[RISCV] Narrow indices of fixed vector gather/scatter nodes

Doing so allows the use of smaller constants overall, and may allow (for some small vector constants) avoiding the constant pool entirely.  This can result in extra VTYPE toggles if we get unlucky.

This was reviewed under PR llvm#66405.

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -11633,15 +11633,39 @@ static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
   return combineSelectAndUseCommutative(N, DAG, /*AllOnes*/ false, Subtarget);
 }
 
-// According to the property that indexed load/store instructions
-// zero-extended their indices, \p narrowIndex tries to narrow the type of index
-// operand if it is matched to pattern (shl (zext x to ty), C) and bits(x) + C <
-// bits(ty).
+/// According to the property that indexed load/store instructions zero-extend
+/// their indices, try to narrow the type of index operand.
 static bool narrowIndex(SDValue &N, ISD::MemIndexType IndexType, SelectionDAG &DAG) {
   if (isIndexTypeSigned(IndexType))
     return false;
 
-  if (N.getOpcode() != ISD::SHL || !N->hasOneUse())
+  if (!N->hasOneUse())
+    return false;
+
+  EVT VT = N.getValueType();
+  SDLoc DL(N);
+
+  // In general, what we're doing here is seeing if we can sink a truncate to
+  // a smaller element type into the expression tree building our index.
+  // TODO: We can generalize this and handle a bunch more cases if useful.
+
+  // Narrow a buildvector to the narrowest element type.  This requires less
+  // work and less register pressure at high LMUL, and creates smaller constants
+  // which may be cheaper to materialize.
+  if (ISD::isBuildVectorOfConstantSDNodes(N.getNode())) {
+    KnownBits Known = DAG.computeKnownBits(N);
+    unsigned ActiveBits = std::max(8u, Known.countMaxActiveBits());
+    LLVMContext &C = *DAG.getContext();
+    EVT ResultVT = EVT::getIntegerVT(C, ActiveBits).getRoundIntegerType(C);
+    if (ResultVT.bitsLT(VT.getVectorElementType())) {
+      N = DAG.getNode(ISD::TRUNCATE, DL,
+                      VT.changeVectorElementType(ResultVT), N);
+      return true;
+    }
+  }
+
+  // Handle the pattern (shl (zext x to ty), C) and bits(x) + C < bits(ty).
+  if (N.getOpcode() != ISD::SHL)
     return false;
 
   SDValue N0 = N.getOperand(0);
@@ -11656,7 +11680,6 @@ static bool narrowIndex(SDValue &N, ISD::MemIndexType IndexType, SelectionDAG &D
   if (!ISD::isConstantSplatVector(N1.getNode(), ShAmt))
     return false;;
 
-  SDLoc DL(N);
   SDValue Src = N0.getOperand(0);
   EVT SrcVT = Src.getValueType();
   unsigned SrcElen = SrcVT.getScalarSizeInBits();

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll

@@ -13027,17 +13027,17 @@ define <8 x i16> @mgather_strided_2xSEW(ptr %base) {
 ; 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:    vle32.v v10, (a1)
-; RV32-NEXT:    vluxei32.v v8, (a0), v10
+; RV32-NEXT:    vle8.v v9, (a1)
+; 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:    vle64.v v12, (a1)
-; RV64V-NEXT:    vluxei64.v v8, (a0), v12
+; RV64V-NEXT:    vle8.v v9, (a1)
+; RV64V-NEXT:    vluxei8.v v8, (a0), v9
 ; RV64V-NEXT:    ret
 ;
 ; RV64ZVE32F-LABEL: mgather_strided_2xSEW:
@@ -13144,17 +13144,17 @@ define <8 x i16> @mgather_gather_2xSEW(ptr %base) {
 ; 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:    vle32.v v10, (a1)
-; RV32-NEXT:    vluxei32.v v8, (a0), v10
+; RV32-NEXT:    vle8.v 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:    vle64.v v12, (a1)
-; RV64V-NEXT:    vluxei64.v v8, (a0), v12
+; RV64V-NEXT:    vle8.v v9, (a1)
+; RV64V-NEXT:    vluxei8.v v8, (a0), v9
 ; RV64V-NEXT:    ret
 ;
 ; RV64ZVE32F-LABEL: mgather_gather_2xSEW: