Thanks Dmitry.

There's already a targets/ONESHELL (ONESHELL is a special target, not a variable) so I added your tests there, except for the first one (we definitely don't want to use echo -n as it's not portable at all; there are no portable arguments to the echo command) which I removed as there are already tests there that can substitute for it.

This is not working on Windows, however.  Windows has some problems because construct_command_argv_internal() assumes that any non-POSIX shell is cmd.exe and creates a cmd.exe batch file which of course does not work with Perl (or Python or whatever).

Really I cannot understand why the command line construction code is so incredibly complicated.  However, of course this new behavior doesn't work in previous versions of make either so it's not a regression.  I'm going to play with fixing this and if it doesn't seem simple enough I may just disable the new ONESHELL tests on Windows, for now.

The patch in the attachment has a fix against the current master and a test.

(file #52654, file #52655)

A user reported the following

.ONESHELL causes quoting in .SHELLFLAGS to not work as expected.

$ cat Makefile
.ONESHELL:
SHELL = /usr/bin/perl
.SHELLFLAGS = -w -E 'use warnings FATAL => "all";' -E
all:
        print 'it works'
$ make all
print 'it works'
Can't find string terminator "'" anywhere before EOF at -e line 1.
make: * [Makefile:5: all] Error 255
$

Without .ONESHELL: the above works as expected.

The original bug report is here
https://lists.gnu.org/archive/html/bug-make/2022-01/msg00024.html.