2 -- The above warning supression flag is a temporary kludge.
3 -- While working on this module you are encouraged to remove it and fix
4 -- any warnings in the module. See
5 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
8 -----------------------------------------------------------------------------
10 -- Generating machine code (instruction selection)
12 -- (c) The University of Glasgow 1996-2004
14 -----------------------------------------------------------------------------
16 -- This is a big module, but, if you pay attention to
17 -- (a) the sectioning, (b) the type signatures, and
18 -- (c) the #if blah_TARGET_ARCH} things, the
19 -- structure should not be too overwhelming.
28 #include "HsVersions.h"
29 #include "nativeGen/NCG.h"
30 #include "../includes/MachDeps.h"
46 -- Our intermediate code:
49 import PprCmm ( pprExpr )
52 import ClosureInfo ( C_SRT(..) )
55 import StaticFlags ( opt_PIC )
56 import ForeignCall ( CCallConv(..) )
59 import qualified Outputable as O
62 import FastBool ( isFastTrue )
63 import Constants ( wORD_SIZE )
66 import Debug.Trace ( trace )
68 import Control.Monad ( mapAndUnzipM )
69 import Data.Maybe ( fromJust )
74 sse2Enabled :: NatM Bool
75 #if x86_64_TARGET_ARCH
76 -- SSE2 is fixed on for x86_64. It would be possible to make it optional,
77 -- but we'd need to fix at least the foreign call code where the calling
78 -- convention specifies the use of xmm regs, and possibly other places.
79 sse2Enabled = return True
82 dflags <- getDynFlagsNat
83 return (dopt Opt_SSE2 dflags)
86 if_sse2 :: NatM a -> NatM a -> NatM a
89 if b then sse2 else x87
94 -> NatM [NatCmmTop Instr]
96 cmmTopCodeGen dynflags
97 (CmmProc info lab params (ListGraph blocks)) = do
98 (nat_blocks,statics) <- mapAndUnzipM basicBlockCodeGen blocks
99 picBaseMb <- getPicBaseMaybeNat
100 let proc = CmmProc info lab params (ListGraph $ concat nat_blocks)
101 tops = proc : concat statics
102 os = platformOS $ targetPlatform dynflags
105 Just picBase -> initializePicBase_x86 ArchX86 os picBase tops
106 Nothing -> return tops
108 cmmTopCodeGen _ (CmmData sec dat) = do
109 return [CmmData sec dat] -- no translation, we just use CmmStatic
114 -> NatM ( [NatBasicBlock Instr]
117 basicBlockCodeGen (BasicBlock id stmts) = do
118 instrs <- stmtsToInstrs stmts
119 -- code generation may introduce new basic block boundaries, which
120 -- are indicated by the NEWBLOCK instruction. We must split up the
121 -- instruction stream into basic blocks again. Also, we extract
124 (top,other_blocks,statics) = foldrOL mkBlocks ([],[],[]) instrs
126 mkBlocks (NEWBLOCK id) (instrs,blocks,statics)
127 = ([], BasicBlock id instrs : blocks, statics)
128 mkBlocks (LDATA sec dat) (instrs,blocks,statics)
129 = (instrs, blocks, CmmData sec dat:statics)
130 mkBlocks instr (instrs,blocks,statics)
131 = (instr:instrs, blocks, statics)
133 return (BasicBlock id top : other_blocks, statics)
136 stmtsToInstrs :: [CmmStmt] -> NatM InstrBlock
138 = do instrss <- mapM stmtToInstrs stmts
139 return (concatOL instrss)
142 stmtToInstrs :: CmmStmt -> NatM InstrBlock
143 stmtToInstrs stmt = case stmt of
144 CmmNop -> return nilOL
145 CmmComment s -> return (unitOL (COMMENT s))
148 | isFloatType ty -> assignReg_FltCode size reg src
149 #if WORD_SIZE_IN_BITS==32
150 | isWord64 ty -> assignReg_I64Code reg src
152 | otherwise -> assignReg_IntCode size reg src
153 where ty = cmmRegType reg
154 size = cmmTypeSize ty
157 | isFloatType ty -> assignMem_FltCode size addr src
158 #if WORD_SIZE_IN_BITS==32
159 | isWord64 ty -> assignMem_I64Code addr src
161 | otherwise -> assignMem_IntCode size addr src
162 where ty = cmmExprType src
163 size = cmmTypeSize ty
165 CmmCall target result_regs args _ _
166 -> genCCall target result_regs args
168 CmmBranch id -> genBranch id
169 CmmCondBranch arg id -> genCondJump id arg
170 CmmSwitch arg ids -> genSwitch arg ids
171 CmmJump arg params -> genJump arg
173 panic "stmtToInstrs: return statement should have been cps'd away"
176 --------------------------------------------------------------------------------
177 -- | 'InstrBlock's are the insn sequences generated by the insn selectors.
178 -- They are really trees of insns to facilitate fast appending, where a
179 -- left-to-right traversal yields the insns in the correct order.
185 -- | Condition codes passed up the tree.
188 = CondCode Bool Cond InstrBlock
191 -- | a.k.a "Register64"
192 -- Reg is the lower 32-bit temporary which contains the result.
193 -- Use getHiVRegFromLo to find the other VRegUnique.
195 -- Rules of this simplified insn selection game are therefore that
196 -- the returned Reg may be modified
204 -- | Register's passed up the tree. If the stix code forces the register
205 -- to live in a pre-decided machine register, it comes out as @Fixed@;
206 -- otherwise, it comes out as @Any@, and the parent can decide which
207 -- register to put it in.
210 = Fixed Size Reg InstrBlock
211 | Any Size (Reg -> InstrBlock)
214 swizzleRegisterRep :: Register -> Size -> Register
215 swizzleRegisterRep (Fixed _ reg code) size = Fixed size reg code
216 swizzleRegisterRep (Any _ codefn) size = Any size codefn
219 -- | Grab the Reg for a CmmReg
220 getRegisterReg :: Bool -> CmmReg -> Reg
222 getRegisterReg use_sse2 (CmmLocal (LocalReg u pk))
223 = let sz = cmmTypeSize pk in
224 if isFloatSize sz && not use_sse2
225 then RegVirtual (mkVirtualReg u FF80)
226 else RegVirtual (mkVirtualReg u sz)
228 getRegisterReg _ (CmmGlobal mid)
229 = case get_GlobalReg_reg_or_addr mid of
230 Left reg -> RegReal $ reg
231 _other -> pprPanic "getRegisterReg-memory" (ppr $ CmmGlobal mid)
232 -- By this stage, the only MagicIds remaining should be the
233 -- ones which map to a real machine register on this
234 -- platform. Hence ...
237 -- | Memory addressing modes passed up the tree.
239 = Amode AddrMode InstrBlock
242 Now, given a tree (the argument to an CmmLoad) that references memory,
243 produce a suitable addressing mode.
245 A Rule of the Game (tm) for Amodes: use of the addr bit must
246 immediately follow use of the code part, since the code part puts
247 values in registers which the addr then refers to. So you can't put
248 anything in between, lest it overwrite some of those registers. If
249 you need to do some other computation between the code part and use of
250 the addr bit, first store the effective address from the amode in a
251 temporary, then do the other computation, and then use the temporary:
255 ... other computation ...
260 -- | Check whether an integer will fit in 32 bits.
261 -- A CmmInt is intended to be truncated to the appropriate
262 -- number of bits, so here we truncate it to Int64. This is
263 -- important because e.g. -1 as a CmmInt might be either
264 -- -1 or 18446744073709551615.
266 is32BitInteger :: Integer -> Bool
267 is32BitInteger i = i64 <= 0x7fffffff && i64 >= -0x80000000
268 where i64 = fromIntegral i :: Int64
271 -- | Convert a BlockId to some CmmStatic data
272 jumpTableEntry :: Maybe BlockId -> CmmStatic
273 jumpTableEntry Nothing = CmmStaticLit (CmmInt 0 wordWidth)
274 jumpTableEntry (Just (BlockId id)) = CmmStaticLit (CmmLabel blockLabel)
275 where blockLabel = mkAsmTempLabel id
278 -- -----------------------------------------------------------------------------
279 -- General things for putting together code sequences
281 -- Expand CmmRegOff. ToDo: should we do it this way around, or convert
282 -- CmmExprs into CmmRegOff?
283 mangleIndexTree :: CmmExpr -> CmmExpr
284 mangleIndexTree (CmmRegOff reg off)
285 = CmmMachOp (MO_Add width) [CmmReg reg, CmmLit (CmmInt (fromIntegral off) width)]
286 where width = typeWidth (cmmRegType reg)
288 -- | The dual to getAnyReg: compute an expression into a register, but
289 -- we don't mind which one it is.
