Re: [PATCH v3 1/9] Allow COND_EXPR and VEC_COND_EXPR condtions to trap

[Ilya Leoshkevich <iii@linux.ibm.com>]

> Am 06.09.2019 um 13:07 schrieb Richard Biener <richard.guenther@gmail.com>:
> 
> On Thu, Sep 5, 2019 at 1:10 PM Ilya Leoshkevich <iii@linux.ibm.com> wrote:
>> 
>> Right now gimplifier does not allow VEC_COND_EXPR's condition to trap
>> and introduces a temporary if this could happen, for example, generating
>> 
>>  _5 = _4 > { 2.0e+0, 2.0e+0, 2.0e+0, 2.0e+0 };
>>  _6 = VEC_COND_EXPR <_5, { -1, -1, -1, -1 }, { 0, 0, 0, 0 }>;
>> 
>> from GENERIC
>> 
>>  VEC_COND_EXPR < (*b > { 2.0e+0, 2.0e+0, 2.0e+0, 2.0e+0 }) ,
>>                  { -1, -1, -1, -1 } ,
>>                  { 0, 0, 0, 0 } >
>> 
>> This is not necessary and makes the resulting GIMPLE harder to analyze.
>> In particular, one of the next patches in series needs to get to
>> VEC_COND_EXPR's comparison code, which is not possible when a temporary
>> is introduced.
>> 
>> This patch takes special care to avoid introducing trapping comparisons
>> in GIMPLE_COND.  They are not allowed, because they would require 3
>> outgoing edges (then, else and EH), which is awkward to say the least.
>> Therefore, computations of such conditions should live in their own basic
>> blocks.
> 
> Comments inline (thanks for the work btw)
> 
>> #endif /* GCC_GIMPLE_EXPR_H */
>> diff --git a/gcc/gimple.c b/gcc/gimple.c
>> index 633ef512a19..fd14fbec15e 100644
>> --- a/gcc/gimple.c
>> +++ b/gcc/gimple.c
>> @@ -2144,6 +2144,8 @@ gimple_could_trap_p_1 (gimple *s, bool include_mem, bool include_stores)
>>       op = gimple_assign_rhs_code (s);
>>       if (get_gimple_rhs_class (op) == GIMPLE_BINARY_RHS)
>>        div = gimple_assign_rhs2 (s);
>> +      else if (op == COND_EXPR || op == VEC_COND_EXPR)
>> +       op = TREE_CODE (gimple_assign_rhs1 (s));
> 
> I think this is not correct since we can have
> 
> int i = fp > 1. ? intval1 : intval2
> 
> and thus FLOAT_TYPE_P (t)  is wrong.  You need to do
> 
>  t = TREE_TYPE (op);
> 
> as well I think.

Doesn't this mean there is a problem with the existing logic too? If `s`
is

    int i = fp > 1.;

then

    t = gimple_expr_type (s);

would give us BOOLEAN_TYPE instead of REAL_TYPE.


Also, the new logic will probably be a bit more complicated, since I
will first have to do:

    tree cond = gimple_assign_rhs1 (s);

then see if `cond` is not e.g. an SSA_NAME, but rather a tcc_comparison,
and only then

    t = TREE_TYPE (TREE_OPERAND (cond, 0))

So I'd rather send a new version before merging this :-)