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.
21 module MachCodeGen ( cmmTopCodeGen, InstrBlock ) where
23 #include "HsVersions.h"
24 #include "nativeGen/NCG.h"
31 import PositionIndependentCode
32 import RegAllocInfo ( mkBranchInstr )
34 -- Our intermediate code:
36 import PprCmm ( pprExpr )
39 import ClosureInfo ( C_SRT(..) )
42 import StaticFlags ( opt_PIC )
43 import ForeignCall ( CCallConv(..) )
46 import qualified Outputable as O
49 import FastBool ( isFastTrue )
50 import Constants ( wORD_SIZE )
52 import Debug.Trace ( trace )
54 import Control.Monad ( mapAndUnzipM )
55 import Data.Maybe ( fromJust )
60 -- -----------------------------------------------------------------------------
61 -- Top-level of the instruction selector
63 -- | 'InstrBlock's are the insn sequences generated by the insn selectors.
64 -- They are really trees of insns to facilitate fast appending, where a
65 -- left-to-right traversal (pre-order?) yields the insns in the correct
68 type InstrBlock = OrdList Instr
70 cmmTopCodeGen :: RawCmmTop -> NatM [NatCmmTop]
71 cmmTopCodeGen (CmmProc info lab params (ListGraph blocks)) = do
72 (nat_blocks,statics) <- mapAndUnzipM basicBlockCodeGen blocks
73 picBaseMb <- getPicBaseMaybeNat
74 let proc = CmmProc info lab params (ListGraph $ concat nat_blocks)
75 tops = proc : concat statics
77 Just picBase -> initializePicBase picBase tops
78 Nothing -> return tops
80 cmmTopCodeGen (CmmData sec dat) = do
81 return [CmmData sec dat] -- no translation, we just use CmmStatic
83 basicBlockCodeGen :: CmmBasicBlock -> NatM ([NatBasicBlock],[NatCmmTop])
84 basicBlockCodeGen (BasicBlock id stmts) = do
85 instrs <- stmtsToInstrs stmts
86 -- code generation may introduce new basic block boundaries, which
87 -- are indicated by the NEWBLOCK instruction. We must split up the
88 -- instruction stream into basic blocks again. Also, we extract
91 (top,other_blocks,statics) = foldrOL mkBlocks ([],[],[]) instrs
93 mkBlocks (NEWBLOCK id) (instrs,blocks,statics)
94 = ([], BasicBlock id instrs : blocks, statics)
95 mkBlocks (LDATA sec dat) (instrs,blocks,statics)
96 = (instrs, blocks, CmmData sec dat:statics)
97 mkBlocks instr (instrs,blocks,statics)
98 = (instr:instrs, blocks, statics)
100 return (BasicBlock id top : other_blocks, statics)
102 stmtsToInstrs :: [CmmStmt] -> NatM InstrBlock
104 = do instrss <- mapM stmtToInstrs stmts
105 return (concatOL instrss)
107 stmtToInstrs :: CmmStmt -> NatM InstrBlock
108 stmtToInstrs stmt = case stmt of
109 CmmNop -> return nilOL
110 CmmComment s -> return (unitOL (COMMENT s))
113 | isFloatType ty -> assignReg_FltCode size reg src
114 #if WORD_SIZE_IN_BITS==32
115 | isWord64 ty -> assignReg_I64Code reg src
117 | otherwise -> assignReg_IntCode size reg src
118 where ty = cmmRegType reg
119 size = cmmTypeSize ty
122 | isFloatType ty -> assignMem_FltCode size addr src
123 #if WORD_SIZE_IN_BITS==32
124 | isWord64 ty -> assignMem_I64Code addr src
126 | otherwise -> assignMem_IntCode size addr src
127 where ty = cmmExprType src
128 size = cmmTypeSize ty
130 CmmCall target result_regs args _ _
131 -> genCCall target result_regs args
133 CmmBranch id -> genBranch id
134 CmmCondBranch arg id -> genCondJump id arg
135 CmmSwitch arg ids -> genSwitch arg ids
136 CmmJump arg params -> genJump arg
138 panic "stmtToInstrs: return statement should have been cps'd away"
140 -- -----------------------------------------------------------------------------
141 -- General things for putting together code sequences
143 -- Expand CmmRegOff. ToDo: should we do it this way around, or convert
144 -- CmmExprs into CmmRegOff?
145 mangleIndexTree :: CmmExpr -> CmmExpr
146 mangleIndexTree (CmmRegOff reg off)
147 = CmmMachOp (MO_Add width) [CmmReg reg, CmmLit (CmmInt (fromIntegral off) width)]
148 where width = typeWidth (cmmRegType reg)
150 -- -----------------------------------------------------------------------------
151 -- Code gen for 64-bit arithmetic on 32-bit platforms
154 Simple support for generating 64-bit code (ie, 64 bit values and 64
155 bit assignments) on 32-bit platforms. Unlike the main code generator
156 we merely shoot for generating working code as simply as possible, and
157 pay little attention to code quality. Specifically, there is no
158 attempt to deal cleverly with the fixed-vs-floating register
159 distinction; all values are generated into (pairs of) floating
160 registers, even if this would mean some redundant reg-reg moves as a
161 result. Only one of the VRegUniques is returned, since it will be
162 of the VRegUniqueLo form, and the upper-half VReg can be determined
163 by applying getHiVRegFromLo to it.
166 data ChildCode64 -- a.k.a "Register64"
169 Reg -- the lower 32-bit temporary which contains the
170 -- result; use getHiVRegFromLo to find the other
171 -- VRegUnique. Rules of this simplified insn
172 -- selection game are therefore that the returned
173 -- Reg may be modified
175 #if WORD_SIZE_IN_BITS==32
176 assignMem_I64Code :: CmmExpr -> CmmExpr -> NatM InstrBlock
177 assignReg_I64Code :: CmmReg -> CmmExpr -> NatM InstrBlock
180 #ifndef x86_64_TARGET_ARCH
181 iselExpr64 :: CmmExpr -> NatM ChildCode64
184 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
188 assignMem_I64Code addrTree valueTree = do
189 Amode addr addr_code <- getAmode addrTree
190 ChildCode64 vcode rlo <- iselExpr64 valueTree
192 rhi = getHiVRegFromLo rlo
194 -- Little-endian store
195 mov_lo = MOV II32 (OpReg rlo) (OpAddr addr)
196 mov_hi = MOV II32 (OpReg rhi) (OpAddr (fromJust (addrOffset addr 4)))
198 return (vcode `appOL` addr_code `snocOL` mov_lo `snocOL` mov_hi)
201 assignReg_I64Code (CmmLocal (LocalReg u_dst pk)) valueTree = do
202 ChildCode64 vcode r_src_lo <- iselExpr64 valueTree
204 r_dst_lo = mkVReg u_dst II32
205 r_dst_hi = getHiVRegFromLo r_dst_lo
206 r_src_hi = getHiVRegFromLo r_src_lo
207 mov_lo = MOV II32 (OpReg r_src_lo) (OpReg r_dst_lo)
208 mov_hi = MOV II32 (OpReg r_src_hi) (OpReg r_dst_hi)
211 vcode `snocOL` mov_lo `snocOL` mov_hi
214 assignReg_I64Code lvalue valueTree
215 = panic "assignReg_I64Code(i386): invalid lvalue"
219 iselExpr64 (CmmLit (CmmInt i _)) = do
220 (rlo,rhi) <- getNewRegPairNat II32
222 r = fromIntegral (fromIntegral i :: Word32)
223 q = fromIntegral ((fromIntegral i `shiftR` 32) :: Word32)
225 MOV II32 (OpImm (ImmInteger r)) (OpReg rlo),
226 MOV II32 (OpImm (ImmInteger q)) (OpReg rhi)
229 return (ChildCode64 code rlo)
231 iselExpr64 (CmmLoad addrTree ty) | isWord64 ty = do
232 Amode addr addr_code <- getAmode addrTree
233 (rlo,rhi) <- getNewRegPairNat II32
235 mov_lo = MOV II32 (OpAddr addr) (OpReg rlo)
236 mov_hi = MOV II32 (OpAddr (fromJust (addrOffset addr 4))) (OpReg rhi)
239 ChildCode64 (addr_code `snocOL` mov_lo `snocOL` mov_hi)
243 iselExpr64 (CmmReg (CmmLocal (LocalReg vu ty))) | isWord64 ty
244 = return (ChildCode64 nilOL (mkVReg vu II32))
246 -- we handle addition, but rather badly
247 iselExpr64 (CmmMachOp (MO_Add _) [e1, CmmLit (CmmInt i _)]) = do
248 ChildCode64 code1 r1lo <- iselExpr64 e1
249 (rlo,rhi) <- getNewRegPairNat II32
251 r = fromIntegral (fromIntegral i :: Word32)
252 q = fromIntegral ((fromIntegral i `shiftR` 32) :: Word32)
253 r1hi = getHiVRegFromLo r1lo
255 toOL [ MOV II32 (OpReg r1lo) (OpReg rlo),
256 ADD II32 (OpImm (ImmInteger r)) (OpReg rlo),
257 MOV II32 (OpReg r1hi) (OpReg rhi),
258 ADC II32 (OpImm (ImmInteger q)) (OpReg rhi) ]
260 return (ChildCode64 code rlo)
262 iselExpr64 (CmmMachOp (MO_Add _) [e1,e2]) = do
263 ChildCode64 code1 r1lo <- iselExpr64 e1
264 ChildCode64 code2 r2lo <- iselExpr64 e2
265 (rlo,rhi) <- getNewRegPairNat II32
267 r1hi = getHiVRegFromLo r1lo
268 r2hi = getHiVRegFromLo r2lo
271 toOL [ MOV II32 (OpReg r1lo) (OpReg rlo),
272 ADD II32 (OpReg r2lo) (OpReg rlo),
273 MOV II32 (OpReg r1hi) (OpReg rhi),
274 ADC II32 (OpReg r2hi) (OpReg rhi) ]
276 return (ChildCode64 code rlo)
278 iselExpr64 (CmmMachOp (MO_UU_Conv _ W64) [expr]) = do
280 r_dst_lo <- getNewRegNat II32
281 let r_dst_hi = getHiVRegFromLo r_dst_lo
284 ChildCode64 (code `snocOL`
285 MOV II32 (OpImm (ImmInt 0)) (OpReg r_dst_hi))
290 = pprPanic "iselExpr64(i386)" (ppr expr)
292 #endif /* i386_TARGET_ARCH */
294 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
296 #if sparc_TARGET_ARCH
298 assignMem_I64Code addrTree valueTree = do
299 Amode addr addr_code <- getAmode addrTree
300 ChildCode64 vcode rlo <- iselExpr64 valueTree
301 (src, code) <- getSomeReg addrTree
303 rhi = getHiVRegFromLo rlo
305 mov_hi = ST II32 rhi (AddrRegImm src (ImmInt 0))
306 mov_lo = ST II32 rlo (AddrRegImm src (ImmInt 4))
307 return (vcode `appOL` code `snocOL` mov_hi `snocOL` mov_lo)
309 assignReg_I64Code (CmmLocal (LocalReg u_dst pk _)) valueTree = do
310 ChildCode64 vcode r_src_lo <- iselExpr64 valueTree
312 r_dst_lo = mkVReg u_dst pk
313 r_dst_hi = getHiVRegFromLo r_dst_lo
314 r_src_hi = getHiVRegFromLo r_src_lo
315 mov_lo = mkMOV r_src_lo r_dst_lo
316 mov_hi = mkMOV r_src_hi r_dst_hi
317 mkMOV sreg dreg = OR False g0 (RIReg sreg) dreg
318 return (vcode `snocOL` mov_hi `snocOL` mov_lo)
319 assignReg_I64Code lvalue valueTree
320 = panic "assignReg_I64Code(sparc): invalid lvalue"
323 -- Don't delete this -- it's very handy for debugging.
325 -- | trace ("iselExpr64: " ++ showSDoc (ppr expr)) False
326 -- = panic "iselExpr64(???)"
328 iselExpr64 (CmmLoad addrTree ty) | isWord64 ty = do
329 Amode (AddrRegReg r1 r2) addr_code <- getAmode addrTree
330 rlo <- getNewRegNat II32
331 let rhi = getHiVRegFromLo rlo
332 mov_hi = LD II32 (AddrRegImm r1 (ImmInt 0)) rhi
333 mov_lo = LD II32 (AddrRegImm r1 (ImmInt 4)) rlo
335 ChildCode64 (addr_code `snocOL` mov_hi `snocOL` mov_lo)
339 iselExpr64 (CmmReg (CmmLocal (LocalReg uq ty))) isWord64 ty = do
340 r_dst_lo <- getNewRegNat b32
341 let r_dst_hi = getHiVRegFromLo r_dst_lo
342 r_src_lo = mkVReg uq b32
343 r_src_hi = getHiVRegFromLo r_src_lo
344 mov_lo = mkMOV r_src_lo r_dst_lo
345 mov_hi = mkMOV r_src_hi r_dst_hi
346 mkMOV sreg dreg = OR False g0 (RIReg sreg) dreg
348 ChildCode64 (toOL [mov_hi, mov_lo]) r_dst_lo
352 = pprPanic "iselExpr64(sparc)" (ppr expr)
354 #endif /* sparc_TARGET_ARCH */
356 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
358 #if powerpc_TARGET_ARCH
360 getI64Amodes :: CmmExpr -> NatM (AddrMode, AddrMode, InstrBlock)
361 getI64Amodes addrTree = do
362 Amode hi_addr addr_code <- getAmode addrTree
363 case addrOffset hi_addr 4 of
364 Just lo_addr -> return (hi_addr, lo_addr, addr_code)
365 Nothing -> do (hi_ptr, code) <- getSomeReg addrTree
366 return (AddrRegImm hi_ptr (ImmInt 0),
367 AddrRegImm hi_ptr (ImmInt 4),
370 assignMem_I64Code addrTree valueTree = do
371 (hi_addr, lo_addr, addr_code) <- getI64Amodes addrTree
372 ChildCode64 vcode rlo <- iselExpr64 valueTree
374 rhi = getHiVRegFromLo rlo
377 mov_hi = ST II32 rhi hi_addr
378 mov_lo = ST II32 rlo lo_addr
380 return (vcode `appOL` addr_code `snocOL` mov_lo `snocOL` mov_hi)
382 assignReg_I64Code (CmmLocal (LocalReg u_dst pk)) valueTree = do
383 ChildCode64 vcode r_src_lo <- iselExpr64 valueTree
385 r_dst_lo = mkVReg u_dst II32
386 r_dst_hi = getHiVRegFromLo r_dst_lo
387 r_src_hi = getHiVRegFromLo r_src_lo
388 mov_lo = MR r_dst_lo r_src_lo
389 mov_hi = MR r_dst_hi r_src_hi
392 vcode `snocOL` mov_lo `snocOL` mov_hi
395 assignReg_I64Code lvalue valueTree
396 = panic "assignReg_I64Code(powerpc): invalid lvalue"
399 -- Don't delete this -- it's very handy for debugging.
401 -- | trace ("iselExpr64: " ++ showSDoc (pprCmmExpr expr)) False
402 -- = panic "iselExpr64(???)"
404 iselExpr64 (CmmLoad addrTree ty) | isWord64 ty = do
405 (hi_addr, lo_addr, addr_code) <- getI64Amodes addrTree
406 (rlo, rhi) <- getNewRegPairNat II32
407 let mov_hi = LD II32 rhi hi_addr
408 mov_lo = LD II32 rlo lo_addr
409 return $ ChildCode64 (addr_code `snocOL` mov_lo `snocOL` mov_hi)
412 iselExpr64 (CmmReg (CmmLocal (LocalReg vu ty))) | isWord64 ty
413 = return (ChildCode64 nilOL (mkVReg vu II32))
415 iselExpr64 (CmmLit (CmmInt i _)) = do
416 (rlo,rhi) <- getNewRegPairNat II32
418 half0 = fromIntegral (fromIntegral i :: Word16)
419 half1 = fromIntegral ((fromIntegral i `shiftR` 16) :: Word16)
420 half2 = fromIntegral ((fromIntegral i `shiftR` 32) :: Word16)
421 half3 = fromIntegral ((fromIntegral i `shiftR` 48) :: Word16)
424 LIS rlo (ImmInt half1),
425 OR rlo rlo (RIImm $ ImmInt half0),
426 LIS rhi (ImmInt half3),
427 OR rlo rlo (RIImm $ ImmInt half2)
430 return (ChildCode64 code rlo)
432 iselExpr64 (CmmMachOp (MO_Add _) [e1,e2]) = do
433 ChildCode64 code1 r1lo <- iselExpr64 e1
434 ChildCode64 code2 r2lo <- iselExpr64 e2
435 (rlo,rhi) <- getNewRegPairNat II32
437 r1hi = getHiVRegFromLo r1lo
438 r2hi = getHiVRegFromLo r2lo
441 toOL [ ADDC rlo r1lo r2lo,
444 return (ChildCode64 code rlo)
446 iselExpr64 (CmmMachOp (MO_UU_Conv W32 W64) [expr]) = do
447 (expr_reg,expr_code) <- getSomeReg expr
448 (rlo, rhi) <- getNewRegPairNat II32
449 let mov_hi = LI rhi (ImmInt 0)
450 mov_lo = MR rlo expr_reg
451 return $ ChildCode64 (expr_code `snocOL` mov_lo `snocOL` mov_hi)
454 = pprPanic "iselExpr64(powerpc)" (ppr expr)
456 #endif /* powerpc_TARGET_ARCH */
459 -- -----------------------------------------------------------------------------
460 -- The 'Register' type
462 -- 'Register's passed up the tree. If the stix code forces the register
463 -- to live in a pre-decided machine register, it comes out as @Fixed@;
464 -- otherwise, it comes out as @Any@, and the parent can decide which
465 -- register to put it in.
468 = Fixed Size Reg InstrBlock
469 | Any Size (Reg -> InstrBlock)
471 swizzleRegisterRep :: Register -> Size -> Register
472 -- Change the width; it's a no-op
473 swizzleRegisterRep (Fixed _ reg code) size = Fixed size reg code
474 swizzleRegisterRep (Any _ codefn) size = Any size codefn
477 -- -----------------------------------------------------------------------------
478 -- Utils based on getRegister, below
480 -- The dual to getAnyReg: compute an expression into a register, but
481 -- we don't mind which one it is.
482 getSomeReg :: CmmExpr -> NatM (Reg, InstrBlock)
484 r <- getRegister expr
487 tmp <- getNewRegNat rep
488 return (tmp, code tmp)
492 -- -----------------------------------------------------------------------------
493 -- Grab the Reg for a CmmReg
495 getRegisterReg :: CmmReg -> Reg
497 getRegisterReg (CmmLocal (LocalReg u pk))
498 = mkVReg u (cmmTypeSize pk)
500 getRegisterReg (CmmGlobal mid)
501 = case get_GlobalReg_reg_or_addr mid of
502 Left (RealReg rrno) -> RealReg rrno
503 _other -> pprPanic "getRegisterReg-memory" (ppr $ CmmGlobal mid)
504 -- By this stage, the only MagicIds remaining should be the
505 -- ones which map to a real machine register on this
506 -- platform. Hence ...
509 -- -----------------------------------------------------------------------------
510 -- Generate code to get a subtree into a Register
512 -- Don't delete this -- it's very handy for debugging.
514 -- | trace ("getRegister: " ++ showSDoc (pprCmmExpr expr)) False
515 -- = panic "getRegister(???)"
517 getRegister :: CmmExpr -> NatM Register
519 #if !x86_64_TARGET_ARCH
520 -- on x86_64, we have %rip for PicBaseReg, but it's not a full-featured
521 -- register, it can only be used for rip-relative addressing.
522 getRegister (CmmReg (CmmGlobal PicBaseReg))
524 reg <- getPicBaseNat wordSize
525 return (Fixed wordSize reg nilOL)
528 getRegister (CmmReg reg)
529 = return (Fixed (cmmTypeSize (cmmRegType reg))
530 (getRegisterReg reg) nilOL)
532 getRegister tree@(CmmRegOff _ _)
533 = getRegister (mangleIndexTree tree)
536 #if WORD_SIZE_IN_BITS==32
537 -- for 32-bit architectuers, support some 64 -> 32 bit conversions:
538 -- TO_W_(x), TO_W_(x >> 32)
540 getRegister (CmmMachOp (MO_UU_Conv W64 W32)
541 [CmmMachOp (MO_U_Shr W64) [x,CmmLit (CmmInt 32 _)]]) = do
542 ChildCode64 code rlo <- iselExpr64 x
543 return $ Fixed II32 (getHiVRegFromLo rlo) code
545 getRegister (CmmMachOp (MO_SS_Conv W64 W32)
546 [CmmMachOp (MO_U_Shr W64) [x,CmmLit (CmmInt 32 _)]]) = do
547 ChildCode64 code rlo <- iselExpr64 x
548 return $ Fixed II32 (getHiVRegFromLo rlo) code
550 getRegister (CmmMachOp (MO_UU_Conv W64 W32) [x]) = do
551 ChildCode64 code rlo <- iselExpr64 x
552 return $ Fixed II32 rlo code
554 getRegister (CmmMachOp (MO_SS_Conv W64 W32) [x]) = do
555 ChildCode64 code rlo <- iselExpr64 x
556 return $ Fixed II32 rlo code
560 -- end of machine-"independent" bit; here we go on the rest...
