https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109849
--- Comment #8 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
We can only SRA if the address is non-escaping. Clang does not seem to need it
to optimize better:
jan@localhost:~> cat t.c
extern void q(int *);
__attribute__ ((noinline))
void
test()
{
for (int a = 0; a < 1000;a++)
if (!(a%100))
q(&a);
}
int
main()
{
for (int a = 0; a < 1000000;a++)
test ();
}
jan@localhost:~> cat t2.c
void q(int *a)
{
}
jan@localhost:~> gcc -O2 t.c t2.c ; perf stat ./a.out
Performance counter stats for './a.out':
2,916.73 msec task-clock:u # 0.999 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
52 page-faults:u # 17.828 /sec
8,344,719,833 cycles:u # 2.861 GHz
13,561,375 stalled-cycles-frontend:u # 0.16% frontend
cycles idle
5,128,112,757 stalled-cycles-backend:u # 61.45% backend
cycles idle
10,050,172,242 instructions:u # 1.20 insn per
cycle
# 0.51 stalled cycles per
insn
2,034,043,082 branches:u # 697.370 M/sec
11,186,312 branch-misses:u # 0.55% of all
branches
2.918344737 seconds time elapsed
2.917844000 seconds user
0.000000000 seconds sys
jan@localhost:~> clang -O2 t.c t2.c ; perf stat ./a.out
Performance counter stats for './a.out':
664.40 msec task-clock:u # 0.999 CPUs
utilized
0 context-switches:u # 0.000 /sec
0 cpu-migrations:u # 0.000 /sec
54 page-faults:u # 81.276 /sec
2,318,095,848 cycles:u # 3.489 GHz
10,417,694 stalled-cycles-frontend:u # 0.45% frontend
cycles idle
1,057,731,301 stalled-cycles-backend:u # 45.63% backend
cycles idle
10,062,172,840 instructions:u # 4.34 insn per
cycle
# 0.11 stalled cycles per
insn
2,034,042,724 branches:u # 3.061 G/sec
10,003,620 branch-misses:u # 0.49% of all
branches
0.665267996 seconds time elapsed
0.665247000 seconds user
0.000000000 seconds sys
We do:
jmp .L3
.p2align 4,,10
.p2align 3
.L2:
movl 12(%rsp), %eax
addl $1, %eax
movl %eax, 12(%rsp)
cmpl $999, %eax
jg .L7
.L3:
imull $-1030792151, %eax, %eax
addl $85899344, %eax
rorl $2, %eax
cmpl $42949672, %eax
ja .L2
leaq 12(%rsp), %rdi
call q
jmp .L2
Which has stupid store-to-load dpendency in the internal loop. Clang keeps the
store but optimizes away the load:
jmp .LBB0_1
.p2align 4, 0x90
.LBB0_3: # in Loop: Header=BB0_1 Depth=1
leal 1(%rax), %ecx
movl %ecx, 12(%rsp)
cmpl $999, %eax # imm = 0x3E7
movl %ecx, %eax
jge .LBB0_4
.LBB0_1: # =>This Inner Loop Header: Depth=1
imull $-1030792151, %eax, %ecx # imm = 0xC28F5C29
addl $85899344, %ecx # imm = 0x51EB850
rorl $2, %ecx
cmpl $42949672, %ecx # imm = 0x28F5C28
ja .LBB0_3
# %bb.2: # in Loop: Header=BB0_1 Depth=1
movq %rbx, %rdi
callq q@PLT
movl 12(%rsp), %eax
jmp .LBB0_3
Wonder what makes clang to think it needs @PLT though.
Why we do not consider the load as partially redundant with itself?
