Add FatLTO support for COFF

clang/lib/Driver/ToolChains/Clang.cpp

@@ -7798,7 +7798,7 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
                                options::OPT_fno_fat_lto_objects)) {
     if (IsUsingLTO && A->getOption().matches(options::OPT_ffat_lto_objects)) {
       assert(LTOMode == LTOK_Full || LTOMode == LTOK_Thin);
-      if (!Triple.isOSBinFormatELF()) {
+      if (!Triple.isOSBinFormatELF() && !Triple.isOSBinFormatCOFF()) {
         D.Diag(diag::err_drv_unsupported_opt_for_target)
             << A->getAsString(Args) << TC.getTripleString();
       }

clang/test/CodeGen/fat-lto-objects.c

@@ -35,6 +35,9 @@
 // RUN: %clang -cc1 -triple x86_64-unknown-linux-gnu -flto=full -ffat-lto-objects -fsplit-lto-unit -S < %s -o - \
 // RUN: | FileCheck %s --check-prefixes=ASM
 
+// RUN: %clang -cc1 -triple i386-pc-win32 -flto=full -ffat-lto-objects -fsplit-lto-unit -S < %s -o - \
+// RUN: | FileCheck %s --check-prefixes=ASM-COFF
+
 /// Make sure that FatLTO generates .llvm.lto sections that are the same as the output from normal LTO compilations
 // RUN: %clang -O2 --target=x86_64-unknown-linux-gnu -fPIE -flto=full -ffat-lto-objects -c %s -o %t.fatlto.full.o
 // RUN: llvm-objcopy --dump-section=.llvm.lto=%t.fatlto.full.bc %t.fatlto.full.o
@@ -43,6 +46,13 @@
 // RUN: llvm-dis < %t.nofat.full.bc -o %t.nofat.full.ll
 // RUN: diff %t.fatlto.full.ll %t.nofat.full.ll
 
+// RUN: %clang -O2 --target=i386-pc-win32 -flto=full -ffat-lto-objects -c %s -o %t.fatlto.full.coff.o
+// RUN: llvm-objcopy --dump-section=.llvm.lto=%t.fatlto.full.coff.bc %t.fatlto.full.coff.o
+// RUN: llvm-dis < %t.fatlto.full.coff.bc -o %t.fatlto.full.coff.ll
+// RUN: %clang -O2 --target=i386-pc-win32 -flto=full -c %s -o %t.nofat.full.coff.bc
+// RUN: llvm-dis < %t.nofat.full.coff.bc -o %t.nofat.full.coff.ll
+// RUN: diff %t.fatlto.full.coff.ll %t.nofat.full.coff.ll
+
 // RUN: %clang -O2 --target=x86_64-unknown-linux-gnu -fPIE -flto=thin -ffat-lto-objects -c %s -o %t.fatlto.thin.o
 // RUN: llvm-objcopy --dump-section=.llvm.lto=%t.fatlto.thin.bc %t.fatlto.thin.o
 // RUN: llvm-dis < %t.fatlto.thin.bc -o %t.fatlto.thin.ll
@@ -67,6 +77,10 @@
 // ASM-NEXT:        .asciz  "BC
 // ASM-NEXT: .size   .Lllvm.embedded.object
 
+//      ASM-COFF: .section .llvm.lto,"ynD"
+// ASM-COFF-NEXT: L_llvm.embedded.object:
+// ASM-COFF-NEXT:        .asciz  "BC
+
 const char* foo = "foo";
 
 int test(void) {

clang/test/Driver/fat-lto-objects.c

@@ -1,10 +1,12 @@
 // RUN: %clang --target=x86_64-unknown-linux-gnu -flto -ffat-lto-objects -### %s -c 2>&1 | FileCheck %s -check-prefix=CHECK-CC
+// RUN: %clang --target=i386-pc-win32 -flto -ffat-lto-objects -### %s -c 2>&1 | FileCheck %s -check-prefix=CHECK-CC
 // CHECK-CC: -cc1
 // CHECK-CC-SAME: -emit-obj
 // CHECK-CC-SAME: -ffat-lto-objects
 
 /// Without -flto -S will just emit normal ASM, so we don't expect -emit-{llvm,obj} or -ffat-lto-objects to be passed to cc1.
 // RUN: %clang --target=x86_64-unknown-linux-gnu -ffat-lto-objects -### %s -S 2>&1 | FileCheck %s -check-prefix=CHECK-CC-S
+// RUN: %clang --target=i386-pc-win32 -ffat-lto-objects -### %s -S 2>&1 | FileCheck %s -check-prefix=CHECK-CC-S
 // CHECK-CC-S: -cc1
 // CHECK-CC-S: -S
 // CHECK-CC-S-NOT: -emit-obj
@@ -13,35 +15,41 @@
 
