[TapirUtils] Fix logic to insert fixups for outputs of taskframes.

- Fix crash when inserting taskframe fixups for a taskframe with no continuation.
- Ensure that allocas for taskframe fixups are properly inserted within the parent taskframe.
- Avoid inserting taskframe load guards for unassociated taskframes.

llvm/include/llvm/Analysis/TapirTaskInfo.h

@@ -29,6 +29,7 @@
 #include "llvm/IR/PassManager.h"
 #include "llvm/Pass.h"
 #include "llvm/Support/Allocator.h"
+#include "llvm/Support/Casting.h"
 #include <utility>
 
 namespace llvm {
@@ -101,6 +102,12 @@ class Spindle {
       ++D;
     return D;
   }
+  BasicBlock::iterator getTaskFrameFirstInsertionPt() {
+    if (Instruction *TFCreate =
+            dyn_cast_or_null<Instruction>(getTaskFrameCreate()))
+      return TFCreate->getNextNode()->getIterator();
+    return getEntry()->getFirstInsertionPt();
+  }
 
   Task *getTaskFromTaskFrame() const;
 

llvm/lib/Transforms/Utils/TapirUtils.cpp

@@ -19,8 +19,10 @@
 #include "llvm/IR/Dominators.h"
 #include "llvm/IR/EHPersonalities.h"
 #include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/Instruction.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/Intrinsics.h"
+#include "llvm/Support/Casting.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
 #include "llvm/Transforms/Utils/Cloning.h"
 #include "llvm/Transforms/Utils/Local.h"
@@ -1003,6 +1005,10 @@ void llvm::SerializeDetach(DetachInst *DI, BasicBlock *ParentEntry,
   // Erase instructions marked to be erased.
   for (Instruction *I : ToErase)
     I->eraseFromParent();
+  if (ReplaceWithTaskFrame && TaskFrame && TaskFrame->use_empty())
+    cast<Instruction>(TaskFrame)->eraseFromParent();
+  if (isa_and_nonnull<Instruction>(SyncRegion) && SyncRegion->use_empty())
+    cast<Instruction>(SyncRegion)->eraseFromParent();
 
   // Update dominator tree.
   if (DT) {
@@ -1765,12 +1771,12 @@ void llvm::fixupTaskFrameExternalUses(Spindle *TF, const TaskInfo &TI,
     return;
   Task *T = TF->getTaskFrameUser();
 
-  LLVM_DEBUG(dbgs() << "fixupTaskFrameExternalUses: spindle@"
-             << TF->getEntry()->getName() << "\n");
   LLVM_DEBUG({
-      if (T)
-        dbgs() << "  used by task@" << T->getEntry()->getName() << "\n";
-    });
+    dbgs() << "fixupTaskFrameExternalUses: spindle@"
+           << TF->getEntry()->getName() << "\n";
+    if (T)
+      dbgs() << "  used by task@" << T->getEntry()->getName() << "\n";
+  });
 
   // Get the set of basic blocks in the taskframe spindles.  At the same time,
   // find the continuation of corresponding taskframe.resume intrinsics.
@@ -1880,12 +1886,11 @@ void llvm::fixupTaskFrameExternalUses(Spindle *TF, const TaskInfo &TI,
   for (auto &TFInstr : ToRewrite) {
     LLVM_DEBUG(dbgs() << "Fixing taskframe output " << *TFInstr.first << "\n");
     // Create an allocation to store the result of the instruction.
-    BasicBlock *ParentEntry;
-    if (Spindle *ParentTF = TF->getTaskFrameParent())
-      ParentEntry = ParentTF->getEntry();
-    else
-      ParentEntry = TF->getParentTask()->getEntry();
-    IRBuilder<> Builder(&*ParentEntry->getFirstInsertionPt());
+    Spindle *ParentS = TF->getTaskFrameParent()
+                           ? TF->getTaskFrameParent()
+                           : TF->getParentTask()->getEntrySpindle();
+    BasicBlock::iterator ParentInsertionPt = ParentS->getTaskFrameFirstInsertionPt();
+    IRBuilder<> Builder(&*ParentInsertionPt);
     Type *TFInstrTy = TFInstr.first->getType();
     AllocaInst *AI = Builder.CreateAlloca(TFInstrTy);
     AI->setName(TFInstr.first->getName());
@@ -1898,13 +1903,18 @@ void llvm::fixupTaskFrameExternalUses(Spindle *TF, const TaskInfo &TI,
       Builder.SetInsertPoint(&*(++TFInstr.first->getIterator()));
     Builder.CreateStore(TFInstr.first, AI);
 
