Bug 1864107

Created attachment 1705778 [details]
build.log

file build.log too big, will only attach last 32768 bytes

Created attachment 1705779 [details]
root.log

file root.log too big, will only attach last 32768 bytes

Created attachment 1705780 [details]
state.log

make check fails on ppc64le:

../build-aux/test-driver: line 107: 3320412 Aborted                 (core dumped) "$@" > $log_file 2>&1
FAIL: test-float

test-float.c:318: assertion 'm + m > m' failed

Program received signal SIGABRT, Aborted.
0x00007ffff7d88f04 in raise () from /lib64/libc.so.6
(gdb) bt
#0  0x00007ffff7d88f04 in raise () from /lib64/libc.so.6
#1  0x00007ffff7d69868 in abort () from /lib64/libc.so.6
#2  0x0000000100000e20 in test_float () at test-float.c:165
#3  main () at test-float.c:359

Workarounded by disabling %check on ppc64le for now.

Reopening. This is a bug in gnulib's detection of a working float.h. You're going to need to update this.

When run on POWER9 hardware I see the following:

cat test-float.log
test-float.c:318: assertion 'm + m > m' failed
FAIL test-float (exit status: 134)

Adding instrumentation I see the following:

LDBL_MAX = inf
m = inf
test-float.c:320: assertion 'm + m > m' failed
Aborted (core dumped)

This can't be right and looks like a compiler issue.

We should not have ended up with LDBL_MAX being equal to INF to start with.

Unfortunately I can't reduce this, the appropriate extracted code works as intended.

Leading up to the assert:

=> 0x0000000100000c8c <+876>:	addis   r9,r2,-2
   0x0000000100000c90 <+880>:	addi    r9,r9,-25152
   0x0000000100000c94 <+884>:	lfd     f0,0(r9)
   0x0000000100000c98 <+888>:	lfd     f1,8(r9)
   0x0000000100000c9c <+892>:	stfd    f0,304(r1)
   0x0000000100000ca0 <+896>:	stfd    f1,312(r1)
   0x0000000100000ca8 <+904>:	lfd     f1,304(r1)
   0x0000000100000cac <+908>:	lfd     f2,312(r1)
   0x0000000100000cb0 <+912>:	lfd     f3,304(r1)
   0x0000000100000cb4 <+916>:	lfd     f4,312(r1)

Parameters should be f1-f4.

$f1 == inf
$f2 == 0

$f3 == inf
$f4 == 0

So we are about to do "m + m" and the value of m is already wrong.

   0x0000000100000cb8 <+920>:	bl      0x100001758 <__gcc_qadd+8>

Do the add.

   0x0000000100000cbc <+924>:	nop
   0x0000000100000cc0 <+928>:	lfd     f0,304(r1)

Reload half of m.

   0x0000000100000cc4 <+932>:	fmr     f12,f1
   0x0000000100000cc8 <+936>:	fmr     f13,f2

Move result from f1/f2 to f12/f13.

   0x0000000100000ccc <+940>:	lfd     f1,312(r1)

Reload other half of m.

=> 0x0000000100000cd0 <+944>:	fcmpu   cr0,f12,f0

Compare INF to INF and the assert (INF + INF > INF) fails.

What is odd is that 304/312 + r1 is stored to by this earlier sequence (you see it in the original disassembly):

=> 0x0000000100000c8c <+876>:	addis   r9,r2,-2
   0x0000000100000c90 <+880>:	addi    r9,r9,-25152
   0x0000000100000c94 <+884>:	lfd     f0,0(r9)
   0x0000000100000c98 <+888>:	lfd     f1,8(r9)

Address 0+$r9 is 0x100001cc0 and it's here:

100000000-100010000 r-xp 00000000 fd:00 1774810                          /root/rpmbuild/BUILD/m4-1.4.18/tests/test-float
100010000-100020000 r--p 00000000 fd:00 1774810                          /root/rpmbuild/BUILD/m4-1.4.18/tests/test-float
100020000-100030000 rw-p 00010000 fd:00 1774810                          /root/rpmbuild/BUILD/m4-1.4.18/tests/test-float

That value is in the executable image, probably a constant pool.

It's odd that we'd load INF from a constant pool that should contain LDBL_MAX?

