transform (gc): track phi nodes in stack slots

torch2424 · Apr 2, 2020 · ae16b2c · ae16b2c
1 parent 6647c43
commit ae16b2c
Show file tree

Hide file tree

Showing 3 changed files with 106 additions and 2 deletions.
diff --git a/transform/gc.go b/transform/gc.go
@@ -150,11 +150,14 @@ func MakeGCStackSlots(mod llvm.Module) bool {
 				continue
 			}
 			switch ptr.InstructionOpcode() {
-			case llvm.PHI, llvm.GetElementPtr:
+			case llvm.GetElementPtr:
 				// These values do not create new values: the values already
 				// existed locally in this function so must have been tracked
 				// already.
 				continue
+			case llvm.PHI:
+				// While the value may have already been tracked, it may be overwritten in a loop.
+				// Therefore, a second copy must be created to ensure that it is tracked over the entirety of its lifetime.
 			case llvm.ExtractValue, llvm.BitCast:
 				// These instructions do not create new values, but their
 				// original value may not be tracked. So keep tracking them for
@@ -228,11 +231,22 @@ func MakeGCStackSlots(mod llvm.Module) bool {
 
 		// Do a store to the stack object after each new pointer that is created.
 		for i, ptr := range pointers {
-			builder.SetInsertPointBefore(llvm.NextInstruction(ptr))
+			// Insert the store after the pointer value is created.
+			insertionPoint := llvm.NextInstruction(ptr)
+			for !insertionPoint.IsAPHINode().IsNil() {
+				// PHI nodes are required to be at the start of the block.
+				// Insert after the last PHI node.
+				insertionPoint = llvm.NextInstruction(insertionPoint)
+			}
+			builder.SetInsertPointBefore(insertionPoint)
+
+			// Extract a pointer to the appropriate section of the stack object.
 			gep := builder.CreateGEP(stackObject, []llvm.Value{
 				llvm.ConstInt(ctx.Int32Type(), 0, false),
 				llvm.ConstInt(ctx.Int32Type(), uint64(2+len(allocas)+i), false),
 			}, "")
+
+			// Store the pointer into the stack slot.
 			builder.CreateStore(ptr, gep)
 		}
 

