https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106557
anlauf at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |anlauf at gcc dot gnu.org
Ever confirmed|0 |1
Status|UNCONFIRMED |NEW
Last reconfirmed| |2022-08-18
--- Comment #1 from anlauf at gcc dot gnu.org ---
Confirmed.
The underlying issue also affects ibclr, and possibly other intrinsics that
are simplified at compile time.
Reduced testcase:
print *, ibset (transfer (0, 0) , 0) ! bad
print *, ibset (transfer (0, 0) + 0, 0) ! OK
print *, ibset (transfer (0, 0) , 0) + 0 ! OK
print *, transfer (ibset (transfer (0, 0), 0), 0) == 1 ! bad
print *, transfer (ibclr (transfer (1, 0), 0), 0) == 0 ! bad
end
The problem arises from the dual representation of the result of the
simplification of TRANSFER in gfc_simplify_transfer(), once as
result->value.integer, and as result->representation.string.
The latter is used e.g. for ensuring safe round-trip for nested
TRANSFER(TRANSFER(...),...).
It happens in gfc_simplify_ibset that value.integer is updated, but not
representation.string, and under the given circumstances either one or the
other is used later.
Possible solutions:
1) conservative hack:
diff --git a/gcc/fortran/simplify.cc b/gcc/fortran/simplify.cc
index fb725994653..6c5ffcdaf86 100644
--- a/gcc/fortran/simplify.cc
+++ b/gcc/fortran/simplify.cc
@@ -3380,6 +3380,11 @@ gfc_simplify_ibclr (gfc_expr *x, gfc_expr *y)
k = gfc_validate_kind (x->ts.type, x->ts.kind, false);
result = gfc_copy_expr (x);
+ if (result->representation.string)
+ {
+ free (result->representation.string);
+ result->representation.string = NULL;
+ }
convert_mpz_to_unsigned (result->value.integer,
gfc_integer_kinds[k].bit_size);
and a similar change to gfc_simplify_ibset. This fixes the issue just for
those two intrinsics. This may miss other cases, although I could not find
them so far.
2) more aggressive hack:
diff --git a/gcc/fortran/simplify.cc b/gcc/fortran/simplify.cc
index fb725994653..3e895e12f8c 100644
--- a/gcc/fortran/simplify.cc
+++ b/gcc/fortran/simplify.cc
@@ -8157,6 +8158,16 @@ gfc_simplify_transfer (gfc_expr *source, gfc_expr *mold,
gfc_expr *size)
/* And read the buffer back into the new expression. */
gfc_target_interpret_expr (buffer, buffer_size, result, false);
+ /* Integer is capable to hold all bits needed for complete round-trip
+ with TRANSFER-in-TRANSFER. Drop separate memory representation so that
+ subsequent simplification of bit manipulation intrinsics of the result
+ of gfc_simplify_transfer does not lead to inconsistencies. */
+ if (result->ts.type == BT_INTEGER)
+ {
+ free (result->representation.string);
+ result->representation.string = NULL;
+ }
+
return result;
}
Both patches regtest cleanly on x86_64-pc-linux-gnu.
