raw-SHA512-free-opencl and XSHA512-free-opencl sometimes fail self-test

On a build made with devtoolset-3 and --enable-asan:

[solar@super run]$ ./john -te -form=raw-SHA512-free-opencl -dev=1
NOTE: This is a debug build, speed will be lower than normal
Device 1:        Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GHz
Benchmarking: raw-SHA512-free-opencl [SHA512 OpenCL (inefficient, development use mostly)]... Binary Build log: Linking started
Linking done
Kernel <kernel_sha512> was not vectorized
Kernel <kernel_cmp> was successfully vectorized
Done.
LWS=256 GWS=4194304 (16384 blocks) DONE
Raw:    7913K c/s real, 653572 c/s virtual

[solar@super run]$ ./john -te -form=XSHA512-free-opencl -dev=1
NOTE: This is a debug build, speed will be lower than normal
Device 1:        Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GHz
Benchmarking: XSHA512-free-opencl, Mac OS X 10.7+ [SHA512 OpenCL (efficient at "many salts" only)]... Binary Build log: Linking started
Linking done
Kernel <kernel_xsha512> was not vectorized
Kernel <kernel_cmp> was successfully vectorized
Done.
FAILED (cmp_all(23))
[solar@super run]$ ./john -te -form=XSHA512-free-opencl -dev=1
NOTE: This is a debug build, speed will be lower than normal
Device 1:        Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GHz
Benchmarking: XSHA512-free-opencl, Mac OS X 10.7+ [SHA512 OpenCL (efficient at "many salts" only)]... Binary Build log: Linking started
Linking done
Kernel <kernel_xsha512> was not vectorized
Kernel <kernel_cmp> was successfully vectorized
Done.
LWS=128 GWS=2097152 (16384 blocks) DONE
Warning: "Many salts" test limited: 21/256
Many salts:     21275K c/s real, 679947 c/s virtual
Only one salt:  8346K c/s real, 640351 c/s virtual

[solar@super run]$ ./john -te -form=raw-SHA512-free-opencl -dev=1
NOTE: This is a debug build, speed will be lower than normal
Device 1:        Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GHz
Benchmarking: raw-SHA512-free-opencl [SHA512 OpenCL (inefficient, development use mostly)]... Binary Build log: Linking started
Linking done
Kernel <kernel_sha512> was not vectorized
Kernel <kernel_cmp> was successfully vectorized
Done.
FAILED (cmp_all(43))
[solar@super run]$ ./john -te -form=raw-SHA512-free-opencl -dev=1
NOTE: This is a debug build, speed will be lower than normal
Device 1:        Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GHz
Benchmarking: raw-SHA512-free-opencl [SHA512 OpenCL (inefficient, development use mostly)]... Binary Build log: Linking started
Linking done
Kernel <kernel_sha512> was not vectorized
Kernel <kernel_cmp> was successfully vectorized
Done.
LWS=256 GWS=2097152 (8192 blocks) DONE
Raw:    7951K c/s real, 652302 c/s virtual

Also:

./john -test -v=5 -form=opencl -dev=1 2>&1 | tee -a log1

resulted in:

Benchmarking: raw-SHA512-free-opencl [SHA512 OpenCL (inefficient, development use mostly)]... Loaded 9 hashes with 1 different salts to test db from test vectors
Options used: -I opencl -cl-mad-enable -D__CPU__ -DDEVICE_INFO=33 -D__SIZEOF_HOST_SIZE_T__=8 -DDEV_VER_MAJOR=1 -DDEV_VER_MINOR=2 -D_OPENCL_COMPILER -DPLAINTEXT_LENGTH=20 $JOHN/opencl/sha512_kernel.cl
Build log: Compilation started
Compilation done
Linking started
Linking done
Kernel <kernel_sha512> was not vectorized
Kernel <kernel_cmp> was successfully vectorized
Done.
binary size 5114
LWS=7 GWS=49 (7 blocks) FAILED (cmp_all(14) $SHA512$fa585d89c851dd338a70dcf535aa2a92fee7836dd6aff1226583e88e0996293f16bc009c652826e0fc5c706695a03cddce372f139eff4d13959da6f1f5d3eabe)

and

Benchmarking: XSHA512-free-opencl, Mac OS X 10.7+ [SHA512 OpenCL (efficient at "many salts" only)]... Loaded 18 hashes with 16 different salts to test db from test vectors
Options used: -I opencl -cl-mad-enable -D__CPU__ -DDEVICE_INFO=33 -D__SIZEOF_HOST_SIZE_T__=8 -DDEV_VER_MAJOR=1 -DDEV_VER_MINOR=2 -D_OPENCL_COMPILER -DPLAINTEXT_LENGTH=20 -DSALT_SIZE=4 $JOHN/opencl/xsha512_kernel.cl
Build log: Compilation started
Compilation done
Linking started
Linking done
Kernel <kernel_xsha512> was not vectorized
Kernel <kernel_cmp> was successfully vectorized
Done.
binary size 5214
LWS=7 GWS=49 (7 blocks) FAILED (cmp_all(8) $LION$52396836b22e1966f14f090fc611ed99916992d6e03bffa86fe77a4993bd0952e706c13acc34edefa97a1dee885c149b34c27b8b4f5b3b611d9e739833b21c5cf772e9e7)

BTW, with formats failing during auto-tune like this, there's no empty line before the next format's "Benchmarking ..." line, which makes it hard to notice the failures. I only noticed from the final:

2 out of 89 tests have FAILED

Start by reproducing the failures with the reported -test commands and device settings, then inspect opencl/sha512_kernel.cl and opencl/xsha512_kernel.cl. Compare the intermittent self-test results and auto-tune output with passing formats; done means both formats pass reliably and failures remain clearly reported.