Merge branch 'main' into xegpu_simt_dist

Author: charithaintc

.ci/monolithic-linux.sh

@@ -28,10 +28,14 @@ if [[ -n "${CLEAR_CACHE:-}" ]]; then
   ccache --clear
 fi
 
+mkdir -p artifacts/reproducers
+
+# Make sure any clang reproducers will end up as artifacts.
+export CLANG_CRASH_DIAGNOSTICS_DIR=`realpath artifacts/reproducers`
+
 function at-exit {
   retcode=$?
 
-  mkdir -p artifacts
   ccache --print-stats > artifacts/ccache_stats.txt
   cp "${BUILD_DIR}"/.ninja_log artifacts/.ninja_log
 
@@ -58,9 +62,18 @@ export PIP_BREAK_SYSTEM_PACKAGES=1
 pip install -q -r "${MONOREPO_ROOT}"/mlir/python/requirements.txt
 pip install -q -r "${MONOREPO_ROOT}"/lldb/test/requirements.txt
 pip install -q -r "${MONOREPO_ROOT}"/.ci/requirements.txt
+
+# Set the system llvm-symbolizer as preferred.
+export LLVM_SYMBOLIZER_PATH=`which llvm-symbolizer`
+[[ ! -f "${LLVM_SYMBOLIZER_PATH}" ]] && echo "llvm-symbolizer not found!"
+
+# Set up all runtimes either way. libcxx is a dependency of LLDB.
+# It will not be built unless it is used.
 cmake -S "${MONOREPO_ROOT}"/llvm -B "${BUILD_DIR}" \
       -D LLVM_ENABLE_PROJECTS="${projects}" \
+      -D LLVM_ENABLE_RUNTIMES="libcxx;libcxxabi;libunwind" \
       -G Ninja \
+      -D CMAKE_PREFIX_PATH="${HOME}/.local" \
       -D CMAKE_BUILD_TYPE=Release \
       -D LLVM_ENABLE_ASSERTIONS=ON \
       -D LLVM_BUILD_EXAMPLES=ON \
@@ -69,7 +82,10 @@ cmake -S "${MONOREPO_ROOT}"/llvm -B "${BUILD_DIR}" \
       -D LLVM_ENABLE_LLD=ON \
       -D CMAKE_CXX_FLAGS=-gmlt \
       -D LLVM_CCACHE_BUILD=ON \
+      -D LIBCXX_CXX_ABI=libcxxabi \
       -D MLIR_ENABLE_BINDINGS_PYTHON=ON \
+      -D LLDB_ENABLE_PYTHON=ON \
+      -D LLDB_ENFORCE_STRICT_TEST_REQUIREMENTS=ON \
       -D CMAKE_INSTALL_PREFIX="${INSTALL_DIR}"
 
 echo "--- ninja"

bolt/lib/Core/BinaryBasicBlock.cpp

@@ -372,8 +372,7 @@ void BinaryBasicBlock::updateJumpTableSuccessors() {
              [](const BinaryBasicBlock *BB1, const BinaryBasicBlock *BB2) {
                return BB1->getInputOffset() < BB2->getInputOffset();
              });
-  SuccessorBBs.erase(std::unique(SuccessorBBs.begin(), SuccessorBBs.end()),
-                     SuccessorBBs.end());
+  SuccessorBBs.erase(llvm::unique(SuccessorBBs), SuccessorBBs.end());
 
   for (BinaryBasicBlock *BB : SuccessorBBs)
     addSuccessor(BB);

bolt/lib/Core/BinaryFunction.cpp

@@ -2376,7 +2376,7 @@ Error BinaryFunction::buildCFG(MCPlusBuilder::AllocatorIdTy AllocatorId) {
   // Without doing jump table value profiling we don't have a use for extra
   // (duplicate) branches.
   llvm::sort(TakenBranches);
-  auto NewEnd = std::unique(TakenBranches.begin(), TakenBranches.end());
+  auto NewEnd = llvm::unique(TakenBranches);
   TakenBranches.erase(NewEnd, TakenBranches.end());
 
   for (std::pair<uint32_t, uint32_t> &Branch : TakenBranches) {

bolt/lib/Core/DebugNames.cpp

@@ -440,8 +440,7 @@ void DWARF5AcceleratorTable::computeBucketCount() {
   for (const auto &E : Entries)
     Uniques.push_back(E.second.HashValue);
   array_pod_sort(Uniques.begin(), Uniques.end());
-  std::vector<uint32_t>::iterator P =
-      std::unique(Uniques.begin(), Uniques.end());
+  std::vector<uint32_t>::iterator P = llvm::unique(Uniques);
 
