Re: [PATCH] range-diff: fix some 'hdr-check' and sparse warnings

From: Johannes Sixt <j6t@kdbg.org>
Date: Jul 13, 2019, 10:44 UTC
Message-ID: <ec635d0d-00ca-2419-3c1a-9b0343b46daa@kdbg.org>
In-Reply-To: <xmqqy313p5hn.fsf@gitster-ct.c.googlers.com>

Am 12.07.19 um 18:44 schrieb Junio C Hamano:

Show 28 quoted lines

> Johannes Sixt <j6t@kdbg.org> writes:
> 
>> Am 12.07.19 um 00:03 schrieb Ramsay Jones:
>>> diff --git a/range-diff.c b/range-diff.c
>>> index ba1e9a4265..0f24a4ad12 100644
>>> --- a/range-diff.c
>>> +++ b/range-diff.c
>>> @@ -102,7 +102,7 @@ static int read_patches(const char *range, struct string_list *list)
>>>  		}
>>>  
>>>  		if (starts_with(line, "diff --git")) {
>>> -			struct patch patch = { 0 };
>>> +			struct patch patch = { NULL };
>>
>> There is nothing wrong with 0 here. IMHO, zero-initialization should
>> *always* be written as = { 0 } and nothing else. Changing 0 to NULL to
>> pacify sparse encourages a wrong style.
> 
> Hmm, care to elaborate a bit?  Certainly, we have a clear preference
> between these two:
> 
> 	struct patch patch;
> 	patch.new_name = 0;
> 	patch.new_name = NULL;
> 
> where the "char *new_name" field is the first one in the structure.
> We always try to write the latter, even though we know they ought to
> be equivalent to the language lawyers.

I'm not questioning this case; the latter form is clearly preferable.

Using only = { 0 } for zero-initialization makes the code immune to rearrangements of the struct members. That is not the case with = { NULL } because it requires that the first member is a pointer; if rearrangement makes the first member a non-pointer, the initializations must be adjusted.

On the other hand, I'm not arguing that

  struct string_list dup_it = { NULL, 0, 0, 1, NULL };

should be written as

  struct string_list dup_it = { 0, 0, 0, 1, 0 };

I'm only complaining about the single-initializer = { 0 } "please initialize this whole struct with zero values" form.

Show 8 quoted lines

> Is the reason why you say 0 is fine here because we consider
> 
> 	struct patch patch, *dpatch;
> 	memset(&patch, 0, sizeof(patch));
> 	dpatch = xcalloc(1, sizeof(patch));
> 
> are perfectly good way to trivially iniitialize an instance of the
> struct?

Absolutely not. Both forms are evildoing as far as struct initialization is concerned because they ignore the types of the members. The memset form should always be replaced by = { 0 }. The correct replacement for the xcalloc form would be

	struct patch zero = { 0 };
	struct patch *dpatch = xmalloc(sizeof(*dpatch));
	*dpatch = zero;

but I do understand that this transformation is unacceptably verbose.

> Do we want to talk to sparse folks about this?

I've no idea which camp they are in. How would they respond to an exceptional case that is also very much a matter of taste?

-- Hannes