Update TargetRegionFlags to mirror OMPTgtExecModeFlags

Author: skatrak

llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp

@@ -6775,7 +6775,8 @@ OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createTargetInit(
   Constant *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
   Constant *IsSPMDVal = ConstantInt::getSigned(Int8, Attrs.ExecFlags);
   Constant *UseGenericStateMachineVal = ConstantInt::getSigned(
-      Int8, Attrs.ExecFlags != omp::OMP_TGT_EXEC_MODE_SPMD);
+      Int8, Attrs.ExecFlags != omp::OMP_TGT_EXEC_MODE_SPMD &&
+                Attrs.ExecFlags != omp::OMP_TGT_EXEC_MODE_SPMD_NO_LOOP);
   Constant *MayUseNestedParallelismVal = ConstantInt::getSigned(Int8, true);
   Constant *DebugIndentionLevelVal = ConstantInt::getSigned(Int16, 0);
 

mlir/include/mlir/Dialect/OpenMP/OpenMPEnums.td

@@ -223,28 +223,21 @@ def ScheduleModifier : OpenMP_I32EnumAttr<
 def ScheduleModifierAttr : OpenMP_EnumAttr<ScheduleModifier, "sched_mod">;
 
 //===----------------------------------------------------------------------===//
-// target_region_flags enum.
+// target_exec_mode enum.
 //===----------------------------------------------------------------------===//
 
-def TargetRegionFlagsNone : I32BitEnumAttrCaseNone<"none">;
-def TargetRegionFlagsSpmd : I32BitEnumAttrCaseBit<"spmd", 0>;
-def TargetRegionFlagsTripCount : I32BitEnumAttrCaseBit<"trip_count", 1>;
-def TargetRegionFlagsNoLoop : I32BitEnumAttrCaseBit<"no_loop", 2>;
-
-def TargetRegionFlags : OpenMP_BitEnumAttr<
-    "TargetRegionFlags",
-    "These flags describe properties of the target kernel. "
-    "TargetRegionFlagsSpmd - denotes SPMD kernel. "
-    "TargetRegionFlagsNoLoop - denotes kernel where "
-    "num_teams * num_threads >= loop_trip_count. It allows the conversion "
-    "of loops into sequential code by ensuring that each team/thread "
-    "executes at most one iteration. "
-    "TargetRegionFlagsTripCount - checks if a singular loop trip count should "
-    "be calculated for the target region.", [
-      TargetRegionFlagsNone,
-      TargetRegionFlagsSpmd,
-      TargetRegionFlagsTripCount,
-      TargetRegionFlagsNoLoop
+def TargetExecModeBare : I32EnumAttrCase<"bare", 0>;
+def TargetExecModeGeneric : I32EnumAttrCase<"generic", 1>;
+def TargetExecModeSpmd : I32EnumAttrCase<"spmd", 2>;
+def TargetExecModeSpmdNoLoop : I32EnumAttrCase<"no_loop", 3>;
+
+def TargetExecMode : OpenMP_I32EnumAttr<
+    "TargetExecMode",
+    "target execution mode, mirroring the `OMPTgtExecModeFlags` LLVM enum", [
+      TargetExecModeBare,
+      TargetExecModeGeneric,
+      TargetExecModeSpmd,
+      TargetExecModeSpmdNoLoop,
     ]>;
 
 //===----------------------------------------------------------------------===//

mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td

@@ -1522,13 +1522,17 @@ def TargetOp : OpenMP_Op<"target", traits = [
     /// operations, the top level one will be the one captured.
     Operation *getInnermostCapturedOmpOp();
 
-    /// Infers the kernel type (Generic, SPMD or Generic-SPMD) based on the
-    /// contents of the target region.
+    /// Infers the kernel type (Bare, Generic or SPMD) based on the contents of
+    /// the target region.
     ///
     /// \param capturedOp result of a still valid (no modifications made to any
     /// nested operations) previous call to `getInnermostCapturedOmpOp()`.
-    static ::mlir::omp::TargetRegionFlags
-    getKernelExecFlags(Operation *capturedOp);
+    /// \param hostEvalTripCount output argument to store whether this kernel
+    /// wraps a loop whose bounds must be evaluated on the host prior to
+    /// launching it.
+    static ::mlir::omp::TargetExecMode
+    getKernelExecFlags(Operation *capturedOp,
+                       bool *hostEvalTripCount = nullptr);
   }] # clausesExtraClassDeclaration;
 
   let assemblyFormat = clausesAssemblyFormat # [{

mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp

@@ -2205,8 +2205,9 @@ LogicalResult TargetOp::verifyRegions() {
     return emitError("target containing multiple 'omp.teams' nested ops");
 