290 getSomeReg :: CmmExpr -> NatM (Reg, InstrBlock)
292 r <- getRegister expr
295 tmp <- getNewRegNat rep
296 return (tmp, code tmp)
304 assignMem_I64Code :: CmmExpr -> CmmExpr -> NatM InstrBlock
305 assignMem_I64Code addrTree valueTree = do
306 Amode addr addr_code <- getAmode addrTree
307 ChildCode64 vcode rlo <- iselExpr64 valueTree
309 rhi = getHiVRegFromLo rlo
311 -- Little-endian store
312 mov_lo = MOV II32 (OpReg rlo) (OpAddr addr)
313 mov_hi = MOV II32 (OpReg rhi) (OpAddr (fromJust (addrOffset addr 4)))
315 return (vcode `appOL` addr_code `snocOL` mov_lo `snocOL` mov_hi)
318 assignReg_I64Code :: CmmReg -> CmmExpr -> NatM InstrBlock
319 assignReg_I64Code (CmmLocal (LocalReg u_dst pk)) valueTree = do
320 ChildCode64 vcode r_src_lo <- iselExpr64 valueTree
322 r_dst_lo = RegVirtual $ mkVirtualReg u_dst II32
323 r_dst_hi = getHiVRegFromLo r_dst_lo
324 r_src_hi = getHiVRegFromLo r_src_lo
325 mov_lo = MOV II32 (OpReg r_src_lo) (OpReg r_dst_lo)
326 mov_hi = MOV II32 (OpReg r_src_hi) (OpReg r_dst_hi)
329 vcode `snocOL` mov_lo `snocOL` mov_hi
332 assignReg_I64Code lvalue valueTree
333 = panic "assignReg_I64Code(i386): invalid lvalue"
338 iselExpr64 :: CmmExpr -> NatM ChildCode64
339 iselExpr64 (CmmLit (CmmInt i _)) = do
340 (rlo,rhi) <- getNewRegPairNat II32
342 r = fromIntegral (fromIntegral i :: Word32)
343 q = fromIntegral ((fromIntegral i `shiftR` 32) :: Word32)
345 MOV II32 (OpImm (ImmInteger r)) (OpReg rlo),
346 MOV II32 (OpImm (ImmInteger q)) (OpReg rhi)
349 return (ChildCode64 code rlo)
351 iselExpr64 (CmmLoad addrTree ty) | isWord64 ty = do
352 Amode addr addr_code <- getAmode addrTree
353 (rlo,rhi) <- getNewRegPairNat II32
355 mov_lo = MOV II32 (OpAddr addr) (OpReg rlo)
356 mov_hi = MOV II32 (OpAddr (fromJust (addrOffset addr 4))) (OpReg rhi)
359 ChildCode64 (addr_code `snocOL` mov_lo `snocOL` mov_hi)
363 iselExpr64 (CmmReg (CmmLocal (LocalReg vu ty))) | isWord64 ty
364 = return (ChildCode64 nilOL (RegVirtual $ mkVirtualReg vu II32))
366 -- we handle addition, but rather badly
367 iselExpr64 (CmmMachOp (MO_Add _) [e1, CmmLit (CmmInt i _)]) = do
368 ChildCode64 code1 r1lo <- iselExpr64 e1
369 (rlo,rhi) <- getNewRegPairNat II32
371 r = fromIntegral (fromIntegral i :: Word32)
372 q = fromIntegral ((fromIntegral i `shiftR` 32) :: Word32)
373 r1hi = getHiVRegFromLo r1lo
375 toOL [ MOV II32 (OpReg r1lo) (OpReg rlo),
376 ADD II32 (OpImm (ImmInteger r)) (OpReg rlo),
377 MOV II32 (OpReg r1hi) (OpReg rhi),
378 ADC II32 (OpImm (ImmInteger q)) (OpReg rhi) ]
380 return (ChildCode64 code rlo)
382 iselExpr64 (CmmMachOp (MO_Add _) [e1,e2]) = do
383 ChildCode64 code1 r1lo <- iselExpr64 e1
384 ChildCode64 code2 r2lo <- iselExpr64 e2
385 (rlo,rhi) <- getNewRegPairNat II32
387 r1hi = getHiVRegFromLo r1lo
388 r2hi = getHiVRegFromLo r2lo
391 toOL [ MOV II32 (OpReg r1lo) (OpReg rlo),
392 ADD II32 (OpReg r2lo) (OpReg rlo),
393 MOV II32 (OpReg r1hi) (OpReg rhi),
394 ADC II32 (OpReg r2hi) (OpReg rhi) ]
396 return (ChildCode64 code rlo)
398 iselExpr64 (CmmMachOp (MO_UU_Conv _ W64) [expr]) = do
400 r_dst_lo <- getNewRegNat II32
401 let r_dst_hi = getHiVRegFromLo r_dst_lo
404 ChildCode64 (code `snocOL`
405 MOV II32 (OpImm (ImmInt 0)) (OpReg r_dst_hi))
410 = pprPanic "iselExpr64(i386)" (ppr expr)
414 --------------------------------------------------------------------------------
415 getRegister :: CmmExpr -> NatM Register
417 #if !x86_64_TARGET_ARCH
418 -- on x86_64, we have %rip for PicBaseReg, but it's not a full-featured
419 -- register, it can only be used for rip-relative addressing.
420 getRegister (CmmReg (CmmGlobal PicBaseReg))
422 reg <- getPicBaseNat archWordSize
423 return (Fixed archWordSize reg nilOL)
426 getRegister (CmmReg reg)
427 = do use_sse2 <- sse2Enabled
429 sz = cmmTypeSize (cmmRegType reg)
430 size | not use_sse2 && isFloatSize sz = FF80
433 return (Fixed sz (getRegisterReg use_sse2 reg) nilOL)
436 getRegister tree@(CmmRegOff _ _)
437 = getRegister (mangleIndexTree tree)
440 #if WORD_SIZE_IN_BITS==32
441 -- for 32-bit architectuers, support some 64 -> 32 bit conversions:
442 -- TO_W_(x), TO_W_(x >> 32)
444 getRegister (CmmMachOp (MO_UU_Conv W64 W32)
445 [CmmMachOp (MO_U_Shr W64) [x,CmmLit (CmmInt 32 _)]]) = do
446 ChildCode64 code rlo <- iselExpr64 x
447 return $ Fixed II32 (getHiVRegFromLo rlo) code
449 getRegister (CmmMachOp (MO_SS_Conv W64 W32)
450 [CmmMachOp (MO_U_Shr W64) [x,CmmLit (CmmInt 32 _)]]) = do
451 ChildCode64 code rlo <- iselExpr64 x
452 return $ Fixed II32 (getHiVRegFromLo rlo) code
454 getRegister (CmmMachOp (MO_UU_Conv W64 W32) [x]) = do
455 ChildCode64 code rlo <- iselExpr64 x
456 return $ Fixed II32 rlo code
458 getRegister (CmmMachOp (MO_SS_Conv W64 W32) [x]) = do
459 ChildCode64 code rlo <- iselExpr64 x
460 return $ Fixed II32 rlo code
465 getRegister (CmmLit lit@(CmmFloat f w)) =
466 if_sse2 float_const_sse2 float_const_x87
472 code dst = unitOL (XOR size (OpReg dst) (OpReg dst))
473 -- I don't know why there are xorpd, xorps, and pxor instructions.
474 -- They all appear to do the same thing --SDM
475 return (Any size code)
478 Amode addr code <- memConstant (widthInBytes w) lit
479 loadFloatAmode True w addr code
481 float_const_x87 = case w of
484 let code dst = unitOL (GLDZ dst)
485 in return (Any FF80 code)
488 let code dst = unitOL (GLD1 dst)
489 in return (Any FF80 code)
492 Amode addr code <- memConstant (widthInBytes w) lit
493 loadFloatAmode False w addr code
495 -- catch simple cases of zero- or sign-extended load
496 getRegister (CmmMachOp (MO_UU_Conv W8 W32) [CmmLoad addr _]) = do
497 code <- intLoadCode (MOVZxL II8) addr
498 return (Any II32 code)
500 getRegister (CmmMachOp (MO_SS_Conv W8 W32) [CmmLoad addr _]) = do
501 code <- intLoadCode (MOVSxL II8) addr
502 return (Any II32 code)
504 getRegister (CmmMachOp (MO_UU_Conv W16 W32) [CmmLoad addr _]) = do
505 code <- intLoadCode (MOVZxL II16) addr
506 return (Any II32 code)
508 getRegister (CmmMachOp (MO_SS_Conv W16 W32) [CmmLoad addr _]) = do
509 code <- intLoadCode (MOVSxL II16) addr
510 return (Any II32 code)
513 #if x86_64_TARGET_ARCH
515 -- catch simple cases of zero- or sign-extended load
516 getRegister (CmmMachOp (MO_UU_Conv W8 W64) [CmmLoad addr _]) = do
517 code <- intLoadCode (MOVZxL II8) addr
518 return (Any II64 code)
520 getRegister (CmmMachOp (MO_SS_Conv W8 W64) [CmmLoad addr _]) = do
521 code <- intLoadCode (MOVSxL II8) addr
522 return (Any II64 code)
524 getRegister (CmmMachOp (MO_UU_Conv W16 W64) [CmmLoad addr _]) = do
525 code <- intLoadCode (MOVZxL II16) addr
526 return (Any II64 code)
528 getRegister (CmmMachOp (MO_SS_Conv W16 W64) [CmmLoad addr _]) = do
529 code <- intLoadCode (MOVSxL II16) addr
530 return (Any II64 code)
532 getRegister (CmmMachOp (MO_UU_Conv W32 W64) [CmmLoad addr _]) = do
533 code <- intLoadCode (MOV II32) addr -- 32-bit loads zero-extend
534 return (Any II64 code)
536 getRegister (CmmMachOp (MO_SS_Conv W32 W64) [CmmLoad addr _]) = do
537 code <- intLoadCode (MOVSxL II32) addr
538 return (Any II64 code)
540 getRegister (CmmMachOp (MO_Add W64) [CmmReg (CmmGlobal PicBaseReg),
541 CmmLit displacement])
542 = return $ Any II64 (\dst -> unitOL $
543 LEA II64 (OpAddr (ripRel (litToImm displacement))) (OpReg dst))
545 #endif /* x86_64_TARGET_ARCH */
551 getRegister (CmmMachOp mop [x]) = do -- unary MachOps
555 | sse2 -> sse2NegCode w x
556 | otherwise -> trivialUFCode FF80 (GNEG FF80) x
558 MO_S_Neg w -> triv_ucode NEGI (intSize w)
559 MO_Not w -> triv_ucode NOT (intSize w)
562 MO_UU_Conv W32 W8 -> toI8Reg W32 x
563 MO_SS_Conv W32 W8 -> toI8Reg W32 x
564 MO_UU_Conv W16 W8 -> toI8Reg W16 x
565 MO_SS_Conv W16 W8 -> toI8Reg W16 x
566 MO_UU_Conv W32 W16 -> toI16Reg W32 x
567 MO_SS_Conv W32 W16 -> toI16Reg W32 x
569 #if x86_64_TARGET_ARCH
570 MO_UU_Conv W64 W32 -> conversionNop II64 x
571 MO_SS_Conv W64 W32 -> conversionNop II64 x
572 MO_UU_Conv W64 W16 -> toI16Reg W64 x
573 MO_SS_Conv W64 W16 -> toI16Reg W64 x
574 MO_UU_Conv W64 W8 -> toI8Reg W64 x
575 MO_SS_Conv W64 W8 -> toI8Reg W64 x
578 MO_UU_Conv rep1 rep2 | rep1 == rep2 -> conversionNop (intSize rep1) x
579 MO_SS_Conv rep1 rep2 | rep1 == rep2 -> conversionNop (intSize rep1) x
582 MO_UU_Conv W8 W32 -> integerExtend W8 W32 MOVZxL x
583 MO_UU_Conv W16 W32 -> integerExtend W16 W32 MOVZxL x
584 MO_UU_Conv W8 W16 -> integerExtend W8 W16 MOVZxL x
586 MO_SS_Conv W8 W32 -> integerExtend W8 W32 MOVSxL x
587 MO_SS_Conv W16 W32 -> integerExtend W16 W32 MOVSxL x
588 MO_SS_Conv W8 W16 -> integerExtend W8 W16 MOVSxL x
590 #if x86_64_TARGET_ARCH
591 MO_UU_Conv W8 W64 -> integerExtend W8 W64 MOVZxL x
592 MO_UU_Conv W16 W64 -> integerExtend W16 W64 MOVZxL x
593 MO_UU_Conv W32 W64 -> integerExtend W32 W64 MOVZxL x
594 MO_SS_Conv W8 W64 -> integerExtend W8 W64 MOVSxL x
595 MO_SS_Conv W16 W64 -> integerExtend W16 W64 MOVSxL x
596 MO_SS_Conv W32 W64 -> integerExtend W32 W64 MOVSxL x
597 -- for 32-to-64 bit zero extension, amd64 uses an ordinary movl.
