Data Flow Analysis (I)

Transcription

1 Compiler Design Data Flow Analysis (I) Hwansoo Han

2 Control Flow Graph What is CFG? Represents program structure for internal use of compilers Used in various program analyses Generated from AST or a sequential list of statements Basic block To build CFG, first partition a program into basic blocks Longest possible sequence of straight-line code no branch-in/out in the middle of basic block single entry at the beginning (label, join-point) single exit at the end (can be branch or non-branch) 2

3 CFG Example CFG = Basic blocks + Control transfers Unreachable code elimination DFS (or BFS) on the CFG can remove unreachable blocks As a result, the declaration of z may also be removed. int test(int n) { int x,y,z,w; if (n < 2) x = 3; else x = 0; goto l; z = x + 5; y = z * 3; l: w = x + 10; y = w * 2; return y; } z=x+5 y=z*3 x=3 test() n<2 w=x+!0 y=w*2 return y goto l x=0 3

4 Data Flow Problem Applying data flow analysis on CFG Identify how data are manipulated/modified Optimization relies on information quality acquired thru DFA DFA is a process to solve the effects of statements, basic blocks or larger segments Typical data flow problems Available expressions, reaching definitions Aliases, live variables, copy propagation Upward exposed uses (live-in for basic block) n<2 x=3 x=0 goto l w=x+!0 y=w*2 return y Analyze data information at each point 4

5 Data Flow Analysis (DFA) Data flow information (DFI) Abstract property of program data (variable, value, name) At various points in the control flow Data flow equation (DFE) Represents effects of statements, basic blocks, or program segments Use systematic equations using following components 1. sets or functions (in, out, gen, kill, f, g, ) 2. operations (,,,) on sets/functions To gather DFI, DFEs are solved by well-defined mathematical procedures. 5

6 DFA on Program Hierarchy Local data flow analysis Composes effect of each statement within a basic block Proceeds sequentially from the leader to the last statement in the basic block Global data flow analysis Composes effect of each program segment (basic blocks, loops or intervals) at the segment boundaries in a CFG Proceeds along the control transfer in a control flow graph 6

7 Local Data Flow Analysis Transfer function (f S ) Represents effects of statement S on DFI within a basic block. Corresponding data flow equation out[s] = f S (in[s]) DFI in[s i ] S i-1 Typical transfer function f S (in[s]) = gen[s] (in[s] kill[s]) gen[s]: DFI newly generated by S kill[s]: DFI that was originally in in[s], but regenerated by S f S i gen[s i ] kill[s i ] out[s i ] S i S i+1 7

8 Example: Local Reaching Definition DFI of reaching definition Which variable definitions reach which points of uses? DFE for each statement S. in[s] : set of definitions that may reach S out[s] : set of definitions that are valid right after S kill[s] : set of definitions that reach S but are killed because S assigns new values to the definitions gen[s] : set of definitions that S newly assigns values to and not subsequently killed by itself Transfer function: f S (in[s]) = gen[s] (in[s] kill[s]) 8

9 Computing local reaching definitions v_s 1 x_s 7 y_s 9 in[b] S 10 x := y+4 S 11 y := y+v S 12 z := x*v in[s 10 ] = in[b]={v_s 1, x_s 7, y_s 9 } in[s 11 ] = out[s 10 ]={v_s 1, x_s 10, y_s 9 } kill[s 10 ] = {x_*}, gen[s 10 ]={x_s 10 } kill[s 11 ] = {y_*}, gen[s 11 ]={y_s 11 } in[s 12 ] = out[s 11 ]={v_s 1, x_s 10, y_s 11 } kill[s 12 ] = {z_*}, gen[s 12 ]={z_s 12 } in[s 13 ] = out[s 12 ]={v_s 1, x_s 10, y_s 11, z_s 12 } S 13 x := x-z kill[s 13 ] = {x_*}, gen[s 13 ]={x_s 13 } out[b] v_s 1 x_s 13 y_s 11 z_s 12 in[s 14 ] = out[s 13 ]={v_s 1, x_s 13, y_s 11, z_s 12 } S 14 goto L on z>0 kill[s 14 ] = {}, gen[s 14 ]={} out[b] = out[s 14 ]={v_s 1, x_s 13, y_s 11,z_S 12 } 9

10 Global Data Flow Analysis Transfer function out[x] = f X (in[x]) in[x] : DFI given to the beginning of X out[x] : output DFI at the end of X, DFI CFG Collects outputs of predecessors in[x] in[x] = out[p] PPred[X] A typical form of a transfer function f X (in[x]) = gen[x] (in[x] kill[x]) gen[x]: DFI newly generated within X f X gen[x] kill[x] out[x] X kill[x]: DFI that was originally in in[x], but regenerated within X 10

