http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59556
Bug ID: 59556
Summary: Floating-point __sec_reduce_add tests not robust in
face of contraction
Product: gcc
Version: 4.9.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: testsuite
Assignee: unassigned at gcc dot gnu.org
Reporter: bhaveet.shah at arm dot com
If fused-mac is available on the target, the following
__sec_reduce_add (taken from
gcc/testsuite/c-c++-common/cilk-plus/AN/builtin_func_double.c) on a
floating-point type could be implemented using a fused-mac:
x = __sec_reduce_add (array3[:] * array4[:]);
The test cases also contain the following loop:
add_value = 0.0000;
mul_value = 1.0000;
for (ii = 0; ii < 10; ii++)
{
add_value += (array3[ii] * array4[ii]);
mul_value *= (array3[ii] * array4[ii]);
}
In this case, a compiler could reasonably do the multiply and addition
separately so that the multiply result could be used for both
add_value and mul_value. If the semantics of the fused-mac
and the separate instructions are not equivalent, the result from this
loop could be different compared to the __sec_reduce_add case.
The test compares the results of these two using a bitwise comparison,
hence this test will fail. This test case should be made more robust to
handle this case.
