[AbstractAttributor] Fold function calls to `__kmpc_is_spmd_exec_mode` if possible

[tianshilei1992]

Summary

In the device runtime there are many function calls to __kmpc_is_spmd_exec_mode
to query the execution mode of current kernels. In many cases, user programs
only contain target region executing in one mode. As a consequence, those runtime
function calls will only return one value. If we can get rid of these function
calls during compliation, it can potentially improve performance.

In this patch, we use AAKernelInfo to analyze kernel execution. Basically, for
each kernel (device) function F, we collect all kernel entries K that can
reach F. In each iteration, we go through all reaching kernel entries and check
their execution mode. If F can only be reached by kernel entries with same mode,
we update a map from CallBase * to Constant * to corresponding value. In
manifest stage, if any entry of the map is not nullptr, which means the function
call can be folded to the Constant *, we replace all uses of the function call
and remove it.

Later we will also add more foldable functions, such as isMainThread.

[jdoerfert]

Test missing. Other than that mostly nits.

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

501

Unsure, I think this way is more natural.

501

2874

use the changeValueAfterManifest interface.

3228–3235

3228–3235

Remove

3271

Add a comment why we would want this (here).

3278–3279

3281

Move this before the kernel handling, then you can exit if it is not a kernel. Alternatively, put the kernel handling in a helper.

3328

Maybe put this into a helper as well. And the "reaching kernels" update. Will make it easier to read in the future.

[tianshilei1992]

fix coments

[JonChesterfield]

Interesting, thanks. For the functions that could be called by generic or spmd kernels, we could clone them (one for each) as (I think) we know at the entry point which it will be.

We probably want this after any transforms that try to convert generic kernels to spmd. Not sure how to order that.

[jdoerfert]

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

2871–2874

[tianshilei1992]

• fixed comments
• added a wrapper for kernel entry set up

[tianshilei1992]

rebase

[tianshilei1992]

fixed rebase issue

[tianshilei1992]

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

3391

This condition should include SPMDization, but what should be used?

[tianshilei1992]

rebase

[tianshilei1992]

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

507

We probably don't need this variable at all.

[tianshilei1992]

rebase

[jdoerfert]

Test missing.

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

538

?

2885–2893

here we need to return changed or we need to set changed to CHANGED if changeToSPMDMode returned true.

3267

Why !assumed() ? Wouldn't that mean we only update if we have unknown reaching kernels?

3341

Early exit please.

3350

I'm not so sure about this assertion. Maybe handle the case conservatively instead.

3351–3352

[tianshilei1992]

fixed comments

[tianshilei1992]

revmoed unrelated change

[jdoerfert]

Test still missing ;)

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

3267

This is better but I'm still not sure if it's correct. When we go from valid to invalid we stop updating the map, which is bad because the content is still in there.

[JonChesterfield]

I wonder if this is separable.

Could we have an attribute, 'openmp_called_in_spmd_mode' or similar, which we apply to functions when that is always true, similar to the above logic.

We could then have a pass which, for functions with that attribute set, folds call instructions within that function appropriately.

edit: Partly thinking of splitting it to simplify testing - we can write various call trees and check the right ones pick up the attribute, and separate tests to check we fold the calls based on that attribute.

[tianshilei1992]

added tests

[tianshilei1992]

llvm/test/Transforms/OpenMP/is_spmd_exec_mode_unfoldable.ll

343 ↗

(On Diff #357330)

This function will not be called if we remove the way to tell if it is in SPMD mode from ident.

[tianshilei1992]

update foldable test

[tianshilei1992]

update the test case with update_test_checks

[tianshilei1992]

update test

[JonChesterfield]

Thought about this a little more. I think there's a dead-simple 90% solution available.

-For each internal global

• If it is undef initialized
• If it is only used by loads and stores (thus no aliasing)
• If every store writes the same constant
• Replace every load with that constant, then delete the stores, then delete the global

That's a very narrow load/store propagation that only applies to globals that are initially undef, which are unusual in most programs, but common where __shared__ or pteam memory is involved.

If I'm reading some disassembled IR right, it would work for the variables used to report is_spmd and is_uninitialized. For modules that only contain one type of kernel, the overall effect would be the same as the current patch, at much reduced complexity.

