[src/trunk]: src/tests/lib/libm on i386 with FLT_EVAL_METHOD=2, the computati...

[maya <maya%NetBSD.org@localhost>]

details:   https://anonhg.NetBSD.org/src/rev/23508a6dbd16
branches:  trunk
changeset: 451550:23508a6dbd16
user:      maya <maya%NetBSD.org@localhost>
date:      Mon May 27 00:10:36 2019 +0000

description:
on i386 with FLT_EVAL_METHOD=2, the computation inside cosf was done in
binary80 arithmetic and never rounded to binary32 at the end, so it
appeared farther from the correctly rounded result in binary32 than expected.

Force binary32 using volatile float.

PR lib/54232: lib/libm/t_cos:cosf_angles regression on i386

>From Riastradh.

diffstat:

 tests/lib/libm/t_cos.c |  17 ++++++++++++++---
 1 files changed, 14 insertions(+), 3 deletions(-)

diffs (36 lines):

diff -r 967801ffd1e2 -r 23508a6dbd16 tests/lib/libm/t_cos.c
--- a/tests/lib/libm/t_cos.c    Sun May 26 23:03:31 2019 +0000
+++ b/tests/lib/libm/t_cos.c    Mon May 27 00:10:36 2019 +0000
@@ -1,4 +1,4 @@
-/* $NetBSD: t_cos.c,v 1.8 2019/04/25 22:58:23 maya Exp $ */
+/* $NetBSD: t_cos.c,v 1.9 2019/05/27 00:10:36 maya Exp $ */
 
 /*-
  * Copyright (c) 2011 The NetBSD Foundation, Inc.
@@ -275,13 +275,24 @@
                float theta = angles[i].x;
                float cos_theta = angles[i].fy;
 
+               /*
+                * Force rounding to float even if FLT_EVAL_METHOD=2,
+                * as is the case on i386.
+                *
+                * The volatile should not be necessary, by C99 Sec.
+                * 5.2.4.2.2. para. 8 on p. 24 which specifies that
+                * assignment and cast remove all extra range and precision,
+                * but seems to be needed to work around a compiler bug.
+                */ 
+               volatile float result = cosf(theta);
+
                if (cos_theta == 999)
                        cos_theta = angles[i].y;
 
                assert(cos_theta != 0);
-               if (!(fabsf((cosf(theta) - cos_theta)/cos_theta) <= eps)) {
+               if (!(fabsf((result - cos_theta)/cos_theta) <= eps)) {
                        atf_tc_fail_nonfatal("cosf(%d deg = %.8g) = %.8g"
-                           " != %.8g", deg, theta, cos(theta), cos_theta);
+                           " != %.8g", deg, theta, result, cos_theta);
                }
        }
 }