gdb/x86: handle stap probe arguments in xmm registers

On x86 machines with xmm register, and with recent versions of
systemtap (and gcc?), it can occur that stap probe arguments will be
placed into xmm registers.

I notice this happening on a current Fedora Rawhide install with the
following package versions installed:

  $ rpm -q glibc systemtap gcc
  glibc-2.35.9000-10.fc37.x86_64
  systemtap-4.7~pre16468670g9f253544-1.fc37.x86_64
  gcc-12.0.1-0.12.fc37.x86_64

If I check the probe data in libc, I see this:

  $ readelf -n /lib64/libc.so.6
  ...
  stapsdt              0x0000004d       NT_STAPSDT (SystemTap probe descriptors)
    Provider: libc
    Name: pthread_start
    Location: 0x0000000000090ac3, Base: 0x00000000001c65c4, Semaphore: 0x0000000000000000
    Arguments: 8@%xmm1 8@1600(%rbx) 8@1608(%rbx)
  stapsdt              0x00000050       NT_STAPSDT (SystemTap probe descriptors)
    Provider: libc
    Name: pthread_create
    Location: 0x00000000000912f1, Base: 0x00000000001c65c4, Semaphore: 0x0000000000000000
    Arguments: 8@%xmm1 8@%r13 8@8(%rsp) 8@16(%rsp)
  ...

Notice that for both of these probes, the first argument is a uint64_t
stored in the xmm1 register.

Unfortunately, if I try to use this probe within GDB, then I can't
view the first argument.  Here's an example session:

  $ gdb $(which gdb)
  (gdb) start
  ...
  (gdb) info probes stap  libc pthread_create
  ...
  (gdb) break *0x00007ffff729e2f1       # Use address of probe.
  (gdb) continue
  ...
  (gdb) p $_probe_arg0
  Invalid cast.

What's going wrong?  If I re-run my session, but this time use 'set
debug stap-expression 1', this is what I see:

  (gdb) set debug stap-expression 1
  (gdb) p $_probe_arg0
  Operation: UNOP_CAST
   Operation: OP_REGISTER
    String: xmm1
   Type: uint64_t
  Operation: UNOP_CAST
   Operation: OP_REGISTER
    String: r13
   Type: uint64_t
  Operation: UNOP_CAST
   Operation: UNOP_IND
    Operation: UNOP_CAST
     Operation: BINOP_ADD
      Operation: OP_LONG
       Type: long
       Constant: 0x0000000000000008
      Operation: OP_REGISTER
       String: rsp
     Type: uint64_t *
   Type: uint64_t
  Operation: UNOP_CAST
   Operation: UNOP_IND
    Operation: UNOP_CAST
     Operation: BINOP_ADD
      Operation: OP_LONG
       Type: long
       Constant: 0x0000000000000010
      Operation: OP_REGISTER
       String: rsp
     Type: uint64_t *
   Type: uint64_t
  Invalid cast.
  (gdb)

The important bit is this:

  Operation: UNOP_CAST
   Operation: OP_REGISTER
    String: xmm1
   Type: uint64_t

Which is where we cast the xmm1 register to uint64_t.  And the final
piece of the puzzle is:

  (gdb) ptype $xmm1
  type = union vec128 {
      v8bf16 v8_bfloat16;
      v4f v4_float;
      v2d v2_double;
      v16i8 v16_int8;
      v8i16 v8_int16;
      v4i32 v4_int32;
      v2i64 v2_int64;
      uint128_t uint128;
  }

So, we are attempting to cast a union type to a scalar type, which is
not supporting in C/C++, and as a consequence GDB's expression
evaluator throws an error when we attempt to do this.

The first approach I considered for solving this problem was to try
and make use of gdbarch_stap_adjust_register.  We already have a
gdbarch method (gdbarch_stap_adjust_register) that allows us to tweak
the name of the register that we access.  Currently only x86
architectures use this to transform things like ax to eax in some
cases.

