Re: [PATCH] cache-tree: do not cache empty trees

[Jonathan Nieder <jrnieder@gmail.com>]

Nguyễn Thái Ngọc Duy wrote:

> Let's do it in a consistent way, always disregard empty trees in
> index. If users choose to create empty trees their own way, they
> should not use index at all.

While this violates some seeming invariants, like

1.
	git reset --hard
	git commit --allow-empty
	git rev-parse HEAD^^{tree} >expect
	git rev-parse HEAD^{tree} >actual
	test_cmp expect actual

2.
	git reset --hard
	git revert HEAD
	if git rev-parse HEAD~2
	then
		git rev-parse HEAD~2^{tree} >expect
		git rev-parse HEAD^{tree} >actual
		test_cmp expect actual
	fi

, I think it's a good change.  Malformed modes in trees already break
those false invariants iiuc.

Thanks.

> --- a/cache-tree.c
> +++ b/cache-tree.c
> @@ -621,9 +621,18 @@ static void prime_cache_tree_rec(struct cache_tree *it, struct tree *tree)
>  			struct tree *subtree = lookup_tree(entry.sha1);
>  			if (!subtree->object.parsed)
>  				parse_tree(subtree);
> +			if (!hashcmp(entry.sha1, (unsigned char *)EMPTY_TREE_SHA1_BIN)) {
> +				warning("empty tree detected! Will be removed in new commits");
> +				cnt = -1;
> +				break;
> +			}

Aside from the warning, this part is an optimization, right?

>  			sub = cache_tree_sub(it, entry.path);
>  			sub->cache_tree = cache_tree();
>  			prime_cache_tree_rec(sub->cache_tree, subtree);
> +			if (sub->cache_tree->entry_count == -1) {
> +				cnt = -1;
> +				break;
> +			}

Would be nice to include a test for this, like so:

	subdir/
		empty1/
		subsubdir/
			empty2/
			empty3/

> --- /dev/null
> +++ b/t/t1013-read-tree-empty.sh
> @@ -0,0 +1,20 @@
> +#!/bin/sh
> +
> +test_description='read-tree with empty trees'
> +
> +. ./test-lib.sh
> +
> +EMPTY_TREE=4b825dc642cb6eb9a060e54bf8d69288fbee4904

If we _have_ to hard-code this (why?) then I'd prefer to do so
in test-lib.sh.

[...]
> +test_expect_success 'write-tree removes empty tree' '
> +	git read-tree `cat tree` &&
> +	git write-tree >actual
> +	echo $EMPTY_TREE >expected
> +	test_cmp expected actual

Broken &&-chain.  Not sure why this test relies on virtual objects
and another (independently nice) patch instead of adding the empty
tree to the object db --- is there something subtle I am missing?

The test does not distinguish between success due to git read-tree
omitting empty trees and success due to git mktree omitting empty
trees.

Hope that helps,
Jonathan