https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113585
Bug ID: 113585
Summary: Poor codegen turning int compare into -1,0,1
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: redbeard0531 at gmail dot com
Target Milestone: ---
https://www.godbolt.org/z/Y31xG7EeT
These two functions should be equivalent, but comp2() produces better code than
comp1() on both arm64 and x86_64
int comp1(int a, int b) {
return a == b ? 0 : (a < b ? -1 : 1);
}
int comp2(int a, int b) {
return a < b ? -1 : (a > b ? 1 : 0);
}
arm64:
comp1(int, int):
cmp w0, w1
mov w0, -1
csinc w0, w0, wzr, lt
csel w0, w0, wzr, ne
ret
comp2(int, int):
cmp w0, w1
cset w0, gt
csinv w0, w0, wzr, ge
ret
x86_64:
comp1(int, int):
xor eax, eax
cmp edi, esi
je .L1
setge al
movzx eax, al
lea eax, [rax-1+rax]
.L1:
ret
comp2(int, int):
xor eax, eax
cmp edi, esi
mov edx, -1
setg al
cmovl eax, edx
ret
In the arm case, I suspect that the perf will be equivalent, although comp1 has
an extra instruction, so will pay a size penalty. On x86, comp2 is branchless
while comp1 has a branch which could be problematic if not predictable. It
seems like there should be a normalization pass to convert these functions (and
other equivalent ones) into a single normalized representation so that they get
the same codegen.
PS: I tried another equivalent function and it produces even worse codegen on
x86_64, comparing the same registers twice:
int comp3(int a, int b) {
return a > b ? 1 : (a == b ? 0 : -1);
}
comp3(int, int):
xor eax, eax
cmp edi, esi
mov edx, 1
setne al
neg eax
cmp edi, esi
cmovg eax, edx
ret
