c++: #pragma GCC unroll C++ fixes [PR112795]
commitb6c78feea08c36e5754818c6a3d7536b3f8913dc
authorJakub Jelinek <jakub@redhat.com>
Mon, 4 Dec 2023 07:59:15 +0000 (4 08:59 +0100)
committerJakub Jelinek <jakub@redhat.com>
Mon, 4 Dec 2023 07:59:15 +0000 (4 08:59 +0100)
tree633fa45ad7f3f8729789424c735c548af6fb0a10
parent9448428b6871a4f7cbc9a1971077c3f9f3742e0d
c++: #pragma GCC unroll C++ fixes [PR112795]

foo in the unroll-5.C testcase ICEs because cp_parser_pragma_unroll
during parsing calls maybe_constant_value unconditionally, which is
fine if !processing_template_decl, but can ICE otherwise.

While just calling fold_non_dependent_expr there instead could be enough
to fix the ICE (and I guess the right thing to do for backports if any),
I don't see a reason why we couldn't handle a dependent #pragma GCC unroll
argument as well, the unrolling isn't done in the FE and all the middle-end
cares about is that ANNOTATE_EXPR has a 1..65534 last operand when it is
annot_expr_unroll_kind.

So, the following patch changes all the unsigned short unroll arguments
to tree unroll (and thus avoids the tree -> unsigned short -> tree
conversions), does the type and value checking during parsing only if
the argument isn't dependent and repeats it during instantiation.

2023-12-04  Jakub Jelinek  <jakub@redhat.com>

PR c++/112795
gcc/cp/
* cp-tree.h (cp_convert_range_for): Change UNROLL type from
unsigned short to tree.
(finish_while_stmt_cond, finish_do_stmt, finish_for_cond): Likewise.
* parser.cc (cp_parser_statement): Pass NULL_TREE rather than 0 to
cp_parser_iteration_statement UNROLL argument.
(cp_parser_for, cp_parser_c_for): Change UNROLL type from
unsigned short to tree.
(cp_parser_range_for): Likewise.  Set RANGE_FOR_UNROLL to just UNROLL
rather than build_int_cst from it.
(cp_convert_range_for, cp_parser_iteration_statement): Change UNROLL
type from unsigned short to tree.
(cp_parser_omp_loop_nest): Pass NULL_TREE rather than 0 to
cp_parser_range_for UNROLL argument.
(cp_parser_pragma_unroll): Return tree rather than unsigned short.
If parsed expression is type dependent, just return it, don't diagnose
issues with value if it is value dependent.
(cp_parser_pragma): Change UNROLL type from unsigned short to tree.
* semantics.cc (finish_while_stmt_cond): Change UNROLL type from
unsigned short to tree.  Build ANNOTATE_EXPR with UNROLL as its last
operand rather than build_int_cst from it.
(finish_do_stmt, finish_for_cond): Likewise.
* pt.cc (tsubst_stmt) <case RANGE_FOR_STMT>: Change UNROLL type from
unsigned short to tree and set it to RECUR on RANGE_FOR_UNROLL (t).
(tsubst_expr) <case ANNOTATE_EXPR>: For annot_expr_unroll_kind repeat
checks on UNROLL value from cp_parser_pragma_unroll.
gcc/testsuite/
* g++.dg/ext/unroll-5.C: New test.
* g++.dg/ext/unroll-6.C: New test.
gcc/cp/cp-tree.h
gcc/cp/parser.cc
gcc/cp/pt.cc
gcc/cp/semantics.cc
gcc/testsuite/g++.dg/ext/unroll-5.C [new file with mode: 0644]
gcc/testsuite/g++.dg/ext/unroll-6.C [new file with mode: 0644]