[Mesa-dev] [PATCH 007/133] nir: add a validation pass

[Eric Anholt]

Connor Abbott <cwabbott0 at gmail.com> writes:

> On Thu, Dec 18, 2014 at 1:49 PM, Eric Anholt <eric at anholt.net> wrote:
>> Connor Abbott <cwabbott0 at gmail.com> writes:
>>
>>> On Thu, Dec 18, 2014 at 2:01 AM, Eric Anholt <eric at anholt.net> wrote:
>>>> Jason Ekstrand <jason at jlekstrand.net> writes:
>>>>
>>>>> From: Connor Abbott <connor.abbott at intel.com>
>>>>>
>>>>> This is similar to ir_validate.cpp.
>>>>>
>>>>> v2: Jason Ekstrand <jason.ekstrand at intel.com>:
>>>>>    whitespace fixes
>>>>
>>>> I have again not reviewed the control flow bits.  Couple of questions I
>>>> had, though:
>>>>
>>>>> +static void
>>>>> +validate_var_use(nir_variable *var, validate_state *state)
>>>>> +{
>>>>> +   if (var->data.mode == nir_var_local) {
>>>>> +      struct hash_entry *entry =
>>>>> +         _mesa_hash_table_search(state->var_defs, _mesa_hash_pointer(var),
>>>>> +                                 var);
>>>>> +
>>>>> +      assert(entry);
>>>>> +      assert((nir_function_impl *) entry->data == state->impl);
>>>>> +   }
>>>>> +}
>>>>
>>>> Is there guaranteed to be a def of a local variable before a use?  It
>>>> would be undefined execution behavior, but not assertion failure
>>>> quality, right?
>>>
>>> Yes, that's correct - there are no guarantees about this for variables
>>> and registers. For SSA values, the definition should always dominate
>>> the use (see the TODO about that) because a lot of SSA algorithms
>>> assume that, so we model the use-before-def case by pointing the use
>>> to a nir_ssa_undef_instr.
>>
>> OK, so it seems like this validation needs to be dropped.
>
> No, this isn't validating what you think it's validating. This
> function is called every time a variable dereference happens, and
> checks that the dereference is in the same function implementation as
> the definition of the variable. Maybe we can add a && "deferencing a
> local variable defined in a different function" to the assert to make
> it clear what what's happening.

Ah, I was confused by the name of the table being "defs".  var_decls
might be better.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 818 bytes
Desc: not available
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20141219/9b81b0ae/attachment.sig>