I am not sure if I have hit work in progress code here....

my arm-embedded app crashes during initialization, the reason is that now exceptions are part of the rtl but there is no properly initialized heapmanager:

Procedure InitExceptions;
{
   Must install uncaught exception handler (ExceptProc)
   and install exceptions for system exceptions or signals.
   (e.g: SIGSEGV -> ESegFault or so.)
}
begin
   ExceptionClass := Exception;
   ;
   // Create objects that may have problems when there is no memory.
   //CRASH BOOM BANG in the following line
   OutOfMemory:=EOutOfMemory.Create(SOutOfMemory);

Creating the object fails because the MemoryManager is not properly initialized yet.

When I put heapmgr unit before sysutils im my program (this should not be necesary, right???) I can debug a little deeper in object initialization but still the root cause exists that there has never been a call to RegisterHeapBlock so objects cannot be initiated because there's no memory block defined that heapmgr can draw memory from.

I can hack the RegisterHeapBlock call into the initialization code of the unit heapmgr but this is more a hack than a proper solution --> seems a bad idea to me.

How can this be properly solved?

Michael

Yes, this is a requirement.

I think allocating a small block (128 bytes for arm?) during the start up of heapmgr should be fine?

var
  InitialBlock: array[0..255] of byte;
...
RegisterHeapBlock(@InitialBlock,sizeof(InitialBlock));

If one includes heapmgr he wants to use the heap after all.

Simple solutions are the best ;-) Thanks for opening my eyes to this easy approach.

I have created a patch for heapmgr.pp, could I ask you to commit it to svn?

For a little more flexible solution I would like to ask you one more questions, it may seem dumb to you but I have never been that deep into a compiler before so I am learning a lot atm.

I can see that the compiler param -Ch can be used to define the heapsize, I can also see in the assembler code of my main program that when I set -Ch I also get a section:

-Ch2048 results in this:

.section .data.n___heapsize
.balign 4
.globl __heapsize
__heapsize:
.long 2048

but I do not seem to be able to access __heapsize in my code, how could I use this information to make the heapsize configurable?

and, it is (as far as I understand) a must to include heapmgr in the code because if not the program will always die on startup because of the object created in Exception class, where would be the right place to do this?

Yes, I can put heapmgr as the first unit included in my main program, but wouldn't it be better to put it in some global initialization loop, like the VariantManager, which always gets initialized without me triggering this in any way.

Michael

Index: rtl/embedded/heapmgr.pp
=================================================================== --- rtl/embedded/heapmgr.pp (revision 23681)
+++ rtl/embedded/heapmgr.pp (working copy)
@@ -34,6 +34,7 @@

      var
        Blocks: PHeapBlock = nil;
+      InitialBlock: array[0..255] of byte;
      procedure InternalFreeMem(Addr: Pointer; Size: ptruint); forward;

@@ -242,6 +243,7 @@

  initialization
    SetMemoryManager(MyMemoryManager);
+  RegisterHeapBlock(@InitialBlock,sizeof(InitialBlock));
  finalization
    //FinalizeHeap;
  end.

Did you try

=== code begin ===

var
   heapsize: LongWord; external name '__heapsize';

=== code end ===

?

Regards,
Sven

This helped me to understand another puzzle piece.

Please correct me if I am wrong:

sram - Memory is divided in several areas on arm-embedded:

_data starts at start of sram ($20000000 in my case) _edata is end of _data segment ($200009b4)
_bss_start defines start of static vars ($200009b4) _bss_end defines end of static vars ($20000e90)

all memory from here on is used for stack, stack grows from end of avaliable sram ($20010000) downto _bss_end

so what I could do is allocate memory for heap starting at address _bss_end with the size of _heapsize. or, perhaps even better:

I start using memory at address _stack_top - _stklen - _heapsize

Correct ??

Michael

That would of course be even better, the compiler could calculate if vars+stack+heap fit in the given memory configuration and issue at minimum a warning or even better a compile error.

But this is definitely something that is out of my league, Florin, could you implement this?

Michael

I'll look into this the next days.

Florian, did you find some spare time to work on this?

Michael

Florian, did you find some spare time to work on this?

Not yet but still on my todo list.

Am 03.03.2013 11:04, schrieb Michael Ring:

In r23853 I committed a patch which creates automatically a heap on the embedded targets if heapmgr is used. Further, sysutils depends now on heapmgr.

Thany you very much, will give it a try later or tomorrow.

Michael