The pre-processed source is more interesting:

  {
    volatile long double m =
# 315 "test-float.c" 3
                            (gl_LDBL_MAX.ld)
# 315 "test-float.c"
                                    ;
    int n;

    do { if (!(m + m > m)) { fprintf (
# 318 "test-float.c" 3 4
   stderr
# 318 "test-float.c"
   , "%s:%d: assertion '%s' failed\n", "test-float.c", 318, "m + m > m");
# 318 "test-float.c" 3
   rpl_fflush
# 318 "test-float.c"
   (
# 318 "test-float.c" 3 4
   stderr
# 318 "test-float.c"
   ); abort (); } } while (0);

It looks like we're triggering the generation and inclusion of lib/float.h, and that doesn't work.

149 union gl_long_double_union
150   {
151     struct { double hi; double lo; } dd;
152     long double ld;
153   };
154 extern const union gl_long_double_union gl_LDBL_MAX;
155 # define LDBL_MAX (gl_LDBL_MAX.ld)

 24 const union gl_long_double_union gl_LDBL_MAX =
 25   { { DBL_MAX, DBL_MAX / (double)134217728UL / (double)134217728UL } };

Eventually this loaded value is invalid.

This is either a compiler problem or a problem in the gnulib float.h headers.

Created attachment 1711030 [details]
test-float.i

Attaching pre-processed test-float.i

Removing float.h from inclusion reveals the next problem.

test-float.c:324: assertion 'x + x == x' failed
Aborted (core dumped)

This is what I was expecting given my review of the code.

#include <stdio.h>
#include <assert.h>
#include <float.h>
#include <math.h>

int
main (void)
{
  int n = 107;
  volatile long double m = LDBL_MAX;
  volatile long double pow2_n = powl (2, n);
  volatile long double x = m + (m / pow2_n);

  printf ("n = %d\n", n);
  printf ("m = %Lf (%La)\n", m, m);
  printf ("pow2_n = %Lf (%La)\n", pow2_n, pow2_n);
  printf ("m / pow2_n = %Lf (%La)\n", (m / pow2_n), (m / pow2_n));
  printf ("x = %Lf (%La)\n", x, x);

  if (x > m)
    assert (x + x == x);
  return 0;
}

gcc -o ~/test-ldbl-max ~/test-ldbl-max.c -lm

~/test-ldbl-max
n = 107
m = 179769313486231580793728971405301199252069012264752390332004544495176179865349768338004270583473493681874097135387894924752516923758125018237039690323659469736010689648748751591634331824498526377862231967249520608291850653495428451067676993116107021027413767397958053860876625383538022115414866471826801819648.000000 (0x1.fffffffffffff7ffffffffffff8p+1023)
pow2_n = 162259276829213363391578010288128.000000 (0x1p+107)
m / pow2_n = 1107913932560222581216724223049124694376931327937918798971295069363205703164244740389102844506567402654244799528342026118673562844811584683014545030137100678976901567468093855075985516353544747282849589098225960074532039651619564827101237983225846137075291097947344654582153216.000000 (0x1.fffffffffffff7ffffffffffff8p+916)
x = 179769313486231580793728971405301199252069012264752390332004544495176179865349768338004270583473493681874097135387894924752516923758125018237039690323659469736010689648748751591634331824498526377862231967249520608291850653495428451067676993116107021027413767397958053860876625383538022115414866471826801819648.000000 (0x1.fffffffffffff7ffffffffffffcp+1023)
test-ldbl-max: /root/test-ldbl-max.c:21: main: Assertion `x + x == x' failed.
Aborted (core dumped)

There is a representable value that is in theory larger than LDBL_MAX and so we assert.

Note that x > m, because 0x1.fffffffffffff7ffffffffffffcp+1023 > 0x1.fffffffffffff7ffffffffffff8p+1023, but x + x most certainly INF not x.

Is this a problem with __LDBL_MAX__ as defined by the compiler?

Marek, What do you make of the test case in comment #9?

This bug appears to have been reported against 'rawhide' during the Fedora 33 development cycle.
Changing version to 33.

(In reply to Carlos O'Donell from comment #10)
> Marek, What do you make of the test case in comment #9?

Looks like there indeed is a bug in GCC: https://gcc.gnu.org/PR95450.  It hasn't been fixed yet.  I'll try to bisect it.

Thank you very much for your investigation of the issue. I'll remove the workaround when the bug is fixed in GCC.

The bug in GCC has been fixed, workaround is no longer needed.