[PATCH] D107082: [X86][RFC] Enable `_Float16` type support on X86 following the psABI

[Jean Perier via Phabricator via llvm-commits]

jeanPerier added a comment.

Hi @pengfei, I am working on flang, and after this patch, we started to see some bugs in Fortran programs using REAL(2) (which is fp16 in flang). I am not an expert in LLVM codegen and the builtins, but I am wondering if there is not issue with how llvm codegen thinks `__truncsfhf2` returns its value and how the runtime actually does return it.

Here is an llvm IR reproducer for a bug we saw:

  define void @bug(ptr %addr, i32 %i) {
    %1 = sitofp i32 %i to half
    store half %1, ptr %addr, align 2
    ret void
  }

After this patch the generated assembly on X86 is:

  bug:                                    # @bug
          push    rbx
          mov     rbx, rdi
          cvtsi2ss        xmm0, esi
          call    __truncsfhf2 at PLT
          pextrw  eax, xmm0, 0
          mov     word ptr [rbx], ax
          pop     rbx
          ret

When running this from a C program to test integers are casted to floats, I am only seeing the bytes of the passed address being set to zero (regardless of the input). It seems to me that there is an issue around the `__truncsfhf2` interface. The `pextrw  eax, xmm0, 0` after the call seems to suggest LLVM codegen is looking for the result in xmm0 register, but it seems that `__truncsfhf2` is only returning it in eax.

Do you have any idea what could be the issue ?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D107082/new/

https://reviews.llvm.org/D107082