GDBserver: Fix "Cond. jump or move depends on uninit value" in x87 code
authorPedro Alves <palves@redhat.com>
Wed, 11 Jul 2018 18:49:19 +0000 (19:49 +0100)
committerPedro Alves <palves@redhat.com>
Wed, 11 Jul 2018 18:49:19 +0000 (19:49 +0100)
commitcb19713281c69c3625dd447a0c2ce539761989e2
tree3a5020d559dba6453f8343726001ba5d3bbaa3ed
parentc597cc3d6eb76802dd079b1262f2d425e07da3eb
GDBserver: Fix "Cond. jump or move depends on uninit value" in x87 code

Running gdbserver under Valgrind I get:

  ==26925== Conditional jump or move depends on uninitialised value(s)
  ==26925==    at 0x473E7F: i387_cache_to_xsave(regcache*, void*) (i387-fp.c:579)
  ==26925==    by 0x46E3ED: x86_fill_xstateregset(regcache*, void*) (linux-x86-low.c:418)
  ==26925==    by 0x45E747: regsets_store_inferior_registers(regsets_info*, regcache*) (linux-low.c:5456)
  ==26925==    by 0x45EEF8: linux_store_registers(regcache*, int) (linux-low.c:5731)
  ==26925==    by 0x426441: regcache_invalidate_thread(thread_info*) (regcache.c:89)
  ==26925==    by 0x45CCAF: linux_resume_one_lwp_throw(lwp_info*, int, int, siginfo_t*) (linux-low.c:4447)
  ==26925==    by 0x45CE2A: linux_resume_one_lwp(lwp_info*, int, int, siginfo_t*) (linux-low.c:4519)
  ==26925==    by 0x45E17C: proceed_one_lwp(thread_info*, lwp_info*) (linux-low.c:5216)
  ==26925==    by 0x45DC81: linux_resume_one_thread(thread_info*, bool) (linux-low.c:5031)
  ==26925==    by 0x45DD34: linux_resume(thread_resume*, unsigned long)::{lambda(thread_info*)#2}::operator()(thread_info*) const (linux-low.c:5095)
  ==26925==    by 0x462907: void for_each_thread<linux_resume(thread_resume*, unsigned long)::{lambda(thread_info*)#2}>(linux_resume(thread_resume*, unsigned long)::{lambda(thread_info*)#2}) (gdbthread.h:150)
  ==26925==    by 0x45DE62: linux_resume(thread_resume*, unsigned long) (linux-low.c:5093)
  ==26925==
  ==26925== Conditional jump or move depends on uninitialised value(s)
  ==26925==    at 0x473EBD: i387_cache_to_xsave(regcache*, void*) (i387-fp.c:586)
  ==26925==    by 0x46E3ED: x86_fill_xstateregset(regcache*, void*) (linux-x86-low.c:418)
  ==26925==    by 0x45E747: regsets_store_inferior_registers(regsets_info*, regcache*) (linux-low.c:5456)
  ==26925==    by 0x45EEF8: linux_store_registers(regcache*, int) (linux-low.c:5731)
  ==26925==    by 0x426441: regcache_invalidate_thread(thread_info*) (regcache.c:89)
  ==26925==    by 0x45CCAF: linux_resume_one_lwp_throw(lwp_info*, int, int, siginfo_t*) (linux-low.c:4447)
  ==26925==    by 0x45CE2A: linux_resume_one_lwp(lwp_info*, int, int, siginfo_t*) (linux-low.c:4519)
  ==26925==    by 0x45E17C: proceed_one_lwp(thread_info*, lwp_info*) (linux-low.c:5216)
  ==26925==    by 0x45DC81: linux_resume_one_thread(thread_info*, bool) (linux-low.c:5031)
  ==26925==    by 0x45DD34: linux_resume(thread_resume*, unsigned long)::{lambda(thread_info*)#2}::operator()(thread_info*) const (linux-low.c:5095)
  ==26925==    by 0x462907: void for_each_thread<linux_resume(thread_resume*, unsigned long)::{lambda(thread_info*)#2}>(linux_resume(thread_resume*, unsigned long)::{lambda(thread_info*)#2}) (gdbthread.h:150)
  ==26925==    by 0x45DE62: linux_resume(thread_resume*, unsigned long) (linux-low.c:5093)

The problem is a type/width mismatch in code like this, in
gdbserver/i387-fp.c:

  /* Some registers are 16-bit.  */
  collect_register_by_name (regcache, "fctrl", &val);
  fp->fctrl = val;

In the above code:

 #1 - 'val' is a 64-bit unsigned long.

 #2 - "fctrl" is 32-bit in the register cache, thus half of 'val' is
      left uninitialized by collect_register_by_name, which works with
      an untyped raw buffer output (i.e., void*).

 #3 - fp->fctrl is an unsigned short (16-bit).  For some such
      registers we're masking off the uninitialized bits with 0xffff,
      but not in all cases.

We end up in such a fragile situation because
collect_registers_by_name works with an untyped output buffer pointer,
making it easy to pass a pointer to a variable of the wrong size.

Fix this by using regcache_raw_get_unsigned instead (actually a new
regcache_raw_get_unsigned_by_name wrapper), which always returns a
zero-extended ULONGEST register value.  It ends up simplifying the
i387-tdep.c code a bit, even.

gdb/gdbserver/ChangeLog:
2018-07-11  Pedro Alves  <palves@redhat.com>

* i387-fp.c (i387_cache_to_fsave, cache_to_fxsave)
(i387_cache_to_xsave): Use regcache_raw_get_unsigned_by_name
instead of collect_register_by_name.
* regcache.c (regcache_raw_get_unsigned_by_name): New.
* regcache.h (regcache_raw_get_unsigned_by_name): New.
gdb/gdbserver/ChangeLog
gdb/gdbserver/i387-fp.c
gdb/gdbserver/regcache.c
gdb/gdbserver/regcache.h