[BOLT] Handle broken .dynsym in stripped binaries

[nhuhuan]

Summary

Strip tools cause a few symbols in .dynsym to have bad section index.
This update safely keeps such broken symbols intact.

Test Plan:

ninja check-bolt

[maksfb]

Looks good, but please add a test case.

[Amir]

We need a test case for this fix: yaml2obj test with a dynamic symbol table with wrong (out of bounds) index. We should check that BOLT keeps this index intact.

[nhuhuan]

Update test case.

[nhuhuan]

Simplified test.

[Amir]

Note for other reviewers: we were unable to get yaml input with a broken symbol section index (yaml2obj refuses to produce an object file). So this approach (yaml2obj + llvm-strip) is the closest we were able to get.

bolt/lib/Rewrite/RewriteInstance.cpp

if (Itr != IslandSizes.end())
       return Itr->second;
     return IslandSizes[&BF] = BF.estimateConstantIslandSize();
   };
 
   // Symbols for the new symbol table.
   std::vector<ELFSymTy> Symbols;
 
   auto getNewSectionIndex = [&](uint32_t OldIndex) {
+    // For dynamic symbol table, the section index could be wrong on the input,
+    // and its value is ignored by the runtime if it's different from
+    // SHN_UNDEF and SHN_ABS.
+    // However, we still need to update dynamic symbol table, so return a
+    // section index, even though the index is broken.
+    if (IsDynSym && OldIndex >= NewSectionIndex.size())
+      return OldIndex;
+
     assert(OldIndex < NewSectionIndex.size() && "section index out of bounds");
     const uint32_t NewIndex = NewSectionIndex[OldIndex];
 
     // We may have stripped the section that dynsym was referencing due to
     // the linker bug. In that case return the old index avoiding marking
     // the symbol as undefined.
     if (IsDynSym && NewIndex != OldIndex && NewIndex == ELF::SHN_UNDEF)
       return OldIndex;

bolt/test/X86/Inputs/broken_dynsym.yaml

+!ELF
+FileHeader:
+  Class: ELFCLASS64
+  Data: ELFDATA2LSB
+  Type: ET_EXEC
+  Machine: EM_X86_64
+ProgramHeaders:
+  - Type: PT_LOAD
+    FirstSec: .a
+    LastSec: .a
+    Align:           0x1000
+Sections:
+  - Name: .a
+    Type: SHT_PROGBITS
+    Content: 00
+    AddressAlign:    0x1
+  - Name: .b
+    Type: 0
+    AddressAlign:    0x1
+  - Name: .bss
+    Type: 0
+    AddressAlign:    0x1
+  - Type: SectionHeaderTable
+    Sections:
+      - Name: .dynsym
+      - Name: .dynstr
+      - Name: .a
+      - Name: .b
+      - Name: .bss
+      - Name: .strtab
+      - Name: .shstrtab
+DynamicSymbols:
+  - Section: .bss

bolt/test/X86/broken_dynsym.test

+# This test checks if BOLT can process stripped binaries, where symbol's section
+# header index is corrupted due to strip tool.
+
+# RUN: yaml2obj %p/Inputs/broken_dynsym.yaml -o %t
+# RUN: llvm-strip -s %t
+# RUN: llvm-bolt %t -o %t.bolt | FileCheck %s
+
+# CHECK-NOT: section index out of bounds