[DAG] Fold logic of zero-checks to multiplication for MinSize

[TelGome]

Addresses #164501

[llvmbot]

@llvm/pr-subscribers-llvm-selectiondag

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

Author: None (TelGome)

Changes

Addresses #164501

---

Full diff: https://github.com/llvm/llvm-project/pull/171805.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+60)
• (added) llvm/test/CodeGen/RISCV/fold-zero-check-minsize.ll (+57)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
old mode 100644
new mode 100755
index 6a99d4e29b64f..7861f0937459d
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -631,6 +631,7 @@ namespace {
 
     SDValue foldAddToAvg(SDNode *N, const SDLoc &DL);
     SDValue foldSubToAvg(SDNode *N, const SDLoc &DL);
+    SDValue foldLogicSetCCToMul(SDNode *N, const SDLoc &DL);
 
     SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
                                          unsigned HiOp);
@@ -6648,6 +6649,59 @@ static unsigned getMinMaxOpcodeForFP(SDValue Operand1, SDValue Operand2,
   return ISD::DELETED_NODE;
 }
 
+// Fold the following patterns for small integers in -Oz mode.
+// (X == 0) || (Y == 0) --> (X * Y) == 0
+// (X != 0) && (Y != 0) --> (X * Y) != 0
+SDValue DAGCombiner::foldLogicSetCCToMul(SDNode *N, const SDLoc &DL) {
+  if (OptLevel == CodeGenOptLevel::None ||
+      !DAG.getMachineFunction().getFunction().hasMinSize())
+    return SDValue();
+
+  unsigned Opcode = N->getOpcode();
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+
+  ISD::CondCode ExpectedCC;
+  if (Opcode == ISD::OR) {
+    ExpectedCC = ISD::SETEQ;
+  } else if (Opcode == ISD::AND) {
+    ExpectedCC = ISD::SETNE;
+  } else {
+    return SDValue();
+  }
+
+  if (N0.getOpcode() != ISD::SETCC || N1.getOpcode() != ISD::SETCC)
+    return SDValue();
+
+  SDValue A = N0.getOperand(0);
+  SDValue B = N1.getOperand(0);
+  SDValue C0 = N0.getOperand(1);
+  SDValue C1 = N1.getOperand(1);
+  ISD::CondCode CC0 = cast<CondCodeSDNode>(N0.getOperand(2))->get();
+  ISD::CondCode CC1 = cast<CondCodeSDNode>(N1.getOperand(2))->get();
+
+  if (CC0 != ExpectedCC || CC1 != ExpectedCC || !isNullConstant(C0) ||
+      !isNullConstant(C1) || A.getValueType() != B.getValueType() ||
+      !A.getValueType().isScalarInteger())
+    return SDValue();
+
+  unsigned BitWidth = A.getValueSizeInBits();
+  KnownBits KnownA = DAG.computeKnownBits(A);
+  KnownBits KnownB = DAG.computeKnownBits(B);
+
+  if (KnownA.countMaxActiveBits() + KnownB.countMaxActiveBits() > BitWidth)
+    return SDValue();
+
+  SDNodeFlags Flags;
+  Flags.setNoUnsignedWrap(true);
+  Flags.setNoSignedWrap(true);
+
+  SDValue Mul = DAG.getNode(ISD::MUL, DL, A.getValueType(), A, B, Flags);
+
+  return DAG.getSetCC(DL, N->getValueType(0), Mul,
+                      DAG.getConstant(0, DL, A.getValueType()), ExpectedCC);
+}
+
 static SDValue foldAndOrOfSETCC(SDNode *LogicOp, SelectionDAG &DAG) {
   using AndOrSETCCFoldKind = TargetLowering::AndOrSETCCFoldKind;
   assert(
@@ -7555,6 +7609,9 @@ SDValue DAGCombiner::visitAND(SDNode *N) {
   if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0), APInt::getAllOnes(BitWidth)))
     return DAG.getConstant(0, DL, VT);
 
+  if (SDValue R = foldLogicSetCCToMul(N, DL))
+    return R;
+
   if (SDValue R = foldAndOrOfSETCC(N, DAG))
     return R;
 
@@ -8520,6 +8577,9 @@ SDValue DAGCombiner::visitOR(SDNode *N) {
   if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
     return N1;
 
+  if (SDValue R = foldLogicSetCCToMul(N, DL))
+    return R;
+
   if (SDValue R = foldAndOrOfSETCC(N, DAG))
     return R;
 
diff --git a/llvm/test/CodeGen/RISCV/fold-zero-check-minsize.ll b/llvm/test/CodeGen/RISCV/fold-zero-check-minsize.ll
new file mode 100755
index 0000000000000..19fc9fb544fb4
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/fold-zero-check-minsize.ll
@@ -0,0 +1,57 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv64 -mattr=+m -verify-machineinstrs < %s \
+; RUN:   | FileCheck %s
+
+define i1 @foldLogicSetCCToMul0(i16 zeroext %a, i16 zeroext %b) minsize {
+; CHECK-LABEL: foldLogicSetCCToMul0:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    mul a0, a0, a1
+; CHECK-NEXT:    seqz a0, a0
+; CHECK-NEXT:    ret
+entry:
+  %cmp1 = icmp eq i16 %a, 0
+  %cmp2 = icmp eq i16 %b, 0
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @foldLogicSetCCToMul1(i16 zeroext %a, i16 zeroext %b) minsize {
+; CHECK-LABEL: foldLogicSetCCToMul1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    mul a0, a0, a1
+; CHECK-NEXT:    snez a0, a0
+; CHECK-NEXT:    ret
+entry:
+  %cmp1 = icmp ne i16 %a, 0
+  %cmp2 = icmp ne i16 %b, 0
+  %and = and i1 %cmp1, %cmp2
+  ret i1 %and
+}
+
+define i1 @foldLogicSetCCToMul2(i64 %a, i64 %b) minsize {
+; CHECK-LABEL: foldLogicSetCCToMul2:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    seqz a0, a0
+; CHECK-NEXT:    seqz a1, a1
+; CHECK-NEXT:    or a0, a0, a1
+; CHECK-NEXT:    ret
+entry:
+  %cmp1 = icmp eq i64 %a, 0
+  %cmp2 = icmp eq i64 %b, 0
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
+
+define i1 @foldLogicSetCCToMul3(i16 zeroext %a, i16 zeroext %b) {
+; CHECK-LABEL: foldLogicSetCCToMul3:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    seqz a0, a0
+; CHECK-NEXT:    seqz a1, a1
+; CHECK-NEXT:    or a0, a0, a1
+; CHECK-NEXT:    ret
+entry:
+  %cmp1 = icmp eq i16 %a, 0
+  %cmp2 = icmp eq i16 %b, 0
+  %or = or i1 %cmp1, %cmp2
+  ret i1 %or
+}
\ No newline at end of file

[lukel97]

Should the title say [DAGCombine]?

[arsenm]

Needs negative tests with mismatched compare operand and opcode, and non-0 constants

[TelGome]

Should the title say [DAGCombine]?

Yes, that makes sense. The issue is labeled llvm:instcombine, so I named it accordingly. But I'm fine with updating the title to [DAGCombine] if that's more appropriate.

[TelGome]

Needs negative tests with mismatched compare operand and opcode, and non-0 constants

I’ve uploaded a commit to address the mentioned issues. Please review when you have a moment.

[dtcxzyw]

Do we have a TLI hook querying whether the target supports predicated instructions?
This optimization looks less profitable on ARM64 since it has conditional compares.