-    // Load the result of the instruction at the continuation.
-    Builder.SetInsertPoint(&*Continuation->getFirstInsertionPt());
-    Builder.CreateCall(
-        Intrinsic::getDeclaration(M, Intrinsic::taskframe_load_guard,
-                                  { AI->getType() }), { AI });
-    LoadInst *ContinVal = Builder.CreateLoad(TFInstrTy, AI);
+    LoadInst *ContinVal = nullptr;
     LoadInst *EHContinVal = nullptr;
+    // Load the result of the instruction at the continuation.
+    if (Continuation) {
+      Builder.SetInsertPoint(&*Continuation->getFirstInsertionPt());
+      if (T)
+        Builder.CreateCall(
+            Intrinsic::getDeclaration(M, Intrinsic::taskframe_load_guard,
+                                      {AI->getType()}),
+            {AI});
+      ContinVal = Builder.CreateLoad(TFInstrTy, AI);
+    }
 
     // For each external use, replace the use with a load from the alloca.
     for (Use *UseToRewrite : TFInstr.second) {
@@ -1919,9 +1929,11 @@ void llvm::fixupTaskFrameExternalUses(Spindle *TF, const TaskInfo &TI,
         // If necessary, load the value at the taskframe.resume continuation.
         if (!EHContinVal) {
           Builder.SetInsertPoint(&*(TFResumeContin->getFirstInsertionPt()));
-          Builder.CreateCall(
-              Intrinsic::getDeclaration(M, Intrinsic::taskframe_load_guard,
-                                        { AI->getType() }), { AI });
+          if (T)
+            Builder.CreateCall(
+                Intrinsic::getDeclaration(M, Intrinsic::taskframe_load_guard,
+                                          {AI->getType()}),
+                {AI});
           EHContinVal = Builder.CreateLoad(TFInstrTy, AI);
         }
 
@@ -1933,9 +1945,11 @@ void llvm::fixupTaskFrameExternalUses(Spindle *TF, const TaskInfo &TI,
       }
 
       // Rewrite to use the value loaded at the continuation.
-      if (UseToRewrite->get()->hasValueHandle())
-        ValueHandleBase::ValueIsRAUWd(*UseToRewrite, ContinVal);
-      UseToRewrite->set(ContinVal);
+      if (ContinVal) {
+        if (UseToRewrite->get()->hasValueHandle())
+          ValueHandleBase::ValueIsRAUWd(*UseToRewrite, ContinVal);
+        UseToRewrite->set(ContinVal);
+      }
     }
   }
 }

