https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107967
--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Ok, I can reproduce this, disabling all the +-*/ handlers fixes it.
It isn libm.so.6 that matters for the failures, not the tests themselves.
So far I've looked at the expm1 stuff, the failures are:
Failure: expm1 (0x1.86ap+16): Exception "Overflow" not set
Failure: expm1 (0x2.c5c4p+12): Exception "Overflow" not set
Failure: expm1 (0xf.ffffffffffff8p+1020): Exception "Overflow" not set
Failure: expm1 (0xf.fffffp+124): Exception "Overflow" not set
Failure: expm1_downward (0x1.86ap+16): Exception "Overflow" not set
Failure: Test: expm1_downward (0x1.86ap+16)
Result:
is: inf inf
should be: 1.7976931348623157e+308 0x1.fffffffffffffp+1023
Failure: expm1_downward (0x2.c5c4p+12): Exception "Overflow" not set
Failure: Test: expm1_downward (0x2.c5c4p+12)
Result:
is: inf inf
should be: 1.7976931348623157e+308 0x1.fffffffffffffp+1023
...
Failure: expm1_upward (0x1.86ap+16): Exception "Overflow" not set
Failure: expm1_upward (0x2.c5c4p+12): Exception "Overflow" not set
Failure: expm1_upward (0xf.ffffffffffff8p+1020): Exception "Overflow" not set
Failure: expm1_upward (0xf.fffffp+124): Exception "Overflow" not set
For all the arguments >= 710.0 or so (non-inf/nan), the path in the source is:
static const double
huge = 1.0e+300,
tiny = 1.0e-300,
o_threshold = 7.09782712893383973096e+02; /* 0x40862E42, 0xFEFA39EF */
...
if (x > o_threshold)
{
__set_errno (ERANGE);
return huge * huge; /* overflow */
}
and the file is compiled with -frounding-math.
So the reduced testcase for at least part of this PR is:
double
foo (void)
{
const double huge = 1.0e+300;
return huge * huge;
}
GCC 12 would compile this into return __builtin_inf (); only with
-fno-trapping-math,
not without it nor with -frounding-math.
Now, GCC trunk compiles this into return __builtin_inf (); with all of
-fno-trapping-math, default or -frounding-math. For the default case, the
problem is the same as in
PR107608. But with -frounding-math, we have an extra problem with the value,
which
actually shouldn't be +INF but DBL_MAX.
E.g.
#include <fenv.h>
#include <stdio.h>
int
main ()
{
volatile double huge = 1.0e+308;
volatile double inf = __builtin_inf ();
fesetround (FE_DOWNWARD);
volatile double r1 = huge + huge;
volatile double r2 = huge * huge;
volatile double r3 = huge + inf;
volatile double r4 = r2 + huge;
volatile double r5 = inf - 1.0;
volatile double r6 = inf - huge;
fesetround (FE_TONEAREST);
printf ("%e %e %e %e %e %e\n", r1, r2, r3, r4, r5, r6);
}
prints 1.797693e+308 1.797693e+308 inf 1.797693e+308 inf inf, so the behavior
seems to be if either operand is already inf, then the result should be inf
even when rounding to -inf (except special cases when it is nan), but if
neither operand is inf, when rounding downward it shouldn't be +inf but max
representable (or when rounding upward not -inf but min representable).
So I assume we should tweak frange_arithmetics for this behavior when
flag_rounding_math.
