win64: speed up test suite by more than 15 minutes

[rainers]

runnable/test17338.d added in #6718 triggers pretty bad identical comdat folding performance by the MS linker.

As this is known to have problems anyway (e.g. https://issues.dlang.org/show_bug.cgi?id=10664) and is disabled in sc.ini (https://github.com/dlang/dmd/blob/master/ini/windows/bin/sc.ini#L23), we can disable it in the tests, too.

[rainers]

The build time for Win_32_64 on win-farm1 has gone down from 43 minutes for other PRs to less than 26 minutes.

[WalterBright]

A comment in the code explaining why this switch is there would be helpful, as I can't see someone guessing its purpose, as the documentation for NOICF doesn't offer a clue.

[rainers]

A comment in the code explaining why this switch is there would be helpful

Amended a comment.

[MartinNowak]

Quite a huge speedup.

Does this only come from runnable/test17338.d?

Of course we should fix Issue 10664 – Win64: exception handling does not work with COMDAT folding at some point, and maybe even explicity use /OPT:REF,ICF for release builds as recommended -OPT (Optimizations).

Using the equivalent for non-Windows linkers would be nice as well.
c++ - GCC(/Clang): Merging functions with identical instructions (COMDAT folding)

[MartinNowak]

Opened an issue for the ld.gold failure Issue 17515 – ld.gold identical comdat folding triggers an assertion.

[rainers]

Does this only come from runnable/test17338.d?

At least most of it. It took more than 15 minutes to link it on my computer.

Opened an issue for the ld.gold failure Issue 17515 – ld.gold identical comdat folding triggers an assertion.

Interesting. Hopefully that helps narrowing down the problem for windows, too.