Hi!
I've missed
+FAIL: libgomp.c/loop-21.c execution test
during testing of the recent patch. The problem is that while
for the number of iterations computation it doesn't matter if we compute
min_inner_iterations as (m2 * first + n2 + (adjusted step) + m1 * first + n1) /
step
or (m2 * last + n2 + (adjusted step) + m1 * last + n1) / step provided that
in the second case we use as factor (m1 - m2) * ostep / step rather than
(m2 - m1) * ostep / step, for the logical to actual iterator values computation
it does matter and in my hand written C implementations of all the cases (outer
vs. inner loop with increasing vs. decreasing iterator) I'm using the same
computation
and it worked well for all the pseudo-random iterators testing it was doing.
It also means min_inner_iterations is misnamed, because it is not really
minimum number of inner iterations, whether the first or last outer iteration
results in the smaller or larger value of this can be (sometimes) only
determined at runtime.
So this patch also renames it to first_inner_iterations.
Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk.
2020-07-15 Jakub Jelinek <ja...@redhat.com>
PR libgomp/96198
* omp-general.h (struct omp_for_data): Rename min_inner_iterations
member to first_inner_iterations, adjust comment.
* omp-general.c (omp_extract_for_data): Adjust for the above change.
Always use n1first and n2first to compute it, rather than depending
on single_nonrect_cond_code. Similarly, always compute factor
as (m2 - m1) * outer_step / inner_step rather than sometimes m1 - m2
depending on single_nonrect_cond_code.
* omp-expand.c (expand_omp_for_init_vars): Rename min_inner_iterations
to first_inner_iterations and min_inner_iterationsd to
first_inner_iterationsd.
--- gcc/omp-general.h.jj 2020-07-14 12:20:01.520110629 +0200
+++ gcc/omp-general.h 2020-07-15 12:54:02.660265973 +0200
@@ -80,10 +80,9 @@ struct omp_for_data
struct omp_for_data_loop *loops;
/* The following are relevant only for non-rectangular loops
where only a single loop depends on an outer loop iterator. */
- tree min_inner_iterations; /* Number of iterations of the inner
- loop with either the first or last
- outer iterator, depending on which
- results in fewer iterations. */
+ tree first_inner_iterations; /* Number of iterations of the inner
+ loop with the first outer iterator
+ (or adjn1, if that is non-NULL). */
tree factor; /* (m2 - m1) * outer_step / inner_step. */
/* Adjusted n1 of the outer loop in such loop nests (if needed). */
tree adjn1;
--- gcc/omp-general.c.jj 2020-07-14 12:20:01.520110629 +0200
+++ gcc/omp-general.c 2020-07-15 12:54:38.793740074 +0200
@@ -212,7 +212,7 @@ omp_extract_for_data (gomp_for *for_stmt
fd->sched_modifiers = 0;
fd->chunk_size = NULL_TREE;
fd->simd_schedule = false;
- fd->min_inner_iterations = NULL_TREE;
+ fd->first_inner_iterations = NULL_TREE;
fd->factor = NULL_TREE;
fd->adjn1 = NULL_TREE;
collapse_iter = NULL;
@@ -726,16 +726,8 @@ omp_extract_for_data (gomp_for *for_stmt
if (loop->m1 || loop->m2)
{
gcc_assert (single_nonrect != -1);
- if (single_nonrect_cond_code == LT_EXPR)
- {
- n1 = n1first;
- n2 = n2first;
- }
- else
- {
- n1 = n1last;
- n2 = n2last;
- }
+ n1 = n1first;
+ n2 = n2first;
}
t = fold_build2 (PLUS_EXPR, itype, t, fold_convert (itype, n2));
t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, n1));
@@ -754,8 +746,6 @@ omp_extract_for_data (gomp_for *for_stmt
or last value of the outer iterator (the one with fewer
iterations).
