While building the main function for tests, we may need to generate
additional blocks while marshalling arugments to the error catcher into
the expected calling convention. This pushes the last block in the main
function down, so that the "entry" block may not be last BB in the
function. Instead, look up the last insertion block before generating
the catcher, and then add a call to the catcher at the end of this last
block.
Closes#2300
Previously, we assumed that a union layout always lived on >= 1 64-bit
boundary when generating an LLVM type for it. For small tags unions,
like `[ Ok i8, Err ]` this need not be the case; indeed, a tag union
like that is actually only 2 bits - 1 bit for the "i8" data, and one bit
of the tag kind.
This led to a discrepancy between what the layout IR and generated LLVM
code would assume about the size of tag unions. In the case above, the
layout IR would assume the tag data is 2 bits wide, and the tag id is 1
bit into the data. But the LLVM code would generate a type that was 65
bits wide, the first 64 bits being for the "i8" data and the last 1 bit
being for the tag kind.
Usually, just running the LLVM-emitted code would not present a problem.
But it does present a problem when we use the layout IR to inspect the
result of LLVM-run code, in particular when we try to look up the tag
ID, as the repl does. This patch fixes that issue.
Note that this bug did not present itself in `test_gen` previously
because the data that most tests check against is stored in the front of
the representation.
Closes#2149
This will simplify parsing and make it possible to have a uniform lexer for the language. Previously unquoted package names were allowed to include '-'s, which aren't valid identifiers.
In the future, we'll distinguish local paths from packages in the package-manager by looking for a ".roc" suffix, which should only be present in local paths.
