LARA

Live-Variable Analysis

For each program point, and each variable approximately compute whether the variable is:

  • definitely not live: its current value will not be used in the future, or
  • it is possibly live: its value may be used in the future

The precise notion of liveness is given by program semantics, as follows.

Definition: a variable $x$ is dynamically live in a concrete program state $(u_i,w_i)$, if there exists a control-flow graph execution

\begin{equation*}
   (u_i,w_i),
   s_i,
   (u_{i+1},w_{i+1}),
   s_{i+1},
   \ldots,
   (u_n,w_n),
   s_n,
\end{equation*}

(where $u_k$ are control-flow graph nodes, $w_k$ are values of variables), such that

  • the statement $s_n$ reads the value of $x$ (i.e. it is a test mentioning $x$ or an assignment statement with $x$ on the right-hand side)
  • none of the statements $s_i, s_{i+1}, \ldots, s_{n-1}$ writes to $x$ (i.e. is an assignment of the form $x = ...$)

We design a data-flow analysis that, for each program point $u$, computes static liveness information for program variables, given by the set of variables $live(u)$.

Correctness statement for liveness analysis: if $x \notin live(u)$ then for every program execution that reaches a state $(u,w)$, the variable $x$ is dynamically live at point $u$.

NOTE: if a static analysis says that the variable is live at program point, the variable may or may not be dynamically live

  • but if we know that it is not statically live, this is useful information, we know that it is not dynamically alive

Consider a sequence of instructions:

code live variables
{z}
x = 42
{x,z}
y = x + 3
{x,y,z}
z = y + z
{x}
y = 3 + x
{}

Applications:

  • allocating space for variables (e.g. register allocation for CPU)
    • if variable not used in future, we can store another variable in the same address
  • an alternative to initialization analysis: must be initialized if it will be used
    • if variable is used in the future before being assigned, it must be initialized now
    • we can do initialization check by checking that no variable is live at CFG entry

Liveness is naturally computed using backwards data-flow analysis

Consider the program execution backwards

  • execution is very non-deterministic (e.g. x=3 goes into all values of x)
  • mathematically equally well-defined
  • introduce an additional state bit to variable, mark it “used” when it is used
  • if a state is reached in backward execution where it is used, then it “will be used”

Corresponding backward analysis:

  • variable uses flow towards their initializations
  • edges in data-flow analysis are interpreted in the opposite way
  • analysis starts from the exit point

Final state at the exit point

  • no variable is live - no more statements at the end, so no future uses
  • this is also the bottom of the lattice

Pointwise Representation

For each variable, store its liveness:

  1. (potentially) live (top)
  2. not live (bottom)

Bottom: map each variable to bottom

Join: pointwise join

  • bottom $\sqcup$ top = top
  • $m_1 \sqcup m_2 = \lambda i. m_1(i) \sqcup m_2(i)$

An Alternative Representation

The set of potentially live variables, instead of

  • $m : \mbox{Var} \to \{\top, \bot\}$ consider
  • $m' \subseteq \mbox{Var}$, given by $m' = \{ x \mid m(x)=\top \}$
  • this is just a different notation for the same thing

Bottom: empty set

Top: set of all variables

Join: union

Semilattice for Live-Variable Analysis

(Semilattice is like lattice but need not have meet.)

Elements are sets $S$ of live variables

  • we assume that variables have been renamed according to scoping rules

Join is union $\cup$

Transfer Functions for Live-Variable Analysis

For each statement st in CFG, we introduce sets of variables

  • use(st) denote variables used in statement
  • def(st) denote variables overwritten in st

For SimpleCFG.scala we have

st use(st) def(st)
x = y op z {y,z} {x}
x = y {y} {x}
Assume[y relOp z] {y,z} {}
print(y) {y}

Examples:

st use(st) def(st)
x = y + 1 {y} {x}
x = x + 1 {x} {x}

In ordinary execution, statement

  • first uses variables in use(st) to compute some value
  • then assigns this value to variables in def(st)

In backward execution, statement

  • marks def(st) as not live
  • marks use(st) as live

in that order

Transfer function

\begin{equation*}
   transFun(st,S) = (S \setminus d{}ef(st)) \cup use(st)
\end{equation*}

We have seen forward analyses, where for each point $v$, we have:

\begin{equation*}
    facts(v) = facts(v) \sqcup \bigsqcup_{e.v2=v} transFun(e.lab, e.v1) \ \ \ (F)
\end{equation*}

in backward analysis, we instead have:

\begin{equation*}
    facts(v) = facts(v) \sqcup \bigsqcup_{e.v1=v} transFun(e.lab, e.v2) \ \ \ (B)
\end{equation*}

References