Re: [PATCH 4/5] commit-graph: be extra careful about mixed generations

[Derrick Stolee <stolee@gmail.com>]

On 2/1/2021 1:04 PM, Taylor Blau wrote:
> On Mon, Feb 01, 2021 at 05:15:06PM +0000, Derrick Stolee via GitGitGadget wrote:
...
>>  	struct topo_level_slab *topo_levels;
>>  	const struct commit_graph_opts *opts;
>> @@ -1452,6 +1453,15 @@ static void compute_generation_numbers(struct write_commit_graph_context *ctx)
>>  		ctx->progress = start_delayed_progress(
>>  					_("Computing commit graph generation numbers"),
>>  					ctx->commits.nr);
>> +
>> +	if (ctx->write_generation_data && !ctx->trust_generation_numbers) {
>> +		for (i = 0; i < ctx->commits.nr; i++) {
>> +			struct commit *c = ctx->commits.list[i];
>> +			repo_parse_commit(ctx->r, c);
>> +			commit_graph_data_at(c)->generation = GENERATION_NUMBER_ZERO;
>> +		}
>> +	}
>> +
> 
> This took me a while to figure out since I spent quite a lot of time
> thinking that you were setting the topological level to zero, _not_ the
> corrected committer date.
> 
> Now that I understand which is which, I agree that this is the right way
> to go forward.
> 
> That said, I do find it unnecessarily complex that we compute both the
> generation number and the topological level in the same loops in
> compute_generation_numbers()...
> 
>>  	for (i = 0; i < ctx->commits.nr; i++) {
>>  		struct commit *c = ctx->commits.list[i];
>>  		uint32_t level;
>> @@ -1480,7 +1490,8 @@ static void compute_generation_numbers(struct write_commit_graph_context *ctx)
>>  				corrected_commit_date = commit_graph_data_at(parent->item)->generation;
>>
>>  				if (level == GENERATION_NUMBER_ZERO ||
>> -				    corrected_commit_date == GENERATION_NUMBER_ZERO) {
>> +				    (ctx->write_generation_data &&
>> +				     corrected_commit_date == GENERATION_NUMBER_ZERO)) {
> 
> ...for exactly reasons like this. It does make sense that they could be
> computed together since their computation is indeed quite similar. But
> in practice I think you end up spending a lot of time reasoning around
> complex conditionals like these.
> 
> So, I feel a little bit like we should spend some effort to split these
> up. I'm OK with a little bit of code duplication (though if we can
> factor out some common routine, that would also be nice). But I think
> there's a tradeoff between DRY-ness and understandability, and that we
> might be on the wrong side of it here.

You're probably right that it is valuable to split the computations.
It would allow us to skip all of the "if (ctx->write_generation_data)"
checks in this implementation and rely on the callers to make that
choice.

>>  					all_parents_computed = 0;
>>  					commit_list_insert(parent->item, &list);
>>  					break;
>> @@ -1500,12 +1511,15 @@ static void compute_generation_numbers(struct write_commit_graph_context *ctx)
>>  					max_level = GENERATION_NUMBER_V1_MAX - 1;
>>  				*topo_level_slab_at(ctx->topo_levels, current) = max_level + 1;
>>
>> -				if (current->date && current->date > max_corrected_commit_date)
>> -					max_corrected_commit_date = current->date - 1;
>> -				commit_graph_data_at(current)->generation = max_corrected_commit_date + 1;
>> -
>> -				if (commit_graph_data_at(current)->generation - current->date > GENERATION_NUMBER_V2_OFFSET_MAX)
>> -					ctx->num_generation_data_overflows++;
>> +				if (ctx->write_generation_data) {
>> +					timestamp_t cur_g;
>> +					if (current->date && current->date > max_corrected_commit_date)
>> +						max_corrected_commit_date = current->date - 1;
>> +					cur_g = commit_graph_data_at(current)->generation
>> +					      = max_corrected_commit_date + 1;
>> +					if (cur_g - current->date > GENERATION_NUMBER_V2_OFFSET_MAX)
>> +						ctx->num_generation_data_overflows++;
>> +				}
> 
> Looks like two things happened here:
> 
>   - A new local variable was introduced to store the value of
>     'commit_graph_data_at(current)->generation' (now called 'cur_g'),
>     and
> 
>   - All of this was guarded by a conditional on
>     'ctx->write_generation_data'.
> 
> The first one is a readability improvement, and the second is the
> substantive one, no?

Yes. Adding these checks and tabs made things super-wide, so cur_g
exists only for readability. If we split the computation, then this
is no longer required.

Thanks,
-Stolee