[Bug c/111506] New: RISC-V: Failed to vectorize conversion from INT64 -> _Float16

[juzhe.zhong at rivai dot ai via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111506

            Bug ID: 111506
           Summary: RISC-V: Failed to vectorize conversion from INT64 ->
                    _Float16
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: juzhe.zhong at rivai dot ai
  Target Milestone: ---

#include <stdint.h>

void foo (int64_t *__restrict a, _Float16 * b, int n)
{
    for (int i = 0; i < n; i++) {
        b[i] = (_Float16)a[i];
    }
}

-march=rv64gcv_zvfh_zfh -O3  -fno-trapping-math -fopt-info-vec-missed
-march=rv64gcv_zvfh_zfh -O3  -ffast-math -fopt-info-vec-missed

<source>:5:23: missed: couldn't vectorize loop
<source>:6:27: missed: not vectorized: no vectype for stmt: _4 = *_3;
 scalar_type: int64_t
Compiler returned: 0

https://godbolt.org/z/orevoq7E1

Consider LLVM can vectorize with -fno-trapping-math.
However, LLVM can not vectorize when -ftrapping-math.

So, we need an explicit patterns from INT64 -> _Float16 with
!flag_trapping_math