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

[Simon Marchi <simon.marchi@polymtl.ca>]

On 2017-01-31 19:30, Pedro Alves wrote:
> ui_file_rewind is a ui_file method that only really works with mem
> buffer files, and is a nop on other ui_file types.  It'd be desirable
> to eliminate it from the base ui_file interface, and move it to the
> "mem_fileopen" subclass of ui_file instead.  A following patch does
> just that.
> 
> Unfortunately, there are a couple references to ui_file_rewind inside
> gdb_disassembler::pretty_print_insn that were made harder to eliminate
> with the recent addition of the gdb_disassembler wrapper.
> 
> Before the gdb_disassembler wrapper was added, in commit
> e47ad6c0bd7aa3 ("Refactor disassembly code"), gdb_pretty_print_insn
> used to be passed a ui_file pointer as argument, and it was simple to
> adjust that pointer be a "mem_fileopen" ui_file pointer instead, since
> there's only one gdb_pretty_print_insn caller.
> 
> That commit made gdb_pretty_print_insn be a method of
> gdb_disassembler, and removed the method's ui_file parameter at the
> same time, replaced by referencing the gdb_disassembler's stream
> instead.  The trouble is that a gdb_disassembler can be instantiated
> with a pointer any kind of ui_file.  Casting the gdb_disassembler's
> stream to a mem_fileopen ui_file inside
> gdb_disassembler::pretty_print_insn in order to call the reset method
> would be gross hack.
> 
> The fix here is to:
> 
>  - make gdb_disassembler::pretty_print_insn a be free function again
>    instead of a method of gdb_disassembler.  I.e., bring back
>    gdb_pretty_print_insn.
> 
>  - but, don't add back the ui_file * parameter.  We'd always be
>    passing in a fresh mem_fileopen anyway, so move the mem_fileopen
>    allocation inside.  That is a better interface, given that the
>    ui_file is only ever used as temporary scratch buffer as an
>    implementation detail of gdb_pretty_print_insn.  The function's
>    real "where to send output" parameter is the ui_out pointer.
> 
>  - don't add back a disassemble_info pointer either.  That used to be
>    necessary for this bit:
> 
> 	  err = m_di.read_memory_func (pc, &data, 1, &m_di);
> 	  if (err != 0)
> 	    m_di.memory_error_func (err, pc, &m_di);
> 
>    ... but AFAIK, it's not really necessary.  We can replace those
>    three lines with a call to read_code.  This seems to fix a
>    regression even, because before commit d8b49cf0c891d0 ("Don't throw
>    exception in dis_asm_memory_error"), that memory_error_func call
>    would throw an error/exception, but now it only records the error
>    in the gdb_disassembler's m_err_memaddr field.  (read_code throws
>    on error.)
> 
> With all these, gdb_pretty_print_insn is completely layered on top of
> gdb_disassembler only using the latter's public API.

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.

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.