Re: [PATCH] range-op-float: Improve binary reverse operations

[Andrew MacLeod <amacleod@redhat.com>]

On 12/5/22 10:33, Jakub Jelinek wrote:
> Hi!
>
> On Mon, Dec 05, 2022 at 02:29:36PM +0100, Aldy Hernandez wrote:
>>> So like this for multiplication op1/2_range if it passes bootstrap/regtest?
>>> For division I'll need to go to a drawing board...
>> Sure, looks good to me.
> Ulrich just filed PR107972, so in the light of that PR the following patch
> attempts to do that differently.
>
> Is this also ok if it passes bootstrap/regtest?


Id actually prefer to avoid passing the tree code around... we're trying 
to avoid that sort of thing even though Aldy temporarily introduced them 
to range-ops. Hes suppose to remove that next stage 1 :-P   Ideally 
anything "special" is locally contained to the specific routine.

It looks like the only time we actually do anything different is for 
divide op2_range?   the divide op1_range seems to still call 
frange_drop_infs() under the same conditions..

In which case, maybe we can just change op2_range to contain all the 
special casing.. frange_drop_infs doesnt seem overly complicated, so 
just specialize it?  ie for divide op2_range do something like this 
instead of the call?:

if (!ret)
   return false;
if (r.known_isnan () || lhs.known_isnan () || r.undefined_p ())
   r.set_varying (type);
else if (!lhs.maybe_isnan ())
   {
     r.clear_nan();
     if (!contains_zero_p (lhs_lb, lhs_ub))
       frange_drop_infs (r, type);
   }
else
   r.update_nan ();
return true;


or whatever the sequence precisely works out to.

