https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64191
Bug ID: 64191
Summary: -march=native messes up dead code elimination in loop
calling dtor
Product: gcc
Version: 5.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: petschy at gmail dot com
Without -march=native, the loops in the 3 fns are eliminated as expected,
resulting in single retq's.
With -march=native, the loop which calls the defined, but empty dtor is
compiled into something rather weird. However, the other empty Nop() call is
optimized away as expected.
g++-5.0.0 -g -O3 -Wall -Wextra -c 20141205-dtor_loop.cpp
g++-5.0.0 -g -O3 -Wall -Wextra -o 20141205-dtor_loop 20141205-dtor_loop.o
Dump of assembler code for function foo_dtor_loop(Foo*, unsigned int):
0x0000000000400570 <+0>: repz retq
Dump of assembler code for function bar_dtor_loop(Bar*, unsigned int):
0x0000000000400580 <+0>: repz retq
Dump of assembler code for function bar_nop_loop(Bar*, unsigned int):
0x0000000000400590 <+0>: repz retq
So far so good.
g++-5.0.0 -g -O3 -march=native -Wall -Wextra -c 20141205-dtor_loop.cpp
g++-5.0.0 -g -O3 -march=native -Wall -Wextra -o 20141205-dtor_loop
20141205-dtor_loop.o
Dump of assembler code for function foo_dtor_loop(Foo*, unsigned int):
0x0000000000400570 <+0>: retq
Dump of assembler code for function bar_dtor_loop(Bar*, unsigned int):
0x0000000000400578 <+0>: test %rdi,%rdi
0x000000000040057b <+3>: je 0x4005b8 <bar_dtor_loop(Bar*, unsigned
int)+64>
0x000000000040057d <+5>: mov %esi,%esi
0x000000000040057f <+7>: lea (%rdi,%rsi,4),%rax
0x0000000000400583 <+11>: cmp %rax,%rdi
0x0000000000400586 <+14>: jae 0x4005b8 <bar_dtor_loop(Bar*, unsigned
int)+64>
0x0000000000400588 <+16>: mov $0x3,%edx
0x000000000040058d <+21>: lea -0x4(%rax),%rsi
0x0000000000400591 <+25>: sub %rdi,%rdx
0x0000000000400594 <+28>: add %rsi,%rdx
0x0000000000400597 <+31>: mov %rdx,%rcx
0x000000000040059a <+34>: shr $0x2,%rcx
0x000000000040059e <+38>: lea 0x1(%rcx),%r8
0x00000000004005a2 <+42>: dec %rcx
0x00000000004005a5 <+45>: shr %rcx
0x00000000004005a8 <+48>: lea 0x2(%rcx,%rcx,1),%rcx
0x00000000004005ad <+53>: cmp $0x2f,%rdx
0x00000000004005b1 <+57>: jbe 0x4005b8 <bar_dtor_loop(Bar*, unsigned
int)+64>
0x00000000004005b3 <+59>: cmp %rcx,%r8
0x00000000004005b6 <+62>: je 0x4005b8 <bar_dtor_loop(Bar*, unsigned
int)+64>
0x00000000004005b8 <+64>: retq
Dump of assembler code for function bar_nop_loop(Bar*, unsigned int):
0x00000000004005c0 <+0>: retq
The bar_dtor_loop() fn is clearly a mess, unfortunately I can't follow the
computation.
The bar_inc_loop() does a single int increment on each object, to see what loop
code is generated if not empty fns are called. It is as expected: the loop is
unrolled 16x times, and the residual part is executed in a tight loop:
0x0000000000400648 <+120>: sub $0x4,%rdx
0x000000000040064c <+124>: incl (%rdx)
0x000000000040064e <+126>: cmp %rdx,%rdi
0x0000000000400651 <+129>: jb 0x400648 <bar_inc_loop(Bar*, unsigned
int)+120>
g++-5.0.0 -v
Using built-in specs.
COLLECT_GCC=g++-5.0.0
COLLECT_LTO_WRAPPER=/usr/local/libexec/gcc/x86_64-unknown-linux-gnu/5.0.0/lto-wrapper
Target: x86_64-unknown-linux-gnu
Configured with: ./configure --enable-languages=c,c++ --disable-multilib
--program-suffix=-5.0.0
Thread model: posix
gcc version 5.0.0 20141027 (experimental) (GCC)
cat /proc/cpuinfo
processor : 0
vendor_id : AuthenticAMD
cpu family : 21
model : 1
model name : AMD FX(tm)-8150 Eight-Core Processor
stepping : 2
microcode : 0x6000626
cpu MHz : 1400.000
cache size : 2048 KB
physical id : 0
siblings : 8
core id : 0
cpu cores : 4
apicid : 16
initial apicid : 0
fpu : yes
fpu_exception : yes
cpuid level : 13
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov
pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb
rdtscp lm constant_tsc rep_good nopl nonstop_tsc extd_apicid aperfmperf pni
pclmulqdq monitor ssse3 cx16 sse4_1 sse4_2 popcnt aes xsave avx lahf_lm
cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs
xop skinit wdt lwp fma4 nodeid_msr topoext perfctr_core perfctr_nb arat cpb
hw_pstate npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid
decodeassists pausefilter pfthreshold
bugs : fxsave_leak
bogomips : 7624.63
TLB size : 1536 4K pages
clflush size : 64
cache_alignment : 64
address sizes : 48 bits physical, 48 bits virtual
power management: ts ttp tm 100mhzsteps hwpstate cpb
Unfortunately, I couldn't test with the latest version since the build fails
with
../.././libcc1/findcomp.cc:20:20: fatal error: config.h: No such file or
directory
a while know, even after deleting everything and doing a git reset --hard HEAD.
----8<----8<----8<----
struct Foo
{
int i;
};
void foo_dtor_loop(Foo* p, unsigned int n)
{
if (p) {
Foo* e = p + n;
while (e > p) {
--e;
e->~Foo();
}
}
}
struct Bar
{
int i;
~Bar() { }
void Nop() { }
void Inc() { ++i; }
};
void bar_dtor_loop(Bar* p, unsigned int n)
{
if (p) {
Bar* e = p + n;
while (e > p) {
--e;
e->~Bar();
}
}
}
void bar_nop_loop(Bar* p, unsigned int n)
{
if (p) {
Bar* e = p + n;
while (e > p) {
--e;
e->Nop();
}
}
}
void bar_inc_loop(Bar* p, unsigned int n)
{
if (p) {
Bar* e = p + n;
while (e > p) {
--e;
e->Inc();
}
}
}
int main()
{
}
