That last bit is out of date (we didn't work much on keeping the web site up-to-date): The front end is mostly C++23-complete (modules notably lagging), and we have several C++26 feature in as well.
The internal documentation comprises about 600 pages. The "External Interface" chapter of that document, covering command-line options, language dialect issues, and the use of language features like templates, is freely available for downloading in PDF format - https://www.edg.com/docs/edg_cpp.pdf (dated Nov 21, 2025 and hence presumably the latest version)
PS: Maybe these essential info (pointed out in my comments and updated by you) should be front and center in the new edgcpp website.
I estimate there were very roughly 50 customers in the end. (I was an EDG engineer and didn't work on admin stuff — so it might be off, but not much. I'm estimating based on support tickets origins.) That's down from maybe double that at our max (again, my estimate). I heard of fairly few customers switching to Clang or other projects (Intel being the most notable exception). Most of our customers we "lost" were due to mergers/acquisitions (i.e., two customers suddenly became just one).
Sorry, but that's nonsense. We've supported plenty of features. We've also occasionally been the first to implement them. In recent years, we spearheaded reflection, supported, first-implemented a bunch of constant-evaluation features (consteval etc.), and supported and first-shipped C++23 "explicit this".
EDG started with two people. When John Spicer joined in 1992, they grew to three. I joined when Mike Anderson retired from software development in 1999, so we were still just three. Mike Miller joined in 2004 to make it four. Mike Herrick in 2007 to make it five. We were briefly seven in the early 2020s, and then fell back to six when Ellen Herrick stepped down. The three youngest of us moved to NVIDIA about a year ago.
Steve Adamczyk, the founder, told me that he created the company to be able to enjoy his work rather than to grow it into something that would rob him of that joy (I hope I paraphrase it right): He is an awesome human!
> Steve Adamczyk, the founder, told me that he created the company to be able to enjoy his work rather than to grow it into something that would rob him of that joy (I hope I paraphrase it right):
No, MSVC is its own front end (the older continuously maintained C++ front end, in fact). However, Visual Studio's C++ IntelliSense uses the EDG front end.
No, we only parse the language, do full semantic analysis, and optionally lower the representation to something that roughly matches C-language semantics. (That lowering can do some minimal inlining if needed, but that's a historical thing.)
We have two "back ends": c_gen_be.c generates C code from the lower IL ("intermediate language"; EDG's term for the AST) and cp_gen_be.c generates C++ code from the unlowered IL.
Several of our customers instrument the AST to add security-probing and/or other dynamic-analysis features.
NVCC uses it to separate out "device" constructs.
That is quite interesting. I haven't thought much about source-to-source transformation within C++.
Do you have any articles/papers/books/etc. you can point us to for understanding this better?
Two usecases i have in mind are;
1) Transforming legacy C++98/C++03 codebases into "modern" C++11/14/17/20/23/26. Is this possible with current edg?
2) Adding verification conditions based on code analysis; both runtime contract asserts and compile time proofs (possible?) so as to convert "normal" C++ code into "somewhat verified" C++ code.
Yes. And this is why the nvcc preprocessor is so effective at working with the various host compilers. It emulates properties of the host compiler so that the device code and host agree on things.
I'd say:
- A single executable can emulate pretty much all versions of MSVC, GCC, and Clang. (E.g., pass the option --gnu_version=80300 and it emulates GCC 8.3.0, including a large number of bugs/idiosynchrasies.)
- An optimized binary is pretty small.
- It builds quite quickly. (A complete "from scratch" build on a modern solid workstation will take just a few seconds.)
- It has some interesting source-to-source transformation capabilities (via cp_gen_be.c).
...
Oh, you had him as a teacher? That's awesome.
I met him and his family 25+ years ago while they were vacationing in California... and they were all great! He was also at the Morristown WG21 meeting, where C++98 was voted out (in November 1997).
(I co-moderated comp.lang.c++.moderated with John. For several years he was the person that kept it alive.)
He was my advisor during undergrad; I had him for everything like C++, Assembly, data structures, operating systems, systems programming, an independent study, and he put me on the path to Unix guru.
His wife Anne was one of my professors as well (Cobol). I promised her I would never get a job writing Cobol. Unbroken.
I would like to think she is happy, but you have to understand the consequences of the bet. Either way she wins in the end. I think she would appreciate that.