[ThinLTO] Detect partially split modules during the thin link

[tejohnson]

Summary

The changes to disable LTO unit splitting by default (r350949) and
detect inconsistently split LTO units (r350948) are causing some crashes
when the inconsistency is detected in multiple threads simultaneously.
Fix that by having the code always look for the inconsistently split
LTO units during the thin link, by checking for the presence of type
tests recorded in the summaries.

Modify test added in r350948 to remove single threading required to fix
a bot failure due to this issue (and some debugging options added in the
process of diagnosing it).

[pcc]

Can we do this in LTO.cpp maybe? If we do that, we can return a nice error message to the user via the Error interface. The message should probably also suggest recompiling with -fsplit-lto-unit.

[tejohnson]

Address suggestion: move error checking to LTO.cpp

[tejohnson]

ping

[tejohnson]

ping

[pcc]

LGTM

lib/LTO/LTO.cpp

828 ↗

(On Diff #184801)

A user who sees this message isn't going to understand what llvm.type.test or llvm.type.checked.load means. Maybe omit this part of the error?

845 ↗

(On Diff #184801)

Same here.

[tejohnson]

Address comment

llvm/trunk/include/llvm/LTO/LTO.h

Error linkRegularLTO(RegularLTOState::AddedModule Mod,
                        bool LivenessFromIndex);
 
   Error addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms,
                    const SymbolResolution *&ResI, const SymbolResolution *ResE);
 
   Error runRegularLTO(AddStreamFn AddStream);
   Error runThinLTO(AddStreamFn AddStream, NativeObjectCache Cache);
 
+  Error checkPartiallySplit();
+
   mutable bool CalledGetMaxTasks = false;
 
   // Use Optional to distinguish false from not yet initialized.
   Optional<bool> EnableSplitLTOUnit;
 };
 
 /// The resolution for a symbol. The linker must provide a SymbolResolution for
 /// each global symbol based on its internal resolution of that symbol.

llvm/trunk/lib/LTO/LTO.cpp

 #include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/Bitcode/BitcodeReader.h"
 #include "llvm/Bitcode/BitcodeWriter.h"
 #include "llvm/CodeGen/Analysis.h"
 #include "llvm/Config/llvm-config.h"
 #include "llvm/IR/AutoUpgrade.h"
 #include "llvm/IR/DiagnosticPrinter.h"
+#include "llvm/IR/Intrinsics.h"
 #include "llvm/IR/LegacyPassManager.h"
 #include "llvm/IR/Mangler.h"
 #include "llvm/IR/Metadata.h"
 #include "llvm/LTO/LTOBackend.h"
 #include "llvm/LTO/SummaryBasedOptimizations.h"
 #include "llvm/Linker/IRMover.h"
 #include "llvm/Object/IRObjectFile.h"
 #include "llvm/Support/Error.h"
Error LTO::addThinLTO(BitcodeModule BM, ArrayRef<InputFile::Symbol> Syms,
   return Error::success();
 }
 
 unsigned LTO::getMaxTasks() const {
   CalledGetMaxTasks = true;
   return RegularLTO.ParallelCodeGenParallelismLevel + ThinLTO.ModuleMap.size();
 }
 
