-
Type:
Bug
-
Resolution: Fixed
-
Priority:
Medium
-
Code Generation Tools
-
CODEGEN-13914
-
-
-
default
-
Reduce the optimization level to -o1 or below. Alternatively, affected code can be rewritten to only use __vpred types and not boolean vector types.
An assignment to a boolean vector element may in some cases not be applied before conversion to a __vpred. An example of affected code is below:
int32_t test_function(void *dst1)
{
__vpred *pvPred = (__vpred *)dst1;
bool_vec vpred_bool = bool_vec(0);
// These writes
for (uint32_t loc = 0; loc < __C7X_VEC_SIZE_BYTES__/4; loc++)
{
vpred_bool.s[loc*4] = true;
vpred_bool.s[loc*4 + 1] = true;
vpred_bool.s[loc*4 + 2] = true;
vpred_bool.s[loc*4 + 3] = true;
}
// Are not correctly applied to *pvPred
*pvPred = __create_vpred(vpred_bool);
uint64_t vpred_as_uint64 = 0;
memcpy(&vpred_as_uint64, dst1, sizeof(__vpred) < sizeof(uint64_t) ? sizeof(__vpred) : sizeof(uint64_t));
// Expect to see something non-zero like "18446744073709551615", but "0" is seen instead
printf("pvPred: %llu\n", vpred_as_uint64);
return 0;
}
