Add patterns to use post-increment addressing for Neon VST1-lane instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@126477 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index a506cffdba3483f7a8493843c0fe50d28f024024..f0d5a7d7c2e7bfec3bb98410ec75abea5f699487 100644 (file)
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -126,6 +126,7 @@ public:
   bool SelectAddrMode5(SDValue N, SDValue &Base,
                        SDValue &Offset);
   bool SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,SDValue &Align);
+  bool SelectAddrMode6Offset(SDNode *Op, SDValue N, SDValue &Offset);
 
   bool SelectAddrModePC(SDValue N, SDValue &Offset, SDValue &Label);
 
@@ -886,6 +887,20 @@ bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,
   return true;
 }
 
+bool ARMDAGToDAGISel::SelectAddrMode6Offset(SDNode *Op, SDValue N,
+                                            SDValue &Offset) {
+  LSBaseSDNode *LdSt = cast<LSBaseSDNode>(Op);
+  ISD::MemIndexedMode AM = LdSt->getAddressingMode();
+  if (AM != ISD::POST_INC)
+    return false;
+  Offset = N;
+  if (ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N)) {
+    if (NC->getZExtValue() * 8 == LdSt->getMemoryVT().getSizeInBits())
+      Offset = CurDAG->getRegister(0, MVT::i32);
+  }
+  return true;
+}
+
 bool ARMDAGToDAGISel::SelectAddrModePC(SDValue N,
                                        SDValue &Offset, SDValue &Label) {
   if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {

diff --git a/lib/Target/ARM/ARMInstrInfo.td b/lib/Target/ARM/ARMInstrInfo.td
index c827ce3da97cb4bea79cb95d5d4579aa1fef21a4..6e3fe2e039f57ca52e448f0857fdd337732de5ac 100644 (file)
--- a/lib/Target/ARM/ARMInstrInfo.td
+++ b/lib/Target/ARM/ARMInstrInfo.td
@@ -561,7 +561,9 @@ def addrmode6 : Operand<i32>,
   let EncoderMethod = "getAddrMode6AddressOpValue";
 }
 
-def am6offset : Operand<i32> {
+def am6offset : Operand<i32>,
+                ComplexPattern<i32, 1, "SelectAddrMode6Offset",
+                               [], [SDNPWantRoot]> {
   let PrintMethod = "printAddrMode6OffsetOperand";
   let MIOperandInfo = (ops GPR);
   let EncoderMethod = "getAddrMode6OffsetOpValue";

diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 1e2e5504e662923c34d7545c1ba2e211346e49b3..dc3d63e26ef5e2865e32c5aa4e2b0bd18fe3aaf2 100644 (file)
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -1402,31 +1402,42 @@ def : Pat<(store (extractelt (v2f32 DPR:$src), imm:$lane), addrmode6:$addr),
 def : Pat<(store (extractelt (v4f32 QPR:$src), imm:$lane), addrmode6:$addr),
           (VST1LNq32Pseudo addrmode6:$addr, QPR:$src, imm:$lane)>;
 
-let mayStore = 1, neverHasSideEffects = 1, hasExtraSrcRegAllocReq = 1 in {
-
 // ...with address register writeback:
-class VST1LNWB<bits<4> op11_8, bits<4> op7_4, string Dt>
+class VST1LNWB<bits<4> op11_8, bits<4> op7_4, string Dt, ValueType Ty,
+               PatFrag StoreOp, SDNode ExtractOp>
   : NLdStLn<1, 0b00, op11_8, op7_4, (outs GPR:$wb),
           (ins addrmode6:$Rn, am6offset:$Rm,
            DPR:$Vd, nohash_imm:$lane), IIC_VST1lnu, "vst1", Dt,
           "\\{$Vd[$lane]\\}, $Rn$Rm",
-          "$Rn.addr = $wb", []>;
+          "$Rn.addr = $wb",
+          [(set GPR:$wb, (StoreOp (ExtractOp (Ty DPR:$Vd), imm:$lane),
+                                  addrmode6:$Rn, am6offset:$Rm))]>;
+class VST1QLNWBPseudo<ValueType Ty, PatFrag StoreOp, SDNode ExtractOp>
+  : VSTQLNWBPseudo<IIC_VST1lnu> {
+  let Pattern = [(set GPR:$wb, (StoreOp (ExtractOp (Ty QPR:$src), imm:$lane),
+                                        addrmode6:$addr, am6offset:$offset))];
+}
 
