[OpenMP] Add the `ompx_attribute` clause for target directives

CUDA and HIP have kernel attributes to tune the code generation (in the
backend). To reuse this functionality for OpenMP target regions we
introduce the `ompx_attribute` clause that takes these kernel
attributes and emits code as if they had been attached to the kernel
fuction (which is implicitly generated).

To limit the impact, we only support three kernel attributes:
`amdgpu_waves_per_eu`, for AMDGPU
`amdgpu_flat_work_group_size`, for AMDGPU
`launch_bounds`, for NVPTX

The existing implementations of those attributes are used for error
checking and code generation. `ompx_attribute` can be attached to any
executable target region and it can hold more than one kernel attribute.

Differential Revision: https://reviews.llvm.org/D156184

1 files changed, 59 insertions, 0 deletions

diff --git a/clang/lib/Parse/ParseOpenMP.cpp b/clang/lib/Parse/ParseOpenMP.cpp
index 96d2e2c..66cabb1 100644
--- a/clang/lib/Parse/ParseOpenMP.cpp
+++ b/clang/lib/Parse/ParseOpenMP.cpp
@@ -3411,6 +3411,9 @@ OMPClause *Parser::ParseOpenMPClause(OpenMPDirectiveKind DKind,
           << getOpenMPClauseName(CKind) << getOpenMPDirectiveName(DKind);
     SkipUntil(tok::comma, tok::annot_pragma_openmp_end, StopBeforeMatch);
     break;
+  case OMPC_ompx_attribute:
+    Clause = ParseOpenMPOMPXAttributesClause(WrongDirective);
+    break;
   default:
     break;
   }
@@ -3691,6 +3694,62 @@ OMPClause *Parser::ParseOpenMPInteropClause(OpenMPClauseKind Kind,
   llvm_unreachable("Unexpected interop variable clause.");
 }
 
+OMPClause *Parser::ParseOpenMPOMPXAttributesClause(bool ParseOnly) {
+  SourceLocation Loc = ConsumeToken();
+  // Parse '('.
+  BalancedDelimiterTracker T(*this, tok::l_paren, tok::annot_pragma_openmp_end);
+  if (T.expectAndConsume(diag::err_expected_lparen_after,
+                         getOpenMPClauseName(OMPC_ompx_attribute).data()))
+    return nullptr;
+
+  ParsedAttributes ParsedAttrs(AttrFactory);
+  ParseAttributes(PAKM_GNU | PAKM_CXX11, ParsedAttrs);
+
+  // Parse ')'.
+  if (T.consumeClose())
+    return nullptr;
+
+  if (ParseOnly)
+    return nullptr;
+
+  SmallVector<Attr *> Attrs;
+  for (const ParsedAttr &PA : ParsedAttrs) {
+    switch (PA.getKind()) {
+    case ParsedAttr::AT_AMDGPUFlatWorkGroupSize:
+      if (!PA.checkExactlyNumArgs(Actions, 2))
+        continue;
+      if (auto *A = Actions.CreateAMDGPUFlatWorkGroupSizeAttr(
+              PA, PA.getArgAsExpr(0), PA.getArgAsExpr(1)))
+        Attrs.push_back(A);
+      continue;
+    case ParsedAttr::AT_AMDGPUWavesPerEU:
+      if (!PA.checkAtLeastNumArgs(Actions, 1) ||
+          !PA.checkAtMostNumArgs(Actions, 2))
+        continue;
+      if (auto *A = Actions.CreateAMDGPUWavesPerEUAttr(
+              PA, PA.getArgAsExpr(0),
+              PA.getNumArgs() > 1 ? PA.getArgAsExpr(1) : nullptr))
+        Attrs.push_back(A);
+      continue;
+    case ParsedAttr::AT_CUDALaunchBounds:
+      if (!PA.checkAtLeastNumArgs(Actions, 1) ||
+          !PA.checkAtMostNumArgs(Actions, 2))
+        continue;
+      if (auto *A = Actions.CreateLaunchBoundsAttr(
+              PA, PA.getArgAsExpr(0),
+              PA.getNumArgs() > 1 ? PA.getArgAsExpr(1) : nullptr))
+        Attrs.push_back(A);
+      continue;
+    default:
+      Diag(Loc, diag::warn_omp_invalid_attribute_for_ompx_attributes) << PA;
+      continue;
+    };
+  }
+
+  return Actions.ActOnOpenMPXAttributeClause(Attrs, Loc, T.getOpenLocation(),
+                                             T.getCloseLocation());
+}
+
 /// Parsing of simple OpenMP clauses like 'default' or 'proc_bind'.
 ///
 ///    default-clause: