[Mono-dev] [PATCH] more support for Google Native Client

Elijah Taylor elijahtaylor at google.com
Tue Dec 21 15:12:53 EST 2010

Greetings Mono developers!

*[tl;dr  very large patch for Native
Client<http://www.chromium.org/nativeclient> support
hosted here <https://github.com/elijahtaylor/mono>, would love feedback and
many eyes to look at it]

I'm back with another round of changes for supporting Google's Native Client
(NaCl), including support for amd64, JIT compilation, and Garbage
Collection.  It's a large set of changes, forked on Dec 14 in github @
https://github.com/elijahtaylor/mono.  I would appreciate feedback on these
changes... to facilitate this, I'll try to explain the largest changes by
feature (please email if clarification is needed):

*1) amd64 codegen*

   - Rules located here:
      - Removed %r15 from register allocation, LMF save/restore, etc.  (r15
      is special and not modifiable by untrusted code)
      - Sandbox all data access through membase address mode.  If not %rsp
      or %rbp relative, re-write as clearing upper 32-bits + memindex
      - align functions, call sites
      - Sandbox returns and all indirect jumps (need to be 32-byte aligned,
      cleared upper 32-bits)
      - Never omit frame pointer as general operations to rbp aren't allowed

*2) NaCl x86-64 is ILP32 (this is the largest set of changes and may make
some mono devs unhappy)*

   - Set SIZEOF_REGISTER == 8 while sizeof(gpointer) == 4 for NaCl amd64 (we
   can use 8-byte instructions, but pointers are 4-bytes)
   - Re-write large portions of mini-amd64.c, tramp-amd64.c,
   exceptions-amd64.c, mini.c, method-to-ir.c to use appropriate sizes
   (SIZEOF_REGISTER, sizeof(gpointer), literal '8').  *These changes are
   disruptive, but ultimately they should be more correct than what was there
   before.  *It's our opinion that these changes actually improve Mono
   despite their impact.
   - We only generate NaCl amd64 code from an ILP32 machine (either a 32-bit
   application for AOT code, or NaCl runtime JIT), so we may not have caught
   all of the [8 <--> SIZEOF_REGISTER] conversions, but we likely caught most
   of the [sizeof(gpointer) <--> SIZEOF_REGISTER] and [8 <--> sizeof(gpointer)]
   changes that are necessary.
   - Change atomic operations and default pointer directives to use 32-bit
   instructions (long instead of quad)
   - Change default operations to use 32-bit integers/pointers (eg,
   OP_LOAD_MEMBASE uses 4-bytes instead of 8)

*3) JIT support for NaCl*

   - Since we're unable to emit code directly in its final executable
   location, we instead:
      - reserve a buffer on the heap
      - create a hash table entry mapping the temp location and final
      - modify all non-local patches relative to the final location
      - request the NaCl runtime to install the created code in the final
   - See mono/utils/mono-codeman.c changes for more detail.
   - For every codeman *reserve*, we must add a codeman *validate* call in
   order to install the method/trampoline/blob in the final location (as well
   as validate it for NaCl, pad it out, etc)
   - We don't delete or reuse code  (we can, but it's icky and the benefits
   don't outweigh the cost)
   - Backpatching changed to use NaCl syscalls to modify existing dynamic

*4) GC support for NaCl (boehm only)*

   - NaCl compiler and Mono code generator both emit instrumentation at GC
   "safe points" (back branches and function prologs), for cooperative thread
   parking (we're not allowed to send and receive signals)
   - Added new opcode OP_NACL_GC_SAFE_POINT to handle mono instrumentation
   - modified pthread_stop_world.c and pthread_support.c somewhat
   extensively to support this new way of stopping the world
   - wrapped pthread_exit because NaCl doesn't support pthread cleanup
   - added machine type NACL to libgc with machine specific defines

*5) Misc bug fixes (not NaCl-specific)*

   - fix *x86_memindex_emit* when disp is 32-bit
   - properly exclude code in libgc/gc_dlopen.c when DYNAMIC_LOADING not
   - properly exclude code based on DISABLE_SOCKETS by including config.h
   before checking define
   - clean up calculation of offset for amd64 AOT specific trampoline args
   - fix bug in *mono_bblock_insert_before_ins* when trying to insert an
   instruction to the beginning of an existing basic block.
   - fix small typo bug in genmdesc.pl which kept amd64 from being able to
   be a target of cross compiling
   - fix struct passing in amd64 with sizeof(struct) == 16 when fields
   aren't 8-byte aligned (eg, first field is 12 bytes, second field is 4
   bytes), pass on stack instead of in registers (mini-amd64.c:*
   - add extra checks to mini-amd64.c:*mono_arch_emit_exceptions* to keep
   exception/R4/R8 emitting from overflowing a buffer silently
   - fix bugs in *new_codechunk* and *mono_code_manager_reserve_align* which
   allowed unaligned code to be allocated.

I know we're close to holidays so I don't have any delusions that these
changes will get in by the end of the year :)  Please feel free to pick
apart these changes and let me know if there are things that should be

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.ximian.com/pipermail/mono-devel-list/attachments/20101221/6daa2505/attachment.html 

More information about the Mono-devel-list mailing list