+// If only some of the modules were split, we cannot correctly handle
+// code that contains type tests or type checked loads.
+Error LTO::checkPartiallySplit() {
+  if (!ThinLTO.CombinedIndex.partiallySplitLTOUnits())
+    return Error::success();
+
+  Function *TypeTestFunc = RegularLTO.CombinedModule->getFunction(
+      Intrinsic::getName(Intrinsic::type_test));
+  Function *TypeCheckedLoadFunc = RegularLTO.CombinedModule->getFunction(
+      Intrinsic::getName(Intrinsic::type_checked_load));
+
+  // First check if there are type tests / type checked loads in the
+  // merged regular LTO module IR.
+  if ((TypeTestFunc && !TypeTestFunc->use_empty()) ||
+      (TypeCheckedLoadFunc && !TypeCheckedLoadFunc->use_empty()))
+    return make_error<StringError>(
+        "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)",
+        inconvertibleErrorCode());
+
+  // Otherwise check if there are any recorded in the combined summary from the
+  // ThinLTO modules.
+  for (auto &P : ThinLTO.CombinedIndex) {
+    for (auto &S : P.second.SummaryList) {
+      auto *FS = dyn_cast<FunctionSummary>(S.get());
+      if (!FS)
+        continue;
+      if (!FS->type_test_assume_vcalls().empty() ||
+          !FS->type_checked_load_vcalls().empty() ||
+          !FS->type_test_assume_const_vcalls().empty() ||
+          !FS->type_checked_load_const_vcalls().empty() ||
+          !FS->type_tests().empty())
+        return make_error<StringError>(
+            "inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)",
+            inconvertibleErrorCode());
+    }
+  }
+  return Error::success();
+}
+
 Error LTO::run(AddStreamFn AddStream, NativeObjectCache Cache) {
   // Compute "dead" symbols, we don't want to import/export these!
   DenseSet<GlobalValue::GUID> GUIDPreservedSymbols;
   DenseMap<GlobalValue::GUID, PrevailingType> GUIDPrevailingResolutions;
   for (auto &Res : GlobalResolutions) {
     // Normally resolution have IR name of symbol. We can do nothing here
     // otherwise. See comments in GlobalResolution struct for more details.
     if (Res.second.IRName.empty())
if (!Conf.StatsFile.empty()) {
     std::error_code EC;
     StatsFile =
         llvm::make_unique<ToolOutputFile>(Conf.StatsFile, EC, sys::fs::F_None);
     if (EC)
       return errorCodeToError(EC);
     StatsFile->keep();
   }
 
+  // Finalize linking of regular LTO modules containing summaries now that
+  // we have computed liveness information.
+  for (auto &M : RegularLTO.ModsWithSummaries)
+    if (Error Err = linkRegularLTO(std::move(M),
+                                   /*LivenessFromIndex=*/true))
+      return Err;
+
+  // Ensure we don't have inconsistently split LTO units with type tests.
+  if (Error Err = checkPartiallySplit())
+    return Err;
+
   Error Result = runRegularLTO(AddStream);
   if (!Result)
     Result = runThinLTO(AddStream, Cache);
 
   if (StatsFile)
     PrintStatisticsJSON(StatsFile->os());
 
   return Result;
 }
 
 Error LTO::runRegularLTO(AddStreamFn AddStream) {
-  for (auto &M : RegularLTO.ModsWithSummaries)
-    if (Error Err = linkRegularLTO(std::move(M),
-                                   /*LivenessFromIndex=*/true))
-      return Err;
-
   // Make sure commons have the right size/alignment: we kept the largest from
   // all the prevailing when adding the inputs, and we apply it here.
   const DataLayout &DL = RegularLTO.CombinedModule->getDataLayout();
   for (auto &I : RegularLTO.Commons) {
     if (!I.second.Prevailing)
       // Don't do anything if no instance of this common was prevailing.
       continue;
     GlobalVariable *OldGV = RegularLTO.CombinedModule->getNamedGlobal(I.first);

llvm/trunk/lib/Transforms/IPO/LowerTypeTests.cpp

for (; UI != E;) {
     if (!isDirectCall(U))
       continue;
 
