apinski--- via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> From: Andrew Pinski <apin...@marvell.com>
>
> This is a simple patch which simplifies the __builtin_aarch64_sqrt* builtins
> into the internal function SQRT which allows for constant folding and other
> optimizations at the gimple level. It was originally suggested we do to
> __builtin_sqrt just for __builtin_aarch64_sqrtdf when -fno-math-errno
> but since r6-4969-g686ee9719a4 we have the internal function SQRT which does
> the same so it makes we don't need to check -fno-math-errno either now.
>
> OK? Bootstrapped and tested on aarch64-linux-gnu with no regressions.
>
> PR target/64821
>
> gcc/ChangeLog:
>
> * config/aarch64/aarch64-builtins.c
> (aarch64_general_gimple_fold_builtin): Handle
> __builtin_aarch64_sqrt* and simplify into SQRT internal
> function.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/aarch64/vsqrt-1.c: New test.
> * gcc.target/aarch64/vsqrt-2.c: New test.
> ---
> gcc/config/aarch64/aarch64-builtins.c | 7 ++++++
> gcc/testsuite/gcc.target/aarch64/vsqrt-1.c | 17 +++++++++++++
> gcc/testsuite/gcc.target/aarch64/vsqrt-2.c | 28 ++++++++++++++++++++++
> 3 files changed, 52 insertions(+)
> create mode 100644 gcc/testsuite/gcc.target/aarch64/vsqrt-1.c
> create mode 100644 gcc/testsuite/gcc.target/aarch64/vsqrt-2.c
>
> diff --git a/gcc/config/aarch64/aarch64-builtins.c
> b/gcc/config/aarch64/aarch64-builtins.c
> index 58bcbd9875f..1bf487477eb 100644
> --- a/gcc/config/aarch64/aarch64-builtins.c
> +++ b/gcc/config/aarch64/aarch64-builtins.c
> @@ -2820,6 +2820,13 @@ aarch64_general_gimple_fold_builtin (unsigned int
> fcode, gcall *stmt,
> gimple_call_set_lhs (new_stmt, gimple_call_lhs (stmt));
> break;
>
> + /* Lower sqrt builtins to gimple/internal function sqrt. */
> + BUILTIN_VHSDF_DF (UNOP, sqrt, 2, FP)
> + new_stmt = gimple_build_call_internal (IFN_SQRT,
> + 1, args[0]);
Sorry for the nit-pick, but: IMO it looks odd to split this over two lines.
> + gimple_call_set_lhs (new_stmt, gimple_call_lhs (stmt));
> + break;
> +
> /*lower store and load neon builtins to gimple. */
> BUILTIN_VALL_F16 (LOAD1, ld1, 0, LOAD)
> BUILTIN_VDQ_I (LOAD1_U, ld1, 0, LOAD)
> diff --git a/gcc/testsuite/gcc.target/aarch64/vsqrt-1.c
> b/gcc/testsuite/gcc.target/aarch64/vsqrt-1.c
> new file mode 100644
> index 00000000000..3207c8774ca
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/vsqrt-1.c
> @@ -0,0 +1,17 @@
> +/* PR target/64821 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-tree-optimized" } */
> +/* Check that we constant fold sqrt(4.0) into 2.0. */
> +/* { dg-final { scan-tree-dump-times ".SQRT" 0 "optimized" } } */
> +/* { dg-final { scan-tree-dump-times "2.0" 1 "optimized" } } */
I think these would be better as scan-tree-dump-not and scan-tree-dump
respectively. The number of 2.0s isn't important and 2.0 has the risk
of matching someone's directory name. Probably worth backslash-quoting
the "."s too.
OK with those changes, thanks.
Richard
> +/* { dg-final { scan-assembler-times "fsqrt" 0 } } */
> +/* We should produce a fmov to d0 with 2.0 but currently don't, see PR
> 103959. */
> +/* { dg-final { scan-assembler-times "\n\tfmov\td0, 2.0e.0" 1 { xfail *-*-*
> } } } */
> +
> +#include <arm_neon.h>
> +
> +float64x1_t f64(void)
> +{
> + float64x1_t a = (float64x1_t){4.0};
> + return vsqrt_f64 (a);
> +}
> diff --git a/gcc/testsuite/gcc.target/aarch64/vsqrt-2.c
> b/gcc/testsuite/gcc.target/aarch64/vsqrt-2.c
> new file mode 100644
> index 00000000000..7465b79f3a7
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/vsqrt-2.c
> @@ -0,0 +1,28 @@
> +/* PR target/64821 */
> +/* { dg-do compile } */
> +/* { dg-options "-fdump-tree-optimized" } */
> +#include <arm_neon.h>
> +
> +/* Check that we lower __builtin_aarch64_sqrt* into the internal function
> SQRT. */
> +/* { dg-final { scan-tree-dump-times "__builtin_aarch64" 0 "optimized" } } */
> +/* { dg-final { scan-tree-dump-times ".SQRT" 4 "optimized" } } */
> +
> +float64x1_t f64(float64x1_t a)
> +{
> + return vsqrt_f64 (a);
> +}
> +
> +float64x2_t f64q(float64x2_t a)
> +{
> + return vsqrtq_f64 (a);
> +}
> +
> +float32x2_t f32(float32x2_t a)
> +{
> + return vsqrt_f32 (a);
> +}
> +
> +float32x4_t f32q(float32x4_t a)
> +{
> + return vsqrtq_f32 (a);
> +}
