2 {-# LANGUAGE OverloadedStrings #-}
3 {-# LANGUAGE ForeignFunctionInterface #-}
5 module Mate.X86CodeGen where
7 import Prelude hiding (and)
9 import Data.BinaryState
12 import qualified Data.Map as M
13 import qualified Data.Set as S
14 import qualified Data.ByteString.Lazy as B
17 import Foreign hiding (xor)
18 import Foreign.C.Types
20 import qualified JVM.Assembler as J
21 import JVM.Assembler hiding (Instruction)
25 import Harpy.X86Disassembler
27 import Mate.BasicBlocks
34 foreign import ccall "dynamic"
35 code_int :: FunPtr (CInt -> CInt -> IO CInt) -> CInt -> CInt -> IO CInt
37 foreign import ccall "getMallocObjectAddr"
38 getMallocObjectAddr :: CUInt
40 type EntryPoint = Ptr Word8
41 type EntryPointOffset = Int
42 type PatchInfo = (BlockID, EntryPointOffset)
44 type BBStarts = M.Map BlockID Int
46 type CompileInfo = (EntryPoint, BBStarts, Int, TrapMap)
49 emitFromBB :: B.ByteString -> Class Direct -> MapBB -> CodeGen e s (CompileInfo, [Instruction])
50 emitFromBB method cls hmap = do
51 llmap <- sequence [newNamedLabel ("bb_" ++ show x) | (x,_) <- M.toList hmap]
52 let lmap = zip (Prelude.fst $ unzip $ M.toList hmap) llmap
56 -- TODO(bernhard): determine a reasonable value.
57 -- e.g. (locals used) * 4
58 sub esp (0x60 :: Word32)
60 (calls, bbstarts) <- efBB (0, hmap M.! 0) M.empty M.empty lmap
63 return ((ep, bbstarts, end, calls), d)
65 getLabel :: BlockID -> [(BlockID, Label)] -> Label
66 getLabel _ [] = error "label not found!"
67 getLabel i ((x,l):xs) = if i==x then l else getLabel i xs
69 efBB :: (BlockID, BasicBlock) -> TrapMap -> BBStarts -> [(BlockID, Label)] -> CodeGen e s (TrapMap, BBStarts)
70 efBB (bid, bb) calls bbstarts lmap =
71 if M.member bid bbstarts then
72 return (calls, bbstarts)
74 bb_offset <- getCodeOffset
75 let bbstarts' = M.insert bid bb_offset bbstarts
76 defineLabel $ getLabel bid lmap
77 cs <- mapM emit' $ code bb
78 let calls' = calls `M.union` M.fromList (catMaybes cs)
80 Return -> return (calls', bbstarts')
81 FallThrough t -> efBB (t, hmap M.! t) calls' bbstarts' lmap
82 OneTarget t -> efBB (t, hmap M.! t) calls' bbstarts' lmap
84 (calls'', bbstarts'') <- efBB (t1, hmap M.! t1) calls' bbstarts' lmap
85 efBB (t2, hmap M.! t2) calls'' bbstarts'' lmap
86 -- TODO(bernhard): also use metainformation
87 -- TODO(bernhard): implement `emit' as function which accepts a list of
88 -- instructions, so we can use patterns for optimizations
90 getCurrentOffset :: CodeGen e s Word32
93 let w32_ep = (fromIntegral $ ptrToIntPtr ep) :: Word32
94 offset <- getCodeOffset
95 return $ w32_ep + fromIntegral offset
97 emitInvoke :: Word16 -> Bool -> CodeGen e s (Maybe (Word32, TrapInfo))
98 emitInvoke cpidx hasThis = do
99 let l = buildMethodID cls cpidx
100 calladdr <- getCurrentOffset
101 newNamedLabel (show l) >>= defineLabel
102 -- causes SIGILL. in the signal handler we patch it to the acutal call.
103 -- place a nop at the end, therefore the disasm doesn't screw up
104 emit32 (0xffff9090 :: Word32) >> emit8 (0x90 :: Word8)
105 -- discard arguments on stack
106 let argcnt = ((if hasThis then 1 else 0) + methodGetArgsCount cls cpidx) * 4
107 when (argcnt > 0) (add esp argcnt)
108 -- push result on stack if method has a return value
109 when (methodHaveReturnValue cls cpidx) (push eax)
110 -- +2 is for correcting eip in trap context
111 return $ Just (calladdr + 2, MI l)
113 emit' :: J.Instruction -> CodeGen e s (Maybe (Word32, TrapInfo))
114 emit' (INVOKESPECIAL cpidx) = emitInvoke cpidx True
115 emit' (INVOKESTATIC cpidx) = emitInvoke cpidx False
116 emit' (INVOKEINTERFACE cpidx _) = do
117 -- get methodInfo entry
118 let mi@(MethodInfo methodname ifacename msig@(MethodSignature args _)) = buildMethodID cls cpidx
119 newNamedLabel (show mi) >>= defineLabel
120 -- objref lives somewhere on the argument stack
121 mov eax (Disp ((*4) $ fromIntegral $ length args), esp)
122 -- get method-table-ptr, keep it in eax (for trap handling)
123 mov eax (Disp 0, eax)
124 -- get interface-table-ptr
125 mov ebx (Disp 0, eax)
127 offset <- liftIO $ getInterfaceMethodOffset ifacename methodname (encode msig)
128 -- make actual (indirect) call
129 calladdr <- getCurrentOffset
130 call (Disp offset, ebx)
131 -- discard arguments on stack (+4 for "this")
132 let argcnt = 4 + 4 * methodGetArgsCount cls cpidx
133 when (argcnt > 0) (add esp argcnt)
134 -- push result on stack if method has a return value
135 when (methodHaveReturnValue cls cpidx) (push eax)
136 -- note, the "mi" has the wrong class reference here.
