gdb: make frame_unwind_got_optimized return a not_lval value

TLDR: frame_unwind_got_optimized uses wrong frame id value, trying to
fix it makes GDB sad, return not_lval value and don't use frame id value
instead.

Longer version:

The `prev_register` method of the `frame_unwind` interface corresponds
to asking the question: "where did this frame - passed as a parameter -
save the value this register had in its caller frame?".  When "this
frame" did not save that register value (DW_CFA_undefined in DWARF), the
implementation can use the `frame_unwind_got_optimized` function to
create a struct value that represents the optimized out / not saved
register.

`frame_unwind_got_optimized` marks the value as fully optimized out,
sets the lval field to lval_register and assigns the required data for
lval_register: the next frame id and the register number.  The problem
is that it uses the frame id from the wrong frame (see below for in
depth explanation).  In practice, this is not problematic because the
frame id is never used: the value is already not lazy (and is marked as
optimized out), so the value is never fetched from the target.

When trying to change it to put the right next frame id in the value, we
bump into problems: computing the frame id for some frame requires
unwinding some register, if that register is not saved / optimized out,
we try to get the frame id that we are currently computing.

This patch addresses the problem by changing
`frame_unwind_got_optimized` to return a not_lval value instead.  Doing
so, we don't need to put a frame id, so we don't hit that problem.  It
may seem like an unnecessary change today, because it looks like we're
fixing something that is not broken (from the user point of view).
However, the bug becomes user visible with the following patches, where
inline frames are involved.  I put this change in its own patch to keep
it logically separate.

Let's now illustrate how we are putting the wrong frame id in the value
returned by `frame_unwind_got_optimized`.  Let's assume this stack:

    frame #0
    frame #1
    frame #2
    frame #3

Let's suppose that we are calling `frame_unwind_register_value` with
frame #2 as the "next_frame" parameter and some register number X as the
regnum parameter.  That is like asking the question "where did frame #2
save frame #3's value for register X".

`frame_unwind_register_value` calls the frame unwinder's `prev_register`
method, which in our case is `dwarf2_frame_prev_register`.  Note that in
`dwarf2_frame_prev_register`, the parameter is now called `this_frame`,
but its value is still frame #2, and we are still looking for where
frame #2 saved frame #3's value of register X.

Let's now suppose that frame #2's CFI explicitly indicates that the
register X is was not saved (DW_CFA_undefined).  We go into
`frame_unwind_got_optimized`.

In `frame_unwind_got_optimized`, the intent is to create a value that
represents register X in frame #3.  An lval_register value requires that
we specify the id of the _next_ frame, that is the frame from which we
would need to unwind in order to get the value.  Therefore, we would
want to put the id of frame #2 in there.

However, `frame_unwind_got_optimized` does:

    VALUE_NEXT_FRAME_ID (val)
      = get_frame_id (get_next_frame_sentinel_okay (frame));

where `frame` is frame #2.  The get_next_frame_sentinel_okay call
returns frame #1, so we end up putting frame #1's id in the value.

Let's now pretend that we try to "fix" it by placing the right frame id,
in other words doing this change:

    --- a/gdb/frame-unwind.c
    +++ b/gdb/frame-unwind.c
    @@ -260,8 +260,7 @@ frame_unwind_got_optimized (struct frame_info *frame, int regnum)
       mark_value_bytes_optimized_out (val, 0, TYPE_LENGTH (type));
       VALUE_LVAL (val) = lval_register;
       VALUE_REGNUM (val) = regnum;
    -  VALUE_NEXT_FRAME_ID (val)
    -    = get_frame_id (get_next_frame_sentinel_okay (frame));
    +  VALUE_NEXT_FRAME_ID (val) = get_frame_id (frame);
       return val;
     }

This makes some tests fails, such as gdb.dwarf2/dw2-undefined-ret-addr.exp,
like so:

    ...
    #9  0x0000557a8ab15a5d in internal_error (file=0x557a8b31ef80 "/home/simark/src/binutils-gdb/gdb/frame.c", line=623, fmt=0x557a8b31efe0 "%s: Assertion `%s' failed.") at /home/simark/src/binutils-gdb/gdbsupport/errors.cc:55
    #10 0x0000557a87f816d6 in get_frame_id (fi=0x62100034bde0) at /home/simark/src/binutils-gdb/gdb/frame.c:623
    #11 0x0000557a87f7cac7 in frame_unwind_got_optimized (frame=0x62100034bde0, regnum=16) at /home/simark/src/binutils-gdb/gdb/frame-unwind.c:264
    #12 0x0000557a87a71a76 in dwarf2_frame_prev_register (this_frame=0x62100034bde0, this_cache=0x62100034bdf8, regnum=16) at /home/simark/src/binutils-gdb/gdb/dwarf2/frame.c:1267
    #13 0x0000557a87f86621 in frame_unwind_register_value (next_frame=0x62100034bde0, regnum=16) at /home/simark/src/binutils-gdb/gdb/frame.c:1288
    #14 0x0000557a87f855d5 in frame_register_unwind (next_frame=0x62100034bde0, regnum=16, optimizedp=0x7fff5f459070, unavailablep=0x7fff5f459080, lvalp=0x7fff5f4590a0, addrp=0x7fff5f4590b0, realnump=0x7fff5f459090, bufferp=0x7fff5f459150 "") at /home/simark/src/binutils-gdb/gdb/frame.c:1191
    #15 0x0000557a87f860ef in frame_unwind_register (next_frame=0x62100034bde0, regnum=16, buf=0x7fff5f459150 "") at /home/simark/src/binutils-gdb/gdb/frame.c:1247
    #16 0x0000557a881875f9 in i386_unwind_pc (gdbarch=0x621000190110, next_frame=0x62100034bde0) at /home/simark/src/binutils-gdb/gdb/i386-tdep.c:1971
    #17 0x0000557a87fe58a5 in gdbarch_unwind_pc (gdbarch=0x621000190110, next_frame=0x62100034bde0) at /home/simark/src/binutils-gdb/gdb/gdbarch.c:3062
    #18 0x0000557a87a6267b in dwarf2_tailcall_sniffer_first (this_frame=0x62100034bde0, tailcall_cachep=0x62100034bee0, entry_cfa_sp_offsetp=0x7fff5f4593f0) at /home/simark/src/binutils-gdb/gdb/dwarf2/frame-tailcall.c:387
    #19 0x0000557a87a70cdf in dwarf2_frame_cache (this_frame=0x62100034bde0, this_cache=0x62100034bdf8) at /home/simark/src/binutils-gdb/gdb/dwarf2/frame.c:1198
    #20 0x0000557a87a711c2 in dwarf2_frame_this_id (this_frame=0x62100034bde0, this_cache=0x62100034bdf8, this_id=0x62100034be40) at /home/simark/src/binutils-gdb/gdb/dwarf2/frame.c:1226
    #21 0x0000557a87f81167 in compute_frame_id (fi=0x62100034bde0) at /home/simark/src/binutils-gdb/gdb/frame.c:587
    #22 0x0000557a87f81803 in get_frame_id (fi=0x62100034bde0) at /home/simark/src/binutils-gdb/gdb/frame.c:635
    #23 0x0000557a87f7efef in scoped_restore_selected_frame::scoped_restore_selected_frame (this=0x7fff5f459920) at /home/simark/src/binutils-gdb/gdb/frame.c:320
    #24 0x0000557a891488ae in print_frame_args (fp_opts=..., func=0x621000183b90, frame=0x62100034bde0, num=-1, stream=0x6030000caa20) at /home/simark/src/binutils-gdb/gdb/stack.c:750
    #25 0x0000557a8914e87a in print_frame (fp_opts=..., frame=0x62100034bde0, print_level=0, print_what=SRC_AND_LOC, print_args=1, sal=...) at /home/simark/src/binutils-gdb/gdb/stack.c:1394
    #26 0x0000557a8914c2ae in print_frame_info (fp_opts=..., frame=0x62100034bde0, print_level=0, print_what=SRC_AND_LOC, print_args=1, set_current_sal=1) at /home/simark/src/binutils-gdb/gdb/stack.c:1119
    ...