11 Example: Global Reaching Definition DFE components for each basic block B in[b] : set of definitions that may reach the beginning of B out[b] : set of definitions that may reach the end of B kill[b] : set of definitions that reach B but are killed by being assigned values in it gen[b] : set of definitions that are newly assigned values in B and not subsequently killed in it Transfer function: f B (in[b]) = gen[b] (in[b] kill[b]) 11

12 Computing Reaching Definition B 0 S 0... B 4 S 50 x := y+4 S 51 y := y+v S 52 z := x*v S 53 x := x-z S 54 goto L on z>0 v_s 12 x_s 27 y_s 39 in[b 4 ] kill[b 4 ] x_* y_* z_* z_s 52 y_s 51 x_s 53 gen[b4 ] v_s 12 x_s 53 y_s 51 out[b 4 ] = gen[b 4 ] (in[b 4 ] kill[b 4 ]) B 8 z_s 52 B 6 S 70 L: t := z+x S 71 s := y*v... S

13 Propagating DFI thru Control Flow Propagate DFI to a join node A node u with multiple predecessors p 1,,p n Need to specify how to combine the DFI propagated from all its predecessors Define a binary meet (or called confluence) operation Find greatest lower bound (glb) on lattice in[u] = out[p 1 ] out[p 2 ] out[p n ] = out[p i ] i = 1,n in[y] initial values y v in[u]=out[v] out[w] u x w out[u]=f u (in[u]) 13

14 Propagating Reaching Definitions The meet operation in the reaching definitions problem is union. B 3 S 40 read(x) S 41 v := z*v S 42 goto L on x=0 x_s 40 z_s 30 v_s 41 out[b 3 ] in[b 8 ] x_s 22 z_s 30 v_s 34 v_s 12, v_s 41 x_s 40, x_s 53 z_s 30, z_s 52 y_s 51 in[b 3 ] B 4 v_s 12 x_s 27 y_s 39 S 50 x := y+4 S 51 y := y+v S 52 z := x*v S 53 x := x-z S 54 goto L on z>0 v_s 12 x_s 53 y_s 51 z_s 52 out[b 4 ] S 0... in[b 8 ] = out[b 3 ] out[b 4 ] in[b 4 ] B 0 in[b 6 ] = out[b 4 ] in[b 6 ] v_s 12 x_s 53 y_s 51 z_s 52 B B 6 8 S 70 L: t := z+x S S s := y*v... 14

15 Iterative Data Flow Analysis Given G=(N,E,S), forward data-flow problem s DFE in[b] = Init (Init = Φ for Reaching Definition) out[b] = f B ( in[b] ) = gen[b] ( in[b] kill[b] ) A typical algorithm for iterative data flow analysis for each program segment X do initialize out[x]; od while changes to any out[x] occur do for each segment X in DFS order do in[x] = out[p X ]; all predecessors P X of X out[x] = gen[x](in[x]-kill[x]); od od 15

16 Unified DFA Framework Most data flow problems need DFI transfer function (or flow function) meet operation A unified framework (V,F,) V : domain of DFI values F: V V : set of f B for each block B : V V V a confluence (or called meet) operator 16

17 Basic Properties of Lattice Lattice, L a poset (partially ordered set) (L, ) is a lattice, if For any two elements x and y of L, glb and lub exist within L Two operations for lattice values meet () : greatest lower bound (glb) join () : least upper bound (lub) For all u,v,w L, commutativity : u v = v u, u v = v u associativity : (u v) w = u (v w), (u v) w = u (v w) idempotence : u u = u, u u = u closure: exist unique x, y L, s.t. u v = x and u v = y Two unique elements bottom () : for all u L, u = top (T) : for all u L, u T = T 17

18 A Lattice For Reaching Definitions Framework for reaching definition problem Suppose a program has three statements S 1,S 2, S 3. V RD = { v v {S 1,S 2,S 3 } } RD = T {S 1 } {S 2 } {S 3 } {S 1,S 2 } {S 1,S 3 } {S 2, S 3 } {S 1, S 2, S 3 } 18

19 Summary Data flow analysis on CFG Data flow information Data flow equation in[b] = out[p] PPred[B] out[b] = f B ( in[b] ) = gen[b] ( in[b] kill[b] ) Define a meet operation for join points Data flow analysis Local DFA: straight line code Global DFA: iterative DFA algorithm Unified DFA on lattice 19