     U.set(New);
   }
 }
 
 bool LowerTypeTestsModule::lower() {
+  // If only some of the modules were split, we cannot correctly perform
+  // this transformation. We already checked for the presense of type tests
+  // with partially split modules during the thin link, and would have emitted
+  // an error if any were found, so here we can simply return.
+  if ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) ||
+      (ImportSummary && ImportSummary->partiallySplitLTOUnits()))
+    return false;
+
   Function *TypeTestFunc =
       M.getFunction(Intrinsic::getName(Intrinsic::type_test));
   Function *ICallBranchFunnelFunc =
       M.getFunction(Intrinsic::getName(Intrinsic::icall_branch_funnel));
   if ((!TypeTestFunc || TypeTestFunc->use_empty()) &&
       (!ICallBranchFunnelFunc || ICallBranchFunnelFunc->use_empty()) &&
       !ExportSummary && !ImportSummary)
     return false;
 
-  // If only some of the modules were split, we cannot correctly handle
-  // code that contains type tests.
-  if (TypeTestFunc && !TypeTestFunc->use_empty() &&
-      ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) ||
-       (ImportSummary && ImportSummary->partiallySplitLTOUnits())))
-    report_fatal_error("inconsistent LTO Unit splitting with llvm.type.test");
-
   if (ImportSummary) {
     if (TypeTestFunc) {
       for (auto UI = TypeTestFunc->use_begin(), UE = TypeTestFunc->use_end();
            UI != UE;) {
         auto *CI = cast<CallInst>((*UI++).getUser());
         importTypeTest(CI);
       }
     }

llvm/trunk/lib/Transforms/IPO/WholeProgramDevirt.cpp

for (auto &&U : NumUnsafeUsesForTypeTest) {
     if (U.second == 0) {
       U.first->replaceAllUsesWith(True);
       U.first->eraseFromParent();
     }
   }
 }
 
 bool DevirtModule::run() {
+  // If only some of the modules were split, we cannot correctly perform
+  // this transformation. We already checked for the presense of type tests
+  // with partially split modules during the thin link, and would have emitted
+  // an error if any were found, so here we can simply return.
+  if ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) ||
+      (ImportSummary && ImportSummary->partiallySplitLTOUnits()))
+    return false;
+
   Function *TypeTestFunc =
       M.getFunction(Intrinsic::getName(Intrinsic::type_test));
   Function *TypeCheckedLoadFunc =
       M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load));
   Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume));
 
-  // If only some of the modules were split, we cannot correctly handle
-  // code that contains type tests or type checked loads.
-  if ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) ||
-      (ImportSummary && ImportSummary->partiallySplitLTOUnits())) {
-    if ((TypeTestFunc && !TypeTestFunc->use_empty()) ||
-        (TypeCheckedLoadFunc && !TypeCheckedLoadFunc->use_empty()))
-      report_fatal_error("inconsistent LTO Unit splitting with llvm.type.test "
-                         "or llvm.type.checked.load");
-    return false;
-  }
-
   // Normally if there are no users of the devirtualization intrinsics in the
   // module, this pass has nothing to do. But if we are exporting, we also need
   // to handle any users that appear only in the function summaries.
   if (!ExportSummary &&
       (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc ||
        AssumeFunc->use_empty()) &&
       (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty()))
     return false;

llvm/trunk/test/ThinLTO/X86/cfi-devirt.ll

 ; RUN:   -r=%t.o,_ZN1C1fEi,p \
 ; RUN:   -r=%t.o,empty,p \
 ; RUN:   -r=%t.o,_ZTV1B, \
 ; RUN:   -r=%t.o,_ZTV1C, \
 ; RUN:   -r=%t.o,_ZN1A1nEi, \
 ; RUN:   -r=%t.o,_ZN1B1fEi, \
 ; RUN:   -r=%t.o,_ZN1C1fEi, \
 ; RUN:   -r=%t.o,_ZTV1B,px \
-; RUN:   -r=%t.o,_ZTV1C,px 2>&1 | FileCheck %s --check-prefix=REMARK -dump-input=always
-; RUN: llvm-dis %t3.1.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-IR -dump-input=always
+; RUN:   -r=%t.o,_ZTV1C,px 2>&1 | FileCheck %s --check-prefix=REMARK
+; RUN: llvm-dis %t3.1.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-IR
 
