2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 % $Id: AbsCSyn.lhs,v 1.44 2002/01/02 12:32:19 simonmar Exp $
6 \section[AbstractC]{Abstract C: the last stop before machine code}
8 This ``Abstract C'' data type describes the raw Spineless Tagless
9 machine model at a C-ish level; it is ``abstract'' in that it only
10 includes C-like structures that we happen to need. The conversion of
11 programs from @StgSyntax@ (basically a functional language) to
12 @AbstractC@ (basically imperative C) is the heart of code generation.
13 From @AbstractC@, one may convert to real C (for portability) or to
14 raw assembler/machine code.
25 mkAbstractCs, mkAbsCStmts, mkAlgAltsCSwitch,
34 MagicId(..), node, infoptr,
39 #include "HsVersions.h"
41 import {-# SOURCE #-} ClosureInfo ( ClosureInfo )
44 import Constants ( mAX_Vanilla_REG, mAX_Float_REG,
45 mAX_Double_REG, spRelToInt )
46 import CostCentre ( CostCentre, CostCentreStack )
47 import Literal ( mkMachInt, Literal(..) )
48 import ForeignCall ( CCallSpec )
49 import PrimRep ( PrimRep(..) )
50 import MachOp ( MachOp(..) )
51 import Unique ( Unique )
52 import StgSyn ( StgOp )
53 import TyCon ( TyCon )
54 import BitSet -- for liveness masks
60 @AbstractC@ is a list of Abstract~C statements, but the data structure
61 is tree-ish, for easier and more efficient putting-together.
67 | AbsCStmts AbstractC AbstractC
69 -- and the individual stmts...
72 A note on @CAssign@: In general, the type associated with an assignment
73 is the type of the lhs. However, when the lhs is a pointer to mixed
74 types (e.g. SpB relative), the type of the assignment is the type of
75 the rhs for float types, or the generic StgWord for all other types.
76 (In particular, a CharRep on the rhs is promoted to IntRep when
77 stored in a mixed type location.)
85 CAddrMode -- Put this in the program counter
86 -- eg `CJump (CReg (VanillaReg PtrRep 1))' puts Ret1 in PC
87 -- Enter can be done by:
88 -- CJump (CVal NodeRel zeroOff)
91 CAddrMode -- Fall through into this routine
92 -- (for the benefit of the native code generators)
93 -- Equivalent to CJump in C land
95 | CReturn -- Perform a return
96 CAddrMode -- Address of a RET_<blah> info table
97 ReturnInfo -- Whether it's a direct or vectored return
100 [(Literal, AbstractC)] -- alternatives
101 AbstractC -- default; if there is no real Abstract C in here
102 -- (e.g., all comments; see function "nonemptyAbsC"),
103 -- then that means the default _cannot_ occur.
104 -- If there is only one alternative & no default code,
105 -- then there is no need to check the tag.
107 -- CSwitch m [(tag,code)] AbsCNop == code
109 | CCodeBlock CLabel AbstractC
110 -- A labelled block of code; this "statement" is not
111 -- executed; rather, the labelled code will be hoisted
112 -- out to the top level (out of line) & it can be
115 | CInitHdr -- to initialise the header of a closure (both fixed/var parts)
117 CAddrMode -- address of the info ptr
118 !CAddrMode -- cost centre to place in closure
119 -- CReg CurCostCentre or CC_HDR(R1.p{-Node-})
120 Int -- size of closure, for profiling
122 -- NEW CASES FOR EXPANDED PRIMOPS
124 | CMachOpStmt -- Machine-level operation
125 (Maybe CAddrMode) -- 0 or 1 results
127 [CAddrMode] -- Arguments
128 (Maybe [MagicId]) -- list of regs which need to be preserved
129 -- across the primop. This is allowed to be Nothing only if
130 -- machOpIsDefinitelyInline returns True. And that in turn may
131 -- only return True if we are absolutely sure that the mach op
132 -- can be done inline on all platforms.
134 | CSequential -- Do the nested AbstractCs sequentially.
