This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[committed] Fix inscan reduction and conditional lastprivate interaction
- From: Jakub Jelinek <jakub at redhat dot com>
- To: gcc-patches at gcc dot gnu dot org
- Date: Thu, 4 Jul 2019 23:43:59 +0200
- Subject: [committed] Fix inscan reduction and conditional lastprivate interaction
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
Hi!
This patch fixes worksharing loops containing both conditional lastprivate
and inscan reduction(s). Furthermore, it for nowait omp for ensures there
is GOMP_loop_end_nowait call at the end after the second loop in scan
and not after the first one.
Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk.
2019-07-04 Jakub Jelinek <jakub@redhat.com>
* omp-expand.c (expand_omp_for_static_nochunk): Don't emit
GOMP_loop_start at the start of second worksharing loop in a scan.
For nowait, don't emit GOMP_loop_end_nowait at the end of first
worksharing loop in a scan even if there are conditional lastprivates,
and do emit GOMP_loop_end_nowait at the end of second worksharing loop.
* testsuite/libgomp.c/scan-9.c: New test.
* testsuite/libgomp.c/scan-10.c: New test.
--- gcc/omp-expand.c.jj 2019-07-03 07:02:16.511989061 +0200
+++ gcc/omp-expand.c 2019-07-04 11:54:22.576366802 +0200
@@ -3744,7 +3744,9 @@ expand_omp_for_static_nochunk (struct om
cond_var = OMP_CLAUSE_DECL (c);
}
if (fd->have_reductemp
- || fd->have_pointer_condtemp
+ /* For scan, we don't want to reinitialize condtemp before the
+ second loop. */
+ || (fd->have_pointer_condtemp && !fd->have_scantemp)
|| fd->have_nonctrl_scantemp)
{
tree t1 = build_int_cst (long_integer_type_node, 0);
@@ -4235,7 +4237,8 @@ expand_omp_for_static_nochunk (struct om
else
gsi_insert_after (&gsi, omp_build_barrier (t), GSI_SAME_STMT);
}
- else if (fd->have_pointer_condtemp)
+ else if ((fd->have_pointer_condtemp || fd->have_scantemp)
+ && !fd->have_nonctrl_scantemp)
{
tree fn = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT);
gcall *g = gimple_build_call (fn, 0);
--- libgomp/testsuite/libgomp.c/scan-9.c.jj 2019-07-04 12:04:48.235360731 +0200
+++ libgomp/testsuite/libgomp.c/scan-9.c 2019-07-04 11:08:54.038069390 +0200
@@ -0,0 +1,116 @@
+/* { dg-require-effective-target size32plus } */
+
+extern void abort (void);
+int r, a[1024], b[1024], x, y, z;
+
+__attribute__((noipa)) void
+foo (int *a, int *b)
+{
+ #pragma omp for reduction (inscan, +:r) lastprivate (conditional: z) firstprivate (x) private (y)
+ for (int i = 0; i < 1024; i++)
+ {
+ { y = a[i]; r += y + x + 12; }
+ #pragma omp scan inclusive(r)
+ { b[i] = r; if ((i & 1) == 0 && i < 937) z = r; }
+ }
+}
+
+__attribute__((noipa)) int
+bar (void)
+{
+ int s = 0;
+ #pragma omp parallel
+ #pragma omp for reduction (inscan, +:s) firstprivate (x) private (y) lastprivate (z)
+ for (int i = 0; i < 1024; i++)
+ {
+ { y = 2 * a[i]; s += y; z = y; }
+ #pragma omp scan inclusive(s)
+ { y = s; b[i] = y + x + 12; }
+ }
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (int *a, int *b)
+{
+ #pragma omp parallel for reduction (inscan, +:r) firstprivate (x) lastprivate (x)
+ for (int i = 0; i < 1024; i++)
+ {
+ { r += a[i] + x + 12; if (i == 1023) x = 29; }
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) int
+qux (void)
+{
+ int s = 0;
+ #pragma omp parallel for reduction (inscan, +:s) lastprivate (conditional: x, y)
+ for (int i = 0; i < 1024; i++)
+ {
+ { s += 2 * a[i]; if ((a[i] & 1) == 1 && i < 825) x = a[i]; }
+ #pragma omp scan inclusive(s)
+ { b[i] = s; if ((a[i] & 1) == 0 && i < 829) y = a[i]; }
+ }
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ x = -12;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b);
+ if (r != 1024 * 1023 / 2 || x != -12 || z != b[936])
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = 25;
+ }
+ if (bar () != 1024 * 1023 || x != -12 || z != 2 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -1;
+ }
+ r = 0;
+ baz (a, b);
+ if (r != 1024 * 1023 / 2 || x != 29)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -25;
+ }
+ if (qux () != 1024 * 1023 || x != 823 || y != 828)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s)
+ abort ();
+ }
+ return 0;
+}
--- libgomp/testsuite/libgomp.c/scan-10.c.jj 2019-07-04 12:09:44.171630459 +0200
+++ libgomp/testsuite/libgomp.c/scan-10.c 2019-07-04 12:11:12.169224358 +0200
@@ -0,0 +1,116 @@
+/* { dg-require-effective-target size32plus } */
+
+extern void abort (void);
+int r, a[1024], b[1024], x, y, z;
+
+__attribute__((noipa)) void
+foo (int *a, int *b)
+{
+ #pragma omp for reduction (inscan, +:r) lastprivate (conditional: z) firstprivate (x) private (y)
+ for (int i = 0; i < 1024; i++)
+ {
+ { b[i] = r; if ((i & 1) == 0 && i < 937) z = r; }
+ #pragma omp scan exclusive(r)
+ { y = a[i]; r += y + x + 12; }
+ }
+}
+
+__attribute__((noipa)) int
+bar (void)
+{
+ int s = 0;
+ #pragma omp parallel
+ #pragma omp for reduction (inscan, +:s) firstprivate (x) private (y) lastprivate (z)
+ for (int i = 0; i < 1024; i++)
+ {
+ { y = s; b[i] = y + x + 12; }
+ #pragma omp scan exclusive(s)
+ { y = 2 * a[i]; s += y; z = y; }
+ }
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (int *a, int *b)
+{
+ #pragma omp parallel for reduction (inscan, +:r) firstprivate (x) lastprivate (x)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ { r += a[i] + x + 12; if (i == 1023) x = 29; }
+ }
+}
+
+__attribute__((noipa)) int
+qux (void)
+{
+ int s = 0;
+ #pragma omp parallel for reduction (inscan, +:s) lastprivate (conditional: x, y)
+ for (int i = 0; i < 1024; i++)
+ {
+ { b[i] = s; if ((a[i] & 1) == 0 && i < 829) y = a[i]; }
+ #pragma omp scan exclusive(s)
+ { s += 2 * a[i]; if ((a[i] & 1) == 1 && i < 825) x = a[i]; }
+ }
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ x = -12;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b);
+ if (r != 1024 * 1023 / 2 || x != -12 || z != b[936])
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = 25;
+ s += i;
+ }
+ if (bar () != 1024 * 1023 || x != -12 || z != 2 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -1;
+ s += 2 * i;
+ }
+ r = 0;
+ baz (a, b);
+ if (r != 1024 * 1023 / 2 || x != 29)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -25;
+ s += i;
+ }
+ if (qux () != 1024 * 1023 || x != 823 || y != 828)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ s += 2 * i;
+ }
+ return 0;
+}
Jakub