 /// When fat LTO is enabled with -S, we expect asm output and -ffat-lto-objects to be passed to cc1.
 // RUN: %clang --target=x86_64-unknown-linux-gnu -flto -ffat-lto-objects -### %s -S 2>&1 | FileCheck %s -check-prefix=CHECK-CC-S-LTO
+// RUN: %clang --target=i386-pc-win32 -flto -ffat-lto-objects -### %s -S 2>&1 | FileCheck %s -check-prefix=CHECK-CC-S-LTO
 // CHECK-CC-S-LTO: -cc1
 // CHECK-CC-S-NOT: -emit-llvm
 // CHECK-CC-S-LTO-SAME: -ffat-lto-objects
 
 /// When fat LTO is enabled with -S and -emit-llvm, we expect IR output and -ffat-lto-objects to be passed to cc1.
 // RUN: %clang --target=x86_64-unknown-linux-gnu -flto -ffat-lto-objects -### %s -S -emit-llvm 2>&1 | FileCheck %s -check-prefix=CHECK-CC-S-EL-LTO
+// RUN: %clang --target=i386-pc-win32 -flto -ffat-lto-objects -### %s -S -emit-llvm 2>&1 | FileCheck %s -check-prefix=CHECK-CC-S-EL-LTO
 // CHECK-CC-S-EL-LTO: -cc1
 // CHECK-CC-S-EL-LTO-SAME: -emit-llvm
 // CHECK-CC-S-EL-LTO-SAME: -ffat-lto-objects
 
 /// When fat LTO is enabled without -S we expect native object output and -ffat-lto-object to be passed to cc1.
 // RUN: %clang --target=x86_64-unknown-linux-gnu -flto -ffat-lto-objects -### %s -c 2>&1 | FileCheck %s -check-prefix=CHECK-CC-C-LTO
+// RUN: %clang --target=i386-pc-win32 -flto -ffat-lto-objects -### %s -c 2>&1 | FileCheck %s -check-prefix=CHECK-CC-C-LTO
 // CHECK-CC-C-LTO: -cc1
 // CHECK-CC-C-LTO-SAME: -emit-obj
 // CHECK-CC-C-LTO-SAME: -ffat-lto-objects
 
 /// When fat LTO is enabled with -c and -emit-llvm we expect bitcode output and -ffat-lto-object to be passed to cc1.
 // RUN: %clang --target=x86_64-unknown-linux-gnu -flto -ffat-lto-objects -### %s -c -emit-llvm 2>&1 | FileCheck %s -check-prefix=CHECK-CC-C-EL-LTO
+// RUN: %clang --target=i386-pc-win32 -flto -ffat-lto-objects -### %s -c -emit-llvm 2>&1 | FileCheck %s -check-prefix=CHECK-CC-C-EL-LTO
 // CHECK-CC-C-EL-LTO: -cc1
 // CHECK-CC-C-EL-LTO-SAME: -emit-llvm-bc
 // CHECK-CC-C-EL-LTO-SAME: -ffat-lto-objects
 
 /// Make sure we don't have a warning for -ffat-lto-objects being unused
 // RUN: %clang --target=x86_64-unknown-linux-gnu -ffat-lto-objects -fdriver-only -Werror -v %s -c 2>&1 | FileCheck %s -check-prefix=CHECK-CC-NOLTO
+// RUN: %clang --target=i386-pc-win32 -ffat-lto-objects -fdriver-only -Werror -v %s -c 2>&1 | FileCheck %s -check-prefix=CHECK-CC-NOLTO
 // CHECK-CC-NOLTO: -cc1
 // CHECK-CC-NOLTO-SAME: -emit-obj
 // CHECK-CC-NOLTO-NOT: -ffat-lto-objects
 
 // RUN: %clang --target=x86_64-unknown-linux-gnu -flto -ffat-lto-objects -fno-fat-lto-objects -### %s -c 2>&1 | FileCheck %s -check-prefix=CHECK-CC-NOLTO
+// RUN: %clang --target=i386-pc-win32 -flto -ffat-lto-objects -fno-fat-lto-objects -### %s -c 2>&1 | FileCheck %s -check-prefix=CHECK-CC-NOLTO
 
 /// We need to pass an additional flag (--fat-lto-objects) to lld when linking w/ -flto -ffat-lto-objects
 /// But it should not be there when LTO is disabled w/ -fno-lto

lld/COFF/Config.h

@@ -207,6 +207,9 @@ struct Configuration {
   // Used for /thinlto-remote-compiler-arg:<arg>
   llvm::SmallVector<llvm::StringRef, 0> dtltoCompilerArgs;
 
