[Bug gdb/34424] New: [gdb] DW_OP_bra incorrectly narrows typed conditions before zero test

https://sourceware.org/bugzilla/show_bug.cgi?id=34424

            Bug ID: 34424
           Summary: [gdb] DW_OP_bra incorrectly narrows typed conditions
                    before zero test
           Product: gdb
           Version: HEAD
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: gdb
          Assignee: unassigned at sourceware dot org
          Reporter: firmiana402 at gmail dot com
  Target Milestone: ---

DWARF v5 Section 2.5.1.5 defines DW_OP_bra as branching when the
popped stack value is not the constant zero.  It does not restrict
the condition to the generic type.  The operation requires only a
zero/nonzero test, not conversion to an 8-byte integer.

GDB currently implements the condition as:

    case DW_OP_bra:
      ...
      val = fetch (0);
      dwarf_require_integral (val->type ());
      if (value_as_long (val) != 0)
        op_ptr += offset;
      pop ();

For example, this expression uses an unsigned 128-bit condition equal
to 2^64 on little-endian x86_64:

    DW_OP_const_type <unsigned __int128>, 16,
        [00 00 00 00 00 00 00 00 01 00 00 00 00 00 00 00]
    DW_OP_bra +5
    DW_OP_const1u 0
    DW_OP_skip +2
    DW_OP_const1u 42
    DW_OP_stack_value

The +5 target is DW_OP_const1u 42.  The condition is unambiguously
nonzero, so the branch must be taken and the final value must be 42.
Instead, value_as_long enters GDB's 8-byte unsigned-integer
extraction path, which rejects the significant ninth byte:

    Value cannot be represented as integer of 8 bytes.

This is specific to the DW_OP_bra condition path.  GDB correctly
evaluates DW_OP_ne on the same 16-byte type: 2^64 != 2^64 produces 0,
while (2^64 + 1) != 2^64 produces 1.  That comparison uses
value_as_mpz rather than narrowing either operand to 8 bytes.

DW_OP_bra should likewise test the condition for zero through a
width-independent integer path instead of value_as_long.

-- 
You are receiving this mail because:
You are on the CC list for the bug.