562 #if alpha_TARGET_ARCH
564 getRegister (StDouble d)
565 = getBlockIdNat `thenNat` \ lbl ->
566 getNewRegNat PtrRep `thenNat` \ tmp ->
567 let code dst = mkSeqInstrs [
568 LDATA RoDataSegment lbl [
569 DATA TF [ImmLab (rational d)]
571 LDA tmp (AddrImm (ImmCLbl lbl)),
572 LD TF dst (AddrReg tmp)]
574 return (Any FF64 code)
576 getRegister (StPrim primop [x]) -- unary PrimOps
578 IntNegOp -> trivialUCode (NEG Q False) x
580 NotOp -> trivialUCode NOT x
582 FloatNegOp -> trivialUFCode FloatRep (FNEG TF) x
583 DoubleNegOp -> trivialUFCode FF64 (FNEG TF) x
585 OrdOp -> coerceIntCode IntRep x
588 Float2IntOp -> coerceFP2Int x
589 Int2FloatOp -> coerceInt2FP pr x
590 Double2IntOp -> coerceFP2Int x
591 Int2DoubleOp -> coerceInt2FP pr x
593 Double2FloatOp -> coerceFltCode x
594 Float2DoubleOp -> coerceFltCode x
596 other_op -> getRegister (StCall fn CCallConv FF64 [x])
598 fn = case other_op of
599 FloatExpOp -> fsLit "exp"
600 FloatLogOp -> fsLit "log"
601 FloatSqrtOp -> fsLit "sqrt"
602 FloatSinOp -> fsLit "sin"
603 FloatCosOp -> fsLit "cos"
604 FloatTanOp -> fsLit "tan"
605 FloatAsinOp -> fsLit "asin"
606 FloatAcosOp -> fsLit "acos"
607 FloatAtanOp -> fsLit "atan"
608 FloatSinhOp -> fsLit "sinh"
609 FloatCoshOp -> fsLit "cosh"
610 FloatTanhOp -> fsLit "tanh"
611 DoubleExpOp -> fsLit "exp"
612 DoubleLogOp -> fsLit "log"
613 DoubleSqrtOp -> fsLit "sqrt"
614 DoubleSinOp -> fsLit "sin"
615 DoubleCosOp -> fsLit "cos"
616 DoubleTanOp -> fsLit "tan"
617 DoubleAsinOp -> fsLit "asin"
618 DoubleAcosOp -> fsLit "acos"
619 DoubleAtanOp -> fsLit "atan"
620 DoubleSinhOp -> fsLit "sinh"
621 DoubleCoshOp -> fsLit "cosh"
622 DoubleTanhOp -> fsLit "tanh"
624 pr = panic "MachCode.getRegister: no primrep needed for Alpha"
626 getRegister (StPrim primop [x, y]) -- dyadic PrimOps
628 CharGtOp -> trivialCode (CMP LTT) y x
629 CharGeOp -> trivialCode (CMP LE) y x
630 CharEqOp -> trivialCode (CMP EQQ) x y
631 CharNeOp -> int_NE_code x y
632 CharLtOp -> trivialCode (CMP LTT) x y
633 CharLeOp -> trivialCode (CMP LE) x y
635 IntGtOp -> trivialCode (CMP LTT) y x
636 IntGeOp -> trivialCode (CMP LE) y x
637 IntEqOp -> trivialCode (CMP EQQ) x y
638 IntNeOp -> int_NE_code x y
639 IntLtOp -> trivialCode (CMP LTT) x y
640 IntLeOp -> trivialCode (CMP LE) x y
642 WordGtOp -> trivialCode (CMP ULT) y x
643 WordGeOp -> trivialCode (CMP ULE) x y
644 WordEqOp -> trivialCode (CMP EQQ) x y
645 WordNeOp -> int_NE_code x y
646 WordLtOp -> trivialCode (CMP ULT) x y
647 WordLeOp -> trivialCode (CMP ULE) x y
649 AddrGtOp -> trivialCode (CMP ULT) y x
650 AddrGeOp -> trivialCode (CMP ULE) y x
651 AddrEqOp -> trivialCode (CMP EQQ) x y
652 AddrNeOp -> int_NE_code x y
653 AddrLtOp -> trivialCode (CMP ULT) x y
654 AddrLeOp -> trivialCode (CMP ULE) x y
656 FloatGtOp -> cmpF_code (FCMP TF LE) EQQ x y
657 FloatGeOp -> cmpF_code (FCMP TF LTT) EQQ x y
658 FloatEqOp -> cmpF_code (FCMP TF EQQ) NE x y
659 FloatNeOp -> cmpF_code (FCMP TF EQQ) EQQ x y
660 FloatLtOp -> cmpF_code (FCMP TF LTT) NE x y
661 FloatLeOp -> cmpF_code (FCMP TF LE) NE x y
663 DoubleGtOp -> cmpF_code (FCMP TF LE) EQQ x y
664 DoubleGeOp -> cmpF_code (FCMP TF LTT) EQQ x y
665 DoubleEqOp -> cmpF_code (FCMP TF EQQ) NE x y
666 DoubleNeOp -> cmpF_code (FCMP TF EQQ) EQQ x y
667 DoubleLtOp -> cmpF_code (FCMP TF LTT) NE x y
668 DoubleLeOp -> cmpF_code (FCMP TF LE) NE x y
670 IntAddOp -> trivialCode (ADD Q False) x y
671 IntSubOp -> trivialCode (SUB Q False) x y
672 IntMulOp -> trivialCode (MUL Q False) x y
673 IntQuotOp -> trivialCode (DIV Q False) x y
674 IntRemOp -> trivialCode (REM Q False) x y
676 WordAddOp -> trivialCode (ADD Q False) x y
677 WordSubOp -> trivialCode (SUB Q False) x y
678 WordMulOp -> trivialCode (MUL Q False) x y
679 WordQuotOp -> trivialCode (DIV Q True) x y
680 WordRemOp -> trivialCode (REM Q True) x y
682 FloatAddOp -> trivialFCode W32 (FADD TF) x y
683 FloatSubOp -> trivialFCode W32 (FSUB TF) x y
684 FloatMulOp -> trivialFCode W32 (FMUL TF) x y
685 FloatDivOp -> trivialFCode W32 (FDIV TF) x y
687 DoubleAddOp -> trivialFCode W64 (FADD TF) x y
688 DoubleSubOp -> trivialFCode W64 (FSUB TF) x y
689 DoubleMulOp -> trivialFCode W64 (FMUL TF) x y
690 DoubleDivOp -> trivialFCode W64 (FDIV TF) x y
692 AddrAddOp -> trivialCode (ADD Q False) x y
693 AddrSubOp -> trivialCode (SUB Q False) x y
694 AddrRemOp -> trivialCode (REM Q True) x y
696 AndOp -> trivialCode AND x y
697 OrOp -> trivialCode OR x y
698 XorOp -> trivialCode XOR x y
699 SllOp -> trivialCode SLL x y
700 SrlOp -> trivialCode SRL x y
702 ISllOp -> trivialCode SLL x y -- was: panic "AlphaGen:isll"
703 ISraOp -> trivialCode SRA x y -- was: panic "AlphaGen:isra"
704 ISrlOp -> trivialCode SRL x y -- was: panic "AlphaGen:isrl"
706 FloatPowerOp -> getRegister (StCall (fsLit "pow") CCallConv FF64 [x,y])
707 DoublePowerOp -> getRegister (StCall (fsLit "pow") CCallConv FF64 [x,y])
709 {- ------------------------------------------------------------
710 Some bizarre special code for getting condition codes into
711 registers. Integer non-equality is a test for equality
712 followed by an XOR with 1. (Integer comparisons always set
713 the result register to 0 or 1.) Floating point comparisons of
714 any kind leave the result in a floating point register, so we
715 need to wrangle an integer register out of things.
717 int_NE_code :: StixTree -> StixTree -> NatM Register
720 = trivialCode (CMP EQQ) x y `thenNat` \ register ->
721 getNewRegNat IntRep `thenNat` \ tmp ->
723 code = registerCode register tmp
724 src = registerName register tmp
725 code__2 dst = code . mkSeqInstr (XOR src (RIImm (ImmInt 1)) dst)
727 return (Any IntRep code__2)
729 {- ------------------------------------------------------------
730 Comments for int_NE_code also apply to cmpF_code
733 :: (Reg -> Reg -> Reg -> Instr)
735 -> StixTree -> StixTree
738 cmpF_code instr cond x y
739 = trivialFCode pr instr x y `thenNat` \ register ->
740 getNewRegNat FF64 `thenNat` \ tmp ->
741 getBlockIdNat `thenNat` \ lbl ->
743 code = registerCode register tmp
744 result = registerName register tmp
746 code__2 dst = code . mkSeqInstrs [
747 OR zeroh (RIImm (ImmInt 1)) dst,
748 BF cond result (ImmCLbl lbl),
749 OR zeroh (RIReg zeroh) dst,
752 return (Any IntRep code__2)
754 pr = panic "trivialU?FCode: does not use PrimRep on Alpha"
755 ------------------------------------------------------------
757 getRegister (CmmLoad pk mem)
758 = getAmode mem `thenNat` \ amode ->
760 code = amodeCode amode
761 src = amodeAddr amode
762 size = primRepToSize pk
763 code__2 dst = code . mkSeqInstr (LD size dst src)
765 return (Any pk code__2)
767 getRegister (StInt i)
770 code dst = mkSeqInstr (OR zeroh (RIImm src) dst)
772 return (Any IntRep code)
775 code dst = mkSeqInstr (LDI Q dst src)
777 return (Any IntRep code)
779 src = ImmInt (fromInteger i)
784 code dst = mkSeqInstr (LDA dst (AddrImm imm__2))
786 return (Any PtrRep code)
789 imm__2 = case imm of Just x -> x
791 #endif /* alpha_TARGET_ARCH */
793 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
797 getRegister (CmmLit (CmmFloat f W32)) = do
798 lbl <- getNewLabelNat
799 dflags <- getDynFlagsNat
800 dynRef <- cmmMakeDynamicReference dflags addImportNat DataReference lbl
801 Amode addr addr_code <- getAmode dynRef
805 CmmStaticLit (CmmFloat f W32)]
806 `consOL` (addr_code `snocOL`
809 return (Any FF32 code)
812 getRegister (CmmLit (CmmFloat d W64))
814 = let code dst = unitOL (GLDZ dst)
815 in return (Any FF64 code)
818 = let code dst = unitOL (GLD1 dst)
819 in return (Any FF64 code)
822 lbl <- getNewLabelNat
823 dflags <- getDynFlagsNat
824 dynRef <- cmmMakeDynamicReference dflags addImportNat DataReference lbl
825 Amode addr addr_code <- getAmode dynRef
829 CmmStaticLit (CmmFloat d W64)]
830 `consOL` (addr_code `snocOL`
833 return (Any FF64 code)
835 #endif /* i386_TARGET_ARCH */
837 #if x86_64_TARGET_ARCH
839 getRegister (CmmLit (CmmFloat 0.0 w)) = do
840 let size = floatSize w
841 code dst = unitOL (XOR size (OpReg dst) (OpReg dst))
842 -- I don't know why there are xorpd, xorps, and pxor instructions.
843 -- They all appear to do the same thing --SDM
844 return (Any size code)
846 getRegister (CmmLit (CmmFloat f w)) = do
847 lbl <- getNewLabelNat
848 let code dst = toOL [
851 CmmStaticLit (CmmFloat f w)],
852 MOV size (OpAddr (ripRel (ImmCLbl lbl))) (OpReg dst)
855 return (Any size code)
856 where size = floatSize w
858 #endif /* x86_64_TARGET_ARCH */
860 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
862 -- catch simple cases of zero- or sign-extended load
863 getRegister (CmmMachOp (MO_UU_Conv W8 W32) [CmmLoad addr _]) = do
864 code <- intLoadCode (MOVZxL II8) addr
865 return (Any II32 code)
867 getRegister (CmmMachOp (MO_SS_Conv W8 W32) [CmmLoad addr _]) = do
868 code <- intLoadCode (MOVSxL II8) addr
869 return (Any II32 code)
871 getRegister (CmmMachOp (MO_UU_Conv W16 W32) [CmmLoad addr _]) = do
872 code <- intLoadCode (MOVZxL II16) addr
873 return (Any II32 code)
875 getRegister (CmmMachOp (MO_SS_Conv W16 W32) [CmmLoad addr _]) = do
876 code <- intLoadCode (MOVSxL II16) addr
877 return (Any II32 code)
881 #if x86_64_TARGET_ARCH
883 -- catch simple cases of zero- or sign-extended load
884 getRegister (CmmMachOp (MO_UU_Conv W8 W64) [CmmLoad addr _]) = do
885 code <- intLoadCode (MOVZxL II8) addr
886 return (Any II64 code)
888 getRegister (CmmMachOp (MO_SS_Conv W8 W64) [CmmLoad addr _]) = do
889 code <- intLoadCode (MOVSxL II8) addr
890 return (Any II64 code)
892 getRegister (CmmMachOp (MO_UU_Conv W16 W64) [CmmLoad addr _]) = do
893 code <- intLoadCode (MOVZxL II16) addr
894 return (Any II64 code)
896 getRegister (CmmMachOp (MO_SS_Conv W16 W64) [CmmLoad addr _]) = do
897 code <- intLoadCode (MOVSxL II16) addr
898 return (Any II64 code)
900 getRegister (CmmMachOp (MO_UU_Conv W32 W64) [CmmLoad addr _]) = do
901 code <- intLoadCode (MOV II32) addr -- 32-bit loads zero-extend
902 return (Any II64 code)
904 getRegister (CmmMachOp (MO_SS_Conv W32 W64) [CmmLoad addr _]) = do
905 code <- intLoadCode (MOVSxL II32) addr
906 return (Any II64 code)
910 #if x86_64_TARGET_ARCH
911 getRegister (CmmMachOp (MO_Add W64) [CmmReg (CmmGlobal PicBaseReg),
912 CmmLit displacement])
913 = return $ Any II64 (\dst -> unitOL $
914 LEA II64 (OpAddr (ripRel (litToImm displacement))) (OpReg dst))
917 #if x86_64_TARGET_ARCH
918 getRegister (CmmMachOp (MO_F_Neg W32) [x]) = do
919 x_code <- getAnyReg x
920 lbl <- getNewLabelNat
922 code dst = x_code dst `appOL` toOL [
923 -- This is how gcc does it, so it can't be that bad:
924 LDATA ReadOnlyData16 [
927 CmmStaticLit (CmmInt 0x80000000 W32),
928 CmmStaticLit (CmmInt 0 W32),
929 CmmStaticLit (CmmInt 0 W32),
930 CmmStaticLit (CmmInt 0 W32)
932 XOR FF32 (OpAddr (ripRel (ImmCLbl lbl))) (OpReg dst)
933 -- xorps, so we need the 128-bit constant
934 -- ToDo: rip-relative
937 return (Any FF32 code)
939 getRegister (CmmMachOp (MO_F_Neg W64) [x]) = do
940 x_code <- getAnyReg x
941 lbl <- getNewLabelNat
943 -- This is how gcc does it, so it can't be that bad:
944 code dst = x_code dst `appOL` toOL [
945 LDATA ReadOnlyData16 [
948 CmmStaticLit (CmmInt 0x8000000000000000 W64),
949 CmmStaticLit (CmmInt 0 W64)
951 -- gcc puts an unpck here. Wonder if we need it.
952 XOR FF64 (OpAddr (ripRel (ImmCLbl lbl))) (OpReg dst)
953 -- xorpd, so we need the 128-bit constant
956 return (Any FF64 code)
959 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
961 getRegister (CmmMachOp mop [x]) -- unary MachOps
964 MO_F_Neg W32 -> trivialUFCode FF32 (GNEG FF32) x
965 MO_F_Neg W64 -> trivialUFCode FF64 (GNEG FF64) x
968 MO_S_Neg w -> triv_ucode NEGI (intSize w)
969 MO_F_Neg w -> triv_ucode NEGI (floatSize w)
970 MO_Not w -> triv_ucode NOT (intSize w)
973 MO_UU_Conv W32 W8 -> toI8Reg W32 x
974 MO_SS_Conv W32 W8 -> toI8Reg W32 x
975 MO_UU_Conv W16 W8 -> toI8Reg W16 x
976 MO_SS_Conv W16 W8 -> toI8Reg W16 x
977 MO_UU_Conv W32 W16 -> toI16Reg W32 x
978 MO_SS_Conv W32 W16 -> toI16Reg W32 x
980 #if x86_64_TARGET_ARCH
981 MO_UU_Conv W64 W32 -> conversionNop II64 x
982 MO_SS_Conv W64 W32 -> conversionNop II64 x
983 MO_UU_Conv W64 W16 -> toI16Reg W64 x
984 MO_SS_Conv W64 W16 -> toI16Reg W64 x
985 MO_UU_Conv W64 W8 -> toI8Reg W64 x
986 MO_SS_Conv W64 W8 -> toI8Reg W64 x
989 MO_UU_Conv rep1 rep2 | rep1 == rep2 -> conversionNop (intSize rep1) x
990 MO_SS_Conv rep1 rep2 | rep1 == rep2 -> conversionNop (intSize rep1) x
993 MO_UU_Conv W8 W32 -> integerExtend W8 W32 MOVZxL x
994 MO_UU_Conv W16 W32 -> integerExtend W16 W32 MOVZxL x
995 MO_UU_Conv W8 W16 -> integerExtend W8 W16 MOVZxL x
997 MO_SS_Conv W8 W32 -> integerExtend W8 W32 MOVSxL x
998 MO_SS_Conv W16 W32 -> integerExtend W16 W32 MOVSxL x
999 MO_SS_Conv W8 W16 -> integerExtend W8 W16 MOVSxL x
1001 #if x86_64_TARGET_ARCH
1002 MO_UU_Conv W8 W64 -> integerExtend W8 W64 MOVZxL x
1003 MO_UU_Conv W16 W64 -> integerExtend W16 W64 MOVZxL x
1004 MO_UU_Conv W32 W64 -> integerExtend W32 W64 MOVZxL x
1005 MO_SS_Conv W8 W64 -> integerExtend W8 W64 MOVSxL x
1006 MO_SS_Conv W16 W64 -> integerExtend W16 W64 MOVSxL x
1007 MO_SS_Conv W32 W64 -> integerExtend W32 W64 MOVSxL x
1008 -- for 32-to-64 bit zero extension, amd64 uses an ordinary movl.
1009 -- However, we don't want the register allocator to throw it
1010 -- away as an unnecessary reg-to-reg move, so we keep it in
1011 -- the form of a movzl and print it as a movl later.
1014 #if i386_TARGET_ARCH
1015 MO_FF_Conv W32 W64 -> conversionNop FF64 x
1016 MO_FF_Conv W64 W32 -> conversionNop FF32 x
1018 MO_FF_Conv W32 W64 -> coerceFP2FP W64 x
1019 MO_FF_Conv W64 W32 -> coerceFP2FP W32 x
1022 MO_FS_Conv from to -> coerceFP2Int from to x
1023 MO_SF_Conv from to -> coerceInt2FP from to x
1025 other -> pprPanic "getRegister" (pprMachOp mop)
1027 triv_ucode :: (Size -> Operand -> Instr) -> Size -> NatM Register
1028 triv_ucode instr size = trivialUCode size (instr size) x
1030 -- signed or unsigned extension.
1031 integerExtend :: Width -> Width
1032 -> (Size -> Operand -> Operand -> Instr)
1033 -> CmmExpr -> NatM Register
1034 integerExtend from to instr expr = do
1035 (reg,e_code) <- if from == W8 then getByteReg expr
1036 else getSomeReg expr
1040 instr (intSize from) (OpReg reg) (OpReg dst)
1041 return (Any (intSize to) code)
1043 toI8Reg :: Width -> CmmExpr -> NatM Register
1044 toI8Reg new_rep expr
1045 = do codefn <- getAnyReg expr
1046 return (Any (intSize new_rep) codefn)
1047 -- HACK: use getAnyReg to get a byte-addressable register.
1048 -- If the source was a Fixed register, this will add the
1049 -- mov instruction to put it into the desired destination.
1050 -- We're assuming that the destination won't be a fixed
1051 -- non-byte-addressable register; it won't be, because all
1052 -- fixed registers are word-sized.
1054 toI16Reg = toI8Reg -- for now
1056 conversionNop :: Size -> CmmExpr -> NatM Register
1057 conversionNop new_size expr
1058 = do e_code <- getRegister expr
1059 return (swizzleRegisterRep e_code new_size)
1062 getRegister e@(CmmMachOp mop [x, y]) -- dyadic MachOps
1064 MO_F_Eq w -> condFltReg EQQ x y
1065 MO_F_Ne w -> condFltReg NE x y
1066 MO_F_Gt w -> condFltReg GTT x y
1067 MO_F_Ge w -> condFltReg GE x y
1068 MO_F_Lt w -> condFltReg LTT x y
1069 MO_F_Le w -> condFltReg LE x y
1071 MO_Eq rep -> condIntReg EQQ x y
1072 MO_Ne rep -> condIntReg NE x y
1074 MO_S_Gt rep -> condIntReg GTT x y
1075 MO_S_Ge rep -> condIntReg GE x y
1076 MO_S_Lt rep -> condIntReg LTT x y
1077 MO_S_Le rep -> condIntReg LE x y
1079 MO_U_Gt rep -> condIntReg GU x y
1080 MO_U_Ge rep -> condIntReg GEU x y
1081 MO_U_Lt rep -> condIntReg LU x y
1082 MO_U_Le rep -> condIntReg LEU x y
1084 #if i386_TARGET_ARCH
1085 MO_F_Add w -> trivialFCode w GADD x y
1086 MO_F_Sub w -> trivialFCode w GSUB x y
1087 MO_F_Quot w -> trivialFCode w GDIV x y
1088 MO_F_Mul w -> trivialFCode w GMUL x y
1091 #if x86_64_TARGET_ARCH
1092 MO_F_Add w -> trivialFCode w ADD x y
1093 MO_F_Sub w -> trivialFCode w SUB x y
1094 MO_F_Quot w -> trivialFCode w FDIV x y
1095 MO_F_Mul w -> trivialFCode w MUL x y
1098 MO_Add rep -> add_code rep x y
1099 MO_Sub rep -> sub_code rep x y
1101 MO_S_Quot rep -> div_code rep True True x y
1102 MO_S_Rem rep -> div_code rep True False x y
1103 MO_U_Quot rep -> div_code rep False True x y
1104 MO_U_Rem rep -> div_code rep False False x y
1106 MO_S_MulMayOflo rep -> imulMayOflo rep x y
1108 MO_Mul rep -> triv_op rep IMUL
1109 MO_And rep -> triv_op rep AND
1110 MO_Or rep -> triv_op rep OR
1111 MO_Xor rep -> triv_op rep XOR
1113 {- Shift ops on x86s have constraints on their source, it
1114 either has to be Imm, CL or 1
1115 => trivialCode is not restrictive enough (sigh.)
1117 MO_Shl rep -> shift_code rep SHL x y {-False-}
1118 MO_U_Shr rep -> shift_code rep SHR x y {-False-}
1119 MO_S_Shr rep -> shift_code rep SAR x y {-False-}
1121 other -> pprPanic "getRegister(x86) - binary CmmMachOp (1)" (pprMachOp mop)
1123 --------------------
1124 triv_op width instr = trivialCode width op (Just op) x y
1125 where op = instr (intSize width)
1127 imulMayOflo :: Width -> CmmExpr -> CmmExpr -> NatM Register
1128 imulMayOflo rep a b = do
1129 (a_reg, a_code) <- getNonClobberedReg a
1130 b_code <- getAnyReg b
1132 shift_amt = case rep of
1135 _ -> panic "shift_amt"
1138 code = a_code `appOL` b_code eax `appOL`
1140 IMUL2 size (OpReg a_reg), -- result in %edx:%eax
1141 SAR size (OpImm (ImmInt shift_amt)) (OpReg eax),
1142 -- sign extend lower part
1143 SUB size (OpReg edx) (OpReg eax)
1144 -- compare against upper
1145 -- eax==0 if high part == sign extended low part
1148 return (Fixed size eax code)
1150 --------------------
1152 -> (Size -> Operand -> Operand -> Instr)
1157 {- Case1: shift length as immediate -}
1158 shift_code width instr x y@(CmmLit lit) = do
1159 x_code <- getAnyReg x
1161 size = intSize width
1163 = x_code dst `snocOL`
1164 instr size (OpImm (litToImm lit)) (OpReg dst)
1166 return (Any size code)
1168 {- Case2: shift length is complex (non-immediate)
1169 * y must go in %ecx.
