Do not reinterpret to wrong alignment

This is undefined behavior in LLVM and can actually segfault on real hardware
including x86 in certain cases.

Also add missing endianess test since it is changed in this commit.

Author: yuyichao

base/bitarray.jl

@@ -283,11 +283,12 @@ function copy_to_bitarray_chunks!(Bc::Vector{UInt64}, pos_d::Int, C::Array{Bool}
     nc8 = (nc >>> 3) << 3
     if nc8 > 0
         ind8 = 1
-        C8 = reinterpret(UInt64, unsafe_wrap(Array, pointer(C, ind), nc8 << 6))
+        P8 = Ptr{UInt64}(pointer(C, ind)) # unaligned i64 pointer
         @inbounds for i = 1:nc8
             c = UInt64(0)
             for j = 0:7
-                c |= (pack8bools(C8[ind8]) << (j<<3))
+                # unaligned load
+                c |= (pack8bools(unsafe_load(P8, ind8)) << (j<<3))
                 ind8 += 1
             end
             Bc[bind] = c

base/dSFMT.jl

@@ -89,9 +89,13 @@ end
 
 # dSFMT jump
 function dsfmt_jump(s::DSFMT_state, jp::AbstractString)
-    index = s.val[end-1]
-    work = zeros(UInt64, JN32>>1)
-    dsfmt = reinterpret(UInt64, copy(s.val))
+    val = s.val
+    nval = length(val)
+    index = val[nval - 1]
+    work = zeros(UInt64, JN32 >> 1)
+    dsfmt = Vector{UInt64}(nval >> 1)
+    ccall(:memcpy, Ptr{Void}, (Ptr{UInt64}, Ptr{Int32}, Csize_t),
+          dsfmt, val, (nval - 1) * sizeof(Int32))
     dsfmt[end] = UInt64(N*2)
 
     for c in jp

base/io.jl

@@ -215,29 +215,33 @@ readlines(s=STDIN; chomp::Bool=true) = collect(eachline(s, chomp=chomp))
 
 ## byte-order mark, ntoh & hton ##
 
+let endian_boms = reinterpret(UInt8, UInt32[0x01020304])
+    global ntoh, hton, ltoh, htol
+    if endian_boms == UInt8[1:4;]
+        ntoh(x) = x
+        hton(x) = x
+        ltoh(x) = bswap(x)
+        htol(x) = bswap(x)
+        const global ENDIAN_BOM = 0x01020304
+    elseif endian_boms == UInt8[4:-1:1;]
+        ntoh(x) = bswap(x)
+        hton(x) = bswap(x)
+        ltoh(x) = x
+        htol(x) = x
+        const global ENDIAN_BOM = 0x04030201
+    else
+        error("seriously? what is this machine?")
+    end
+end
+
 """
     ENDIAN_BOM
 
 The 32-bit byte-order-mark indicates the native byte order of the host machine.
 Little-endian machines will contain the value `0x04030201`. Big-endian machines will contain
 the value `0x01020304`.
 """
-const ENDIAN_BOM = reinterpret(UInt32,UInt8[1:4;])[1]
-
-if ENDIAN_BOM == 0x01020304
-    ntoh(x) = x
-    hton(x) = x
-    ltoh(x) = bswap(x)
-    htol(x) = bswap(x)
-elseif ENDIAN_BOM == 0x04030201
-    ntoh(x) = bswap(x)
-    hton(x) = bswap(x)
-    ltoh(x) = x
-    htol(x) = x
-else
-    error("seriously? what is this machine?")
-end
-
+ENDIAN_BOM
 
 """
     ntoh(x)

base/pcre.jl

@@ -69,18 +69,18 @@ const OPTIONS_MASK = COMPILE_MASK | EXECUTE_MASK
 
 const UNSET = ~Csize_t(0)  # Indicates that an output vector element is unset
 
-function info(regex::Ptr{Void}, what::Integer, T)
-    buf = zeros(UInt8,sizeof(T))
+function info(regex::Ptr{Void}, what::Integer, ::Type{T}) where T
+    buf = Ref{T}()
     ret = ccall((:pcre2_pattern_info_8, PCRE_LIB), Int32,
-                (Ptr{Void}, Int32, Ptr{UInt8}),
+                (Ptr{Void}, Int32, Ptr{Void}),
                 regex, what, buf)
     if ret != 0
         error(ret == ERROR_NULL      ? "NULL regex object" :
               ret == ERROR_BADMAGIC  ? "invalid regex object" :
               ret == ERROR_BADOPTION ? "invalid option flags" :
                                        "unknown error $ret")
     end
-    reinterpret(T,buf)[1]
+    buf[]
 end
 
 function get_ovec(match_data)

test/core.jl

@@ -4848,12 +4848,15 @@ end
 let ni128 = sizeof(FP128test) ÷ sizeof(Int),
     ns128 = sizeof(FP128align) ÷ sizeof(Int),
     nbit = sizeof(Int) * 8,
-    arr = reinterpret(FP128align, collect(Int, 1:(2 * ns128))),
+    arr = Vector{FP128align}(2),
     offset = Base.datatype_alignment(FP128test) ÷ sizeof(Int),
     little,
-    expected
+    expected,
+    arrint = reinterpret(Int, arr)
+
+    @test length(arrint) == 2 * ns128
+    arrint .= 1:(2 * ns128)
     @test sizeof(FP128test) == 16
-    @test length(arr) == 2
     @test arr[1].i == 1
     @test arr[2].i == 1 + ns128
     expected = UInt128(0)

test/misc.jl

@@ -678,3 +678,26 @@ let foo() = begin
     end
     @test foo() == 2
 end
+
+# Endian tests
+# For now, we only support little endian.
+# Add an `Sys.ARCH` test for big endian when/if we add support for that.
+# Do **NOT** use `ENDIAN_BOM` to figure out the endianess
+# since that's exactly what we want to test.
+@test ENDIAN_BOM == 0x04030201
+@test ntoh(0x1) == 0x1
+@test hton(0x1) == 0x1
+@test ltoh(0x1) == 0x1
+@test htol(0x1) == 0x1
+@test ntoh(0x102) == 0x201
+@test hton(0x102) == 0x201
+@test ltoh(0x102) == 0x102
+@test htol(0x102) == 0x102
+@test ntoh(0x1020304) == 0x4030201
+@test hton(0x1020304) == 0x4030201
+@test ltoh(0x1020304) == 0x1020304
+@test htol(0x1020304) == 0x1020304
+@test ntoh(0x102030405060708) == 0x807060504030201
+@test hton(0x102030405060708) == 0x807060504030201
+@test ltoh(0x102030405060708) == 0x102030405060708
+@test htol(0x102030405060708) == 0x102030405060708