Fixed MC encoding for index_align for VLD1/VST1 (single element from one lane) for size 32

llvm-svn: 131085

4 files changed, 60 insertions, 2 deletions

diff --git a/llvm/lib/Target/ARM/ARMCodeEmitter.cpp b/llvm/lib/Target/ARM/ARMCodeEmitter.cpp
index d0c0e952512..8397e3bf1d7 100644
--- a/llvm/lib/Target/ARM/ARMCodeEmitter.cpp
+++ b/llvm/lib/Target/ARM/ARMCodeEmitter.cpp
@@ -221,6 +221,9 @@ namespace {
       const { return 0; }
     unsigned getAddrMode6AddressOpValue(const MachineInstr &MI, unsigned Op)
       const { return 0; }
+    unsigned getAddrMode6OneLane32AddressOpValue(const MachineInstr &MI,
+                                                 unsigned Op)
+      const { return 0; }
     unsigned getAddrMode6DupAddressOpValue(const MachineInstr &MI, unsigned Op)
       const { return 0; }
     unsigned getAddrMode6OffsetOpValue(const MachineInstr &MI, unsigned Op)
diff --git a/llvm/lib/Target/ARM/ARMInstrInfo.td b/llvm/lib/Target/ARM/ARMInstrInfo.td
index 2bfd76bfc5b..4e81401eb7e 100644
--- a/llvm/lib/Target/ARM/ARMInstrInfo.td
+++ b/llvm/lib/Target/ARM/ARMInstrInfo.td
@@ -588,6 +588,15 @@ def am6offset : Operand<i32>,
   let EncoderMethod = "getAddrMode6OffsetOpValue";
 }
 
