The compiler is allowed to use intermediate higher precision for float

arithmetic, which may cause differences smaller than float precision
but still much larger than eps = 1e-30.
Forcing intermediate results to volatile variables removes the excess
precision.
This commit is contained in:
mlelstv 2011-11-19 12:46:41 +00:00
parent 8d134405ac
commit ea28fb54a5
1 changed files with 4 additions and 4 deletions

View File

@ -1,4 +1,4 @@
/* $NetBSD: t_sqrt.c,v 1.1 2011/10/16 08:25:40 jruoho Exp $ */
/* $NetBSD: t_sqrt.c,v 1.2 2011/11/19 12:46:41 mlelstv Exp $ */
/*-
* Copyright (c) 2011 The NetBSD Foundation, Inc.
@ -29,7 +29,7 @@
* POSSIBILITY OF SUCH DAMAGE.
*/
#include <sys/cdefs.h>
__RCSID("$NetBSD: t_sqrt.c,v 1.1 2011/10/16 08:25:40 jruoho Exp $");
__RCSID("$NetBSD: t_sqrt.c,v 1.2 2011/11/19 12:46:41 mlelstv Exp $");
#include <atf-c.h>
#include <math.h>
@ -177,7 +177,7 @@ ATF_TC_BODY(sqrtf_powf, tc)
#ifndef __vax__
const float x[] = { 0.0, 0.005, 1.0, 99.0, 123.123, 9999.9999 };
const float eps = 1.0e-30;
float y, z;
volatile float y, z;
size_t i;
for (i = 0; i < __arraycount(x); i++) {
@ -187,7 +187,7 @@ ATF_TC_BODY(sqrtf_powf, tc)
if (fabsf(y - z) > eps)
atf_tc_fail_nonfatal("sqrtf(%0.03f) != "
"powf(%0.03f, 1/3)\n", x[i], x[i]);
"powf(%0.03f, 1/2)\n", x[i], x[i]);
}
#endif
}