Agenda

CS738: Advanced Compiler Optimizations

Data Flow Analysis

Amey Karkare ◮ Intraprocedural Data Flow Analysis: Classical Examples

[email protected] ◮ Last lecture: Reaching Definitions
◮ Today: Available Expressions
http://www.cse.iitk.ac.in/~karkare/cs738 ◮ Discussion about the similarities/differences
Department of CSE, IIT Kanpur

Available Expressions Analysis AvE Analysis of a Structured Program

IN(s1 )
◮ An expression e is available at a point p if
d :x =y +z s1
◮ Every path from the Entry to p has at least one evaluation
of e
◮ There is no assignment to any component variable of e OUT(s1 )
after the last evaluation of e prior to p
◮ Expression e is generated by its evaluation
OUT(s1 ) = IN(s1 ) − KILL(s1 ) ∪ GEN(s1 )
◮ Expression e is killed by assignment to its component
GEN(s1 ) = {y + z}
variables
KILL(s1 ) = Ex
where Ex : set of all expression having x as a component
This may not work in general – WHY?
AvE Analysis of a Structured Program AvE Analysis of a Structured Program

IN(s1 )
IN(s1 )
x =x +z s1
lhs = rhs s1
OUT(s1 )
OUT(s1 )

OUT(s1 ) = IN(s1 ) − KILL(s1 ) ∪ GEN(s1 )

GEN(s1 ) = {x + z} OUT(s1 ) = IN(s1 ) − KILL(s1 ) ∪ GEN(s1 )
KILL(s1 ) = Ex GEN(s1 ) = {rhs | lhs is not part of rhs}
Incorrectly marks x + z as available after s1 KILL(s1 ) = Elhs
GEN(s1 ) = ∅ for this case

AvE Analysis of a Structured Program AvE Analysis of a Structured Program

IN(S)
IN(S)
S
s1 S

s1 s2
s2
OUT(S)
OUT(S)

GEN(S) = GEN(s1 ) − KILL(s2 ) ∪ GEN(s2 )

GEN(S) = GEN(s1 ) ∩ GEN(s2 )
KILL(S) = KILL(s1 ) − GEN(s2 ) ∪ KILL(s2 )
KILL(S) = KILL(s1 ) ∪ KILL(s2 )
IN(s1 ) = IN(S)
IN(s1 ) = IN(s2 ) = IN(S)
IN(s2 ) = OUT(s1 )
OUT(S) = OUT(s1 ) ∩ OUT(s2 )
OUT(S) = OUT(s2 )
AvE Analysis of a Structured Program AvE Analysis of a Structured Program

IN(S)
S
IN(S) = {x + y }
s1
S
s1 : nop
OUT(S)

GEN(S) = GEN(s1 ) OUT(S) =?

KILL(S) = KILL(s1 )
OUT(S) = OUT(s1 ) Is x + y available at OUT(S)?
IN(s1 ) = IN(S) ∩ GEN(s1 ) ?
IN(s1 ) = IN(S) ∩ OUT(s1 )??

AvE Analysis is Approximate AvE Analysis is Approximate

IN(S) IN(S)
S S
s1 s2 s1 s2

OUT(S) OUT(S)
◮ Assumption: All paths are feasible. ◮ Thus,
◮ Example: true GEN(S) ⊇ analysis GEN(S)
if (true) s1; true KILL(S) ⊆ analysis KILL(S)
else s2; ◮ Fewer expressions marked available than actually do!
Fact Computed Actual ◮ Later we shall see that this is SAFE approximation
GEN(S) = GEN(s1 ) ∩ GEN(s2 ) ⊆ GEN(s1 ) ◮ prevents optimizations
◮ but NO wrong optimization
KILL(S) = KILL(s1 ) ∪ KILL(s2 ) ⊇ KILL(s1 )
AvE for Basic Blocks Solving AvE Constraints

◮ Expr e is available at the start of a block if

◮ It is available at the end of all predecessors
\
IN(B) = OUT(P) ◮ KILL & GEN known for each BB.
P∈PRED(B) ◮ A program with N BBs has 2N equations with 2N
◮ Expr e is available at the end of a block if unknowns.
◮ Either it is generated by the block ◮ Solution is possible.
◮ Or it is available at the start of the block and not killed by ◮ Iterative approach (on the next slide).
the block
OUT(B) = IN(B) − KILL(B) ∪ GEN(B)