598 -- However, we don't want the register allocator to throw it
599 -- away as an unnecessary reg-to-reg move, so we keep it in
600 -- the form of a movzl and print it as a movl later.
604 | sse2 -> coerceFP2FP W64 x
605 | otherwise -> conversionNop FF80 x
608 | sse2 -> coerceFP2FP W32 x
609 | otherwise -> conversionNop FF80 x
611 MO_FS_Conv from to -> coerceFP2Int from to x
612 MO_SF_Conv from to -> coerceInt2FP from to x
614 other -> pprPanic "getRegister" (pprMachOp mop)
616 triv_ucode :: (Size -> Operand -> Instr) -> Size -> NatM Register
617 triv_ucode instr size = trivialUCode size (instr size) x
619 -- signed or unsigned extension.
620 integerExtend :: Width -> Width
621 -> (Size -> Operand -> Operand -> Instr)
622 -> CmmExpr -> NatM Register
623 integerExtend from to instr expr = do
624 (reg,e_code) <- if from == W8 then getByteReg expr
629 instr (intSize from) (OpReg reg) (OpReg dst)
630 return (Any (intSize to) code)
632 toI8Reg :: Width -> CmmExpr -> NatM Register
634 = do codefn <- getAnyReg expr
635 return (Any (intSize new_rep) codefn)
636 -- HACK: use getAnyReg to get a byte-addressable register.
637 -- If the source was a Fixed register, this will add the
638 -- mov instruction to put it into the desired destination.
639 -- We're assuming that the destination won't be a fixed
640 -- non-byte-addressable register; it won't be, because all
641 -- fixed registers are word-sized.
643 toI16Reg = toI8Reg -- for now
645 conversionNop :: Size -> CmmExpr -> NatM Register
646 conversionNop new_size expr
647 = do e_code <- getRegister expr
648 return (swizzleRegisterRep e_code new_size)
651 getRegister e@(CmmMachOp mop [x, y]) = do -- dyadic MachOps
654 MO_F_Eq w -> condFltReg EQQ x y
655 MO_F_Ne w -> condFltReg NE x y
656 MO_F_Gt w -> condFltReg GTT x y
657 MO_F_Ge w -> condFltReg GE x y
658 MO_F_Lt w -> condFltReg LTT x y
659 MO_F_Le w -> condFltReg LE x y
661 MO_Eq rep -> condIntReg EQQ x y
662 MO_Ne rep -> condIntReg NE x y
664 MO_S_Gt rep -> condIntReg GTT x y
665 MO_S_Ge rep -> condIntReg GE x y
666 MO_S_Lt rep -> condIntReg LTT x y
667 MO_S_Le rep -> condIntReg LE x y
669 MO_U_Gt rep -> condIntReg GU x y
670 MO_U_Ge rep -> condIntReg GEU x y
671 MO_U_Lt rep -> condIntReg LU x y
672 MO_U_Le rep -> condIntReg LEU x y
674 MO_F_Add w | sse2 -> trivialFCode_sse2 w ADD x y
675 | otherwise -> trivialFCode_x87 w GADD x y
676 MO_F_Sub w | sse2 -> trivialFCode_sse2 w SUB x y
677 | otherwise -> trivialFCode_x87 w GSUB x y
678 MO_F_Quot w | sse2 -> trivialFCode_sse2 w FDIV x y
679 | otherwise -> trivialFCode_x87 w GDIV x y
680 MO_F_Mul w | sse2 -> trivialFCode_sse2 w MUL x y
681 | otherwise -> trivialFCode_x87 w GMUL x y
683 MO_Add rep -> add_code rep x y
684 MO_Sub rep -> sub_code rep x y
686 MO_S_Quot rep -> div_code rep True True x y
687 MO_S_Rem rep -> div_code rep True False x y
688 MO_U_Quot rep -> div_code rep False True x y
689 MO_U_Rem rep -> div_code rep False False x y
691 MO_S_MulMayOflo rep -> imulMayOflo rep x y
693 MO_Mul rep -> triv_op rep IMUL
694 MO_And rep -> triv_op rep AND
695 MO_Or rep -> triv_op rep OR
696 MO_Xor rep -> triv_op rep XOR
698 {- Shift ops on x86s have constraints on their source, it
699 either has to be Imm, CL or 1
700 => trivialCode is not restrictive enough (sigh.)
702 MO_Shl rep -> shift_code rep SHL x y {-False-}
703 MO_U_Shr rep -> shift_code rep SHR x y {-False-}
704 MO_S_Shr rep -> shift_code rep SAR x y {-False-}
706 other -> pprPanic "getRegister(x86) - binary CmmMachOp (1)" (pprMachOp mop)
709 triv_op width instr = trivialCode width op (Just op) x y
710 where op = instr (intSize width)
712 imulMayOflo :: Width -> CmmExpr -> CmmExpr -> NatM Register
713 imulMayOflo rep a b = do
714 (a_reg, a_code) <- getNonClobberedReg a
715 b_code <- getAnyReg b
717 shift_amt = case rep of
720 _ -> panic "shift_amt"
723 code = a_code `appOL` b_code eax `appOL`
725 IMUL2 size (OpReg a_reg), -- result in %edx:%eax
726 SAR size (OpImm (ImmInt shift_amt)) (OpReg eax),
727 -- sign extend lower part
728 SUB size (OpReg edx) (OpReg eax)
729 -- compare against upper
730 -- eax==0 if high part == sign extended low part
733 return (Fixed size eax code)
737 -> (Size -> Operand -> Operand -> Instr)
742 {- Case1: shift length as immediate -}
743 shift_code width instr x y@(CmmLit lit) = do
744 x_code <- getAnyReg x
748 = x_code dst `snocOL`
749 instr size (OpImm (litToImm lit)) (OpReg dst)
751 return (Any size code)
753 {- Case2: shift length is complex (non-immediate)
755 * we cannot do y first *and* put its result in %ecx, because
756 %ecx might be clobbered by x.
757 * if we do y second, then x cannot be
758 in a clobbered reg. Also, we cannot clobber x's reg
759 with the instruction itself.
761 - do y first, put its result in a fresh tmp, then copy it to %ecx later
762 - do y second and put its result into %ecx. x gets placed in a fresh
763 tmp. This is likely to be better, becuase the reg alloc can
764 eliminate this reg->reg move here (it won't eliminate the other one,
765 because the move is into the fixed %ecx).
767 shift_code width instr x y{-amount-} = do
768 x_code <- getAnyReg x
769 let size = intSize width
770 tmp <- getNewRegNat size
771 y_code <- getAnyReg y
773 code = x_code tmp `appOL`
775 instr size (OpReg ecx) (OpReg tmp)
777 return (Fixed size tmp code)
780 add_code :: Width -> CmmExpr -> CmmExpr -> NatM Register
781 add_code rep x (CmmLit (CmmInt y _))
782 | is32BitInteger y = add_int rep x y
783 add_code rep x y = trivialCode rep (ADD size) (Just (ADD size)) x y
784 where size = intSize rep
787 sub_code :: Width -> CmmExpr -> CmmExpr -> NatM Register
788 sub_code rep x (CmmLit (CmmInt y _))
789 | is32BitInteger (-y) = add_int rep x (-y)
790 sub_code rep x y = trivialCode rep (SUB (intSize rep)) Nothing x y
792 -- our three-operand add instruction:
793 add_int width x y = do
794 (x_reg, x_code) <- getSomeReg x
797 imm = ImmInt (fromInteger y)
801 (OpAddr (AddrBaseIndex (EABaseReg x_reg) EAIndexNone imm))
804 return (Any size code)
806 ----------------------
807 div_code width signed quotient x y = do
808 (y_op, y_code) <- getRegOrMem y -- cannot be clobbered
809 x_code <- getAnyReg x
812 widen | signed = CLTD size
813 | otherwise = XOR size (OpReg edx) (OpReg edx)
815 instr | signed = IDIV
818 code = y_code `appOL`
820 toOL [widen, instr size y_op]
822 result | quotient = eax
826 return (Fixed size result code)
829 getRegister (CmmLoad mem pk)
832 Amode addr mem_code <- getAmode mem
833 use_sse2 <- sse2Enabled
834 loadFloatAmode use_sse2 (typeWidth pk) addr mem_code
837 getRegister (CmmLoad mem pk)
840 code <- intLoadCode instr mem
841 return (Any size code)
845 instr = case width of
848 -- We always zero-extend 8-bit loads, if we
849 -- can't think of anything better. This is because
850 -- we can't guarantee access to an 8-bit variant of every register
851 -- (esi and edi don't have 8-bit variants), so to make things
852 -- simpler we do our 8-bit arithmetic with full 32-bit registers.
855 #if x86_64_TARGET_ARCH
856 -- Simpler memory load code on x86_64
857 getRegister (CmmLoad mem pk)
859 code <- intLoadCode (MOV size) mem
860 return (Any size code)
861 where size = intSize $ typeWidth pk
864 getRegister (CmmLit (CmmInt 0 width))
868 -- x86_64: 32-bit xor is one byte shorter, and zero-extends to 64 bits
869 adj_size = case size of II64 -> II32; _ -> size
870 size1 = IF_ARCH_i386( size, adj_size )
872 = unitOL (XOR size1 (OpReg dst) (OpReg dst))
874 return (Any size code)
876 #if x86_64_TARGET_ARCH
877 -- optimisation for loading small literals on x86_64: take advantage
878 -- of the automatic zero-extension from 32 to 64 bits, because the 32-bit
879 -- instruction forms are shorter.
880 getRegister (CmmLit lit)
881 | isWord64 (cmmLitType lit), not (isBigLit lit)
884 code dst = unitOL (MOV II32 (OpImm imm) (OpReg dst))
886 return (Any II64 code)
888 isBigLit (CmmInt i _) = i < 0 || i > 0xffffffff
890 -- note1: not the same as (not.is32BitLit), because that checks for
891 -- signed literals that fit in 32 bits, but we want unsigned
893 -- note2: all labels are small, because we're assuming the
894 -- small memory model (see gcc docs, -mcmodel=small).
897 getRegister (CmmLit lit)
899 size = cmmTypeSize (cmmLitType lit)
901 code dst = unitOL (MOV size (OpImm imm) (OpReg dst))
903 return (Any size code)
905 getRegister other = pprPanic "getRegister(x86)" (ppr other)
908 intLoadCode :: (Operand -> Operand -> Instr) -> CmmExpr
909 -> NatM (Reg -> InstrBlock)
910 intLoadCode instr mem = do
911 Amode src mem_code <- getAmode mem
912 return (\dst -> mem_code `snocOL` instr (OpAddr src) (OpReg dst))
914 -- Compute an expression into *any* register, adding the appropriate
915 -- move instruction if necessary.
916 getAnyReg :: CmmExpr -> NatM (Reg -> InstrBlock)
918 r <- getRegister expr
921 anyReg :: Register -> NatM (Reg -> InstrBlock)
922 anyReg (Any _ code) = return code
923 anyReg (Fixed rep reg fcode) = return (\dst -> fcode `snocOL` reg2reg rep reg dst)
925 -- A bit like getSomeReg, but we want a reg that can be byte-addressed.
926 -- Fixed registers might not be byte-addressable, so we make sure we've
927 -- got a temporary, inserting an extra reg copy if necessary.
928 getByteReg :: CmmExpr -> NatM (Reg, InstrBlock)
929 #if x86_64_TARGET_ARCH
930 getByteReg = getSomeReg -- all regs are byte-addressable on x86_64
933 r <- getRegister expr
936 tmp <- getNewRegNat rep
937 return (tmp, code tmp)
939 | isVirtualReg reg -> return (reg,code)
941 tmp <- getNewRegNat rep
942 return (tmp, code `snocOL` reg2reg rep reg tmp)
943 -- ToDo: could optimise slightly by checking for byte-addressable
944 -- real registers, but that will happen very rarely if at all.
947 -- Another variant: this time we want the result in a register that cannot
948 -- be modified by code to evaluate an arbitrary expression.
949 getNonClobberedReg :: CmmExpr -> NatM (Reg, InstrBlock)
950 getNonClobberedReg expr = do
951 r <- getRegister expr
954 tmp <- getNewRegNat rep
955 return (tmp, code tmp)
957 -- only free regs can be clobbered
958 | RegReal (RealRegSingle rr) <- reg
959 , isFastTrue (freeReg rr)
961 tmp <- getNewRegNat rep
962 return (tmp, code `snocOL` reg2reg rep reg tmp)
966 reg2reg :: Size -> Reg -> Reg -> Instr
968 | size == FF80 = GMOV src dst
969 | otherwise = MOV size (OpReg src) (OpReg dst)
972 --------------------------------------------------------------------------------
973 getAmode :: CmmExpr -> NatM Amode
974 getAmode tree@(CmmRegOff _ _) = getAmode (mangleIndexTree tree)
976 #if x86_64_TARGET_ARCH
978 getAmode (CmmMachOp (MO_Add W64) [CmmReg (CmmGlobal PicBaseReg),
979 CmmLit displacement])
980 = return $ Amode (ripRel (litToImm displacement)) nilOL
985 -- This is all just ridiculous, since it carefully undoes
986 -- what mangleIndexTree has just done.
987 getAmode (CmmMachOp (MO_Sub rep) [x, CmmLit lit@(CmmInt i _)])
989 -- ASSERT(rep == II32)???
990 = do (x_reg, x_code) <- getSomeReg x
991 let off = ImmInt (-(fromInteger i))
992 return (Amode (AddrBaseIndex (EABaseReg x_reg) EAIndexNone off) x_code)
994 getAmode (CmmMachOp (MO_Add rep) [x, CmmLit lit@(CmmInt i _)])
996 -- ASSERT(rep == II32)???
997 = do (x_reg, x_code) <- getSomeReg x
998 let off = ImmInt (fromInteger i)
999 return (Amode (AddrBaseIndex (EABaseReg x_reg) EAIndexNone off) x_code)
1001 -- Turn (lit1 << n + lit2) into (lit2 + lit1 << n) so it will be
1002 -- recognised by the next rule.
1003 getAmode (CmmMachOp (MO_Add rep) [a@(CmmMachOp (MO_Shl _) _),
1005 = getAmode (CmmMachOp (MO_Add rep) [b,a])
1007 getAmode (CmmMachOp (MO_Add rep) [x, CmmMachOp (MO_Shl _)
1008 [y, CmmLit (CmmInt shift _)]])
1009 | shift == 0 || shift == 1 || shift == 2 || shift == 3
1010 = x86_complex_amode x y shift 0
1012 getAmode (CmmMachOp (MO_Add rep)
1013 [x, CmmMachOp (MO_Add _)
1014 [CmmMachOp (MO_Shl _) [y, CmmLit (CmmInt shift _)],
1015 CmmLit (CmmInt offset _)]])
1016 | shift == 0 || shift == 1 || shift == 2 || shift == 3
1017 && is32BitInteger offset
1018 = x86_complex_amode x y shift offset
1020 getAmode (CmmMachOp (MO_Add rep) [x,y])
1021 = x86_complex_amode x y 0 0
1023 getAmode (CmmLit lit) | is32BitLit lit
1024 = return (Amode (ImmAddr (litToImm lit) 0) nilOL)
1027 (reg,code) <- getSomeReg expr
1028 return (Amode (AddrBaseIndex (EABaseReg reg) EAIndexNone (ImmInt 0)) code)
1031 x86_complex_amode :: CmmExpr -> CmmExpr -> Integer -> Integer -> NatM Amode
1032 x86_complex_amode base index shift offset
1033 = do (x_reg, x_code) <- getNonClobberedReg base
1034 -- x must be in a temp, because it has to stay live over y_code
1035 -- we could compre x_reg and y_reg and do something better here...
1036 (y_reg, y_code) <- getSomeReg index
1038 code = x_code `appOL` y_code
1039 base = case shift of 0 -> 1; 1 -> 2; 2 -> 4; 3 -> 8
1040 return (Amode (AddrBaseIndex (EABaseReg x_reg) (EAIndex y_reg base) (ImmInt (fromIntegral offset)))
1046 -- -----------------------------------------------------------------------------
1047 -- getOperand: sometimes any operand will do.
1049 -- getNonClobberedOperand: the value of the operand will remain valid across
1050 -- the computation of an arbitrary expression, unless the expression
1051 -- is computed directly into a register which the operand refers to
1052 -- (see trivialCode where this function is used for an example).
1054 getNonClobberedOperand :: CmmExpr -> NatM (Operand, InstrBlock)
1055 getNonClobberedOperand (CmmLit lit) = do
1056 use_sse2 <- sse2Enabled
1057 if use_sse2 && isSuitableFloatingPointLit lit
1059 let CmmFloat _ w = lit
1060 Amode addr code <- memConstant (widthInBytes w) lit
1061 return (OpAddr addr, code)
1064 if is32BitLit lit && not (isFloatType (cmmLitType lit))
1065 then return (OpImm (litToImm lit), nilOL)
1066 else getNonClobberedOperand_generic (CmmLit lit)
1068 getNonClobberedOperand (CmmLoad mem pk) = do
1069 use_sse2 <- sse2Enabled
1070 if (not (isFloatType pk) || use_sse2)
1071 && IF_ARCH_i386(not (isWord64 pk), True)
1073 Amode src mem_code <- getAmode mem
1075 if (amodeCouldBeClobbered src)
1077 tmp <- getNewRegNat archWordSize
1078 return (AddrBaseIndex (EABaseReg tmp) EAIndexNone (ImmInt 0),
1079 unitOL (LEA II32 (OpAddr src) (OpReg tmp)))
1082 return (OpAddr src', save_code `appOL` mem_code)
1084 getNonClobberedOperand_generic (CmmLoad mem pk)
1086 getNonClobberedOperand e = getNonClobberedOperand_generic e
1088 getNonClobberedOperand_generic :: CmmExpr -> NatM (Operand, InstrBlock)
1089 getNonClobberedOperand_generic e = do
1090 (reg, code) <- getNonClobberedReg e
1091 return (OpReg reg, code)
1093 amodeCouldBeClobbered :: AddrMode -> Bool
1094 amodeCouldBeClobbered amode = any regClobbered (addrModeRegs amode)
1096 regClobbered (RegReal (RealRegSingle rr)) = isFastTrue (freeReg rr)
1097 regClobbered _ = False
1099 -- getOperand: the operand is not required to remain valid across the
1100 -- computation of an arbitrary expression.
