• Teodor Dutu (18/18) Nov 27 2023 Hi,

Teodor Dutu <teodor.dutu gmail.com> writes:

Hi,

Last week I started to convert 
[`_d_arrayliteralTX`](https://github.com/dlang/dmd/blob/41f3577d7d8f63a7221a9ff2359fe4e5163643cd/druntime/src/rt/lif
time.d#L2099-L2140) to a template. The hook is somewhat similar to
[`_d_newarrayT`](https://github.com/dlang/dmd/pull/15299) in that it's used to
allocate array literals like so `[1, 2, 3]`. Note that unlike `_d_newarrayT`,
which also initialises the newly created array, `_d_arrayliteralTX` just
allocates memory for it. The reason for this choice is to limit the number of
arguments passed to the hook. If it were to also initialise the array, it would
have to receive its contents as an extra variadic argument. I am going to stick
to this decision because now with the template hook, an extra variadic template
argument would mean a lot more instances for something that is already handled
just as fine and with less code by the compiler.

One annoying fact about array literals is that they suffer lots 
of optimisations after semantic analysis (so after the 
introduction of the hook). In short, they are either eliminated 
altogether, or coalesced with concatenations into larger 
literals. For example, `[1, 2, 3] ~ 4` is optimised by the 
compiler to `[1, 2, 3, 4]`. This becomes a nuisance because the 
hook receives the length of the array as an argument and this 
length changes in the previous example. Since the lowering is 
introduced during [semantic 
analysis](https://github.com/dlang/dmd/blob/41f3577d7d8f63a7221a9ff2359fe4e5163643cd/compiler/src/dmd/ex
ressionsem.d#L4243) and this optimisation (which is not the only one that may
happen) takes place during [optimize.d, so after semantic
analysis](https://github.com/dlang/dmd/blob/41f3577d7d8f63a7221a9ff2359fe4e5163643cd/compiler/src/dm
/optimize.d#L1288), I'd have to make changes to `ArrayLiteralExp`s in multiple
places, which does not scale and will introduce bugs if another change to
`ArrayLiteralExp` is introduced without taking the hook into account.

A cleaner solution which I'm going to try this week is to inspect 
the call to the hook once more during [IR 
generation](https://github.com/dlang/dmd/blob/41f3577d7d8f63a7221a9ff2359fe4e5163643cd/compiler/s
c/dmd/e2ir.d#L3957) and update the array length argument according to the new
size of the array.

Thanks,
Teodor