Re: [PATCH] middle-end/57245 - honor -frounding-math in real truncation

[Richard Biener]

On October 27, 2021 4:44:53 PM GMT+02:00, Jakub Jelinek <jakub@redhat.com> wrote:
>On Wed, Oct 27, 2021 at 04:29:38PM +0200, Richard Biener wrote:
>> So something like the following below?  Note I have to fix 
>> simplify_const_unary_operation to not perform the invalid constant
>> folding with (not worrying about the exact conversion case - I doubt
>> any of the constant folding is really relevant on RTL these days,
>> maybe we should simply punt for all unary float-float ops when either
>> mode has sign dependent rounding modes)
>> 
>> diff --git a/gcc/simplify-rtx.c b/gcc/simplify-rtx.c
>> index bbbd6b74942..9522a31570e 100644
>> --- a/gcc/simplify-rtx.c
>> +++ b/gcc/simplify-rtx.c
>> @@ -2068,6 +2073,9 @@ simplify_const_unary_operation (enum rtx_code code, 
>> machine_mode mode,
>>              and the operand is a signaling NaN.  */
>>           if (HONOR_SNANS (mode) && REAL_VALUE_ISSIGNALING_NAN (d))
>>             return NULL_RTX;
>> +         /* Or if flag_rounding_math is on.  */
>> +         if (HONOR_SIGN_DEPENDENT_ROUNDING (mode))
>> +           return NULL_RTX;
>>           d = real_value_truncate (mode, d);
>>           break;
>
>Won't this stop folding of truncations that are never a problem?
>I mean at least if the wider float mode constant is exactly representable
>in the narrower float mode, no matter what rounding mode is used the value
>will be always the same...
>And people use
>  float f = 1.0;
>or
>  float f = 1.25;
>etc. a lot.

Yes, but I do expect any such opportunities to be realized on GENERIC/GIMPLE? 

>So perhaps again
>	if (HONOR_SIGN_DEPENDENT_ROUNDING (mode)
>	    && !exact_real_truncate (mode, &d))
>	  return NULL_RTX;
>?

Sure, for this case it's short and straight forward. 

>
>> /* PR57245 */
>> /* { dg-do run } */
>> /* { dg-require-effective-target fenv } */
>> /* { dg-additional-options "-frounding-math" } */
>> 
>> #include <fenv.h>
>> #include <stdlib.h>
>> 
>> int
>> main ()
>> {
>
>Roughly yes.  Some tests also do #ifdef FE_*, so in your case
>> #if __DBL_MANT_DIG__ == 53 && __FLT_MANT_DIG__ == 24
>+#ifdef FE_UPWARD

Ah, OK. Will fix. 

Richard. 

>>   fesetround (FE_UPWARD);
>>   float f = 1.3;
>>   if (f != 0x1.4ccccep+0f)
>>     __builtin_abort ();
>+#endif
>+#ifdef FE_TONEAREST
>etc.
>>   fesetround (FE_TONEAREST);
>>   /* Use different actual values so the bogus CSE we perform does not
>>      break things.  */
>>   f = 1.33;
>>   if (f != 0x1.547ae2p+0f)
>>     abort ();
>>   fesetround (FE_DOWNWARD);
>>   f = 1.333;
>>   if (f != 0x1.553f7cp+0f)
>>     abort ();
>>   fesetround (FE_TOWARDZERO);
>>   f = 1.3333;
>>   if (f != 0x1.555326p+0f)
>>     abort ();
>> #endif
>>   return 0;
>> }
>
>	Jakub
>