Re: [Patch-86512]: Subnormal float support in armv7(with -msoft-float) for intrinsics

[Umesh Kalappa <umesh dot kalappa0 at gmail dot com>]

Hi Richard,

We tested on the SP and yes the problem persist on the SP too and
attached patch will fix the both SP and DP issues for the  denormal
resultant.
We bootstrapped the compiler ,look ok to us with minimal testing ,

Any floating point test-suite to test for the attached patch ? any
recommendations or inputs  ?

Thank you again
~Umesh


On Mon, Jul 23, 2018 at 3:28 PM, Richard Earnshaw (lists)
<Richard.Earnshaw@arm.com> wrote:
> So why is this only changing the double-precision implementation?
> Surely, a problem like this will normally be common to both SP and DP
> floating-point computations.
>
> R.
>
> On 23/07/18 08:46, Umesh Kalappa wrote:
>> Thank you Wilco for the inputs and we agree that the  fix break down
>> for the case.
>>
>> Meanwhile ,attached patch will take care the inputs and we are testing
>> the patch vigorously ,would you recommended any test-suite out there
>> for the same ?
>>
>> Thank you
>> ~Umesh
>>
>>
>> On Fri, Jul 20, 2018 at 10:04 PM, Wilco Dijkstra <Wilco.Dijkstra@arm.com> wrote:
>>> Umesh Kalappa wrote:
>>>
>>>> We tried some of the normalisation numbers and the fix works and please
>>>> could you help us with the input ,where  if you see that fix breaks down.
>>>
>>> Well try any set of inputs which require normalisation. You'll find these no
>>> longer get normalised and so will get incorrect results. Try basic cases like
>>> 1.0 - 0.75 which I think will return 0.625...
>>>
>>> A basic test would be to run old vs new on a large set of inputs to verify
>>> there aren't any obvious differences.
>>>
>>> Wilco
>>>
>>>
>>> pr86512.patch
>>>
>>>
>>> Index: libgcc/config/arm/ieee754-df.S
>>> ===================================================================
>>> --- libgcc/config/arm/ieee754-df.S   (revision 262850)
>>> +++ libgcc/config/arm/ieee754-df.S   (working copy)
>>> @@ -203,6 +203,7 @@
>>>  #endif
>>>
>>>      @ Determine how to normalize the result.
>>> +    @ if result is denormal i.e (exp)=0,then don't normalise the result,
>>>  LSYM(Lad_p):
>>>      cmp     xh, #0x00100000
>>>      bcc     LSYM(Lad_a)
>>> @@ -235,6 +236,8 @@
>>>
>>>      @ Result must be shifted left and exponent adjusted.
>>>  LSYM(Lad_a):
>>> +    cmp     r4,#0x0
>>> +    beq     LSYM(Lad_e)
>>>      movs    ip, ip, lsl #1
>>>      adcs    xl, xl, xl
>>>      adc     xh, xh, xh
>>> Index: gcc/testsuite/gcc.target/arm/pr86512.c
>>> ===================================================================
>>> --- gcc/testsuite/gcc.target/arm/pr86512.c   (nonexistent)
>>> +++ gcc/testsuite/gcc.target/arm/pr86512.c   (working copy)
>>> @@ -0,0 +1,28 @@
>>> +/* { dg-do run } */
>>> +/* { dg-options "-O0 -msoft-float" } */
>>> +
>>> +#include<stdlib.h>
>>> +#include<stdint.h>
>>> +
>>> +typedef union
>>> +{
>>> +    double d;
>>> +    uint64_t i;
>>> +} u;
>>> +
>>> +int main()
>>> +{
>>> +  u smallestPositiveNormal, smallesPositiveSubnormal, expectedResult, result;
>>> +
>>> +  smallesPositiveSubnormal.i = 1;
>>> +
>>> +  smallestPositiveNormal.i = 0x0010000000000000;
>>> +  expectedResult.i = 0x000fffffffffffff;
>>> +  result.d = smallestPositiveNormal.d - smallesPositiveSubnormal.d;
>>> +
>>> +  if (result.i != expectedResult.i)
>>> +    abort();
>>> +
>>> +  return 0;
>>> +}
>>> +
>

Attachment: pr86512.patch
Description: Binary data