The cleanups actually improve the performance from 80,458,208,346 to 80,430,759,717 in the proposed step from #4171.

And simply using SmallVector without Boost reduces it from 80,430,759,717 to 80,152,494,488. Some changes necessary to make it compile (it doesn't like the initializer lists anymore) make it no longer use ReferenceToken* std::__uninitialized_copy<>::__uninit_copy<>(ReferenceToken const*, ReferenceToken const*, ReferenceToken*) (cppcheck: stl_uninitialized.h, ...).

Using Boost it reduces from 80,152,494,488 to 79,232,522,651.

Better numbers for both compilers soon.

This is finally ready for review. 🥳

current -> SmallVector -> SmallVector + Boost 1.74

Performing a --enable=all --inconclusive scan on mame_regtest.

With DISABLE_VALUEFLOW=1:

Clang 14 2,026,151,486 -> 2,020,278,735 -> 1,999,090,088
GCC 12 1,994,489,219 -> 1,985,524,671 -> 1,964,385,704

With Valueflow enabled:

Clang 14 51,130,767,151 -> 51,083,225,640 -> 50,940,612,547
GCC 12 54,989,579,042 -> 54,923,205,508 -> 54,760,703,625

And scanning cli with --enable=all --inconclusive -Ilib -D __GNUC__ and DISABLE_VALUEFLOW=1

Clang 14 2,935,467,131 -> 2,922,870,135 -> 2,889,963,287
GCC 12 3,009,733,699 -> 2,995,173,269 -> 2,961,009,395

There's still a lot of Ir being used by the SmallVector. It is hard to tell from which code it is actually coming from though.

There's still a lot of Ir being used by the SmallVector. It is hard to tell from which code it is actually coming from though.

This is addressed in #4526.