This is the mail archive of the gdb-patches@sourceware.org mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [RFC][patch] Allow to disassemble line.


On Mon, Oct 19, 2009 at 09:37:25PM +0200, Eli Zaretskii wrote:
> It's really hard to believe that 1 or 2 or 3 extra columns can make
> such a difference.  Listing the source, GDB is unable to show most of
> the lines in the call stack without wrapping them, and we live with
> that.  How on Earth can disassembly do worse?

Disassembly can do worse because so much of the line is usually
unintersting.  The three columns won't be a huge difference, even
though they're a noticeable percentage of the screen real estate;
mostly they prompted me to start this conversation.

I usually use a 104 column terminal.  Here's a sample from GDB; the
places most likely to wrap and be hard to write are calls:

0x080ec87f <handle_inferior_event+3087>:        call   0x8101d90 <gdbarch_deprecated_function_start_offset>

GDB wrote that on one line and let it hard wrap.

Or one from GNU libstdc++:

0xc6bc7 <_ZSt17rethrow_exceptionNSt15__exception_ptr13exception_ptrE+87>:
    callq  0x58e20 <_Unwind_RaiseException@plt>

GDB wrapped that before the callq.  I can't figure out when it does
this versus when it lets things run off the end of the line; I've seen
both today.

Neither of these are unusual functions.

> > We've got this layout here with nicely aligned columns but tons of
> > whitespace.  In a halfway modern world we could do this with color.
> > Or bold the address of the current instruction.
> 
> I have nothing against this, but it will have to be an option, because
> dumb terminals, non-Posix emulations of termcap, and some front ends
> will need to turn it off (and then the arrow should re-appear).

Agreed.  I hope that we could do this automatically in 99% of cases,
using the termcap capabilities and isatty.  We do already rely on
readline which requires tgetent.

> > Another possibility would be to factor out the name of the function.
> > Something like this:
> > 
> > (top) disas
> > Dump of assembler code for function gdb_main:
> >    0x0000000000454c9e (+0):     push   %rbp
> >    0x0000000000454c9f (+1):     mov    %rsp,%rbp
> >    0x0000000000454ca2 (+4):     sub    $0x10,%rsp
> >    0x0000000000454ca6 (+8):     mov    %rdi,-0x8(%rbp)
> > => 0x0000000000454caa (+12):    mov    -0x8(%rbp),%rax
> >    0x0000000000454cae (+16):    mov    0x10(%rax),%eax
> 
> I don't like this: disassembly should look like it looks elsewhere.
> However, we could offer an optional removal of the leftmost column,
> the address, and the next one, which shows the symbol and offset from
> it.  Stripping leading zeros, as you point out, will help even more,
> especially on 64-bit platforms.

I use both of those pieces of information... it's the function name
that I don't use.

Joel raised a good point about the function name in x/i though.

-- 
Daniel Jacobowitz
CodeSourcery


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]