1 {-# LANGUAGE OverloadedStrings #-}
2 {-# LANGUAGE ForeignFunctionInterface #-}
3 module Mate.X86CodeGen where
5 import Prelude hiding (and, div)
7 import Data.BinaryState
10 import Data.List (genericLength)
11 import qualified Data.Map as M
12 import qualified Data.ByteString.Lazy as B
14 import Control.Applicative
16 import Foreign hiding (xor)
17 import Foreign.C.Types
19 import qualified JVM.Assembler as J
20 import JVM.Assembler hiding (Instruction)
24 import Harpy.X86Disassembler
26 import Mate.BasicBlocks
27 import Mate.NativeSizes
31 import Mate.ClassHierarchy
32 import {-# SOURCE #-} Mate.MethodPool
36 foreign import ccall "&mallocObjectGC"
37 mallocObjectAddr :: FunPtr (Int -> IO CPtrdiff)
39 type EntryPoint = Ptr Word8
40 type EntryPointOffset = Int
41 type PatchInfo = (BlockID, EntryPointOffset)
43 type BBStarts = M.Map BlockID Int
45 type CompileInfo = (EntryPoint, Int, TrapMap)
48 emitFromBB :: Class Direct -> MethodInfo -> RawMethod -> CodeGen e JpcNpcMap (CompileInfo, [Instruction])
49 emitFromBB cls miThis method = do
50 let keys = M.keys hmap
51 llmap <- mapM (newNamedLabel . (++) "bb_" . show) keys
52 let lmap = zip keys llmap
56 sub esp (fromIntegral (rawLocals method) * ptrSize :: Word32)
57 calls <- M.fromList . catMaybes . concat <$> mapM (efBB lmap) keys
60 return ((ep, end, calls), d)
62 hmap = rawMapBB method
64 getLabel :: BlockID -> [(BlockID, Label)] -> Label
65 getLabel bid [] = error $ "label " ++ show bid ++ " not found"
66 getLabel i ((x,l):xs) = if i==x then l else getLabel i xs
68 efBB :: [(BlockID, Label)] -> BlockID -> CodeGen e JpcNpcMap [(Maybe (Word32, TrapCause))]
70 defineLabel $ getLabel bid lmap
71 retval <- mapM emit'' $ code bb
74 -- TODO(bernhard): le dirty hax. see java/lang/Integer.toString(int, int)
81 forceRegDump :: CodeGen e s ()
84 mov esi (0x13371234 :: Word32)
88 getCurrentOffset :: CodeGen e s Word32
90 ep <- (fromIntegral . ptrToIntPtr) <$> getEntryPoint
91 offset <- fromIntegral <$> getCodeOffset
94 emitInvoke :: Word16 -> Bool -> CodeGen e s (Maybe (Word32, TrapCause))
95 emitInvoke cpidx hasThis = do
96 let l = buildMethodID cls cpidx
97 newNamedLabel (show l) >>= defineLabel
98 -- like: call $0x01234567
99 calladdr <- emitSigIllTrap 5
100 let patcher reip = do
101 (entryAddr, _) <- liftIO $ getMethodEntry l
102 call (fromIntegral (entryAddr - (reip + 5)) :: NativeWord)
104 -- discard arguments on stack
105 let argcnt = ((if hasThis then 1 else 0) + methodGetArgsCount (methodNameTypeByIdx cls cpidx)) * ptrSize
106 when (argcnt > 0) (add esp argcnt)
107 -- push result on stack if method has a return value
108 when (methodHaveReturnValue cls cpidx) (push eax)
109 return $ Just (calladdr, StaticMethod patcher)
111 virtualCall :: Word16 -> Bool -> CodeGen e s (Maybe (Word32, TrapCause))
112 virtualCall cpidx isInterface = do
113 let mi@(MethodInfo methodname objname msig@(MethodSignature args _)) = buildMethodID cls cpidx
114 newNamedLabel (show mi) >>= defineLabel
115 -- get method offset for call @ runtime
116 let offset = if isInterface
117 then getInterfaceMethodOffset objname methodname (encode msig)
118 else getMethodOffset objname (methodname `B.append` encode msig)
119 let argsLen = genericLength args
120 -- objref lives somewhere on the argument stack
121 mov ebx (Disp (argsLen * ptrSize), esp)
123 mov ebx (Disp 0, ebx) -- get method-table-ptr, keep it in ebx
124 -- get method-table-ptr (or interface-table-ptr)
125 mov eax (Disp 0, ebx)
126 -- make actual (indirect) call
127 calladdr <- getCurrentOffset
128 -- will be patched to this: call (Disp 0xXXXXXXXX, eax)
130 -- discard arguments on stack (`+1' for "this")
131 let argcnt = ptrSize * (1 + methodGetArgsCount (methodNameTypeByIdx cls cpidx))
132 when (argcnt > 0) (add esp argcnt)
133 -- push result on stack if method has a return value
134 when (methodHaveReturnValue cls cpidx) (push eax)
135 -- note, that "mi" has the wrong class reference here.