1101 getOperand :: CmmExpr -> NatM (Operand, InstrBlock)
1103 getOperand (CmmLit lit) = do
1104 use_sse2 <- sse2Enabled
1105 if (use_sse2 && isSuitableFloatingPointLit lit)
1107 let CmmFloat _ w = lit
1108 Amode addr code <- memConstant (widthInBytes w) lit
1109 return (OpAddr addr, code)
1112 if is32BitLit lit && not (isFloatType (cmmLitType lit))
1113 then return (OpImm (litToImm lit), nilOL)
1114 else getOperand_generic (CmmLit lit)
1116 getOperand (CmmLoad mem pk) = do
1117 use_sse2 <- sse2Enabled
1118 if (not (isFloatType pk) || use_sse2) && IF_ARCH_i386(not (isWord64 pk), True)
1120 Amode src mem_code <- getAmode mem
1121 return (OpAddr src, mem_code)
1123 getOperand_generic (CmmLoad mem pk)
1125 getOperand e = getOperand_generic e
1127 getOperand_generic e = do
1128 (reg, code) <- getSomeReg e
1129 return (OpReg reg, code)
1131 isOperand :: CmmExpr -> Bool
1132 isOperand (CmmLoad _ _) = True
1133 isOperand (CmmLit lit) = is32BitLit lit
1134 || isSuitableFloatingPointLit lit
1137 memConstant :: Int -> CmmLit -> NatM Amode
1138 memConstant align lit = do
1139 #ifdef x86_64_TARGET_ARCH
1140 lbl <- getNewLabelNat
1141 let addr = ripRel (ImmCLbl lbl)
1144 lbl <- getNewLabelNat
1145 dflags <- getDynFlagsNat
1146 dynRef <- cmmMakeDynamicReference dflags addImportNat DataReference lbl
1147 Amode addr addr_code <- getAmode dynRef
1155 return (Amode addr code)
1158 loadFloatAmode :: Bool -> Width -> AddrMode -> InstrBlock -> NatM Register
1159 loadFloatAmode use_sse2 w addr addr_code = do
1160 let size = floatSize w
1161 code dst = addr_code `snocOL`
1163 then MOV size (OpAddr addr) (OpReg dst)
1164 else GLD size addr dst
1166 return (Any (if use_sse2 then size else FF80) code)
1169 -- if we want a floating-point literal as an operand, we can
1170 -- use it directly from memory. However, if the literal is
1171 -- zero, we're better off generating it into a register using
1173 isSuitableFloatingPointLit (CmmFloat f _) = f /= 0.0
1174 isSuitableFloatingPointLit _ = False
1176 getRegOrMem :: CmmExpr -> NatM (Operand, InstrBlock)
1177 getRegOrMem e@(CmmLoad mem pk) = do
1178 use_sse2 <- sse2Enabled
1179 if (not (isFloatType pk) || use_sse2) && IF_ARCH_i386(not (isWord64 pk), True)
1181 Amode src mem_code <- getAmode mem
1182 return (OpAddr src, mem_code)
1184 (reg, code) <- getNonClobberedReg e
1185 return (OpReg reg, code)
1187 (reg, code) <- getNonClobberedReg e
1188 return (OpReg reg, code)
1190 #if x86_64_TARGET_ARCH
1191 is32BitLit (CmmInt i W64) = is32BitInteger i
1192 -- assume that labels are in the range 0-2^31-1: this assumes the
1193 -- small memory model (see gcc docs, -mcmodel=small).
1200 -- Set up a condition code for a conditional branch.
1202 getCondCode :: CmmExpr -> NatM CondCode
1204 -- yes, they really do seem to want exactly the same!
1206 getCondCode (CmmMachOp mop [x, y])
1209 MO_F_Eq W32 -> condFltCode EQQ x y
1210 MO_F_Ne W32 -> condFltCode NE x y
1211 MO_F_Gt W32 -> condFltCode GTT x y
1212 MO_F_Ge W32 -> condFltCode GE x y
1213 MO_F_Lt W32 -> condFltCode LTT x y
1214 MO_F_Le W32 -> condFltCode LE x y
1216 MO_F_Eq W64 -> condFltCode EQQ x y
1217 MO_F_Ne W64 -> condFltCode NE x y
1218 MO_F_Gt W64 -> condFltCode GTT x y
1219 MO_F_Ge W64 -> condFltCode GE x y
1220 MO_F_Lt W64 -> condFltCode LTT x y
1221 MO_F_Le W64 -> condFltCode LE x y
1223 MO_Eq rep -> condIntCode EQQ x y
1224 MO_Ne rep -> condIntCode NE x y
1226 MO_S_Gt rep -> condIntCode GTT x y
1227 MO_S_Ge rep -> condIntCode GE x y
1228 MO_S_Lt rep -> condIntCode LTT x y
1229 MO_S_Le rep -> condIntCode LE x y
1231 MO_U_Gt rep -> condIntCode GU x y
1232 MO_U_Ge rep -> condIntCode GEU x y
1233 MO_U_Lt rep -> condIntCode LU x y
1234 MO_U_Le rep -> condIntCode LEU x y
1236 other -> pprPanic "getCondCode(x86,x86_64,sparc)" (ppr (CmmMachOp mop [x,y]))
1238 getCondCode other = pprPanic "getCondCode(2)(x86,sparc)" (ppr other)
1243 -- @cond(Int|Flt)Code@: Turn a boolean expression into a condition, to be
1244 -- passed back up the tree.
1246 condIntCode :: Cond -> CmmExpr -> CmmExpr -> NatM CondCode
1248 -- memory vs immediate
1249 condIntCode cond (CmmLoad x pk) (CmmLit lit) | is32BitLit lit = do
1250 Amode x_addr x_code <- getAmode x
1253 code = x_code `snocOL`
1254 CMP (cmmTypeSize pk) (OpImm imm) (OpAddr x_addr)
1256 return (CondCode False cond code)
1258 -- anything vs zero, using a mask
1259 -- TODO: Add some sanity checking!!!!
1260 condIntCode cond (CmmMachOp (MO_And rep) [x,o2]) (CmmLit (CmmInt 0 pk))
1261 | (CmmLit lit@(CmmInt mask pk2)) <- o2, is32BitLit lit
1263 (x_reg, x_code) <- getSomeReg x
1265 code = x_code `snocOL`
1266 TEST (intSize pk) (OpImm (ImmInteger mask)) (OpReg x_reg)
1268 return (CondCode False cond code)
1271 condIntCode cond x (CmmLit (CmmInt 0 pk)) = do
1272 (x_reg, x_code) <- getSomeReg x
1274 code = x_code `snocOL`
1275 TEST (intSize pk) (OpReg x_reg) (OpReg x_reg)
1277 return (CondCode False cond code)
1279 -- anything vs operand
1280 condIntCode cond x y | isOperand y = do
1281 (x_reg, x_code) <- getNonClobberedReg x
1282 (y_op, y_code) <- getOperand y
1284 code = x_code `appOL` y_code `snocOL`
1285 CMP (cmmTypeSize (cmmExprType x)) y_op (OpReg x_reg)
1287 return (CondCode False cond code)
1289 -- anything vs anything
1290 condIntCode cond x y = do
1291 (y_reg, y_code) <- getNonClobberedReg y
1292 (x_op, x_code) <- getRegOrMem x
1294 code = y_code `appOL`
1296 CMP (cmmTypeSize (cmmExprType x)) (OpReg y_reg) x_op
1298 return (CondCode False cond code)
1302 --------------------------------------------------------------------------------
1303 condFltCode :: Cond -> CmmExpr -> CmmExpr -> NatM CondCode
1305 condFltCode cond x y
1306 = if_sse2 condFltCode_sse2 condFltCode_x87
1310 = ASSERT(cond `elem` ([EQQ, NE, LE, LTT, GE, GTT])) do
1311 (x_reg, x_code) <- getNonClobberedReg x
1312 (y_reg, y_code) <- getSomeReg y
1313 use_sse2 <- sse2Enabled
1315 code = x_code `appOL` y_code `snocOL`
1316 GCMP cond x_reg y_reg
1317 -- The GCMP insn does the test and sets the zero flag if comparable
1318 -- and true. Hence we always supply EQQ as the condition to test.
1319 return (CondCode True EQQ code)
1321 -- in the SSE2 comparison ops (ucomiss, ucomisd) the left arg may be
1322 -- an operand, but the right must be a reg. We can probably do better
1323 -- than this general case...
1324 condFltCode_sse2 = do
1325 (x_reg, x_code) <- getNonClobberedReg x
1326 (y_op, y_code) <- getOperand y
1328 code = x_code `appOL`
1330 CMP (floatSize $ cmmExprWidth x) y_op (OpReg x_reg)
1331 -- NB(1): we need to use the unsigned comparison operators on the
1332 -- result of this comparison.
1334 return (CondCode True (condToUnsigned cond) code)
1336 -- -----------------------------------------------------------------------------
1337 -- Generating assignments
1339 -- Assignments are really at the heart of the whole code generation
1340 -- business. Almost all top-level nodes of any real importance are
1341 -- assignments, which correspond to loads, stores, or register
1342 -- transfers. If we're really lucky, some of the register transfers
1343 -- will go away, because we can use the destination register to
1344 -- complete the code generation for the right hand side. This only
1345 -- fails when the right hand side is forced into a fixed register
1346 -- (e.g. the result of a call).
1348 assignMem_IntCode :: Size -> CmmExpr -> CmmExpr -> NatM InstrBlock
1349 assignReg_IntCode :: Size -> CmmReg -> CmmExpr -> NatM InstrBlock
1351 assignMem_FltCode :: Size -> CmmExpr -> CmmExpr -> NatM InstrBlock
1352 assignReg_FltCode :: Size -> CmmReg -> CmmExpr -> NatM InstrBlock
1355 -- integer assignment to memory
1357 -- specific case of adding/subtracting an integer to a particular address.
1358 -- ToDo: catch other cases where we can use an operation directly on a memory
1360 assignMem_IntCode pk addr (CmmMachOp op [CmmLoad addr2 _,
1361 CmmLit (CmmInt i _)])
1362 | addr == addr2, pk /= II64 || is32BitInteger i,
1363 Just instr <- check op
1364 = do Amode amode code_addr <- getAmode addr
1365 let code = code_addr `snocOL`
1366 instr pk (OpImm (ImmInt (fromIntegral i))) (OpAddr amode)
1369 check (MO_Add _) = Just ADD
1370 check (MO_Sub _) = Just SUB
1375 assignMem_IntCode pk addr src = do
1376 Amode addr code_addr <- getAmode addr
1377 (code_src, op_src) <- get_op_RI src
1379 code = code_src `appOL`
1381 MOV pk op_src (OpAddr addr)
1382 -- NOTE: op_src is stable, so it will still be valid
1383 -- after code_addr. This may involve the introduction
1384 -- of an extra MOV to a temporary register, but we hope
1385 -- the register allocator will get rid of it.
