On 11/22/21 12:16, Richard Henderson wrote:
On 11/22/21 10:43 AM, Richard Henderson wrote:
On 11/21/21 6:47 PM, Cédric Le Goater wrote:
I am getting an error with this test. See below.
...
### Rounding to nearest
from single: f32(-nan:0xffa00000)
- to double: f64(-nan:0x00fff4000000000000) (INVALID)
+ to double: f64(-nan:0x00fff4000000000000) (OK)
Well that's disconcerting.
I can replicate this failure on an x86_64 host, but do not see the same error
on a power9 ppc64le host.
Bah. The test case is buggy.
It reads the fpscr for the flags *after* having gone through the printf for the
result, at which point you are at the mercy of whatever other fp arithmetic
libc chooses to do.
Fixed with
--- a/tests/tcg/multiarch/float_convs.c
+++ b/tests/tcg/multiarch/float_convs.c
@@ -51,8 +51,8 @@ static void convert_single_to_double(float input)
output = input;
- out_fmt = fmt_f64(output);
flag_fmt = fmt_flags();
+ out_fmt = fmt_f64(output);
printf(" to double: %s (%s)\n", out_fmt, flag_fmt);
free(out_fmt);
free(flag_fmt);
But this alone of course causes other "failures", because we've got some
incorrect reference files.
The only one I have seen so far is on hexagon:
https://gitlab.com/legoater/qemu/-/jobs/1829273672
C.