1170 * we cannot do y first *and* put its result in %ecx, because
1171 %ecx might be clobbered by x.
1172 * if we do y second, then x cannot be
1173 in a clobbered reg. Also, we cannot clobber x's reg
1174 with the instruction itself.
1176 - do y first, put its result in a fresh tmp, then copy it to %ecx later
1177 - do y second and put its result into %ecx. x gets placed in a fresh
1178 tmp. This is likely to be better, becuase the reg alloc can
1179 eliminate this reg->reg move here (it won't eliminate the other one,
1180 because the move is into the fixed %ecx).
1182 shift_code width instr x y{-amount-} = do
1183 x_code <- getAnyReg x
1184 let size = intSize width
1185 tmp <- getNewRegNat size
1186 y_code <- getAnyReg y
1188 code = x_code tmp `appOL`
1190 instr size (OpReg ecx) (OpReg tmp)
1192 return (Fixed size tmp code)
1194 --------------------
1195 add_code :: Width -> CmmExpr -> CmmExpr -> NatM Register
1196 add_code rep x (CmmLit (CmmInt y _))
1197 | is32BitInteger y = add_int rep x y
1198 add_code rep x y = trivialCode rep (ADD size) (Just (ADD size)) x y
1199 where size = intSize rep
1201 --------------------
1202 sub_code :: Width -> CmmExpr -> CmmExpr -> NatM Register
1203 sub_code rep x (CmmLit (CmmInt y _))
1204 | is32BitInteger (-y) = add_int rep x (-y)
1205 sub_code rep x y = trivialCode rep (SUB (intSize rep)) Nothing x y
1207 -- our three-operand add instruction:
1208 add_int width x y = do
1209 (x_reg, x_code) <- getSomeReg x
1211 size = intSize width
1212 imm = ImmInt (fromInteger y)
1216 (OpAddr (AddrBaseIndex (EABaseReg x_reg) EAIndexNone imm))
1219 return (Any size code)
1221 ----------------------
1222 div_code width signed quotient x y = do
1223 (y_op, y_code) <- getRegOrMem y -- cannot be clobbered
1224 x_code <- getAnyReg x
1226 size = intSize width
1227 widen | signed = CLTD size
1228 | otherwise = XOR size (OpReg edx) (OpReg edx)
1230 instr | signed = IDIV
1233 code = y_code `appOL`
1235 toOL [widen, instr size y_op]
1237 result | quotient = eax
1241 return (Fixed size result code)
1244 getRegister (CmmLoad mem pk)
1247 Amode src mem_code <- getAmode mem
1249 size = cmmTypeSize pk
1250 code dst = mem_code `snocOL`
1251 IF_ARCH_i386(GLD size src dst,
1252 MOV size (OpAddr src) (OpReg dst))
1253 return (Any size code)
1255 #if i386_TARGET_ARCH
1256 getRegister (CmmLoad mem pk)
1259 code <- intLoadCode instr mem
1260 return (Any size code)
1262 width = typeWidth pk
1263 size = intSize width
1264 instr = case width of
1267 -- We always zero-extend 8-bit loads, if we
1268 -- can't think of anything better. This is because
1269 -- we can't guarantee access to an 8-bit variant of every register
1270 -- (esi and edi don't have 8-bit variants), so to make things
1271 -- simpler we do our 8-bit arithmetic with full 32-bit registers.
1274 #if x86_64_TARGET_ARCH
1275 -- Simpler memory load code on x86_64
1276 getRegister (CmmLoad mem pk)
1278 code <- intLoadCode (MOV size) mem
1279 return (Any size code)
1280 where size = intSize $ typeWidth pk
1283 getRegister (CmmLit (CmmInt 0 width))
1285 size = intSize width
1287 -- x86_64: 32-bit xor is one byte shorter, and zero-extends to 64 bits
1288 adj_size = case size of II64 -> II32; _ -> size
1289 size1 = IF_ARCH_i386( size, adj_size )
1291 = unitOL (XOR size1 (OpReg dst) (OpReg dst))
1293 return (Any size code)
1295 #if x86_64_TARGET_ARCH
1296 -- optimisation for loading small literals on x86_64: take advantage
1297 -- of the automatic zero-extension from 32 to 64 bits, because the 32-bit
1298 -- instruction forms are shorter.
1299 getRegister (CmmLit lit)
1300 | isWord64 (cmmLitType lit), not (isBigLit lit)
1303 code dst = unitOL (MOV II32 (OpImm imm) (OpReg dst))
1305 return (Any II64 code)
1307 isBigLit (CmmInt i _) = i < 0 || i > 0xffffffff
1309 -- note1: not the same as (not.is32BitLit), because that checks for
1310 -- signed literals that fit in 32 bits, but we want unsigned
1312 -- note2: all labels are small, because we're assuming the
1313 -- small memory model (see gcc docs, -mcmodel=small).
1316 getRegister (CmmLit lit)
1318 size = cmmTypeSize (cmmLitType lit)
1320 code dst = unitOL (MOV size (OpImm imm) (OpReg dst))
1322 return (Any size code)
1324 getRegister other = pprPanic "getRegister(x86)" (ppr other)
1327 intLoadCode :: (Operand -> Operand -> Instr) -> CmmExpr
1328 -> NatM (Reg -> InstrBlock)
1329 intLoadCode instr mem = do
1330 Amode src mem_code <- getAmode mem
1331 return (\dst -> mem_code `snocOL` instr (OpAddr src) (OpReg dst))
1333 -- Compute an expression into *any* register, adding the appropriate
1334 -- move instruction if necessary.
1335 getAnyReg :: CmmExpr -> NatM (Reg -> InstrBlock)
1337 r <- getRegister expr
1340 anyReg :: Register -> NatM (Reg -> InstrBlock)
1341 anyReg (Any _ code) = return code
1342 anyReg (Fixed rep reg fcode) = return (\dst -> fcode `snocOL` reg2reg rep reg dst)
1344 -- A bit like getSomeReg, but we want a reg that can be byte-addressed.
1345 -- Fixed registers might not be byte-addressable, so we make sure we've
1346 -- got a temporary, inserting an extra reg copy if necessary.
1347 getByteReg :: CmmExpr -> NatM (Reg, InstrBlock)
1348 #if x86_64_TARGET_ARCH
1349 getByteReg = getSomeReg -- all regs are byte-addressable on x86_64
1351 getByteReg expr = do
1352 r <- getRegister expr
1355 tmp <- getNewRegNat rep
1356 return (tmp, code tmp)
1358 | isVirtualReg reg -> return (reg,code)
1360 tmp <- getNewRegNat rep
1361 return (tmp, code `snocOL` reg2reg rep reg tmp)
1362 -- ToDo: could optimise slightly by checking for byte-addressable
1363 -- real registers, but that will happen very rarely if at all.
1366 -- Another variant: this time we want the result in a register that cannot
1367 -- be modified by code to evaluate an arbitrary expression.
1368 getNonClobberedReg :: CmmExpr -> NatM (Reg, InstrBlock)
1369 getNonClobberedReg expr = do
1370 r <- getRegister expr
1373 tmp <- getNewRegNat rep
1374 return (tmp, code tmp)
1376 -- only free regs can be clobbered
1377 | RealReg rr <- reg, isFastTrue (freeReg rr) -> do
1378 tmp <- getNewRegNat rep
1379 return (tmp, code `snocOL` reg2reg rep reg tmp)
1383 reg2reg :: Size -> Reg -> Reg -> Instr
1384 reg2reg size src dst
1385 #if i386_TARGET_ARCH
1386 | isFloatSize size = GMOV src dst
1388 | otherwise = MOV size (OpReg src) (OpReg dst)
1390 #endif /* i386_TARGET_ARCH || x86_64_TARGET_ARCH */
1392 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
1394 #if sparc_TARGET_ARCH
1396 getRegister (CmmLit (CmmFloat f W32)) = do
1397 lbl <- getNewLabelNat
1398 let code dst = toOL [
1401 CmmStaticLit (CmmFloat f W32)],
1402 SETHI (HI (ImmCLbl lbl)) dst,
1403 LD FF32 (AddrRegImm dst (LO (ImmCLbl lbl))) dst]
1404 return (Any FF32 code)
1406 getRegister (CmmLit (CmmFloat d W64)) = do
1407 lbl <- getNewLabelNat
1408 let code dst = toOL [
1411 CmmStaticLit (CmmFloat d W64)],
1412 SETHI (HI (ImmCLbl lbl)) dst,
1413 LD FF64 (AddrRegImm dst (LO (ImmCLbl lbl))) dst]
1414 return (Any FF64 code)
1416 getRegister (CmmMachOp mop [x]) -- unary MachOps
1418 MO_F_Neg W32 -> trivialUFCode FF32 (FNEG FF32) x
1419 MO_F_Neg W64 -> trivialUFCode FF64 (FNEG FF64) x
1421 MO_S_Neg rep -> trivialUCode (intSize rep) (SUB False False g0) x
1422 MO_Not rep -> trivialUCode (intSize rep) (XNOR False g0) x
1424 MO_FF_Conv W64 W32-> coerceDbl2Flt x
1425 MO_FF_Conv W32 W64-> coerceFlt2Dbl x
1427 MO_FS_Conv from to -> coerceFP2Int from to x
1428 MO_SF_Conv from to -> coerceInt2FP from to x
1430 -- Conversions which are a nop on sparc
1432 | from == to -> conversionNop to x
1433 MO_UU_Conv W32 W8 -> trivialCode I8 (AND False) x (CmmLit (CmmInt 255 I8))
1434 MO_UU_Conv W32 to -> conversionNop to x
1435 MO_SS_Conv W32 to -> conversionNop to x
1438 MO_UU_Conv W8 W32 -> integerExtend False W8 W32 x
1439 MO_UU_Conv W16 W32 -> integerExtend False W16 W32 x
1440 MO_UU_Conv W8 W16 -> integerExtend False W8 W16 x
1441 MO_SS_Conv W16 W32 -> integerExtend True W16 W32 x
1443 other_op -> panic "Unknown unary mach op"
1446 integerExtend signed from to expr = do
1447 (reg, e_code) <- getSomeReg expr
1451 ((if signed then SRA else SRL)
1452 reg (RIImm (ImmInt 0)) dst)
1453 return (Any (intSize to) code)
1454 conversionNop new_rep expr
1455 = do e_code <- getRegister expr
1456 return (swizzleRegisterRep e_code new_rep)
1458 getRegister (CmmMachOp mop [x, y]) -- dyadic PrimOps
1460 MO_Eq rep -> condIntReg EQQ x y
1461 MO_Ne rep -> condIntReg NE x y
1463 MO_S_Gt rep -> condIntReg GTT x y
1464 MO_S_Ge rep -> condIntReg GE x y
1465 MO_S_Lt rep -> condIntReg LTT x y
1466 MO_S_Le rep -> condIntReg LE x y
1468 MO_U_Gt W32 -> condIntReg GTT x y
1469 MO_U_Ge W32 -> condIntReg GE x y
1470 MO_U_Lt W32 -> condIntReg LTT x y
1471 MO_U_Le W32 -> condIntReg LE x y
1473 MO_U_Gt W16 -> condIntReg GU x y
1474 MO_U_Ge W16 -> condIntReg GEU x y
1475 MO_U_Lt W16 -> condIntReg LU x y
1476 MO_U_Le W16 -> condIntReg LEU x y
1478 MO_Add W32 -> trivialCode W32 (ADD False False) x y
1479 MO_Sub W32 -> trivialCode W32 (SUB False False) x y
1481 MO_S_MulMayOflo rep -> imulMayOflo rep x y
1483 -- ToDo: teach about V8+ SPARC div instructions
1484 MO_S_Quot W32 -> idiv FSLIT(".div") x y
1485 MO_S_Rem W32 -> idiv FSLIT(".rem") x y
1486 MO_U_Quot W32 -> idiv FSLIT(".udiv") x y
1487 MO_U_Rem W32 -> idiv FSLIT(".urem") x y
1490 MO_F_Eq w -> condFltReg EQQ x y
1491 MO_F_Ne w -> condFltReg NE x y
1493 MO_F_Gt w -> condFltReg GTT x y
1494 MO_F_Ge w -> condFltReg GE x y
1495 MO_F_Lt w -> condFltReg LTT x y
1496 MO_F_Le w -> condFltReg LE x y
1498 MO_F_Add w -> trivialFCode w FADD x y
1499 MO_F_Sub w -> trivialFCode w FSUB x y
1500 MO_F_Mul w -> trivialFCode w FMUL x y
1501 MO_F_Quot w -> trivialFCode w FDIV x y
1503 MO_And rep -> trivialCode rep (AND False) x y
1504 MO_Or rep -> trivialCode rep (OR False) x y
1505 MO_Xor rep -> trivialCode rep (XOR False) x y
1507 MO_Mul rep -> trivialCode rep (SMUL False) x y
1509 MO_Shl rep -> trivialCode rep SLL x y
1510 MO_U_Shr rep -> trivialCode rep SRL x y
1511 MO_S_Shr rep -> trivialCode rep SRA x y
1514 MO_F32_Pwr -> getRegister (StCall (Left (fsLit "pow")) CCallConv FF64
1515 [promote x, promote y])
1516 where promote x = CmmMachOp MO_F32_to_Dbl [x]
1517 MO_F64_Pwr -> getRegister (StCall (Left (fsLit "pow")) CCallConv FF64
1520 other -> pprPanic "getRegister(sparc) - binary CmmMachOp (1)" (pprMachOp mop)
1522 --idiv fn x y = getRegister (StCall (Left fn) CCallConv II32 [x, y])
1524 --------------------
1525 imulMayOflo :: Width -> CmmExpr -> CmmExpr -> NatM Register
1526 imulMayOflo rep a b = do
1527 (a_reg, a_code) <- getSomeReg a
1528 (b_reg, b_code) <- getSomeReg b
1529 res_lo <- getNewRegNat b32
1530 res_hi <- getNewRegNat b32
1532 shift_amt = case rep of
1535 _ -> panic "shift_amt"
1536 code dst = a_code `appOL` b_code `appOL`
1538 SMUL False a_reg (RIReg b_reg) res_lo,
1540 SRA res_lo (RIImm (ImmInt shift_amt)) res_lo,
1541 SUB False False res_lo (RIReg res_hi) dst
1543 return (Any II32 code)
1545 getRegister (CmmLoad mem pk) = do
1546 Amode src code <- getAmode mem
1548 code__2 dst = code `snocOL` LD pk src dst
1549 return (Any pk code__2)
1551 getRegister (CmmLit (CmmInt i _))
1554 src = ImmInt (fromInteger i)
1555 code dst = unitOL (OR False g0 (RIImm src) dst)
1557 return (Any II32 code)
1559 getRegister (CmmLit lit)
1560 = let rep = cmmLitType lit
1564 OR False dst (RIImm (LO imm)) dst]
1565 in return (Any II32 code)
1567 #endif /* sparc_TARGET_ARCH */
1569 #if powerpc_TARGET_ARCH
1570 getRegister (CmmLoad mem pk)
1573 Amode addr addr_code <- getAmode mem
1574 let code dst = ASSERT((regClass dst == RcDouble) == isFloatType pk)
1575 addr_code `snocOL` LD size dst addr
1576 return (Any size code)
1577 where size = cmmTypeSize pk
1579 -- catch simple cases of zero- or sign-extended load
1580 getRegister (CmmMachOp (MO_UU_Conv W8 W32) [CmmLoad mem _]) = do
1581 Amode addr addr_code <- getAmode mem
1582 return (Any II32 (\dst -> addr_code `snocOL` LD II8 dst addr))
1584 -- Note: there is no Load Byte Arithmetic instruction, so no signed case here
1586 getRegister (CmmMachOp (MO_UU_Conv W16 W32) [CmmLoad mem _]) = do
1587 Amode addr addr_code <- getAmode mem
1588 return (Any II32 (\dst -> addr_code `snocOL` LD II16 dst addr))
1590 getRegister (CmmMachOp (MO_SS_Conv W16 W32) [CmmLoad mem _]) = do
1591 Amode addr addr_code <- getAmode mem
1592 return (Any II32 (\dst -> addr_code `snocOL` LA II16 dst addr))
1594 getRegister (CmmMachOp mop [x]) -- unary MachOps
1596 MO_Not rep -> triv_ucode_int rep NOT
1598 MO_F_Neg w -> triv_ucode_float w FNEG
1599 MO_S_Neg w -> triv_ucode_int w NEG
1601 MO_FF_Conv W64 W32 -> trivialUCode FF32 FRSP x
1602 MO_FF_Conv W32 W64 -> conversionNop FF64 x
1604 MO_FS_Conv from to -> coerceFP2Int from to x
1605 MO_SF_Conv from to -> coerceInt2FP from to x
1608 | from == to -> conversionNop (intSize to) x
1610 -- narrowing is a nop: we treat the high bits as undefined
1611 MO_SS_Conv W32 to -> conversionNop (intSize to) x
1612 MO_SS_Conv W16 W8 -> conversionNop II8 x
1613 MO_SS_Conv W8 to -> triv_ucode_int to (EXTS II8)
1614 MO_SS_Conv W16 to -> triv_ucode_int to (EXTS II16)
1617 | from == to -> conversionNop (intSize to) x
1618 -- narrowing is a nop: we treat the high bits as undefined
1619 MO_UU_Conv W32 to -> conversionNop (intSize to) x
1620 MO_UU_Conv W16 W8 -> conversionNop II8 x
1621 MO_UU_Conv W8 to -> trivialCode to False AND x (CmmLit (CmmInt 255 W32))
1622 MO_UU_Conv W16 to -> trivialCode to False AND x (CmmLit (CmmInt 65535 W32))
1625 triv_ucode_int width instr = trivialUCode (intSize width) instr x
1626 triv_ucode_float width instr = trivialUCode (floatSize width) instr x
1628 conversionNop new_size expr
1629 = do e_code <- getRegister expr
1630 return (swizzleRegisterRep e_code new_size)
1632 getRegister (CmmMachOp mop [x, y]) -- dyadic PrimOps
1634 MO_F_Eq w -> condFltReg EQQ x y
1635 MO_F_Ne w -> condFltReg NE x y
1636 MO_F_Gt w -> condFltReg GTT x y
1637 MO_F_Ge w -> condFltReg GE x y
1638 MO_F_Lt w -> condFltReg LTT x y
1639 MO_F_Le w -> condFltReg LE x y
1641 MO_Eq rep -> condIntReg EQQ (extendUExpr rep x) (extendUExpr rep y)
1642 MO_Ne rep -> condIntReg NE (extendUExpr rep x) (extendUExpr rep y)
1644 MO_S_Gt rep -> condIntReg GTT (extendSExpr rep x) (extendSExpr rep y)
1645 MO_S_Ge rep -> condIntReg GE (extendSExpr rep x) (extendSExpr rep y)
1646 MO_S_Lt rep -> condIntReg LTT (extendSExpr rep x) (extendSExpr rep y)
1647 MO_S_Le rep -> condIntReg LE (extendSExpr rep x) (extendSExpr rep y)
1649 MO_U_Gt rep -> condIntReg GU (extendUExpr rep x) (extendUExpr rep y)
1650 MO_U_Ge rep -> condIntReg GEU (extendUExpr rep x) (extendUExpr rep y)
1651 MO_U_Lt rep -> condIntReg LU (extendUExpr rep x) (extendUExpr rep y)
1652 MO_U_Le rep -> condIntReg LEU (extendUExpr rep x) (extendUExpr rep y)
1654 MO_F_Add w -> triv_float w FADD
1655 MO_F_Sub w -> triv_float w FSUB
1656 MO_F_Mul w -> triv_float w FMUL
1657 MO_F_Quot w -> triv_float w FDIV
1659 -- optimize addition with 32-bit immediate
1663 CmmLit (CmmInt imm immrep) | Just _ <- makeImmediate W32 True (-imm)
1664 -> trivialCode W32 True ADD x (CmmLit $ CmmInt imm immrep)
1667 (src, srcCode) <- getSomeReg x
1668 let imm = litToImm lit
1669 code dst = srcCode `appOL` toOL [
1670 ADDIS dst src (HA imm),
1671 ADD dst dst (RIImm (LO imm))
1673 return (Any II32 code)
1674 _ -> trivialCode W32 True ADD x y
1676 MO_Add rep -> trivialCode rep True ADD x y
1678 case y of -- subfi ('substract from' with immediate) doesn't exist
1679 CmmLit (CmmInt imm immrep) | Just _ <- makeImmediate rep True (-imm)
1680 -> trivialCode rep True ADD x (CmmLit $ CmmInt (-imm) immrep)
1681 _ -> trivialCodeNoImm' (intSize rep) SUBF y x
1683 MO_Mul rep -> trivialCode rep True MULLW x y
1685 MO_S_MulMayOflo W32 -> trivialCodeNoImm' II32 MULLW_MayOflo x y
1687 MO_S_MulMayOflo rep -> panic "S_MulMayOflo (rep /= II32): not implemented"
1688 MO_U_MulMayOflo rep -> panic "U_MulMayOflo: not implemented"
1690 MO_S_Quot rep -> trivialCodeNoImm' (intSize rep) DIVW (extendSExpr rep x) (extendSExpr rep y)
1691 MO_U_Quot rep -> trivialCodeNoImm' (intSize rep) DIVWU (extendUExpr rep x) (extendUExpr rep y)
1693 MO_S_Rem rep -> remainderCode rep DIVW (extendSExpr rep x) (extendSExpr rep y)
1694 MO_U_Rem rep -> remainderCode rep DIVWU (extendUExpr rep x) (extendUExpr rep y)
1696 MO_And rep -> trivialCode rep False AND x y
1697 MO_Or rep -> trivialCode rep False OR x y
1698 MO_Xor rep -> trivialCode rep False XOR x y
1700 MO_Shl rep -> trivialCode rep False SLW x y
1701 MO_S_Shr rep -> trivialCode rep False SRAW (extendSExpr rep x) y
1702 MO_U_Shr rep -> trivialCode rep False SRW (extendUExpr rep x) y
1704 triv_float :: Width -> (Size -> Reg -> Reg -> Reg -> Instr) -> NatM Register
1705 triv_float width instr = trivialCodeNoImm (floatSize width) instr x y
1707 getRegister (CmmLit (CmmInt i rep))
1708 | Just imm <- makeImmediate rep True i
1710 code dst = unitOL (LI dst imm)
1712 return (Any (intSize rep) code)
1714 getRegister (CmmLit (CmmFloat f frep)) = do
1715 lbl <- getNewLabelNat
1716 dflags <- getDynFlagsNat
1717 dynRef <- cmmMakeDynamicReference dflags addImportNat DataReference lbl
1718 Amode addr addr_code <- getAmode dynRef
1719 let size = floatSize frep
1721 LDATA ReadOnlyData [CmmDataLabel lbl,
1722 CmmStaticLit (CmmFloat f frep)]
1723 `consOL` (addr_code `snocOL` LD size dst addr)
1724 return (Any size code)
1726 getRegister (CmmLit lit)
1727 = let rep = cmmLitType lit
1731 ADD dst dst (RIImm (LO imm))
1733 in return (Any (cmmTypeSize rep) code)
1735 getRegister other = pprPanic "getRegister(ppc)" (pprExpr other)
1737 -- extend?Rep: wrap integer expression of type rep
1738 -- in a conversion to II32
1739 extendSExpr W32 x = x
1740 extendSExpr rep x = CmmMachOp (MO_SS_Conv rep W32) [x]
1741 extendUExpr W32 x = x
1742 extendUExpr rep x = CmmMachOp (MO_UU_Conv rep W32) [x]
1744 #endif /* powerpc_TARGET_ARCH */
1747 -- -----------------------------------------------------------------------------
1748 -- The 'Amode' type: Memory addressing modes passed up the tree.
