diff --git a/Makefile b/Makefile
index bc79b8b3..fa458470 100644
--- a/Makefile
+++ b/Makefile
@@ -9,6 +9,13 @@ include pkgconfig.mk	# package version
 
 LIBNAME = unicorn
 UNAME_S := $(shell uname -s)
+# SMP_MFLAGS is used for controlling the amount of parallelism used
+# in external 'make' invocations. If the user doesn't override it, it
+# does "-j4". That is, it uses 4 job threads. If you want to use more or less,
+# pass in a different -jX, with X being the number of threads.
+# For example, to completely disable parallel building, pass "-j1".
+# If you want to use 16 job threads, use "-j16".
+SMP_MFLAGS := -j4
 
 GENOBJ = $(shell find qemu/$(1) -name "*.o" 2>/dev/null) 
 
@@ -204,7 +211,7 @@ qemu/config-host.h-timestamp:
 	cd qemu && \
 	./configure --cc="${CC}" --extra-cflags="$(UNICORN_CFLAGS)" --target-list="$(UNICORN_TARGETS)" ${UNICORN_QEMU_FLAGS}
 	printf "$(UNICORN_ARCHS)" > config.log
-	$(MAKE) -C qemu -j 4
+	$(MAKE) -C qemu $(SMP_MFLAGS)
 	$(eval UC_TARGET_OBJ += $$(wildcard qemu/util/*.o) $$(wildcard qemu/*.o) $$(wildcard qemu/qom/*.o) $$(wildcard qemu/hw/core/*.o) $$(wildcard qemu/qapi/*.o) $$(wildcard qemu/qobject/*.o))
 
 unicorn: $(LIBRARY) $(ARCHIVE)
diff --git a/bindings/Makefile b/bindings/Makefile
index 14e88df8..1e3c179a 100644
--- a/bindings/Makefile
+++ b/bindings/Makefile
@@ -42,7 +42,7 @@ python:
 	$(ENV_VARS) python python/$(@:%.txt=%) > $@
 
 %.py.test: %.c.txt %.py.txt
-	$(DIFF) $(@:%.py.test=%.c.txt) $(@:%.py.test=%.py.txt)
+	$(DIFF) -u $(@:%.py.test=%.c.txt) $(@:%.py.test=%.py.txt)
 
 clean:
 #	rm -rf *.txt
diff --git a/bindings/dotnet/UnicornManaged/Const/X86.fs b/bindings/dotnet/UnicornManaged/Const/X86.fs
index d0526f87..2dc3f61f 100644
--- a/bindings/dotnet/UnicornManaged/Const/X86.fs
+++ b/bindings/dotnet/UnicornManaged/Const/X86.fs
@@ -257,7 +257,8 @@ module X86 =
     let UC_X86_REG_TR = 245
     let UC_X86_REG_FPCW = 246
     let UC_X86_REG_FPTAG = 247
-    let UC_X86_REG_ENDING = 248
+    let UC_X86_REG_MSR = 248
+    let UC_X86_REG_ENDING = 249
 
     // X86 instructions
 
diff --git a/bindings/go/unicorn/x86_const.go b/bindings/go/unicorn/x86_const.go
index ef18c08e..bd47e3e2 100644
--- a/bindings/go/unicorn/x86_const.go
+++ b/bindings/go/unicorn/x86_const.go
@@ -252,7 +252,8 @@ const (
 	X86_REG_TR = 245
 	X86_REG_FPCW = 246
 	X86_REG_FPTAG = 247
-	X86_REG_ENDING = 248
+	X86_REG_MSR = 248
+	X86_REG_ENDING = 249
 
 // X86 instructions
 
diff --git a/bindings/haskell/.gitignore b/bindings/haskell/.gitignore
index 3b5b7859..9966d825 100644
--- a/bindings/haskell/.gitignore
+++ b/bindings/haskell/.gitignore
@@ -21,3 +21,4 @@ SampleMips
 SampleSparc
 SampleX86
 Shellcode
+SampleBatchReg
diff --git a/bindings/haskell/samples/SampleBatchReg.hs b/bindings/haskell/samples/SampleBatchReg.hs
new file mode 100644
index 00000000..9b40b3f5
--- /dev/null
+++ b/bindings/haskell/samples/SampleBatchReg.hs
@@ -0,0 +1,99 @@
+import Unicorn
+import Unicorn.Hook
+import qualified Unicorn.CPU.X86 as X86
+
+import Control.Monad.Trans.Class (lift)
+import qualified Data.ByteString as BS
+import Data.Int
+import Data.List (intercalate)
+import Data.Word
+import qualified Numeric as N (showHex)
+import System.IO (hPutStrLn, stderr)
+
+syscallABI :: [X86.Register]
+syscallABI = [ X86.Rax
+             , X86.Rdi
+             , X86.Rsi
+             , X86.Rdx
+             , X86.R10
+             , X86.R8
+             , X86.R9
+             ]
+
+vals :: [Int64]
+vals = [ 200
+       , 10
+       , 11
+       , 12
+       , 13
+       , 14
+       , 15
+       ]
+
+ucPerror :: Error
+         -> IO ()
+ucPerror err =
+    hPutStrLn stderr $ "Error " ++ ": " ++ strerror err
+
+base :: Word64
+base = 0x10000
+
+-- mov rax, 100; mov rdi, 1; mov rsi, 2; mov rdx, 3; mov r10, 4; mov r8, 5; mov r9, 6; syscall
+code :: BS.ByteString
+code = BS.pack [ 0x48, 0xc7, 0xc0, 0x64, 0x00, 0x00, 0x00, 0x48, 0xc7, 0xc7
+               , 0x01, 0x00, 0x00, 0x00, 0x48, 0xc7, 0xc6, 0x02, 0x00, 0x00
+               , 0x00, 0x48, 0xc7, 0xc2, 0x03, 0x00, 0x00, 0x00, 0x49, 0xc7
+               , 0xc2, 0x04, 0x00, 0x00, 0x00, 0x49, 0xc7, 0xc0, 0x05, 0x00
+               , 0x00, 0x00, 0x49, 0xc7, 0xc1, 0x06, 0x00, 0x00, 0x00, 0x0f
+               , 0x05
+               ]
+
+-- Pretty-print integral as hex
+showHex :: (Integral a, Show a) => a -> String
+showHex i =
+    N.showHex (fromIntegral i :: Word64) ""
+
+-- Write a string (with a newline character) to standard output in the emulator
+emuPutStrLn :: String -> Emulator ()
+emuPutStrLn =
+    lift . putStrLn
+
+hookSyscall :: SyscallHook ()
+hookSyscall uc _ = do
+    runEmulator $ do
+        readVals <- regReadBatch uc syscallABI
+        emuPutStrLn $ "syscall: {"
+                      ++ intercalate ", " (map show readVals)
+                      ++ "}"
+    return ()
+
+hookCode :: CodeHook ()
+hookCode _ addr size _ = do
+    putStrLn $ "HOOK_CODE: 0x" ++ showHex addr ++ ", 0x" ++
+               maybe "0" showHex size
+
+main :: IO ()
+main = do
+    result <- runEmulator $ do
+        uc <- open ArchX86 [Mode64]
+
+        -- regWriteBatch
+        emuPutStrLn "regWriteBatch {200, 10, 11, 12, 13, 14, 15}"
+        regWriteBatch uc syscallABI vals
+
+        readVals <- regReadBatch uc syscallABI
+    
+        emuPutStrLn $ "regReadBatch = {"
+                      ++ intercalate ", " (map show readVals)
+                      ++ "}"
+
+        -- syscall
+        emuPutStrLn "running syscall shellcode"
+        syscallHookAdd uc hookSyscall () 1 0
+        memMap uc base (0x1000) [ProtAll]
+        memWrite uc base code
+        let codeLen = fromIntegral $ BS.length code
+        start uc base (base + codeLen) Nothing Nothing
+    case result of
+        Right _  -> return ()
+        Left err -> ucPerror err
diff --git a/bindings/haskell/samples/SampleX86.hs b/bindings/haskell/samples/SampleX86.hs
index e395ed75..5f4640e6 100644
--- a/bindings/haskell/samples/SampleX86.hs
+++ b/bindings/haskell/samples/SampleX86.hs
@@ -98,7 +98,7 @@ hookCode :: CodeHook ()
 hookCode uc addr size _ = do
     runEmulator $ do
         emuPutStrLn $ ">>> Tracing instruction at 0x" ++ showHex addr ++
-                      ", instruction size = 0x" ++ (maybe "0" showHex size)
+                      ", instruction size = 0x" ++ maybe "0" showHex size
 
         eflags <- regRead uc X86.Eflags
         emuPutStrLn $ ">>> --- EFLAGS is 0x" ++ showHex eflags
diff --git a/bindings/haskell/src/Unicorn.hs b/bindings/haskell/src/Unicorn.hs
index 19c07f8c..a605015d 100644
--- a/bindings/haskell/src/Unicorn.hs
+++ b/bindings/haskell/src/Unicorn.hs
@@ -25,6 +25,8 @@ module Unicorn
       -- * Register operations
     , regWrite
     , regRead
+    , regWriteBatch
+    , regReadBatch
 
       -- * Memory operations
     , MemoryPermission(..)
@@ -140,13 +142,11 @@ stop uc = do
 -- | Write to register.
 regWrite :: Reg r
          => Engine      -- ^ 'Unicorn' engine handle
-         -> r           -- ^ Register ID to write to
+         -> r           -- ^ Register to write to
          -> Int64       -- ^ Value to write to register
          -> Emulator () -- ^ An 'Error' on failure
-regWrite uc regId value = do
-    err <- lift . alloca $ \ptr -> do
-        poke ptr value
-        ucRegWrite uc regId ptr
+regWrite uc reg value = do
+    err <- lift $ ucRegWrite uc reg value
     if err == ErrOk then
         right ()
     else
@@ -155,16 +155,49 @@ regWrite uc regId value = do
 -- | Read register value.
 regRead :: Reg r
         => Engine           -- ^ 'Unicorn' engine handle
-        -> r                -- ^ Register ID to read from
+        -> r                -- ^ Register to read from
         -> Emulator Int64   -- ^ The value read from the register on success,
                             -- or an 'Error' on failure
-regRead uc regId = do
-    (err, val) <- lift $ ucRegRead uc regId
+regRead uc reg = do
+    (err, val) <- lift $ ucRegRead uc reg
     if err == ErrOk then
         right val
     else
         left err
 
+-- | Write multiple register values.
+regWriteBatch :: Reg r
+              => Engine         -- ^ 'Unicorn' engine handle
+              -> [r]            -- ^ List of registers to write to
+              -> [Int64]        -- ^ List of values to write to the registers
+              -> Emulator ()    -- ^ An 'Error' on failure
+regWriteBatch uc regs vals = do
+    err <- lift $ ucRegWriteBatch uc regs vals (length regs)
+    if err == ErrOk then
+        right ()
+    else
+        left err
+
+-- | Read multiple register values.
+regReadBatch ::  Reg r
+             => Engine              -- ^ 'Unicorn' engine handle
+             -> [r]                 -- ^ List of registers to read from
+             -> Emulator [Int64]    -- ^ A list of register values on success,
+                                    -- or an 'Error' on failure
+regReadBatch uc regs = do
+    -- Allocate an array of the given size
+    let size = length regs
+    result <- lift . allocaArray size $ \array -> do
+        err <- ucRegReadBatch uc regs array size
+        if err == ErrOk then
+            -- If ucRegReadBatch completed successfully, pack the contents of
+            -- the array into a list and return it
+            liftM Right (peekArray size array)
+        else
+            -- Otherwise return the error
+            return $ Left err
+    hoistEither result
+
 -------------------------------------------------------------------------------
 -- Memory operations
 -------------------------------------------------------------------------------
@@ -190,12 +223,12 @@ memRead :: Engine                       -- ^ 'Unicorn' engine handle
                                         -- an 'Error' on failure
 memRead uc address size = do
     -- Allocate an array of the given size
-    result <- lift . allocaArray size $ \ptr -> do
-        err <- ucMemRead uc address ptr size
+    result <- lift . allocaArray size $ \array -> do
+        err <- ucMemRead uc address array size
         if err == ErrOk then
             -- If ucMemRead completed successfully, pack the contents of the
             -- array into a ByteString and return it
-            liftM (Right . pack) (peekArray size ptr)
+            liftM (Right . pack) (peekArray size array)
         else
             -- Otherwise return the error
             return $ Left err
diff --git a/bindings/haskell/src/Unicorn/Internal/Hook.chs b/bindings/haskell/src/Unicorn/Internal/Hook.chs
index affed872..261adf12 100644
--- a/bindings/haskell/src/Unicorn/Internal/Hook.chs
+++ b/bindings/haskell/src/Unicorn/Internal/Hook.chs
@@ -314,7 +314,7 @@ marshalMemoryHook memoryHook =
             maybeValue = case memAccess of
                              MemRead  -> Nothing
                              MemWrite -> Just $ fromIntegral value
-                             _        -> undefined  -- XX Handle this?
+                             _        -> error "Invalid memory access"
         memoryHook uc memAccess address (fromIntegral size) maybeValue userData
 
 -- | Callback function for hooking memory reads.
@@ -390,7 +390,7 @@ marshalMemoryEventHook eventMemoryHook =
                              MemReadProt      -> Nothing
                              MemWriteUnmapped -> Just $ fromIntegral value
                              MemWriteProt     -> Just $ fromIntegral value
-                             _                -> undefined  -- XX Handle this?
+                             _                -> error "Invalid memory access"
         res <- eventMemoryHook uc memAccess address (fromIntegral size)
                                maybeValue userData
         return $ boolToInt res
diff --git a/bindings/haskell/src/Unicorn/Internal/Unicorn.chs b/bindings/haskell/src/Unicorn/Internal/Unicorn.chs
index bfe8bd37..db975520 100644
--- a/bindings/haskell/src/Unicorn/Internal/Unicorn.chs
+++ b/bindings/haskell/src/Unicorn/Internal/Unicorn.chs
@@ -28,6 +28,8 @@ module Unicorn.Internal.Unicorn
     , ucEmuStop
     , ucRegWrite
     , ucRegRead
+    , ucRegWriteBatch
+    , ucRegReadBatch
     , ucMemWrite
     , ucMemRead
     , ucMemMap
@@ -154,7 +156,8 @@ mkContext ptr =
    , `Word64'
    , `Word64'
    , `Int'
-   , `Int'} -> `Error'
+   , `Int'
+   } -> `Error'
 #}
 
 {# fun uc_emu_stop as ^
@@ -166,19 +169,37 @@ mkContext ptr =
 -- Register operations
 -------------------------------------------------------------------------------
 
-{# fun uc_reg_write as ^
+{# fun uc_reg_write_wrapper as ucRegWrite
    `Reg r' =>
    { `Engine'
    , enumToNum `r'
-   , castPtr `Ptr Int64'
+   , withIntegral* `Int64'
    } -> `Error'
 #}
 
-{# fun uc_reg_read as ^
+{# fun uc_reg_read_wrapper as ucRegRead
    `Reg r' =>
    { `Engine'
    , enumToNum `r'
-   , allocaInt64ToVoid- `Int64' castPtrAndPeek*
+   , alloca- `Int64' castPtrAndPeek*
+   } -> `Error'
+#}
+
+{# fun uc_reg_write_batch_wrapper as ucRegWriteBatch
+   `Reg r' =>
+   { `Engine'
+   , withEnums* `[r]'
+   , integralListToArray* `[Int64]'
+   , `Int'
+   } -> `Error'
+#}
+
+{# fun uc_reg_read_batch_wrapper as ucRegReadBatch
+   `Reg r' =>
+   { `Engine'
+   , withEnums* `[r]'
+   , castPtr `Ptr Int64'
+   , `Int'
    } -> `Error'
 #}
 
@@ -197,7 +218,8 @@ mkContext ptr =
    { `Engine'
    , `Word64'
    , castPtr `Ptr Word8'
-   , `Int'} -> `Error'
+   , `Int'
+   } -> `Error'
 #}
 
 {# fun uc_mem_map as ^
@@ -205,7 +227,8 @@ mkContext ptr =
    , `Word64'
    , `Int'
    , combineEnums `[MemoryPermission]'
-   } -> `Error' #}
+   } -> `Error'
+#}
 
 {# fun uc_mem_unmap as ^
    { `Engine'
@@ -296,13 +319,32 @@ expandMemPerms perms =
         checkRWE _ [] =
             []
 
-allocaInt64ToVoid :: (Ptr () -> IO b)
-                  -> IO b
-allocaInt64ToVoid f =
-    alloca $ \(ptr :: Ptr Int64) -> poke ptr 0 >> f (castPtr ptr)
+withIntegral :: (Integral a, Num b, Storable b)
+             => a
+             -> (Ptr b -> IO c)
+             -> IO c
+withIntegral =
+    with . fromIntegral
 
-withByteStringLen :: ByteString
-                  -> ((Ptr (), CULong) -> IO a)
-                  -> IO a
+withByteStringLen :: Integral a
+                  => ByteString
+                  -> ((Ptr (), a) -> IO b)
+                  -> IO b
 withByteStringLen bs f =
     useAsCStringLen bs $ \(ptr, len) -> f (castPtr ptr, fromIntegral len)
+
+withEnums :: Enum a
+          => [a]
+          -> (Ptr b -> IO c)
+          -> IO c
+withEnums l f =
+    let ints :: [CInt] = map enumToNum l in
+    withArray ints $ \ptr -> f (castPtr ptr)
+
+integralListToArray :: (Integral a, Storable b, Num b)
+                    => [a]
+                    -> (Ptr b -> IO c)
+                    -> IO c
+integralListToArray l f =
+    let l' = map fromIntegral l in
+    withArray l' $ \array -> f array
diff --git a/bindings/haskell/src/cbits/unicorn_wrapper.c b/bindings/haskell/src/cbits/unicorn_wrapper.c
index 92251a7c..878518ee 100644
--- a/bindings/haskell/src/cbits/unicorn_wrapper.c
+++ b/bindings/haskell/src/cbits/unicorn_wrapper.c
@@ -1,3 +1,5 @@
+#include <stdlib.h>
+
 #include "unicorn_wrapper.h"
 
 void uc_close_wrapper(uc_engine *uc) {
@@ -7,6 +9,42 @@ void uc_close_wrapper(uc_engine *uc) {
 void uc_close_dummy(uc_engine *uc) {
 }
 
+uc_err uc_reg_write_wrapper(uc_engine *uc, int regid, const int64_t *value) {
+    return uc_reg_write(uc, regid, (const void*) value);
+}
+
+uc_err uc_reg_read_wrapper(uc_engine *uc, int regid, int64_t *value) {
+    return uc_reg_read(uc, regid, (void*) value);
+}
+
+uc_err uc_reg_write_batch_wrapper(uc_engine *uc, int *regs, int64_t *vals, int count) {
+    void **valsPtr = malloc(sizeof(void*) * count);
+    int i;
+
+    for (i = 0; i < count; ++i) {
+        valsPtr[i] = (void*) &vals[i];
+    }
+
+    uc_err ret = uc_reg_write_batch(uc, regs, (void *const*) valsPtr, count);
+    free(valsPtr);
+
+    return ret;
+}
+
+uc_err uc_reg_read_batch_wrapper(uc_engine *uc, int *regs, int64_t *vals, int count) {
+    void **valsPtr = malloc(sizeof(void*) * count);
+    int i;
+
+    for (i = 0; i < count; ++i) {
+        valsPtr[i] = (void*) &vals[i];
+    }
+
+    uc_err ret = uc_reg_read_batch(uc, regs, valsPtr, count);
+    free(valsPtr);
+
+    return ret;
+}
+
 void uc_free_wrapper(void *mem) {
     uc_free(mem);
 }
diff --git a/bindings/haskell/src/include/unicorn_wrapper.h b/bindings/haskell/src/include/unicorn_wrapper.h
index 77174750..31757163 100644
--- a/bindings/haskell/src/include/unicorn_wrapper.h
+++ b/bindings/haskell/src/include/unicorn_wrapper.h
@@ -1,6 +1,7 @@
 #ifndef UNICORN_WRAPPER_H
 #define UNICORN_WRAPPER_H
 
+#include <stdint.h>
 #include <unicorn/unicorn.h>
 
 /*
@@ -13,6 +14,14 @@ void uc_close_wrapper(uc_engine *uc);
  */
 void uc_close_dummy(uc_engine *uc);
 
+/*
+ * Wrappers for register read/write functions that accept int64_t pointers.
+ */
+uc_err uc_reg_write_wrapper(uc_engine *uc, int regid, const int64_t *value);
+uc_err uc_reg_read_wrapper(uc_engine *uc, int regid, int64_t *value);
+uc_err uc_reg_write_batch_wrapper(uc_engine *uc, int *regs, int64_t *vals, int count);
+uc_err uc_reg_read_batch_wrapper(uc_engine *uc, int *regs, int64_t *vals, int count);
+
 /*
  * Wrap Unicorn's uc_free function and ignore the returned error code.
  */
diff --git a/bindings/java/unicorn/X86Const.java b/bindings/java/unicorn/X86Const.java
index 7afed1f9..c41ccfc9 100644
--- a/bindings/java/unicorn/X86Const.java
+++ b/bindings/java/unicorn/X86Const.java
@@ -254,7 +254,8 @@ public interface X86Const {
    public static final int UC_X86_REG_TR = 245;
    public static final int UC_X86_REG_FPCW = 246;
    public static final int UC_X86_REG_FPTAG = 247;
-   public static final int UC_X86_REG_ENDING = 248;
+   public static final int UC_X86_REG_MSR = 248;
+   public static final int UC_X86_REG_ENDING = 249;
 
 // X86 instructions
 
diff --git a/bindings/python/sample_arm.py b/bindings/python/sample_arm.py
index 402ba57a..5fae6b6b 100755
--- a/bindings/python/sample_arm.py
+++ b/bindings/python/sample_arm.py
@@ -41,7 +41,8 @@ def test_arm():
         mu.reg_write(UC_ARM_REG_R0, 0x1234)
         mu.reg_write(UC_ARM_REG_R2, 0x6789)
         mu.reg_write(UC_ARM_REG_R3, 0x3333)
-
+        mu.reg_write(UC_ARM_REG_APSR, 0xFFFFFFFF) #All application flags turned on
+   
         # tracing all basic blocks with customized callback
         mu.hook_add(UC_HOOK_BLOCK, hook_block)
 
diff --git a/bindings/python/unicorn/unicorn.py b/bindings/python/unicorn/unicorn.py
index 18751ae4..fa50ad1f 100644
--- a/bindings/python/unicorn/unicorn.py
+++ b/bindings/python/unicorn/unicorn.py
@@ -9,7 +9,7 @@ import os.path
 import sys
 import weakref
 
-from . import x86_const, unicorn_const as uc
+from . import x86_const, arm64_const, unicorn_const as uc
 
 if not hasattr(sys.modules[__name__], "__file__"):
     __file__ = inspect.getfile(inspect.currentframe())
@@ -202,6 +202,11 @@ class uc_x86_mmr(ctypes.Structure):
         ("flags", ctypes.c_uint32),     # not used by GDTR and IDTR
     ]
 
+class uc_x86_msr(ctypes.Structure):
+    _fields_ = [
+        ("rid", ctypes.c_uint32),
+        ("value", ctypes.c_uint64),
+    ]
 
 class uc_x86_float80(ctypes.Structure):
     """Float80"""
@@ -218,6 +223,13 @@ class uc_x86_xmm(ctypes.Structure):
         ("high_qword", ctypes.c_uint64),
     ]
 
+class uc_arm64_neon128(ctypes.Structure):
+    """128-bit neon register"""
+    _fields_ = [
+        ("low_qword", ctypes.c_uint64),
+        ("high_qword", ctypes.c_uint64),
+    ]
+
 # Subclassing ref to allow property assignment.
 class UcRef(weakref.ref):
     pass
@@ -282,7 +294,7 @@ class Uc(object):
             raise UcError(status)
 
     # return the value of a register
-    def reg_read(self, reg_id):
+    def reg_read(self, reg_id, opt=None):
         if self._arch == uc.UC_ARCH_X86:
             if reg_id in [x86_const.UC_X86_REG_IDTR, x86_const.UC_X86_REG_GDTR, x86_const.UC_X86_REG_LDTR, x86_const.UC_X86_REG_TR]:
                 reg = uc_x86_mmr()
@@ -302,6 +314,23 @@ class Uc(object):
                 if status != uc.UC_ERR_OK:
                     raise UcError(status)
                 return reg.low_qword | (reg.high_qword << 64)
+            if reg_id is x86_const.UC_X86_REG_MSR:
+                if opt is None:
+                    raise UcError(uc.UC_ERR_ARG)
+                reg = uc_x86_msr()
+                reg.rid = opt
+                status = _uc.uc_reg_read(self._uch, reg_id, ctypes.byref(reg))
+                if status != uc.UC_ERR_OK:
+                    raise UcError(status)
+                return reg.value
+
+        if self._arch == uc.UC_ARCH_ARM64:
+            if reg_id in range(arm64_const.UC_ARM64_REG_Q0, arm64_const.UC_ARM64_REG_Q31+1) or range(arm64_const.UC_ARM64_REG_V0, arm64_const.UC_ARM64_REG_V31+1):
+                reg = uc_arm64_neon128()
+                status = _uc.uc_reg_read(self._uch, reg_id, ctypes.byref(reg))
+                if status != uc.UC_ERR_OK:
+                    raise UcError(status)
+                return reg.low_qword | (reg.high_qword << 64)
 
         # read to 64bit number to be safe
         reg = ctypes.c_uint64(0)
@@ -330,6 +359,16 @@ class Uc(object):
                 reg = uc_x86_xmm()
                 reg.low_qword = value & 0xffffffffffffffff
                 reg.high_qword = value >> 64
+            if reg_id is x86_const.UC_X86_REG_MSR:
+                reg = uc_x86_msr()
+                reg.rid = value[0]
+                reg.value = value[1]
+
+        if self._arch == uc.UC_ARCH_ARM64:
+            if reg_id in range(arm64_const.UC_ARM64_REG_Q0, arm64_const.UC_ARM64_REG_Q31+1) or range(arm64_const.UC_ARM64_REG_V0, arm64_const.UC_ARM64_REG_V31+1):
+                reg = uc_arm64_neon128()
+                reg.low_qword = value & 0xffffffffffffffff
+                reg.high_qword = value >> 64
 
         if reg is None:
             # convert to 64bit number to be safe
@@ -339,6 +378,14 @@ class Uc(object):
         if status != uc.UC_ERR_OK:
             raise UcError(status)
 
+    # read from MSR - X86 only
+    def msr_read(self, msr_id):
+        return self.reg_read(x86_const.UC_X86_REG_MSR, msr_id)
+
+    # write to MSR - X86 only
+    def msr_write(self, msr_id, value):
+        return self.reg_write(x86_const.UC_X86_REG_MSR, (msr_id, value))
+
     # read data from memory
     def mem_read(self, address, size):
         data = ctypes.create_string_buffer(size)
diff --git a/bindings/python/unicorn/x86_const.py b/bindings/python/unicorn/x86_const.py
index ee5ba38f..5642c543 100644
--- a/bindings/python/unicorn/x86_const.py
+++ b/bindings/python/unicorn/x86_const.py
@@ -250,7 +250,8 @@ UC_X86_REG_LDTR = 244
 UC_X86_REG_TR = 245
 UC_X86_REG_FPCW = 246
 UC_X86_REG_FPTAG = 247
-UC_X86_REG_ENDING = 248
+UC_X86_REG_MSR = 248
+UC_X86_REG_ENDING = 249
 
 # X86 instructions
 
diff --git a/bindings/ruby/unicorn_gem/lib/unicorn/x86_const.rb b/bindings/ruby/unicorn_gem/lib/unicorn/x86_const.rb
index 2226de17..f85ae21a 100644
--- a/bindings/ruby/unicorn_gem/lib/unicorn/x86_const.rb
+++ b/bindings/ruby/unicorn_gem/lib/unicorn/x86_const.rb
@@ -252,7 +252,8 @@ module Unicorn
 	UC_X86_REG_TR = 245
 	UC_X86_REG_FPCW = 246
 	UC_X86_REG_FPTAG = 247
-	UC_X86_REG_ENDING = 248
+	UC_X86_REG_MSR = 248
+	UC_X86_REG_ENDING = 249
 
 # X86 instructions
 
diff --git a/include/uc_priv.h b/include/uc_priv.h
index 4251b06b..e16ad943 100644
--- a/include/uc_priv.h
+++ b/include/uc_priv.h
@@ -253,3 +253,4 @@ struct uc_context {
 MemoryRegion *memory_mapping(struct uc_struct* uc, uint64_t address);
 
