[LoongArch] Custom legalize i32 operations for LA64 to reduce signed extensions (#93811)

2 files changed, 21 insertions, 4 deletions

diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index 8a87c82..51384f2 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -121,6 +121,8 @@ LoongArchTargetLowering::LoongArchTargetLowering(const TargetMachine &TM,
   // Set operations for LA64 only.
 
   if (Subtarget.is64Bit()) {
+    setOperationAction(ISD::ADD, MVT::i32, Custom);
+    setOperationAction(ISD::SUB, MVT::i32, Custom);
     setOperationAction(ISD::SHL, MVT::i32, Custom);
     setOperationAction(ISD::SRA, MVT::i32, Custom);
     setOperationAction(ISD::SRL, MVT::i32, Custom);
@@ -1723,6 +1725,18 @@ static SDValue customLegalizeToWOp(SDNode *N, SelectionDAG &DAG, int NumOp,
   return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), NewRes);
 }
 
+// Converts the given 32-bit operation to a i64 operation with signed extension
+// semantic to reduce the signed extension instructions.
+static SDValue customLegalizeToWOpWithSExt(SDNode *N, SelectionDAG &DAG) {
+  SDLoc DL(N);
+  SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
+  SDValue NewOp1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
+  SDValue NewWOp = DAG.getNode(N->getOpcode(), DL, MVT::i64, NewOp0, NewOp1);
+  SDValue NewRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, NewWOp,
+                               DAG.getValueType(MVT::i32));
+  return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes);
+}
+
 // Helper function that emits error message for intrinsics with/without chain
 // and return a UNDEF or and the chain as the results.
 static void emitErrorAndReplaceIntrinsicResults(
@@ -1846,6 +1860,12 @@ void LoongArchTargetLowering::ReplaceNodeResults(
   switch (N->getOpcode()) {
   default:
     llvm_unreachable("Don't know how to legalize this operation");
+  case ISD::ADD:
+  case ISD::SUB:
+    assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
+           "Unexpected custom legalisation");
+    Results.push_back(customLegalizeToWOpWithSExt(N, DAG));
+    break;
   case ISD::UDIV:
   case ISD::UREM:
     assert(VT == MVT::i32 && Subtarget.is64Bit() &&
diff --git a/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td
index 35ea9f0..66bd74e 100644
--- a/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td
+++ b/llvm/lib/Target/LoongArch/LoongArchInstrInfo.td
@@ -441,7 +441,7 @@ def simm16_lsl16 : Operand<GRLenVT>,
 // A 32-bit signed immediate expressible with a pair of `addu16i.d + addi` for
 // use in additions.
 def simm32_hi16_lo12: Operand<GRLenVT>, ImmLeaf<GRLenVT, [{
-  return isShiftedInt<16, 16>(Imm - SignExtend64<12>(Imm));
+  return !isInt<12>(Imm) && isShiftedInt<16, 16>(Imm - SignExtend64<12>(Imm));
 }]>;
 
 def BareSymbol : AsmOperandClass {
@@ -1106,11 +1106,8 @@ foreach Idx = 1...3 in {
 
 let Predicates = [IsLA64] in {
 def : PatGprGpr<add, ADD_D>;
-def : PatGprGpr_32<add, ADD_W>;
 def : PatGprImm<add, ADDI_D, simm12>;
-def : PatGprImm_32<add, ADDI_W, simm12>;
 def : PatGprGpr<sub, SUB_D>;
-def : PatGprGpr_32<sub, SUB_W>;
 def : PatGprGpr<sdiv, DIV_D>;
 def : PatGprGpr_32<sdiv, DIV_W>;
 def : PatGprGpr<udiv, DIV_DU>;