136 -- we figure that out at run-time, in the methodpool,
137 -- depending on the method-table-ptr
138 return $ Just (calladdr, VirtualCall isInterface mi offset)
140 emit'' :: J.Instruction -> CodeGen e JpcNpcMap (Maybe (Word32, TrapCause))
142 ep <- (fromIntegral . ptrToIntPtr) <$> getEntryPoint
144 setState (M.insert ep bid jpcrpc)
145 newNamedLabel ("jvm_insn: " ++ show insn) >>= defineLabel >> emit' insn
147 emit' :: J.Instruction -> CodeGen e s (Maybe (Word32, TrapCause))
148 emit' (INVOKESPECIAL cpidx) = emitInvoke cpidx True
149 emit' (INVOKESTATIC cpidx) = emitInvoke cpidx False
150 emit' (INVOKEINTERFACE cpidx _) = virtualCall cpidx True
151 emit' (INVOKEVIRTUAL cpidx) = virtualCall cpidx False
153 emit' (PUTSTATIC cpidx) = do
155 trapaddr <- getCurrentOffset
156 mov (Addr 0x00000000) eax -- it's a trap
157 return $ Just (trapaddr, StaticField $ buildStaticFieldID cls cpidx)
158 emit' (GETSTATIC cpidx) = do
159 trapaddr <- getCurrentOffset
160 mov eax (Addr 0x00000000) -- it's a trap
162 return $ Just (trapaddr, StaticField $ buildStaticFieldID cls cpidx)
164 emit' (GETFIELD x) = do
165 pop eax -- this pointer
166 -- like: 099db064 ff b0 e4 14 00 00 pushl 5348(%eax)
167 trapaddr <- emitSigIllTrap 6
168 let patcher reip = do
169 let (cname, fname) = buildFieldOffset cls x
170 offset <- liftIO $ fromIntegral <$> getFieldOffset cname fname
171 push32RelEax (Disp offset) -- get field
173 return $ Just (trapaddr, ObjectField patcher)
174 emit' (PUTFIELD x) = do
175 pop ebx -- value to write
176 pop eax -- this pointer
177 -- like: 4581fc6b 89 98 30 7b 00 00 movl %ebx,31536(%eax)
178 trapaddr <- emitSigIllTrap 6
179 let patcher reip = do
180 let (cname, fname) = buildFieldOffset cls x
181 offset <- liftIO $ fromIntegral <$> getFieldOffset cname fname
182 mov32RelEbxEax (Disp offset) -- set field
184 return $ Just (trapaddr, ObjectField patcher)
186 emit' (INSTANCEOF cpidx) = do
188 -- place something like `mov edx $mtable_of_objref' instead
189 trapaddr <- emitSigIllTrap 4
191 let patcher reax reip = do
193 let classname = buildClassID cls cpidx
194 check <- liftIO $ isInstanceOf (fromIntegral reax) classname
196 then push (1 :: Word32)
197 else push (0 :: Word32)
199 return $ Just (trapaddr, InstanceOf patcher)
200 emit' (NEW objidx) = do
201 let objname = buildClassID cls objidx
202 -- place something like `push $objsize' instead
203 trapaddr <- emitSigIllTrap 5
205 -- 0x13371337 is just a placeholder; will be replaced with mtable ptr
206 mov (Disp 0, eax) (0x13371337 :: Word32)
207 mov (Disp 4, eax) (0x1337babe :: Word32)
208 let patcher reip = do
209 objsize <- liftIO $ getObjectSize objname
212 mtable <- liftIO $ getMethodTable objname
213 mov (Disp 0, eax) mtable
214 mov (Disp 4, eax) (0x1337babe :: Word32)
216 return $ Just (trapaddr, NewObject patcher)
219 trapaddr <- emitSigIllTrap 2
220 let patcher resp reip = do
