Missing stacktrace on memory allocation failure

[Mathias LANG]

On Wednesday, 2 June 2021 at 17:52:12 UTC, mw wrote:
> On Thursday, 12 November 2020 at 13:16:21 UTC, Steven 
> Schveighoffer wrote:
>> On 11/12/20 4:22 AM, Per Nordlöw wrote:
>>> Why don't I get a stack trace on Memory allocation exceptions?
>>> 
>>> In my case I only get:
>>> 
>>> src/core/exception.d(647): [unittest] Memory allocation failed
>>> core.exception.OutOfMemoryError at src/core/exception.d(647): 
>>> Memory allocation failed
>>> 
>>
>> Certain errors are flagged as not able to allocate memory. So 
>> you don't get a stack trace.
>>
>> I think it's a huge mistake, and we don't need memory 
>> allocation to do stack trace printing. But that's the way it 
>> is.
>
> Hit this one today, since it's a mistake, what does it take to 
> fix it? or you are saying it's unfix-able in the current D 
> runtime?

It is fixable, but it'll take a lot of work. The main blocker is 
that the interface that is exposed to the user relies on memory 
allocation.
Namely, the `TraceInfo` interface is: 
https://github.com/dlang/druntime/blob/2d8b28da39e8bc3bc3172c69bb96c35d77f40d2a/src/object.d#L2257-L2261

Note that:
1) `toString` returns a `string` (allocation);
2) The `opApply`'s delegate use a `char[]` as parameter;

What we *should* have is an `opApply` on a struct that contains 
all necessary info, which itself writes to a sink / OutputRange.
Such struct exists 
(https://github.com/dlang/druntime/blob/a17bb23b418405e1ce8e4a317651039758013f39/src/core/internal/backtrace/dwarf.d#L77-L160) but need to be exposed to user code.
Then *printing* the stack trace will no longer allocate. However, 
`defaultTraceHandler` still returns an object, so acquiring the 
handler will also need to be addressed. There have been talks of 
deprecating chained Exceptions, and that would make such a task 
much simpler.