[clang] [llvm] Add support for flag output operand "=@cc" for SystemZ. (PR #125970)

[Ulrich Weigand via llvm-commits]

================
@@ -8030,7 +8077,431 @@ SDValue SystemZTargetLowering::combineBSWAP(
   return SDValue();
 }
 
+// Combine IPM sequence for flag output operands.
+static bool combineSRL_IPM_CCMask(SDValue &CCReg, int &CCValid, int &CCMask) {
+  // Convert CCVal to CCMask and update it along with  CCValid.
+  const auto convertCCValToCCMask = [&CCMask, &CCValid](int CCVal) {
+    bool Invert = false;
+    if (CCMask == SystemZ::CCMASK_CMP_NE)
+      Invert = !Invert;
+    if (CCMask == SystemZ::CCMASK_CMP_EQ || CCMask == SystemZ::CCMASK_CMP_NE) {
+      CCMask = CCMaskForSystemZCCVal(CCVal);
+      if (Invert)
+        CCMask ^= SystemZ::CCMASK_ANY;
+      CCValid = SystemZ::CCMASK_ANY;
+      return true;
+    } else if (CCMask == SystemZ::CCMASK_CMP_LT) {
+      // CC in range [0, CCVal).
+      CCMask = ((~0U << (4 - CCVal)) & SystemZ::CCMASK_ANY);
+      CCValid = SystemZ::CCMASK_ANY;
+      return true;
+    } else if (CCMask == SystemZ::CCMASK_CMP_GT) {
+      // CC in range (CCVal, 3].
+      CCMask = (~(~0U << (3 - CCVal))) & SystemZ::CCMASK_ANY;
+      CCValid = SystemZ::CCMASK_ANY;
+      return true;
+    }
+    return false;
+  };
+  // Check (SRL (IPM (CC))) and update CCReg to combine.
+  const auto isSRL_IPM_CCSequence = [&CCReg](SDNode *N) {
+    if (!N || N->getOpcode() != ISD::SRL)
+      return false;
+    auto *SRLCount = dyn_cast<ConstantSDNode>(N->getOperand(1));
+    if (!SRLCount || SRLCount->getZExtValue() != SystemZ::IPM_CC)
+      return false;
+    auto *IPM = N->getOperand(0).getNode();
+    if (!IPM || IPM->getOpcode() != SystemZISD::IPM)
+      return false;
+    auto *IPMOp0 = IPM->getOperand(0).getNode();
+    if (!IPMOp0 || IPMOp0->getNumOperands() < 2)
+      return false;
+    auto *RN = dyn_cast<RegisterSDNode>(IPMOp0->getOperand(1));
+    // Check if operand 1 is SystemZ::CC. Also, it avoids srl/ipm/tbegin and
+    // srl/ipm/tend kind of sequences.
----------------
uweigand wrote:

Why is this check needed?   The logic should be correct no matter the operand of `IPM`.  In fact, we might *want* to optimize these sequences for `tbegin` or `tend`  (or, more importantly, some of the vector intrinsics).

https://github.com/llvm/llvm-project/pull/125970