Fix issue 4520 due to storage model mismatch

[sklam]

as titled.

[stuartarchibald]

Looks good, thanks for the patch.

[esc]

@sklam this is dark. Can you give us a --verbose on why this fixes the issue?

[sklam]

@esc, yes, i'll try to explain this.

Background

Numba has this concept called datamodel (which I decided to call in storage model in the PR title because it is probably more descriptive). The datamodels describe how a type is represented in different situations:

• value representation is for in-registers data or temporary in-stack data that does not need to be addressable.
• data representation is for heap data that needs to be addressable.
• argument and return representations need to be ABI compatible for the corresponding function boundary.

Most types have the same value and data representation but a bool is stored as a single bit in the value representation and a single byte in the data representation. The reason being that LLVM uses a i1 (1-bit integer) as the predicate for the branch instruction br, but a i1 is not addressable. The smaller addressable unit is a byte i8.

Related code: https://github.com/numba/numba/blob/master/numba/datamodel/models.py

The bug

This patch fixed the incref/decref wrapper function for the containers which is taking a pointer to the container elements using the data representation, but the underlying incref/decref operations are expecting the value representation. Elements in containers must be stored in the data representation because it needs to be addressable. To convert the data representation to value representation, the patch used the datamodel.load_from_data_pointer() instead of the builder.load().