[PATCH] D94557: [ARM] Fixed incorrect lowering when using GNUEABI (libgcc) and 16bit floats

[Fernando De La Garza via Phabricator via llvm-commits]

fdelagarza added a comment.

Here's some observations that might help figure out what path to take

According to the discussion on this ticket <https://reviews.llvm.org/D12413> , there are only four arm `__aeabi` prefixed functions in compiter-rt that aren't present in libgcc:

  __aeabi_memcpy
  __aeabi_memmove
  __aeabi_memset
  __aeabi_d2h

`memcpy`, `memmove` and `memset` default to being lowered to the non `__aebi` prefixed functions in `include/llvm/IR/RuntimeLibcalls.def`:

  HANDLE_LIBCALL(MEMCPY, "memcpy")
  HANDLE_LIBCALL(MEMMOVE, "memmove")
  HANDLE_LIBCALL(MEMSET, "memset")

A code path in `ARMISelLowering.cpp` switches them to `__aeabi_memcpy`, `__aeabi_memmove` and `__aeabi_memset` if the "EABIVersion" option is set to `EABI4` or `EABI5`.

Also in `include/llvm/IR/RuntimeLibcalls.def`, the  `FPROUND_F32_F16` and `FPROUND_F32_F16` library calls default to `__gnu` prefixed functions:

  HANDLE_LIBCALL(FPEXT_F16_F32, "__gnu_h2f_ieee")
  HANDLE_LIBCALL(FPROUND_F32_F16, "__gnu_f2h_ieee")
  HANDLE_LIBCALL(FPROUND_F64_F16, "__truncdfhf2")

Code in `ARMISelLowering.cpp`  is currently unconditionally switching the `FPROUND_F64_F16` default lowering to `__aeabi_d2h`. Kevin's change would make it default to `__gnu_d2h_ieee`

There is a code path in ` ARMISelLowering.cpp` that changes  `FPROUND_F32_F16`, `FPROUND_F64_F16`, and `FPEXT_F16_F32` to the `__eabui` prefixed functions when `EABI4` or `EABI5` is requested (mentioned in Kevin's last comment)

To me it seems the proposed change would make the default `FPROUND_F64_F16` lowering more consistent with that of  `FPEXT_F16_F32` and `FPROUND_F32_F16`


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D94557/new/

https://reviews.llvm.org/D94557