[PATCH] D127115: [RFC][DAGCombine] Make sure combined nodes are added back to the worklist in topological order.
Fraser Cormack via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Tue Jun 7 07:38:59 PDT 2022
frasercrmck added inline comments.
================
Comment at: llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vxor-vp.ll:166
+; RV32-NEXT: vxor.vv v8, v8, v10, v0.t
+; RV32-NEXT: ret
+;
----------------
RKSimon wrote:
> @craig.topper Might all of these vrgather regressions be due to a missing combine for splat cases?
Something like that. I see that we have the initial pattern:
```
t9: v2i64 = insert_vector_elt undef:v2i64, Constant:i64<-1>, Constant:i64<0>
t10: v2i64 = vector_shuffle<0,0> t9, undef:v2i64
```
on RV64 that's combined:
```
Combining: t9: v2i64 = insert_vector_elt undef:v2i64, Constant:i64<-1>, Constant:i64<0>
... into: t14: v2i64 = BUILD_VECTOR Constant:i64<-1>, undef:i64
Combining: t14: v2i64 = BUILD_VECTOR Constant:i64<-1>, undef:i64
Combining: t10: v2i64 = vector_shuffle<0,0> t14, undef:v2i64
... into: t15: v2i64 = BUILD_VECTOR Constant:i64<-1>, Constant:i64<-1>
t15: v2i64 = BUILD_VECTOR Constant:i64<-1>, Constant:i64<-1>
t11: v2i64 = sub t5, t15
```
On RV32 it's:
```
Combining: t8: v2i64 = insert_vector_elt undef:v2i64, Constant:i64<-1>, Constant:i32<0>
... into: t15: v2i64 = BUILD_VECTOR Constant:i64<-1>, undef:i64
Combining: t15: v2i64 = BUILD_VECTOR Constant:i64<-1>, undef:i64
... into: t16: v2i64 = splat_vector Constant:i64<-1>
Combining: t16: v2i64 = splat_vector Constant:i64<-1>
Combining: t9: v2i64 = vector_shuffle<0,0> t16, undef:v2i64
t16: v2i64 = splat_vector Constant:i64<-1>
t9: v2i64 = vector_shuffle<0,0> t16, undef:v2i64
t10: v2i64 = sub t5, t9
```
I'm not able to say what //exactly// is going wrong, though. We're definitely missing combines for shuffles of splats, but do we really want to see `splat_vector` for fixed-length vectors here?
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D127115/new/
https://reviews.llvm.org/D127115
More information about the llvm-commits
mailing list