+  // Used for /fat-lto-objects
+  bool fatLTOObjects = false;
+
   // Used for /opt:[no]ltodebugpassmanager
   bool ltoDebugPassManager = false;
 

lld/COFF/Driver.cpp

@@ -28,6 +28,7 @@
 #include "llvm/Config/llvm-config.h"
 #include "llvm/LTO/LTO.h"
 #include "llvm/Object/COFFImportFile.h"
+#include "llvm/Object/IRObjectFile.h"
 #include "llvm/Option/Arg.h"
 #include "llvm/Option/ArgList.h"
 #include "llvm/Option/Option.h"
@@ -256,6 +257,23 @@ MemoryBufferRef LinkerDriver::takeBuffer(std::unique_ptr<MemoryBuffer> mb) {
   return mbref;
 }
 
+InputFile *LinkerDriver::tryCreateFatLTOFile(MemoryBufferRef mb,
+                                             StringRef archiveName,
+                                             uint64_t offsetInArchive,
+                                             bool lazy) {
+  if (!ctx.config.fatLTOObjects)
+    return nullptr;
+
+  Expected<MemoryBufferRef> fatLTOData =
+      IRObjectFile::findBitcodeInMemBuffer(mb);
+
+  if (errorToBool(fatLTOData.takeError()))
+    return nullptr;
+
+  return BitcodeFile::create(ctx, *fatLTOData, archiveName, offsetInArchive,
+                             lazy);
+}
+
 void LinkerDriver::addBuffer(std::unique_ptr<MemoryBuffer> mb,
                              bool wholeArchive, bool lazy) {
   StringRef filename = mb->getBufferIdentifier();
@@ -289,7 +307,14 @@ void LinkerDriver::addBuffer(std::unique_ptr<MemoryBuffer> mb,
   case file_magic::bitcode:
     addFile(BitcodeFile::create(ctx, mbref, "", 0, lazy));
     break;
-  case file_magic::coff_object:
+  case file_magic::coff_object: {
+    InputFile *obj = tryCreateFatLTOFile(mbref, "", 0, lazy);
+    if (obj)
+      addFile(obj);
+    else
+      addFile(ObjFile::create(ctx, mbref, lazy));
+    break;
+  }
   case file_magic::coff_import_library:
     addFile(ObjFile::create(ctx, mbref, lazy));
     break;
@@ -374,7 +399,9 @@ void LinkerDriver::addArchiveBuffer(MemoryBufferRef mb, StringRef symName,
 
   InputFile *obj;
   if (magic == file_magic::coff_object) {
-    obj = ObjFile::create(ctx, mb);
+    obj = tryCreateFatLTOFile(mb, parentName, offsetInArchive, /*lazy=*/false);
+    if (!obj)
+      obj = ObjFile::create(ctx, mb);
   } else if (magic == file_magic::bitcode) {
     obj = BitcodeFile::create(ctx, mb, parentName, offsetInArchive,
                               /*lazy=*/false);
@@ -2121,6 +2148,10 @@ void LinkerDriver::linkerMain(ArrayRef<const char *> argsArr) {
   config->dtltoCompilerArgs =
       args::getStrings(args, OPT_thinlto_remote_compiler_arg);
 
+  // Handle /fat-lto-objects
+  config->fatLTOObjects =
+      args.hasFlag(OPT_fat_lto_objects, OPT_fat_lto_objects_no, false);
+
   // Handle /dwodir
   config->dwoDir = args.getLastArgValue(OPT_dwodir);
 

lld/COFF/Driver.h

@@ -169,6 +169,8 @@ class LinkerDriver {
 
   std::set<std::string> visitedLibs;
 
+  InputFile *tryCreateFatLTOFile(MemoryBufferRef mb, StringRef archiveName,
+                                 uint64_t offsetInArchive, bool lazy);
   void addBuffer(std::unique_ptr<MemoryBuffer> mb, bool wholeArchive,
                  bool lazy);
   void addArchiveBuffer(MemoryBufferRef mbref, StringRef symName,

lld/COFF/Options.td

@@ -298,6 +298,9 @@ def thinlto_remote_compiler_prepend_arg : P<"thinlto-remote-compiler-prepend-arg
 def thinlto_remote_compiler_arg : P<"thinlto-remote-compiler-arg",
   "Compiler arguments for the ThinLTO distributor to pass for ThinLTO backend "
   "compilations">;
+defm fat_lto_objects: B<"fat-lto-objects",
+    "Use the .llvm.lto section, which contains LLVM bitcode, in fat LTO object files to perform LTO.",
+    "Ignore the .llvm.lto section in relocatable object files (default).">;
 def lto_obj_path : P<
     "lto-obj-path",
     "output native object for merged LTO unit to this path">;

