Re: redirections preceded by a word with builtins inconsistent

[Dan Douglas]

On Sun, May 22, 2016 at 1:06 PM, Chet Ramey <chet.ramey@case.edu> wrote:
> On 5/21/16 5:16 PM, adonis papaderos wrote:
>
>> Bash Version: 4.3
>> Patch Level: 42
>> Release Status: release
>>
>> Description:
>>     When using redirections preceded by a word on builtins
>>         'i.e. : {varname}<&1'
>>         the redirection persists to the current shell.
>
> This is by design, and has been discussed previously.  One such discussion
> starts at http://lists.gnu.org/archive/html/bug-bash/2016-01/msg00156.html.
> Maybe it should only do it when preceded by `exec', but it never has.
>
>>         Later using {varname}<&- does not close fd.
>
> You don't precede it with an `exec', so it doesn't affect the current
> shell.  So-called `null' commands with redirections are executed in
> subshells.  This is where you need the `exec'.

I know POSIX allows for a subshell but don't know what the point is. I
was confused by this when I first hit it. The subshell isn't obvious.

 $ bash -c 'echo "$BASHPID"
{x[$((pid=BASHPID))$(echo "2: $pid" >&2),0]}<<<"$((pid=BASHPID))$(echo
"1: $pid" >&2)"'
3907
1: 3908
2: 3908
 $ bash -c 'echo "$BASHPID"
{x[$((pid=BASHPID))$(echo "2: $pid" >&2),0]}<<<"$((pid=BASHPID))$(echo
"1: $pid" >&2)"'
3911
1: 3911
2: 3911

ksh being the only other shell that supports auto-fd assignment for
null commands doesn't create one.

 $ bash -c 'typeset -li y=2
{x[y*=2]}<&$((y*=2,0)); typeset -p x y'
bash: line 1: typeset: x: not found
declare -il y="2"
 $ bash -c 'typeset -li y=2
: {x[y*=2]}<&$((y*=2,0)); typeset -p x y'
declare -a x='([8]="10")'
declare -il y="8"
 $ ksh -c 'typeset -li y=2
{x[y*=2]}<&$((y*=2,0)); typeset -p x y'
typeset -a -i x=([8]=11)
typeset -l -i y=8
 $ ksh -c 'typeset -li y=2
: {x[y*=2]}<&$((y*=2,0)); typeset -p x y'
typeset -a -i x=([8]=11)
typeset -l -i y=8

ksh and zsh are hard to test since they do a horrible job parsing the
{fd} in all but trivial cases.

It would be way simpler to only allow redirects at the end of a
command and require each redirect to be a separate word. That
eliminates a ton of ambiguities for {fd}>x, &>x and >&x.

Maybe add a `set -k` work-alike for legacy scripts and people that insist on
`echo >&2 i like surprise redirec{t}<ion`.