Re: [patch] Fix PHI optimization issue with boolean types

[Jeff Law]

On 10/18/2016 02:35 AM, Richard Biener wrote:
On Tue, Oct 18, 2016 at 8:35 AM, Eric Botcazou <ebotca...@adacore.com> wrote:
Hi,
this is a regression present on the mainline and 6 branch: the compiler now
generates wrong code for the attached testcase at -O because of an internal
conflict about boolean types.  The sequence is as follows.  In .mergephi3:

  boolean _22;
  p__enum res;

  <bb 9>:
  if (_22 != 0)
    goto <bb 10>;
  else
    goto <bb 11>;

  <bb 10>:

  <bb 11>:
  # res_17 = PHI <2(8), 0(9), 1(10)>

is turned into:

COND_EXPR in block 9 and PHI in block 11 converted to straightline code.
PHI res_17 changed to factor conversion out from COND_EXPR.
New stmt with CAST that defines res_17.

  boolean _33;

  <bb 10>:
  # _33 = PHI <2(8), _22(9)>
  res_17 = (p__enum) _33;

[...]

  <bb 12>:
  if (res_17 != 0)
    goto <bb 13>;
  else
    goto <bb 14>;

  <bb 13>:
  _12 = res_17 == 2;
  _13 = (integer) _12

in .phiopt1.  So boolean _33 can have value 2.  Later forwprop3 propagates _33
into the uses of res_17:

  <bb 12>:
  if (_33 != 0)
    goto <bb 13>;
  else
    goto <bb 14>;

  <bb 13>:
  _12 = _33 == 2;
  _13 = (integer) _12;

and DOM3 deduces:

  <bb 13>:
  _12 = 0;
  _13 = 0;

because it computes that _33 has value 1 in BB 13 since it's a boolean.

The problem was introduced by the new factor_out_conditional_conversion:

      /* If arg1 is an INTEGER_CST, fold it to new type.  */
      if (INTEGRAL_TYPE_P (TREE_TYPE (new_arg0))
          && int_fits_type_p (arg1, TREE_TYPE (new_arg0)))
        {
          if (gimple_assign_cast_p (arg0_def_stmt))
            new_arg1 = fold_convert (TREE_TYPE (new_arg0), arg1);
          else
            return NULL;
        }
      else
        return NULL

int_fits_type_p is documented as taking only INTEGER_TYPE, but is invoked
on constant 2 and a BOOLEAN_TYPE and returns true.

BOOLEAN_TYPE is special in Ada: it has precision 8 and range [0;255] so the
outcome of int_fits_type_p is not unreasonable.  But this goes against the
various transformations applied to boolean types in the compiler, which all
assume that they can only take values 0 or 1.

Hence the attached fix (which should be a no-op except for Ada), tested on
x86_64-suse-linux, OK for mainline and 6 branch?

Hmm, I wonder if the patch is a good approach as you are likely only papering
over a single of possibly very many affected places (wi::fits_to_tree_p comes
immediately to my mind).  I suppose a better way would be for Ada to not
lie about those types and not use BOOLEAN_TYPE but INTEGER_TYPE.
Because BOOLEAN_TYPE types only have two values as documented in
tree.def:

/* Boolean type (true or false are the only values).  Looks like an
   INTEGRAL_TYPE.  */
DEFTREECODE (BOOLEAN_TYPE, "boolean_type", tcc_type, 0)

There are not many references to BOOLEAN_TYPE in ada/gcc-interface
thus it shouldn't be hard to change ... (looks like Ada might even prefer
ENUMERAL_TYPE here).

Thanks,
Richard.


2016-10-18  Eric Botcazou  <ebotca...@adacore.com>

        * tree.c (int_fits_type_p): Accept only 0 and 1 for boolean types.
Alternately we could check the precision/range in the optimizers.  We do 
that in various places because of this exact issue, I could well have 
missed one or more.

Though I would have a preference for fixing int_fits_type_p which seems 
like it'll catch the cases we care about without having to twiddle every 
optimizer.

jeff