llvm 3.2 regression

Project Infrastructure LLVM Dev List Archives
1

There seems to be a new regression in current llvm/compiler-rt 3.2 branch.
On 86_64-apple-darwin12, I am seeing the failure...

[100%] Running all regression tests
lit.py: lit.common.cfg:19: note: using clang: '/sw/src/fink.build/llvm32-3.2-0/llvm-3.2/build/bin/clang'
lit.py: lit.cfg:171: note: using clang: '/sw/src/fink.build/llvm32-3.2-0/llvm-3.2/build/bin/./clang'
FAIL: AddressSanitizer-Unit :: Release/AsanTest/AddressSanitizerMac.GCDSourceCancel (162 of 12631)
******************** TEST 'AddressSanitizer-Unit :: Release/AsanTest/AddressSanitizerMac.GCDSourceCancel' FAILED ********************
Note: Google Test filter = AddressSanitizerMac.GCDSourceCancel
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from AddressSanitizerMac
[ RUN ] AddressSanitizerMac.GCDSourceCancel
/sw/src/fink.build/llvm32-3.2-0/llvm-3.2/projects/compiler-rt/lib/asan/tests/asan_test.cc:2052: Failure
Death test: TestGCDSourceCancel()
    Result: died but not with expected error.
  Expected: Shadow byte and word
Actual msg:
[ DEATH ] =================================================================
[ DEATH ] ==88865== ERROR: AddressSanitizer: heap-buffer-overflow on address 0x00010cf70e4a at pc 0x10ab6a741 bp 0x10d4f3e30 sp 0x10d4f3e28
[ DEATH ] WRITE of size 1 at 0x00010cf70e4a thread T16777215
[ DEATH ] 0x00010cf70e4a is located 0 bytes to the right of 10-byte region [0x00010cf70e40,0x00010cf70e4a)
[ DEATH ] ==88865== AddressSanitizer CHECK failed: /sw/src/fink.build/llvm32-3.2-0/llvm-3.2/projects/compiler-rt/lib/asan/asan_report.cc:216 "((t)) != (0)" (0x0, 0x0)
[ DEATH ] Stats: 0M malloced (0M for red zones) by 2252 calls
[ DEATH ] Stats: 0M realloced by 0 calls
[ DEATH ] Stats: 0M freed by 1633 calls
[ DEATH ] Stats: 0M really freed by 0 calls
[ DEATH ] Stats: 4M (1027 full pages) mmaped in 8 calls
[ DEATH ] mmaps by size class: 7:4095; 8:2047; 9:1023; 10:511; 11:255; 12:128; 13:64; 14:32;
[ DEATH ] mallocs by size class: 7:1895; 8:202; 9:43; 10:79; 11:26; 12:2; 13:2; 14:3;
[ DEATH ] frees by size class: 7:1452; 8:38; 9:42; 10:77; 11:22; 13:2;
[ DEATH ] rfrees by size class:
[ DEATH ] Stats: malloc large: 5 small slow: 20
[ DEATH ]
[ FAILED ] AddressSanitizerMac.GCDSourceCancel (8742 ms)
[----------] 1 test from AddressSanitizerMac (8742 ms total)

[----------] Global test environment tear-down
[==========] 1 test from 1 test case ran. (8742 ms total)
[ PASSED ] 0 tests.
[ FAILED ] 1 test, listed below:
[ FAILED ] AddressSanitizerMac.GCDSourceCancel

1 FAILED TEST
  YOU HAVE 27 DISABLED TESTS

Can this be fixed before the 3.2 release?
           Jack

2

Can this be fixed before the 3.2 release?

Given that release is tomorrow, then - no, unfortunately.

3

Well if you can't fix the bug by tomorrow, can't you at least revert the
failing part of...

https://llvm.org/viewvc/llvm-project/compiler-rt/trunk/lib/asan/tests/asan_test.cc?sortby=log&r1=145953&r2=145952&pathrev=145953

with...

Index: lib/asan/tests/asan_test.cc

4

+glider

Weird, we haven’t seen problems with this test for quite a while.

5

Neither we have them now.
Unfortunately disabling the test in the 3.2 branch won't fix the
problem, and the actual fix may require cherry-picking a number of
recent CLs, which may be rather intrusive.

6

Neither we have them now.

Alexander,
   Do you mean to say that you can't reproduce this failure locally on x86_64 darwin?
I am testing with current llvm 3.2 branches for llvm/compiler-rt/clang/polly/testsuite
against Xcode 4.5.2 on x86_64-apple-darwin12 using a cmake build with...

-DLLVM_ENABLE_ASSERTIONS=OFF -DCMAKE_BUILD_TYPE=Release ..

and have CC set to clang and CXX to clang++.
               Jack