[rs4gc] Optionally directly relocated vector of pointers

[reames]

Summary

This patch teaches rewrite-statepoints-for-gc to relocate vector-of-pointers directly rather than trying to split them. This builds on the recent lowering/IR changes to allow vector typed gc.relocates.

The motivation for this is that we recently found a bug in the vector splitting code where depending on visit order, a vector might not be relocated at some safepoint. Specifically, the bug is that the splitting code wasn't updating the side tables (live vector) of other safepoints. As a result, a vector which was live at two safepoints might not be updated at one of them. However, if you happened to visit safepoints in post order over the dominator tree, everything worked correctly. Weirdly, it turns out that post order is actually an incredibly common order to visit instructions in in practice. Frustratingly, I have not managed to write a test case which actually hits this. I can only reproduce it in large IR files produced by actual applications.

Rather than continue to make this code more complicated, we can remove all of the complexity by just representing the relocation of the entire vector natively in the IR.

At the moment, the new functionality is hidden behind a flag. To use this code, you need to disable use "-rs4gc-split-vector-values=0". Once I have a chance to stress test with this option and get feedback from other users, my plan is to flip the default and remove the original splitting code. I would just remove it now, but given the rareness of the bug, I figured it was better to leave it in place until the new approach has been stress tested.

[mjacob]

lib/Transforms/Scalar/RewriteStatepointsForGC.cpp

1331–1342

You can hoist the construction of the i8 pointer in the correct address space before the conditional by using Type::getInt8PtrTy(M->getContext(), Ty->getPointerAddressSpace()). Then you could further simplify the code by sinking the call to Intrinsic::getDeclaration() to a common path.

The resulting code would be similar to this:

Type *Int8Ptr = Type::getInt8PtrTy(M->getContext(), Ty->getPointerAddressSpace());
if (VectorType *VT = dyn_cast<VectorType>(Ty))
  Int8Ptr = VectorType::get(Int8Ptr, VT->getNumElements());
return Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, {Int8Ptr});

[reames]

address manuel's comments - thanks, that's a lot cleaner!

[mjacob]

LGTM.

Maybe you could move the caching to the lambda. However I'm fine with it if you had good reasons not to do so.

lib/Transforms/Scalar/RewriteStatepointsForGC.cpp

1322–1327

This comment is a bit inaccurate now.

[reames]

In D15982#322785, @mjacob wrote:

Maybe you could move the caching to the lambda. However I'm fine with it if you had good reasons not to do so.

I generally find that having the generation function separate from the caching layer leads to fewer bugs. Maybe it's just me, but I've seen and written a *lot* of bugs where something didn't get cached correctly. I like to keep that portion of the code as simple and obviously correct as possible. I could use a second layer of lambda, but that seemed like a overkill in this situation.

lib/Transforms/Scalar/RewriteStatepointsForGC.cpp

1322–1327

Will tweak before submission.

lib/Transforms/Scalar/RewriteStatepointsForGC.cpp

static void CreateGCRelocates(ArrayRef<Value *> LiveVariables,
 
   auto FindIndex = [](ArrayRef<Value *> LiveVec, Value *Val) {
     auto ValIt = std::find(LiveVec.begin(), LiveVec.end(), Val);
     assert(ValIt != LiveVec.end() && "Val not found in LiveVec!");
     size_t Index = std::distance(LiveVec.begin(), ValIt);
     assert(Index < LiveVec.size() && "Bug in std::find?");
     return Index;
   };
+  Module *M = StatepointToken->getModule();
   
   // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
   // unique declarations for each pointer type, but this proved problematic
   // because the intrinsic mangling code is incomplete and fragile.  Since
   // we're moving towards a single unified pointer type anyways, we can just
   // cast everything to an i8* of the right address space.  A bitcast is added
   // later to convert gc_relocate to the actual value's type. 
-  Module *M = StatepointToken->getModule();
-  auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
-  Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
-  Value *GCRelocateDecl =
-    Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
+  auto getGCRelocateDecl = [&] (Type *Ty) {
+    assert(isHandledGCPointerType(Ty));
+    auto AS = Ty->getScalarType()->getPointerAddressSpace();
+    Type *NewTy = Type::getInt8PtrTy(M->getContext(), AS);
+    if (auto *VT = dyn_cast<VectorType>(Ty))
+      NewTy = VectorType::get(NewTy, VT->getNumElements());
+    return Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate,
+                                     {NewTy});
+  };
+
+  // Lazily populated map from input types to the canonicalized form mentioned
+  // in the comment above.  This should probably be cached somewhere more
+  // broadly.
+  DenseMap<Type*, Value*> TypeToDeclMap;
 
   for (unsigned i = 0; i < LiveVariables.size(); i++) {
     // Generate the gc.relocate call and save the result
     Value *BaseIdx =
       Builder.getInt32(LiveStart + FindIndex(LiveVariables, BasePtrs[i]));
     Value *LiveIdx = Builder.getInt32(LiveStart + i);
 