221 (_, jnmap) <- liftIO $ getMethodEntry miThis
222 error "no athrow for you, sorry"
225 return $ Just (trapaddr, ThrowException patcher)
227 emit' insn = emit insn >> return Nothing
229 emit :: J.Instruction -> CodeGen e s ()
230 emit POP = add esp (ptrSize :: Word32) -- drop value
231 emit DUP = push (Disp 0, esp)
232 emit DUP_X1 = do pop eax; pop ebx; push eax; push ebx; push eax
233 emit DUP_X2 = do pop eax; pop ebx; pop ecx; push eax; push ecx; push ebx; push eax
234 emit AASTORE = emit IASTORE
238 add ebx (1 :: Word32)
240 mov (ecx, ebx, S4) eax
244 add ebx (1 :: Word32)
246 mov (ecx, ebx, S1) eax -- TODO(bernhard): char is two byte
247 emit AALOAD = emit IALOAD
250 add ebx (1 :: Word32)
255 add ebx (1 :: Word32)
257 push (ecx, ebx, S1) -- TODO(bernhard): char is two byte
258 emit ARRAYLENGTH = do
261 emit (ANEWARRAY _) = emit (NEWARRAY 10) -- 10 == T_INT
262 emit (NEWARRAY typ) = do
263 let tsize = case decodeS (0 :: Integer) (B.pack [typ]) of
266 _ -> error "newarray: type not implemented yet"
267 -- get length from stack, but leave it there
268 mov eax (Disp 0, esp)
269 mov ebx (tsize :: Word32)
270 -- multiple amount with native size of one element
271 mul ebx -- result is in eax
272 add eax (ptrSize :: Word32) -- for "length" entry
273 -- push amount of bytes to allocate
276 pop eax -- ref to arraymemory
278 mov (Disp 0, eax) ebx -- store length at offset 0
279 push eax -- push ref again
281 emit (CHECKCAST _) = nop -- TODO(bernhard): ...
284 and eax (0x000000ff :: Word32)
286 emit (BIPUSH val) = push (fromIntegral val :: Word32)
287 emit (SIPUSH val) = push (fromIntegral (fromIntegral val :: Int16) :: Word32)
288 emit ACONST_NULL = push (0 :: Word32)
289 emit (ICONST_M1) = push ((-1) :: Word32)
290 emit (ICONST_0) = push (0 :: Word32)
291 emit (ICONST_1) = push (1 :: Word32)
292 emit (ICONST_2) = push (2 :: Word32)
293 emit (ICONST_3) = push (3 :: Word32)
294 emit (ICONST_4) = push (4 :: Word32)
295 emit (ICONST_5) = push (5 :: Word32)
297 emit (ALOAD_ x) = emit (ILOAD_ x)
298 emit (ILOAD_ x) = emit (ILOAD $ cArgs_ x)
299 emit (ALOAD x) = emit (ILOAD x)
300 emit (ILOAD x) = push (Disp (cArgs x), ebp)
302 emit (ASTORE_ x) = emit (ISTORE_ x)
303 emit (ISTORE_ x) = emit (ISTORE $ cArgs_ x)
304 emit (ASTORE x) = emit (ISTORE x)
307 mov (Disp (cArgs x), ebp) eax
309 emit (LDC1 x) = emit (LDC2 $ fromIntegral x)
311 value <- case constsPool cls M.! x of
312 (CString s) -> liftIO $ getUniqueStringAddr s
313 (CInteger i) -> liftIO $ return i
314 e -> error $ "LDCI... missing impl.: " ++ show e
317 emit IADD = do pop ebx; pop eax; add eax ebx; push eax
318 emit ISUB = do pop ebx; pop eax; sub eax ebx; push eax
319 emit IMUL = do pop ebx; pop eax; mul ebx; push eax
320 emit IDIV = do pop ebx; pop eax; xor edx edx; div ebx; push eax
321 emit IREM = do pop ebx; pop eax; xor edx edx; div ebx; push edx
