[clang][Driver][HIP] Change OffloadingActionBuilder to respect the --no-gpu-bundle-output flag

bolt/include/bolt/Core/BinaryContext.h

@@ -190,6 +190,9 @@ class BinaryContext {
   /// Unique build ID if available for the binary.
   std::optional<std::string> FileBuildID;
 
+  /// GNU property note indicating AArch64 BTI.
+  bool UsesBTI{false};
+
   /// Set of all sections.
   struct CompareSections {
     bool operator()(const BinarySection *A, const BinarySection *B) const {
@@ -384,6 +387,9 @@ class BinaryContext {
   }
   void setFileBuildID(StringRef ID) { FileBuildID = std::string(ID); }
 
+  bool usesBTI() const { return UsesBTI; }
+  void setUsesBTI(bool Value) { UsesBTI = Value; }
+
   bool hasSymbolsWithFileName() const { return HasSymbolsWithFileName; }
   void setHasSymbolsWithFileName(bool Value) { HasSymbolsWithFileName = Value; }
 

bolt/include/bolt/Passes/SplitFunctions.h

@@ -18,25 +18,6 @@
 namespace llvm {
 namespace bolt {
 
-/// Strategy used to partition blocks into fragments.
-enum SplitFunctionsStrategy : char {
-  /// Split each function into a hot and cold fragment using profiling
-  /// information.
-  Profile2 = 0,
-  /// Split each function into a hot, warm, and cold fragment using
-  /// profiling information.
-  CDSplit,
-  /// Split each function into a hot and cold fragment at a randomly chosen
-  /// split point (ignoring any available profiling information).
-  Random2,
-  /// Split each function into N fragments at a randomly chosen split points
-  /// (ignoring any available profiling information).
-  RandomN,
-  /// Split all basic blocks of each function into fragments such that each
-  /// fragment contains exactly a single basic block.
-  All
-};
-
 class SplitStrategy {
 public:
   using BlockIt = BinaryFunction::BasicBlockOrderType::iterator;

bolt/include/bolt/Rewrite/MetadataRewriters.h

@@ -27,6 +27,8 @@ std::unique_ptr<MetadataRewriter> createPseudoProbeRewriter(BinaryContext &);
 
 std::unique_ptr<MetadataRewriter> createSDTRewriter(BinaryContext &);
 
+std::unique_ptr<MetadataRewriter> createGNUPropertyRewriter(BinaryContext &);
+
 } // namespace bolt
 } // namespace llvm
 

bolt/include/bolt/Utils/CommandLineOpts.h

@@ -29,6 +29,25 @@ enum HeatmapModeKind {
   HM_Optional   // perf2bolt --heatmap
 };
 
+/// Strategy used to partition blocks into fragments.
+enum SplitFunctionsStrategy : char {
+  /// Split each function into a hot and cold fragment using profiling
+  /// information.
+  Profile2 = 0,
+  /// Split each function into a hot, warm, and cold fragment using
+  /// profiling information.
+  CDSplit,
+  /// Split each function into a hot and cold fragment at a randomly chosen
+  /// split point (ignoring any available profiling information).
+  Random2,
+  /// Split each function into N fragments at a randomly chosen split points
+  /// (ignoring any available profiling information).
+  RandomN,
+  /// Split all basic blocks of each function into fragments such that each
+  /// fragment contains exactly a single basic block.
+  All
+};
+
 using HeatmapBlockSizes = std::vector<unsigned>;
 struct HeatmapBlockSpecParser : public llvm::cl::parser<HeatmapBlockSizes> {
   explicit HeatmapBlockSpecParser(llvm::cl::Option &O)
@@ -78,6 +97,7 @@ extern llvm::cl::opt<std::string> OutputFilename;
 extern llvm::cl::opt<std::string> PerfData;
 extern llvm::cl::opt<bool> PrintCacheMetrics;
 extern llvm::cl::opt<bool> PrintSections;
+extern llvm::cl::opt<SplitFunctionsStrategy> SplitStrategy;
 
