llvm
8614a89d - [SYCL] tighten known_identity to finite_math macro (#18247)

Commit
345 days ago
[SYCL] tighten known_identity to finite_math macro (#18247) Some time ago we fixed a bug in `sycl::known_identity` where it was incorrectly using `std::numeric_limits<T>::infinity()` even in cases where `::infinity()` should not have been available. At the time, the fix was to simply test for `__FAST_MATH__`. That worked, but was not ideal. Now we are tightening this to use the `__FINITE_MATH_ONLY__` macro which is more accurate. see https://github.com/intel/llvm/issues/13813 `__FINITE_MATH_ONLY__` will be set when both the `-fno-honor-infinities` and `-fno-honor-nans` flags are used. It was felt by the FE team that there is no reasonable case where one of the flags would be used without the other and therefore `__FINITE_MATH_ONLY__` is sufficient for this problem ( and superior to the `__FAST_MATH__` macro we were using before). --------- Signed-off-by: Chris Perkins <chris.perkins@intel.com>
Author
Parents
Loading