   // Check that host_eval values are only used in legal ways.
+  bool hostEvalTripCount;
   Operation *capturedOp = getInnermostCapturedOmpOp();
-  TargetRegionFlags execFlags = getKernelExecFlags(capturedOp);
+  TargetExecMode execMode = getKernelExecFlags(capturedOp, &hostEvalTripCount);
   for (Value hostEvalArg :
        cast<BlockArgOpenMPOpInterface>(getOperation()).getHostEvalBlockArgs()) {
     for (Operation *user : hostEvalArg.getUsers()) {
@@ -2221,7 +2222,7 @@ LogicalResult TargetOp::verifyRegions() {
                                 "and 'thread_limit' in 'omp.teams'";
       }
       if (auto parallelOp = dyn_cast<ParallelOp>(user)) {
-        if (bitEnumContainsAny(execFlags, TargetRegionFlags::spmd) &&
+        if (execMode == TargetExecMode::spmd &&
             parallelOp->isAncestor(capturedOp) &&
             hostEvalArg == parallelOp.getNumThreads())
           continue;
@@ -2231,8 +2232,7 @@ LogicalResult TargetOp::verifyRegions() {
                   "'omp.parallel' when representing target SPMD";
       }
       if (auto loopNestOp = dyn_cast<LoopNestOp>(user)) {
-        if (bitEnumContainsAny(execFlags, TargetRegionFlags::trip_count) &&
-            loopNestOp.getOperation() == capturedOp &&
+        if (hostEvalTripCount && loopNestOp.getOperation() == capturedOp &&
             (llvm::is_contained(loopNestOp.getLoopLowerBounds(), hostEvalArg) ||
              llvm::is_contained(loopNestOp.getLoopUpperBounds(), hostEvalArg) ||
              llvm::is_contained(loopNestOp.getLoopSteps(), hostEvalArg)))
@@ -2362,7 +2362,9 @@ static bool canPromoteToNoLoop(Operation *capturedOp, TeamsOp teamsOp,
          ompFlags.getAssumeThreadsOversubscription();
 }
 
-TargetRegionFlags TargetOp::getKernelExecFlags(Operation *capturedOp) {
+TargetExecMode TargetOp::getKernelExecFlags(Operation *capturedOp,
+                                            bool *hostEvalTripCount) {
+  // TODO: Support detection of bare kernel mode.
   // A non-null captured op is only valid if it resides inside of a TargetOp
   // and is the result of calling getInnermostCapturedOmpOp() on it.
   TargetOp targetOp =
@@ -2371,9 +2373,12 @@ TargetRegionFlags TargetOp::getKernelExecFlags(Operation *capturedOp) {
           (targetOp && targetOp.getInnermostCapturedOmpOp() == capturedOp)) &&
          "unexpected captured op");
 
+  if (hostEvalTripCount)
+    *hostEvalTripCount = false;
+
   // If it's not capturing a loop, it's a default target region.
   if (!isa_and_present<LoopNestOp>(capturedOp))
-    return TargetRegionFlags::none;
+    return TargetExecMode::generic;
 
   // Get the innermost non-simd loop wrapper.
   SmallVector<LoopWrapperInterface> loopWrappers;
@@ -2386,59 +2391,63 @@ TargetRegionFlags TargetOp::getKernelExecFlags(Operation *capturedOp) {
 
   auto numWrappers = std::distance(innermostWrapper, loopWrappers.end());
   if (numWrappers != 1 && numWrappers != 2)
-    return TargetRegionFlags::none;
+    return TargetExecMode::generic;
 
   // Detect target-teams-distribute-parallel-wsloop[-simd].
   if (numWrappers == 2) {
     WsloopOp *wsloopOp = dyn_cast<WsloopOp>(innermostWrapper);
     if (!wsloopOp)
-      return TargetRegionFlags::none;
+      return TargetExecMode::generic;
 
     innermostWrapper = std::next(innermostWrapper);
     if (!isa<DistributeOp>(innermostWrapper))
-      return TargetRegionFlags::none;
+      return TargetExecMode::generic;
 
     Operation *parallelOp = (*innermostWrapper)->getParentOp();
     if (!isa_and_present<ParallelOp>(parallelOp))
-      return TargetRegionFlags::none;
+      return TargetExecMode::generic;
 
     TeamsOp teamsOp = dyn_cast<TeamsOp>(parallelOp->getParentOp());
     if (!teamsOp)
-      return TargetRegionFlags::none;
+      return TargetExecMode::generic;
 
     if (teamsOp->getParentOp() == targetOp.getOperation()) {
-      TargetRegionFlags result =
-          TargetRegionFlags::spmd | TargetRegionFlags::trip_count;
+      TargetExecMode result = TargetExecMode::spmd;
       if (canPromoteToNoLoop(capturedOp, teamsOp, wsloopOp))
-        result = result | TargetRegionFlags::no_loop;
+        result = TargetExecMode::no_loop;
+      if (hostEvalTripCount)
+        *hostEvalTripCount = true;
       return result;
     }
   }
   // Detect target-teams-distribute[-simd] and target-teams-loop.
   else if (isa<DistributeOp, LoopOp>(innermostWrapper)) {
     Operation *teamsOp = (*innermostWrapper)->getParentOp();
     if (!isa_and_present<TeamsOp>(teamsOp))
-      return TargetRegionFlags::none;
+      return TargetExecMode::generic;
 