llvm/test/Transforms/Tapir/alloca-insert-split-taskframe.ll

@@ -0,0 +1,161 @@
+; RUN: opt < %s -passes="tapir2target" -tapir-target=opencilk -use-opencilk-runtime-bc=false -debug-abi-calls -S | FileCheck %s
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+%"struct.parlay::sequence_internal::sequence_base<double, parlay::allocator<double>, false>::storage_impl::capacitated_buffer::header" = type { i64, %union.anon.229 }
+%union.anon.229 = type { [1 x i8], [7 x i8] }
+
+; Function Attrs: nounwind willreturn memory(argmem: readwrite)
+declare token @llvm.syncregion.start() #0
+
+; Function Attrs: nounwind willreturn memory(argmem: readwrite)
+declare token @llvm.tapir.runtime.start() #0
+
+; Function Attrs: nounwind willreturn memory(argmem: readwrite)
+declare token @llvm.taskframe.create() #0
+
+; Function Attrs: nounwind willreturn memory(argmem: readwrite)
+declare void @llvm.tapir.runtime.end(token) #0
+
+; Function Attrs: nounwind willreturn memory(argmem: readwrite)
+declare void @llvm.taskframe.end(token) #0
+
+define fastcc void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2(ptr %call.i.i.i.i31.i874.ls2) personality ptr null {
+invoke.cont61.tf.i.i.ls2:
+  %syncreg15.i.i.i.i.i.ls2 = tail call token @llvm.syncregion.start()
+  %0 = tail call token @llvm.syncregion.start()
+  %tf.i.i.i.i.ls2 = call token @llvm.taskframe.create()
+  %1 = call token @llvm.tapir.runtime.start()
+  call void @llvm.tapir.runtime.end(token %1)
+  %tf.i872.ls2 = call token @llvm.taskframe.create()
+  %2 = getelementptr %"struct.parlay::sequence_internal::sequence_base<double, parlay::allocator<double>, false>::storage_impl::capacitated_buffer::header", ptr %call.i.i.i.i31.i874.ls2, i64 0, i32 1
+  %3 = call token @llvm.tapir.runtime.start()
+  detach within %syncreg15.i.i.i.i.i.ls2, label %det.achd.i.i.ls2, label %det.cont.i.i.ls2 unwind label %lpad17.i.i.ls2
+
+vector.ph3079.ls2:                                ; No predecessors!
+  %4 = getelementptr double, ptr %2, i64 0
+  br label %invoke.cont.i.i.i.i.i266.i.i.ls2
+
+invoke.cont.i.i.i.i.i266.i.i.ls2:                 ; preds = %det.cont.i.i.ls2, %vector.ph3079.ls2
+  call void @llvm.taskframe.end(token %tf.i872.ls2)
+  %tf.i1536.ls2 = call token @llvm.taskframe.create()
+  %5 = call token @llvm.tapir.runtime.start()
+  detach within %syncreg15.i.i.i.i.i.ls2, label %det.achd.i.i1532.ls2, label %det.cont.i.i1526.ls2 unwind label %lpad17.i.i1523.ls2
+
+det.cont.i.i1526.ls2:                             ; preds = %det.achd.i.i1532.ls2, %invoke.cont.i.i.i.i.i266.i.i.ls2
+  sync within %syncreg15.i.i.i.i.i.ls2, label %sync.continue28.i.i1530.ls2
+
+sync.continue28.i.i1530.ls2:                      ; preds = %det.cont.i.i1526.ls2
+  call void @llvm.taskframe.end(token %tf.i1536.ls2)
+  call void @llvm.taskframe.end(token %tf.i.i.i.i.ls2)
+  ret void
+
+lpad17.i.i1523.ls2:                               ; preds = %invoke.cont.i.i.i.i.i266.i.i.ls2
+  %6 = landingpad { ptr, i32 }
+          cleanup
+  call void @llvm.tapir.runtime.end(token %5)
+  unreachable
+
+det.cont.i.i.ls2:                                 ; preds = %det.achd.i.i.ls2, %invoke.cont61.tf.i.i.ls2
+  sync within %syncreg15.i.i.i.i.i.ls2, label %invoke.cont.i.i.i.i.i266.i.i.ls2
+
+lpad17.i.i.ls2:                                   ; preds = %invoke.cont61.tf.i.i.ls2
+  %7 = landingpad { ptr, i32 }
+          cleanup
+  call void @llvm.tapir.runtime.end(token %3)
+  unreachable
+
+det.achd.i.i1532.ls2:                             ; preds = %invoke.cont.i.i.i.i.i266.i.i.ls2
+  reattach within %syncreg15.i.i.i.i.i.ls2, label %det.cont.i.i1526.ls2
+
+det.achd.i.i.ls2:                                 ; preds = %invoke.cont61.tf.i.i.ls2
+  reattach within %syncreg15.i.i.i.i.i.ls2, label %det.cont.i.i.ls2
+}
+
+; CHECK-LABEL: define {{.*}}void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2(ptr %call.i.i.i.i31.i874.ls2)
+; CHECK: invoke.cont61.tf.i.i.ls2:
+; CHECK-NEXT: %[[FIXUP_ALLOCA:.+]] = alloca ptr
+; CHECK-NEXT: call token @llvm.syncregion.start()
