moulins

The bug is reproducible on all 4 backends. For completeness, I also tested the `read` and `bufread` encoders, and their `total_in` method correctly returns the size including the header.

This means there's an asymmetry between `write::ZlibEncoder` and `read::ZlibEncoder` though: the `write` encoder doesn't include the header, but the `read` encoder does! Personally, I would prefer this to be fixed,...

**Note:** I just found out that my explanation was slightly wrong, as the ZLIB header is only 2 bytes, not 6. The 6 extra bytes actually correspond to: - The...

> Right now the value returned by banana violates the validity invariants for the reference types and I don't see how the alignment niche bits could be masked out anywhere,...

> one possible way to spell `Aligned` is `&'unsafe T` That could actually be pretty elegant; it would remove the need for defining the new niches *via* custom attributes.

> I remember people some time ago floating the idea of adding yet more pointer niches by recognising the fact that under some ABIs, certain pointer ranges are reserved and...

> Back on the topic of `ptr::WellFormed`... the problem is that the invariants of `ptr::WellFormed` are basically that the pointer points to a place that was valid to dereference previously,...

> But more importantly, we had concluded back then that transmuting from Foo to Foo (given T and U both Sized) should always succeed, and thus we can't rely on...

To solve the circular queries issue without having to explicitly detect and break cycles, I propose separating the layout computation in three phases: - **First phase** - Compute the (exact)...

I have stumbled onto an other use case for this feature, where `ManuallyDrop` isn't enough: ```rs pub struct Unescaper { pub fn new(buf: &'a mut String) -> Self { Self...