x86 Denormal flag

x86 Denormal flag

[Michael Morrell]

Should fetestexcept be able to test for the denormal flag in the MXCSR?

My test case is:

define _GNU_SOURCE
#include <fenv.h>
#include <stdio.h>
#include <xmmintrin.h>

volatile float two = 2.0f;
volatile float den = 0x1.0p-130;

int main()
{
    volatile float r;

    r = den / two;

    if (_mm_getcsr() & _MM_EXCEPT_DENORM)
                printf("mxcsr has DENORM bit set\n");
    if (fetestexcept(__FE_DENORM))
                printf("DENORM is set\n");
    return 0;
}

And it shows that the Denormal flag is set in MXCSR, but fetestexcept returns 0 here.

It seems like fetestexcept only works for the 5 standard IEEE flags, but then why is __FE_DENORM even defined?

   Michael

Re: x86 Denormal flag

[Joseph Myers]

On Sat, 19 Jun 2021, Michael Morrell wrote:

> It seems like fetestexcept only works for the 5 standard IEEE flags, but 
> then why is __FE_DENORM even defined?

It would probably be reasonable to move __FE_DENORM from an installed 
header to an internal one that's only used by the fesetenv and fesetmode 
implementations that need to work on the environment (or the control mode 
parts of the environment) as a whole, including implementation-defined 
parts of the environment.

I think including the "denormal operand" flag in FE_ALL_EXCEPT would be a 
bad idea, since it doesn't indicate any kind of exceptional condition in 
standard terms, and given that it's not in FE_ALL_EXCEPT, treating it as 
an exception in other standard functions would be problematic.

-- 
Joseph S. Myers
joseph@codesourcery.com

RE: x86 Denormal flag

[Michael Morrell]

I guess I'm not convinced that adding the denormal operand flag in FE_ALL_EXCEPT is bad.  From a quick grep, I think this flag is defined in the following architectures: ia64 (FE_UNNORMAL), alpha (FE_DENORMAL), csky (__FE_DENORMAL), and x86 (__FE_DENORM).

AARCH64 also has this flag (the IDC bit in FPSR), but glibc does not define an FE_ constant for it.

For ia64, FE_ALL_EXCEPT includes FE_UNNORMAL, but none of the others seems to include it so that is kind of inconsistent.

   Michael

-----Original Message-----
From: Joseph Myers <joseph@codesourcery.com> 
Sent: Monday, June 21, 2021 1:01 PM
To: Michael Morrell <mmorrell@tachyum.com>
Cc: libc-alpha@sourceware.org
Subject: Re: x86 Denormal flag

On Sat, 19 Jun 2021, Michael Morrell wrote:

> It seems like fetestexcept only works for the 5 standard IEEE flags, 
> but then why is __FE_DENORM even defined?

It would probably be reasonable to move __FE_DENORM from an installed header to an internal one that's only used by the fesetenv and fesetmode implementations that need to work on the environment (or the control mode parts of the environment) as a whole, including implementation-defined parts of the environment.

I think including the "denormal operand" flag in FE_ALL_EXCEPT would be a bad idea, since it doesn't indicate any kind of exceptional condition in standard terms, and given that it's not in FE_ALL_EXCEPT, treating it as an exception in other standard functions would be problematic.

--
Joseph S. Myers
joseph@codesourcery.com