On Wed, Nov 14, 2018 at 1:21 PM Jakub Jelinek <ja...@redhat.com> wrote:
>
> On Wed, Nov 14, 2018 at 09:29:15AM +0100, Richard Biener wrote:
> > So where do the zero_extracts come from? Can we somehow avoid
> > zero-sized bit-extracts at expansion time by folding them to zero?
>
> The following patch implements the missing folding.
> Note, as I've mentioned, it is just an optimization, as if we are unlucky
> enough, a constant could be propagated to the insn only at RTL optimization
> time. So the other patch is needed, even when this patch also fixes the
> bmi2-bzhi-2.c testcase execution.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2018-11-14 Jakub Jelinek <ja...@redhat.com>
>
> PR rtl-optimization/87817
> * config/i386/i386.c (ix86_fold_builtin): For _bzhi_u{32,64} if
> last argument has low 8 bits clear, fold to 0.
>
> * gcc.target/i386/bmi2-bzhi-3.c (main): Add a couple of new tests.
OK.
Thanks,
Uros.
> --- gcc/config/i386/i386.c.jj 2018-11-14 00:55:50.199357949 +0100
> +++ gcc/config/i386/i386.c 2018-11-14 10:28:29.492491865 +0100
> @@ -32671,6 +32671,8 @@ ix86_fold_builtin (tree fndecl, int n_ar
> unsigned int idx = tree_to_uhwi (args[1]) & 0xff;
> if (idx >= TYPE_PRECISION (TREE_TYPE (args[0])))
> return args[0];
> + if (idx == 0)
> + return build_int_cst (TREE_TYPE (TREE_TYPE (fndecl)), 0);
> if (!tree_fits_uhwi_p (args[0]))
> break;
> unsigned HOST_WIDE_INT res = tree_to_uhwi (args[0]);
> --- gcc/testsuite/gcc.target/i386/bmi2-bzhi-3.c.jj 2016-10-31
> 13:28:07.961422421 +0100
> +++ gcc/testsuite/gcc.target/i386/bmi2-bzhi-3.c 2018-11-14 10:38:40.352417418
> +0100
> @@ -58,7 +58,11 @@ main ()
> link_error ();
> if (_bzhi_u32 (c, 32) != c
> || _bzhi_u32 (c, 64) != c
> - || _bzhi_u32 (c, 255) != c)
> + || _bzhi_u32 (c, 255) != c
> + || _bzhi_u32 (c, 544) != c
> + || _bzhi_u32 (c, 0) != 0
> + || _bzhi_u32 (c, 256) != 0
> + || _bzhi_u32 (c, 1024) != 0)
> link_error ();
> #ifdef __x86_64__
> if (f21 () != 0 || f22 (-1ULL) != 0
> @@ -70,7 +74,11 @@ main ()
> || f33 () != -1ULL || f34 (-1ULL) != -1ULL)
> link_error ();
> if (_bzhi_u64 (d, 64) != d
> - || _bzhi_u64 (d, 255) != d)
> + || _bzhi_u64 (d, 255) != d
> + || _bzhi_u64 (d, 576) != d
> + || _bzhi_u64 (d, 0) != 0
> + || _bzhi_u64 (d, 256) != 0
> + || _bzhi_u64 (d, 512) != 0)
> link_error ();
> #endif
> return 0;
>
>
> Jakub
