Archive
Lifetime of coding mistakes in the Linux kernel
What is the lifetime of coding mistakes in the Linux kernel? Some coding mistakes result in fault reports (some of which are fixed), while many are removed when the source that contains them is deleted/changed during ongoing development.
After fixing the coding mistake(s) in the kernel that generated a reported fault, developer(s) log the commit that introduced the coding mistake, along with the commit that fixed it. This logging started in 2013, and I only found out about it this week. To be exact, I discovered the repo: A dataset of Linux Kernel commits created by Maes Bermejo, Gonzalez-Barahona, Gallego, and Robles.
The log contains the commit hashes for the 90,760 fixes made to the 63 mainline kernel versions from 3.12 to 6.13. The complete log of 1,233,421 commits has to be searched to extract the details, e.g., date, lines added, etc.
The kernel development process involves regular release cycles of around 80 days. Developers submit the code they want to be included in the next release, this goes through a series of reviews, with Linus making the final decision.
The following analysis is based on the coding mistakes introduced between successive kernel releases, e.g., version 3.13 coding mistakes are those introduced into the source between 4 Nov 2013 (the day after version 3.12 was released) and 19 Jan 2014 (when version 3.13 was released). Code will have been worked on, and mistakes created/fixed, before it reached the kernel, which ensures some level of maturity.
The number of people working with pre-release code is likely to be tiny, compared to the number running released kernels. Consequently, the characteristics of coding mistake lifetime is expected to be different pre/post release, if only because more users are likely to report more faults.
The plot below shows the pre-release daily mistake fixed density against days since start of work on the current release, the red line is a fitted regression line mapped to density (fitted regression is a biexponential; code and data):

For all versions, the prior to release daily fix rate follows a consistent pattern: Most fixes occur in the first few days, with roughly an exponential decline to the release date.
The following analysis builds a broad brush model of cumulative fixes over time across 53 mainline kernel releases (the final 10 releases were not included because of their relatively short history).
The number of users of a new kernel takes time to increase as it percolates onto systems, e.g., adopted by Linux distributions and then installed by users, or installed by cloud providers. Eventually, code first included in a particular version will be running on most systems.
The post release daily fix rate is best modelled using the cumulative number of fixes, i.e., total number of fixes up to a given day since release. The models fitted below are based on dividing the post release cumulative fixes into before/after 200 days since release. The 200-day division is a round number (technically, a nearby value may provide a better fit) that supports the fitting of good quality before/after regression models. Averaged over all releases, 42% of fixes occurred within 200-days, and 58% after 200-days.
The plot below shows the cumulative number of post-release fixed faults, in red, for various kernel versions, with fitted regression lines in green and blue (grey line is at 200-days; code and data):

The equation fitted to the before 200-days fixes had the following form:

where:
is a kernel version specific constant; see plot below.
The equation fitted to the after 200-days fixes had the following form:

where:
is a kernel version specific constant; see plot below.
Approximately, after release, the cumulative fix rate starts out quadratic in elapsed days, with the rate decreasing over time, until after 200-days the rate settles down to following the cube-root of days.
Comparing the number of post-release fixes across versions, there is a lot more variability in the first 200-days (i.e., the model fit to the data is sometimes very poor), relatively to after 200-days (where the model fit is consistently good).
Each kernel release has its own characteristics, parameterised by the values
, and
in the above equations. The plot below shows these values across versions, with red for
, blue/green for
, and grey line showing normalised LOC added/changed in the release (code and data):

The plot clearly shows a large increase in the number of fixes between kernel version 3.14 and later versions. The before 200-days rate (blue/green) increase by a factor of seven, while the after 200-days rate increased by a factor of three.
Is this increase driven by some underlying factor in kernel development, or is it an external factor such as an increase in the number of users (more users leads to more faults reports), or the extensive post-release fuzz testing that is now common.
The number of lines of code added/changed, indicated by the grey line (shifted to fit plot axes) cannot be added to the fitted models because they exactly correlate with their respective version.
What is driving the long-term rate of fixes, i.e., cube-root of elapsed days?
Actually, what people are really want to know is what can be done to reduce the number of fixes required after release. When people ask me this, my usual reply is: “Spend more on testing”.
The probability of a coding mistake causing a fault report is decreasing: fixes reduce the number of remaining mistakes, and source added in one kernel version may be removed in a later version.
Perhaps the set of input behaviors is growing, producing the distinct conditions needed to trigger different coding mistakes, or the faults are occurring but are only reported when experienced by a small subset of users.
As always, more data is needed.
Decline in downloads of once popular packages
What happens to the popularity of Open source packages, measured in monthly downloads, once they cease to be updated or attract new users?
If the software does not have any competition within its domain, there is no reason why its popularity should decline. In practice, there are usually alternative packages offering the same or similar functionality. Even when alternatives are available, existing practice and sunk costs can slow migration. A year or so after I started using Asciidoc to write by Software Engineering book, the author announced that he was no longer going to update the software; initially there was no alternative, but the software did what I wanted, and I have been happily using it over the last 12 years.
The paper: Do All Software Projects Die When Not Maintained? Analyzing Developer Maintenance to Predict OSS Usage by Emily Nguyen measured the monthly downloads, commits and other characteristics of 38K GitHub packages having at least 10K downloads during any month between January 2015 and December 2020. The data made available (more here) is a subset, i.e., downloads for 1,583 projects starting in May 2015.
The author investigated the connection between various project characteristics (focusing on commits or lack thereof in particular) and downloads by fitting a Cox proportional hazards model.
The plot below shows the 67 monthly downloads for a selection of packages; the red line is a fitted local regression used to smooth the data (code and data):

