Re: [Patch] Optimize condition reductions where the result is an integer induction variable

[Richard Biener <richard dot guenther at gmail dot com>]

On Wed, Nov 11, 2015 at 1:22 PM, Alan Hayward <alan.hayward@arm.com> wrote:
> Hi,
> I hoped to post this in time for Mondayâs cut off date, but circumstances
> delayed me until today. Hoping if possible this patch will still be able
> to go in.
>
>
> This patch builds upon the change for PR65947, and reduces the amount of
> code produced in a vectorized condition reduction where operand 2 of the
> COND_EXPR is an assignment of a increasing integer induction variable that
> won't wrap.
>
>
> For example (assuming all types are ints), this is a match:
>
> last = 5;
> for (i = 0; i < N; i++)
>   if (a[i] < min_v)
>     last = i;
>
> Whereas, this is not because the result is based off a memory access:
> last = 5;
> for (i = 0; i < N; i++)
>   if (a[i] < min_v)
>     last = a[i];
>
> In the integer induction variable case we can just use a MAX reduction and
> skip all the code I added in my vectorized condition reduction patch - the
> additional induction variables in vectorizable_reduction () and the
> additional checks in vect_create_epilog_for_reduction (). From the patch
> diff only, it's not immediately obvious that those parts will be skipped
> as there is no code changes in those areas.
>
> The initial value of the induction variable is force set to zero, as any
> other value could effect the result of the induction. At the end of the
> loop, if the result is zero, then we restore the original initial value.

+static bool
+is_integer_induction (gimple *stmt, struct loop *loop)

is_nonwrapping_integer_induction?

+  tree lhs_max = TYPE_MAX_VALUE (TREE_TYPE (gimple_phi_result (stmt)));

don't use TYPE_MAX_VALUE.

+  /* Check that the induction increments.  */
+  if (tree_int_cst_compare (step, size_zero_node) <= 0)
+    return false;

tree_int_cst_sgn (step) == -1

+  /* Check that the max size of the loop will not wrap.  */
+
+  if (! max_loop_iterations (loop, &ni))
+    return false;
+  /* Convert backedges to iterations.  */
+  ni += 1;

just use max_stmt_executions (loop, &ni) which properly checks for overflow
of the +1.

+  max_loop_value = wi::add (wi::to_widest (base),
+                           wi::mul (wi::to_widest (step), ni));
+
+  if (wi::gtu_p (max_loop_value, wi::to_widest (lhs_max)))
+    return false;

you miss a check for the wi::add / wi::mul to overflow.  You can use
extra args to determine this.

Instead of TYPE_MAX_VALUE use wi::max_value (precision, sign).

I wonder if you want to skip all the overflow checks for TYPE_OVERFLOW_UNDEFINED
IV types?

Thanks,
Richard.

>
>
>
> Cheers,
> Alan.
>