[RISCV] Combine a gather to a larger element type

[preames]

If we have a gather load whose indices correspond to valid offsets for a gather with element type twice that our source, we can reduce the number of indices and perform the operation at the larger element type.

This is generally profitable since we half VL - and these operations are linear in VL. This may require some additional VL/VTYPE toggles, but this appears to be worthwhile on the whole.

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Changes

If we have a gather load whose indices correspond to valid offsets for a gather with element type twice that our source, we can reduce the number of indices and perform the operation at the larger element type.

This is generally profitable since we half VL - and these operations are linear in VL. This may require some additional VL/VTYPE toggles, but this appears to be worthwhile on the whole.

---

Full diff: https://github.com/llvm/llvm-project/pull/66694.diff

2 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVISelLowering.cpp (+73)
• (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll (+16-16)

diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index e74d184c0a35d04..ae2bac69f396682 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -13589,6 +13589,52 @@ static bool matchIndexAsShuffle(EVT VT, SDValue Index, SDValue Mask,
   return ActiveLanes.all();
 }
 
+/// Match the index of a gather or scatter operation as an operation
+/// with twice the element width and half the number of elements.  This is
+/// generally profitable (if legal) because these operations are linear
+/// in VL, so even if we cause some extract VTYPE/VL toggles, we still
+/// come out ahead.
+static bool matchIndexAsWiderOp(EVT VT, SDValue Index, SDValue Mask,
+                                Align BaseAlign, const RISCVSubtarget &ST) {
+  if (!ISD::isConstantSplatVectorAllOnes(Mask.getNode()))
+    return false;
+  if (!ISD::isBuildVectorOfConstantSDNodes(Index.getNode()))
+    return false;
+
+  // Attempt a doubling.  If we can use a element type 4x or 8x in
+  // size, this will happen via multiply iterations of the transform.
+  const unsigned NumElems = VT.getVectorNumElements();
+  if (NumElems % 2 != 0)
+    return false;
+
+  const unsigned ElementSize = VT.getScalarStoreSize();
+  const unsigned WiderElementSize = ElementSize * 2;
+  if (WiderElementSize > ST.getELen()/8)
+    return false;
+
+  if (!ST.enableUnalignedVectorMem() && BaseAlign < WiderElementSize)
+    return false;
+
+  for (unsigned i = 0; i < Index->getNumOperands(); i++) {
+    // TODO: We've found an active bit of UB, and could be
+    // more aggressive here if desired.
+    if (Index->getOperand(i)->isUndef())
+      return false;
+    // TODO: This offset check is too strict if we support fully
+    // misaligned memory operations.
+    uint64_t C = Index->getConstantOperandVal(i);
+    if (C % ElementSize != 0)
+      return false;
+    if (i % 2 == 0)
+      continue;
+    uint64_t Last = Index->getConstantOperandVal(i-1);
+    if (C != Last + ElementSize)
+      return false;
+  }
+  return true;
+}
+
+
 SDValue RISCVTargetLowering::PerformDAGCombine(SDNode *N,
                                                DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
@@ -14020,6 +14066,33 @@ SDValue RISCVTargetLowering::PerformDAGCombine(SDNode *N,
         DAG.getVectorShuffle(VT, DL, Load, DAG.getUNDEF(VT), ShuffleMask);
       return DAG.getMergeValues({Shuffle, Load.getValue(1)}, DL);
     }
+
+    if (MGN->getExtensionType() == ISD::NON_EXTLOAD &&
+        matchIndexAsWiderOp(VT, Index, MGN->getMask(),
+                            MGN->getMemOperand()->getBaseAlign(), Subtarget)) {
+      SmallVector<SDValue> NewIndices;
+      for (unsigned i = 0; i < Index->getNumOperands(); i += 2)
+        NewIndices.push_back(Index.getOperand(i));
+      EVT IndexVT = Index.getValueType()
+        .getHalfNumVectorElementsVT(*DAG.getContext());
+      Index = DAG.getBuildVector(IndexVT, DL, NewIndices);
+