+; CHECK-NEXT: call token @llvm.syncregion.start()
+; CHECK-NEXT: call {{.*}}void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2.outline_invoke.cont61.tf.i.i.ls2.tf.otf0(ptr %call.i.i.i.i31.i874.ls2, ptr %[[FIXUP_ALLOCA]])
+; CHECK-NEXT: br label %sync.continue28.i.i1530.ls2.tfend
+
+; CHECK: sync.continue28.i.i1530.ls2.tfend:
+; CHECK: ret void
+
+; CHECK-LABEL: define {{.*}}void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2.outline_det.achd.i.i.ls2.otd1
+; CHECK: (ptr %[[ARG:.+]])
+; CHECK: call void @__cilkrts_detach(
+; CHECK: ret void
+
+; CHECK-LABEL: define {{.*}}void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2.outline_invoke.cont61.tf.i.i.ls2.tf.tf.otf1(ptr {{.*}}%call.i.i.i.i31.i874.ls2.otf1,
+; CHECK: ptr {{.*}}%[[ARG:.+]])
+
+; CHECK: invoke.cont61.tf.i.i.ls2.tf.tf.otf1:
+; CHECK: %[[ADDR:.+]] = getelementptr %"struct.parlay::sequence_internal::sequence_base<double, parlay::allocator<double>, false>::storage_impl::capacitated_buffer::header", ptr %call.i.i.i.i31.i874.ls2.otf1, i64 0, i32 1
+; CHECK-NEXT: store ptr %[[ADDR]], ptr %[[ARG]]
+; CHECK-NEXT: call void @__cilkrts_enter_frame(
+; CHECK-NEXT: %[[TAPIR_RT_START:.+]] = call token @llvm.tapir.runtime.start()
+; CHECK-NEXT: call i32 @__cilk_prepare_spawn(
+
+; CHECK: invoke {{.*}}void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2.outline_det.achd.i.i.ls2.otd1(
+; CHECK-NEXT: to label %det.cont.i.i.ls2.otf1 unwind label %lpad17.i.i.ls2.otf1
+
+; CHECK: lpad17.i.i.ls2.otf1:
+; CHECK: landingpad
+; CHECK: call void @llvm.tapir.runtime.end(token %[[TAPIR_RT_START]])
+; CHECK-NEXT: unreachable
+
+; CHECK: det.cont.i.i.ls2.otf1:
+; CHECK-NEXT: call void @__cilk_sync(
+
+; CHECK: ret void
+
+; CHECK-LABEL: define {{.*}}void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2.outline_det.achd.i.i1532.ls2.otd1(
+; CHECK: (ptr %[[ARG:.+]])
+; CHECK: call void @__cilkrts_detach(
+; CHECK: ret void
+
+; CHECK-LABEL: define {{.*}}void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2.outline_invoke.cont.i.i.i.i.i266.i.i.ls2.tf.otf1()
+; CHECK: invoke.cont.i.i.i.i.i266.i.i.ls2.tf.otf1:
+; CHECK: call void @__cilkrts_enter_frame(
+; CHECK-NEXT: %[[TAPIR_RT_START:.+]] = call token @llvm.tapir.runtime.start()
+; CHECK-NEXT: call i32 @__cilk_prepare_spawn(
+
+; CHECK: invoke {{.*}}void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2.outline_det.achd.i.i1532.ls2.otd1(
+; CHECK-NEXT: to label %det.cont.i.i1526.ls2.otf1 unwind label %lpad17.i.i1523.ls2.otf1
+
+; CHECK: lpad17.i.i1523.ls2.otf1:
+; CHECK-NEXT: landingpad
+; CHECK: call void @llvm.tapir.runtime.end(token %[[TAPIR_RT_START]])
+; CHECK-NEXT: unreachable
+
+; CHECK: det.cont.i.i1526.ls2.otf1:
+; CHECK-NEXT: call void @__cilk_sync(
+
+; CHECK: ret void
+
+; CHECK-LABEL: define {{.*}}void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2.outline_invoke.cont61.tf.i.i.ls2.tf.otf0(ptr align 1 %call.i.i.i.i31.i874.ls2.otf0,
+; CHECK: ptr {{.*}}%[[ARG:.+]])
+; CHECK: %[[FIXUP_ALLOCA:.+]] = alloca ptr
+
+; CHECK: call void @llvm.lifetime.start.p0(i64 8, ptr %[[FIXUP_ALLOCA]])
+; CHECK-NEXT: call void @__cilkrts_enter_frame(
+; CHECK-NEXT: %[[TAPIR_RT_START:.+]] = call token @llvm.tapir.runtime.start()
+; CHECK-NEXT: call void @__cilk_parent_epilogue(
+; CHECK-NEXT: call void @llvm.tapir.runtime.end(token %[[TAPIR_RT_START]])
+; CHECK-NEXT: call {{.*}}void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2.outline_invoke.cont61.tf.i.i.ls2.tf.tf.otf1(ptr %call.i.i.i.i31.i874.ls2.otf0, ptr %[[FIXUP_ALLOCA]])
+
+; CHECK: %[[FIXUP_LOAD:.+]] = load ptr, ptr %[[FIXUP_ALLOCA]]
+; CHECK-NEXT: store ptr %[[FIXUP_LOAD]], ptr %[[ARG]]
+; CHECK-NEXT: call {{.*}}void @_Z6kmeansIFdRKN6parlay8sequenceIdNS0_9allocatorIdEELb0EEES6_EEDaRNS1_IS4_NS2_IS4_EELb0EEEiRT_d.outline_pfor.cond.i.i.i182.ls2.outline_invoke.cont.i.i.i.i.i266.i.i.ls2.tf.otf1()
+
+; CHECK: call void @llvm.lifetime.end.p0(i64 8, ptr %[[FIXUP_ALLOCA]])
+
+; CHECK: ret void
+
+; uselistorder directives
+uselistorder ptr null, { 1, 2, 0 }
+
+attributes #0 = { nounwind willreturn memory(argmem: readwrite) }

