1 files changed, 71 insertions, 0 deletions
diff --git a/llvm/lib/Target/X86/X86MCInstLower.cpp b/llvm/lib/Target/X86/X86MCInstLower.cpp
index 6ed9a533d51..9de2d18e0be 100644
--- a/llvm/lib/Target/X86/X86MCInstLower.cpp
+++ b/llvm/lib/Target/X86/X86MCInstLower.cpp
@@ -1680,6 +1680,77 @@ void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
   case X86::TLS_base_addr64:
     return LowerTlsAddr(MCInstLowering, *MI);
 
+  // Loading/storing mask pairs requires two kmov operations. The second one of these
+  // needs a 2 byte displacement relative to the specified address (with 32 bit spill
+  // size). The pairs of 1bit masks up to 16 bit masks all use the same spill size,
+  // they all are stored using MASKPAIR16STORE, loaded using MASKPAIR16LOAD.
+  //
+  // The displacement value might wrap around in theory, thus the asserts in both
+  // cases.
+  case X86::MASKPAIR16LOAD: {
+    int64_t Disp = MI->getOperand(1 + X86::AddrDisp).getImm();
+    assert(Disp >= 0 && Disp <= INT32_MAX - 2 && "Unexpected displacement");
+    const X86RegisterInfo *RI =
+      MF->getSubtarget<X86Subtarget>().getRegisterInfo();
+    unsigned Reg = MI->getOperand(0).getReg();
+    unsigned Reg0 = RI->getSubReg(Reg, X86::sub_mask_0);
+    unsigned Reg1 = RI->getSubReg(Reg, X86::sub_mask_1);
+
+    // Load the first mask register
+    MCInstBuilder MIB = MCInstBuilder(X86::KMOVWkm);
+    MIB.addReg(Reg0);
+    for (int i = 0; i < X86::AddrNumOperands; ++i) {
+      auto Op = MCInstLowering.LowerMachineOperand(MI, MI->getOperand(1 + i));
+      MIB.addOperand(Op.getValue());
+    }
+    EmitAndCountInstruction(MIB);
+
+    // Load the second mask register of the pair
+    MIB = MCInstBuilder(X86::KMOVWkm);
+    MIB.addReg(Reg1);
+    for (int i = 0; i < X86::AddrNumOperands; ++i) {
+      if (i == X86::AddrDisp) {
+        MIB.addImm(Disp + 2);
+      } else {
+        auto Op = MCInstLowering.LowerMachineOperand(MI, MI->getOperand(1 + i));
+        MIB.addOperand(Op.getValue());
+      }
+    }
+    EmitAndCountInstruction(MIB);
+    return;
+  }
+
+  case X86::MASKPAIR16STORE: {
+    int64_t Disp = MI->getOperand(X86::AddrDisp).getImm();
+    assert(Disp >= 0 && Disp <= INT32_MAX - 2 && "Unexpected displacement");
+    const X86RegisterInfo *RI =
+      MF->getSubtarget<X86Subtarget>().getRegisterInfo();
+    unsigned Reg = MI->getOperand(X86::AddrNumOperands).getReg();
+    unsigned Reg0 = RI->getSubReg(Reg, X86::sub_mask_0);
+    unsigned Reg1 = RI->getSubReg(Reg, X86::sub_mask_1);
+
+    // Store the first mask register
+    MCInstBuilder MIB = MCInstBuilder(X86::KMOVWmk);
+    for (int i = 0; i < X86::AddrNumOperands; ++i)
+      MIB.addOperand(MCInstLowering.LowerMachineOperand(MI, MI->getOperand(i)).getValue());
+    MIB.addReg(Reg0);
+    EmitAndCountInstruction(MIB);
+
+    // Store the second mask register of the pair
+    MIB = MCInstBuilder(X86::KMOVWmk);
+    for (int i = 0; i < X86::AddrNumOperands; ++i) {
+      if (i == X86::AddrDisp) {
+        MIB.addImm(Disp + 2);
+      } else {
+        auto Op = MCInstLowering.LowerMachineOperand(MI, MI->getOperand(0 + i));
+        MIB.addOperand(Op.getValue());
+      }
+    }
+    MIB.addReg(Reg1);
+    EmitAndCountInstruction(MIB);
+    return;
+  }
+
   case X86::MOVPC32r: {
     // This is a pseudo op for a two instruction sequence with a label, which
     // looks like: