merge main into amd-staging (#346)

Author: ronlieb

clang/docs/AllocToken.rst

@@ -37,8 +37,8 @@ The default mode to calculate tokens is:
   pointers.
 
 Other token ID assignment modes are supported, but they may be subject to
-change or removal. These may (experimentally) be selected with ``-mllvm
--alloc-token-mode=<mode>``:
+change or removal. These may (experimentally) be selected with ``-Xclang
+-falloc-token-mode=<mode>``:
 
 * ``typehash``: This mode assigns a token ID based on the hash of the allocated
   type's name.

clang/include/clang/AST/StmtOpenACC.h

@@ -815,6 +815,17 @@ class OpenACCAtomicConstruct final
   Stmt *getAssociatedStmt() {
     return OpenACCAssociatedStmtConstruct::getAssociatedStmt();
   }
+
+  // A struct to represent a broken-down version of the associated statement,
+  // providing the information specified in OpenACC3.3 Section 2.12.
+  struct StmtInfo {
+    const Expr *V;
+    const Expr *X;
+    // TODO: OpenACC: We should expand this as we're implementing the other
+    // atomic construct kinds.
+  };
+
+  const StmtInfo getAssociatedStmtInfo() const;
 };
 
 } // namespace clang

clang/include/clang/Basic/CodeGenOptions.h

@@ -455,10 +455,6 @@ class CodeGenOptions : public CodeGenOptionsBase {
 
   std::optional<double> AllowRuntimeCheckSkipHotCutoff;
 
-  /// Maximum number of allocation tokens (0 = no max), nullopt if none set (use
-  /// pass default).
-  std::optional<uint64_t> AllocTokenMax;
-
   /// List of backend command-line options for -fembed-bitcode.
   std::vector<uint8_t> CmdArgs;
 

clang/include/clang/Basic/LangOptions.h

@@ -25,6 +25,7 @@
 #include "llvm/ADT/FloatingPointMode.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/BinaryFormat/DXContainer.h"
+#include "llvm/Support/AllocToken.h"
 #include "llvm/TargetParser/Triple.h"
 #include <optional>
 #include <string>
@@ -565,6 +566,13 @@ class LangOptions : public LangOptionsBase {
   bool AtomicFineGrainedMemory = false;
   bool AtomicIgnoreDenormalMode = false;
 
+  /// Maximum number of allocation tokens (0 = no max), nullopt if none set (use
+  /// target default).
+  std::optional<uint64_t> AllocTokenMax;
+
+  /// The allocation token mode.
+  std::optional<llvm::AllocTokenMode> AllocTokenMode;
+
   LangOptions();
 
   /// Set language defaults for the given input language and

clang/include/clang/Driver/Options.td

@@ -2760,6 +2760,10 @@ def falloc_token_max_EQ : Joined<["-"], "falloc-token-max=">,
   MetaVarName<"<N>">,
   HelpText<"Limit to maximum N allocation tokens (0 = no max)">;
 
+def falloc_token_mode_EQ : Joined<["-"], "falloc-token-mode=">,
+  Group<f_Group>, Visibility<[CC1Option]>,
+  HelpText<"Set the allocation token mode (experimental)">;
+
 def fallow_runtime_check_skip_hot_cutoff_EQ
     : Joined<["-"], "fallow-runtime-check-skip-hot-cutoff=">,
       Group<f_clang_Group>,

clang/lib/AST/StmtOpenACC.cpp

@@ -12,7 +12,9 @@
 
 #include "clang/AST/StmtOpenACC.h"
 #include "clang/AST/ASTContext.h"
+#include "clang/AST/ExprCXX.h"
 #include "clang/AST/StmtCXX.h"
+
 using namespace clang;
 
 OpenACCComputeConstruct *
@@ -322,6 +324,38 @@ OpenACCAtomicConstruct *OpenACCAtomicConstruct::Create(
   return Inst;
 }
 
+const OpenACCAtomicConstruct::StmtInfo
+OpenACCAtomicConstruct::getAssociatedStmtInfo() const {
+  // This ends up being a vastly simplified version of SemaOpenACCAtomic, since
+  // it doesn't have to worry about erroring out, but we should do a lot of
+  // asserts to ensure we don't get off into the weeds.
+  assert(getAssociatedStmt() && "invalid associated stmt?");
+
+  switch (AtomicKind) {
+  case OpenACCAtomicKind::None:
+  case OpenACCAtomicKind::Write:
+  case OpenACCAtomicKind::Update:
+  case OpenACCAtomicKind::Capture:
+    assert(false && "Only 'read' has been implemented here");
+    return {};
+  case OpenACCAtomicKind::Read: {
+    // Read only supports the format 'v = x'; where both sides are a scalar
+    // expression. This can come in 2 forms; BinaryOperator or
+    // CXXOperatorCallExpr (rarely).
+    const Expr *AssignExpr = cast<const Expr>(getAssociatedStmt());
+    if (const auto *BO = dyn_cast<BinaryOperator>(AssignExpr)) {
+      assert(BO->getOpcode() == BO_Assign);
+      return {BO->getLHS()->IgnoreImpCasts(), BO->getRHS()->IgnoreImpCasts()};
+    }
+
+    const auto *OO = cast<CXXOperatorCallExpr>(AssignExpr);
+    assert(OO->getOperator() == OO_Equal);
+
+    return {OO->getArg(0)->IgnoreImpCasts(), OO->getArg(1)->IgnoreImpCasts()};
+  }
+  }
+}
+
 OpenACCCacheConstruct *OpenACCCacheConstruct::CreateEmpty(const ASTContext &C,
                                                           unsigned NumVars) {
   void *Mem =

