Remove 256-bit AVX non-temporal store intrinsics. Similar was previously done for 128-bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@156375 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 57 insertions, 22 deletions

diff --git a/include/llvm/IntrinsicsX86.td b/include/llvm/IntrinsicsX86.td
index 1abeb9e555..805d3667d1 100644
--- a/include/llvm/IntrinsicsX86.td
+++ b/include/llvm/IntrinsicsX86.td
@@ -1282,16 +1282,6 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
         Intrinsic<[], [llvm_ptr_ty, llvm_v32i8_ty], []>;
 }
 
-// Cacheability support ops
-let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_avx_movnt_dq_256 : GCCBuiltin<"__builtin_ia32_movntdq256">,
-        Intrinsic<[], [llvm_ptr_ty, llvm_v4i64_ty], []>;
-  def int_x86_avx_movnt_pd_256 : GCCBuiltin<"__builtin_ia32_movntpd256">,
-        Intrinsic<[], [llvm_ptr_ty, llvm_v4f64_ty], []>;
-  def int_x86_avx_movnt_ps_256 : GCCBuiltin<"__builtin_ia32_movntps256">,
-        Intrinsic<[], [llvm_ptr_ty, llvm_v8f32_ty], []>;
-}
-
 // Conditional load ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_avx_maskload_pd : GCCBuiltin<"__builtin_ia32_maskloadpd">,
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index d6755cea41..a518997b78 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -3336,13 +3336,6 @@ let AddedComplexity = 400 in { // Prefer non-temporal versions
                                                 IIC_SSE_MOVNT>, VEX;
 }
 
-def : Pat<(int_x86_avx_movnt_dq_256 addr:$dst, VR256:$src),
-          (VMOVNTDQYmr addr:$dst, VR256:$src)>;
-def : Pat<(int_x86_avx_movnt_pd_256 addr:$dst, VR256:$src),
-          (VMOVNTPDYmr addr:$dst, VR256:$src)>;
-def : Pat<(int_x86_avx_movnt_ps_256 addr:$dst, VR256:$src),
-          (VMOVNTPSYmr addr:$dst, VR256:$src)>;
-
 let AddedComplexity = 400 in { // Prefer non-temporal versions
 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
                     "movntps\t{$src, $dst|$dst, $src}",
diff --git a/lib/VMCore/AutoUpgrade.cpp b/lib/VMCore/AutoUpgrade.cpp
index 2e16372fb1..522b07e454 100644
--- a/lib/VMCore/AutoUpgrade.cpp
+++ b/lib/VMCore/AutoUpgrade.cpp
@@ -57,7 +57,10 @@ static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
         Name.startswith("x86.sse2.pcmpgt.") ||
         Name.startswith("x86.avx2.pcmpeq.") ||
         Name.startswith("x86.avx2.pcmpgt.") ||
-        Name.startswith("x86.avx.vpermil.")) {
+        Name.startswith("x86.avx.vpermil.") ||
+        Name == "x86.avx.movnt.dq.256" ||
+        Name == "x86.avx.movnt.pd.256" ||
+        Name == "x86.avx.movnt.ps.256") {
       NewFn = 0;
       return true;
     }
@@ -118,15 +121,40 @@ void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
                                   "pcmpgt");
       // need to sign extend since icmp returns vector of i1
       Rep = Builder.CreateSExt(Rep, CI->getType(), "");
+    } else if (Name == "llvm.x86.avx.movnt.dq.256" ||
+               Name == "llvm.x86.avx.movnt.ps.256" ||
+               Name == "llvm.x86.avx.movnt.pd.256") {
+      IRBuilder<> Builder(C);
+      Builder.SetInsertPoint(CI->getParent(), CI);
+
+      Module *M = F->getParent();
+      SmallVector<Value *, 1> Elts;
+      Elts.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
+      MDNode *Node = MDNode::get(C, Elts);
+
+      Value *Arg0 = CI->getArgOperand(0);
+      Value *Arg1 = CI->getArgOperand(1);
+
+      // Convert the type of the pointer to a pointer to the stored type.
+      Value *BC = Builder.CreateBitCast(Arg0,
+                                        PointerType::getUnqual(Arg1->getType()),
+                                        "cast");
+      StoreInst *SI = Builder.CreateStore(Arg1, BC);
+      SI->setMetadata(M->getMDKindID("nontemporal"), Node);
+      SI->setAlignment(16);
+
+      // Remove intrinsic.
+      CI->eraseFromParent();
+      return;
     } else {
       bool PD128 = false, PD256 = false, PS128 = false, PS256 = false;
-      if (Name.startswith("llvm.x86.avx.vpermil.pd.256"))
+      if (Name == "llvm.x86.avx.vpermil.pd.256")
         PD256 = true;
-      else if (Name.startswith("llvm.x86.avx.vpermil.pd"))
+      else if (Name == "llvm.x86.avx.vpermil.pd")
         PD128 = true;
-      else if (Name.startswith("llvm.x86.avx.vpermil.ps.256"))
+      else if (Name == "llvm.x86.avx.vpermil.ps.256")
         PS256 = true;
-      else if (Name.startswith("llvm.x86.avx.vpermil.ps"))
+      else if (Name == "llvm.x86.avx.vpermil.ps")
         PS128 = true;
 
       if (PD256 || PD128 || PS256 || PS128) {
diff --git a/test/CodeGen/X86/avx-intrinsics-x86.ll b/test/CodeGen/X86/avx-intrinsics-x86.ll
index b33493252a..4cca693d38 100644
--- a/test/CodeGen/X86/avx-intrinsics-x86.ll
+++ b/test/CodeGen/X86/avx-intrinsics-x86.ll
@@ -2555,3 +2555,27 @@ define i32 @crc32_32_32(i32 %a, i32 %b) nounwind {
   ret i32 %tmp
 }
 declare i32 @llvm.x86.sse42.crc32.32.32(i32, i32) nounwind
+
+; CHECK: movntdq
+define void @movnt_dq(i8* %p, <4 x i64> %a1) nounwind {
+  %a2 = add <4 x i64> %a1, <i64 1, i64 1, i64 1, i64 1>
+  tail call void @llvm.x86.avx.movnt.dq.256(i8* %p, <4 x i64> %a2) nounwind
+  ret void
+}
+declare void @llvm.x86.avx.movnt.dq.256(i8*, <4 x i64>) nounwind
+
+; CHECK: movntps
+define void @movnt_ps(i8* %p, <8 x float> %a) nounwind {
+  tail call void @llvm.x86.avx.movnt.ps.256(i8* %p, <8 x float> %a) nounwind
+  ret void
+}
+declare void @llvm.x86.avx.movnt.ps.256(i8*, <8 x float>) nounwind
+
+; CHECK: movntpd
+define void @movnt_pd(i8* %p, <4 x double> %a1) nounwind {
+  ; add operation forces the execution domain.
+  %a2 = fadd <4 x double> %a1, <double 0x0, double 0x0, double 0x0, double 0x0>
+  tail call void @llvm.x86.avx.movnt.pd.256(i8* %p, <4 x double> %a2) nounwind
+  ret void
+}
+declare void @llvm.x86.avx.movnt.pd.256(i8*, <4 x double>) nounwind