From 761073412e8df3c66ff9efa48078c74a2dfbe7eb Mon Sep 17 00:00:00 2001
From: Akash Kothari <akashk4@tyler.cs.illinois.edu>
Date: Tue, 28 Jan 2020 03:46:15 -0600
Subject: [PATCH] Adding more regression tests

---
 .../BuildDFG/AllocationNode.ll                | 229 +++++++++++++
 .../{CreateNode.genvisc.ll => CreateNode.ll}  |   0
 ...ndEdge.genvisc.ll => CreateNodeAndEdge.ll} |   0
 ...eafBindEdge.genvisc.ll => LeafBindEdge.ll} |   0
 ...eGetters.genvisc.ll => LeafNodeGetters.ll} |   0
 .../{ThreeLevel.genvisc.ll => ThreeLevel.ll}  |   0
 .../BuildDFG/ThreeLevelEdge.genvisc.ll        | 248 ---------------
 .../BuildDFG/ThreeLevelEdge.ll                | 257 ++++++++-------
 .../{TwoLaunch.genvisc.ll => TwoLaunch.ll}    |   0
 hpvm/test/regressionTests/BuildDFG/TwoNode.ll | 197 ------------
 .../ThreeLevel.constmem.nvptxTox86.ll         | 276 ++++++++++++++++
 .../ClearDFG/ThreeLevel.nvptxTox86.ll         | 256 +++++++++++++++
 .../ClearDFG/ThreeLevel.x86.ll                | 225 +++++++++++++
 ...atomic.genvisc.ll => ThreeLevel.atomic.ll} |   2 +-
 .../DFG2LLVM_NVPTX/ThreeLevel.cond.const.ll   | 259 +++++++++++++++
 .../DFG2LLVM_NVPTX/ThreeLevel.cond.ll         | 258 +++++++++++++++
 ....opt.genvisc.ll => ThreeLevel.constmem.ll} |   2 +-
 .../{ThreeLevel.genvisc.ll => ThreeLevel.ll}  |   2 +-
 ...{CreateNode.dfg.ll => OneLevel.codeGen.ll} |  12 +-
 ...oneLaunchAlloca.dfg.ll => OneRootBasic.ll} |   6 +-
 ...hreeLevel.dfg.ll => ThreeLevel.codeGen.ll} |  20 +-
 .../DFG2LLVM_X86/ThreeLevel.cond.ll           | 300 ++++++++++++++++++
 .../DFG2LLVM_X86/ThreeLevel.constmem.nvptx.ll | 270 ++++++++++++++++
 .../DFG2LLVM_X86/ThreeLevel.nvptx.ll          | 250 +++++++++++++++
 .../{TwoLevel.dfg.ll => TwoLevel.codeGen.ll}  |  16 +-
 .../{TwoLaunch.dfg.ll => TwoRoot.ll}          |  26 +-
 .../regressionTests/GenHPVM/AllocationNode.ll | 187 +++++++++++
 .../{GenVISC => GenHPVM}/CreateNode.ll        |   0
 .../{GenVISC => GenHPVM}/CreateNodeAndEdge.ll |   0
 .../{GenVISC => GenHPVM}/LeafBindEdge.ll      |   0
 .../{GenVISC => GenHPVM}/LeafNodeGetters.ll   |   0
 .../{GenVISC => GenHPVM}/PopAndPush.ll        |   0
 .../{GenVISC => GenHPVM}/ThreeLevel.ll        |   0
 .../{GenVISC => GenHPVM}/ThreeLevelEdge.ll    |   0
 .../{GenVISC => GenHPVM}/TwoLaunch.ll         |   0
 .../{GenVISC => GenHPVM}/TwoLevel.ll          |   0
 .../{GenVISC => GenHPVM}/oneLaunchAlloca.ll   |   0
 .../{GenVISC => GenHPVM}/oneLaunchMalloc.ll   |   0
 .../regressionTests/GenVISC/AllocationNode.ll | 206 ------------
 39 files changed, 2701 insertions(+), 803 deletions(-)
 create mode 100644 hpvm/test/regressionTests/BuildDFG/AllocationNode.ll
 rename hpvm/test/regressionTests/BuildDFG/{CreateNode.genvisc.ll => CreateNode.ll} (100%)
 rename hpvm/test/regressionTests/BuildDFG/{CreateNodeAndEdge.genvisc.ll => CreateNodeAndEdge.ll} (100%)
 rename hpvm/test/regressionTests/BuildDFG/{LeafBindEdge.genvisc.ll => LeafBindEdge.ll} (100%)
 rename hpvm/test/regressionTests/BuildDFG/{LeafNodeGetters.genvisc.ll => LeafNodeGetters.ll} (100%)
 rename hpvm/test/regressionTests/BuildDFG/{ThreeLevel.genvisc.ll => ThreeLevel.ll} (100%)
 delete mode 100644 hpvm/test/regressionTests/BuildDFG/ThreeLevelEdge.genvisc.ll
 rename hpvm/test/regressionTests/BuildDFG/{TwoLaunch.genvisc.ll => TwoLaunch.ll} (100%)
 delete mode 100644 hpvm/test/regressionTests/BuildDFG/TwoNode.ll
 create mode 100644 hpvm/test/regressionTests/ClearDFG/ThreeLevel.constmem.nvptxTox86.ll
 create mode 100644 hpvm/test/regressionTests/ClearDFG/ThreeLevel.nvptxTox86.ll
 create mode 100644 hpvm/test/regressionTests/ClearDFG/ThreeLevel.x86.ll
 rename hpvm/test/regressionTests/DFG2LLVM_NVPTX/{ThreeLevel.atomic.genvisc.ll => ThreeLevel.atomic.ll} (99%)
 create mode 100644 hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.cond.const.ll
 create mode 100644 hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.cond.ll
 rename hpvm/test/regressionTests/DFG2LLVM_NVPTX/{ThreeLevel.opt.genvisc.ll => ThreeLevel.constmem.ll} (99%)
 rename hpvm/test/regressionTests/DFG2LLVM_NVPTX/{ThreeLevel.genvisc.ll => ThreeLevel.ll} (99%)
 rename hpvm/test/regressionTests/DFG2LLVM_X86/{CreateNode.dfg.ll => OneLevel.codeGen.ll} (95%)
 rename hpvm/test/regressionTests/DFG2LLVM_X86/{oneLaunchAlloca.dfg.ll => OneRootBasic.ll} (95%)
 rename hpvm/test/regressionTests/DFG2LLVM_X86/{ThreeLevel.dfg.ll => ThreeLevel.codeGen.ll} (95%)
 create mode 100644 hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.cond.ll
 create mode 100644 hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.constmem.nvptx.ll
 create mode 100644 hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.nvptx.ll
 rename hpvm/test/regressionTests/DFG2LLVM_X86/{TwoLevel.dfg.ll => TwoLevel.codeGen.ll} (95%)
 rename hpvm/test/regressionTests/DFG2LLVM_X86/{TwoLaunch.dfg.ll => TwoRoot.ll} (93%)
 create mode 100644 hpvm/test/regressionTests/GenHPVM/AllocationNode.ll
 rename hpvm/test/regressionTests/{GenVISC => GenHPVM}/CreateNode.ll (100%)
 rename hpvm/test/regressionTests/{GenVISC => GenHPVM}/CreateNodeAndEdge.ll (100%)
 rename hpvm/test/regressionTests/{GenVISC => GenHPVM}/LeafBindEdge.ll (100%)
 rename hpvm/test/regressionTests/{GenVISC => GenHPVM}/LeafNodeGetters.ll (100%)
 rename hpvm/test/regressionTests/{GenVISC => GenHPVM}/PopAndPush.ll (100%)
 rename hpvm/test/regressionTests/{GenVISC => GenHPVM}/ThreeLevel.ll (100%)
 rename hpvm/test/regressionTests/{GenVISC => GenHPVM}/ThreeLevelEdge.ll (100%)
 rename hpvm/test/regressionTests/{GenVISC => GenHPVM}/TwoLaunch.ll (100%)
 rename hpvm/test/regressionTests/{GenVISC => GenHPVM}/TwoLevel.ll (100%)
 rename hpvm/test/regressionTests/{GenVISC => GenHPVM}/oneLaunchAlloca.ll (100%)
 rename hpvm/test/regressionTests/{GenVISC => GenHPVM}/oneLaunchMalloc.ll (100%)
 delete mode 100644 hpvm/test/regressionTests/GenVISC/AllocationNode.ll

diff --git a/hpvm/test/regressionTests/BuildDFG/AllocationNode.ll b/hpvm/test/regressionTests/BuildDFG/AllocationNode.ll
new file mode 100644
index 0000000000..2bbe422e72
--- /dev/null
+++ b/hpvm/test/regressionTests/BuildDFG/AllocationNode.ll
@@ -0,0 +1,229 @@
+; RUN: opt -load LLVMBuildDFG.so -S < %s | FileCheck %s
+; ModuleID = 'AllocationNode.ll'
+source_filename = "ThreeLevel.allocation.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%struct.Root = type { i32*, i64, i32*, i64 }
+%struct.out.Allocation = type <{ i8*, i64 }>
+%emptyStruct = type <{}>
+%emptyStruct.0 = type <{}>
+%emptyStruct.1 = type <{}>
+%emptyStruct.2 = type <{}>
+
+
+; CHECK-LABEL: %struct.out.Allocation =
+
+; CHECK-LABEL: void @Launch(
+; CHECK: call i8* @llvm.hpvm.launch(i8*
+; CHECK-NEXT: call void @llvm.hpvm.wait(i8*
+
+; CHECK-LABEL: i32 @main(
+; CHECK: call void @llvm.hpvm.init()
+; CHECK-NEXT: tail call void @Launch(
+; CHECK-NEXT: call void @llvm.hpvm.cleanup()
+
+; CHECK-LABEL: @Allocation_cloned(
+; CHECK: call i8* @llvm.hpvm.malloc(i64
+
+; CHECK-LABEL: @Func1_cloned(
+
+; CHECK-LABEL: @Func3_cloned(
+; CHECK: %Func1_cloned.node = call i8* @llvm.hpvm.createNode2D(i8*
+; CHECK: %Allocation_cloned.node = call i8* @llvm.hpvm.createNode(i8*
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Allocation_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node
+; CHECK-NEXT: call i8* @llvm.hpvm.createEdge(i8* %Allocation_cloned.node
+; CHECK-NEXTL call i8* @llvm.hpvm.createEdge(i8* %Allocation_cloned.node
+
+; CHECK-LABEL: @Func2_cloned(
+; CHECK: %Func3_cloned.node = call i8* @llvm.hpvm.createNode2D(
+; CHECK-NEXT:  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+; CHECK-NEXT:  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+; CHECK-NEXT:  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+; CHECK-NEXT:  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+
+; CHECK-LABEL: @PipeRoot_cloned(
+; CHECK: %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8*
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+
+
+declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__malloc(i64) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__return(i32, ...) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__edge(i8*, i8*, i32, i32, i32, i32) local_unnamed_addr #0
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local void @Launch() local_unnamed_addr #2 {
+entry:
+  %RootArgs = alloca %struct.Root, align 8
+  %0 = bitcast %struct.Root* %RootArgs to i8*
+  call void @llvm.lifetime.start.p0i8(i64 32, i8* nonnull %0) #5
+  %call = tail call noalias i8* @malloc(i64 1024) #5
+  %1 = bitcast %struct.Root* %RootArgs to i8**
+  store i8* %call, i8** %1, align 8, !tbaa !6
+  %Insize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
+  store i64 1024, i64* %Insize, align 8, !tbaa !12
+  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
+  %call1 = tail call noalias i8* @malloc(i64 1024) #5
+  %2 = bitcast i32** %output to i8**
+  store i8* %call1, i8** %2, align 8, !tbaa !13
+  %Outsize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 3
+  store i64 1024, i64* %Outsize, align 8, !tbaa !14
+  %3 = bitcast %struct.Root* %RootArgs to i8*
+  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned to i8*), i8* %3, i1 false)
+  call void @llvm.hpvm.wait(i8* %graphID)
+  call void @llvm.lifetime.end.p0i8(i64 32, i8* nonnull %0) #5
+  ret void
+}
+
+; Function Attrs: nofree nounwind
+declare dso_local noalias i8* @malloc(i64) local_unnamed_addr #3
+
+declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
+
+; Function Attrs: nounwind uwtable
+define dso_local i32 @main() local_unnamed_addr #4 {
+entry:
+  call void @llvm.hpvm.init()
+  tail call void @Launch()
+  call void @llvm.hpvm.cleanup()
+  ret i32 0
+}
+
+declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
+
+declare i8* @llvm_hpvm_initializeTimerSet()
+
+declare void @llvm_hpvm_switchToTimer(i8**, i32)
+
+declare void @llvm_hpvm_printTimerSet(i8**, i8*)
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.malloc(i64) #5
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.Allocation @Allocation_cloned(i64 %block) #4 {
+entry:
+  %call1 = call i8* @llvm.hpvm.malloc(i64 %block)
+  %returnStruct = insertvalue %struct.out.Allocation undef, i8* %call1, 0
+  %returnStruct2 = insertvalue %struct.out.Allocation %returnStruct, i64 %block, 1
+  ret %struct.out.Allocation %returnStruct2
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode2D(i8*, i64, i64) #5
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct @Func1_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  ret %emptyStruct undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode(i8*) #5
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #5
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createEdge(i8*, i8*, i1, i32, i32, i1) #5
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.0 @Func3_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func1_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned to i8*), i64 3, i64 5)
+  %Allocation_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Allocation (i64)* @Allocation_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Allocation_cloned.node, i32 1, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 3, i32 3, i1 false)
+  %output = call i8* @llvm.hpvm.createEdge(i8* %Allocation_cloned.node, i8* %Func1_cloned.node, i1 true, i32 0, i32 0, i1 false)
+  %output1 = call i8* @llvm.hpvm.createEdge(i8* %Allocation_cloned.node, i8* %Func1_cloned.node, i1 true, i32 1, i32 1, i1 false)
+  ret %emptyStruct.0 undef
+}
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.1 @Func2_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func3_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.1 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #5
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.2 @PipeRoot_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.2 undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.wait(i8*) #5
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.init() #5
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.cleanup() #5
+
+attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { noinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #3 = { nofree nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #4 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #5 = { nounwind }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+!hpvm_hint_gpu = !{!2}
+!hpvm_hint_cpu = !{!3, !4, !5}
+!hpvm_hint_cpu_gpu = !{}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 3551132592a00cab6c966df508ab511598269f78)"}
+!2 = !{%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned}
+!3 = !{%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned}
+!4 = !{%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned}
+!5 = !{%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned}
+!6 = !{!7, !8, i64 0}
+!7 = !{!"Root", !8, i64 0, !11, i64 8, !8, i64 16, !11, i64 24}
+!8 = !{!"any pointer", !9, i64 0}
+!9 = !{!"omnipotent char", !10, i64 0}
+!10 = !{!"Simple C/C++ TBAA"}
+!11 = !{!"long", !9, i64 0}
+!12 = !{!7, !11, i64 8}
+!13 = !{!7, !8, i64 16}
+!14 = !{!7, !11, i64 24}
diff --git a/hpvm/test/regressionTests/BuildDFG/CreateNode.genvisc.ll b/hpvm/test/regressionTests/BuildDFG/CreateNode.ll
similarity index 100%
rename from hpvm/test/regressionTests/BuildDFG/CreateNode.genvisc.ll
rename to hpvm/test/regressionTests/BuildDFG/CreateNode.ll
diff --git a/hpvm/test/regressionTests/BuildDFG/CreateNodeAndEdge.genvisc.ll b/hpvm/test/regressionTests/BuildDFG/CreateNodeAndEdge.ll
similarity index 100%
rename from hpvm/test/regressionTests/BuildDFG/CreateNodeAndEdge.genvisc.ll
rename to hpvm/test/regressionTests/BuildDFG/CreateNodeAndEdge.ll
diff --git a/hpvm/test/regressionTests/BuildDFG/LeafBindEdge.genvisc.ll b/hpvm/test/regressionTests/BuildDFG/LeafBindEdge.ll
similarity index 100%
rename from hpvm/test/regressionTests/BuildDFG/LeafBindEdge.genvisc.ll
rename to hpvm/test/regressionTests/BuildDFG/LeafBindEdge.ll
diff --git a/hpvm/test/regressionTests/BuildDFG/LeafNodeGetters.genvisc.ll b/hpvm/test/regressionTests/BuildDFG/LeafNodeGetters.ll
similarity index 100%
rename from hpvm/test/regressionTests/BuildDFG/LeafNodeGetters.genvisc.ll
rename to hpvm/test/regressionTests/BuildDFG/LeafNodeGetters.ll
diff --git a/hpvm/test/regressionTests/BuildDFG/ThreeLevel.genvisc.ll b/hpvm/test/regressionTests/BuildDFG/ThreeLevel.ll
similarity index 100%
rename from hpvm/test/regressionTests/BuildDFG/ThreeLevel.genvisc.ll
rename to hpvm/test/regressionTests/BuildDFG/ThreeLevel.ll
diff --git a/hpvm/test/regressionTests/BuildDFG/ThreeLevelEdge.genvisc.ll b/hpvm/test/regressionTests/BuildDFG/ThreeLevelEdge.genvisc.ll
deleted file mode 100644
index 8c76da27e8..0000000000
--- a/hpvm/test/regressionTests/BuildDFG/ThreeLevelEdge.genvisc.ll
+++ /dev/null
@@ -1,248 +0,0 @@
-; RUN: opt -load LLVMBuildDFG.so -S <  %s | FileCheck %s
-; ModuleID = 'ThreeLevelEdge.ll'
-source_filename = "ThreeLevelEdge.c"
-target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
-target triple = "x86_64-unknown-linux-gnu"
-
-%struct.Root = type { i32*, i32*, i32* }
-%struct.out.Func1 = type <{ i32* }>
-%struct.out.Func4 = type <{ i32* }>
-%struct.out.Func5 = type <{ i32* }>
-%struct.out.Func3 = type <{ i32* }>
-%struct.out.Func2 = type <{ i32* }>
-%struct.out.PipeRoot = type <{ i32* }>
-
-; CHECK-LABEL: struct.Root =
-
-; CHECK-LABEL: %struct.out.Func1 =
-; CHECK-LABEL: %struct.out.Func4 =
-; CHECK-LABEL: %struct.out.Func5 =
-; CHECK-LABEL: %struct.out.Func3 =
-; CHECK-LABEL: %struct.out.Func2 =
-; CHECK-LABEL: %struct.out.PipeRoot =
-
-; CHECK-LABEL: i32 @main(
-; CHECK: [[ALLOCA:%[1-9a-zA-Z]+]] = alloca %struct.Root
-; CHECK: call void @llvm.hpvm.init()
-; CHECK:  [[REGISTER:%[1-9]+]] = bitcast %struct.Root* [[ALLOCA]] to i8*
-; CHECK: call i8* @llvm.hpvm.launch(i8* bitcast (%struct.out.PipeRoot (i32*, i32*, i32*)* @PipeRoot_cloned to i8*), i8* [[REGISTER]],
-; CHECK-NEXT: call void @llvm.hpvm.wait(i8*
-
-; CHECK-LABEL: @Func1_cloned(
-; CHECK: [[RET1:%[1-9a-zA-Z]+]] = insertvalue %struct.out.Func1 undef,
-; CHECK-NEXT: ret %struct.out.Func1 [[RET1]]
-
-; CHECK-LABEL: @Func4_cloned(
-; CHECK: [[RET4:%[1-9a-zA-Z]+]] = insertvalue %struct.out.Func4 undef,
-; CHECK-NEXT: ret %struct.out.Func4 [[RET4]]
-
-; CHECK-LABEL: @Func5_cloned(
-; CHECK: [[RET5:%[1-9a-zA-Z]+]] = insertvalue %struct.out.Func5 undef,
-; CHECK-NEXT: ret %struct.out.Func5 [[RET5]]
-
-; CHECK-LABEL: @Func3_cloned(
-; CHECK: %Func4_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%struct.out.Func4 (i32*
-; CHECK-NEXT: %Func5_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%struct.out.Func5 (i32*
-; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func4_cloned.node,
-; CHECK-NEXT: call i8* @llvm.hpvm.createEdge(i8* %Func4_cloned.node, i8* %Func5_cloned.node,
-; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func5_cloned.node,
-
-; CHECK-LABEL: @Func2_cloned(
-; CHECK: %Func3_cloned.node = call i8* @llvm.hpvm.createNode1D(i8* bitcast (%struct.out.Func3 (i32*
-; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node,
-; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func3_cloned.node,
-
-; CHECK-LABEL: @PipeRoot_cloned(i32*
-; CHECK: %Func1_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func1 (i32*
-; CHECK-NEXT: %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func2 (i32*
-; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node,
-; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node,
-; CHECK-NEXT: call i8* @llvm.hpvm.createEdge(i8* %Func1_cloned.node, i8* %Func2_cloned.node,
-; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func2_cloned.node,
-
-
-declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__return(i32, ...) local_unnamed_addr #0
-
-; Function Attrs: argmemonly nounwind
-declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
-
-declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
-
-declare dso_local i8* @__hpvm__edge(i8*, i8*, i32, i32, i32, i32) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__bindOut(i8*, i32, i32, i32) local_unnamed_addr #0
-
-; Function Attrs: argmemonly nounwind
-declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
-
-; Function Attrs: nounwind uwtable
-define dso_local i32 @main() local_unnamed_addr #2 {
-entry:
-  %In1 = alloca i32, align 4
-  %In2 = alloca i32, align 4
-  %Out = alloca i32, align 4
-  %RootArgs = alloca %struct.Root, align 8
-  %0 = bitcast i32* %In1 to i8*
-  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %0) #3
-  store i32 1, i32* %In1, align 4, !tbaa !8
-  %1 = bitcast i32* %In2 to i8*
-  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %1) #3
-  store i32 2, i32* %In2, align 4, !tbaa !8
-  %2 = bitcast i32* %Out to i8*
-  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %2) #3
-  store i32 0, i32* %Out, align 4, !tbaa !8
-  %3 = bitcast %struct.Root* %RootArgs to i8*
-  call void @llvm.lifetime.start.p0i8(i64 24, i8* nonnull %3) #3
-  %input1 = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 0
-  store i32* %In1, i32** %input1, align 8, !tbaa !12
-  %intput2 = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
-  store i32* %In2, i32** %intput2, align 8, !tbaa !15
-  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
-  store i32* %Out, i32** %output, align 8, !tbaa !16
-  call void @llvm.hpvm.init()
-  %4 = bitcast %struct.Root* %RootArgs to i8*
-  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%struct.out.PipeRoot (i32*, i32*, i32*)* @PipeRoot_cloned to i8*), i8* %4, i1 false)
-  call void @llvm.hpvm.wait(i8* %graphID)
-  call void @llvm.hpvm.cleanup()
-  call void @llvm.lifetime.end.p0i8(i64 24, i8* nonnull %3) #3
-  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %2) #3
-  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %1) #3
-  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %0) #3
-  ret i32 0
-}
-
-declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
-
-declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
-
-declare i8* @llvm_hpvm_initializeTimerSet()
-
-declare void @llvm_hpvm_switchToTimer(i8**, i32)
-
-declare void @llvm_hpvm_printTimerSet(i8**, i8*)
-
-; Function Attrs: nounwind uwtable
-define dso_local %struct.out.Func1 @Func1_cloned(i32* in %In, i32* out %Out) #2 {
-entry:
-  %returnStruct = insertvalue %struct.out.Func1 undef, i32* %Out, 0
-  ret %struct.out.Func1 %returnStruct
-}
-
-; Function Attrs: nounwind uwtable
-define dso_local %struct.out.Func4 @Func4_cloned(i32* in %In, i32* out %Out) #2 {
-entry:
-  %returnStruct = insertvalue %struct.out.Func4 undef, i32* %Out, 0
-  ret %struct.out.Func4 %returnStruct
-}
-
-; Function Attrs: nounwind uwtable
-define dso_local %struct.out.Func5 @Func5_cloned(i32* in %In1, i32* in %In2, i32* out %Out) #2 {
-entry:
-  %returnStruct = insertvalue %struct.out.Func5 undef, i32* %Out, 0
-  ret %struct.out.Func5 %returnStruct
-}
-
-; Function Attrs: nounwind
-declare i8* @llvm.hpvm.createNode2D(i8*, i64, i64) #3
-
-; Function Attrs: nounwind
-declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #3
-
-; Function Attrs: nounwind
-declare i8* @llvm.hpvm.createEdge(i8*, i8*, i1, i32, i32, i1) #3
-
-; Function Attrs: nounwind
-declare void @llvm.hpvm.bind.output(i8*, i32, i32, i1) #3
-
-; Function Attrs: nounwind uwtable
-define dso_local %struct.out.Func3 @Func3_cloned(i32* in %In, i32* out %Out) #2 {
-entry:
-  %Func4_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%struct.out.Func4 (i32*, i32*)* @Func4_cloned to i8*), i64 3, i64 6)
-  %Func5_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%struct.out.Func5 (i32*, i32*, i32*)* @Func5_cloned to i8*), i64 4, i64 5)
-  call void @llvm.hpvm.bind.input(i8* %Func4_cloned.node, i32 0, i32 0, i1 false)
-  %output = call i8* @llvm.hpvm.createEdge(i8* %Func4_cloned.node, i8* %Func5_cloned.node, i1 false, i32 1, i32 1, i1 false)
-  call void @llvm.hpvm.bind.output(i8* %Func5_cloned.node, i32 0, i32 0, i1 false)
-  ret %struct.out.Func3 undef
-}
-
-; Function Attrs: nounwind
-declare i8* @llvm.hpvm.createNode1D(i8*, i64) #3
-
-; Function Attrs: nounwind uwtable
-define dso_local %struct.out.Func2 @Func2_cloned(i32* in %BindIn, i32* in %SrcIn, i32* out %Out) #2 {
-entry:
-  %Func3_cloned.node = call i8* @llvm.hpvm.createNode1D(i8* bitcast (%struct.out.Func3 (i32*, i32*)* @Func3_cloned to i8*), i64 3)
-  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
-  call void @llvm.hpvm.bind.output(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
-  ret %struct.out.Func2 undef
-}
-
-; Function Attrs: nounwind
-declare i8* @llvm.hpvm.createNode(i8*) #3
-
-; Function Attrs: nounwind uwtable
-define dso_local %struct.out.PipeRoot @PipeRoot_cloned(i32* in %In1, i32* in %In2, i32* out %Out) #2 {
-entry:
-  %Func1_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func1 (i32*, i32*)* @Func1_cloned to i8*))
-  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func2 (i32*, i32*, i32*)* @Func2_cloned to i8*))
-  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 0, i32 0, i1 false)
-  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 0, i1 false)
-  %output = call i8* @llvm.hpvm.createEdge(i8* %Func1_cloned.node, i8* %Func2_cloned.node, i1 false, i32 1, i32 1, i1 false)
-  call void @llvm.hpvm.bind.output(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
-  ret %struct.out.PipeRoot undef
-}
-
-; Function Attrs: nounwind
-declare void @llvm.hpvm.init() #3
-
-; Function Attrs: nounwind
-declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #3
-
-; Function Attrs: nounwind
-declare void @llvm.hpvm.wait(i8*) #3
-
-; Function Attrs: nounwind
-declare void @llvm.hpvm.cleanup() #3
-
-attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
-attributes #1 = { argmemonly nounwind }
-attributes #2 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
-attributes #3 = { nounwind }
-
-!llvm.module.flags = !{!0}
-!llvm.ident = !{!1}
-!hpvm_hint_cpu = !{!2, !3, !4, !5, !6, !7}
-!hpvm_hint_gpu = !{}
-!hpvm_hint_spir = !{}
-!hpvm_hint_cudnn = !{}
-!hpvm_hint_promise = !{}
-!hpvm_hint_cpu_gpu = !{}
-!hpvm_hint_cpu_spir = !{}
-
-!0 = !{i32 1, !"wchar_size", i32 4}
-!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 6690f9e7e8b46b96aea222d3e85315cd63545953)"}
-!2 = !{%struct.out.Func1 (i32*, i32*)* @Func1_cloned}
-!3 = !{%struct.out.Func4 (i32*, i32*)* @Func4_cloned}
-!4 = !{%struct.out.Func5 (i32*, i32*, i32*)* @Func5_cloned}
-!5 = !{%struct.out.Func3 (i32*, i32*)* @Func3_cloned}
-!6 = !{%struct.out.Func2 (i32*, i32*, i32*)* @Func2_cloned}
-!7 = !{%struct.out.PipeRoot (i32*, i32*, i32*)* @PipeRoot_cloned}
-!8 = !{!9, !9, i64 0}
-!9 = !{!"int", !10, i64 0}
-!10 = !{!"omnipotent char", !11, i64 0}
-!11 = !{!"Simple C/C++ TBAA"}
-!12 = !{!13, !14, i64 0}
-!13 = !{!"Root", !14, i64 0, !14, i64 8, !14, i64 16}
-!14 = !{!"any pointer", !10, i64 0}
-!15 = !{!13, !14, i64 8}
-!16 = !{!13, !14, i64 16}
diff --git a/hpvm/test/regressionTests/BuildDFG/ThreeLevelEdge.ll b/hpvm/test/regressionTests/BuildDFG/ThreeLevelEdge.ll
index 4e97523049..8c76da27e8 100644
--- a/hpvm/test/regressionTests/BuildDFG/ThreeLevelEdge.ll
+++ b/hpvm/test/regressionTests/BuildDFG/ThreeLevelEdge.ll
@@ -1,10 +1,17 @@
-; RUN: opt -load LLVMGenHPVM.so -S -genhpvm <  %s | FileCheck %s
-; ModuleID = 'ThreeLevelEdge.c'
+; RUN: opt -load LLVMBuildDFG.so -S <  %s | FileCheck %s
+; ModuleID = 'ThreeLevelEdge.ll'
 source_filename = "ThreeLevelEdge.c"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 %struct.Root = type { i32*, i32*, i32* }
+%struct.out.Func1 = type <{ i32* }>
+%struct.out.Func4 = type <{ i32* }>
+%struct.out.Func5 = type <{ i32* }>
+%struct.out.Func3 = type <{ i32* }>
+%struct.out.Func2 = type <{ i32* }>
+%struct.out.PipeRoot = type <{ i32* }>
+
 ; CHECK-LABEL: struct.Root =
 