Andrew


> As for testcase, I've tried both attached testcases, but unfortunately it
> seems that in neither of the cases we actually figure out that res range
> is finite (or for last function non-zero ordered).  So there is further
> work needed on that.
>
> 2022-12-05  Jakub Jelinek  <jakub@redhat.com>
>
> 	PR tree-optimization/107972
> 	* range-op-float.cc (frange_drop_infs): New function.
> 	(float_binary_op_range_finish): Add OP argument.  If OP is not
> 	RDIV_EXPR and lhs is finite, r must be finite too.
> 	(foperator_plus::op1_range, foperator_minus::op1_range,
> 	foperator_minus::op2_range, foperator_mult::op1_range): Pass
> 	operation code to float_binary_op_range_finish.
> 	(foperator_div::op1_range): Pass RDIV_EXPR to
> 	float_binary_op_range_finish.  If lhs is finite, r must be finite
> 	too.
> 	(foperator_div::op2_range): Pass RDIV_EXPR to
> 	float_binary_op_range_finish.  If lhs is not NAN nor zero, r must
> 	be finite.
>
> --- gcc/range-op-float.cc.jj	2022-12-05 11:17:34.900573272 +0100
> +++ gcc/range-op-float.cc	2022-12-05 16:13:54.414845672 +0100
> @@ -330,6 +330,18 @@ frange_drop_ninf (frange &r, tree type)
>     r.intersect (tmp);
>   }
>   
> +// Crop R to [MIN, MAX] where MAX is the maximum representable number
> +// for TYPE and MIN the minimum representable number for TYPE.
> +
> +static inline void
> +frange_drop_infs (frange &r, tree type)
> +{
> +  REAL_VALUE_TYPE max = real_max_representable (type);
> +  REAL_VALUE_TYPE min = real_min_representable (type);
> +  frange tmp (type, min, max);
> +  r.intersect (tmp);
> +}
> +
>   // If zero is in R, make sure both -0.0 and +0.0 are in the range.
>   
>   static inline void
> @@ -1883,7 +1895,7 @@ foperator_unordered_equal::op1_range (fr
>   
>   static bool
>   float_binary_op_range_finish (bool ret, frange &r, tree type,
> -			      const frange &lhs)
> +			      const frange &lhs, enum tree_code op)
>   {
>     if (!ret)
>       return false;
> @@ -1904,7 +1916,16 @@ float_binary_op_range_finish (bool ret,
>     // If lhs isn't NAN, then neither operand could be NAN,
>     // even if the reverse operation does introduce a maybe_nan.
>     if (!lhs.maybe_isnan ())
> -    r.clear_nan ();
> +    {
> +      r.clear_nan ();
> +      if (op != RDIV_EXPR
> +	  && !real_isinf (&lhs.lower_bound ())
> +	  && !real_isinf (&lhs.upper_bound ()))
> +	// For reverse + or - or *, if result is finite, then r must
> +	// be finite too, as X + INF or X - INF or X * INF is always
> +	// +-INF or NAN.  For division handle it in the caller.
> +	frange_drop_infs (r, type);
> +    }
>     // If lhs is a maybe or known NAN, the operand could be
>     // NAN.
>     else
> @@ -2020,7 +2041,7 @@ public:
>       if (!minus)
>         return false;
>       return float_binary_op_range_finish (minus.fold_range (r, type, lhs, op2),
> -					 r, type, lhs);
> +					 r, type, lhs, PLUS_EXPR);
>     }
>     virtual bool op2_range (frange &r, tree type,
>   			  const frange &lhs,
> @@ -2067,7 +2088,7 @@ public:
>         return false;
>       return float_binary_op_range_finish (fop_plus.fold_range (r, type, lhs,
>   							      op2),
> -					 r, type, lhs);
> +					 r, type, lhs, MINUS_EXPR);
>     }
>     virtual bool op2_range (frange &r, tree type,
>   			  const frange &lhs,
> @@ -2077,7 +2098,7 @@ public:
>       if (lhs.undefined_p ())
>         return false;
>       return float_binary_op_range_finish (fold_range (r, type, op1, lhs),
> -					 r, type, lhs);
> +					 r, type, lhs, MINUS_EXPR);
>     }
>   private:
>     void rv_fold (REAL_VALUE_TYPE &lb, REAL_VALUE_TYPE &ub, bool &maybe_nan,
> @@ -2166,7 +2187,7 @@ public:
>       // or if lhs must be zero and op2 doesn't include zero, it would be
>       // UNDEFINED, while rdiv.fold_range computes a zero or singleton INF
>       // range.  Those are supersets of UNDEFINED, so let's keep that way.
> -    return float_binary_op_range_finish (ret, r, type, lhs);
> +    return float_binary_op_range_finish (ret, r, type, lhs, MULT_EXPR);
>     }
>     virtual bool op2_range (frange &r, tree type,
>   			  const frange &lhs,
> @@ -2313,7 +2334,12 @@ public:
>   	zero_to_inf_range (lb, ub, signbit_known);
>   	r.set (type, lb, ub);
>         }
> -    return float_binary_op_range_finish (ret, r, type, lhs);
> +    // If lhs must be finite (can't be +-INF nor NAN), then op1 must be
> +    // finite too - +-INF / anything is either +-INF or NAN (NAN if op2 is
> +    // +-INF or NAN).
> +    if (!real_isinf (&lhs_lb) && !real_isinf (&lhs_ub) && !lhs.maybe_isnan ())
> +      frange_drop_infs (r, type);
> +    return float_binary_op_range_finish (ret, r, type, lhs, RDIV_EXPR);
>     }
>     virtual bool op2_range (frange &r, tree type,
>   			  const frange &lhs,
> @@ -2341,7 +2367,11 @@ public:
>   	zero_to_inf_range (lb, ub, signbit_known);
>   	r.set (type, lb, ub);
>         }
> -    return float_binary_op_range_finish (ret, r, type, lhs);
> +    // If lhs can't be zero nor NAN, then op2 must be finite - anything / +-INF
> +    // is either +-0 or NAN (NAN if op1 is +-INF or NAN).
> +    if (!contains_zero_p (lhs_lb, lhs_ub) && !lhs.maybe_isnan ())
> +      frange_drop_infs (r, type);
> +    return float_binary_op_range_finish (ret, r, type, lhs, RDIV_EXPR);
>     }
>   private:
>     void rv_fold (REAL_VALUE_TYPE &lb, REAL_VALUE_TYPE &ub, bool &maybe_nan,
>
>
> 	Jakub