[ASan] Tentative fix for 'non-ABS relocation R_386_GOTOFF' in 'debug_*'

[hctim]

Summary

Looks like GCC used to create non-PC relative relocations in debug
sections, which is invalid.

Seems like a well known problem:

• https://reviews.llvm.org/D43351

Unfortunately some of the tests were recently flipped to -fuse-ld=lld,
and this broke the bots. This is a tentative fix for:

• http://lab.llvm.org:8011/#/builders/37/builds/1109/steps/9/logs/stdio

This should probably be fixed in LLD, to make LLD be permissive of
GOTOFF relocations like the GOTPC relocation fix upstream. However, it's
hard for me to test this as the earliest gcc I have access to is 8.4.0,
which doesn't seem to exhibit these problems.

[hctim]

(update description)

[MaskRay]

non-PC relative relocations in debug sections, which is invalid.

non-absolute relocations in debug sections, which is invalid

https://reviews.llvm.org/D43351

LLD worked around the R_386_GOTPC problem in D43351 but it seems that incorrect R_386_GOTOFF can be produced as well. It may or may not be the same as https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82630

As chatted offline, it'd be great to have a translation unit to exhibit the R_386_GOTPC in .rel.debug_* problem and follow up on https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82630 asking whether it has been fixed in a newer GCC 8 release.

[hctim]

(reverted 140808768d3e5c0f4e52dd42094650f5d282e34a in 1d03a54d9460fa122f0be9bf9018b9a4358745a9 as discussed offline)

compiler-rt/test/asan/TestCases/Linux/global-overflow-lld.cpp

 // Test that gc-sections-friendly instrumentation of globals does not introduce
 // false negatives with the LLD linker.
 // RUN: %clangxx_asan -fuse-ld=lld -Wl,-gc-sections -ffunction-sections -fdata-sections -O0 %s -o %t && not %run %t 2>&1 | FileCheck %s
-// REQUIRES: lld
+// REQUIRES: lld-available
+
+// Looks like GCC ended up creating PIC relocations in the debug section for a
+// while. Mark these tests as unsupported on the broken versions, as they're
+// using lld which rejects such constructs.
+// UNSUPPORTED: gcc-4, gcc-5, gcc-6, gcc-7, gcc-8
 
 #include <string.h>
 int main(int argc, char **argv) {
   static char XXX[10];
   static char YYY[10];
   static char ZZZ[10];
   memset(XXX, 0, 10);
   memset(YYY, 0, 10);
   memset(ZZZ, 0, 10);
   int res = YYY[argc * 10];  // BOOOM
   // CHECK: {{READ of size 1 at}}
   // CHECK: {{located 0 bytes to the right of global variable}}
   res += XXX[argc] + ZZZ[argc];
   return res;
 }

compiler-rt/test/asan/TestCases/Linux/globals-gc-sections-lld.cpp

 /// Without instrumenting globals, --gc-sections drops the undefined symbol.
 // RUN: %clangxx_asan %s -o /dev/null -Wl,--gc-sections -fuse-ld=lld -ffunction-sections -fdata-sections -mllvm -asan-globals=0
 /// With -fsanitize-address-globals-dead-stripping and -fdata-sections, a garbage
 /// collectable custom metadata section is used for instrumented globals.
 // RUN: %clangxx_asan %s -o /dev/null -Wl,--gc-sections -fuse-ld=lld -ffunction-sections -fdata-sections -fsanitize-address-globals-dead-stripping
 
 // https://github.com/google/sanitizers/issues/260
 // REQUIRES: lld-available
+
+// Looks like GCC ended up creating PIC relocations in the debug section for a
+// while. Mark these tests as unsupported on the broken versions, as they're
+// using lld which rejects such constructs.
+// UNSUPPORTED: gcc-4, gcc-5, gcc-6, gcc-7, gcc-8
 int undefined();
 
 // On i386 clang adds --export-dynamic when linking with ASan, which adds all
 // non-hidden globals to GC roots.
 __attribute__((visibility("hidden"))) int (*unused)() = undefined;
 
 int main() {
         return 0;
 }