[Bug target/95471] New: vrndvq_f32 not supported on armv8

[Bug target/95471] New: vrndvq_f32 not supported on armv8

[evan@coeus-group.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95471

            Bug ID: 95471
           Summary: vrndvq_f32 not supported on armv8
           Product: gcc
           Version: 9.2.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: evan@coeus-group.com
  Target Milestone: ---

With -march=armv8-a -mfpu=neon, arm_neon.h exposes vrndnq_f32, but the
intrinsic used to implement it (__builtin_neon_vrintnv4sf) emits an "error:
this builtin is not supported for this target".

Here is a trivial test case:

  #include <arm_neon.h>

  #if defined(__ARM_NEON)
  float32x4_t foo(float32x4_t a) {
    return vrndnq_f32(a);
  }
  #endif

On Compiler Explorer: https://godbolt.org/z/ThfJQe

Relevant documentation from ARM:
<https://developer.arm.com/architectures/instruction-sets/simd-isas/neon/intrinsics?search=vrndnq_f32>.

[Bug target/95471] vrndvq_f32 not supported on armv8

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95471

ktkachov at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|---                         |INVALID
                 CC|                            |ktkachov at gcc dot gnu.org

--- Comment #1 from ktkachov at gcc dot gnu.org ---
-mfpu=neon doesn't support the Neon features from Armv8-A.
The recommended way is to use -mfpu=auto and use an appropriate -mcpu or -march
option.
In this case, using -march=armv8-a+simd -mfpu=auto generates the instruction.

[Bug target/95471] vrndvq_f32 not supported on armv8

[evan@coeus-group.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95471

--- Comment #2 from Evan Nemerson <evan@coeus-group.com> ---
In that case shouldn't the header be adjusted to not define the vrndvq_f32
function unless it is enabled?

It is already guarded by a check for __ARM_ARCH >= 8 (see
<https://github.com/gcc-mirror/gcc/blob/95827968e5f3e472bbef0ad8b65b0a7537c9cee3/gcc/config/arm/arm_neon.h#L1757>,
but obviously that isn't sufficient.  If nothing else it would help document
the actual requirements, which would be great since the flags and macros on ARM
are a bit of a mess.