[PATCH] D113107: Support of expression granularity for _Float16.
John McCall via Phabricator via cfe-commits
cfe-commits at lists.llvm.org
Wed Jun 22 12:44:45 PDT 2022
rjmccall added a comment.
In D113107#3602478 <https://reviews.llvm.org/D113107#3602478>, @zahiraam wrote:
> Do we want to add an -fexcess-precision option?
Well, generally we try to offer the same options GCC does. However, it looks like GCC's `-fexcess-precision` does not line up with the options we're thinking of offering: GCC offers only `-fexcess-precision=standard` (essentially what you're implementing) and `-fexcess-precision=fast` (which I think would require some sort of optimizer annotation to permit folding truncations across statement boundaries), with no option to get operation-by-operation fidelity (our current behavior). And I think we're very unlikely to default to `-fexcess-precision=fast` the way that GCC apparently does. I'm also not sure if anyone cares enough to enable this for anything other than `_Float16`; on GCC, `-fexcess-precision` is much broader, e.g. affecting `float` and `double` math on non-SSE x86 targets. So I think we should hold off on adding this option until we know what behavior we actually want to expose via it.
CHANGES SINCE LAST ACTION
More information about the cfe-commits