[llvm] [X86][SelectionDAG] Fix the Gather's base and index by modifying the Scale value (PR #137813)

[Simon Pilgrim via llvm-commits]

================
@@ -56652,13 +56652,56 @@ static SDValue combineGatherScatter(SDNode *N, SelectionDAG &DAG,
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
   if (DCI.isBeforeLegalize()) {
+    // Attempt to move shifted index into the address scale, allows further
+    // index truncation below.
+    if (Index.getOpcode() == ISD::SHL && isa<ConstantSDNode>(Scale)) {
+      unsigned BitWidth = Index.getScalarValueSizeInBits();
+      unsigned ScaleAmt = Scale->getAsZExtVal();
+      assert(isPowerOf2_32(ScaleAmt) && "Scale must be a power of 2");
+      unsigned Log2ScaleAmt = Log2_32(ScaleAmt);
+      unsigned MaskBits = BitWidth - Log2ScaleAmt;
+      APInt DemandedBits = APInt::getLowBitsSet(BitWidth, MaskBits);
+      if (TLI.SimplifyDemandedBits(Index, DemandedBits, DCI)) {
+        if (N->getOpcode() != ISD::DELETED_NODE)
+          DCI.AddToWorklist(N);
+        return SDValue(N, 0);
+      }
+      if (auto MinShAmt = DAG.getValidMinimumShiftAmount(Index)) {
+        if (*MinShAmt >= 1 && (*MinShAmt + Log2ScaleAmt) < 4 &&
+            DAG.ComputeNumSignBits(Index.getOperand(0)) > 1) {
+          SDValue ShAmt = Index.getOperand(1);
+          SDValue NewShAmt =
+              DAG.getNode(ISD::SUB, DL, ShAmt.getValueType(), ShAmt,
+                          DAG.getConstant(1, DL, ShAmt.getValueType()));
+          SDValue NewIndex = DAG.getNode(ISD::SHL, DL, Index.getValueType(),
+                                         Index.getOperand(0), NewShAmt);
+          SDValue NewScale =
+              DAG.getConstant(ScaleAmt * 2, DL, Scale.getValueType());
+          return rebuildGatherScatter(GorS, NewIndex, Base, NewScale, DAG);
+        }
+      }
+    }
     unsigned IndexWidth = Index.getScalarValueSizeInBits();
 
+    // If the index is a left shift, \ComputeNumSignBits we are recomputing the
+    // number of sign bits from the shifted value. We are trying to enable the
+    // optimization in which we can shrink indices if they are larger than
+    // 32-bits. Using the existing fold techniques implemented below.
+    unsigned ComputeNumSignBits = DAG.ComputeNumSignBits(Index);
+    if (Index.getOpcode() == ISD::SHL) {
+      if (auto MinShAmt = DAG.getValidMinimumShiftAmount(Index)) {
+        if (DAG.ComputeNumSignBits(Index.getOperand(0)) > 1) {
+          ComputeNumSignBits += *MinShAmt;
+        }
+      }
+    }
+
     // Shrink indices if they are larger than 32-bits.
     // Only do this before legalize types since v2i64 could become v2i32.
     // FIXME: We could check that the type is legal if we're after legalize
     // types, but then we would need to construct test cases where that happens.
-    if (IndexWidth > 32 && DAG.ComputeNumSignBits(Index) > (IndexWidth - 32)) {
+    // \ComputeNumSignBits value is recomputed for the shift Index
+    if (IndexWidth > 32 && ComputeNumSignBits > (IndexWidth - 32)) {
----------------
RKSimon wrote:

Probably better if you can move everything inside `if (IndexWidth > 32)`

https://github.com/llvm/llvm-project/pull/137813