1 Introduction

Programming is a humbling job which requires acknowledging that we will make untold errors in our quest to perfect a program. Most troubling are semantic errors, where we intended the program to do one thing, but it does another. Less troubling, but often no less irritating, are syntax errors, which are generally minor deviances from the exacting syntax required by a compiler. So common are syntax errors that parsers in modern compilers are designed to cope with us making several: rather than stop on the first syntax error, they attempt to recover from it. This allows them to report, and us to fix, all our syntax errors in one go.

When error recovery works well, it is a useful productivity gain. Unfortunately, most current error recovery approaches are simplistic. The most common grammar-neutral approach to error recovery are those algorithms described as ‘panic mode’ (e.g. [13, p. 348]) which skip input until the parser finds something it is able to parse. A more grammar-specific variation of this idea is to skip input until a pre-determined synchronisation token (e.g. ‘;’ in Java) is reached [8 , p. 3], or to try inserting a single synchronisation token. Such strategies are often unsuccessful, leading to a cascade of spurious syntax errors (see Figure 1 for an example). Programmers quickly learn that only the location of the first error in a file – not the reported repair, nor the location of subsequent errors – can be relied upon to be accurate.

---

---

It is possible to hand-craft error recovery algorithms for a specific language. These generally allow better recovery from errors, but are challenging to create. For example, the Java error recovery approach in the Eclipse IDE is 5KLoC long, making it only slightly smaller than a modern version of Berkeley Yacc — a complete parsing system! Unsurprisingly, few real-world parsers contain effective hand-written error recovery algorithms.

Most of us are so used to these trade-offs (cheap generic algorithms and poor recovery vs. expensive hand-written algorithms and reasonable recovery) that we assume them to be inevitable. However, there is a long line of work on more advanced generic error recovery algorithms. Probably the earliest such algorithm is Aho and Peterson [1], which, upon encountering an error, creates on-the-fly an alternative (possibly ambiguous) grammar which allows the parser to recover. This algorithm has fallen out of favour in programming language circles, probably because of its implementation complexity and the difficulty of explaining to users what recovery has been used. A simpler family of algorithms, which trace their roots to Fischer et al. [11], instead try to find a single minimum cost repair sequence of token insertions and deletions which allow the parser to recover. Algorithms in this family are much better at recovering from errors than naive approaches and can communicate the repairs they find in a way that humans can easily replicate. However, such algorithms have seen little practical use because their typical performance is seen as poor and their worst case unbounded [17, p. 14]. We add a further complaint: such approaches only report a single repair sequence to users. In general – and especially in syntactically rich languages – there are multiple reasonable repair sequences for a given error location, and the algorithm has no way of knowing which best matches the user’s intentions.

In this paper we introduce a new error recovery algorithm in the Fischer et al. family, CPCT+. This takes the approach of Corchuelo et al. [5] as a base, corrects it, expands it, and optimises its implementation. CPCT+ is simple to implement (under 500 lines of Rust code), is able to repair nearly all errors in reasonable time, reports the complete set of minimum cost repair sequences to users, and does so in less time than Corchuelo et al..

We validate CPCT+ on a corpus of 200,000 real, syntactically incorrect, Java programs (Section 6). CPCT+ is able to recover 98.37%±0.017% of files within a 0.5s timeout and does so while reporting fewer than half the error locations as a traditional panic mode algorithm: in other words, CPCT+ substantially reduces the cascading error problem. We also show – for, as far as we know, the first time – that advanced error recovery can be naturally added to a Yacc-esque system, allowing users to make fine-grained decisions about what to do when error recovery has been applied to an input (Section 7). We believe that this shows that algorithms such as CPCT+ are ready for wider usage, either on their own, or as part of a multi-phase recovery system.

1.1 Defining the problem

Formally speaking, we first test the following hypothesis:

H1

The complete set of minimum cost repair sequences can be found in acceptable time.

The only work we are aware of with a similar concept of ‘acceptable time’ is [6], who define it as the total time spent in error recovery per file, with a threshold of 1s. We use that definition with one change: Since many compilers are able to fully execute in less than 1s, we felt that a tighter threshold is more appropriate: we use 0.5s since we think that even the most demanding users will tolerate such a delay. We strongly validate this hypothesis.

The complete set of minimum cost repair sequences makes it more likely that the programmer will see a repair sequence that matches their original intention (see Figure 1 for an example; Appendix A contains further examples in Java, Lua, and PHP). It also opens up a new opportunity. Previous error recovery algorithms find a single repair sequence, apply that to the input, and then continue parsing. While that repair sequence may have been a reasonable local choice, it may cause cascading errors later. Since we have the complete set of minimum cost repair sequences available, we can select from that set a repair sequence which causes fewer cascading errors. We thus rank repair sequences by how far they allow parsing to continue successfully (up to a threshold — parsing the whole file would, in general, be too costly), and choose from the subset that gets furthest (note that the time required to do this is included in the 0.5s timeout). We thus also test a second hypothesis:

H2

Ranking the complete set of minimum cost repair sequences by how far they allow parsing to continue locally reduces the global cascading error problem.

We also strongly validate this hypothesis. We do this by comparing ‘normal’ CPCT+ with a simple variant CPCT+rev which reverses the ranking process, always selecting from amongst the worst performing minimum cost repair sequence. CPCT+rev models the worst case of previous approaches in the Fischer et al. family, which non-deterministically select a single minimum cost repair sequence. CPCT+rev leads to 31.93%±0.289% more errors being reported (i.e. it substantially worsens the cascading error problem).

This paper is structured as follows. We describe the Corchuelo et al. algorithm (Section 4), filling in missing details from the original description and correcting its definition. We then expand the algorithm into CPCT+ (Section 5). Finally, we validate CPCT+ on a corpus of 200,000 real, syntactically incorrect, Java programs comparing it to implementations of panic mode and Corchuelo et al. (Section 6). To emphasise that our algorithms are grammar-neutral, we show examples of error recovery on different grammars in Appendix A.

2 Background

We assume a high-level understanding of the mechanics of parsing in this paper, but in this section we provide a handful of definitions, and a brief refresher of relevant low-level details, needed to understand the rest of this paper. Although the parsing tool we created for this paper is written in Rust, we appreciate that this is still an unfamiliar language to many readers: algorithms are therefore given in Python which, we hope, is familiar to most.

Although there are many flavours of parsing, the Fischer et al. family of error recovery algorithms are designed to be used with LR(k) parsers [16]. LR parsing remains one of the most widely used parsing approaches due to the ubiquity of Yacc [14] and its descendants (which include the Rust parsing tool we created for this paper). We use Yacc syntax throughout this paper so that examples can easily be tested in Yacc-compatible parsing tools.

Yacc-like tools take in a Context-Free Grammar (CFG) and produce a parser from it. The CFG has one or more rules; each rule has a name and one or more productions (often called ‘alternatives’); each production contains one or more symbols; and a symbol references either a token type or a grammar rule. One rule is designated the start rule. The resulting parser takes as input a stream of tokens, each of which has a type (e.g. INT) and a value (e.g. 123) – we assume the existence of a Lex-like tool which can split a string into a stream of tokens. Strictly speaking, parsing is the act of determining whether a stream of tokens is correct with respect to the underlying grammar. Since this is rarely useful on its own, Yacc-like tools allow grammars to specify ‘semantic actions’ which are executed when a production in the grammar is successfully matched. Except where stated otherwise, we assume that the semantic actions build a parse tree, ordering the tokens into a tree of nonterminal nodes (which can have children) and terminal nodes (which cannot have children).