 #endif
+/* vim: set ts=4 noet:  */
diff --git a/include/unicorn/x86.h b/include/unicorn/x86.h
index cb3f1d93..40e3e3c7 100644
--- a/include/unicorn/x86.h
+++ b/include/unicorn/x86.h
@@ -19,64 +19,72 @@ typedef struct uc_x86_mmr {
     uint32_t flags;     /* not used by GDTR and IDTR */
 } uc_x86_mmr;
 
+// Model-Specific Register structure, use this with UC_X86_REG_MSR (as the register ID) in
+// call to uc_reg_write/uc_reg_read() to manipulate MSRs.
+typedef struct uc_x86_msr {
+    uint32_t rid;   // MSR id
+    uint64_t value; // MSR value
+} uc_x86_msr;
+
 // Callback function for tracing SYSCALL/SYSENTER (for uc_hook_intr())
 // @user_data: user data passed to tracing APIs.
 typedef void (*uc_cb_insn_syscall_t)(struct uc_struct *uc, void *user_data);
 
 //> X86 registers
 typedef enum uc_x86_reg {
-    UC_X86_REG_INVALID = 0,
-    UC_X86_REG_AH, UC_X86_REG_AL, UC_X86_REG_AX, UC_X86_REG_BH, UC_X86_REG_BL,
-    UC_X86_REG_BP, UC_X86_REG_BPL, UC_X86_REG_BX, UC_X86_REG_CH, UC_X86_REG_CL,
-    UC_X86_REG_CS, UC_X86_REG_CX, UC_X86_REG_DH, UC_X86_REG_DI, UC_X86_REG_DIL,
-    UC_X86_REG_DL, UC_X86_REG_DS, UC_X86_REG_DX, UC_X86_REG_EAX, UC_X86_REG_EBP,
-    UC_X86_REG_EBX, UC_X86_REG_ECX, UC_X86_REG_EDI, UC_X86_REG_EDX, UC_X86_REG_EFLAGS,
-    UC_X86_REG_EIP, UC_X86_REG_EIZ, UC_X86_REG_ES, UC_X86_REG_ESI, UC_X86_REG_ESP,
-    UC_X86_REG_FPSW, UC_X86_REG_FS, UC_X86_REG_GS, UC_X86_REG_IP, UC_X86_REG_RAX,
-    UC_X86_REG_RBP, UC_X86_REG_RBX, UC_X86_REG_RCX, UC_X86_REG_RDI, UC_X86_REG_RDX,
-    UC_X86_REG_RIP, UC_X86_REG_RIZ, UC_X86_REG_RSI, UC_X86_REG_RSP, UC_X86_REG_SI,
-    UC_X86_REG_SIL, UC_X86_REG_SP, UC_X86_REG_SPL, UC_X86_REG_SS, UC_X86_REG_CR0,
-    UC_X86_REG_CR1, UC_X86_REG_CR2, UC_X86_REG_CR3, UC_X86_REG_CR4, UC_X86_REG_CR5,
-    UC_X86_REG_CR6, UC_X86_REG_CR7, UC_X86_REG_CR8, UC_X86_REG_CR9, UC_X86_REG_CR10,
-    UC_X86_REG_CR11, UC_X86_REG_CR12, UC_X86_REG_CR13, UC_X86_REG_CR14, UC_X86_REG_CR15,
-    UC_X86_REG_DR0, UC_X86_REG_DR1, UC_X86_REG_DR2, UC_X86_REG_DR3, UC_X86_REG_DR4,
-    UC_X86_REG_DR5, UC_X86_REG_DR6, UC_X86_REG_DR7, UC_X86_REG_DR8, UC_X86_REG_DR9,
-    UC_X86_REG_DR10, UC_X86_REG_DR11, UC_X86_REG_DR12, UC_X86_REG_DR13, UC_X86_REG_DR14,
-    UC_X86_REG_DR15, UC_X86_REG_FP0, UC_X86_REG_FP1, UC_X86_REG_FP2, UC_X86_REG_FP3,
-    UC_X86_REG_FP4, UC_X86_REG_FP5, UC_X86_REG_FP6, UC_X86_REG_FP7,
-    UC_X86_REG_K0, UC_X86_REG_K1, UC_X86_REG_K2, UC_X86_REG_K3, UC_X86_REG_K4,
-    UC_X86_REG_K5, UC_X86_REG_K6, UC_X86_REG_K7, UC_X86_REG_MM0, UC_X86_REG_MM1,
-    UC_X86_REG_MM2, UC_X86_REG_MM3, UC_X86_REG_MM4, UC_X86_REG_MM5, UC_X86_REG_MM6,
-    UC_X86_REG_MM7, UC_X86_REG_R8, UC_X86_REG_R9, UC_X86_REG_R10, UC_X86_REG_R11,
-    UC_X86_REG_R12, UC_X86_REG_R13, UC_X86_REG_R14, UC_X86_REG_R15,
-    UC_X86_REG_ST0, UC_X86_REG_ST1, UC_X86_REG_ST2, UC_X86_REG_ST3,
-    UC_X86_REG_ST4, UC_X86_REG_ST5, UC_X86_REG_ST6, UC_X86_REG_ST7,
-    UC_X86_REG_XMM0, UC_X86_REG_XMM1, UC_X86_REG_XMM2, UC_X86_REG_XMM3, UC_X86_REG_XMM4,
-    UC_X86_REG_XMM5, UC_X86_REG_XMM6, UC_X86_REG_XMM7, UC_X86_REG_XMM8, UC_X86_REG_XMM9,
-    UC_X86_REG_XMM10, UC_X86_REG_XMM11, UC_X86_REG_XMM12, UC_X86_REG_XMM13, UC_X86_REG_XMM14,
-    UC_X86_REG_XMM15, UC_X86_REG_XMM16, UC_X86_REG_XMM17, UC_X86_REG_XMM18, UC_X86_REG_XMM19,
-    UC_X86_REG_XMM20, UC_X86_REG_XMM21, UC_X86_REG_XMM22, UC_X86_REG_XMM23, UC_X86_REG_XMM24,
-    UC_X86_REG_XMM25, UC_X86_REG_XMM26, UC_X86_REG_XMM27, UC_X86_REG_XMM28, UC_X86_REG_XMM29,
-    UC_X86_REG_XMM30, UC_X86_REG_XMM31, UC_X86_REG_YMM0, UC_X86_REG_YMM1, UC_X86_REG_YMM2,
-    UC_X86_REG_YMM3, UC_X86_REG_YMM4, UC_X86_REG_YMM5, UC_X86_REG_YMM6, UC_X86_REG_YMM7,
-    UC_X86_REG_YMM8, UC_X86_REG_YMM9, UC_X86_REG_YMM10, UC_X86_REG_YMM11, UC_X86_REG_YMM12,
-    UC_X86_REG_YMM13, UC_X86_REG_YMM14, UC_X86_REG_YMM15, UC_X86_REG_YMM16, UC_X86_REG_YMM17,
-    UC_X86_REG_YMM18, UC_X86_REG_YMM19, UC_X86_REG_YMM20, UC_X86_REG_YMM21, UC_X86_REG_YMM22,
-    UC_X86_REG_YMM23, UC_X86_REG_YMM24, UC_X86_REG_YMM25, UC_X86_REG_YMM26, UC_X86_REG_YMM27,
-    UC_X86_REG_YMM28, UC_X86_REG_YMM29, UC_X86_REG_YMM30, UC_X86_REG_YMM31, UC_X86_REG_ZMM0,
-    UC_X86_REG_ZMM1, UC_X86_REG_ZMM2, UC_X86_REG_ZMM3, UC_X86_REG_ZMM4, UC_X86_REG_ZMM5,
-    UC_X86_REG_ZMM6, UC_X86_REG_ZMM7, UC_X86_REG_ZMM8, UC_X86_REG_ZMM9, UC_X86_REG_ZMM10,
-    UC_X86_REG_ZMM11, UC_X86_REG_ZMM12, UC_X86_REG_ZMM13, UC_X86_REG_ZMM14, UC_X86_REG_ZMM15,
-    UC_X86_REG_ZMM16, UC_X86_REG_ZMM17, UC_X86_REG_ZMM18, UC_X86_REG_ZMM19, UC_X86_REG_ZMM20,
-    UC_X86_REG_ZMM21, UC_X86_REG_ZMM22, UC_X86_REG_ZMM23, UC_X86_REG_ZMM24, UC_X86_REG_ZMM25,
-    UC_X86_REG_ZMM26, UC_X86_REG_ZMM27, UC_X86_REG_ZMM28, UC_X86_REG_ZMM29, UC_X86_REG_ZMM30,
-    UC_X86_REG_ZMM31, UC_X86_REG_R8B, UC_X86_REG_R9B, UC_X86_REG_R10B, UC_X86_REG_R11B,
-    UC_X86_REG_R12B, UC_X86_REG_R13B, UC_X86_REG_R14B, UC_X86_REG_R15B, UC_X86_REG_R8D,
-    UC_X86_REG_R9D, UC_X86_REG_R10D, UC_X86_REG_R11D, UC_X86_REG_R12D, UC_X86_REG_R13D,
-    UC_X86_REG_R14D, UC_X86_REG_R15D, UC_X86_REG_R8W, UC_X86_REG_R9W, UC_X86_REG_R10W,
-    UC_X86_REG_R11W, UC_X86_REG_R12W, UC_X86_REG_R13W, UC_X86_REG_R14W, UC_X86_REG_R15W,
-    UC_X86_REG_IDTR, UC_X86_REG_GDTR, UC_X86_REG_LDTR, UC_X86_REG_TR, UC_X86_REG_FPCW,
-    UC_X86_REG_FPTAG,
+	UC_X86_REG_INVALID = 0,
+	UC_X86_REG_AH, UC_X86_REG_AL, UC_X86_REG_AX, UC_X86_REG_BH, UC_X86_REG_BL,
+	UC_X86_REG_BP, UC_X86_REG_BPL, UC_X86_REG_BX, UC_X86_REG_CH, UC_X86_REG_CL,
+	UC_X86_REG_CS, UC_X86_REG_CX, UC_X86_REG_DH, UC_X86_REG_DI, UC_X86_REG_DIL,
+	UC_X86_REG_DL, UC_X86_REG_DS, UC_X86_REG_DX, UC_X86_REG_EAX, UC_X86_REG_EBP,
+	UC_X86_REG_EBX, UC_X86_REG_ECX, UC_X86_REG_EDI, UC_X86_REG_EDX, UC_X86_REG_EFLAGS,
+	UC_X86_REG_EIP, UC_X86_REG_EIZ, UC_X86_REG_ES, UC_X86_REG_ESI, UC_X86_REG_ESP,
+	UC_X86_REG_FPSW, UC_X86_REG_FS, UC_X86_REG_GS, UC_X86_REG_IP, UC_X86_REG_RAX,
+	UC_X86_REG_RBP, UC_X86_REG_RBX, UC_X86_REG_RCX, UC_X86_REG_RDI, UC_X86_REG_RDX,
+	UC_X86_REG_RIP, UC_X86_REG_RIZ, UC_X86_REG_RSI, UC_X86_REG_RSP, UC_X86_REG_SI,
+	UC_X86_REG_SIL, UC_X86_REG_SP, UC_X86_REG_SPL, UC_X86_REG_SS, UC_X86_REG_CR0,
+	UC_X86_REG_CR1, UC_X86_REG_CR2, UC_X86_REG_CR3, UC_X86_REG_CR4, UC_X86_REG_CR5,
+	UC_X86_REG_CR6, UC_X86_REG_CR7, UC_X86_REG_CR8, UC_X86_REG_CR9, UC_X86_REG_CR10,
+	UC_X86_REG_CR11, UC_X86_REG_CR12, UC_X86_REG_CR13, UC_X86_REG_CR14, UC_X86_REG_CR15,
+	UC_X86_REG_DR0, UC_X86_REG_DR1, UC_X86_REG_DR2, UC_X86_REG_DR3, UC_X86_REG_DR4,
+	UC_X86_REG_DR5, UC_X86_REG_DR6, UC_X86_REG_DR7, UC_X86_REG_DR8, UC_X86_REG_DR9,
+	UC_X86_REG_DR10, UC_X86_REG_DR11, UC_X86_REG_DR12, UC_X86_REG_DR13, UC_X86_REG_DR14,
+	UC_X86_REG_DR15, UC_X86_REG_FP0, UC_X86_REG_FP1, UC_X86_REG_FP2, UC_X86_REG_FP3,
+	UC_X86_REG_FP4, UC_X86_REG_FP5, UC_X86_REG_FP6, UC_X86_REG_FP7,
+	UC_X86_REG_K0, UC_X86_REG_K1, UC_X86_REG_K2, UC_X86_REG_K3, UC_X86_REG_K4,
+	UC_X86_REG_K5, UC_X86_REG_K6, UC_X86_REG_K7, UC_X86_REG_MM0, UC_X86_REG_MM1,
+	UC_X86_REG_MM2, UC_X86_REG_MM3, UC_X86_REG_MM4, UC_X86_REG_MM5, UC_X86_REG_MM6,
+	UC_X86_REG_MM7, UC_X86_REG_R8, UC_X86_REG_R9, UC_X86_REG_R10, UC_X86_REG_R11,
+	UC_X86_REG_R12, UC_X86_REG_R13, UC_X86_REG_R14, UC_X86_REG_R15,
+	UC_X86_REG_ST0, UC_X86_REG_ST1, UC_X86_REG_ST2, UC_X86_REG_ST3,
+	UC_X86_REG_ST4, UC_X86_REG_ST5, UC_X86_REG_ST6, UC_X86_REG_ST7,
+	UC_X86_REG_XMM0, UC_X86_REG_XMM1, UC_X86_REG_XMM2, UC_X86_REG_XMM3, UC_X86_REG_XMM4,
+	UC_X86_REG_XMM5, UC_X86_REG_XMM6, UC_X86_REG_XMM7, UC_X86_REG_XMM8, UC_X86_REG_XMM9,
+	UC_X86_REG_XMM10, UC_X86_REG_XMM11, UC_X86_REG_XMM12, UC_X86_REG_XMM13, UC_X86_REG_XMM14,
+	UC_X86_REG_XMM15, UC_X86_REG_XMM16, UC_X86_REG_XMM17, UC_X86_REG_XMM18, UC_X86_REG_XMM19,
+	UC_X86_REG_XMM20, UC_X86_REG_XMM21, UC_X86_REG_XMM22, UC_X86_REG_XMM23, UC_X86_REG_XMM24,
+	UC_X86_REG_XMM25, UC_X86_REG_XMM26, UC_X86_REG_XMM27, UC_X86_REG_XMM28, UC_X86_REG_XMM29,
+	UC_X86_REG_XMM30, UC_X86_REG_XMM31, UC_X86_REG_YMM0, UC_X86_REG_YMM1, UC_X86_REG_YMM2,
+	UC_X86_REG_YMM3, UC_X86_REG_YMM4, UC_X86_REG_YMM5, UC_X86_REG_YMM6, UC_X86_REG_YMM7,
+	UC_X86_REG_YMM8, UC_X86_REG_YMM9, UC_X86_REG_YMM10, UC_X86_REG_YMM11, UC_X86_REG_YMM12,
+	UC_X86_REG_YMM13, UC_X86_REG_YMM14, UC_X86_REG_YMM15, UC_X86_REG_YMM16, UC_X86_REG_YMM17,
+	UC_X86_REG_YMM18, UC_X86_REG_YMM19, UC_X86_REG_YMM20, UC_X86_REG_YMM21, UC_X86_REG_YMM22,
+	UC_X86_REG_YMM23, UC_X86_REG_YMM24, UC_X86_REG_YMM25, UC_X86_REG_YMM26, UC_X86_REG_YMM27,
+	UC_X86_REG_YMM28, UC_X86_REG_YMM29, UC_X86_REG_YMM30, UC_X86_REG_YMM31, UC_X86_REG_ZMM0,
+	UC_X86_REG_ZMM1, UC_X86_REG_ZMM2, UC_X86_REG_ZMM3, UC_X86_REG_ZMM4, UC_X86_REG_ZMM5,
+	UC_X86_REG_ZMM6, UC_X86_REG_ZMM7, UC_X86_REG_ZMM8, UC_X86_REG_ZMM9, UC_X86_REG_ZMM10,
+	UC_X86_REG_ZMM11, UC_X86_REG_ZMM12, UC_X86_REG_ZMM13, UC_X86_REG_ZMM14, UC_X86_REG_ZMM15,
+	UC_X86_REG_ZMM16, UC_X86_REG_ZMM17, UC_X86_REG_ZMM18, UC_X86_REG_ZMM19, UC_X86_REG_ZMM20,
+	UC_X86_REG_ZMM21, UC_X86_REG_ZMM22, UC_X86_REG_ZMM23, UC_X86_REG_ZMM24, UC_X86_REG_ZMM25,
+	UC_X86_REG_ZMM26, UC_X86_REG_ZMM27, UC_X86_REG_ZMM28, UC_X86_REG_ZMM29, UC_X86_REG_ZMM30,
+	UC_X86_REG_ZMM31, UC_X86_REG_R8B, UC_X86_REG_R9B, UC_X86_REG_R10B, UC_X86_REG_R11B,
+	UC_X86_REG_R12B, UC_X86_REG_R13B, UC_X86_REG_R14B, UC_X86_REG_R15B, UC_X86_REG_R8D,
+	UC_X86_REG_R9D, UC_X86_REG_R10D, UC_X86_REG_R11D, UC_X86_REG_R12D, UC_X86_REG_R13D,
+	UC_X86_REG_R14D, UC_X86_REG_R15D, UC_X86_REG_R8W, UC_X86_REG_R9W, UC_X86_REG_R10W,
+	UC_X86_REG_R11W, UC_X86_REG_R12W, UC_X86_REG_R13W, UC_X86_REG_R14W, UC_X86_REG_R15W,
+	UC_X86_REG_IDTR, UC_X86_REG_GDTR, UC_X86_REG_LDTR, UC_X86_REG_TR, UC_X86_REG_FPCW,
+	UC_X86_REG_FPTAG,
+    UC_X86_REG_MSR, // Model-Specific Register
 
