答复: [PATCH] [RISC-V]: Handling optimized prologue

[<lifang_xia@linux.alibaba.com>]

Hi Andrew,
Thanks for reviewing.
See below.


> -----邮件原件-----
> 发件人: Andrew Burgess <aburgess@redhat.com>
> 发送时间: 2022年3月1日 18:34
> 收件人: Lifang Xia <lifang_xia@c-sky.com>; gdb-patches@sourceware.org
> 抄送: Lifang Xia <lifang_xia@linux.alibaba.com>
> 主题: Re: [PATCH] [RISC-V]: Handling optimized prologue
> 
> Lifang Xia <lifang_xia@c-sky.com> writes:
> 
> Hello Lifang,
> 
> Thanks for working on this.
> 
> > From: Lifang Xia <lifang_xia@linux.alibaba.com>
> >
> > The optimizer might shove anything into the prologue, if we build up
> > cache (cache != NULL) from scanning prologue, we just skip what we
> > don't recognize and scan further to make cache as complete as
> > possible.
> 
> My concern here would be, what if the thing you are skipping over modifies
> the state in such a way, that when we think we are building the cache,
we're
> really doing the wrong thing in some way...
> 
> 
> >                                      However, if we skip prologue,
> > we'll stop immediately on unrecognized instruction.
> >
> > The case is:
> >
> > ----------------------------
> > Disassembly of section .text:
> >
> > 0000000000010078 <foo>:
> >    10078:       00100513                li      a0,1
> >    1007c:       008000ef                jal     ra,10084 <bar>
> >    10080:       00008067                ret
> >
> > 0000000000010084 <bar>:
> >    10084:       00051263                bnez    a0,10088 <bar+0x4>
> >    10088:       ff010113                addi    sp,sp,-16
> >    1008c:       00813023                sd      s0,0(sp)
> >    10090:       00050413                mv      s0,a0
> >    10094:       00113423                sd      ra,8(sp)
> >    10098:       014000ef                jal     ra,100ac <aaa>
> >    1009c:       00813083                ld      ra,8(sp)
> >    100a0:       00013403                ld      s0,0(sp)
> >    100a4:       01010113                addi    sp,sp,16
> >    100a8:       00008067                ret
> >
> > 00000000000100ac <aaa>:
> >    100ac:       00000013                nop
> >>> 100b0:       00008067                ret
> > ----------------------
> > (gdb) bt
> > #0  0x00000000000100b0 in aaa ()
> > #1  0x000000000001009c in bar ()
> > Backtrace stopped: frame did not save the PC
> 
> I assume it's the bnez that's causing the problem for the prologue scan
here?
> 
> Not that I really understand what that instruction is about, it appears to
be
> branching to the next instruction.  I wonder if this objdump output is for
the
> object file, rather than the executable?  I'm guessing the bnez has a
reloc
> which points at, maybe, the ret at 0x100a8?
> 
> I wonder if a better solution in this case would be to allow the prologue
scan
> to skip over forward branches within the prologue, maybe with the
limitation
> that the destination has to be within the function bounds?
> 

Yes, You are right. bnez cause the prologue scan failed.
It's one of the cases I have met. As mentioned before, The optimizer of
compiler might shove anything to the prologue.
It could be multiple instructions, shift instructions, condition branches
...
We can't handle all of instructions from  the ISA specs or vendor extensions
here.

Actually, if cache is not nullptr, that means we need to build the frame
information from prologue. Before the  loop, we get the base address and the
prologue_end,
 traversing all the prologue is not a bad choice to build the frame(if cache
is not nullptr). 

> >
> > gdb/
> > 	* riscv-tdep.c (riscv_scan_prologue): Keep scaning if cache is not
> NULL.
> > ---
> >  gdb/riscv-tdep.c | 10 +++++++++-
> >  1 file changed, 9 insertions(+), 1 deletion(-)
> >
> > diff --git a/gdb/riscv-tdep.c b/gdb/riscv-tdep.c index
> > 886996c..e46d441 100644
> > --- a/gdb/riscv-tdep.c
> > +++ b/gdb/riscv-tdep.c
> > @@ -1987,7 +1987,15 @@ riscv_scan_prologue (struct gdbarch *gdbarch,
> >        else
> >  	{
> >  	  end_prologue_addr = cur_pc;
> > -	  break;
> > +
> > +	  /* The optimizer might shove anything into the prologue, if
> > +	     we build up cache (cache != NULL) from scanning prologue,
> > +	     we just skip what we don't recognize and scan further to
> > +	     make cache as complete as possible.  However, if we skip
> > +	     prologue, we'll stop immediately on unrecognized
> > +	     instruction.  */
> > +	  if (cache == NULL)
> 
> Use nullptr not NULL please.
> 
> Thanks,
> Andrew
> 
> 
> 
> > +	    break;
> >  	}
> >      }
> >
> > --
> > 2.7.4