CodeGen: extend RHS when splitting ATOMIC_CMP_SWAP_WITH_SUCCESS.

If the operation's type has been promoted during type legalization, we
need to account for the fact that the high bits of the comparison
operand are likely unspecified.

The LHS is usually zero-extended, but MIPS sign extends it, so we have
to be slightly careful.

Patch by Simon Dardis.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@264296 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/Target/TargetLowering.h

@@ -1172,6 +1172,11 @@ public:
     return nullptr;
   }
 
+  /// Returns true if the platform's atomic operations are sign extended.
+  virtual bool hasSignExtendedAtomicOps() const {
+    return false;
+  }
+
   /// Returns true if we should normalize
   /// select(N0&N1, X, Y) => select(N0, select(N1, X, Y), Y) and
   /// select(N0|N1, X, Y) => select(N0, select(N1, X, Y, Y)) if it is likely

lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

@@ -3118,10 +3118,25 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
         cast<AtomicSDNode>(Node)->getFailureOrdering(),
         cast<AtomicSDNode>(Node)->getSynchScope());
 
-    SDValue Success = DAG.getSetCC(SDLoc(Node), Node->getValueType(1),
+    SDValue LHS = Res;
-                                   Res, Node->getOperand(2), ISD::SETEQ);
+    SDValue RHS = Node->getOperand(1);
 
-    Results.push_back(Res.getValue(0));
+    EVT AtomicType = cast<AtomicSDNode>(Node)->getMemoryVT();
+    EVT OuterType = Node->getValueType(0);
+    if (TLI.hasSignExtendedAtomicOps()) {
+      LHS = DAG.getNode(ISD::AssertSext, dl, OuterType, Res,
+                        DAG.getValueType(AtomicType));
+      RHS = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, OuterType,
+                        Node->getOperand(2), DAG.getValueType(AtomicType));
+    } else {
+      LHS = DAG.getNode(ISD::AssertZext, dl, OuterType, Res, DAG.getValueType(AtomicType));
+      RHS = DAG.getNode(ISD::ZERO_EXTEND, dl, OuterType, Node->getOperand(2));
+    }
+
+    SDValue Success =
+        DAG.getSetCC(dl, Node->getValueType(1), LHS, RHS, ISD::SETEQ);
+
+    Results.push_back(LHS.getValue(0));
     Results.push_back(Success);
     Results.push_back(Res.getValue(1));
     break;

lib/Target/Mips/MipsISelLowering.h

@@ -238,6 +238,10 @@ namespace llvm {
     bool isCheapToSpeculateCttz() const override;
     bool isCheapToSpeculateCtlz() const override;
 
+    bool hasSignExtendedAtomicOps() const override {
+      return true;
+    }
+
     void LowerOperationWrapper(SDNode *N,
                                SmallVectorImpl<SDValue> &Results,
                                SelectionDAG &DAG) const override;

test/CodeGen/Mips/atomic.ll

@@ -357,10 +357,17 @@ entry:
 ; NO-SEB-SEH:    sll     $[[R18:[0-9]+]], $[[R17]], 24
 ; NO-SEB-SEH:    sra     $[[R19:[0-9]+]], $[[R18]], 24
 
+; FIXME: -march=mips produces a redundant sign extension here...
+; NO-SEB-SEH:    sll     $[[R20:[0-9]+]], $5, 24
+; NO-SEB-SEH:    sra     $[[R20]], $[[R20]], 24
+
 ; HAS-SEB-SEH:   seb     $[[R19:[0-9]+]], $[[R17]]
 
-; ALL:           xor     $[[R20:[0-9]+]], $[[R19]], $5
+; FIXME: ...Leading to this split check.
-; ALL:           sltiu   $2, $[[R20]], 1
+; NO-SEB-SEH:    xor     $[[R21:[0-9]+]], $[[R19]], $[[R20]]
+; HAS-SEB-SEH:   xor     $[[R21:[0-9]+]], $[[R19]], $5
+
+; ALL: sltiu   $2, $[[R21]], 1
 }
 
 ; Check one i16 so that we cover the seh sign extend
@@ -407,6 +414,49 @@ entry:
 ; MIPS32R2:      seh     $2, $[[R16]]
 }
 
+; Test that the i16 return value from cmpxchg is recognised as signed,
+; so that setCC doesn't end up comparing an unsigned value to a signed
+; value.
+; The rest of the functions here are testing the atomic expansion, so
+; we just match the end of the function.
+define {i16, i1} @foo(i16* %addr, i16 %l, i16 %r, i16 %new) {
+  %desired = add i16 %l, %r
+  %res = cmpxchg i16* %addr, i16 %desired, i16 %new seq_cst seq_cst
+  ret {i16, i1} %res
+
+; ALL-LABEL: foo
+; MIPSR6:        addu    $[[R2:[0-9]+]], $[[R1:[0-9]+]], $[[R0:[0-9]+]]
+; NOT-MICROMIPS: addu    $[[R2:[0-9]+]], $[[R1:[0-9]+]], $[[R0:[0-9]+]]
+; MICROMIPS:     addu16  $[[R2:[0-9]+]], $[[R1:[0-9]+]], $[[R0:[0-9]+]]
+
+; ALL:           sync
+
+; ALL:           andi    $[[R3:[0-9]+]], $[[R2]], 65535
+; ALL:       $[[BB0:[A-Z_0-9]+]]:
+; ALL:           ll      $[[R4:[0-9]+]], 0($[[R5:[0-9]+]])
+; ALL:           and     $[[R6:[0-9]+]], $[[R4]], $
+; ALL:           and     $[[R7:[0-9]+]], $[[R4]], $
+; ALL:           or      $[[R8:[0-9]+]], $[[R7]], $
+; ALL:           sc      $[[R8]], 0($[[R5]])
+; NOT-MICROMIPS: beqz    $[[R8]], $[[BB0]]
+; MICROMIPS:     beqzc   $[[R8]], $[[BB0]]
+; MIPSR6:        beqzc   $[[R8]], $[[BB0]]
+
+; ALL:           srlv    $[[R9:[0-9]+]], $[[R6]], $
+
+; NO-SEB-SEH:    sll     $[[R10:[0-9]+]], $[[R9]], 16
+; NO-SEB-SEH:    sra     $[[R11:[0-9]+]], $[[R10]], 16
+
+; NO-SEB-SEH:    sll     $[[R12:[0-9]+]], $[[R2]], 16
+; NO-SEB-SEH:    sra     $[[R13:[0-9]+]], $[[R12]], 16
+
+; HAS-SEB-SEH:   seh     $[[R11:[0-9]+]], $[[R9]]
+; HAS-SEB-SEH:   seh     $[[R13:[0-9]+]], $[[R2]]
+
+; ALL:           xor     $[[R12:[0-9]+]], $[[R11]], $[[R13]]
+; ALL:           sltiu   $3, $[[R12]], 1
+; ALL:           sync
+}
 
 @countsint = common global i32 0, align 4