+// Special version of addrmode6 to handle alignment encoding for VST1/VLD1
+// (single element from one lane) for size 32.
+def addrmode6oneL32 : Operand<i32>,
+                ComplexPattern<i32, 2, "SelectAddrMode6", [], [SDNPWantParent]>{
+  let PrintMethod = "printAddrMode6Operand";
+  let MIOperandInfo = (ops GPR:$addr, i32imm);
+  let EncoderMethod = "getAddrMode6OneLane32AddressOpValue";
+}
+
 // Special version of addrmode6 to handle alignment encoding for VLD-dup
 // instructions, specifically VLD4-dup.
 def addrmode6dup : Operand<i32>,
diff --git a/llvm/lib/Target/ARM/ARMInstrNEON.td b/llvm/lib/Target/ARM/ARMInstrNEON.td
index e34d69a44d9..2f8912f1ec5 100644
--- a/llvm/lib/Target/ARM/ARMInstrNEON.td
+++ b/llvm/lib/Target/ARM/ARMInstrNEON.td
@@ -531,6 +531,17 @@ class VLD1LN<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
                                          imm:$lane))]> {
   let Rm = 0b1111;
 }
+class VLD1LN32<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
+             PatFrag LoadOp>
+  : NLdStLn<1, 0b10, op11_8, op7_4, (outs DPR:$Vd),
+          (ins addrmode6oneL32:$Rn, DPR:$src, nohash_imm:$lane),
+          IIC_VLD1ln, "vld1", Dt, "\\{$Vd[$lane]\\}, $Rn",
+          "$src = $Vd",
+          [(set DPR:$Vd, (vector_insert (Ty DPR:$src),
+                                         (i32 (LoadOp addrmode6oneL32:$Rn)),
+                                         imm:$lane))]> {
+  let Rm = 0b1111;
+}
 class VLD1QLNPseudo<ValueType Ty, PatFrag LoadOp> : VLDQLNPseudo<IIC_VLD1ln> {
   let Pattern = [(set QPR:$dst, (vector_insert (Ty QPR:$src),
                                                (i32 (LoadOp addrmode6:$addr)),
@@ -544,7 +555,7 @@ def VLD1LNd16 : VLD1LN<0b0100, {?,?,0,?}, "16", v4i16, extloadi16> {
   let Inst{7-6} = lane{1-0};
   let Inst{4}   = Rn{4};
 }
-def VLD1LNd32 : VLD1LN<0b1000, {?,0,?,?}, "32", v2i32, load> {
+def VLD1LNd32 : VLD1LN32<0b1000, {?,0,?,?}, "32", v2i32, load> {
   let Inst{7} = lane{0};
   let Inst{5} = Rn{4};
   let Inst{4} = Rn{4};
@@ -1371,6 +1382,14 @@ class VST1LN<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
           [(StoreOp (ExtractOp (Ty DPR:$Vd), imm:$lane), addrmode6:$Rn)]> {
   let Rm = 0b1111;
 }
+class VST1LN32<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
+             PatFrag StoreOp, SDNode ExtractOp>
+  : NLdStLn<1, 0b00, op11_8, op7_4, (outs),
+          (ins addrmode6oneL32:$Rn, DPR:$Vd, nohash_imm:$lane),
+          IIC_VST1ln, "vst1", Dt, "\\{$Vd[$lane]\\}, $Rn", "",
+          [(StoreOp (ExtractOp (Ty DPR:$Vd), imm:$lane), addrmode6oneL32:$Rn)]> {
+  let Rm = 0b1111;
+}
 class VST1QLNPseudo<ValueType Ty, PatFrag StoreOp, SDNode ExtractOp>
   : VSTQLNPseudo<IIC_VST1ln> {
   let Pattern = [(StoreOp (ExtractOp (Ty QPR:$src), imm:$lane),
@@ -1386,7 +1405,8 @@ def VST1LNd16 : VST1LN<0b0100, {?,?,0,?}, "16", v4i16, truncstorei16,
   let Inst{7-6} = lane{1-0};
   let Inst{4}   = Rn{5};
 }
-def VST1LNd32 : VST1LN<0b1000, {?,0,?,?}, "32", v2i32, store, extractelt> {
+
+def VST1LNd32 : VST1LN32<0b1000, {?,0,?,?}, "32", v2i32, store, extractelt> {
   let Inst{7}   = lane{0};
   let Inst{5-4} = Rn{5-4};
 }
diff --git a/llvm/lib/Target/ARM/ARMMCCodeEmitter.cpp b/llvm/lib/Target/ARM/ARMMCCodeEmitter.cpp
index 10607b17c53..2b6d9fcfb33 100644
--- a/llvm/lib/Target/ARM/ARMMCCodeEmitter.cpp
+++ b/llvm/lib/Target/ARM/ARMMCCodeEmitter.cpp
@@ -273,6 +273,8 @@ public:
                                   SmallVectorImpl<MCFixup> &Fixups) const;
   unsigned getAddrMode6AddressOpValue(const MCInst &MI, unsigned Op,
                                       SmallVectorImpl<MCFixup> &Fixups) const;
+  unsigned getAddrMode6OneLane32AddressOpValue(const MCInst &MI, unsigned Op,
+                                        SmallVectorImpl<MCFixup> &Fixups) const;
   unsigned getAddrMode6DupAddressOpValue(const MCInst &MI, unsigned Op,
                                         SmallVectorImpl<MCFixup> &Fixups) const;
   unsigned getAddrMode6OffsetOpValue(const MCInst &MI, unsigned Op,
@@ -1178,6 +1180,30 @@ getAddrMode6AddressOpValue(const MCInst &MI, unsigned Op,
   return RegNo | (Align << 4);
 }
 
+/// getAddrMode6OneLane32AddressOpValue - Encode an addrmode6 register number
+/// along  with the alignment operand for use in VST1 and VLD1 with size 32.
+unsigned ARMMCCodeEmitter::
+getAddrMode6OneLane32AddressOpValue(const MCInst &MI, unsigned Op,
+                                    SmallVectorImpl<MCFixup> &Fixups) const {
+  const MCOperand &Reg = MI.getOperand(Op);
+  const MCOperand &Imm = MI.getOperand(Op + 1);
+
+  unsigned RegNo = getARMRegisterNumbering(Reg.getReg());
+  unsigned Align = 0;
+
+  switch (Imm.getImm()) {
+  default: break;
+  case 2:
+  case 4:
+  case 8:
+  case 16: Align = 0x00; break;
+  case 32: Align = 0x03; break;
+  }
+
+  return RegNo | (Align << 4);
+}
+
+
 /// getAddrMode6DupAddressOpValue - Encode an addrmode6 register number and
 /// alignment operand for use in VLD-dup instructions.  This is the same as
 /// getAddrMode6AddressOpValue except for the alignment encoding, which is