On Mon, Jan 23, 2017 at 11:23:48AM +0000, James Greenhalgh wrote:
>
> Hi,
>
> As subject, we have an oversight in aarch64_simd_container_mode for
> HFmode inputs. This results in trunk only autovectorizing to a 64-bit vector,
> rather than a full 128-bit vector.
>
> The fix is obvious, we just need to handle HFmode, and return an
> appropriate vector mode.
>
> Tested on aarch64-none-elf with no issues. This patch looks low risk
> for this development stage to me, though it fixes an oversight rather
> than a regression.
*Ping*
Thanks,
James
> gcc/
>
> 2017-01-23 James Greenhalgh <james.greenha...@arm.com>
>
> * config/aarch64/aarch64.c (aarch64_simd_container_mode): Handle
> HFmode.
>
> gcc/testsuite/
>
> 2017-01-23 James Greenhalgh <james.greenha...@arm.com>
>
> * gcc.target/aarch64/vect_fp16_1.c: New.
>
> diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
> index 0cf7d12..7efc1f2 100644
> --- a/gcc/config/aarch64/aarch64.c
> +++ b/gcc/config/aarch64/aarch64.c
> @@ -10777,6 +10777,8 @@ aarch64_simd_container_mode (machine_mode mode,
> unsigned width)
> return V2DFmode;
> case SFmode:
> return V4SFmode;
> + case HFmode:
> + return V8HFmode;
> case SImode:
> return V4SImode;
> case HImode:
> @@ -10793,6 +10795,8 @@ aarch64_simd_container_mode (machine_mode mode,
> unsigned width)
> {
> case SFmode:
> return V2SFmode;
> + case HFmode:
> + return V4HFmode;
> case SImode:
> return V2SImode;
> case HImode:
> diff --git a/gcc/testsuite/gcc.target/aarch64/vect_fp16_1.c
> b/gcc/testsuite/gcc.target/aarch64/vect_fp16_1.c
> new file mode 100644
> index 0000000..da0cd81
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/vect_fp16_1.c
> @@ -0,0 +1,30 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O3 -fno-vect-cost-model" } */
> +
> +/* Check that we vectorize to a full 128-bit vector for _Float16 and __fp16
> + types. */
> +
> +/* Enable ARMv8.2-A+fp16 so we have access to the vector instructions. */
> +#pragma GCC target ("arch=armv8.2-a+fp16")
> +
> +_Float16
> +sum_Float16 (_Float16 *__restrict__ __attribute__ ((__aligned__ (16))) a,
> + _Float16 *__restrict__ __attribute__ ((__aligned__ (16))) b,
> + _Float16 *__restrict__ __attribute__ ((__aligned__ (16))) c)
> +{
> + for (int i = 0; i < 256; i++)
> + a[i] = b[i] + c[i];
> +}
> +
> +_Float16
> +sum_fp16 (__fp16 *__restrict__ __attribute__ ((__aligned__ (16))) a,
> + __fp16 *__restrict__ __attribute__ ((__aligned__ (16))) b,
> + __fp16 *__restrict__ __attribute__ ((__aligned__ (16))) c)
> +{
> + for (int i = 0; i < 256; i++)
> + a[i] = b[i] + c[i];
> +}
> +
> +/* Two FADD operations on "8h" data widths, one from sum_Float16, one from
> + sum_fp16. */
> +/* { dg-final { scan-assembler-times "fadd\tv\[0-9\]\+.8h" 2 } } */