lld/docs/ReleaseNotes.rst

@@ -38,6 +38,9 @@ Breaking changes
 COFF Improvements
 -----------------
 
+* ``/fat-lto-objects`` addded to support FatLTO. Without ``/fat-lto-objects`` or with ``/fat-lto-objects:no``, LLD will link LLVM FatLTO objects using the relocatable object file.
+  (`#165529 <https://github.com/llvm/llvm-project/pull/165529>`_)
+
 MinGW Improvements
 ------------------
 

lld/test/COFF/fatlto/fatlto-invalid.s

@@ -0,0 +1,8 @@
+# RUN: llvm-mc -filetype=obj -triple=i386-pc-win32 %s -o %t
+# RUN: env LLD_IN_TEST=1 not lld-link %t /out:/dev/null /fat-lto-objects 2>&1 | FileCheck %s
+
+# CHECK: error:{{.*}} Invalid bitcode signature
+
+.section        .llvm.lto,"ynD"
+L_llvm.embedded.object:
+        .asciz  "BC\300\3365\000"

lld/test/COFF/fatlto/fatlto.test

@@ -0,0 +1,111 @@
+;; Basic FatLTO tests.
+; REQUIRES: x86
+
+; RUN: rm -rf %t && split-file %s %t && cd %t
+
+;; Ensure that input files contain .llvm.lto section.
+; RUN: llc a-LTO.ll --filetype=obj -o a-fatLTO.o --relocation-model=pic
+; RUN: opt < a-LTO.ll --module-summary -o a-fatLTO.bc
+; RUN: llvm-objcopy --add-section=.llvm.lto=a-fatLTO.bc --set-section-flags=.llvm.lto=exclude a-fatLTO.o
+
+; RUN: llc main-LTO.ll --filetype=obj -o main-fatLTO.o --relocation-model=pic
+; RUN: opt < main-LTO.ll --module-summary -o main-fatLTO.bc
+; RUN: llvm-objcopy --add-section=.llvm.lto=main-fatLTO.bc --set-section-flags=.llvm.lto=exclude main-fatLTO.o
+
+; RUN: llvm-readobj --sections a-fatLTO.o | FileCheck --check-prefix=HAS_LLVM_LTO %s
+; RUN: llvm-readobj --sections main-fatLTO.o | FileCheck --check-prefix=HAS_LLVM_LTO %s
+
+;; Make sure that the section flags are set correctly
+; HAS_LLVM_LTO:          Name: .llvm.lto (2F 34 00 00 00 00 00 00)
+; HAS_LLVM_LTO:          Characteristics [
+; HAS_LLVM_LTO-SAME:     (0xC0000800)
+; HAS_LLVM_LTO-NEXT:       IMAGE_SCN_LNK_REMOVE (0x800)
+; HAS_LLVM_LTO-NEXT:       IMAGE_SCN_MEM_READ (0x40000000)
+; HAS_LLVM_LTO-NEXT:       IMAGE_SCN_MEM_WRITE (0x80000000)
+; HAS_LLVM_LTO-NEXT:     ]
+
+;; Final executable should not have .llvm.lto section no matter what the target is.
+; RUN: lld-link /timestamp:0 /out:foo-fatLTO /entry:main a-fatLTO.o main-fatLTO.o -fat-lto-objects
+; RUN: llvm-readobj --sections foo-fatLTO | FileCheck --check-prefix=CHECK-LTO-TARGET %s
+
+;; Check that fat objects work w/ --start-lib.
+; RUN: lld-link /out:foo-fatLTO.start_lib /entry:main -start-lib a-fatLTO.o main-fatLTO.o -fat-lto-objects
+; RUN: llvm-readobj --sections foo-fatLTO.start_lib | FileCheck --check-prefix=CHECK-LTO-TARGET %s
+
+;; Check if .llvm.lto section gets aggregated in LTO target.
+; CHECK-LTO-TARGET-NOT: Name: .llvm.lto
+
+;; Final executable should not have .llvm.lto section no matter what the target is.
+; RUN: lld-link /timestamp:0 /out:foo-fatNoLTO /entry:main a-fatLTO.o %/t/main-fatLTO.o
+; RUN: llvm-readobj --sections foo-fatNoLTO | FileCheck --check-prefix=CHECK-NON-LTO-TARGET %s
+
+;; Check if .llvm.lto section gets aggregated in non-LTO target.
+; CHECK-NON-LTO-TARGET-NOT: Name: .llvm.lto
+
+;; Check if the LTO target executable produced from FatLTO object file is
+;; identical to the one produced from LTO modules.
