2 {-# LANGUAGE OverloadedStrings #-}
3 {-# LANGUAGE ForeignFunctionInterface #-}
5 module Mate.X86CodeGen where
7 import Prelude hiding (and, div)
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
37 foreign import ccall "&mallocObject"
38 mallocObjectAddr :: FunPtr (Int -> IO 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 -> MethodSignature -> Class Direct -> MapBB -> CodeGen e s (CompileInfo, [Instruction])
50 emitFromBB method sig 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 let imm8 = is8BitOffset offset
140 return $ Just (calladdr + (if imm8 then 3 else 6), II imm8 mi)
141 emit' (INVOKEVIRTUAL cpidx) = do
142 -- get methodInfo entry
143 let mi@(MethodInfo methodname objname msig@(MethodSignature args _)) = buildMethodID cls cpidx
144 newNamedLabel (show mi) >>= defineLabel
145 -- objref lives somewhere on the argument stack
146 mov eax (Disp ((*4) $ fromIntegral $ length args), esp)
147 -- get method-table-ptr
148 mov eax (Disp 0, eax)
150 let nameAndSig = methodname `B.append` encode msig
151 offset <- liftIO $ getMethodOffset objname nameAndSig
152 -- make actual (indirect) call
153 calladdr <- getCurrentOffset
154 call (Disp offset, eax)
155 -- discard arguments on stack (+4 for "this")
156 let argcnt = 4 + 4 * methodGetArgsCount cls cpidx
157 when (argcnt > 0) (add esp argcnt)
158 -- push result on stack if method has a return value
159 when (methodHaveReturnValue cls cpidx) (push eax)
160 -- note, the "mi" has the wrong class reference here.
161 -- we figure that out at run-time, in the methodpool,
162 -- depending on the method-table-ptr
163 let imm8 = is8BitOffset offset
164 return $ Just (calladdr + (if imm8 then 3 else 6), VI imm8 mi)
165 emit' (PUTSTATIC cpidx) = do
167 trapaddr <- getCurrentOffset
168 mov (Addr 0x00000000) eax -- it's a trap
169 return $ Just (trapaddr, SFI $ buildStaticFieldID cls cpidx)
170 emit' (GETSTATIC cpidx) = do
171 trapaddr <- getCurrentOffset
172 mov eax (Addr 0x00000000) -- it's a trap
174 return $ Just (trapaddr, SFI $ buildStaticFieldID cls cpidx)
175 emit' insn = emit insn >> return Nothing
177 emit :: J.Instruction -> CodeGen e s ()
178 emit POP = add esp (4 :: Word32) -- drop value
179 emit DUP = push (Disp 0, esp)
180 emit DUP_X1 = do pop eax; pop ebx; push eax; push ebx; push eax
181 emit AASTORE = emit IASTORE
185 add ebx (1 :: Word32)
187 mov (ecx, ebx, S4) eax
191 add ebx (1 :: Word32)
193 mov (ecx, ebx, S1) eax -- TODO(bernhard): char is two byte
194 emit AALOAD = emit IALOAD
197 add ebx (1 :: Word32)
202 add ebx (1 :: Word32)
204 push (ecx, ebx, S1) -- TODO(bernhard): char is two byte
205 emit ARRAYLENGTH = do
208 emit (ANEWARRAY _) = emit (NEWARRAY 10) -- 10 == T_INT
209 emit (NEWARRAY typ) = do
210 let tsize = case decodeS (0 :: Integer) (B.pack [typ]) of
213 _ -> error "newarray: type not implemented yet"
214 -- get length from stack, but leave it there
215 mov eax (Disp 0, esp)
216 mov ebx (tsize :: Word32)
217 -- multiple amount with native size of one element
218 mul ebx -- result is in eax
219 add eax (4 :: Word32) -- for "length" entry
220 -- push amount of bytes to allocate
223 pop eax -- ref to arraymemory
225 mov (Disp 0, eax) ebx -- store length at offset 0
226 push eax -- push ref again
227 emit (NEW objidx) = do
228 let objname = buildClassID cls objidx
229 amount <- liftIO $ getObjectSize objname
230 push (amount :: Word32)
232 -- TODO(bernhard): save reference somewhere for GC
233 -- set method table pointer
234 mtable <- liftIO $ getMethodTable objname
235 mov (Disp 0, eax) mtable
236 emit (CHECKCAST _) = nop -- TODO(bernhard): ...
237 -- TODO(bernhard): ...
238 emit (INSTANCEOF _) = do
241 emit ATHROW = nop -- TODO(bernhard): ...
