https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128697

            Bug ID: 128697
           Summary: var-tracking does not terminate with -g -Os (dataflow
                    oscillates in endless loop block)
           Product: gcc
           Version: 16.2.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: debug
          Assignee: unassigned at gcc dot gnu.org
          Reporter: frank.mehnert at googlemail dot com
  Target Milestone: ---

The compiler hangs in the var-tracking pass (100% CPU, never finishes) when
compiling the following test case with -g and -Os. Without -g, it compiles
instantly.

Test case (t.c):

struct cpu { unsigned v; };
static void raw_exit(struct cpu c) { (void)c; }
static void (*exit_question)(struct cpu) = raw_exit;
void cov_print(void) __attribute__((weak));

void set_exit(void (*eq)(struct cpu)) { exit_question = eq; }

void
terminate_from(struct cpu cpu)
{
  if (exit_question && !cov_print)
    exit_question(cpu);

  if (cov_print)
    cov_print();

  for (;;)
    __asm__ volatile ("nop");
}

Command:

  aarch64-linux-gnu-gcc -c -g -Os t.c

Results with aarch64-linux-gnu-gcc:

  11.5.0  -g -Os                              ok
  13.5.0  -g -Os                              hangs
  14.4.0  -g -Os                              hangs
  15.3.0  -g -Os                              hangs
  16.2.0  -g -Os                              hangs
  16.2.0  -g -O2, -g -O3                      hang
  16.2.0  -g -O1, -g -Og                      ok
  16.2.0  -Os                                 ok
  16.2.0  -g -Os -fno-var-tracking            ok
  16.2.0  -g -Os -fno-var-tracking-assignments ok

The same test case (with "wait" or "nop" in the asm) also hangs on MIPS
(-mabi=32 -march=mips32r6 -mno-abicalls -fno-pic -g -Os, also mips32r2 and
mips64r6) with GCC 13.4.0 and 14.2.0. GCC 12 was not tested. It does not hang
on arm-linux-gnueabihf-gcc 16.2.0 or on x86-64 GCC 16.2.0.

All of the following are needed to trigger the hang:
- the struct passed by value (an int parameter compiles fine),
- the test "!cov_print" of the weak symbol in the first condition,
- the call cov_print(),
- the endless loop at the end.

The original code comes from an OS kernel (terminate function that ends in an
endless "wait" loop) and was reduced manually.

Workaround: -fno-var-tracking-assignments, or restructuring the code so that
cov_print is tested only once (if (cov_print) ... else ...).

Reply via email to