1389 get_op_RI :: CmmExpr -> NatM (InstrBlock,Operand) -- code, operator
1390 get_op_RI (CmmLit lit) | is32BitLit lit
1391 = return (nilOL, OpImm (litToImm lit))
1393 = do (reg,code) <- getNonClobberedReg op
1394 return (code, OpReg reg)
1397 -- Assign; dst is a reg, rhs is mem
1398 assignReg_IntCode pk reg (CmmLoad src _) = do
1399 load_code <- intLoadCode (MOV pk) src
1400 return (load_code (getRegisterReg False{-no sse2-} reg))
1402 -- dst is a reg, but src could be anything
1403 assignReg_IntCode pk reg src = do
1404 code <- getAnyReg src
1405 return (code (getRegisterReg False{-no sse2-} reg))
1408 -- Floating point assignment to memory
1409 assignMem_FltCode pk addr src = do
1410 (src_reg, src_code) <- getNonClobberedReg src
1411 Amode addr addr_code <- getAmode addr
1412 use_sse2 <- sse2Enabled
1414 code = src_code `appOL`
1416 if use_sse2 then MOV pk (OpReg src_reg) (OpAddr addr)
1417 else GST pk src_reg addr
1420 -- Floating point assignment to a register/temporary
1421 assignReg_FltCode pk reg src = do
1422 use_sse2 <- sse2Enabled
1423 src_code <- getAnyReg src
1424 return (src_code (getRegisterReg use_sse2 reg))
1427 genJump :: CmmExpr{-the branch target-} -> NatM InstrBlock
1429 genJump (CmmLoad mem pk) = do
1430 Amode target code <- getAmode mem
1431 return (code `snocOL` JMP (OpAddr target))
1433 genJump (CmmLit lit) = do
1434 return (unitOL (JMP (OpImm (litToImm lit))))
1437 (reg,code) <- getSomeReg expr
1438 return (code `snocOL` JMP (OpReg reg))
1441 -- -----------------------------------------------------------------------------
1442 -- Unconditional branches
1444 genBranch :: BlockId -> NatM InstrBlock
1445 genBranch = return . toOL . mkJumpInstr
1449 -- -----------------------------------------------------------------------------
1450 -- Conditional jumps
1453 Conditional jumps are always to local labels, so we can use branch
1454 instructions. We peek at the arguments to decide what kind of
1457 I386: First, we have to ensure that the condition
1458 codes are set according to the supplied comparison operation.
1462 :: BlockId -- the branch target
1463 -> CmmExpr -- the condition on which to branch
1466 genCondJump id bool = do
1467 CondCode is_float cond cond_code <- getCondCode bool
1468 use_sse2 <- sse2Enabled
1469 if not is_float || not use_sse2
1471 return (cond_code `snocOL` JXX cond id)
1473 lbl <- getBlockIdNat
1475 -- see comment with condFltReg
1476 let code = case cond of
1482 plain_test = unitOL (
1485 or_unordered = toOL [
1489 and_ordered = toOL [
1495 return (cond_code `appOL` code)
1498 -- -----------------------------------------------------------------------------
1499 -- Generating C calls
1501 -- Now the biggest nightmare---calls. Most of the nastiness is buried in
1502 -- @get_arg@, which moves the arguments to the correct registers/stack
1503 -- locations. Apart from that, the code is easy.
1505 -- (If applicable) Do not fill the delay slots here; you will confuse the
1506 -- register allocator.
1509 :: CmmCallTarget -- function to call
1510 -> HintedCmmFormals -- where to put the result
1511 -> HintedCmmActuals -- arguments (of mixed type)
1514 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
1516 #if i386_TARGET_ARCH
1518 genCCall (CmmPrim MO_WriteBarrier) _ _ = return nilOL
1519 -- write barrier compiles to no code on x86/x86-64;
1520 -- we keep it this long in order to prevent earlier optimisations.
1522 -- we only cope with a single result for foreign calls
1523 genCCall (CmmPrim op) [CmmHinted r _] args = do
1524 l1 <- getNewLabelNat
1525 l2 <- getNewLabelNat
1529 outOfLineFloatOp op r args
1531 MO_F32_Sqrt -> actuallyInlineFloatOp GSQRT FF32 args
1532 MO_F64_Sqrt -> actuallyInlineFloatOp GSQRT FF64 args
1534 MO_F32_Sin -> actuallyInlineFloatOp (\s -> GSIN s l1 l2) FF32 args
1535 MO_F64_Sin -> actuallyInlineFloatOp (\s -> GSIN s l1 l2) FF64 args
1537 MO_F32_Cos -> actuallyInlineFloatOp (\s -> GCOS s l1 l2) FF32 args
1538 MO_F64_Cos -> actuallyInlineFloatOp (\s -> GCOS s l1 l2) FF64 args
1540 MO_F32_Tan -> actuallyInlineFloatOp (\s -> GTAN s l1 l2) FF32 args
1541 MO_F64_Tan -> actuallyInlineFloatOp (\s -> GTAN s l1 l2) FF64 args
1543 other_op -> outOfLineFloatOp op r args
1546 actuallyInlineFloatOp instr size [CmmHinted x _]
1547 = do res <- trivialUFCode size (instr size) x
1549 return (any (getRegisterReg False (CmmLocal r)))
1551 genCCall target dest_regs args = do
1553 sizes = map (arg_size . cmmExprType . hintlessCmm) (reverse args)
1554 #if !darwin_TARGET_OS
1555 tot_arg_size = sum sizes
1557 raw_arg_size = sum sizes
1558 tot_arg_size = roundTo 16 raw_arg_size
1559 arg_pad_size = tot_arg_size - raw_arg_size
1560 delta0 <- getDeltaNat
1561 setDeltaNat (delta0 - arg_pad_size)
1564 use_sse2 <- sse2Enabled
1565 push_codes <- mapM (push_arg use_sse2) (reverse args)
1566 delta <- getDeltaNat
1569 -- deal with static vs dynamic call targets
1570 (callinsns,cconv) <-
1573 CmmCallee (CmmLit (CmmLabel lbl)) conv
1574 -> -- ToDo: stdcall arg sizes
1575 return (unitOL (CALL (Left fn_imm) []), conv)
1576 where fn_imm = ImmCLbl lbl
1578 -> do { (dyn_c, dyn_r) <- get_op expr
1579 ; ASSERT( isWord32 (cmmExprType expr) )
1580 return (dyn_c `snocOL` CALL (Right dyn_r) [], conv) }
1583 #if darwin_TARGET_OS
1585 = toOL [SUB II32 (OpImm (ImmInt arg_pad_size)) (OpReg esp),
1586 DELTA (delta0 - arg_pad_size)]
1587 `appOL` concatOL push_codes
1590 = concatOL push_codes
1591 call = callinsns `appOL`
1593 -- Deallocate parameters after call for ccall;
1594 -- but not for stdcall (callee does it)
1595 (if cconv == StdCallConv || tot_arg_size==0 then [] else
1596 [ADD II32 (OpImm (ImmInt tot_arg_size)) (OpReg esp)])
1598 [DELTA (delta + tot_arg_size)]
1601 setDeltaNat (delta + tot_arg_size)
1604 -- assign the results, if necessary
1605 assign_code [] = nilOL
1606 assign_code [CmmHinted dest _hint]
1609 then let tmp_amode = AddrBaseIndex (EABaseReg esp)
1613 in toOL [ SUB II32 (OpImm (ImmInt b)) (OpReg esp),
1614 GST sz fake0 tmp_amode,
1615 MOV sz (OpAddr tmp_amode) (OpReg r_dest),
1616 ADD II32 (OpImm (ImmInt b)) (OpReg esp)]
1617 else unitOL (GMOV fake0 r_dest)
1618 | isWord64 ty = toOL [MOV II32 (OpReg eax) (OpReg r_dest),
1619 MOV II32 (OpReg edx) (OpReg r_dest_hi)]
1620 | otherwise = unitOL (MOV (intSize w) (OpReg eax) (OpReg r_dest))
1622 ty = localRegType dest
1625 r_dest_hi = getHiVRegFromLo r_dest
1626 r_dest = getRegisterReg use_sse2 (CmmLocal dest)
1627 assign_code many = pprPanic "genCCall.assign_code - too many return values:" (ppr many)
1629 return (push_code `appOL`
1631 assign_code dest_regs)
1634 arg_size :: CmmType -> Int -- Width in bytes
1635 arg_size ty = widthInBytes (typeWidth ty)
1637 roundTo a x | x `mod` a == 0 = x
1638 | otherwise = x + a - (x `mod` a)
1641 push_arg :: Bool -> HintedCmmActual {-current argument-}
1642 -> NatM InstrBlock -- code
1644 push_arg use_sse2 (CmmHinted arg _hint) -- we don't need the hints on x86
1645 | isWord64 arg_ty = do
1646 ChildCode64 code r_lo <- iselExpr64 arg
1647 delta <- getDeltaNat
1648 setDeltaNat (delta - 8)
1650 r_hi = getHiVRegFromLo r_lo
1652 return ( code `appOL`
1653 toOL [PUSH II32 (OpReg r_hi), DELTA (delta - 4),
1654 PUSH II32 (OpReg r_lo), DELTA (delta - 8),
1659 (code, reg) <- get_op arg
1660 delta <- getDeltaNat
1661 let size = arg_size arg_ty -- Byte size
1662 setDeltaNat (delta-size)
1663 if (isFloatType arg_ty)
1664 then return (code `appOL`
1665 toOL [SUB II32 (OpImm (ImmInt size)) (OpReg esp),
1667 let addr = AddrBaseIndex (EABaseReg esp)
1670 size = floatSize (typeWidth arg_ty)
1673 then MOV size (OpReg reg) (OpAddr addr)
1674 else GST size reg addr
1677 else return (code `snocOL`
1678 PUSH II32 (OpReg reg) `snocOL`
1682 arg_ty = cmmExprType arg
1685 get_op :: CmmExpr -> NatM (InstrBlock, Reg) -- code, reg
1687 (reg,code) <- getSomeReg op
1690 #elif x86_64_TARGET_ARCH
1692 genCCall (CmmPrim MO_WriteBarrier) _ _ = return nilOL
1693 -- write barrier compiles to no code on x86/x86-64;
1694 -- we keep it this long in order to prevent earlier optimisations.
1697 genCCall (CmmPrim op) [CmmHinted r _] args =
1698 outOfLineFloatOp op r args
1700 genCCall target dest_regs args = do
1702 -- load up the register arguments
1703 (stack_args, aregs, fregs, load_args_code)
1704 <- load_args args allArgRegs allFPArgRegs nilOL
1707 fp_regs_used = reverse (drop (length fregs) (reverse allFPArgRegs))
1708 int_regs_used = reverse (drop (length aregs) (reverse allArgRegs))
1709 arg_regs = [eax] ++ int_regs_used ++ fp_regs_used
1710 -- for annotating the call instruction with
1712 sse_regs = length fp_regs_used
1714 tot_arg_size = arg_size * length stack_args
1716 -- On entry to the called function, %rsp should be aligned
1717 -- on a 16-byte boundary +8 (i.e. the first stack arg after
1718 -- the return address is 16-byte aligned). In STG land
1719 -- %rsp is kept 16-byte aligned (see StgCRun.c), so we just
1720 -- need to make sure we push a multiple of 16-bytes of args,
1721 -- plus the return address, to get the correct alignment.
1722 -- Urg, this is hard. We need to feed the delta back into
1723 -- the arg pushing code.
1724 (real_size, adjust_rsp) <-
1725 if tot_arg_size `rem` 16 == 0
1726 then return (tot_arg_size, nilOL)
1727 else do -- we need to adjust...
