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 -> RawMethod -> CodeGen e s (CompileInfo, [Instruction])
49 emitFromBB cls 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 s [(Maybe (Word32, TrapCause))]
70 defineLabel $ getLabel bid lmap
71 ret <- 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 s (Maybe (Word32, TrapCause))
141 emit'' insn = newNamedLabel ("jvm_insn: " ++ show insn) >>= defineLabel >> emit' insn
143 emit' :: J.Instruction -> CodeGen e s (Maybe (Word32, TrapCause))
144 emit' (INVOKESPECIAL cpidx) = emitInvoke cpidx True
145 emit' (INVOKESTATIC cpidx) = emitInvoke cpidx False
146 emit' (INVOKEINTERFACE cpidx _) = virtualCall cpidx True
147 emit' (INVOKEVIRTUAL cpidx) = virtualCall cpidx False
149 emit' (PUTSTATIC cpidx) = do
151 trapaddr <- getCurrentOffset
152 mov (Addr 0x00000000) eax -- it's a trap
153 return $ Just (trapaddr, StaticField $ buildStaticFieldID cls cpidx)
154 emit' (GETSTATIC cpidx) = do
155 trapaddr <- getCurrentOffset
156 mov eax (Addr 0x00000000) -- it's a trap
158 return $ Just (trapaddr, StaticField $ buildStaticFieldID cls cpidx)
160 emit' (GETFIELD x) = do
161 pop eax -- this pointer
162 -- like: 099db064 ff b0 e4 14 00 00 pushl 5348(%eax)
163 trapaddr <- emitSigIllTrap 6
164 let patcher reip = do
165 let (cname, fname) = buildFieldOffset cls x
166 offset <- liftIO $ fromIntegral <$> getFieldOffset cname fname
167 push32RelEax (Disp offset) -- get field
169 return $ Just (trapaddr, ObjectField patcher)
170 emit' (PUTFIELD x) = do
171 pop ebx -- value to write
172 pop eax -- this pointer
173 -- like: 4581fc6b 89 98 30 7b 00 00 movl %ebx,31536(%eax)
174 trapaddr <- emitSigIllTrap 6
175 let patcher reip = do
176 let (cname, fname) = buildFieldOffset cls x
177 offset <- liftIO $ fromIntegral <$> getFieldOffset cname fname
178 mov32RelEbxEax (Disp offset) -- set field
180 return $ Just (trapaddr, ObjectField patcher)
182 emit' (INSTANCEOF cpidx) = do
184 -- place something like `mov edx $mtable_of_objref' instead
185 trapaddr <- emitSigIllTrap 4
187 let patcher reax reip = do
189 let classname = buildClassID cls cpidx
190 check <- liftIO $ isInstanceOf (fromIntegral reax) classname
192 then push (1 :: Word32)
193 else push (0 :: Word32)
195 return $ Just (trapaddr, InstanceOf patcher)
196 emit' (NEW objidx) = do
197 let objname = buildClassID cls objidx
198 -- place something like `push $objsize' instead
199 trapaddr <- emitSigIllTrap 5
201 -- 0x13371337 is just a placeholder; will be replaced with mtable ptr
202 mov (Disp 0, eax) (0x13371337 :: Word32)
203 mov (Disp 4, eax) (0x1337babe :: Word32)
204 let patcher reip = do
205 objsize <- liftIO $ getObjectSize objname
208 mtable <- liftIO $ getMethodTable objname
209 mov (Disp 0, eax) mtable
210 mov (Disp 4, eax) (0x1337babe :: Word32)
212 return $ Just (trapaddr, NewObject patcher)
215 trapaddr <- emitSigIllTrap 2
216 let patcher resp reip = do
217 error "no athrow for you, sorry"
220 return $ Just (trapaddr, ThrowException patcher)
222 emit' insn = emit insn >> return Nothing
224 emit :: J.Instruction -> CodeGen e s ()
225 emit POP = add esp (ptrSize :: Word32) -- drop value
226 emit DUP = push (Disp 0, esp)
227 emit DUP_X1 = do pop eax; pop ebx; push eax; push ebx; push eax
228 emit DUP_X2 = do pop eax; pop ebx; pop ecx; push eax; push ecx; push ebx; push eax
229 emit AASTORE = emit IASTORE
233 add ebx (1 :: Word32)
235 mov (ecx, ebx, S4) eax
239 add ebx (1 :: Word32)
241 mov (ecx, ebx, S1) eax -- TODO(bernhard): char is two byte
242 emit AALOAD = emit IALOAD
245 add ebx (1 :: Word32)
250 add ebx (1 :: Word32)
252 push (ecx, ebx, S1) -- TODO(bernhard): char is two byte
253 emit ARRAYLENGTH = do
256 emit (ANEWARRAY _) = emit (NEWARRAY 10) -- 10 == T_INT
257 emit (NEWARRAY typ) = do
258 let tsize = case decodeS (0 :: Integer) (B.pack [typ]) of
261 _ -> error "newarray: type not implemented yet"
262 -- get length from stack, but leave it there
263 mov eax (Disp 0, esp)
264 mov ebx (tsize :: Word32)
265 -- multiple amount with native size of one element
266 mul ebx -- result is in eax
267 add eax (ptrSize :: Word32) -- for "length" entry
268 -- push amount of bytes to allocate
271 pop eax -- ref to arraymemory
273 mov (Disp 0, eax) ebx -- store length at offset 0
274 push eax -- push ref again
276 emit (CHECKCAST _) = nop -- TODO(bernhard): ...