 ; CHECK-LABEL: %struct.out.Func1 =
@@ -54,99 +61,28 @@ target triple = "x86_64-unknown-linux-gnu"
 ; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func2_cloned.node,
 
 
-; Function Attrs: nounwind uwtable
-define dso_local void @Func1(i32* %In, i32* %Out) #0 {
-; CHECK-NOT: @Func1(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 1, i32* %In, i32 1, i32* %Out) #3
-  tail call void (i32, ...) @__hpvm__return(i32 1, i32* %Out) #3
-  ret void
-}
-
-declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #1
-
-declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #1
+declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
 
-declare dso_local void @__hpvm__return(i32, ...) local_unnamed_addr #1
-
-; Function Attrs: nounwind uwtable
-define dso_local void @Func4(i32* %In, i32* %Out) #0 {
-; CHECK-NOT: @Func4(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 1, i32* %In, i32 1, i32* %Out) #3
-  tail call void (i32, ...) @__hpvm__return(i32 1, i32* %Out) #3
-  ret void
-}
+declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
 
-; Function Attrs: nounwind uwtable
-define dso_local void @Func5(i32* %In1, i32* %In2, i32* %Out) #0 {
-; CHECK-NOT: @Func3(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 2, i32* %In1, i32* %In2, i32 1, i32* %Out) #3
-  tail call void (i32, ...) @__hpvm__return(i32 1, i32* %Out) #3
-  ret void
-}
-
-; Function Attrs: nounwind uwtable
-define dso_local void @Func3(i32* %In, i32* %Out) #0 {
-; CHECK-NOT: @Func3(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 1, i32* %In, i32 1, i32* %Out) #3
-  %call = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 2, void (i32*, i32*)* nonnull @Func4, i64 3, i64 6) #3
-  %call1 = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 2, void (i32*, i32*, i32*)* nonnull @Func5, i64 4, i64 5) #3
-  tail call void @__hpvm__bindIn(i8* %call, i32 0, i32 0, i32 0) #3
-  %call2 = tail call i8* @__hpvm__edge(i8* %call, i8* %call1, i32 0, i32 1, i32 1, i32 0) #3
-  tail call void @__hpvm__bindOut(i8* %call1, i32 0, i32 0, i32 0) #3
-  ret void
-}
+declare dso_local void @__hpvm__return(i32, ...) local_unnamed_addr #0
 
 ; Function Attrs: argmemonly nounwind
-declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #2
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
 
-declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #1
+declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
 
-declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #1
+declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
 
-declare dso_local i8* @__hpvm__edge(i8*, i8*, i32, i32, i32, i32) local_unnamed_addr #1
+declare dso_local i8* @__hpvm__edge(i8*, i8*, i32, i32, i32, i32) local_unnamed_addr #0
 
-declare dso_local void @__hpvm__bindOut(i8*, i32, i32, i32) local_unnamed_addr #1
+declare dso_local void @__hpvm__bindOut(i8*, i32, i32, i32) local_unnamed_addr #0
 
 ; Function Attrs: argmemonly nounwind
-declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #2
-
-; Function Attrs: nounwind uwtable
-define dso_local void @Func2(i32* %BindIn, i32* %SrcIn, i32* %Out) #0 {
-; CHECK-NOT: @Func2(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 2, i32* %BindIn, i32* %SrcIn, i32 1, i32* %Out) #3
-  %call = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 1, void (i32*, i32*)* nonnull @Func3, i64 3) #3
-  tail call void @__hpvm__bindIn(i8* %call, i32 0, i32 0, i32 0) #3
-  tail call void @__hpvm__bindOut(i8* %call, i32 0, i32 0, i32 0) #3
-  ret void
-}
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
 
 ; Function Attrs: nounwind uwtable
-define dso_local void @PipeRoot(i32* %In1, i32* %In2, i32* %Out) #0 {
-; CHECK-NOT: @PipeRoot(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 2, i32* %In1, i32* %In2, i32 1, i32* %Out) #3
-  %call = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 0, void (i32*, i32*)* nonnull @Func1) #3
-  %call1 = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 0, void (i32*, i32*, i32*)* nonnull @Func2) #3
-  tail call void @__hpvm__bindIn(i8* %call, i32 0, i32 0, i32 0) #3
-  tail call void @__hpvm__bindIn(i8* %call1, i32 1, i32 0, i32 0) #3
-  %call2 = tail call i8* @__hpvm__edge(i8* %call, i8* %call1, i32 0, i32 1, i32 1, i32 0) #3
-  tail call void @__hpvm__bindOut(i8* %call1, i32 0, i32 0, i32 0) #3
-  ret void
-}
-
-; Function Attrs: nounwind uwtable
-define dso_local i32 @main() local_unnamed_addr #0 {
+define dso_local i32 @main() local_unnamed_addr #2 {
 entry:
   %In1 = alloca i32, align 4
   %In2 = alloca i32, align 4
@@ -154,25 +90,26 @@ entry:
   %RootArgs = alloca %struct.Root, align 8
   %0 = bitcast i32* %In1 to i8*
   call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %0) #3
-  store i32 1, i32* %In1, align 4, !tbaa !2
+  store i32 1, i32* %In1, align 4, !tbaa !8
   %1 = bitcast i32* %In2 to i8*
   call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %1) #3
-  store i32 2, i32* %In2, align 4, !tbaa !2
+  store i32 2, i32* %In2, align 4, !tbaa !8
   %2 = bitcast i32* %Out to i8*
   call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %2) #3
-  store i32 0, i32* %Out, align 4, !tbaa !2
+  store i32 0, i32* %Out, align 4, !tbaa !8
   %3 = bitcast %struct.Root* %RootArgs to i8*
   call void @llvm.lifetime.start.p0i8(i64 24, i8* nonnull %3) #3
   %input1 = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 0
-  store i32* %In1, i32** %input1, align 8, !tbaa !6
+  store i32* %In1, i32** %input1, align 8, !tbaa !12
   %intput2 = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
-  store i32* %In2, i32** %intput2, align 8, !tbaa !9
+  store i32* %In2, i32** %intput2, align 8, !tbaa !15
   %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
-  store i32* %Out, i32** %output, align 8, !tbaa !10
-  call void (...) @__hpvm__init() #3
-  %call = call i8* (i32, ...) @__hpvm__launch(i32 0, void (i32*, i32*, i32*)* nonnull @PipeRoot, %struct.Root* nonnull %RootArgs) #3
-  call void @__hpvm__wait(i8* %call) #3
-  call void (...) @__hpvm__cleanup() #3
+  store i32* %Out, i32** %output, align 8, !tbaa !16
+  call void @llvm.hpvm.init()
+  %4 = bitcast %struct.Root* %RootArgs to i8*
+  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%struct.out.PipeRoot (i32*, i32*, i32*)* @PipeRoot_cloned to i8*), i8* %4, i1 false)
+  call void @llvm.hpvm.wait(i8* %graphID)
+  call void @llvm.hpvm.cleanup()
   call void @llvm.lifetime.end.p0i8(i64 24, i8* nonnull %3) #3
   call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %2) #3
   call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %1) #3
@@ -180,30 +117,132 @@ entry:
   ret i32 0
 }
 
