https://issues.dlang.org/show_bug.cgi?id=12606
Andrej Mitrovic <[email protected]> changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |[email protected] --- Comment #2 from Andrej Mitrovic <[email protected]> --- Kenji: The way I see it, we need some way of returning back information on why some implicit conversion didn't work. For example, in 'implicitConvTo' in the visitor functions 'result' is simply initialized to MATCHnomatch if the dimensions of the two slices don't match. But this isn't propagate back to the caller side. We could vastly improve diagnostics for type mismatches, length mismatches, template instance failures, if we had a way of return back information. I think one way we could do it is to turn the MATCH enum into a struct, which internally holds more state. For example: struct MATCH { /// equivalent to what MATCH was before: MATCHnomatch, MATCHconvert... MATCHLevel level; /// call this to emit informative diagnostic const char*(*error_fp)(); } In internal matching routines the code could do (sorry this is a bit mixed-C/D code syntax): ----- if (e->elements->dim != tsa->dim->toInteger()) { result = MATCH(MATCHnomatch, { return format("Array length mismatch: %s %s", e->elements->dim, tsa->dim->toInteger()); } } ----- So the idea is the routine initializes a pointer, and then the call site can decide whether to call it to extract a diagnostic, or whether to ignore it. For example the 'resolveFuncCall' function typically walks through several template instances before it decides on emitting diagnostics, so having a lazy way of extracting a diagnostic is cruical. On the call site (e.g. 'resolveFuncCall') the code could look like: ----- if (matchresult.level == MATCHnomatch) { error("Failed to match: %s", matchresult.error_fp()); } ----- What do you think about this design? I'm not sure whether we can create delegates in C++ like we can in D though, I'm not too familiar with what's supported there. --
