Re: [PATCH v4 1/2] Add back gdb_pretty_print_insn

[Pedro Alves <palves@redhat.com>]

On 02/01/2017 06:09 PM, Simon Marchi wrote:

> I don't think I understand the situation fully, but what you suggest
> looks good to me.  I was confused by the fact that the gdb_disassembler
> constructor accepts a stream, but the pretty_print_insn method takes a
> uiout.  Which one is used for printing then?  I think that your patch
> clears that up.
> 

Let me try to clear up a bit.  v3, which predated the gdb_disassembler
changes, did this:

 gdb_pretty_print_insn (struct gdbarch *gdbarch, struct ui_out *uiout,
                       struct disassemble_info * di,
                       const struct disasm_insn *insn, int flags,
-                      struct ui_file *stb)
+                      string_file &stb)
 {
   /* parts of the symbolic representation of the address */
   int unmapped;
@@ -240,7 +240,7 @@ gdb_pretty_print_insn (struct gdbarch *gdbarch, struct ui_out *uiout,
   if (name != NULL)
     xfree (name);
 
-  ui_file_rewind (stb);
+  stb.clear ();

   [...]

   uiout->field_stream ("inst", stb);
-  ui_file_rewind (stb);
+  stb.clear ();
   do_cleanups (ui_out_chain);
   uiout->text ("\n");


while in current master we have:

int
gdb_disassembler::pretty_print_insn (struct ui_out *uiout,
                                     const struct disasm_insn *insn,
                                     int flags)
{
  [...]
  struct ui_file *stb = stream ();

  [...]
  ui_file_rewind (stb);
  [...]


So we can no longer do the same thing v3 did, because
"stream ();" is generic.  Looking at the callers of
pretty_print_insn, we know that the ui_file returned by
"stream ();" here is "string_file *", but doing:

-  struct ui_file *stb = stream ();
+  string_file *stb = (string_file *) stream ();

would be a gross hack, for baking in knowledge of who
are the current callers.

> The only possible issue I can see is that in your version, one
> gdb_disassembler and one string_file object are constructed for each
> disassembled instruction, rather than re-using them for as long as we
> need to disassemble.  I don't know how much impact it has on the
> performance (probably negligible), but something to keep in mind.

Yeah.  It's simple to add a string_file parameter to gdb_pretty_print_insn,
in order to pass in a buffer that is reused, like it used to be,
if found necessary.

gdb_disassembler is on the stack so practically doesn't
count, in overhead terms.  I think for this series it may end
up balanced by allocating fewer cleanups, and also I suspect
most disassembled instructions fit std::string's
"small string optimization", meaning no heap allocation.

Thanks,
Pedro Alves