Handle DW_OP_form_tls_address
authorTom Tromey <tom@tromey.com>
Mon, 22 Aug 2016 22:56:52 +0000 (16:56 -0600)
committerTom Tromey <tom@tromey.com>
Fri, 2 Sep 2016 17:33:00 +0000 (11:33 -0600)
commit4aa4e28bdcf5f0d733def62b542fea11d5f219d5
tree3a5f878a8dd275ea370cf6e8172f8ed7f92accd1
parent1b2d85b255f11d0fce6c323f9e113673445c961a
Handle DW_OP_form_tls_address

Currently gdb supports DW_OP_GNU_push_tls_address, but not
DW_OP_form_tls_address.  I think it would be better if the toolchain
as a whole moved to using the standard opcode, and the prerequisite to
this is getting gdb to recognize it.

GCC can sometimes emit DW_OP_form_tls_address for emultls targets.  As
far as I know, nobody has ever tried this with gdb (since it wouldn't
work at all).

I don't think there's a major drawback to using a single opcode for
all targets, because computing the location of a thread-local is
already target specific.

This is PR gdb/11616.

I don't know how to write a test case for this; though it's worth
noting that there aren't explicit tests for DW_OP_GNU_push_tls_address
either -- and if I change GCC, these paths will be tested to the same
extent they are now.

2016-09-02  Tom Tromey  <tom@tromey.com>

PR gdb/11616:
* dwarf2read.c (decode_locdesc): Handle DW_OP_form_tls_address.
* dwarf2loc.c (dwarf2_compile_expr_to_ax): Handle
DW_OP_form_tls_address.
(locexpr_describe_location_piece): Likewise.
* dwarf2expr.h (struct dwarf_expr_context_funcs): Update comment.
* dwarf2expr.c (execute_stack_op): Handle DW_OP_form_tls_address.
(ctx_no_get_tls_address): Mention DW_OP_form_tls_address.
* compile/compile-loc2c.c (struct insn_info): Update comment.
(compute_stack_depth_worker): Handle DW_OP_form_tls_address.
gdb/ChangeLog
gdb/compile/compile-loc2c.c
gdb/dwarf2expr.c
gdb/dwarf2expr.h
gdb/dwarf2loc.c
gdb/dwarf2read.c