Memory sanitizer support revisited

[Keno]

With a new, stable LLVM patch for thread local storage support, I decided to revisit memory sanitizer support for julia. With this branch and a few upstream commits which are currently pending review, I am able to run julia's test suite under msan, without having msan report any failures (though there was a stream error which may or may not be related):

   JULIA test/all
    From worker 14:      * linalg/lu            in  19.81 seconds
    From worker 18:      * keywordargs          in  39.09 seconds
    From worker 12:      * linalg/givens        in  90.51 seconds
    From worker 12:      * remote               in   7.42 seconds
    From worker 12:      * iobuffer             in  27.20 seconds
    From worker 12:      * staged               in  20.33 seconds
    From worker 6:       * linalg/lapack        in 178.97 seconds
    From worker 5:       * linalg4              in 183.56 seconds
    From worker 14:      * dict                 in 181.65 seconds
    From worker 6:       * tuple                in  35.45 seconds
    From worker 8:       * linalg/tridiag       in 231.06 seconds
    From worker 18:      * hashing              in 208.93 seconds
    From worker 11:      * linalg/pinv          in 256.57 seconds
    From worker 18:      * intfuncs             in  44.19 seconds
    From worker 11:      * simdloop             in  44.31 seconds
    From worker 15:      * linalg/arnoldi       in 324.30 seconds
    From worker 14:      * reduce               in 128.32 seconds
    From worker 16:      * linalg/symmetric     in 331.67 seconds
    From worker 17:      * core                 in 336.32 seconds
    From worker 17:      * copy                 in  36.32 seconds
    From worker 10:      * linalg/diagonal      in 413.32 seconds
    From worker 11:      * fft                  in 121.43 seconds
    From worker 18:      * blas                 in 134.04 seconds
    From worker 18:      * operators            in  11.05 seconds
    From worker 18:      * path                 in   2.18 seconds
    From worker 11:      * functional           in  39.98 seconds
    From worker 15:      * dsp                  in 147.04 seconds
    From worker 9:       * linalg/bidiag        in 506.74 seconds
    From worker 20:      * strings              in 511.06 seconds
    From worker 18:      * ccall                in  79.63 seconds
    From worker 18:      * backtrace            in   6.94 seconds
    From worker 6:       * reducedim            in 329.53 seconds
    From worker 20:      * spawn                      [stdio passthrough ok]
    From worker 18:      * priorityqueue        in  32.88 seconds
    From worker 11:      * bigint               in 116.71 seconds
    From worker 4:       * linalg3              in 611.40 seconds
    From worker 18:      * version              in  55.95 seconds
    From worker 10:      * fastmath             in 215.24 seconds
    From worker 8:       * random               in 400.20 seconds
    From worker 4:       * pollfd               in  30.37 seconds
    From worker 6:       * file                 in 101.88 seconds
    From worker 20:  in 145.61 seconds
    From worker 6:       * floatapprox          in  19.65 seconds
The following 'Returned code...' warnings indicate normal behavior:
Warning: Returned code may not match what actually runs.
Warning: Returned code may not match what actually runs.
Warning: Returned code may not match what actually runs.
Warning: Returned code may not match what actually runs.
    From worker 6:       * reflection           in  24.44 seconds
    From worker 6:       * regex                in  19.87 seconds
    From worker 4:       * socket               in  67.92 seconds
    From worker 17:      * math                 in 362.75 seconds
    From worker 11:      * resolve              in 163.58 seconds
    From worker 17:      * sysinfo              in  13.89 seconds
    From worker 6:       * float16              in  69.50 seconds
    From worker 11:      * rounding             in  43.59 seconds
    From worker 6:       * mod2pi               in  16.28 seconds
    From worker 9:       * statistics           in 309.23 seconds
    From worker 4:       * combinatorics        in 111.36 seconds
    From worker 11:      * euler                in  42.55 seconds
    From worker 20:      * readdlm              in 203.69 seconds
    From worker 8:       * complex              in 236.80 seconds
    From worker 6:       * show                 in  81.90 seconds
    From worker 20:      * replutil             in  26.62 seconds
    From worker 20:      * goto                 in   3.87 seconds
    From worker 20:      * llvmcall             in   4.51 seconds
    From worker 6:       * test                 in  21.35 seconds
    From worker 9:       * lineedit             in 116.17 seconds
    From worker 9:       * meta                 in  14.65 seconds
    From worker 13:      * linalg/cholesky      in 944.91 seconds
    From worker 13:      * libgit2              in   1.26 seconds
    From worker 13:      * docs                 in   4.92 seconds
    From worker 18:      * mpfr                 in 353.37 seconds
    From worker 4:       * replcompletions      in 148.67 seconds
    From worker 4:       * parser               in  12.80 seconds
    From worker 18:      * base64               in  15.39 seconds
    From worker 9:       * profile              in  46.44 seconds
    From worker 8:       * sets                 in 142.41 seconds
    From worker 11:      * repl                 in 181.91 seconds
    From worker 9:       * char                 in  33.61 seconds
    From worker 18:      * functors             in  59.25 seconds
    From worker 6:       * nullable             in 145.81 seconds
    From worker 18:      * i18n                 in   4.07 seconds
    From worker 8:       * misc                 in  45.78 seconds
    From worker 10:      * broadcast            in 427.78 seconds
    From worker 13:      * markdown             in 126.58 seconds
    From worker 4:       * serialize            in 132.69 seconds
    From worker 12:      * arrayops             in 987.30 seconds
    From worker 18:      * unicode              in 126.44 seconds
    From worker 11:      * enums                in 197.65 seconds
    From worker 20:      * grisu                in 327.78 seconds
    From worker 21:      * dates                in 1228.74 seconds
    From worker 3:       * linalg2              in 1249.12 seconds
ArgumentError("stream is closed or unusable")
    From worker 6:       * examples             in 228.83 seconds
WARNING: Forcibly interrupting busy workers
WARNING: rmprocs: process 1 not removed
    From worker 2:       * linalg1              in 1347.09 seconds
    From worker 17:      * ranges               in 645.29 seconds
    From worker 16:      * bitarray             in 1109.57 seconds
    From worker 19:      * numbers              in 1452.90 seconds
    From worker 9:       * cmdlineargs          in 659.74 seconds
    From worker 15:      * sorting              in 1231.75 seconds
    From worker 14:      * sparse               in 1521.88 seconds
    From worker 5:       * subarray             in 4307.21 seconds
    From worker 7:       * linalg/triangular    in 5029.93 seconds
``

[StefanKarpinski]

Nice. This should be very helpful for stability.

[tkelman]

should this only apply conditionally?

[Keno]

oops. Left over debugging flag. Thanks for catching that. Should be deleted.