Re: [PATCH 1/2] patience diff: remove unnecessary string comparisons

From: Phillip Wood <phillip.wood123@gmail.com>
Date: May 5, 2021, 09:34 UTC
Message-ID: <87001425-8043-4c66-dbc2-637f05a7229f@gmail.com>
In-Reply-To: <xmqqpmy658e1.fsf@gitster.g>

On 05/05/2021 01:31, Junio C Hamano wrote:

Show 26 quoted lines

> "Phillip Wood via GitGitGadget" <gitgitgadget@gmail.com> writes:
> 
>> From: Phillip Wood <phillip.wood@dunelm.org.uk>
>>
>> xdl_prepare_env() calls xdl_classify_record() which arranges for the
>> hashes of non-matching lines to be different so lines can be tested
>> for equality by comparing just their hashes.
> 
> Hmph, that is a bit different from what I read from the comment in
> the post context of the first hunk, though.
> 
> 	/*
> 	 * After xdl_prepare_env() (or more precisely, due to
> 	 * xdl_classify_record()), the "ha" member of the records (AKA lines)
> 	 * is _not_ the hash anymore, but a linearized version of it.  In
> 	 * other words, the "ha" member is guaranteed to start with 0 and
> 	 * the second record's ha can only be 0 or 1, etc.
> 	 *
> 	 * So we multiply ha by 2 in the hope that the hashing was
> 	 * "unique enough".
> 	 */
> 
> The words "home" and "enough" hints to me that the "ha" member is
> not hash, but "lineralized version of it" (whatever it means) does
> not guarantee that two records with the same "ha" are identical, or
> does it?

By "hashes" I meant "the value of record->ha". That comment is a bit confusing. I think "linearized version of it" is referring to xdl_classify_record() assigning a unique integer to each unique input line starting from zero and increasing by one for each unique input line (the function is fairly easy to follow). I assume "unique enough" is referring to the line below the comment which takes the modulus of record->ha and record->ha is not evenly distributed over the whole integer range but bunched at the lower end.

The Myers implementation calls xdl_classify_record() and then only ever compares record->ha, it does not call xdl_recmatch() while computing the diff.

> Well, I should just go read xdl_classify_record() to see what it
> really does, but if it eliminates collisions, then the patch is a
> clear and obvious improvement.

Thanks

Phillip

Show 38 quoted lines

> Thanks.
> 
>> diff --git a/xdiff/xpatience.c b/xdiff/xpatience.c
>> index 20699a6f6054..db2d53e89cb0 100644
>> --- a/xdiff/xpatience.c
>> +++ b/xdiff/xpatience.c
>> @@ -90,7 +90,7 @@ static void insert_record(xpparam_t const *xpp, int line, struct hashmap *map,
>>   {
>>   	xrecord_t **records = pass == 1 ?
>>   		map->env->xdf1.recs : map->env->xdf2.recs;
>> -	xrecord_t *record = records[line - 1], *other;
>> +	xrecord_t *record = records[line - 1];
>>   	/*
>>   	 * After xdl_prepare_env() (or more precisely, due to
>>   	 * xdl_classify_record()), the "ha" member of the records (AKA lines)
>> @@ -104,11 +104,7 @@ static void insert_record(xpparam_t const *xpp, int line, struct hashmap *map,
>>   	int index = (int)((record->ha << 1) % map->alloc);
>>   
>>   	while (map->entries[index].line1) {
>> -		other = map->env->xdf1.recs[map->entries[index].line1 - 1];
>> -		if (map->entries[index].hash != record->ha ||
>> -				!xdl_recmatch(record->ptr, record->size,
>> -					other->ptr, other->size,
>> -					map->xpp->flags)) {
>> +		if (map->entries[index].hash != record->ha) {
>>   			if (++index >= map->alloc)
>>   				index = 0;
>>   			continue;
>> @@ -253,8 +249,7 @@ static int match(struct hashmap *map, int line1, int line2)
>>   {
>>   	xrecord_t *record1 = map->env->xdf1.recs[line1 - 1];
>>   	xrecord_t *record2 = map->env->xdf2.recs[line2 - 1];
>> -	return xdl_recmatch(record1->ptr, record1->size,
>> -		record2->ptr, record2->size, map->xpp->flags);
>> +	return record1->ha == record2->ha;
>>   }
>>   
>>   static int patience_diff(mmfile_t *file1, mmfile_t *file2,