---

---

The CFG is first transformed into a stategraph, a statemachine where each node contains one or more items (describing the valid parse states at that point) and edges are labelled with terminals or nonterminals. Since even on a modern machine, a canonical (i.e. unmerged) LR stategraph can take several seconds to build, and a surprising amount of memory to store, we use the algorithm of [21] to merge together compatible states¹ . The effect of this is significant, reducing the Java grammar we use later from 8908 to 1148 states. The stategraph is then transformed into a statetable with one row per state. Each row has a possibly empty action (shift, reduce, or accept) for each terminal and a possibly empty goto state for each nonterminal. Figure 2 shows an example grammar, its stategraph, and statetable.

The statetable allows us to define a simple, efficient, parsing process. We first define two functions relative to the statetable: action(s,t) returns the action for the state s and token t or error if no such action exists; and goto(s,N) returns the goto state for the state s and the nonterminal N or error if no such goto state exists. We then define a reduction relation →LR for (parsing stack,token list) pairs with two reduction rules as shown in Figure 3. A full LR parse →*LR repeatedly applies the two →LR rules until neither applies, which means that action(s_n,t₀) is either: accept (i.e. the input has been fully parsed); or error(i.e. an error has been detected at the terminal t₀). A full parse takes a starting pair of ([0], [t₀ … t_n,$]), where state 0 is expected to represent the entry point into the stategraph, t₀ … t_n is the sequence of input tokens, and ‘$’ is the special End-Of-File (EOF) token.

---

---

3 Panic mode

Error recovery algorithms are invoked by a parser when it has yet to finish but there is no apparent way to continue parsing (i.e. when action(s_n,t₀) = error). Error recovery algorithms are thus called with a parsing stack and a sequence of remaining input (which we represent as a list of tokens): they can modify either or both of the parsing stack and the input in their quest to get parsing back on track. The differences between algorithms are thus in what modifications they can carry out (e.g. altering the parse stack; deleting input; inserting input), and how they carry such modifications out.

The simplest grammar-neutral error recovery algorithms are widely called ‘panic mode’ algorithms (the origin of this family of algorithms seems lost in time). While there are several members of this family for LL parsing, there is only one fundamental way of creating a grammar-neutral panic mode algorithm for LR parsing. We take our formulation from Holub [13, p. 348]² . The algorithm works by popping elements off the parsing stack to see if an earlier part of the stack is able to parse the next input symbol. If no element in the stack is capable of parsing the next input symbol, the next input symbol is skipped, the stack restored, and the process repeated. At worst, this algorithm guarantees to find a match at the EOF token. Figure 4 shows a more formal version of this algorithm.

---

---

The advantage of this algorithm is its simplicity and speed. For example, consider the grammar from Figure 2 and the input ‘2 + + 3’. The parser encounters an error on the second ‘+’ token, leaving it with a parsing stack of [0, 2, 7] and the input ‘+ 3’ remaining. The error recovery algorithm now starts. It first tries action(7, ‘+’) which (by definition, since it is the place the parser encountered an error) returns error; it then pops the top element from the parsing stack and tries action(2, ‘+’), which returns shift. This is enough for the error recovery algorithm to complete, and parsing resumes with a stack [0, 2].

The fundamental problem with error recovery can be seen from the above example: by popping from the parsing stack, it implicitly deletes input from before the error location (in this case the first ‘+’) in order to find a way of parsing input after the error location. This often leads to panic mode throwing away huge portions of the input in its quest to find a repair. Not only can the resulting recovery appear as a Deus ex machina, but the more input that is skipped, the more likely that a cascade of further parsing errors ensues (as we will see later in Section 6.2).

4 Corchuelo et al.

There have been many attempts to create better LR error recovery algorithms than panic mode. Most numerous are those error recovery algorithms in what we call the Fischer et al. family. Indeed, there are far too many members of this family of algorithms to cover in one paper. We therefore start with one of the most recent – Corchuelo et al. [5]. We first explain the original algorithm (Section 4.1), although we use different notation than the original, fill in several missing details, and provide a more formal definition. We then make two correctness fixes to ensure that the algorithm always finds minimum cost repair sequences (Section 4.2). Since the original gives few details as to how the algorithm might best be implemented, we then explain our approach to making a fast implementation (Section 4.3).

4.1 The original algorithm

Intuitively, the Corchuelo et al. algorithm starts at the error state and tries to find a minimum cost repair sequence consisting of: insert T (‘insert a token of type T’), delete (‘delete the token at the current offset’), or shift (‘parse the token at the current offset’). The algorithm completes: successfully if it reaches an accept state or shifts ‘enough’ tokens (N_shifts, set at 3 in Corchuelo et al.); or unsuccessfully if a repair sequence contains too many delete or insert repairs (set at 3 and 4 respectively in Corchuelo et al.) or spans ‘too much’ input (N_total, set at 10 in Corchuelo et al.). Repair sequences are reported back to users with trailing shift repairs pruned i.e. [insert x, shift y, delete z, shift a, shift b, shift c] is reported as [insert x, shift y, delete z].

In order to find repair sequences, the algorithm keeps a breadth-first queue of configurations, each of which represents a different search state; configurations are queried for their neighbours which are put into the queue; and the search terminates when a successful configuration is found. The cost of a configuration is the sum of the repair costs in its repair sequence. By definition, a configuration’s neighbours have the same, or greater, cost to it.

---

---

---

---

As with the original, we explain the approach in two parts. First is a new reduction relation →CR which defines a configuration’s neighbours (Figure 5). Second is an algorithm which makes use of the →CR relation to generate neighbours, and determines when a successful configuration has been found or if error recovery has failed (Figure 6). As well as several changes for clarity, the biggest difference is that Figure 6 captures semi-formally what Corchuelo et al. explain in prose (spread amongst several topics over several pages): perhaps inevitably we have had to fill in several missing details. For example, Corchuelo et al. do not define what the cost of repairs is: for simplicities sake, we define the cost of insert and delete as 1, and shift as 0.³

4.2 Ensuring that minimum cost repair sequences aren’t missed

CR Shift 1 (see Figure 5) has two flaws which prevent it from generating all possible minimum cost repair sequences.

---

---

---

---

First, CR Shift 1 always consumes input, missing intermediate configurations (including accept states!) that only require reductions/gotos to be performed. CR Shift 2 in Figure 7 shows the two-phase fix which addresses this problem. We first change the condition 0 < j ≤N_shifts to 0 ≤j ≤N_shifts so that the parser can make progress without consuming input. However, this opens the possibility of an infinite loop, so we then add a condition that if no input is consumed, the parsing stack must have changed. In other words, we require progress to be made, whether or not that progress involved consuming input.

Second, CR Shift 1 and CR Shift 2 generate multiple shifts at a time. This causes them to skip intermediate configurations from which minimum cost repair sequences may be found. The solution⁴ is simple: at most one shift can be generated at any one time. CR Shift 3 in Figure 7 (as well as incorporating the fix from CR Shift 2) generates at most one shift repair at a time. Relative to CR Shift 1, it is simpler, though it also inevitably slows down the search, as more configurations are generated.

The problems with CR Shift 1, in particular, can be severe. Figure 8 shows an example input where CR Shift 1 is unable to find any repair sequences, CR Shift 2 some, and CR Shift 3 all minimum cost repair sequences.

4.3 Implementation considerations

The definitions we have given thus far do not obviously lead to an efficient implementation and Corchuelo et al. give few useful hints. We found that two techniques were both effective at improving performance while being simple to implement.

First, Corchuelo et al. suggest using a breadth-first search but give no further details. It was clear to us that the most natural way to model the search is as an instance of Dijkstra’s algorithm. However, rather than use a general queue data-structure (probably based on a tree) to discover which element to search next, we use a similar queue data-structure to [4, p. 25]. This consists of one sub-list per cost (i.e. the first sub-list contains configurations of cost 0, the second sub-list configurations of cost 1 and so on). Since we always know what cost we are currently investigating, finding the next todo element requires only a single pop (line 8 of Figure 6). Similarly, adding elements requires only an append to the relevant sub-list (lines 18, 21, 22). This data-structure is a good fit because costs in our setting are always small (double digits is unusual for real-world grammars) and each neighbour generated from a configuration with cost c has a cost ≥ c.

Second, since error recovery frequently adjusts and resets parsing stacks and repair sequences, during error recovery we do not represent these as lists (as is the case during normal parsing). We found that lists consume noticeably more memory, and are slightly less efficient, than using parent pointer trees (often called ‘cactuses’). Every node in such a tree has a reference to a single parent (or null for the root node) but no references to child nodes. Since our implementation is written in Rust – a language without garbage collection – nodes are reference counted (i.e. a parent is only freed when it is not in a todo list and no children point to it). When the error recovery algorithm starts, it converts the main parsing stack (a list) into a parent pointer tree; and repair sequences start as empty parent pointer trees. The →CR part of our implementation thus operates exclusively on parent pointer trees. Although this does mean that neighbouring configurations are scattered throughout memory, the memory sharing involved seems to more than compensate for poor cache behaviour; it also seems to be a good fit with modern malloc implementations, which are particularly efficient when allocating and freeing objects of the same size. This representation is likely to be a reasonable choice in most contexts, although it is difficult to know from our experience whether it will always be the best choice (e.g for garbage collected languages).

One seemingly obvious improvement is to split the search into parallel threads. However, we found that the nature of the problem means that parallelisation is more tricky, and less productive, than might be expected. There are two related problems: we cannot tell in advance if a given configuration will have huge numbers of successors or none at all; and configurations are, in general, searched for successors extremely quickly. Thus if we attempt to seed threads with initial sets of configurations, some threads quickly run out of work whilst others have ever growing queues. If, alternatively, we have a single global queue then significant amounts of time can be spent adding or removing configurations in a thread-safe manner. A work stealing algorithm might solve this problem but, as we shall see in Section 6, CPCT+ runs fast enough that the additional complexity of such an approach is not, in our opinion, justified.

5 CPCT+

In this section, we extend the Corchuelo et al. algorithm to become what we call CPCT+. First we enhance the algorithm to find the complete set of minimum cost repair sequences (Section 5.1). Since this slows down the search, we optimise by merging together compatible configurations (Section 5.2). The complete set of minimum cost repair sequences allows us to make an algorithm less susceptible to the cascading error problem (Section 5.3). We then change the criteria for terminating error recovery (Section 5.4).

5.1 Finding the complete set of minimum cost repair sequences

The basic Corchuelo et al. algorithm non-deterministically completes as soon as it has found a single minimum cost repair sequence. This is confusing in two different ways: the successful repair sequence found can vary from run to run; and the successful repair sequence might not match the user’s intention.

We therefore introduce the idea of the complete set of minimum cost repair sequences: that is, all equivalently good repair sequences. Although we will refine the concept of ‘equivalently good’ in Section 5.3, at this stage we consider all successful repair sequences with minimum cost c to be equivalently good. In other words, as soon as we find the first successful repair sequence, its cost c defines the minimum cost.

An algorithm to generate this set is then simple: when a repair sequence of cost c is found to be successful, we discard all repair sequences with cost > c, and continue exploring configurations in cost c (including, transitively, all neighbours that are also of cost c; those with cost > c are immediately discarded). Each successful configuration is recorded and, when all configurations in c have been explored, the set of successful configurations is returned. One of these successful configurations is then non-deterministically chosen, applied to the input, and parsing continued.

5.2 Merging compatible configurations

Relative to finding a single solution, finding the complete set of repair sequences can be extremely expensive because there may be many remaining configurations in c, which may, transitively, have many neighbours. Our solution to this performance problem is to merge together compatible configurations on-the-fly, preserving their distinct repair sequences while reducing the search space. Two configurations are compatible if:

1.

their parsing stacks are identical,

2.

they both have an identical amount of input remaining,

3.

and their repair sequences are compatible.

Two repair sequences are compatible:

1.

if they both end in the same number (n ≥0) of shifts,

2.

and, if one repair sequence ends in a delete, the other repair sequence also ends in a delete.

The first of these conditions is a direct consequence of the fact that a configuration is deemed successful if it ends in N_shifts shift repairs. When we merge configurations, one part of the merge is ‘dominant’ (i.e. checked for N_shifts) and the other ‘subsumed’: we have to maintain symmetry between the two to prevent the dominant part accidentally preventing the subsumed part from being recorded as successful. In other words, if the dominant part of the merge had fewer shifts at the end of its repair sequence than the subsumed part, then the N_shifts check (line 10, Figure 6) would fail, even though reversing the dominant and subsumed parts may have lead to success. It is therefore only safe to merge repair sequences which end in the same number of shifts.

The second condition relates to the weak form of compatible merging inherited from [5, p. 8]: delete repairs are never followed by an insert (see Figure 6) since [delete, insert x] always leads to the same configuration as [insert x, delete]. Although we get much of the same effect through compatible configuration merging, we keep it as a separate optimisation because: it is such a frequent case; our use of the todo list means that we would not catch every case; the duplicate repair sequences are uninteresting from a user perspective, so we would have to filter them out later anyway; and each additional merge costs memory. We thus have to make sure that merged repair sequences don’t accidentally suppress insert repairs because one part of the repair sequence ends in a delete while the other does not. The simplest way of solving this problem is thus to forbid merging repair sequences if one sequence ends in a delete and the other does not.

Fortunately, implementing compatible configuration merging is simple. We first modify the todo data-structure to be a list-of-ordered-hashsets⁵. This has near-identical append / pop performance to a normal list, but filters out duplicates with near-identical performance to an unordered hashset. We then make use of a simple property of hashsets: an object’s hashing behaviour need only be a non-strict subset of its equality behaviour. Configuration hashing is based solely on a configuration’s parsing stack and remaining input, giving us a fast way of finding configurations that are compatible under conditions #1 (identical parsing stacks) and #2 (identical input remaining). As well as checking those two conditions, configuration equality also checks condition #3 (compatible repair sequences).

