[Hexagon] Add support for nontemporal loads and stores on HVX

Patch by Michael Wu.

Differential Revision: https://reviews.llvm.org/D35104

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@307671 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp b/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp
index e5f49ca77a91263907c0c5d2031120922c114227..0163b2e2bdc46ee56598ba58ccdc723b775e7b7c 100644 (file)
--- a/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp
+++ b/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp
@@ -241,22 +241,31 @@ void HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, const SDLoc &dl) {
   case MVT::v32i16:
   case MVT::v16i32:
   case MVT::v8i64:
-    if (isAlignedMemNode(LD))
-      Opcode = IsValidInc ? Hexagon::V6_vL32b_pi : Hexagon::V6_vL32b_ai;
-    else
+    if (isAlignedMemNode(LD)) {
+      if (LD->isNonTemporal())
+        Opcode = IsValidInc ? Hexagon::V6_vL32b_nt_pi : Hexagon::V6_vL32b_nt_ai;
+      else
+        Opcode = IsValidInc ? Hexagon::V6_vL32b_pi : Hexagon::V6_vL32b_ai;
+    } else {
       Opcode = IsValidInc ? Hexagon::V6_vL32Ub_pi : Hexagon::V6_vL32Ub_ai;
+    }
     break;
   // 128B
   case MVT::v128i8:
   case MVT::v64i16:
   case MVT::v32i32:
   case MVT::v16i64:
-    if (isAlignedMemNode(LD))
-      Opcode = IsValidInc ? Hexagon::V6_vL32b_pi_128B
-                          : Hexagon::V6_vL32b_ai_128B;
-    else
+    if (isAlignedMemNode(LD)) {
+      if (LD->isNonTemporal())
+        Opcode = IsValidInc ? Hexagon::V6_vL32b_nt_pi_128B
+                            : Hexagon::V6_vL32b_nt_ai_128B;
+      else
+        Opcode = IsValidInc ? Hexagon::V6_vL32b_pi_128B
+                            : Hexagon::V6_vL32b_ai_128B;
+    } else {
       Opcode = IsValidInc ? Hexagon::V6_vL32Ub_pi_128B
                           : Hexagon::V6_vL32Ub_ai_128B;
+    }
     break;
   default:
     llvm_unreachable("Unexpected memory type in indexed load");
@@ -529,22 +538,31 @@ void HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, const SDLoc &dl) {
   case MVT::v32i16:
   case MVT::v16i32:
   case MVT::v8i64:
-    if (isAlignedMemNode(ST))
-      Opcode = IsValidInc ? Hexagon::V6_vS32b_pi : Hexagon::V6_vS32b_ai;
-    else
+    if (isAlignedMemNode(ST)) {
+      if (ST->isNonTemporal())
+        Opcode = IsValidInc ? Hexagon::V6_vS32b_nt_pi : Hexagon::V6_vS32b_nt_ai;
+      else
+        Opcode = IsValidInc ? Hexagon::V6_vS32b_pi : Hexagon::V6_vS32b_ai;
+    } else {
       Opcode = IsValidInc ? Hexagon::V6_vS32Ub_pi : Hexagon::V6_vS32Ub_ai;
+    }
     break;
   // 128B
   case MVT::v128i8:
   case MVT::v64i16:
   case MVT::v32i32:
   case MVT::v16i64:
-    if (isAlignedMemNode(ST))
-      Opcode = IsValidInc ? Hexagon::V6_vS32b_pi_128B
-                          : Hexagon::V6_vS32b_ai_128B;
-    else
+    if (isAlignedMemNode(ST)) {
+      if (ST->isNonTemporal())
+        Opcode = IsValidInc ? Hexagon::V6_vS32b_nt_pi_128B
+                            : Hexagon::V6_vS32b_nt_ai_128B;
+      else
+        Opcode = IsValidInc ? Hexagon::V6_vS32b_pi_128B
+                            : Hexagon::V6_vS32b_ai_128B;
+    } else {
       Opcode = IsValidInc ? Hexagon::V6_vS32Ub_pi_128B
                           : Hexagon::V6_vS32Ub_ai_128B;
+    }
     break;
   default:
     llvm_unreachable("Unexpected memory type in indexed store");

