[EVM] Adding LLVMLinkEVM C-API.

Please, note this is a temporary patch. It adds initial support
of dependencies, but it doesn't work in a general case.
A full solution will be more sophisticated and will likely be
implemented on the FE driver side without a need of the LLD usage.

lld/include/lld-c/LLDAsLibraryC.h

@@ -123,6 +123,32 @@ char **LLVMGetUndefinedLinkerSymbolsEraVM(LLVMMemoryBufferRef inBuffer,
  *  LLVMGetUndefinedSymbolsEraVM(). */
 void LLVMDisposeUndefinedLinkerSymbolsEraVM(char *linkerSymbolNames[],
                                             uint64_t numLinkerSymbols);
+
+/** Links the deploy and runtime ELF object files using the information about
+ *  dependencies.
+ *  \p inBuffers - array of input memory buffers with following structure:
+ *
+ *   inBuffers[0] - deploy ELF object code
+ *   inBuffers[1] - deplyed (runtime) ELF object code
+ *   --------------------------
+ *   inBuffers[2] - 1-st sub-contract (final EVM bytecode)
+ *   ...
+ *   inBuffers[N] - N-st sub-contract (final EVM bytecode)
+ *
+ *  Sub contracts are optional. They should have the same ordering as in
+ *  the YUL layout.
+ *
+ *  \p inBuffersIDs - array of string identifiers of the buffers. IDs correspond
+ *  to the object names in the YUL layout.
+ *  On success, outBuffers[0] will contain the deploy bytecode and outBuffers[1]
+ *  the runtime bytecode.
+ *  In case of an error the function returns 'true' and the error message is
+ *  passes in \p errorMessage. The message should be disposed by
+ *  'LLVMDisposeMessage'. */
+LLVMBool LLVMLinkEVM(LLVMMemoryBufferRef *inBuffers, const char *inBuffersIDs[],
+                     uint64_t numInBuffers, LLVMMemoryBufferRef outBuffers[2],
+                     char **errorMessage);
+
 LLVM_C_EXTERN_C_END
 