-declare dso_local void @__hpvm__init(...) local_unnamed_addr #1
+declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
+
+declare i8* @llvm_hpvm_initializeTimerSet()
+
+declare void @llvm_hpvm_switchToTimer(i8**, i32)
+
+declare void @llvm_hpvm_printTimerSet(i8**, i8*)
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.Func1 @Func1_cloned(i32* in %In, i32* out %Out) #2 {
+entry:
+  %returnStruct = insertvalue %struct.out.Func1 undef, i32* %Out, 0
+  ret %struct.out.Func1 %returnStruct
+}
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.Func4 @Func4_cloned(i32* in %In, i32* out %Out) #2 {
+entry:
+  %returnStruct = insertvalue %struct.out.Func4 undef, i32* %Out, 0
+  ret %struct.out.Func4 %returnStruct
+}
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.Func5 @Func5_cloned(i32* in %In1, i32* in %In2, i32* out %Out) #2 {
+entry:
+  %returnStruct = insertvalue %struct.out.Func5 undef, i32* %Out, 0
+  ret %struct.out.Func5 %returnStruct
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode2D(i8*, i64, i64) #3
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #3
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createEdge(i8*, i8*, i1, i32, i32, i1) #3
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.output(i8*, i32, i32, i1) #3
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.Func3 @Func3_cloned(i32* in %In, i32* out %Out) #2 {
+entry:
+  %Func4_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%struct.out.Func4 (i32*, i32*)* @Func4_cloned to i8*), i64 3, i64 6)
+  %Func5_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%struct.out.Func5 (i32*, i32*, i32*)* @Func5_cloned to i8*), i64 4, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func4_cloned.node, i32 0, i32 0, i1 false)
+  %output = call i8* @llvm.hpvm.createEdge(i8* %Func4_cloned.node, i8* %Func5_cloned.node, i1 false, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.output(i8* %Func5_cloned.node, i32 0, i32 0, i1 false)
+  ret %struct.out.Func3 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode1D(i8*, i64) #3
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.Func2 @Func2_cloned(i32* in %BindIn, i32* in %SrcIn, i32* out %Out) #2 {
+entry:
+  %Func3_cloned.node = call i8* @llvm.hpvm.createNode1D(i8* bitcast (%struct.out.Func3 (i32*, i32*)* @Func3_cloned to i8*), i64 3)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.output(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  ret %struct.out.Func2 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode(i8*) #3
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.PipeRoot @PipeRoot_cloned(i32* in %In1, i32* in %In2, i32* out %Out) #2 {
+entry:
+  %Func1_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func1 (i32*, i32*)* @Func1_cloned to i8*))
+  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func2 (i32*, i32*, i32*)* @Func2_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 0, i1 false)
+  %output = call i8* @llvm.hpvm.createEdge(i8* %Func1_cloned.node, i8* %Func2_cloned.node, i1 false, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.output(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
+  ret %struct.out.PipeRoot undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.init() #3
 
-declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #1
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #3
 
-declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #1
+; Function Attrs: nounwind
+declare void @llvm.hpvm.wait(i8*) #3
 
-declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #1
+; Function Attrs: nounwind
+declare void @llvm.hpvm.cleanup() #3
 
-attributes #0 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
-attributes #1 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
-attributes #2 = { argmemonly nounwind }
+attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
 attributes #3 = { nounwind }
 
 !llvm.module.flags = !{!0}
 !llvm.ident = !{!1}
+!hpvm_hint_cpu = !{!2, !3, !4, !5, !6, !7}
+!hpvm_hint_gpu = !{}
+!hpvm_hint_spir = !{}
+!hpvm_hint_cudnn = !{}
+!hpvm_hint_promise = !{}
+!hpvm_hint_cpu_gpu = !{}
+!hpvm_hint_cpu_spir = !{}
 
