2.088 build fails on i386 due to test failures

[sebastic]

The Debian package build on i386 for PDL 2.088 failed due to test failures:

    #   Failed test 'hist works'
    #   at t/primitive-misc.t line 32.
    #          got: '[3 5 1 2 2 1 2 1 0 4]'
    #     expected: '[3 5 1 2 1 2 2 1 0 4]'
    # Looks like you failed 1 test of 1.
#   Failed test 'hist'
#   at t/primitive-misc.t line 33.
# Looks like you failed 1 test of 7.
[...]
Test Summary Report
-------------------
t/pdl_from_string.t      (Wstat: 0 Tests: 144 Failed: 0)
  TODO passed:   60-62
t/primitive-misc.t       (Wstat: 256 (exited 1) Tests: 7 Failed: 1)
  Failed test:  2
  Non-zero exit status: 1
t/primitive-random.t     (Wstat: 0 Tests: 3 Failed: 0)
  TODO passed:   1
Files=45, Tests=2267, 72 wallclock secs ( 0.40 usr  0.12 sys + 65.80 cusr  8.38 csys = 74.70 CPU)
Result: FAIL
Failed 1/45 test programs. 1/2267 subtests failed.

Full buildlog: https://salsa.debian.org/perl-team/modules/packages/pdl/-/jobs/5625325

Same failure on the buildds: i386, hurd-i386

The other 32bit architectures like armel & armhf are OK, unlike last time (#469).

[mohawk2]

This has boiled down to this:

$ make core && perl -Mblib -MPDL -e 'print pdl( 0.5000 )->histogram(0.1, 0, 10), "\n"'
# on i686:
# [0 0 0 0 1 0 0 0 0 0]
# on x64:
# [0 0 0 0 0 1 0 0 0 0]

This is because histogram does this:

    PDL_Indx j = ($in()-min)/step;

0.5 / 0.1 when truncated to an integer on x64 was producing a 5. For some reason, quite possibly a bug, on i686, it was producing a 4. This is despite trying to capture the double value and printf it with %1.70g, it still only showed a 5 (I was expecting a 4.999999999999999 etc).

[mohawk2]

@sebastic The above linked commit makes it work on my 32-bit box. Are you amenable to applying the above commit as a patch on 2.088?

[sebastic]

Patch added in commit 9bb7f548.

The package build succeeded locally in an i386 chroot, in Salsa CI, and on the build daemons.

[mohawk2]

Thank you for the report, as always!

[sisyphus]

Just for fun, one way I could get 0.5 / 0.1 to evaluate (in C) to less than 5 is if 0.1 is single precision:

#include <stdio.h>
int main(void) {
float f1 = 0.5, f2 = 0.1;
double x;
x = (double)f1 / f2;
printf("%.17g %a\n", x, x);
return 0;
}

/* Outputs: 4.9999999254941949 0x1.3fffffb000001p+2 */

(Confirmed using MPFR library.)
Using 64-bit gcc-13.2.0, if that cast to double was removed, then the output changed to 5 0x1.4p+2, but using 32-bit gcc-13.2.0, it mattered not whether that cast was specified.