1750 data Amode = Amode AddrMode InstrBlock
1753 Now, given a tree (the argument to an CmmLoad) that references memory,
1754 produce a suitable addressing mode.
1756 A Rule of the Game (tm) for Amodes: use of the addr bit must
1757 immediately follow use of the code part, since the code part puts
1758 values in registers which the addr then refers to. So you can't put
1759 anything in between, lest it overwrite some of those registers. If
1760 you need to do some other computation between the code part and use of
1761 the addr bit, first store the effective address from the amode in a
1762 temporary, then do the other computation, and then use the temporary:
1766 ... other computation ...
1770 getAmode :: CmmExpr -> NatM Amode
1771 getAmode tree@(CmmRegOff _ _) = getAmode (mangleIndexTree tree)
1773 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
1775 #if alpha_TARGET_ARCH
1777 getAmode (StPrim IntSubOp [x, StInt i])
1778 = getNewRegNat PtrRep `thenNat` \ tmp ->
1779 getRegister x `thenNat` \ register ->
1781 code = registerCode register tmp
1782 reg = registerName register tmp
1783 off = ImmInt (-(fromInteger i))
1785 return (Amode (AddrRegImm reg off) code)
1787 getAmode (StPrim IntAddOp [x, StInt i])
1788 = getNewRegNat PtrRep `thenNat` \ tmp ->
1789 getRegister x `thenNat` \ register ->
1791 code = registerCode register tmp
1792 reg = registerName register tmp
1793 off = ImmInt (fromInteger i)
1795 return (Amode (AddrRegImm reg off) code)
1799 = return (Amode (AddrImm imm__2) id)
1802 imm__2 = case imm of Just x -> x
1805 = getNewRegNat PtrRep `thenNat` \ tmp ->
1806 getRegister other `thenNat` \ register ->
1808 code = registerCode register tmp
1809 reg = registerName register tmp
1811 return (Amode (AddrReg reg) code)
1813 #endif /* alpha_TARGET_ARCH */
1815 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
1817 #if x86_64_TARGET_ARCH
1819 getAmode (CmmMachOp (MO_Add W64) [CmmReg (CmmGlobal PicBaseReg),
1820 CmmLit displacement])
1821 = return $ Amode (ripRel (litToImm displacement)) nilOL
1825 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
1827 -- This is all just ridiculous, since it carefully undoes
1828 -- what mangleIndexTree has just done.
1829 getAmode (CmmMachOp (MO_Sub rep) [x, CmmLit lit@(CmmInt i _)])
1831 -- ASSERT(rep == II32)???
1832 = do (x_reg, x_code) <- getSomeReg x
1833 let off = ImmInt (-(fromInteger i))
1834 return (Amode (AddrBaseIndex (EABaseReg x_reg) EAIndexNone off) x_code)
1836 getAmode (CmmMachOp (MO_Add rep) [x, CmmLit lit@(CmmInt i _)])
1838 -- ASSERT(rep == II32)???
1839 = do (x_reg, x_code) <- getSomeReg x
1840 let off = ImmInt (fromInteger i)
1841 return (Amode (AddrBaseIndex (EABaseReg x_reg) EAIndexNone off) x_code)
1843 -- Turn (lit1 << n + lit2) into (lit2 + lit1 << n) so it will be
1844 -- recognised by the next rule.
1845 getAmode (CmmMachOp (MO_Add rep) [a@(CmmMachOp (MO_Shl _) _),
1847 = getAmode (CmmMachOp (MO_Add rep) [b,a])
1849 getAmode (CmmMachOp (MO_Add rep) [x, CmmMachOp (MO_Shl _)
1850 [y, CmmLit (CmmInt shift _)]])
1851 | shift == 0 || shift == 1 || shift == 2 || shift == 3
1852 = x86_complex_amode x y shift 0
1854 getAmode (CmmMachOp (MO_Add rep)
1855 [x, CmmMachOp (MO_Add _)
1856 [CmmMachOp (MO_Shl _) [y, CmmLit (CmmInt shift _)],
1857 CmmLit (CmmInt offset _)]])
1858 | shift == 0 || shift == 1 || shift == 2 || shift == 3
1859 && is32BitInteger offset
1860 = x86_complex_amode x y shift offset
1862 getAmode (CmmMachOp (MO_Add rep) [x,y])
1863 = x86_complex_amode x y 0 0
1865 getAmode (CmmLit lit) | is32BitLit lit
1866 = return (Amode (ImmAddr (litToImm lit) 0) nilOL)
1869 (reg,code) <- getSomeReg expr
1870 return (Amode (AddrBaseIndex (EABaseReg reg) EAIndexNone (ImmInt 0)) code)
1873 x86_complex_amode :: CmmExpr -> CmmExpr -> Integer -> Integer -> NatM Amode
1874 x86_complex_amode base index shift offset
1875 = do (x_reg, x_code) <- getNonClobberedReg base
1876 -- x must be in a temp, because it has to stay live over y_code
1877 -- we could compre x_reg and y_reg and do something better here...
1878 (y_reg, y_code) <- getSomeReg index
1880 code = x_code `appOL` y_code
1881 base = case shift of 0 -> 1; 1 -> 2; 2 -> 4; 3 -> 8
1882 return (Amode (AddrBaseIndex (EABaseReg x_reg) (EAIndex y_reg base) (ImmInt (fromIntegral offset)))
1885 #endif /* i386_TARGET_ARCH || x86_64_TARGET_ARCH */
1887 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
1889 #if sparc_TARGET_ARCH
1891 getAmode (CmmMachOp (MO_Sub rep) [x, CmmLit (CmmInt i _)])
1894 (reg, code) <- getSomeReg x
1896 off = ImmInt (-(fromInteger i))
1897 return (Amode (AddrRegImm reg off) code)
1900 getAmode (CmmMachOp (MO_Add rep) [x, CmmLit (CmmInt i _)])
1903 (reg, code) <- getSomeReg x
1905 off = ImmInt (fromInteger i)
1906 return (Amode (AddrRegImm reg off) code)
1908 getAmode (CmmMachOp (MO_Add rep) [x, y])
1910 (regX, codeX) <- getSomeReg x
1911 (regY, codeY) <- getSomeReg y
1913 code = codeX `appOL` codeY
1914 return (Amode (AddrRegReg regX regY) code)
1916 -- XXX Is this same as "leaf" in Stix?
1917 getAmode (CmmLit lit)
1919 tmp <- getNewRegNat b32
1921 code = unitOL (SETHI (HI imm__2) tmp)
1922 return (Amode (AddrRegImm tmp (LO imm__2)) code)
1924 imm__2 = litToImm lit
1928 (reg, code) <- getSomeReg other
1931 return (Amode (AddrRegImm reg off) code)
1933 #endif /* sparc_TARGET_ARCH */
1935 #ifdef powerpc_TARGET_ARCH
1936 getAmode (CmmMachOp (MO_Sub W32) [x, CmmLit (CmmInt i _)])
1937 | Just off <- makeImmediate W32 True (-i)
1939 (reg, code) <- getSomeReg x
1940 return (Amode (AddrRegImm reg off) code)
1943 getAmode (CmmMachOp (MO_Add W32) [x, CmmLit (CmmInt i _)])
1944 | Just off <- makeImmediate W32 True i
1946 (reg, code) <- getSomeReg x
1947 return (Amode (AddrRegImm reg off) code)
1949 -- optimize addition with 32-bit immediate
1951 getAmode (CmmMachOp (MO_Add W32) [x, CmmLit lit])
1953 tmp <- getNewRegNat II32
1954 (src, srcCode) <- getSomeReg x
1955 let imm = litToImm lit
1956 code = srcCode `snocOL` ADDIS tmp src (HA imm)
1957 return (Amode (AddrRegImm tmp (LO imm)) code)
1959 getAmode (CmmLit lit)
1961 tmp <- getNewRegNat II32
1962 let imm = litToImm lit
1963 code = unitOL (LIS tmp (HA imm))
1964 return (Amode (AddrRegImm tmp (LO imm)) code)
1966 getAmode (CmmMachOp (MO_Add W32) [x, y])
1968 (regX, codeX) <- getSomeReg x
1969 (regY, codeY) <- getSomeReg y
1970 return (Amode (AddrRegReg regX regY) (codeX `appOL` codeY))
1974 (reg, code) <- getSomeReg other
1977 return (Amode (AddrRegImm reg off) code)
1978 #endif /* powerpc_TARGET_ARCH */
1980 -- -----------------------------------------------------------------------------
1981 -- getOperand: sometimes any operand will do.
1983 -- getNonClobberedOperand: the value of the operand will remain valid across
1984 -- the computation of an arbitrary expression, unless the expression
1985 -- is computed directly into a register which the operand refers to
1986 -- (see trivialCode where this function is used for an example).
1988 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
1990 getNonClobberedOperand :: CmmExpr -> NatM (Operand, InstrBlock)
1991 #if x86_64_TARGET_ARCH
1992 getNonClobberedOperand (CmmLit lit)
1993 | isSuitableFloatingPointLit lit = do
1994 lbl <- getNewLabelNat
1995 let code = unitOL (LDATA ReadOnlyData [CmmDataLabel lbl,
1997 return (OpAddr (ripRel (ImmCLbl lbl)), code)
1999 getNonClobberedOperand (CmmLit lit)
2000 | is32BitLit lit && not (isFloatType (cmmLitType lit)) =
2001 return (OpImm (litToImm lit), nilOL)
2002 getNonClobberedOperand (CmmLoad mem pk)
2003 | IF_ARCH_i386(not (isFloatType pk) && not (isWord64 pk), True) = do
2004 Amode src mem_code <- getAmode mem
2006 if (amodeCouldBeClobbered src)
2008 tmp <- getNewRegNat wordSize
2009 return (AddrBaseIndex (EABaseReg tmp) EAIndexNone (ImmInt 0),
2010 unitOL (LEA II32 (OpAddr src) (OpReg tmp)))
2013 return (OpAddr src', save_code `appOL` mem_code)
2014 getNonClobberedOperand e = do
2015 (reg, code) <- getNonClobberedReg e
2016 return (OpReg reg, code)
2018 amodeCouldBeClobbered :: AddrMode -> Bool
2019 amodeCouldBeClobbered amode = any regClobbered (addrModeRegs amode)
2021 regClobbered (RealReg rr) = isFastTrue (freeReg rr)
2022 regClobbered _ = False
2024 -- getOperand: the operand is not required to remain valid across the
2025 -- computation of an arbitrary expression.
2026 getOperand :: CmmExpr -> NatM (Operand, InstrBlock)
2027 #if x86_64_TARGET_ARCH
2028 getOperand (CmmLit lit)
2029 | isSuitableFloatingPointLit lit = do
2030 lbl <- getNewLabelNat
2031 let code = unitOL (LDATA ReadOnlyData [CmmDataLabel lbl,
2033 return (OpAddr (ripRel (ImmCLbl lbl)), code)
2035 getOperand (CmmLit lit)
2036 | is32BitLit lit && not (isFloatType (cmmLitType lit)) = do
2037 return (OpImm (litToImm lit), nilOL)
2038 getOperand (CmmLoad mem pk)
2039 | IF_ARCH_i386(not (isFloatType pk) && not (isWord64 pk), True) = do
2040 Amode src mem_code <- getAmode mem
2041 return (OpAddr src, mem_code)
2043 (reg, code) <- getSomeReg e
2044 return (OpReg reg, code)
2046 isOperand :: CmmExpr -> Bool
2047 isOperand (CmmLoad _ _) = True
2048 isOperand (CmmLit lit) = is32BitLit lit
2049 || isSuitableFloatingPointLit lit
2052 -- if we want a floating-point literal as an operand, we can
2053 -- use it directly from memory. However, if the literal is
2054 -- zero, we're better off generating it into a register using
2056 isSuitableFloatingPointLit (CmmFloat f _) = f /= 0.0
2057 isSuitableFloatingPointLit _ = False
2059 getRegOrMem :: CmmExpr -> NatM (Operand, InstrBlock)
2060 getRegOrMem (CmmLoad mem pk)
2061 | IF_ARCH_i386(not (isFloatType pk) && not (isWord64 pk), True) = do
2062 Amode src mem_code <- getAmode mem
2063 return (OpAddr src, mem_code)
2065 (reg, code) <- getNonClobberedReg e
2066 return (OpReg reg, code)
2068 #if x86_64_TARGET_ARCH
2069 is32BitLit (CmmInt i W64) = is32BitInteger i
2070 -- assume that labels are in the range 0-2^31-1: this assumes the
2071 -- small memory model (see gcc docs, -mcmodel=small).
2076 is32BitInteger :: Integer -> Bool
2077 is32BitInteger i = i64 <= 0x7fffffff && i64 >= -0x80000000
2078 where i64 = fromIntegral i :: Int64
2079 -- a CmmInt is intended to be truncated to the appropriate
2080 -- number of bits, so here we truncate it to Int64. This is
2081 -- important because e.g. -1 as a CmmInt might be either
2082 -- -1 or 18446744073709551615.
2084 -- -----------------------------------------------------------------------------
2085 -- The 'CondCode' type: Condition codes passed up the tree.
2087 data CondCode = CondCode Bool Cond InstrBlock
2089 -- Set up a condition code for a conditional branch.
2091 getCondCode :: CmmExpr -> NatM CondCode
2093 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2095 #if alpha_TARGET_ARCH
2096 getCondCode = panic "MachCode.getCondCode: not on Alphas"
2097 #endif /* alpha_TARGET_ARCH */
2099 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2101 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH || sparc_TARGET_ARCH
2102 -- yes, they really do seem to want exactly the same!
2104 getCondCode (CmmMachOp mop [x, y])
2107 MO_F_Eq W32 -> condFltCode EQQ x y
2108 MO_F_Ne W32 -> condFltCode NE x y
2109 MO_F_Gt W32 -> condFltCode GTT x y
2110 MO_F_Ge W32 -> condFltCode GE x y
2111 MO_F_Lt W32 -> condFltCode LTT x y
2112 MO_F_Le W32 -> condFltCode LE x y
2114 MO_F_Eq W64 -> condFltCode EQQ x y
2115 MO_F_Ne W64 -> condFltCode NE x y
2116 MO_F_Gt W64 -> condFltCode GTT x y
2117 MO_F_Ge W64 -> condFltCode GE x y
2118 MO_F_Lt W64 -> condFltCode LTT x y
2119 MO_F_Le W64 -> condFltCode LE x y
2121 MO_Eq rep -> condIntCode EQQ x y
2122 MO_Ne rep -> condIntCode NE x y
2124 MO_S_Gt rep -> condIntCode GTT x y
2125 MO_S_Ge rep -> condIntCode GE x y
2126 MO_S_Lt rep -> condIntCode LTT x y
2127 MO_S_Le rep -> condIntCode LE x y
2129 MO_U_Gt rep -> condIntCode GU x y
2130 MO_U_Ge rep -> condIntCode GEU x y
2131 MO_U_Lt rep -> condIntCode LU x y
2132 MO_U_Le rep -> condIntCode LEU x y
2134 other -> pprPanic "getCondCode(x86,x86_64,sparc)" (ppr (CmmMachOp mop [x,y]))
2136 getCondCode other = pprPanic "getCondCode(2)(x86,sparc)" (ppr other)
2138 #elif powerpc_TARGET_ARCH
2140 -- almost the same as everywhere else - but we need to
2141 -- extend small integers to 32 bit first
2143 getCondCode (CmmMachOp mop [x, y])
2145 MO_F_Eq W32 -> condFltCode EQQ x y
2146 MO_F_Ne W32 -> condFltCode NE x y
2147 MO_F_Gt W32 -> condFltCode GTT x y
2148 MO_F_Ge W32 -> condFltCode GE x y
2149 MO_F_Lt W32 -> condFltCode LTT x y
2150 MO_F_Le W32 -> condFltCode LE x y
2152 MO_F_Eq W64 -> condFltCode EQQ x y
2153 MO_F_Ne W64 -> condFltCode NE x y
2154 MO_F_Gt W64 -> condFltCode GTT x y
2155 MO_F_Ge W64 -> condFltCode GE x y
2156 MO_F_Lt W64 -> condFltCode LTT x y
2157 MO_F_Le W64 -> condFltCode LE x y
2159 MO_Eq rep -> condIntCode EQQ (extendUExpr rep x) (extendUExpr rep y)
2160 MO_Ne rep -> condIntCode NE (extendUExpr rep x) (extendUExpr rep y)
2162 MO_S_Gt rep -> condIntCode GTT (extendSExpr rep x) (extendSExpr rep y)
2163 MO_S_Ge rep -> condIntCode GE (extendSExpr rep x) (extendSExpr rep y)
2164 MO_S_Lt rep -> condIntCode LTT (extendSExpr rep x) (extendSExpr rep y)
2165 MO_S_Le rep -> condIntCode LE (extendSExpr rep x) (extendSExpr rep y)
2167 MO_U_Gt rep -> condIntCode GU (extendUExpr rep x) (extendUExpr rep y)
2168 MO_U_Ge rep -> condIntCode GEU (extendUExpr rep x) (extendUExpr rep y)
2169 MO_U_Lt rep -> condIntCode LU (extendUExpr rep x) (extendUExpr rep y)
2170 MO_U_Le rep -> condIntCode LEU (extendUExpr rep x) (extendUExpr rep y)
2172 other -> pprPanic "getCondCode(powerpc)" (pprMachOp mop)
2174 getCondCode other = panic "getCondCode(2)(powerpc)"
2180 -- @cond(Int|Flt)Code@: Turn a boolean expression into a condition, to be
2181 -- passed back up the tree.
2183 condIntCode, condFltCode :: Cond -> CmmExpr -> CmmExpr -> NatM CondCode
2185 #if alpha_TARGET_ARCH
2186 condIntCode = panic "MachCode.condIntCode: not on Alphas"
2187 condFltCode = panic "MachCode.condFltCode: not on Alphas"
2188 #endif /* alpha_TARGET_ARCH */
2190 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2191 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
2193 -- memory vs immediate
2194 condIntCode cond (CmmLoad x pk) (CmmLit lit) | is32BitLit lit = do
2195 Amode x_addr x_code <- getAmode x
2198 code = x_code `snocOL`
2199 CMP (cmmTypeSize pk) (OpImm imm) (OpAddr x_addr)
2201 return (CondCode False cond code)
2203 -- anything vs zero, using a mask
2204 -- TODO: Add some sanity checking!!!!
2205 condIntCode cond (CmmMachOp (MO_And rep) [x,o2]) (CmmLit (CmmInt 0 pk))
2206 | (CmmLit lit@(CmmInt mask pk2)) <- o2, is32BitLit lit
2208 (x_reg, x_code) <- getSomeReg x
2210 code = x_code `snocOL`
2211 TEST (intSize pk) (OpImm (ImmInteger mask)) (OpReg x_reg)
2213 return (CondCode False cond code)
2216 condIntCode cond x (CmmLit (CmmInt 0 pk)) = do
2217 (x_reg, x_code) <- getSomeReg x
2219 code = x_code `snocOL`
2220 TEST (intSize pk) (OpReg x_reg) (OpReg x_reg)
2222 return (CondCode False cond code)
2224 -- anything vs operand
2225 condIntCode cond x y | isOperand y = do
2226 (x_reg, x_code) <- getNonClobberedReg x
2227 (y_op, y_code) <- getOperand y
2229 code = x_code `appOL` y_code `snocOL`
2230 CMP (cmmTypeSize (cmmExprType x)) y_op (OpReg x_reg)
2232 return (CondCode False cond code)
2234 -- anything vs anything
2235 condIntCode cond x y = do
2236 (y_reg, y_code) <- getNonClobberedReg y
2237 (x_op, x_code) <- getRegOrMem x
2239 code = y_code `appOL`
2241 CMP (cmmTypeSize (cmmExprType x)) (OpReg y_reg) x_op
2243 return (CondCode False cond code)
2246 #if i386_TARGET_ARCH
2247 condFltCode cond x y
2248 = ASSERT(cond `elem` ([EQQ, NE, LE, LTT, GE, GTT])) do
2249 (x_reg, x_code) <- getNonClobberedReg x
2250 (y_reg, y_code) <- getSomeReg y
2252 code = x_code `appOL` y_code `snocOL`
2253 GCMP cond x_reg y_reg
2254 -- The GCMP insn does the test and sets the zero flag if comparable
2255 -- and true. Hence we always supply EQQ as the condition to test.
2256 return (CondCode True EQQ code)
2257 #endif /* i386_TARGET_ARCH */
2259 #if x86_64_TARGET_ARCH
2260 -- in the SSE2 comparison ops (ucomiss, ucomisd) the left arg may be
2261 -- an operand, but the right must be a reg. We can probably do better
2262 -- than this general case...
