[Lldb-commits] [PATCH] Register Number in DWARFCallFrameInfo.cpp

[Tong Shen]

Source code:
int f() { return 5; }
int main() {
  f();
  return 0;
}

Can you execute the following commands on lldb:

log enable lldb unwind
b f
r
bt

and see what it outputs?


On Tue, Aug 12, 2014 at 3:51 PM, Jason Molenda <jmolenda at apple.com> wrote:

> Hm, on my Mac I'm getting the correct registers from the eh_frame.
>
> * thread #1: tid = 0xce36, 0x00001f30 a.out`main, queue =
> 'com.apple.main-thread', stop reason = breakpoint 1.1
>     #0: 0x00001f30 a.out`main
> a.out`main:
> -> 0x1f30:  pushl  %ebp
>    0x1f31:  movl   %esp, %ebp
>    0x1f33:  subl   $0x8, %esp
>    0x1f36:  calll  0x1f3b                    ; main + 11
> (lldb) image show-unwind -n main
> <lldb.driver.main-thread> (i386) /private/tmp/a.out: Reading EH frame info
> First non-prologue instruction is at address 0x1f36 or offset 6 into the
> function.
>
> Asynchronous (not restricted to call-sites) UnwindPlan for a.out`main
> (start addr 0x1f30):
> This UnwindPlan originally sourced from assembly insn profiling
> Address range of this UnwindPlan: [a.out.__TEXT.__text + 0-0x00000029)
> row[0]: 0x00000000: CFA=esp +4 => esp=esp+4 eip=[esp]
> row[1]: 0x00000001: CFA=esp +8 => ebp=[esp] esp=esp+8 eip=[esp+4]
> row[2]: 0x00000003: CFA=ebp +8 => ebp=[ebp] esp=ebp+8 eip=[ebp+4]
> row[3]: 0x00000028: CFA=esp +4 => esp=esp+4 eip=[esp]
>
> Synchronous (restricted to call-sites) UnwindPlan for a.out`main (start
> addr 0x1f30):
> This UnwindPlan originally sourced from eh_frame CFI
> Address range of this UnwindPlan: [a.out.__TEXT.__text + 0-0x00000029)
> row[0]: 0x00000000: CFA=esp +4 => eip=[esp]
> row[1]: 0x00000001: CFA=esp +8 => ebp=[esp] eip=[esp+4]
> row[2]: 0x00000003: CFA=ebp +8 => ebp=[ebp] eip=[ebp+4]
>
>
> The UnwindAssembly_x86 generated UnwindPlan and the clang-generated
> eh_frame cfi match up...
>
>
>
> > On Aug 12, 2014, at 3:38 PM, Tong Shen <endlessroad at google.com> wrote:
> >
> > In i386 .eh_frame, esp=4, ebp=5. I confirmed it by cfi directives gcc
> generates and objdump -W.
> >
> > So it's either we chose the wrong eRegisterKindXXX, or we put {dwarf,
> gcc}_{esp, ebp}_i386 in wrong order in
> source/Plugins/Process/Utility/RegisterContext_x86.h :-)
> >
> >
> > On Tue, Aug 12, 2014 at 3:23 PM, Jason Molenda <jmolenda at apple.com>
> wrote:
> > Hi Tong, sorry for not replying earlier.
> >
> > I don't understand why the current code is wrong.
> >
> >         eRegisterKindGCC = 0,    // the register numbers seen in eh_frame
> >         eRegisterKindDWARF,      // the register numbers seen DWARF
> >
> > and
> >
> >         SectionSP sect = sl->FindSectionByType (eSectionTypeEHFrame,
> true);
> >         if (sect.get())
> >         {
> >             m_eh_frame = new DWARFCallFrameInfo(m_object_file, sect,
> eRegisterKindGCC, true);
> >         }
> >
> > We create a DWARFCallFrameInfo object to read the eh_frame information
> and we tell it that the register kind to use for the reg numbers is
> eRegisterKindGCC - aka the eh_frame register numbering domain.
> >
> > This all looks correct to me, and designed specifically to handle the
> case of that ancient gcc register numbering bug in the eh_frame section for
> i386.
> >
> > We never read the actual DWARF debug_frame (today) because I've never
> seen a case where it had more information than the eh_frame unwind info.
> >
> >
> >
> > > On Aug 12, 2014, at 1:36 PM, Tong Shen <endlessroad at google.com> wrote:
> > >
> > > Aha, got it.
> > > Patch confirmed and attached.
> > >
> > >
> > > On Tue, Aug 12, 2014 at 12:51 PM, Ed Maste <emaste at freebsd.org> wrote:
> > > On 12 August 2014 15:43, Todd Fiala <tfiala at google.com> wrote:
> > > >
> > > > Hey Tong,
> > > >
> > > > Ah interesting.  The register infos in lldb have a vector of
> different register numbers (I think we call them "register kinds") based on
> the type of encoding you're using.  Let's have a look at that to see if we
> can adjust the code to query for the right register kind.  We'll look at
> this today!
> > >
> > > There was a thread on this topic on this list a while back, and I
> > > think the two register sets in question here are DWARF and GCC, which
> > > are identical in all cases except that i386 swaps esp and ebp between
> > > the two.  This is all the legacy of an ancient GCC bug.  So I think
> > > you're right Todd, just picking the other set should solve this.
> > >
> > >
> > >
> > > --
> > > Best Regards, Tong Shen
> > >
> <fix_register_kind.patch>_______________________________________________
> > > lldb-commits mailing list
> > > lldb-commits at cs.uiuc.edu
> > > http://lists.cs.uiuc.edu/mailman/listinfo/lldb-commits
> >
> >
> >
> >
> > --
> > Best Regards, Tong Shen
>
>


-- 
Best Regards, Tong Shen
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/lldb-commits/attachments/20140812/95bd9d4b/attachment.html>