     UC_X86_REG_ENDING		// <-- mark the end of the list of registers
 } uc_x86_reg;
diff --git a/qemu/aarch64.h b/qemu/aarch64.h
index 6cb47c34..7a7a5ea7 100644
--- a/qemu/aarch64.h
+++ b/qemu/aarch64.h
@@ -2,6 +2,7 @@
 #ifndef UNICORN_AUTOGEN_AARCH64_H
 #define UNICORN_AUTOGEN_AARCH64_H
 #define aarch64_tb_set_jmp_target aarch64_tb_set_jmp_target_aarch64
+#define ppc_tb_set_jmp_target ppc_tb_set_jmp_target_aarch64
 #define use_idiv_instructions_rt use_idiv_instructions_rt_aarch64
 #define tcg_target_deposit_valid tcg_target_deposit_valid_aarch64
 #define helper_power_down helper_power_down_aarch64
diff --git a/qemu/arm.h b/qemu/arm.h
index 5ab27471..80225bdd 100644
--- a/qemu/arm.h
+++ b/qemu/arm.h
@@ -2,6 +2,7 @@
 #ifndef UNICORN_AUTOGEN_ARM_H
 #define UNICORN_AUTOGEN_ARM_H
 #define aarch64_tb_set_jmp_target aarch64_tb_set_jmp_target_arm
+#define ppc_tb_set_jmp_target ppc_tb_set_jmp_target_arm
 #define use_idiv_instructions_rt use_idiv_instructions_rt_arm
 #define tcg_target_deposit_valid tcg_target_deposit_valid_arm
 #define helper_power_down helper_power_down_arm
diff --git a/qemu/header_gen.py b/qemu/header_gen.py
index 0561817d..047f6358 100644
--- a/qemu/header_gen.py
+++ b/qemu/header_gen.py
@@ -8,6 +8,7 @@ import sys
 
 symbols = (
     'aarch64_tb_set_jmp_target',
+    'ppc_tb_set_jmp_target',
     'use_idiv_instructions_rt',
     'tcg_target_deposit_valid',
     'helper_power_down',
diff --git a/qemu/include/elf.h b/qemu/include/elf.h
index c9b16056..b55aaa6f 100644
--- a/qemu/include/elf.h
+++ b/qemu/include/elf.h
@@ -126,11 +126,15 @@
 /* Bits present in AT_HWCAP for Sparc.  */
 #define HWCAP_SPARC_VIS3        0x00020000
 
