| Issue |
207882
|
| Summary |
[LV][libc++] failed to vectorize std::bitset in SPEC17 541.leela_r
|
| Labels |
libc++,
vectorizers,
missed-optimization
|
| Assignees |
|
| Reporter |
Andarwinux
|
https://godbolt.org/z/evMv3qxfG
found from https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121744
> 541.leela_r contains sth along the following:
```c++
#include <bitset>
#include <vector>
class Foo
{
public:
void fun (std::bitset<21*21>& blacksq);
std::vector<int> m_mcowner;
};
void Foo::fun(std::bitset<21*21>& blacksq)
{
for (unsigned int i = 0; i < blacksq.size(); i++)
if (blacksq[i])
m_mcowner[i]++;
}
```
gcc libstdc++:
```asm
"Foo::fun(std::bitset<441ul>&)":
mov r8, QWORD PTR [rdi]
vmovdqa xmm3, XMMWORD PTR .LC0[rip]
mov eax, 1
vpcmpeqd xmm9, xmm9, xmm9
vpbroadcastq xmm5, rax
mov eax, 3
vpsrld xmm8, xmm9, 26
kmovb k2, eax
mov eax, 4
vpxor xmm6, xmm6, xmm6
vpbroadcastd xmm7, eax
lea rdi, [r8+1760]
mov rdx, r8
.L3:
vpmovzxdq xmm2, xmm3
vpsrldq xmm1, xmm3, 8
vpmovzxdq xmm1, xmm1
vpandd xmm0, xmm3, xmm8
vpsrlq xmm2, xmm2, 6
vpextrq rax, xmm2, 1
vmovq rcx, xmm2
vpsrlq xmm1, xmm1, 6
vmovq xmm2, QWORD PTR [rsi+rcx*8]
vmovq rcx, xmm1
vpinsrq xmm2, xmm2, QWORD PTR [rsi+rax*8], 1
vpextrq rax, xmm1, 1
vmovq xmm1, QWORD PTR [rsi+rcx*8]
vpinsrq xmm1, xmm1, QWORD PTR [rsi+rax*8], 1
vpmovzxdq xmm4, xmm0
vpsrldq xmm0, xmm0, 8
vpmovzxdq xmm0, xmm0
vpsllvq xmm4, xmm5, xmm4
vpsllvq xmm0, xmm5, xmm0
vpandq xmm2, xmm2, xmm4
vpandq xmm0, xmm1, xmm0
vpcmpq k0, xmm2, xmm6, 4
vpcmpq k1, xmm0, xmm6, 4
kandb k0, k0, k2
kshiftlb k1, k1, 2
korb k3, k0, k1
kmovb eax, k3
and eax, 15
kmovb k1, eax
je .L2
vmovdqu32 xmm0{k1}{z}, XMMWORD PTR [rdx]
vpsubd xmm0, xmm0, xmm9
vmovdqu32 XMMWORD PTR [rdx]{k1}, xmm0
.L2:
add rdx, 16
vpaddd xmm3, xmm3, xmm7
cmp rdi, rdx
jne .L3
test BYTE PTR [rsi+55], 1
je .L10
inc DWORD PTR [r8+1760]
.L10:
ret
.LC0:
.long 0
.long 1
.long 2
.long 3
```
clang libstdc++ or libc++:
```asm
Foo::fun(std::bitset<441ul>&):
mov rax, qword ptr [rdi]
mov ecx, 6
jmp .LBB0_1
.LBB0_15:
add rcx, 7
cmp rcx, 447
je .LBB0_16
.LBB0_1:
lea rdx, [rcx - 6]
mov rdi, rdx
shr rdi, 6
mov rdi, qword ptr [rsi + 8*rdi]
bt rdi, rdx
jae .LBB0_3
inc dword ptr [rax + 4*rcx - 24]
.LBB0_3:
lea rdx, [rcx - 5]
mov rdi, rdx
shr rdi, 6
mov rdi, qword ptr [rsi + 8*rdi]
bt rdi, rdx
jae .LBB0_5
inc dword ptr [rax + 4*rcx - 20]
.LBB0_5:
lea rdx, [rcx - 4]
mov rdi, rdx
shr rdi, 6
mov rdi, qword ptr [rsi + 8*rdi]
bt rdi, rdx
jae .LBB0_7
inc dword ptr [rax + 4*rcx - 16]
.LBB0_7:
lea rdx, [rcx - 3]
mov rdi, rdx
shr rdi, 6
mov rdi, qword ptr [rsi + 8*rdi]
bt rdi, rdx
jae .LBB0_9
inc dword ptr [rax + 4*rcx - 12]
.LBB0_9:
lea rdx, [rcx - 2]
mov rdi, rdx
shr rdi, 6
mov rdi, qword ptr [rsi + 8*rdi]
bt rdi, rdx
jae .LBB0_11
inc dword ptr [rax + 4*rcx - 8]
.LBB0_11:
lea rdx, [rcx - 1]
mov rdi, rdx
shr rdi, 6
mov rdi, qword ptr [rsi + 8*rdi]
bt rdi, rdx
jae .LBB0_13
inc dword ptr [rax + 4*rcx - 4]
.LBB0_13:
mov rdx, rcx
shr rdx, 6
mov rdx, qword ptr [rsi + 8*rdx]
bt rdx, rcx
jae .LBB0_15
inc dword ptr [rax + 4*rcx]
jmp .LBB0_15
.LBB0_16:
ret
```
Also reported as a libc++ issue, since vectorization (manual) is also missing with libc++.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