Compute t2 = ((m2 - m1) * ostep) / step
- (for single_nonrect_cond_code GT_EXPR
- t2 = ((m1 - m2) * ostep) / step instead)
and niters = outer_count * t
+ t2 * ((outer_count - 1) * outer_count / 2)
*/
@@ -763,11 +753,7 @@ omp_extract_for_data (gomp_for *for_stmt
tree m2 = loop->m2 ? loop->m2 : integer_zero_node;
m1 = fold_convert (itype, m1);
m2 = fold_convert (itype, m2);
- tree t2;
- if (single_nonrect_cond_code == LT_EXPR)
- t2 = fold_build2 (MINUS_EXPR, itype, m2, m1);
- else
- t2 = fold_build2 (MINUS_EXPR, itype, m1, m2);
+ tree t2 = fold_build2 (MINUS_EXPR, itype, m2, m1);
t2 = fold_build2 (MULT_EXPR, itype, t2, ostep);
if (TYPE_UNSIGNED (itype) && loop->cond_code == GT_EXPR)
t2 = fold_build2 (TRUNC_DIV_EXPR, itype,
@@ -776,7 +762,7 @@ omp_extract_for_data (gomp_for *for_stmt
else
t2 = fold_build2 (TRUNC_DIV_EXPR, itype, t2, step);
t2 = fold_convert (llutype, t2);
- fd->min_inner_iterations = t;
+ fd->first_inner_iterations = t;
fd->factor = t2;
t = fold_build2 (MULT_EXPR, llutype, t,
single_nonrect_count);
@@ -834,11 +820,11 @@ omp_extract_for_data (gomp_for *for_stmt
if (count)
{
*collapse_count = fold_convert_loc (loc, iter_type, count);
- if (fd->min_inner_iterations && fd->factor)
+ if (fd->first_inner_iterations && fd->factor)
{
t = make_tree_vec (4);
TREE_VEC_ELT (t, 0) = *collapse_count;
- TREE_VEC_ELT (t, 1) = fd->min_inner_iterations;
+ TREE_VEC_ELT (t, 1) = fd->first_inner_iterations;
TREE_VEC_ELT (t, 2) = fd->factor;
TREE_VEC_ELT (t, 3) = fd->adjn1;
*collapse_count = t;
@@ -856,7 +842,7 @@ omp_extract_for_data (gomp_for *for_stmt
if (TREE_CODE (fd->loop.n2) == TREE_VEC)
{
gcc_assert (fd->non_rect);
- fd->min_inner_iterations = TREE_VEC_ELT (fd->loop.n2, 1);
+ fd->first_inner_iterations = TREE_VEC_ELT (fd->loop.n2, 1);
fd->factor = TREE_VEC_ELT (fd->loop.n2, 2);
fd->adjn1 = TREE_VEC_ELT (fd->loop.n2, 3);
fd->loop.n2 = TREE_VEC_ELT (fd->loop.n2, 0);
--- gcc/omp-expand.c.jj 2020-07-14 12:20:01.505110847 +0200
+++ gcc/omp-expand.c 2020-07-15 12:55:29.994994878 +0200
@@ -2264,7 +2264,7 @@ expand_omp_for_init_vars (struct omp_for
{
tree outer_n1 = fd->adjn1 ? fd->adjn1 : fd->loops[i - 1].n1;
tree itype = TREE_TYPE (fd->loops[i].v);
- tree min_inner_iterations = fd->min_inner_iterations;
+ tree first_inner_iterations = fd->first_inner_iterations;
tree factor = fd->factor;
gcond *cond_stmt
= gimple_build_cond (NE_EXPR, factor,
@@ -2282,21 +2282,21 @@ expand_omp_for_init_vars (struct omp_for
stopvalull
= force_gimple_operand_gsi (gsi, stopvalull, true, NULL_TREE,
false, GSI_CONTINUE_LINKING);
- min_inner_iterations
- = fold_convert (slltype, min_inner_iterations);
- min_inner_iterations
- = force_gimple_operand_gsi (gsi, min_inner_iterations, true,
+ first_inner_iterations
+ = fold_convert (slltype, first_inner_iterations);
+ first_inner_iterations
+ = force_gimple_operand_gsi (gsi, first_inner_iterations, true,
NULL_TREE, false,
GSI_CONTINUE_LINKING);
factor = fold_convert (slltype, factor);
factor
= force_gimple_operand_gsi (gsi, factor, true, NULL_TREE,
false, GSI_CONTINUE_LINKING);
- tree min_inner_iterationsd
+ tree first_inner_iterationsd
= fold_build1 (FLOAT_EXPR, double_type_node,
- min_inner_iterations);
- min_inner_iterationsd
- = force_gimple_operand_gsi (gsi, min_inner_iterationsd, true,
+ first_inner_iterations);
+ first_inner_iterationsd
+ = force_gimple_operand_gsi (gsi, first_inner_iterationsd, true,
NULL_TREE, false,
GSI_CONTINUE_LINKING);
tree factord = fold_build1 (FLOAT_EXPR, double_type_node,
@@ -2318,7 +2318,7 @@ expand_omp_for_init_vars (struct omp_for
t = fold_build2 (RDIV_EXPR, double_type_node, factord,
build_real (double_type_node, dconst2));
tree t3 = fold_build2 (MINUS_EXPR, double_type_node,
- min_inner_iterationsd, t);
+ first_inner_iterationsd, t);
t3 = force_gimple_operand_gsi (gsi, t3, true, NULL_TREE, false,
GSI_CONTINUE_LINKING);
t = fold_build2 (MULT_EXPR, double_type_node, factord,
@@ -2356,12 +2356,12 @@ expand_omp_for_init_vars (struct omp_for
t = fold_build2 (RSHIFT_EXPR, ulltype, t, integer_one_node);
t = fold_build2 (MULT_EXPR, ulltype, fd->factor, t);
tree t2 = fold_build2 (MULT_EXPR, ulltype, c,
- fd->min_inner_iterations);
+ fd->first_inner_iterations);
t = fold_build2 (PLUS_EXPR, ulltype, t, t2);
expand_omp_build_assign (gsi, d, t, true);
t = fold_build2 (MULT_EXPR, ulltype, fd->factor, c);
t = fold_build2 (PLUS_EXPR, ulltype,
- t, fd->min_inner_iterations);
+ t, fd->first_inner_iterations);
t2 = force_gimple_operand_gsi (gsi, t, true, NULL_TREE, false,
GSI_CONTINUE_LINKING);
cond_stmt = gimple_build_cond (GE_EXPR, stopvalull, d,
Jakub
