[PATCH] D125546: [RISCV] Use tail agnostic if inserting subvector/element at the end of the vector.
Roger Ferrer Ibanez via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Thu May 19 00:17:28 PDT 2022
rogfer01 added inline comments.
================
Comment at: llvm/lib/Target/RISCV/RISCVISelLowering.cpp:4429
+ int PolicyVal = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED;
+ // We could use tail agnostic if it's inserting to the latest of a vector.
+ if (VecVT.isFixedLengthVector() && isa<ConstantSDNode>(Idx) &&
----------------
I'd use "at the end of the vector" (I think "latest" is about time)
================
Comment at: llvm/lib/Target/RISCV/RISCVISelLowering.cpp:5360
+ if (VecVT.isFixedLengthVector() &&
+ (OrigIdx + SubVecVT.getVectorNumElements() + 1) ==
+ VecVT.getVectorNumElements())
----------------
I'm a bit confused here, probably I misunderstood what we mean by the end of the vector.
In the following testcase taken from `fixed-vectors-fp-shuffles.ll`
```lang=llvm
define void @insert_v32i1_v8i1_16(<32 x i1>* %vp, <8 x i1>* %svp) #0 {
%v = load <32 x i1>, <32 x i1>* %vp, align 4
%sv = load <8 x i1>, <8 x i1>* %svp, align 1
%c = call <32 x i1> @llvm.experimental.vector.insert.v32i1.v8i1(<32 x i1> %v, <8 x i1> %sv, i64 16)
store <32 x i1> %c, <32 x i1>* %vp, align 4
ret void
}
declare <32 x i1> @llvm.experimental.vector.insert.v32i1.v8i1(<32 x i1> %v, <8 x i1> %sv, i64 %idx)
```
Once the mask types are promoted to integer types we have: `OrigIdx == 2`, `VecVT == v4i8`, and `SubVecVT == v1i8`. My impression is that the end of the vector in this case would be when `OrigIdx == 3`. Does this align with your expectation or my notion of "end" is incorrect?
The condition check does `2 + 1 + 1 == 4` so we make this tail agnostic. Perhaps, not sure, the last `+ 1` is not needed (it was needed for the insert element case because you were checking the index in the vector).
================
Comment at: llvm/lib/Target/RISCV/RISCVInstrInfoVVLPatterns.td:1843
-def SDTRVVSlide : SDTypeProfile<1, 5, [
+// We use vsludeup with different tail policy to implement insert subvector or
+// element operations.
----------------
Typo `vslideup`
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D125546/new/
https://reviews.llvm.org/D125546
More information about the llvm-commits
mailing list