https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91616
Bug ID: 91616
Summary: Incorrect data address computation in very simple code
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: mpoulhies at kalray dot eu
Target Milestone: ---
The following code is miscompiled in O2 only (0,1 and 3 are correct):
typedef unsigned long long uint64_t;
uint64_t data_array[16] = {0};
unsigned int i = 0;
int main(void) {
unsigned long long seed = 0xcafebebedeadbeef;
for (i = 0; i < 16; i++)
data_array[i] = (seed++);
return 0;
}
It emits:
main:
movabs rax, -3819405706974609681
movabs rsi, OFFSET FLAT:data_array-6338242491622225792 // <- this is not
correct
movabs rcx, -3819405706974609665
.L2:
mov rdx, rax
add rax, 1
mov QWORD PTR [rsi+rax*8], rdx
cmp rax, rcx
jne .L2
mov DWORD PTR i[rip], 16
xor eax, eax
ret
i:
.zero 4
data_array:
.zero 128
After looking at compilers available on godbolt, all GCC starting from 8.* are
showing this behavior, see https://godbolt.org/z/OX0s10
The pass that seems to introduce this wrong address is 'ivopts' (first seen in
164t.ivopts in 9.2.1)
