https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115520
Bug ID: 115520
Summary: Loop vectorization depends on variable names
Product: gcc
Version: 14.1.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: max.sagebaum at scicomp dot uni-kl.de
Target Milestone: ---
Created attachment 58447
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58447&action=edit
Reproducer
I asked this question on gcc-help and got the reply that this is probably a
bug.
(https://gcc.gnu.org/pipermail/gcc-help/2024-June/143496.html)
Compiler options: -O3 -std=c++20 -fopt-info-vec-missed
Version: 14.1 (Compiler Explorer)
Reproducer on Compiler Explorer: https://godbolt.org/z/YEPzfx3eT
With the above options this loop vectorizes:
```
struct match_return {
char* pos;
bool matched;
};
bool func_is_vectorized (char*& cur, context& ctx) {
match_return r {cur, true};
constexpr char str1[] =
"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa";
if (!(std::distance(r.pos, ctx.end) < 50)) {return false;}
for(int i = 0; i < 50; i += 1) {
if (str1[i] != r.pos[i]) { r.matched = false; }
}
if (r.matched) {
r.pos += 50;
}
return r.matched;
}
```
If I remove the match_return structure, then the loop no longer vectorizes. The
message is:
not vectorized: relevant phi not supported: matched_21 = PHI <_20(6),
1(5)>
The "new" code is:
```
bool func_is_also_not_vectorized (char*& cur, context& ctx) {
bool matched = true;
constexpr char str1[] =
"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa";
if (!(std::distance(cur, ctx.end) < 50)) {return false;}
for(int i = 0; i < 50; i += 1) {
if (str1[i] != cur[i]) { matched = false; }
}
if (matched) {
cur += 50;
}
return matched;
}
```
The interesting part is. If I only replace `r.matched` with `matched`
and introduce the proper variable. The vectorizer is already failing.
