Bottom-up parsing

Goal of parser : build a derivation Top-down parser : build a derivation by working from the start symbol towards the input. Builds parse tree from root to leaves Builds leftmost derivation Bottom-up parser : build a derivation by working from the input back toward the start symbol Builds parse tree from leaves to root Builds reverse rightmost derivation

Bottom-up parsing
The parser looks for a substring of the parse tree's

frontier...

...that matches the rhs of a production and ...whose reduction to the non-terminal on the lhs represents on step along the reverse of a rightmost derivation

Such a substring is called a handle.

Important: Not all substrings that match a rhs are handles.

Bottom-up parsing techniques

Shift-reduce parsing Shift input symbols until a handle is found. Then, reduce the substring to the non-terminal on the lhs of the corresponding production. Operator-precedence parsing Based on shift-reduce parsing. Identifies handles based on precedence rules.

Example: Shift-reduce parsing

STACK ACTION Shift Reduce (rule 5)

Grammar: 1. 2. 3. 4. 5. S E E E E E E+E E*E num id

$ $ id1

$E
$E+

Shift
Shift

$ E + num
$E+E $E+E* $ E + E * id2 $E+E*E $E+E $E $S

Reduce (rule 4)
Shift Shift Reduce (rule 5) Reduce (rule 3) Reduce (rule 2) Reduce (rule 1) Accept
4

Input to parse: id1 + num * id2 Handles:

underlined

Shift-Reduce parsing
A shift-reduce parser has 4 actions: Shift -- next input symbol is shifted onto the stack Reduce -- handle is at top of stack pop handle push appropriate lhs Accept -- stop parsing & report success Error -- call error reporting/recovery routine

Shift-Reduce parsing
How can we know when we have found a handle? Analyze the grammar beforehand. Build tables Look ahead in the input LR(1) parsers recognize precisely those languages in

which one symbol of look-ahead is enough to determine whether to reduce or shift.

L : for left-to-right parse of the input R : for reverse rightmost derivation 1: for one symbol of lookahead

How does it work?

Read input, one token at a time

Use stack to keep track of current state The state at the top of the stack summarizes the information below. The stack contains information about what has been parsed so far.
Use parsing table to determine action based on

current state and look-ahead symbol. How do we build a parsing table?

LR parsing techniques
SLR (not in the book) Simple LR parsing Easy to implement, not strong enough Uses LR(0) items Canonical LR Larger parser but powerful Uses LR(1) items LALR (not in the book) Condensed version of canonical LR May introduce conflicts Uses LR(1) items
8

Class examples
E' E E E+T S' S S L=R

E
T T

T
T*F F

S
L L

R
*R id

id

Finding handles
As a shift/reduce parser processes the input, it must

keep track of all potential handles. For example, consider the usual expression grammar and the input string x+y.

Suppose the parser has processed x and reduced it to E. Then, the current state can be represented by E +E where means that an E has already been parsed and that +E is a potential suffix, which, if found, will result in a successful parse. Our goal is to eventually reach state E+E, which represents an actual handle and should result in the reduction EE+E
10

LR parsing
Typically, LR parsing works by building an automaton where

each state represents what has been parsed so far and what we hope to parse in the future.

In other words, states contain productions with dots, as described

earlier. Such productions are called items

States containing handles (meaning the dot is all the way to the

right end of the production) lead to actual reductions depending on the lookahead.

11

SLR parsing
SLR parsers build automata where states contain

items (a.k.a. LR(0) items) and reductions are decided based on FOLLOW set information. We will build an SLR table for the augmented grammar
S'S S L=R SR L *R L id RL
12

SLR parsing
When parsing begins, we have not parsed any input at all and

we hope to parse an S. This is represented by S'S.

Note that in order to parse that S, we must either parse an L=R or an R. This is represented by SL=R and SR

closure of a state: if AaBb represents the current state and B is a production, then add B to the state. Justification: aBb means that we hope to see a B next. But parsing a B is equivalent to parsing a , so we can say that we hope to see a next

13

SLR parsing
Use the closure operation to define states containing

LR(0) items. The first state will be:

S' S S L=R SR L *R L id RL

From this state, if we parse, say, an id, then we go to L id state If, after some steps we parse input that reduces to S L =R an L, then we go to state R L

14

SLR parsing
Continuing the same way, we define all LR(0) item

states:

I0

S' S S L=R SR L *R L id RL id R L id

I3

R I6 S L= R S L=R RL I9 id L *R L I3 S L =R = L id I2 R L L * * L * R I5 R L L R L I7 L id R id L * R L *R I8
S' S *
15

I1

I4 S R

SLR parsing
The automaton and the FOLLOW sets tell us how to build the

parsing table: Shift actions If from state i, you can go to state j when parsing a token t, then slot [i,t] of the table should contain action "shift and go to state j", written sj Reduce actions If a state i contains a handle A, then slot [i, t] of the table should contain action "reduce using A", for all tokens t that are in FOLLOW (A). This is written r(A)

The reasoning is that if the lookahead is a symbol that may follow A, then a reduction A should lead closer to a successful parse.
16

continued on next slide

SLR parsing
The automaton and the FOLLOW sets tell us how to build the

parsing table: Reduce actions, continued Transitions on non-terminals represent several steps together that have resulted in a reduction. For example, if we are in state 0 and parse a bit of input that ends up being reduced to an L, then we should go to state 2. Such actions are recorded in a separate part of the parsing table, called the GOTO part.