Reasons for a decline from a peak number of downloads include: competition from alternative packages, change of fashion, and market saturation, or perhaps the peak was caused by a one-off event. Whatever the reason for a peak+decline, my interest is learning about patterns in the rate of decline.
Some of the monthly package downloads in the above plot have an obvious peak and decline, with others continually increasing, and others having multiple peaks. The following algorithm was used to select packages having a peak followed by a decline, based on the predicted values from a fitted loess model:
- find the month with the most downloads, this is the primary peak,
- if this month is within 10 months of the end of the measurement period, this is not a peak/decline package,
- does a secondary peak exist? A secondary peak is a month containing the most downloads from 10 months after the end of the primary peak, where the number of downloads is within 66% of the primary peak downloads,
- the secondary peak becomes the primary peak, provided it is not within 10 months of the end of the measurement period.
The final fraction of the primary peak is the average monthly download during the last three months divided by the peak month downloads.
The plot below shows the 693 packages whose final fraction of peak was below 0.6 against months from peak to the last month (at the end of 2020), with the red line showing a fitted regression of the form
(code and data):

As the above plot shows, there don’t appear to be any patterns in the decline of package downloads, and
is a poor predictor of fraction of peak.
Perhaps a more sophisticated peak+decline selection algorithm will uncover some patterns. Both ChatGPT (its generated python script failed) and Grok (very wrong answers) failed miserably at classifying the plots. Deepseek will only process images to extract text.
Occurrence of binary operator overloading in C++
Operator overloading, like many programming language constructs, was first supported in the 1960s (Algol 68 also provided a means to specify a precedence for the operator). C++ is perhaps the most widely used language supporting operator overloading; but not redefining their precedence.
I have always thought that operator overloading was more talked about than actually used (despite its long history, I have not been able to find any published usage information). A previous post noted that the CodeQL databases hosted by GitHub provides the data needed to measure usage, and having wrestled with the documentation (ql scripts used), C++ operator overload usage data is available.
The table below shows the total uses of overloaded and ‘usual’ binary operators in the source code (excluding headers) of 77 C++ repositories on GitHub (the 100 repositories C/C+ MRVA). The table is ordered by total occurrences of overloads, with the Percentage column showing the percentage use of overloaded operators against the total for the respective operator (i.e.,
; code and data):
Binary Overload Usual Total Percentage << 103,855 20,463 124,318 83.5 == 21,845 118,037 139,882 15.6 != 14,749 69,273 84,022 17.6 * 12,849 57,906 70,755 18.2 + 10,928 103,072 114,000 9.6 && 8,183 64,148 72,331 11.3 - 5,064 77,775 82,839 6.1 <= 3,960 18,344 22,304 17.8 & 3,320 27,388 30,708 10.8 < 1,351 93,393 94,744 1.4 >> 1,082 11,038 12,120 8.9 / 1,062 29,023 30,085 3.5 > 537 44,556 45,093 1.2 >= 473 27,738 28,211 1.7 | 293 13,959 14,252 2.0 ^ 71 1,248 1,319 5.4 <=> 13 12 25 52.0 % 11 9,338 9,349 0.1 || 9 53,829 53,838 0.017 |
Use of the overloaded << operator is driven by standard library I/O, rather than left shifting.
There are seven operators where 10-20% of the usage is overloaded, which is a lot higher than I was expecting (not that I am a C++ expert).
How much does overloaded binary operator usage vary across projects? In the plot below, each vertical colored violin plot shows the distribution of overload usage for one operator across all 77 projects (the central black lines denote the range of the central 50% of the points; code and data):

