Fix thread local rng implementation in C so it's correctly unbiased and actually random.

Author: gbaraldi

src/julia_internal.h

@@ -1308,18 +1308,31 @@ JL_DLLEXPORT size_t jl_maxrss(void);
 
 STATIC_INLINE uint64_t cong(uint64_t max, uint64_t *seed) JL_NOTSAFEPOINT
 {
-    if (max == 0)
+    if (max < 2)
         return 0;
     uint64_t mask = ~(uint64_t)0;
-    --max;
-    mask >>= __builtin_clzll(max|1);
-    uint64_t x;
+    int zeros = __builtin_clzll(max);
+    int bits = CHAR_BIT * sizeof(uint64_t) - zeros;
+    mask = mask >> zeros;
     do {
-        *seed = 69069 * (*seed) + 362437;
-        x = *seed & mask;
-    } while (x > max);
-    return x;
+        uint64_t value = 69069 * (*seed) + 362437;
+        *seed = value;
+        uint64_t x = value & mask;
+        if (x < max) {
+			return x;
+		}
+        int bits_left = zeros;
+		while (bits_left >= bits) {
+			value >>= bits;
+			x = value & mask;
+			if (x < max) {
+				return x;
+			}
+			bits_left -= bits;
+		}
+    } while (1);
 }
+
 JL_DLLEXPORT uint64_t jl_rand(void) JL_NOTSAFEPOINT;
 JL_DLLEXPORT void jl_srand(uint64_t) JL_NOTSAFEPOINT;
 JL_DLLEXPORT void jl_init_rand(void);