llvm dev - Jul 2005 - [LLVMdev] Source level debugging

Hi,

I am planning to use LLVM for a compiler validation
project that checks if the optimized version of the
code is the correct translation of the unoptimized
code. And I have a couple of questions.

I need two pieces of information from the compiler: 

1) Some kind of mapping between the structure of the
unoptimized and optimized code. It seems, that can be
obtained via the llvm.dbg.stoppoint intrinsic
functions.

2) Mapping between the variables of the source code
(or  unoptimized LLVM code) and  the
variables/registers in the optimized code. My
understanding is that such mapping could be a part of
the debug information, but has not been implemented
yet. It seems that LLVM symbol table is not designed
to keep such information either.
Any suggestions on how to obtain this mapping? 

As a side note, does anyone have examples of tail call
elimination performed with debugging information
enabled? It seems that it should be possible according
to the Source Level Debugging with LLVM:
"you get accurate stack traces despite tail call
elimination and inlining),". However, in my examples, 
%llvm.dbg.stoppoint and %llvm.dbg.region.end get
inserted between the tail call and the return
statement, which stop tailcallelim unless I remove
them manually.    

Thank you for your help,
Anna.

On Fri, 29 Jul 2005, Anna Zaks wrote:
> I am planning to use LLVM for a compiler validation
> project that checks if the optimized version of the
> code is the correct translation of the unoptimized
> code. And I have a couple of questions.
ok
> I need two pieces of information from the compiler:
>
> 1) Some kind of mapping between the structure of the
> unoptimized and optimized code. It seems, that can be
> obtained via the llvm.dbg.stoppoint intrinsic
> functions.
This is very hard.
> 2) Mapping between the variables of the source code
> (or  unoptimized LLVM code) and  the
> variables/registers in the optimized code. My
> understanding is that such mapping could be a part of
> the debug information, but has not been implemented
> yet. It seems that LLVM symbol table is not designed
> to keep such information either.
> Any suggestions on how to obtain this mapping?
Not really.  You could implement debug info, that's really the only way to 
get what you want.
> As a side note, does anyone have examples of tail call
> elimination performed with debugging information
> enabled? It seems that it should be possible according
> to the Source Level Debugging with LLVM:
> "you get accurate stack traces despite tail call
> elimination and inlining),". However, in my examples,
> %llvm.dbg.stoppoint and %llvm.dbg.region.end get
> inserted between the tail call and the return
> statement, which stop tailcallelim unless I remove
> them manually.
I haven't tried it with tail call elim, though inlining certainly works.

-Chris

-- 
http://nondot.org/sabre/
http://llvm.org/