c/4346: wrong float multiply with Sparc

c/4346: wrong float multiply with Sparc

[tassart]

>Number:         4346
>Category:       c
>Synopsis:       wrong float multiply with Sparc
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          wrong-code
>Submitter-Id:   net
>Arrival-Date:   Tue Sep 18 05:56:03 PDT 2001
>Closed-Date:
>Last-Modified:
>Originator:     Stephan Tassart
>Release:        gcc version 2.95.2 19991024 (release)
>Organization:
>Environment:
SunOS 5.7 Generic_106541-16 sun4u sparc SUNW,Ultra-5_10
>Description:
#include <stdio.h>

int main(int argv, char *argc[]) {
  float a1;
  double acc;
  a1 = -1.83351052f;
  acc = a1*a1;
  printf("%.40g\n",acc);
  return 0;
}

/* This code compiles wrong the floating multiplication 
on Sparc. It produces 
	ld	[%fp-20], %f2
	ld	[%fp-20], %f3
	fmuls	%f2, %f3, %f2
	fstod	%f2, %f4

i.e. 
  float32 x float32 -> float32
  cast(float32)     -> float64

Instead of

	ld	[%fp-20], %f2
	ld	[%fp-20], %f3
	fsmuld	%f2, %f3, %f4

i.e.
 float32 x float32 -> float64 


This produces a loss of a precision in the floating
point multiplication. (and because of this loss, same
code compiled on another plateform produces a different
output)
 */


>How-To-Repeat:
compile on sparc the forementionned code with 

gcc -S foo.c -o foo.s 
>Fix:
Dunno
>Release-Note:
>Audit-Trail:
>Unformatted:

Re: c/4346: wrong float multiply with Sparc

[aoliva]

Synopsis: wrong float multiply with Sparc

State-Changed-From-To: open->closed
State-Changed-By: aoliva
State-Changed-When: Fri Sep 28 21:26:28 2001
State-Changed-Why:
    GCC is correct, as far as the code you wrote goes.  There's no reason to do the multiplication that outputs a double; in fact, this could even give incorrect results in case the algorithm depended on the multiplication being performed in float, and only then converted to double, which is what the code asks for.  Perhaps if you convert the operands of the multiplication to double, GCC will do the multiplication the way you expect.

http://gcc.gnu.org/cgi-bin/gnatsweb.pl?cmd=view&pr=4346&database=gcc