Merge pull request #301 from nh2/no-random_r-flagHEADmaster

Allow disabling random_r() usage manually.

2 files changed, 31 insertions, 10 deletions

diff --git a/packages/base/hmatrix.cabal b/packages/base/hmatrix.cabal
index 76e3825..4dc62e5 100644
--- a/packages/base/hmatrix.cabal
+++ b/packages/base/hmatrix.cabal
@@ -36,6 +36,11 @@ flag disable-default-paths
     default:        False
     manual: True
 
+flag no-random_r
+    description:    When enabled, don't depend on the random_r() C function.
+    default:        False
+    manual: True
+
 library
 
     Build-Depends:      base >= 4.9 && < 5,
@@ -99,6 +104,9 @@ library
         cc-options:     -msse2
 
 
+    if flag(no-random_r)
+        cc-options: -DNO_RANDOM_R
+
     if os(OSX)
         if flag(openblas)
             if !flag(disable-default-paths)
diff --git a/packages/base/src/Internal/C/vector-aux.c b/packages/base/src/Internal/C/vector-aux.c
index dcd6c0b..73455a5 100644
--- a/packages/base/src/Internal/C/vector-aux.c
+++ b/packages/base/src/Internal/C/vector-aux.c
@@ -932,20 +932,33 @@ int vectorScan(char * file, int* n, double**pp){
 
 ////////////////////////////////////////////////////////////////////////////////
 
-#if defined (__APPLE__) || (__FreeBSD__)
-/* FreeBSD and Mac OS X do not provide random_r(), thread safety cannot be
-   guaranteed.
+#if defined (__APPLE__) || (__FreeBSD__) || defined(NO_RANDOM_R) || defined(_WIN32) || defined(WIN32)
+/* Windows use thread-safe random
+   See: http://stackoverflow.com/questions/143108/is-windows-rand-s-thread-safe
+*/
+#if defined (__APPLE__) || (__FreeBSD__) || defined(NO_RANDOM_R)
+
+/* For FreeBSD, Mac OS X, and other libcs (like `musl`) that do not provide
+   random_r(), or if the use of random_r() is explicitly disabled, thread safety
+   cannot be guaranteed.
+   As per current understanding, this should at worst lead to less "random"
+   numbers being generated, in particular
+     * if another thread somebody calls lcong48() at the same time as nrand48()
+       is called
+     * in addition to that, for glibc with NO_RANDOM_R enabled when ndrand48()
+       is called for the first time by multiple threads in parallel due to the
+       initialisation function placed within it
+   See: http://www.evanjones.ca/random-thread-safe.html
+
    For FreeBSD and Mac OS X, nrand48() is much better than random().
    See: http://www.evanjones.ca/random-thread-safe.html
-*/
-#pragma message "randomVector is not thread-safe in OSX and FreeBSD"
-#endif
 
-#if defined (__APPLE__) || (__FreeBSD__) || defined(_WIN32) || defined(WIN32)
-/* Windows use thread-safe random
-   See: http://stackoverflow.com/questions/143108/is-windows-rand-s-thread-safe
+   TODO: As mentioned in the linked article, this could be fixed:
+         "the best solution for truly portable applications is to include
+          your own random number generator implementation,
+          and not rely on the system's C library".
 */
-#if defined (__APPLE__) || (__FreeBSD__)
+#pragma message "randomVector is not thread-safe in OSX and FreeBSD or with NO_RANDOM_R; this likely leads to less random numbers at worst; see http://www.evanjones.ca/random-thread-safe.html"
 
 inline double urandom() {
     /* the probalility of matching will be theoretically p^3(in fact, it is not)