On 10/28/21 9:51 AM, Tobias Burnus wrote:
Before this patch, running
#define TEST(T) T
#define PARALLEL(X) TEST(X)
PARALLEL(
for (int i = 0; i < N; i++) { \
_Pragma("omp ordered") \
S[0] += C[i] + D[i]; \
})
through 'gcc -E' yielded
#pragma omp ordered
for (int i = 0; i < N; i++) { S[0] += C[i] + D[i]; }
Note that the '#pragma omp ordered' is now above the loop, i.e. before
all macro arguments (and macro expansions).
With the patch, the result is the following, which matches Clang and I
assume GCC before 4.2 or 4.3, but I have no GCC 4.x available:
for (int i = 0; i < N; i++) {
#pragma omp ordered
S[0] += C[i] + D[i]; }
There are a number of bug reports of _Pragma not working right
in macros, including (and especially) to control diagnostics:
https://gcc.gnu.org/bugzilla/buglist.cgi?quicksearch=_Pragma%20macro&list_id=328003
Just by the description this change seems like it could also
fix some of them. It would be helpful to check to see if it
does and if so, add tests and resolve the bugs it fixes (I'm
willing to help with that in stage 3).
Martin
The reason seems to be the addition done for PR34692 in r131819, which
added code to avoid an ICE with
FOO(
#pragma GCC diagnostic
)
There is a length description in macro.c about what it does and the
pragma_buff which is passed around, including to the now modified
collect_args. Namely, the comment above enter_macro_context states:
If there were additionally any unexpanded deferred #pragma
directives among macro arguments, push another context containing
the pragma tokens before the yet-to-be-rescanned replacement list
and return two.
While that seems to work fine with #pragma, it obviously does not do
what it should for _Pragma. The solution in the patch was to add a
flag to distinguish the CPP_PRAGMA coming from the _Pragma operator
(alias BT_PRAGMA) from the CPP_PRAGMA coming from a user's #pragma.
OK for mainline? – It is a long-standing regression, but it hasn't
been reported for a while. Thus: how do you feel about backporting?
I did test it with a full bootstrap + regtesting. I also tested
omptests (cf. PR).
Tobias
PS: I had the hope that it would fix some of the other _Pragma related
PRs (see e.g. refs in this PR102409 or search Bugzilla), but it does
not seem to help for those. I do note that most of them are related to
diagnostic. In particular, for PR91669, the output of gcc -E is the
same for GCC 10, for a patched GCC and for clang-11, which makes the
result (issue unaffected by this patch) not that surprising.
-----------------
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