137 -- we figure that out at run-time, in the methodpool,
138 -- depending on the method-table-ptr
139 return $ Just (calladdr, II mi)
140 emit' (INVOKEVIRTUAL cpidx) = do
141 -- get methodInfo entry
142 let mi@(MethodInfo methodname objname msig@(MethodSignature args _)) = buildMethodID cls cpidx
143 newNamedLabel (show mi) >>= defineLabel
144 -- objref lives somewhere on the argument stack
145 mov eax (Disp ((*4) $ fromIntegral $ length args), esp)
146 -- get method-table-ptr
147 mov eax (Disp 0, eax)
149 let nameAndSig = methodname `B.append` encode msig
150 offset <- liftIO $ getMethodOffset objname nameAndSig
151 -- make actual (indirect) call
152 calladdr <- getCurrentOffset
153 call (Disp offset, eax)
154 -- discard arguments on stack (+4 for "this")
155 let argcnt = 4 + 4 * methodGetArgsCount cls cpidx
156 when (argcnt > 0) (add esp argcnt)
157 -- push result on stack if method has a return value
158 when (methodHaveReturnValue cls cpidx) (push eax)
159 -- note, the "mi" has the wrong class reference here.
160 -- we figure that out at run-time, in the methodpool,
161 -- depending on the method-table-ptr
162 return $ Just (calladdr, VI mi)
163 emit' (PUTSTATIC cpidx) = do
165 trapaddr <- getCurrentOffset
166 mov (Addr 0x00000000) eax -- it's a trap
167 return $ Just (trapaddr, SFI $ buildStaticFieldID cls cpidx)
168 emit' (GETSTATIC cpidx) = do
169 trapaddr <- getCurrentOffset
170 mov eax (Addr 0x00000000) -- it's a trap
172 return $ Just (trapaddr, SFI $ buildStaticFieldID cls cpidx)
173 emit' insn = emit insn >> return Nothing
175 emit :: J.Instruction -> CodeGen e s ()
176 emit POP = add esp (4 :: Word32) -- drop value
177 emit DUP = push (Disp 0, esp)
178 emit AASTORE = emit IASTORE
182 add ebx (1 :: Word32)
184 mov (ecx, ebx, S4) eax
188 add ebx (1 :: Word32)
190 mov (ecx, ebx, S1) eax -- TODO(bernhard): char is two byte
191 emit AALOAD = emit IALOAD
194 add ebx (1 :: Word32)
199 add ebx (1 :: Word32)
201 push (ecx, ebx, S1) -- TODO(bernhard): char is two byte
202 emit ARRAYLENGTH = do
205 emit (ANEWARRAY _) = emit (NEWARRAY 10) -- 10 == T_INT
206 emit (NEWARRAY typ) = do
207 let tsize = case decodeS (0 :: Integer) (B.pack [typ]) of
210 _ -> error "newarray: type not implemented yet"
211 -- get length from stack, but leave it there
212 mov eax (Disp 0, esp)
213 mov ebx (tsize :: Word32)
214 -- multiple amount with native size of one element
215 mul ebx -- result is in eax
216 add eax (4 :: Word32) -- for "length" entry
217 -- push amount of bytes to allocate
220 pop eax -- ref to arraymemory
222 mov (Disp 0, eax) ebx -- store length at offset 0
223 push eax -- push ref again
224 emit (NEW objidx) = do
225 let objname = buildClassID cls objidx
226 amount <- liftIO $ getObjectSize objname
227 push (amount :: Word32)
229 -- TODO(bernhard): save reference somewhere for GC
230 -- set method table pointer
231 mtable <- liftIO $ getMethodTable objname
232 mov (Disp 0, eax) mtable
233 emit (CHECKCAST _) = nop -- TODO(bernhard): ...