   UniqueHashCount = std::distance(Uniques.begin(), P);
 

bolt/lib/Profile/YAMLProfileWriter.cpp

@@ -133,12 +133,9 @@ std::vector<yaml::bolt::PseudoProbeInfo>
 YAMLProfileWriter::convertNodeProbes(NodeIdToProbes &NodeProbes) {
   struct BlockProbeInfoHasher {
     size_t operator()(const yaml::bolt::PseudoProbeInfo &BPI) const {
-      auto HashCombine = [](auto &Range) {
-        return llvm::hash_combine_range(Range.begin(), Range.end());
-      };
-      return llvm::hash_combine(HashCombine(BPI.BlockProbes),
-                                HashCombine(BPI.CallProbes),
-                                HashCombine(BPI.IndCallProbes));
+      return llvm::hash_combine(llvm::hash_combine_range(BPI.BlockProbes),
+                                llvm::hash_combine_range(BPI.CallProbes),
+                                llvm::hash_combine_range(BPI.IndCallProbes));
     }
   };
 

bolt/lib/Rewrite/RewriteInstance.cpp

@@ -237,6 +237,12 @@ UseGnuStack("use-gnu-stack",
   cl::ZeroOrMore,
   cl::cat(BoltCategory));
 
+static cl::opt<uint64_t> CustomAllocationVMA(
+    "custom-allocation-vma",
+    cl::desc("use a custom address at which new code will be put, "
+             "bypassing BOLT's logic to detect where to put code"),
+    cl::Hidden, cl::cat(BoltCategory));
+
 static cl::opt<bool>
 SequentialDisassembly("sequential-disassembly",
   cl::desc("performs disassembly sequentially"),
@@ -592,6 +598,25 @@ Error RewriteInstance::discoverStorage() {
 
   FirstNonAllocatableOffset = NextAvailableOffset;
 
+  if (opts::CustomAllocationVMA) {
+    // If user specified a custom address where we should start writing new
+    // data, honor that.
+    NextAvailableAddress = opts::CustomAllocationVMA;
+    // Sanity check the user-supplied address and emit warnings if something
+    // seems off.
+    for (const ELF64LE::Phdr &Phdr : PHs) {
+      switch (Phdr.p_type) {
+      case ELF::PT_LOAD:
+        if (NextAvailableAddress >= Phdr.p_vaddr &&
+            NextAvailableAddress < Phdr.p_vaddr + Phdr.p_memsz) {
+          BC->errs() << "BOLT-WARNING: user-supplied allocation vma 0x"
+                     << Twine::utohexstr(NextAvailableAddress)
+                     << " conflicts with ELF segment at 0x"
+                     << Twine::utohexstr(Phdr.p_vaddr) << "\n";
+        }
+      }
+    }
+  }
   NextAvailableAddress = alignTo(NextAvailableAddress, BC->PageAlign);
   NextAvailableOffset = alignTo(NextAvailableOffset, BC->PageAlign);
 
@@ -1055,10 +1080,11 @@ void RewriteInstance::discoverFileObjects() {
       continue;
     }
 
