Objectives

1. Objectives • Identify advantages (and disadvantages ?) of optimizing in SSA form • Given a CFG in SSA form, perform • Global Constant Propagation • Dead code elimination • Global Value Numbering • Conversion of SSA back to CFG form

2. Constant Propagation • Operate on sparse graph (SSA) • Incorporate branch folding • Meet operations at -nodes

3. EvalStmt Algorithm EvalStmt( Stmt ) if Stmt is arithmetic or -node evaluate Stmt if result is “lowered” foreach j  Uses (Stmt.Lval())) propagate result if j.block  Visited Work = { j } else if Stmt is a branch foreach possible destination DB of Stmt if Edge(Stmt, DB) not executable mark Edge(Stmt, DB) executable Blocks U= DB

4. Constant Prop Algorithm Mark all CFG Edges not executable Init node of SSA graph to T Work = Visited = Ø Blocks = {Entry} While Work  Ø  Blocks Ø While Work  Ø Choose Stmt from Work EvalStmt (Stmt) While Blocks  Ø Choose BB from Blocks foreach Stmt in  (BB) EvalStmt (Stmt) if BB Visited Visited = {BB} foreach Stmt in  (BB) EvalStmt (Stmt)

5. Dead Code Elimination • Use SSA to detect dead code • Method • Remove statement(s) that do not directly or indirectly use data observable outside the procedure • Eliminate branches never taken • Uses control dependence

6. EliminateDeadCode () Worklist = Necessary = Ø foreach BB  CFG foreach StmtBB if (Stmt defines external data )  (Stmt is and I/O instruction)  (Stmt is function call) Necessary U= { Stmt } WorkList U= { Stmt } endif endfor endfor

7. Eliminate Dead Code Algorithm while Worklist  Ø choose Stmt from Worklist BB = Stmt’s basic block foreach block, CB, upon which BB is control dependent J = CB’s branch statement Necessary U= J Worklist U= J foreach T  Stmt’s operands Def = Stmt’s definition Necessary U= {Def } WorkList U= {Def } foreach BB  CFG foreach StmtinBB if Stmt  Necessary removeStmt else if Stmt isa branch Stmt  Necessary Stmt.target=ipdom(BB)

8. Global Value Numbering • Apply value numbering at function scope • Associate a field (for value number) with each temporary. • Temps with same value number are = • No loops ==> reverse postorder traversal suffices (operands defined before used.)

9. Strongly Connected Components • What about loops ? • Value numbers for -nodes ambiguous • Assume best case • Iterate on the SCC in reverse postorder until no further changes

10. Processing -nodes • If -node has a value number, VN, assign VN to temp defined by -node • Consider operands with non-Null value numbers: If two have different values, assign a new value for -node. • If all operands have same value number assign that to temp defined by -node

11. Efficiency Refinements • If -node has non-Null value number different from its operands, don’t need a new value number • For SCC, use a temp value number table called scratch table. Once that table “stabilizes”, copy to value table.

12. SSA to CFG • -nodes “not executable” directly • Must convert SSA back to CFG • “Semantics” of -nodes is that they are “executed” simultaneously upon entry to block • Insert code for each operand of -node in appropriate predecessor block.