+/* Bits present in AT_HWCAP for PowerPC.  */
+#define PPC_FEATURE_ARCH_2_06           0x00000100
+
 /* Symbolic values for the entries in the auxiliary table
   put on the initial stack */
 #define AT_PLATFORM 15  /* string identifying CPU for optimizations */
 #define AT_HWCAP  16    /* arch dependent hints at CPU capabilities */
-
+#define AT_DCACHEBSIZE   19  /* data cache block size */
+#define AT_ICACHEBSIZE 20  /* instruction cache block size */
 
 /*
  * 68k ELF relocation types
diff --git a/qemu/m68k.h b/qemu/m68k.h
index 9a265549..f40a8dc4 100644
--- a/qemu/m68k.h
+++ b/qemu/m68k.h
@@ -2,6 +2,7 @@
 #ifndef UNICORN_AUTOGEN_M68K_H
 #define UNICORN_AUTOGEN_M68K_H
 #define aarch64_tb_set_jmp_target aarch64_tb_set_jmp_target_m68k
+#define ppc_tb_set_jmp_target ppc_tb_set_jmp_target_m68k
 #define use_idiv_instructions_rt use_idiv_instructions_rt_m68k
 #define tcg_target_deposit_valid tcg_target_deposit_valid_m68k
 #define helper_power_down helper_power_down_m68k
diff --git a/qemu/mips.h b/qemu/mips.h
index 57bb375a..1b8e9e2b 100644
--- a/qemu/mips.h
+++ b/qemu/mips.h
@@ -2,6 +2,7 @@
 #ifndef UNICORN_AUTOGEN_MIPS_H
 #define UNICORN_AUTOGEN_MIPS_H
 #define aarch64_tb_set_jmp_target aarch64_tb_set_jmp_target_mips
+#define ppc_tb_set_jmp_target ppc_tb_set_jmp_target_mips
 #define use_idiv_instructions_rt use_idiv_instructions_rt_mips
 #define tcg_target_deposit_valid tcg_target_deposit_valid_mips
 #define helper_power_down helper_power_down_mips
diff --git a/qemu/mips64.h b/qemu/mips64.h
index a6af4eeb..06b30666 100644
--- a/qemu/mips64.h
+++ b/qemu/mips64.h
@@ -2,6 +2,7 @@
 #ifndef UNICORN_AUTOGEN_MIPS64_H
 #define UNICORN_AUTOGEN_MIPS64_H
 #define aarch64_tb_set_jmp_target aarch64_tb_set_jmp_target_mips64
+#define ppc_tb_set_jmp_target ppc_tb_set_jmp_target_mips64
 #define use_idiv_instructions_rt use_idiv_instructions_rt_mips64
 #define tcg_target_deposit_valid tcg_target_deposit_valid_mips64
 #define helper_power_down helper_power_down_mips64
diff --git a/qemu/mips64el.h b/qemu/mips64el.h
index c453400d..87cc74c8 100644
--- a/qemu/mips64el.h
+++ b/qemu/mips64el.h
@@ -2,6 +2,7 @@
 #ifndef UNICORN_AUTOGEN_MIPS64EL_H
 #define UNICORN_AUTOGEN_MIPS64EL_H
 #define aarch64_tb_set_jmp_target aarch64_tb_set_jmp_target_mips64el
+#define ppc_tb_set_jmp_target ppc_tb_set_jmp_target_mips64el
 #define use_idiv_instructions_rt use_idiv_instructions_rt_mips64el
 #define tcg_target_deposit_valid tcg_target_deposit_valid_mips64el
 #define helper_power_down helper_power_down_mips64el
diff --git a/qemu/mipsel.h b/qemu/mipsel.h
index 4ad7be8d..a8a3fb4f 100644
--- a/qemu/mipsel.h
+++ b/qemu/mipsel.h
@@ -2,6 +2,7 @@
 #ifndef UNICORN_AUTOGEN_MIPSEL_H
 #define UNICORN_AUTOGEN_MIPSEL_H
 #define aarch64_tb_set_jmp_target aarch64_tb_set_jmp_target_mipsel
+#define ppc_tb_set_jmp_target ppc_tb_set_jmp_target_mipsel
 #define use_idiv_instructions_rt use_idiv_instructions_rt_mipsel
 #define tcg_target_deposit_valid tcg_target_deposit_valid_mipsel
 #define helper_power_down helper_power_down_mipsel
diff --git a/qemu/powerpc.h b/qemu/powerpc.h
index 80520561..fd831d98 100644
--- a/qemu/powerpc.h
+++ b/qemu/powerpc.h
@@ -2,6 +2,7 @@
 #ifndef UNICORN_AUTOGEN_POWERPC_H
 #define UNICORN_AUTOGEN_POWERPC_H
 #define aarch64_tb_set_jmp_target aarch64_tb_set_jmp_target_powerpc
+#define ppc_tb_set_jmp_target ppc_tb_set_jmp_target_powerpc
 #define use_idiv_instructions_rt use_idiv_instructions_rt_powerpc
 #define tcg_target_deposit_valid tcg_target_deposit_valid_powerpc
 #define helper_power_down helper_power_down_powerpc
diff --git a/qemu/sparc.h b/qemu/sparc.h
index b3c6d853..fda21bf4 100644
--- a/qemu/sparc.h
+++ b/qemu/sparc.h
@@ -2,6 +2,7 @@
 #ifndef UNICORN_AUTOGEN_SPARC_H
 #define UNICORN_AUTOGEN_SPARC_H
 #define aarch64_tb_set_jmp_target aarch64_tb_set_jmp_target_sparc
+#define ppc_tb_set_jmp_target ppc_tb_set_jmp_target_sparc
 #define use_idiv_instructions_rt use_idiv_instructions_rt_sparc
 #define tcg_target_deposit_valid tcg_target_deposit_valid_sparc
 #define helper_power_down helper_power_down_sparc
diff --git a/qemu/sparc64.h b/qemu/sparc64.h
index c8860863..aaafd634 100644
--- a/qemu/sparc64.h
+++ b/qemu/sparc64.h
@@ -2,6 +2,7 @@
 #ifndef UNICORN_AUTOGEN_SPARC64_H
 #define UNICORN_AUTOGEN_SPARC64_H
 #define aarch64_tb_set_jmp_target aarch64_tb_set_jmp_target_sparc64
+#define ppc_tb_set_jmp_target ppc_tb_set_jmp_target_sparc64
 #define use_idiv_instructions_rt use_idiv_instructions_rt_sparc64
 #define tcg_target_deposit_valid tcg_target_deposit_valid_sparc64
 #define helper_power_down helper_power_down_sparc64
diff --git a/qemu/target-arm/unicorn_aarch64.c b/qemu/target-arm/unicorn_aarch64.c
index 3631e9ce..9a884b70 100644
--- a/qemu/target-arm/unicorn_aarch64.c
+++ b/qemu/target-arm/unicorn_aarch64.c
@@ -52,10 +52,27 @@ int arm64_reg_read(struct uc_struct *uc, unsigned int *regs, void **vals, int co
     for (i = 0; i < count; i++) {
         unsigned int regid = regs[i];
         void *value = vals[i];
+        // V & Q registers are the same
+        if (regid >= UC_ARM64_REG_V0 && regid <= UC_ARM64_REG_V31) {
+            regid += UC_ARM64_REG_Q0 - UC_ARM64_REG_V0;
+        }
         if (regid >= UC_ARM64_REG_X0 && regid <= UC_ARM64_REG_X28) {
             *(int64_t *)value = ARM_CPU(uc, mycpu)->env.xregs[regid - UC_ARM64_REG_X0];
         } else if (regid >= UC_ARM64_REG_W0 && regid <= UC_ARM64_REG_W30) {
             *(int32_t *)value = READ_DWORD(ARM_CPU(uc, mycpu)->env.xregs[regid - UC_ARM64_REG_W0]);
+        } else if (regid >= UC_ARM64_REG_Q0 && regid <= UC_ARM64_REG_Q31) {
+            float64 *dst = (float64*) value;
+            uint32_t reg_index = 2*(regid - UC_ARM64_REG_Q0);
+            dst[0] = ARM_CPU(uc, mycpu)->env.vfp.regs[reg_index];
+            dst[1] = ARM_CPU(uc, mycpu)->env.vfp.regs[reg_index+1];
+        } else if (regid >= UC_ARM64_REG_D0 && regid <= UC_ARM64_REG_D31) {
+            *(float64*)value = ARM_CPU(uc, mycpu)->env.vfp.regs[2*(regid - UC_ARM64_REG_D0)];
+        } else if (regid >= UC_ARM64_REG_S0 && regid <= UC_ARM64_REG_S31) {
+            *(int32_t*)value = READ_DWORD(ARM_CPU(uc, mycpu)->env.vfp.regs[2*(regid - UC_ARM64_REG_S0)]);
+        } else if (regid >= UC_ARM64_REG_H0 && regid <= UC_ARM64_REG_H31) {
+            *(int16_t*)value = READ_WORD(ARM_CPU(uc, mycpu)->env.vfp.regs[2*(regid - UC_ARM64_REG_H0)]);
+        } else if (regid >= UC_ARM64_REG_B0 && regid <= UC_ARM64_REG_B31) {
+            *(int8_t*)value = READ_BYTE_L(ARM_CPU(uc, mycpu)->env.vfp.regs[2*(regid - UC_ARM64_REG_B0)]);
         } else {
             switch(regid) {
                 default: break;
@@ -86,10 +103,26 @@ int arm64_reg_write(struct uc_struct *uc, unsigned int *regs, void* const* vals,
     for (i = 0; i < count; i++) {
         unsigned int regid = regs[i];
         const void *value = vals[i];
+        if (regid >= UC_ARM64_REG_V0 && regid <= UC_ARM64_REG_V31) {
+            regid += UC_ARM64_REG_Q0 - UC_ARM64_REG_V0;
+        }
         if (regid >= UC_ARM64_REG_X0 && regid <= UC_ARM64_REG_X28) {
             ARM_CPU(uc, mycpu)->env.xregs[regid - UC_ARM64_REG_X0] = *(uint64_t *)value;
         } else if (regid >= UC_ARM64_REG_W0 && regid <= UC_ARM64_REG_W30) {
             WRITE_DWORD(ARM_CPU(uc, mycpu)->env.xregs[regid - UC_ARM64_REG_W0], *(uint32_t *)value);
+        } else if (regid >= UC_ARM64_REG_Q0 && regid <= UC_ARM64_REG_Q31) {
+            float64 *src = (float64*) value;
+            uint32_t reg_index = 2*(regid - UC_ARM64_REG_Q0);
+            ARM_CPU(uc, mycpu)->env.vfp.regs[reg_index] = src[0];
+            ARM_CPU(uc, mycpu)->env.vfp.regs[reg_index+1] = src[1];
+        } else if (regid >= UC_ARM64_REG_D0 && regid <= UC_ARM64_REG_D31) {
+            ARM_CPU(uc, mycpu)->env.vfp.regs[2*(regid - UC_ARM64_REG_D0)] = * (float64*) value;
+        } else if (regid >= UC_ARM64_REG_S0 && regid <= UC_ARM64_REG_S31) {
+            WRITE_DWORD(ARM_CPU(uc, mycpu)->env.vfp.regs[2*(regid - UC_ARM64_REG_S0)], *(int32_t*) value);
+        } else if (regid >= UC_ARM64_REG_H0 && regid <= UC_ARM64_REG_H31) {
+            WRITE_WORD(ARM_CPU(uc, mycpu)->env.vfp.regs[2*(regid - UC_ARM64_REG_H0)], *(int16_t*) value);
+        } else if (regid >= UC_ARM64_REG_B0 && regid <= UC_ARM64_REG_B31) {
+            WRITE_BYTE_L(ARM_CPU(uc, mycpu)->env.vfp.regs[2*(regid - UC_ARM64_REG_B0)], *(int8_t*) value);
         } else {
             switch(regid) {
                 default: break;
diff --git a/qemu/target-arm/unicorn_arm.c b/qemu/target-arm/unicorn_arm.c
index a7ecc249..3d07683f 100644
--- a/qemu/target-arm/unicorn_arm.c
+++ b/qemu/target-arm/unicorn_arm.c
@@ -64,6 +64,9 @@ int arm_reg_read(struct uc_struct *uc, unsigned int *regs, void **vals, int coun
             *(float64 *)value = ARM_CPU(uc, mycpu)->env.vfp.regs[regid - UC_ARM_REG_D0];
         else {
             switch(regid) {
+                case UC_ARM_REG_APSR:
+                    *(int32_t *)value = cpsr_read(&ARM_CPU(uc, mycpu)->env) & CPSR_NZCV;
+                    break;
                 case UC_ARM_REG_CPSR:
                     *(int32_t *)value = cpsr_read(&ARM_CPU(uc, mycpu)->env);
                     break;
@@ -109,6 +112,9 @@ int arm_reg_write(struct uc_struct *uc, unsigned int *regs, void* const* vals, i
             ARM_CPU(uc, mycpu)->env.vfp.regs[regid - UC_ARM_REG_D0] = *(float64 *)value;
         else {
             switch(regid) {
+                case UC_ARM_REG_APSR:
+                    cpsr_write(&ARM_CPU(uc, mycpu)->env, *(uint32_t *)value, CPSR_NZCV);
+                    break;
                 case UC_ARM_REG_CPSR:
                     cpsr_write(&ARM_CPU(uc, mycpu)->env, *(uint32_t *)value, ~0);
                     break;
diff --git a/qemu/target-i386/unicorn.c b/qemu/target-i386/unicorn.c
index 1c1aee92..8af79649 100644
--- a/qemu/target-i386/unicorn.c
+++ b/qemu/target-i386/unicorn.c
@@ -17,6 +17,10 @@ static void load_seg_16_helper(CPUX86State *env, int seg, uint32_t selector)
     cpu_x86_load_seg_cache(env, seg, selector, (selector << 4), 0xffff, X86_NON_CS_FLAGS);
 }
 