-    if (!Section->isText()) {
+    if (!Section->isText() || Section->isVirtual()) {
       assert(SymbolType != SymbolRef::ST_Function &&
              "unexpected function inside non-code section");
-      LLVM_DEBUG(dbgs() << "BOLT-DEBUG: rejecting as symbol is not in code\n");
+      LLVM_DEBUG(dbgs() << "BOLT-DEBUG: rejecting as symbol is not in code or "
+                           "is in nobits section\n");
       registerName(SymbolSize);
       continue;
     }

bolt/lib/Target/AArch64/AArch64MCPlusBuilder.cpp

@@ -2336,7 +2336,7 @@ class AArch64MCPlusBuilder : public MCPlusBuilder {
   std::optional<Relocation>
   createRelocation(const MCFixup &Fixup,
                    const MCAsmBackend &MAB) const override {
-    const MCFixupKindInfo &FKI = MAB.getFixupKindInfo(Fixup.getKind());
+    MCFixupKindInfo FKI = MAB.getFixupKindInfo(Fixup.getKind());
 
     assert(FKI.TargetOffset == 0 && "0-bit relocation offset expected");
     const uint64_t RelOffset = Fixup.getOffset();

bolt/lib/Target/X86/X86MCPlusBuilder.cpp

@@ -2441,7 +2441,7 @@ class X86MCPlusBuilder : public MCPlusBuilder {
   std::optional<Relocation>
   createRelocation(const MCFixup &Fixup,
                    const MCAsmBackend &MAB) const override {
-    const MCFixupKindInfo &FKI = MAB.getFixupKindInfo(Fixup.getKind());
+    MCFixupKindInfo FKI = MAB.getFixupKindInfo(Fixup.getKind());
 
     assert(FKI.TargetOffset == 0 && "0-bit relocation offset expected");
     const uint64_t RelOffset = Fixup.getOffset();

bolt/test/X86/high-segments.s

@@ -0,0 +1,46 @@
+// Check that we are able to rewrite binaries when we fail to identify a
+// suitable location to put new code and user supplies a custom one via
+// --custom-allocation-vma. This happens more obviously if the binary has
+// segments mapped to very high addresses.
+
+// In this example, my.reserved.section is mapped to a segment to be loaded
+// at address 0x10000000000, while regular text should be at 0x200000. We
+// pick a vma in the middle at 0x700000 to carve space for BOLT to put data,
+// since BOLT's usual route of allocating after the last segment will put
+// code far away and that will blow up relocations from main.
+
+// RUN: split-file %s %t
+// RUN: llvm-mc -filetype=obj -triple x86_64-unknown-unknown %t/main.s -o %t.o
+// RUN: %clang %cflags -no-pie %t.o -o %t.exe -Wl,-T %t/main.ls
+// RUN: llvm-bolt %t.exe -o %t.bolt --custom-allocation-vma=0x700000
+
+//--- main.s
+  .type            reserved_space,@object
+  .section        .my.reserved.section,"awx",@nobits
+  .globl           reserved_space
+  .p2align         4, 0x0
+reserved_space:
+  .zero  0x80000000
+  .size   reserved_space, 0x80000000
+
+	.text
+  .globl main
+  .globl _start
+  .type main, %function
+_start:
+main:
+	.cfi_startproc
+  nop
+  nop
+  nop
+  retq
+	.cfi_endproc
+.size main, .-main
+
+//--- main.ls
+SECTIONS
+{
+    .my.reserved.section 1<<40 : {
+      *(.my.reserved.section);
+    }
+} INSERT BEFORE .comment;

bolt/test/X86/nobits-symbol.s

@@ -0,0 +1,24 @@
+## Check that llvm-bolt doesn't choke on symbols defined in nobits sections.
+
+# RUN: llvm-mc -filetype=obj -triple x86_64-unknown-unknown %s -o %t.o
+# RUN: %clang %cflags %t.o -o %t.exe
+# RUN: llvm-bolt %t.exe -o %t
+#
+
+  .type            symbol_in_nobits,@object
+  .section        .my.nobits.section,"awx",@nobits
+  .globl           symbol_in_nobits
+  .p2align         4, 0x0
+symbol_in_nobits:
+  .zero  0x100000
+  .size   symbol_in_nobits, 0x100000
+
+	.text
+  .globl main
+  .type main, %function
+main:
+	.cfi_startproc
+.LBB06:
+  retq
+	.cfi_endproc
+.size main, .-main