244 and eax (0x000000ff :: Word32)
246 emit (BIPUSH val) = push (fromIntegral val :: Word32)
247 emit (SIPUSH val) = push (fromIntegral (fromIntegral val :: Int16) :: Word32)
248 emit ACONST_NULL = push (0 :: Word32)
249 emit (ICONST_M1) = push ((-1) :: Word32)
250 emit (ICONST_0) = push (0 :: Word32)
251 emit (ICONST_1) = push (1 :: Word32)
252 emit (ICONST_2) = push (2 :: Word32)
253 emit (ICONST_3) = push (3 :: Word32)
254 emit (ICONST_4) = push (4 :: Word32)
255 emit (ICONST_5) = push (5 :: Word32)
256 emit (ALOAD_ x) = emit (ILOAD_ x)
257 emit (ILOAD_ x) = push (Disp (cArgs_ x), ebp)
258 emit (ALOAD x) = emit (ILOAD x)
259 emit (ILOAD x) = push (Disp (cArgs x), ebp)
260 emit (ASTORE_ x) = emit (ISTORE_ x)
261 emit (ISTORE_ x) = do
263 mov (Disp (cArgs_ x), ebp) eax
264 emit (ASTORE x) = emit (ISTORE x)
267 mov (Disp (cArgs x), ebp) eax
269 emit (LDC1 x) = emit (LDC2 $ fromIntegral x)
271 value <- case constsPool cls M.! x of
272 (CString s) -> liftIO $ getUniqueStringAddr s
273 e -> error $ "LDCI... missing impl.: " ++ show e
275 emit (GETFIELD x) = do
276 offset <- emitFieldOffset x
277 push (Disp (fromIntegral offset), eax) -- get field
278 emit (PUTFIELD x) = do
279 pop ebx -- value to write
280 offset <- emitFieldOffset x
281 mov (Disp (fromIntegral offset), eax) ebx -- set field
283 emit IADD = do pop ebx; pop eax; add eax ebx; push eax
284 emit ISUB = do pop ebx; pop eax; sub eax ebx; push eax
285 emit IMUL = do pop ebx; pop eax; mul ebx; push eax
286 emit IDIV = do pop ebx; pop eax; xor edx edx; div ebx; push eax
287 emit IREM = do pop ebx; pop eax; xor edx edx; div ebx; push edx
288 emit IXOR = do pop ebx; pop eax; xor eax ebx; push eax
289 emit INEG = do pop eax; neg eax; push eax
291 add (Disp (cArgs x), ebp) (s8_w32 imm)
293 emit (IFNONNULL x) = emit (IF C_NE x)
294 emit (IFNULL x) = emit (IF C_EQ x)
295 emit (IF_ACMP cond x) = emit (IF_ICMP cond x)
296 emit (IF_ICMP cond _) = do
299 cmp ebx eax -- intel syntax is swapped (TODO(bernhard): test that plz)
302 emit (IF cond _) = do
304 cmp eax (0 :: Word32) -- TODO(bernhard): test that plz
308 let sid = case successor bb of OneTarget t -> t; _ -> error "bad"
309 jmp $ getLabel sid lmap
311 emit RETURN = do mov esp ebp; pop ebp; ret
312 emit ARETURN = emit IRETURN
313 emit IRETURN = do pop eax; emit RETURN
314 emit invalid = error $ "insn not implemented yet: " ++ show invalid
316 emitFieldOffset :: Word16 -> CodeGen e s Int32
317 emitFieldOffset x = do
318 pop eax -- this pointer
319 let (cname, fname) = buildFieldOffset cls x
320 liftIO $ getFieldOffset cname fname
322 emitIF :: CMP -> CodeGen e s ()
324 sid = case successor bb of TwoTarget _ t -> t; _ -> error "bad"
325 l = getLabel sid lmap
327 C_EQ -> je l; C_NE -> jne l
328 C_LT -> jl l; C_GT -> jg l
329 C_GE -> jge l; C_LE -> jle l
331 callMalloc :: CodeGen e s ()
333 call mallocObjectAddr
334 add esp (4 :: Word32)
337 -- for locals we use a different storage
338 cArgs :: Word8 -> Word32
339 cArgs x = if x' >= thisMethodArgCnt
340 -- TODO(bernhard): maybe s/(-4)/(-8)/
341 then fromIntegral $ (-4) * (x' - thisMethodArgCnt + 1)
342 else 4 + (thisMethodArgCnt * 4) - (4 * x')
343 where x' = fromIntegral x
345 cArgs_ :: IMM -> Word32
346 cArgs_ x = cArgs $ case x of I0 -> 0; I1 -> 1; I2 -> 2; I3 -> 3
348 thisMethodArgCnt :: Word32
349 thisMethodArgCnt = isNonStatic + fromIntegral (length args)
351 (Just m) = lookupMethodSig method sig cls
352 (MethodSignature args _) = sig
353 isNonStatic = if S.member ACC_STATIC (methodAccessFlags m)
354 then 0 else 1 -- one argument for the this pointer
357 -- sign extension from w8 to w32 (over s8)
358 -- unfortunately, hs-java is using Word8 everywhere (while
359 -- it should be Int8 actually)
360 s8_w32 :: Word8 -> Word32
361 s8_w32 w8 = fromIntegral s8
362 where s8 = fromIntegral w8 :: Int8
364 is8BitOffset :: Word32 -> Bool
365 is8BitOffset w32 = s32 < 128 && s32 > (-127)
366 where s32 = fromIntegral w32 :: Int32