 !0 = !{i32 1, !"wchar_size", i32 4}
 !1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 6690f9e7e8b46b96aea222d3e85315cd63545953)"}
-!2 = !{!3, !3, i64 0}
-!3 = !{!"int", !4, i64 0}
-!4 = !{!"omnipotent char", !5, i64 0}
-!5 = !{!"Simple C/C++ TBAA"}
-!6 = !{!7, !8, i64 0}
-!7 = !{!"Root", !8, i64 0, !8, i64 8, !8, i64 16}
-!8 = !{!"any pointer", !4, i64 0}
-!9 = !{!7, !8, i64 8}
-!10 = !{!7, !8, i64 16}
+!2 = !{%struct.out.Func1 (i32*, i32*)* @Func1_cloned}
+!3 = !{%struct.out.Func4 (i32*, i32*)* @Func4_cloned}
+!4 = !{%struct.out.Func5 (i32*, i32*, i32*)* @Func5_cloned}
+!5 = !{%struct.out.Func3 (i32*, i32*)* @Func3_cloned}
+!6 = !{%struct.out.Func2 (i32*, i32*, i32*)* @Func2_cloned}
+!7 = !{%struct.out.PipeRoot (i32*, i32*, i32*)* @PipeRoot_cloned}
+!8 = !{!9, !9, i64 0}
+!9 = !{!"int", !10, i64 0}
+!10 = !{!"omnipotent char", !11, i64 0}
+!11 = !{!"Simple C/C++ TBAA"}
+!12 = !{!13, !14, i64 0}
+!13 = !{!"Root", !14, i64 0, !14, i64 8, !14, i64 16}
+!14 = !{!"any pointer", !10, i64 0}
+!15 = !{!13, !14, i64 8}
+!16 = !{!13, !14, i64 16}
diff --git a/hpvm/test/regressionTests/BuildDFG/TwoLaunch.genvisc.ll b/hpvm/test/regressionTests/BuildDFG/TwoLaunch.ll
similarity index 100%
rename from hpvm/test/regressionTests/BuildDFG/TwoLaunch.genvisc.ll
rename to hpvm/test/regressionTests/BuildDFG/TwoLaunch.ll
diff --git a/hpvm/test/regressionTests/BuildDFG/TwoNode.ll b/hpvm/test/regressionTests/BuildDFG/TwoNode.ll
deleted file mode 100644
index 49b9b35205..0000000000
--- a/hpvm/test/regressionTests/BuildDFG/TwoNode.ll
+++ /dev/null
@@ -1,197 +0,0 @@
-; RUN: opt -load LLVMBuildDFG.so -S < %s | FileCheck %s
-; ModuleID = 'CreateNodeAndEdge.ll'
-source_filename = "CreateNodeAndEdge.c"
-target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
-target triple = "x86_64-unknown-linux-gnu"
-
-%struct.Root = type { i32*, i32*, i32* }
-%struct.out.Func1 = type <{ i32* }>
-%struct.out.Func2 = type <{ i32* }>
-%struct.out.PipeRoot = type <{ i32* }>
-
-; CHECK-LABEL: struct.Root =
-
-; CHECK-LABEL: %struct.out.Func1 =
-; CHECK-LABEL: %struct.out.Func3 =
-; CHECK-LABEL: %struct.out.Func2 =
-; CHECK-LABEL: %struct.out.PipeRoot =
-
-; CHECK-LABEL: i32 @main(
-; CHECK: [[ALLOCA:%[1-9a-zA-Z]+]] = alloca %struct.Root
-; CHECK: call void @llvm.hpvm.init()
-; CHECK:  [[REGISTER:%[1-9]+]] = bitcast %struct.Root* [[ALLOCA]] to i8*
-; CHECK: call i8* @llvm.hpvm.launch(i8* bitcast (%struct.out.PipeRoot (i32*, i32*, i32*)* @PipeRoot_cloned to i8*), i8* %4,
-; CHECK-NEXT: call void @llvm.hpvm.wait(i8*
-
-; CHECK-LABEL: @Func1_cloned(
-; CHECK: [[RET1:%[1-9a-zA-Z]+]] = insertvalue %struct.out.Func1 undef,
-; CHECK-NEXT: ret %struct.out.Func1 [[RET1]]
-
-; CHECK-LABEL: @Func3_cloned(
-; CHECK: [[RET3:%[1-9a-zA-Z]+]] = insertvalue %struct.out.Func3 undef,
-; CHECK-NEXT: ret %struct.out.Func3 [[RET3]]
-
-; CHECK-LABEL: @Func2_cloned(
-; CHECK: %Func3_cloned.node = call i8* @llvm.hpvm.createNode1D(i8* bitcast (%struct.out.Func3 (i32*
-; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node,
-; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func3_cloned.node,
-
-; CHECK-LABEL: @PipeRoot_cloned(i32*
-; CHECK: %Func1_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func1 (i32*
-; CHECK-NEXT: %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func2 (i32*
-; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node,
-; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node,
-; CHECK-NEXT: call i8* @llvm.hpvm.createEdge(i8* %Func1_cloned.node, i8* %Func2_cloned.node,
-; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func2_cloned.node,
-
-
-
-declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__return(i32, ...) local_unnamed_addr #0
-
-; Function Attrs: argmemonly nounwind
-declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
-
-declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
-
-declare dso_local i8* @__hpvm__edge(i8*, i8*, i32, i32, i32, i32) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__bindOut(i8*, i32, i32, i32) local_unnamed_addr #0
-
-; Function Attrs: argmemonly nounwind
-declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
-
-; Function Attrs: nounwind uwtable
-define dso_local i32 @main() local_unnamed_addr #2 {
-entry:
-  %In1 = alloca i32, align 4
-  %In2 = alloca i32, align 4
-  %Out = alloca i32, align 4
-  %RootArgs = alloca %struct.Root, align 8
-  %0 = bitcast i32* %In1 to i8*
-  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %0) #3
-  store i32 1, i32* %In1, align 4, !tbaa !5
-  %1 = bitcast i32* %In2 to i8*
-  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %1) #3
-  store i32 2, i32* %In2, align 4, !tbaa !5
-  %2 = bitcast i32* %Out to i8*
-  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %2) #3
-  store i32 0, i32* %Out, align 4, !tbaa !5
-  %3 = bitcast %struct.Root* %RootArgs to i8*
-  call void @llvm.lifetime.start.p0i8(i64 24, i8* nonnull %3) #3
-  %input1 = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 0
-  store i32* %In1, i32** %input1, align 8, !tbaa !9
-  %input2 = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
-  store i32* %In2, i32** %input2, align 8, !tbaa !12
-  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
-  store i32* %Out, i32** %output, align 8, !tbaa !13
-  call void @llvm.hpvm.init()
-  %4 = bitcast %struct.Root* %RootArgs to i8*
-  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%struct.out.PipeRoot (i32*, i32*, i32*)* @PipeRoot_cloned to i8*), i8* %4, i1 false)
-  call void @llvm.hpvm.wait(i8* %graphID)
-  call void @llvm.hpvm.cleanup()
-  call void @llvm.lifetime.end.p0i8(i64 24, i8* nonnull %3) #3
-  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %2) #3
-  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %1) #3
-  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %0) #3
-  ret i32 0
-}
-
-declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
-
-declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
-
-declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
-
-declare i8* @llvm_hpvm_initializeTimerSet()
-
-declare void @llvm_hpvm_switchToTimer(i8**, i32)
-
-declare void @llvm_hpvm_printTimerSet(i8**, i8*)
-
-; Function Attrs: nounwind uwtable
-define dso_local %struct.out.Func1 @Func1_cloned(i32* in %In, i32* out %Out) #2 {
-entry:
-  %returnStruct = insertvalue %struct.out.Func1 undef, i32* %Out, 0
-  ret %struct.out.Func1 %returnStruct
-}
-
-; Function Attrs: nounwind uwtable
-define dso_local %struct.out.Func2 @Func2_cloned(i32* in %BindIn, i32* in %SrcIn, i32* out %Out) #2 {
-entry:
-  %returnStruct = insertvalue %struct.out.Func2 undef, i32* %Out, 0
-  ret %struct.out.Func2 %returnStruct
-}
-
-; Function Attrs: nounwind
-declare i8* @llvm.hpvm.createNode(i8*) #3
-
-; Function Attrs: nounwind
-declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #3
-
-; Function Attrs: nounwind
-declare i8* @llvm.hpvm.createEdge(i8*, i8*, i1, i32, i32, i1) #3
-
-; Function Attrs: nounwind
-declare void @llvm.hpvm.bind.output(i8*, i32, i32, i1) #3
-
-; Function Attrs: nounwind uwtable
-define dso_local %struct.out.PipeRoot @PipeRoot_cloned(i32* in %In1, i32* in %In2, i32* out %Out) #2 {
-entry:
-  %Func1_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func1 (i32*, i32*)* @Func1_cloned to i8*))
-  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func2 (i32*, i32*, i32*)* @Func2_cloned to i8*))
-  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 0, i32 0, i1 false)
-  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 0, i1 false)
-  %output = call i8* @llvm.hpvm.createEdge(i8* %Func1_cloned.node, i8* %Func2_cloned.node, i1 false, i32 1, i32 1, i1 false)
-  call void @llvm.hpvm.bind.output(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
-  ret %struct.out.PipeRoot undef
-}
-
-; Function Attrs: nounwind
-declare void @llvm.hpvm.init() #3
-
-; Function Attrs: nounwind
-declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #3
-
-; Function Attrs: nounwind
-declare void @llvm.hpvm.wait(i8*) #3
-
-; Function Attrs: nounwind
-declare void @llvm.hpvm.cleanup() #3
-
-attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
-attributes #1 = { argmemonly nounwind }
-attributes #2 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
-attributes #3 = { nounwind }
-
-!llvm.module.flags = !{!0}
-!llvm.ident = !{!1}
-!hpvm_hint_cpu = !{!2, !3, !4}
-!hpvm_hint_gpu = !{}
-!hpvm_hint_spir = !{}
-!hpvm_hint_cudnn = !{}
-!hpvm_hint_promise = !{}
-!hpvm_hint_cpu_gpu = !{}
-!hpvm_hint_cpu_spir = !{}
-
-!0 = !{i32 1, !"wchar_size", i32 4}
-!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 6690f9e7e8b46b96aea222d3e85315cd63545953)"}
-!2 = !{%struct.out.Func1 (i32*, i32*)* @Func1_cloned}
-!3 = !{%struct.out.Func2 (i32*, i32*, i32*)* @Func2_cloned}
-!4 = !{%struct.out.PipeRoot (i32*, i32*, i32*)* @PipeRoot_cloned}
-!5 = !{!6, !6, i64 0}
-!6 = !{!"int", !7, i64 0}
-!7 = !{!"omnipotent char", !8, i64 0}
-!8 = !{!"Simple C/C++ TBAA"}
-!9 = !{!10, !11, i64 0}
-!10 = !{!"Root", !11, i64 0, !11, i64 8, !11, i64 16}
-!11 = !{!"any pointer", !7, i64 0}
-!12 = !{!10, !11, i64 8}
-!13 = !{!10, !11, i64 16}
diff --git a/hpvm/test/regressionTests/ClearDFG/ThreeLevel.constmem.nvptxTox86.ll b/hpvm/test/regressionTests/ClearDFG/ThreeLevel.constmem.nvptxTox86.ll
new file mode 100644
index 0000000000..f6f0af5433
--- /dev/null
+++ b/hpvm/test/regressionTests/ClearDFG/ThreeLevel.constmem.nvptxTox86.ll
@@ -0,0 +1,276 @@
+; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_NVPTX.so -load LLVMDFG2LLVM_X86.so -load LLVMClearDFG.so -S -localmem -dfg2llvm-nvptx -dfg2llvm-x86 -clearDFG <  %s | FileCheck %s
+; ModuleID = 'ThreeLevel.opt.ll'
+source_filename = "ThreeLevel.opt.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%struct.Root = type { i32*, i64, i32*, i64 }
+%struct.out.Allocation = type <{ i8*, i64 }>
+%emptyStruct = type <{}>
+%emptyStruct.0 = type <{}>
+%emptyStruct.1 = type <{}>
+%emptyStruct.2 = type <{}>
+
+declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__getNode(...) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__getParentNode(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_y(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_y(i8*) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__malloc(i64) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__return(i32, ...) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__edge(i8*, i8*, i32, i32, i32, i32) local_unnamed_addr #0
+
+; CHECK-LABEL: @Launch(
+; CHECK: call i8* @llvm_hpvm_x86_launch(i8*
+; CHECK-NOT: call i8* @llvm.hpvm.launch(i8*
+; CHECK: call void @llvm_hpvm_x86_wait(i8*
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local void @Launch() local_unnamed_addr #2 {
+entry:
+  %RootArgs = alloca %struct.Root, align 8
+  %0 = bitcast %struct.Root* %RootArgs to i8*
+  call void @llvm.lifetime.start.p0i8(i64 32, i8* nonnull %0) #6
+  %call = tail call noalias i8* @malloc(i64 1024) #6
+  %1 = bitcast %struct.Root* %RootArgs to i8**
+  store i8* %call, i8** %1, align 8, !tbaa !6
+  %Insize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
+  store i64 1024, i64* %Insize, align 8, !tbaa !12
+  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
+  %call1 = tail call noalias i8* @malloc(i64 1024) #6
+  %2 = bitcast i32** %output to i8**
+  store i8* %call1, i8** %2, align 8, !tbaa !13
+  %Outsize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 3
+  store i64 1024, i64* %Outsize, align 8, !tbaa !14
+  %3 = bitcast %struct.Root* %RootArgs to i8*
+  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned to i8*), i8* %3, i1 false)
+  call void @llvm.hpvm.wait(i8* %graphID)
+  call void @llvm.lifetime.end.p0i8(i64 32, i8* nonnull %0) #6
+  ret void
+}
+
+; Function Attrs: nofree nounwind
+declare dso_local noalias i8* @malloc(i64) local_unnamed_addr #3
+
+declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
+
+; CHECK-LABEL: @main(
+; CHECK-NOT: call void @llvm.hpvm.init(
+; CHECK-NOT: call void @llvm.hpvm.cleanup(
+; CHECK: call i8* @llvm_hpvm_ocl_initContext(i32
+; CHECK: call i8* @llvm_hpvm_ocl_launch(i8*
+; CHECK: call void @llvm_hpvm_ocl_clearContext(i8*
+
+; CHECK-LABEL: @Func2_cloned.3_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK: call i8* @llvm_hpvm_ocl_argument_ptr(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_shared(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_executeNode(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_free(i8*
+
+; CHECK-LABEL: @PipeRoot_cloned.4(
+; CHECK: call void @llvm_hpvm_x86_dstack_push(
+; CHECK-NEXT: @Func2_cloned.3_cloned_cloned_cloned_cloned_cloned_clone
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop(
+
+; CHECK-LABEL: define i8* @LaunchDataflowGraph(i8*
+
+
+; Function Attrs: nounwind uwtable
+define dso_local i32 @main() local_unnamed_addr #4 {
+entry:
+  call void @llvm.hpvm.init()
+  tail call void @Launch()
+  call void @llvm.hpvm.cleanup()
+  ret i32 0
+}
+
+declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
+
+declare i8* @llvm_hpvm_initializeTimerSet()
+
+declare void @llvm_hpvm_switchToTimer(i8**, i32)
+
+declare void @llvm_hpvm_printTimerSet(i8**, i8*)
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getNode() #5
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getParentNode(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.y(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.y(i8*) #5
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.malloc(i64) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.Allocation @Allocation_cloned(i64 %block) #4 {
+entry:
+  %call1 = call i8* @llvm.hpvm.malloc(i64 %block)
+  %returnStruct = insertvalue %struct.out.Allocation undef, i8* %call1, 0
+  %returnStruct2 = insertvalue %struct.out.Allocation %returnStruct, i64 %block, 1
+  ret %struct.out.Allocation %returnStruct2
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode2D(i8*, i64, i64) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct @Func1_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+; CHECK-NOT: @Func1_cloned(
+entry:
+  %call4 = call i8* @llvm.hpvm.getNode()
+  %call15 = call i8* @llvm.hpvm.getParentNode(i8* %call4)
+  %call26 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call4)
+  %call37 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call4)
+  %call58 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call15)
+  %call79 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call15)
+  %call910 = call i64 @llvm.hpvm.getNumNodeInstances.x(i8* %call4)
+  %call1111 = call i64 @llvm.hpvm.getNumNodeInstances.y(i8* %call4)
+  %mul = mul i64 %call910, %call58
+  %add = add i64 %mul, %call26
+  %arrayidx = getelementptr inbounds i32, i32* %In, i64 3
+  %0 = load i32, i32* %arrayidx, align 4, !tbaa !15
+  %sext = shl i64 %add, 32
+  %idxprom = ashr exact i64 %sext, 32
+  %arrayidx15 = getelementptr inbounds i32, i32* %Out, i64 %idxprom
+  %1 = load i32, i32* %arrayidx15, align 4, !tbaa !15
+  %add16 = add nsw i32 %1, %0
+  store i32 %add16, i32* %arrayidx15, align 4, !tbaa !15
+  ret %emptyStruct undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode(i8*) #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #6
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createEdge(i8*, i8*, i1, i32, i32, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.0 @Func3_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+; CHECK-NOT: @Func3_cloned(
+entry:
+  %Func1_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned to i8*), i64 3, i64 5)
+  %Allocation_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Allocation (i64)* @Allocation_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Allocation_cloned.node, i32 1, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 3, i32 3, i1 false)
+  %output = call i8* @llvm.hpvm.createEdge(i8* %Allocation_cloned.node, i8* %Func1_cloned.node, i1 true, i32 0, i32 0, i1 false)
+  %output1 = call i8* @llvm.hpvm.createEdge(i8* %Allocation_cloned.node, i8* %Func1_cloned.node, i1 true, i32 1, i32 1, i1 false)
+  ret %emptyStruct.0 undef
+}
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.1 @Func2_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+; CHECK-NOT: @Func2_cloned(
+entry:
+  %Func3_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.1 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.2 @PipeRoot_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+; CHECK-NOT: @PipeRoot_cloned(
+entry:
+  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.2 undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.wait(i8*) #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.init() #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.cleanup() #6
+
+attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { noinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #3 = { nofree nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #4 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #5 = { nounwind readnone }
+attributes #6 = { nounwind }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+!hpvm_hint_gpu = !{!2}
+!hpvm_hint_cpu = !{!3, !4, !5}
+!hpvm_hint_spir = !{}
+!hpvm_hint_cudnn = !{}
+!hpvm_hint_promise = !{}
+!hpvm_hint_cpu_gpu = !{}
+!hpvm_hint_cpu_spir = !{}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 6690f9e7e8b46b96aea222d3e85315cd63545953)"}
+!2 = !{%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned}
+!3 = !{%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned}
+!4 = !{%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned}
+!5 = !{%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned}
+!6 = !{!7, !8, i64 0}
+!7 = !{!"Root", !8, i64 0, !11, i64 8, !8, i64 16, !11, i64 24}
+!8 = !{!"any pointer", !9, i64 0}
+!9 = !{!"omnipotent char", !10, i64 0}
+!10 = !{!"Simple C/C++ TBAA"}
+!11 = !{!"long", !9, i64 0}
+!12 = !{!7, !11, i64 8}
+!13 = !{!7, !8, i64 16}
+!14 = !{!7, !11, i64 24}
+!15 = !{!16, !16, i64 0}
+!16 = !{!"int", !9, i64 0}
diff --git a/hpvm/test/regressionTests/ClearDFG/ThreeLevel.nvptxTox86.ll b/hpvm/test/regressionTests/ClearDFG/ThreeLevel.nvptxTox86.ll
new file mode 100644
index 0000000000..163c4fc93c
--- /dev/null
+++ b/hpvm/test/regressionTests/ClearDFG/ThreeLevel.nvptxTox86.ll
@@ -0,0 +1,256 @@
+; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_NVPTX.so -load LLVMDFG2LLVM_X86.so -load LLVMClearDFG.so -S -localmem -dfg2llvm-nvptx -dfg2llvm-x86 -clearDFG <  %s | FileCheck %s
+; ModuleID = 'ThreeLevel.ll'
+source_filename = "ThreeLevel.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%struct.Root = type { i32*, i64, i32*, i64 }
+%emptyStruct = type <{}>
+%emptyStruct.0 = type <{}>
+%emptyStruct.1 = type <{}>
+%emptyStruct.2 = type <{}>
+
+declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__getNode(...) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__getParentNode(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_y(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_y(i8*) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
+
+; CHECK-LABEL: @Launch(
+; CHECK: call i8* @llvm_hpvm_x86_launch(i8*
+; CHECK-NOT: call i8* @llvm.hpvm.launch(i8*
+; CHECK: call void @llvm_hpvm_x86_wait(i8*
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local void @Launch() local_unnamed_addr #2 {
+entry:
+  %RootArgs = alloca %struct.Root, align 8
+  %0 = bitcast %struct.Root* %RootArgs to i8*
+  call void @llvm.lifetime.start.p0i8(i64 32, i8* nonnull %0) #6
+  %call = tail call noalias i8* @malloc(i64 1024) #6
+  %1 = bitcast %struct.Root* %RootArgs to i8**
+  store i8* %call, i8** %1, align 8, !tbaa !6
+  %Insize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
+  store i64 1024, i64* %Insize, align 8, !tbaa !12
+  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
+  %call1 = tail call noalias i8* @malloc(i64 1024) #6
+  %2 = bitcast i32** %output to i8**
+  store i8* %call1, i8** %2, align 8, !tbaa !13
+  %Outsize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 3
+  store i64 1024, i64* %Outsize, align 8, !tbaa !14
+  %3 = bitcast %struct.Root* %RootArgs to i8*
+  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned to i8*), i8* %3, i1 false)
+  call void @llvm.hpvm.wait(i8* %graphID)
+  call void @llvm.lifetime.end.p0i8(i64 32, i8* nonnull %0) #6
+  ret void
+}
+
+; Function Attrs: nofree nounwind
+declare dso_local noalias i8* @malloc(i64) local_unnamed_addr #3
+
+declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
+
+; CHECK-LABEL: @main(
+; CHECK-NOT: call void @llvm.hpvm.init(
+; CHECK-NOT: call void @llvm.hpvm.cleanup(
+; CHECK: call i8* @llvm_hpvm_ocl_initContext(i32
+; CHECK: call i8* @llvm_hpvm_ocl_launch(i8*
+; CHECK: call void @llvm_hpvm_ocl_clearContext(i8*
+
+; CHECK-LABEL: @Func2_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK: call i8* @llvm_hpvm_ocl_argument_ptr(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_argument_ptr(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_executeNode(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_free(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_free(i8*
+
+; CHECK-LABEL: @PipeRoot_cloned.3(
+; CHECK: call void @llvm_hpvm_x86_dstack_push(
+; CHECK-NEXT: @Func2_cloned.2_cloned_cloned_cloned_cloned_cloned_clone
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop(
+
+; CHECK-LABEL: define i8* @LaunchDataflowGraph(i8*
+
+; Function Attrs: nounwind uwtable
+define dso_local i32 @main() local_unnamed_addr #4 {
+entry:
+  call void @llvm.hpvm.init()
+  tail call void @Launch()
+  call void @llvm.hpvm.cleanup()
+  ret i32 0
+}
+
+declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
+
+declare i8* @llvm_hpvm_initializeTimerSet()
+
+declare void @llvm_hpvm_switchToTimer(i8**, i32)
+
+declare void @llvm_hpvm_printTimerSet(i8**, i8*)
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getNode() #5
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getParentNode(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.y(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.y(i8*) #5
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode2D(i8*, i64, i64) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct @Func1_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+; CHECK-NOT: @Func1_cloned(
+entry:
+  %call4 = call i8* @llvm.hpvm.getNode()
+  %call15 = call i8* @llvm.hpvm.getParentNode(i8* %call4)
+  %call26 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call4)
+  %call37 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call4)
+  %call58 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call15)
+  %call79 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call15)
+  %call910 = call i64 @llvm.hpvm.getNumNodeInstances.x(i8* %call4)
+  %call1111 = call i64 @llvm.hpvm.getNumNodeInstances.y(i8* %call4)
+  %mul = mul i64 %call910, %call58
+  %add = add i64 %mul, %call26
+  %mul13 = mul i64 %call1111, %call79
+  %add14 = add i64 %mul13, %call37
+  %sext = shl i64 %add14, 32
+  %idxprom = ashr exact i64 %sext, 32
+  %arrayidx = getelementptr inbounds i32, i32* %In, i64 %idxprom
+  %0 = load i32, i32* %arrayidx, align 4, !tbaa !15
+  %sext36 = shl i64 %add, 32
+  %idxprom15 = ashr exact i64 %sext36, 32
+  %arrayidx16 = getelementptr inbounds i32, i32* %Out, i64 %idxprom15
+  %1 = load i32, i32* %arrayidx16, align 4, !tbaa !15
+  %add17 = add nsw i32 %1, %0
+  store i32 %add17, i32* %arrayidx16, align 4, !tbaa !15
+  ret %emptyStruct undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.0 @Func3_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+; CHECK-NOT: @Func3_cloned(
+entry:
+  %Func1_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.0 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode(i8*) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.1 @Func2_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+; CHECK-NOT: @Func2_cloned(
+entry:
+  %Func3_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.1 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.2 @PipeRoot_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+; CHECK-NOT: @PipeRoot_cloned(
+entry:
+  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.2 undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.wait(i8*) #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.init() #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.cleanup() #6
+
+attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { noinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #3 = { nofree nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #4 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #5 = { nounwind readnone }
+attributes #6 = { nounwind }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+!hpvm_hint_gpu = !{!2}
+!hpvm_hint_cpu = !{!3, !4, !5}
+!hpvm_hint_spir = !{}
+!hpvm_hint_cudnn = !{}
+!hpvm_hint_promise = !{}
+!hpvm_hint_cpu_gpu = !{}
+!hpvm_hint_cpu_spir = !{}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 6690f9e7e8b46b96aea222d3e85315cd63545953)"}
+!2 = !{%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned}
+!3 = !{%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned}
+!4 = !{%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned}
+!5 = !{%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned}
+!6 = !{!7, !8, i64 0}
+!7 = !{!"Root", !8, i64 0, !11, i64 8, !8, i64 16, !11, i64 24}
+!8 = !{!"any pointer", !9, i64 0}
+!9 = !{!"omnipotent char", !10, i64 0}
+!10 = !{!"Simple C/C++ TBAA"}
+!11 = !{!"long", !9, i64 0}
+!12 = !{!7, !11, i64 8}
+!13 = !{!7, !8, i64 16}
+!14 = !{!7, !11, i64 24}
+!15 = !{!16, !16, i64 0}
+!16 = !{!"int", !9, i64 0}
diff --git a/hpvm/test/regressionTests/ClearDFG/ThreeLevel.x86.ll b/hpvm/test/regressionTests/ClearDFG/ThreeLevel.x86.ll
new file mode 100644
index 0000000000..2bd337e8ab
--- /dev/null
+++ b/hpvm/test/regressionTests/ClearDFG/ThreeLevel.x86.ll
@@ -0,0 +1,225 @@
+; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_X86.so -load LLVMClearDFG.so -S -dfg2llvm-x86 -clearDFG <  %s | FileCheck %s
+; ModuleID = 'ThreeLevel.ll'
+source_filename = "ThreeLevel.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%struct.Root = type { i32*, i64, i32*, i64, i32*, i64 }
+%struct.out.Func1 = type <{ i32* }>
+%struct.out.Func3 = type <{ i32* }>
+%struct.out.Func2 = type <{ i32* }>
+%struct.out.PipeRoot = type <{ i32* }>
+
+
+; CHECK-LABEL: i32 @main(
+; CHECK-NOT: call void @llvm.hpvm.init()
+; CHECK: call i8* @llvm_hpvm_x86_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
+; CHECK-NOT: call i8* @llvm.hpvm.launch(i8*
+; CHECK: call void @llvm_hpvm_x86_wait(i8*
+
+; CHECK-LABEL: @Func1_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned
+; CHECK: call i8* @llvm_hpvm_x86_argument_ptr(
+
+; CHECK-LABEL: @Func3_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK-LABEL: for.body1:
+; CHECK: %index.y = phi i64 [ 0, %for.body ], [ %index.y.inc, %for.body1 ]
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_push(
+; CHECK-NEXT: @Func1_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
+
+; CHECK-LABEL: @Func2_cloned.3_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK-LABEL: for.body:
+; CHECK-NEXT: %index.x = phi i64 [ 0, %entry ], [ %index.x.inc, %for.body ]
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_push(
+; CHECK-NEXT: @Func3_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
+
+; CHECK-LABEL: @PipeRoot_cloned.4(
+; CHECK: call void @llvm_hpvm_x86_dstack_push(
+; CHECK-NEXT: @Func2_cloned.3_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
+
+; CHECK-LABEL: @LaunchDataflowGraph(
+; CHECK: call %struct.out.PipeRoot @PipeRoot_cloned.4(
+
+declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__return(i32, ...) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__bindOut(i8*, i32, i32, i32) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
+
+; Function Attrs: nounwind uwtable
+define dso_local i32 @main() local_unnamed_addr #2 {
+entry:
+  %In1 = alloca i32, align 4
+  %In2 = alloca i32, align 4
+  %Out = alloca i32, align 4
+  %RootArgs = alloca %struct.Root, align 8
+  %0 = bitcast i32* %In1 to i8*
+  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %0) #3
+  store i32 1, i32* %In1, align 4, !tbaa !6
+  %1 = bitcast i32* %In2 to i8*
+  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %1) #3
+  store i32 2, i32* %In2, align 4, !tbaa !6
+  %2 = bitcast i32* %Out to i8*
+  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %2) #3
+  store i32 0, i32* %Out, align 4, !tbaa !6
+  %3 = bitcast %struct.Root* %RootArgs to i8*
+  call void @llvm.lifetime.start.p0i8(i64 48, i8* nonnull %3) #3
+  %input1 = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 0
+  store i32* %In1, i32** %input1, align 8, !tbaa !10
+  %Insize1 = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
+  store i64 32, i64* %Insize1, align 8, !tbaa !14
+  %input2 = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
+  store i32* %In2, i32** %input2, align 8, !tbaa !15
+  %Insize2 = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 3
+  store i64 32, i64* %Insize2, align 8, !tbaa !16
+  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 4
+  store i32* %Out, i32** %output, align 8, !tbaa !17
+  %Outsize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 5
+  store i64 32, i64* %Outsize, align 8, !tbaa !18
+  call void @llvm.hpvm.init()
+  %4 = bitcast %struct.Root* %RootArgs to i8*
+  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%struct.out.PipeRoot (i32*, i64, i32*, i64, i32*, i64)* @PipeRoot_cloned to i8*), i8* %4, i1 false)
+  call void @llvm.hpvm.wait(i8* %graphID)
+  call void @llvm.hpvm.cleanup()
+  call void @llvm.lifetime.end.p0i8(i64 48, i8* nonnull %3) #3
+  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %2) #3
+  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %1) #3
+  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %0) #3
+  ret i32 0
+}
+
+declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
+
+declare i8* @llvm_hpvm_initializeTimerSet()
+
+declare void @llvm_hpvm_switchToTimer(i8**, i32)
+
+declare void @llvm_hpvm_printTimerSet(i8**, i8*)
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.Func1 @Func1_cloned(i32* in %In, i64 %Insize, i32* out %Out, i64 %Outsize) #2 {
+entry:
+  %returnStruct = insertvalue %struct.out.Func1 undef, i32* %Out, 0
+  ret %struct.out.Func1 %returnStruct
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode2D(i8*, i64, i64) #3
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #3
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.output(i8*, i32, i32, i1) #3
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.Func3 @Func3_cloned(i32* in %In, i64 %Insize, i32* out %Out, i64 %Outsize) #2 {
+; CHECK-NOT: @Func3_cloned
+entry:
+  %Func1_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%struct.out.Func1 (i32*, i64, i32*, i64)* @Func1_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 3, i32 3, i1 false)
+  call void @llvm.hpvm.bind.output(i8* %Func1_cloned.node, i32 0, i32 0, i1 false)
+  ret %struct.out.Func3 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode1D(i8*, i64) #3
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.Func2 @Func2_cloned(i32* in %In, i64 %Insize, i32* out %Out, i64 %Outsize) #2 {
+; CHECK-NOT: @Func2_cloned
+entry:
+  %Func3_cloned.node = call i8* @llvm.hpvm.createNode1D(i8* bitcast (%struct.out.Func3 (i32*, i64, i32*, i64)* @Func3_cloned to i8*), i64 3)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 3, i32 3, i1 false)
+  call void @llvm.hpvm.bind.output(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  ret %struct.out.Func2 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode(i8*) #3
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.PipeRoot @PipeRoot_cloned(i32* in %In1, i64 %Insize1, i32* in %In2, i64 %InSize2, i32* out %Out, i64 %Outsize) #2 {
+; CHECK-NOT: @PipeRoot_cloned
+entry:
+  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func2 (i32*, i64, i32*, i64)* @Func2_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 3, i32 3, i1 false)
+  call void @llvm.hpvm.bind.output(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
+  ret %struct.out.PipeRoot undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.init() #3
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #3
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.wait(i8*) #3
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.cleanup() #3
+
+attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #3 = { nounwind }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+!hpvm_hint_cpu = !{!2, !3, !4, !5}
+!hpvm_hint_gpu = !{}
+!hpvm_hint_spir = !{}
+!hpvm_hint_cudnn = !{}
+!hpvm_hint_promise = !{}
+!hpvm_hint_cpu_gpu = !{}
+!hpvm_hint_cpu_spir = !{}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 6690f9e7e8b46b96aea222d3e85315cd63545953)"}
+!2 = !{%struct.out.Func1 (i32*, i64, i32*, i64)* @Func1_cloned}
+!3 = !{%struct.out.Func3 (i32*, i64, i32*, i64)* @Func3_cloned}
+!4 = !{%struct.out.Func2 (i32*, i64, i32*, i64)* @Func2_cloned}
+!5 = !{%struct.out.PipeRoot (i32*, i64, i32*, i64, i32*, i64)* @PipeRoot_cloned}
+!6 = !{!7, !7, i64 0}
+!7 = !{!"int", !8, i64 0}
+!8 = !{!"omnipotent char", !9, i64 0}
+!9 = !{!"Simple C/C++ TBAA"}
+!10 = !{!11, !12, i64 0}
+!11 = !{!"Root", !12, i64 0, !13, i64 8, !12, i64 16, !13, i64 24, !12, i64 32, !13, i64 40}
+!12 = !{!"any pointer", !8, i64 0}
+!13 = !{!"long", !8, i64 0}
+!14 = !{!11, !13, i64 8}
+!15 = !{!11, !12, i64 16}
+!16 = !{!11, !13, i64 24}
+!17 = !{!11, !12, i64 32}
+!18 = !{!11, !13, i64 40}
diff --git a/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.atomic.genvisc.ll b/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.atomic.ll
similarity index 99%
rename from hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.atomic.genvisc.ll
rename to hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.atomic.ll
index ea6ec14d10..451035b21e 100644
--- a/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.atomic.genvisc.ll
+++ b/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.atomic.ll
@@ -1,4 +1,4 @@
-; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_OpenCL.so -S -localmem -dfg2llvm-opencl <  %s | FileCheck %s
+; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_NVPTX.so -S -localmem -dfg2llvm-nvptx <  %s | FileCheck %s
 ; ModuleID = 'ThreeLevel.atomic.ll'
 source_filename = "ThreeLevel.constmem.c"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
diff --git a/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.cond.const.ll b/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.cond.const.ll
new file mode 100644
index 0000000000..3f275587c4
--- /dev/null
+++ b/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.cond.const.ll
@@ -0,0 +1,259 @@
+; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_NVPTX.so -S -localmem -dfg2llvm-nvptx <  %s | FileCheck %s
+; ModuleID = 'ThreeLevel.cond.genhpvm.gpu.ll'
+source_filename = "ThreeLevel.cond.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%struct.Root = type { i32*, i64, i32*, i64 }
+%emptyStruct = type <{}>
+%emptyStruct.0 = type <{}>
+%emptyStruct.1 = type <{}>
+%emptyStruct.2 = type <{}>
+
+declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__getNode(...) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__getParentNode(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_y(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_y(i8*) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
+
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local void @Launch() local_unnamed_addr #2 {
+entry:
+  %RootArgs = alloca %struct.Root, align 8
+  %0 = bitcast %struct.Root* %RootArgs to i8*
+  call void @llvm.lifetime.start.p0i8(i64 32, i8* nonnull %0) #6
+  %call = tail call noalias i8* @malloc(i64 1024) #6
+  %1 = bitcast %struct.Root* %RootArgs to i8**
+  store i8* %call, i8** %1, align 8, !tbaa !6
+  %Insize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
+  store i64 1024, i64* %Insize, align 8, !tbaa !12
+  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
+  %call1 = tail call noalias i8* @malloc(i64 1024) #6
+  %2 = bitcast i32** %output to i8**
+  store i8* %call1, i8** %2, align 8, !tbaa !13
+  %Outsize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 3
+  store i64 1024, i64* %Outsize, align 8, !tbaa !14
+  %3 = bitcast %struct.Root* %RootArgs to i8*
+  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned to i8*), i8* %3, i1 false)
+  call void @llvm.hpvm.wait(i8* %graphID)
+  call void @llvm.lifetime.end.p0i8(i64 32, i8* nonnull %0) #6
+  ret void
+}
+
+; Function Attrs: nofree nounwind
+declare dso_local noalias i8* @malloc(i64) local_unnamed_addr #3
+
+declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
+
+; CHECK-LABEL: @main(
+; CHECK: call i8* @llvm_hpvm_ocl_initContext(i32
+; CHECK: call i8* @llvm_hpvm_ocl_launch(i8*
+; CHECK: call void @llvm.hpvm.init(
+; CHECK: call void @llvm_hpvm_ocl_clearContext(i8*
+; CHECK: call void @llvm.hpvm.cleanup(
+
+; CHECK-LABEL: @Func1_cloned(
+; CHECK: br i1 %cmp, label %if.end, label %if.then
+; CHECK-LABEL: if.then:
+; CHECK-LABEL: if.end:
+
+; CHECK-LABEL: @Func2_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned
+; CHECK-NOT: call void @llvm_hpvm_ocl_argument_shared(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_argument_ptr(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK-NOT: call void @llvm_hpvm_ocl_argument_shared(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_argument_ptr(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_executeNode(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_free(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_free(i8*
+
+
+
+define dso_local i32 @main() local_unnamed_addr #4 {
+entry:
+  call void @llvm.hpvm.init()
+  tail call void @Launch()
+  call void @llvm.hpvm.cleanup()
+  ret i32 0
+}
+
+declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
+
+declare i8* @llvm_hpvm_initializeTimerSet()
+
+declare void @llvm_hpvm_switchToTimer(i8**, i32)
+
+declare void @llvm_hpvm_printTimerSet(i8**, i8*)
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getNode() #5
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getParentNode(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.y(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.y(i8*) #5
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode2D(i8*, i64, i64) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct @Func1_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %call4 = call i8* @llvm.hpvm.getNode()
+  %call15 = call i8* @llvm.hpvm.getParentNode(i8* %call4)
+  %call26 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call4)
+  %conv = trunc i64 %call26 to i32
+  %call37 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call4)
+  %conv4 = trunc i64 %call37 to i32
+  %call58 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call15)
+  %conv6 = trunc i64 %call58 to i32
+  %call79 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call15)
+  %conv8 = trunc i64 %call79 to i32
+  %call910 = call i64 @llvm.hpvm.getNumNodeInstances.x(i8* %call4)
+  %conv10 = trunc i64 %call910 to i32
+  %call1111 = call i64 @llvm.hpvm.getNumNodeInstances.y(i8* %call4)
+  %conv12 = trunc i64 %call1111 to i32
+  %mul = mul nsw i32 %conv10, %conv6
+  %add = add nsw i32 %mul, %conv
+  %mul13 = mul nsw i32 %conv12, %conv8
+  %add14 = add nsw i32 %mul13, %conv4
+  %cmp = icmp eq i32 %add, %add14
+  br i1 %cmp, label %if.end, label %if.then
+
+if.then:                                          ; preds = %entry
+  %arrayidx = getelementptr inbounds i32, i32* %In, i64 3
+  %0 = load i32, i32* %arrayidx, align 4, !tbaa !15
+  %idxprom = sext i32 %add to i64
+  %arrayidx16 = getelementptr inbounds i32, i32* %Out, i64 %idxprom
+  %1 = load i32, i32* %arrayidx16, align 4, !tbaa !15
+  %add17 = add nsw i32 %1, %0
+  store i32 %add17, i32* %arrayidx16, align 4, !tbaa !15
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %entry
+  ret %emptyStruct undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.0 @Func3_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func1_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.0 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode(i8*) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.1 @Func2_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func3_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.1 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.2 @PipeRoot_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.2 undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.wait(i8*) #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.init() #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.cleanup() #6
+
+attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { noinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #3 = { nofree nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #4 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #5 = { nounwind readnone }
+attributes #6 = { nounwind }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+!hpvm_hint_gpu = !{!2}
+!hpvm_hint_cpu = !{!3, !4, !5}
+!hpvm_hint_spir = !{}
+!hpvm_hint_cudnn = !{}
+!hpvm_hint_promise = !{}
+!hpvm_hint_cpu_gpu = !{}
+!hpvm_hint_cpu_spir = !{}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 3551132592a00cab6c966df508ab511598269f78)"}
+!2 = !{%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned}
+!3 = !{%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned}
+!4 = !{%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned}
+!5 = !{%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned}
+!6 = !{!7, !8, i64 0}
+!7 = !{!"Root", !8, i64 0, !11, i64 8, !8, i64 16, !11, i64 24}
+!8 = !{!"any pointer", !9, i64 0}
+!9 = !{!"omnipotent char", !10, i64 0}
+!10 = !{!"Simple C/C++ TBAA"}
+!11 = !{!"long", !9, i64 0}
+!12 = !{!7, !11, i64 8}
+!13 = !{!7, !8, i64 16}
+!14 = !{!7, !11, i64 24}
+!15 = !{!16, !16, i64 0}
+!16 = !{!"int", !9, i64 0}
diff --git a/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.cond.ll b/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.cond.ll
new file mode 100644
index 0000000000..e5f99dd971
--- /dev/null
+++ b/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.cond.ll
@@ -0,0 +1,258 @@
+; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_NVPTX.so -S -localmem -dfg2llvm-nvptx <  %s | FileCheck %s
+; ModuleID = 'ThreeLevel.cond.ll'
+source_filename = "ThreeLevel.cond.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%struct.Root = type { i32*, i64, i32*, i64 }
+%emptyStruct = type <{}>
+%emptyStruct.0 = type <{}>
+%emptyStruct.1 = type <{}>
+%emptyStruct.2 = type <{}>
+
+declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__getNode(...) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__getParentNode(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_y(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_y(i8*) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local void @Launch() local_unnamed_addr #2 {
+entry:
+  %RootArgs = alloca %struct.Root, align 8
+  %0 = bitcast %struct.Root* %RootArgs to i8*
+  call void @llvm.lifetime.start.p0i8(i64 32, i8* nonnull %0) #6
+  %call = tail call noalias i8* @malloc(i64 1024) #6
+  %1 = bitcast %struct.Root* %RootArgs to i8**
+  store i8* %call, i8** %1, align 8, !tbaa !6
+  %Insize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
+  store i64 1024, i64* %Insize, align 8, !tbaa !12
+  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
+  %call1 = tail call noalias i8* @malloc(i64 1024) #6
+  %2 = bitcast i32** %output to i8**
+  store i8* %call1, i8** %2, align 8, !tbaa !13
+  %Outsize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 3
+  store i64 1024, i64* %Outsize, align 8, !tbaa !14
+  %3 = bitcast %struct.Root* %RootArgs to i8*
+  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned to i8*), i8* %3, i1 false)
+  call void @llvm.hpvm.wait(i8* %graphID)
+  call void @llvm.lifetime.end.p0i8(i64 32, i8* nonnull %0) #6
+  ret void
+}
+
+; Function Attrs: nofree nounwind
+declare dso_local noalias i8* @malloc(i64) local_unnamed_addr #3
+
+declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
+
+; CHECK-LABEL: @main(
+; CHECK: call i8* @llvm_hpvm_ocl_initContext(i32
+; CHECK: call i8* @llvm_hpvm_ocl_launch(i8*
+; CHECK: call void @llvm.hpvm.init(
+; CHECK: call void @llvm_hpvm_ocl_clearContext(i8*
+; CHECK: call void @llvm.hpvm.cleanup(
+
+; CHECK-LABEL: @Func1_cloned(
+; CHECK: br i1 %cmp, label %if.end, label %if.then
+; CHECK-LABEL: if.then:
+; CHECK-LABEL: if.end:
+
+; CHECK-LABEL: @Func2_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned
+; CHECK-NOT: call void @llvm_hpvm_ocl_argument_shared(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_argument_ptr(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK-NOT: call void @llvm_hpvm_ocl_argument_shared(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_argument_ptr(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_executeNode(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_free(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_free(i8*
+
+; Function Attrs: nounwind uwtable
+define dso_local i32 @main() local_unnamed_addr #4 {
+entry:
+  call void @llvm.hpvm.init()
+  tail call void @Launch()
+  call void @llvm.hpvm.cleanup()
+  ret i32 0
+}
+
+declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
+
+declare i8* @llvm_hpvm_initializeTimerSet()
+
+declare void @llvm_hpvm_switchToTimer(i8**, i32)
+
+declare void @llvm_hpvm_printTimerSet(i8**, i8*)
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getNode() #5
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getParentNode(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.y(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.y(i8*) #5
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode2D(i8*, i64, i64) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct @Func1_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %call4 = call i8* @llvm.hpvm.getNode()
+  %call15 = call i8* @llvm.hpvm.getParentNode(i8* %call4)
+  %call26 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call4)
+  %conv = trunc i64 %call26 to i32
+  %call37 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call4)
+  %conv4 = trunc i64 %call37 to i32
+  %call58 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call15)
+  %conv6 = trunc i64 %call58 to i32
+  %call79 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call15)
+  %conv8 = trunc i64 %call79 to i32
+  %call910 = call i64 @llvm.hpvm.getNumNodeInstances.x(i8* %call4)
+  %conv10 = trunc i64 %call910 to i32
+  %call1111 = call i64 @llvm.hpvm.getNumNodeInstances.y(i8* %call4)
+  %conv12 = trunc i64 %call1111 to i32
+  %mul = mul nsw i32 %conv10, %conv6
+  %add = add nsw i32 %mul, %conv
+  %mul13 = mul nsw i32 %conv12, %conv8
+  %add14 = add nsw i32 %mul13, %conv4
+  %cmp = icmp eq i32 %add, %add14
+  br i1 %cmp, label %if.end, label %if.then
+
+if.then:                                          ; preds = %entry
+  %idxprom = sext i32 %add14 to i64
+  %arrayidx = getelementptr inbounds i32, i32* %In, i64 %idxprom
+  %0 = load i32, i32* %arrayidx, align 4, !tbaa !15
+  %idxprom16 = sext i32 %add to i64
+  %arrayidx17 = getelementptr inbounds i32, i32* %Out, i64 %idxprom16
+  %1 = load i32, i32* %arrayidx17, align 4, !tbaa !15
+  %add18 = add nsw i32 %1, %0
+  store i32 %add18, i32* %arrayidx17, align 4, !tbaa !15
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %entry
+  ret %emptyStruct undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.0 @Func3_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func1_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.0 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode(i8*) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.1 @Func2_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func3_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.1 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.2 @PipeRoot_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.2 undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.wait(i8*) #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.init() #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.cleanup() #6
+
+attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { noinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #3 = { nofree nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #4 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #5 = { nounwind readnone }
+attributes #6 = { nounwind }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+!hpvm_hint_gpu = !{!2}
+!hpvm_hint_cpu = !{!3, !4, !5}
+!hpvm_hint_spir = !{}
+!hpvm_hint_cudnn = !{}
+!hpvm_hint_promise = !{}
+!hpvm_hint_cpu_gpu = !{}
+!hpvm_hint_cpu_spir = !{}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 3551132592a00cab6c966df508ab511598269f78)"}
+!2 = !{%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned}
+!3 = !{%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned}
+!4 = !{%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned}
+!5 = !{%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned}
+!6 = !{!7, !8, i64 0}
+!7 = !{!"Root", !8, i64 0, !11, i64 8, !8, i64 16, !11, i64 24}
+!8 = !{!"any pointer", !9, i64 0}
+!9 = !{!"omnipotent char", !10, i64 0}
+!10 = !{!"Simple C/C++ TBAA"}
+!11 = !{!"long", !9, i64 0}
+!12 = !{!7, !11, i64 8}
+!13 = !{!7, !8, i64 16}
+!14 = !{!7, !11, i64 24}
+!15 = !{!16, !16, i64 0}
+!16 = !{!"int", !9, i64 0}
diff --git a/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.opt.genvisc.ll b/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.constmem.ll
similarity index 99%
rename from hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.opt.genvisc.ll
rename to hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.constmem.ll
index 5de9fd4e33..060608fdc5 100644
--- a/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.opt.genvisc.ll
+++ b/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.constmem.ll
@@ -1,4 +1,4 @@
-; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_OpenCL.so -S -localmem -dfg2llvm-opencl <  %s | FileCheck %s
+; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_NVPTX.so -S -localmem -dfg2llvm-nvptx <  %s | FileCheck %s
 ; ModuleID = 'ThreeLevel.opt.ll'
 source_filename = "ThreeLevel.opt.c"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
diff --git a/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.genvisc.ll b/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.ll
similarity index 99%
rename from hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.genvisc.ll
rename to hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.ll
index 9af2e48674..ed99bee9f7 100644
--- a/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.genvisc.ll
+++ b/hpvm/test/regressionTests/DFG2LLVM_NVPTX/ThreeLevel.ll
@@ -1,4 +1,4 @@
-; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_OpenCL.so -S -localmem -dfg2llvm-opencl <  %s | FileCheck %s
+; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_NVPTX.so -S -localmem -dfg2llvm-nvptx <  %s | FileCheck %s
 ; ModuleID = 'ThreeLevel.ll'
 source_filename = "ThreeLevel.c"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
diff --git a/hpvm/test/regressionTests/DFG2LLVM_X86/CreateNode.dfg.ll b/hpvm/test/regressionTests/DFG2LLVM_X86/OneLevel.codeGen.ll
similarity index 95%
rename from hpvm/test/regressionTests/DFG2LLVM_X86/CreateNode.dfg.ll
rename to hpvm/test/regressionTests/DFG2LLVM_X86/OneLevel.codeGen.ll
index f7f943a255..1373d13159 100644
--- a/hpvm/test/regressionTests/DFG2LLVM_X86/CreateNode.dfg.ll
+++ b/hpvm/test/regressionTests/DFG2LLVM_X86/OneLevel.codeGen.ll
@@ -1,4 +1,4 @@
-; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_CPU.so -S -dfg2llvm-cpu <  %s | FileCheck %s
+; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_X86.so -S -dfg2llvm-x86 <  %s | FileCheck %s
 ; ModuleID = 'CreateNode.ll'
 source_filename = "CreateNode.c"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
@@ -10,9 +10,9 @@ target triple = "x86_64-unknown-linux-gnu"
 