llvm/test/Transforms/Tapir/inline-detach-unwind.ll

@@ -4,10 +4,13 @@
 target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
 target triple = "arm64-apple-macosx15.0.0"
 
+; Function Attrs: nounwind willreturn memory(argmem: readwrite)
+declare token @llvm.syncregion.start() #0
+
 ; CHECK: define void @_ZNK5Graph17pbfs_walk_PennantEP7PennantIiERH3BagIiEjPj()
 
 ; CHECK: pfor.cond:
-; CHECK-NEXT: detach within none, label %pfor.body.entry, label %pfor.cond unwind label %lpad59
+; CHECK-NEXT: detach within %syncreg, label %pfor.body.entry, label %pfor.cond unwind label %lpad59
 
 ; CHECK: pfor.body.entry:
 ; CHECK: br label %[[INLINED_PFOR_COND:.+]]
@@ -25,18 +28,19 @@ target triple = "arm64-apple-macosx15.0.0"
 
 define void @_ZNK5Graph17pbfs_walk_PennantEP7PennantIiERH3BagIiEjPj() personality ptr null {
 entry:
+  %syncreg = tail call token @llvm.syncregion.start()
   %0 = call token @llvm.tapir.runtime.start()
   br label %pfor.cond
 
 pfor.cond:                                        ; preds = %pfor.preattach, %pfor.cond, %entry
-  detach within none, label %pfor.body.entry, label %pfor.cond unwind label %lpad59
+  detach within %syncreg, label %pfor.body.entry, label %pfor.cond unwind label %lpad59
 
 pfor.body.entry:                                  ; preds = %pfor.cond
   invoke fastcc void @_ZL14pbfs_proc_NodePKiiRH3BagIiEjPjS0_S0_(ptr null, i32 0, ptr null, i32 0, ptr null, ptr null, ptr null)
           to label %pfor.preattach unwind label %lpad49
 
 pfor.preattach:                                   ; preds = %pfor.body.entry
-  reattach within none, label %pfor.cond
+  reattach within %syncreg, label %pfor.cond
 
 lpad49:                                           ; preds = %pfor.body.entry
   %1 = landingpad { ptr, i32 }
@@ -52,10 +56,11 @@ lpad59:                                           ; preds = %pfor.cond
 
