On Thu, Aug 29, 2024 at 3:37 AM Andrew Pinski <quic_apin...@quicinc.com> wrote:
>
> After adding popcount{qi,hi}2 to the aarch64 backend, I noticed that
> the expansion for popcount==1 was no longer trying to do the trick
> of handling popcount==1 as `(arg ^ (arg - 1)) > arg - 1`. The problem
> is the expansion was using OPTAB_DIRECT, when using OPTAB_WIDEN
> will allow modes which are smaller than SImode (in the aarch64 case).
>
> Note QImode's cost still needs some improvements so part of popcnt-eq-1.c
> is xfailed. Though there is a check to make sure the costs are compared now.
>
> Built and tested on aarch64-linux-gnu.
OK.
Richard.
> PR middle-end/116508
>
> gcc/ChangeLog:
>
> * internal-fn.cc (expand_POPCOUNT): Use OPTAB_WIDEN for PLUS and
> XOR/AND expansion.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/aarch64/popcnt-eq-1.c: New test.
>
> Signed-off-by: Andrew Pinski <quic_apin...@quicinc.com>
> ---
> gcc/internal-fn.cc | 4 +-
> .../gcc.target/aarch64/popcnt-eq-1.c | 45 +++++++++++++++++++
> 2 files changed, 47 insertions(+), 2 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/aarch64/popcnt-eq-1.c
>
> diff --git a/gcc/internal-fn.cc b/gcc/internal-fn.cc
> index 78997ef056a..4e33db365ac 100644
> --- a/gcc/internal-fn.cc
> +++ b/gcc/internal-fn.cc
> @@ -5332,11 +5332,11 @@ expand_POPCOUNT (internal_fn fn, gcall *stmt)
> start_sequence ();
> rtx op0 = expand_normal (arg);
> rtx argm1 = expand_simple_binop (mode, PLUS, op0, constm1_rtx, NULL_RTX,
> - 1, OPTAB_DIRECT);
> + 1, OPTAB_WIDEN);
> if (argm1 == NULL_RTX)
> goto fail;
> rtx argxorargm1 = expand_simple_binop (mode, nonzero_arg ? AND : XOR, op0,
> - argm1, NULL_RTX, 1, OPTAB_DIRECT);
> + argm1, NULL_RTX, 1, OPTAB_WIDEN);
> if (argxorargm1 == NULL_RTX)
> goto fail;
> rtx cmp;
> diff --git a/gcc/testsuite/gcc.target/aarch64/popcnt-eq-1.c
> b/gcc/testsuite/gcc.target/aarch64/popcnt-eq-1.c
> new file mode 100644
> index 00000000000..bb9e2bf0a54
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/popcnt-eq-1.c
> @@ -0,0 +1,45 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-rtl-expand-details" } */
> +/* { dg-final { check-function-bodies "**" "" } } */
> +/* PR middle-end/116508 */
> +
> +#pragma GCC target "+nocssc"
> +
> +/*
> +** h16:
> +** sub w([0-9]+), w0, #1
> +** eor w([0-9]+), w0, w\1
> +** and w([0-9]+), w\1, 65535
> +** cmp w\3, w\2, uxth
> +** cset w0, cc
> +** ret
> +*/
> +
> +/* when expanding popcount == 1, should use
> + `(arg ^ (arg - 1)) > arg - 1` as that has a lower latency
> + than doing the popcount then comparing against 1.
> + The popcount/addv can be costly. */
> +unsigned h16 (const unsigned short a) {
> + return __builtin_popcountg (a) == 1;
> +}
> +
> +/* unsigned char should also do the same trick */
> +/* Currently xfailed since the cost does not take into account the
> + moving between gprs and vector regs correctly. */
> +/*
> +** h8: { xfail *-*-* }
> +** sub w([0-9]+), w0, #1
> +** eor w([0-9]+), w0, w\1
> +** and w([0-9]+), w\1, 255
> +** cmp w\3, w\2, uxtb
> +** cset w0, cc
> +** ret
> +*/
> +
> +
> +unsigned h8 (const unsigned char a) {
> + return __builtin_popcountg (a) == 1;
> +}
> +
> +/* There should be printing out the costs for h8 and h16's popcount == 1 */
> +/* { dg-final { scan-rtl-dump-times "popcount == 1:" 2 "expand"} } */
> --
> 2.43.0
>