+
+extern void helper_wrmsr(CPUX86State *env);
+extern void helper_rdmsr(CPUX86State *env);
+
 const int X86_REGS_STORAGE_SIZE = offsetof(CPUX86State, tlb_table);
 
 static void x86_set_pc(struct uc_struct *uc, uint64_t address)
@@ -156,6 +160,49 @@ void x86_reg_reset(struct uc_struct *uc)
     }
 }
 
+static int x86_msr_read(struct uc_struct *uc, uc_x86_msr *msr) 
+{
+    CPUX86State *env = (CPUX86State *)uc->cpu->env_ptr;
+    uint64_t ecx = env->regs[R_ECX];
+    uint64_t eax = env->regs[R_EAX];
+    uint64_t edx = env->regs[R_EDX];
+
+    env->regs[R_ECX] = msr->rid;
+    helper_rdmsr(env);
+
+    msr->value = ((uint32_t)env->regs[R_EAX]) |
+    	((uint64_t)((uint32_t)env->regs[R_EDX]) << 32);
+
+    env->regs[R_EAX] = eax;
+    env->regs[R_ECX] = ecx;
+    env->regs[R_EDX] = edx;
+
+    /* The implementation doesn't throw exception or return an error if there is one, so
+     * we will return 0.  */
+    return 0;
+}
+
+static int x86_msr_write(struct uc_struct *uc, uc_x86_msr *msr)
+{
+    CPUX86State *env = (CPUX86State *)uc->cpu->env_ptr;
+    uint64_t ecx = env->regs[R_ECX];
+    uint64_t eax = env->regs[R_EAX];
+    uint64_t edx = env->regs[R_EDX];
+
+    env->regs[R_ECX] = msr->rid;
+    env->regs[R_EAX] = (unsigned int)msr->value;
+    env->regs[R_EDX] = (unsigned int)(msr->value >> 32);
+    helper_wrmsr(env);
+
+    env->regs[R_ECX] = ecx;
+    env->regs[R_EAX] = eax;
+    env->regs[R_EDX] = edx;
+
+    /* The implementation doesn't throw exception or return an error if there is one, so
+     * we will return 0.  */
+    return 0;
+}
+
 int x86_reg_read(struct uc_struct *uc, unsigned int *regs, void **vals, int count)
 {
     CPUState *mycpu = uc->cpu;
@@ -401,6 +448,9 @@ int x86_reg_read(struct uc_struct *uc, unsigned int *regs, void **vals, int coun
                         ((uc_x86_mmr *)value)->selector = (uint16_t)X86_CPU(uc, mycpu)->env.tr.selector;
                         ((uc_x86_mmr *)value)->flags = X86_CPU(uc, mycpu)->env.tr.flags;
                         break;
+                    case UC_X86_REG_MSR:
+                        x86_msr_read(uc, (uc_x86_msr *)value);
+                        break;
                 }
                 break;
 
@@ -680,6 +730,9 @@ int x86_reg_read(struct uc_struct *uc, unsigned int *regs, void **vals, int coun
                         ((uc_x86_mmr *)value)->selector = (uint16_t)X86_CPU(uc, mycpu)->env.tr.selector;
                         ((uc_x86_mmr *)value)->flags = X86_CPU(uc, mycpu)->env.tr.flags;
                         break;
+                    case UC_X86_REG_MSR:
+                        x86_msr_read(uc, (uc_x86_msr *)value);
+                        break;
                 }
                 break;
 #endif
@@ -924,6 +977,9 @@ int x86_reg_write(struct uc_struct *uc, unsigned int *regs, void *const *vals, i
                         X86_CPU(uc, mycpu)->env.tr.selector = (uint16_t)((uc_x86_mmr *)value)->selector;
                         X86_CPU(uc, mycpu)->env.tr.flags = ((uc_x86_mmr *)value)->flags;
                         break;
+                    case UC_X86_REG_MSR:
+                        x86_msr_write(uc, (uc_x86_msr *)value);
+                        break;
                 }
                 break;
 
@@ -1213,6 +1269,9 @@ int x86_reg_write(struct uc_struct *uc, unsigned int *regs, void *const *vals, i
                         X86_CPU(uc, mycpu)->env.tr.selector = (uint16_t)((uc_x86_mmr *)value)->selector;
                         X86_CPU(uc, mycpu)->env.tr.flags = ((uc_x86_mmr *)value)->flags;
                         break;
+                    case UC_X86_REG_MSR:
+                        x86_msr_write(uc, (uc_x86_msr *)value);
+                        break;
                 }
                 break;
 #endif
@@ -1257,3 +1316,5 @@ void x86_uc_init(struct uc_struct* uc)
     uc->stop_interrupt = x86_stop_interrupt;
     uc_common_init(uc);
 }
+
+/* vim: set ts=4 sts=4 sw=4 et:  */
diff --git a/qemu/target-mips/unicorn.c b/qemu/target-mips/unicorn.c
index 86dd5582..3d821528 100644
--- a/qemu/target-mips/unicorn.c
+++ b/qemu/target-mips/unicorn.c
@@ -18,6 +18,12 @@ const int MIPS_REGS_STORAGE_SIZE = offsetof(CPUMIPSState, tlb_table);
 #endif
 #endif
 