 ; CHECK-LABEL: i32 @main(
 ; CHECK: call void @llvm.hpvm.init()
-; CHECK: call i8* @llvm_hpvm_cpu_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
+; CHECK: call i8* @llvm_hpvm_x86_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
 ; CHECK-NEXT: call i8* @llvm.hpvm.launch(i8*
-; CHECK-NEXT: call void @llvm_hpvm_cpu_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_x86_wait(i8*
 
 ; CHECK-LABEL: @PipeRoot_cloned(
 ; CHECK: call i8* @llvm.hpvm.createNode(
@@ -23,12 +23,12 @@ target triple = "x86_64-unknown-linux-gnu"
 ; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func_cloned.node
 
 ; CHECK-LABEL: @Func_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned
-; CHECK: call i8* @llvm_hpvm_cpu_argument_ptr(
+; CHECK: call i8* @llvm_hpvm_x86_argument_ptr(
 
 ; CHECK-LABEL: @PipeRoot_cloned.2(
-; CHECK: call void @llvm_hpvm_cpu_dstack_push(
+; CHECK: call void @llvm_hpvm_x86_dstack_push(
 ; CHECK-NEXT: @Func_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_pop()
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
 
 ; CHECK-LABEL: @LaunchDataflowGraph(i8*
 ; call %struct.out.PipeRoot @PipeRoot_cloned.2(
diff --git a/hpvm/test/regressionTests/DFG2LLVM_X86/oneLaunchAlloca.dfg.ll b/hpvm/test/regressionTests/DFG2LLVM_X86/OneRootBasic.ll
similarity index 95%
rename from hpvm/test/regressionTests/DFG2LLVM_X86/oneLaunchAlloca.dfg.ll
rename to hpvm/test/regressionTests/DFG2LLVM_X86/OneRootBasic.ll
index 1bfa5f0c0b..a0f0f6ecfc 100644
--- a/hpvm/test/regressionTests/DFG2LLVM_X86/oneLaunchAlloca.dfg.ll
+++ b/hpvm/test/regressionTests/DFG2LLVM_X86/OneRootBasic.ll
@@ -1,4 +1,4 @@
-; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_CPU.so -S -dfg2llvm-cpu <  %s | FileCheck %s
+; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_X86.so -S -dfg2llvm-x86 <  %s | FileCheck %s
 ; ModuleID = 'oneLaunchAlloca.ll'
 source_filename = "oneLaunchAlloca.c"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
@@ -13,9 +13,9 @@ declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
 
 ; CHECK-LABEL: i32 @main(
 ; CHECK: call void @llvm.hpvm.init()
-; CHECK: call i8* @llvm_hpvm_cpu_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
+; CHECK: call i8* @llvm_hpvm_x86_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
 ; CHECK-NEXT: call i8* @llvm.hpvm.launch(i8*
-; CHECK-NEXT: call void @llvm_hpvm_cpu_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_x86_wait(i8*
 
 ; CHECK-LABEL: @PipeRoot_cloned.1(
 
diff --git a/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.dfg.ll b/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.codeGen.ll
similarity index 95%
rename from hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.dfg.ll
rename to hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.codeGen.ll
index 35d63f16ef..a60f28a08a 100644
--- a/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.dfg.ll
+++ b/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.codeGen.ll
@@ -1,4 +1,4 @@
-; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_CPU.so -S -dfg2llvm-cpu <  %s | FileCheck %s
+; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_X86.so -S -dfg2llvm-x86 <  %s | FileCheck %s
 ; ModuleID = 'ThreeLevel.ll'
 source_filename = "ThreeLevel.c"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
@@ -13,9 +13,9 @@ target triple = "x86_64-unknown-linux-gnu"
 
 ; CHECK-LABEL: i32 @main(
 ; CHECK: call void @llvm.hpvm.init()
-; CHECK: call i8* @llvm_hpvm_cpu_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
+; CHECK: call i8* @llvm_hpvm_x86_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
 ; CHECK-NEXT: call i8* @llvm.hpvm.launch(i8*
-; CHECK-NEXT: call void @llvm_hpvm_cpu_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_x86_wait(i8*
 
