Breakpoints, don't skip prologue of ifunc resolvers with debug info
authorPedro Alves <palves@redhat.com>
Thu, 26 Apr 2018 12:01:26 +0000 (13:01 +0100)
committerPedro Alves <palves@redhat.com>
Thu, 26 Apr 2018 12:06:53 +0000 (13:06 +0100)
commit76af0f26356580771a18c37de4ebccdfbc449356
treec8f893101273c3ac277b3b6ed1f8058a6dd90aaf
parent3467ec66bc1f30cf3ed7f9fe75234c96fc9c92d5
Breakpoints, don't skip prologue of ifunc resolvers with debug info

Without this patch, some of the tests added to gdb.base/gnu-ifunc.exp
by a following patch fail like so:

  FAIL: gdb.base/gnu-ifunc.exp: resolver_attr=0: resolver_debug=1: resolved_debug=0: set-break: before resolving: break gnu_ifunc
  FAIL: gdb.base/gnu-ifunc.exp: resolver_attr=0: resolver_debug=1: resolved_debug=0: set-break: before resolving: info breakpoints
  FAIL: gdb.base/gnu-ifunc.exp: resolver_attr=0: resolver_debug=1: resolved_debug=0: set-break: after resolving: break gnu_ifunc
  FAIL: gdb.base/gnu-ifunc.exp: resolver_attr=0: resolver_debug=1: resolved_debug=0: set-break: after resolving: info breakpoints
  FAIL: gdb.base/gnu-ifunc.exp: resolver_attr=0: resolver_debug=1: resolved_debug=1: set-break: before resolving: break gnu_ifunc
  FAIL: gdb.base/gnu-ifunc.exp: resolver_attr=0: resolver_debug=1: resolved_debug=1: set-break: before resolving: info breakpoints
  FAIL: gdb.base/gnu-ifunc.exp: resolver_attr=0: resolver_debug=1: resolved_debug=1: set-break: after resolving: break gnu_ifunc
  FAIL: gdb.base/gnu-ifunc.exp: resolver_attr=0: resolver_debug=1: resolved_debug=1: set-break: after resolving: info breakpoints

All of them trigger iff:

 - you have debug info for the ifunc resolver.
 - the resolver and the user-visible symbol have the same name.

If you have an ifunc that has a resolver with the same name as the
user visible symbol, debug info for the resolver masks out the ifunc
minsym.  When you set a breakpoint by name on the user visible symbol,
GDB finds the DWARF symbol for the resolver, and thinking that it's a
regular function, sets a breakpoint location past its prologue.

Like so, location 1.2, before the ifunc is resolved by the inferior:

  (gdb) break gnu_ifunc
  Breakpoint 2 at 0x7ffff7bd36ea (2 locations)
  (gdb) info breakpoints
  Num     Type           Disp Enb Address            What
  1       breakpoint     keep y   <MULTIPLE>
  1.1                         y     0x00007ffff7bd36ea <gnu_ifunc>
  1.2                         y     0x00007ffff7bd36f2 in gnu_ifunc at src/gdb/testsuite/gdb.base/gnu-ifunc-lib.c:34
  (gdb)

And like so, location 2.2, if you set the breakpoint after the ifunc
is resolved by the inferior (to "final"):

  (gdb) break gnu_ifunc
  Breakpoint 5 at 0x400757 (2 locations)
  (gdb) info breakpoints
  Num     Type           Disp Enb Address            What
  2       breakpoint     keep y   <MULTIPLE>
  2.1                         y     0x000000000040075a in final at src/gdb/testsuite/gdb.base/gnu-ifunc-resd.c:21
  2.2                         y     0x00007ffff7bd36f2 in gnu_ifunc at src/gdb/testsuite/gdb.base/gnu-ifunc-lib.c:34
  (gdb)

I don't think this is right because when users set a breakpoint at an
ifunc, they don't care about debugging the resolver.  Instead what you
should is a single location for the ifunc in the first case, and a
single location of the ifunc target in the second case.

gdb/ChangeLog:
2018-04-26  Pedro Alves  <palves@redhat.com>

* linespec.c (struct bound_minimal_symbol_search_key): New.
(convert_linespec_to_sals): Sort minimal symbols earlier.  Don't
skip first line if we found a GNU ifunc minimal symbol by name.
(compare_msymbols): Change parameters to work with a destructured
lhs minsym.
(compare_msymbols_for_qsort, compare_msymbols_for_bsearch): New
functions.
gdb/ChangeLog
gdb/linespec.c