https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105739
--- Comment #8 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
After inlning I see:
IPA function summary for rcu_tasks_trace_pertask/5350 inlinable
global time: 13.535950
self size: 11
global size: 16
min size: 11
self stack: 0
global stack: 0
estimated growth:5
size:8.000000, time:5.807250
size:3.000000, time:2.000000, executed if:(not inlined)
size:2.000000, time:2.000000, nonconst if:(op0 changed)
calls:
rcu_tasks_trace_pertask.part.0/5788 inlined
freq:0.62
Stack frame offset 0, callee self size 0
trc_wait_for_one_reader/5799 inlined
freq:0.62
Stack frame offset 0, callee self size 0
__builtin_unreachable/5800 unreachable
freq:0.00 loop depth: 0 size: 0 time: 0 predicate: (false)
op0 is compile time invariant
op1 is compile time invariant
trc_wait_for_one_reader.part.0/5784 --param max-inline-insns-auto limit
reached
freq:0.31 loop depth: 0 size: 3 time: 12 callee size:100 stack: 0
__builtin_expect/5421 function body not available
freq:1.00 loop depth: 0 size: 0 time: 0
op1 is compile time invariant
So it seems that we determine call in trc_wait_for_one_reader unreachable.
It originally calls printk:
IPA function summary for trc_wait_for_one_reader/5348 inlinable
global time: 13.500000
self size: 20
global size: 20
min size: 1
self stack: 0
global stack: 0
size:1.000000, time:1.000000
size:3.000000, time:2.000000, executed if:(not inlined)
size:0.500000, time:0.500000, executed if:(not inlined), nonconst
if:(op0[ref offset: 8672] changed) && (not inlined)
size:2.500000, time:2.500000, nonconst if:(op0[ref offset: 8672] changed)
size:1.500000, time:0.250000, executed if:(op0[ref offset: 8672] != -1) &&
(not inlined)
size:3.500000, time:1.250000, executed if:(op0[ref offset: 8672] != -1)
calls:
trc_wait_for_one_reader.part.0/5784 function not considered for inlining
freq:0.50 loop depth: 0 size: 3 time: 12 callee size:55 stack: 0
predicate: (op0[ref offset: 8672] == -1)
_printk/5452 function body not available
freq:0.00 loop depth: 0 size: 5 time: 14 predicate: (op0[ref offset:
8672] != -1)
op0 is compile time invariant
So we somehow figure out (op0[ref offset: 8672] != -1) here:
Enqueueing calls in rcu_tasks_trace_pertask.part.0/5788.
Estimating body: trc_wait_for_one_reader/5348
Known to be false: not inlined, op0[ref offset: 8672] != -1, op0[ref offset:
8672] changed
size:4 time:7.000000 nonspec time:12.000000
This seems to be based on:
Jump functions:
Jump functions of caller rcu_tasks_trace_pertask.part.0/5788:
callsite rcu_tasks_trace_pertask.part.0/5788 ->
trc_wait_for_one_reader/5348 :
param 0: PASS THROUGH: 0, op nop_expr
Aggregate passed by reference:
offset: 8672, type: int, CONST: -1
value: 0x0, mask: 0xffffffffffffffff
Unknown VR
param 1: PASS THROUGH: 1, op nop_expr
value: 0x0, mask: 0xffffffffffffffff
Unknown VR
Which seems correct:
rcu_tasks_trace_pertask.part.0 (struct task_struct * t, struct list_head * hop)
{
struct task_struct * D.58527;
u64 pfo_val__;
_Bool _1;
long int _2;
long int _3;
struct task_struct * _14;
<bb 4> [local count: 1073741824]:
<bb 2> [local count: 1073741824]:
__asm__ __volatile__("" : : : "memory");
MEM[(volatile u8 *)t_1(D) + 1089B] ={v} 0;
__asm__ __volatile__("lock; addl $0,-4(%%rsp)" : : : "cc", "memory");
t_1(D)->trc_ipi_to_cpu = -1;
trc_wait_for_one_reader (t_1(D), hop_2(D));
<bb 3> [local count: 1073741824]:
return;
}
so there is really store to trc_ipi_to_cpu. The code uses it as:
_60 ={v} t_59(D)->trc_ipi_to_cpu;
__asm__ __volatile__("" : : : "memory");
if (_60 != -1)
goto <bb 3>; [50.00%]
else
goto <bb 6>; [50.00%]
So bug may be that we ignore volatile flag when determining IPA predicates?
problem goes away with -fno-partial-inlining.
