1 CS 201 Compiler Construction Code Optimizations: Partial Dead Code Elimination

Dead Code Elimination Dead Code is code that is either never executed or, if it is executed, its result is never used by the program. Therefore dead code can be eliminated from the program. 2 dead code x = … y = y+1 x = y+1 write x dead code A = 1 B = 5 If A<B ……. write x ……. TF

Dead Code: More Examples 3 Values used only by dead statements. x = … y = x+1 read y write y+1 x = … …..…x = x+1

Partially Dead Code Value computed by partially dead statement is sometimes used and sometimes not used. 4 x is partially dead here x is live here x is dead here x =... x = x +1 write x x = y +1

Code Motion Reqd. for Elimination 5 x =... x = x +1 write x x = y +1 x = … x = x +1 write x x = … x = y +1 x = … x = x +1 write x x = y +1

Removal of Code From Loops 6 … = x.…..… x = y+1 no use of x x = y+1 … = x.…..… …….....

Critical Edges 7 X = …. …… = XX = ….. critical edge should be split ……... …… = X X = ….. X = ….

Second Order Effects 8 a=c+d cannot be moved till x=a+b is moved

Algorithm Repeat Perform dead/faint variable analysis and eliminate dead/faint code Perform delayability analysis for sinking candidates Place statements in new positions Until Program Stabilizes 9

Definitions Dead Code: S: x = t is dead, if x is dead following S, i.e. every path from S to every right hand side occurrence of x following S is preceded by a redefinition of x. Faint Code: S: x = t is faint, if x is faint following S, i.e. every path from S to every right hand side occurrence of x following S is either preceded by a redefinition of x or is in a statement whose left hand side variable is faint. 10

Local Predicates USED s (x) – x is rhs variable of statement s RELV-USED s (x) – x is a rhs variable of relevant statement s; relevant statements are those that cannot be eliminated – conditionals like if x<y and output statements like write x. ASSIGN-USED s (x) – x is a rhs variable of assignment statement s. MOD s (x) – x is lhs variable of statement s. 11

Dead Variable Analysis Bit vector analysis for all variables. 12 N-DEAD s = USED s ( X-DEAD s V MOD s ) V V X-DEAD s = N-DEAD i i ε succ(s)

Faint Variable Analysis Not bit vector analysis. Perform following analysis for each x 13 V X-FAINT s (x) = N-FAINT i (x) i ε succ(s) N-FAINT s (x) = RELV-USED s (x) ∧ ( X-FAINT s (x) ∨ MOD s (x) ) ∧ ( X-FAINT s (lhs s ) ∨ ASSIGN-USED s (x) )

Analysis of Equation 14 N-FAINT S (x) is false if any of the following conditions is true: 1.the statement s uses x and s cannot ever be marked faint, i.e. 2. the statement s does not modify x and x is not faint on exit, i.e. 3. x is used by statement s and lhs of s is not faint, i.e. REL-USED s (x) is true ( X-FAINT s (x) ∧ MOD s (x) ) is true ( X-FAINT s (lhs) ∧ ASSIGN-USED s (x) ) is true REL-USED s (x) ∨ ( X-FAINT s (x) ∧ MOD s (x) ) ∨ ( X-FAINT s (lhs) ∧ ASSIGN-USED s (x) ) N-FAINT s (x) =

Elimination Following dead/faint variable analysis we can eliminate dead/faint code  Eliminate n: x=.. assignment if x is dead/faint at exit of n. The above elimination rule eliminates fully dead/faint code. To eliminate partially dead code, we develop delayability analysis  using this analysis partially dead statement is moved to program points where it is fully dead or fully live. Iterative application of dead/faint code elimination and delaying partially dead assignments leads to optimization. 15

Delayability 16 x = a+b ……… write x x = c+d ………. …….… write x x = c+d x = a+b Delayability Analysis fully dead fully live

Delayability Analysis 17 LOCDELAYED n (α): α is a sinking candidate that ε n. LOCBLOCKED n (α): sinking of α is blocked by code in n. N-DELAYED n /X-DELAYED n indicate whether α can be delayed to the entry/exit point of n. N-DELAYED n (α) = X-DELAYED n (α) = LOCDELAYED n (α) ∨ ( N-DELAYED n (α) ∧ LOCBLOCKED n (α) ) falseif n = s otherwise ∧ X-DELAYED m (α) m ε pred(n)

Contd.. 18 Placing α at its new positions. If N-INSERT n (α)/X-INSERT n (α) is true then place α at n’s entry/exit point. Essentially we are moving α to the latest points to which it can be delayed. N-INSERT n (α) = N-DELAYED n (α) ∧ LOCBLOCKED n (α) X-INSERT n (α) = X-DELAYED n (α) ∧ ( ∨ N-DELAYED m (α) ) m ε succ(s)

Sample Problem Apply the PDE algorithm to the given code example. 19 P=P+1 B=P+1 P = 1 output B P=E+1 input E output A A=B+1