While there is some variation between these 77 projects, in most cases a non-trivial percentage of an operator's usage is overloaded.
Fifth anniversary of Evidence-based Software Engineering book
Yesterday was the 5th anniversary of the publication of my book Evidence-based Software Engineering.
The general research trajectory I was expecting in the 2020s (e.g., more sophisticated statistical analysis and more evidence based studies) has been derailed by the arrival of LLMs three years ago. Almost all software engineering researchers have jumped on the LLM bandwagon, studying whatever LLM use case is likely to result in a published paper. While I have noticed more papers using statistical techniques discovered after the digital computer was invented (perhaps influenced by the second half of the book), there seems to be a lot fewer evidence based papers being published. I don’t expect researches studying software engineering to jump off the LLM bandwagon in the next few years.
The net result of this lack of new research findings is that the book contents are not yet in need of an update.
On a positive note, LLMs’ mathematical problem-solving capabilities have significantly reduced the time needed to analyse models of software engineering processes.
Had today’s LLMs been available while I was writing the book, the text would probably have included many more theoretical models and their analysis. ‘Probably’, because sometimes the analysis finds that a model does not provide meaningfully mimic reality, so it’s possible that only a few more models would have been included.
My plan for the next year is to use LLM’s mathematical problem-solving capabilities to help me analyse models of software engineering processes. A discussion of any interested results found will appear on this blog. I’m hoping that there will be active conversations on the evidence based software engineering Discord channel.
It makes sense to hone my model analysis skills by starting with the subject I am most familiar with, i.e., source code. It also helps that tools are available for obtaining more source measurement data.
I will continue to write about any interesting papers that appear on the arXiv lists cs.se and cs.PL, as well as the major conferences. There won’t be time to track the minor conferences.
Questions raised during model analysis sometimes suggest ideas that, when searched for, lead to new data being discovered. Discovering new data using a previously untried search phrase is always surprising.
Best tool for measuring lots of source code
Human written source code contains various common usage patterns. This blog has analysed a variety of these patterns, and in a few cases built models of processes that replicate these patterns. The data for this analysis has primarily comes from programs written in C and Java, because these are the languages that researchers most often study (tool availability and herd mentality).
Do these common usage patterns occur in other languages, or at least other C/Java like languages? I think so, and have set out to collect the necessary data. Obtaining this data requires large quantities of code written in many languages, and the ability to analyse code written in these languages.
GitHub contains huge quantities of code. There are two freely available source code analysis tools supporting many languages: Opengrep (the Open source version of semgrep) and CodeQL.
CodeQL’s method of operation had previously put me off trying it. The method is a two stage process: First a database of information is created by extracting information during a project’s build process (e.g., running existing makefiles and host compilers), followed by querying this database using a declarative language (think minimalist SQL with lots of built-in functions). This approach has the huge advantage of not having to worry about handling compiler dialects/options, however, I’m an ingrained user of tools that process individual files.
From the research perspective, CodeQL has a major feature that is not available with other tools. GitHub, who now owns CodeQL, host thousands of project databases and GitHub Actions allows third-parties to scan up to 1,000 databases of the most popular projects. Access to existing CodeQL databases removes the need to download repo/build project/store database locally.
CodeQL, like other static analysis tools, was designed to find issues/problems in code, and so might not support the kind of functionality I needed to extract source code measurements. The best way to find out if the data of interest could be extracted is to try and do it.
In the best developer tradition, I downloaded a prebuilt release (available for Linux, Windows and Mac; called CodeQL Bundles), skimmed the documentation, ran a simple QL script and spent an hour or two trying to figure out why I was getting Java runtime errors, e.g., “no String-argument constructor/factory method to deserialize from String value“.
Progress would have been faster if I had used Visual Studio Code, available free from the owners of GitHub, rather than the command line. The documentation is not command line oriented. Visual Studio handles details like creating a qlpack.yml file (whose necessary existence I eventually found out about). Also, the harmless looking metadata appearing in comments is necessary and had better match the output parameters of the query. How hard is it to warn that a file could not be found, or that metadata is missing?
The code databases are queried using the declarative language QL, which is a kind of minimal SQL (with the select appearing last, rather than first). The import statement specifies the language, or rather the name of a library module.
The imported library contains classes for each language construct (e.g., BlockStmt, Function, ArrayExpr, etc). In the query below, the line “from LocalScopeVariable lv” extracts all local scope variables, which can subsequently be referred to via the name lv. The where line lists conditions that must be met (in this example, not be a parameter and not be accessed; testing for unused variables). The select line invokes methods that return various kinds of information about the class, e.g., the name of the variable, and location within the source.
/** * @id compound-stmt * @kind problem * @problem.severity warning */ import cpp from LocalScopeVariable lv where not lv instanceof Parameter and not exists(lv.getAnAccess()) select "", ""+lv.getName()+ ","+lv.getLocation().getStartLine()+ ","+lv.getLocation().getEndLine()+ ","+lv.getEnclosingFunction()+","+bs.getFile() |
The output generated is driven by the select, whose number/kind of arguments must match that specified by the metadata.
Developers can write and call functions, such as this one:
predicate header_suffix(string fstr) { fstr = "h" or fstr = "H" or fstr = "hpp" } |
The QL language is a declarative logical query language with roots in Datalog (subset of Prolog). The claim that it is an object-oriented language is technically correct, in that it groups functions into things called classes and supports various constructs usually found in object-oriented languages. The language has the feel of an academic project that happened to be used in a tool that was in the right place at the right time. Using host compilers to enable the tool to support many languages must have been very attractive to GitHub.
Coding in a declarative logic language requires a major mindset change. There are no loops, if statements or assignments. The query is one, potentially very long and complicated, predicate. A mindset change is necessary, but not sufficient, some fluency with the library of functions available is also needed. For instance, the isSideEffectFree predicate is true/false, but does not return a value (so there is nothing to print). I wanted to output 0/1, depending on whether a function was side effect free or not. When asked, all the LLMs questioned insisted that QL supported if-statements and assignment, just like other languages. After lots of dead-ends, an LLM claimed that “CodeQL automatically treats boolean expressions in count as 1/0″, and a test run showed this to be the case:
count(int dummy | dummy = 1 and func.isSideEffectFree() | dummy) |
The QL scripts needed to extract all the data of immediate interest to me were easily implemented. Looking at existing scripts has given me some ideas for more patterns I might measure. CodeQL currently supports 10 languages, and their classes appear to be slightly different (my initial focus is C, C++, Java and Python).
Visual Studio Code is required to run multi-repository variant analysis, i.e., scan up to 1,000 project databases on GitHub. It was after installing the CodeQL extension that I discovered how much smoother the process is within this IDE, compared to the command line (and off course the output is slightly different). There may be alternatives to Visual Studio, but I’m sticking with what the official documentation says.
Stepping back, is CodeQL a useful tool?
For me it is currently very useful, because of the large number of project databases. Some practice is needed to achieve some fluency in the use of a declarative logic language, not a major hurdle.
The need to run queries against a project database may be a major inconvenience for some developers, depending on working practices. Those practicing continuous integration should be ok.
Distribution of method chains in Java and Python
Some languages support three different ways of organizing a sequence of functions/methods, with calls taking as their first argument the value returned by the immediately prior call. For instance, Java supports the following possibilities:
r1=f1(val); r2=f2(r1); r3=f3(r2); // Sequential calls r3=f3(f2(f1(val))); // Nested calls, read right to left r3=val.f1().f2().f3(); // Method chain, read left to right |
Simula 67 was the first language to support the dot-call syntax used to code method chains. Ten years later Smalltalk-76 supported sending a message to the result of a prior send, which could be seen as a method chain rather than a nested call (because it is read left to right; Smalltalk makes minimal use of punctuator characters, so the syntax is not distinguishable).
How common are method chains in source code, and what is the distribution of chain length? Two studies have investigated this question: An Empirical Study of Method Chaining in Java by Nakamaru (PhD thesis), Matsunaga, Akiyama, Yamazaki, and Chiba, and Method Chaining Redux: An Empirical Study of Method Chaining in Java, Kotlin, and Python by Keshk, and Dyer.
The plot below shows the number of Java method chains having a given length, for code available in a given year. The red line is a fitted regression line for 2018, based on a model fitted to the complete dataset (code and data):