2263 condFltCode cond x y = do
2264 (x_reg, x_code) <- getNonClobberedReg x
2265 (y_op, y_code) <- getOperand y
2267 code = x_code `appOL`
2269 CMP (floatSize $ cmmExprWidth x) y_op (OpReg x_reg)
2270 -- NB(1): we need to use the unsigned comparison operators on the
2271 -- result of this comparison.
2273 return (CondCode True (condToUnsigned cond) code)
2276 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2278 #if sparc_TARGET_ARCH
2280 condIntCode cond x (CmmLit (CmmInt y rep))
2283 (src1, code) <- getSomeReg x
2285 src2 = ImmInt (fromInteger y)
2286 code' = code `snocOL` SUB False True src1 (RIImm src2) g0
2287 return (CondCode False cond code')
2289 condIntCode cond x y = do
2290 (src1, code1) <- getSomeReg x
2291 (src2, code2) <- getSomeReg y
2293 code__2 = code1 `appOL` code2 `snocOL`
2294 SUB False True src1 (RIReg src2) g0
2295 return (CondCode False cond code__2)
2298 condFltCode cond x y = do
2299 (src1, code1) <- getSomeReg x
2300 (src2, code2) <- getSomeReg y
2301 tmp <- getNewRegNat FF64
2303 promote x = FxTOy FF32 FF64 x tmp
2310 code1 `appOL` code2 `snocOL`
2311 FCMP True pk1 src1 src2
2312 else if typeWidth pk1 == W32 then
2313 code1 `snocOL` promote src1 `appOL` code2 `snocOL`
2314 FCMP True FF64 tmp src2
2316 code1 `appOL` code2 `snocOL` promote src2 `snocOL`
2317 FCMP True FF64 src1 tmp
2318 return (CondCode True cond code__2)
2320 #endif /* sparc_TARGET_ARCH */
2322 #if powerpc_TARGET_ARCH
2323 -- ###FIXME: I16 and I8!
2324 condIntCode cond x (CmmLit (CmmInt y rep))
2325 | Just src2 <- makeImmediate rep (not $ condUnsigned cond) y
2327 (src1, code) <- getSomeReg x
2329 code' = code `snocOL`
2330 (if condUnsigned cond then CMPL else CMP) II32 src1 (RIImm src2)
2331 return (CondCode False cond code')
2333 condIntCode cond x y = do
2334 (src1, code1) <- getSomeReg x
2335 (src2, code2) <- getSomeReg y
2337 code' = code1 `appOL` code2 `snocOL`
2338 (if condUnsigned cond then CMPL else CMP) II32 src1 (RIReg src2)
2339 return (CondCode False cond code')
2341 condFltCode cond x y = do
2342 (src1, code1) <- getSomeReg x
2343 (src2, code2) <- getSomeReg y
2345 code' = code1 `appOL` code2 `snocOL` FCMP src1 src2
2346 code'' = case cond of -- twiddle CR to handle unordered case
2347 GE -> code' `snocOL` CRNOR ltbit eqbit gtbit
2348 LE -> code' `snocOL` CRNOR gtbit eqbit ltbit
2351 ltbit = 0 ; eqbit = 2 ; gtbit = 1
2352 return (CondCode True cond code'')
2354 #endif /* powerpc_TARGET_ARCH */
2356 -- -----------------------------------------------------------------------------
2357 -- Generating assignments
2359 -- Assignments are really at the heart of the whole code generation
2360 -- business. Almost all top-level nodes of any real importance are
2361 -- assignments, which correspond to loads, stores, or register
2362 -- transfers. If we're really lucky, some of the register transfers
2363 -- will go away, because we can use the destination register to
2364 -- complete the code generation for the right hand side. This only
2365 -- fails when the right hand side is forced into a fixed register
2366 -- (e.g. the result of a call).
2368 assignMem_IntCode :: Size -> CmmExpr -> CmmExpr -> NatM InstrBlock
2369 assignReg_IntCode :: Size -> CmmReg -> CmmExpr -> NatM InstrBlock
2371 assignMem_FltCode :: Size -> CmmExpr -> CmmExpr -> NatM InstrBlock
2372 assignReg_FltCode :: Size -> CmmReg -> CmmExpr -> NatM InstrBlock
2374 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2376 #if alpha_TARGET_ARCH
2378 assignIntCode pk (CmmLoad dst _) src
2379 = getNewRegNat IntRep `thenNat` \ tmp ->
2380 getAmode dst `thenNat` \ amode ->
2381 getRegister src `thenNat` \ register ->
2383 code1 = amodeCode amode []
2384 dst__2 = amodeAddr amode
2385 code2 = registerCode register tmp []
2386 src__2 = registerName register tmp
2387 sz = primRepToSize pk
2388 code__2 = asmSeqThen [code1, code2] . mkSeqInstr (ST sz src__2 dst__2)
2392 assignIntCode pk dst src
2393 = getRegister dst `thenNat` \ register1 ->
2394 getRegister src `thenNat` \ register2 ->
2396 dst__2 = registerName register1 zeroh
2397 code = registerCode register2 dst__2
2398 src__2 = registerName register2 dst__2
2399 code__2 = if isFixed register2
2400 then code . mkSeqInstr (OR src__2 (RIReg src__2) dst__2)
2405 #endif /* alpha_TARGET_ARCH */
2407 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2409 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
2411 -- integer assignment to memory
2413 -- specific case of adding/subtracting an integer to a particular address.
2414 -- ToDo: catch other cases where we can use an operation directly on a memory
2416 assignMem_IntCode pk addr (CmmMachOp op [CmmLoad addr2 _,
2417 CmmLit (CmmInt i _)])
2418 | addr == addr2, pk /= II64 || is32BitInteger i,
2419 Just instr <- check op
2420 = do Amode amode code_addr <- getAmode addr
2421 let code = code_addr `snocOL`
2422 instr pk (OpImm (ImmInt (fromIntegral i))) (OpAddr amode)
2425 check (MO_Add _) = Just ADD
2426 check (MO_Sub _) = Just SUB
2431 assignMem_IntCode pk addr src = do
2432 Amode addr code_addr <- getAmode addr
2433 (code_src, op_src) <- get_op_RI src
2435 code = code_src `appOL`
2437 MOV pk op_src (OpAddr addr)
2438 -- NOTE: op_src is stable, so it will still be valid
2439 -- after code_addr. This may involve the introduction
2440 -- of an extra MOV to a temporary register, but we hope
2441 -- the register allocator will get rid of it.
2445 get_op_RI :: CmmExpr -> NatM (InstrBlock,Operand) -- code, operator
2446 get_op_RI (CmmLit lit) | is32BitLit lit
2447 = return (nilOL, OpImm (litToImm lit))
2449 = do (reg,code) <- getNonClobberedReg op
2450 return (code, OpReg reg)
2453 -- Assign; dst is a reg, rhs is mem
2454 assignReg_IntCode pk reg (CmmLoad src _) = do
2455 load_code <- intLoadCode (MOV pk) src
2456 return (load_code (getRegisterReg reg))
2458 -- dst is a reg, but src could be anything
2459 assignReg_IntCode pk reg src = do
2460 code <- getAnyReg src
2461 return (code (getRegisterReg reg))
2463 #endif /* i386_TARGET_ARCH */
2465 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2467 #if sparc_TARGET_ARCH
2469 assignMem_IntCode pk addr src = do
2470 (srcReg, code) <- getSomeReg src
2471 Amode dstAddr addr_code <- getAmode addr
2472 return $ code `appOL` addr_code `snocOL` ST pk srcReg dstAddr
2474 assignReg_IntCode pk reg src = do
2475 r <- getRegister src
2477 Any _ code -> code dst
2478 Fixed _ freg fcode -> fcode `snocOL` OR False g0 (RIReg dst) freg
2480 dst = getRegisterReg reg
2483 #endif /* sparc_TARGET_ARCH */
2485 #if powerpc_TARGET_ARCH
2487 assignMem_IntCode pk addr src = do
2488 (srcReg, code) <- getSomeReg src
2489 Amode dstAddr addr_code <- getAmode addr
2490 return $ code `appOL` addr_code `snocOL` ST pk srcReg dstAddr
2492 -- dst is a reg, but src could be anything
2493 assignReg_IntCode pk reg src
2495 r <- getRegister src
2497 Any _ code -> code dst
2498 Fixed _ freg fcode -> fcode `snocOL` MR dst freg
2500 dst = getRegisterReg reg
2502 #endif /* powerpc_TARGET_ARCH */
2505 -- -----------------------------------------------------------------------------
2506 -- Floating-point assignments
2508 #if alpha_TARGET_ARCH
2510 assignFltCode pk (CmmLoad dst _) src
2511 = getNewRegNat pk `thenNat` \ tmp ->
2512 getAmode dst `thenNat` \ amode ->
2513 getRegister src `thenNat` \ register ->
2515 code1 = amodeCode amode []
2516 dst__2 = amodeAddr amode
2517 code2 = registerCode register tmp []
2518 src__2 = registerName register tmp
2519 sz = primRepToSize pk
2520 code__2 = asmSeqThen [code1, code2] . mkSeqInstr (ST sz src__2 dst__2)
2524 assignFltCode pk dst src
2525 = getRegister dst `thenNat` \ register1 ->
2526 getRegister src `thenNat` \ register2 ->
2528 dst__2 = registerName register1 zeroh
2529 code = registerCode register2 dst__2
2530 src__2 = registerName register2 dst__2
2531 code__2 = if isFixed register2
2532 then code . mkSeqInstr (FMOV src__2 dst__2)
2537 #endif /* alpha_TARGET_ARCH */
2539 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2541 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
2543 -- Floating point assignment to memory
2544 assignMem_FltCode pk addr src = do
2545 (src_reg, src_code) <- getNonClobberedReg src
2546 Amode addr addr_code <- getAmode addr
2548 code = src_code `appOL`
2550 IF_ARCH_i386(GST pk src_reg addr,
2551 MOV pk (OpReg src_reg) (OpAddr addr))
2554 -- Floating point assignment to a register/temporary
2555 assignReg_FltCode pk reg src = do
2556 src_code <- getAnyReg src
2557 return (src_code (getRegisterReg reg))
2559 #endif /* i386_TARGET_ARCH */
2561 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2563 #if sparc_TARGET_ARCH
2565 -- Floating point assignment to memory
2566 assignMem_FltCode pk addr src = do
2567 Amode dst__2 code1 <- getAmode addr
2568 (src__2, code2) <- getSomeReg src
2569 tmp1 <- getNewRegNat pk
2571 pk__2 = cmmExprType src
2572 code__2 = code1 `appOL` code2 `appOL`
2574 then unitOL (ST pk src__2 dst__2)
2575 else toOL [FxTOy pk__2 pk src__2 tmp1, ST pk tmp1 dst__2]
2578 -- Floating point assignment to a register/temporary
2579 -- ToDo: Verify correctness
2580 assignReg_FltCode pk reg src = do
2581 r <- getRegister src
2582 v1 <- getNewRegNat pk
2584 Any _ code -> code dst
2585 Fixed _ freg fcode -> fcode `snocOL` FMOV pk freg v1
2587 dst = getRegisterReg reg
2589 #endif /* sparc_TARGET_ARCH */
2591 #if powerpc_TARGET_ARCH
2594 assignMem_FltCode = assignMem_IntCode
2595 assignReg_FltCode = assignReg_IntCode
2597 #endif /* powerpc_TARGET_ARCH */
2600 -- -----------------------------------------------------------------------------
2601 -- Generating an non-local jump
2603 -- (If applicable) Do not fill the delay slots here; you will confuse the
2604 -- register allocator.
2606 genJump :: CmmExpr{-the branch target-} -> NatM InstrBlock
2608 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2610 #if alpha_TARGET_ARCH
2612 genJump (CmmLabel lbl)
2613 | isAsmTemp lbl = returnInstr (BR target)
2614 | otherwise = returnInstrs [LDA pv (AddrImm target), JMP zeroh (AddrReg pv) 0]
2616 target = ImmCLbl lbl
2619 = getRegister tree `thenNat` \ register ->
2620 getNewRegNat PtrRep `thenNat` \ tmp ->
2622 dst = registerName register pv
2623 code = registerCode register pv
2624 target = registerName register pv
2626 if isFixed register then
2627 returnSeq code [OR dst (RIReg dst) pv, JMP zeroh (AddrReg pv) 0]
2629 return (code . mkSeqInstr (JMP zeroh (AddrReg pv) 0))
2631 #endif /* alpha_TARGET_ARCH */
2633 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2635 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
2637 genJump (CmmLoad mem pk) = do
2638 Amode target code <- getAmode mem
2639 return (code `snocOL` JMP (OpAddr target))
2641 genJump (CmmLit lit) = do
2642 return (unitOL (JMP (OpImm (litToImm lit))))
2645 (reg,code) <- getSomeReg expr
2646 return (code `snocOL` JMP (OpReg reg))
2648 #endif /* i386_TARGET_ARCH */
2650 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2652 #if sparc_TARGET_ARCH
2654 genJump (CmmLit (CmmLabel lbl))
2655 = return (toOL [CALL (Left target) 0 True, NOP])
2657 target = ImmCLbl lbl
2661 (target, code) <- getSomeReg tree
2662 return (code `snocOL` JMP (AddrRegReg target g0) `snocOL` NOP)
2664 #endif /* sparc_TARGET_ARCH */
2666 #if powerpc_TARGET_ARCH
2667 genJump (CmmLit (CmmLabel lbl))
2668 = return (unitOL $ JMP lbl)
2672 (target,code) <- getSomeReg tree
2673 return (code `snocOL` MTCTR target `snocOL` BCTR [])
2674 #endif /* powerpc_TARGET_ARCH */
2677 -- -----------------------------------------------------------------------------
2678 -- Unconditional branches
2680 genBranch :: BlockId -> NatM InstrBlock
2682 genBranch = return . toOL . mkBranchInstr
2684 -- -----------------------------------------------------------------------------
2685 -- Conditional jumps
2688 Conditional jumps are always to local labels, so we can use branch
2689 instructions. We peek at the arguments to decide what kind of
2692 ALPHA: For comparisons with 0, we're laughing, because we can just do
2693 the desired conditional branch.
2695 I386: First, we have to ensure that the condition
2696 codes are set according to the supplied comparison operation.
2698 SPARC: First, we have to ensure that the condition codes are set
2699 according to the supplied comparison operation. We generate slightly
2700 different code for floating point comparisons, because a floating
2701 point operation cannot directly precede a @BF@. We assume the worst
2702 and fill that slot with a @NOP@.
2704 SPARC: Do not fill the delay slots here; you will confuse the register
2710 :: BlockId -- the branch target
2711 -> CmmExpr -- the condition on which to branch
2714 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2716 #if alpha_TARGET_ARCH
2718 genCondJump id (StPrim op [x, StInt 0])
2719 = getRegister x `thenNat` \ register ->
2720 getNewRegNat (registerRep register)
2723 code = registerCode register tmp
2724 value = registerName register tmp
2725 pk = registerRep register
2726 target = ImmCLbl lbl
2728 returnSeq code [BI (cmpOp op) value target]
2730 cmpOp CharGtOp = GTT
2732 cmpOp CharEqOp = EQQ
2734 cmpOp CharLtOp = LTT
2743 cmpOp WordGeOp = ALWAYS
2744 cmpOp WordEqOp = EQQ
2746 cmpOp WordLtOp = NEVER
2747 cmpOp WordLeOp = EQQ
2749 cmpOp AddrGeOp = ALWAYS
2750 cmpOp AddrEqOp = EQQ
2752 cmpOp AddrLtOp = NEVER
2753 cmpOp AddrLeOp = EQQ
2755 genCondJump lbl (StPrim op [x, StDouble 0.0])
2756 = getRegister x `thenNat` \ register ->
2757 getNewRegNat (registerRep register)
2760 code = registerCode register tmp
2761 value = registerName register tmp
2762 pk = registerRep register
2763 target = ImmCLbl lbl
2765 return (code . mkSeqInstr (BF (cmpOp op) value target))
2767 cmpOp FloatGtOp = GTT
2768 cmpOp FloatGeOp = GE
2769 cmpOp FloatEqOp = EQQ
2770 cmpOp FloatNeOp = NE
2771 cmpOp FloatLtOp = LTT
2772 cmpOp FloatLeOp = LE
2773 cmpOp DoubleGtOp = GTT
2774 cmpOp DoubleGeOp = GE
2775 cmpOp DoubleEqOp = EQQ
2776 cmpOp DoubleNeOp = NE
2777 cmpOp DoubleLtOp = LTT
2778 cmpOp DoubleLeOp = LE
2780 genCondJump lbl (StPrim op [x, y])
2782 = trivialFCode pr instr x y `thenNat` \ register ->
2783 getNewRegNat FF64 `thenNat` \ tmp ->
2785 code = registerCode register tmp
2786 result = registerName register tmp
2787 target = ImmCLbl lbl
2789 return (code . mkSeqInstr (BF cond result target))
2791 pr = panic "trivialU?FCode: does not use PrimRep on Alpha"
2793 fltCmpOp op = case op of
2807 (instr, cond) = case op of
2808 FloatGtOp -> (FCMP TF LE, EQQ)
2809 FloatGeOp -> (FCMP TF LTT, EQQ)
2810 FloatEqOp -> (FCMP TF EQQ, NE)
2811 FloatNeOp -> (FCMP TF EQQ, EQQ)
2812 FloatLtOp -> (FCMP TF LTT, NE)
2813 FloatLeOp -> (FCMP TF LE, NE)
2814 DoubleGtOp -> (FCMP TF LE, EQQ)
2815 DoubleGeOp -> (FCMP TF LTT, EQQ)
2816 DoubleEqOp -> (FCMP TF EQQ, NE)
2817 DoubleNeOp -> (FCMP TF EQQ, EQQ)
2818 DoubleLtOp -> (FCMP TF LTT, NE)
2819 DoubleLeOp -> (FCMP TF LE, NE)
2821 genCondJump lbl (StPrim op [x, y])
2822 = trivialCode instr x y `thenNat` \ register ->
2823 getNewRegNat IntRep `thenNat` \ tmp ->
2825 code = registerCode register tmp
2826 result = registerName register tmp
2827 target = ImmCLbl lbl
2829 return (code . mkSeqInstr (BI cond result target))
2831 (instr, cond) = case op of
2832 CharGtOp -> (CMP LE, EQQ)
2833 CharGeOp -> (CMP LTT, EQQ)
2834 CharEqOp -> (CMP EQQ, NE)
2835 CharNeOp -> (CMP EQQ, EQQ)
2836 CharLtOp -> (CMP LTT, NE)
2837 CharLeOp -> (CMP LE, NE)
2838 IntGtOp -> (CMP LE, EQQ)
2839 IntGeOp -> (CMP LTT, EQQ)
2840 IntEqOp -> (CMP EQQ, NE)
2841 IntNeOp -> (CMP EQQ, EQQ)
2842 IntLtOp -> (CMP LTT, NE)
2843 IntLeOp -> (CMP LE, NE)
2844 WordGtOp -> (CMP ULE, EQQ)
2845 WordGeOp -> (CMP ULT, EQQ)
2846 WordEqOp -> (CMP EQQ, NE)
2847 WordNeOp -> (CMP EQQ, EQQ)
2848 WordLtOp -> (CMP ULT, NE)
2849 WordLeOp -> (CMP ULE, NE)
2850 AddrGtOp -> (CMP ULE, EQQ)
2851 AddrGeOp -> (CMP ULT, EQQ)
2852 AddrEqOp -> (CMP EQQ, NE)
2853 AddrNeOp -> (CMP EQQ, EQQ)
2854 AddrLtOp -> (CMP ULT, NE)
2855 AddrLeOp -> (CMP ULE, NE)
2857 #endif /* alpha_TARGET_ARCH */
2859 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2861 #if i386_TARGET_ARCH
2863 genCondJump id bool = do
2864 CondCode _ cond code <- getCondCode bool
2865 return (code `snocOL` JXX cond id)
2869 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2871 #if x86_64_TARGET_ARCH
2873 genCondJump id bool = do
2874 CondCode is_float cond cond_code <- getCondCode bool
2877 return (cond_code `snocOL` JXX cond id)
2879 lbl <- getBlockIdNat
2881 -- see comment with condFltReg
2882 let code = case cond of
2888 plain_test = unitOL (
2891 or_unordered = toOL [
2895 and_ordered = toOL [
2901 return (cond_code `appOL` code)
2905 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2907 #if sparc_TARGET_ARCH
2909 genCondJump (BlockId id) bool = do
2910 CondCode is_float cond code <- getCondCode bool
2915 then [NOP, BF cond False (ImmCLbl (mkAsmTempLabel id)), NOP]
2916 else [BI cond False (ImmCLbl (mkAsmTempLabel id)), NOP]
2920 #endif /* sparc_TARGET_ARCH */
2923 #if powerpc_TARGET_ARCH
2925 genCondJump id bool = do
2926 CondCode is_float cond code <- getCondCode bool
2927 return (code `snocOL` BCC cond id)
2929 #endif /* powerpc_TARGET_ARCH */
2932 -- -----------------------------------------------------------------------------
2933 -- Generating C calls
2935 -- Now the biggest nightmare---calls. Most of the nastiness is buried in
2936 -- @get_arg@, which moves the arguments to the correct registers/stack
2937 -- locations. Apart from that, the code is easy.
2939 -- (If applicable) Do not fill the delay slots here; you will confuse the
2940 -- register allocator.
2943 :: CmmCallTarget -- function to call
2944 -> HintedCmmFormals -- where to put the result
2945 -> HintedCmmActuals -- arguments (of mixed type)
2948 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
2950 #if alpha_TARGET_ARCH
2954 genCCall fn cconv result_regs args
2955 = mapAccumLNat get_arg (allArgRegs, eXTRA_STK_ARGS_HERE) args
2956 `thenNat` \ ((unused,_), argCode) ->
2958 nRegs = length allArgRegs - length unused
2959 code = asmSeqThen (map ($ []) argCode)
2962 LDA pv (AddrImm (ImmLab (ptext fn))),
2963 JSR ra (AddrReg pv) nRegs,
2964 LDGP gp (AddrReg ra)]
2966 ------------------------
2967 {- Try to get a value into a specific register (or registers) for
2968 a call. The first 6 arguments go into the appropriate
2969 argument register (separate registers for integer and floating
2970 point arguments, but used in lock-step), and the remaining
2971 arguments are dumped to the stack, beginning at 0(sp). Our
2972 first argument is a pair of the list of remaining argument
2973 registers to be assigned for this call and the next stack
2974 offset to use for overflowing arguments. This way,
2975 @get_Arg@ can be applied to all of a call's arguments using
2979 :: ([(Reg,Reg)], Int) -- Argument registers and stack offset (accumulator)
2980 -> StixTree -- Current argument
2981 -> NatM (([(Reg,Reg)],Int), InstrBlock) -- Updated accumulator and code
2983 -- We have to use up all of our argument registers first...