-def VST1LNd8_UPD  : VST1LNWB<0b0000, {?,?,?,0}, "8"> {
+def VST1LNd8_UPD  : VST1LNWB<0b0000, {?,?,?,0}, "8", v8i8, post_truncsti8,
+                             NEONvgetlaneu> {
   let Inst{7-5} = lane{2-0};
 }
-def VST1LNd16_UPD : VST1LNWB<0b0100, {?,?,0,?}, "16"> {
+def VST1LNd16_UPD : VST1LNWB<0b0100, {?,?,0,?}, "16", v4i16, post_truncsti16,
+                             NEONvgetlaneu> {
   let Inst{7-6} = lane{1-0};
   let Inst{4}   = Rn{5};
 }
-def VST1LNd32_UPD : VST1LNWB<0b1000, {?,0,?,?}, "32"> {
+def VST1LNd32_UPD : VST1LNWB<0b1000, {?,0,?,?}, "32", v2i32, post_store,
+                             extractelt> {
   let Inst{7}   = lane{0};
   let Inst{5-4} = Rn{5-4};
 }
 
-def VST1LNq8Pseudo_UPD  : VSTQLNWBPseudo<IIC_VST1lnu>;
-def VST1LNq16Pseudo_UPD : VSTQLNWBPseudo<IIC_VST1lnu>;
-def VST1LNq32Pseudo_UPD : VSTQLNWBPseudo<IIC_VST1lnu>;
+def VST1LNq8Pseudo_UPD  : VST1QLNWBPseudo<v16i8, post_truncsti8, NEONvgetlaneu>;
+def VST1LNq16Pseudo_UPD : VST1QLNWBPseudo<v8i16, post_truncsti16,NEONvgetlaneu>;
+def VST1LNq32Pseudo_UPD : VST1QLNWBPseudo<v4i32, post_store, extractelt>;
+
+let mayStore = 1, neverHasSideEffects = 1, hasExtraSrcRegAllocReq = 1 in {
 
 //   VST2LN   : Vector Store (single 2-element structure from one lane)
 class VST2LN<bits<4> op11_8, bits<4> op7_4, string Dt>

diff --git a/test/CodeGen/ARM/vstlane.ll b/test/CodeGen/ARM/vstlane.ll
index 6cc052bbeb1cad75eac25af61c5646110349aa83..d1bc15ad576da687cbafe01ebf7330ed3319ea31 100644 (file)
--- a/test/CodeGen/ARM/vstlane.ll
+++ b/test/CodeGen/ARM/vstlane.ll
@@ -10,6 +10,19 @@ define void @vst1lanei8(i8* %A, <8 x i8>* %B) nounwind {
        ret void
 }
 
+;Check for a post-increment updating store.
+define void @vst1lanei8_update(i8** %ptr, <8 x i8>* %B) nounwind {
+;CHECK: vst1lanei8_update:
+;CHECK: vst1.8 {d16[3]}, [r2]!
+       %A = load i8** %ptr
+       %tmp1 = load <8 x i8>* %B
+       %tmp2 = extractelement <8 x i8> %tmp1, i32 3
+       store i8 %tmp2, i8* %A, align 8
+       %tmp3 = getelementptr i8* %A, i32 1
+       store i8* %tmp3, i8** %ptr
+       ret void
+}
+
 define void @vst1lanei16(i16* %A, <4 x i16>* %B) nounwind {
 ;CHECK: vst1lanei16:
 ;Check the alignment value.  Max for this instruction is 16 bits:
@@ -66,6 +79,19 @@ define void @vst1laneQi32(i32* %A, <4 x i32>* %B) nounwind {
        ret void
 }
 
+;Check for a post-increment updating store.
+define void @vst1laneQi32_update(i32** %ptr, <4 x i32>* %B) nounwind {
+;CHECK: vst1laneQi32_update:
+;CHECK: vst1.32 {d17[1]}, [r1, :32]!
+       %A = load i32** %ptr
+       %tmp1 = load <4 x i32>* %B
+       %tmp2 = extractelement <4 x i32> %tmp1, i32 3
+       store i32 %tmp2, i32* %A, align 8
+       %tmp3 = getelementptr i32* %A, i32 1
+       store i32* %tmp3, i32** %ptr
+       ret void
+}
+
 define void @vst1laneQf(float* %A, <4 x float>* %B) nounwind {
 ;CHECK: vst1laneQf:
 ;CHECK: vst1.32 {d17[1]}, [r0]