diff --git a/lib/Target/Hexagon/HexagonInstrInfo.cpp b/lib/Target/Hexagon/HexagonInstrInfo.cpp
index ff99a1047396e84077eb4abbda81a379a6639dad..c77c669f4ca75e6a8e2b9d9899c015274cf8fc22 100644 (file)
--- a/lib/Target/Hexagon/HexagonInstrInfo.cpp
+++ b/lib/Target/Hexagon/HexagonInstrInfo.cpp
@@ -250,15 +250,19 @@ unsigned HexagonInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
   case Hexagon::L2_loadri_io:
   case Hexagon::L2_loadrd_io:
   case Hexagon::V6_vL32b_ai:
+  case Hexagon::V6_vL32b_nt_ai:
   case Hexagon::V6_vL32b_ai_128B:
+  case Hexagon::V6_vL32b_nt_ai_128B:
   case Hexagon::V6_vL32Ub_ai:
   case Hexagon::V6_vL32Ub_ai_128B:
   case Hexagon::LDriw_pred:
   case Hexagon::LDriw_mod:
   case Hexagon::PS_vloadrq_ai:
   case Hexagon::PS_vloadrw_ai:
+  case Hexagon::PS_vloadrw_nt_ai:
   case Hexagon::PS_vloadrq_ai_128B:
-  case Hexagon::PS_vloadrw_ai_128B: {
+  case Hexagon::PS_vloadrw_ai_128B:
+  case Hexagon::PS_vloadrw_nt_ai_128B: {
     const MachineOperand OpFI = MI.getOperand(1);
     if (!OpFI.isFI())
       return 0;
@@ -2473,20 +2477,28 @@ bool HexagonInstrInfo::isValidOffset(unsigned Opcode, int Offset,
   switch (Opcode) {
   case Hexagon::PS_vstorerq_ai:
   case Hexagon::PS_vstorerw_ai:
+  case Hexagon::PS_vstorerw_nt_ai:
   case Hexagon::PS_vloadrq_ai:
   case Hexagon::PS_vloadrw_ai:
+  case Hexagon::PS_vloadrw_nt_ai:
   case Hexagon::V6_vL32b_ai:
   case Hexagon::V6_vS32b_ai:
+  case Hexagon::V6_vL32b_nt_ai:
+  case Hexagon::V6_vS32b_nt_ai:
   case Hexagon::V6_vL32Ub_ai:
   case Hexagon::V6_vS32Ub_ai:
     return isShiftedInt<4,6>(Offset);
 
   case Hexagon::PS_vstorerq_ai_128B:
   case Hexagon::PS_vstorerw_ai_128B:
+  case Hexagon::PS_vstorerw_nt_ai_128B:
   case Hexagon::PS_vloadrq_ai_128B:
   case Hexagon::PS_vloadrw_ai_128B:
+  case Hexagon::PS_vloadrw_nt_ai_128B:
   case Hexagon::V6_vL32b_ai_128B:
   case Hexagon::V6_vS32b_ai_128B:
+  case Hexagon::V6_vL32b_nt_ai_128B:
+  case Hexagon::V6_vS32b_nt_ai_128B:
   case Hexagon::V6_vL32Ub_ai_128B:
   case Hexagon::V6_vS32Ub_ai_128B:
     return isShiftedInt<4,7>(Offset);
@@ -3198,11 +3210,19 @@ int HexagonInstrInfo::getDotCurOp(const MachineInstr &MI) const {
     return Hexagon::V6_vL32b_cur_pi;
   case Hexagon::V6_vL32b_ai:
     return Hexagon::V6_vL32b_cur_ai;
+  case Hexagon::V6_vL32b_nt_pi:
+    return Hexagon::V6_vL32b_nt_cur_pi;
+  case Hexagon::V6_vL32b_nt_ai:
+    return Hexagon::V6_vL32b_nt_cur_ai;
   //128B
   case Hexagon::V6_vL32b_pi_128B:
     return Hexagon::V6_vL32b_cur_pi_128B;
   case Hexagon::V6_vL32b_ai_128B:
     return Hexagon::V6_vL32b_cur_ai_128B;
+  case Hexagon::V6_vL32b_nt_pi_128B:
+    return Hexagon::V6_vL32b_nt_cur_pi_128B;
+  case Hexagon::V6_vL32b_nt_ai_128B:
+    return Hexagon::V6_vL32b_nt_cur_ai_128B;
   }
   return 0;
 }
@@ -3215,11 +3235,19 @@ int HexagonInstrInfo::getNonDotCurOp(const MachineInstr &MI) const {
     return Hexagon::V6_vL32b_pi;
   case Hexagon::V6_vL32b_cur_ai:
     return Hexagon::V6_vL32b_ai;
+  case Hexagon::V6_vL32b_nt_cur_pi:
+    return Hexagon::V6_vL32b_nt_pi;
+  case Hexagon::V6_vL32b_nt_cur_ai:
+    return Hexagon::V6_vL32b_nt_ai;
   //128B
   case Hexagon::V6_vL32b_cur_pi_128B:
     return Hexagon::V6_vL32b_pi_128B;
   case Hexagon::V6_vL32b_cur_ai_128B:
     return Hexagon::V6_vL32b_ai_128B;
+  case Hexagon::V6_vL32b_nt_cur_pi_128B:
+    return Hexagon::V6_vL32b_nt_pi_128B;
+  case Hexagon::V6_vL32b_nt_cur_ai_128B:
+    return Hexagon::V6_vL32b_nt_ai_128B;
   }
   return 0;
 }

diff --git a/lib/Target/Hexagon/HexagonPatterns.td b/lib/Target/Hexagon/HexagonPatterns.td
index 44475d2e73661a9b72266aa84e3e773b14c021b2..ba98b8994937f6ee629d00da81070bcac8b1735b 100644 (file)
--- a/lib/Target/Hexagon/HexagonPatterns.td
+++ b/lib/Target/Hexagon/HexagonPatterns.td
@@ -2770,6 +2770,9 @@ def unalignedstore : PatFrag<(ops node:$val, node:$addr), (store $val, $addr), [
 
 multiclass vS32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
   // Aligned stores
+  def : Pat<(alignednontemporalstore (VTSgl VectorRegs:$src1), IntRegs:$addr),
+            (V6_vS32b_nt_ai IntRegs:$addr, 0, (VTSgl VectorRegs:$src1))>,
+            Requires<[UseHVXSgl]>;
   def : Pat<(alignedstore (VTSgl VectorRegs:$src1), IntRegs:$addr),
             (V6_vS32b_ai IntRegs:$addr, 0, (VTSgl VectorRegs:$src1))>,
             Requires<[UseHVXSgl]>;
@@ -2778,6 +2781,9 @@ multiclass vS32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
             Requires<[UseHVXSgl]>;
 
   // 128B Aligned stores
+  def : Pat<(alignednontemporalstore (VTDbl VectorRegs128B:$src1), IntRegs:$addr),
+            (V6_vS32b_nt_ai_128B IntRegs:$addr, 0, (VTDbl VectorRegs128B:$src1))>,
+            Requires<[UseHVXDbl]>;
   def : Pat<(alignedstore (VTDbl VectorRegs128B:$src1), IntRegs:$addr),
             (V6_vS32b_ai_128B IntRegs:$addr, 0, (VTDbl VectorRegs128B:$src1))>,
             Requires<[UseHVXDbl]>;
@@ -2787,6 +2793,11 @@ multiclass vS32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
 