279 and eax (0x000000ff :: Word32)
281 emit (BIPUSH val) = push (fromIntegral val :: Word32)
282 emit (SIPUSH val) = push (fromIntegral (fromIntegral val :: Int16) :: Word32)
283 emit ACONST_NULL = push (0 :: Word32)
284 emit (ICONST_M1) = push ((-1) :: Word32)
285 emit (ICONST_0) = push (0 :: Word32)
286 emit (ICONST_1) = push (1 :: Word32)
287 emit (ICONST_2) = push (2 :: Word32)
288 emit (ICONST_3) = push (3 :: Word32)
289 emit (ICONST_4) = push (4 :: Word32)
290 emit (ICONST_5) = push (5 :: Word32)
292 emit (ALOAD_ x) = emit (ILOAD_ x)
293 emit (ILOAD_ x) = emit (ILOAD $ cArgs_ x)
294 emit (ALOAD x) = emit (ILOAD x)
295 emit (ILOAD x) = push (Disp (cArgs x), ebp)
297 emit (ASTORE_ x) = emit (ISTORE_ x)
298 emit (ISTORE_ x) = emit (ISTORE $ cArgs_ x)
299 emit (ASTORE x) = emit (ISTORE x)
302 mov (Disp (cArgs x), ebp) eax
304 emit (LDC1 x) = emit (LDC2 $ fromIntegral x)
306 value <- case constsPool cls M.! x of
307 (CString s) -> liftIO $ getUniqueStringAddr s
308 (CInteger i) -> liftIO $ return i
309 e -> error $ "LDCI... missing impl.: " ++ show e
312 emit IADD = do pop ebx; pop eax; add eax ebx; push eax
313 emit ISUB = do pop ebx; pop eax; sub eax ebx; push eax
314 emit IMUL = do pop ebx; pop eax; mul ebx; push eax
315 emit IDIV = do pop ebx; pop eax; xor edx edx; div ebx; push eax
316 emit IREM = do pop ebx; pop eax; xor edx edx; div ebx; push edx
317 emit IXOR = do pop ebx; pop eax; xor eax ebx; push eax
318 emit IUSHR = do pop ecx; pop eax; sar eax cl; push eax
319 emit INEG = do pop eax; neg eax; push eax
321 add (Disp (cArgs x), ebp) (s8_w32 imm)
323 emit (IFNONNULL x) = emit (IF C_NE x)
324 emit (IFNULL x) = emit (IF C_EQ x)
325 emit (IF_ACMP cond x) = emit (IF_ICMP cond x)
326 emit (IF_ICMP cond _) = do
329 cmp ebx eax -- intel syntax is swapped (TODO(bernhard): test that plz)
332 emit (IF cond _) = do
334 cmp eax (0 :: Word32) -- TODO(bernhard): test that plz
338 let sid = case successor bb of OneTarget t -> t; _ -> error "bad"
339 jmp $ getLabel sid lmap
341 emit RETURN = do mov esp ebp; pop ebp; ret
342 emit ARETURN = emit IRETURN
343 emit IRETURN = do pop eax; emit RETURN
344 emit invalid = error $ "insn not implemented yet: " ++ show invalid
346 emitIF :: CMP -> CodeGen e s ()
348 sid = case successor bb of TwoTarget _ t -> t; _ -> error "bad"
349 l = getLabel sid lmap
350 sid2 = case successor bb of TwoTarget t _ -> t; _ -> error "bad"
351 l2 = getLabel sid2 lmap
354 C_EQ -> je l; C_NE -> jne l
355 C_LT -> jl l; C_GT -> jg l
356 C_GE -> jge l; C_LE -> jle l
357 -- TODO(bernhard): ugly workaround, to get broken emitBB working
358 -- (it didn't work for gnu/classpath/SystemProperties.java)
361 emitSigIllTrap :: Int -> CodeGen e s NativeWord
362 emitSigIllTrap traplen = do
363 when (traplen < 2) (error "emitSigIllTrap: trap len too short")
364 trapaddr <- getCurrentOffset
365 -- 0xffff causes SIGILL
366 emit8 (0xff :: Word8); emit8 (0xff :: Word8)
367 -- fill rest up with NOPs
368 sequence_ [nop | _ <- [1 .. (traplen - 2)]]
372 -- for locals we use a different storage
373 cArgs :: Word8 -> Word32
374 cArgs x = ptrSize * (argcount - x' + isLocal)
377 argcount = rawArgCount method
378 isLocal = if x' >= argcount then (-1) else 1
380 cArgs_ :: IMM -> Word8
381 cArgs_ x = case x of I0 -> 0; I1 -> 1; I2 -> 2; I3 -> 3
384 -- sign extension from w8 to w32 (over s8)
385 -- unfortunately, hs-java is using Word8 everywhere (while
386 -- it should be Int8 actually)
387 s8_w32 :: Word8 -> Word32
388 s8_w32 w8 = fromIntegral s8
389 where s8 = fromIntegral w8 :: Int8
391 callMalloc :: CodeGen e s ()
393 call mallocObjectAddr
394 add esp (ptrSize :: Word32)
398 -- harpy tries to cut immediates (or displacements), if they fit in 8bit.
399 -- however, this is bad for patching so we want to put always 32bit.
402 push32 :: Word32 -> CodeGen e s ()
403 push32 imm32 = emit8 0x68 >> emit32 imm32
406 call32Eax :: Disp -> CodeGen e s ()
407 call32Eax (Disp disp32) = emit8 0xff >> emit8 0x90 >> emit32 disp32
410 push32RelEax :: Disp -> CodeGen e s ()
411 push32RelEax (Disp disp32) = emit8 0xff >> emit8 0xb0 >> emit32 disp32
413 -- mov %ebx, disp32(%eax)
414 mov32RelEbxEax :: Disp -> CodeGen e s ()
415 mov32RelEbxEax (Disp disp32) = emit8 0x89 >> emit8 0x98 >> emit32 disp32