https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109849
Jan Hubicka <hubicka at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Blocks| |109811
CC| |mjambor at suse dot cz
--- Comment #6 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Here is slightly improved testcase which actually pushes into stack and
measures something. It test loops 1000 times and returns. It also makes stack
to be local variable so race conditions are not a problem.
#include <vector>
typedef unsigned int uint32_t;
std::pair<uint32_t, uint32_t> pair;
void
test()
{
std::vector<std::pair<uint32_t, uint32_t>> stack;
stack.push_back (pair);
while (!stack.empty()) {
std::pair<uint32_t, uint32_t> cur = stack.back();
stack.pop_back();
if (!cur.first)
{
cur.second++;
stack.push_back (cur);
}
if (cur.second > 10000)
break;
}
}
int
main()
{
for (int i = 0; i < 10000; i++)
test();
}
Clang code is about twice as fast
jan@localhost:/tmp> clang++ -O2 tt.C -fno-exceptions
jan@localhost:/tmp> g++ -O2 tt.C -fno-exceptions -o a.out-gcc
jan@localhost:/tmp> perf stat ./a.out
Performance counter stats for './a.out':
434.24 msec task-clock:u # 0.997 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
129 page-faults:u # 297.073 /sec
1,003,191,657 cycles:u # 2.310 GHz
68,927 stalled-cycles-frontend:u # 0.01% frontend
cycles idle
800,792,619 stalled-cycles-backend:u # 79.82% backend
cycles idle
1,904,682,933 instructions:u # 1.90 insn per
cycle
# 0.42 stalled cycles per
insn
500,912,196 branches:u # 1.154 G/sec
23,144 branch-misses:u # 0.00% of all
branches
0.435340389 seconds time elapsed
0.431409000 seconds user
0.003994000 seconds sys
jan@localhost:/tmp> perf stat ./a.out-gcc
Performance counter stats for './a.out-gcc':
1,197.28 msec task-clock:u # 0.999 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
131 page-faults:u # 109.415 /sec
2,903,995,656 cycles:u # 2.425 GHz
86,204 stalled-cycles-frontend:u # 0.00% frontend
cycles idle
2,690,907,052 stalled-cycles-backend:u # 92.66% backend
cycles idle
2,005,212,311 instructions:u # 0.69 insn per
cycle
# 1.34 stalled cycles per
insn
401,007,320 branches:u # 334.932 M/sec
23,290 branch-misses:u # 0.01% of all
branches
1.198388186 seconds time elapsed
1.198450000 seconds user
0.000000000 seconds sys
The problem seems to be, like in first example, that we keep updating in-memory
stack in the main loop.
.L39:
movl 12(%rsp), %ebx
.L30:
movq 16(%rsp), %rax
cmpl $10000, %ebx
ja .L33
.L40:
movq 24(%rsp), %rdi
cmpq %rdi, %rax
je .L28
.L34:
movq -8(%rdi), %rax
leaq -8(%rdi), %rsi
movq %rsi, 24(%rsp)
movq %rax, 8(%rsp)
testl %eax, %eax
jne .L39
While clang does:
.LBB0_1: # in Loop: Header=BB0_4 Depth=1
movq %rax, %r14
.LBB0_2: # in Loop: Header=BB0_4 Depth=1
movq %rbx, %r12
movq %r12, %rbx
cmpl $10001, %r13d # imm = 0x2711
jae .LBB0_27
.LBB0_4: # =>This Loop Header: Depth=1
# Child Loop BB0_16 Depth 2
# Child Loop BB0_21 Depth 2
cmpq %r14, %rbx
je .LBB0_26
# %bb.5: # in Loop: Header=BB0_4 Depth=1
leaq -8(%r14), %rax
movq -8(%r14), %rcx
movq %rcx, %r13
shrq $32, %r13
testl %ecx, %ecx
jne .LBB0_1
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109811
[Bug 109811] libjxl 0.7 is a lot slower in GCC 13.1 vs Clang 16