 ; CHECK-LABEL: @Func3_cloned(
 ; CHECK: call i8* @llvm.hpvm.createNode2D(
@@ -42,26 +42,26 @@ target triple = "x86_64-unknown-linux-gnu"
 ; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func2_cloned.node
 
 ; CHECK-LABEL: @Func1_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned
-; CHECK: call i8* @llvm_hpvm_cpu_argument_ptr(
+; CHECK: call i8* @llvm_hpvm_x86_argument_ptr(
 
 ; CHECK-LABEL: @Func3_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
 ; CHECK-LABEL: for.body1:
 ; CHECK: %index.y = phi i64 [ 0, %for.body ], [ %index.y.inc, %for.body1 ]
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_push(
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_push(
 ; CHECK-NEXT: @Func1_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_pop()
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
 
 ; CHECK-LABEL: @Func2_cloned.3_cloned_cloned_cloned_cloned_cloned_cloned(
 ; CHECK-LABEL: for.body:
 ; CHECK-NEXT: %index.x = phi i64 [ 0, %entry ], [ %index.x.inc, %for.body ]
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_push(
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_push(
 ; CHECK-NEXT: @Func3_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_pop()
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
 
 ; CHECK-LABEL: @PipeRoot_cloned.4(
-; CHECK: call void @llvm_hpvm_cpu_dstack_push(
+; CHECK: call void @llvm_hpvm_x86_dstack_push(
 ; CHECK-NEXT: @Func2_cloned.3_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_pop()
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
 
 ; CHECK-LABEL: @LaunchDataflowGraph(
 ; CHECK: call %struct.out.PipeRoot @PipeRoot_cloned.4(
diff --git a/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.cond.ll b/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.cond.ll
new file mode 100644
index 0000000000..2057fcbcb8
--- /dev/null
+++ b/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.cond.ll
@@ -0,0 +1,300 @@
+; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_X86.so -S -dfg2llvm-x86 <  %s | FileCheck %s
+; ModuleID = 'ThreeLevel.cond.ll'
+source_filename = "ThreeLevel.cond.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%struct.Root = type { i32*, i64, i32*, i64 }
+%emptyStruct = type <{}>
+%emptyStruct.0 = type <{}>
+%emptyStruct.1 = type <{}>
+%emptyStruct.2 = type <{}>
+
+; CHECK-LABEL: @Launch(
+; CHECK: call i8* @llvm_hpvm_x86_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
+; CHECK-NEXT: call i8* @llvm.hpvm.launch(i8*
+; CHECK-NEXT: call void @llvm_hpvm_x86_wait(i8*
+
+; CHECK-LABEL: i32 @main(
+; CHECK: call void @llvm.hpvm.init()
+; CHECK: call void @llvm.hpvm.cleanup()
+
+; CHECK-LABEL: @Func3_cloned(
+; CHECK: call i8* @llvm.hpvm.createNode2D(
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node
+
+; CHECK-LABEL: @Func2_cloned(
+; CHECK: call i8* @llvm.hpvm.createNode2D(
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+
+; CHECK-LABEL: @PipeRoot_cloned(
+; CHECK: call i8* @llvm.hpvm.createNode(
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+
+; CHECK-LABEL: @Func1_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned
+; CHECK: call i8* @llvm_hpvm_x86_argument_ptr(
+; CHECK: call i8* @llvm_hpvm_x86_argument_ptr(
+; CHECK: call i64 @llvm_hpvm_x86_getDimInstance(
+; CHECK: call i64 @llvm_hpvm_x86_getDimInstance(
+
+
+; CHECK-LABEL: @Func3_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK: call void @llvm_hpvm_x86_dstack_push(
+; CHECK-NEXT: @Func1_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
+; CHECK: br i1 %cond.y, label %for.body1, label %for.end2
+; CHECK-LABEL: for.end2:
+; CHECK:  br i1 %cond.x, label %for.body, label %for.end
+; CHECK-LABEL: for.end:
+
+; CHECK-LABEL: @Func2_cloned.3_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK-LABEL: for.body:
+; CHECK-NEXT: %index.x = phi i64 [ 0, %entry ], [ %index.x.inc, %for.end2 ]
+; CHECK-LABEL: for.body1:
+; CHECK-NEXT: %index.y = phi i64 [ 0, %for.body ], [ %index.y.inc, %for.body1 ]
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_push(
+; CHECK-NEXT: @Func3_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
+; CHECK: br i1 %cond.y, label %for.body1, label %for.end2
+; CHECK-LABEL: for.end2:
+; CHECK: br i1 %cond.x, label %for.body, label %for.end
+; CHECK-LABEL: for.end:
+
+; CHECK-LABEL: @PipeRoot_cloned.4(
+; CHECK: call void @llvm_hpvm_x86_dstack_push(
+; CHECK-NEXT: @Func2_cloned.3_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
+
+; CHECK-LABEL: @LaunchDataflowGraph(
+; CHECK: call %emptyStruct.2 @PipeRoot_cloned.4(
+
+declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__getNode(...) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__getParentNode(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_y(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_y(i8*) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local void @Launch() local_unnamed_addr #2 {
+entry:
+  %RootArgs = alloca %struct.Root, align 8
+  %0 = bitcast %struct.Root* %RootArgs to i8*
+  call void @llvm.lifetime.start.p0i8(i64 32, i8* nonnull %0) #6
+  %call = tail call noalias i8* @malloc(i64 1024) #6
+  %1 = bitcast %struct.Root* %RootArgs to i8**
+  store i8* %call, i8** %1, align 8, !tbaa !6
+  %Insize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
+  store i64 1024, i64* %Insize, align 8, !tbaa !12
+  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
+  %call1 = tail call noalias i8* @malloc(i64 1024) #6
+  %2 = bitcast i32** %output to i8**
+  store i8* %call1, i8** %2, align 8, !tbaa !13
+  %Outsize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 3
+  store i64 1024, i64* %Outsize, align 8, !tbaa !14
+  %3 = bitcast %struct.Root* %RootArgs to i8*
+  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned to i8*), i8* %3, i1 false)
+  call void @llvm.hpvm.wait(i8* %graphID)
+  call void @llvm.lifetime.end.p0i8(i64 32, i8* nonnull %0) #6
+  ret void
+}
+
+; Function Attrs: nofree nounwind
+declare dso_local noalias i8* @malloc(i64) local_unnamed_addr #3
+
+declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
+
+; Function Attrs: nounwind uwtable
+define dso_local i32 @main() local_unnamed_addr #4 {
+entry:
+  call void @llvm.hpvm.init()
+  tail call void @Launch()
+  call void @llvm.hpvm.cleanup()
+  ret i32 0
+}
+
+declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
+
+declare i8* @llvm_hpvm_initializeTimerSet()
+
+declare void @llvm_hpvm_switchToTimer(i8**, i32)
+
+declare void @llvm_hpvm_printTimerSet(i8**, i8*)
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getNode() #5
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getParentNode(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.y(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.y(i8*) #5
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode2D(i8*, i64, i64) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct @Func1_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %call4 = call i8* @llvm.hpvm.getNode()
+  %call15 = call i8* @llvm.hpvm.getParentNode(i8* %call4)
+  %call26 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call4)
+  %conv = trunc i64 %call26 to i32
+  %call37 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call4)
+  %conv4 = trunc i64 %call37 to i32
+  %call58 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call15)
+  %conv6 = trunc i64 %call58 to i32
+  %call79 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call15)
+  %conv8 = trunc i64 %call79 to i32
+  %call910 = call i64 @llvm.hpvm.getNumNodeInstances.x(i8* %call4)
+  %conv10 = trunc i64 %call910 to i32
+  %call1111 = call i64 @llvm.hpvm.getNumNodeInstances.y(i8* %call4)
+  %conv12 = trunc i64 %call1111 to i32
+  %mul = mul nsw i32 %conv10, %conv6
+  %add = add nsw i32 %mul, %conv
+  %mul13 = mul nsw i32 %conv12, %conv8
+  %add14 = add nsw i32 %mul13, %conv4
+  %cmp = icmp eq i32 %add, %add14
+  br i1 %cmp, label %if.end, label %if.then
+
+if.then:                                          ; preds = %entry
+  %arrayidx = getelementptr inbounds i32, i32* %In, i64 3
+  %0 = load i32, i32* %arrayidx, align 4, !tbaa !15
+  %idxprom = sext i32 %add to i64
+  %arrayidx16 = getelementptr inbounds i32, i32* %Out, i64 %idxprom
+  %1 = load i32, i32* %arrayidx16, align 4, !tbaa !15
+  %add17 = add nsw i32 %1, %0
+  store i32 %add17, i32* %arrayidx16, align 4, !tbaa !15
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %entry
+  ret %emptyStruct undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.0 @Func3_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func1_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.0 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode(i8*) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.1 @Func2_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func3_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.1 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.2 @PipeRoot_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.2 undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.wait(i8*) #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.init() #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.cleanup() #6
+
+attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { noinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #3 = { nofree nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #4 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #5 = { nounwind readnone }
+attributes #6 = { nounwind }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+!hpvm_hint_cpu = !{!2, !3, !4, !5}
+!hpvm_hint_gpu = !{}
+!hpvm_hint_spir = !{}
+!hpvm_hint_cudnn = !{}
+!hpvm_hint_promise = !{}
+!hpvm_hint_cpu_gpu = !{}
+!hpvm_hint_cpu_spir = !{}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 3551132592a00cab6c966df508ab511598269f78)"}
+!2 = !{%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned}
+!3 = !{%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned}
+!4 = !{%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned}
+!5 = !{%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned}
+!6 = !{!7, !8, i64 0}
+!7 = !{!"Root", !8, i64 0, !11, i64 8, !8, i64 16, !11, i64 24}
+!8 = !{!"any pointer", !9, i64 0}
+!9 = !{!"omnipotent char", !10, i64 0}
+!10 = !{!"Simple C/C++ TBAA"}
+!11 = !{!"long", !9, i64 0}
+!12 = !{!7, !11, i64 8}
+!13 = !{!7, !8, i64 16}
+!14 = !{!7, !11, i64 24}
+!15 = !{!16, !16, i64 0}
+!16 = !{!"int", !9, i64 0}
diff --git a/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.constmem.nvptx.ll b/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.constmem.nvptx.ll
new file mode 100644
index 0000000000..c051cedcf5
--- /dev/null
+++ b/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.constmem.nvptx.ll
@@ -0,0 +1,270 @@
+; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_NVPTX.so -load LLVMDFG2LLVM_X86.so -S -localmem -dfg2llvm-nvptx -dfg2llvm-x86 <  %s | FileCheck %s
+; ModuleID = 'ThreeLevel.opt.ll'
+source_filename = "ThreeLevel.opt.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%struct.Root = type { i32*, i64, i32*, i64 }
+%struct.out.Allocation = type <{ i8*, i64 }>
+%emptyStruct = type <{}>
+%emptyStruct.0 = type <{}>
+%emptyStruct.1 = type <{}>
+%emptyStruct.2 = type <{}>
+
+declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__getNode(...) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__getParentNode(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_y(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_y(i8*) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__malloc(i64) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__return(i32, ...) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__edge(i8*, i8*, i32, i32, i32, i32) local_unnamed_addr #0
+
+; CHECK-LABEL: @Launch(
+; CHECK: call i8* @llvm_hpvm_x86_launch(i8*
+; CHECK-NEXT: call i8* @llvm.hpvm.launch(i8*
+; CHECK-NEXT: call void @llvm_hpvm_x86_wait(i8*
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local void @Launch() local_unnamed_addr #2 {
+entry:
+  %RootArgs = alloca %struct.Root, align 8
+  %0 = bitcast %struct.Root* %RootArgs to i8*
+  call void @llvm.lifetime.start.p0i8(i64 32, i8* nonnull %0) #6
+  %call = tail call noalias i8* @malloc(i64 1024) #6
+  %1 = bitcast %struct.Root* %RootArgs to i8**
+  store i8* %call, i8** %1, align 8, !tbaa !6
+  %Insize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
+  store i64 1024, i64* %Insize, align 8, !tbaa !12
+  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
+  %call1 = tail call noalias i8* @malloc(i64 1024) #6
+  %2 = bitcast i32** %output to i8**
+  store i8* %call1, i8** %2, align 8, !tbaa !13
+  %Outsize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 3
+  store i64 1024, i64* %Outsize, align 8, !tbaa !14
+  %3 = bitcast %struct.Root* %RootArgs to i8*
+  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned to i8*), i8* %3, i1 false)
+  call void @llvm.hpvm.wait(i8* %graphID)
+  call void @llvm.lifetime.end.p0i8(i64 32, i8* nonnull %0) #6
+  ret void
+}
+
+; Function Attrs: nofree nounwind
+declare dso_local noalias i8* @malloc(i64) local_unnamed_addr #3
+
+declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
+
+; CHECK-LABEL: @main(
+; CHECK: call i8* @llvm_hpvm_ocl_initContext(i32
+; CHECK: call i8* @llvm_hpvm_ocl_launch(i8*
+; CHECK: call void @llvm_hpvm_ocl_clearContext(i8*
+
+; CHECK-LABEL: @Func2_cloned.3_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK: call i8* @llvm_hpvm_ocl_argument_ptr(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_shared(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_executeNode(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_free(i8*
+
+; CHECK-LABEL: @PipeRoot_cloned.4(
+; CHECK: call void @llvm_hpvm_x86_dstack_push(
+; CHECK-NEXT: @Func2_cloned.3_cloned_cloned_cloned_cloned_cloned_clone
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop(
+
+; CHECK-LABEL: define i8* @LaunchDataflowGraph(i8*
+
+
+; Function Attrs: nounwind uwtable
+define dso_local i32 @main() local_unnamed_addr #4 {
+entry:
+  call void @llvm.hpvm.init()
+  tail call void @Launch()
+  call void @llvm.hpvm.cleanup()
+  ret i32 0
+}
+
+declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
+
+declare i8* @llvm_hpvm_initializeTimerSet()
+
+declare void @llvm_hpvm_switchToTimer(i8**, i32)
+
+declare void @llvm_hpvm_printTimerSet(i8**, i8*)
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getNode() #5
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getParentNode(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.y(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.y(i8*) #5
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.malloc(i64) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %struct.out.Allocation @Allocation_cloned(i64 %block) #4 {
+entry:
+  %call1 = call i8* @llvm.hpvm.malloc(i64 %block)
+  %returnStruct = insertvalue %struct.out.Allocation undef, i8* %call1, 0
+  %returnStruct2 = insertvalue %struct.out.Allocation %returnStruct, i64 %block, 1
+  ret %struct.out.Allocation %returnStruct2
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode2D(i8*, i64, i64) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct @Func1_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %call4 = call i8* @llvm.hpvm.getNode()
+  %call15 = call i8* @llvm.hpvm.getParentNode(i8* %call4)
+  %call26 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call4)
+  %call37 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call4)
+  %call58 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call15)
+  %call79 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call15)
+  %call910 = call i64 @llvm.hpvm.getNumNodeInstances.x(i8* %call4)
+  %call1111 = call i64 @llvm.hpvm.getNumNodeInstances.y(i8* %call4)
+  %mul = mul i64 %call910, %call58
+  %add = add i64 %mul, %call26
+  %arrayidx = getelementptr inbounds i32, i32* %In, i64 3
+  %0 = load i32, i32* %arrayidx, align 4, !tbaa !15
+  %sext = shl i64 %add, 32
+  %idxprom = ashr exact i64 %sext, 32
+  %arrayidx15 = getelementptr inbounds i32, i32* %Out, i64 %idxprom
+  %1 = load i32, i32* %arrayidx15, align 4, !tbaa !15
+  %add16 = add nsw i32 %1, %0
+  store i32 %add16, i32* %arrayidx15, align 4, !tbaa !15
+  ret %emptyStruct undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode(i8*) #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #6
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createEdge(i8*, i8*, i1, i32, i32, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.0 @Func3_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func1_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned to i8*), i64 3, i64 5)
+  %Allocation_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Allocation (i64)* @Allocation_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Allocation_cloned.node, i32 1, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 3, i32 3, i1 false)
+  %output = call i8* @llvm.hpvm.createEdge(i8* %Allocation_cloned.node, i8* %Func1_cloned.node, i1 true, i32 0, i32 0, i1 false)
+  %output1 = call i8* @llvm.hpvm.createEdge(i8* %Allocation_cloned.node, i8* %Func1_cloned.node, i1 true, i32 1, i32 1, i1 false)
+  ret %emptyStruct.0 undef
+}
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.1 @Func2_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func3_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.1 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.2 @PipeRoot_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.2 undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.wait(i8*) #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.init() #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.cleanup() #6
+
+attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { noinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #3 = { nofree nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #4 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #5 = { nounwind readnone }
+attributes #6 = { nounwind }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+!hpvm_hint_gpu = !{!2}
+!hpvm_hint_cpu = !{!3, !4, !5}
+!hpvm_hint_spir = !{}
+!hpvm_hint_cudnn = !{}
+!hpvm_hint_promise = !{}
+!hpvm_hint_cpu_gpu = !{}
+!hpvm_hint_cpu_spir = !{}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 6690f9e7e8b46b96aea222d3e85315cd63545953)"}
+!2 = !{%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned}
+!3 = !{%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned}
+!4 = !{%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned}
+!5 = !{%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned}
+!6 = !{!7, !8, i64 0}
+!7 = !{!"Root", !8, i64 0, !11, i64 8, !8, i64 16, !11, i64 24}
+!8 = !{!"any pointer", !9, i64 0}
+!9 = !{!"omnipotent char", !10, i64 0}
+!10 = !{!"Simple C/C++ TBAA"}
+!11 = !{!"long", !9, i64 0}
+!12 = !{!7, !11, i64 8}
+!13 = !{!7, !8, i64 16}
+!14 = !{!7, !11, i64 24}
+!15 = !{!16, !16, i64 0}
+!16 = !{!"int", !9, i64 0}
diff --git a/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.nvptx.ll b/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.nvptx.ll
new file mode 100644
index 0000000000..a8804734f7
--- /dev/null
+++ b/hpvm/test/regressionTests/DFG2LLVM_X86/ThreeLevel.nvptx.ll
@@ -0,0 +1,250 @@
+; RUN: opt -load LLVMBuildDFG.so -load LLVMLocalMem.so -load LLVMDFG2LLVM_NVPTX.so -load LLVMDFG2LLVM_X86.so -S -localmem -dfg2llvm-nvptx -dfg2llvm-x86 <  %s | FileCheck %s
+; ModuleID = 'ThreeLevel.ll'
+source_filename = "ThreeLevel.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%struct.Root = type { i32*, i64, i32*, i64 }
+%emptyStruct = type <{}>
+%emptyStruct.0 = type <{}>
+%emptyStruct.1 = type <{}>
+%emptyStruct.2 = type <{}>
+
+declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__getNode(...) local_unnamed_addr #0
+
+declare dso_local i8* @__hpvm__getParentNode(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNodeInstanceID_y(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_x(i8*) local_unnamed_addr #0
+
+declare dso_local i64 @__hpvm__getNumNodeInstances_y(i8*) local_unnamed_addr #0
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #1
+
+declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #0
+
+; CHECK-LABEL: @Launch(
+; CHECK: call i8* @llvm_hpvm_x86_launch(i8*
+; CHECK-NEXT: call i8* @llvm.hpvm.launch(i8*
+; CHECK-NEXT: call void @llvm_hpvm_x86_wait(i8*
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local void @Launch() local_unnamed_addr #2 {
+entry:
+  %RootArgs = alloca %struct.Root, align 8
+  %0 = bitcast %struct.Root* %RootArgs to i8*
+  call void @llvm.lifetime.start.p0i8(i64 32, i8* nonnull %0) #6
+  %call = tail call noalias i8* @malloc(i64 1024) #6
+  %1 = bitcast %struct.Root* %RootArgs to i8**
+  store i8* %call, i8** %1, align 8, !tbaa !6
+  %Insize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
+  store i64 1024, i64* %Insize, align 8, !tbaa !12
+  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
+  %call1 = tail call noalias i8* @malloc(i64 1024) #6
+  %2 = bitcast i32** %output to i8**
+  store i8* %call1, i8** %2, align 8, !tbaa !13
+  %Outsize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 3
+  store i64 1024, i64* %Outsize, align 8, !tbaa !14
+  %3 = bitcast %struct.Root* %RootArgs to i8*
+  %graphID = call i8* @llvm.hpvm.launch(i8* bitcast (%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned to i8*), i8* %3, i1 false)
+  call void @llvm.hpvm.wait(i8* %graphID)
+  call void @llvm.lifetime.end.p0i8(i64 32, i8* nonnull %0) #6
+  ret void
+}
+
+; Function Attrs: nofree nounwind
+declare dso_local noalias i8* @malloc(i64) local_unnamed_addr #3
+
+declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #0
+
+; CHECK-LABEL: @main(
+; CHECK: call i8* @llvm_hpvm_ocl_initContext(i32
+; CHECK: call i8* @llvm_hpvm_ocl_launch(i8*
+; CHECK: call void @llvm_hpvm_ocl_clearContext(i8*
+
+; CHECK-LABEL: @Func2_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
+; CHECK: call i8* @llvm_hpvm_ocl_argument_ptr(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_argument_ptr(i8*
+; CHECK: call void @llvm_hpvm_ocl_argument_scalar(i8*
+; CHECK: call i8* @llvm_hpvm_ocl_executeNode(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_free(i8*
+; CHECK-NEXT: call void @llvm_hpvm_ocl_free(i8*
+
+; CHECK-LABEL: @PipeRoot_cloned.3(
+; CHECK: call void @llvm_hpvm_x86_dstack_push(
+; CHECK-NEXT: @Func2_cloned.2_cloned_cloned_cloned_cloned_cloned_clone
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop(
+
+; CHECK-LABEL: define i8* @LaunchDataflowGraph(i8*
+
+; Function Attrs: nounwind uwtable
+define dso_local i32 @main() local_unnamed_addr #4 {
+entry:
+  call void @llvm.hpvm.init()
+  tail call void @Launch()
+  call void @llvm.hpvm.cleanup()
+  ret i32 0
+}
+
+declare dso_local void @__hpvm__init(...) local_unnamed_addr #0
+
+declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #0
+
+declare i8* @llvm_hpvm_initializeTimerSet()
+
+declare void @llvm_hpvm_switchToTimer(i8**, i32)
+
+declare void @llvm_hpvm_printTimerSet(i8**, i8*)
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getNode() #5
+
+; Function Attrs: nounwind readnone
+declare i8* @llvm.hpvm.getParentNode(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNodeInstanceID.y(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.x(i8*) #5
+
+; Function Attrs: nounwind readnone
+declare i64 @llvm.hpvm.getNumNodeInstances.y(i8*) #5
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode2D(i8*, i64, i64) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct @Func1_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %call4 = call i8* @llvm.hpvm.getNode()
+  %call15 = call i8* @llvm.hpvm.getParentNode(i8* %call4)
+  %call26 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call4)
+  %call37 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call4)
+  %call58 = call i64 @llvm.hpvm.getNodeInstanceID.x(i8* %call15)
+  %call79 = call i64 @llvm.hpvm.getNodeInstanceID.y(i8* %call15)
+  %call910 = call i64 @llvm.hpvm.getNumNodeInstances.x(i8* %call4)
+  %call1111 = call i64 @llvm.hpvm.getNumNodeInstances.y(i8* %call4)
+  %mul = mul i64 %call910, %call58
+  %add = add i64 %mul, %call26
+  %mul13 = mul i64 %call1111, %call79
+  %add14 = add i64 %mul13, %call37
+  %sext = shl i64 %add14, 32
+  %idxprom = ashr exact i64 %sext, 32
+  %arrayidx = getelementptr inbounds i32, i32* %In, i64 %idxprom
+  %0 = load i32, i32* %arrayidx, align 4, !tbaa !15
+  %sext36 = shl i64 %add, 32
+  %idxprom15 = ashr exact i64 %sext36, 32
+  %arrayidx16 = getelementptr inbounds i32, i32* %Out, i64 %idxprom15
+  %1 = load i32, i32* %arrayidx16, align 4, !tbaa !15
+  %add17 = add nsw i32 %1, %0
+  store i32 %add17, i32* %arrayidx16, align 4, !tbaa !15
+  ret %emptyStruct undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.bind.input(i8*, i32, i32, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.0 @Func3_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func1_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.0 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.createNode(i8*) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.1 @Func2_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func3_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned to i8*), i64 3, i64 5)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.1 undef
+}
+
+; Function Attrs: nounwind
+declare i8* @llvm.hpvm.launch(i8*, i8*, i1) #6
+
+; Function Attrs: nounwind uwtable
+define dso_local %emptyStruct.2 @PipeRoot_cloned(i32* in %In, i64 %Insize, i32* in out %Out, i64 %Outsize) #4 {
+entry:
+  %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned to i8*))
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 0, i32 0, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 1, i32 1, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 2, i32 2, i1 false)
+  call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node, i32 3, i32 3, i1 false)
+  ret %emptyStruct.2 undef
+}
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.wait(i8*) #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.init() #6
+
+; Function Attrs: nounwind
+declare void @llvm.hpvm.cleanup() #6
+
+attributes #0 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { noinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #3 = { nofree nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #4 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
+attributes #5 = { nounwind readnone }
+attributes #6 = { nounwind }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+!hpvm_hint_gpu = !{!2}
+!hpvm_hint_cpu = !{!3, !4, !5}
+!hpvm_hint_spir = !{}
+!hpvm_hint_cudnn = !{}
+!hpvm_hint_promise = !{}
+!hpvm_hint_cpu_gpu = !{}
+!hpvm_hint_cpu_spir = !{}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 6690f9e7e8b46b96aea222d3e85315cd63545953)"}
+!2 = !{%emptyStruct (i32*, i64, i32*, i64)* @Func1_cloned}
+!3 = !{%emptyStruct.0 (i32*, i64, i32*, i64)* @Func3_cloned}
+!4 = !{%emptyStruct.1 (i32*, i64, i32*, i64)* @Func2_cloned}
+!5 = !{%emptyStruct.2 (i32*, i64, i32*, i64)* @PipeRoot_cloned}
+!6 = !{!7, !8, i64 0}
+!7 = !{!"Root", !8, i64 0, !11, i64 8, !8, i64 16, !11, i64 24}
+!8 = !{!"any pointer", !9, i64 0}
+!9 = !{!"omnipotent char", !10, i64 0}
+!10 = !{!"Simple C/C++ TBAA"}
+!11 = !{!"long", !9, i64 0}
+!12 = !{!7, !11, i64 8}
+!13 = !{!7, !8, i64 16}
+!14 = !{!7, !11, i64 24}
+!15 = !{!16, !16, i64 0}
+!16 = !{!"int", !9, i64 0}
diff --git a/hpvm/test/regressionTests/DFG2LLVM_X86/TwoLevel.dfg.ll b/hpvm/test/regressionTests/DFG2LLVM_X86/TwoLevel.codeGen.ll
similarity index 95%
rename from hpvm/test/regressionTests/DFG2LLVM_X86/TwoLevel.dfg.ll
rename to hpvm/test/regressionTests/DFG2LLVM_X86/TwoLevel.codeGen.ll
index f8ee61f1a7..b218b70fd0 100644
--- a/hpvm/test/regressionTests/DFG2LLVM_X86/TwoLevel.dfg.ll
+++ b/hpvm/test/regressionTests/DFG2LLVM_X86/TwoLevel.codeGen.ll
@@ -1,4 +1,4 @@
-; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_CPU.so -S -dfg2llvm-cpu <  %s | FileCheck %s
+; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_X86.so -S -dfg2llvm-x86 <  %s | FileCheck %s
 ; ModuleID = 'TwoLevel.ll'
 source_filename = "TwoLevel.c"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
@@ -11,9 +11,9 @@ target triple = "x86_64-unknown-linux-gnu"
 
