1. Lazy Code Motion C OMP 512 Rice University Houston, Texas Fall 2003 Copyright 2003, Keith D. Cooper & Linda Torczon, all rights reserved. Students enrolled in Comp 512 at Rice University have explicit permission to make copies of these materials for their personal use. This lecture draws on two papers: “Lazy Code Motion,” J. Knoop, O. Ruthing, & B. Steffen, in PLDI 92 “A Variation of Knoop, Ruthing, and Steffen’s Lazy Code Motion,” K. Drechsler & M. Stadel, SIGPLAN Notices, 28(5), May 1993

2. COMP 512, Fall 20032 Where are we? Machine Independent Redundancy Redundancy elim. Partial red. elim. Consolidation Code motion Loop-invariant c.m. Consolidation Global Scheduling [Click] Constant propagation Useless code Dead code elim. Partial d.c.e. Constant propagation Algebraic identities Create opportunities Reassociation Replication Specialization Replication Strength Reduction Constant Propagation Method caching Heap  stack allocation Tail recursion elimination * Today (lazy code motion)

3. COMP 512, Fall 20033 Some occurrences of b+c are redundant Redundant Expression An expression is redundant at point p if, on every path to p 1. It is evaluated before reaching p, and 2. Non of its constitutent values is redefined before p Example a  b + c b  b + 1 a  b + c *

4. COMP 512, Fall 20034 b  b + 1 a  b + c Inserting a copy of “a  b + c” after the definition of b can make it redundant fully redundant? Partially Redundant Expression An expression is partially redundant at p if it is redundant along some, but not all, paths reaching p Example b  b + 1a  b + c *

5. COMP 512, Fall 20035 Loop Invariant Expression Another example Loop invariant expressions are partially redundant Partial redundancy elimination performs code motion Major part of the work is figuring out where to insert operations x  y * z a  b * c b+c is partially redundant here x  y * z a  b * c

6. COMP 512, Fall 20036 Lazy Code Motion The concept Solve data-flow problems that reveal limits of code motion Compute I NSERT & D ELETE sets from solutions Linear pass over the code to rewrite it (using I NSERT & D ELETE ) The history Partial redundancy elimination ( Morel & Renvoise, CACM, 1979 ) Improvements by Drechsler & Stadel, Joshi & Dhamdhere, Chow, Knoop, Ruthing & Steffen, Dhamdhere, Sorkin, … All versions of PRE optimize placement  Guarantee that no path is lengthened L CM was invented by Knoop et al. in P LDI, 1992 We will look at a variation by Drechsler & Stadel SIGPLAN Notices, 28(5), May, 1993

7. COMP 512, Fall 20037 Lazy Code Motion The intuitions Compute available expressions Compute anticipable expressions These lead to an earliest placement for each expression Push expressions down the CFG until it changes behavior Assumptions Uses a lexical notion of identity ( not value identity ) I LOC -style code with unlimited name space Consistent, disciplined use of names  Identical expressions define the same name  No other expression defines that name } Avoids copies Result serves as proxy

8. COMP 512, Fall 20038 Lazy Code Motion The Name Space r i + r j  r k, always ( hash to find k ) We can refer to r i + r j by r k ( bit-vector sets ) Variables must be set by copies  No consistent definition for a variable  Break the rule for this case, but require r source < r destination  To achieve this, assign register names to variables first Without this name space L CM must insert copies to preserve redundant values L CM must compute its own map of expressions to unique ids Digression in Chapter 5 of EAC: “The impact of naming”

9. COMP 512, Fall 20039 Lazy Code Motion Local Predicates D EF (b) contains expressions defined in b that survive to the end of b. e  DEF(b)  evaluating e at the end of b produces the same value for e A NT L OC (b) contains expressions defined in b that have upward exposed arguments (both args). e  A NT L OC (b)  evaluating e at the start of b produces the same value for e NotKilled(b) contains those expressions whose arguments are not defined in b. e  N oT K ILLED (b)  evaluating e produces the same result at the start and end of b D EF and N OT K ILLED are from A VAIL

