https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123343
Bug ID: 123343
Summary: Loop unrolling before vectorization produces
suboptimal RISC-V code
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: chenzhongyao.hit at gmail dot com
Target Milestone: ---
Target: riscv
testcase is extracted from x264.
https://godbolt.org/z/doreGfo79
The inner loop gets unrolled before the vectorizer sees it, leads to
inefficient code with vslide/vcompress permutation instructions.
Potential approaches (seeking feedback):
1. Skip unrolling - Don't unroll innermost loops with memory loads + type
promotion when vectorization is enabled.
2. SLP reorganization - Improve SLP vectorization to detect and reorganize
interleaved loads back into contiguous accesses after unrolling.
Or other better approaches ?