I wondered, what if we change gdbarch_stap_adjust_register to do more
than just change the register names?  What if this method instead
became gdbarch_stap_read_register.  This new method would return a
operation_up, and would take the register name, and the type we are
trying to read from the register, and return the operation that
actually reads the register.

The default implementation of this method would just use
user_reg_map_name_to_regnum, and then create a register_operation,
like we already do in stap_parse_register_operand.  But, for x86
architectures this method would fist possibly adjust the register
name, then do the default action to read the register.  Finally, for
x86 this method would spot when we were accessing an xmm register,
and, based on the type being pulled from the register, would extract
the correct field from the union.

The benefit of this approach is that it would work with the expression
types that GDB currently supports.  The draw back would be that this
approach would not be very generic.  We'd need code to handle each
sub-field size with an xmm register.  If other architectures started
using vector registers for probe arguments, those architectures would
have to create their own gdbarch_stap_read_register method.  And
finally, the type of the xmm registers comes from the type defined in
the target description, there's a risk that GDB might end up
hard-coding the names of type sub-fields, then if a target uses a
different target description, with different field names for xmm
registers, the stap probes would stop working.

And so, based on all the above draw backs, I rejected this first
approach.

My second plan involves adding a new expression type to GDB called
unop_extract_operation.  This new expression takes a value and a type,
during evaluation the value contents are fetched, and then a new value
is extracted from the value contents (based on type).  This is similar
to the following C expression:

  result_value = *((output_type *) &input_value);

Obviously we can't actually build this expression in this case, as the
input_value is in a register, but hopefully the above makes it clearer
what I'm trying to do.

The benefit of the new expression approach is that this code can be
shared across all architectures, and it doesn't care about sub-field
names within the union type.

The draw-backs that I see are potential future problems if arguments
are not stored within the least significant bytes of the register.
However if/when that becomes an issue we can adapt the
gdbarch_stap_read_register approach to allow architectures to control
how a value is extracted.

For testing, I've extended the existing gdb.base/stap-probe.exp test
to include a function that tries to force an argument into an xmm
register.  Obviously, that will only work on a x86 target, so I've
guarded the new function with an appropriate GCC define.  In the exp
script we use readelf to check if the probe exists, and is using the
xmm register.

If the probe doesn't exist then the associated tests are skipped.

If the probe exists, put isn't using the xmm register (which will
depend on systemtap/gcc versions), then again, the tests are skipped.

Otherwise, we can run the test.  I think the cost of running readelf
is pretty low, so I don't feel too bad making all the non-xmm targets
running this step.

I found that on a Fedora 35 install, with these packages installed, I
was able to run this test and have the probe argument be placed in an
xmm register:

  $ rpm -q systemtap gcc glibc
  systemtap-4.6-4.fc35.x86_64
  gcc-11.2.1-9.fc35.x86_64
  glibc-2.34-7.fc35.x86_64

Finally, as this patch adds a new operation type, then I need to
consider how to generate an agent expression for the new operation
type.

I have kicked the can down the road a bit on this.  In the function
stap_parse_register_operand, I only create a unop_extract_operation in
the case where the register type is non-scalar, this means that in
most cases I don't need to worry about generating an agent expression
at all.

In the xmm register case, when an unop_extract_operation will be
created, I have sketched out how the agent expression could be
handled, however, this code is currently not reached.  When we try to
generate the agent expression to place the xmm register on the stack,
GDB hits this error:

  (gdb) trace -probe-stap test:xmmreg
  Tracepoint 1 at 0x401166
  (gdb) actions
  Enter actions for tracepoint 1, one per line.
  End with a line saying just "end".
  >collect $_probe_arg0
  Value not scalar: cannot be an rvalue.

This is because GDB doesn't currently support placing non-scalar types
on the agent expression evaluation stack.  Solving this is clearly
related to the original problem, but feels a bit like a second
problem.  I'd like to get feedback on whether my approach to solving
the original problem is acceptable or not before I start looking at
how to handle xmm registers within agent expressions.

0 files changed, 0 insertions, 0 deletions