Fortran/OpenMP: Fix strictly structured blocks parsing
Commit Message
This patch is about error diagnostic + an ICE for invalid code.
Before the patch, gfortran/f951 showed:
...
Error: 'ancestor' device modifier not preceded by 'requires' directive with 'reverse_offload' clause
18 | end block
| 1
Error: Expecting END PROGRAM statement at (1)
gfortran: internal compiler error: Segmentation fault signal terminated program f951
* The first error is for a user error and fine.
* A follow-up error is expected (due to the now dangling '!$omp end target'),
but the error location and wording is misleading
* And the ICE is plainly wrong.
With the patch, there is no ICE and the the second error reads:
16 | !$omp end target
Error: Unexpected !$OMP END TARGET statement at (1)
The problem was due to nesting '!$omp target' (= same directive),
where the outer one was a strictly structured block - and parsing
the inner '!$omp target' failed with MATCH_ERROR, which ignored that
line - while '!$omp end target' remained.
(So far, so good, but then the parsing-code did run into a bug.)
For the blocks, the following applies. OpenMP permits either
* strictly structured blocks (with optional END_ST == 'end target')
!$omp target
block
...
end block
!$omp end target ! << this line is optional
* loosely structured block
!$omp target
... ! may not start with 'block' (and hence cannot end with 'end block')
!$omp end target ! << required
The parsing issue is in the following code,
which first takes care of the 'strictly': 'end block' + optional 'end target'
and then of the 'loosely structured' case with just: 'end target':
else if (block_construct && st == ST_END_BLOCK)
...
st = next_statement ();
if (st == omp_end_st)
accept_statement (st);
...
else if (st != omp_end_st)
{
unexpected_statement (st);
st = next_statement ();
}
The fix is to change the second if condition to:
else if (st != omp_end_st || (block_construct && st == omp_end_st))
or rather to the following equivalent code:
else if (st != omp_end_st || block_construct)
OK for mainline and GCC 12?*
Tobias
*strictly structured blocks were added in r12-4592.
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955
Comments
Hi Tobias,
Am 24.08.22 um 19:47 schrieb Tobias Burnus:
> This patch is about error diagnostic + an ICE for invalid code.
[...]
> (So far, so good, but then the parsing-code did run into a bug.)
>
> For the blocks, the following applies. OpenMP permits either
> * strictly structured blocks (with optional END_ST == 'end target')
> !$omp target
> block
> ...
> end block
> !$omp end target ! << this line is optional
> * loosely structured block
> !$omp target
> ... ! may not start with 'block' (and hence cannot end with 'end
> block')
> !$omp end target ! << required
>
>
> The parsing issue is in the following code,
> which first takes care of the 'strictly': 'end block' + optional 'end
> target'
> and then of the 'loosely structured' case with just: 'end target':
>
> else if (block_construct && st == ST_END_BLOCK)
> ...
> st = next_statement ();
> if (st == omp_end_st)
> accept_statement (st);
> ...
> else if (st != omp_end_st)
> {
> unexpected_statement (st);
> st = next_statement ();
> }
>
> The fix is to change the second if condition to:
> else if (st != omp_end_st || (block_construct && st == omp_end_st))
>
> or rather to the following equivalent code:
> else if (st != omp_end_st || block_construct)
LGTM.
> OK for mainline and GCC 12?*
Yes for both.
Thanks for the patch!
Harald
> Tobias
>
> *strictly structured blocks were added in r12-4592.
>
>
> -----------------
> Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201,
> 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer:
> Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München;
> Registergericht München, HRB 106955
Fortran/OpenMP: Fix strictly structured blocks parsing
gcc/fortran/ChangeLog:
* parse.cc (parse_omp_structured_block): When parsin strictly
structured blocks, issue an error if the end-directive comes
before the 'end block'.
gcc/testsuite/ChangeLog:
* gfortran.dg/gomp/strictly-structured-block-4.f90: New test.
gcc/fortran/parse.cc | 2 +-
.../gomp/strictly-structured-block-4.f90 | 21 +++++++++++++++++++++
2 files changed, 22 insertions(+), 1 deletion(-)
@@ -5709,7 +5709,7 @@ parse_omp_structured_block (gfc_statement omp_st, bool workshare_stmts_only)
}
return st;
}
- else if (st != omp_end_st)
+ else if (st != omp_end_st || block_construct)
{
unexpected_statement (st);
st = next_statement ();
new file mode 100644
@@ -0,0 +1,21 @@
+! { dg-do compile }
+implicit none
+integer ::x,z
+x = 42
+print '(*(z16:" "))', loc(x)
+!$omp target map(x, z)
+block
+ integer :: y
+ x = 123
+ y = 99
+ !$omp target device(ancestor:1) map(always,tofrom:x) map(y) ! { dg-error "'ancestor' device modifier not preceded by 'requires' directive with 'reverse_offload' clause" }
+ print '(*(z16:" "))', loc(x), loc(y)
+ print * ,x, y
+ x = -x
+ y = -y
+ !$omp end target ! { dg-error "Unexpected ..OMP END TARGET statement" }
+ z = y
+end block
+ print * ,x !, z
+end
+