alexfh added a comment.
No reduced test case, but one observation is that most samples in the slow version are around this (preexisting code):
// The reduction nodes (stored in UserIgnoreList) also should stay scalar.
for (Value *V : VL) {
if (is_contained(UserIgnoreList, V)) {
0.00 │ 7aa: testq %rdx, %rdx ▒
│ leaq -536(%rbp), %r15 ▒
│ ; llvm-project/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:4296 ▒
│ ; for (Value *V : VL) { ◆
│ ↓ je 80a ▒
│ movq -80(%rbp), %rax ▒
│ movq 2088(%rax), %rbx ▒
│ movq 2096(%rax), %rcx ▒
│ leaq (%rbx,%rcx,8), %r10 ▒
│ leaq (,%rcx,8), %r8 ▒
│ ↓ jmp 7e8 ▒
│ ; llvm-project/llvm/include/llvm/ADT/STLExtras.h:1671 ▒
│ ; return std::find(adl_begin(Range), adl_end(Range), Element) != adl_end(Range); ▒
│ 7d6: cmpq %r10, %rax ▒
│ ; llvm-project/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:4297 ▒
│ ; if (is_contained(UserIgnoreList, V)) { ▒
│ ↓ jne a98 ▒
│ ; llvm-project/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:4296 ▒
│ ; for (Value *V : VL) { ▒
0.12 │ 7df: addq $8, %rsi ▒
0.01 │ cmpq %r9, %rsi ▒
│ ↓ je 80a ▒
│ 7e8: movq %rbx, %rax ▒
│ testq %rcx, %rcx ▒
│ ; include/c++/v1/__algorithm/find.h:24 ▒
│ ; for (; __first != __last; ++__first) ▒
│ ↑ je 7d6 ▒
0.10 │ movq (%rsi), %rdi ▒
│ movq %r8, %rdx ▒
│ movq %rbx, %rax ▒
│ ; include/c++/v1/__algorithm/find.h:25 ▒
│ ; if (*__first == __value_) ▒
99.02 │ 7f9: cmpq %rdi, (%rax) ▒
│ ↑ je 7d6 ▒
│ ; include/c++/v1/__algorithm/find.h:24 ▒
│ ; for (; __first != __last; ++__first) ▒
│ addq $8, %rax ▒
│ addq $-8, %rdx ▒
0.03 │ ↑ jne 7f9 ▒
0.28 │ ↑ jmp 7df ▒
│ ; llvm-project/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:4308 ▒
│ ; auto *VL0 = cast<Instruction>(S.OpValue); ▒
│ 80a: movq -136(%rbp), %r14 ▒
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D114171/new/
https://reviews.llvm.org/D114171