7 ms·

Domain-specific compilers are a bigger field, where there's more opportunity to do something interesting Take C (or some other language) and its mature optimiz

Domain-specific compilers are a bigger field, where there's more opportunity to do something interesting Take C (or some other language) and its mature optimizing compilers as a foundation, and write a program that compiles a high-level description of a domain-specific program into C For example, NN-512 (https://NN-512.com https://NN-512.com) is a compiler that takes as input a simple text description of a convolutional neural net inference graph and produces as output a stand-alone C implementation of that graph GNU Bison (https://www.gnu.org/software/bison/ https://www.gnu.org/software/bison/) is another example, where the output is a parser: Bison reads a specification of a context-free language, warns about any parsing ambiguities, and generates a parser (either in C, C++, or Java) which reads sequences of tokens and decides whether the sequence conforms to the syntax specified by the grammar Halide (https://halide-lang.org/ https://halide-lang.org/) is a language for fast, portable computation on images and tensors, and it originally generated human-readable C++ code as output (but not anymore, now it feeds directly into LLVM) Can anyone provide other examples? (Code generators for network packet processing?)

chrisseaton 6y ago
> Take C (or some other language) and its mature optimizing compilers as a foundation, and write a program that compiles a high-level description of a program into C. Don't you still need many of the techniques taught in this class to build something like that? The material doesn't look inherently targeted at low-level languages or machine-code as a target.
- dbcurtis 6y ago
  One way to think of it is to split the "compiler" problem into three big pieces. 1) Front-end, 2) back-end, 3) tooling. Front end is lexing, parsing, building the AST (in whole or just keeping pieces of it around), semantic analysis. When that is done, you can say "Yup, valid program. It is possible to generate code." So, to your question, yes. Those techniques from the course are 100% in play. Back-end is turning the AST into something you can optimized and generate code from. (Sometimes there is a "middle-end" in there....) High-level optimizations, low level optimizations, memory allocation, register allocation. Tooling is everything that keeps your life from being miserable as a user -- starting with debug symbols and these days people expect library and package management, etc, etc. So if you are interested in exploring the Next Great Syntax or some semantic problems that can be re-written into C, then doing a C-front is a great way to do that. Let the C compiler handle all the really-really-really-hard back-end issues for you, and avoid all that distraction. But.... expect that getting debug symbols from your spiffy language into the ready-to-link object file is going to be, as they say, "non-trivial". So... great for experiments and a great learning exercise, but it's hard to do a production compiler that way.
  - samps 6y ago
    Hi! This course is about the "middle end," FWIW. We do not do parsing or codegen in 6120, and there is no goal (as there is in many undergrad-level compilers courses) of making a complete, end-to-end compiler for a C-like language.
kodoque2 6y ago
Do you have a structured course for Domain-specific compilers available online?
klelatti 6y ago
Not open source but I'm working on a system that turns a DSL (embedded into Jupyter notebooks) into either c or c with opencl for a niche finance application. Not a huge project (<10k lines of code). Nature of the problem means that DSL can be very simple (it's aimed at non-programmers). This approach also means that could easily add other targets - e.g. cuda or wasm.
- 37ef_ced3 6y ago
  Awesome. Link?
  - klelatti 6y ago
    Thanks and sorry - not open source and in testing with single client at the moment.
davidpolberger 6y ago
I have just spent the last 13 months significantly reworking a compiler that turns Excel-like formulas into JavaScript. The compiler is used by an app designer, and the formula language is meant to be easy to pick up for people used to spreadsheets. A significant difference compared to Excel is that our compiler features static typing, enabling helpful error messages to be produced at build-time. (Spreadsheets, on the other hand, tend to try very hard to make sense of types at runtime, using type coercion to turn values of the wrong type into something of the expected type they can work with. That produces surprising results and confuses users. =IF(2, 3, 4) will generally produce 3, because 2 -- which should be a logical/boolean value -- is interpreted as TRUE. We think it's a better idea to warn the user that the first parameter is wrong, and to do that as soon as possible.) By far the largest challenge with reworking the compiler has been the type system. I have expanded our type system from the four standard spreadsheet types to encompass more than 80, as well as added support for lambdas (for use with functions like SUMIF and FILTER), type inference, parameterized types and union types. The biggest addition is probably an imperative mode, that executes formulas with side-effects in response to events being fired. This is all in service of features we have wanted to add for years, that have been held back by our (previously) simplistic type system. I did take a compiler class at university, more than 15 years ago, but while that proved very useful when writing the formula parser (an operator-precedence parser, initially straight out of a textbook), type theory was not part of the class. I had to pick up bits and pieces by reading the Java Language Specification (JLS) and by experimenting. The Cornell class doesn't seem to feature type theory either, which I think is unfortunate. Now that the pendulum has swung, yet again, firmly in the direction of static typing, perhaps static typing will become a bigger part of CS compiler curriculums.
- tomcam 6y ago
  Your project sounds like fun. What/for whom is the project? What level of Excel formula compatibility are you shooting for, e.g. just in the neighborhood like Google Sheets, reasonably close as in Libre Office, or full compatibility with the latest version? If the latter you must have one gnarly test suite. Would love to hear more about it.
  - 6y ago
ritter2a 6y ago
AnyDSL (https://anydsl.github.io/ https://anydsl.github.io/) might be worth mentioning, it is a framework for creating domain specific languages using partial evaluation to give powerful abstractions to the user while still producing high-performance code for different target platforms. On the website, you can find that they already have DSLs for ray tracing and image stencil codes as well as works for bioinformatics.
tardismechanic 6y ago
MATLAB Coder - Generate C and C++ code from MATLAB code https://www.mathworks.com/products/matlab-coder.html https://www.mathworks.com/products/matlab-coder.html Actually the whole family of embedded code generation products from MathWorks https://www.mathworks.com/solutions/embedded-code-generation.html https://www.mathworks.com/solutions/embedded-code-generation...