 #endif // LLD_C_LLDASLIBRARYC_H

lld/lld-c/LLDAsLibraryC.cpp

@@ -443,3 +443,184 @@ void LLVMDisposeUndefinedLinkerSymbolsEraVM(char *linkerSymbolNames[],
     std::free(linkerSymbolNames[idx]);
   std::free(linkerSymbolNames);
 }
+
+//----------------------------------------------------------------------------//
+
+/// This function generates a linker script for EVM architecture.
+/// \p memBufs - array of input memory buffers with following structure:
+///
+///   memBufs[0] - deploy object code
+///   memBufs[1] - deplyed object code
+///   --------------------------
+///   memBufs[2] - 1-st sub-contract (final EVM bytecode)
+///   ...
+///   memBufs[N] - N-st sub-contract (final EVM bytecode)
+///
+/// Sub contracts are optional. They should have the same ordering as in
+/// the YUL layout.
+///
+/// \p bufIDs - array of string identifiers of the buffers. IDs correspond
+/// to the object names in the YUL layout.
+///
+/// For example, the YUL object:
+///
+///   |--D_105_deploy --||--D_105_deployed --||-- B_40 --|
+///
+///   __datasize_B_40 = 1384;
+///   SECTIONS {
+///     . = 0;
+///     .text : SUBALIGN(1) {
+///       D_105(.text);
+///       __dataoffset_D_105_deployed = .;
+///       D_105_deployed(.text);
+///       __datasize_D_105_deployed = . - __dataoffset_D_105_deployed;
+///       __dataoffset_B_40 = .;
+///       __datasize_D_105 = __dataoffset_B_40 + __datasize_B_40;
+///       LONG(__dataoffset_D_105_deployed);
+///     }
+///
+/// The dot '.' denotes current location in the resulting file.
+/// The purpose of the script is to define datasize/dataoffset absolute symbols
+/// that reflect the YUL layour.
+static std::string creteEVMLinkerScript(ArrayRef<LLVMMemoryBufferRef> memBufs,
+                                        ArrayRef<const char *> bufIDs) {
+  assert(memBufs.size() == bufIDs.size());
+  Twine dataSizePrefix("__datasize_");
+  Twine dataOffsetPrefix("__dataoffset_");
+
+  // Define the script part related to the top-level contract.
+  StringRef topName(bufIDs[0]);
+  StringRef deployed(bufIDs[1]);
+
+  Twine topLevel = topName + "(.text);\n" + dataOffsetPrefix + deployed +
+                   " = .;\n" + deployed + "(.text);\n" + dataSizePrefix +
+                   deployed + " = . - " + dataOffsetPrefix + deployed + ";\n";
+
+  std::string symDatasizeDeps;
+  std::string symDataOffsetDeps;
+  if (memBufs.size() > 2) {
+    // Define datasize symbols for the dependent contracts. They start after
+    // {deploy, deployed} pair of the top-level contract,.i.e. at index 2.
+    for (unsigned idx = 2; idx < bufIDs.size(); ++idx)
+      symDatasizeDeps += (dataSizePrefix + bufIDs[idx] + " = " +
+                          Twine(LLVMGetBufferSize(memBufs[idx])) + ";\n")
+                             .str();
+
+    // Define dataoffset symbols for the dependent contracts.
+    symDataOffsetDeps = (dataOffsetPrefix + bufIDs[2] + " = .;\n").str();
+    for (unsigned idx = 3; idx < bufIDs.size(); ++idx) {
+      symDataOffsetDeps +=
+          (dataOffsetPrefix + bufIDs[idx] + " = " + dataOffsetPrefix +
+           bufIDs[idx - 1] + " + " + dataSizePrefix + bufIDs[idx - 1] + ";\n")
+              .str();
+    }
+  }
+
+  // Define datasize symbol for top-level contract.
+  std::string symDatasizeTop = (dataSizePrefix + topName + " = ").str();
+  if (memBufs.size() > 2)
+    symDatasizeTop += (dataOffsetPrefix + bufIDs.back() + " + " +
+                       dataSizePrefix + bufIDs.back() + ";\n")
+                          .str();
+  else
+    symDatasizeTop += ".;\n";
+
+  // Emit size of the deploy code as an 8-byte usingned integer.
+  Twine deploySize = "LONG(" + dataOffsetPrefix + deployed + ");\n";
+
+  Twine script = formatv("{0}\n\
+ENTRY(0);\n\
+SECTIONS {\n\
+  . = 0;\n\
+  .code : SUBALIGN(1) {\n\
+{1}\
+{2}\
+{3}\
+{4}\
+  }\n\
+}\n\
+",
+                         symDatasizeDeps, topLevel, symDataOffsetDeps,
+                         symDatasizeTop, deploySize);
+
+  return script.str();
+}
+
+LLVMBool LLVMLinkEVM(LLVMMemoryBufferRef inBuffers[],
+                     const char *inBuffersIDs[], uint64_t numInBuffers,
+                     LLVMMemoryBufferRef outBuffers[2], char **errorMessage) {
+  assert(numInBuffers > 1);
+  SmallVector<MemoryBufferRef> localInMemBufRefs(3);
+  SmallVector<std::unique_ptr<MemoryBuffer>> localInMemBufs(3);
+  for (unsigned idx = 0; idx < 2; ++idx) {
+    MemoryBufferRef Ref = *unwrap(inBuffers[idx]);
+    localInMemBufs[idx] =
+        MemoryBuffer::getMemBuffer(Ref.getBuffer(), inBuffersIDs[idx],
+                                   /*RequiresNullTerminator*/ false);
+    localInMemBufRefs[idx] = localInMemBufs[idx]->getMemBufferRef();
+  }
+
+  std::string linkerScript = creteEVMLinkerScript(
+      ArrayRef(inBuffers, numInBuffers), ArrayRef(inBuffersIDs, numInBuffers));
+  std::unique_ptr<MemoryBuffer> ScriptBuf =
+      MemoryBuffer::getMemBuffer(linkerScript, "script.x");
+  localInMemBufRefs[2] = ScriptBuf->getMemBufferRef();
+
+  SmallVector<const char *, 16> lldArgs;
+  lldArgs.push_back("ld.lld");
+  lldArgs.push_back("-T");
+  lldArgs.push_back("script.x");
+
+  // Use remapping of file names (a linker feature) to replace file names with
+  // indexes in the array of memory buffers.
+  Twine remapStr("--remap-inputs=");
+  std::string remapDeployStr = (remapStr + inBuffersIDs[0] + "=0").str();
+  lldArgs.push_back(remapDeployStr.c_str());
+
+  std::string remapDeployedStr = (remapStr + inBuffersIDs[1] + "=1").str();
+  lldArgs.push_back(remapDeployedStr.c_str());
+
+  lldArgs.push_back("--remap-inputs=script.x=2");
+
+  // Deploy code
+  lldArgs.push_back(inBuffersIDs[0]);
+  // Deployed code
+  lldArgs.push_back(inBuffersIDs[1]);
+
+  lldArgs.push_back("--oformat=binary");
+
+  SmallString<0> codeString;
+  raw_svector_ostream ostream(codeString);
+  SmallString<0> errorString;
+  raw_svector_ostream errorOstream(errorString);
+
+  // Lld-as-a-library is not thread safe, as it has a global state,
+  // so we need to protect lld from simultaneous access from different threads.
+  std::unique_lock<std::mutex> lock(lldMutex);
+  const lld::Result s =
+      lld::lldMainMemBuf(localInMemBufRefs, &ostream, lldArgs, outs(),
+                         errorOstream, {{lld::Gnu, &lld::elf::linkMemBuf}});
+  lock.unlock();
+
+  bool Ret = !s.retCode && s.canRunAgain;
+  if (!Ret) {
+    *errorMessage = strdup(errorString.c_str());
+    return true;
+  }
+
+  StringRef data = ostream.str();
+  // Linker script adds size of the deploy code as a 8-byte BE unsigned to the
+  // end of .text section. Knowing this, we can extract final deploy and
+  // deployed codes.
+  assert(data.size() > 4);
+  size_t deploySize = support::endian::read32be(data.data() + data.size() - 4);
+  assert(deploySize < data.size());
+  size_t deployedSize = data.size() - deploySize - 4;
+
+  outBuffers[0] = LLVMCreateMemoryBufferWithMemoryRangeCopy(
+      data.data(), deploySize, "deploy");
+  outBuffers[1] = LLVMCreateMemoryBufferWithMemoryRangeCopy(
+      data.data() + deploySize, deployedSize, "deployed");
+
+  return false;
+}