We end up calling get_frame_id (in the hunk above, frame #10)  while we are
computing it (frame #21), and that's not good.

Now, the question is how do we fix this.  I suggest making the unwinder
return a not_lval value in this case.

The reason why we return an lval_register here is to make sure that this
is printed as "not saved" and not "optimized out" down the line.  See
these two commits:

1. 901461f8eb40 ("Print registers not saved in the frame as "<not saved>"
   instead of "<optimized out>".").
2. 6bd273ae450b ("Make "set debug frame 1" output print <not saved> instead of
   <optimized out>.")

The current design (introduced by the first commit) is to check the
value's lval to choose which one to print (see val_print_optimized_out).

Making the unwinder return not_lval instead of lval_register doesn't
break "not saved" when doing "print $rax" or "info registers", because
value_fetch_lazy_register only consumes the contents and optimized-out
property from the value the unwinder returned.  The value being
un-lazified stays an lval_register.

I believe that this is a correct technical solution (and not just
papering over the problem), because what we expect of unwinders is to
tell us where a given register's value is saved.  If the value is saved
in memory, -> lval_memory.  If the value is saved in some other register
of the next frame, -> lval_register.  If the value is not saved, it
doesn't really make sense to return an lval_register value.  not_lval
would be more appropriate.  If the code then wants to represent an
optimized out register value (like value_fetch_lazy_register does), then
it's a separate concern which shouldn't involve the unwinder.

This change breaks the output of "set debug frame 1" though (introduced
by the second commit), since that logging statement consumes the return
value of the unwinder directly.  To keep the correct behavior, just make
`frame_unwind_register_value` call `val_print_not_saved` directly,
instead of `val_print_optimized_out`.  This is fine because we know in
this context that we are always talking about a register value, and that
we want to show "not saved" for those.

I augmented the gdb.dwarf2/dw2-reg-undefined.exp test case to test some
cases I stumbled on while working on this, which I think are not tested
anywhere:

- the "set debug frame 1" debug output mentioned above.  It's just debug
  output, but if we want to make sure it doesn't change, it should be
  tested
- printing not-saved register values from the history (should print not
  saved)
- copying a not-saved register value in a convenience variable.  In this
  case, we expect that printing the convenience variable shows
  "optimized out", because we copied the value, not the property of
  where the value came from.

gdb/ChangeLog:

	* frame-unwind.c (frame_unwind_got_optimized): Don't set
	regnum/frame in value.  Call allocate_value_lazy.
	* frame.c (frame_unwind_register_value): Use
	val_print_not_saved.

gdb/testsuite/ChangeLog:

	* gdb.dwarf2/dw2-reg-undefined.exp: Test "set debug frame 1"
	output, printing a "not saved" value from history and printing a
	convenience variable created from a "not saved" value.

Change-Id: If451739a3ef7a5b453b1f50707e21ce16d74807e

1 files changed, 3 insertions, 0 deletions

diff --git a/gdb/frame-unwind.h b/gdb/frame-unwind.h
index c80ddc8..25a601c 100644
--- a/gdb/frame-unwind.h
+++ b/gdb/frame-unwind.h
@@ -133,6 +133,9 @@ typedef void (frame_this_id_ftype) (struct frame_info *this_frame,
    may be a lazy reference to memory, a lazy reference to the value of
    a register in THIS frame, or a non-lvalue.
 
+   If the previous frame's register was not saved by THIS_FRAME and is
+   therefore undefined, return a not_lval wholly optimized-out value.
+
    THIS_PROLOGUE_CACHE can be used to share any prolog analysis data
    with the other unwind methods.  Memory for that cache should be
    allocated using FRAME_OBSTACK_ZALLOC().  */