Should I type that up? I think we're only looking to reduce the cost of the current/legacy devicertl with this patch. When more serious aliasing analysis comes online it would obsolete the trivial IR transform above, at which point we can delete said transform and keep the verification tests. I particularly like the path to cleanly deleting the pass when it becomes obsolete.

edit: Tried to write a test case for this, but everything I've tried so far gets stripped out by the optimiser already. i.e. the is_spmd_exec_mode calls are already removed. Can you share the test case/application where this is not the case?

[jdoerfert]

In D105246#2866459, @JonChesterfield wrote:

Thought about this a little more. I think there's a dead-simple 90% solution available.

-For each internal global

• If it is undef initialized
• If it is only used by loads and stores (thus no aliasing)
• If every store writes the same constant
• Replace every load with that constant, then delete the stores, then delete the global

And if you make it "field-sensitive" now you are almost at https://reviews.llvm.org/D104432 :)
Not sure I'd call that dead-simple but I'm with you.

That's a very narrow load/store propagation that only applies to globals that are initially undef, which are unusual in most programs, but common where __shared__ or pteam memory is involved.

If I'm reading some disassembled IR right, it would work for the variables used to report is_spmd and is_uninitialized. For modules that only contain one type of kernel, the overall effect would be the same as the current patch, at much reduced complexity.

Should I type that up? I think we're only looking to reduce the cost of the current/legacy devicertl with this patch. When more serious aliasing analysis comes online it would obsolete the trivial IR transform above, at which point we can delete said transform and keep the verification tests. I particularly like the path to cleanly deleting the pass when it becomes obsolete.

edit: Tried to write a test case for this, but everything I've tried so far gets stripped out by the optimiser already. i.e. the is_spmd_exec_mode calls are already removed. Can you share the test case/application where this is not the case?

[tianshilei1992]

Will move the change out of AAKernelInfo.

[tianshilei1992]

Move to D105787

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

struct KernelInfoState : AbstractState {
   /// The __kmpc_target_init call in this kernel, if any. If we find more than
   /// one we abort as the kernel is malformed.
   CallBase *KernelInitCB = nullptr;
 
   /// The __kmpc_target_deinit call in this kernel, if any. If we find more than
   /// one we abort as the kernel is malformed.
   CallBase *KernelDeinitCB = nullptr;
 
+  /// A map from a function to its constant return value. If the value is
+  /// nullptr, the function cannot be folded.
+  SmallDenseMap<CallBase *, Constant *> FoldableFunctions;
+
+  /// Flag to indicate that we may reach a parallel region that is not tracked
+  /// in the ParallelRegions set above.
+  SmallPtrSet<CallBase *, 4> UnknownParallelRegions;
+
+  /// Flag to indicate that we may reach a kernel entry that is not tracked.
+  bool MaybeReachedByUnknownKernel = false;
+
+  /// Flag to indicate if the associated function can be executed in SPMD mode.
+  /// Note that it could be set to false if the function is already in SPMD
+  /// mode. Use \p willBeExecutedInSPMDMode to query the status instead.
+  bool IsSPMDCompatible = true;
+
+  /// Flag to indicate if the associated function is in SPMD mode. This is
+  /// different from \p IsSPMDCompatible. This variable is only set to true if
+  /// it is indeed in a SPMD mode, not from being spmdized.
+  bool IsSPMDMode = false;
+
+  /// Flag to indicate if the associated function is a kernel entry.
+  bool IsKernelEntry = false;
+
+  /// Kernels that can reach the associated function.
+  SmallPtrSet<Kernel, 4> ReachingKernels;
+
   /// Abstract State interface
   ///{
 
   KernelInfoState() {}
   KernelInfoState(bool BestState) {
     if (!BestState)
       indicatePessimisticFixpoint();
   }
 
   /// See AbstractState::isValidState(...)
   bool isValidState() const override { return true; }
 
   /// See AbstractState::isAtFixpoint(...)
   bool isAtFixpoint() const override { return IsAtFixpoint; }
 
   /// See AbstractState::indicatePessimisticFixpoint(...)
   ChangeStatus indicatePessimisticFixpoint() override {
     IsAtFixpoint = true;
     SPMDCompatibilityTracker.indicatePessimisticFixpoint();
     ReachedUnknownParallelRegions.indicatePessimisticFixpoint();
+    IsSPMDCompatible = false;
+    UnknownParallelRegions.insert(nullptr);
+    MaybeReachedByUnknownKernel = true;
     return ChangeStatus::CHANGED;
   }
 