 ; CHECK-LABEL: i32 @main(
 ; CHECK: call void @llvm.hpvm.init()
-; CHECK: call i8* @llvm_hpvm_cpu_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
+; CHECK: call i8* @llvm_hpvm_x86_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
 ; CHECK-NEXT: call i8* @llvm.hpvm.launch(i8* 
-; CHECK-NEXT: call void @llvm_hpvm_cpu_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_x86_wait(i8*
 
 ; CHECK-LABEL: @Func2_cloned(
 ; CHECK: call i8* @llvm.hpvm.createNode1D(
@@ -32,19 +32,19 @@ target triple = "x86_64-unknown-linux-gnu"
 ; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func2_cloned.node
 
 ; CHECK-LABEL: @Func1_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK: call i8* @llvm_hpvm_cpu_argument_ptr(
+; CHECK: call i8* @llvm_hpvm_x86_argument_ptr(
 
 ; CHECK-LABEL: @Func2_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
 ; CHECK-LABEL: for.body
 ; CHECK: %index.x = phi i64 [ 0, %entry ], [ %index.x.inc, %for.body ]
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_push(
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_push(
 ; CHECK-NEXT: @Func1_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_pop()
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
 
 ; CHECK-LABEL: @PipeRoot_cloned.3(
-; CHECK: call void @llvm_hpvm_cpu_dstack_push(
+; CHECK: call void @llvm_hpvm_x86_dstack_push(
 ; CHECK-NEXT: @Func2_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_pop()
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
 
 ; CHECK-LABEL: @LaunchDataflowGraph(i8*
 ; call %struct.out.PipeRoot @PipeRoot_cloned.3(
diff --git a/hpvm/test/regressionTests/DFG2LLVM_X86/TwoLaunch.dfg.ll b/hpvm/test/regressionTests/DFG2LLVM_X86/TwoRoot.ll
similarity index 93%
rename from hpvm/test/regressionTests/DFG2LLVM_X86/TwoLaunch.dfg.ll
rename to hpvm/test/regressionTests/DFG2LLVM_X86/TwoRoot.ll
index 3f74a190f3..5ce7a58e21 100644
--- a/hpvm/test/regressionTests/DFG2LLVM_X86/TwoLaunch.dfg.ll
+++ b/hpvm/test/regressionTests/DFG2LLVM_X86/TwoRoot.ll
@@ -1,4 +1,4 @@
-; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_CPU.so -S -dfg2llvm-cpu <  %s | FileCheck %s
+; RUN: opt -load LLVMBuildDFG.so -load LLVMDFG2LLVM_X86.so -S -dfg2llvm-x86 <  %s | FileCheck %s
 ; ModuleID = 'TwoLaunch.ll'
 source_filename = "TwoLaunch.c"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
@@ -11,12 +11,12 @@ target triple = "x86_64-unknown-linux-gnu"
 
 ; CHECK-LABEL: i32 @main(
 ; CHECK: call void @llvm.hpvm.init()
-; CHECK: @llvm_hpvm_cpu_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
+; CHECK: @llvm_hpvm_x86_launch(i8* (i8*)* @LaunchDataflowGraph, i8*
 ; CHECK-NEXT: call i8* @llvm.hpvm.launch(i8*
-; CHECK: @llvm_hpvm_cpu_launch(i8* (i8*)* @LaunchDataflowGraph.7, i8*
+; CHECK: @llvm_hpvm_x86_launch(i8* (i8*)* @LaunchDataflowGraph.7, i8*
 ; CHECK-NEXT: call i8* @llvm.hpvm.launch(i8*
-; CHECK-NEXT: call void @llvm_hpvm_cpu_wait(i8*
-; CHECK-NEXT: call void @llvm_hpvm_cpu_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_x86_wait(i8*
+; CHECK-NEXT: call void @llvm_hpvm_x86_wait(i8*
 
 ; CHECK-LABEL: @Func2_cloned(
 ; CHECK: call i8* @llvm.hpvm.createNode1D(
@@ -35,30 +35,30 @@ target triple = "x86_64-unknown-linux-gnu"
 ; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func2_cloned.node
 
 ; CHECK-LABEL: @Func1_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK: call i8* @llvm_hpvm_cpu_argument_ptr(
+; CHECK: call i8* @llvm_hpvm_x86_argument_ptr(
 
 ; CHECK-LABEL: @Func2_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
 ; CHECK: %index.x = phi i64 [ 0, %entry ], [ %index.x.inc, %for.body ]
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_push(
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_push(
 ; CHECK-NEXT: @Func1_cloned.1_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_pop()
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
 
 ; CHECK-LABEL: @PipeRoot_cloned.3(
-; CHECK: call void @llvm_hpvm_cpu_dstack_push(
+; CHECK: call void @llvm_hpvm_x86_dstack_push(
 ; CHECK-NEXT: @Func2_cloned.2_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_pop()
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
 
 ; CHECK-LABEL: @LaunchDataflowGraph(i8*
 ; CHECK: call %struct.out.PipeRoot @PipeRoot_cloned.3(
 
 ; CHECK-LABEL: @Func1_cloned.4_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK: @llvm_hpvm_cpu_argument_ptr(
+; CHECK: @llvm_hpvm_x86_argument_ptr(
 
 ; CHECK-LABEL: @Func2_cloned.5_cloned_cloned_cloned_cloned_cloned_cloned(
 ; CHECK: %index.x = phi i64 [ 0, %entry ], [ %index.x.inc, %for.body ]
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_push(
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_push(
 ; CHECK-NEXT: @Func1_cloned.4_cloned_cloned_cloned_cloned_cloned_cloned(
-; CHECK-NEXT: call void @llvm_hpvm_cpu_dstack_pop()
+; CHECK-NEXT: call void @llvm_hpvm_x86_dstack_pop()
 
