[pushed] Change inline frame breakpoint skipping logic (fix gdb.gdb/selftest.exp)

  On 06/25/2018 10:04 PM, Joel Brobecker wrote:
> Hello,
> 
>> gdb/ChangeLog:
>> 2018-06-19  Pedro Alves  <palves@redhat.com>
>>
>> 	* inline-frame.c (stopped_by_user_bp_inline_frame): Replace PC
>> 	parameter with a block parameter.  Compare location's block symbol
>> 	with the frame's block instead of addresses.
>> 	(skip_inline_frames): Pass the current block instead of the
>> 	frame's address.  Break out as soon as we determine the frame
>> 	should not be skipped.
>>
>> gdb/testsuite/ChangeLog:
>> 2018-06-19  Pedro Alves  <palves@redhat.com>
>>
>> 	* gdb.opt/inline-break.c (func_inline_callee, func_inline_caller)
>> 	(func_extern_caller): New.
>> 	(main): Call func_extern_caller.
>> 	* gdb.opt/inline-break.exp: Add tests for inline frame skipping
>> 	logic change.
> 
> it looks like this patch is causing a crash with the following
> example program:
> 
>     $ cat -n r.h
>          1	/* r.h */
>          2	int counter = 42;
>          3
>          4	inline void
>          5	callee () {
>          6	  counter = 0;  /* break here */
>          7	}
>     $ cat -n r.c
>          1	/* r.c */
>          2	#include "r.h"
>          3
>          4	int
>          5	main ()
>          6	{
>          7	  callee ();
>          8	}
> 
> I compiled it using the following commands:
> 
>     $ gcc -c -g -O2 r.c
>     $ gcc -o r r.o
> 
> Then, trying to put a breakpoint on r.h:6 (inside "callee") causes
> a SEGV for me:
> 
>     $ gdb -q r
>     Reading symbols from r...done.
>     (gdb) b r.h:6
>     Breakpoint 1 at 0x4003c0: file r.h, line 6.
>     (gdb) run
>     Starting program: /[...]/r
>     [1]    75618 segmentation fault  /[...]/gdb -q r
> 
> Prior to this commit, the behavior is the following for the "run"
> command:
> 
>     (gdb) run
>     Starting program: /[...]/r
> 
>     Breakpoint 1, callee () at r.h:6
>     6	  counter = 0;  /* break here */
> 
> The problem occurs because we apparently assume that a bp_location's
> symbols is not NULL:
> 
>     Program received signal SIGSEGV, Segmentation fault.
>     0x00000000006f42bb in stopped_by_user_bp_inline_frame (
>         stop_chain=<optimized out>, frame_block=<optimized out>)
>         at /homes/brobecke/act/gdb/gdb-head/gdb/inline-frame.c:305
>     305		      && frame_block == SYMBOL_BLOCK_VALUE (loc->symbol))
>     (gdb) p loc->symbol
>     $1 = (const symbol *) 0x0

Whoops.  I remember actually thinking about loc->symbol potentially
being null, but somehow convinced myself that SYMBOL_BLOCK_VALUE
would return null in that case...  :-P

> 
> I don't know yet whether that's a valid assumption or something
> occurred earlier in the process. Any thoughts on this before I start
> looking deeper?
> 
> I'm using a version of GCC 7.3.1 on x86_64-linux if anyone wants to
> reproduce.

If we just add a "loc->symbol != NULL" check, then we end up
presenting the stop at the caller of the inline function, where
the inline function was inlined, which is not what we want, since that's
not where the user set the breakpoint.

Recording the symbol in the location (it is copied from the sal
that linespec.c creates into the location by
add_location_to_breakpoint), like in the patch below, makes gdb present
the stop at  the "break here" line successfully.

I tried to come up with a more complicated testcase that included
more nested blocks inside the inlined function, to see if we would
need to record the inner inlined block in the sal instead of the
function's symbol (does that actually make sense for inlined functions?),
but all I came up with worked with this patch as is.  So maybe we
can defer thinking about that until we find a testcase?

From ac746927c3a8078098729dc3256010b2c1e617f8 Mon Sep 17 00:00:00 2001
From: Pedro Alves <palves@redhat.com>
Date: Tue, 26 Jun 2018 19:14:41 +0100
Subject: [PATCH] inline

---
 gdb/inline-frame.c | 1 +
 gdb/linespec.c     | 1 +
 2 files changed, 2 insertions(+)
  

Hi Pedro,

> >     Program received signal SIGSEGV, Segmentation fault.
> >     0x00000000006f42bb in stopped_by_user_bp_inline_frame (
> >         stop_chain=<optimized out>, frame_block=<optimized out>)
> >         at /homes/brobecke/act/gdb/gdb-head/gdb/inline-frame.c:305
> >     305		      && frame_block == SYMBOL_BLOCK_VALUE (loc->symbol))
> >     (gdb) p loc->symbol
> >     $1 = (const symbol *) 0x0
> 
> Whoops.  I remember actually thinking about loc->symbol potentially
> being null, but somehow convinced myself that SYMBOL_BLOCK_VALUE
> would return null in that case...  :-P
> 
> > 
> > I don't know yet whether that's a valid assumption or something
> > occurred earlier in the process. Any thoughts on this before I start
> > looking deeper?
> > 
> > I'm using a version of GCC 7.3.1 on x86_64-linux if anyone wants to
> > reproduce.
> 
> If we just add a "loc->symbol != NULL" check, then we end up
> presenting the stop at the caller of the inline function, where
> the inline function was inlined, which is not what we want, since that's
> not where the user set the breakpoint.
> 
> Recording the symbol in the location (it is copied from the sal
> that linespec.c creates into the location by
> add_location_to_breakpoint), like in the patch below, makes gdb present
> the stop at  the "break here" line successfully.
> 
> I tried to come up with a more complicated testcase that included
> more nested blocks inside the inlined function, to see if we would
> need to record the inner inlined block in the sal instead of the
> function's symbol (does that actually make sense for inlined functions?),
> but all I came up with worked with this patch as is.  So maybe we
> can defer thinking about that until we find a testcase?

