[llvm-dev] GPL Software

[jacob navia via llvm-dev]

Le 25/08/2015 20:39, Renato Golin a écrit :
> I'm also interested in knowing what you did. I think most people here
> would be glad to know your peril and how you solved it.

Stack layout:

C++ code calls
    |
    |
JITted C code, compiled and linked on the fly tha calls
    |
    |
C++ code that throws

The throw must pass through the JITted code to arrive at the catch at 
the top.

So, the JIT must generate and install the CIE/FDE DWARF4 debug info to 
inform the stack unwinding virtual machine how to do it.

Problems solved:

1) Do not read the DWARF docs. They give you a general idea but nobody 
follows it.
2) Do not try to *understand* anything. It is hopeless and you will just 
lose your time.
3) Try to generate EXACTLY what llvm is generating.

So, I extracted the __eh_frame section from a simple C program.

otool -s __TEXT __eh_frame hello.o

I figure out from the hexadecimal dump where the CIE lies. Then, I 
generate exactly the same.

Then, generate the FDEs for each function. Again, use the dump of the 
__eh_frame section as a guide but add your own virtual machine 
instructions. Of course do not copy blindly what llvm generates since 
some fields like length, etc mst be changed!

Then the problem arises:

You have to inform the running program that a new piece of code has ben 
JITted and add the eh_frame info dynamically. That is the question I 
asked here a month ago.

Since there were no answers, I tried the code that uses the

  _keymgr_get_and_lock_processwide_ptr

as specified in some Apple's program source.

Do not do that. It will just never work, it returns always zero.

Then I tried to use what I use on Linux:

register_table_bases with
void __register_frame_info_table_bases (void *begin, struct object *ob, 
void *tbase, void *dbase);
This function requires a different format than register_frame_info, 
quite difficult to generate.

Do not bother, it doesn't work, even if it is advertised as "working in 
OS X now".

Then I tried to use
register_frame_info with yet another format. Didn't work either. 
Desperate to know WHY it wasn't working I followed the machine on 
assembly, instruction by instruction until I got there. The reason was 
obvious: That function consists of just...

a return statement. Nothing more. It is decoy to avoid giving you a 
linker error.

Then  I was TRULY desperate and asked Apple. That worked. The friendly 
people at the Apple groups told me to look into the code of the dynamic 
loader that does exactly what I was trying to do OF COURSE!

I compiled that. Do not use the xcode project coming with the software 
because I got incredible strange debugger problems and could not debug 
it. Using a simple Makefile works.

Then, I could follow ALL the process of calling __register_frame(); and 
that function works. It expects an FDE (not a CIE as I thought) but 
reads the CIE using the CIE pointer in the given FDE for each FDE.

Now, the throw of the C++ code below my JIT code is catched by the C++ 
code above.

GOSH!

This is just a quick description. I do not speak about all the wrong 
starts, the dead ends, the sheer frustration at the lack of docs, etc. 
This is the result of this part of the compiler being very complex by 
design (DWARF is incredibly complex), and a total lack of documentation. 
For instance the entry for __register_frame_info in GNU docs is just:

"DOCUMENT THIS!"

On top of this, Apple has modified the DWARF specs to compress the 
information, and it uses a public domain library (libunwind) that has 
been heavily modified. If this isn't a MESS I do not know what a mess is.

Anyway, I got there in just a month of work. I am getting better, under 
linux it took me three months, but that was the first time.

jacob

P.S. Sorry for the message about GNU software. I was mad at you, but 
actually is not your fault. You have also your share of hard work. I 
suppose that at a certain level of complexity there are always VERY FEW 
people that know anything, and you have to figure it out yourself 
without any help.