322 emit IXOR = do pop ebx; pop eax; xor eax ebx; push eax
323 emit IUSHR = do pop ecx; pop eax; sar eax cl; push eax
324 emit INEG = do pop eax; neg eax; push eax
326 add (Disp (cArgs x), ebp) (s8_w32 imm)
328 emit (IFNONNULL x) = emit (IF C_NE x)
329 emit (IFNULL x) = emit (IF C_EQ x)
330 emit (IF_ACMP cond x) = emit (IF_ICMP cond x)
331 emit (IF_ICMP cond _) = do
334 cmp ebx eax -- intel syntax is swapped (TODO(bernhard): test that plz)
337 emit (IF cond _) = do
339 cmp eax (0 :: Word32) -- TODO(bernhard): test that plz
343 let sid = case successor bb of OneTarget t -> t; _ -> error "bad"
344 jmp $ getLabel sid lmap
346 emit RETURN = do mov esp ebp; pop ebp; ret
347 emit ARETURN = emit IRETURN
348 emit IRETURN = do pop eax; emit RETURN
349 emit invalid = error $ "insn not implemented yet: " ++ show invalid
351 emitIF :: CMP -> CodeGen e s ()
353 sid = case successor bb of TwoTarget _ t -> t; _ -> error "bad"
354 l = getLabel sid lmap
355 sid2 = case successor bb of TwoTarget t _ -> t; _ -> error "bad"
356 l2 = getLabel sid2 lmap
359 C_EQ -> je l; C_NE -> jne l
360 C_LT -> jl l; C_GT -> jg l
361 C_GE -> jge l; C_LE -> jle l
362 -- TODO(bernhard): ugly workaround, to get broken emitBB working
363 -- (it didn't work for gnu/classpath/SystemProperties.java)
366 emitSigIllTrap :: Int -> CodeGen e s NativeWord
367 emitSigIllTrap traplen = do
368 when (traplen < 2) (error "emitSigIllTrap: trap len too short")
369 trapaddr <- getCurrentOffset
370 -- 0xffff causes SIGILL
371 emit8 (0xff :: Word8); emit8 (0xff :: Word8)
372 -- fill rest up with NOPs
373 sequence_ [nop | _ <- [1 .. (traplen - 2)]]
377 -- for locals we use a different storage
378 cArgs :: Word8 -> Word32
379 cArgs x = ptrSize * (argcount - x' + isLocal)
382 argcount = rawArgCount method
383 isLocal = if x' >= argcount then (-1) else 1
385 cArgs_ :: IMM -> Word8
386 cArgs_ x = case x of I0 -> 0; I1 -> 1; I2 -> 2; I3 -> 3
389 -- sign extension from w8 to w32 (over s8)
390 -- unfortunately, hs-java is using Word8 everywhere (while
391 -- it should be Int8 actually)
392 s8_w32 :: Word8 -> Word32
393 s8_w32 w8 = fromIntegral s8
394 where s8 = fromIntegral w8 :: Int8
396 callMalloc :: CodeGen e s ()
398 call mallocObjectAddr
399 add esp (ptrSize :: Word32)
403 -- harpy tries to cut immediates (or displacements), if they fit in 8bit.
404 -- however, this is bad for patching so we want to put always 32bit.
407 push32 :: Word32 -> CodeGen e s ()
408 push32 imm32 = emit8 0x68 >> emit32 imm32
411 call32Eax :: Disp -> CodeGen e s ()
412 call32Eax (Disp disp32) = emit8 0xff >> emit8 0x90 >> emit32 disp32
415 push32RelEax :: Disp -> CodeGen e s ()
416 push32RelEax (Disp disp32) = emit8 0xff >> emit8 0xb0 >> emit32 disp32
418 -- mov %ebx, disp32(%eax)
419 mov32RelEbxEax :: Disp -> CodeGen e s ()
420 mov32RelEbxEax (Disp disp32) = emit8 0x89 >> emit8 0x98 >> emit32 disp32