https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111894
Bug ID: 111894
Summary: Missed vectorization opportunity
Product: gcc
Version: 13.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: Mark_B53 at yahoo dot com
Target Milestone: ---
Consider the following code that uses views to implement a two dimensional
`iota`:
#include <array>
#include <ranges>
template<std::integral T>
std::ranges::range auto iota2D(T xbound, T ybound) {
auto fn = [=](T idx) { return std::tuple{idx / ybound, idx % ybound};
};
return std::views::iota(T{0}, xbound * ybound) |
std::views::transform(fn);
}
constexpr std::size_t N = 20;
std::array<std::array<int, N>, N> data;
__attribute__((noinline)) void init1() {
for (auto i : std::views::iota(size_t{}, N)) {
for (auto j : std::views::iota(size_t{}, N)) {
data[i][j] = 123;
}
}
}
__attribute__((noinline)) void init2() {
for (auto [i, j] : iota2D(N,N)) {
data[i][j] = 123;
}
}
Using gcc 13.2 with -O3, we see that the code using a nested loop is nicely
vectorized:
init1():
movdqa xmm0, XMMWORD PTR .LC0[rip]
mov eax, OFFSET FLAT:data
.L2:
movaps XMMWORD PTR [rax], xmm0
add rax, 80
movaps XMMWORD PTR [rax-64], xmm0
movaps XMMWORD PTR [rax-48], xmm0
movaps XMMWORD PTR [rax-32], xmm0
movaps XMMWORD PTR [rax-16], xmm0
cmp rax, OFFSET FLAT:data+1600
jne .L2
ret
The code using iota2D is not vectorized:
init2():
xor eax, eax
.L6:
mov DWORD PTR data[0+rax*4], 123
add rax, 1
cmp rax, 400
jne .L6
ret
Although GCC 13 produces much higher quality assembly than previous versions,
it fails to vectorize the loop.