Hi!
The parsing of !$omp workshare requires only small precise set of statements
that can appear inside, which the inlining of matmul assign breaks.
Fixed by disabling the inlining in !$omp workshare regions.
Long term, it would be nice to inline those in !$omp workshare again and
actually parallelize, but I guess in that case it should not be done during
frontend passes, but during workshare translation into generic - and would
need to arrange for proper worksharing of both of the actions (clearing of
the target array as well as actually computing it).
Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk
and 6.2.
2016-05-11 Jakub Jelinek <ja...@redhat.com>
PR fortran/70855
* frontend-passes.c (inline_matmul_assign): Disable in !$omp workshare.
* gfortran.dg/gomp/pr70855.f90: New test.
--- gcc/fortran/frontend-passes.c.jj 2016-02-29 19:33:09.000000000 +0100
+++ gcc/fortran/frontend-passes.c 2016-05-11 11:54:43.747298277 +0200
@@ -2812,6 +2812,12 @@ inline_matmul_assign (gfc_code **c, int
if (in_where)
return 0;
+ /* For now don't do anything in OpenMP workshare, it confuses
+ its translation, which expects only the allowed statements in there.
+ We should figure out how to parallelize this eventually. */
+ if (in_omp_workshare)
+ return 0;
+
expr1 = co->expr1;
expr2 = co->expr2;
if (expr2->expr_type != EXPR_FUNCTION
--- gcc/testsuite/gfortran.dg/gomp/pr70855.f90.jj 2016-05-11
12:03:37.627977013 +0200
+++ gcc/testsuite/gfortran.dg/gomp/pr70855.f90 2016-05-11 12:04:06.927575148
+0200
@@ -0,0 +1,18 @@
+! PR fortran/70855
+! { dg-do compile }
+! { dg-additional-options "-O2" }
+
+program pr70855
+ integer, parameter :: m = 4
+ integer, parameter :: n = 2
+ real :: a(m,n)
+ real :: x(n)
+ real :: y(m)
+ a = 1.0
+ x = 1.0
+!$omp parallel
+!$omp workshare
+ y(1:m) = matmul ( a(1:m,1:n), x(1:n) )
+!$omp end workshare
+!$omp end parallel
+end program pr70855
Jakub
