[Mono-dev] How To Convince Mono To Allocate Big Arrays

Luis F. Ortiz LuisOrtiz at Verizon.Net
Mon May 5 19:51:59 EDT 2008

Back in September ("Big Arrays, Many Changes --- Request for Advice")  
I asked folks for advice on how to go about adding the capability to  
Mono to allocate arrays with more than Int32.MaxValue elements.  After  
playing around with it for a few months, I'm at the point where I have  
an implementation that mostly works, with a couple of warts which  
could probably be quickly fixed by someone who knows more than I do  
about Mono internals. I spoke with Miguel about these patches, and he  
encouraged me to post them to mono-dev as soon as I got them to pass  
the "make check" test suite.  So here I am a week later.

I want to start by going over the changes themselves, what  
alternatives there might be to what I had done and what flaws I know  
to exist in the implementation.

First off, though Microsoft chose for some reason NOT to implement  
large array allocation, the necessary APIs are in the .NET  
specification.  For example, in the System.Array class, we find:

	public long GetLongLength(int dimension);
	public long LongLength { get; }
	public static Array CreateInstance(Type elementType, params long[]  
	public Object GetValue(long index);
	public void SetValue(Object value, long index);
	... (other overloads omitted, but there)
and we find that the Newarr  opcode takes a natural int or an Int32 as  
the length, so the bytecode level is ready too.

Mono as of 1.2.6 already implemented most (all?) of the necessary  
interfaces in mcs/class/corlib/System/Array.cs, but they all cast down  
their long arguments down to integers as the underlying implementation  
was not there.

So the first set of changes were to:

In object.h I made three changes:

	1) Changed MonoArrayBounds to use gsize instead of guint32 as the  
type for length and lower_bound,
	2) Changed MonoArray to use gsize instead of guint32 as the type of  
	3) Changed the prototypes for mono_array_new(),  
mono_array_new_full(), and mono_array_new_specific() to
	take gsize's instead of guint32's for their size and bounds parameters.
	-mono_array_new (MonoDomain *domain, MonoClass *eclass, guint32 n);
	+mono_array_new (MonoDomain *domain, MonoClass *eclass, gsize n);

	 mono_array_new_full        (MonoDomain *domain, MonoClass  
	-                            guint32 *lengths, guint32 *lower_bounds);
	+                            gsize *lengths, gsize *lower_bounds);

  	MonoArray *
	-mono_array_new_specific            (MonoVTable *vtable, guint32 n);
	+mono_array_new_specific            (MonoVTable *vtable, gsize n);

This is the first place an alternative shows up.  ¿Which type is  
better: gsize or gssize?  The unsigned type gsize better matches the  
type memory allocation functions use (size_t or some variant) and the  
existing guint32, but the signed type better matches the interface  
presented at the top level (i.e. CreateInstance).  I chose the  
unsigned alternative, but an argument could be made for the signed  
type.  Another alternative would be to create 64 bit versions of the  
mono_array_new(), mono_array_new_full(), and mono_array_new_specific()  
functions, but that seemed to be too much work.

The changes in object.c add the implementations of the modified  
mono_array_new(), mono_array_new_full(), and mono_array_new_specific()  
functions.  There was some confusing #defines around MYGUINT32_MAX  
that I did not like, but rather than replace that cruft, I extended it.

The changes in icall-def.h add two new method calls to the array  
object, CreateInstanceImpl64() and GetLongLength().  It might be  
possible to avoid the CreateInstanceImpl64() definition and make it an  
overload of CreateInstanceImpl() with long parameters, if I was sure  
on how to do that.

The changes to icall.c tweak the implementation of  
ves_icall_System_Array_CreateInstanceImpl() to change the type of the  
sizes array and add the implementations of  
ves_icall_System_Array_CreateInstanceImpl64() and  

The change to socket-io.c was to tweak its usage of  
mono_array_new_full to use gssize instead of int for for the array of  

So all these changes get us to the point where the basic foundation is  
laid down, but there is still the JIT to contend with.  It requires a  
few more files to be changed:

The changes in mini.c change the signature of mono_array_new and  
mono_array_new_specific to take "int" instead of "int32".

The changes in jit-calls.c do the boring change of guint32's into  

The changes in exceptions.cs split the test case for test_0_array_size  
into a 32 and 64 bit variant, because an allocation of Int32.MaxValue  
can succeed after these changes are applied.

There was only one touch-up needed in the the C# compiler:  the  
GetLength property code special-inlined code-generation needed to be  
tweaked since it is now possible to get an array length that will not  
fit into an I4.  Changing mcs/mcs/ecore.cs and mcs/mcs/expression.cs  
to use OpCodes.Conv_Ovf_I4 after OpCodes.Ldlen instead of  
OpCodes.Conv_I4  fixed that.

Oh, yeah, and ALL the long method calls in mcs/class/corlib/System/ 
Array.cs needed to be converted over to use CreateInstanceImpl64() and  
GetLongLength().  The SetValue() and GetValue() implementations still  
need work, but since there are no unit tests for those methods, I put  
them off.

That gets us to the point where we can allocate a large array, but it  
does not let us index a large array.   I changed the following files  
to start the process of converting the indexing operations to do  
bounds checking against the now 32/64 bit length of arrays and to  
index using a 64/32 bit index:

In inssel-amd64.brg, I changed the macro  
OP_AMD64_ICOMPARE_MEMBASE_REG and hacked CEE_LDELEMA to no longer use   
the faster OP_X86_LEA because I was not sure how to generalize it.  
Perhaps MONO_EMIT_NEW_AMD64_ICOMPARE_MEMBASE_REG should be retired and  
explicit I4 and I8 versions substituted where appropriate.

In mini-amd64.c I added the OP_AMD64_COMPARE_MEMBASE_REG_I8 opcode as  
being the same as OP_AMD64_ICOMPARE_MEMBASE_REG, except with an  
operand size of 8 bytes instead of 4.

In mini-ops.h and cpu-amd64.md I added an entries for  
OP_AMD64_COMPARE_MEMBASE_REG_I8 and amd64_compare_membase_reg_i8.

These changes seem to get 64 bit indexing working, and passed all the  
regression tests in 1.2.6, but in 1.9.1 a regression test called  
test_0_regress_75832() breaks.  It could be that the changes I made in  
mono/mini are incorrect.  I am sure the changes are incomplete, and I  
have not considered what to do to other architectures.

Advice or assistance is greatly appreciated.

Luis F. Ortiz
Official Mono Modifier
Interactive Supercomputing,  Inc.

PS:  Here are the changes proper:

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.ximian.com/pipermail/mono-devel-list/attachments/20080505/8f8eb396/attachment-0002.html 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: BigArrays.patch
Type: application/octet-stream
Size: 27156 bytes
Desc: not available
Url : http://lists.ximian.com/pipermail/mono-devel-list/attachments/20080505/8f8eb396/attachment-0001.obj 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.ximian.com/pipermail/mono-devel-list/attachments/20080505/8f8eb396/attachment-0003.html 

More information about the Mono-devel-list mailing list