 define fastcc void @_ZL14pbfs_proc_NodePKiiRH3BagIiEjPjS0_S0_(ptr %n, i32 %fillSize, ptr %next, i32 %newdist, ptr %distances, ptr %nodes, ptr %edges) personality ptr null {
 entry:
+  %syncreg = tail call token @llvm.syncregion.start()
   br label %pfor.cond
 
 pfor.cond:                                        ; preds = %pfor.cond, %entry
-  detach within none, label %pfor.body.entry, label %pfor.cond unwind label %lpad20
+  detach within %syncreg, label %pfor.body.entry, label %pfor.cond unwind label %lpad20
 
 pfor.body.entry:                                  ; preds = %pfor.cond
   br label %for.cond

llvm/test/Transforms/Tapir/inline-nounwind-detach-into-invoked-taskframe.ll

@@ -22,12 +22,12 @@ for.cond86:                                       ; preds = %for.cond86, %entry
 
 ; CHECK: [[IF_ELSE_I]]:
 ; CHECK-NEXT: %[[TF_I:.+]] = call token @llvm.taskframe.create()
-; CHECK-NEXT: %syncreg19.i.i3 = call token @llvm.syncregion.start()
-; CHECK-NEXT: detach within %syncreg19.i.i3, label %det.achd.i.i, label
+; CHECK: %syncreg19.i.i5 = call token @llvm.syncregion.start()
+; CHECK-NEXT: detach within %syncreg19.i.i5, label %det.achd.i.i, label
 ; CHECK-NOT: unwind label
 
 ; CHECK: det.achd.i.i:
-; CHECK-NEXT: reattach within %syncreg19.i.i3, label
+; CHECK-NEXT: reattach within %syncreg19.i.i5, label
 
 lpad90:                                           ; preds = %for.cond86
   %0 = landingpad { ptr, i32 }
@@ -53,27 +53,29 @@ if.else:                                          ; preds = %entry
 
 define linkonce_odr void @_ZN6parlay12parallel_forIZN4gbbs9vertexMapINS1_16vertexSubsetDataINS1_5emptyEEENS1_2bc37SSBetweennessCentrality_Back_Vertex_FINS_8sequenceIbSaIbELb0EEENS8_IdSaIdELb0EEEEELi0EEEvRT_T0_mEUlmE_EEvmmOSE_lb() {
 entry:
-  detach within none, label %pfor.body.entry, label %pfor.inc
+  %syncreg = tail call token @llvm.syncregion.start()
+  detach within %syncreg, label %pfor.body.entry, label %pfor.inc
 
 pfor.body.entry:                                  ; preds = %entry
-  reattach within none, label %pfor.inc
+  reattach within %syncreg, label %pfor.inc
 
 pfor.inc:                                         ; preds = %pfor.body.entry, %entry
   ret void
 }
 
 define linkonce_odr void @_ZN6parlay12parallel_forIZN4gbbs9vertexMapINS1_16vertexSubsetDataINS1_5emptyEEENS1_2bc37SSBetweennessCentrality_Back_Vertex_FINS_8sequenceIbSaIbELb0EEENS8_IdSaIdELb0EEEEELi0EEEvRT_T0_mEUlmE0_EEvmmOSE_lb(i64 %granularity) {
 entry:
+  %syncreg = tail call token @llvm.syncregion.start()
   %syncreg19 = call token @llvm.syncregion.start()
   %cmp = icmp eq i64 %granularity, 0
   br i1 %cmp, label %pfor.cond, label %if.else
 
 pfor.cond:                                        ; preds = %pfor.body.entry, %pfor.cond, %entry
-  detach within none, label %pfor.body.entry, label %pfor.cond
+  detach within %syncreg, label %pfor.body.entry, label %pfor.cond
 
 pfor.body.entry:                                  ; preds = %pfor.cond
   call void @_ZZN4gbbs9vertexMapINS_16vertexSubsetDataINS_5emptyEEENS_2bc37SSBetweennessCentrality_Back_Vertex_FIN6parlay8sequenceIbSaIbELb0EEENS7_IdSaIdELb0EEEEELi0EEEvRT_T0_mENKUlmE0_clEm()
-  reattach within none, label %pfor.cond
+  reattach within %syncreg, label %pfor.cond
 
 if.else:                                          ; preds = %entry
   detach within %syncreg19, label %det.achd, label %det.cont