Re: [PATCH v4 2/5] t: move reftable/tree_test.c to the unit testing framework

[Chandra Pratap <chandrapratap3519@gmail.com>]

On Thu, 18 Jul 2024 at 13:40, Karthik Nayak <karthik.188@gmail.com> wrote:
>
> Chandra Pratap <chandrapratap3519@gmail.com> writes:
>
> > On Thu, 18 Jul 2024 at 03:45, Justin Tobler <jltobler@gmail.com> wrote:
> >>
> >> On 24/07/17 08:00PM, Chandra Pratap wrote:
> >> > On Wed, 17 Jul 2024 at 18:09, Karthik Nayak <karthik.188@gmail.com> wrote:
> >> > >
> >> > > Chandra Pratap <chandrapratap3519@gmail.com> writes:
> >> > >
> >> > > > +struct curry {
> >> > > > +     void *last;
> >> > > > +};
> >> > > > +
> >> > > > +static void check_increasing(void *arg, void *key)
> >> > > > +{
> >> > > > +     struct curry *c = arg;
> >> > > > +     if (c->last)
> >> > > > +             check_int(t_compare(c->last, key), <, 0);
> >> > > > +     c->last = key;
> >> > > > +}
> >> > > > +
> >> > > > +static void t_tree(void)
> >> > > > +{
> >> > > > +     struct tree_node *root = NULL;
> >> > > > +     void *values[11] = { 0 };
> >> > >
> >> > > Although we were comparing 'char' above, here we have a 'void *' array.
> >> > > Why?
> >> >
> >> > The array is passed as a parameter to the 'tree_search()' function which
> >> > requires a void * parameter (i.e. a generic pointer). In the comparison
> >> > function (also passed as a parameter), we cast it to our expected type
> >> > (a character pointer) and then perform the required comparison.
> >>
> >> The point of `values` is to provide a set of values of type `void **` to
> >> be inserted in the tree. As far as I can tell, there is no reason for
> >> `values` to be initialized to begin with and is a bit misleading. Might
> >> be reasonable to remove its initialization here.
> >
> > The thing is, the values[] array being 0-initialized makes debugging
> > a lot easier in the case of a test failure, so I'm not very sure about
> > getting rid of the initialization here.
> >
> >> > > > +     struct tree_node *nodes[11] = { 0 };
> >> > > > +     size_t i = 1;
> >> > > > +     struct curry c = { 0 };
> >> > > > +
> >> > > > +     do {
> >> > > > +             nodes[i] = tree_search(values + i, &root, &t_compare, 1);
> >> > > > +             i = (i * 7) % 11;
> >> > >
> >> > > It gets weirder, we calculate 'i' as {7, 5, 2, 3, 10, 4, 6, 9, 8, 1}. We
> >> > > use that to index 'values', but values is '0' initialized, so we always
> >> > > send '0' to tree_search? Doesn't that make this whole thing a moot? Or
> >> > > did I miss something?
> >> >
> >> > We don't use 'i' to index 'values[]', we use it to calculate the next pointer
> >> > address to be passed to the 'tree_search()' function (the pointer that is 'i'
> >> > ahead of the pointer 'values'), which isn't 0.
> >>
> >> The `i = (i * 7) % 11;` is used to deterministically generate numbers
> >> 1-10 in a psuedo-random fashion. These numbers are used as memory
> >> offsets to be inserted into the tree. I suspect the psuedo-randomness is
> >> useful keys should be ordered when inserted into the tree and that is
> >> later validated as part of the in-order traversal that is performed.
> >
> > That's right, the randomness of the insertion order is helpful in validating
> > that the tree-functions 'tree_search()' and 'infix_walk()' work according
> > to their defined behaviour.
> >
> >> While rather compact, I find the test setup here to rather difficult to
> >> parse. It might be a good idea to either provide comments explaining
> >> this test setup or consider refactoring it. Honestly, I'd personally
> >> perfer the tree setup be done more explicitly as I think it would make
> >> understanding the test much easier.
> >
> > This probably ties in with the comments by Patrick on the previous iteration
> > of this patch, that using 'tree_search()' to insert tree nodes leads to
> > confusion. Solving that would require efforts outside the scope of this
> > patch series though, so I'm more inclined towards providing comments
> > and other ways of simplifying this subroutine.
>
> Agreed that refactoring `tree_search()` probably is out of scope here.
> But rewriting the test is definitely something we can do.
>
> Perhaps:
>
> static void t_tree(void)
> {
>         struct tree_node *root = NULL;
>         int values[11] = {7, 5, 2, 3, 10, 4, 6, 9, 8, 1};
>         struct tree_node *nodes[11] = { 0 };
>         size_t i = 1;
>         struct curry c = { 0 };
>
>     // Insert values to the tree by passing '1' as the last argument.
>     for (i = 1; i < ARRAY_SIZE(values); i++) {
>                 nodes[i] = tree_search(&values[i], &root, &t_compare, 1);
>     }
>
>         for (i = 1; i < ARRAY_SIZE(nodes); i++) {
>                 check_pointer_eq(values[i], nodes[i]->key);
>                 check_pointer_eq(nodes[i], tree_search(values + i, &root, &t_compare, 0));
>         }
>
>         infix_walk(root, check_increasing, &c);
>         tree_free(root);
> }
>
> Wouldn't this have the same effect while making it much easier to read?

I agree that the change 'values + i -> &values[i]' is a net positive, I had this
change in mind as well. This comment on the other hand,
>     // Insert values to the tree by passing '1' as the last argument.
has already been stated in the commit message of the 3rd patch
as was suggested by Patrick earlier:

'Note that the last parameter in the tree_search() function is
'int insert' which when set, inserts the key if it is not found
in the tree. Otherwise, the function returns NULL for such cases.'

So I was thinking of adding something along the lines of:
'// pseudo-randomly insert pointers to elements between values[1]
and values[10] in the tree'