[PATCH] D50685: [AArch64] Support conversion between fp16 and fp128

[Sjoerd Meijer via Phabricator via llvm-commits]

SjoerdMeijer added inline comments.


================
Comment at: compiler-rt/lib/builtins/fp_extend.h:43
 #elif defined SRC_HALF
+#if defined __aarch64__ && defined __ARM_FP16_ARGS
+// Use __fp16 if available.
----------------
efriedma wrote:
> Can we implement this some other way?  _Float16 should have the right calling convention on all targets, I think.
That was exactly the first thing I was thinking of.  But then I also got confused about e.g. changing this:

  float __aeabi_h2f(uint16_t a)

to:  

  float __aeabi_h2f(src a)

where src can now be __fp16. First I was wondering if changing the signature would be problematic somehow. Probably not, and probably it  is more correct. Looking at the RT ABI for h2f, the descriptions reads:

    IEEE 754 binary16 storage format (VFP half precision) to binary32 (float)  conversion

>From a first glance, I could image that __fp16 would capture this behaviour, but I would need to check again what the binary16 storage format exactly is, and that's why I wasn't sure about all this.



Repository:
  rL LLVM

https://reviews.llvm.org/D50685