Hi,
This patch series improves SLP reduction vectorization in two areas:
1. **IFN commutative operand swap**: Extends the existing commutative swap
logic to internal functions like .COND_ADD. If-conversion can generate
COND_ADD with reduction operands at different positions (arg1 vs arg2),
which previously caused SLP rejection. The fix uses
first_commutative_argument()
to recognize and swap IFN operands when needed.
2. **Single-lane SLP fallback**: Moves the SLP discovery limit check to only
guard chain analysis, allowing single-lane reductions to vectorize even
when the limit is exhausted. This prevents missed optimizations in loops
with many independent reductions.
For 731.astcenc_r (SPEC 2026) single-copy:
- EMR : +7.1% (patch 1), +3.8% (patch 2) → ~11% total
- Znver5: +2.5% (patch 1), +1.4% (patch 2) → ~3.9% total
No regressions observed in other benchmarks.
Bootstrapped and tested on x86_64-linux-gnu. OK for trunk?
Thanks,
Lili.
Lili Cui (2):
tree-optimization/vect: Allow commutative operand swap for IFN in SLP
reduction
tree-optimization/vect: Allow single-lane SLP fallback when limit is
exhausted
gcc/testsuite/gcc.dg/vect/slp-reduc-15.c | 81 ++++++++++++++++++++++++
gcc/tree-vect-slp.cc | 17 +++--
2 files changed, 92 insertions(+), 6 deletions(-)
create mode 100644 gcc/testsuite/gcc.dg/vect/slp-reduc-15.c
--
2.34.1
