initial dietc commit

1 files changed, 76 insertions, 0 deletions

diff --git a/README.txt b/README.txt
new file mode 100644
index 0000000..ca59161
--- /dev/null
+++ b/README.txt
@@ -0,0 +1,76 @@
+# dietC: a C backend for chibicc
+
+What: desugar C code
+
+Why: easier program analysis and transformation
+
+How: fork of Rui Ueyama's great chibicc compiler, replacing the x86 assembly
+backend with a C backend
+
+Quick Example:
+
+  $ cat tests/super_simple.c
+  int main(void) {
+    int x = 1;
+    int y = 2;
+    return y;
+  }
+  $ make
+  ...
+  $ ./dietc tests/super_simple.c
+  #include "/home/matthew/repos/dietc/scripts/dietc_helpers.h"
+  typedef int Type_1 ;
+  ...
+  extern Type_5 main ;
+  ...
+  Type_1 main ( ) {
+          Type_1 t1 ;
+          ...
+          t3 = 1 ;
+          * t5 = t3 ;
+          ...
+          return t1 ;
+  }
+
+Selling points:
+
+  - Guaranteed space-separated tokens
+  - Extremely restricted grammar, super easy to parse; see docs/LANGUAGE.txt
+  - Maintains all type information
+  - Comes with a wrapper to replace GCC in existing build scripts; see
+    docs/DIETCC.txt
+
+Known to be unsupported:
+
+- dietC is ... optimistic about identifier conflicts with its generated
+  labels. It likely will not work if you run it on its own output. This is very
+  fixable, will do soon.
+- chibicc does not have good support for long double; specifically, long
+  doubles cannot be used in constant expressions. This is fixable, but requires
+  non-trivial modifications to the chibicc side of the code.
+- chibicc does not parse const, volatile, etc. type modifiers. it would not be
+  too hard to support this if desired.
+
+Why not cilly?
+
+  - cilly performs both not enough and too much lowering.
+  - it often compiles a struct field acces a->foo into a literal offset like
+    *(a + 16).
+  - even after simplification it has multiple control flow constructs: loop,
+    etc.
+  - it's written in OCAML, and hard to parse the result outside of OCAML
+
+Why not C--, assembly, QBE?
+
+  - all these are strictly lower level than cilly. They preserve almost no type
+    information at all.
+
+Why not LLVM?
+
+  - LLVM has breaking changes approximately every 5ns and preservation of
+    high-level type information is not guaranteed.
+  - mostly forced to use LLVM's C++ interfaces and then compile with LLVM
+
+License:
+  - the underlying chibicc code is licensed under the MIT license
+  - my modifications are licensed under the AGPLv3