2985 get_arg ((iDst,fDst):dsts, offset) arg
2986 = getRegister arg `thenNat` \ register ->
2988 reg = if isFloatType pk then fDst else iDst
2989 code = registerCode register reg
2990 src = registerName register reg
2991 pk = registerRep register
2994 if isFloatType pk then
2995 ((dsts, offset), if isFixed register then
2996 code . mkSeqInstr (FMOV src fDst)
2999 ((dsts, offset), if isFixed register then
3000 code . mkSeqInstr (OR src (RIReg src) iDst)
3003 -- Once we have run out of argument registers, we move to the
3006 get_arg ([], offset) arg
3007 = getRegister arg `thenNat` \ register ->
3008 getNewRegNat (registerRep register)
3011 code = registerCode register tmp
3012 src = registerName register tmp
3013 pk = registerRep register
3014 sz = primRepToSize pk
3016 return (([], offset + 1), code . mkSeqInstr (ST sz src (spRel offset)))
3018 #endif /* alpha_TARGET_ARCH */
3020 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
3022 #if i386_TARGET_ARCH
3024 genCCall (CmmPrim MO_WriteBarrier) _ _ = return nilOL
3025 -- write barrier compiles to no code on x86/x86-64;
3026 -- we keep it this long in order to prevent earlier optimisations.
3028 -- we only cope with a single result for foreign calls
3029 genCCall (CmmPrim op) [CmmHinted r _] args = do
3030 l1 <- getNewLabelNat
3031 l2 <- getNewLabelNat
3033 MO_F32_Sqrt -> actuallyInlineFloatOp GSQRT FF32 args
3034 MO_F64_Sqrt -> actuallyInlineFloatOp GSQRT FF64 args
3036 MO_F32_Sin -> actuallyInlineFloatOp (\s -> GSIN s l1 l2) FF32 args
3037 MO_F64_Sin -> actuallyInlineFloatOp (\s -> GSIN s l1 l2) FF64 args
3039 MO_F32_Cos -> actuallyInlineFloatOp (\s -> GCOS s l1 l2) FF32 args
3040 MO_F64_Cos -> actuallyInlineFloatOp (\s -> GCOS s l1 l2) FF64 args
3042 MO_F32_Tan -> actuallyInlineFloatOp (\s -> GTAN s l1 l2) FF32 args
3043 MO_F64_Tan -> actuallyInlineFloatOp (\s -> GTAN s l1 l2) FF64 args
3045 other_op -> outOfLineFloatOp op r args
3047 actuallyInlineFloatOp instr size [CmmHinted x _]
3048 = do res <- trivialUFCode size (instr size) x
3050 return (any (getRegisterReg (CmmLocal r)))
3052 genCCall target dest_regs args = do
3054 sizes = map (arg_size . cmmExprType . hintlessCmm) (reverse args)
3055 #if !darwin_TARGET_OS
3056 tot_arg_size = sum sizes
3058 raw_arg_size = sum sizes
3059 tot_arg_size = roundTo 16 raw_arg_size
3060 arg_pad_size = tot_arg_size - raw_arg_size
3061 delta0 <- getDeltaNat
3062 setDeltaNat (delta0 - arg_pad_size)
3065 push_codes <- mapM push_arg (reverse args)
3066 delta <- getDeltaNat
3069 -- deal with static vs dynamic call targets
3070 (callinsns,cconv) <-
3073 CmmCallee (CmmLit (CmmLabel lbl)) conv
3074 -> -- ToDo: stdcall arg sizes
3075 return (unitOL (CALL (Left fn_imm) []), conv)
3076 where fn_imm = ImmCLbl lbl
3078 -> do { (dyn_c, dyn_r) <- get_op expr
3079 ; ASSERT( isWord32 (cmmExprType expr) )
3080 return (dyn_c `snocOL` CALL (Right dyn_r) [], conv) }
3083 #if darwin_TARGET_OS
3085 = toOL [SUB II32 (OpImm (ImmInt arg_pad_size)) (OpReg esp),
3086 DELTA (delta0 - arg_pad_size)]
3087 `appOL` concatOL push_codes
3090 = concatOL push_codes
3091 call = callinsns `appOL`
3093 -- Deallocate parameters after call for ccall;
3094 -- but not for stdcall (callee does it)
3095 (if cconv == StdCallConv || tot_arg_size==0 then [] else
3096 [ADD II32 (OpImm (ImmInt tot_arg_size)) (OpReg esp)])
3098 [DELTA (delta + tot_arg_size)]
3101 setDeltaNat (delta + tot_arg_size)
3104 -- assign the results, if necessary
3105 assign_code [] = nilOL
3106 assign_code [CmmHinted dest _hint]
3107 | isFloatType ty = unitOL (GMOV fake0 r_dest)
3108 | isWord64 ty = toOL [MOV II32 (OpReg eax) (OpReg r_dest),
3109 MOV II32 (OpReg edx) (OpReg r_dest_hi)]
3110 | otherwise = unitOL (MOV (intSize w) (OpReg eax) (OpReg r_dest))
3112 ty = localRegType dest
3114 r_dest_hi = getHiVRegFromLo r_dest
3115 r_dest = getRegisterReg (CmmLocal dest)
3116 assign_code many = panic "genCCall.assign_code many"
3118 return (push_code `appOL`
3120 assign_code dest_regs)
3123 arg_size :: CmmType -> Int -- Width in bytes
3124 arg_size ty = widthInBytes (typeWidth ty)
3126 roundTo a x | x `mod` a == 0 = x
3127 | otherwise = x + a - (x `mod` a)
3130 push_arg :: HintedCmmActual {-current argument-}
3131 -> NatM InstrBlock -- code
3133 push_arg (CmmHinted arg _hint) -- we don't need the hints on x86
3134 | isWord64 arg_ty = do
3135 ChildCode64 code r_lo <- iselExpr64 arg
3136 delta <- getDeltaNat
3137 setDeltaNat (delta - 8)
3139 r_hi = getHiVRegFromLo r_lo
3141 return ( code `appOL`
3142 toOL [PUSH II32 (OpReg r_hi), DELTA (delta - 4),
3143 PUSH II32 (OpReg r_lo), DELTA (delta - 8),
3148 (code, reg) <- get_op arg
3149 delta <- getDeltaNat
3150 let size = arg_size arg_ty -- Byte size
3151 setDeltaNat (delta-size)
3152 if (isFloatType arg_ty)
3153 then return (code `appOL`
3154 toOL [SUB II32 (OpImm (ImmInt size)) (OpReg esp),
3156 GST (floatSize (typeWidth arg_ty))
3157 reg (AddrBaseIndex (EABaseReg esp)
3161 else return (code `snocOL`
3162 PUSH II32 (OpReg reg) `snocOL`
3166 arg_ty = cmmExprType arg
3169 get_op :: CmmExpr -> NatM (InstrBlock, Reg) -- code, reg
3171 (reg,code) <- getSomeReg op
3174 #endif /* i386_TARGET_ARCH */
3176 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
3178 outOfLineFloatOp :: CallishMachOp -> CmmFormal -> HintedCmmActuals
3180 outOfLineFloatOp mop res args
3182 dflags <- getDynFlagsNat
3183 targetExpr <- cmmMakeDynamicReference dflags addImportNat CallReference lbl
3184 let target = CmmCallee targetExpr CCallConv
3186 if isFloat64 (localRegType res)
3188 stmtToInstrs (CmmCall target [CmmHinted res NoHint] args CmmUnsafe CmmMayReturn)
3192 tmp = LocalReg uq f64
3194 code1 <- stmtToInstrs (CmmCall target [CmmHinted tmp NoHint] args CmmUnsafe CmmMayReturn)
3195 code2 <- stmtToInstrs (CmmAssign (CmmLocal res) (CmmReg (CmmLocal tmp)))
3196 return (code1 `appOL` code2)
3198 lbl = mkForeignLabel fn Nothing False
3201 MO_F32_Sqrt -> fsLit "sqrtf"
3202 MO_F32_Sin -> fsLit "sinf"
3203 MO_F32_Cos -> fsLit "cosf"
3204 MO_F32_Tan -> fsLit "tanf"
3205 MO_F32_Exp -> fsLit "expf"
3206 MO_F32_Log -> fsLit "logf"
3208 MO_F32_Asin -> fsLit "asinf"
3209 MO_F32_Acos -> fsLit "acosf"
3210 MO_F32_Atan -> fsLit "atanf"
3212 MO_F32_Sinh -> fsLit "sinhf"
3213 MO_F32_Cosh -> fsLit "coshf"
3214 MO_F32_Tanh -> fsLit "tanhf"
3215 MO_F32_Pwr -> fsLit "powf"
3217 MO_F64_Sqrt -> fsLit "sqrt"
3218 MO_F64_Sin -> fsLit "sin"
3219 MO_F64_Cos -> fsLit "cos"
3220 MO_F64_Tan -> fsLit "tan"
3221 MO_F64_Exp -> fsLit "exp"
3222 MO_F64_Log -> fsLit "log"
3224 MO_F64_Asin -> fsLit "asin"
3225 MO_F64_Acos -> fsLit "acos"
3226 MO_F64_Atan -> fsLit "atan"
3228 MO_F64_Sinh -> fsLit "sinh"
3229 MO_F64_Cosh -> fsLit "cosh"
3230 MO_F64_Tanh -> fsLit "tanh"
3231 MO_F64_Pwr -> fsLit "pow"
3233 #endif /* i386_TARGET_ARCH || x86_64_TARGET_ARCH */
3235 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
3237 #if x86_64_TARGET_ARCH
3239 genCCall (CmmPrim MO_WriteBarrier) _ _ = return nilOL
3240 -- write barrier compiles to no code on x86/x86-64;
3241 -- we keep it this long in order to prevent earlier optimisations.
3244 genCCall (CmmPrim op) [CmmHinted r _] args =
3245 outOfLineFloatOp op r args
3247 genCCall target dest_regs args = do
3249 -- load up the register arguments
3250 (stack_args, aregs, fregs, load_args_code)
3251 <- load_args args allArgRegs allFPArgRegs nilOL
3254 fp_regs_used = reverse (drop (length fregs) (reverse allFPArgRegs))
3255 int_regs_used = reverse (drop (length aregs) (reverse allArgRegs))
3256 arg_regs = [eax] ++ int_regs_used ++ fp_regs_used
3257 -- for annotating the call instruction with
3259 sse_regs = length fp_regs_used
3261 tot_arg_size = arg_size * length stack_args
3263 -- On entry to the called function, %rsp should be aligned
3264 -- on a 16-byte boundary +8 (i.e. the first stack arg after
3265 -- the return address is 16-byte aligned). In STG land
3266 -- %rsp is kept 16-byte aligned (see StgCRun.c), so we just
3267 -- need to make sure we push a multiple of 16-bytes of args,
3268 -- plus the return address, to get the correct alignment.
3269 -- Urg, this is hard. We need to feed the delta back into
3270 -- the arg pushing code.
3271 (real_size, adjust_rsp) <-
3272 if tot_arg_size `rem` 16 == 0
3273 then return (tot_arg_size, nilOL)
3274 else do -- we need to adjust...
3275 delta <- getDeltaNat
3276 setDeltaNat (delta-8)
3277 return (tot_arg_size+8, toOL [
3278 SUB II64 (OpImm (ImmInt 8)) (OpReg rsp),
3282 -- push the stack args, right to left
3283 push_code <- push_args (reverse stack_args) nilOL
3284 delta <- getDeltaNat
3286 -- deal with static vs dynamic call targets
3287 (callinsns,cconv) <-
3290 CmmCallee (CmmLit (CmmLabel lbl)) conv
3291 -> -- ToDo: stdcall arg sizes
3292 return (unitOL (CALL (Left fn_imm) arg_regs), conv)
3293 where fn_imm = ImmCLbl lbl
3295 -> do (dyn_r, dyn_c) <- getSomeReg expr
3296 return (dyn_c `snocOL` CALL (Right dyn_r) arg_regs, conv)
3299 -- The x86_64 ABI requires us to set %al to the number of SSE
3300 -- registers that contain arguments, if the called routine
3301 -- is a varargs function. We don't know whether it's a
3302 -- varargs function or not, so we have to assume it is.
3304 -- It's not safe to omit this assignment, even if the number
3305 -- of SSE regs in use is zero. If %al is larger than 8
3306 -- on entry to a varargs function, seg faults ensue.
3307 assign_eax n = unitOL (MOV II32 (OpImm (ImmInt n)) (OpReg eax))
3309 let call = callinsns `appOL`
3311 -- Deallocate parameters after call for ccall;
3312 -- but not for stdcall (callee does it)
3313 (if cconv == StdCallConv || real_size==0 then [] else
3314 [ADD (intSize wordWidth) (OpImm (ImmInt real_size)) (OpReg esp)])
3316 [DELTA (delta + real_size)]
3319 setDeltaNat (delta + real_size)
3322 -- assign the results, if necessary
3323 assign_code [] = nilOL
3324 assign_code [CmmHinted dest _hint] =
3325 case typeWidth rep of
3326 W32 | isFloatType rep -> unitOL (MOV (floatSize W32) (OpReg xmm0) (OpReg r_dest))
3327 W64 | isFloatType rep -> unitOL (MOV (floatSize W32) (OpReg xmm0) (OpReg r_dest))
3328 _ -> unitOL (MOV (cmmTypeSize rep) (OpReg rax) (OpReg r_dest))
3330 rep = localRegType dest
3331 r_dest = getRegisterReg (CmmLocal dest)
3332 assign_code many = panic "genCCall.assign_code many"
3334 return (load_args_code `appOL`
3337 assign_eax sse_regs `appOL`
3339 assign_code dest_regs)
3342 arg_size = 8 -- always, at the mo
3344 load_args :: [CmmHinted CmmExpr]
3345 -> [Reg] -- int regs avail for args
3346 -> [Reg] -- FP regs avail for args
3348 -> NatM ([CmmHinted CmmExpr],[Reg],[Reg],InstrBlock)
3349 load_args args [] [] code = return (args, [], [], code)
3350 -- no more regs to use
3351 load_args [] aregs fregs code = return ([], aregs, fregs, code)
3352 -- no more args to push
3353 load_args ((CmmHinted arg hint) : rest) aregs fregs code
3354 | isFloatType arg_rep =
3358 arg_code <- getAnyReg arg
3359 load_args rest aregs rs (code `appOL` arg_code r)
3364 arg_code <- getAnyReg arg
3365 load_args rest rs fregs (code `appOL` arg_code r)
3367 arg_rep = cmmExprType arg
3370 (args',ars,frs,code') <- load_args rest aregs fregs code
3371 return ((CmmHinted arg hint):args', ars, frs, code')
3373 push_args [] code = return code
3374 push_args ((CmmHinted arg hint):rest) code
3375 | isFloatType arg_rep = do
3376 (arg_reg, arg_code) <- getSomeReg arg
3377 delta <- getDeltaNat
3378 setDeltaNat (delta-arg_size)
3379 let code' = code `appOL` arg_code `appOL` toOL [
3380 SUB (intSize wordWidth) (OpImm (ImmInt arg_size)) (OpReg rsp) ,
3381 DELTA (delta-arg_size),
3382 MOV (floatSize width) (OpReg arg_reg) (OpAddr (spRel 0))]
3383 push_args rest code'
3386 -- we only ever generate word-sized function arguments. Promotion
3387 -- has already happened: our Int8# type is kept sign-extended
3388 -- in an Int#, for example.
3389 ASSERT(width == W64) return ()
3390 (arg_op, arg_code) <- getOperand arg
3391 delta <- getDeltaNat
3392 setDeltaNat (delta-arg_size)
3393 let code' = code `appOL` toOL [PUSH II64 arg_op,
3394 DELTA (delta-arg_size)]
3395 push_args rest code'
3397 arg_rep = cmmExprType arg
3398 width = typeWidth arg_rep
3401 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
3403 #if sparc_TARGET_ARCH
3405 The SPARC calling convention is an absolute
3406 nightmare. The first 6x32 bits of arguments are mapped into
3407 %o0 through %o5, and the remaining arguments are dumped to the
3408 stack, beginning at [%sp+92]. (Note that %o6 == %sp.)
3410 If we have to put args on the stack, move %o6==%sp down by
3411 the number of words to go on the stack, to ensure there's enough space.
3413 According to Fraser and Hanson's lcc book, page 478, fig 17.2,
3414 16 words above the stack pointer is a word for the address of
3415 a structure return value. I use this as a temporary location
3416 for moving values from float to int regs. Certainly it isn't
3417 safe to put anything in the 16 words starting at %sp, since
3418 this area can get trashed at any time due to window overflows
3419 caused by signal handlers.
3421 A final complication (if the above isn't enough) is that
3422 we can't blithely calculate the arguments one by one into
3423 %o0 .. %o5. Consider the following nested calls:
3427 Naive code moves a into %o0, and (fff b c) into %o1. Unfortunately
3428 the inner call will itself use %o0, which trashes the value put there
3429 in preparation for the outer call. Upshot: we need to calculate the
3430 args into temporary regs, and move those to arg regs or onto the
3431 stack only immediately prior to the call proper. Sigh.
3434 genCCall target dest_regs argsAndHints = do
3436 args = map hintlessCmm argsAndHints
3437 argcode_and_vregs <- mapM arg_to_int_vregs args
3439 (argcodes, vregss) = unzip argcode_and_vregs
3440 n_argRegs = length allArgRegs
3441 n_argRegs_used = min (length vregs) n_argRegs
3442 vregs = concat vregss
3443 -- deal with static vs dynamic call targets
3444 callinsns <- (case target of
3445 CmmCallee (CmmLit (CmmLabel lbl)) conv -> do
3446 return (unitOL (CALL (Left (litToImm (CmmLabel lbl))) n_argRegs_used False))
3447 CmmCallee expr conv -> do
3448 (dyn_c, [dyn_r]) <- arg_to_int_vregs expr
3449 return (dyn_c `snocOL` CALL (Right dyn_r) n_argRegs_used False)
3451 (res, reduce) <- outOfLineFloatOp mop
3452 lblOrMopExpr <- case res of
3454 return (unitOL (CALL (Left (litToImm (CmmLabel lbl))) n_argRegs_used False))
3456 (dyn_c, [dyn_r]) <- arg_to_int_vregs mopExpr
3457 return (dyn_c `snocOL` CALL (Right dyn_r) n_argRegs_used False)
3458 if reduce then panic "genCCall(sparc): can not reduce" else return lblOrMopExpr
3462 argcode = concatOL argcodes
3463 (move_sp_down, move_sp_up)
3464 = let diff = length vregs - n_argRegs
3465 nn = if odd diff then diff + 1 else diff -- keep 8-byte alignment
3468 else (unitOL (moveSp (-1*nn)), unitOL (moveSp (1*nn)))
3470 = toOL (move_final vregs allArgRegs eXTRA_STK_ARGS_HERE)
3471 return (argcode `appOL`
3472 move_sp_down `appOL`
3473 transfer_code `appOL`
3478 -- move args from the integer vregs into which they have been
3479 -- marshalled, into %o0 .. %o5, and the rest onto the stack.
3480 move_final :: [Reg] -> [Reg] -> Int -> [Instr]
3482 move_final [] _ offset -- all args done
3485 move_final (v:vs) [] offset -- out of aregs; move to stack
3486 = ST II32 v (spRel offset)
3487 : move_final vs [] (offset+1)
3489 move_final (v:vs) (a:az) offset -- move into an arg (%o[0..5]) reg
3490 = OR False g0 (RIReg v) a
3491 : move_final vs az offset
3493 -- generate code to calculate an argument, and move it into one
3494 -- or two integer vregs.
3495 arg_to_int_vregs :: CmmExpr -> NatM (OrdList Instr, [Reg])
3496 arg_to_int_vregs arg
3497 | isWord64 (cmmExprType arg)
3499 (ChildCode64 code r_lo) <- iselExpr64 arg
3501 r_hi = getHiVRegFromLo r_lo
3502 return (code, [r_hi, r_lo])
3505 (src, code) <- getSomeReg arg
3506 tmp <- getNewRegNat (cmmExprType arg)
3508 pk = cmmExprType arg
3511 v1 <- getNewRegNat II32
3512 v2 <- getNewRegNat II32
3515 FMOV FF64 src f0 `snocOL`
3516 ST FF32 f0 (spRel 16) `snocOL`
3517 LD II32 (spRel 16) v1 `snocOL`
3518 ST FF32 (fPair f0) (spRel 16) `snocOL`
3519 LD II32 (spRel 16) v2
3524 v1 <- getNewRegNat II32
3527 ST FF32 src (spRel 16) `snocOL`
3528 LD II32 (spRel 16) v1
3533 v1 <- getNewRegNat II32
3535 code `snocOL` OR False g0 (RIReg src) v1
3539 outOfLineFloatOp mop =
3541 dflags <- getDynFlagsNat
3542 mopExpr <- cmmMakeDynamicReference dflags addImportNat CallReference $
3543 mkForeignLabel functionName Nothing True
3544 let mopLabelOrExpr = case mopExpr of
3545 CmmLit (CmmLabel lbl) -> Left lbl
3547 return (mopLabelOrExpr, reduce)
3549 (reduce, functionName) = case mop of
3550 MO_F32_Exp -> (True, fsLit "exp")
3551 MO_F32_Log -> (True, fsLit "log")
3552 MO_F32_Sqrt -> (True, fsLit "sqrt")
3554 MO_F32_Sin -> (True, fsLit "sin")
3555 MO_F32_Cos -> (True, fsLit "cos")
3556 MO_F32_Tan -> (True, fsLit "tan")
3558 MO_F32_Asin -> (True, fsLit "asin")
3559 MO_F32_Acos -> (True, fsLit "acos")
3560 MO_F32_Atan -> (True, fsLit "atan")
3562 MO_F32_Sinh -> (True, fsLit "sinh")
3563 MO_F32_Cosh -> (True, fsLit "cosh")
3564 MO_F32_Tanh -> (True, fsLit "tanh")
3566 MO_F64_Exp -> (False, fsLit "exp")
3567 MO_F64_Log -> (False, fsLit "log")
3568 MO_F64_Sqrt -> (False, fsLit "sqrt")
3570 MO_F64_Sin -> (False, fsLit "sin")
3571 MO_F64_Cos -> (False, fsLit "cos")
3572 MO_F64_Tan -> (False, fsLit "tan")
3574 MO_F64_Asin -> (False, fsLit "asin")
3575 MO_F64_Acos -> (False, fsLit "acos")
3576 MO_F64_Atan -> (False, fsLit "atan")
3578 MO_F64_Sinh -> (False, fsLit "sinh")
3579 MO_F64_Cosh -> (False, fsLit "cosh")
3580 MO_F64_Tanh -> (False, fsLit "tanh")
3582 other -> pprPanic "outOfLineFloatOp(sparc) "
3583 (pprCallishMachOp mop)
3585 #endif /* sparc_TARGET_ARCH */
3587 #if powerpc_TARGET_ARCH
3589 #if darwin_TARGET_OS || linux_TARGET_OS
3591 The PowerPC calling convention for Darwin/Mac OS X
3592 is described in Apple's document
3593 "Inside Mac OS X - Mach-O Runtime Architecture".
