This is the mail archive of the mailing list for the GDB project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [RFC] Use address_from_register in dwarf2-frame.c:read_addr_from_reg

Andrew Pinski wrote:

> I think this patch broke MIPS64 n32 big-endian support.  We assert here:
> gdb_assert (!gdbarch_convert_register_p (gdbarch, regnum, type));
> The convert_register_p code for MIPS does:
>   return (register_size (gdbarch, regnum) == 8
>           && regnum % num_regs > 0 && regnum % num_regs < 32
>           && TYPE_LENGTH (type) < 8);
> Since the register size is 8 byte wide (MIPS64) and the type length is
> 4 (pointer), we return true.  In MIPS64, the registers are stored
> 64bits but pointers are 32bits.
> Here is the code that is used by mips_register_to_value:
>       int len = TYPE_LENGTH (type);
>       CORE_ADDR offset;
>       offset = gdbarch_byte_order (gdbarch) == BFD_ENDIAN_BIG ? 8 - len : 0;
>       if (!get_frame_register_bytes (frame, regnum, offset, len, to,
>     optimizedp, unavailablep))
> return 0;
>       *optimizedp = *unavailablep = 0;
>       return 1;

Huh, I wasn't aware of that conversion.  Note that for the register_to_value
case, I don't actually see any difference to the default behavior; it's the
value_to_register routine that's really special (because of the sign-extension
in performs).

> Is there a way to fix this in a target neutral way?  (I might need a
> way like this for AARCH64 ILP32 also).

I guess it isn't too hard to support gdbarch_convert_register_p in that
routine as well; I just didn't have any target to test on.

Can you try whether something along the following lines works for you?



	* findvar.c (address_from_register): Handle targets requiring
	a special conversion routine even for plain pointer types.

diff --git a/gdb/findvar.c b/gdb/findvar.c
index 41887de..ba3dd4d 100644
--- a/gdb/findvar.c
+++ b/gdb/findvar.c
@@ -764,11 +764,28 @@ address_from_register (int regnum, struct frame_info *frame)
      would therefore abort in get_frame_id.  However, since we only need
      a temporary value that is never used as lvalue, we actually do not
      really need to set its VALUE_FRAME_ID.  Therefore, we re-implement
-     the core of value_from_register, but use the null_frame_id.
+     the core of value_from_register, but use the null_frame_id.  */
-     This works only if we do not require a special conversion routine,
-     which is true for plain pointer types for all current targets.  */
-  gdb_assert (!gdbarch_convert_register_p (gdbarch, regnum, type));
+  /* Some targets require a special conversion routine even for plain
+     pointer types.  Avoid constructing a value object in those cases.  */
+  if (gdbarch_convert_register_p (gdbarch, regnum, type))
+    {
+      gdb_byte *buf = alloca (TYPE_LENGTH (type));
+      int optim, unavail, ok;
+      ok = gdbarch_register_to_value (gdbarch, frame, regnum, type,
+				      buf, &optim, &unavail);
+      if (!ok)
+	{
+	  /* This function is used while computing a location expression.
+	     Complain about the value being optimized out, rather than
+	     letting value_as_address complain about some random register
+	     the expression depends on not being saved.  */
+	  error_value_optimized_out ();
+	}
+      return unpack_long (type, buf);
+    }
   value = gdbarch_value_from_register (gdbarch, type, regnum, null_frame_id);
   read_frame_register_value (value, frame);

  Dr. Ulrich Weigand
  GNU/Linux compilers and toolchain

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]