Hum, silly me - That can't work because of bootstrapping. @MartinNowak : How are you compiling the releases ? Could we do this in 2 steps, by first introducing this switch, and enabling it later on ?

I don't think this adds much of any value to -release. Note that -release has the following effects:

if (global.params.release)
{
    global.params.useInvariants = false;
    global.params.useIn = false;
    global.params.useOut = false;
    global.params.useAssert = false;
    global.params.useSwitchError = false;
}
if (global.params.useUnitTests)
    global.params.useAssert = true;

It's also becoming clear that turning asserts on and off with the command line is a bit blunt, especially when projects tend to be all files on one command line. Some asserts are more equal than others. Individual asserts can be left on with:

if (!condition) assert(0);

but writing code that way is awkward and just doesn't look good.

I propose instead to do something analogous to the accepted pragma(inline, true/false), such as:

pragma(assert); // default
pragma(assert, true); // turn asserts on
pragma(assert, false); // turn asserts off

Can you write a DIP for it?

I don't think this adds much of any value to -release.

Yeah, we could allow more fine-grained control over what is enabled / disabled, but in any case, it won't solve the problem it's trying to solve, since we need to compile with the oldest C++ DMD.
Instead of changing D behavior, I'm considering just replacing all assert calls with throw new IceError()... Would that work, or do you have other suggestions ?

Would that work

I don't see how it would be any different.

-release enables several checks, as I listed above. Probably a decent step is to try turning them on/off individually, and see which one contributes most to the 8% slowdown.

I don't see how it would be any different.

Well the problem it would introduce is that we won't be able to use this new feature directly for a while.
And as we keep compatibility with the last C++ DMD, it only makes matter worse.
Using an Error type would solve that issue.

Regarding pragma(assert), I'm not sure I understand your position completely.
It would allow one function (/ module / aggregate) to force assert to stay, right ? The point of this switch is to unconditionally do that for all module, rather than allowing fine-grained control (which is already possible by other means, such as the assert(0) trick you described, or a manual throw Error).

The issue itself (15962) needs another fix.
The functionality (more control over assert than what debug / release provide) will be covered by dlang/DIPs#54