Fix VLDMQ and VSTMQ instructions to use the correct encoding and address modes.

These instructions are only needed for codegen, so I've removed all the
explicit encoding bits for now; they should be set in the same way as the for
VLDMD and VSTMD whenever we add encodings for VFP.  The use of addrmode5
requires that the instructions be custom-selected so that the number of
registers can be set in the AM5Opc value.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@99309 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 59 insertions, 30 deletions

diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index e96e2fac7a..88306633db 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -745,7 +745,9 @@ storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
     } else {
       AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQ)).
                      addReg(SrcReg, getKillRegState(isKill))
-                     .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
+                     .addFrameIndex(FI)
+                     .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4))
+                     .addMemOperand(MMO));
     }
   }
 }
@@ -793,7 +795,9 @@ loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
                      .addMemOperand(MMO));
     } else {
       AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQ), DestReg)
-                     .addFrameIndex(FI).addImm(0).addMemOperand(MMO));
+                     .addFrameIndex(FI)
+                     .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4))
+                     .addMemOperand(MMO));
     }
   }
 }
diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index d0d940addc..7d48663189 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -1701,6 +1701,35 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
       ResNode = SelectARMIndexedLoad(N);
     if (ResNode)
       return ResNode;
+
+    // VLDMQ must be custom-selected for "v2f64 load" to set the AM5Opc value.
+    if (Subtarget->hasVFP2() &&
+        N->getValueType(0).getSimpleVT().SimpleTy == MVT::v2f64) {
+      SDValue Chain = N->getOperand(0);
+      SDValue AM5Opc =
+        CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
+      SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
+      SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
+      SDValue Ops[] = { N->getOperand(1), AM5Opc, Pred, PredReg, Chain };
+      return CurDAG->getMachineNode(ARM::VLDMQ, dl, MVT::v2f64, MVT::Other,
+                                    Ops, 5);
+    }
+    // Other cases are autogenerated.
+    break;
+  }
+  case ISD::STORE: {
+    // VSTMQ must be custom-selected for "v2f64 store" to set the AM5Opc value.
+    if (Subtarget->hasVFP2() &&
+        N->getOperand(1).getValueType().getSimpleVT().SimpleTy == MVT::v2f64) {
+      SDValue Chain = N->getOperand(0);
+      SDValue AM5Opc =
+        CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::ia, 4), MVT::i32);
+      SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32);
+      SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
+      SDValue Ops[] = { N->getOperand(1), N->getOperand(2),
+                        AM5Opc, Pred, PredReg, Chain };
+      return CurDAG->getMachineNode(ARM::VSTMQ, dl, MVT::Other, Ops, 6);
+    }
     // Other cases are autogenerated.
     break;
   }
diff --git a/lib/Target/ARM/ARMInstrFormats.td b/lib/Target/ARM/ARMInstrFormats.td
index 43dcf32a7d..b311125d1c 100644
--- a/lib/Target/ARM/ARMInstrFormats.td
+++ b/lib/Target/ARM/ARMInstrFormats.td
@@ -1503,12 +1503,6 @@ class NI<dag oops, dag iops, InstrItinClass itin, string opc, string asm,
            pattern> {
 }
 
-class NI4<dag oops, dag iops, InstrItinClass itin, string opc,
-          string asm, list<dag> pattern>
-  : NeonXI<oops, iops, AddrMode4, IndexModeNone, itin, opc, asm, "",
-          pattern> {
-}
-
 class NLdSt<bit op23, bits<2> op21_20, bits<4> op11_8, bits<4> op7_4,
             dag oops, dag iops, InstrItinClass itin,
             string opc, string dt, string asm, string cstr, list<dag> pattern>
diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 34e011a30b..55a658993d 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -115,19 +115,20 @@ def h64imm : Operand<i64> {
 // NEON load / store instructions
 //===----------------------------------------------------------------------===//
 
+let mayLoad = 1 in {
 // Use vldmia to load a Q register as a D register pair.
-// This is equivalent to VLDMD except that it has a Q register operand.
-def VLDMQ : NI4<(outs QPR:$dst), (ins addrmode4:$addr), IIC_fpLoadm,
-                "vldmia", "$addr, ${dst:dregpair}",
-                [(set QPR:$dst, (v2f64 (load addrmode4:$addr)))]> {
-  let Inst{27-25} = 0b110;
-  let Inst{24}    = 0; // P bit
-  let Inst{23}    = 1; // U bit
-  let Inst{20}    = 1;
-  let Inst{11-8}  = 0b1011;
-}
+// This is equivalent to VLDMD except that it has a Q register operand
+// instead of a pair of D registers.
+def VLDMQ
+  : AXDI5<(outs QPR:$dst), (ins addrmode5:$addr, pred:$p),
+          IndexModeNone, IIC_fpLoadm,
+          "vldm${addr:submode}${p}\t${addr:base}, ${dst:dregpair}", "", []>;
+def VLDMQ_UPD
+  : AXDI5<(outs QPR:$dst, GPR:$wb), (ins addrmode5:$addr, pred:$p),
+          IndexModeUpd, IIC_fpLoadm,
+          "vldm${addr:submode}${p}\t${addr:base}!, ${dst:dregpair}",
+          "$addr.base = $wb", []>;
 
-let mayLoad = 1 in {
 // Use vld1 to load a Q register as a D register pair.
 // This alternative to VLDMQ allows an alignment to be specified.
 // This is equivalent to VLD1q64 except that it has a Q register operand.
@@ -140,19 +141,20 @@ def VLD1q_UPD
           "${dst:dregpair}, $addr$offset", "$addr.addr = $wb", []>;
 } // mayLoad = 1
 
+let mayStore = 1 in {
 // Use vstmia to store a Q register as a D register pair.
-// This is equivalent to VSTMD except that it has a Q register operand.
-def VSTMQ : NI4<(outs), (ins QPR:$src, addrmode4:$addr), IIC_fpStorem,
-                "vstmia", "$addr, ${src:dregpair}",
-                [(store (v2f64 QPR:$src), addrmode4:$addr)]> {
-  let Inst{27-25} = 0b110;
-  let Inst{24}    = 0; // P bit
-  let Inst{23}    = 1; // U bit
-  let Inst{20}    = 0;
-  let Inst{11-8}  = 0b1011;
-}
+// This is equivalent to VSTMD except that it has a Q register operand
+// instead of a pair of D registers.
+def VSTMQ
+  : AXDI5<(outs), (ins QPR:$src, addrmode5:$addr, pred:$p),
+          IndexModeNone, IIC_fpStorem,
+          "vstm${addr:submode}${p}\t${addr:base}, ${src:dregpair}", "", []>;
+def VSTMQ_UPD
+  : AXDI5<(outs GPR:$wb), (ins QPR:$src, addrmode5:$addr, pred:$p),
+          IndexModeUpd, IIC_fpStorem,
+          "vstm${addr:submode}${p}\t${addr:base}!, ${src:dregpair}",
+          "$addr.base = $wb", []>;
 
-let mayStore = 1 in {
 // Use vst1 to store a Q register as a D register pair.
 // This alternative to VSTMQ allows an alignment to be specified.
 // This is equivalent to VST1q64 except that it has a Q register operand.