135 [AbstractC] -- In particular, as far as the AbsCUtils.doSimultaneously
136 -- is concerned, these stmts are to be treated as atomic
137 -- and are not to be reordered.
139 -- end of NEW CASES FOR EXPANDED PRIMOPS
142 [CAddrMode] -- Results
144 [CAddrMode] -- Arguments
145 [MagicId] -- Potentially volatile/live registers
146 -- (to save/restore around the call/op)
148 -- INVARIANT: When a PrimOp which can cause GC is used, the
149 -- only live data is tidily on the STG stacks or in the STG
150 -- registers (the code generator ensures this).
152 -- Why this? Because if the arguments were arbitrary
153 -- addressing modes, they might be things like (Hp+6) which
154 -- will get utterly spongled by GC.
156 | CSimultaneous -- Perform simultaneously all the statements
157 AbstractC -- in the nested AbstractC. They are only
158 -- allowed to be CAssigns, COpStmts and AbsCNops, so the
159 -- "simultaneous" part just concerns making
160 -- sure that permutations work.
161 -- For example { a := b, b := a }
162 -- needs to go via (at least one) temporary
164 | CCheck -- heap or stack checks, or both.
165 CCheckMacro -- These might include some code to fill in tags
166 [CAddrMode] -- on the stack, so we can't use CMacroStmt below.
169 | CRetDirect -- Direct return
170 !Unique -- for making labels
171 AbstractC -- return code
173 Liveness -- stack liveness at the return point
175 -- see the notes about these next few; they follow below...
176 | CMacroStmt CStmtMacro [CAddrMode]
177 | CCallProfCtrMacro FAST_STRING [CAddrMode]
178 | CCallProfCCMacro FAST_STRING [CAddrMode]
180 {- The presence of this constructor is a makeshift solution;
181 it being used to work around a gcc-related problem of
182 handling typedefs within statement blocks (or, rather,
183 the inability to do so.)
185 The AbstractC flattener takes care of lifting out these
186 typedefs if needs be (i.e., when generating .hc code and
187 compiling 'foreign import dynamic's)
189 | CCallTypedef Bool {- True => use "typedef"; False => use "extern"-}
190 CCallSpec Unique [CAddrMode] [CAddrMode]
192 -- *** the next three [or so...] are DATA (those above are CODE) ***
195 ClosureInfo -- Todo: maybe info_lbl & closure_lbl instead?
196 CAddrMode -- cost centre identifier to place in closure
197 [CAddrMode] -- free vars; ptrs, then non-ptrs.
199 | CSRT CLabel [CLabel] -- SRT declarations: basically an array of
200 -- pointers to static closures.
202 | CBitmap CLabel LivenessMask -- A bitmap to be emitted if and only if
203 -- it is larger than a target machine word.
205 | CClosureInfoAndCode
206 ClosureInfo -- Explains placement and layout of closure
207 AbstractC -- Slow entry point code
209 -- Fast entry point code, if any
210 String -- Closure description; NB we can't get this
211 -- from ClosureInfo, because the latter refers
212 -- to the *right* hand side of a defn, whereas
213 -- the "description" refers to *left* hand side
215 | CRetVector -- A labelled block of static data
219 Liveness -- stack liveness at the return point
221 | CClosureTbl -- table of constructors for enumerated types
222 TyCon -- which TyCon this table is for
224 | CModuleInitBlock -- module initialisation block
225 CLabel -- label for init block
226 AbstractC -- initialisation code
228 | CCostCentreDecl -- A cost centre *declaration*
229 Bool -- True <=> local => full declaration
230 -- False <=> extern; just say so
233 | CCostCentreStackDecl -- A cost centre stack *declaration*
234 CostCentreStack -- this is the declaration for a
235 -- pre-defined singleton CCS (see
238 | CSplitMarker -- Split into separate object modules here
240 -- C_SRT is what StgSyn.SRT gets translated to...
241 -- we add a label for the table, and expect only the 'offset/length' form
244 | C_SRT CLabel !Int{-offset-} !Int{-length-}
246 needsSRT :: C_SRT -> Bool
247 needsSRT NoC_SRT = False
248 needsSRT (C_SRT _ _ _) = True
251 About @CMacroStmt@, etc.: notionally, they all just call some
252 arbitrary C~macro or routine, passing the @CAddrModes@ as arguments.