 // The format to use with -o in aggregation mode (perf2bolt)
 enum ProfileFormatKind { PF_Fdata, PF_YAML };

bolt/lib/Passes/LongJmp.cpp

@@ -895,6 +895,10 @@ void LongJmpPass::relaxLocalBranches(BinaryFunction &BF) {
 
 Error LongJmpPass::runOnFunctions(BinaryContext &BC) {
 
+  assert((opts::CompactCodeModel ||
+          opts::SplitStrategy != opts::SplitFunctionsStrategy::CDSplit) &&
+         "LongJmp cannot work with functions split in more than two fragments");
+
   if (opts::CompactCodeModel) {
     BC.outs()
         << "BOLT-INFO: relaxing branches for compact code model (<128MB)\n";

bolt/lib/Passes/SplitFunctions.cpp

@@ -86,29 +86,6 @@ static cl::opt<unsigned> SplitThreshold(
              "increase after splitting."),
     cl::init(0), cl::Hidden, cl::cat(BoltOptCategory));
 
-static cl::opt<SplitFunctionsStrategy> SplitStrategy(
-    "split-strategy", cl::init(SplitFunctionsStrategy::Profile2),
-    cl::values(clEnumValN(SplitFunctionsStrategy::Profile2, "profile2",
-                          "split each function into a hot and cold fragment "
-                          "using profiling information")),
-    cl::values(clEnumValN(SplitFunctionsStrategy::CDSplit, "cdsplit",
-                          "split each function into a hot, warm, and cold "
-                          "fragment using profiling information")),
-    cl::values(clEnumValN(
-        SplitFunctionsStrategy::Random2, "random2",
-        "split each function into a hot and cold fragment at a randomly chosen "
-        "split point (ignoring any available profiling information)")),
-    cl::values(clEnumValN(
-        SplitFunctionsStrategy::RandomN, "randomN",
-        "split each function into N fragments at a randomly chosen split "
-        "points (ignoring any available profiling information)")),
-    cl::values(clEnumValN(
-        SplitFunctionsStrategy::All, "all",
-        "split all basic blocks of each function into fragments such that each "
-        "fragment contains exactly a single basic block")),
-    cl::desc("strategy used to partition blocks into fragments"),
-    cl::cat(BoltOptCategory));
-
 static cl::opt<double> CallScale(
     "call-scale",
     cl::desc("Call score scale coefficient (when --split-strategy=cdsplit)"),
@@ -724,14 +701,14 @@ Error SplitFunctions::runOnFunctions(BinaryContext &BC) {
   // If split strategy is not CDSplit, then a second run of the pass is not
   // needed after function reordering.
   if (BC.HasFinalizedFunctionOrder &&
-      opts::SplitStrategy != SplitFunctionsStrategy::CDSplit)
+      opts::SplitStrategy != opts::SplitFunctionsStrategy::CDSplit)
     return Error::success();
 
   std::unique_ptr<SplitStrategy> Strategy;
   bool ForceSequential = false;
 
   switch (opts::SplitStrategy) {
-  case SplitFunctionsStrategy::CDSplit:
+  case opts::SplitFunctionsStrategy::CDSplit:
     // CDSplit runs two splitting passes: hot-cold splitting (SplitPrfoile2)
     // before function reordering and hot-warm-cold splitting
     // (SplitCacheDirected) after function reordering.
@@ -742,21 +719,21 @@ Error SplitFunctions::runOnFunctions(BinaryContext &BC) {
     opts::AggressiveSplitting = true;
     BC.HasWarmSection = true;
     break;
-  case SplitFunctionsStrategy::Profile2:
+  case opts::SplitFunctionsStrategy::Profile2:
     Strategy = std::make_unique<SplitProfile2>();
     break;
-  case SplitFunctionsStrategy::Random2:
+  case opts::SplitFunctionsStrategy::Random2:
     Strategy = std::make_unique<SplitRandom2>();
     // If we split functions randomly, we need to ensure that across runs with
     // the same input, we generate random numbers for each function in the same
     // order.
     ForceSequential = true;
     break;
-  case SplitFunctionsStrategy::RandomN:
+  case opts::SplitFunctionsStrategy::RandomN:
     Strategy = std::make_unique<SplitRandomN>();
     ForceSequential = true;
     break;
-  case SplitFunctionsStrategy::All:
+  case opts::SplitFunctionsStrategy::All:
     Strategy = std::make_unique<SplitAll>();
     break;
   }

bolt/lib/Rewrite/CMakeLists.txt

@@ -25,6 +25,7 @@ add_llvm_library(LLVMBOLTRewrite
   PseudoProbeRewriter.cpp
   RewriteInstance.cpp
   SDTRewriter.cpp
+  GNUPropertyRewriter.cpp
 
   NO_EXPORT
   DISABLE_LLVM_LINK_LLVM_DYLIB

bolt/lib/Rewrite/GNUPropertyRewriter.cpp

@@ -0,0 +1,147 @@
+//===- bolt/Rewrite/GNUPropertyRewriter.cpp -------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Read the .note.gnu.property section.
+//
+//===----------------------------------------------------------------------===//
+
+#include "bolt/Rewrite/MetadataRewriter.h"
+#include "bolt/Rewrite/MetadataRewriters.h"
+#include "llvm/Support/Errc.h"
+
+using namespace llvm;
+using namespace bolt;
+
+namespace {
+
+class GNUPropertyRewriter final : public MetadataRewriter {
+
+  Expected<uint32_t> decodeGNUPropertyNote(StringRef Desc);
+
+public:
+  GNUPropertyRewriter(StringRef Name, BinaryContext &BC)
+      : MetadataRewriter(Name, BC) {}
+
+  Error sectionInitializer() override;
+};
+
+Error GNUPropertyRewriter::sectionInitializer() {
+
+  ErrorOr<BinarySection &> Sec =
+      BC.getUniqueSectionByName(".note.gnu.property");
+  if (!Sec)
+    return Error::success();
+
+  // Accumulate feature bits
+  uint32_t FeaturesAcc = 0;
+
+  StringRef Buf = Sec->getContents();
+  DataExtractor DE(Buf, BC.AsmInfo->isLittleEndian(),
+                   BC.AsmInfo->getCodePointerSize());
+  DataExtractor::Cursor Cursor(0);
+  while (Cursor && !DE.eof(Cursor)) {
+    const uint32_t NameSz = DE.getU32(Cursor);
+    const uint32_t DescSz = DE.getU32(Cursor);
+    const uint32_t Type = DE.getU32(Cursor);
+
+    StringRef Name =
+        NameSz ? Buf.slice(Cursor.tell(), Cursor.tell() + NameSz) : "<empty>";
+    Cursor.seek(alignTo(Cursor.tell() + NameSz, 4));
+
+    const uint64_t DescOffset = Cursor.tell();
+    StringRef Desc =
+        DescSz ? Buf.slice(DescOffset, DescOffset + DescSz) : "<empty>";
+    Cursor.seek(alignTo(DescOffset + DescSz, 4));
+    if (!Cursor)
+      return createStringError(
+          errc::executable_format_error,
+          "out of bounds while reading .note.gnu.property section: %s",
+          toString(Cursor.takeError()).c_str());
+
+    if (Type == ELF::NT_GNU_PROPERTY_TYPE_0 && Name.starts_with("GNU") &&
+        DescSz) {
+      auto Features = decodeGNUPropertyNote(Desc);
+      if (!Features)
+        return Features.takeError();
+      FeaturesAcc |= *Features;
+    }
+  }
+
+  if (BC.isAArch64()) {
+    BC.setUsesBTI(FeaturesAcc & llvm::ELF::GNU_PROPERTY_AARCH64_FEATURE_1_BTI);
+    if (BC.usesBTI())
+      BC.outs() << "BOLT-WARNING: binary is using BTI. Optimized binary may be "
+                   "corrupted\n";
+  }
+
+  return Error::success();
+}
+
+/// \p Desc contains an array of property descriptors. Each member has the
+/// following structure:
+/// typedef struct {
+///   Elf_Word pr_type;
+///   Elf_Word pr_datasz;
+///   unsigned char pr_data[PR_DATASZ];
+///   unsigned char pr_padding[PR_PADDING];
+/// } Elf_Prop;
+///
+/// As there is no guarantee that the features are encoded in which element of
+/// the array, we have to read all, and OR together the result.
+Expected<uint32_t> GNUPropertyRewriter::decodeGNUPropertyNote(StringRef Desc) {
+  DataExtractor DE(Desc, BC.AsmInfo->isLittleEndian(),
+                   BC.AsmInfo->getCodePointerSize());
+  DataExtractor::Cursor Cursor(0);
+  const uint32_t Align = DE.getAddressSize();
+
+  std::optional<uint32_t> Features = 0;
+  while (Cursor && !DE.eof(Cursor)) {
+    const uint32_t PrType = DE.getU32(Cursor);
+    const uint32_t PrDataSz = DE.getU32(Cursor);
+
+    const uint64_t PrDataStart = Cursor.tell();
+    const uint64_t PrDataEnd = PrDataStart + PrDataSz;
+    Cursor.seek(PrDataEnd);
+    if (!Cursor)
+      return createStringError(
+          errc::executable_format_error,
+          "out of bounds while reading .note.gnu.property section: %s",
+          toString(Cursor.takeError()).c_str());
+
+    if (PrType == llvm::ELF::GNU_PROPERTY_AARCH64_FEATURE_1_AND) {
+      if (PrDataSz != 4) {
+        return createStringError(
+            errc::executable_format_error,
+            "Property descriptor size has to be 4 bytes on AArch64\n");
+      }
+      DataExtractor::Cursor Tmp(PrDataStart);
+      // PrDataSz = 4 -> PrData is uint32_t
+      const uint32_t FeaturesItem = DE.getU32(Tmp);
+      if (!Tmp)
+        return createStringError(
+            errc::executable_format_error,
+            "failed to read property from .note.gnu.property section: %s",
+            toString(Tmp.takeError()).c_str());
+      Features = Features ? (*Features | FeaturesItem) : FeaturesItem;
+    }
+
+    Cursor.seek(alignTo(PrDataEnd, Align));
+    if (!Cursor)
+      return createStringError(errc::executable_format_error,
+                               "out of bounds while reading property array in "
+                               ".note.gnu.property section: %s",
+                               toString(Cursor.takeError()).c_str());
+  }
+  return Features.value_or(0u);
+}
+} // namespace
+
+std::unique_ptr<MetadataRewriter>
+llvm::bolt::createGNUPropertyRewriter(BinaryContext &BC) {
+  return std::make_unique<GNUPropertyRewriter>("gnu-property-rewriter", BC);
+}