236 and eax (0x000000ff :: Word32)
238 emit (BIPUSH val) = push (fromIntegral val :: Word32)
239 emit (SIPUSH val) = push (fromIntegral (fromIntegral val :: Int16) :: Word32)
240 emit (ICONST_0) = push (0 :: Word32)
241 emit (ICONST_1) = push (1 :: Word32)
242 emit (ICONST_2) = push (2 :: Word32)
243 emit (ICONST_3) = push (3 :: Word32)
244 emit (ICONST_4) = push (4 :: Word32)
245 emit (ICONST_5) = push (5 :: Word32)
246 emit (ALOAD_ x) = emit (ILOAD_ x)
247 emit (ILOAD_ x) = push (Disp (cArgs_ x), ebp)
248 emit (ALOAD x) = emit (ILOAD x)
249 emit (ILOAD x) = push (Disp (cArgs x), ebp)
250 emit (ASTORE_ x) = emit (ISTORE_ x)
251 emit (ISTORE_ x) = do
253 mov (Disp (cArgs_ x), ebp) eax
254 emit (ASTORE x) = emit (ISTORE x)
257 mov (Disp (cArgs x), ebp) eax
259 emit (LDC1 x) = emit (LDC2 $ fromIntegral x)
261 value <- case constsPool cls M.! x of
262 (CString s) -> liftIO $ getUniqueStringAddr s
263 _ -> error "LDCI... missing impl."
265 emit (GETFIELD x) = do
266 offset <- emitFieldOffset x
267 push (Disp (fromIntegral offset), eax) -- get field
268 emit (PUTFIELD x) = do
269 pop ebx -- value to write
270 offset <- emitFieldOffset x
271 mov (Disp (fromIntegral offset), eax) ebx -- set field
273 emit IADD = do pop ebx; pop eax; add eax ebx; push eax
274 emit ISUB = do pop ebx; pop eax; sub eax ebx; push eax
275 emit IMUL = do pop ebx; pop eax; mul ebx; push eax
276 emit IXOR = do pop ebx; pop eax; xor eax ebx; push eax
278 add (Disp (cArgs x), ebp) (s8_w32 imm)
280 emit (IF_ACMP cond x) = emit (IF_ICMP cond x)
281 emit (IF_ICMP cond _) = do
284 cmp ebx eax -- intel syntax is swapped (TODO(bernhard): test that plz)
287 emit (IF cond _) = do
289 cmp eax (0 :: Word32) -- TODO(bernhard): test that plz
293 let sid = case successor bb of OneTarget t -> t; _ -> error "bad"
294 jmp $ getLabel sid lmap
296 emit RETURN = do mov esp ebp; pop ebp; ret
297 emit ARETURN = emit IRETURN
298 emit IRETURN = do pop eax; emit RETURN
299 emit invalid = error $ "insn not implemented yet: " ++ show invalid
301 emitFieldOffset :: Word16 -> CodeGen e s Int32
302 emitFieldOffset x = do
303 pop eax -- this pointer
304 let (cname, fname) = buildFieldOffset cls x
305 liftIO $ getFieldOffset cname fname
307 emitIF :: CMP -> CodeGen e s ()
309 sid = case successor bb of TwoTarget _ t -> t; _ -> error "bad"
310 l = getLabel sid lmap
312 C_EQ -> je l; C_NE -> jne l
313 C_LT -> jl l; C_GT -> jg l
314 C_GE -> jge l; C_LE -> jle l
316 callMalloc :: CodeGen e s ()
318 calladdr <- getCurrentOffset
319 let w32_calladdr = 5 + calladdr
320 let malloaddr = fromIntegral getMallocObjectAddr :: Word32
321 call (malloaddr - w32_calladdr)
322 add esp (4 :: Word32)
325 -- for locals we use a different storage
326 cArgs :: Word8 -> Word32
327 cArgs x = if x' >= thisMethodArgCnt
328 -- TODO(bernhard): maybe s/(-4)/(-8)/
329 then fromIntegral $ (-4) * (x' - thisMethodArgCnt + 1)
330 else 4 + (thisMethodArgCnt * 4) - (4 * x')
331 where x' = fromIntegral x
333 cArgs_ :: IMM -> Word32
334 cArgs_ x = cArgs $ case x of I0 -> 0; I1 -> 1; I2 -> 2; I3 -> 3
336 thisMethodArgCnt :: Word32
337 thisMethodArgCnt = isNonStatic + fromIntegral (length args)
339 (Just m) = lookupMethod method cls
340 (MethodSignature args _) = methodSignature m
341 isNonStatic = if S.member ACC_STATIC (methodAccessFlags m)
342 then 0 else 1 -- one argument for the this pointer
345 -- sign extension from w8 to w32 (over s8)
346 -- unfortunately, hs-java is using Word8 everywhere (while
347 -- it should be Int8 actually)
348 s8_w32 :: Word8 -> Word32
349 s8_w32 w8 = fromIntegral s8
350 where s8 = fromIntegral w8 :: Int8