Conceptually, merging two configurations together is simple: each configuration needs to store a set of repair sequences, each of which is updated as further repairs are found. However, this is an extremely inefficient representation as the sets involved need to be copied and extended as each new repair is found. Instead, we reuse the idea of graph-structured stacks from GLR parsing [28, p. 4] which allows us to avoid copying whenever possible. The basic idea is that configurations no longer reference a parent pointer tree of repairs directly, but instead a parent pointer tree of repair merges. A repair merge is a pair (repair, merged) where repair is a plain repair and merged is a (possibly null) set of repair merge sequences. This structure has two advantages. First, the N_shifts check can be performed solely using the first element of repair merge pairs. Second, we avoid allocating memory for configurations which have not yet been subject to a merge. The small downside to this scheme is that expanding configurations into repair sequences requires recursively expanding both the normal parent pointer tree of the first element as well as the merged parent pointer trees of the second element.

Compatible configuration merging is particularly effective in complex cases, even though it can only merge configurations in the todo list (i.e. we cannot detect all possible compatible merges). An example of compatible configuration merging can be seen in Figure 9.

5.3 Ranking repair sequences

---

---

---

---

In nearly all cases, members of the complete set of minimum cost repair sequences end with N_shifts (the only exception being error locations near the end of an input where recovery leads to an accept state). Thus while the repair sequences we find are all equivalently good within the range of N_shifts , some, but not others, may perform poorly beyond that range. This problem is exacerbated by the fact that N_shifts has to be a fairly small integer (we use 3, the value suggested by Corchuelo et al.) since each additional token searched exponentially increases the search space. Thus while all repair sequences found may be locally equivalent, when considered in the context of the entire input, some may be better than others. While it is, in general, impractically slow to determine which repair sequences are the global best, we can quickly determine which are better under a wider definition of ‘local’.

We thus rank configurations which represent the complete set of minimum cost repair sequences by how far they allow parsing to continue, up to a limit of N_try tokens (which we somewhat arbitrarily set at 250). Taking the furthest-parse point as our top rank, we then discard all configurations which parsed less input than this. The reason why we rank the configurations, and not the repair sequences, is that we only need to rank one of the repair sequences from each merged configuration, a small but useful optimisation. We then expand the top ranked configurations into repair sequences and remove shifts from the end of those repair sequences. Since the earlier merging of compatible configurations is imprecise (it misses configurations that have already been processed), there can be some remaining duplicate repair sequences: we thus perform a final purge of duplicate repair sequences. Figure 9 shows a visualisation of CPCT+ in action.

Particularly on real-world grammars, selecting the top-ranked repair sequences substantially decreases cascading errors (see Figure 10 for an example). It also does so for very little additional computational cost, as the complete set of minimum cost repair sequences typically contains only a small number of items. However, it cannot entirely reduce the cascading error problem. Since, from our perspective, each member of the top-ranked set is equivalently good, we non-deterministically select one of its members to repair the input and allow parsing to continue. This can mean that we select a repair sequence which performs less well beyond N_try tokens than other repair sequences in the top-ranked set.

5.4 Timeout

The final part of CPCT+ relates to the use of N_total in Corchuelo et al.. As with all members of the Fischer et al. family, CPCT+ is not only unbounded in time [17, p. 14], but also unbounded in memory. In an attempt to combat this, Corchuelo et al. limits repair sequences to a maximum of 3 deletes and 4 inserts and a span of at most 10 tokens, attempting to stop the search from going too far. Unfortunately it is impossible to find good values for these constants, as ‘too far’ is entirely dependent on the grammar and erroneous input: Java’s grammar, for example, is large with a commensurately large search space (requiring smaller constants) while Lua’s grammar is small with a commensurately small search space (which can cope with larger constants).

This problem can be easily seen on inputs with unbalanced brackets (e.g. expressions such as ‘x = f(();’): each additional unmatched bracket exponentially increases the search space. On a modern machine with a Java 7 grammar, CPCT+ takes about 0.3s to find the complete set of minimum cost repair sequences for 3 unmatched brackets, 3s for 4 unmatched brackets, and 6 unmatched brackets caused our 32GiB test machine to run out of RAM.

The only sensible alternative is a timeout: up to several seconds is safe in our experience. We thus remove N_total from CPCT+ and rely entirely on a timeout which, in this paper, is defined to be 0.5s.

6 Experiment

In order to understand the performance of CPCT+, we conducted a large experiment on real-world Java code. In this section we outline our methodology (Section 6.1) and results (Section 6.2). Our experiment is fully repeatable and downloadable from https://archive.org/download/error_recovery_experiment/0.4/. The results from our particular run of the experiment can also be downloaded from the same location.

6.1 Methodology

In order to evaluate error recovery implementations, we need a concrete implementation. We created a new Yacc-compatible parsing system grmtools in Rust which we use for our experiments. Including associated libraries for LR table generation and so on, grmtools is around 13KLoC. Although intended as a production library, it has accidentally played a part as a flexible test bed for experimenting with, and understanding, error recovery algorithms. We added a simple front-end nimbleparse which produces the output seen in e.g. Figure 1.

There are two standard problems when evaluating error recovery algorithms: how to determine if a good job has been done on an individual example; and how to obtain sufficient examples to get a wide perspective on an algorithm’s performance. Unfortunately, solutions to these problems are mutually exclusive, since the only way to tell if a good job has been done on a particular is to manually evaluate it, which means that it is only practical to use a small set of input programs. Most papers we are aware of use at most 200 source files (e.g. [5]), with one using a single source file with minor variants [15]. [4] was the first to use a large-scale corpus of approximately 60,000 Java source files. Early in the development of our methodology, we performed some rough experiments which suggested that statistics only start to stabilise once a corpus exceeds 10,000 source files. We therefore prefer to use a much larger corpus than most previous studies. We are fortunate to have access to the Blackbox project [3], an opt-in data collection facility for the BlueJ editor, which records major editing events (e.g. compiling a file) and sends them to a central repository. Crucially, one can see the source code associated with each event. What makes Blackbox most appealing as a data source is its scale and diversity: it has hundreds of thousands of users, and a huge collection of source code.

We first obtained a Java 1.5 Yacc grammar and updated it to support Java 1.7.⁶ We then randomly selected source files from Blackbox’s database (following the lead of [27], we selected data from Blackbox’s beginning until the end of 2017-12-31). We then ran such source files through our Java 1.7 lexer. We immediately rejected files which didn’t lex, to ensure we were dealing solely with parsing errors⁷ (see Section 7.4). We then parsed candidate files with our Java grammar and rejected any which did parse successfully, since there is little point running an error recovery algorithm on correct input. The final corpus consists of 200,000 source files (collectively a total of 401MiB). Since Blackbox, quite reasonably, requires each person with access to the source files to register with them, we cannot distribute the source files directly; instead, we distribute the (inherently anonymised) identifiers necessary to extract the source files for those who register with Blackbox.

The size of our corpus means that we cannot manually evaluate repairs for quality. Instead, we report several other metrics, of which the number of error locations is perhaps the closest proxy for perceived quality. However, this number has to be treated with caution for two reasons. First, it is affected by differences in the failure rate: if a particular error recovery algorithm cannot repair an entire file then it may not have had time to find all the ‘true’ error locations. Second, the number of error locations only allows relative comparisons. Although we know that the corpus contains at least 200,000 manually created errors (i.e. at least one per file), we cannot know if, or how many, files contain more than one error. Since we cannot know the true number of error locations, we are unable to evaluate algorithms in an absolute sense.

In order to test hypothesis H1 we ran each error recovery algorithm against the entire Java corpus, collecting for each file: the time spent in recovery (in seconds); whether error recovery on the file succeeded or failed (where failure is due to either the timeout being exceeded or no repair sequences being found for an error location); the number of error locations; the cost of repair sequences at each error location; and the proportion of tokens skipped by error recovery (i.e. how many delete repairs were applied). We measure the time spent in error recovery with a monotonic wall-clock timer, covering the time from when the main parser first invokes error recovery until an updated parsing stack and parsing index are returned along with minimum cost repair sequences. The timer is suspended when normal parsing restarts and resumed if error recovery is needed again (i.e. the timeout applies to the file as a whole).

We evaluate three main error recovery algorithms: Corchuelo et al., CPCT+, and panic mode. Our implementation of Corchuelo et al. is to some extent a ‘best effort’ since we have had to fill in several implementation details ourselves. As per the description, we: use the same limits on repair sequences (repair sequences can contain at most 3 delete or 4 insert repairs, and cannot span more than 10 tokens in the input); complete as soon as a single successful repair sequence is found; and, when no available repair sequence is found, fall back on panic mode. In addition, we impose the same 0.5s timeout on this algorithm, as it is otherwise unbounded in length, and sometimes exhausts available RAM. Panic mode implements the algorithm from Section 3. We do not report the average cost size for panic mode or Corchuelo et al. (which falls back on panic mode) since they do not (always) report repair sequences (see Section 3). Although panic mode can implicitly delete input from before the error location, we only include the input it explicitly skips in the proportion of tokens skipped.

In order to test hypothesis H2, we created a variant of CPCT+ called CPCT+rev. Instead of selecting from the minimum cost repair sequences which allow parsing to continue furthest, CPCT+rev selects from those which allow parsing to continue the least far. This models the worst case for other members of the Fischer et al. family which non-deterministically select a single minimum cost repair sequence. In other words, it allows us to understand how many more errors could be reported to users of other members of the Fischer et al. family compared to CPCT+ .

Configuration merging (see Section 5.2) is the most complex part of CPCT+. To understand whether this complexity leads to better performance, we created another variant of CPCT+ called CPCT+ DM which disables configuration merging.

We bootstrap [10] our results 10,000 times to produce 99% confidence intervals. However, as Figure 12 shows, our distribution is heavy-tailed, so we cannot bootstrap naively. Instead, we ran each error recovery algorithm 30 times on each source file; when bootstrapping we randomly sample one of the 30 values collected (i.e. our bootstrapped data contains an entry for every file in the experiment; that entry is one of the 30 values collected for that file).

All experiments were run on an otherwise unloaded Intel Xeon E3-1240 v6 with 32GiB RAM running Debian 10. We disabled hyperthreading and turbo boost and ran experiments serially. Our experiments took approximately 15 days to complete. We used Rust 1.43.1 to compile grmtools (the Cargo.lock file necessary to reproduce the build is included in our experimental repository).

6.2 Results

---

---

---

---

Figure 11 shows a summary of the results of our experiment. Comparing the different algorithms requires care as a higher failure rate tends to lower the cost size, tokens skipped, and number of error locations simply because files are not completely parsed. For example, although Corchuelo et al. reports fewer error locations than CPCT+, that is probably due to Corchuelo et al.’s higher failure rate; however, as we shall see in Section 6.3, panic mode’s much higher number of error locations relative to CPCT+ might better be explained by other factors.

With that caution in mind, the overall conclusions are fairly clear. CPCT+ is able to repair nearly all input files within the 0.5s timeout. While panic mode is able to repair every file within the 0.5s timeout, it reports well over twice as many error locations as CPCT+– in other words, panic mode substantially worsens the cascading error problem. As well as producing more detailed and accurate output, CPCT+ has a lower failure rate, median, and mean time than Corchuelo et al.. The fact that the median recovery time for CPCT+ is two orders of magnitude lower than its mean recovery time suggests that only a small number of outliers cause error recovery to take long enough to be perceptible to humans; this is confirmed by the histogram in Figure 12. These results strongly validate Hypothesis H1.

Corchuelo et al.’s poor performance may be surprising, as it produces at most one (possibly non-minimum cost) repair sequence whereas CPCT+ produces the complete set of minimum cost repair sequences – in other words, CPCT+ is doing more work, more accurately, and in less time than Corchuelo et al.. There are three main reasons for Corchuelo et al.’s poor performance. First, the use of CR Shift 1 causes the search to miss intermediate nodes that would lead to success being detected earlier. Second, the heuristics used to stop the search from going too far (e.g. limiting a repair sequence’s number of inserts and deletes) are not well-suited to a large grammar such as Java’s: the main part of the search often exceeds the timeout, leaving no time for the fallback mechanism of panic mode to be used. Finally, Corchuelo et al. lacks configuration merging, causing it to perform needless duplicate work.

---

---

CPCT+ ranks the complete set of minimum cost repair sequences by how far parsing can continue and chooses from those which allow parsing to continue furthest. CPCT+rev, in contrast, selects from those which allow parsing to continue the least far. CPCT+rev shows that the ranking technique used in CPCT+ substantially reduces the potential for cascading errors: CPCT+rev leads to 31.93%±0.289% more error locations being reported to users relative to CPCT+. We visualise this in the histogram of Figure 13 which shows all files with 1–50 error locations (a complete histogram can be found in Figure 18 in the Appendix). Note that files where error recovery did not complete and no error locations were found (which happens occasionally with CPCT+rev) are excluded from this histogram (since we know that every file in the corpus has at least one error), but files where error recovery did not complete but some error locations were found are included (since this gives us, at least, a lower bound on the number of error locations). As Figure 13 shows, the distribution of error locations in CPCT+ and CPCT+rev is similar, with the latter simply shifted slightly to the right. In other words, CPCT+rev makes error recovery slightly worse in a number of files (rather than making error recovery in a small number of files a lot worse). This strongly validates Hypothesis H2.

Interestingly, and despite its higher failure rate, CPCT+rev has a noticeably higher mean cost of repair sequences relative to CPCT+. In other words, CPCT+rev not only causes more error locations to be reported, but those additional error locations have longer repair sequences. This suggests that there is a double whammy from cascading errors: not only are more error locations reported, but the poorer quality repair sequences chosen make subsequent error locations disproportionately harder for the error recovery algorithm to recover from.

CPCT+ DM shows that configuration merging has a significant effect on the failure rate, in our opinion justifying both its conceptual complexity and the less than 100LoC Rust code taken to implement it. The slowdown in the mean and median time for CPCT+ DM suggests that configuration merging is particularly effective on files with complex or numerous errors.

