Hi,

Last week I fixed this bug introduced by _d_arrayappendcTX. The fix meant first storing the lowering to a new field inside CatAssignExp instead of modifying the original array.

This in turn caused vthiss for some functions because of the inliner. The solution was to inline the lowering when present.

After this, I ran into a familiar error: the lowering needs no codegen at CTFE. In fact, the lowering shouldn't even be introduced during semantic if the CatAssignExp is only needed at CTFE. However, this problem is familiar because I ran into it in the past. The same restrictions hold for CatExps, but because telling if an expression needs codegen or not is a mess during semantic, I couldn't figure out when not to perform the lowering to _d_arraycatnTX. The solution in both cases was to handle this case during IR generation. Therefore, unfortunately, both lowerings are made even in some situations when they're not required and then e2ir.d checks of using the GC is enabled. If not, it outputs an error. I'd love to do this more nicely in the future and I'm looking for advice on how to figure out when to not generate lowerings during semantic analysis. If you have any advice on the topic, please let me know.

Furthermore, I'm almost done converting _d_newarray{mTX, miTX, Op} to a single template. I created this draft PR with my changes. I still need to fix some final tests in dmd and Phobos before turning it into a legit PR. I need to say that the previous implementations of _d_newarraym{,i}TX were really messy because they called _d_newarrayOp with a funcion argument with which to initialise the arrays (one of _d_newarray{,i}T). But now that I've converted these 2 to a single template as well, I was able to coalesce all hooks for multi-dimensional arrays into a single one much more smoothly.

Thanks,
Teodor