std.algorithm.searching: no mapping-specialization for extremum

[wilzbach]

In #4265 I showed that extremum (aka minElement and maxElement) can be made a lot faster if we provide a special path for the identity case, because the compiler can then make use of SSE instructions.
In theory the compiler should be smart enough to figure it out himself, but well neither dmd nor ldc are that smart :/

For reference here are the benchmarks from #4265:

dmd -release -O foo.d && ./foo
0 4 secs, 610 ms, 609 μs, and 5 hnsecs
1 3 secs, 769 ms, 89 μs, and 4 hnsecs

ldc -release -O3 foo.d && ./foo
0 1 sec, 345 ms, and 914 μs
1 216 ms, 347 μs, and 1 hnsec

Source for benchmarks: http://sprunge.us/QSWP

This PR does the specialization via overloads whereas #4265 does them via comparison of strings.

[wilzbach]

The problem here is that with string mixins a[0] can be a valid unary and binary function:

 pragma(msg, is(typeof(unaryFun!`a[0]`([0])))); // true
 pragma(msg, is(typeof(binaryFun!`a[0]`([0], [0])))); // true

[andralex]

this appears uncovered, why?

[wilzbach]

Because most tests were using orderd arrays and the default predicate. It's easy to cover though:

58b1c0b

[wilzbach]

In #4265 I showed that extremum (aka minElement and maxElement) can be made a lot faster if we provide a special path for the identity case, because the compiler can then make use of SSE instructions.

I added this to the changelog.
FYI: it's now possible to preview the changelog, e.g.:

http://dtest.dlang.io/artifact/website-90f9ffa1188c64a9b85c20374f22679f2d3b1a6a-3ccd57ed79e89ad113b428236ca14115/web/changelog/2.074.0.html

(there seems to be some whitespace issues in ddoc, but that's the fault of the changelog generation script)

[wilzbach]

@andralex andralex added the auto-merge-squash label 10 minutes ago
@dlang-bot dlang-bot removed the auto-merge-squash label 4 minutes ago

FYI: This is a new security feature to protect from unapproved changes sneaking in.

[andralex]

@wilzbach cool! so what did I do wrong?

[wilzbach]

@wilzbach cool! so what did I do wrong?

Nothing - I added the changelog entry.

[andralex]

Ah, got it. So now I'll re-review and reapply.

[wilzbach]

Ah, got it. So now I'll re-review and reapply.

Yep, that was the idea behind this feature. The auto-tester allows such "silent changes".

[andralex]

"a lot" -> "considerably"

better yet:

"considerably (almost 2x in microbenchmarks)" (or whatever)

i.e. we're trying to avoid colloquialisms

[andralex]

thx!!

[wilzbach]

[SQUASH] Add shell % prefix to the DMD test

Simple "one char" change in the changelog. Sorry

(there seems to be some whitespace issues in ddoc, but that's the fault of the changelog generation script)

PR: dlang/tools#223

[JackStouffer]

Removing the check for length here will allow an infinite random access range to use this branch

[wilzbach]

!isInfinite!Range is an excluding constraint above as extremum will be an endless loop on infinite ranges.

[JackStouffer]

needs a check for length here too

[wilzbach]

@JackStouffer I had another look at this PR and discovered that there were still some uncovered bits due to the combinatorial explosion of the alias (aka non-string comparison) approach.
Ideally CodeCov won't complain now :)