https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112612
Bug ID: 112612
Summary: [Missed Optimization] Holding on the loop variable
rather than a derived value which can replace it
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: eyalroz1 at gmx dot com
Target Milestone: ---
Consider the following function:
void foo(int* __restrict__ a) {
int i, val;
for (i = 0; i < 100; i++) {
val = 2 * i;
a[i] = val;
}
}
When compiling it for x86_64 with -O3 -fno-unroll-loops -fno-tree-vectorize,
GCC 7.2 used to give:
foo:
xor eax, eax
.L2:
mov DWORD PTR [rdi], eax
add eax, 2
add rdi, 4
cmp eax, 200
jne .L2
rep ret
which was rather wasteful, as eax and rdi - eax are linearly related. With GCC
13.2 or trunk on GodBolt as of today, this improves, but not really:
foo:
xor eax, eax
.L2:
lea edx, [rax+rax]
mov DWORD PTR [rdi+rax*4], edx
add rax, 1
cmp rax, 100
jne .L2
ret
So, we don't increment two things; but - we do have an addition-via-lea in each
iteration. Is that really necessary? I mean, instead of keeping the i variable
(in rax), we could keep v = 2 * i, and that's good enough for both addressing
and condition checking. Indeed, clang 17 emits:
foo: # @foo
xor eax, eax
.LBB0_1: # =>This Inner Loop Header: Depth=1
mov dword ptr [rdi + 2*rax], eax
add rax, 2
cmp rax, 200
jne .LBB0_1
ret
which is almost the same, except that it holds v = 2 * i rather than i. (clang
has produced this code since v3.0.0 at least.)
GodBolt link: https://gcc.godbolt.org/z/MjzTbr831
Originally discussed in this SO question:
https://stackoverflow.com/q/48354636/1593077