[ThinLTO] Handle variable with twice promoted name (Rust)

[tejohnson]

Summary

Ensure that we can internalize values produced from two rounds of
promotion.

Note that this cannot happen currently via clang, but in other use cases
such as the Rust compiler which does a first round of ThinLTO on library
code, producing bitcode, and a second round on the final binary.

In particular this can happen if a function is exported and promoted,
ending up with a ".llvm.${hash}" suffix, and then goes through a round
of optimization creating an internal switch table expansion variable
that is internal and contains the promoted name of the enclosing
function. This variable will be promoted in the second round of ThinLTO
if @foo is imported again, and therefore ends up with two
".llvm.${hash}" suffixes. Only the final one should be stripped when
consulting the index to locate the summary.

[wmi]

LGTM.

llvm/include/llvm/IR/ModuleSummaryIndex.h

1271

I see that the patch and test descriptions are already very informative about why two rounds of promotion is possible and why rightmost split off is needed for internalizing. Maybe briefly mention two rounds of promotion here to let the reader conveniently know which side chose to split off is not arbitrary.

[tejohnson]

Address comments

llvm/include/llvm/IR/ModuleSummaryIndex.h

static std::string getGlobalNameForLocal(StringRef Name, ModuleHash ModHash) {
     SmallString<256> NewName(Name);
     NewName += ".llvm.";
     NewName += utostr((uint64_t(ModHash[0]) << 32) |
                       ModHash[1]); // Take the first 64 bits
     return NewName.str();
   }
 
   /// Helper to obtain the unpromoted name for a global value (or the original
-  /// name if not promoted).
+  /// name if not promoted). Split off the rightmost ".llvm.${hash}" suffix.
   static StringRef getOriginalNameBeforePromote(StringRef Name) {
-    std::pair<StringRef, StringRef> Pair = Name.split(".llvm.");
+    std::pair<StringRef, StringRef> Pair = Name.rsplit(".llvm.");
     return Pair.first;
   }
 
   typedef ModulePathStringTableTy::value_type ModuleInfo;
 
   /// Add a new module with the given \p Hash, mapped to the given \p
   /// ModID, and return a reference to the module.
   ModuleInfo *addModule(StringRef ModPath, uint64_t ModId,

llvm/test/ThinLTO/X86/Inputs/thinlto-internalize-doublepromoted.ll

+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.15.0"
+
+define i32 @main() {
+entry:
+  call void @foo.llvm.123()
+  ret i32 0
+}
+
+declare void @foo.llvm.123()

llvm/test/ThinLTO/X86/thinlto-internalize-doublepromoted.ll

+; Test to ensure that we can internalize values produced from two rounds
+; of ThinLTO promotion, so they end up with two ".llvm.${hash}" suffixes.
+; Only the second should be stripped when consulting the index to locate the
+; summary.
+;
+; Note that this cannot happen currently via clang, but in other use cases such
+; as the Rust compiler which does a first round of ThinLTO on library code,
+; producing bitcode, and a second round on the final binary.
+;
+; In this case we assume a prior round of ThinLTO has promoted @foo, and
+; subsequent optimization created an internal switch table expansion variable
+; that is internal and contains the promoted name of the enclosing function.
+; This variable will be promoted in the second round of ThinLTO if @foo is
+; imported again.
+
+; RUN: opt -module-summary -o %t.bc %s
+; RUN: opt -module-summary -o %t-main.bc %S/Inputs/thinlto-internalize-doublepromoted.ll
+; RUN: llvm-lto -thinlto-action=thinlink %t.bc %t-main.bc -o %t-index.bc
+; RUN: llvm-lto -thinlto-action=internalize -exported-symbol=main -thinlto-index %t-index.bc %t.bc -o %t.internalize.bc
+; RUN: llvm-dis %t.internalize.bc -o - | FileCheck %s
+
+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.15.0"
+
+; CHECK: @switch.table.foo.llvm.123.llvm.0 = hidden unnamed_addr constant
[e-mail].123 = private unnamed_addr constant [10 x i8] c"\00\01\02\03\00\00\00\00\00\09", align 1
+
+; CHECK: define hidden void @foo.llvm.123()
+define hidden void @foo.llvm.123() {
+  %1 = getelementptr inbounds [10 x i8], [10 x i8]* @switch.table.foo.llvm.123, i64 0, i64 0
+  store i8 1, i8* %1, align 8
+  ret void
+}