Fix regression with trap on left hand side of pipe.
Fix forced return from shell structure within nested function.
Fix tests exiting too early.
Add new test case.
If executing a command not at the end of a pipeline, and
not optimised in the calling code, we can fork before
"prefork" substitutions, at the same point as
for background commands.
Move fork before our preliminary scan of arguments to find
a command: in the cases of early fork
we don't need this information before forking.
Ensure we _exit if forked in execcmd_exec().
Rationalise use of forks and pipes.
Ensure we _exit instead of returning from execcmd_exec() if we
have forked. Before the optimisation code after the fork always ran
to the check at the end, but that code is overkill for the logic
between the early fork and the existing one.
Remove old workaround to fork in caller of execcmd for current shell
constructs as no longer needed with early fork below.
Close input of newly created pipe on fork (destined for RHS of pipe
which we never execute): this replaces a workaround from
zsh-workers/32171, commit 9887fc3d7b.
Set last1 on early fork as needed by some instances of shell
constructs on LHS of pipeline to know they are exiting.
Don't close associated file descriptors in the closem()
tidy up function as they should remain visible to external
processes. Override if about to exit.
Unit test for the failing case: note this relies on the
existence of /proc/self/fd or equivalent.
In execcmd the case of running the last command in a pipeline
asynchronously for the purpose of & and &! is easy to work out,
and we can avoid side effects and unnecessary execution time in
the parent shell by forking earlier.
If process group leader exits, allow a newly forked process to become
process leader. If a foreground job, reattach the shell to the
terminal until that happens.
Unblock signals when reading output for command subsitution so that
we can do this reattaching immediately.
Functions defined inside other fucntions needs file line number
adding. Particularly useful for anonymous fucntions.
Add flag to indicate a function is anonymous. Done up to now
by comparing the name to a pointer but this is more consistent.
Move detection of key/value pairs down into prefork().
Detect normal array assignment and [key]=val array assignemnt
separately. Mark key / value pairs with Marker and pass up flag. Deal
with marked triads specially later on.
Works for both normal and typeset case, also var+=...
Still to do: allow to be mixed with straight array assignment,
improve typeset -p, implement [key]+=value.
There was an exception to the usual ERR_EXIT pattern that causes
problems when executing a function in an else branch. It seems
the exception is no longer needed as the regression tests pass
without it.
This was turned off explicitly as it was assumed not to be needed,
but it is when code in a sourced file is suspended.
We'll need to see if this makes start up inefficient.
If typeset family builtins are recognised as keywords then the value
is handled as a separate expansion and we don't need the old
magic behaviour, even if the option is set.
If doing "autoload -X", the path present might actually be location
of file containing the function with the autoload -X. Add
an explicit flag to say it's a directory for autoload.
When updating shfunc structure to change filename we should be
careful to free whatever's there already, we may now be the
directory in which to find the function definition. After loading
the field contains the full name of the file.
path.
-d defaults to normal fpath
-r remembers the path without actually loading. May be combined with -d.
-R does the same but it's an error if not found
-X can now take a directory path: this is used to output not yet loaded
functions that have an associated path.
