Re: [PATCH v2] parallel-checkout: avoid dash local bug in tests

["René Scharfe" <l.s.r@web.de>]

Am 05.06.21 um 21:09 schrieb SZEDER Gábor:
> On Sat, Jun 05, 2021 at 08:11:24PM +0200, René Scharfe wrote:
>> The parallel checkout tests fail when run with /bin/dash on MacOS 11.4,
>> reporting the following error:
>>
>>    ./t2080-parallel-checkout-basics.sh: 33: local: 0: bad variable name
>>
>> That's because wc's output contains leading spaces and this version of
>> dash erroneously expands the variable declaration as "local workers= 0",
>> i.e. it tries to set the "workers" variable to the empty string and also
>> declare a variable named "0", which not a valid name.  This is a known
>> dash bug (https://bugs.launchpad.net/ubuntu/+source/dash/+bug/139097).
>
> Perhaps a more accurate wording for this bug would be:
>
>   ... and even fairly recent versions of dash erroneously perform
>   field splitting on the expansion of the command substitution before
>   assigning it to a local variable.

OK.

>
> I think the relevant part of POSIX is section 2.9.1 Simple Commands:
>
>   4. Each variable assignment shall be expanded for tilde expansion,
>      parameter expansion, command substitution, arithmetic expansion,
>      and quote removal prior to assigning the value.
>
>   https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html#tag_18_09_01
>
> Note that it didn't mention field splitting; though POSIX doesn't
> specifies local variables in the first place, so...
>
> Anyway, this bug has been fixed in v0.5.11 (2020-06-01).
> This is an old bug, it was already present in v0.5.5 (2009-01-13); I
> didn't check earlier versions.
>
>> Work around it by passing the command output directly to test instead of
>> storing it in a variable first.  While at it, let grep count the number
>> of lines instead of piping its output to wc, which is a bit shorter and
>> more efficient.
>
> A more debug-friendly alternative would be to save 'grep's output to a
> temporary file and use 'test_line_count = $expected_workers'.

Yes, but that would cement the use of grep and wc -l and I still can't
let go of the idea that grep -c would be slightly quicker.  And it would
add file I/O.

Something like this would be more efficient in the expected case:

	if test $expected_count -ne $(grep -c -e "$pattern" "$file")
	then
		echo "Expected $expected_count lines matching $patter, but got:"
		grep -e "$pattern" "$file"
		return 1
	fi

I have no performance numbers to show, just the vague feeling that the
test suite takes way too long already.

Anyway, for now a minimal fix should do.  Debug features can be added to
this case and several similar ones later.

>
>> Helped-by: Matheus Tavares Bernardino <matheus.bernardino@usp.br>
>> Signed-off-by: René Scharfe <l.s.r@web.de>
>> ---
>> Changes since v1:
>> - Explain the root cause.
>> - Get rid of the local variable "workers".
>> - Adjust title accordingly.
>> - Still use grep -c, though.
>> - Remove input redirection.
>>
>>  t/lib-parallel-checkout.sh | 3 +--
>>  1 file changed, 1 insertion(+), 2 deletions(-)
>>
>> diff --git a/t/lib-parallel-checkout.sh b/t/lib-parallel-checkout.sh
>> index 21f5759732..66350d5207 100644
>> --- a/t/lib-parallel-checkout.sh
>> +++ b/t/lib-parallel-checkout.sh
>> @@ -27,8 +27,7 @@ test_checkout_workers () {
>>  	rm -f "$trace_file" &&
>>  	GIT_TRACE2="$(pwd)/$trace_file" "$@" 2>&8 &&
>>
>> -	local workers=$(grep "child_start\[..*\] git checkout--worker" "$trace_file" | wc -l) &&
>> -	test $workers -eq $expected_workers &&
>> +	test $(grep -c "child_start\[..*\] git checkout--worker" "$trace_file") -eq $expected_workers &&
>>  	rm "$trace_file"
>>  } 8>&2 2>&4
>>
>> --
>> 2.31.1
>>