   // Fold Add R+OFF into vector store.
   let AddedComplexity = 10 in {
+    def : Pat<(alignednontemporalstore (VTSgl VectorRegs:$src1),
+                     (add IntRegs:$src2, Iss4_6:$offset)),
+              (V6_vS32b_nt_ai IntRegs:$src2, Iss4_6:$offset,
+                           (VTSgl VectorRegs:$src1))>,
+              Requires<[UseHVXSgl]>;
     def : Pat<(alignedstore (VTSgl VectorRegs:$src1),
                      (add IntRegs:$src2, Iss4_6:$offset)),
               (V6_vS32b_ai IntRegs:$src2, Iss4_6:$offset,
@@ -2799,6 +2810,11 @@ multiclass vS32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
               Requires<[UseHVXSgl]>;
 
     // Fold Add R+OFF into vector store 128B.
+    def : Pat<(alignednontemporalstore (VTDbl VectorRegs128B:$src1),
+                     (add IntRegs:$src2, Iss4_7:$offset)),
+              (V6_vS32b_nt_ai_128B IntRegs:$src2, Iss4_7:$offset,
+                                (VTDbl VectorRegs128B:$src1))>,
+              Requires<[UseHVXDbl]>;
     def : Pat<(alignedstore (VTDbl VectorRegs128B:$src1),
                      (add IntRegs:$src2, Iss4_7:$offset)),
               (V6_vS32b_ai_128B IntRegs:$src2, Iss4_7:$offset,
@@ -2820,6 +2836,9 @@ defm : vS32b_ai_pats <v8i64,  v16i64>;
 
 multiclass vL32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
   // Aligned loads
+  def : Pat < (VTSgl (alignednontemporalload IntRegs:$addr)),
+              (V6_vL32b_nt_ai IntRegs:$addr, 0) >,
+              Requires<[UseHVXSgl]>;
   def : Pat < (VTSgl (alignedload IntRegs:$addr)),
               (V6_vL32b_ai IntRegs:$addr, 0) >,
               Requires<[UseHVXSgl]>;
@@ -2828,6 +2847,9 @@ multiclass vL32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
               Requires<[UseHVXSgl]>;
 
   // 128B Load
+  def : Pat < (VTDbl (alignednontemporalload IntRegs:$addr)),
+              (V6_vL32b_nt_ai_128B IntRegs:$addr, 0) >,
+              Requires<[UseHVXDbl]>;
   def : Pat < (VTDbl (alignedload IntRegs:$addr)),
               (V6_vL32b_ai_128B IntRegs:$addr, 0) >,
               Requires<[UseHVXDbl]>;
@@ -2837,6 +2859,9 @@ multiclass vL32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
 
   // Fold Add R+OFF into vector load.
   let AddedComplexity = 10 in {
+    def : Pat<(VTDbl (alignednontemporalload (add IntRegs:$src2, Iss4_7:$offset))),
+              (V6_vL32b_nt_ai_128B IntRegs:$src2, Iss4_7:$offset)>,
+               Requires<[UseHVXDbl]>;
     def : Pat<(VTDbl (alignedload (add IntRegs:$src2, Iss4_7:$offset))),
               (V6_vL32b_ai_128B IntRegs:$src2, Iss4_7:$offset)>,
                Requires<[UseHVXDbl]>;
@@ -2844,6 +2869,9 @@ multiclass vL32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
               (V6_vL32Ub_ai_128B IntRegs:$src2, Iss4_7:$offset)>,
                Requires<[UseHVXDbl]>;
 