llvm/lib/Target/EVM/MCTargetDesc/EVMTargetStreamer.cpp

@@ -11,14 +11,22 @@
 //===----------------------------------------------------------------------===//
 
 #include "MCTargetDesc/EVMTargetStreamer.h"
+#include "llvm/BinaryFormat/ELF.h"
+#include "llvm/MC/MCSymbolELF.h"
+#include "llvm/Support/Casting.h"
 
 using namespace llvm;
 
 // EVMTargetStreamer implementations
 
 EVMTargetStreamer::EVMTargetStreamer(MCStreamer &S) : MCTargetStreamer(S) {}
 
-EVMTargetStreamer::~EVMTargetStreamer() = default;
+void EVMTargetStreamer::emitLabel(MCSymbol *Symbol) {
+  // This is mostly a workaround for the current linking scheme.
+  // Mark all the symbols as local to their translation units.
+  auto ELFSymbol = cast<MCSymbolELF>(Symbol);
+  ELFSymbol->setBinding(ELF::STB_LOCAL);
+}
 
 EVMTargetObjStreamer::EVMTargetObjStreamer(MCStreamer &S)
     : EVMTargetStreamer(S) {}

llvm/lib/Target/EVM/MCTargetDesc/EVMTargetStreamer.h

@@ -24,7 +24,8 @@ class EVMTargetStreamer : public MCTargetStreamer {
   EVMTargetStreamer(EVMTargetStreamer &&) = delete;
   EVMTargetStreamer &operator=(const EVMTargetStreamer &) = delete;
   EVMTargetStreamer &operator=(EVMTargetStreamer &&) = delete;
-  ~EVMTargetStreamer() override;
+
+  void emitLabel(MCSymbol *Symbol) override;
 };
 
 /// This part is for ASCII assembly output