The fact that you parse an entire subtree before you know the kind of its parent is exactly what makes it bottom-up. Pratt is certainly easy to integrate into a

The fact that you parse an entire subtree before you know the kind of its parent is exactly what makes it bottom-up. Pratt is certainly easy to integrate into an otherwise-top-down parser, because you can enter it from the top and it can enter its "leaves" from the top, but internally it is absolutely bottom-up, and you can generalize this to any recursive ascent parser. To make this more explicit: The first thing a Pratt parser does before entering the loop is based on the lookahead token's "nud." This corresponds to an LR automaton in its initial state shifting that token and advancing into the leaf-most rule(s). The "nud" typically switches back to top-down internally, where pure bottom-up would not, but the choice of "nud" is necessarily made in a bottom-up way. Once "nud" returns, or the LR automaton reduces back to the initial state, the next thing it does is based on the operator token's "led." This corresponds to the LR automaton shifting that token and advancing into the parent rule determined by the operator. This resolution is just as "eager" in LR, where it is represented by the new state containing only items from the chosen rule. This is the hallmark of bottom-up parsing: as you progress through the children, you narrow down the possibilities for the parent. Finally, note that none of this has anything to do with the use of function calls vs a table of integers and explicit stack. Both top-down and bottom-up parsers can be implemented using either recursion or a table and stack.