3 The Mono runtime implements a JIT engine for the CIL virtual
4 machine (as well as a byte code interpreter, this is to
5 quickly port it to new systems), the class loader, the garbage
6 collector, threading system and metadata access libraries.
8 We currently have two runtimes:
11 * <b>mono:</b> The Just In Time compiler implemented
12 using a BURS instruction selector. We only support
13 x86 machines in the JIT engine at this point.
15 * <b>mint:</b> The Mono interpreter. This is an
16 easy-to-port runtime engine.
19 Currently we are using the Bohem conservative garbage
20 collector, but we working on incorporating the ORP GC engine.
22 ** Executing MSIL/CIL images
24 The code will load an executable and map the references to
25 external assemblies to our own version of the assemblies on
28 Our roadmap looks like this, this has been updated as of
33 * Milestone 1: <b>Done</b> Fully read and parse all CIL byte-codes
34 and metadata tokens (ie, a disassembler).
36 * Milestone 2: <b>Done</b> Complete an interpreter for CIL byte
37 codes. This interpreter can be used temporarly to
38 run CIL byte code on a system where no JIT is
41 * Milestone 3: <b>Done</b>Define an <i>lburg</i>-like
42 instruction selector for the JITer for Intel.
44 * Milestone 4: <b>Done</b> Implement JITer. This is where our
45 current efforts are focused on, the JITer currently runs
46 all of the code we have tested on it. The major limitation
47 is that our class libraries are not complete, and hence not
48 every application can be ran.
50 * Milestone 5: Port of the JITer to non IA32 systems.
53 A setup similar to the Kaffe JIT engine will be used to
54 layout the code to support non-IA32 architectures. Our work
55 will be focused on getting a IA32 version running first.
57 The JIT engine works on Linux and Win32, although you
58 will need to install the CygWin32 development tools to get a
59 Unix-like compilation environment (mostly we use GNU make in
60 a few of the makefiles).
62 ** JIT Engine (<b>updated, July 8th, 2002</b>)
64 The JIT engine uses a code-generator generator approach for
65 compilation. Given the properties of CIL byte codes, we can
66 take full advantage of a real instruction selector for our
69 The JIT engine implements a number of optimizations:
72 * Opcode cost estimates (our architecture allows
73 us to generate different code paths depending
74 on the target CPU dynamically).
80 Although compilers typically do
81 constant folding, the combination of inlining with
82 constant folding gives some very good results.
84 * Linear scan register allocation. In the past,
85 register allocation was our achilles heel, but now
86 we have left this problem behind.
89 There are a couple of books that deal with this technique: "A
90 Retargetable C Compiler" and "Advanced Compiler Design and
91 Implementation" are good references. You can also get a
92 technical description of <a
93 href="http://research.microsoft.com/copyright/accept.asp?path=http://www.research.microsoft.com/~drh/pubs/iburg.pdf&pub=ACM">lbrug</a>.
95 A few papers that describe the instruction selector:
98 * <a href="http://research.microsoft.com/copyright/accept.asp?path=http://www.research.microsoft.com/~drh/pubs/interface.pdf&pub=wiley">A code generation interface for ANSI C</a>
101 * <a href="http://research.microsoft.com/copyright/accept.asp?path=http://www.research.microsoft.com/~drh/pubs/iburg.pdf&pub=ACM">Engineering efficient code generators using tree matching and dynamic programming.</a>
107 We are evaluating the future directions for the JIT engine:
108 both from our needs (optimizations like inlining, better register allocation,
109 instruction scheduling, and porting to other CPUs).
111 We have not yet decided how we will evolve the JIT engine. We
112 might just upgrade our current architecture, and provide optimizations as
115 ** Garbage Collection
117 Currently we are using the Boehm conservative GC. Although our plans
118 are to move to the Intel ORP GC engine, our plans on a next generation
119 dual-JIT engine have to be taken into account.
121 We will be using the Intel ORP GC engine as it provides a precise
122 garbage collector engine, similar to what is available on the
125 Although using a conservative garbage collector like Bohem's
126 would work, all the type information is available at runtime,
127 so we can actually implement a better collector than a
128 conservative collector.
131 * Garbage collection list and FAQ:<br>
132 <a href="http://www.iecc.com/gclist/">http://www.iecc.com/gclist/</a>
134 * "GC points in a Threaded Environment":<br>
135 <a href="http://research.sun.com/techrep/1998/abstract-70.html">
136 http://research.sun.com/techrep/1998/abstract-70.html</a>
138 * "A Generational Mostly-concurrent Garbage Collector":
139 <a href="http://research.sun.com/techrep/2000/abstract-88.html">
140 http://research.sun.com/techrep/2000/abstract-88.html</a>
142 * Details on The Microsoft .NET Garbage Collection Implementation:<br>
143 <a href="http://msdn.microsoft.com/library/default.asp?url=/library/en-us/dnmag00/html/GCI.asp">http://msdn.microsoft.com/library/default.asp?url=/library/en-us/dnmag00/html/GCI.asp</a>
144 <a href="http://msdn.microsoft.com/library/default.asp?url=/library/en-us/dnmag00/html/GCI2.asp">http://msdn.microsoft.com/library/default.asp?url=/library/en-us/dnmag00/html/GCI2.asp</a>
149 The ECMA runtime and the .NET runtime assume an IO model and a
150 threading model that is very similar to the Win32 API.
152 Dick Porter has been working on the Mono abstraction layer
153 that allows our runtime to execute code that depend on this
158 Paolo Molaro found a few interesting links:
161 * On compilation of stack-based languages:<br>
162 <a href="http://www.complang.tuwien.ac.at/projects/rafts.html">
163 http://www.complang.tuwien.ac.at/projects/rafts.html</a>
165 * A paper on fast JIT compilation of a stack-based language:<br>
166 <a href="http://www.research.microsoft.com/~cwfraser/pldi99codegen.pdf">
167 http://www.research.microsoft.com/~cwfraser/pldi99codegen.pdf</a>
169 * Vmgen generates much of the code for efficient virtual machine (VM)
170 interpreters from simple descriptions of the VM instructions:<br>
171 <a href="http://www.complang.tuwien.ac.at/anton/vmgen/">
172 http://www.complang.tuwien.ac.at/anton/vmgen</a>
177 PInvoke is the mechanism we are using to wrap Unix API calls
178 as well as talking to system libraries.
180 Initially we used libffi, but it was fairly slow, so we have
181 reused parts of the JIT work to create efficient PInvoke trampolines.
185 Mono has support for remoting and proxy objects, just like
186 .NET does. The runtime provides these facilities.
190 If you are interested in porting the Mono runtime to other
191 platforms, you might find the pre-compiled <a
192 href="archive/mono-tests.tar.gz">Mono regression test
193 suite</a> useful to debug your implementation.
197 We plan on adding support for XPCOM on Unix and COM on Microsoft
198 Windows later in our development process.