1728 delta <- getDeltaNat
1729 setDeltaNat (delta-8)
1730 return (tot_arg_size+8, toOL [
1731 SUB II64 (OpImm (ImmInt 8)) (OpReg rsp),
1735 -- push the stack args, right to left
1736 push_code <- push_args (reverse stack_args) nilOL
1737 delta <- getDeltaNat
1739 -- deal with static vs dynamic call targets
1740 (callinsns,cconv) <-
1743 CmmCallee (CmmLit (CmmLabel lbl)) conv
1744 -> -- ToDo: stdcall arg sizes
1745 return (unitOL (CALL (Left fn_imm) arg_regs), conv)
1746 where fn_imm = ImmCLbl lbl
1748 -> do (dyn_r, dyn_c) <- getSomeReg expr
1749 return (dyn_c `snocOL` CALL (Right dyn_r) arg_regs, conv)
1752 -- The x86_64 ABI requires us to set %al to the number of SSE2
1753 -- registers that contain arguments, if the called routine
1754 -- is a varargs function. We don't know whether it's a
1755 -- varargs function or not, so we have to assume it is.
1757 -- It's not safe to omit this assignment, even if the number
1758 -- of SSE2 regs in use is zero. If %al is larger than 8
1759 -- on entry to a varargs function, seg faults ensue.
1760 assign_eax n = unitOL (MOV II32 (OpImm (ImmInt n)) (OpReg eax))
1762 let call = callinsns `appOL`
1764 -- Deallocate parameters after call for ccall;
1765 -- but not for stdcall (callee does it)
1766 (if cconv == StdCallConv || real_size==0 then [] else
1767 [ADD (intSize wordWidth) (OpImm (ImmInt real_size)) (OpReg esp)])
1769 [DELTA (delta + real_size)]
1772 setDeltaNat (delta + real_size)
1775 -- assign the results, if necessary
1776 assign_code [] = nilOL
1777 assign_code [CmmHinted dest _hint] =
1778 case typeWidth rep of
1779 W32 | isFloatType rep -> unitOL (MOV (floatSize W32) (OpReg xmm0) (OpReg r_dest))
1780 W64 | isFloatType rep -> unitOL (MOV (floatSize W64) (OpReg xmm0) (OpReg r_dest))
1781 _ -> unitOL (MOV (cmmTypeSize rep) (OpReg rax) (OpReg r_dest))
1783 rep = localRegType dest
1784 r_dest = getRegisterReg True (CmmLocal dest)
1785 assign_code many = panic "genCCall.assign_code many"
1787 return (load_args_code `appOL`
1790 assign_eax sse_regs `appOL`
1792 assign_code dest_regs)
1795 arg_size = 8 -- always, at the mo
1797 load_args :: [CmmHinted CmmExpr]
1798 -> [Reg] -- int regs avail for args
1799 -> [Reg] -- FP regs avail for args
1801 -> NatM ([CmmHinted CmmExpr],[Reg],[Reg],InstrBlock)
1802 load_args args [] [] code = return (args, [], [], code)
1803 -- no more regs to use
1804 load_args [] aregs fregs code = return ([], aregs, fregs, code)
1805 -- no more args to push
1806 load_args ((CmmHinted arg hint) : rest) aregs fregs code
1807 | isFloatType arg_rep =
1811 arg_code <- getAnyReg arg
1812 load_args rest aregs rs (code `appOL` arg_code r)
1817 arg_code <- getAnyReg arg
1818 load_args rest rs fregs (code `appOL` arg_code r)
1820 arg_rep = cmmExprType arg
1823 (args',ars,frs,code') <- load_args rest aregs fregs code
1824 return ((CmmHinted arg hint):args', ars, frs, code')
1826 push_args [] code = return code
1827 push_args ((CmmHinted arg hint):rest) code
1828 | isFloatType arg_rep = do
1829 (arg_reg, arg_code) <- getSomeReg arg
1830 delta <- getDeltaNat
1831 setDeltaNat (delta-arg_size)
1832 let code' = code `appOL` arg_code `appOL` toOL [
1833 SUB (intSize wordWidth) (OpImm (ImmInt arg_size)) (OpReg rsp) ,
1834 DELTA (delta-arg_size),
1835 MOV (floatSize width) (OpReg arg_reg) (OpAddr (spRel 0))]
1836 push_args rest code'
1839 -- we only ever generate word-sized function arguments. Promotion
1840 -- has already happened: our Int8# type is kept sign-extended
1841 -- in an Int#, for example.
1842 ASSERT(width == W64) return ()
1843 (arg_op, arg_code) <- getOperand arg
1844 delta <- getDeltaNat
1845 setDeltaNat (delta-arg_size)
1846 let code' = code `appOL` arg_code `appOL` toOL [
1848 DELTA (delta-arg_size)]
1849 push_args rest code'
1851 arg_rep = cmmExprType arg
1852 width = typeWidth arg_rep
1855 genCCall = panic "X86.genCCAll: not defined"
1857 #endif /* x86_64_TARGET_ARCH */
1862 outOfLineFloatOp :: CallishMachOp -> CmmFormal -> HintedCmmActuals -> NatM InstrBlock
1863 outOfLineFloatOp mop res args
1865 dflags <- getDynFlagsNat
1866 targetExpr <- cmmMakeDynamicReference dflags addImportNat CallReference lbl
1867 let target = CmmCallee targetExpr CCallConv
1869 stmtToInstrs (CmmCall target [CmmHinted res NoHint] args CmmUnsafe CmmMayReturn)
1871 -- Assume we can call these functions directly, and that they're not in a dynamic library.
1872 -- TODO: Why is this ok? Under linux this code will be in libm.so
1873 -- Is is because they're really implemented as a primitive instruction by the assembler?? -- BL 2009/12/31
1874 lbl = mkForeignLabel fn Nothing ForeignLabelInThisPackage IsFunction
1877 MO_F32_Sqrt -> fsLit "sqrtf"
1878 MO_F32_Sin -> fsLit "sinf"
1879 MO_F32_Cos -> fsLit "cosf"
1880 MO_F32_Tan -> fsLit "tanf"
1881 MO_F32_Exp -> fsLit "expf"
1882 MO_F32_Log -> fsLit "logf"
1884 MO_F32_Asin -> fsLit "asinf"
1885 MO_F32_Acos -> fsLit "acosf"
1886 MO_F32_Atan -> fsLit "atanf"
1888 MO_F32_Sinh -> fsLit "sinhf"
1889 MO_F32_Cosh -> fsLit "coshf"
1890 MO_F32_Tanh -> fsLit "tanhf"
1891 MO_F32_Pwr -> fsLit "powf"
1893 MO_F64_Sqrt -> fsLit "sqrt"
1894 MO_F64_Sin -> fsLit "sin"
1895 MO_F64_Cos -> fsLit "cos"
1896 MO_F64_Tan -> fsLit "tan"
1897 MO_F64_Exp -> fsLit "exp"
1898 MO_F64_Log -> fsLit "log"
1900 MO_F64_Asin -> fsLit "asin"
1901 MO_F64_Acos -> fsLit "acos"
1902 MO_F64_Atan -> fsLit "atan"
1904 MO_F64_Sinh -> fsLit "sinh"
1905 MO_F64_Cosh -> fsLit "cosh"
1906 MO_F64_Tanh -> fsLit "tanh"
1907 MO_F64_Pwr -> fsLit "pow"
1913 -- -----------------------------------------------------------------------------
1914 -- Generating a table-branch
1916 genSwitch :: CmmExpr -> [Maybe BlockId] -> NatM InstrBlock
1921 (reg,e_code) <- getSomeReg expr
1922 lbl <- getNewLabelNat
1923 dflags <- getDynFlagsNat
1924 dynRef <- cmmMakeDynamicReference dflags addImportNat DataReference lbl
1925 (tableReg,t_code) <- getSomeReg $ dynRef
1927 jumpTable = map jumpTableEntryRel ids
1929 jumpTableEntryRel Nothing
1930 = CmmStaticLit (CmmInt 0 wordWidth)
1931 jumpTableEntryRel (Just (BlockId id))
1932 = CmmStaticLit (CmmLabelDiffOff blockLabel lbl 0)
1933 where blockLabel = mkAsmTempLabel id
1935 op = OpAddr (AddrBaseIndex (EABaseReg tableReg)
1936 (EAIndex reg wORD_SIZE) (ImmInt 0))
1938 #if x86_64_TARGET_ARCH
1939 #if darwin_TARGET_OS
1940 -- on Mac OS X/x86_64, put the jump table in the text section
1941 -- to work around a limitation of the linker.
1942 -- ld64 is unable to handle the relocations for
1944 -- if L0 is not preceded by a non-anonymous label in its section.
1946 code = e_code `appOL` t_code `appOL` toOL [
1947 ADD (intSize wordWidth) op (OpReg tableReg),
1948 JMP_TBL (OpReg tableReg) [ id | Just id <- ids ],
1949 LDATA Text (CmmDataLabel lbl : jumpTable)
1952 -- HACK: On x86_64 binutils<2.17 is only able to generate PC32
1953 -- relocations, hence we only get 32-bit offsets in the jump
1954 -- table. As these offsets are always negative we need to properly
1955 -- sign extend them to 64-bit. This hack should be removed in
1956 -- conjunction with the hack in PprMach.hs/pprDataItem once
1957 -- binutils 2.17 is standard.
1958 code = e_code `appOL` t_code `appOL` toOL [
1959 LDATA ReadOnlyData (CmmDataLabel lbl : jumpTable),
1961 (OpAddr (AddrBaseIndex (EABaseReg tableReg)
1962 (EAIndex reg wORD_SIZE) (ImmInt 0)))
1964 ADD (intSize wordWidth) (OpReg reg) (OpReg tableReg),
1965 JMP_TBL (OpReg tableReg) [ id | Just id <- ids ]
1969 code = e_code `appOL` t_code `appOL` toOL [
1970 LDATA ReadOnlyData (CmmDataLabel lbl : jumpTable),
1971 ADD (intSize wordWidth) op (OpReg tableReg),
1972 JMP_TBL (OpReg tableReg) [ id | Just id <- ids ]
1978 (reg,e_code) <- getSomeReg expr
1979 lbl <- getNewLabelNat
1981 jumpTable = map jumpTableEntry ids
1982 op = OpAddr (AddrBaseIndex EABaseNone (EAIndex reg wORD_SIZE) (ImmCLbl lbl))
1983 code = e_code `appOL` toOL [
1984 LDATA ReadOnlyData (CmmDataLabel lbl : jumpTable),
1985 JMP_TBL op [ id | Just id <- ids ]
1991 -- -----------------------------------------------------------------------------
1992 -- 'condIntReg' and 'condFltReg': condition codes into registers
1994 -- Turn those condition codes into integers now (when they appear on
1995 -- the right hand side of an assignment).