bolt/lib/Rewrite/RewriteInstance.cpp

@@ -2115,6 +2115,13 @@ void RewriteInstance::adjustCommandLineOptions() {
     opts::SplitEH = false;
   }
 
+  if (BC->isAArch64() && !opts::CompactCodeModel &&
+      opts::SplitStrategy == opts::SplitFunctionsStrategy::CDSplit) {
+    BC->errs() << "BOLT-ERROR: CDSplit is not supported with LongJmp. Try with "
+                  "'--compact-code-model'\n";
+    exit(1);
+  }
+
   if (opts::StrictMode && !BC->HasRelocations) {
     BC->errs()
         << "BOLT-WARNING: disabling strict mode (-strict) in non-relocation "
@@ -3331,6 +3338,8 @@ void RewriteInstance::initializeMetadataManager() {
   MetadataManager.registerRewriter(createPseudoProbeRewriter(*BC));
 
   MetadataManager.registerRewriter(createSDTRewriter(*BC));
+
+  MetadataManager.registerRewriter(createGNUPropertyRewriter(*BC));
 }
 
 void RewriteInstance::processSectionMetadata() {

bolt/lib/Utils/CommandLineOpts.cpp

@@ -104,6 +104,29 @@ ExecutionCountThreshold("execution-count-threshold",
   cl::Hidden,
   cl::cat(BoltOptCategory));
 
+cl::opt<SplitFunctionsStrategy> SplitStrategy(
+    "split-strategy", cl::init(SplitFunctionsStrategy::Profile2),
+    cl::values(clEnumValN(SplitFunctionsStrategy::Profile2, "profile2",
+                          "split each function into a hot and cold fragment "
+                          "using profiling information")),
+    cl::values(clEnumValN(SplitFunctionsStrategy::CDSplit, "cdsplit",
+                          "split each function into a hot, warm, and cold "
+                          "fragment using profiling information")),
+    cl::values(clEnumValN(
+        SplitFunctionsStrategy::Random2, "random2",
+        "split each function into a hot and cold fragment at a randomly chosen "
+        "split point (ignoring any available profiling information)")),
+    cl::values(clEnumValN(
+        SplitFunctionsStrategy::RandomN, "randomN",
+        "split each function into N fragments at a randomly chosen split "
+        "points (ignoring any available profiling information)")),
+    cl::values(clEnumValN(
+        SplitFunctionsStrategy::All, "all",
+        "split all basic blocks of each function into fragments such that each "
+        "fragment contains exactly a single basic block")),
+    cl::desc("strategy used to partition blocks into fragments"),
+    cl::cat(BoltOptCategory));
+
 bool HeatmapBlockSpecParser::parse(cl::Option &O, StringRef ArgName,
                                    StringRef Arg, HeatmapBlockSizes &Val) {
   // Parses a human-readable suffix into a shift amount or nullopt on error.