| Issue |
141630
|
| Summary |
Converting loop pointer bump to index increases register pressure
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
dzaima
|
https://godbolt.org/z/oEzY8G6PK
The code:
```c
#include<stddef.h>
__attribute__((noreturn))
void unexpected(size_t i);
void external(int,int,int,int,int);
void foo(int* start, size_t n, int arg1, int arg2, int arg3, int arg4) {
int* curr = start;
int* end = start + n;
while (curr < end) {
int val = *curr;
if (__builtin_expect(val < 0, 0)) {
unexpected(curr - start);
}
external(val, arg1, arg2, arg3, arg4);
curr++;
}
}
```
compiled with `-O3` gets compiled to this core loop:
```asm
.LBB0_2:
mov edi, dword ptr [r12 + r13]
test edi, edi
js .LBB0_5
mov esi, r15d
mov edx, r14d
mov ecx, ebp
mov r8d, dword ptr [rsp + 4] ; avoidable!
call external
lea rax, [r12 + r13]
add rax, 4
add r13, 4
cmp rax, rbx
jb .LBB0_2
```
which contains a stack reload, as a consequence of LLVM replacing the pointer bump with a (scaled) index, thus taking two registers to store `curr`. Whereas gcc produces:
```asm
.L3:
mov r8d, r13d
mov ecx, r12d
mov edx, ebp
mov esi, ebx
call "external"
add r15, 4
cmp r15, r14
jnb .L1
.L4: mov edi, DWORD PTR [r15]
test edi, edi
jns .L3
```
which doesn't contain the reload, and also saves two instructions on computing `curr` for the comparison (by not having to do it at all, whereas LLVM does it in an inefficient way (the `lea; add; add` could trivially be `add; lea`, but this is unrelated to the main issue)), and as such is two instructions shorter.
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
