On Fri, 10 Oct 2014, Jakub Jelinek wrote:
> Hi!
>
> This patch adds a small optimization to emit_case_bit_tests,
> instead of emitting (for high, low, mask all constants)
> (x - low) <= (high - low) && ((1 << (x - low)) & mask)
> if high is smaller than BITS_PER_WORD and low > 0 we can emit
> x <= high && ((1 << x) & (mask << low))
> and avoid subtraction. Do this only if mask << low isn't more costly than
> mask.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
But isn't this a general optimization? Also testing for RTX costs
this early sounds bogus.
Thanks,
Richard.
> 2014-10-10 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/63464
> * tree-switch-conversion.c (struct case_bit_test): Remove
> hi and lo fields, add wide_int mask field.
> (emit_case_bit_tests): Add MAXVAL argument, rewrite uses of
> hi/lo fields into wide_int mask operations, optimize by pretending
> minval to be 0 if maxval is small enough.
> (process_switch): Adjust caller.
>
> --- gcc/tree-switch-conversion.c.jj 2014-08-01 09:24:12.000000000 +0200
> +++ gcc/tree-switch-conversion.c 2014-10-07 12:36:29.438181857 +0200
> @@ -246,8 +246,7 @@ This transformation was contributed by R
>
> struct case_bit_test
> {
> - HOST_WIDE_INT hi;
> - HOST_WIDE_INT lo;
> + wide_int mask;
> edge target_edge;
> tree label;
> int bits;
> @@ -294,13 +293,14 @@ case_bit_test_cmp (const void *p1, const
> are not guaranteed to be of the same type as INDEX_EXPR
> (the gimplifier doesn't change the type of case label values,
> and MINVAL and RANGE are derived from those values).
> + MAXVAL is MINVAL + RANGE.
>
> There *MUST* be MAX_CASE_BIT_TESTS or less unique case
> node targets. */
>
> static void
> emit_case_bit_tests (gimple swtch, tree index_expr,
> - tree minval, tree range)
> + tree minval, tree range, tree maxval)
> {
> struct case_bit_test test[MAX_CASE_BIT_TESTS];
> unsigned int i, j, k;
> @@ -324,6 +324,8 @@ emit_case_bit_tests (gimple swtch, tree
> tree word_type_node = lang_hooks.types.type_for_mode (word_mode, 1);
> tree word_mode_zero = fold_convert (word_type_node, integer_zero_node);
> tree word_mode_one = fold_convert (word_type_node, integer_one_node);
> + int prec = TYPE_PRECISION (word_type_node);
> + wide_int wone = wi::one (prec);
>
> memset (&test, 0, sizeof (test));
>
> @@ -348,8 +350,7 @@ emit_case_bit_tests (gimple swtch, tree
> if (k == count)
> {
> gcc_checking_assert (count < MAX_CASE_BIT_TESTS);
> - test[k].hi = 0;
> - test[k].lo = 0;
> + test[k].mask = wi::zero (prec);
> test[k].target_edge = e;
> test[k].label = label;
> test[k].bits = 1;
> @@ -367,14 +368,39 @@ emit_case_bit_tests (gimple swtch, tree
> CASE_HIGH (cs), minval));
>
> for (j = lo; j <= hi; j++)
> - if (j >= HOST_BITS_PER_WIDE_INT)
> - test[k].hi |= (HOST_WIDE_INT) 1 << (j - HOST_BITS_PER_INT);
> - else
> - test[k].lo |= (HOST_WIDE_INT) 1 << j;
> + test[k].mask |= wi::lshift (wone, j);
> }
>
> qsort (test, count, sizeof (*test), case_bit_test_cmp);
>
> + /* If all values are in the 0 .. BITS_PER_WORD-1 range, we can get rid of
> + the minval subtractions, but it might make the mask constants more
> + expensive. So, compare the costs. */
> + if (compare_tree_int (minval, 0) > 0
> + && compare_tree_int (maxval, GET_MODE_BITSIZE (word_mode)) < 0)
> + {
> + int cost_diff;
> + HOST_WIDE_INT m = tree_to_uhwi (minval);
> + rtx reg = gen_raw_REG (word_mode, 10000);
> + bool speed_p = optimize_bb_for_speed_p (gimple_bb (swtch));
> + cost_diff = set_rtx_cost (gen_rtx_PLUS (word_mode, reg,
> + GEN_INT (-m)), speed_p);
> + for (i = 0; i < count; i++)
> + {
> + rtx r = immed_wide_int_const (test[i].mask, word_mode);
> + cost_diff += set_src_cost (gen_rtx_AND (word_mode, reg, r), speed_p);
> + r = immed_wide_int_const (wi::lshift (test[i].mask, m), word_mode);
> + cost_diff -= set_src_cost (gen_rtx_AND (word_mode, reg, r), speed_p);
> + }
> + if (cost_diff > 0)
> + {
> + for (i = 0; i < count; i++)
> + test[i].mask = wi::lshift (test[i].mask, m);
> + minval = build_zero_cst (TREE_TYPE (minval));
> + range = maxval;
> + }
> + }
> +
> /* We generate two jumps to the default case label.
> Split the default edge, so that we don't have to do any PHI node
> updating. */
> @@ -446,13 +472,7 @@ emit_case_bit_tests (gimple swtch, tree
> if (const & csui) goto target */
> for (k = 0; k < count; k++)
> {
> - HOST_WIDE_INT a[2];
> -
> - a[0] = test[k].lo;
> - a[1] = test[k].hi;
> - tmp = wide_int_to_tree (word_type_node,
> - wide_int::from_array (a, 2,
> - TYPE_PRECISION
> (word_type_node)));
> + tmp = wide_int_to_tree (word_type_node, test[k].mask);
> tmp = fold_build2 (BIT_AND_EXPR, word_type_node, csui, tmp);
> tmp = force_gimple_operand_gsi (&gsi, tmp,
> /*simple=*/true, NULL_TREE,
> @@ -1369,8 +1389,8 @@ process_switch (gimple swtch)
> {
> if (dump_file)
> fputs (" expanding as bit test is preferable\n", dump_file);
> - emit_case_bit_tests (swtch, info.index_expr,
> - info.range_min, info.range_size);
> + emit_case_bit_tests (swtch, info.index_expr, info.range_min,
> + info.range_size, info.range_max);
> loops_state_set (LOOPS_NEED_FIXUP);
> return NULL;
> }
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE / SUSE Labs
SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
GF: Jeff Hawn, Jennifer Guild, Felix Imend"orffer
