Hi,

My project wants to replace dmd hooks to druntime functions such as _d_arrayctor, _d_arrayappendT and _d_arraycatT with templates. This will make lowerings to these functions faster at runtime, as their templated versions will not require all the indirection used by the hooks. Speaking of indirection, at the end of this project, the TypeInfo class will be removed as all its information will be used at compile time, when instantiating the aforementioned templates. As a result, the maintainability of the compiler will be improved as lowerings will now take place in the semantic analysis phase, as opposed to the IR phase (where the hooks are currently present).

Dan Printzell worked on this in the past, as part of his GSoC 2019 project. I'll be picking up the work that he left and then I'll continue with the remaining hooks.

During my first week, I worked on finishing Dan's PR for the _d_array{,set}ctor functions:

• My mentors (Eduard Stăniloiu and Răzvan Nițu) and I fixed some of Dan's logic for choosing when to perform the lowering or not. We skipped lowering in expressionsem.d for snippets like S[3] arr = [S(1), S(2), S(3)];, as e2ir.d already handles such cases more elegantly.
• The tests that are currently failing are runnable/testassign.d and compilable/interpret3.d. For example, runnable/testassign.d fails because in expressions like funcVal(a = b) the temporary variable used as the actual argument for funcVal is not lowered to use _d_arraysetctor, when it actually should. I am still working on fixing it.

Thanks,
Teodor

On Thursday, 23 September 2021 at 15:15:32 UTC, Teodor Dutu wrote:

This is bad news for all non-DMD compilers, as e2ir.d is DMD only, and LDC and GDC would have to duplicate thew lowerings in the glue layer now. Please don't.

On Thursday, 23 September 2021 at 16:06:57 UTC, kinke wrote:

Just to clarify, there is no lowering happening in e2ir. What happens right now (before replacing the hook with template) is that in some situations the hook is bypassed (i.e. no call is made to druntime). We mimicked this logic in our implementation.

To clarify:

S[1] a = [S(1)];

would be lowered to:

_d_arrayCtor(a, [S(1)]);

which in turn gets rewritten to:

S[1] __arraystack_literal = [S(1)];
_d_arratCtor(a, __arraystack_literal);

Now, the first line, if we are not careful, will be lowered
to _d_arrayCtor hence the infinite loop.

Looking at the current state of e2ir [1], we saw that in the case
where rhs is an array literal the lowering is not performed
anymore. Instead, the literal is assigned element-wise.

What Teo meant is that in this situation we are not lowering to
_d_arrayctor, rather we just let the compiler handle the case as
it did up until now.

[1] https://github.com/dlang/dmd/blob/master/src/dmd/e2ir.d#L2519

On Thursday, 23 September 2021 at 15:15:32 UTC, Teodor Dutu wrote:

With what it will be replaced then?

Will I be able to do this:

int[] id;

typeid(id).writeln;

The typeinfo itself would be best to remain imho. Though functionality that is only related to those hooks should be removed.

Best regards,
Alexandru.

On Thursday, 23 September 2021 at 16:30:57 UTC, RazvanN wrote:

Okay that sounds better, thx for clarifying.

On Thursday, 23 September 2021 at 17:04:51 UTC, Alexandru Ermicioi wrote:

int[] id;

typeid(id).writeln;

Yes, you're right. What I meant to say is that TypeInfo will only be removed from the hooks' ecosystem (together with the hooks themselves) and in its entirety.

Teodor

On Thursday, 23 September 2021 at 18:28:31 UTC, Teodor Dutu wrote:

int[] id;

typeid(id).writeln;

... and not in its entirety.

Teodor

On Thursday, 23 September 2021 at 18:38:47 UTC, Teodor Dutu wrote:

Awesome, keep up doing the good work!

Regards,
Alexandru.