Improve compilation speed by taking advantage of memoization

[jonathanpoelen]

I put in bulk the benchmarks and tests that can be found in each commit (issue #77)

(result of /usr/bin/time --format='%Es - %MK' $compiler ...)

mp_transform_if

compiler	gcc-12	clang-15
before	0:00.13s - 60432K	0:00.16s - 100956K
after	0:00.11s - 47300K	0:00.16s - 100484K

template<class T> using p = mp_bool<T::value & 1>;
template<class T> using f = mp_size_t<T::value + 1>;
template<class I> using test = mp_transform_if<p, f, mp_iota<I>>;

using r1 = mp_transform<test, mp_iota_c<50>>;

mp_filter

compiler	gcc-12	clang-15
before	0:00.16s - 72460K	0:00.16s - 102068K
after	0:00.14s - 61884K	0:00.16s - 101828K

template<class T> using p = mp_bool<T::value & 1>;
template<class I> using test = mp_filter<p, mp_iota<I>>;

using r1 = mp_transform<test, mp_iota_c<50>>;

mp_drop

compiler	gcc-12	clang-15
before	0:00.48s - 192464K	0:00.48s - 154480K
after	0:00.43s - 172472K	0:00.46s - 150404K

template<class L> struct f { template<class I> using g = mp_drop<L, I>; };
template<class I, class L = mp_iota<I>> using test = mp_transform<f<L>::template g, L>;

using r1 = mp_transform<test, mp_iota_c<50>>;

mp_sort

compiler	gcc-12	clang-15
before	0:00.46s - 198848K	0:00.49s - 132632K
after	0:00.42s - 183384K	0:00.48s - 131748K

template<class I> using test = mp_sort<mp_iota<I>, mp_less>;

using r1 = mp_transform<test, mp_iota_c<25>>;

mp_replace_at

compiler	gcc-12	clang-15
before	0:00.34s - 134428K	0:00.33s - 124720K
after	0:00.29s - 116920K	0:00.32s - 122568K

template<class I> using f = mp_replace_at<mp_iota<I>, I, void>;
template<class I> using test = mp_transform<f, mp_iota<I>>;

using r1 = mp_transform<test, mp_iota_c<50>>;

mp_power_set

compiler	gcc-12	clang-15
before	0:01.35s - 538732K	0:00.65s - 203924K
after	0:01.02s - 396032K	0:00.65s - 203080K

template<class I> using test = mp_power_set<mp_iota<I>>;

using r1 = mp_transform<test, mp_iota_c<15>>;

mp_map_find

compiler	gcc-12	clang-15
before	0:00.29s - 128600K	0:00.25s - 120604K
after	0:00.22s - 99908K	0:00.23s - 116084K

template<class L, class M = mp_transform<mp_list, L>>
struct f { template<class I> using g = mp_map_find<M, I>; };

template<class I, class L = mp_iota<I>> using test
  = mp_transform<f<L>::template g, L>;

using r1 = mp_transform<test, mp_iota_c<50>>;

mp_map_update

compiler	gcc-12	clang-15
before	0:00.73s - 270224K	0:00.54s - 156112K
after	0:00.19s - 85004K	0:00.29s - 119232K

template<class L, class M = mp_transform<mp_list, L>>
struct f { template<class I> using g = mp_map_update<M, mp_list<I>, mp_list>; };

template<class I, class L = mp_iota<I>> using test
  = mp_transform<f<L>::template g, L>;

using r1 = mp_transform<test, mp_iota_c<20>>;

mp_map_erase

compiler	gcc-12	clang-15
before	0:00.52s - 197128K	0:00.44s - 140780K
after	0:00.47s - 183968K	0:00.43s - 139336K

template<class L, class M = mp_transform<mp_list, L>>
struct f { template<class I> using g = mp_map_erase<M, I>; };

template<class I, class L = mp_iota<I>> using test
  = mp_transform<f<L>::template g, L>;

using r1 = mp_transform<test, mp_iota_c<30>>;

mp_map_replace

compiler	gcc-12	clang-15
before	0:00.46s - 187160K	0:00.38s - 129104K
after	0:00.21s - 80236K	0:00.33s - 123644K

template<class L, class M = mp_transform<mp_list, L>>
struct f { template<class I> using g = mp_map_replace<M, mp_list<I, I>>; };

template<class I, class L = mp_iota<I>> using test
  = mp_transform<f<L>::template g, L>;

using r1 = mp_transform<test, mp_iota_c<30>>;

[jonathanpoelen]

@pdimov would you be able to take a look, please?

[pdimov]

If you use fn as the name of the inner alias, instead of _f et al, these helpers will become quoted metafunctions and you'll be able to use the _q variants of the algorithms instead of spelling the ::template _f.

[pdimov]

This looks good to me. I can only wonder why I failed to look at it two years ago when you submitted it originally; it probably came in an inopportune time when I was too busy.

I see a couple more things. It looks like you started to also optimize mp_map_replace_impl by moving the msvc workaround into an ifdef, but then didn't extract _f into a helper template.

Also, in the two cases where the _impl class is reduced to just having a single using type = ... (mp_transform_if_impl and mp_map_erase_impl), you can just remove the _impl class altogether and move the ... part into the primary alias.

The 1.90 beta cutoff is the day after tomorrow, so I won't be able to merge this now, but to make sure I don't drop the ball again (sorry), please remind me after 1.90 is released.

And thank you for this work.

[jonathanpoelen]

I see a couple more things. It looks like you started to also optimize mp_map_replace_impl by moving the msvc workaround into an ifdef, but then didn't extract _f into a helper template.

I went back to msvc because it wasn't working: 87719f6. However, after two years, the logs are no longer available.

Also, in the two cases where the _impl class is reduced to just having a single using type = ... (mp_transform_if_impl and mp_map_erase_impl), you can just remove the _impl class altogether and move the ... part into the primary alias.

Done.

The 1.90 beta cutoff is the day after tomorrow, so I won't be able to merge this now, but to make sure I don't drop the ball again (sorry), please remind me after 1.90 is released.

I will try to remember :)