+    Type *Ty = LiveVariables[i]->getType();
+    if (!TypeToDeclMap.count(Ty))
+      TypeToDeclMap[Ty] = getGCRelocateDecl(Ty);
+    Value *GCRelocateDecl = TypeToDeclMap[Ty];
+
     // only specify a debug name if we can give a useful one
     CallInst *Reloc = Builder.CreateCall(
         GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
         suffixed_name_or(LiveVariables[i], ".relocated", ""));
     // Trick CodeGen into thinking there are lots of free registers at this
     // fake call.
     Reloc->setCallingConv(CallingConv::Cold);
   }
#ifndef NDEBUG
     }
 #endif
   }
   unique_unsorted(Live);
 
 #ifndef NDEBUG
   // sanity check
   for (auto *Ptr : Live)
-    assert(isGCPointerType(Ptr->getType()) && "must be a gc pointer type");
+    assert(isHandledGCPointerType(Ptr->getType()) &&
+           "must be a gc pointer type");
 #endif
 
   relocationViaAlloca(F, DT, Live, Records);
   return !Records.empty();
 }
 
 // Handles both return values and arguments for Functions and CallSites.
 template <typename AttrHolder>

test/Transforms/RewriteStatepointsForGC/deopt-bundles/live-vector-nosplit.ll

+; Test that we can correctly handle vectors of pointers in statepoint 
+; rewriting.  
+; RUN: opt %s -rewrite-statepoints-for-gc -rs4gc-use-deopt-bundles -rs4gc-split-vector-values=0 -S | FileCheck  %s
+
+; A non-vector relocation for comparison
+define i64 addrspace(1)* @test(i64 addrspace(1)* %obj) gc "statepoint-example" {
+; CHECK-LABEL: test
+; CHECK: gc.statepoint
+; CHECK-NEXT: gc.relocate
+; CHECK-NEXT: bitcast
+; CHECK-NEXT: ret i64 addrspace(1)*
+; A base vector from a argument
+entry:
+  call void @do_safepoint() [ "deopt"() ]
+  ret i64 addrspace(1)* %obj
+}
+
+; A vector argument
+define <2 x i64 addrspace(1)*> @test2(<2 x i64 addrspace(1)*> %obj) gc "statepoint-example" {
+; CHECK-LABEL: test2
+; CHECK-NEXT: gc.statepoint
+; CHECK-NEXT: gc.relocate
+; CHECK-NEXT: bitcast
+; CHECK-NEXT: ret <2 x i64 addrspace(1)*>
+  call void @do_safepoint() [ "deopt"() ]
+  ret <2 x i64 addrspace(1)*> %obj
+}
+
+; A load
+define <2 x i64 addrspace(1)*> @test3(<2 x i64 addrspace(1)*>* %ptr) gc "statepoint-example" {
+; CHECK-LABEL: test3
+; CHECK: load
+; CHECK-NEXT: gc.statepoint
+; CHECK-NEXT: gc.relocate
+; CHECK-NEXT: bitcast
+; CHECK-NEXT: ret <2 x i64 addrspace(1)*>
+entry:
+  %obj = load <2 x i64 addrspace(1)*>, <2 x i64 addrspace(1)*>* %ptr
+  call void @do_safepoint() [ "deopt"() ]
+  ret <2 x i64 addrspace(1)*> %obj
+}
+
+declare i32 @fake_personality_function()
+
+; When a statepoint is an invoke rather than a call
+define <2 x i64 addrspace(1)*> @test4(<2 x i64 addrspace(1)*>* %ptr) gc "statepoint-example" personality i32 ()* @fake_personality_function {
+; CHECK-LABEL: test4
+; CHECK: load
+; CHECK-NEXT: gc.statepoint
+entry:
+  %obj = load <2 x i64 addrspace(1)*>, <2 x i64 addrspace(1)*>* %ptr
+  invoke void @do_safepoint() [ "deopt"() ]
+          to label %normal_return unwind label %exceptional_return
+
+normal_return:                                    ; preds = %entry
+; CHECK-LABEL: normal_return:
+; CHECK: gc.relocate