+#ifdef TARGET_MIPS64
+typedef uint64_t mipsreg_t;
+#else
+typedef uint32_t mipsreg_t;
+#endif
+
 static uint64_t mips_mem_redirect(uint64_t address)
 {
     // kseg0 range masks off high address bit
@@ -91,7 +97,7 @@ int mips_reg_read(struct uc_struct *uc, unsigned int *regs, void **vals, int cou
             switch(regid) {
                 default: break;
                 case UC_MIPS_REG_PC:
-                         *(int32_t *)value = MIPS_CPU(uc, mycpu)->env.active_tc.PC;
+                         *(mipsreg_t *)value = MIPS_CPU(uc, mycpu)->env.active_tc.PC;
                          break;
             }
         }
@@ -109,12 +115,12 @@ int mips_reg_write(struct uc_struct *uc, unsigned int *regs, void *const *vals,
         unsigned int regid = regs[i];
         const void *value = vals[i];
         if (regid >= UC_MIPS_REG_0 && regid <= UC_MIPS_REG_31)
-            MIPS_CPU(uc, mycpu)->env.active_tc.gpr[regid - UC_MIPS_REG_0] = *(uint32_t *)value;
+            MIPS_CPU(uc, mycpu)->env.active_tc.gpr[regid - UC_MIPS_REG_0] = *(mipsreg_t *)value;
         else {
             switch(regid) {
                 default: break;
                 case UC_MIPS_REG_PC:
-                         MIPS_CPU(uc, mycpu)->env.active_tc.PC = *(uint32_t *)value;
+                         MIPS_CPU(uc, mycpu)->env.active_tc.PC = *(mipsreg_t *)value;
                          // force to quit execution and flush TB
                          uc->quit_request = true;
                          uc_emu_stop(uc);
diff --git a/qemu/tcg/tcg.c b/qemu/tcg/tcg.c
index 56e1da4b..c82e33b7 100644
--- a/qemu/tcg/tcg.c
+++ b/qemu/tcg/tcg.c
@@ -717,8 +717,8 @@ void tcg_gen_callN(TCGContext *s, void *func, TCGArg ret,
             int is_64bit = sizemask & (1 << (i+1)*2);
             if (is_64bit) {
                 TCGv_i64 orig = MAKE_TCGV_I64(args[i]);
-                TCGv_i32 h = tcg_temp_new_i32();
-                TCGv_i32 l = tcg_temp_new_i32();
+                TCGv_i32 h = tcg_temp_new_i32(s);
+                TCGv_i32 l = tcg_temp_new_i32(s);
                 tcg_gen_extr_i64_i32(l, h, orig);
                 split_args[real_args++] = GET_TCGV_I32(h);
                 split_args[real_args++] = GET_TCGV_I32(l);
@@ -738,9 +738,9 @@ void tcg_gen_callN(TCGContext *s, void *func, TCGArg ret,
             TCGv_i64 temp = tcg_temp_new_i64(s);
             TCGv_i64 orig = MAKE_TCGV_I64(args[i]);
             if (is_signed) {
-                tcg_gen_ext32s_i64(temp, orig);
+                tcg_gen_ext32s_i64(s, temp, orig);
             } else {
-                tcg_gen_ext32u_i64(temp, orig);
+                tcg_gen_ext32u_i64(s, temp, orig);
             }
             args[i] = GET_TCGV_I64(temp);
         }
@@ -834,8 +834,8 @@ void tcg_gen_callN(TCGContext *s, void *func, TCGArg ret,
         if (is_64bit) {
             TCGv_i32 h = MAKE_TCGV_I32(args[real_args++]);
             TCGv_i32 l = MAKE_TCGV_I32(args[real_args++]);
-            tcg_temp_free_i32(h);
-            tcg_temp_free_i32(l);
+            tcg_temp_free_i32(s, h);
+            tcg_temp_free_i32(s, l);
         } else {
             real_args++;
         }
@@ -845,15 +845,15 @@ void tcg_gen_callN(TCGContext *s, void *func, TCGArg ret,
            Note that describing these as TCGv_i64 eliminates an unnecessary
            zero-extension that tcg_gen_concat_i32_i64 would create.  */
         tcg_gen_concat32_i64(MAKE_TCGV_I64(ret), retl, reth);
-        tcg_temp_free_i64(retl);
-        tcg_temp_free_i64(reth);
+        tcg_temp_free_i64(s, retl);
+        tcg_temp_free_i64(s, reth);
     }
 #elif defined(TCG_TARGET_EXTEND_ARGS) && TCG_TARGET_REG_BITS == 64
     for (i = 0; i < nargs; ++i) {
         int is_64bit = sizemask & (1 << (i+1)*2);
         if (!is_64bit) {
             TCGv_i64 temp = MAKE_TCGV_I64(args[i]);
-            tcg_temp_free_i64(temp);
+            tcg_temp_free_i64(s, temp);
         }
     }
 #endif /* TCG_TARGET_EXTEND_ARGS */
diff --git a/qemu/x86_64.h b/qemu/x86_64.h
index df86439d..71025141 100644
--- a/qemu/x86_64.h
+++ b/qemu/x86_64.h
@@ -2,6 +2,7 @@
 #ifndef UNICORN_AUTOGEN_X86_64_H
 #define UNICORN_AUTOGEN_X86_64_H
 #define aarch64_tb_set_jmp_target aarch64_tb_set_jmp_target_x86_64
+#define ppc_tb_set_jmp_target ppc_tb_set_jmp_target_x86_64
 #define use_idiv_instructions_rt use_idiv_instructions_rt_x86_64
 #define tcg_target_deposit_valid tcg_target_deposit_valid_x86_64
 #define helper_power_down helper_power_down_x86_64
diff --git a/samples/sample_all.sh b/samples/sample_all.sh
index 502171d4..026e0107 100755
--- a/samples/sample_all.sh
+++ b/samples/sample_all.sh
@@ -42,3 +42,18 @@ if test -e $DIR/sample_m68k; then
   echo "=========================="
   $DIR/sample_m68k
 fi
+
+if test -e $DIR/mem_apis; then
+  echo "=========================="
+  $DIR/mem_apis
+fi
+
+if test -e $DIR/sample_batch_reg; then
+  echo "=========================="
+  $DIR/sample_batch_reg
+fi
+
+if test -e $DIR/sample_x86_32_gdt_and_seg_regs; then
+  echo "=========================="
+  $DIR/sample_x86_32_gdt_and_seg_regs
+fi
diff --git a/tests/regress/.gitignore b/tests/regress/.gitignore
index 00144a35..ca97ad9b 100644
--- a/tests/regress/.gitignore
+++ b/tests/regress/.gitignore
@@ -1,5 +1,6 @@
 !*.c
 
+arm_enable_vfp
 map_crash
 sigill
 sigill2
diff --git a/tests/regress/eflags_noset.c b/tests/regress/eflags_noset.c
index 9d4f2291..7f1b9710 100644
--- a/tests/regress/eflags_noset.c
+++ b/tests/regress/eflags_noset.c
@@ -23,26 +23,32 @@ typedef uint64_t puint;
 uint32_t realEflags()
 {
     puint val = 0;
-    puint i = 0xFFFFFEFF; //attempt to set ALL bits except trap flag.
 
 #if defined(__i386__)
+    puint i = 0xFFFFFEFF; //attempt to set ALL bits except trap flag.
+
     __asm__("pushf\n\t"
     "push %0\n\t"
     "popf\n\t" 
     "pushf\n\t"
     "pop %0\n\t"
     "popf"
-#else
+    : "=r"(val)
+    : "r"(i)
+    : "%0");
+#elif defined(__x86_64__)
+    puint i = 0xFFFFFEFF; //attempt to set ALL bits except trap flag.
+
     __asm__("pushfq\n\t"
     "pushq %0\n\t"
     "popfq\n\t" 
     "pushfq\n\t"
     "popq %0\n\t"
     "popfq"
-#endif
     : "=r"(val)
     : "r"(i)
     : "%0"); 
+#endif
 
     printf("Real system eflags: 0x%08"PRIX3264"\n", val);
 
@@ -51,7 +57,7 @@ uint32_t realEflags()
 
 static void VM_exec()
 {
-
+#if defined(__i386__) || defined(__x86_64__)
     uc_engine *uc;
     uc_err err;
     unsigned int r_eax, eflags, r_esp, realflags = 0;
@@ -110,6 +116,7 @@ static void VM_exec()
     realflags = realEflags();
 
     assert(r_eax == realflags);
+#endif
 }
 
 int main(int argc, char *argv[])
diff --git a/uc.c b/uc.c
index 65e82087..be0d288e 100644
--- a/uc.c
+++ b/uc.c
@@ -380,14 +380,12 @@ uc_err uc_reg_read(uc_engine *uc, int regid, void *value)
     return uc_reg_read_batch(uc, &regid, &value, 1);
 }
 
-
 UNICORN_EXPORT
 uc_err uc_reg_write(uc_engine *uc, int regid, const void *value)
 {
     return uc_reg_write_batch(uc, &regid, (void *const *)&value, 1);
 }
 
-
 // check if a memory area is mapped
 // this is complicated because an area can overlap adjacent blocks
 static bool check_mem_area(uc_engine *uc, uint64_t address, size_t size)