+    def : Pat<(VTSgl (alignednontemporalload (add IntRegs:$src2, Iss4_6:$offset))),
+              (V6_vL32b_nt_ai IntRegs:$src2, Iss4_6:$offset)>,
+              Requires<[UseHVXSgl]>;
     def : Pat<(VTSgl (alignedload (add IntRegs:$src2, Iss4_6:$offset))),
               (V6_vL32b_ai IntRegs:$src2, Iss4_6:$offset)>,
               Requires<[UseHVXSgl]>;
@@ -2859,6 +2887,9 @@ defm : vL32b_ai_pats <v16i32, v32i32>;
 defm : vL32b_ai_pats <v8i64,  v16i64>;
 
 multiclass STrivv_pats <ValueType VTSgl, ValueType VTDbl> {
+  def : Pat<(alignednontemporalstore (VTSgl VecDblRegs:$src1), IntRegs:$addr),
+            (PS_vstorerw_nt_ai IntRegs:$addr, 0, (VTSgl VecDblRegs:$src1))>,
+           Requires<[UseHVXSgl]>;
   def : Pat<(alignedstore (VTSgl VecDblRegs:$src1), IntRegs:$addr),
             (PS_vstorerw_ai IntRegs:$addr, 0, (VTSgl VecDblRegs:$src1))>,
            Requires<[UseHVXSgl]>;
@@ -2866,6 +2897,10 @@ multiclass STrivv_pats <ValueType VTSgl, ValueType VTDbl> {
             (PS_vstorerwu_ai IntRegs:$addr, 0, (VTSgl VecDblRegs:$src1))>,
            Requires<[UseHVXSgl]>;
 
+  def : Pat<(alignednontemporalstore (VTDbl VecDblRegs128B:$src1), IntRegs:$addr),
+            (PS_vstorerw_nt_ai_128B IntRegs:$addr, 0,
+                  (VTDbl VecDblRegs128B:$src1))>,
+            Requires<[UseHVXDbl]>;
   def : Pat<(alignedstore (VTDbl VecDblRegs128B:$src1), IntRegs:$addr),
             (PS_vstorerw_ai_128B IntRegs:$addr, 0,
                   (VTDbl VecDblRegs128B:$src1))>,
@@ -2882,6 +2917,9 @@ defm : STrivv_pats <v32i32, v64i32>;
 defm : STrivv_pats <v16i64, v32i64>;
 
 multiclass LDrivv_pats <ValueType VTSgl, ValueType VTDbl> {
+  def : Pat<(VTSgl (alignednontemporalload I32:$addr)),
+            (PS_vloadrw_nt_ai I32:$addr, 0)>,
+           Requires<[UseHVXSgl]>;
   def : Pat<(VTSgl (alignedload I32:$addr)),
             (PS_vloadrw_ai I32:$addr, 0)>,
            Requires<[UseHVXSgl]>;
@@ -2889,6 +2927,9 @@ multiclass LDrivv_pats <ValueType VTSgl, ValueType VTDbl> {
             (PS_vloadrwu_ai I32:$addr, 0)>,
            Requires<[UseHVXSgl]>;
 
+  def : Pat<(VTDbl (alignednontemporalload I32:$addr)),
+            (PS_vloadrw_nt_ai_128B I32:$addr, 0)>,
+           Requires<[UseHVXDbl]>;
   def : Pat<(VTDbl (alignedload I32:$addr)),
             (PS_vloadrw_ai_128B I32:$addr, 0)>,
            Requires<[UseHVXDbl]>;

diff --git a/lib/Target/Hexagon/HexagonPseudo.td b/lib/Target/Hexagon/HexagonPseudo.td
index 93fb688fc1c0aea25ad84253a6b1d63c329daa7c..b42c1ab975a80ee28370a0119e8f8f95e03ae05b 100644 (file)
--- a/lib/Target/Hexagon/HexagonPseudo.td
+++ b/lib/Target/Hexagon/HexagonPseudo.td
@@ -407,6 +407,11 @@ def PS_vstorerw_ai: STrivv_template<VecDblRegs, V6_vS32b_ai>,
 def PS_vstorerw_ai_128B: STrivv_template<VecDblRegs128B, V6_vS32b_ai_128B>,
       Requires<[HasV60T,UseHVXDbl]>;
 
