The root cause for these test suite failures seems to be the different
precision of the C language type 'long double' on ppc architectures.
They are using 106-bit precision compared to the x86 precision of 80-bit.
For example the following program gives different results:
<<<< C program <<<<
#include <stdio.h>
int main() {
printf("IEC 559 support: ");
#ifdef __STDC_IEC_559__
printf("yes\n");
#else
printf("no\n");
#endif
printf("sizeof(long double) = %zu\n", sizeof(long double));
long double a = 0.0005L;
long double b = 0.0004L;
printf("0.0005 - 0.0004 < 0.0001 = %s\n", a - b < 0.0001L ? "true"
: "false");
printf("(0.0005 - 0.0004) = %.50Lf hex: ", a);
for (unsigned j = 0; j < sizeof(long double); j++) {
printf("%2x ", ((unsigned char *)&a)[j]);
}
printf("\n");
return 0;
}
>>>> C program >>>>
<<<< amd64 output <<<<
IEC 559 support: yes
sizeof(long double) = 16
0.0005 - 0.0004 < 0.0001 = true
(0.0005 - 0.0004) =
0.00049999999999999999997924769279226964169282609873 hex: 3b df 4f 8d
97 6e 12 83 f4 3f 40 0 0 0 0 0
>>>> amd64 output >>>>
<<<< ppc64el output <<<<
IEC 559 support: yes
sizeof(long double) = 16
0.0005 - 0.0004 < 0.0001 = false
(0.0005 - 0.0004) =
0.00050000000000000000000000000000000279259841936149 hex: fc a9 f1 d2
4d 62 40 3f f8 7e 6a bc 74 93 c8 bb
>>>> ppc64el output >>>>
The test suite of check depends on the precision of long double, e.g.
https://sources.debian.org/src/check/0.12.0-0.2/tests/check_check_sub.c/#L783
.
I think there is no problem with check handling the type 'long double'
in general, so the test suite failures should not lead to regressions
in caller code.