   /// See AbstractState::indicateOptimisticFixpoint(...)
   ChangeStatus indicateOptimisticFixpoint() override {
     IsAtFixpoint = true;
     return ChangeStatus::UNCHANGED;
   }
 
   /// Return the assumed state
   KernelInfoState &getAssumed() { return *this; }
   const KernelInfoState &getAssumed() const { return *this; }
 
   bool operator==(const KernelInfoState &RHS) const {
     if (SPMDCompatibilityTracker != RHS.SPMDCompatibilityTracker)
       return false;
     if (ReachedKnownParallelRegions != RHS.ReachedKnownParallelRegions)
       return false;
-    if (ReachedUnknownParallelRegions != RHS.ReachedUnknownParallelRegions)
+
+    if (ReachingKernels != RHS.ReachingKernels)
+      return false;
+
+    if (FoldableFunctions != RHS.FoldableFunctions)
       return false;
+
     return true;
   }
-
   /// Return empty set as the best state of potential values.
   static KernelInfoState getBestState() { return KernelInfoState(true); }
 
   static KernelInfoState getBestState(KernelInfoState &KIS) {
     return getBestState();
   }
 
   /// Return full set as the worst state of potential values.
KernelInfoState operator^=(const KernelInfoState &KIS) {
     if (KIS.KernelDeinitCB) {
       if (KernelDeinitCB && KernelDeinitCB != KIS.KernelDeinitCB)
         indicatePessimisticFixpoint();
       KernelDeinitCB = KIS.KernelDeinitCB;
     }
     SPMDCompatibilityTracker ^= KIS.SPMDCompatibilityTracker;
     ReachedKnownParallelRegions ^= KIS.ReachedKnownParallelRegions;
     ReachedUnknownParallelRegions ^= KIS.ReachedUnknownParallelRegions;
+    UnknownParallelRegions.insert(KIS.UnknownParallelRegions.begin(),
+                                  KIS.UnknownParallelRegions.end());
+    MaybeReachedByUnknownKernel |= KIS.MaybeReachedByUnknownKernel;
+    IsSPMDCompatible &= KIS.IsSPMDCompatible;
     return *this;
   }
 
   KernelInfoState operator&=(const KernelInfoState &KIS) {
     return (*this ^= KIS);
   }
 
   ///}
ConstantInt *IsSPMDArg =
         dyn_cast<ConstantInt>(KernelInitCB->getArgOperand(InitIsSPMDArgNo));
     if (IsSPMDArg && !IsSPMDArg->isZero())
       SPMDCompatibilityTracker.indicateOptimisticFixpoint();
   }
 
   /// Modify the IR based on the KernelInfoState as the fixpoint iteration is
   /// finished now.
   ChangeStatus manifest(Attributor &A) override {
-    // If we are not looking at a kernel with __kmpc_target_init and
-    // __kmpc_target_deinit call we cannot actually manifest the information.
-    if (!KernelInitCB || !KernelDeinitCB)
-      return ChangeStatus::UNCHANGED;
+    ChangeStatus Change = ChangeStatus::UNCHANGED;
 
-    // Known SPMD-mode kernels need no manifest changes.
-    if (SPMDCompatibilityTracker.isKnown())
-      return ChangeStatus::UNCHANGED;
+    // This is a somewhat unrelated modification. We basically flatten the
+    // function that was reached from a kernel completely by asking the inliner
+    // to inline everything it can. This should live in a separate AA though as
+    // it should also run on parallel regions and other GPU functions.
+    auto CheckCallInst = [&](Instruction &I) {
+      auto &CB = cast<CallBase>(I);
+      CB.addAttribute(AttributeList::FunctionIndex, Attribute::AlwaysInline);
+      return true;
+    };
+    A.checkForAllCallLikeInstructions(CheckCallInst, *this);
+
+    // Fold all valid foldable functions
+    for (std::pair<CallBase *, Constant *> &P : FoldableFunctions) {
+      if (P.second == nullptr)
+        continue;
 
-    // If we can we change the execution mode to SPMD-mode otherwise we build a
-    // custom state machine.
-    if (!changeToSPMDMode(A))
-      buildCustomStateMachine(A);
+      A.changeValueAfterManifest(*P.first, *P.second);
+      A.deleteAfterManifest(*P.first);
 