17

SLR parsing
Before we can build the parsing table, we need to compute the

FOLLOW sets:

S' S S L=R S R L *R L id R L

FOLLOW(S') = {$} FOLLOW(S) = {$} FOLLOW(L) = {$, =} FOLLOW(R) = {$, =}

18

SLR parsing
state 0 1 2 3 4 5 6 7 8 9 id s3 = s6/r(RL) r(Lid) s3 s3 s5 s5 action * s5 $ accept r(Lid) r(SR) r(RL) r(L*R) r(SL=R) goto S L R 1 2 4

r(RL) r(L*R)

7 7

8 9

Note the shift/reduce conflict on state 2 when the lookahead is an =

19

Conflicts in LR parsing
There are two types of conflicts in LR parsing: shift/reduce On some particular lookahead it is possible to shift or reduce The if/else ambiguity would give rise to a shift/reduce conflict reduce/reduce This occurs when a state contains more than one handle that may be reduced on the same lookahead.

20

Conflicts in SLR parsing

The parser we built has a shift/reduce conflict.

Does that mean that the original grammar was

ambiguous? Not necessarily. Let's examine the conflict:

it seems to occur when we have parsed an L and are seeing an =. A reduce at that point would turn the L into an R. However, note that a reduction at that point would never actually lead to a successful parse. In practice, L should only be reduced to an R when the lookahead is EOF ($). An easy way to understand this is by considering that L represents l-values while R represents r-values.

21

Conflicts in SLR parsing

The conflict occurred because we made a decision about when

to reduce based on what token may follow a non-terminal at any time. However, the fact that a token t may follow a non-terminal N in some derivation does not necessarily imply that t will follow N in some other derivation. SLR parsing does not make a distinction.

22

Conflicts in SLR parsing

SLR parsing is weak.

Solution : instead of using general FOLLOW

information, try to keep track of exactly what tokens many follow a non-terminal in each possible derivation and perform reductions based on that knowledge. Save this information in the states. This gives rise to LR(1) items:

items where we also save the possible lookaheads.

23

Canonical LR(1) parsing

In the beginning, all we know is that we have not

read any input (S'S), we hope to parse an S and after that we should expect to see a $ as lookahead. We write this as: S'S, $ Now, consider a general item A, x. It means that we have parsed an , we hope to parse and after those we should expect an x. Recall that if there is a production , we should add to the state. What kind of lookahead should we expect to see after we have parsed ?

We should expect to see whatever starts a . If is empty or can vanish, then we should expect to see an x after we have parsed (and reduced it to B)
24

Canonical LR(1) parsing

The closure function for LR(1) items is then defined

as follows:

For each item A, x in state I, each production in the grammar, and each terminal b in FIRST(x), add , b to I If a state contains core item with multiple possible lookaheads b1, b2,..., we write , b1/b2 as shorthand for , b1 and , b2
25

Canonical LR(1) parsing

I6 S' S , $ I0 S' S, $ S L=R, $ L S R, $ S L =R, $ = L *R, =/$ I2 R L , $ L id, =/$ * R L, $ L *R, =/$ id R I5 R L, =/$ I5' L id, =/$ I3 L id , =/$ id L *R, =/$ L R * I4 S R, =/$ I8 L *R , =/$ S I1 S L= R, $ R SL=R, $ R L, $ id L *R, $ Lid, $ I3' L id, $ * L R L, $ I7' L *R, $ L id I3' R L, $ L id, $ L *R , $ L *R, $ R I8' * I9

R L, =/$ I7
26

Canonical LR(1) parsing

The table is created in the same way as SLR, except

we now use the possible lookahead tokens saved in each state, instead of the FOLLOW sets. Note that the conflict that had appeared in the SLR parser is now gone. However, the LR(1) parser has many more states. This is not very practical.

It may be possible to merge states!

27

LALR(1) parsing
This is the result of an effort to reduce the number of

states in an LR(1) parser. We notice that some states in our LR(1) automaton have the same core items and differ only in the possible lookahead information. Furthermore, their transitions are similar.

States I3 and I3', I5 and I5', I7 and I7', I8 and I8'

We shrink our parser by merging such states. SLR : 10 states, LR(1): 14 states, LALR(1) : 10 states

28

LALR(1) parsing
I6 S' S , $ I0 S' S, $ S L=R, $ L S R, $ S L =R, $ = L *R, =/$ I2 R L , $ L id, =/$ * R L, $ L *R, =/$ id R I5 R L, =/$ L id, =/$ I3 L id , =/$ id L *R, =/$ L I4 S R, =/$ S I1 S L= R, $ R SL=R, $ R L, $ id I L *R, $ 3 L id, $ * L I9

R L, =/$ I7

I8 L *R , =/$
29

Conflicts in LALR(1) parsing

Note that the conflict that had vanished when we

created the LR(1) parser has not reappeared. Can LALR(1) parsers introduce conflicts that did not exist in the LR(1) parser? Unfortunately YES. BUT, only reduce/reduce conflicts.

30

Conflicts in LALR(1) parsing

LALR(1) parsers cannot introduce shift/reduce conflicts.

Such conflicts are caused when a lookahead is the same as a token on which we can shift. They depend on the core of the item. But we only merge states that had the same core to begin with. The only way for an LALR(1) parser to have a shift/reduce conflict is if one existed already in the LR(1) parser. LALR(1) parsers can introduce reduce/reduce conflicts. Here's a situation when this might happen:

A B , x A C , y

merges with

A B,y to give: A C , x

A B , x/y A C , x/y

31