The fitted regression model is:

Why is the number of chains of all lengths growing by around 46% per year? I think this growth is driven by the growth in the amount of source measured. Measurements show that the percentage of source lines containing a method call is roughly constant. In the plot above, the number of unchained methods (i.e., chains of length one) increases in-step with the growth of chained methods. All chain lengths will grow at the same rate, if the source that contains them is growing.
What is responsible for the step change in the number of chains at around 10 methods? Nakamaru classified a random sample of 280 chains, and found that roughly 80% of chains longer than eight methods built an object, e.g., the following chain:
MoreObjects.toStringHelper(this) .add("iLine" , iLine) .add("lastK" , lastK) .add("spacesPending", spacesPending) .add("newlinesPending", newlinesPending) .add("blankLines", blankLines) .add("super", super.toString()) .toString() |
Are these chain usage patterns present in Python? The plot below shows the number of Python method chains having a given length, for code available in a given year. The red line is a fitted regression line for 2020, based on a model fitted to the complete dataset (code and data):

The fitted regression model is:

While this model is almost identical to the model fitted to the Java data (the annual growth rate is 39%), the above plot shows a large step change after chains of length two. Keshk’s paper focuses on replicating Nakamaru’s Java results, and then briefly discusses Python. I have an assortment of explanations, but nothing stands out.
Within code, how are method calls split between single calls and a chain of two or more calls?
The fractions in the plot below are calculated as the ratio of chains of length one (i.e., single method call) against chains containing two or more methods. The “j” shows Java ratios, and “p” Python ratios. The red lines show the fraction based on the total number of method calls, and the blue/green lines are based on occurrences of chains, i.e., chain of one vs chain of many (code and data):

