[clang] [llvm] [AArch64][SME] Remove immediate argument restriction for svldr and svstr (PR #68565)

Sam Tebbs via llvm-commits llvm-commits at lists.llvm.org
Tue Nov 7 08:20:52 PST 2023


================
@@ -4825,6 +4827,72 @@ SDValue AArch64TargetLowering::getPStateSM(SelectionDAG &DAG, SDValue Chain,
                      Mask);
 }
 
+SDValue LowerSMELdrStr(SDValue N, SelectionDAG &DAG, bool IsLoad) {
+  // Lower an SME LDR/STR ZA intrinsic to LDR_ZA_PSEUDO or STR_ZA.
+  // If the vector number is an immediate between 0 and 15 inclusive then we can
+  // put that directly into the immediate field of the instruction. If it's
+  // outside of that range then we modify the base and slice by the greatest
+  // multiple of 15 smaller than that number and put the remainder in the
+  // instruction field. If it's not an immediate then we modify the base and
+  // slice registers by that number and put 0 in the instruction.
+  SDLoc DL(N);
+
+  SDValue TileSlice = N->getOperand(2);
+  SDValue Base = N->getOperand(3);
+  SDValue VecNum = N->getOperand(4);
+  SDValue Remainder = DAG.getTargetConstant(0, DL, MVT::i32);
+
+  // true if the base and slice registers need to me modified
+  bool NeedsAdd = true;
+  if (auto ImmNode = dyn_cast<ConstantSDNode>(VecNum)) {
+    int Imm = ImmNode->getSExtValue();
+    if (Imm >= 0 && Imm <= 15) {
+      Remainder = DAG.getTargetConstant(Imm, DL, MVT::i32);
+      NeedsAdd = false;
+    } else {
+      Remainder = DAG.getTargetConstant(Imm % 15, DL, MVT::i32);
+      NeedsAdd = true;
+      VecNum = DAG.getConstant(Imm - (Imm % 15), DL, MVT::i32);
+    }
+  } else if (VecNum.getOpcode() == ISD::ADD) {
+    // If the vnum is an add, we can fold that add into the instruction if the
+    // operand is an immediate in range
+    if (auto ImmNode = dyn_cast<ConstantSDNode>(VecNum.getOperand(1))) {
+      int Imm = ImmNode->getSExtValue();
+      if (Imm >= 0 && Imm <= 15) {
----------------
SamTebbs33 wrote:

My latest two commits do this, which improves the codegen as well since we properly fold successive loads/stores with adds of large immediates. See the test [here](https://media.tenor.com/vYhPH4gA184AAAAC/vroom-car.gif). Thanks for the idea!

https://github.com/llvm/llvm-project/pull/68565


More information about the llvm-commits mailing list