| Issue |
171870
|
| Summary |
LLVM emits late return-value initialization instead of hoisting it before branch conditions
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
foxtran
|
For the following code, one can get a quite nice ASM output both for x86-64 and Arm v8 (`-O3`). See the code here: https://godbolt.org/z/3xjhKf79T. RISC-V have similar ASM.
```cpp
#include <vector>
int sum(const std::vector<int> vec) {
int total = 0;
for(std::size_t index = 0; index != vec.size(); index++) {
total += vec[index];
}
return total;
}
```
However, in both cases, at the entering to function, I see the following pattern:
```asm
sum(std::vector<int, std::allocator<int>>):
mov rcx, qword ptr [rdi]
mov rdx, qword ptr [rdi + 8]
sub rdx, rcx
je .LBB0_1
sar rdx, 2
cmp rdx, 8
jae .LBB0_6
xor esi, esi
xor eax, eax <- set return register to 0
jmp .LBB0_5
.LBB0_1:
xor eax, eax <- set return register to 0
jmp .LBB0_2
.LBB0_6:
```
Later, this register is used only for keeping `total`. So, then, instruction can be resorted as follows:
```asm
sum(std::vector<int, std::allocator<int>>):
mov rcx, qword ptr [rdi]
mov rdx, qword ptr [rdi + 8]
sub rdx, rcx
xor eax, eax <- set return register to 0
je .LBB0_1
sar rdx, 2
cmp rdx, 8
jae .LBB0_6
xor esi, esi
jmp .LBB0_5
.LBB0_1:
jmp .LBB0_2
.LBB0_6:
```
And with optimized jumps:
```asm
sum(std::vector<int, std::allocator<int>>):
mov rcx, qword ptr [rdi]
mov rdx, qword ptr [rdi + 8]
sub rdx, rcx
xor eax, eax <- set return register to 0
je .LBB0_2
sar rdx, 2
cmp rdx, 8
jae .LBB0_6
xor esi, esi
jmp .LBB0_5
.LBB0_6:
```
The same approach can be used also for Arm v8 and RISC-V.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
