[Mono-dev] Proposed Patch - Google Native Client

[Rafael Teixeira]

Looks promising, thanks for the effort, and specially the stricter checks
result summary.

Rafael "Monoman" Teixeira
---------------------------------------
"To be creative means to be in love with life. You can be creative only if
you love life enough that you want to enhance its beauty, you want to bring
a little more music to it, a little more poetry to it, a little more dance
to it."
Osho


On Tue, Jun 22, 2010 at 2:29 PM, Elijah Taylor <elijahtaylor at google.com>wrote:

> Greetings Mono Developers,
>
> Attached is a patch to support 32-bit x86 code generation for Google Native
> Client (http://code.google.com/p/nativeclient/).  I encourage you to
> browse our project for more information if you're curious.  I apologize for
> the large diff, let me try to explain the highlights to make it easier to
> digest.
>
> There is a code generation component (define: __native_client_codegen__)
> which affects the Mono bytecode -> native code generation for x86-32.  There
> are a set of alignment restrictions, illegal instructions, and replacement
> instructions we use for Native Client to ensure proper control-flow
> sandboxing.  Please see
> http://nativeclient.googlecode.com/svn/data/docs_tarball/nacl/googleclient/native_client/documentation/nacl_paper.pdffor more details.
>
> There is also a runtime component (define: __native_client__) which
> modifies or disables some functionality to be compatible with the Native
> Client runtime.
>
> We also had to modify some code that doesn't fall under either of the above
> defines.  Most of these changes revolved around type safety.  The modified
> version of gcc we use to compile Native Client modules is more strict about
> types, and it caught what look like legitimate issues with the Mono
> codebase.  The largest issue in terms of number of errors was the use of
> mono_bool and gboolean interchangeably between declaration and definition of
> many functions.  gboolean is defined as an "int" but mono_bool is defined as
> int32_t.  Other type issues are listed directly below.  Feedback is
> appreciated on these changes because of our unfamiliarity with this code,
> but I modified these in the way that seemed most "right" at the time.
>
> mono/metadata/decimal.h:47 mono_decimal2string int -> gint32
> mono/metadata/filewatcher.h:28 gboolean -> int
> mono/metadata/filewatcher.c:158 int32 -> gint32
> mono/metadata/threads-type.h:64 int -> gint32
>
> mono/mini/mini.h:1546  gboolean sort_end -> int sort_type
> mono/mini/mini.h:1733  gboolean fp -> int bank
>
> The last bit of modification is to genmdesc and the Makefiles in general.
> We added a new flag to genmdesc called "nacl" which overrides the given max
> length of an instruction.  Native Client code tends to be larger because of
> some of the instruction requirements we have, so some of the instructions in
> cpu-x86.md had to be modified.  This is all tied to a new configure flag
> called "--enable-nacl-codegen", which enables the codegen define, and sets
> up calls to genmdesc with a --nacl flag.  It also modifies the mono-wrapper
> script so one aspect of our code generation rules (masking jump targets to
> 32-byte boundaries) is turned off while compiling and testing mono from the
> Makefiles, which is required when testing outside of the Native Client
> environment.  We're also including a standalone check "fsacheck" which tests
> mono code generation as full AOT and a the library linked into a fully
> static executable.
>
>
> I look forward to your comments, questions, and suggestions.
>
>
> -Elijah Taylor
> Google Native Client Team
>
> _______________________________________________
> Mono-devel-list mailing list
> Mono-devel-list at lists.ximian.com
> http://lists.ximian.com/mailman/listinfo/mono-devel-list
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.ximian.com/pipermail/mono-devel-list/attachments/20100625/d8b3eee7/attachment.html