+      unsigned ElementSize = VT.getScalarStoreSize();
+      EVT WideScalarVT = MVT::getIntegerVT(ElementSize * 8 * 2);
+      EVT WideVT = VT.changeVectorElementType(WideScalarVT)
+        .getHalfNumVectorElementsVT(*DAG.getContext());
+      SDValue Passthru = DAG.getBitcast(WideVT, MGN->getPassThru());
+      SDValue Mask = DAG.getSplat(WideVT.changeVectorElementType(MVT::i1), DL,
+                                  DAG.getConstant(1, DL, MVT::i1));
+
+      SDValue Gather =
+        DAG.getMaskedGather(DAG.getVTList(WideVT, MVT::Other), WideVT, DL,
+                            {MGN->getChain(), Passthru, Mask, MGN->getBasePtr(),
+                             Index, ScaleOp},
+                            MGN->getMemOperand(), IndexType, ISD::NON_EXTLOAD);
+      SDValue Result = DAG.getBitcast(VT, Gather.getValue(0));
+      return DAG.getMergeValues({Result, Gather.getValue(1)}, DL);
+    }
     break;
   }
   case ISD::MSCATTER:{
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll
index ac5c11ca88df51a..130d2c7613b326b 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll
@@ -13024,19 +13024,19 @@ define <4 x i32> @mgather_narrow_edge_case(ptr %base) {
 define <8 x i16> @mgather_strided_2xSEW(ptr %base) {
 ; RV32-LABEL: mgather_strided_2xSEW:
 ; RV32:       # %bb.0:
-; RV32-NEXT:    lui a1, %hi(.LCPI107_0)
-; RV32-NEXT:    addi a1, a1, %lo(.LCPI107_0)
-; RV32-NEXT:    vsetivli zero, 8, e16, m1, ta, ma
-; RV32-NEXT:    vle8.v v9, (a1)
+; RV32-NEXT:    vsetivli zero, 4, e8, mf4, ta, ma
+; RV32-NEXT:    vid.v v8
+; RV32-NEXT:    vsll.vi v9, v8, 3
+; RV32-NEXT:    vsetvli zero, zero, e32, m1, ta, ma
 ; RV32-NEXT:    vluxei8.v v8, (a0), v9
 ; RV32-NEXT:    ret
 ;
 ; RV64V-LABEL: mgather_strided_2xSEW:
 ; RV64V:       # %bb.0:
-; RV64V-NEXT:    lui a1, %hi(.LCPI107_0)
-; RV64V-NEXT:    addi a1, a1, %lo(.LCPI107_0)
-; RV64V-NEXT:    vsetivli zero, 8, e16, m1, ta, ma
-; RV64V-NEXT:    vle8.v v9, (a1)
+; RV64V-NEXT:    vsetivli zero, 4, e8, mf4, ta, ma
+; RV64V-NEXT:    vid.v v8
+; RV64V-NEXT:    vsll.vi v9, v8, 3
+; RV64V-NEXT:    vsetvli zero, zero, e32, m1, ta, ma
 ; RV64V-NEXT:    vluxei8.v v8, (a0), v9
 ; RV64V-NEXT:    ret
 ;
@@ -13141,19 +13141,19 @@ define <8 x i16> @mgather_strided_2xSEW(ptr %base) {
 define <8 x i16> @mgather_gather_2xSEW(ptr %base) {
 ; RV32-LABEL: mgather_gather_2xSEW:
 ; RV32:       # %bb.0:
-; RV32-NEXT:    lui a1, %hi(.LCPI108_0)
-; RV32-NEXT:    addi a1, a1, %lo(.LCPI108_0)
-; RV32-NEXT:    vsetivli zero, 8, e16, m1, ta, ma
-; RV32-NEXT:    vle8.v v9, (a1)
+; RV32-NEXT:    lui a1, 82176
+; RV32-NEXT:    addi a1, a1, 1024
+; RV32-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
+; RV32-NEXT:    vmv.s.x v9, a1
 ; RV32-NEXT:    vluxei8.v v8, (a0), v9
 ; RV32-NEXT:    ret
 ;
 ; RV64V-LABEL: mgather_gather_2xSEW:
 ; RV64V:       # %bb.0:
-; RV64V-NEXT:    lui a1, %hi(.LCPI108_0)
-; RV64V-NEXT:    addi a1, a1, %lo(.LCPI108_0)
-; RV64V-NEXT:    vsetivli zero, 8, e16, m1, ta, ma
-; RV64V-NEXT:    vle8.v v9, (a1)
+; RV64V-NEXT:    lui a1, 82176
+; RV64V-NEXT:    addiw a1, a1, 1024
+; RV64V-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
+; RV64V-NEXT:    vmv.s.x v9, a1
 ; RV64V-NEXT:    vluxei8.v v8, (a0), v9
 ; RV64V-NEXT:    ret
 ;

[topperc]

I would avoid changeVectorElementType. If VT is representable as an MVT and WideVT isn't it will crash. The MVT won't have a Type* and we would need an LLVMContext to create the EVT for the WideVT.

[preames]

What would you think of explicitly constructing an EVT from VT, and then using this idiom? Alternatively, do you have something better to recommend?

Edit: Also, shouldn't the no-such MVT case be disallowed by the element width less than ELEN check done inside the helper?

[topperc]

That doesn't help. VT already is an EVT, but EVT internally is an MVT or a Type*. The constructor uses MVT when possible.

You can use roughly EVT::getVectorVT(*DAG.getContext(), WideScalarVT, VT.getVectorElementCount()).

Or we should add a context argument to changeVectorElementType. I'm tired of finding this bug.

[topperc]

LGTM