3595 PowerPC Linux uses the System V Release 4 Calling Convention
3596 for PowerPC. It is described in the
3597 "System V Application Binary Interface PowerPC Processor Supplement".
3599 Both conventions are similar:
3600 Parameters may be passed in general-purpose registers starting at r3, in
3601 floating point registers starting at f1, or on the stack.
3603 But there are substantial differences:
3604 * The number of registers used for parameter passing and the exact set of
3605 nonvolatile registers differs (see MachRegs.lhs).
3606 * On Darwin, stack space is always reserved for parameters, even if they are
3607 passed in registers. The called routine may choose to save parameters from
3608 registers to the corresponding space on the stack.
3609 * On Darwin, a corresponding amount of GPRs is skipped when a floating point
3610 parameter is passed in an FPR.
3611 * SysV insists on either passing I64 arguments on the stack, or in two GPRs,
3612 starting with an odd-numbered GPR. It may skip a GPR to achieve this.
3613 Darwin just treats an I64 like two separate II32s (high word first).
3614 * I64 and FF64 arguments are 8-byte aligned on the stack for SysV, but only
3615 4-byte aligned like everything else on Darwin.
3616 * The SysV spec claims that FF32 is represented as FF64 on the stack. GCC on
3617 PowerPC Linux does not agree, so neither do we.
3619 According to both conventions, The parameter area should be part of the
3620 caller's stack frame, allocated in the caller's prologue code (large enough
3621 to hold the parameter lists for all called routines). The NCG already
3622 uses the stack for register spilling, leaving 64 bytes free at the top.
3623 If we need a larger parameter area than that, we just allocate a new stack
3624 frame just before ccalling.
3628 genCCall (CmmPrim MO_WriteBarrier) _ _
3629 = return $ unitOL LWSYNC
3631 genCCall target dest_regs argsAndHints
3632 = ASSERT (not $ any (`elem` [II8,II16]) $ map cmmTypeSize argReps)
3633 -- we rely on argument promotion in the codeGen
3635 (finalStack,passArgumentsCode,usedRegs) <- passArguments
3637 allArgRegs allFPArgRegs
3641 (labelOrExpr, reduceToFF32) <- case target of
3642 CmmCallee (CmmLit (CmmLabel lbl)) conv -> return (Left lbl, False)
3643 CmmCallee expr conv -> return (Right expr, False)
3644 CmmPrim mop -> outOfLineFloatOp mop
3646 let codeBefore = move_sp_down finalStack `appOL` passArgumentsCode
3647 codeAfter = move_sp_up finalStack `appOL` moveResult reduceToFF32
3652 `snocOL` BL lbl usedRegs
3655 (dynReg, dynCode) <- getSomeReg dyn
3657 `snocOL` MTCTR dynReg
3659 `snocOL` BCTRL usedRegs
3662 #if darwin_TARGET_OS
3663 initialStackOffset = 24
3664 -- size of linkage area + size of arguments, in bytes
3665 stackDelta _finalStack = roundTo 16 $ (24 +) $ max 32 $ sum $
3666 map (widthInBytes . typeWidth) argReps
3667 #elif linux_TARGET_OS
3668 initialStackOffset = 8
3669 stackDelta finalStack = roundTo 16 finalStack
3671 args = map hintlessCmm argsAndHints
3672 argReps = map cmmExprType args
3674 roundTo a x | x `mod` a == 0 = x
3675 | otherwise = x + a - (x `mod` a)
3677 move_sp_down finalStack
3679 toOL [STU II32 sp (AddrRegImm sp (ImmInt (-delta))),
3682 where delta = stackDelta finalStack
3683 move_sp_up finalStack
3685 toOL [ADD sp sp (RIImm (ImmInt delta)),
3688 where delta = stackDelta finalStack
3691 passArguments [] _ _ stackOffset accumCode accumUsed = return (stackOffset, accumCode, accumUsed)
3692 passArguments ((arg,arg_ty):args) gprs fprs stackOffset
3693 accumCode accumUsed | isWord64 arg_ty =
3695 ChildCode64 code vr_lo <- iselExpr64 arg
3696 let vr_hi = getHiVRegFromLo vr_lo
3698 #if darwin_TARGET_OS
3703 (accumCode `appOL` code
3704 `snocOL` storeWord vr_hi gprs stackOffset
3705 `snocOL` storeWord vr_lo (drop 1 gprs) (stackOffset+4))
3706 ((take 2 gprs) ++ accumUsed)
3708 storeWord vr (gpr:_) offset = MR gpr vr
3709 storeWord vr [] offset = ST II32 vr (AddrRegImm sp (ImmInt offset))
3711 #elif linux_TARGET_OS
3712 let stackOffset' = roundTo 8 stackOffset
3713 stackCode = accumCode `appOL` code
3714 `snocOL` ST II32 vr_hi (AddrRegImm sp (ImmInt stackOffset'))
3715 `snocOL` ST II32 vr_lo (AddrRegImm sp (ImmInt (stackOffset'+4)))
3716 regCode hireg loreg =
3717 accumCode `appOL` code
3718 `snocOL` MR hireg vr_hi
3719 `snocOL` MR loreg vr_lo
3722 hireg : loreg : regs | even (length gprs) ->
3723 passArguments args regs fprs stackOffset
3724 (regCode hireg loreg) (hireg : loreg : accumUsed)
3725 _skipped : hireg : loreg : regs ->
3726 passArguments args regs fprs stackOffset
3727 (regCode hireg loreg) (hireg : loreg : accumUsed)
3728 _ -> -- only one or no regs left
3729 passArguments args [] fprs (stackOffset'+8)
3733 passArguments ((arg,rep):args) gprs fprs stackOffset accumCode accumUsed
3734 | reg : _ <- regs = do
3735 register <- getRegister arg
3736 let code = case register of
3737 Fixed _ freg fcode -> fcode `snocOL` MR reg freg
3738 Any _ acode -> acode reg
3742 #if darwin_TARGET_OS
3743 -- The Darwin ABI requires that we reserve stack slots for register parameters
3744 (stackOffset + stackBytes)
3745 #elif linux_TARGET_OS
3746 -- ... the SysV ABI doesn't.
3749 (accumCode `appOL` code)
3752 (vr, code) <- getSomeReg arg
3756 (stackOffset' + stackBytes)
3757 (accumCode `appOL` code `snocOL` ST (cmmTypeSize rep) vr stackSlot)
3760 #if darwin_TARGET_OS
3761 -- stackOffset is at least 4-byte aligned
3762 -- The Darwin ABI is happy with that.
3763 stackOffset' = stackOffset
3765 -- ... the SysV ABI requires 8-byte alignment for doubles.
3766 stackOffset' | isFloatType rep && typeWidth rep == W64 =
3767 roundTo 8 stackOffset
3768 | otherwise = stackOffset
3770 stackSlot = AddrRegImm sp (ImmInt stackOffset')
3771 (nGprs, nFprs, stackBytes, regs) = case cmmTypeSize rep of
3772 II32 -> (1, 0, 4, gprs)
3773 #if darwin_TARGET_OS
3774 -- The Darwin ABI requires that we skip a corresponding number of GPRs when
3776 FF32 -> (1, 1, 4, fprs)
3777 FF64 -> (2, 1, 8, fprs)
3778 #elif linux_TARGET_OS
3779 -- ... the SysV ABI doesn't.
3780 FF32 -> (0, 1, 4, fprs)
3781 FF64 -> (0, 1, 8, fprs)
3784 moveResult reduceToFF32 =
3787 [CmmHinted dest _hint]
3788 | reduceToFF32 && isFloat32 rep -> unitOL (FRSP r_dest f1)
3789 | isFloat32 rep || isFloat64 rep -> unitOL (MR r_dest f1)
3790 | isWord64 rep -> toOL [MR (getHiVRegFromLo r_dest) r3,
3792 | otherwise -> unitOL (MR r_dest r3)
3793 where rep = cmmRegType (CmmLocal dest)
3794 r_dest = getRegisterReg (CmmLocal dest)
3796 outOfLineFloatOp mop =
3798 dflags <- getDynFlagsNat
3799 mopExpr <- cmmMakeDynamicReference dflags addImportNat CallReference $
3800 mkForeignLabel functionName Nothing True
3801 let mopLabelOrExpr = case mopExpr of
3802 CmmLit (CmmLabel lbl) -> Left lbl
3804 return (mopLabelOrExpr, reduce)
3806 (functionName, reduce) = case mop of
3807 MO_F32_Exp -> (fsLit "exp", True)
3808 MO_F32_Log -> (fsLit "log", True)
3809 MO_F32_Sqrt -> (fsLit "sqrt", True)
3811 MO_F32_Sin -> (fsLit "sin", True)
3812 MO_F32_Cos -> (fsLit "cos", True)
3813 MO_F32_Tan -> (fsLit "tan", True)
3815 MO_F32_Asin -> (fsLit "asin", True)
3816 MO_F32_Acos -> (fsLit "acos", True)
3817 MO_F32_Atan -> (fsLit "atan", True)
3819 MO_F32_Sinh -> (fsLit "sinh", True)
3820 MO_F32_Cosh -> (fsLit "cosh", True)
3821 MO_F32_Tanh -> (fsLit "tanh", True)
3822 MO_F32_Pwr -> (fsLit "pow", True)
3824 MO_F64_Exp -> (fsLit "exp", False)
3825 MO_F64_Log -> (fsLit "log", False)
3826 MO_F64_Sqrt -> (fsLit "sqrt", False)
3828 MO_F64_Sin -> (fsLit "sin", False)
3829 MO_F64_Cos -> (fsLit "cos", False)
3830 MO_F64_Tan -> (fsLit "tan", False)
3832 MO_F64_Asin -> (fsLit "asin", False)
3833 MO_F64_Acos -> (fsLit "acos", False)
3834 MO_F64_Atan -> (fsLit "atan", False)
3836 MO_F64_Sinh -> (fsLit "sinh", False)
3837 MO_F64_Cosh -> (fsLit "cosh", False)
3838 MO_F64_Tanh -> (fsLit "tanh", False)
3839 MO_F64_Pwr -> (fsLit "pow", False)
3840 other -> pprPanic "genCCall(ppc): unknown callish op"
3841 (pprCallishMachOp other)
3843 #endif /* darwin_TARGET_OS || linux_TARGET_OS */
3845 #endif /* powerpc_TARGET_ARCH */
3848 -- -----------------------------------------------------------------------------
3849 -- Generating a table-branch
3851 genSwitch :: CmmExpr -> [Maybe BlockId] -> NatM InstrBlock
3853 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
3857 (reg,e_code) <- getSomeReg expr
3858 lbl <- getNewLabelNat
3859 dflags <- getDynFlagsNat
3860 dynRef <- cmmMakeDynamicReference dflags addImportNat DataReference lbl
3861 (tableReg,t_code) <- getSomeReg $ dynRef
3863 jumpTable = map jumpTableEntryRel ids
3865 jumpTableEntryRel Nothing
3866 = CmmStaticLit (CmmInt 0 wordWidth)
3867 jumpTableEntryRel (Just (BlockId id))
3868 = CmmStaticLit (CmmLabelDiffOff blockLabel lbl 0)
3869 where blockLabel = mkAsmTempLabel id
3871 op = OpAddr (AddrBaseIndex (EABaseReg tableReg)
3872 (EAIndex reg wORD_SIZE) (ImmInt 0))
3874 #if x86_64_TARGET_ARCH
3875 #if darwin_TARGET_OS
3876 -- on Mac OS X/x86_64, put the jump table in the text section
3877 -- to work around a limitation of the linker.
3878 -- ld64 is unable to handle the relocations for
3880 -- if L0 is not preceded by a non-anonymous label in its section.
3882 code = e_code `appOL` t_code `appOL` toOL [
3883 ADD (intSize wordWidth) op (OpReg tableReg),
3884 JMP_TBL (OpReg tableReg) [ id | Just id <- ids ],
3885 LDATA Text (CmmDataLabel lbl : jumpTable)
3888 -- HACK: On x86_64 binutils<2.17 is only able to generate PC32
3889 -- relocations, hence we only get 32-bit offsets in the jump
3890 -- table. As these offsets are always negative we need to properly
3891 -- sign extend them to 64-bit. This hack should be removed in
3892 -- conjunction with the hack in PprMach.hs/pprDataItem once
3893 -- binutils 2.17 is standard.
3894 code = e_code `appOL` t_code `appOL` toOL [
3895 LDATA ReadOnlyData (CmmDataLabel lbl : jumpTable),
3897 (OpAddr (AddrBaseIndex (EABaseReg tableReg)
3898 (EAIndex reg wORD_SIZE) (ImmInt 0)))
3900 ADD (intSize wordWidth) (OpReg reg) (OpReg tableReg),
3901 JMP_TBL (OpReg tableReg) [ id | Just id <- ids ]
3905 code = e_code `appOL` t_code `appOL` toOL [
3906 LDATA ReadOnlyData (CmmDataLabel lbl : jumpTable),
3907 ADD (intSize wordWidth) op (OpReg tableReg),
3908 JMP_TBL (OpReg tableReg) [ id | Just id <- ids ]
3914 (reg,e_code) <- getSomeReg expr
3915 lbl <- getNewLabelNat
3917 jumpTable = map jumpTableEntry ids
3918 op = OpAddr (AddrBaseIndex EABaseNone (EAIndex reg wORD_SIZE) (ImmCLbl lbl))
3919 code = e_code `appOL` toOL [
3920 LDATA ReadOnlyData (CmmDataLabel lbl : jumpTable),
3921 JMP_TBL op [ id | Just id <- ids ]
3925 #elif powerpc_TARGET_ARCH
3929 (reg,e_code) <- getSomeReg expr
3930 tmp <- getNewRegNat II32
3931 lbl <- getNewLabelNat
3932 dflags <- getDynFlagsNat
3933 dynRef <- cmmMakeDynamicReference dflags addImportNat DataReference lbl
3934 (tableReg,t_code) <- getSomeReg $ dynRef
3936 jumpTable = map jumpTableEntryRel ids
3938 jumpTableEntryRel Nothing
3939 = CmmStaticLit (CmmInt 0 wordWidth)
3940 jumpTableEntryRel (Just (BlockId id))
3941 = CmmStaticLit (CmmLabelDiffOff blockLabel lbl 0)
3942 where blockLabel = mkAsmTempLabel id
3944 code = e_code `appOL` t_code `appOL` toOL [
3945 LDATA ReadOnlyData (CmmDataLabel lbl : jumpTable),
3946 SLW tmp reg (RIImm (ImmInt 2)),
3947 LD II32 tmp (AddrRegReg tableReg tmp),
3948 ADD tmp tmp (RIReg tableReg),
3950 BCTR [ id | Just id <- ids ]
3955 (reg,e_code) <- getSomeReg expr
3956 tmp <- getNewRegNat II32
3957 lbl <- getNewLabelNat
3959 jumpTable = map jumpTableEntry ids
3961 code = e_code `appOL` toOL [
3962 LDATA ReadOnlyData (CmmDataLabel lbl : jumpTable),
3963 SLW tmp reg (RIImm (ImmInt 2)),
3964 ADDIS tmp tmp (HA (ImmCLbl lbl)),
3965 LD II32 tmp (AddrRegImm tmp (LO (ImmCLbl lbl))),
3967 BCTR [ id | Just id <- ids ]
3971 #error "ToDo: genSwitch"
3974 jumpTableEntry Nothing = CmmStaticLit (CmmInt 0 wordWidth)
3975 jumpTableEntry (Just (BlockId id)) = CmmStaticLit (CmmLabel blockLabel)
3976 where blockLabel = mkAsmTempLabel id
3978 -- -----------------------------------------------------------------------------
3980 -- -----------------------------------------------------------------------------
3983 -- -----------------------------------------------------------------------------
3984 -- 'condIntReg' and 'condFltReg': condition codes into registers
3986 -- Turn those condition codes into integers now (when they appear on
3987 -- the right hand side of an assignment).
3989 -- (If applicable) Do not fill the delay slots here; you will confuse the
3990 -- register allocator.
3992 condIntReg, condFltReg :: Cond -> CmmExpr -> CmmExpr -> NatM Register
3994 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
3996 #if alpha_TARGET_ARCH
3997 condIntReg = panic "MachCode.condIntReg (not on Alpha)"
3998 condFltReg = panic "MachCode.condFltReg (not on Alpha)"
3999 #endif /* alpha_TARGET_ARCH */
4001 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
4003 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
4005 condIntReg cond x y = do
4006 CondCode _ cond cond_code <- condIntCode cond x y
4007 tmp <- getNewRegNat II8
4009 code dst = cond_code `appOL` toOL [
4010 SETCC cond (OpReg tmp),
4011 MOVZxL II8 (OpReg tmp) (OpReg dst)
4014 return (Any II32 code)
4018 #if i386_TARGET_ARCH
4020 condFltReg cond x y = do
4021 CondCode _ cond cond_code <- condFltCode cond x y
4022 tmp <- getNewRegNat II8
4024 code dst = cond_code `appOL` toOL [
4025 SETCC cond (OpReg tmp),
4026 MOVZxL II8 (OpReg tmp) (OpReg dst)
4029 return (Any II32 code)
4033 #if x86_64_TARGET_ARCH
4035 condFltReg cond x y = do
4036 CondCode _ cond cond_code <- condFltCode cond x y
4037 tmp1 <- getNewRegNat wordSize
4038 tmp2 <- getNewRegNat wordSize
4040 -- We have to worry about unordered operands (eg. comparisons
4041 -- against NaN). If the operands are unordered, the comparison
4042 -- sets the parity flag, carry flag and zero flag.
4043 -- All comparisons are supposed to return false for unordered
4044 -- operands except for !=, which returns true.
4046 -- Optimisation: we don't have to test the parity flag if we
4047 -- know the test has already excluded the unordered case: eg >
4048 -- and >= test for a zero carry flag, which can only occur for
4049 -- ordered operands.
4051 -- ToDo: by reversing comparisons we could avoid testing the
4052 -- parity flag in more cases.
4057 NE -> or_unordered dst
4058 GU -> plain_test dst
4059 GEU -> plain_test dst
4060 _ -> and_ordered dst)
4062 plain_test dst = toOL [
4063 SETCC cond (OpReg tmp1),
4064 MOVZxL II8 (OpReg tmp1) (OpReg dst)
4066 or_unordered dst = toOL [
4067 SETCC cond (OpReg tmp1),
4068 SETCC PARITY (OpReg tmp2),
4069 OR II8 (OpReg tmp1) (OpReg tmp2),
4070 MOVZxL II8 (OpReg tmp2) (OpReg dst)
4072 and_ordered dst = toOL [
4073 SETCC cond (OpReg tmp1),
4074 SETCC NOTPARITY (OpReg tmp2),
4075 AND II8 (OpReg tmp1) (OpReg tmp2),
4076 MOVZxL II8 (OpReg tmp2) (OpReg dst)
4079 return (Any II32 code)
4083 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
4085 #if sparc_TARGET_ARCH
4087 condIntReg EQQ x (CmmLit (CmmInt 0 d)) = do
4088 (src, code) <- getSomeReg x
4089 tmp <- getNewRegNat II32
4091 code__2 dst = code `appOL` toOL [
4092 SUB False True g0 (RIReg src) g0,
4093 SUB True False g0 (RIImm (ImmInt (-1))) dst]
4094 return (Any II32 code__2)
4096 condIntReg EQQ x y = do
4097 (src1, code1) <- getSomeReg x
4098 (src2, code2) <- getSomeReg y
4099 tmp1 <- getNewRegNat II32
4100 tmp2 <- getNewRegNat II32
4102 code__2 dst = code1 `appOL` code2 `appOL` toOL [
4103 XOR False src1 (RIReg src2) dst,
4104 SUB False True g0 (RIReg dst) g0,
4105 SUB True False g0 (RIImm (ImmInt (-1))) dst]
4106 return (Any II32 code__2)
4108 condIntReg NE x (CmmLit (CmmInt 0 d)) = do
4109 (src, code) <- getSomeReg x
4110 tmp <- getNewRegNat II32
4112 code__2 dst = code `appOL` toOL [
4113 SUB False True g0 (RIReg src) g0,
4114 ADD True False g0 (RIImm (ImmInt 0)) dst]
4115 return (Any II32 code__2)
4117 condIntReg NE x y = do
4118 (src1, code1) <- getSomeReg x
4119 (src2, code2) <- getSomeReg y
4120 tmp1 <- getNewRegNat II32
4121 tmp2 <- getNewRegNat II32
4123 code__2 dst = code1 `appOL` code2 `appOL` toOL [
4124 XOR False src1 (RIReg src2) dst,
4125 SUB False True g0 (RIReg dst) g0,
4126 ADD True False g0 (RIImm (ImmInt 0)) dst]
4127 return (Any II32 code__2)
4129 condIntReg cond x y = do
4130 BlockId lbl1 <- getBlockIdNat
4131 BlockId lbl2 <- getBlockIdNat
4132 CondCode _ cond cond_code <- condIntCode cond x y
4134 code__2 dst = cond_code `appOL` toOL [
4135 BI cond False (ImmCLbl (mkAsmTempLabel lbl1)), NOP,
4136 OR False g0 (RIImm (ImmInt 0)) dst,
4137 BI ALWAYS False (ImmCLbl (mkAsmTempLabel lbl2)), NOP,
4138 NEWBLOCK (BlockId lbl1),
4139 OR False g0 (RIImm (ImmInt 1)) dst,
4140 NEWBLOCK (BlockId lbl2)]
4141 return (Any II32 code__2)
4143 condFltReg cond x y = do
4144 BlockId lbl1 <- getBlockIdNat
4145 BlockId lbl2 <- getBlockIdNat
4146 CondCode _ cond cond_code <- condFltCode cond x y
4148 code__2 dst = cond_code `appOL` toOL [
4150 BF cond False (ImmCLbl (mkAsmTempLabel lbl1)), NOP,
4151 OR False g0 (RIImm (ImmInt 0)) dst,
4152 BI ALWAYS False (ImmCLbl (mkAsmTempLabel lbl2)), NOP,
4153 NEWBLOCK (BlockId lbl1),
4154 OR False g0 (RIImm (ImmInt 1)) dst,
4155 NEWBLOCK (BlockId lbl2)]
4156 return (Any II32 code__2)
4158 #endif /* sparc_TARGET_ARCH */
4160 #if powerpc_TARGET_ARCH
4161 condReg getCond = do
4162 lbl1 <- getBlockIdNat
4163 lbl2 <- getBlockIdNat
4164 CondCode _ cond cond_code <- getCond
4166 {- code dst = cond_code `appOL` toOL [
4175 code dst = cond_code
4179 RLWINM dst dst (bit + 1) 31 31
4182 negate_code | do_negate = unitOL (CRNOR bit bit bit)
4185 (bit, do_negate) = case cond of
4199 return (Any II32 code)
4201 condIntReg cond x y = condReg (condIntCode cond x y)
4202 condFltReg cond x y = condReg (condFltCode cond x y)
4203 #endif /* powerpc_TARGET_ARCH */
4206 -- -----------------------------------------------------------------------------
4207 -- 'trivial*Code': deal with trivial instructions
4209 -- Trivial (dyadic: 'trivialCode', floating-point: 'trivialFCode',
4210 -- unary: 'trivialUCode', unary fl-pt:'trivialUFCode') instructions.