 ; CHECK-LABEL: @LaunchDataflowGraph.7(i8*
 ; call %struct.out.PipeRoot @PipeRoot_cloned.6(
diff --git a/hpvm/test/regressionTests/GenHPVM/AllocationNode.ll b/hpvm/test/regressionTests/GenHPVM/AllocationNode.ll
new file mode 100644
index 0000000000..8f23a425cc
--- /dev/null
+++ b/hpvm/test/regressionTests/GenHPVM/AllocationNode.ll
@@ -0,0 +1,187 @@
+; RUN: opt -load LLVMGenHPVM.so -S -genhpvm <  %s | FileCheck %s
+; ModuleID = 'ThreeLevel.allocation.c'
+source_filename = "ThreeLevel.allocation.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%struct.Root = type { i32*, i64, i32*, i64 }
+; CHECK-LABEL: %struct.out.Allocation =
+
+; CHECK-LABEL: void @Launch(
+; CHECK: call i8* @llvm.hpvm.launch(i8*
+; CHECK-NEXT: call void @llvm.hpvm.wait(i8*
+
+; CHECK-LABEL: i32 @main(
+; CHECK: call void @llvm.hpvm.init()
+; CHECK-NEXT: tail call void @Launch(
+; CHECK-NEXT: call void @llvm.hpvm.cleanup()
+
+; CHECK-LABEL: @Allocation_cloned(
+; CHECK: call i8* @llvm.hpvm.malloc(i64
+
+; CHECK-LABEL: @Func1_cloned(
+
+; CHECK-LABEL: @Func3_cloned(
+; CHECK: %Func1_cloned.node = call i8* @llvm.hpvm.createNode2D(i8*
+; CHECK: %Allocation_cloned.node = call i8* @llvm.hpvm.createNode(i8*
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Allocation_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node
+; CHECK-NEXT: call i8* @llvm.hpvm.createEdge(i8* %Allocation_cloned.node
+; CHECK-NEXTL call i8* @llvm.hpvm.createEdge(i8* %Allocation_cloned.node
+
+; CHECK-LABEL: @Func2_cloned(
+; CHECK: %Func3_cloned.node = call i8* @llvm.hpvm.createNode2D(
+; CHECK-NEXT:  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+; CHECK-NEXT:  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+; CHECK-NEXT:  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+; CHECK-NEXT:  call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node
+
+; CHECK-LABEL: @PipeRoot_cloned(
+; CHECK: %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8*
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node
+
+; Function Attrs: nounwind uwtable
+define dso_local void @Func1(i32* %In, i64 %Insize, i32* %Out, i64 %Outsize) #0 {
+entry:
+  tail call void @__hpvm__hint(i32 2) #5
+  tail call void (i32, ...) @__hpvm__attributes(i32 2, i32* %In, i32* %Out, i32 1, i32* %Out) #5
+  ret void
+}
+
+declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #1
+
+declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #1
+
+; Function Attrs: nounwind uwtable
+define dso_local void @Allocation(i64 %block) #0 {
+entry:
+  %call = tail call i8* @__hpvm__malloc(i64 %block) #5
+  tail call void (i32, ...) @__hpvm__return(i32 2, i8* %call, i64 %block) #5
+  ret void
+}
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #2
+
+declare dso_local i8* @__hpvm__malloc(i64) local_unnamed_addr #1
+
+declare dso_local void @__hpvm__return(i32, ...) local_unnamed_addr #1
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #2
+
+; Function Attrs: nounwind uwtable
+define dso_local void @Func3(i32* %In, i64 %Insize, i32* %Out, i64 %Outsize) #0 {
+entry:
+  tail call void @__hpvm__hint(i32 1) #5
+  tail call void (i32, ...) @__hpvm__attributes(i32 2, i32* %In, i32* %Out, i32 1, i32* %Out) #5
+  %call = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 2, void (i32*, i64, i32*, i64)* nonnull @Func1, i64 3, i64 5) #5
+  %call1 = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 0, void (i64)* nonnull @Allocation) #5
+  tail call void @__hpvm__bindIn(i8* %call1, i32 1, i32 0, i32 0) #5
+  tail call void @__hpvm__bindIn(i8* %call, i32 2, i32 2, i32 0) #5
+  tail call void @__hpvm__bindIn(i8* %call, i32 3, i32 3, i32 0) #5
+  %call2 = tail call i8* @__hpvm__edge(i8* %call1, i8* %call, i32 1, i32 0, i32 0, i32 0) #5
+  %call3 = tail call i8* @__hpvm__edge(i8* %call1, i8* %call, i32 1, i32 1, i32 1, i32 0) #5
+  ret void
+}
+
+declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #1
+
+declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #1
+
+declare dso_local i8* @__hpvm__edge(i8*, i8*, i32, i32, i32, i32) local_unnamed_addr #1
+
+; Function Attrs: nounwind uwtable
+define dso_local void @Func2(i32* %In, i64 %Insize, i32* %Out, i64 %Outsize) #0 {
+entry:
+  tail call void @__hpvm__hint(i32 1) #5
+  tail call void (i32, ...) @__hpvm__attributes(i32 2, i32* %In, i32* %Out, i32 1, i32* %Out) #5
+  %call = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 2, void (i32*, i64, i32*, i64)* nonnull @Func3, i64 3, i64 5) #5
+  tail call void @__hpvm__bindIn(i8* %call, i32 0, i32 0, i32 0) #5
+  tail call void @__hpvm__bindIn(i8* %call, i32 1, i32 1, i32 0) #5
+  tail call void @__hpvm__bindIn(i8* %call, i32 2, i32 2, i32 0) #5
+  tail call void @__hpvm__bindIn(i8* %call, i32 3, i32 3, i32 0) #5
+  ret void
+}
+
+; Function Attrs: nounwind uwtable
+define dso_local void @PipeRoot(i32* %In, i64 %Insize, i32* %Out, i64 %Outsize) #0 {
+entry:
+  tail call void @__hpvm__hint(i32 1) #5
+  tail call void (i32, ...) @__hpvm__attributes(i32 2, i32* %In, i32* %Out, i32 1, i32* %Out) #5
+  %call = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 0, void (i32*, i64, i32*, i64)* nonnull @Func2) #5
+  tail call void @__hpvm__bindIn(i8* %call, i32 0, i32 0, i32 0) #5
+  tail call void @__hpvm__bindIn(i8* %call, i32 1, i32 1, i32 0) #5
+  tail call void @__hpvm__bindIn(i8* %call, i32 2, i32 2, i32 0) #5
+  tail call void @__hpvm__bindIn(i8* %call, i32 3, i32 3, i32 0) #5
+  ret void
+}
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local void @Launch() local_unnamed_addr #3 {
+entry:
+  %RootArgs = alloca %struct.Root, align 8
+  %0 = bitcast %struct.Root* %RootArgs to i8*
+  call void @llvm.lifetime.start.p0i8(i64 32, i8* nonnull %0) #5
+  %call = tail call noalias i8* @malloc(i64 1024) #5
+  %1 = bitcast %struct.Root* %RootArgs to i8**
+  store i8* %call, i8** %1, align 8, !tbaa !2
+  %Insize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
+  store i64 1024, i64* %Insize, align 8, !tbaa !8
+  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 2
+  %call1 = tail call noalias i8* @malloc(i64 1024) #5
+  %2 = bitcast i32** %output to i8**
+  store i8* %call1, i8** %2, align 8, !tbaa !9
+  %Outsize = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 3
+  store i64 1024, i64* %Outsize, align 8, !tbaa !10
+  %call2 = call i8* (i32, ...) @__hpvm__launch(i32 0, void (i32*, i64, i32*, i64)* nonnull @PipeRoot, %struct.Root* nonnull %RootArgs) #5
+  call void @__hpvm__wait(i8* %call2) #5
+  call void @llvm.lifetime.end.p0i8(i64 32, i8* nonnull %0) #5
+  ret void
+}
+
+; Function Attrs: nofree nounwind
+declare dso_local noalias i8* @malloc(i64) local_unnamed_addr #4
+
+declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #1
+
+declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #1
+
+; Function Attrs: nounwind uwtable
+define dso_local i32 @main() local_unnamed_addr #0 {
+entry:
+  tail call void (...) @__hpvm__init() #5
+  tail call void @Launch()
+  tail call void (...) @__hpvm__cleanup() #5
+  ret i32 0
+}
+
+declare dso_local void @__hpvm__init(...) local_unnamed_addr #1
+
+declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #1
+
+attributes #0 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #1 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #2 = { argmemonly nounwind }
+attributes #3 = { noinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #4 = { nofree nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #5 = { nounwind }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 3551132592a00cab6c966df508ab511598269f78)"}
+!2 = !{!3, !4, i64 0}
+!3 = !{!"Root", !4, i64 0, !7, i64 8, !4, i64 16, !7, i64 24}
+!4 = !{!"any pointer", !5, i64 0}
+!5 = !{!"omnipotent char", !6, i64 0}
+!6 = !{!"Simple C/C++ TBAA"}
+!7 = !{!"long", !5, i64 0}
+!8 = !{!3, !7, i64 8}
+!9 = !{!3, !4, i64 16}
+!10 = !{!3, !7, i64 24}
diff --git a/hpvm/test/regressionTests/GenVISC/CreateNode.ll b/hpvm/test/regressionTests/GenHPVM/CreateNode.ll
similarity index 100%
rename from hpvm/test/regressionTests/GenVISC/CreateNode.ll
rename to hpvm/test/regressionTests/GenHPVM/CreateNode.ll
diff --git a/hpvm/test/regressionTests/GenVISC/CreateNodeAndEdge.ll b/hpvm/test/regressionTests/GenHPVM/CreateNodeAndEdge.ll
similarity index 100%
rename from hpvm/test/regressionTests/GenVISC/CreateNodeAndEdge.ll
rename to hpvm/test/regressionTests/GenHPVM/CreateNodeAndEdge.ll
diff --git a/hpvm/test/regressionTests/GenVISC/LeafBindEdge.ll b/hpvm/test/regressionTests/GenHPVM/LeafBindEdge.ll
similarity index 100%
rename from hpvm/test/regressionTests/GenVISC/LeafBindEdge.ll
rename to hpvm/test/regressionTests/GenHPVM/LeafBindEdge.ll
diff --git a/hpvm/test/regressionTests/GenVISC/LeafNodeGetters.ll b/hpvm/test/regressionTests/GenHPVM/LeafNodeGetters.ll
similarity index 100%
rename from hpvm/test/regressionTests/GenVISC/LeafNodeGetters.ll
rename to hpvm/test/regressionTests/GenHPVM/LeafNodeGetters.ll
diff --git a/hpvm/test/regressionTests/GenVISC/PopAndPush.ll b/hpvm/test/regressionTests/GenHPVM/PopAndPush.ll
similarity index 100%
rename from hpvm/test/regressionTests/GenVISC/PopAndPush.ll
rename to hpvm/test/regressionTests/GenHPVM/PopAndPush.ll
diff --git a/hpvm/test/regressionTests/GenVISC/ThreeLevel.ll b/hpvm/test/regressionTests/GenHPVM/ThreeLevel.ll
similarity index 100%
rename from hpvm/test/regressionTests/GenVISC/ThreeLevel.ll
rename to hpvm/test/regressionTests/GenHPVM/ThreeLevel.ll
diff --git a/hpvm/test/regressionTests/GenVISC/ThreeLevelEdge.ll b/hpvm/test/regressionTests/GenHPVM/ThreeLevelEdge.ll
similarity index 100%
rename from hpvm/test/regressionTests/GenVISC/ThreeLevelEdge.ll
rename to hpvm/test/regressionTests/GenHPVM/ThreeLevelEdge.ll
diff --git a/hpvm/test/regressionTests/GenVISC/TwoLaunch.ll b/hpvm/test/regressionTests/GenHPVM/TwoLaunch.ll
similarity index 100%
rename from hpvm/test/regressionTests/GenVISC/TwoLaunch.ll
rename to hpvm/test/regressionTests/GenHPVM/TwoLaunch.ll
diff --git a/hpvm/test/regressionTests/GenVISC/TwoLevel.ll b/hpvm/test/regressionTests/GenHPVM/TwoLevel.ll
similarity index 100%
rename from hpvm/test/regressionTests/GenVISC/TwoLevel.ll
rename to hpvm/test/regressionTests/GenHPVM/TwoLevel.ll
diff --git a/hpvm/test/regressionTests/GenVISC/oneLaunchAlloca.ll b/hpvm/test/regressionTests/GenHPVM/oneLaunchAlloca.ll
similarity index 100%
rename from hpvm/test/regressionTests/GenVISC/oneLaunchAlloca.ll
rename to hpvm/test/regressionTests/GenHPVM/oneLaunchAlloca.ll
diff --git a/hpvm/test/regressionTests/GenVISC/oneLaunchMalloc.ll b/hpvm/test/regressionTests/GenHPVM/oneLaunchMalloc.ll
similarity index 100%
rename from hpvm/test/regressionTests/GenVISC/oneLaunchMalloc.ll
rename to hpvm/test/regressionTests/GenHPVM/oneLaunchMalloc.ll
diff --git a/hpvm/test/regressionTests/GenVISC/AllocationNode.ll b/hpvm/test/regressionTests/GenVISC/AllocationNode.ll
deleted file mode 100644
index db5f3ebbfc..0000000000
--- a/hpvm/test/regressionTests/GenVISC/AllocationNode.ll
+++ /dev/null
@@ -1,206 +0,0 @@
-; RUN: opt -load LLVMGenHPVM.so -S -genhpvm <  %s | FileCheck %s
-; ModuleID = 'AllocationNode.c'
-source_filename = "AllocationNode.c"
-target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
-target triple = "x86_64-unknown-linux-gnu"
-
-%struct.Root = type { i64*, i64* }
-; CHECK-LABEL: struct.Root =
-
-; CHECK-LABEL: %struct.out.Func1 =
-; CHECK-LABEL: %struct.out.Func4 =
-; CHECK-LABEL: %struct.out.Func5 =
-; CHECK-LABEL: %struct.out.Func3 =
-; CHECK-LABEL: %struct.out.Func2 =
-; CHECK-LABEL: %struct.out.PipeRoot =
-
-; CHECK-LABEL: i32 @main(
-; CHECK: [[ALLOCA:%[1-9a-zA-Z]+]] = alloca %struct.Root
-; CHECK: call void @llvm.hpvm.init()
-; CHECK:  [[REGISTER:%[1-9]+]] = bitcast %struct.Root* [[ALLOCA]] to i8*
-; CHECK: call i8* @llvm.hpvm.launch(i8* bitcast (%struct.out.PipeRoot (i64*, i64*)* @PipeRoot_cloned to i8*), i8* [[REGISTER]], i1 false)
-; CHECK-NEXT: call void @llvm.hpvm.wait(i8*
-
-; CHECK-LABEL: @Func1_cloned(
-; CHECK: [[RET1:%[1-9a-zA-Z]+]] = insertvalue %struct.out.Func1 undef,
-; CHECK-NEXT: ret %struct.out.Func1 [[RET1]]
-
-; CHECK-LABEL: @Func4_cloned(
-; CHECK: [[RET4:%[1-9a-zA-Z]+]] = call i8* @llvm.hpvm.malloc(i64
-; CHECK: [[RET4:%[1-9a-zA-Z]+]] = insertvalue %struct.out.Func4 undef, i8* [[RET4]], 0
-; CHECK-NEXT: ret %struct.out.Func4 [[RET4]]
-
-; CHECK-LABEL: @Func5_cloned(
-; CHECK: [[RET5:%[1-9a-zA-Z]+]] = insertvalue %struct.out.Func5 undef,
-; CHECK-NEXT: ret %struct.out.Func5 [[RET5]]
-
-; CHECK-LABEL: @Func3_cloned(
-; CHECK: %Func4_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%struct.out.Func4 (i64*
-; CHECK-NEXT: %Func5_cloned.node = call i8* @llvm.hpvm.createNode2D(i8* bitcast (%struct.out.Func5 (i64*
-; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func4_cloned.node,
-; CHECK-NEXT: call i8* @llvm.hpvm.createEdge(i8* %Func4_cloned.node, i8* %Func5_cloned.node,
-; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func4_cloned.node,
-
-; CHECK-LABEL: @Func2_cloned(
-; CHECK: %Func3_cloned.node = call i8* @llvm.hpvm.createNode1D(i8* bitcast (%struct.out.Func3 (i64*
-; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func3_cloned.node,
-; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func3_cloned.node,
-
-; CHECK-LABEL: @PipeRoot_cloned(i64*
-; CHECK: %Func1_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func1 (i64*
-; CHECK-NEXT: %Func2_cloned.node = call i8* @llvm.hpvm.createNode(i8* bitcast (%struct.out.Func2 (i64*
-; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func1_cloned.node,
-; CHECK-NEXT: call void @llvm.hpvm.bind.input(i8* %Func2_cloned.node,
-; CHECK-NEXT: call i8* @llvm.hpvm.createEdge(i8* %Func1_cloned.node, i8* %Func2_cloned.node,
-; CHECK-NEXT: call void @llvm.hpvm.bind.output(i8* %Func1_cloned.node,
-
-
-; Function Attrs: nounwind uwtable
-define dso_local void @Func1(i64* %In, i64* %Out) #0 {
-; CHECK-NOT: @Func1(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 1, i64* %In, i32 1, i64* %Out) #3
-  tail call void (i32, ...) @__hpvm__return(i32 1, i64* %Out) #3
-  ret void
-}
-
-declare dso_local void @__hpvm__hint(i32) local_unnamed_addr #1
-
-declare dso_local void @__hpvm__attributes(i32, ...) local_unnamed_addr #1
-
-declare dso_local void @__hpvm__return(i32, ...) local_unnamed_addr #1
-
-; Function Attrs: nounwind uwtable
-define dso_local void @Func4(i64* %In, i64* %Out) #0 {
-; CHECK-NOT: @Func4(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 1, i64* %In, i32 1, i64* %Out) #3
-  %0 = load i64, i64* %In, align 8, !tbaa !2
-  %call = tail call i8* @__hpvm__malloc(i64 %0) #3
-  tail call void (i32, ...) @__hpvm__return(i32 1, i8* %call) #3
-  ret void
-}
-
-declare dso_local i8* @__hpvm__malloc(i64) local_unnamed_addr #1
-
-; Function Attrs: nounwind uwtable
-define dso_local void @Func5(i64* %In, i64* %Out) #0 {
-; CHECK-NOT: @Func5(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 1, i64* %In, i32 1, i64* %Out) #3
-  tail call void (i32, ...) @__hpvm__return(i32 1, i64* %Out) #3
-  ret void
-}
-
-; Function Attrs: nounwind uwtable
-define dso_local void @Func3(i64* %In, i64* %Out) #0 {
-; CHECK-NOT: @Func3(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 1, i64* %In, i32 1, i64* %Out) #3
-  %call = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 2, void (i64*, i64*)* nonnull @Func4, i64 3, i64 6) #3
-  %call1 = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 2, void (i64*, i64*)* nonnull @Func5, i64 4, i64 5) #3
-  tail call void @__hpvm__bindIn(i8* %call, i32 0, i32 0, i32 0) #3
-  %call2 = tail call i8* @__hpvm__edge(i8* %call, i8* %call1, i32 1, i32 0, i32 1, i32 0) #3
-  tail call void @__hpvm__bindOut(i8* %call, i32 0, i32 0, i32 0) #3
-  ret void
-}
-
-; Function Attrs: argmemonly nounwind
-declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #2
-
-declare dso_local i8* @__hpvm__createNodeND(i32, ...) local_unnamed_addr #1
-
-declare dso_local void @__hpvm__bindIn(i8*, i32, i32, i32) local_unnamed_addr #1
-
-declare dso_local i8* @__hpvm__edge(i8*, i8*, i32, i32, i32, i32) local_unnamed_addr #1
-
-declare dso_local void @__hpvm__bindOut(i8*, i32, i32, i32) local_unnamed_addr #1
-
-; Function Attrs: argmemonly nounwind
-declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #2
-
-; Function Attrs: nounwind uwtable
-define dso_local void @Func2(i64* %BindIn, i64* %SrcIn, i64* %Out) #0 {
-; CHECK-NOT: @Func2(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 2, i64* %BindIn, i64* %SrcIn, i32 1, i64* %Out) #3
-  %call = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 1, void (i64*, i64*)* nonnull @Func3, i64 3) #3
-  tail call void @__hpvm__bindIn(i8* %call, i32 0, i32 0, i32 0) #3
-  tail call void @__hpvm__bindOut(i8* %call, i32 0, i32 0, i32 0) #3
-  ret void
-}
-
-; Function Attrs: nounwind uwtable
-define dso_local void @PipeRoot(i64* %In, i64* %Out) #0 {
-; CHECK-NOT: @PipeRoot(
-entry:
-  tail call void @__hpvm__hint(i32 1) #3
-  tail call void (i32, ...) @__hpvm__attributes(i32 1, i64* %In, i32 1, i64* %Out) #3
-  %call = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 0, void (i64*, i64*)* nonnull @Func1) #3
-  %call1 = tail call i8* (i32, ...) @__hpvm__createNodeND(i32 0, void (i64*, i64*, i64*)* nonnull @Func2) #3
-  tail call void @__hpvm__bindIn(i8* %call, i32 0, i32 0, i32 0) #3
-  tail call void @__hpvm__bindIn(i8* %call1, i32 0, i32 0, i32 0) #3
-  %call2 = tail call i8* @__hpvm__edge(i8* %call, i8* %call1, i32 1, i32 0, i32 1, i32 0) #3
-  tail call void @__hpvm__bindOut(i8* %call, i32 0, i32 0, i32 0) #3
-  ret void
-}
-
-; Function Attrs: nounwind uwtable
-define dso_local i32 @main() local_unnamed_addr #0 {
-entry:
-  %In = alloca i64, align 8
-  %Out = alloca i64, align 8
-  %RootArgs = alloca %struct.Root, align 8
-  %0 = bitcast i64* %In to i8*
-  call void @llvm.lifetime.start.p0i8(i64 8, i8* nonnull %0) #3
-  store i64 1, i64* %In, align 8, !tbaa !2
-  %1 = bitcast i64* %Out to i8*
-  call void @llvm.lifetime.start.p0i8(i64 8, i8* nonnull %1) #3
-  store i64 0, i64* %Out, align 8, !tbaa !2
-  %2 = bitcast %struct.Root* %RootArgs to i8*
-  call void @llvm.lifetime.start.p0i8(i64 16, i8* nonnull %2) #3
-  %input = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 0
-  store i64* %In, i64** %input, align 8, !tbaa !6
-  %output = getelementptr inbounds %struct.Root, %struct.Root* %RootArgs, i64 0, i32 1
-  store i64* %Out, i64** %output, align 8, !tbaa !9
-  call void (...) @__hpvm__init() #3
-  %call = call i8* (i32, ...) @__hpvm__launch(i32 0, void (i64*, i64*)* nonnull @PipeRoot, %struct.Root* nonnull %RootArgs) #3
-  call void @__hpvm__wait(i8* %call) #3
-  call void (...) @__hpvm__cleanup() #3
-  call void @llvm.lifetime.end.p0i8(i64 16, i8* nonnull %2) #3
-  call void @llvm.lifetime.end.p0i8(i64 8, i8* nonnull %1) #3
-  call void @llvm.lifetime.end.p0i8(i64 8, i8* nonnull %0) #3
-  ret i32 0
-}
-
-declare dso_local void @__hpvm__init(...) local_unnamed_addr #1
-
-declare dso_local i8* @__hpvm__launch(i32, ...) local_unnamed_addr #1
-
-declare dso_local void @__hpvm__wait(i8*) local_unnamed_addr #1
-
-declare dso_local void @__hpvm__cleanup(...) local_unnamed_addr #1
-
-attributes #0 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-jump-tables"="false" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
-attributes #1 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="true" "use-soft-float"="false" }
-attributes #2 = { argmemonly nounwind }
-attributes #3 = { nounwind }
-
-!llvm.module.flags = !{!0}
-!llvm.ident = !{!1}
-
-!0 = !{i32 1, !"wchar_size", i32 4}
-!1 = !{!"clang version 9.0.0 (https://gitlab.engr.illinois.edu/llvm/hpvm.git 6690f9e7e8b46b96aea222d3e85315cd63545953)"}
-!2 = !{!3, !3, i64 0}
-!3 = !{!"long", !4, i64 0}
-!4 = !{!"omnipotent char", !5, i64 0}
-!5 = !{!"Simple C/C++ TBAA"}
-!6 = !{!7, !8, i64 0}
-!7 = !{!"Root", !8, i64 0, !8, i64 8}
-!8 = !{!"any pointer", !4, i64 0}
-!9 = !{!7, !8, i64 8}
-- 
GitLab