https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123230
Bug ID: 123230
Summary: Missed cross-function optimization involving a loop
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: 652023330028 at smail dot nju.edu.cn
Target Milestone: ---
Hello,
For the reduced test case below, GCC at -O3 appears to miss a cross-function
optimization opportunity in function j.
Since h[*] is always zero, j() can be optimized to a direct call to a().
If the loop condition is changed from l < 6 to l < 5 (as in j_optimized), GCC
performs the optimization as expected.
Reduced code (from ABC project
https://github.com/berkeley-abc/abc/blob/64637b8395a94aa68e67b211d4ecc21840accde1/src/map/if/ifDec07.c#L409):
https://godbolt.org/z/9oza5o7W3
void a();
void c(int e, int h[]) {
while (e < 6 && h[e])
e++;
if (1 == e)
a();
}
void j() {
int h[6] = {0};
for (int l = 0; l < 6; l++){
c(l, h);
}
}
void j_optimized() {
int h[6] = {0};
for (int l = 0; l < 5; l++){
c(l, h);
}
}
Expected (Clang -O3):
j():
jmp a()@PLT
GCC -O3:
"j()":
push rbp
pxor xmm0, xmm0
push rbx
mov ebx, 1
sub rsp, 40
mov QWORD PTR [rsp+16], 0
mov rbp, rsp
movaps XMMWORD PTR [rsp], xmm0
jmp .L21
.L24:
mov esi, DWORD PTR [rbp+4]
test esi, esi
je .L23
cmp ebx, 5
je .L11
mov ecx, DWORD PTR [rbp+8]
test ecx, ecx
je .L11
cmp ebx, 4
je .L11
mov edx, DWORD PTR [rbp+12]
test edx, edx
je .L11
cmp ebx, 3
je .L11
.L11:
add rbx, 1
add rbp, 4
.L21:
mov edi, DWORD PTR [rbp+0]
test edi, edi
je .L14
cmp ebx, 6
jne .L24
.L12:
add rbx, 1
cmp rbx, 7
jne .L25
add rsp, 40
pop rbx
pop rbp
ret
.L14:
cmp ebx, 2
jne .L12
call "a()"
jmp .L12
.L23:
cmp ebx, 1
jne .L11
call "a()"
add rbx, 1
add rbp, 4
jmp .L21
.L25:
add rbp, 4
jmp .L21
