The ContextAnalysis visitor for braced blocks was not updated during void support implementation which is why I do it now. Although we are re-using the functionCall production rule a different visitor is required for visiting function calls directly inside braced blocks. This is because the other place where functionCalls are used the functionCall is annotated with a label with wraps the functionCall context inside a label context. Fortunately this is in fact simple wrapping so we just implement a visitor for the wrapping class that unwraps the functionCall and passes it to the visitor that actually implements the functionCall.
This was a tricky one. In order to call malloc the stack needs to be 16 byte aligned. GenASM up until this point was 8 byte aligned. This commit changes the code generation so that it is guaranteed that the stack is 16 byte aligned before visiting a child node. We need to do this before visiting *any* child node because we can not know whether somewhere downstream some child node calls malloc.
To make it possible for a visitor method to determine if it needs to align the stack it has to assume that the stack is currently aligned. This is the other reason we ensure that the stack is aligned before visiting any child node.
On top of that some visitor methods did not clean up after themselves. I am not sure why this did not result in segfaults already, but I changed the code so that each visitor method leaves the stack in the same state after completion.
This enables us to add all local variable Definitions to the DAST so that downstream visitors don't need to compute the local variables of a function again.
The GenASM visitor only uses quad mnemonics so each manipulation to the stack is done in 8 byte chunks. This means our stack is always 8 byte aligned which is fine when only calling KLang code. libc functions like 'malloc' or 'free' require the stack to be 16 byte aligned which isn't automatically guaranteed when being 8 byte aligned. This is why we need to be careful when manipulating the stack (e.g. pushq, popq, subq x, %rsp). By assuming that we are stack aligned at the beginning of execution and by tracking the number of pushq operations we can deduce whether we misaligned the stack and align it again by adding another 8 byte to the stack.
This re-alignment of the stack is done directly before entering a function body or calling malloc. Since there are no stack manipulations involved when generating the code to calling 'free' there is reason to check for alignment there.
We generate our own main function that executes the user's specified expression at the end of his file. This auto generated function has to be called "main" in order for it to be executed on startup. If the user chooses to call one of his own functions "main" our auto generated function would collide with the user's function. That's why we quietly rewrite the user's function to "main_by_user". This way we prevent a collision.
Note for the future: We should change our Syntax to allow for no expression as the last definition of a file. This way the user can choose if a particular source file needs to contain a main function or not (just like c does it). This is also one of the requirements for modules to work.
This allows us to store the index of the enum value along the name. The index can be used to compare two enum values in assembler.
Later on this might be used to enable users of KLang to set arbitrary values as the index of an enum value.
nitrix