another IEE related problem

another IEE related problem

[Ulrich Drepper]

Hi,

Compiling this little piece of code on a ix86 machine

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
extern inline float
bar (float x)
{
  float res;
  asm ("fsqrt" : "=t" (res) : "0" (x));
  return res;
}
extern float fabsf (float);

float
foo (float x)
{
  float res = fabsf (bar (x));
  return res;
}
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

The generated code with -O3 -momit-leaf-frame-pointer

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
foo:
        flds 4(%esp)
#APP
        fsqrt
#NO_APP
        ret
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

This is wrong.  In the real code I added the `fabs' in full knowledge
of the situation.  The problem is that according to IEEE

	sqrt(-0.0) == -0.0

and so I get a wrong result.  I don't know why gcc thinks the
fabs call is redundant (it should not be able to analyze the asm
statement) but in any case it is plainly wrong.

Another bug preventing the glibc test suite to be passed.

-- Uli
---------------.      drepper at gnu.org  ,-.   Rubensstrasse 5
Ulrich Drepper  \    ,-------------------'   \  76149 Karlsruhe/Germany
Cygnus Solutions `--' drepper at cygnus.com   `------------------------

Re: another IEE related problem

[Bernd Schmidt]

> Compiling this little piece of code on a ix86 machine
> 
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> extern inline float
> bar (float x)
> {
>   float res;
>   asm ("fsqrt" : "=t" (res) : "0" (x));
>   return res;
> }
> extern float fabsf (float);
> 
> float
> foo (float x)
> {
>   float res = fabsf (bar (x));
>   return res;
> }
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> 
> The generated code with -O3 -momit-leaf-frame-pointer
> 
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> foo:
>         flds 4(%esp)
> #APP
>         fsqrt
> #NO_APP
>         ret
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
[the fabs statement is missing]

This is a bug in the combiner. When trying to simplify an
(ABS:SF (ASM_OPERANDS ...)) expression it thinks that the ASM_OPERANDS is
always positive.
This can be fixed with the patch I have appended.

> Another bug preventing the glibc test suite to be passed.

Ulrich, do you have more unsolved bugs that affect glibc?

Bernd

*** ./combine.c.orig-1	Tue Dec  9 10:53:39 1997
--- ./combine.c	Tue Dec  9 10:55:52 1997
*************** simplify_rtx (x, op0_mode, last, in_dest
*** 4092,4101 ****
      case XOR:
        return simplify_logical (x, last);
  
!     case ABS:
        /* (abs (neg <foo>)) -> (abs <foo>) */
        if (GET_CODE (XEXP (x, 0)) == NEG)
  	SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
  
        /* If operand is something known to be positive, ignore the ABS.  */
        if (GET_CODE (XEXP (x, 0)) == FFS || GET_CODE (XEXP (x, 0)) == ABS
--- 4092,4106 ----
      case XOR:
        return simplify_logical (x, last);
  
!     case ABS:      
        /* (abs (neg <foo>)) -> (abs <foo>) */
        if (GET_CODE (XEXP (x, 0)) == NEG)
  	SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
+ 
+       /* If the mode of the operand is VOIDmode (i.e. if it is ASM_OPERANDS),
+          do nothing.  */
+       if (GET_MODE (XEXP (x, 0)) == VOIDmode)
+ 	break;
  
        /* If operand is something known to be positive, ignore the ABS.  */
        if (GET_CODE (XEXP (x, 0)) == FFS || GET_CODE (XEXP (x, 0)) == ABS

Re: another IEE related problem

[Ulrich Drepper]

Bernd Schmidt <crux@starsky.Informatik.RWTH-Aachen.DE> writes:

> This is a bug in the combiner. When trying to simplify an
> (ABS:SF (ASM_OPERANDS ...)) expression it thinks that the ASM_OPERANDS is
> always positive.
> This can be fixed with the patch I have appended.

Thanks a lot, I'll try the patch.

> Ulrich, do you have more unsolved bugs that affect glibc?

Yes, the current egcs snapshot fails to compile the dynamic linker.
So far I'm not able to reproduce it in less than 100kB source code.
I'm working on this.

Thanks,

-- Uli
---------------.      drepper at gnu.org  ,-.   Rubensstrasse 5
Ulrich Drepper  \    ,-------------------'   \  76149 Karlsruhe/Germany
Cygnus Solutions `--' drepper at cygnus.com   `------------------------

Re: another IEE related problem

[Ulrich Drepper]

Bernd Schmidt <crux@starsky.Informatik.RWTH-Aachen.DE> writes:

> This can be fixed with the patch I have appended.

Yes, it fixes at least my problems.  Jeff, could you add the patch?

-- Uli
---------------.      drepper at gnu.org  ,-.   Rubensstrasse 5
Ulrich Drepper  \    ,-------------------'   \  76149 Karlsruhe/Germany
Cygnus Solutions `--' drepper at cygnus.com   `------------------------

Re: another IEE related problem

[Ulrich Drepper]

Jeffrey A Law <law@hurl.cygnus.com> writes:

> Ulrich can you try this patch instead?
> 
> 	* combine.c (simplify_rtx, case ABS): Do nothing for ASM_OPERANDS.

This patch works as well.

-- Uli
---------------.      drepper at gnu.org  ,-.   Rubensstrasse 5
Ulrich Drepper  \    ,-------------------'   \  76149 Karlsruhe/Germany
Cygnus Solutions `--' drepper at cygnus.com   `------------------------

Re: another IEE related problem

[Jeffrey A Law]

  In message < Pine.SOL.3.90.971210122825.25453C-100000@starsky.informatik.rwth-aachen.de >you write:
It seems to me that this code would be cleaner if you checked directly for
ASM_OPERANDS instead of indirectly via a VOIDmode operand.

I also think your change would prevent optimizing away an abs if you
had something like this:

(set (reg x) (const_int 5))
(set (reg y) (abs (reg x))

Ulrich can you try this patch instead?

	* combine.c (simplify_rtx, case ABS): Do nothing for ASM_OPERANDS.

Index: combine.c
===================================================================
RCS file: /cvs/cvsfiles/egcs/gcc/combine.c,v
retrieving revision 1.9
diff -c -3 -p -r1.9 combine.c
*** combine.c	1997/12/07 00:28:07	1.9
--- combine.c	1997/12/11 01:09:22
*************** simplify_rtx (x, op0_mode, last, in_dest
*** 4097,4102 ****
--- 4097,4105 ----
        if (GET_CODE (XEXP (x, 0)) == NEG)
  	SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0));
  
+       if (GET_CODE (XEXP (x, 0)) == ASM_OPERANDS)
+ 	return x;
+ 
        /* If operand is something known to be positive, ignore the ABS.  */
        if (GET_CODE (XEXP (x, 0)) == FFS || GET_CODE (XEXP (x, 0)) == ABS
  	  || ((GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0)))