Just a quick message that the patch makes sense to me, and that
I was just able to run it through AdaCore's testsuite with succes.
Or, I should qualify that - there is one tiny change that I haven't
had time to analyze, but from the surface, it is exactly what you
explained about why you need the second hunk.

I haven't had a chance to run it through the official testsuite,
however, as I have to go ... I am so laaaaate!

I can do that tomorrow, or if you prefer to just finish the patch
up and push it, it'd be perfect. I think the patch is good.

Thanks again!

> 
> >From ac746927c3a8078098729dc3256010b2c1e617f8 Mon Sep 17 00:00:00 2001
> From: Pedro Alves <palves@redhat.com>
> Date: Tue, 26 Jun 2018 19:14:41 +0100
> Subject: [PATCH] inline
> 
> ---
>  gdb/inline-frame.c | 1 +
>  gdb/linespec.c     | 1 +
>  2 files changed, 2 insertions(+)
> 
> diff --git a/gdb/inline-frame.c b/gdb/inline-frame.c
> index 896b0004e4a..3d07f8d0970 100644
> --- a/gdb/inline-frame.c
> +++ b/gdb/inline-frame.c
> @@ -302,6 +302,7 @@ stopped_by_user_bp_inline_frame (const block *frame_block, bpstat stop_chain)
>  
>  	  if ((t == bp_loc_software_breakpoint
>  	       || t == bp_loc_hardware_breakpoint)
> +	      && loc->symbol != nullptr
>  	      && frame_block == SYMBOL_BLOCK_VALUE (loc->symbol))
>  	    return true;
>  	}
> diff --git a/gdb/linespec.c b/gdb/linespec.c
> index ae0200b8133..93e66c389f7 100644
> --- a/gdb/linespec.c
> +++ b/gdb/linespec.c
> @@ -2196,6 +2196,7 @@ create_sals_line_offset (struct linespec_state *self,
>  
>  	    if (self->funfirstline)
>  	      skip_prologue_sal (&intermediate_results[i]);
> +	    intermediate_results[i].symbol = sym;
>  	    add_sal_to_sals (self, &values, &intermediate_results[i],
>  			     sym ? SYMBOL_NATURAL_NAME (sym) : NULL, 0);
>  	  }
> -- 
> 2.14.4
> 
> Thanks,
> Pedro Alves
  

Hi Joel,

On 06/26/2018 11:02 PM, Joel Brobecker wrote:

> Just a quick message that the patch makes sense to me, and that
> I was just able to run it through AdaCore's testsuite with succes.
> Or, I should qualify that - there is one tiny change that I haven't
> had time to analyze, but from the surface, it is exactly what you
> explained about why you need the second hunk.
> 
> I haven't had a chance to run it through the official testsuite,
> however, as I have to go ... I am so laaaaate!
> 
> I can do that tomorrow, or if you prefer to just finish the patch
> up and push it, it'd be perfect. I think the patch is good.
> 
> Thanks again!

FYI, I'm starting to look at this now.

Thanks,
Pedro Alves
  

On 06/27/2018 05:28 PM, Pedro Alves wrote:
> Hi Joel,
> 
> On 06/26/2018 11:02 PM, Joel Brobecker wrote:
> 
>> Just a quick message that the patch makes sense to me, and that
>> I was just able to run it through AdaCore's testsuite with succes.
>> Or, I should qualify that - there is one tiny change that I haven't
>> had time to analyze, but from the surface, it is exactly what you
>> explained about why you need the second hunk.
>>
>> I haven't had a chance to run it through the official testsuite,
>> however, as I have to go ... I am so laaaaate!
>>
>> I can do that tomorrow, or if you prefer to just finish the patch
>> up and push it, it'd be perfect. I think the patch is good.
>>
>> Thanks again!
> 
> FYI, I'm starting to look at this now.

So while poking some more at this, noticed that setting a breakpoint
by address crashes in the same way, like "b *ADDRESS".  So I thought
that maybe it would be better to make stopped_by_user_bp_inline_frame
return true if the location has no symbol instead of returning
false like in the version I sent before.  That preserves the previous
behavior of showing the stop at the inline function if we miss
setting the sal's symbol somewhere.

However, playing with that made me notice something else
unrelated to my "Change inline frame breakpoint skipping"
patch:

  (gdb) b *0x40062f
  Breakpoint 2 at 0x40062f: file inline-break.c, line 32.
  (gdb) info breakpoints
  Num     Type           Disp Enb Address            What
  2       breakpoint     keep y   0x000000000040062f in main at inline-break.c:32
   (gdb) r
   ....
  Breakpoint 2, func1 (x=1) at inline-break.c:32
  32        return x * 23; /* break here */

Notice that above "info break" says "in main":

   in main at inline-break.c:32
      ^^^^

Since we say "inline-break.c:32" everywhere, and present the
stop at the inline function, I think that "info break" should say instead:

   in func1 at inline-break.c:32
      ^^^^^

Fixing that ends up going back to setting the symbol in the sal
again, but I decided to do that in a separate patch, and still
make "loc->symbol == nullptr" in stopped_by_user_bp_inline_frame return
true, unlike the previous version of the patch.

I'll be sending two patches in response to this email.

Thanks,
Pedro Alves