[MC] Clang integrated-as doesn't support local aliases to global variables (GNU as does).

[llvmbot]

Bugzilla Link	24486
Resolution	FIXED
Resolved on	Oct 05, 2015 07:49
Version	unspecified
OS	Linux
Reporter	LLVM Bugzilla Contributor

Extended Description

Trying to resolve this https://code.google.com/p/address-sanitizer/issues/detail?id=398 issue, I found out, that clang integrated-as doesn't support local aliases to global variables (GNU assembler does). Consider the following testcase:

$ cat test.s
 
	.data
	.global foo
foo:
	.long
bar = foo
	.long	bar

$ clang -c tmp.s && readelf -r tmp.o

Relocation section '.rela.data' at offset 0x90 contains 1 entries:
  Offset          Info           Type           Sym. Value    Sym. Name + Addend
000000000000  00020000000a R_X86_64_32       0000000000000000 foo + 0

$ gcc -c tmp.s && readelf -r tmp.o

Relocation section '.rela.data' at offset 0x318 contains 1 entries:
  Offset          Info           Type           Sym. Value    Sym. Name + Addend
000000000000  00020000000a R_X86_64_32       0000000000000000 .data + 0

While GNU as generates relocation for "bar", addressing local .data storage, clang integrated as addresses it to its aliasee, global variable "foo", that might be located in completely different place (e.g. when shared library and executable define the same global variable).

After some investigation, I found out, that the code responsible for Sym. Name entry in '.rela.data' section located here:

$ cat lib/MC/MCExpr.cpp

bool MCExpr::evaluateAsRelocatableImpl(MCValue &Res, const MCAssembler *Asm,
                                       const MCAsmLayout *Layout,
                                       const MCFixup *Fixup,
                                       const SectionAddrMap *Addrs,
                                       bool InSet) const {
.........................
    
  case SymbolRef: {
    const MCSymbolRefExpr *SRE = cast<MCSymbolRefExpr>(this);
    const MCSymbol &Sym = SRE->getSymbol();
    
    // Evaluate recursively if this is a variable.
    if (Sym.isVariable() && SRE->getKind() == MCSymbolRefExpr::VK_None &&
        canExpand(Sym, Asm, InSet)) {
      bool IsMachO = SRE->hasSubsectionsViaSymbols();
      if (Sym.getVariableValue()->evaluateAsRelocatableImpl(
              Res, Asm, Layout, Fixup, Addrs, InSet || IsMachO)) {
........................
  } 

Is this the right place to fix the issue? Could someone more familiar with Clang/LLVM code base help me to cook proper fix?

[llvmbot]

I think the real problem is that for variables we store just the section they are in. This means that

SA.getFragment() == SB.getFragment()

will fail if SA is a variable pointing to SB. This is turn means that the expression evaluation has to be aggressive to compensate.

The fix would probably be to change MCSymbol's SectionOrFragmentAndHasName to store a BaseSymbol instead of the section for variables. For example, given

foo = bar + 1

The base of foo is bar.

Given that we should be able to directly get

• The base symbol
• Atom
• Fragment

from symbols and the expression evaluation should be able to stop walking past variables.

[llvmbot]

Fixed by r249303.

Thanks!