+; RUN: opt < a-LTO.ll --module-summary -o a-LTO.bc
+; RUN: opt < main-LTO.ll --module-summary -o main-LTO.bc
+; RUN: lld-link /timestamp:0 /out:foo-LTO /entry:main a-LTO.bc main-LTO.bc
+; RUN: cmp foo-fatLTO foo-LTO
+
+;; Check if the no-LTO target executable produced from FatLTO object file is
+;; identical to the one produced from regular object files.
+; RUN: llc a-LTO.ll --filetype=obj -o a.o
+; RUN: llc main-LTO.ll --filetype=obj -o main.o
+; RUN: lld-link /timestamp:0 /out:foo-noLTO /entry:main a.o main.o
+; RUN: cmp foo-fatNoLTO foo-noLTO
+
+;; Check archive support.
+; RUN: llvm-ar rcs a.a a-fatLTO.o
+; RUN: lld-link /timestamp:0 /out:foo-fatLTO.archive /entry:main /wholearchive a.a main-LTO.bc -fat-lto-objects
+; RUN: cmp foo-fatLTO.archive foo-LTO
+
+;--- a-LTO.ll
+target datalayout = "e-m:x-p:32:32-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32-a:0:32-S32"
+target triple = "i386-pc-windows-msvc19.33.0"
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local i32 @_start() #0 {
+entry:
+  ret i32 0
+}
+
+attributes #0 = { noinline nounwind uwtable }
+
+!llvm.module.flags = !{!0, !1, !2, !3, !4, !5, !6}
+
+!0 = !{i32 1, !"NumRegisterParameters", i32 0}
+!1 = !{i32 2, !"Debug Info Version", i32 3}
+!2 = !{i32 1, !"wchar_size", i32 2}
+!3 = !{i32 7, !"frame-pointer", i32 2}
+!4 = !{i32 1, !"MaxTLSAlign", i32 65536}
+!5 = !{i32 1, !"ThinLTO", i32 0}
+!6 = !{i32 1, !"EnableSplitLTOUnit", i32 1}
+
+;--- main-LTO.ll
+target datalayout = "e-m:x-p:32:32-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32-a:0:32-S32"
+target triple = "i386-pc-windows-msvc19.33.0"
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local i32 @main() #0 {
+entry:
+  %retval = alloca i32, align 4
+  store i32 0, ptr %retval, align 4
+  %call = call i32 (...) @_start()
+  ret i32 %call
+}
+
+declare i32 @_start(...)
+
+attributes #0 = { noinline nounwind uwtable }
+
+!llvm.module.flags = !{!0, !1, !2, !3, !4, !5, !6}
+
+!0 = !{i32 1, !"NumRegisterParameters", i32 0}
+!1 = !{i32 2, !"Debug Info Version", i32 3}
+!2 = !{i32 1, !"wchar_size", i32 2}
+!3 = !{i32 7, !"frame-pointer", i32 2}
+!4 = !{i32 1, !"MaxTLSAlign", i32 65536}
+!5 = !{i32 1, !"ThinLTO", i32 0}
+!6 = !{i32 1, !"EnableSplitLTOUnit", i32 1}

llvm/lib/Transforms/IPO/EmbedBitcodePass.cpp

@@ -27,9 +27,9 @@ PreservedAnalyses EmbedBitcodePass::run(Module &M, ModuleAnalysisManager &AM) {
     reportFatalUsageError("Can only embed the module once");
 
   Triple T(M.getTargetTriple());
-  if (T.getObjectFormat() != Triple::ELF)
-    reportFatalUsageError(
-        "EmbedBitcode pass currently only supports ELF object format");
+  if (T.getObjectFormat() != Triple::ELF && T.getObjectFormat() != Triple::COFF)
+    reportFatalUsageError("EmbedBitcode pass currently only supports COFF and "
+                          "ELF object formats");
 
   std::string Data;
   raw_string_ostream OS(Data);

llvm/test/Transforms/EmbedBitcode/embed-unsupported-object-format.ll

@@ -2,4 +2,4 @@
 
 @a = global i32 1
 
-; CHECK: LLVM ERROR: EmbedBitcode pass currently only supports ELF object format
+; CHECK: LLVM ERROR: EmbedBitcode pass currently only supports COFF and ELF object formats