[PATCH][gdb/testsuite] Fix gdb.dwarf2/dw2-dir-file-name.exp

[Tom de Vries]

On 8/17/22 14:01, Ulrich Weigand wrote:
> Carl Love <cel@us.ibm.com> wrote:
> 
>> PowerPC has two entry points, local and global.  The test used to set
>> the breakpoint for the function at the local entry point.  With your
>> changes, the breakpoint is now being set at the global breakpoint
> which
>> is before the local breakpoint.  The function is actually entered at
>> the local breakpoint thus gdb never "sees" the breakpoint that was
> set.
>> Specfically, here is the objdump for the test:
> 
>> 00000000100006e0 <compdir_missing__ldir_missing__file_basename>:
>>     100006e0:   02 10 40 3c     lis     r2,4098                      <-
>> Global entry point
>>     100006e4:   00 7f 42 38     addi    r2,r2,32512
> The local enty point is actually here:
>>     100006e8:   f8 ff e1 fb     std     r31,-8(r1)
>>     100006ec:   d1 ff 21 f8     stdu    r1,-48(r1)
>>     100006f0:   78 0b 3f 7c     mr      r31,r1
>>     100006f4:   00 00 00 60     nop                                  <-
>> Local entry point
>>     100006f8:   28 81 22 39     addi    r9,r2,-32472
> 
> Not here. This point might be the end of the prologue, which is a
> separate question from the local vs. global entry point issue.
> 
>> Perhaps Ulrich has some ideas???
> 
> This:
> +       gdb_breakpoint *$func
> is nearly always wrong, and test cases shouldn't be using it.
> 

Ack, I've posted a fix that doesn't use that construct anymore.

> I know that in the past, this construct was sometimes used with an
> intended meaning of "set a breakpoint at the start of a function
> without skipping the prologue", but it does *not* actually mean that.
> 
> The "*" operator tells the breakpoint logic to set a breakpoint on an
> absolute address.  If followed by a symbol, that symbol's value is used
> as that absolute address.  Now, on many platform, that symbol value
> matches the address of the first instruction of a function, so the
> "break *func" does more or less what's intended above.
> 
> But there are other platforms where this is not true, and the
> relationship between the function symbol value and the address of the
> first executed instruction is more complex.  This specifically applies
> to targets that implement gdbarch_deprecated_function_start_offset
> and/or gdbarch_skip_entrypoint.  (ppc64le uses the latter.)
> 

I think we could add a note to the docs explaining this.

> Note that these are *independent* of prologue skipping, and on
> platforms where it matters, they have to be used even when avoiding
> prologue skipping, in order to correctly find the first instruction of
> a function to set a breakpoint on.   "break *func" ignores this,
> leading to failures on such platforms.
> 
> 
> As I said initially, I think this construct should never be used in
> test cases (at least not in scenerios where it is intended to set a
> breakpoint that will actually be hit).
> 
> For this specific test, if the underlying problem is a bug in some
> architecture's prologue parser, then ideally this bug simply should be
> fixed. 

Yes, it should, and a test-case should be added for that.

In this test-case, we try to test a feature that has nothing to do with 
architecture-specific prologue skipping, so given that there can be 
problems with it on various targets, the test-case should simply skip it.

Thanks,
- Tom

> If we need to actually avoid prologue skipping for some real
> underlying reason, we should use a way that still handles function
> start offsets and entrypoint skipping.  For example, in a function with
> debug info including valid location lists, GDB will always avoid
> prologue skipping.  We could also think of adding an explicit linespec
> modifier that would allow to explicitly set a breakpoint on the first
> instruction of a function without skipping prologue ...
> 
> 
> Bye,
> Ulrich
>