Some Issues

for each block B {

OUT(B) = U ; U = “universal” set of all exprs
}
OUT(Entry ) = ∅; // remember reaching defs? ◮ What is U – the set of all expressions?
change = true;
◮ How to compute it efficiently?
while (change) {
change = false; ◮ Why Entry block is initialized differently?
for each blockT B other than Entry {
IN(B) = P∈PRED(B) OUT(P);
oldOut = OUT(B);
OUT(B) = IN(B) − KILL(B) ∪ GEN(B);
if (OUT(B) 6=oldOut) then {
change = true;
}
}
}
Available Expressions: Example Available Expressions: Example

Pass# Pt B1 B2 B3 B4
Init IN - - - -
BB GEN KILL OUT U U U U
B1 {a*b, c+d} {} 1 IN ∅ a*b, c+d c+d
c+d
B2 {c+d} {a*b} OUT a*b, c+d a*b, a*b
B3 {a*b} {} c+d c+d
2 IN ∅ a*b c+d c+d
B4 {a*b} {c+d} OUT a*b, c+d a*b, a*b
c+d c+d
3 IN ∅ a*b c+d c+d
U = {a*b, c+d} OUT a*b, c+d a*b, a*b
c+d c+d

Available Expressions: Bitvectors Available Expressions: Bitvectors

◮ Set-theoretic definitions:
\
a bit for each expression: IN(B) = OUT(P)
a*b c+d
P∈PRED(B)
Pass# Pt B1 B2 B3 B4
Init IN - - - - OUT(B) = IN(B) − KILL(B) ∪ GEN(B)
OUT 11 11 11 11
1 IN 00 11 01 01 ◮ Bitvector definitions:
OUT 11 01 11 10
2 IN 00 10 01 01
^
OUT 11 01 11 10 IN(B) = OUT(P)
3 IN 00 10 01 01 P∈PRED(B)
OUT 11 01 11 10
OUT(B) = IN(B) ∧ ¬KILL(B) ∨ GEN(B)
◮ Bitwise ∨, ∧, ¬ operators
Available Expressions: Application Comparison of RD and AvE

◮ Common subexpression elimination in a block B ◮ Some vs. All path property

◮ Expression e available at the entry of B
◮ Meet operator: ∪ vs. ∩
◮ e is also computed at a point p in B
◮ Components of e are not modified from entry of B to p ◮ Initialization of Entry: ∅
◮ e is “upward exposed” in B ◮ Initialization of other BBs: ∅ vs. U
◮ Expressions generated in B are “downward exposed” ◮ Safety: “More” RD vs. “Fewer” AvE

AvE: alternate Initialization Live Variables

◮ What if we Initialize:
◮ A variable x is live at a point p if
OUT(B) = ∅, ∀B including Entry ◮ There is a point p′ along some path in the flow graph
starting at p to the Exit
◮ Would we find “extra” available expressions? ◮ Value of x could be used at p′
◮ More opportunity to optimize? ◮ There is no definition of x between p and p′ along this path
◮ OR would we miss some expressions that are available? ◮ Otherwise x is dead at p
◮ Loose on opportunity to optimize?
Live Variables: GEN Live Variables: KILL

◮ GEN(B): Set of variables whose values may be used in ◮ KILL(B): Set of variables defined in block B prior to any
block B prior to any definition use
◮ Also called “use(B)” ◮ Also called “def(B)”
◮ “upward exposed use” of a variable in B ◮ “upward exposed definition” of a variable in B

Live Variables: Equations Very Busy Expressions

◮ Set-theoretic definitions:
[
OUT(B) = IN(S)
S∈SUCC(B) ◮ Expression e is very busy at a point p if
◮ Every path from p to Exit has at least one evaluation of e
IN(B) = OUT(B) − KILL(B) ∪ GEN(B) ◮ On every path, there is no assignment to any component
◮ Bitvector definitions: variable of e before the first evaluation of e following p
_ ◮ Also called Anticipable expression
OUT(B) = OUT(S)
S∈SUCC(B)

IN(B) = OUT(B) ∧ ¬KILL(B) ∨ GEN(B)

◮ Bitwise ∨, ∧, ¬ operators
QQ

◮ Expression e is very busy at a point p if

◮ Every path from p to Exit has at least one evaluation of e
and there is no assignment to any component variable of e
before the first evaluation of e following p on these paths.
◮ Set up the data flow equations for Very Busy Expressions
(VBE). You have to give equations for GEN, KILL, IN, and
OUT.
◮ Think of an optimization/transformation that uses VBE
analysis. Briefly describe it (2-3 lines only)
◮ Will your optimization be safe if we replace “Every” by
“Some” in the definition of VBE?