The ratio of Java chains containing two or more methods vs one method, grew by around 6% a year between 2006 and 2018, which is only a small part of the overall 46% annual Java growth.
Method chaining is three times more common in Java than Python. In 2020 around a quarter of all method calls were in a chain of two or more, and single method calls were around ten times more common than multi-call chains.
In Python, the use of method chains has roughly remained unchanged over 15 years, with around 5% of all method calls appearing in a chain.
I don’t have a good idea for why method chains are three times more common in Java than Python. Are nested calls the more common usage in Python, or do developers use a sequence of calls communicating using temporary variables?
What of languages that don’t support method chaining, e.g., C. Is the distribution of the number of nested calls (or sequence of calls using temporaries) a power law with an exponent close to 3.7?
Suggestions and pointers to more data welcome.
Finding links between gcc source code and the C Standard
How close is the agreement between the behavior of a compiler and its corresponding language specification?
In the previous century, some Standards’ bodies offered a compiler validation service. However, even when the number of commercial compilers numbered in the hundreds, this service was not commercially viable. These days there are only a handful of industrial strength compilers.
The availability of huge quantities of Open source, for some languages, has created a new language specification. Being able to turn much of this source into executable programs has become an effective measure of compiler correctness.
Those working on C/C++ compilers (Open source or otherwise), often claim that they implement the requirements contained in the corresponding ISO Standard. Some are active in the ISO Standards’ process, and I believe that they do strive to implement the requirements contained in the language standard.
How confident can we be that all the requirements contained in a language standard are correctly implemented by a compiler?
There is a cottage industry of testing compiler runtime behavior, often using fuzzers, and sometimes a compiler is one of the programs chosen to test new fuzzing techniques. This research checks optimization and code generation.
This runtime testing is all well and good, but a large percentage of the text in a language specification contains requirements on the syntax and semantics. The quality of syntax/semantic testing depends on how well the people writing the tests understand the language semantics. It takes a year or two of detailed study to achieve an effective compiler-level of understanding of these ‘front-end’ requirements.
The approach taken by the Model Implementation C Checker to show syntax/semantic correctness was to cross-referenced every if-statement in the front-end to one or more lines in the C90 Standard (the 1990 edition of the ISO C Standard), or an internal house-keeping reference (the source contained 3K references to 1.3K requirements in the C Standard). This compiler/checker was formally validated by BSI. As far as I know, this is the only compiler source cross-referenced at the level of individual lines/if-statements; there are compilers whose source contains cross-references to the sections of a language specification.
The main benefit of this cross-referencing process is insuring that every requirement in the C Standard is addressed by the compiler (correctly or otherwise). Other benefits include providing packets of wording for targeted tests and the ability to generate a runtime trace of all language features involved in compiling a given translation unit.
Replicating this cross-referencing for the gcc or llvm C compiler front-ends would be a huge amount of work for somebody who already has a detailed knowledge of the C Standard, along with some knowledge of compilers. The number of pages in the Standard relating to the C language has grown from 101 pages in C90 to 190 pages in C23. At an average of 14 cross-referenceable lines per page, the expected number of cross-references is now likely to be around 2,700.
LLMs are great at extracting information from text, can generate impressive quality C conformance tests, and are much, much cheaper than C language experts. How good are LLMs at working out which section(s)/paragraph(s) in the C Standard are implemented by the source code of a function contained in, say, gcc?
The LLM prompt I used to test out the idea was:
List the section and paragraph number of the latest ISO C standard containing the requirements implemented in the following source code from a C compiler |
Links to all responses from: ChatGPT5 Thinking, Grok3 Fast and Gok3 Expert.
It is very likely that text from the various versions of the C Standard were in the training data of the larger LLMs, so I did not include this material in the prompt. ChatGPT5 Thinking listed the 2024 draft N3220 as its source document.
The source code was copied from the C front-end code of gcc (i.e., the directory: gcc-15.1.0/gcc/c), which contains 67KLOC, with 6.8K if-statements.
Starting small, the following code looked like an easy test (I removed the comment containing text from the C Standard and obscured the big clue in the original function name: null_pointer_constant_p):
bool check_is_n_ptr (const_tree expr) { /* This should really operate on c_expr structures, but they aren't yet available everywhere required. */ tree type = TREE_TYPE (expr); if (expr == nullptr_node) return true; return (TREE_CODE (expr) == INTEGER_CST && !TREE_OVERFLOW (expr) && integer_zerop (expr) && (INTEGRAL_TYPE_P (type) || (TREE_CODE (type) == POINTER_TYPE && VOID_TYPE_P (TREE_TYPE (type)) && TYPE_QUALS (TREE_TYPE (type)) == TYPE_UNQUALIFIED))); } |
Wow, ChatGPT5 Thinking’s response even included a note on why the code checks for an unqualified type:
Here are the exact C23 (ISO/IEC 9899:2024) clauses that your function is implementing:
6.3.2.3 Pointers, ¶3 — Definition of a null pointer constant: “An integer constant expression
with the value 0, such an expression cast to type void *, or the predefined constant nullptr
is called a null pointer constant.”
6.6 Constant expressions, ¶8 — What qualifies as an integer constant expression
(your INTEGER_CST / INTEGRAL_TYPE_P checks).
6.6 Constant expressions, ¶4 — Range requirement for constant expressions
(motivates the !TREE_OVERFLOW (expr) check).
6.4.4.6 Predefined constants, ¶¶1–3 — nullptr is a predefined constant (token classification;
ties to your expr == nullptr_node fast-path).
6.3.2.4 nullptr_t, ¶¶1–2 — Conversions of nullptr_t (confirms nullptr converts to any pointer
type yielding the null pointer value).
Notes
• The “cast to type void *” wording in 6.3.2.3 means unqualified void *,
which is why your code insists the pointed-to void is unqualified.
Grok3 Fast’s response was even more expansive (and almost instantaneous), listing examples of section/paragraphs containing uses of null pointers, and pointing out that the wording was unchanged until C23 (when the nullptr keyword was introduced, along with its type nullptr_t). It also pointed out gcc specific features of the implementation (the prompt did not identify the compiler).
That’s enough of the easy questions. The following code (comments removed, function name unchanged) is essentially asking a question: What is the promoted type of the argument?
tree c_type_promotes_to (tree type) { tree ret = NULL_TREE; if (TYPE_MAIN_VARIANT (type) == float_type_node) ret = double_type_node; else if (c_promoting_integer_type_p (type)) { if (TYPE_UNSIGNED (type) && (TYPE_PRECISION (type) == TYPE_PRECISION (integer_type_node))) ret = unsigned_type_node; else ret = integer_type_node; } if (ret != NULL_TREE) return (TYPE_ATOMIC (type) ? c_build_qualified_type (ret, TYPE_QUAL_ATOMIC) : ret); return type; } |
ChatGPT5 listed six references. Three were good, and the other three were closely related, but I would not have cited them. The seven Grok3 references came from several documents using slightly different section numbers. Updating the prompt to explicitly name N3220 as the document to use did not change Grok3’s cited references (for this question).
All the code in the previous questions was there because of text in the C Standard. How do ChatGPT5/Grok3 handle the presence of code that does not have standard associated text?
The following function contains code to handle named address spaces (defined in a 2005 Technical Report: TR 18037 Extensions to support embedded processors).
static tree qualify_type (tree type, tree like) { addr_space_t as_type = TYPE_ADDR_SPACE (type); addr_space_t as_like = TYPE_ADDR_SPACE (like); addr_space_t as_common; /* If the two named address spaces are different, determine the common superset address space. If there isn't one, raise an error. */ if (!addr_space_superset (as_type, as_like, &as_common)) { as_common = as_type; error ("%qT and %qT are in disjoint named address spaces", type, like); } return c_build_qualified_type (type, TYPE_QUALS_NO_ADDR_SPACE (type) | TYPE_QUALS_NO_ADDR_SPACE_NO_ATOMIC (like) | ENCODE_QUAL_ADDR_SPACE (as_common)); } |
ChatGPT5 listed six good references and pointed out the association between the named address space code and TR 18037. Grok3 Fast hallucinated extensive quoted text/references from TR 18037 related to named address spaces. Grok3 Expert pointed out that the Standard does not contain any requirements related to named address spaces and listed two reasonable references.
Finding appropriate cross-references is the time-consuming first step. Next, I want the LLM to add them as comments next to the corresponding code.
I picked a 312 line function, and updated the prompt to add comments to the attached file:
Find the section and paragraph numbers in the ISO C standard, specified in document N3220, containing the requirements implemented in the source code contained in the attached file, and add these section and paragraph numbers at the corresponding places in the code as comment |
ChatGPT5 Thinking thought for 5 min 46 secs (output), and Grok3 Expert thought for 3 mins 4 secs (output).
Both ChatGPT5 and Grok3 modified the existing code, either by joining adjacent lines, changing variable names, or deleting lines. ChatGPT made far fewer changes, while the Grok3 output was 65 lines shorter than the original (including the added comments).
Both LLMs added comments to blocks of if-statements (my fault for not explicitly specifying that every if should be cross-referenced), with ChatGPT5 adding the most cross-references.
One way to stop the LLMs making unasked for changes to the source is to have them focus on the added comments, i.e., ask for a diff that can be fed into patch. The updated prompt is:
Find the section and paragraph numbers in the ISO C standard, specified in document N3220, containing the requirements implemented by each if statement in the source code contained in the attached file. Create a diff file that patch can use to add these section and paragraph numbers as comments at the corresponding lines in the original code |
ChatGPT5 Thinking thought for around 4 min (it reported inconsistent values (output), and Grok3 Expert thought for 5 min 1 sec (output).
The ChatGPT5 patch contained many more cross-references than its earlier output, with comments on more if-statements. The Grok3 patch was a third the size of the ChatGPT5 patch.
How well did the LLMs perform?
ChatGPT5 did very well, and its patch output would be a good starting point for a detailed human expert edit. Perhaps an improved prompt, or some form of fine-tuning would useful improve performance.
Grok3 Fast does not appear to be usable, but Grok3 Expert could be used as an independent check against ChatGPT5 output.
Working at the section/paragraph level it is not always possible to give the necessary detailed cross-reference because some paragraphs contain multiple requirements. It might be easier to split the C Standard text into smaller chunks, rather than trying to get LLMs to give line offsets within a paragraph.
Modeling the distribution of method sizes
The number of lines of code in a method/function follows the same pattern in the three languages for which I have measurements: C, Java, Pharo (derived from Smalltalk-80).
The number of methods containing a given number of lines is a power law, with an exponent of 2.8 for C, 2.7 for Java and 2.6 for Pharo.
This behavior does not appear to be consistent with a simplistic model of method growth, in lines of code, based on the following three kinds of steps over a 2-D lattice: moving right with probability
, moving up and to the right with probability
, and moving down and to the right with probability
. The start of an if or for statement are examples of coding constructs that produce a
step followed by a
step at the end of the statement;
steps are any non-compound statement. The image below shows the distinct paths for a method containing four statements:

For this model, if
the probability of returning to the origin after taking
is a complicated expression with an exponentially decaying tail, and the case
is a well studied problem in 1-D random walks (the probability of returning to the origin after taking
steps is
).
Possible changes to this model to more closely align its behavior with source statement production include:
- include terms for the correlation between statements, e.g., assigning to a local variable implies a later statement that reads from that variable,
- include context terms in the up/down probabilities, e.g., nesting level.
Measuring statement correlation requires handling lots of special cases, while measurements of up/down steps is easily obtained.
How can
/
probabilities be written such that step length has a power law with an exponent greater than two?
ChatGPT 5 told me that the Langevin equation and Fokker–Planck equation could be used to derive probabilities that produced a power law exponent greater than two. I had no idea had they might be used, so I asked ChatGPT, Grok, Deepseek and Kimi to suggest possible equations for the 
/
probabilities.
The physics model corresponding to this code related problem involves the trajectories of particles at the bottom of a well, with the steepness of the wall varying with height. This model is widely studied in physics, where it is known as a potential well.
Reaching a possible solution involved refining the questions I asked, following suggestions that turned out to be hallucinations, and trying to work out what a realistic solution might look like.
One ChatGPT suggestion that initially looked promising used a Metropolis–Hastings approach, and a logarithmic potential well. However, it eventually dawned on me that
, where
is nesting level, and
some constant, is unlikely to be realistic (I expect the probability of stepping up to decrease with nesting level).
Kimi proposed a model based on what it called algebraic divergence:

where:
normalises the probabilities to equal one,
,
is the up probability at nesting 0,
is the down probability at nesting 0, and
is the desired power law exponent (e.g., 2.8).
For C,
, giving 
The average length of a method, in LOC, is given by:
, where: 
For C, the mean function length is 26.4 lines, and the values of
,
, and
need to be chosen subject to the constraint
.
Combining the normalization factor
with the requirement
, shows that as
increases,
slowly decreases and
slowly increases.
One way to judge how closely a model matches reality is to use it to make predictions about behavior patterns that were not used to create the model. The behavior patterns used to build this model were: function/method length is a power law with exponent greater than 2. The mean length,
, is a tuneable parameter.
Ideally a model works across many languages, but to start, given the ease of measuring C source (using Coccinelle), this one language will be the focus.
I need to think of measurable source code patterns that are not an immediate consequence of the power law pattern used to create the model. Suggestions welcome.
It’s possible that the impact of factors not included in this model (e.g., statement correlation) is large enough to hide any nesting related patterns that are there. While different kinds of compound statements (e.g., if vs. for) may have different step probabilities, in C, and I suspect other languages, if-statement use dominates (Table 1713.1: if 16%, for 4.6% while 2.1%, non-compound statements 66%).
Early research on economies of scale for computer systems
Before microprocessor cost/performance wiped out (in the early 1990s) other cpu platforms (e.g., mainframes and minis), people argued that computer hardware benefited from economies of scale.
The claimed benefit was more bang for the buck, i.e., more compute for less money.
Checking this claim requires treating pre-microprocessor computer systems and the later microprocessor-based systems as two separate cases, because many of the factors driving costs and performance are very different.
Today’s large microprocessor-based computer systems achieve economies of scale through discounts from bulk purchases and spreading fixed costs across multiple systems. The data is available, and the economic analysis is straight forward.
A lack of reliable data on the costs of designing/building pre-microprocessor computer systems rules out an economic analysis of cost/performance from first principles. The data that was/is available is the cost of computer systems and some indicators of performance (such as instruction timings or benchmarks).
Now, the observed fact that the cost of compute was decreasing over time is unrelated to the claim that the cost of compute decreases as the size of the computer increases.
Assuming a power law relationship between computer cost,
, and size,
, at a point in time, we have:
, where
is some constant. Economies of scale occur when: 
In his detailed cost/performance analysis of computers between 1944-1967, Kenneth Knight treated computers launched in the same year as effectively occurring at the same time. He also built a single model, with year included as an explanatory variable, which means the fitted rate of decrease is the same over all years (rather than varying between years).
The plot below uses Knight’s 1953-1961 data, and shows operations per second against seconds per dollar (a confusing combination, but what Knight used), with fitted regression lines for three years using Knight’s model (code and data)

