[Bug target/113484] Add support for _Float16 type on PowerPC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113484 Andrew Pinski changed: What|Removed |Added Severity|normal |enhancement
[Bug target/113484] Add support for _Float16 type on PowerPC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113484 --- Comment #1 from Joseph S. Myers --- It would of course be necessary to define the ABI used for _Float16 (and _Complex _Float16) argument passing and return (in each PowerPC ABI for which we support use of this feature).
[Bug target/113484] Add support for _Float16 type on PowerPC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113484 --- Comment #2 from John Platts --- (In reply to Joseph S. Myers from comment #1) > It would of course be necessary to define the ABI used for _Float16 (and > _Complex _Float16) argument passing and return (in each PowerPC ABI for > which we support use of this feature). Should scalar _Float16 and _Complex _Float16 function parameters be first passed in the F1-F13 registers on PPC, following the same rules as float, double, _Complex _Float, and _Complex double function parameters (and with the _Float16 function parameters that are passed into F1-F13 be converted to double-precision floating point)? Should a scalar _Float16 result be returned in the F1 register on PPC, with the _Float16 result in the F1 register in double-precision floating point format (but rounded to the precision and range of a _Float16)?
[Bug target/113484] Add support for _Float16 type on PowerPC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113484 Segher Boessenkool changed: What|Removed |Added Status|UNCONFIRMED |RESOLVED Resolution|--- |INVALID --- Comment #3 from Segher Boessenkool --- Yup, it needs to be added to the ABI first. And that is unlikely to get any priority soon, since bf16 is pretty useless anyway, and we have no instructions that efficiently round to it. There are instructions to convert from it (by simply adding a bunch of zero bits to the right) (xvcvbf16spn), and more proper conversion to it (from SP) (xvcvspbf16). Those are ISA 3.1 insns btw, so POWER10, not POWER9! What instructions did you mean? This isn't good enough to use it as an actual computational type. It is fine for what it is meant for: do all of your computations as SP float, and only load and store as bfloat16. We do have builtin functions to do that already.
[Bug target/113484] Add support for _Float16 type on PowerPC
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113484 --- Comment #4 from Segher Boessenkool --- Ah, this was about *actual* half-precision float, which indeed is 3.0 (Power9). But all the same holds: it needs to be added to the ABI before we can have a type for it, and it still won't be terribly useful (it will make for *slowe code*, not using fewer resources either).
