https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119164
--- Comment #3 from Vineet Gupta <vineetg at gcc dot gnu.org> ---
However the first part of prev fix alone can't pass a glibc build - hitting
bunch of ICE.
Following reduced test (w/ my patch applied) actually segfaults
build: -O2 -c -march=rv64gcv_zvl256b_zba_zbb_zbs_zicond -fexceptions
void a();
typedef struct {
void *b;
} c;
void d(int **);
c e;
void f() {
int *g __attribute__((cleanup(d)h);
if (e.b )
{
int i;
__atomic_compare_exchange_n(&i, f, 1, 0, 2, 0);
a();
}
}
So we end up a bunch of BBs with calls and ensuing code_labels.
Entering mode_sw we have 11 BBs (max index 10), a new FRM read is created in
stack below:
df_worklist_dataflow_doublequeue: n_basic_blocks 11 n_edges 12 count 11 (
1)
[uid 23] frm mode needed : Call
scanning new insn with uid = 64.
changing bb of uid 65 from 6 to 11
unscanned insn
riscv_frm_mode_needed
riscv_frm_emit_after_bb_end
commit_edge_insertions
However it seems to also create a new BB:
commit_one_edge_insertion
split_edge
rtl_split_edge
create_basic_block
...
df_insn_change_bb
Which runs afoul of generic mode_sw whose bitmap vectors are sized for 11 BBs
and haven't been adjusted for this new one.
for (i = 0; i < no_mode; i++)
{
int m = targetm.mode_switching.priority (entity_map[j], i);
FOR_EACH_BB_FN (bb, cfun)
{
if (!bitmap_bit_p (transp_all, bb->index)) <-- bb->index 11
clear_mode_bit (transp[bb->index], j, m); <-- segfault here
So I don't know if the general direction of fix is wrong or that we need to
make some adj to generic mode switch as well.
