PR34148: Do not assume we can use a copy relocation for an `external_weak` global

[rsmith]

Summary

An external_weak global may be intended to resolve as a null pointer if it's not defined, so it doesn't make sense to use a copy relocation for it.

See also https://reviews.llvm.org/D36280 (which fixes an overlapping issue for X86).

[• rafael]

LGTM as this fixes a bug, but even this I think is incomplete.

Thinking a bit more about it I see that I was wrong. Even if we have a hidden weak_external we should not be using lea. For example, the normal sequence for accessing a hidden variable when linked with bfd becomes

00000150 <_start>:
150: e8 00 00 00 00 call 155 <_start+0x5>
155: 58 pop %eax
156: 81 c0 ab 1e 00 00 add $0x1eab,%eax
15c: 8d 80 00 e0 ff ff lea -0x2000(%eax),%eax
162: c3 ret

The problem is that there are no relocations. This is resolving the undefined weak symbol to the load address, not to 0.

lib/Target/TargetMachine.cpp

       return true;
 
     bool IsTLS = GV && GV->isThreadLocal();
-    bool IsAccessViaCopyRelocs =
-        Options.MCOptions.MCPIECopyRelocations && GV && isa<GlobalVariable>(GV);
+    bool IsAccessViaCopyRelocs = Options.MCOptions.MCPIECopyRelocations && GV &&
+                                 isa<GlobalVariable>(GV) &&
+                                 !GV->hasExternalWeakLinkage();
     Triple::ArchType Arch = TT.getArch();
     bool IsPPC =
         Arch == Triple::ppc || Arch == Triple::ppc64 || Arch == Triple::ppc64le;

test/CodeGen/X86/global-access-pie-copyrelocs.ll

   ret i32 %0
 }
 
+; ExternalWeak Linkage
+@e = extern_weak global i32, align 4
+
+define i32* @my_access_global_d() #0 {
+; X32-LABEL: my_access_global_d:
+; X32:       addl $_GLOBAL_OFFSET_TABLE_{{.*}}, %eax
+; X32:       movl e@GOT(%eax), %eax
+; X64-LABEL: my_access_global_d:
+; X64:       movq e@GOTPCREL(%rip), %rax
+
+entry:
+  ret i32* @e
+}
+
 ; External Linkage, only declaration, store a value.
 
 define i32 @my_access_global_store_d() #0 {