6.3 The impact of skipping input

The number of error locations reported by panic mode is well over twice that of CPCT+; even given CPCT+ ’s higher failure rate relative to panic mode, this seemed hard to explain. We thus made an additional hypothesis:

H3

The greater the proportion of tokens that are skipped, the greater the number of error locations.

The intuition underlying this hypothesis is that, in general, the user’s input is very close to being correct and that the more input that error recovery skips, the less likely it is to get back to a successful parse. We added the ability to record the proportion of tokens skipped as the result of delete repairs during error recovery. The results in Figure 11 show a general correlation between the proportion of tokens skipped and the number of error locations (e.g. CPCT+ skips very little of the user’s input; CPCT+rev skips a little more; and panic mode skips an order of magnitude more). However, Corchuelo et al. does not obviously follow this pattern: relative to the other algorithms, its number of error locations does not correlate with the proportion of input skipped. This is mostly explained by its high mean time and high failure rate: Corchuelo et al. tends to fail on files with large numbers of error locations, underreporting the ‘true’ number of error locations simply because it cannot make it all of the way through such files before the timeout. However, this outlier means that we consider Hypothesis H3 to be only weakly validated.

7 Using error recovery in practice

Although several members of the Fischer et al. family were implemented in parsing tools of the day, to the best of our knowledge none of those implementations have survived. Equally, we are not aware of any member of the Fischer et al. family which explains how error recovery should be used or, indeed, if it has any implications for users at all.

We are therefore forced to treat the following as an open question: can one sensibly use error recovery in the Fischer et al. family in practice? In particular, given that the most common way to use LR grammars is to execute semantic actions as each production is reduced, what should semantic actions do when parts of the input have been altered by error recovery? This latter question is important for real-world systems (e.g. compilers) which can still perform useful computations (e.g. running a type checker) in the face of syntax errors.

While different languages are likely to lend themselves to different solutions, in this section we show that grmtools allows sensible integration of error recovery in a Rust context. Readers who prefer to avoid Rust-specific details may wish to move immediately to Section 8.

7.1 A basic solution

---

---

Figure 14 shows a naive grmtools version of the grammar from Figure 2 that can evaluate numeric results as parsing occurs (i.e. given the input 2 + 3 *4 it returns 14). This grammar should mostly be familiar to Yacc users: each production has a semantic action (i.e. Rust code that is executed when the production is reduced); and symbols in the production are available to the semantic action as pseudo-variables named $n (a production of n symbols has n pseudo-variables with the first symbol connected to $1 and so on). A minor difference from traditional Yacc is that grmtools allows rules to specify a different return type, an approach shared with other modern parsers such as ANTLR [22].

A more significant difference relates to the $n pseudo-variables: if they reference a rule R, then their type is R’s return type; if they reference a token T, then their type is (slightly simplified) Result<Lexeme, Lexeme>. We will explain the reasons for this shortly, but at this stage it suffices to note that, unless a token was inserted by error recovery, we can extract tokens by calling $1.unwrap(), and obtain the actual string the user passed by using the globally available $lexer.span_str function.

7.2 Can semantic action execution continue in the face of error recovery?

---

---

In Yacc, semantic actions can assume that each symbol in the production has ‘normal’ data attached to it (either a rule’s value or the string matching a token; Yacc’s error recovery is implicitly expected to maintain this guarantee) whereas, in our setting, inserted tokens have a type but no value. Given the input ‘(2 + 3’, the inserted close bracket is not hugely important, and our calculator returns the value 5. However, given the input ‘2 +’, CPCT+ finds a single repair sequence [Insert Int]: what should a calculator do with an inserted integer? Our naive calculator simply panics (which is roughly equivalent to ‘raises an exception and then exits’) in such a situation (the unwrap in Figure 14 on line 17). However, there are two alternatives to this rather extreme outcome: the semantic action can assume a default value or stop further execution of semantic values while allowing parsing to continue. Determining which is the right action in the face of inserted tokens is inherently situation specific. We therefore need a pragmatic way for users to control what happens in such cases.

The approach we take is to allow users to easily differentiate normal vs. inserted tokens in a semantic action. Pseudo-variables that reference tokens have (slightly simplified) the Rust type Result<Lexeme, Lexeme>. Rust’s Result type⁸ is a sum type which represents success (Ok(…)) or error (Err(…)) conditions. We use the Ok case to represent ‘normal’ tokens created from user input and the Err case to represent tokens inserted by error recovery. Since the Result type is widely used in Rust code, users can avail themselves of standard idioms.

For example, we can then alter our calculator grammar to continue parsing, but stop executing meaningful semantic action code, when an inserted integer is encountered. We change grammar rules from returning type T to Result<T, Box<dyn Error>> (where Box<dyn Error> is roughly equivalent to ‘can return any type of error’). It is then, deliberately, fairly easy to use with the Result<Lexeme, Lexeme> type: for tokens whose value we absolutely require, we use Rust’s ‘?’ operator (which passes Ok values through unchanged but returns Err values to the function caller) to percolate our unwillingness to continue evaluation upwards. While Box<dyn Error> is slightly verbose, it is a widely understood Rust idiom. Figure 15 shows that changing the grammar to make use of this idiom requires relatively little extra code.

7.3 Avoiding insert repairs when possible

Although we now have a reasonable mechanism for dealing with inserted tokens, there are cases where we can bypass them entirely. For example, consider the input ‘2 + + 3’, which has two repair sequences [Delete +], [Insert Int]: evaluation of the expression can continue with the former repair sequence, but not the latter. However, as presented thus far, these repair sequences are ranked equally and one non-deterministically selected.

We therefore added an optional declaration %avoid_insert to grmtools which allows users to specify those tokens which, if inserted by error recovery, are likely to prevent semantic actions from continuing execution. In practise, this is synonymous with those tokens whose values (and not just their types) are important. In the calculator grammar only the INT token satisfies this criteria, so we add %avoid_insert "INT" to the grammar. We then make a simple change to the repair sequence ranking of Section 5.3 such that the final list of repair sequences is sorted with inserts of such tokens at the bottom of the list. In our case, this means that we always select Delete + as the repair sequence to apply to the input ‘2 + + 3’ (i.e. the Insert Int repair sequence is always presented as the second option).

7.4 Turning lexing errors into parsing errors

---

---

In most traditional parsing systems, lexing errors are distinct from parsing errors: only files which can be fully lexed can be parsed. This is confusing for users, who are often unaware of the distinction between these two phases. To avoid this, we lightly adapt the idea of error tokens from incremental parsing [31, p. 99]. In essence, any input which cannot be lexed is put into a token whose type is not referenced in the normal grammar. This guarantees that all possible input lexes without error and, when the parser encounters an error token, normal error recovery commences⁹ .

A basic, but effective, version of this requires no special support from grmtools (see Figure 16). First, we add a new toke type to the lexer that matches each otherwise invalid input character. Second, since grmtools requires that all tokens defined in the lexer are referenced in the grammar, we add a dummy rule to the grammar that references the token (making sure not to reference this rule elsewhere in the grammar). These two steps are sufficient to ensure that users always see the same style of error messages, and the same style of error recovery, no matter whether they make a lexing or a parsing error.