 ; New PM
 ; FIXME: Fix machine verifier issues and remove -verify-machineinstrs=0. PR39436.
 ; RUN: llvm-lto2 run %t.o -save-temps -use-new-pm -pass-remarks=. \
 ; RUN:   -verify-machineinstrs=0 \
 ; RUN:   -o %t3 \
 ; RUN:   -r=%t.o,test,px \
 ; RUN:   -r=%t.o,_ZN1A1nEi,p \
 ; RUN:   -r=%t.o,_ZN1B1fEi,p \
 ; RUN:   -r=%t.o,_ZN1C1fEi,p \
 ; RUN:   -r=%t.o,empty,p \
 ; RUN:   -r=%t.o,_ZTV1B, \
 ; RUN:   -r=%t.o,_ZTV1C, \
 ; RUN:   -r=%t.o,_ZN1A1nEi, \
 ; RUN:   -r=%t.o,_ZN1B1fEi, \
 ; RUN:   -r=%t.o,_ZN1C1fEi, \
 ; RUN:   -r=%t.o,_ZTV1B,px \
-; RUN:   -r=%t.o,_ZTV1C,px 2>&1 | FileCheck %s --check-prefix=REMARK -dump-input=always
-; RUN: llvm-dis %t3.1.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-IR -dump-input=always
+; RUN:   -r=%t.o,_ZTV1C,px 2>&1 | FileCheck %s --check-prefix=REMARK
+; RUN: llvm-dis %t3.1.4.opt.bc -o - | FileCheck %s --check-prefix=CHECK-IR
 
 ; REMARK: single-impl: devirtualized a call to _ZN1A1nEi
 
 ; Next check that we emit an error when trying to LTO link this module
 ; containing an llvm.type.checked.load (with a split LTO Unit) with one
-; that does not have a split LTO Unit.
+; that does not have a split LTO Unit. Use -thinlto-distributed-indexes
+; to ensure it is being caught in the thin link.
 ; RUN: opt -thinlto-bc -o %t2.o %S/Inputs/empty.ll
-; RUN: not llvm-lto2 run %t.o %t2.o -save-temps -pass-remarks=. \
-; RUN:   -verify-machineinstrs=0 -thinlto-threads=1 \
+; RUN: not llvm-lto2 run %t.o %t2.o -thinlto-distributed-indexes \
+; RUN:   -verify-machineinstrs=0 \
 ; RUN:   -o %t3 \
 ; RUN:   -r=%t.o,test,px \
 ; RUN:   -r=%t.o,_ZN1A1nEi,p \
 ; RUN:   -r=%t.o,_ZN1B1fEi,p \
 ; RUN:   -r=%t.o,_ZN1C1fEi,p \
 ; RUN:   -r=%t.o,empty,p \
 ; RUN:   -r=%t.o,_ZTV1B, \
 ; RUN:   -r=%t.o,_ZTV1C, \
 ; RUN:   -r=%t.o,_ZN1A1nEi, \
 ; RUN:   -r=%t.o,_ZN1B1fEi, \
 ; RUN:   -r=%t.o,_ZN1C1fEi, \
 ; RUN:   -r=%t.o,_ZTV1B,px \
-; RUN:   -r=%t.o,_ZTV1C,px 2>&1 | FileCheck %s --check-prefix=ERROR -dump-input=always
-; ERROR: LLVM ERROR: inconsistent LTO Unit splitting with llvm.type.test or llvm.type.checked.load
+; RUN:   -r=%t.o,_ZTV1C,px 2>&1 | FileCheck %s --check-prefix=ERROR
+; ERROR: failed: inconsistent LTO Unit splitting (recompile with -fsplit-lto-unit)
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-grtev4-linux-gnu"
 
 %struct.A = type { i32 (...)** }
 %struct.B = type { %struct.A }
 %struct.C = type { %struct.A }