https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64653
Bug ID: 64653
Summary: Incomplete emission of D1/D2 destructors when
optimising -- link failure with gold
Product: gcc
Version: 5.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: thiago at kde dot org
Created attachment 34470
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=34470&action=edit
Test project
When the attached project is built with GCC 5 (x86-64, SVN build from 219682)
and linked with gold, the linker exits with:
/usr/bin/ld.gold: error: src2.o: requires dynamic R_X86_64_PC32 reloc against
'_ZN5QListI7QStringED1Ev' which may overflow at runtime; recompile with -fPIC
src2.cpp:3: warning: relocation refers to discarded section
In other words: the linker wants to use a symbol that is present in the same .o
file itself, but it has discarded that section. Since now the symbol is
out-of-library, it prints an error about the relocation.
It appears that the issue is highly dependent on optimisations applied to
src1.cpp, which is why I am also attaching the generated assembly by my version
of GCC.
If I compare that generated assembly with the one generated by GCC 4.9, the
following diff stands out:
--- src1-4.9.s 2015-01-17 17:31:02.254643092 -0800
+++ src1-5.0.s 2015-01-17 17:30:50.795714657 -0800
@@ -38,9 +38,6 @@
.LCOLDE0:
.section
.text._ZN5QListI7QStringED2Ev,"axG",@progbits,_ZN5QListI7QStringED5Ev,comdat
.LHOTE0:
- .weak _ZN5QListI7QStringED1Ev
- .hidden _ZN5QListI7QStringED1Ev
- .set _ZN5QListI7QStringED1Ev,_ZN5QListI7QStringED2Ev
.section .text.unlikely,"ax",@progbits
In other words, the compiler failed to emit the symbol for the D1 destructor in
src1.o. In this file, GCC emitted calls to the "D2" destructor only.
If we look at src2.o, the compiler emitted a call to the "D1" destructor, not
to "D2". It did emit the correct section with the "D1" symbol, but gold
discarded this section, as it chose the one from src1.o. The regular ld.bfd
does not seem to have a problem.
It's not apparent to me why GCC chose D1 over D2 or vice-versa. This class has
no base, let alone a virtual base, so both destructors are identical.
Clang 3.5 and 3.6 emit only the "D2" destructor; it doesn't seem to emit D1 at
all, not even as an alias, in either source file. I expected the same problem
to occur if I compiled src1 with clang and src2 with GCC, but there was no
linker error. The only other difference was that Clang did not emit the D5
symbol.
ICC emits both and calls "D1" in src1.o (under -O0) and emits D1 only in
src2.o, calling it. D5 is also missing.
