Merged main:186a4b3b6578 into amd-gfx:7ad30b56268e

Local branch amd-gfx 7ad30b5 Merged main:3f0bddb56ac3 into amd-gfx:32727405140d
Remote branch main 186a4b3 [LLVM][OpenMP] Allow OpenMPOpt to handle non-OpenMP target regions (llvm#67075)

Author: SC llvm team

llvm/include/llvm/Config/llvm-config.h.cmake

@@ -16,7 +16,7 @@
 
 /* Indicate that this is LLVM compiled from the amd-gfx branch. */
 #define LLVM_HAVE_BRANCH_AMD_GFX
-#define LLVM_MAIN_REVISION 475744
+#define LLVM_MAIN_REVISION 475745
 
 /* Define if LLVM_ENABLE_DUMP is enabled */
 #cmakedefine LLVM_ENABLE_DUMP

llvm/include/llvm/Transforms/IPO/OpenMPOpt.h

@@ -29,8 +29,9 @@ bool containsOpenMP(Module &M);
 /// Helper to determine if \p M is a OpenMP target offloading device module.
 bool isOpenMPDevice(Module &M);
 
-/// Return true iff \p Fn is a GPU kernel; \p Fn has the "kernel" attribute.
-bool isKernel(Function &Fn);
+/// Return true iff \p Fn is an OpenMP GPU kernel; \p Fn has the "kernel"
+/// attribute.
+bool isOpenMPKernel(Function &Fn);
 
 /// Get OpenMP device kernels in \p M.
 KernelSet getDeviceKernels(Module &M);

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

@@ -158,6 +158,8 @@ STATISTIC(NumOpenMPRuntimeFunctionUsesIdentified,
           "Number of OpenMP runtime function uses identified");
 STATISTIC(NumOpenMPTargetRegionKernels,
           "Number of OpenMP target region entry points (=kernels) identified");
+STATISTIC(NumNonOpenMPTargetRegionKernels,
+          "Number of non-OpenMP target region kernels identified");
 STATISTIC(NumOpenMPTargetRegionKernelsSPMD,
           "Number of OpenMP target region entry points (=kernels) executed in "
           "SPMD-mode instead of generic-mode");
@@ -989,7 +991,7 @@ struct OpenMPOpt {
   /// Print OpenMP GPU kernels for testing.
   void printKernels() const {
     for (Function *F : SCC) {
-      if (!omp::isKernel(*F))
+      if (!omp::isOpenMPKernel(*F))
         continue;
 
       auto Remark = [&](OptimizationRemarkAnalysis ORA) {
@@ -2030,7 +2032,7 @@ Kernel OpenMPOpt::getUniqueKernelFor(Function &F) {
     // TODO: We should use an AA to create an (optimistic and callback
     //       call-aware) call graph. For now we stick to simple patterns that
     //       are less powerful, basically the worst fixpoint.
-    if (isKernel(F)) {
+    if (isOpenMPKernel(F)) {
       CachedKernel = Kernel(&F);
       return *CachedKernel;
     }
@@ -2721,7 +2723,7 @@ struct AAExecutionDomainFunction : public AAExecutionDomain {
       HandleAlignedBarrier(CB);
 
     // Handle the "kernel end barrier" for kernels too.
-    if (omp::isKernel(*getAnchorScope()))
+    if (omp::isOpenMPKernel(*getAnchorScope()))
       HandleAlignedBarrier(nullptr);
 
     return Changed;
@@ -2974,7 +2976,7 @@ bool AAExecutionDomainFunction::handleCallees(Attributor &A,
   } else {
     // We could not find all predecessors, so this is either a kernel or a
     // function with external linkage (or with some other weird uses).
-    if (omp::isKernel(*getAnchorScope())) {
+    if (omp::isOpenMPKernel(*getAnchorScope())) {
       EntryBBED.IsExecutedByInitialThreadOnly = false;
       EntryBBED.IsReachedFromAlignedBarrierOnly = true;
       EntryBBED.EncounteredNonLocalSideEffect = false;
@@ -3028,7 +3030,7 @@ ChangeStatus AAExecutionDomainFunction::updateImpl(Attributor &A) {
 
   Function *F = getAnchorScope();
   BasicBlock &EntryBB = F->getEntryBlock();
-  bool IsKernel = omp::isKernel(*F);
+  bool IsKernel = omp::isOpenMPKernel(*F);
 
   SmallVector<Instruction *> SyncInstWorklist;
   for (auto &RIt : *RPOT) {
@@ -4167,7 +4169,7 @@ struct AAKernelInfoFunction : AAKernelInfo {
       auto *CB = cast<CallBase>(Kernel->user_back());
       Kernel = CB->getCaller();
     }
-    assert(omp::isKernel(*Kernel) && "Expected kernel function!");
+    assert(omp::isOpenMPKernel(*Kernel) && "Expected kernel function!");
 
     // Check if the kernel is already in SPMD mode, if so, return success.
     ConstantStruct *ExistingKernelEnvC =
@@ -5804,7 +5806,9 @@ PreservedAnalyses OpenMPOptCGSCCPass::run(LazyCallGraph::SCC &C,
   return PreservedAnalyses::all();
 }
 
-bool llvm::omp::isKernel(Function &Fn) { return Fn.hasFnAttribute("kernel"); }
+bool llvm::omp::isOpenMPKernel(Function &Fn) {
+  return Fn.hasFnAttribute("kernel");
+}
 
 KernelSet llvm::omp::getDeviceKernels(Module &M) {
   // TODO: Create a more cross-platform way of determining device kernels.
@@ -5826,10 +5830,13 @@ KernelSet llvm::omp::getDeviceKernels(Module &M) {
     if (!KernelFn)
       continue;
 
-    assert(isKernel(*KernelFn) && "Inconsistent kernel function annotation");
-    ++NumOpenMPTargetRegionKernels;
-
-    Kernels.insert(KernelFn);
+    // We are only interested in OpenMP target regions. Others, such as kernels
+    // generated by CUDA but linked together, are not interesting to this pass.
+    if (isOpenMPKernel(*KernelFn)) {
+      ++NumOpenMPTargetRegionKernels;
+      Kernels.insert(KernelFn);
+    } else
+      ++NumNonOpenMPTargetRegionKernels;
   }
 
   return Kernels;

llvm/test/Transforms/OpenMP/bug66687.ll

@@ -0,0 +1,29 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-attributes --version 3
+; RUN: opt < %s -S -passes=openmp-opt | FileCheck %s
+
+source_filename = "bug66687.ll"
+target datalayout = "e-i64:64-i128:128-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64-nvidia-cuda"
+
+define weak void @openmp_kernel() "kernel" {
+; CHECK-LABEL: define weak void @openmp_kernel(
+; CHECK-SAME: ) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:    ret void
+;
+  ret void
+}
+
+define weak_odr void @non_openmp_kernel() {
+; CHECK-LABEL: define weak_odr void @non_openmp_kernel() {
+; CHECK-NEXT:    ret void
+;
+  ret void
+}
+
+!llvm.module.flags = !{!0, !1}
+!nvvm.annotations = !{!2, !3}
+
+!0 = !{i32 7, !"openmp", i32 51}
+!1 = !{i32 7, !"openmp-device", i32 51}
+!2 = !{ptr @openmp_kernel, !"kernel", i32 1}
+!3 = !{ptr @non_openmp_kernel, !"kernel", i32 1}