     if (teamsOp->getParentOp() != targetOp.getOperation())
-      return TargetRegionFlags::none;
+      return TargetExecMode::generic;
+
+    if (hostEvalTripCount)
+      *hostEvalTripCount = true;
 
     if (isa<LoopOp>(innermostWrapper))
-      return TargetRegionFlags::spmd | TargetRegionFlags::trip_count;
+      return TargetExecMode::spmd;
 
-    return TargetRegionFlags::trip_count;
+    return TargetExecMode::generic;
   }
   // Detect target-parallel-wsloop[-simd].
   else if (isa<WsloopOp>(innermostWrapper)) {
     Operation *parallelOp = (*innermostWrapper)->getParentOp();
     if (!isa_and_present<ParallelOp>(parallelOp))
-      return TargetRegionFlags::none;
+      return TargetExecMode::generic;
 
     if (parallelOp->getParentOp() == targetOp.getOperation())
-      return TargetRegionFlags::spmd;
+      return TargetExecMode::spmd;
   }
 
-  return TargetRegionFlags::none;
+  return TargetExecMode::generic;
 }
 
 //===----------------------------------------------------------------------===//

mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp

@@ -2601,11 +2601,8 @@ convertOmpWsloop(Operation &opInst, llvm::IRBuilderBase &builder,
     // for every omp.wsloop nested inside a no-loop SPMD target region, even if
     // that loop is not the top-level SPMD one.
     if (loopOp == targetCapturedOp) {
-      omp::TargetRegionFlags kernelFlags =
-          targetOp.getKernelExecFlags(targetCapturedOp);
-      if (omp::bitEnumContainsAll(kernelFlags,
-                                  omp::TargetRegionFlags::spmd |
-                                      omp::TargetRegionFlags::no_loop))
+      if (targetOp.getKernelExecFlags(targetCapturedOp) ==
+          omp::TargetExecMode::no_loop)
         noLoopMode = true;
     }
   }
@@ -5435,14 +5432,21 @@ initTargetDefaultAttrs(omp::TargetOp targetOp, Operation *capturedOp,
   }
 
   // Update kernel bounds structure for the `OpenMPIRBuilder` to use.
-  omp::TargetRegionFlags kernelFlags = targetOp.getKernelExecFlags(capturedOp);
-  attrs.ExecFlags =
-      omp::bitEnumContainsAny(kernelFlags, omp::TargetRegionFlags::spmd)
-          ? omp::bitEnumContainsAny(kernelFlags, omp::TargetRegionFlags::no_loop)
-                ? llvm::omp::OMP_TGT_EXEC_MODE_SPMD_NO_LOOP
-                : llvm::omp::OMP_TGT_EXEC_MODE_SPMD
-          : llvm::omp::OMP_TGT_EXEC_MODE_GENERIC;
-
+  omp::TargetExecMode execMode = targetOp.getKernelExecFlags(capturedOp);
+  switch (execMode) {
+  case omp::TargetExecMode::bare:
+    attrs.ExecFlags = llvm::omp::OMP_TGT_EXEC_MODE_BARE;
+    break;
+  case omp::TargetExecMode::generic:
+    attrs.ExecFlags = llvm::omp::OMP_TGT_EXEC_MODE_GENERIC;
+    break;
+  case omp::TargetExecMode::spmd:
+    attrs.ExecFlags = llvm::omp::OMP_TGT_EXEC_MODE_SPMD;
+    break;
+  case omp::TargetExecMode::no_loop:
+    attrs.ExecFlags = llvm::omp::OMP_TGT_EXEC_MODE_SPMD_NO_LOOP;
+    break;
+  }
   attrs.MinTeams = minTeamsVal;
   attrs.MaxTeams.front() = maxTeamsVal;
   attrs.MinThreads = 1;
@@ -5492,8 +5496,9 @@ initTargetRuntimeAttrs(llvm::IRBuilderBase &builder,
   if (numThreads)
     attrs.MaxThreads = moduleTranslation.lookupValue(numThreads);
 
-  if (omp::bitEnumContainsAny(targetOp.getKernelExecFlags(capturedOp),
-                              omp::TargetRegionFlags::trip_count)) {
+  bool hostEvalTripCount;
+  targetOp.getKernelExecFlags(capturedOp, &hostEvalTripCount);
+  if (hostEvalTripCount) {
     llvm::OpenMPIRBuilder *ompBuilder = moduleTranslation.getOpenMPBuilder();
     attrs.LoopTripCount = nullptr;