253 However, we distinguish between various flavours of these things,
254 mostly just to keep things somewhat less wild and wooly.
258 Some {\em essential} bits of the STG execution model are done with C
259 macros. An example is @STK_CHK@, which checks for stack-space
260 overflow. This enumeration type lists all such macros:
263 = ARGS_CHK -- arg satisfaction check
264 | ARGS_CHK_LOAD_NODE -- arg check for top-level functions
265 | UPD_CAF -- update CAF closure with indirection
266 | UPD_BH_UPDATABLE -- eager backholing
267 | UPD_BH_SINGLE_ENTRY -- more eager blackholing
268 | PUSH_UPD_FRAME -- push update frame
269 | PUSH_SEQ_FRAME -- push seq frame
270 | UPDATE_SU_FROM_UPD_FRAME -- pull Su out of the update frame
271 | SET_TAG -- set TagReg if it exists
273 | REGISTER_FOREIGN_EXPORT -- register a foreign exported fun
274 | REGISTER_IMPORT -- register an imported module
275 | REGISTER_DIMPORT -- register an imported module from
278 | GRAN_FETCH -- for GrAnSim only -- HWL
279 | GRAN_RESCHEDULE -- for GrAnSim only -- HWL
280 | GRAN_FETCH_AND_RESCHEDULE -- for GrAnSim only -- HWL
281 | THREAD_CONTEXT_SWITCH -- for GrAnSim only -- HWL
282 | GRAN_YIELD -- for GrAnSim only -- HWL
285 Heap/Stack checks. There are far too many of these.
290 = HP_CHK_NP -- heap/stack checks when
291 | STK_CHK_NP -- node points to the closure
293 | HP_CHK_SEQ_NP -- for 'seq' style case alternatives
295 | HP_CHK -- heap/stack checks when
296 | STK_CHK -- node doesn't point
298 -- case alternative heap checks:
300 | HP_CHK_NOREGS -- no registers live
301 | HP_CHK_UNPT_R1 -- R1 is boxed/unlifted
302 | HP_CHK_UNBX_R1 -- R1 is unboxed
303 | HP_CHK_F1 -- FloatReg1 (only) is live
304 | HP_CHK_D1 -- DblReg1 (only) is live
305 | HP_CHK_L1 -- LngReg1 (only) is live
306 | HP_CHK_UT_ALT -- unboxed tuple return.
308 | HP_CHK_GEN -- generic heap check
311 \item[@CCallProfCtrMacro@:]
312 The @String@ names a macro that, if \tr{#define}d, will bump one/some
313 of the STG-event profiling counters.
315 \item[@CCallProfCCMacro@:]
316 The @String@ names a macro that, if \tr{#define}d, will perform some
317 cost-centre-profiling-related action.
320 %************************************************************************
322 \subsection[CAddrMode]{C addressing modes}
324 %************************************************************************
328 = CVal RegRelative PrimRep
329 -- On RHS of assign: Contents of Magic[n]
330 -- On LHS of assign: location Magic[n]
331 -- (ie at addr Magic+n)
334 -- On RHS of assign: Address of Magic[n]; ie Magic+n
335 -- n=0 gets the Magic location itself
336 -- (NB: n=0 case superceded by CReg)
337 -- On LHS of assign: only sensible if n=0,
338 -- which gives the magic location itself
339 -- (NB: superceded by CReg)
341 | CReg MagicId -- To replace (CAddr MagicId 0)
343 | CTemp !Unique !PrimRep -- Temporary locations
344 -- ``Temporaries'' correspond to local variables in C, and registers in
347 | CLbl CLabel -- Labels in the runtime system, etc.