-    return ChangeStatus::CHANGED;
+      Change = ChangeStatus::CHANGED;
+    }
+
+    Change = Change | buildCustomStateMachine(A);
+
+    return Change;
   }
 
   bool changeToSPMDMode(Attributor &A) {
     if (!SPMDCompatibilityTracker.isAssumed()) {
       for (Instruction *NonCompatibleI : SPMDCompatibilityTracker) {
         if (!NonCompatibleI)
           continue;
         auto Remark = [&](OptimizationRemarkAnalysis ORA) {
ChangeStatus buildCustomStateMachine(Attributor &A) {
 
     CallInst::Create(BarrierFn, {Ident, GTid}, "", StateMachineDoneBarrierBB)
         ->setDebugLoc(DLoc);
     BranchInst::Create(StateMachineBeginBB, StateMachineDoneBarrierBB)
         ->setDebugLoc(DLoc);
 
     return ChangeStatus::CHANGED;
   }
 
+  /// Returns true if the associated function will be executed in SPMD mode, no
+  /// matter whether it is initially in SPMD mode or spmdized.
+  bool willBeExecutedInSPMDMode() const {
+    return IsSPMDMode || IsSPMDCompatible;
+  }
+
   /// Fixpoint iteration update function. Will be called every time a dependence
   /// changed its state (and in the beginning).
   ChangeStatus updateImpl(Attributor &A) override {
     KernelInfoState StateBefore = getState();
 
     // Callback to check a read/write instruction.
     auto CheckRWInst = [&](Instruction &I) {
       // We handle calls later.
       if (isa<CallBase>(I))
auto CheckRWInst = [&](Instruction &I) {
       }
       // For now we give up on everything but stores.
       SPMDCompatibilityTracker.insert(&I);
       return true;
     };
     if (!A.checkForAllReadWriteInstructions(CheckRWInst, *this))
       SPMDCompatibilityTracker.indicatePessimisticFixpoint();
 
+    updateReachingKernels(A);
+
+    // Update info regarding execution mode.
+    if (!MaybeReachedByUnknownKernel)
+      updateSPMDFolding(A);
+
     // Callback to check a call instruction.
     auto CheckCallInst = [&](Instruction &I) {
       auto &CB = cast<CallBase>(I);
       auto &CBAA = A.getAAFor<AAKernelInfo>(
           *this, IRPosition::callsite_function(CB), DepClassTy::OPTIONAL);
       if (CBAA.getState().isValidState())
         getState() ^= CBAA.getState();
+      Function *Callee = CB.getCalledFunction();
+      if (Callee) {
+        // We need to propagate information to the callee, but since the
+        // construction of AA always starts with kernel entries, we have to
+        // create AAKernelInfoFunction for all called functions. However, here
+        // the caller doesn't depend on the callee.
+        // TODO: We might want to change the dependence here later if we need
+        // information from callee to caller.
+        A.getOrCreateAAFor<AAKernelInfo>(IRPosition::function(*Callee), this,
+                                         DepClassTy::NONE);
+
+        auto &CBAA = A.getAAFor<AAKernelInfo>(
+            *this, IRPosition::callsite_function(CB), DepClassTy::REQUIRED);
+        if (CBAA.getState().isValidState()) {
+          getState() ^= CBAA.getState();
+          return true;
+        }
+      }
+
+      // The callee is not known, not ipo-amendable (e.g., due to linkage), or
+      // we can for some other reason not analyze it. If we cannot gather
+      // information, e.g., the state of the AAKernelInfo we got is invalid, we
+      // don't have to completely give up here. It basically means we have no
+      // idea what the effects of the call might be, for now the worst that can
+      // happen are unknown parallel regions hide in the callee.
+      IsSPMDCompatible = false;
+      UnknownParallelRegions.insert(&CB);
       return true;
     };
 
     if (!A.checkForAllCallLikeInstructions(CheckCallInst, *this))
       return indicatePessimisticFixpoint();
 