diff --git a/transform/testdata/gc-stackslots.ll b/transform/testdata/gc-stackslots.ll
@@ -50,3 +50,37 @@ define i8* @noAllocatingFunction() {
   call void @runtime.trackPointer(i8* %ptr)
   ret i8* %ptr
 }
+
+define i8* @fibNext(i8* %x, i8* %y) {
+  %x.val = load i8, i8* %x
+  %y.val = load i8, i8* %y
+  %out.val = add i8 %x.val, %y.val
+  %out.alloc = call i8* @runtime.alloc(i32 1)
+  call void @runtime.trackPointer(i8* %out.alloc)
+  store i8 %out.val, i8* %out.alloc
+  ret i8* %out.alloc
+}
+
+define i8* @allocLoop() {
+entry:
+  %entry.x = call i8* @runtime.alloc(i32 1)
+  call void @runtime.trackPointer(i8* %entry.x)
+  %entry.y = call i8* @runtime.alloc(i32 1)
+  call void @runtime.trackPointer(i8* %entry.y)
+  store i8 1, i8* %entry.y
+  br label %loop
+
+loop:
+  %prev.y = phi i8* [ %entry.y, %entry ], [ %prev.x, %loop ]
+  %prev.x = phi i8* [ %entry.x, %entry ], [ %next.x, %loop ]
+  call void @runtime.trackPointer(i8* %prev.x)
+  call void @runtime.trackPointer(i8* %prev.y)
+  %next.x = call i8* @fibNext(i8* %prev.x, i8* %prev.y)
+  call void @runtime.trackPointer(i8* %next.x)
+  %next.x.val = load i8, i8* %next.x
+  %loop.done = icmp ult i8 40, %next.x.val
+  br i1 %loop.done, label %end, label %loop
+
+end:
+  ret i8* %next.x
+}
diff --git a/transform/testdata/gc-stackslots.out.ll b/transform/testdata/gc-stackslots.out.ll
@@ -56,3 +56,59 @@ define i8* @noAllocatingFunction() {
   %ptr = call i8* @getPointer()
   ret i8* %ptr
 }
+
+define i8* @fibNext(i8* %x, i8* %y) {
+  %gc.stackobject = alloca { %runtime.stackChainObject*, i32, i8* }
+  store { %runtime.stackChainObject*, i32, i8* } { %runtime.stackChainObject* null, i32 1, i8* null }, { %runtime.stackChainObject*, i32, i8* }* %gc.stackobject
+  %1 = load %runtime.stackChainObject*, %runtime.stackChainObject** @runtime.stackChainStart
+  %2 = getelementptr { %runtime.stackChainObject*, i32, i8* }, { %runtime.stackChainObject*, i32, i8* }* %gc.stackobject, i32 0, i32 0
+  store %runtime.stackChainObject* %1, %runtime.stackChainObject** %2
+  %3 = bitcast { %runtime.stackChainObject*, i32, i8* }* %gc.stackobject to %runtime.stackChainObject*
+  store %runtime.stackChainObject* %3, %runtime.stackChainObject** @runtime.stackChainStart
+  %x.val = load i8, i8* %x
+  %y.val = load i8, i8* %y
+  %out.val = add i8 %x.val, %y.val
+  %out.alloc = call i8* @runtime.alloc(i32 1)
+  %4 = getelementptr { %runtime.stackChainObject*, i32, i8* }, { %runtime.stackChainObject*, i32, i8* }* %gc.stackobject, i32 0, i32 2
+  store i8* %out.alloc, i8** %4
+  store i8 %out.val, i8* %out.alloc
+  store %runtime.stackChainObject* %1, %runtime.stackChainObject** @runtime.stackChainStart
+  ret i8* %out.alloc
+}
+
+define i8* @allocLoop() {
+entry:
+  %gc.stackobject = alloca { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }
+  store { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* } { %runtime.stackChainObject* null, i32 5, i8* null, i8* null, i8* null, i8* null, i8* null }, { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }* %gc.stackobject
+  %0 = load %runtime.stackChainObject*, %runtime.stackChainObject** @runtime.stackChainStart
+  %1 = getelementptr { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }, { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }* %gc.stackobject, i32 0, i32 0
+  store %runtime.stackChainObject* %0, %runtime.stackChainObject** %1
+  %2 = bitcast { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }* %gc.stackobject to %runtime.stackChainObject*
+  store %runtime.stackChainObject* %2, %runtime.stackChainObject** @runtime.stackChainStart
+  %entry.x = call i8* @runtime.alloc(i32 1)
+  %3 = getelementptr { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }, { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }* %gc.stackobject, i32 0, i32 2
+  store i8* %entry.x, i8** %3
+  %entry.y = call i8* @runtime.alloc(i32 1)
+  %4 = getelementptr { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }, { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }* %gc.stackobject, i32 0, i32 3
+  store i8* %entry.y, i8** %4
+  store i8 1, i8* %entry.y
+  br label %loop
+
+loop:
+  %prev.y = phi i8* [ %entry.y, %entry ], [ %prev.x, %loop ]
+  %prev.x = phi i8* [ %entry.x, %entry ], [ %next.x, %loop ]
+  %5 = getelementptr { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }, { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }* %gc.stackobject, i32 0, i32 5
+  store i8* %prev.y, i8** %5
+  %6 = getelementptr { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }, { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }* %gc.stackobject, i32 0, i32 4
+  store i8* %prev.x, i8** %6
+  %next.x = call i8* @fibNext(i8* %prev.x, i8* %prev.y)
+  %7 = getelementptr { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }, { %runtime.stackChainObject*, i32, i8*, i8*, i8*, i8*, i8* }* %gc.stackobject, i32 0, i32 6
+  store i8* %next.x, i8** %7
+  %next.x.val = load i8, i8* %next.x
+  %loop.done = icmp ult i8 40, %next.x.val
+  br i1 %loop.done, label %end, label %loop
+
+end:
+  store %runtime.stackChainObject* %0, %runtime.stackChainObject** @runtime.stackChainStart
+  ret i8* %next.x
+}