Valgrind showed a leak of 1640 bytes per iteration of one of the jit
testcases (I think this is per-function in a compile):
8,200 bytes in 5 blocks are definitely lost in loss record 214 of 223
at 0x4A0645D: malloc (in /usr/lib64/valgrind/vgpreload_memcheck-amd64-linux.so)
by 0x5D75D1F: xrealloc (xmalloc.c:177)
by 0x4E0C94E: void va_heap::reserve<int>(vec<int, va_heap, vl_embed>*&, unsigned int, bool) (vec.h:310)
by 0x4E0C7A3: vec<int, va_heap, vl_ptr>::reserve(unsigned int, bool) (vec.h:1428)
by 0x4E0C55A: vec<int, va_heap, vl_ptr>::reserve_exact(unsigned int) (vec.h:1448)
by 0x4E0C2C6: vec<int, va_heap, vl_ptr>::create(unsigned int) (vec.h:1463)
by 0x519B316: lra_create_live_ranges(bool) (lra-lives.c:1248)
by 0x5179BD7: lra(_IO_FILE*) (lra.c:2297)
by 0x5126C4E: do_reload() (ira.c:5380)
by 0x5127098: (anonymous namespace)::pass_reload::execute(function*) (ira.c:5550)
by 0x52235BD: execute_one_pass(opt_pass*) (passes.c:2306)
by 0x5223834: execute_pass_list_1(opt_pass*) (passes.c:2358)
which is this line:
point_freq_vec.create (get_max_uid () * 2);
point_freq_vec's buffer is released in lra_clear_live_ranges.
Am unfamiliar with LRA, but my reading of the lra code is that the
"live_p" boolean tracks whether or not we need to call
lra_clear_live_ranges.
Debugging calls to the above line and calls to lra_clear_live_ranges
showed a mismatch; the call to lra_create_live_ranges before
lra_inheritance wasn't setting live_p.
Setting it after that callsite fixes the leak (though perhaps the code
could be cleaned up by having live_p be set/cleared by the
allocation/clear functions themselves, rather than by their callers?
Vladimir?)
gcc/ChangeLog:
PR jit/63854
* lra.c (lra): After creating live ranges in preparation for call
to lra_inheritance, set live_p to true.