     return StateBefore == getState() ? ChangeStatus::UNCHANGED
                                      : ChangeStatus::CHANGED;
   }
+
+private:
+  /// Do some setup if the associated function is a kernel entry.
+  void setupForKernelEntry(Attributor &A, Function *F) {
+    // Add itself to the reaching kernel and set IsKernelEntry.
+    ReachingKernels.insert(F);
+    IsKernelEntry = true;
+
+    // Check the callsite of function __kmpc_target_init to get the argument of
+    // SPMD mode and set IsSPMDMode accordingly.
+    CallBase *CB = nullptr;
+
+    auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache());
+    OMPInformationCache::RuntimeFunctionInfo &TargetInitRFI =
+        OMPInfoCache.RFIs[OMPRTL___kmpc_target_init];
+
+    TargetInitRFI.foreachUse(
+        [&](Use &U, Function &Caller) {
+          assert(CB == nullptr && "__kmpc_target_init is called more than "
+                                  "once in one kernel entry");
+
+          assert(dyn_cast<CallBase>(U.getUser()) &&
+                 "Use of __kmpc_target_init is not a CallBase");
+
+          CB = cast<CallBase>(U.getUser());
+
+          return true;
+        },
+        F);
+
+    assert(CB && "Call to __kmpc_target_init is missing");
+
+    constexpr const int InitIsSPMDArgNo = 1;
+
+    ConstantInt *IsSPMDArg =
+        dyn_cast<ConstantInt>(CB->getArgOperand(InitIsSPMDArgNo));
+
+    if (IsSPMDArg && !IsSPMDArg->isZero())
+      IsSPMDMode = true;
+  }
+
+  /// Update info regarding reaching kernels.
+  void updateReachingKernels(Attributor &A) {
+    if (!IsKernelEntry) {
+      auto PredCallSite = [&](AbstractCallSite ACS) {
+        Function *Caller = ACS.getInstruction()->getFunction();
+
+        assert(Caller && "Caller is nullptr");
+
+        auto &CAA =
+            A.getOrCreateAAFor<AAKernelInfo>(IRPosition::function(*Caller));
+        if (CAA.isValidState()) {
+          ReachingKernels.insert(CAA.ReachingKernels.begin(),
+                                 CAA.ReachingKernels.end());
+          return true;
+        }
+
+        // We lost track of the caller of the associated function, any kernel
+        // could reach now.
+        MaybeReachedByUnknownKernel = true;
+
+        return true;
+      };
+
+      bool AllCallSitesKnown;
+      if (!A.checkForAllCallSites(PredCallSite, *this,
+                                  true /* RequireAllCallSites */,
+                                  AllCallSitesKnown))
+        MaybeReachedByUnknownKernel = true;
+    }
+  }
+
+  /// Update information regarding folding SPMD mode function calls.
+  void updateSPMDFolding(Attributor &A) {
+    unsigned Count = 0;
+
+    for (Kernel K : ReachingKernels) {
+      auto &AA = A.getAAFor<AAKernelInfo>(*this, IRPosition::function(*K),
+                                          DepClassTy::REQUIRED);
+      assert(AA.isValidState() && "AA should be valid here");
+      if (AA.IsSPMDMode)
+        ++Count;
+    }
+
+    // Assume reaching kernels are in a mixture of SPMD and non-SPMD mode.
+    // Update all function calls to __kmpc_is_spmd_exec_mode to nullptr.
+    Constant *C = nullptr;
+
+    auto &Ctx = getAnchorValue().getContext();
+
+    if (Count == 0) {
+      // All reaching kernels are in non-SPMD mode. Update all function
+      // calls to __kmpc_is_spmd_exec_mode to 0.
+      C = ConstantInt::get(Type::getInt8Ty(Ctx), 0);
+    } else if (Count == ReachingKernels.size()) {
+      // All reaching kernels are in SPMD mode. Update all function calls to
+      // __kmpc_is_spmd_exec_mode to 1.
+      C = ConstantInt::get(Type::getInt8Ty(Ctx), 1);
+    }
+
+    auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache());
+    OMPInformationCache::RuntimeFunctionInfo &IsSPMDExecModeRFI =
+        OMPInfoCache.RFIs[OMPRTL___kmpc_is_spmd_exec_mode];
+
+    for (std::pair<CallBase *, Constant *> &P : FoldableFunctions) {
+      CallBase *CB = P.first;
+      if (CB->getCalledFunction() == IsSPMDExecModeRFI.Declaration)
+        P.second = C;
+    }
+  }
 };
 
 /// The call site kernel info abstract attribute, basically, what can we say
 /// about a call site with regards to the KernelInfoState. For now this simply
 /// forwards the information from the callee.
 struct AAKernelInfoCallSite : AAKernelInfo {
   AAKernelInfoCallSite(const IRPosition &IRP, Attributor &A)
       : AAKernelInfo(IRP, A) {}