https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127889
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Target| |x86_64-*-*
Ever confirmed|0 |1
Keywords| |compile-time-hog, ra
CC| |vmakarov at gcc dot gnu.org
Status|UNCONFIRMED |NEW
Last reconfirmed| |2026-10-14
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed. With -O3 I see
LRA hard reg assignment : 51.99 ( 53%) 0 ( 0%)
load CSE after reload : 17.76 ( 18%) 1584 ( 0%)
TOTAL : 98.76 292M
with -O2+ it's documented that we enable parts of the compiler that scale badly
with program or function size. In particular comparing the IL with -O2 vs -O3
at RA shows:
-O2: 151276 a/t.c.329r.ira
-O3: 586385 b/t.c.329r.ira
there is -fno-expensive-optimizations that should gate the above mentioned
parts (but it does not).
On trunk the figure above adds
reload CSE regs : 10.28 ( 8%) 15M ( 5%)
perf reports
Samples: 414K of event 'cycles:P', Event count (approx.): 450714822924
Overhead Samples Command Shared Object Symbol
22.26% 93514 cc1 cc1 [.] lra_assign(bool&)
◆
11.15% 47516 cc1 cc1 [.]
bitmap_clear_bit(bitmap_head*, int) ▒
6.89% 28977 cc1 cc1 [.]
bitmap_set_bit(bitmap_head*, int) ▒
5.07% 20477 cc1 cc1 [.]
refs_may_alias_p_2(ao_ref*, ao_ref*, bool) ▒
4.92% 20863 cc1 cc1 [.] update_lives(int,
bool) ▒
3.81% 15434 cc1 cc1 [.]
get_ref_base_and_extent(tree_node*, poly_int<1u, long>*, po▒
2.77% 11215 cc1 cc1 [.]
ao_ref_from_mem(ao_ref*, rtx_def const*)
and lra_assign points to spill_for doing
EXECUTE_IF_SET_IN_BITMAP (&spill_pseudos_bitmap, 0, spill_regno, bi)
{
if (bitmap_bit_p (&insn_conflict_pseudos, spill_regno))
insn_pseudos_num++;
if (spill_regno >= (unsigned int) lra_bad_spill_regno_start)
bad_spills_num++;
for (r = lra_reg_info[spill_regno].live_ranges;
r != NULL;
r = r->next)
{
for (p = r->start; p <= r->finish; p++)
{
lra_live_range_t r2;
for (r2 = start_point_ranges[p];
r2 != NULL;
r2 = r2->start_next)
if (r2->regno >= lra_constraint_new_regno_start)
sparseset_set_bit (live_range_reload_inheritance_pseudos,
r2->regno);
where we likely run into out-of-cache behavior after crossing some critical
working set size (and that will depend on your CPU).
