On Thursday, 5 February 2015 at 03:00:53 UTC, Andrei Alexandrescu
wrote:
On 2/2/15 2:42 PM, "Ulrich =?UTF-8?B?S8O8dHRsZXIi?=
<kuett...@gmail.com>" wrote:
On Friday, 30 January 2015 at 23:17:09 UTC, Andrei
Alexandrescu wrote:
Sorry, I thought that was in the bag. Keep current semantics,
call it
chunkBy. Add the key to each group when the predicate is
unary. Make
sure aggregate() works nice with chunkBy().
I might miss some information on this, so please forgive my
naive
question. Your requirements seem to be contradictory to me.
1. aggregate expects a range of ranges
Probably we need to change that because aggregate should
integrate seamlessly with chunkBy.
2. you ask chunkBy to return something that is not a range of
ranges
Yah.
3. you ask chunkBy to play along nicely with aggregate
Yah.
There are certainly ways to make this work. Adding a special
version of
aggregate comes to mind. However, I fail to see the rational
behind this.
Rationale as discussed is that the key value for each group is
useful information. Returning a range of ranges would waste
that information forcing e.g. its recomputation.
I understand and agree. My suggestion aims to avoid this
particular waste. See below.
To me the beauty of range is the composibility of "simple"
constructs to
create complex behavior. The current chunkBy does not need to
be changed
to "add the key to each group when the predicate is unary":
r.map!(pred, "a")
.chunkBy!("a[0]")
.map!(inner => tuple(inner.front[0], inner.map!"a[1]"));
So I'd like to know why the above is inferior to a rework of
the
chunkBy's implementation. Maybe this is a question for D.learn.
Wouldn't that force recomputation if a more complex expression
replaced a[0]?
I do not think you ever want to replace a[0] here. In the code
above the (original) predicate to chunkBy is pred. The idea is to
evaluate the predicate outside of chunkBy. Create a range of
tuples from the original range, chunk the range of tuples and
construct the desired result from the chunked range of tuples.
// create a range of `tuple(pred(a), a)`
r.map!(pred, "a")
// chunk the range of tuples based of the first tuple element
// this results in a range of ranges of tuples
.chunkBy!("a[0]")
// convert the inner ranges of tuples to a tuple of the predicate
applied and the appropriate range
.map!(inner => tuple(inner.front[0], inner.map!"a[1]"));
The construction of a range of tuples is not for free. On the
bright side:
* you only do it when you need it
* if your predicate is that heavy, you might want to precompute
it anyway
* a modified chunkBy is not exactly free either (and you pay the
price even if you do not need the key value)
Now I learned that map is very lazy and applies the function
inside front(). Thus, the above might actually result in multiple
evaluations of the predicate. Luckily, there is the new cache
function:
auto chunkByStar(alias pred, Range)(Range r)
{
return r.map!(pred, "a")
.cache
.chunkBy!("a[0]")
.map!(inner => tuple(inner.front[0], inner.map!"a[1]"));
}
My point here is, we can construct a version of chunkBy that does
not waste the key value with modest means. With great power comes
great flexibility. I wanted to sneak this in as an example,
because it is not clear what eventual users might actually need.
On the other hand there is no limit to the special cases we could
add. aggregate might not be the only function to work with
chunkBy. And even an aggregate function that takes a tuple of a
range and something else and only uses the range seems wrong to
me, given expressive the power D has. The transformation of the
range is just on map away:
chunkByStar!(...)(r).map!"a[1]".aggregate!max
Then again, I might be missing something huge here.
