[llvm] [NVPTX] Add 3-operand fmin/fmax DAGCombines (PR #159729)

[Lewis Crawford via llvm-commits]

================
@@ -5316,6 +5320,56 @@ static SDValue PerformFADDCombine(SDNode *N,
   return PerformFADDCombineWithOperands(N, N1, N0, DCI, OptLevel);
 }
 
+/// Get 3-input version of a 2-input min/max opcode
+static NVPTXISD::NodeType getMinMax3Opcode(unsigned MinMax2Opcode) {
+  switch (MinMax2Opcode) {
+  case ISD::FMAXNUM:
+  case ISD::FMAXIMUMNUM:
+    return NVPTXISD::FMAXNUM3;
+  case ISD::FMINNUM:
+  case ISD::FMINIMUMNUM:
+    return NVPTXISD::FMINNUM3;
+  case ISD::FMAXIMUM:
+    return NVPTXISD::FMAXIMUM3;
+  case ISD::FMINIMUM:
+    return NVPTXISD::FMINIMUM3;
+  default:
+    llvm_unreachable("Invalid 2-input min/max opcode");
+  }
+}
+
+/// PerformFMinMaxCombine - Combine (fmaxnum (fmaxnum a, b), c) into
+/// (fmaxnum3 a, b, c). Also covers other llvm min/max intrinsics.
+static SDValue PerformFMinMaxCombine(SDNode *N,
+                                     TargetLowering::DAGCombinerInfo &DCI,
+                                     unsigned PTXVersion, unsigned SmVersion) {
+
+  // 3-input min/max requires PTX 8.8+ and SM_100+, and only supports f32s
+  EVT VT = N->getValueType(0);
+  if (VT != MVT::f32 || PTXVersion < 88 || SmVersion < 100)
+    return SDValue();
+
+  SDValue Op0 = N->getOperand(0);
+  SDValue Op1 = N->getOperand(1);
+  unsigned MinMaxOp2 = N->getOpcode();
+  NVPTXISD::NodeType MinMaxOp3 = getMinMax3Opcode(MinMaxOp2);
+
+  if (Op0.getOpcode() == MinMaxOp2 && Op0.hasOneUse()) {
+    // (maxnum (maxnum a, b), c) -> (maxnum3 a, b, c)
+    SDValue A = Op0.getOperand(0);
+    SDValue B = Op0.getOperand(1);
+    SDValue C = Op1;
+    return DCI.DAG.getNode(MinMaxOp3, SDLoc(N), VT, A, B, C, N->getFlags());
+  } else if (Op1->getOpcode() == MinMaxOp2 && Op1->hasOneUse()) {
----------------
LewisCrawford wrote:

There was no good reason, so I've changed it to use `.` in both places now.
Both `->` and `.` are equivalent in this case, as `*` returns the inner SDNode from an SDValue, and `getOperand` on an SDValue just calls `getOperand` on the inner SDNode anyway, but I've changed it to use the same approach everywhere now for consistency.

https://github.com/llvm/llvm-project/pull/159729