Issue 161206.1: Clarify placement of register pieces

Author: Andreas Arnez
Champion:
Date submitted: 2016-12-06
Date revised:
Date closed:
Type: Clarification
Status: Rejected
DWARF Version: 5
Section 2.6.1.2, pg 42
The current definition of the placement of a DW_OP_bit_piece within a
register reads:

      If the location is a register, the offset is from the
      least significant bit end of the register.

This is ambiguous and may not be appropriate in some cases:

1. The term "least significant bit end" is not defined, and there are
   cases where its definition is not obvious, particularly on
   big-endian platforms.  For instance, consider vector- or
   floating-point registers.  (An attempt for defining a register's
   least significant bit might look like this: "The least significant
   bit of the first/last byte in the register's natural memory
   representation, on little-/big-endian platforms, respectively.")
2. The definition does not allow for widening a register set beyond
   each of its registers' "least significant bit", such as it occurred
   with z/Architecture when the FP registers were expanded to vector
   registers.
3. Assuming that DW_OP_piece is just an abbreviation for a
   DW_OP_bit_piece operation with offset zero, this would preclude the
   use of DW_OP_piece for cases like single-precision FP values in a
   double-precision FP register on big-endian platforms.

One way of clearing this up would be to leave the placement to the
ABI.  This would also reduce the need for a more formal definition of
the terms "register", "natural memory representation", "least
significant bit end", etc.

E.g., change the definition to the following:

      If the location is a register, the placement of the piece
      within that register for the given size and offset is
      defined by the ABI.

--

Rejected - 2/2/2017

1.  Least significant bit is well defined for registers and values,
and is not dependent on the endianity of the memory representation
for these values.  Some architectures allow a value to be stored
in either big- or little-endian format; this does not alter the
representation in the registers.

2.  "Widening a register" is a concept out of the scope of the
DWARF Specification.  This appears to be a Quality of Implementation
issue.

3.  How a value is loaded into the register (perhaps by loading a
single-precision FP value into a double-precision FP register as
mentioned) does not change the register format, nor does it change
the interpretation of DW_OP_piece or DW_OP_bit_piece.  How a
producer generates a DWARF expression is a Quality of Implementation
issue.