1	Dealing with Software Complexity The Discovery of Structure in Programs
2	Software Development Designing new code Understanding old code Written by others Written by current developer some time ago Maintenance starts from day two Code understanding—the most important and least automated part of development
3	Software Understanding The trivial structure Lexical Grammatical Alphabetical list of classes and functions Diagrams (Booch, etc.) High-level structure Relationships between classes (next slide) Design ideas, patterns Emerging structure
4	Class Dependencies Class A uses class B Strongly: it requires #include B.h Weakly: it requires forward declaration of B Dependency graph analysis Levels of abstraction Tree-like structure (next slide) Cycles and how to break them
5	Hierarchy of Classes
6	Neurology Brain can only work on tree-like structures There are 7 +/- 2 “registers” There is a cache that must be pre-loaded There is long-term, slow-search, memory Reality is infinite, brain is finite. Something must be discarded. Abstracting—creating an item that can fit in a register
7	Programmer’s Brain Context switching—re-loading the brain cache (multitasking is expensive) Small program, easy to restart the brain Maintenance requires code understanding—can take hours or days to fill the cache + a lot is left to debugging and testing Few tools to speed up cache loading Needed: ready-made abstractions that fit the registers
8	Abstractions Subtracting “irrelevant” features What are features? Which features are irrelevant? Abstracting and Categorizing Biology and evolution (next slide)
9	The Origin of Abstraction Primitive organism has access to a featureless input stream—”reality stream” Some things in the stream influence metabolism—selection for primitive detectors Data from detectors are “features”—evolution decides which features are relevant First abstractions: food and danger. Particular combinations of features
10	Imitating Life Cellular automata Image processing, discovering features (lines, squares, faces) Genetic algorithms, training Automata feeding on source code Lexing automata Parsing automata—require infinite number of states
11	Scope Discovery Cellular automata that can count 2-d state space Vertical counts parentheses Horizontal counts braces
12	Bubble Diagrams Glue together matching parentheses Glue together matching braces
13	Document Processing Pre-defined features: documents and words (word breakers) Statistics: distribution of words among documents Relevant words: words that occur more often in a given document than in the rest of the corpus
14	Clustering Documents with similar relevant word profiles form clusters Categorization of documents based of statistical features Categorization—automatic generation of abstractions
15	Clustering Pick a representative set of N relevant words/phrases from the whole corpus Each document is a point in N-dim space Distance between documents in N dim Add gravitational attraction (potential) Documents will start clustering just like galaxies in the early universe
16	Example of Clustering
17	Statistics & Abstractions Nature gathers statistics in a very slow process Science is based on statistics Physics Math Statistical methods (clustering) in program analysis
18	Programs as Documents High-level structure of program Reflects programmers’ ideas Information encoded in vocabulary: names, comments Strong influence of problem domain Depends on personal style Can be processed like document corpus Relevant words, clusters
19	Statistical Discovery Intended structures, for instance Model, View, Controller Hidden structures, for instance separation of UI from the “model” Horizontal structures, aspects Vertical structures, exceptions, exception specification Copy and paste programming, code reuse
20	Integration with Tools Overnight automatic program analysis Creating a map of abstractions Creating various views (bubble diagrams, trees, layers) Discovering hidden structures The ecosphere of a program Evolving automata, feeding on code and on statistical abstractions Maintenance by programmer Encapsulating hidden structures Improving existing abstractions Adding new abstractions