1 CS 201 Compiler Construction Lecture 4 Data Flow Framework.

1

CS 201Compiler Construction

Lecture 4Data Flow Framework

2

Data Flow Framework

The various problems considered have things in common:– Transfer functions

– Confluence Operator

– Direction of Propagation

These problems can be treated in a unified way data flow framework is an algebraic structure used to encode and solve data flow problems.

3

Monotone Data Flow Framework

Components of the framework:1.Information Set: L2.Effect of joining paths: ∧ (meet operator)3.Effect of basic blocks: fn (monotone transfer func.)4.Iterative Solution: can be shown to terminate

(L,∧) is a semilattice st ∨ a,b,c εL1.a ∧ a = a (idempotent)2.a ∧ b = b ∧ a (commutative)3.a ∧ (b ∧ c) = (a ∧ b) ∧ c (assocative)

Bottom Element st ∨ a ε L, a ∧ = Top Element Τ st ∨ a ε L, a ∧ Τ = aIf top & bottom elements exist, they are unique.

Contd..

Relation ≤ is a partial order on L

a ≤ b ≅ a ∧ b = a

Can similarly define <, >, ≥ relations

A semilattice is bounded iff ∨ a εL there exists a constant ca st length of chain beginning at a is at most ca.

4

Max ca

5

Monotonic Functions

Effect of each basic block is modeled by a transfer function f: L L. Function f must be monotonic.

A total function f: LL is monotonic iff ∨ a,b ε L f(a∧b) ≤ f(a) ∧ f(b)

Distributive function: f(a∧b) = f(a) ∧ f(b)

For monotonic functions: a ≤ b => f(a) ≤ f(b)

6

Contd..

For monotonic functions: a ≤ b => f(a) ≤ f(b)

Proof:

f(a∧b) ≤ f(a) ∧ f(b) Defn. of Monotonicityf(a∧b)∧f(a)∧f(b) = f(a∧b) Defn. of ≤:(a∧b=a)f(a)∧f(a)∧f(b)= f(a) Given a ≤ b: a ∧ b = af(a) ∧ f(b) = f(a) f(a)∧f(a) = f(a) idempotencef(a) ≤ f(b) Defn. of ≤

7

Fixpoint

A fixpoint of a monotonic function f: L L is a valuea ε L such that f(a) = a

Τ > f (Τ) > f ( f (Τ) ) > f ( f ( f (Τ) ) ) ……..

There exists t such that f ( ft (Τ) ) = ft (Τ)

ft (Τ) is the greatest fixpoint of f.

8

Monotone Function SpaceA monotone function space for a semilattice is a set F of monotonic functions which:

1.Contains the identity function (id)-- basic blocks may not modify information

2.Is closed under function composition-- to model the effects of paths

3.For each a ε L, there exists fεF st f( ) = a-- to model gen functions

A distributive function space is a monotone function space in which all functions are distributive.

9

A Monotone Data Flow System

A monotone data flow system is a tuple < L, ∧, F, G, FM >

1.(L,∧) is a bounded semilattice with Τ & 2.F is the monotone function space3.G = (N, E, s) is the program flow graph4.FM: N F is a total function that associates a function from F with each basic block.

10

Meet Over All Paths SolutionMeet over all paths solution (MOP) of a data flow system – MOP: N L

MOP(s) = NULL (NULL is the element in L which represents “no information”)

Ffπ is composition of functions from nodes along path π excluding node n. n1n2n3….nk-1nk fnk o fnk-1 o….o fn2 o fn1

11

MOP SolutionFinding MOP solution is undecidable, i.e. there does not exist a general algorithm that computes MOP solution for all monotone data flow systems.

Let X: N L denote a total function that associates nodes with lattice elements.X is conservative or safe iff ∨n εN, X(n) ≤ MOP(n)

Iterative algorithm computes conservative approximation of MOP. For distributive data flow systems, it computes solution that is identical to MOP solution.

12

Iterative Algorithm

13

Reaching Definitions

14

Contd…

15

Dominators

16

Constant Propagation

f (X)={(a,2),(b,3),(c,5)}f (Y)={(a,3),(b,2),(c,5)}f (X) ∧ f (Y) = {(a,not-const), (b, not-const),

(c,5)}

X ∧ Y = {(a,not-const),(b,not-const),(c,undef)}f (X∧ Y) = {(a,not-const),(b,not-const),(c,not-

const)}

f (X ∧ Y) ≤ f(X) ∧ f(Y)