The fitted exponent for this form of x/y axis maps to a value which has
, i.e., there are economies of scale.
It so happens that the value of the Knight’s fitted exponent is close to that proposed in a 1953 paper (“High speed arithmetic: The digital computer as a research tool”, no online copy):
It used to cost one cent to do a multiplication on a desk calculator; now it is more like four cents; but with these big machines we can do a million in an hour for $400, and that means twenty-five multiplications for a cent! I believe that there is a fundamental rule, which I modestly call Grosch's law, giving added economy only as the square root of the increase in speed-that is, to do a calculation ten times as cheaply you must do it one hundred times as fast. |
which did indeed become widely known as Grosch’s law.
Having been given a lucky kick-start by Knight (fitted individually, years are not close to Grosch’s law), checking for agreement with Grosch’s law became a focus for later studies. While various papers highlighted problems with the later data analysis (e.g., the regression techniques and sample noise producing mathematical artifacts), Grosch’s law ceased being a thing because mainframes/minicomputers ceased being a thing.
Did mainframe/mincomputers have economies of scale in the years after Knight’s data? It’s difficult to tell, the publicly available data is too sparse to support reliable analysis.
Data+code for book: The New C Standard
All the data+code from my book The New C Standard: An Economic and Cultural Commentary is now available on GitHub. For many years I have been meaning to create an easy way to map from a graph/table in the book to the file containing the data, which has blocked me adding the data to GitHub. I have unblocked by releasing this minimal viable product, i.e., it is essentially a copy of the usage subdirectory in the book’s directory.
While the five stage process to get from graph/table to data is tedious, at least there is a process that provides the data. The caption of the graphs in my Evidence-based Software Engineering book contain a link to the corresponding file on GitHub. This was not possible for the C book because GitHub was still 3-years in the future when the book was published (in 2005).
Work on the book started in late 1999 and measurements of C usage was an integral component. Publicly available source code was still a novelty and large Open source projects were rare (SourceForge was launched at the end of 1999). The large projects with C source available to measure were: Linux, Netscape, Gcc, PostgresSQL, OpenAFS, and OpenMotif. Several popular projects originally written in C had migrated to using C++, and were therefore not applicable.
As the book was completed in 2005, evidence-based software engineering restarted, 20-years after the fall of Rome. Or rather, I have nominated 2005 as the year this happened. Feel free to quibble plus/minus a few years.
Search engines were an essential tool for obtaining research papers, reports, and occasionally downloading data. In 2000 the search engine of choice was AltaVista, but a few years later Google had become the best.
While writing the book, I was a regular visitor to bricks and mortar buildings called libraries. Back then, university libraries contained tens of thousands of physical books, and researchers would photocopy papers of interest. Little did I know that this research practice would soon be dead.
In 2005, I had this to say about software evolution:
Measuring the characteristics of software that change over many releases (software evolution) is a relatively new research topic. Software evolution is discussed in a few sentences, and any future major revision ought to cover this important topic in substantially more detail. |
How might C source code characteristics have changed in the last 20 years?
- The use of K&R style function definitions is probably very rare; it was well on the way out in 1999,
- big software systems have gotten bigger, i.e., more lines of code and more
#includes, - A lot more code using 32-bit integers and 64-bit pointers,
- More storage allocated (memory capacity has increased) because it’s faster to do everything in memory, and there is more data.
Recent Comments