RuntimeDyldELF: don't create thunk if not needed

This patch doesn't create thunk for branch operation when following conditions are met:
- Architecture is AArch64
- Relocation target is in the same object file
- Relocation target is close enough to be encoded in immediate offset

In such case we branch directly to the target instead of branching to thunk

Differential revision: https://reviews.llvm.org/D28108

llvm-svn: 291431

llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp

@@ -896,6 +896,49 @@ uint32_t RuntimeDyldELF::getMatchingLoRelocation(uint32_t RelType,
   return ELF::R_MIPS_NONE;
 }
 
+// Sometimes we don't need to create thunk for a branch.
+// This typically happens when branch target is located 
+// in the same object file. In such case target is either
+// a weak symbol or symbol in a different executable section.
+// This function checks if branch target is located in the
+// same object file and if distance between source and target
+// fits R_AARCH64_CALL26 relocation. If both conditions are
+// met, it emits direct jump to the target and returns true.
+// Otherwise false is returned and thunk is created.
+bool RuntimeDyldELF::resolveAArch64ShortBranch(
+    unsigned SectionID, relocation_iterator RelI,
+    const RelocationValueRef &Value) {
+  uint64_t Address;
+  if (Value.SymbolName) {
+    auto Loc = GlobalSymbolTable.find(Value.SymbolName);
+
+    // Don't create direct branch for external symbols.
+    if (Loc == GlobalSymbolTable.end())
+      return false;
+
+    const auto &SymInfo = Loc->second;
+    Address = reinterpret_cast<uint64_t>(
+        Sections[SymInfo.getSectionID()].getLoadAddressWithOffset(
+            SymInfo.getOffset()));
+  } else {
+    Address =
+        reinterpret_cast<uint64_t>(Sections[Value.SectionID].getLoadAddress());
+  }
+  uint64_t Offset = RelI->getOffset();
+  uint64_t SourceAddress = Sections[SectionID].getLoadAddressWithOffset(Offset);
+
+  // R_AARCH64_CALL26 requires immediate to be in range -2^27 <= imm < 2^27
+  // If distance between source and target is out of range then we should
+  // create thunk.
+  if (!isInt<28>(Address + Value.Addend - SourceAddress))
+    return false;
+
+  resolveRelocation(Sections[SectionID], Offset, Address, RelI->getType(),
+                    Value.Addend);
+
+  return true;
+}
+
 Expected<relocation_iterator>
 RuntimeDyldELF::processRelocationRef(
     unsigned SectionID, relocation_iterator RelI, const ObjectFile &O,
@@ -1003,7 +1046,7 @@ RuntimeDyldELF::processRelocationRef(
                         (uint64_t)Section.getAddressWithOffset(i->second),
                         RelType, 0);
       DEBUG(dbgs() << " Stub function found\n");
-    } else {
+    } else if (!resolveAArch64ShortBranch(SectionID, RelI, Value)) {
       // Create a new stub function.
       DEBUG(dbgs() << " Create a new stub function\n");
       Stubs[Value] = Section.getStubOffset();

llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.h

@@ -40,6 +40,9 @@ class RuntimeDyldELF : public RuntimeDyldImpl {
   void resolveAArch64Relocation(const SectionEntry &Section, uint64_t Offset,
                                 uint64_t Value, uint32_t Type, int64_t Addend);
 
+  bool resolveAArch64ShortBranch(unsigned SectionID, relocation_iterator RelI,
+                                 const RelocationValueRef &Value);
+
   void resolveARMRelocation(const SectionEntry &Section, uint64_t Offset,
                             uint32_t Value, uint32_t Type, int32_t Addend);
 

llvm/test/ExecutionEngine/RuntimeDyld/AArch64/ELF_ARM64_local_branch.s

@@ -0,0 +1,14 @@
+# RUN: llvm-mc -triple=arm64-none-linux-gnu -filetype=obj -o %T/branch.o %s
+# RUN: llvm-rtdyld -triple=arm64-none-linux-gnu -verify -check=%s %T/branch.o
+
+.globl _main
+.weak _label1
+
+.section .text.1,"ax"
+_label1:
+  nop
+_main:
+  b _label1
+
+## Branch 1 instruction back from _main
+# rtdyld-check: *{4}(_main) = 0x17ffffff