_error(kind, reason, stacktrace) do
     {banner, rest} = format_error_parts(kind, reason, stacktrace)
-    IO.write(:stderr, [IO.ANSI.format([:red, banner]), rest])
+    error = [IO.ANSI.format([:red, banner]), rest]
+
+    # stderr may not be writable (a closed descriptor, or a pipe whose reader
+    # has exited): the error cannot be shown, but the exit status still applies
+    try do
+      IO.write(:stderr, error)
+    rescue
+      _ -> :ok
+    end
   end
 
   defp blame_match(%{match?: true, node: node}), do: blame_ansi(:normal, "+", node)

lib/elixir/test/elixir/kernel/cli_test.exs

@@ -184,6 +184,21 @@ defmodule Kernel.CLI.ExecutableTest do
              "false\n"
   end
 
+  @tag :unix
+  @tag :tmp_dir
+  test "exits with status 1 when the error cannot be written to stderr",
+       %{tmp_dir: tmp_dir, cli_extension: cli_extension} do
+    # stderr is a pipe whose reader has exited, so printing the error fails;
+    # the status must still be the error's, and no crash dump written
+    cmd =
+      ~c"cd #{tmp_dir} && (#{elixir_executable(cli_extension)} -e \"raise ArgumentError\" " ++
+        ~c"2>&1 >/dev/null; echo $? > status) | true"
+
+    :os.cmd(cmd)
+    assert File.read!(Path.join(tmp_dir, "status")) == "1\n"
+    refute File.exists?(Path.join(tmp_dir, "erl_crash.dump"))
+  end
+
   stderr_test "blames exceptions", %{cli_extension: cli_extension} do
     error = elixir(~c"-e \"Access.fetch :foo, :bar\"", cli_extension)
     assert error =~ "** (FunctionClauseError) no function clause matching in Access.fetch/2"