10. COMP 512, Fall 200310 Lazy Code Motion Availability Initialize A VAIL I N (n) to the set of all names, except at n 0 Set A VAIL I N (n 0 ) to Ø Interpreting A VAIL e  A VAIL O UT (b)  evaluating e at end of b produces the same value for e. A VAIL O UT tells the compiler how far forward e can move This differs from the way we talk about A VAIL in global redundancy elimination A VAIL I N (n) =  m  preds(n) A VAIL O UT (m), n ≠ n 0 A VAIL O UT (m) = D EF (m)  (A VAIL I N (m)  N OT K ILLED (m))

11. COMP 512, Fall 200311 Lazy Code Motion Anticipability Initialize A NT O UT (n) to the set of all names, except at exit blocks Set A NT O UT (n) to Ø, for each exit block n Interpreting A NT O UT e  A NT I N (b)  evaluating e at start of b produces the same value for e. A NT I N tells the compiler how far backward e can move This view shows that anticipability is, in some sense, the inverse of availablilty (& explains the new interpretation of A VAIL ) A NT O UT (n) =  m  succs(n) A NT I N (m), n not an exit block A NT I N (m) = A NT L OC (m)  (A NT O UT (m)  N OT K ILLED (m))

12. COMP 512, Fall 200312 Lazy Code Motion Earliest placement E ARLIEST is a predicate Computed for edges rather than nodes ( placement ) e  EARLIEST(i,j) if  It can move to head of j,  It is not available at the end of i and  either it cannot move to the head of i (N OT K ILLED (i)) or another edge leaving i prevents its placement in i (A NT O UT (i)) E ARLIEST (i,j) = A NT I N (j)  A VAIL O UT (i)  (N OT K ILLED (i)  A NT O UT (i)) E ARLIEST (n 0,j) = A NT I N (j)  A VAIL O UT (n 0 )

13. COMP 512, Fall 200313 Lazy Code Motion Later (than earliest) placement Initialize L ATER I N (n 0 ) to Ø x  L ATER I N (k)  every path that reaches k has x  E ARLIEST (m) for some block m, and the path from m to k is x-clear & does not evaluate x  the compiler can move x through k without losing any benefit x  L ATER (i,j)  is its earliest placement, or it can be moved forward from i (L ATER (i)) and placement at entry to i does not anticipate a use in i ( moving it across the edge exposes that use ) L ATER I N (j) =  m  pred(n) L ATER (i,j), j ≠ n 0 L ATER (i,j) = E ARLIEST (i,j)  (L ATER I N (i)  A NT L OC (i))

14. COMP 512, Fall 200314 Lazy Code Motion Rewriting the code I NSERT & D ELETE are predicates Compiler uses them to guide the rewrite step x  I NSERT (i,j)  insert x at start of i, end of j, or new block x  D ELETE (k)  delete first evaluation of x in k I NSERT (i,j) = L ATER (i,j)  L ATER I N (j) D ELETE (k) = A NT L OC (k)  L ATER I N (k), k ≠ n 0 If local redundancy elimination has already been performed, only one copy of x exists. Otherwise, remove all upward exposed copies of x *

15. COMP 512, Fall 200315 Lazy Code Motion Edge placement x  I NSERT (i,j) Three cases |succs(i)| = 1  insert at end of i | succs(i)| > 1, but |preds(j)| = 1  insert at start of j | succs(i)| > 1, & |preds(j)| > 1  create new block in for x BiBi BjBj |succs(i)| = 1 x |preds(j)| = 1 BiBi BjBj BkBk x |succs(i) > 1 |preds(j)| > 1 BiBi BjBj BkBk BhBh x

16. COMP 512, Fall 200316 Lazy Code Motion Example B 1 : r 1  1 r 2  r 0 + @ m if r 1 < r 2  B 2, B 3 B 2 : … r 20  r 17 * r 18... r 4  r 1 + 1 r 1  r 4 if r 1 < r 2  B 2, B 3 B 3 :... B1B1 B2B2 B3B3 Example is too small to show off Later Insert(1,2) = { r 20 } Delete(2) = { r 20 }