+; CHECK-NEXT: bitcast
+; CHECK-NEXT: ret <2 x i64 addrspace(1)*>
+  ret <2 x i64 addrspace(1)*> %obj
+
+exceptional_return:                               ; preds = %entry
+; CHECK-LABEL: exceptional_return:
+; CHECK: gc.relocate
+; CHECK-NEXT: bitcast
+; CHECK-NEXT: ret <2 x i64 addrspace(1)*>
+  %landing_pad4 = landingpad token
+          cleanup
+  ret <2 x i64 addrspace(1)*> %obj
+}
+
+; A newly created vector
+define <2 x i64 addrspace(1)*> @test5(i64 addrspace(1)* %p) gc "statepoint-example" {
+; CHECK-LABEL: test5
+; CHECK: insertelement
+; CHECK-NEXT: gc.statepoint
+; CHECK-NEXT: gc.relocate
+; CHECK-NEXT: bitcast
+; CHECK-NEXT: ret <2 x i64 addrspace(1)*> %vec.relocated.casted
+entry:
+  %vec = insertelement <2 x i64 addrspace(1)*> undef, i64 addrspace(1)* %p, i32 0
+  call void @do_safepoint() [ "deopt"() ]
+  ret <2 x i64 addrspace(1)*> %vec
+}
+
+; A merge point
+define <2 x i64 addrspace(1)*> @test6(i1 %cnd, <2 x i64 addrspace(1)*>* %ptr) gc "statepoint-example" {
+; CHECK-LABEL: test6
+entry:
+  br i1 %cnd, label %taken, label %untaken
+
+taken:                                            ; preds = %entry
+  %obja = load <2 x i64 addrspace(1)*>, <2 x i64 addrspace(1)*>* %ptr
+  br label %merge
+
+untaken:                                          ; preds = %entry
+  %objb = load <2 x i64 addrspace(1)*>, <2 x i64 addrspace(1)*>* %ptr
+  br label %merge
+
+merge:                                            ; preds = %untaken, %taken
+; CHECK-LABEL: merge:
+; CHECK-NEXT: = phi
+; CHECK-NEXT: gc.statepoint
+; CHECK-NEXT: gc.relocate
+; CHECK-NEXT: bitcast
+; CHECK-NEXT: ret <2 x i64 addrspace(1)*>
+  %obj = phi <2 x i64 addrspace(1)*> [ %obja, %taken ], [ %objb, %untaken ]
+  call void @do_safepoint() [ "deopt"() ]
+  ret <2 x i64 addrspace(1)*> %obj
+}
+
+declare void @do_safepoint()

test/Transforms/RewriteStatepointsForGC/deopt-bundles/live-vector.ll

 ; Test that we can correctly handle vectors of pointers in statepoint 
 ; rewriting.  Currently, we scalarize, but that's an implementation detail.
-; RUN: opt %s -rewrite-statepoints-for-gc -rs4gc-use-deopt-bundles -S | FileCheck  %s
+; RUN: opt %s -rewrite-statepoints-for-gc -rs4gc-use-deopt-bundles -rs4gc-split-vector-values -S | FileCheck  %s
 
 ; A non-vector relocation for comparison
 
 define i64 addrspace(1)* @test(i64 addrspace(1)* %obj) gc "statepoint-example" {
 ; CHECK-LABEL: test
 ; CHECK: gc.statepoint
 ; CHECK-NEXT: gc.relocate
 ; CHECK-NEXT: bitcast

test/Transforms/RewriteStatepointsForGC/live-vector.ll

 ; Test that we can correctly handle vectors of pointers in statepoint 
 ; rewriting.  Currently, we scalarize, but that's an implementation detail.
-; RUN: opt %s -rewrite-statepoints-for-gc -S | FileCheck  %s
+; RUN: opt %s -rewrite-statepoints-for-gc -rs4gc-split-vector-values -S | FileCheck  %s
 
 ; A non-vector relocation for comparison
 define i64 addrspace(1)* @test(i64 addrspace(1)* %obj) gc "statepoint-example" {
 ; CHECK-LABEL: test
 ; CHECK: gc.statepoint
 ; CHECK-NEXT: gc.relocate
 ; CHECK-NEXT: bitcast
 ; CHECK-NEXT: ret i64 addrspace(1)* %obj.relocated.casted