https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119014
Bug ID: 119014
Summary: Extending _Float16 constant at compile and run time
differs
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: stefansf at gcc dot gnu.org
Target Milestone: ---
__attribute__ ((noipa)) _Float16
test (void)
{
return 42.42f16;
}
int
main (void)
{
_Float16 x = test ();
if (x != 42.42f16)
__builtin_abort ();
return 0;
}
Naively I would expect that the condition evaluates to true. However, on
x86_64 with -mno-avx512fp16 this evaluates to false. It seems to me that the
front end chose to statically extend 42.42f16 to a float which differs from
extending the result of test() at runtime. Optimized tree dump:
__attribute__((noipa, noinline, noclone, no_icf))
_Float16 test ()
{
<bb 2> [local count: 1073741824]:
return 4.240625e+1;
}
int main ()
{
_Float16 x;
float _1;
<bb 2> [local count: 1073741824]:
x_4 = test ();
_1 = (float) x_4;
if (_1 != 4.24199981689453125e+1)
goto <bb 3>; [0.00%]
else
goto <bb 4>; [100.00%]
<bb 3> [count: 0]:
__builtin_abort ();
<bb 4> [local count: 1073741824]:
return 0;
}
The float cast results in a call to __extendhfsf2 where the result is unequal
to 4.24199981689453125e+1.
Changing the initial example slightly by not comparing directly with a constant
but assigning it first to variable y:
__attribute__ ((noipa)) _Float16
test (void)
{
return 42.42f16;
}
int
main (void)
{
_Float16 x = test ();
_Float16 y = 42.42f16;
if (x != y)
__builtin_abort ();
return 0;
}
gives us the expected result:
__attribute__((noipa, noinline, noclone, no_icf))
_Float16 test ()
{
<bb 2> [local count: 1073741824]:
return 4.240625e+1;
}
int main ()
{
_Float16 x;
<bb 2> [local count: 1073741824]:
x_3 = test ();
if (x_3 != 4.240625e+1)
goto <bb 3>; [0.00%]
else
goto <bb 4>; [100.00%]
<bb 3> [count: 0]:
__builtin_abort ();
<bb 4> [local count: 1073741824]:
return 0;
}
I'm not sure whether this is "expected" and is covered by some C/IEEE-754
rules. Maybe someone can shed some light on it?