chrisseaton 6y ago

> Take C (or some other language) and its mature optimizing compilers as a foundation, and write a program that compiles a high-level description of a program into C. Don't you still need many of the techniques taught in this class to build something like that? The material doesn't look inherently targeted at low-level languages or machine-code as a target.

dbcurtis 6y ago
One way to think of it is to split the "compiler" problem into three big pieces. 1) Front-end, 2) back-end, 3) tooling. Front end is lexing, parsing, building the AST (in whole or just keeping pieces of it around), semantic analysis. When that is done, you can say "Yup, valid program. It is possible to generate code." So, to your question, yes. Those techniques from the course are 100% in play. Back-end is turning the AST into something you can optimized and generate code from. (Sometimes there is a "middle-end" in there....) High-level optimizations, low level optimizations, memory allocation, register allocation. Tooling is everything that keeps your life from being miserable as a user -- starting with debug symbols and these days people expect library and package management, etc, etc. So if you are interested in exploring the Next Great Syntax or some semantic problems that can be re-written into C, then doing a C-front is a great way to do that. Let the C compiler handle all the really-really-really-hard back-end issues for you, and avoid all that distraction. But.... expect that getting debug symbols from your spiffy language into the ready-to-link object file is going to be, as they say, "non-trivial". So... great for experiments and a great learning exercise, but it's hard to do a production compiler that way.
- samps 6y ago
  Hi! This course is about the "middle end," FWIW. We do not do parsing or codegen in 6120, and there is no goal (as there is in many undergrad-level compilers courses) of making a complete, end-to-end compiler for a C-like language.

kodoque2 6y ago

Do you have a structured course for Domain-specific compilers available online?

klelatti 6y ago

Not open source but I'm working on a system that turns a DSL (embedded into Jupyter notebooks) into either c or c with opencl for a niche finance application. Not a huge project (<10k lines of code). Nature of the problem means that DSL can be very simple (it's aimed at non-programmers). This approach also means that could easily add other targets - e.g. cuda or wasm.

37ef_ced3 6y ago
Awesome. Link?
- klelatti 6y ago
  Thanks and sorry - not open source and in testing with single client at the moment.

davidpolberger 6y ago

I have just spent the last 13 months significantly reworking a compiler that turns Excel-like formulas into JavaScript. The compiler is used by an app designer, and the formula language is meant to be easy to pick up for people used to spreadsheets. A significant difference compared to Excel is that our compiler features static typing, enabling helpful error messages to be produced at build-time. (Spreadsheets, on the other hand, tend to try very hard to make sense of types at runtime, using type coercion to turn values of the wrong type into something of the expected type they can work with. That produces surprising results and confuses users. =IF(2, 3, 4) will generally produce 3, because 2 -- which should be a logical/boolean value -- is interpreted as TRUE. We think it's a better idea to warn the user that the first parameter is wrong, and to do that as soon as possible.) By far the largest challenge with reworking the compiler has been the type system. I have expanded our type system from the four standard spreadsheet types to encompass more than 80, as well as added support for lambdas (for use with functions like SUMIF and FILTER), type inference, parameterized types and union types. The biggest addition is probably an imperative mode, that executes formulas with side-effects in response to events being fired. This is all in service of features we have wanted to add for years, that have been held back by our (previously) simplistic type system. I did take a compiler class at university, more than 15 years ago, but while that proved very useful when writing the formula parser (an operator-precedence parser, initially straight out of a textbook), type theory was not part of the class. I had to pick up bits and pieces by reading the Java Language Specification (JLS) and by experimenting. The Cornell class doesn't seem to feature type theory either, which I think is unfortunate. Now that the pendulum has swung, yet again, firmly in the direction of static typing, perhaps static typing will become a bigger part of CS compiler curriculums.

tomcam 6y ago
Your project sounds like fun. What/for whom is the project? What level of Excel formula compatibility are you shooting for, e.g. just in the neighborhood like Google Sheets, reasonably close as in Libre Office, or full compatibility with the latest version? If the latter you must have one gnarly test suite. Would love to hear more about it.
- 6y ago

ritter2a 6y ago

AnyDSL (https://anydsl.github.io/ https://anydsl.github.io/) might be worth mentioning, it is a framework for creating domain specific languages using partial evaluation to give powerful abstractions to the user while still producing high-performance code for different target platforms. On the website, you can find that they already have DSLs for ray tracing and image stencil codes as well as works for bioinformatics.

tardismechanic 6y ago

MATLAB Coder - Generate C and C++ code from MATLAB code https://www.mathworks.com/products/matlab-coder.html https://www.mathworks.com/products/matlab-coder.html Actually the whole family of embedded code generation products from MathWorks https://www.mathworks.com/solutions/embedded-code-generation.html https://www.mathworks.com/solutions/embedded-code-generation...