ASan instrumentation should work with -O0

[GoogleCodeExporter]

What steps will reproduce the problem?

$ cat try.c
#include <stdio.h>
int main() { printf("Ok\n"); return(0); }

$ clang -o try -g3 -D__APPLE__ -faddress-sanitizer try.c
$ ./try
Ok==33560== 
CHECK failed: summary at asan_thread_registry.cc:99, pthread_self=0x7fff70989cc0
Segmentation fault

$ gdb ./try
(gdb) r
(gdb) bt
#0  0x0000000000000000 in ?? ()
#1  0x000000010000b3af in __asan::AsanThread::ClearShadowForThreadStack 
(this=0x10242a9a0) at asan_thread.cc:51
#2  0x000000010000b6b1 in __asan::AsanThread::~AsanThread (this=0x10242a9a0) at 
asan_thread.cc:45
#3  0x000000010000c3d0 in __asan::AsanThreadRegistry::~AsanThreadRegistry 
(this=0x10042a9a0) at asan_thread_registry.h:29
#4  0x000000010000fe15 in __tcf_0 () at asan_thread_registry.cc:24
#5  0x00007fff8640a374 in __cxa_finalize ()
#6  0x00007fff8640a28c in exit ()
#7  0x0000000100000f53 in start () at asan_thread_registry.h:28

(gdb) up
(gdb) p *this
$1 = {
  static kInvalidTid = -1, 
  summary_ = 0x0, 
  start_routine_ = 0, 
  arg_ = 0x0, 
  stack_top_ = 0, 
  stack_bottom_ = 0, 
  fake_stack_ = {
    static kMinStackFrameSizeLog = 9, 
    static kMaxStackFrameSizeLog = 16, 
    static kMaxStackMallocSize = 65536, 
    static kNumberOfSizeClasses = 8, 
    stack_size_ = 0, 
    alive_ = false, 
    allocated_size_classes_ = {0, 0, 0, 0, 0, 0, 0, 0}, 
...

$ clang --version
clang version 3.1 (trunk 144800)
Target: x86_64-apple-darwin10.8.0
Thread model: posix

The attached patch fixed it for me

Original issue reported on code.google.com by reini.urban on 22 Nov 2011 at 2:21

Attachments:

• protect-empty-prog.patch

[GoogleCodeExporter]

Original comment by konstant...@gmail.com on 22 Nov 2011 at 2:36

• Changed state: Accepted

[GoogleCodeExporter]

This happens because ASan does not currently work with -O0. In fact the module 
remains uninstrumented, thus __asan_init() is not called and the runtime 
library is in inconsistent state. This problem cannot be fixed without calling 
__asan_init(), so the proposed patch is invalid.

The short-term workaround is to pass an additional -O1 option to the compiler:
  $ clang try.c -faddress-sanitizer -o try -O1
  $ ./try
  Ok

A proper fix on the ASan side is to allow performing the instrumentation with 
-O0. ASan instrumentation is conceptually different from code optimization, so 
this is correct.

Original comment by ramosian.glider@gmail.com on 22 Nov 2011 at 8:18

• Changed title: ASan instrumentation should work with -O0
• Added labels: OpSys-All, Priority-High
• Removed labels: Priority-Medium

[GoogleCodeExporter]

Thanks. This explains why I found no single error in the Perl testsuite and a 
fair number of CPAN modules so far. Too good to be true. No single SIGSEGV. 
With -O1 it works fine

Original comment by reini.urban on 22 Nov 2011 at 2:18

[GoogleCodeExporter]

Attached is a patch for Clang that fixes the problem with -O0 (this is not a 
replacement for llvm/clang.patch, it should be applied separately)

Once someone on the LLVM-dev approves it, we'll commit it into Clang.

Original comment by ramosian.glider@gmail.com on 23 Nov 2011 at 8:11

Attachments:

• clang.patch

[GoogleCodeExporter]

Works for me, thanks. At least for the perl Configure probes.

Got now into other problems unrelated to this patch (fails with and without) to 
be handled in a separate ticket.

Original comment by reini.urban on 23 Nov 2011 at 4:41

[GoogleCodeExporter]

The O0 patch has landed. 
The build instructions updated. (note, they are unstable, nothing is guaranteed 
to work until we are done integrating with llvm). 

Original comment by konstant...@gmail.com on 1 Dec 2011 at 5:56

• Changed state: Fixed