8 Threats to validity

Although it might not be obvious at first, CPCT+ is non-deterministic, which can lead to different results from one run to the next. The root cause of this problem is that multiple repair sequences may have identical effects up to N_try tokens, but cause different effects after that value. By running each file through each error recovery multiple times and reporting confidence intervals, we are able to give a good – though inevitably imperfect – sense of the likely variance induced by this non-determinism.

Our implementation of Corchuelo et al. is a ‘best effort’. The description in the paper is incomplete in places and, to the best of our knowledge, the accompanying source code is no longer available. We thus may not have faithfully implemented the intended algorithm.

Blackbox contains an astonishingly large amount of source code but has two inherent limitations. First, it only contains Java source code. This means that our main experiment is limited to one grammar: it is possible that our techniques do not generalise beyond the Java grammar (though, as Appendix A suggests, our techniques do appear to work well on other grammars). Although [4, p. 109] suggests that different grammars make relatively little difference to the performance of such error recovery algorithms, we are not aware of an equivalent repository for other language’s source code. One solution is to mutate correct source files (e.g. randomly deleting tokens), thus obtaining incorrect inputs which we can later test: however, it is difficult to uncover and then emulate the numerous, sometimes surprising, ways that humans make syntax errors, particularly as some are language specific (though there is some early work in this area [7]). Second, Blackbox’s data comes largely from students, who are more likely than average to be somewhat novice programmers. It is clear that novice programmers make some different syntax errors – and, probably, make some syntax errors more often – relative to advanced programmers. For example, many of the files with the greatest number of syntax errors are caused by erroneous fragments repeated with variants (i.e. it is likely that the programmer wrote a line of code, copy and pasted it, edited it, and repeated that multiple times before deciding to test for syntactic validity). It is thus possible that a corpus consisting solely of programs from advanced programmers would lead to different results. We consider this a minor worry, partly because a good error recovery algorithm should aim to perform well with inputs from users of different experience levels.

Our corpus was parsed using a Java 1.7 grammar, but some members of the corpus were almost certainly written using Java 1.8 or later features. Many – though not all – post-1.7 Java features require a new keyword: such candidate source files would thus have failed our initial lexing test and not been included in our corpus. However, some Java 1.8 files will have made it through our checks. Arguably these are still a valid test of our error recovery algorithms. It is even likely that they may be a little more challenging on average, since they are likely to be further away from being valid syntax than files intended for Java 1.7.

9 Related work

Error recovery techniques are so numerous that there is no definitive reference or overview of them. However, [8] contains an overall historical analysis and [4] an excellent overview of much of the Fischer et al. family. Both must be supplemented with more recent works.

The biggest limitation of error recovery algorithms in the Fischer et al. family (including CPCT+ ) is that they find repairs at the point that an error is discovered, which may be later in the file than the cause of the error. Thus even when they successfully recover from an error, the repair sequence reported may be very different from the fix the user considers appropriate (note that this is distinct from the cascading error problem, which our ranking of repair sequences in Section 5.3 partly addresses). A common, frustrating, example of this is a missing ‘}’ character in C/Java-like languages. Some approaches are able to backtrack from the source of the error in order to try and find more appropriate repairs. However, there are two challenges to this: first, the cost of maintaining the necessary state to backtrack slows down normal parsing (e.g. [6] only stores the relevant state at each line encountered to reduce this cost), whereas we add no overhead at all to normal parsing; second, the search-space is so hugely increased that it can be harder to find any repairs at all [8].

One approach to global error recovery is to use machine learning to train a system on syntactically correct programs [27]: when a syntax error is encountered, the resulting model is used to suggest appropriate global fixes. Although [27] also use data from Blackbox, their experimental methodology is both stricter – aiming to find exactly the same repair as the human user applied – and looser – they only consider errors which can be fixed by a single token, discarding 42% of the data [27, p. 8]) whereas we attempt to fix errors which span multiple tokens. It is thus difficult to directly compare our results to theirs. However, by the high bar they have set themselves, they are able to repair 52% of single-token errors.

As CPCT+rev emphasises, choosing an inappropriate repair sequence during error recovery leads to cascading errors. The noncorrecting error recovery approach proposed by [26] explicitly addresses this weakness, eschewing repairs entirely. When a syntax error is discovered, noncorrecting error recovery attempts to discover all further syntax errors by checking whether the remaining input (after the point an error is detected) is a valid suffix in the language. This is achieved by creating a recogniser that can identify all valid suffixes in the language. Any errors identified in the suffix parse are guaranteed to be genuine syntax errors because they are uninfluenced by errors in the (discarded) prefix (though this does mean that some genuine syntax errors are missed that would not have been valid suffixes at that point in the user’s input had the original syntax error not been present). There seem to be two main reasons why noncorrecting error recovery has not been adopted. First, building an appropriate recogniser is surprisingly tricky and we are not currently aware of one that can handle the full class of LR grammars (though the full class of LL grammars has been tackled [30]), though we doubt that this problem is insoluble. Second, as soon as a syntax error is encountered, noncorrecting error recovery is unable to execute semantic actions, since it lacks the execution context they need.

Although one of our paper’s aims is to find the complete set of minimum cost repair sequences, it is unclear how best to present them to users, leading to questions such as: should they be simplified? should a subset be presented? and so on. Although rare, there are some surprising edge cases. For example, the (incorrect) Java 1.7 expression ‘x = f(""a""b);’ leads to 23,067 minimum cost repair sequences being found, due to the large number of Java keywords that are valid in several parts of this expression leading to a combinatorial explosion: even the most diligent user is unlikely to find such a volume of information valuable. In a different vein, the success condition of ‘reached an accept’ state is encountered rarely enough that we sometimes forgot that it could happen and were confused by an apparently unexplained difference in the repair sequences reported for the same syntax chunk when it was moved from the end to the middle of a file. There is a body of work which has tried to understand how best to structure compiler error messages (normally in the context of those learning to program). However, the results are hard to interpret: some studies find that more complex error messages are not useful [20], while others suggest they are [24]. It is unclear to us what the right approach might be, or how it could be applied in our context.

The approach of [17] is similar to Corchuelo et al., although the former cannot incorporate shift repairs. It tries harder than CPCT+ to prune out pointless search configurations [17, p. 12], such as cycles in the parsing stack, although this leads to some minimum cost repairs being skipped [2]. A number of interlocking, sophisticated pruning mechanisms which build on this are described in [4]. These are significantly more complex than our merging of compatible configurations: since this gives us acceptable performance in practise, we have not investigated other pruning mechanisms.

The most radical member of the Fischer et al. family is that of [15]¹⁰ . This generates repair sequences in the vein of Corchuelo et al. using the A* algorithm and a precomputed distance table. [15] works exclusively on the stategraph, assuming that it is unambiguous. However, Yacc systems allow ambiguous stategraphs and provide a means for resolving those ambiguities when creating the statetable. Many real-world grammars (e.g. Lua, PHP) make use of ambiguity resolution. In an earlier online draft, we created MF, a statetable-based algorithm which extends CPCT+ with ideas from [15] at the cost of significant additional complexity. With the benefit of hindsight, we do not consider MF’s relatively small benefits (e.g. reducing the failure rate by approximately an additional 0.5%) to be worth that additional complexity.

