CS553 Lecture Reuse Optimization: Common SubExpr Elim 3

Transcription

1 Reuse Optimizatio Last time Value umberig Today Commo subexpressio elimiatio (CSE) CS553 Lecture Reuse Optimizatio: Commo SubExpr Elim 2 Commo Subexpressio Elimiatio Fid commo subexpressios whose rage spas the same basic blocks ad elimiate uecessary re-evaluatios Leverage available expressios Recall available expressios A expressio (e.g., x+y) is available at ode if every path from the etry ode to evaluates x+y, ad there are o defiitios of x or y after the last evaluatio alog that path Strategy If a expressio is available at a poit where it is evaluated, it eed ot be recomputed CS553 Lecture Reuse Optimizatio: Commo SubExpr Elim 3

2 CSE 2 i := i + b := 4 * i i := j a := 4 * i 3 c := 4 * i 2 i := i + t := 4 * i b := t i := j t := 4 * i a := t 3 c := t Will value umberig fid this redudacy? No; value umberig operates o values CSE operates o expressios CS553 Lecture Reuse Optimizatio: Commo SubExpr Elim 4 Aother CSE Before CSE c := a + b d := m & e := b + d f := a + b g := -b h := b + a a := j + a k := m & j := b + d a := -b if m & goto L2 istructios 2 variables 9 biary operators After CSE t := a + b c := t t2 := m & d := t2 t3 := b + d e := t3 f := t g := -b h := t a := j + a k := t2 j := t3 a := -b if t2 goto L2 4 istructios 5 variables 4 biary operators CS553 Lecture Reuse Optimizatio: Commo SubExpr Elim 5 2

3 CSE Approach Notatio Avail(b) is the set of expressios available at block b Ge(b) is the set of expressios geerated ad ot killed at block b If we use e ad e Avail(b) Allocate a ew ame Search backward from b (i CFG) to fid statemets (oe for each path) that most recetly geerate e Isert copy to after geerators Problems Backward search for each use is expesive Geerates uique ame for each use ames Uses > Avail Each geerator may have may copies t := a t2 := a e := t b + c f := t2 b + c CS553 Lecture Reuse Optimizatio: Commo SubExpr Elim 6 CSE Approach 2 Reduce umber of copies by assigig a uique ame to each uique expressio e Name[e] = uassiged if we use e ad e Avail(b) if Name[e]=uassiged, allocate ew ame ad Name[e] = else = Name[e] I a subsequet traversal of block b, if e Ge(b) ad Name[e] uassiged, the isert a copy to Name[e] after the geerator of e Problem May still isert uecessary copies Requires two passes over the code t := a CS553 Lecture Reuse Optimizatio: Commo SubExpr Elim 7 3

4 CSE Approach 3 Do t worry about temporaries Create oe temporary for each uique expressio Let subsequet pass elimiate uecessary temporaries At a evaluatio of e Hash e to a ame,, i a table Isert a assigmet of e to At a use of e i b, if e Avail(b) Lookup e s ame i the hash table (call this ame ) Problems Iserts more copies tha approach 2 (but extra copies are dead) Still requires two passes (2 d pass is very geeral) CS553 Lecture Reuse Optimizatio: Commo SubExpr Elim 8 Extraeous Copies Extraeous copies degrade performace Let other trasformatios deal with them Dead code elimiatio Coalescig Coalesce assigmets to t ad t2 ito a sigle statemet t := b + c t2 := t Greatly simplifies CSE CS553 Lecture Reuse Optimizatio: Commo SubExpr Elim 9 4

5 Partial Redudacy Elimiatio (PRE) Partial Redudacy A expressio (e.g., x+y) is partially redudat at ode if some path from the etry ode to evaluates x+y, ad there are o defiitios of x or y betwee the last evaluatio of x+y ad Elimiatio Discover partially redudat expressios Covert them to fully redudat expressios Remove redudacy PRE subsumes CSE ad loop ivariat code motio CS553 Lecture Reuse Optimizatio: Commo SubExpr Elim 0 Loop Ivariace PRE removes loop ivariats A ivariat expressio is partially redudat PRE coverts this partial redudacy to full redudacy PRE removes the redudacy x := y * z x := y * z x := y * z CS553 Lecture Reuse Optimizatio: Commo SubExpr Elim 5