clang/lib/CIR/CodeGen/CIRGenOpenACCClause.cpp

@@ -553,12 +553,15 @@ class OpenACCClauseCIREmitter final
   }
 
   void VisitIfClause(const OpenACCIfClause &clause) {
-    if constexpr (isOneOfTypes<OpTy, mlir::acc::ParallelOp, mlir::acc::SerialOp,
-                               mlir::acc::KernelsOp, mlir::acc::InitOp,
-                               mlir::acc::ShutdownOp, mlir::acc::SetOp,
-                               mlir::acc::DataOp, mlir::acc::WaitOp,
-                               mlir::acc::HostDataOp, mlir::acc::EnterDataOp,
-                               mlir::acc::ExitDataOp, mlir::acc::UpdateOp>) {
+    if constexpr (isOneOfTypes<
+                      OpTy, mlir::acc::ParallelOp, mlir::acc::SerialOp,
+                      mlir::acc::KernelsOp, mlir::acc::InitOp,
+                      mlir::acc::ShutdownOp, mlir::acc::SetOp,
+                      mlir::acc::DataOp, mlir::acc::WaitOp,
+                      mlir::acc::HostDataOp, mlir::acc::EnterDataOp,
+                      mlir::acc::ExitDataOp, mlir::acc::UpdateOp,
+                      mlir::acc::AtomicReadOp, mlir::acc::AtomicWriteOp,
+                      mlir::acc::AtomicUpdateOp, mlir::acc::AtomicCaptureOp>) {
       operation.getIfCondMutable().append(
           createCondition(clause.getConditionExpr()));
     } else if constexpr (isCombinedType<OpTy>) {
@@ -1144,6 +1147,10 @@ EXPL_SPEC(mlir::acc::HostDataOp)
 EXPL_SPEC(mlir::acc::EnterDataOp)
 EXPL_SPEC(mlir::acc::ExitDataOp)
 EXPL_SPEC(mlir::acc::UpdateOp)
+EXPL_SPEC(mlir::acc::AtomicReadOp)
+EXPL_SPEC(mlir::acc::AtomicWriteOp)
+EXPL_SPEC(mlir::acc::AtomicCaptureOp)
+EXPL_SPEC(mlir::acc::AtomicUpdateOp)
 #undef EXPL_SPEC
 
 template <typename ComputeOp, typename LoopOp>

clang/lib/CIR/CodeGen/CIRGenStmtOpenACC.cpp

@@ -306,6 +306,29 @@ CIRGenFunction::emitOpenACCCacheConstruct(const OpenACCCacheConstruct &s) {
 
 mlir::LogicalResult
 CIRGenFunction::emitOpenACCAtomicConstruct(const OpenACCAtomicConstruct &s) {
-  cgm.errorNYI(s.getSourceRange(), "OpenACC Atomic Construct");
-  return mlir::failure();
+  // For now, we are only support 'read', so diagnose. We can switch on the kind
+  // later once we start implementing the other 3 forms.
+  if (s.getAtomicKind() != OpenACCAtomicKind::Read) {
+    cgm.errorNYI(s.getSourceRange(), "OpenACC Atomic Construct");
+    return mlir::failure();
+  }
+
+  // While Atomic is an 'associated statement' construct, it 'steals' the
+  // expression it is associated with rather than emitting it inside of it.  So
+  // it has custom emit logic.
+  mlir::Location start = getLoc(s.getSourceRange().getBegin());
+  OpenACCAtomicConstruct::StmtInfo inf = s.getAssociatedStmtInfo();
+  // Atomic 'read' only permits 'v = x', where v and x are both scalar L values.
+  // The getAssociatedStmtInfo strips off implicit casts, which includes
+  // implicit conversions and L-to-R-Value conversions, so we can just emit it
+  // as an L value.  The Flang implementation has no problem with different
+  // types, so it appears that the dialect can handle the conversions.
+  mlir::Value v = emitLValue(inf.V).getPointer();
+  mlir::Value x = emitLValue(inf.X).getPointer();
+  mlir::Type resTy = convertType(inf.V->getType());
+  auto op = mlir::acc::AtomicReadOp::create(builder, start, x, v, resTy,
+                                            /*ifCond=*/{});
+  emitOpenACCClauses(op, s.getDirectiveKind(), s.getDirectiveLoc(),
+                     s.clauses());
+  return mlir::success();
 }

clang/lib/CodeGen/BackendUtil.cpp

@@ -234,9 +234,12 @@ class EmitAssemblyHelper {
 };
 } // namespace
 
-static AllocTokenOptions getAllocTokenOptions(const CodeGenOptions &CGOpts) {
+static AllocTokenOptions getAllocTokenOptions(const LangOptions &LangOpts,
+                                              const CodeGenOptions &CGOpts) {
   AllocTokenOptions Opts;
-  Opts.MaxTokens = CGOpts.AllocTokenMax;
+  if (LangOpts.AllocTokenMode)
+    Opts.Mode = *LangOpts.AllocTokenMode;
+  Opts.MaxTokens = LangOpts.AllocTokenMax;
   Opts.Extended = CGOpts.SanitizeAllocTokenExtended;
   Opts.FastABI = CGOpts.SanitizeAllocTokenFastABI;
   return Opts;
@@ -802,7 +805,7 @@ static void addSanitizers(const Triple &TargetTriple,
         // memory allocation function detection.
         MPM.addPass(InferFunctionAttrsPass());
       }
-      MPM.addPass(AllocTokenPass(getAllocTokenOptions(CodeGenOpts)));
+      MPM.addPass(AllocTokenPass(getAllocTokenOptions(LangOpts, CodeGenOpts)));
     }
   };
   if (ClSanitizeOnOptimizerEarlyEP) {

clang/lib/Frontend/CompilerInvocation.cpp

@@ -1841,10 +1841,6 @@ void CompilerInvocationBase::GenerateCodeGenArgs(const CodeGenOptions &Opts,
        serializeSanitizerKinds(Opts.SanitizeAnnotateDebugInfo))
     GenerateArg(Consumer, OPT_fsanitize_annotate_debug_info_EQ, Sanitizer);
 
-  if (Opts.AllocTokenMax)
-    GenerateArg(Consumer, OPT_falloc_token_max_EQ,
-                std::to_string(*Opts.AllocTokenMax));
-
   if (!Opts.EmitVersionIdentMetadata)
     GenerateArg(Consumer, OPT_Qn);
 
@@ -2358,15 +2354,6 @@ bool CompilerInvocation::ParseCodeGenArgs(CodeGenOptions &Opts, ArgList &Args,
     }
   }
 
-  if (const auto *Arg = Args.getLastArg(options::OPT_falloc_token_max_EQ)) {
-    StringRef S = Arg->getValue();
-    uint64_t Value = 0;
-    if (S.getAsInteger(0, Value))
-      Diags.Report(diag::err_drv_invalid_value) << Arg->getAsString(Args) << S;
-    else
-      Opts.AllocTokenMax = Value;
-  }
-
   Opts.EmitVersionIdentMetadata = Args.hasFlag(OPT_Qy, OPT_Qn, true);
 
   if (!LangOpts->CUDAIsDevice)
@@ -4037,6 +4024,29 @@ void CompilerInvocationBase::GenerateLangArgs(const LangOptions &Opts,
 
   if (!Opts.RandstructSeed.empty())
     GenerateArg(Consumer, OPT_frandomize_layout_seed_EQ, Opts.RandstructSeed);
+
+  if (Opts.AllocTokenMax)
+    GenerateArg(Consumer, OPT_falloc_token_max_EQ,
+                std::to_string(*Opts.AllocTokenMax));
+
+  if (Opts.AllocTokenMode) {
+    StringRef S;
+    switch (*Opts.AllocTokenMode) {
+    case llvm::AllocTokenMode::Increment:
+      S = "increment";
+      break;
+    case llvm::AllocTokenMode::Random:
+      S = "random";
+      break;
+    case llvm::AllocTokenMode::TypeHash:
+      S = "typehash";
+      break;
+    case llvm::AllocTokenMode::TypeHashPointerSplit:
+      S = "typehashpointersplit";
+      break;
+    }
+    GenerateArg(Consumer, OPT_falloc_token_mode_EQ, S);
+  }
 }
 
 bool CompilerInvocation::ParseLangArgs(LangOptions &Opts, ArgList &Args,
@@ -4673,6 +4683,23 @@ bool CompilerInvocation::ParseLangArgs(LangOptions &Opts, ArgList &Args,
   if (const Arg *A = Args.getLastArg(OPT_frandomize_layout_seed_EQ))
     Opts.RandstructSeed = A->getValue(0);
 
+  if (const auto *Arg = Args.getLastArg(options::OPT_falloc_token_max_EQ)) {
+    StringRef S = Arg->getValue();
+    uint64_t Value = 0;
+    if (S.getAsInteger(0, Value))
+      Diags.Report(diag::err_drv_invalid_value) << Arg->getAsString(Args) << S;
+    else
+      Opts.AllocTokenMax = Value;
+  }
+
+  if (const auto *Arg = Args.getLastArg(options::OPT_falloc_token_mode_EQ)) {
+    StringRef S = Arg->getValue();
+    if (auto Mode = getAllocTokenModeFromString(S))
+      Opts.AllocTokenMode = Mode;
+    else
+      Diags.Report(diag::err_drv_invalid_value) << Arg->getAsString(Args) << S;
+  }
+
   // Validate options for HLSL
   if (Opts.HLSL) {
     // TODO: Revisit restricting SPIR-V to logical once we've figured out how to