On Fri, Oct 16, 2015 at 5:25 PM, Alan Lawrence <[email protected]> wrote:
> This lets the vectorizer handle some simple strides expressed using left-shift
> rather than mul, e.g. a[i << 1] (whereas previously only a[i * 2] would have
> been handled).
>
> This patch does *not* handle the general case of shifts - neither a[i << j]
> nor a[1 << i] will be handled; that would be a significantly bigger patch
> (probably duplicating or generalizing much of chrec_fold_multiply and
> chrec_fold_multiply_poly_poly in tree-chrec.c), and would probably also only
> be applicable to machines with gather-load support.
>
> Bootstrapped+check-gcc,g++,gfortran on x86_64, AArch64 and ARM, also Ada on
> x86_64.
>
> Is this OK for trunk?
>
> gcc/ChangeLog:
>
> PR tree-optimization/65963
> * tree-scalar-evolution.c (interpret_rhs_expr): Handle some
> LSHIFT_EXPRs
> as equivalent MULT_EXPRs.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.dg/vect/vect-strided-shift-1.c: New.
> ---
> gcc/testsuite/gcc.dg/vect/vect-strided-shift-1.c | 33
> ++++++++++++++++++++++++
> gcc/tree-scalar-evolution.c | 18 +++++++++++++
> 2 files changed, 51 insertions(+)
> create mode 100644 gcc/testsuite/gcc.dg/vect/vect-strided-shift-1.c
>
> diff --git a/gcc/testsuite/gcc.dg/vect/vect-strided-shift-1.c
> b/gcc/testsuite/gcc.dg/vect/vect-strided-shift-1.c
> new file mode 100644
> index 0000000..b1ce2ec
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/vect-strided-shift-1.c
> @@ -0,0 +1,33 @@
> +/* PR tree-optimization/65963. */
> +#include "tree-vect.h"
> +
> +#define N 512
> +
> +int in[2*N], out[N];
> +
> +__attribute__ ((noinline)) void
> +loop (void)
> +{
> + for (int i = 0; i < N; i++)
> + out[i] = in[i << 1] + 7;
> +}
> +
> +int
> +main (int argc, char **argv)
> +{
> + check_vect ();
> + for (int i = 0; i < 2*N; i++)
> + {
> + in[i] = i;
> + __asm__ volatile ("" : : : "memory");
> + }
> + loop ();
> + __asm__ volatile ("" : : : "memory");
> + for (int i = 0; i < N; i++)
> + {
> + if (out[i] != i*2 + 7)
> + abort ();
> + }
> + return 0;
> +}
> +/* { dg-final { scan-tree-dump-times "vectorized 1 loops in function" 1
> "vect" { target { vect_strided2 } } } } */
> diff --git a/gcc/tree-scalar-evolution.c b/gcc/tree-scalar-evolution.c
> index 0753bf3..e478b0e 100644
> --- a/gcc/tree-scalar-evolution.c
> +++ b/gcc/tree-scalar-evolution.c
> @@ -1831,12 +1831,30 @@ interpret_rhs_expr (struct loop *loop, gimple
> *at_stmt,
> break;
>
> case MULT_EXPR:
> + case LSHIFT_EXPR:
> + /* Handle A<<B as A * (1<<B). */
> chrec1 = analyze_scalar_evolution (loop, rhs1);
> chrec2 = analyze_scalar_evolution (loop, rhs2);
> chrec1 = chrec_convert (type, chrec1, at_stmt);
> chrec2 = chrec_convert (type, chrec2, at_stmt);
> chrec1 = instantiate_parameters (loop, chrec1);
> chrec2 = instantiate_parameters (loop, chrec2);
> + if (code == LSHIFT_EXPR)
> + {
> + /* Do the shift in the larger size, as in e.g. (long) << (int)32,
> + we must do 1<<32 as a long or we'd overflow. */
Err, you should always do the shift in the type of rhs1. You should also
avoid the chrec_convert of rhs2 above for shifts. I think globbing
shifts and multiplies together doesn't make the code any clearer.
Richard.
> + tree type = TREE_TYPE (chrec2);
> + if (TYPE_PRECISION (TREE_TYPE (chrec1)) > TYPE_PRECISION (type))
> + type = TREE_TYPE (chrec1);
> + if (TYPE_PRECISION (type) == 0)
> + {
> + res = chrec_dont_know;
> + break;
> + }
> + chrec2 = fold_build2 (LSHIFT_EXPR, type,
> + build_int_cst (type, 1),
> + chrec2);
> + }
> res = chrec_fold_multiply (type, chrec1, chrec2);
> break;
>
> --
> 1.9.1
>