4211 -- Only look for constants on the right hand side, because that's
4212 -- where the generic optimizer will have put them.
4214 -- Similarly, for unary instructions, we don't have to worry about
4215 -- matching an StInt as the argument, because genericOpt will already
4216 -- have handled the constant-folding.
4219 :: Width -- Int only
4220 -> IF_ARCH_alpha((Reg -> RI -> Reg -> Instr)
4221 ,IF_ARCH_i386 ((Operand -> Operand -> Instr)
4222 -> Maybe (Operand -> Operand -> Instr)
4223 ,IF_ARCH_x86_64 ((Operand -> Operand -> Instr)
4224 -> Maybe (Operand -> Operand -> Instr)
4225 ,IF_ARCH_sparc((Reg -> RI -> Reg -> Instr)
4226 ,IF_ARCH_powerpc(Bool -> (Reg -> Reg -> RI -> Instr)
4228 -> CmmExpr -> CmmExpr -- the two arguments
4231 #ifndef powerpc_TARGET_ARCH
4233 :: Width -- Floating point only
4234 -> IF_ARCH_alpha((Reg -> Reg -> Reg -> Instr)
4235 ,IF_ARCH_sparc((Size -> Reg -> Reg -> Reg -> Instr)
4236 ,IF_ARCH_i386 ((Size -> Reg -> Reg -> Reg -> Instr)
4237 ,IF_ARCH_x86_64 ((Size -> Operand -> Operand -> Instr)
4239 -> CmmExpr -> CmmExpr -- the two arguments
4245 -> IF_ARCH_alpha((RI -> Reg -> Instr)
4246 ,IF_ARCH_i386 ((Operand -> Instr)
4247 ,IF_ARCH_x86_64 ((Operand -> Instr)
4248 ,IF_ARCH_sparc((RI -> Reg -> Instr)
4249 ,IF_ARCH_powerpc((Reg -> Reg -> Instr)
4251 -> CmmExpr -- the one argument
4254 #ifndef powerpc_TARGET_ARCH
4257 -> IF_ARCH_alpha((Reg -> Reg -> Instr)
4258 ,IF_ARCH_i386 ((Reg -> Reg -> Instr)
4259 ,IF_ARCH_x86_64 ((Reg -> Reg -> Instr)
4260 ,IF_ARCH_sparc((Reg -> Reg -> Instr)
4262 -> CmmExpr -- the one argument
4266 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
4268 #if alpha_TARGET_ARCH
4270 trivialCode instr x (StInt y)
4272 = getRegister x `thenNat` \ register ->
4273 getNewRegNat IntRep `thenNat` \ tmp ->
4275 code = registerCode register tmp
4276 src1 = registerName register tmp
4277 src2 = ImmInt (fromInteger y)
4278 code__2 dst = code . mkSeqInstr (instr src1 (RIImm src2) dst)
4280 return (Any IntRep code__2)
4282 trivialCode instr x y
4283 = getRegister x `thenNat` \ register1 ->
4284 getRegister y `thenNat` \ register2 ->
4285 getNewRegNat IntRep `thenNat` \ tmp1 ->
4286 getNewRegNat IntRep `thenNat` \ tmp2 ->
4288 code1 = registerCode register1 tmp1 []
4289 src1 = registerName register1 tmp1
4290 code2 = registerCode register2 tmp2 []
4291 src2 = registerName register2 tmp2
4292 code__2 dst = asmSeqThen [code1, code2] .
4293 mkSeqInstr (instr src1 (RIReg src2) dst)
4295 return (Any IntRep code__2)
4298 trivialUCode instr x
4299 = getRegister x `thenNat` \ register ->
4300 getNewRegNat IntRep `thenNat` \ tmp ->
4302 code = registerCode register tmp
4303 src = registerName register tmp
4304 code__2 dst = code . mkSeqInstr (instr (RIReg src) dst)
4306 return (Any IntRep code__2)
4309 trivialFCode _ instr x y
4310 = getRegister x `thenNat` \ register1 ->
4311 getRegister y `thenNat` \ register2 ->
4312 getNewRegNat FF64 `thenNat` \ tmp1 ->
4313 getNewRegNat FF64 `thenNat` \ tmp2 ->
4315 code1 = registerCode register1 tmp1
4316 src1 = registerName register1 tmp1
4318 code2 = registerCode register2 tmp2
4319 src2 = registerName register2 tmp2
4321 code__2 dst = asmSeqThen [code1 [], code2 []] .
4322 mkSeqInstr (instr src1 src2 dst)
4324 return (Any FF64 code__2)
4326 trivialUFCode _ instr x
4327 = getRegister x `thenNat` \ register ->
4328 getNewRegNat FF64 `thenNat` \ tmp ->
4330 code = registerCode register tmp
4331 src = registerName register tmp
4332 code__2 dst = code . mkSeqInstr (instr src dst)
4334 return (Any FF64 code__2)
4336 #endif /* alpha_TARGET_ARCH */
4338 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
4340 #if i386_TARGET_ARCH || x86_64_TARGET_ARCH
4343 The Rules of the Game are:
4345 * You cannot assume anything about the destination register dst;
4346 it may be anything, including a fixed reg.
4348 * You may compute an operand into a fixed reg, but you may not
4349 subsequently change the contents of that fixed reg. If you
4350 want to do so, first copy the value either to a temporary
4351 or into dst. You are free to modify dst even if it happens
4352 to be a fixed reg -- that's not your problem.
4354 * You cannot assume that a fixed reg will stay live over an
4355 arbitrary computation. The same applies to the dst reg.
4357 * Temporary regs obtained from getNewRegNat are distinct from
4358 each other and from all other regs, and stay live over
4359 arbitrary computations.
4361 --------------------
4363 SDM's version of The Rules:
4365 * If getRegister returns Any, that means it can generate correct
4366 code which places the result in any register, period. Even if that
4367 register happens to be read during the computation.
4369 Corollary #1: this means that if you are generating code for an
4370 operation with two arbitrary operands, you cannot assign the result
4371 of the first operand into the destination register before computing
4372 the second operand. The second operand might require the old value
4373 of the destination register.
4375 Corollary #2: A function might be able to generate more efficient
4376 code if it knows the destination register is a new temporary (and
4377 therefore not read by any of the sub-computations).
4379 * If getRegister returns Any, then the code it generates may modify only:
4380 (a) fresh temporaries
4381 (b) the destination register
4382 (c) known registers (eg. %ecx is used by shifts)
4383 In particular, it may *not* modify global registers, unless the global
4384 register happens to be the destination register.
4387 trivialCode width instr (Just revinstr) (CmmLit lit_a) b
4388 | is32BitLit lit_a = do
4389 b_code <- getAnyReg b
4392 = b_code dst `snocOL`
4393 revinstr (OpImm (litToImm lit_a)) (OpReg dst)
4395 return (Any (intSize width) code)
4397 trivialCode width instr maybe_revinstr a b
4398 = genTrivialCode (intSize width) instr a b
4400 -- This is re-used for floating pt instructions too.
4401 genTrivialCode rep instr a b = do
4402 (b_op, b_code) <- getNonClobberedOperand b
4403 a_code <- getAnyReg a
4404 tmp <- getNewRegNat rep
4406 -- We want the value of b to stay alive across the computation of a.
4407 -- But, we want to calculate a straight into the destination register,
4408 -- because the instruction only has two operands (dst := dst `op` src).
4409 -- The troublesome case is when the result of b is in the same register
4410 -- as the destination reg. In this case, we have to save b in a
4411 -- new temporary across the computation of a.
4413 | dst `regClashesWithOp` b_op =
4415 unitOL (MOV rep b_op (OpReg tmp)) `appOL`
4417 instr (OpReg tmp) (OpReg dst)
4421 instr b_op (OpReg dst)
4423 return (Any rep code)
4425 reg `regClashesWithOp` OpReg reg2 = reg == reg2
4426 reg `regClashesWithOp` OpAddr amode = any (==reg) (addrModeRegs amode)
4427 reg `regClashesWithOp` _ = False
4431 trivialUCode rep instr x = do
4432 x_code <- getAnyReg x
4437 return (Any rep code)
4441 #if i386_TARGET_ARCH
4443 trivialFCode width instr x y = do
4444 (x_reg, x_code) <- getNonClobberedReg x -- these work for float regs too
4445 (y_reg, y_code) <- getSomeReg y
4447 size = floatSize width
4451 instr size x_reg y_reg dst
4452 return (Any size code)
4456 #if x86_64_TARGET_ARCH
4457 trivialFCode pk instr x y
4458 = genTrivialCode size (instr size) x y
4459 where size = floatSize pk
4464 trivialUFCode size instr x = do
4465 (x_reg, x_code) <- getSomeReg x
4471 return (Any size code)
4473 #endif /* i386_TARGET_ARCH */
4475 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
4477 #if sparc_TARGET_ARCH
4479 trivialCode pk instr x (CmmLit (CmmInt y d))
4482 (src1, code) <- getSomeReg x
4483 tmp <- getNewRegNat II32
4485 src2 = ImmInt (fromInteger y)
4486 code__2 dst = code `snocOL` instr src1 (RIImm src2) dst
4487 return (Any II32 code__2)
4489 trivialCode pk instr x y = do
4490 (src1, code1) <- getSomeReg x
4491 (src2, code2) <- getSomeReg y
4492 tmp1 <- getNewRegNat II32
4493 tmp2 <- getNewRegNat II32
4495 code__2 dst = code1 `appOL` code2 `snocOL`
4496 instr src1 (RIReg src2) dst
4497 return (Any II32 code__2)
4500 trivialFCode pk instr x y = do
4501 (src1, code1) <- getSomeReg x
4502 (src2, code2) <- getSomeReg y
4503 tmp1 <- getNewRegNat (cmmExprType x)
4504 tmp2 <- getNewRegNat (cmmExprType y)
4505 tmp <- getNewRegNat FF64
4507 promote x = FxTOy FF32 FF64 x tmp
4514 code1 `appOL` code2 `snocOL`
4515 instr (floatSize pk) src1 src2 dst
4516 else if typeWidth pk1 == W32 then
4517 code1 `snocOL` promote src1 `appOL` code2 `snocOL`
4518 instr FF64 tmp src2 dst
4520 code1 `appOL` code2 `snocOL` promote src2 `snocOL`
4521 instr FF64 src1 tmp dst
4522 return (Any (if pk1 == pk2 then pk1 else cmmFloat W64) code__2)
4525 trivialUCode size instr x = do
4526 (src, code) <- getSomeReg x
4527 tmp <- getNewRegNat size
4529 code__2 dst = code `snocOL` instr (RIReg src) dst
4530 return (Any size code__2)
4533 trivialUFCode pk instr x = do
4534 (src, code) <- getSomeReg x
4535 tmp <- getNewRegNat pk
4537 code__2 dst = code `snocOL` instr src dst
4538 return (Any pk code__2)
4540 #endif /* sparc_TARGET_ARCH */
4542 #if powerpc_TARGET_ARCH
4545 Wolfgang's PowerPC version of The Rules:
4547 A slightly modified version of The Rules to take advantage of the fact
4548 that PowerPC instructions work on all registers and don't implicitly
4549 clobber any fixed registers.
4551 * The only expression for which getRegister returns Fixed is (CmmReg reg).
4553 * If getRegister returns Any, then the code it generates may modify only:
4554 (a) fresh temporaries
4555 (b) the destination register
4556 It may *not* modify global registers, unless the global
4557 register happens to be the destination register.
4558 It may not clobber any other registers. In fact, only ccalls clobber any
4560 Also, it may not modify the counter register (used by genCCall).
4562 Corollary: If a getRegister for a subexpression returns Fixed, you need
4563 not move it to a fresh temporary before evaluating the next subexpression.
4564 The Fixed register won't be modified.
4565 Therefore, we don't need a counterpart for the x86's getStableReg on PPC.
4567 * SDM's First Rule is valid for PowerPC, too: subexpressions can depend on
4568 the value of the destination register.
4571 trivialCode rep signed instr x (CmmLit (CmmInt y _))
4572 | Just imm <- makeImmediate rep signed y
4574 (src1, code1) <- getSomeReg x
4575 let code dst = code1 `snocOL` instr dst src1 (RIImm imm)
4576 return (Any (intSize rep) code)
4578 trivialCode rep signed instr x y = do
4579 (src1, code1) <- getSomeReg x
4580 (src2, code2) <- getSomeReg y
4581 let code dst = code1 `appOL` code2 `snocOL` instr dst src1 (RIReg src2)
4582 return (Any (intSize rep) code)
4584 trivialCodeNoImm' :: Size -> (Reg -> Reg -> Reg -> Instr)
4585 -> CmmExpr -> CmmExpr -> NatM Register
4586 trivialCodeNoImm' size instr x y = do
4587 (src1, code1) <- getSomeReg x
4588 (src2, code2) <- getSomeReg y
4589 let code dst = code1 `appOL` code2 `snocOL` instr dst src1 src2
4590 return (Any size code)
4592 trivialCodeNoImm :: Size -> (Size -> Reg -> Reg -> Reg -> Instr)
4593 -> CmmExpr -> CmmExpr -> NatM Register
4594 trivialCodeNoImm size instr x y = trivialCodeNoImm' size (instr size) x y
4596 trivialUCode rep instr x = do
4597 (src, code) <- getSomeReg x
4598 let code' dst = code `snocOL` instr dst src
4599 return (Any rep code')
4601 -- There is no "remainder" instruction on the PPC, so we have to do
4603 -- The "div" parameter is the division instruction to use (DIVW or DIVWU)
4605 remainderCode :: Width -> (Reg -> Reg -> Reg -> Instr)
4606 -> CmmExpr -> CmmExpr -> NatM Register
4607 remainderCode rep div x y = do
4608 (src1, code1) <- getSomeReg x
4609 (src2, code2) <- getSomeReg y
4610 let code dst = code1 `appOL` code2 `appOL` toOL [
4612 MULLW dst dst (RIReg src2),
4615 return (Any (intSize rep) code)
4617 #endif /* powerpc_TARGET_ARCH */
4620 -- -----------------------------------------------------------------------------
4621 -- Coercing to/from integer/floating-point...
4623 -- When going to integer, we truncate (round towards 0).
4625 -- @coerce(Int2FP|FP2Int)@ are more complicated integer/float
4626 -- conversions. We have to store temporaries in memory to move
4627 -- between the integer and the floating point register sets.
4629 -- @coerceDbl2Flt@ and @coerceFlt2Dbl@ are done this way because we
4630 -- pretend, on sparc at least, that double and float regs are seperate
4631 -- kinds, so the value has to be computed into one kind before being
4632 -- explicitly "converted" to live in the other kind.
4634 coerceInt2FP :: Width -> Width -> CmmExpr -> NatM Register
4635 coerceFP2Int :: Width -> Width -> CmmExpr -> NatM Register
4637 #if sparc_TARGET_ARCH
4638 coerceDbl2Flt :: CmmExpr -> NatM Register
4639 coerceFlt2Dbl :: CmmExpr -> NatM Register
4642 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
4644 #if alpha_TARGET_ARCH
4647 = getRegister x `thenNat` \ register ->
4648 getNewRegNat IntRep `thenNat` \ reg ->
4650 code = registerCode register reg
4651 src = registerName register reg
4653 code__2 dst = code . mkSeqInstrs [
4655 LD TF dst (spRel 0),
4658 return (Any FF64 code__2)
4662 = getRegister x `thenNat` \ register ->
4663 getNewRegNat FF64 `thenNat` \ tmp ->
4665 code = registerCode register tmp
4666 src = registerName register tmp
4668 code__2 dst = code . mkSeqInstrs [
4670 ST TF tmp (spRel 0),
4673 return (Any IntRep code__2)
4675 #endif /* alpha_TARGET_ARCH */
4677 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
4679 #if i386_TARGET_ARCH
4681 coerceInt2FP from to x = do
4682 (x_reg, x_code) <- getSomeReg x
4684 opc = case to of W32 -> GITOF; W64 -> GITOD
4685 code dst = x_code `snocOL` opc x_reg dst
4686 -- ToDo: works for non-II32 reps?
4687 return (Any (floatSize to) code)
4691 coerceFP2Int from to x = do
4692 (x_reg, x_code) <- getSomeReg x
4694 opc = case from of W32 -> GFTOI; W64 -> GDTOI
4695 code dst = x_code `snocOL` opc x_reg dst
4696 -- ToDo: works for non-II32 reps?
4698 return (Any (intSize to) code)
4700 #endif /* i386_TARGET_ARCH */
4702 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
4704 #if x86_64_TARGET_ARCH
4706 coerceFP2Int from to x = do
4707 (x_op, x_code) <- getOperand x -- ToDo: could be a safe operand
4709 opc = case from of W32 -> CVTTSS2SIQ; W64 -> CVTTSD2SIQ
4710 code dst = x_code `snocOL` opc x_op dst
4712 return (Any (intSize to) code) -- works even if the destination rep is <II32
4714 coerceInt2FP from to x = do
4715 (x_op, x_code) <- getOperand x -- ToDo: could be a safe operand
4717 opc = case to of W32 -> CVTSI2SS; W64 -> CVTSI2SD
4718 code dst = x_code `snocOL` opc x_op dst
4720 return (Any (floatSize to) code) -- works even if the destination rep is <II32
4722 coerceFP2FP :: Width -> CmmExpr -> NatM Register
4723 coerceFP2FP to x = do
4724 (x_reg, x_code) <- getSomeReg x
4726 opc = case to of W32 -> CVTSD2SS; W64 -> CVTSS2SD
4727 code dst = x_code `snocOL` opc x_reg dst
4729 return (Any (floatSize to) code)
4732 -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
4734 #if sparc_TARGET_ARCH
4736 coerceInt2FP pk1 pk2 x = do
4737 (src, code) <- getSomeReg x
4739 code__2 dst = code `appOL` toOL [
4740 ST pk1 src (spRel (-2)),
4741 LD pk1 (spRel (-2)) dst,
4742 FxTOy pk1 pk2 dst dst]
4743 return (Any pk2 code__2)
4746 coerceFP2Int pk fprep x = do
4747 (src, code) <- getSomeReg x
4748 reg <- getNewRegNat fprep
4749 tmp <- getNewRegNat pk
4751 code__2 dst = ASSERT(fprep == FF64 || fprep == FF32)
4753 FxTOy fprep pk src tmp,
4754 ST pk tmp (spRel (-2)),
4755 LD pk (spRel (-2)) dst]
4756 return (Any pk code__2)
4759 coerceDbl2Flt x = do
4760 (src, code) <- getSomeReg x
4761 return (Any FF32 (\dst -> code `snocOL` FxTOy FF64 FF32 src dst))
4764 coerceFlt2Dbl x = do
4765 (src, code) <- getSomeReg x
4766 return (Any FF64 (\dst -> code `snocOL` FxTOy FF32 FF64 src dst))
4768 #endif /* sparc_TARGET_ARCH */
4770 #if powerpc_TARGET_ARCH
4771 coerceInt2FP fromRep toRep x = do
4772 (src, code) <- getSomeReg x
4773 lbl <- getNewLabelNat
4774 itmp <- getNewRegNat II32
4775 ftmp <- getNewRegNat FF64
4776 dflags <- getDynFlagsNat
4777 dynRef <- cmmMakeDynamicReference dflags addImportNat DataReference lbl
4778 Amode addr addr_code <- getAmode dynRef
4780 code' dst = code `appOL` maybe_exts `appOL` toOL [
4783 CmmStaticLit (CmmInt 0x43300000 W32),
4784 CmmStaticLit (CmmInt 0x80000000 W32)],
4785 XORIS itmp src (ImmInt 0x8000),
4786 ST II32 itmp (spRel 3),
4787 LIS itmp (ImmInt 0x4330),
4788 ST II32 itmp (spRel 2),
4789 LD FF64 ftmp (spRel 2)
4790 ] `appOL` addr_code `appOL` toOL [
4792 FSUB FF64 dst ftmp dst
4793 ] `appOL` maybe_frsp dst
4795 maybe_exts = case fromRep of
4796 W8 -> unitOL $ EXTS II8 src src
4797 W16 -> unitOL $ EXTS II16 src src
4799 maybe_frsp dst = case toRep of
4800 W32 -> unitOL $ FRSP dst dst
4802 return (Any (floatSize toRep) code')
4804 coerceFP2Int fromRep toRep x = do
4805 -- the reps don't really matter: F*->FF64 and II32->I* are no-ops
4806 (src, code) <- getSomeReg x
4807 tmp <- getNewRegNat FF64
4809 code' dst = code `appOL` toOL [
4810 -- convert to int in FP reg
4812 -- store value (64bit) from FP to stack
4813 ST FF64 tmp (spRel 2),
4814 -- read low word of value (high word is undefined)
4815 LD II32 dst (spRel 3)]
4816 return (Any (intSize toRep) code')
4817 #endif /* powerpc_TARGET_ARCH */
4820 -- -----------------------------------------------------------------------------
4821 -- eXTRA_STK_ARGS_HERE
4823 -- We (allegedly) put the first six C-call arguments in registers;
4824 -- where do we start putting the rest of them?
4826 -- Moved from MachInstrs (SDM):
4828 #if alpha_TARGET_ARCH || sparc_TARGET_ARCH
4829 eXTRA_STK_ARGS_HERE :: Int
4831 = IF_ARCH_alpha(0, IF_ARCH_sparc(23, ???))