Fix gdb.dwarf2/amd64-entry-value-param.exp with -fPIE/-pie

  Hi,

When running gdb.dwarf2/amd64-entry-value-param.exp with target board
unix/-fPIE/-pie, we get:
...
FAIL: gdb.arch/amd64-entry-value-param.exp: call 1: p y
...

The problem is that read_call_site_scope attempts to put relocated addresses
in cu->call_site_htab, for both the pc field:
...
  baseaddr = ANOFFSET (objfile->section_offsets, SECT_OFF_TEXT (objfile));
  ...
  pc = attr_value_as_address (attr) + baseaddr;
  pc = gdbarch_adjust_dwarf2_addr (gdbarch, pc);
  ...
  call_site->pc = pc;
...
and the target field:
...
             lowpc = gdbarch_adjust_dwarf2_addr (gdbarch, lowpc + baseaddr);
             SET_FIELD_PHYSADDR (call_site->target, lowpc);
...
but fails to do so because it is called before objfile_relocate, making
baseaddr 0.

Fix this by eliminating baseaddr from read_call_site_scope, and handling the
relocation offset at the use sites in call_site_for_pc and
call_site_to_target_addr.

Tested on x86_64-linux, both with and without -fPIE/-pie.

OK for trunk?

Thanks,
- Tom

[gdb] Fix gdb.dwarf2/amd64-entry-value-param.exp with -fPIE/-pie

gdb/ChangeLog:

2019-08-09  Tom de Vries  <tdevries@suse.de>

	PR gdb/24892
	* block.c (call_site_for_pc): Substract relocation offset before
	finding pc in COMPUNIT_CALL_SITE_HTAB.
	* dwarf2loc.c (call_site_to_target_addr): Add relocation offset to
	FIELD_STATIC_PHYSADDR (call_site->target).
	* dwarf2read.c (read_call_site_scope): Eliminate baseaddr.

---
 gdb/block.c      | 17 ++++++++++++++++-
 gdb/dwarf2loc.c  | 12 +++++++++++-
 gdb/dwarf2read.c | 12 +++---------
 3 files changed, 30 insertions(+), 11 deletions(-)
  

On Fri, 09 Aug 2019 09:54:26 +0200, Tom de Vries wrote:
> Fix this by eliminating baseaddr from read_call_site_scope, and handling the
> relocation offset at the use sites in call_site_for_pc and
> call_site_to_target_addr.

It sure looks correct to me that way, thanks.  This is not a commit approval.


Thanks,
Jan Kratochvil
  

>>>>> "Tom" == Tom de Vries <tdevries@suse.de> writes:

Tom> Fix this by eliminating baseaddr from read_call_site_scope, and handling the
Tom> relocation offset at the use sites in call_site_for_pc and
Tom> call_site_to_target_addr.

I like this approach, since it represents some small progress on the
objfile splitting project.

Tom> +    {
Tom> +      struct obj_section *sec;
Tom> +      sec = find_pc_section (pc);
Tom> +      if (sec != NULL)
Tom> +	{
Tom> +	  struct objfile *objfile = sec->objfile;
Tom> +	  CORE_ADDR baseaddr
Tom> +	    = ANOFFSET (objfile->section_offsets, SECT_OFF_TEXT (objfile));

Why SECT_OFF_TEXT rather than the section "sec"?

Tom> +	  CORE_ADDR pc_unrelocated
Tom> +	    = gdbarch_adjust_dwarf2_addr (gdbarch, pc - baseaddr);

I am not sure gdbarch_adjust_dwarf2_addr can be used "bidirectionally"
like this.  It is probably fine in practice but I wonder about the
documented contract.

Tom> +	CORE_ADDR baseaddr
Tom> +	  = ANOFFSET (objfile->section_offsets, SECT_OFF_TEXT (objfile));

I guess this assumes the text section - but then can the call to
find_pc_section give anything else?  Maybe it's just something to
comment and move on.

Tom> -  pc = attr_value_as_address (attr) + baseaddr;
Tom> -  pc = gdbarch_adjust_dwarf2_addr (gdbarch, pc);
Tom> +  pc = attr_value_as_address (attr);

The approach taken elsewhere in dwarf2read.c is to bias, adjust, then
unbias:

      CORE_ADDR low
	= (gdbarch_adjust_dwarf2_addr (gdbarch, best_lowpc + baseaddr)
	   - baseaddr);

Maybe this would be better here as well, or at least consistent.

Tom