1997 -- (If applicable) Do not fill the delay slots here; you will confuse the
1998 -- register allocator.
2000 condIntReg :: Cond -> CmmExpr -> CmmExpr -> NatM Register
2002 condIntReg cond x y = do
2003 CondCode _ cond cond_code <- condIntCode cond x y
2004 tmp <- getNewRegNat II8
2006 code dst = cond_code `appOL` toOL [
2007 SETCC cond (OpReg tmp),
2008 MOVZxL II8 (OpReg tmp) (OpReg dst)
2011 return (Any II32 code)
2015 condFltReg :: Cond -> CmmExpr -> CmmExpr -> NatM Register
2016 condFltReg cond x y = if_sse2 condFltReg_sse2 condFltReg_x87
2019 CondCode _ cond cond_code <- condFltCode cond x y
2020 tmp <- getNewRegNat II8
2022 code dst = cond_code `appOL` toOL [
2023 SETCC cond (OpReg tmp),
2024 MOVZxL II8 (OpReg tmp) (OpReg dst)
2027 return (Any II32 code)
2029 condFltReg_sse2 = do
2030 CondCode _ cond cond_code <- condFltCode cond x y
2031 tmp1 <- getNewRegNat archWordSize
2032 tmp2 <- getNewRegNat archWordSize
2034 -- We have to worry about unordered operands (eg. comparisons
2035 -- against NaN). If the operands are unordered, the comparison
2036 -- sets the parity flag, carry flag and zero flag.
2037 -- All comparisons are supposed to return false for unordered
2038 -- operands except for !=, which returns true.
2040 -- Optimisation: we don't have to test the parity flag if we
2041 -- know the test has already excluded the unordered case: eg >
2042 -- and >= test for a zero carry flag, which can only occur for
2043 -- ordered operands.
2045 -- ToDo: by reversing comparisons we could avoid testing the
2046 -- parity flag in more cases.
2051 NE -> or_unordered dst
2052 GU -> plain_test dst
2053 GEU -> plain_test dst
2054 _ -> and_ordered dst)
2056 plain_test dst = toOL [
2057 SETCC cond (OpReg tmp1),
2058 MOVZxL II8 (OpReg tmp1) (OpReg dst)
2060 or_unordered dst = toOL [
2061 SETCC cond (OpReg tmp1),
2062 SETCC PARITY (OpReg tmp2),
2063 OR II8 (OpReg tmp1) (OpReg tmp2),
2064 MOVZxL II8 (OpReg tmp2) (OpReg dst)
2066 and_ordered dst = toOL [
2067 SETCC cond (OpReg tmp1),
2068 SETCC NOTPARITY (OpReg tmp2),
2069 AND II8 (OpReg tmp1) (OpReg tmp2),
2070 MOVZxL II8 (OpReg tmp2) (OpReg dst)
2073 return (Any II32 code)
2076 -- -----------------------------------------------------------------------------
2077 -- 'trivial*Code': deal with trivial instructions
2079 -- Trivial (dyadic: 'trivialCode', floating-point: 'trivialFCode',
2080 -- unary: 'trivialUCode', unary fl-pt:'trivialUFCode') instructions.
2081 -- Only look for constants on the right hand side, because that's
2082 -- where the generic optimizer will have put them.
2084 -- Similarly, for unary instructions, we don't have to worry about
2085 -- matching an StInt as the argument, because genericOpt will already
2086 -- have handled the constant-folding.
2090 The Rules of the Game are:
2092 * You cannot assume anything about the destination register dst;
2093 it may be anything, including a fixed reg.
2095 * You may compute an operand into a fixed reg, but you may not
2096 subsequently change the contents of that fixed reg. If you
2097 want to do so, first copy the value either to a temporary
2098 or into dst. You are free to modify dst even if it happens
2099 to be a fixed reg -- that's not your problem.
2101 * You cannot assume that a fixed reg will stay live over an
2102 arbitrary computation. The same applies to the dst reg.
2104 * Temporary regs obtained from getNewRegNat are distinct from
2105 each other and from all other regs, and stay live over
2106 arbitrary computations.
2108 --------------------
2110 SDM's version of The Rules:
2112 * If getRegister returns Any, that means it can generate correct
2113 code which places the result in any register, period. Even if that
2114 register happens to be read during the computation.
2116 Corollary #1: this means that if you are generating code for an
2117 operation with two arbitrary operands, you cannot assign the result
2118 of the first operand into the destination register before computing
2119 the second operand. The second operand might require the old value
2120 of the destination register.
2122 Corollary #2: A function might be able to generate more efficient
2123 code if it knows the destination register is a new temporary (and
2124 therefore not read by any of the sub-computations).
2126 * If getRegister returns Any, then the code it generates may modify only:
2127 (a) fresh temporaries
2128 (b) the destination register
2129 (c) known registers (eg. %ecx is used by shifts)
2130 In particular, it may *not* modify global registers, unless the global
2131 register happens to be the destination register.
2134 trivialCode width instr (Just revinstr) (CmmLit lit_a) b
2135 | is32BitLit lit_a = do
2136 b_code <- getAnyReg b
2139 = b_code dst `snocOL`
2140 revinstr (OpImm (litToImm lit_a)) (OpReg dst)
2142 return (Any (intSize width) code)
2144 trivialCode width instr maybe_revinstr a b
2145 = genTrivialCode (intSize width) instr a b
2147 -- This is re-used for floating pt instructions too.
2148 genTrivialCode rep instr a b = do
2149 (b_op, b_code) <- getNonClobberedOperand b
2150 a_code <- getAnyReg a
2151 tmp <- getNewRegNat rep
2153 -- We want the value of b to stay alive across the computation of a.
2154 -- But, we want to calculate a straight into the destination register,
2155 -- because the instruction only has two operands (dst := dst `op` src).
2156 -- The troublesome case is when the result of b is in the same register
2157 -- as the destination reg. In this case, we have to save b in a
2158 -- new temporary across the computation of a.
2160 | dst `regClashesWithOp` b_op =
2162 unitOL (MOV rep b_op (OpReg tmp)) `appOL`
2164 instr (OpReg tmp) (OpReg dst)
2168 instr b_op (OpReg dst)
2170 return (Any rep code)
2172 reg `regClashesWithOp` OpReg reg2 = reg == reg2
2173 reg `regClashesWithOp` OpAddr amode = any (==reg) (addrModeRegs amode)
2174 reg `regClashesWithOp` _ = False
2178 trivialUCode rep instr x = do
2179 x_code <- getAnyReg x
2184 return (Any rep code)
2188 trivialFCode_x87 width instr x y = do
2189 (x_reg, x_code) <- getNonClobberedReg x -- these work for float regs too
2190 (y_reg, y_code) <- getSomeReg y
2192 size = FF80 -- always, on x87
2196 instr size x_reg y_reg dst
2197 return (Any size code)
2199 trivialFCode_sse2 pk instr x y
2200 = genTrivialCode size (instr size) x y
2201 where size = floatSize pk
2204 trivialUFCode size instr x = do
2205 (x_reg, x_code) <- getSomeReg x
2211 return (Any size code)
2214 --------------------------------------------------------------------------------
2215 coerceInt2FP :: Width -> Width -> CmmExpr -> NatM Register
2216 coerceInt2FP from to x = if_sse2 coerce_sse2 coerce_x87
2219 (x_reg, x_code) <- getSomeReg x
2221 opc = case to of W32 -> GITOF; W64 -> GITOD
2222 code dst = x_code `snocOL` opc x_reg dst
2223 -- ToDo: works for non-II32 reps?
2224 return (Any FF80 code)
2227 (x_op, x_code) <- getOperand x -- ToDo: could be a safe operand
2229 opc = case to of W32 -> CVTSI2SS; W64 -> CVTSI2SD
2230 code dst = x_code `snocOL` opc (intSize from) x_op dst
2232 return (Any (floatSize to) code)
2233 -- works even if the destination rep is <II32
2235 --------------------------------------------------------------------------------
2236 coerceFP2Int :: Width -> Width -> CmmExpr -> NatM Register
2237 coerceFP2Int from to x = if_sse2 coerceFP2Int_sse2 coerceFP2Int_x87
2239 coerceFP2Int_x87 = do
2240 (x_reg, x_code) <- getSomeReg x
2242 opc = case from of W32 -> GFTOI; W64 -> GDTOI
2243 code dst = x_code `snocOL` opc x_reg dst
2244 -- ToDo: works for non-II32 reps?
2246 return (Any (intSize to) code)
2248 coerceFP2Int_sse2 = do
2249 (x_op, x_code) <- getOperand x -- ToDo: could be a safe operand
2251 opc = case from of W32 -> CVTTSS2SIQ; W64 -> CVTTSD2SIQ
2252 code dst = x_code `snocOL` opc (intSize to) x_op dst
2254 return (Any (intSize to) code)
2255 -- works even if the destination rep is <II32
2258 --------------------------------------------------------------------------------
2259 coerceFP2FP :: Width -> CmmExpr -> NatM Register
2260 coerceFP2FP to x = do
2261 (x_reg, x_code) <- getSomeReg x
2263 opc = case to of W32 -> CVTSD2SS; W64 -> CVTSS2SD
2264 code dst = x_code `snocOL` opc x_reg dst
2266 return (Any (floatSize to) code)
2268 --------------------------------------------------------------------------------
2270 sse2NegCode :: Width -> CmmExpr -> NatM Register
2271 sse2NegCode w x = do
2272 let sz = floatSize w
2273 x_code <- getAnyReg x
2274 -- This is how gcc does it, so it can't be that bad:
2276 const | FF32 <- sz = CmmInt 0x80000000 W32
2277 | otherwise = CmmInt 0x8000000000000000 W64
2278 Amode amode amode_code <- memConstant (widthInBytes w) const
2279 tmp <- getNewRegNat sz
2281 code dst = x_code dst `appOL` amode_code `appOL` toOL [
2282 MOV sz (OpAddr amode) (OpReg tmp),
2283 XOR sz (OpReg tmp) (OpReg dst)
2286 return (Any sz code)