Re: Backtraces broken on i386 by unorthodox encoding of push %ebp

[Richard Smith <richard at ex-parrot dot com>]

Jan Kratochvil wrote:

> You should always provide unwind information and then it is irrelevant which
> instructions and/or which their coding you use.

Yes, you're right, I could quite easily.  But a design 
criterion for this project is that it is easy for 
third-parties to produce backtraces from production code 
(i.e. once debugging information has been stripped).

I think on balance, I should just fix the prologue to use 
the one-byte push instruction.  (My reluctance to do that 
was because the in question code is just-in-time compiled 
code produced by a lightweight compiler/assembler that says 
resident in memory.  I'm not keen to change the assembler 
component from the upstream version as the change to fix it 
looks more intrusive than you might imagine.)


Joel Brobecker wrote:

> But if he cannot get the unwind information with the 
> binary, it might indeed be easy enough to improve the 
> prologue analyzer for x86. On x86_64, on the other hand, 
> we made the conscious design decision to rely more on 
> unwind info, so it's a little more questionable whether 
> we'd want to do that for this target.
>
> The thing I wanted to add to Richard is that there is a 
> very high chance that the only way to get this fixed in 
> GDB is by submitting a patch himself (to gdb-patches).

I'll certainly take a look at it to see if it's as 
straightforward as it seems.

However, I expect the same problem will arise on x86_64, and 
I can understand not wanting to complicate unnecessarily its 
prologue analyzer.

Richard