+def PS_vstorerw_nt_ai: STrivv_template<VecDblRegs, V6_vS32b_nt_ai>,
+      Requires<[HasV60T,UseHVXSgl]>;
+def PS_vstorerw_nt_ai_128B: STrivv_template<VecDblRegs128B, V6_vS32b_nt_ai_128B>,
+      Requires<[HasV60T,UseHVXDbl]>;
+
 def PS_vstorerwu_ai: STrivv_template<VecDblRegs, V6_vS32Ub_ai>,
       Requires<[HasV60T,UseHVXSgl]>;
 def PS_vstorerwu_ai_128B: STrivv_template<VecDblRegs128B, V6_vS32Ub_ai_128B>,
@@ -433,6 +438,11 @@ def PS_vloadrw_ai: LDrivv_template<VecDblRegs, V6_vL32b_ai>,
 def PS_vloadrw_ai_128B: LDrivv_template<VecDblRegs128B, V6_vL32b_ai_128B>,
       Requires<[HasV60T,UseHVXDbl]>;
 
+def PS_vloadrw_nt_ai: LDrivv_template<VecDblRegs, V6_vL32b_nt_ai>,
+      Requires<[HasV60T,UseHVXSgl]>;
+def PS_vloadrw_nt_ai_128B: LDrivv_template<VecDblRegs128B, V6_vL32b_nt_ai_128B>,
+      Requires<[HasV60T,UseHVXDbl]>;
+
 def PS_vloadrwu_ai: LDrivv_template<VecDblRegs, V6_vL32Ub_ai>,
       Requires<[HasV60T,UseHVXSgl]>;
 def PS_vloadrwu_ai_128B: LDrivv_template<VecDblRegs128B, V6_vL32Ub_ai_128B>,

diff --git a/test/CodeGen/Hexagon/hvx-nontemporal.ll b/test/CodeGen/Hexagon/hvx-nontemporal.ll
new file mode 100644 (file)
index 0000000..98c5ef4
--- /dev/null
+++ b/test/CodeGen/Hexagon/hvx-nontemporal.ll
@@ -0,0 +1,28 @@
+; RUN: llc -march=hexagon < %s | FileCheck %s
+target triple = "hexagon"
+
+; Function Attrs: norecurse nounwind
+define void @test(<32 x i32>* nocapture readonly %x, <32 x i32>* nocapture readnone %y, <32 x i32>* nocapture %a, <32 x i32>* nocapture %b) #0 {
+entry:
+; CHECK: v0 = vmem(r0+#7):nt
+  %add.ptr = getelementptr inbounds <32 x i32>, <32 x i32>* %x, i32 7
+  %0 = load <32 x i32>, <32 x i32>* %add.ptr, align 128, !tbaa !1, !nontemporal !4
+
+; CHECK: v1.cur = vmem(r2+#0):nt
+  %1 = load <32 x i32>, <32 x i32>* %a, align 128, !tbaa !1, !nontemporal !4
+
+; CHECK: vmem(r3+#3):nt = v1
+  %add.ptr2 = getelementptr inbounds <32 x i32>, <32 x i32>* %b, i32 3
+  store <32 x i32> %1, <32 x i32>* %add.ptr2, align 128, !tbaa !1, !nontemporal !4
+
+; CHECK: vmem(r2+#0):nt = v0
+  store <32 x i32> %0, <32 x i32>* %a, align 128, !tbaa !1, !nontemporal !4
+  ret void
+}
+
+attributes #0 = { norecurse nounwind "target-cpu"="hexagonv60" "target-features"="+hvx,+hvx-double" }
+
+!1 = !{!2, !2, i64 0}
+!2 = !{!"omnipotent char", !3, i64 0}
+!3 = !{!"Simple C/C++ TBAA"}
+!4 = !{i32 1}