CPCT+ takes only the grammar and token types into account. However, it is possible to use additional information, such as nesting (e.g. taking into account curly brackets) and indentation when recovering from errors. This has two aims: reducing the size of the search space (i.e. speeding up error recovery); and making it more likely that the repairs reported matched the user’s intentions. The most sophisticated approach in this vein we are aware of is that of [6]. At its core, this approach uses GLR parsing: after a grammar is suitably annotated by the user, it is then transformed into a ‘permissive’ grammar which can parse likely erroneous inputs; strings which match the permissive parts of the grammar can then be transformed into a non-permissive counterpart. In all practical cases, the transformed grammar will be ambiguous, hence the need for generalised parsing. Our use of parent-pointer trees in configuration merging gives that part of our algorithm a similar feel to GLR parsing (even though we do not generate ambiguous strings). However, there are major differences: LR parsers are much simpler than GLR parsers; and the Fischer et al. family of algorithms do not require manually annotating, or increasing the size of, the grammar.

A different approach to error recovery is that taken by [23]: rather than try and recover from errors directly, it reports in natural language how the user’s input caused the parser to reach an error state (e.g. “I read an open bracket followed by an expression, so I was expecting a close bracket here”), and possible routes out of the error (e.g. “A function or variable declaration is valid here”). This involves significant manual work, as every parser state (1148 in the Java grammar we use) in which an error can occur needs to be manually marked up, though the approach has various techniques to lessen the problem of maintaining messages as a grammar evolves.

Many compilers and editors have hand-written parsers with hand-written error recovery. Though generally ad-hoc in their approach, it is possible, with sufficient effort, to make them perform well. However, this comes at a cost. For example, the hand-written error recovery routines in the Eclipse IDE are approximately 5KLoC and are solely for use with Java code: CPCT+ is approximately 500LoC and can be applied to any LR grammar.

Although error recovery approaches have, historically, been mostly LR based, there are several non-LR approaches. A full overview is impractical, though a few pointers are useful. When LL parsers encounter an error, they generally skip input until a token in the follow set is encountered (an early example is [29]). Although this outperforms the simple panic mode of Section 3, it will, in general, clearly skip more input than CPCT+, which is undesirable. LL parsers do, however, make it somewhat easier to express grammar-specific error recovery rules. The most advanced LL approach that we are aware of is IntelliJ’s Grammar-Kit, which allows users to annotate their grammars for error recovery. Perhaps the most interesting annotation is that certain rules can be considered as fully matched even if only a prefix is matched (e.g. a partially completed function is parsed as if it was complete). It might be possible to add similar ideas to a successor of CPCT+, though this is more awkward to express in an LR approach. Error recovery for PEG grammars is much more challenging, because the non-LL parts of the grammar mean that there is not always a clearly defined point at which an error is determined to have occurred. PEG error recovery has thus traditionally required extensive manual annotations in order to achieve good quality recovery. [18] tackles this problem by automatically adding many (though not necessarily all) of the annotations needed for good PEG error recovery. However, deciding when to add, and when not to add, annotations is a difficult task and the two algorithms presented have different trade-offs: the Standard algorithm adds more annotations, leading to better quality error recovery, but can change the input language accepted; the Unique algorithm adds fewer annotations, leading to poorer quality error recovery, but does not affect the language accepted. The quality of error recovery of the Unique algorithm, in particular, is heavily dependent on the input grammar: it works well on some (e.g. Pascal) but less well on others (e.g. Java). In cases where it performs less well, it can lead to parsers which skip large portions (sometimes the remainder) of the input.

While the field of programming languages has largely forgotten the approach of [1], there are a number of successor works (e.g. [25]). These improve the time complexity, though none that we are aware of address the issue of how to present to the user what has been done.

We are not aware of any error recovery algorithms that are formally verified. Indeed, as shown in this paper, some have serious flaws. We are only aware of two works which have begun to consider what correctness for such algorithms might mean: [32] provides a brief philosophical justification of the need and [12] provides an outline of an approach. Until such time as someone verifies a full error recovery algorithm, it is difficult to estimate the effort involved, or what issues may be uncovered.

10 Conclusions

In this paper we have shown that error recovery algorithms in the Fischer et al. family can run fast enough to be usable in the real world. Extending such algorithms to produce the complete set of minimum cost repair sequences allows parsers to provide better feedback to users, as well as significantly reducing the cascading error problem. The CPCT+ algorithm is simple to implement (less than 500LoC in our Rust system) and still has good performance.

Looking to the future, we, perhaps immodestly, suggest that CPCT+ might be ‘good enough’ to serve as a common representative of the Fischer et al. family. However, we do not think that it is the perfect solution. We suspect that, in the future, multi-phase solutions will be developed. For example, one may use noncorrecting error recovery (e.g. [26]) to identify syntax errors, and then use a combination of machine-learning (e.g. [27]) and CPCT+ to discover those repair sequences that do not lead to additional error locations being encountered.

Acknowledgements: We are grateful to the Blackbox developers for allowing us access to their data, and particularly to Neil Brown for help in extracting a relevant corpus. We thank Edd Barrett for helping to set up our benchmarking machine and for comments on the paper. We also thank Carl Friedrich Bolz-Tereick, Sérgio Queiroz de Medeiros, Sarah Mount, François Pottier, Christian Urban, and Naveneetha Vasudevan for comments. This research was funded by the EPSRC Lecture (EP/L02344X/1) Fellowship.

References

Appendix (not peer-reviewed)

A Curated examples

In this section we show several examples of error recovery using CPCT+ in different languages, to give some idea of what error recovery looks like in practise, and to emphasise that the algorithms in this paper are grammar neutral. All of these examples use the output from the nimbleparse tool that is part of grmtools.

A.1 Java 7

Example 1 input:

Example 1 output:

Example 2 input:

Example 2 output:

Example 3 (taken from [6, p. 10]) input:

Example 3 output:

Example 4 (taken from [6, p. 16]) input:

Example 4 output:

Example 5 (taken from [19, p. 2]):

Example 5:

Example 6:

Example 6 output, showing the timeout being exceeded and error recovery unable to complete:

A.2 Lua 5.3

Example 1 input:

Example 1 output:

Example 2 input. Note that ‘=’ in Lua is the assignment operator, which is not valid in conditionals; and that if/then/else blocks must be terminated by ‘end’.

Example 2 output:

Examples 3 and 4 (both derived from the Lua 5.3 reference manual) show that CPCT+ naturally deals with an inherent ambiguity in Lua’s Yacc grammar involving function calls and assignments (which, following the Lua specification, is resolved by Yacc in favour of function calls). Example 3 shows the ‘unambiguous’ case (i.e. if Lua forced users to use ‘;’ everywhere, the grammar would have no ambiguities):

Example 3 output:

Example 4 shows what happens in the ‘ambiguous’ case (which Lua’s grammar resolves in favour of viewing the code below as a function call to c):

Example 4 output:

Example 5 (taken from [19, p. 7]):

Example 5 output:

A.3 PHP 7.3

Example 1 input:

Example 1 output:

Example 2 input:

Example 2 output:

Example 3 input:

Example 3 output:

---

---

---

---