Assignment 5: Hundred-pacer: Register Allocation

Due: Sun 11/05 at 11:59pm

git clone

The popular name “hundred pacer” refers to a local belief that, after being bitten, the victim will only be able to walk 100 steps before dying. Incorporating register allocation should allow your compiled code to complete its paces faster than before

This assignment will be a bit different from the others: there are no new features to implement and we won’t be testing your compiler’s input/output behavior directly. Rather this time you will be evaluated on how well you implement the component analyses of a register allocator. Incorporating the register allocation into your compiler is optional and will not be graded.

We break the assignment down into two parts. First, we need to analyze the code to determine which variables cannot be stored in the same register. Second, we use Chaitin’s graph coloring algorithm with a perfect elimination ordering derived from the function’s syntax to assign registers (or spilled stack slots) to all variables.

1 Liveness and Conflict Analysis

First need to figure out, for each global function definition in the program, which variables conflict with each other. We break this down into liveness analysis, which determines when a variable’s value is needed and conflict analysis which determines which variables are live at the same time with possibly different values.

We implement liveness analysis as a compiler pass, annotating all expressions with the free variables live at that point. It has the following type:

fn liveness<Ann>(e: &SeqExp<Ann>, params: &HashSet<String>) -> SeqExp<HashSet<String>> {
}

The parameter set params are the parameters in the current function body. You need this information because you should not treat these parameters as part of your liveness analysis since where the parameters are stored is dictated by the calling convention and so is not subject to register allocation1more sophisticated register allocators handle this by "pre-coloring" the parameters with the register dictated by the calling convention. The output of this function is an expression with the same structure as e but with each sub-expression annotated with the variables that are live in that sub-expression. So for example, if the input expression were

let x = 3 in f(x, y, a)

where a is a parameter, y is not and f is a globally defined function, then your liveness function would return the expression annotated as follows:

Let { var: "x",
      bound_exp: Imm(Num(3), { }),
      body: ExternalCall{ fun_name: "f", args: [Var("x"), Var("y"), Var("a")], is_tail: true, ann: {"x", "y"}},
      ann: {"y"},
    }

The inner ExternalCall has both x and y live because both values are used as arguments. a is not live since it is a parameter. The outer Let only has y live in it because x is written to in the let.

The liveness function returns a SeqExp<HashSet<String>>, an expression annotated with its liveness information. To extract the liveness information for an expression you can use the .ann() method implemented in syntax.rs.

As discussed in the Monday, October 30th lecture, the algorithm we gave for liveness analysis only works correctly when the SeqExps are "fully flattened", meaning the bound_exp of a let expression can only ever be an immediate, call or primitive operation, and not a let, if or fundef expression. The autograder will only test your analysis on examples that are fully flattened in this way, but if you want to incorporate register allocation in your compiler you should add a pass to flatten your expressions first.

Next, given our liveness analysis, we go through and determine what the conflicts are. This generates a graph that will be the input to our graph coloring function. The conflicts function has the following type.

fn conflicts(e: &SeqExp<HashSet<String>>) -> Graph<String>

The output of the conflicts function is a Graph<String> representing the conflicts between variables. The Graph datatype is provided in the graph.rs module, and represents an undirected graph. Each method is documented with a brief description. Note that other than Graph::new(), these are all methods on a graph, so for instance to insert a vertex v into a graph g using the insert_vertex method, you invoke the method as g.insert_vertex(v).

Make sure you insert all let-bound variables and arguments to local function definitions into the graph you produce so that they are all assigned a register, even if they have no conflicts. To ensure that you don’t put any parameters in your output graph, add the variables into your output when you see them defined, rather than when they are used.

2 Graph Coloring

Now that we have our conflict graph, we attempt to assign each vertex a "color", i.e., register. As shown in class, we will use Chaitin’s algorithm, but using the perfect elimination order derived from the syntax being in SSA form.

So first you should implement

fn elimination_order<Ann>(e: &SeqExp<Ann>) -> Vec<String>

which produces an ordering of all of the variables defined in let and local function definitions so that if x is in scope when y is defined, then x occurs before y in the output vector. So for example in

let x = 5 in
let y = x * 3 in
y * x

fn graph_color(
    conflicts: Graph<String>,
    elimination_order: &[String],
    registers: &[Reg],
) -> HashMap<String, VarLocation>

You are given the register interference graph, a vector of variables representing the elimination order and a slice of registers to be used in register allocation and you should return a HashMap mapping the variables in the input graph to a VarLocation, which is either a register VarLocation::Reg(r) or a stack offset VarLocation::Spill(i).

Chaitin’s algorithm is defined recursively by removing a vertex from the graph, recursively coloring the rest of the graph and then adding the vertex back in and coloring it with the next available color. You must remove the vertices in the order given by the elimination ordering to ensure that you color the graph optimally.

We parameterize the graph coloring algorithm by which registers we are using so that we can test spilling more easily by passing in fewer registers. You may find this useful for incrementally implementing your use of registers as well2For instance you might first use no registers, then use only caller-save registers and finally both caller-save and callee-save registers.

3 Recommended TODO List

We suggest handling the graph coloring routines in the following order:

1. liveness

2. conflict analysis

3. elimination order

4. graph coloring

4 List of Deliverables

• your analysis.rs, in particular the liveness, conflicts and allocate_registers functions

• any additional rust files you saw fit to write

• the Cargo.toml

5 Grading Standards

This assignment will be solely autograded. There will not be any hidden tests.

There are two kinds of tests:

• Liveness/Conflict analysis tests to test the soundness and precision of your analyses

• Full register allocator tests that ensure you always give valid register/spill assignments and don’t spill when unnecessary

6 Submission

Wait! Please read the assignment again and verify that you have not forgotten anything!

1more sophisticated register allocators handle this by "pre-coloring" the parameters with the register dictated by the calling convention

2For instance you might first use no registers, then use only caller-save registers and finally both caller-save and callee-save registers