Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
[WIP][interp] Unify execution and valuetype stacks (#43681)
Before this change, an InterpFrame contained 3 regions of data : args + locals, valuetype stack, execution stack. Each entry on the execution stack was a stackval structure. The space for valuetypes, was allocated separately, since they have various sizes. When pushing a valuetype on the stack, we allocated first the space for it on the vtstack and then pushed the address of the region on the execution stack. This change merges the execution stack with the valuetype stack, meaning we push now variable sized data on the execution stack. In order to keep track of the current stack location, whenever we push a type on stack, during transform phase, we also keep track of the offset where this value will reside on stack, as well as the size it occupies. All callsites need to be informed how much they need to pop the stack for the arguments. While called code can access this space normally (the args are special locals belonging to the frame and are accessed directly as such), external code needs a new mechanism to detect each argument for a given frame. This is achieved with the lazily initialized arg_offsets array on an InterpMethod. The method doesn't need to be compiled for this array to be correctly initialized. Why : - this simplifies handling of valuetypes, their storage follows the same rules as a normal objref/primitive type - removes the common use of the vt_sp variable. The compiler no longer needs to reserve it in a register during the switch loop, we no longer need to save it with each call. The sp and ip become now the only variables describing the execution state in a method. - the flow of the data on the execution stack is well behaved now (with the exception of a few opcodes that update directly based on the stack offset). We were using the vtstack for some magic storage (MINT_VTRESULT for example) - this makes it such that the stack offset of every value is easily known at compile time, making it possible to completely drop the execution stack approach, and have every opcode have a unique dreg and a list of sregs (which are mapped to a certain stack offset). This will enable more advanced optimizations during compile stage. Co-authored-by: BrzVlad <[email protected]>
- Loading branch information