348 PrimRep -- the kind is so we can generate accurate C decls
350 | CCharLike CAddrMode -- The address of a static char-like closure for
351 -- the specified character. It is guaranteed to be in
352 -- the range mIN_CHARLIKE..mAX_CHARLIKE
354 | CIntLike CAddrMode -- The address of a static int-like closure for the
355 -- specified small integer. It is guaranteed to be in
356 -- the range mIN_INTLIKE..mAX_INTLIKE
360 | CJoinPoint -- This is used as the amode of a let-no-escape-bound
362 VirtualSpOffset -- Sp value after any volatile free vars
363 -- of the rhs have been saved on stack.
364 -- Just before the code for the thing is jumped to,
365 -- Sp will be set to this value,
366 -- and then any stack-passed args pushed,
367 -- then the code for this thing will be entered
369 !PrimRep -- the kind of the result
370 CExprMacro -- the macro to generate a value
371 [CAddrMode] -- and its arguments
373 | CMem PrimRep -- A value :: PrimRep, in memory, at the
374 CAddrMode -- specified address
376 | CBytesPerWord -- Word size, in bytes, on this platform
377 -- required for: half-word loads (used in fishing tags
378 -- out of info tables), and sizeofByteArray#.
381 Various C macros for values which are dependent on the back-end layout.
387 | ARG_TAG -- stack argument tagging
388 | GET_TAG -- get current constructor tag
394 Convenience functions:
397 mkIntCLit :: Int -> CAddrMode
398 mkIntCLit i = CLit (mkMachInt (toInteger i))
400 mkCString :: FAST_STRING -> CAddrMode
401 mkCString s = CLit (MachStr s)
403 mkCCostCentre :: CostCentre -> CAddrMode
404 mkCCostCentre cc = CLbl (mkCC_Label cc) DataPtrRep
406 mkCCostCentreStack :: CostCentreStack -> CAddrMode
407 mkCCostCentreStack ccs = CLbl (mkCCS_Label ccs) DataPtrRep
410 %************************************************************************
412 \subsection[RegRelative]{@RegRelatives@: ???}
414 %************************************************************************
419 | SpRel FastInt -- }- offsets in StgWords
420 | NodeRel FastInt -- }
421 | CIndex CAddrMode CAddrMode PrimRep -- pointer arithmetic :-)
422 -- CIndex a b k === (k*)a[b]
425 = DirectReturn -- Jump directly, if possible
426 | StaticVectoredReturn Int -- Fixed tag, starting at zero
427 | DynamicVectoredReturn CAddrMode -- Dynamic tag given by amode, starting at zero
429 hpRel :: VirtualHeapOffset -- virtual offset of Hp
430 -> VirtualHeapOffset -- virtual offset of The Thing
431 -> RegRelative -- integer offset
432 hpRel hp off = HpRel (iUnbox (hp - off))
434 spRel :: VirtualSpOffset -- virtual offset of Sp
435 -> VirtualSpOffset -- virtual offset of The Thing
436 -> RegRelative -- integer offset
437 spRel sp off = SpRel (iUnbox (spRelToInt sp off))
439 nodeRel :: VirtualHeapOffset
441 nodeRel off = NodeRel (iUnbox off)
445 %************************************************************************
447 \subsection[Liveness]{Liveness Masks}
449 %************************************************************************
451 We represent liveness bitmaps as a BitSet (whose internal
452 representation really is a bitmap). These are pinned onto case return
453 vectors to indicate the state of the stack for the garbage collector.
455 In the compiled program, liveness bitmaps that fit inside a single
456 word (StgWord) are stored as a single word, while larger bitmaps are
457 stored as a pointer to an array of words. When we compile via C
458 (especially when we bootstrap via HC files), we generate identical C
459 code regardless of whether words are 32- or 64-bit on the target
460 machine, by postponing the decision of how to store each liveness
461 bitmap to C compilation time (or rather, C preprocessing time).
464 type LivenessMask = [BitSet]
466 data Liveness = Liveness CLabel LivenessMask
469 %************************************************************************
471 \subsection[HeapOffset]{@Heap Offsets@}
473 %************************************************************************
475 This used to be a grotesquely complicated datatype in an attempt to
476 hide the details of header sizes from the compiler itself. Now these
477 constants are imported from the RTS, and we deal in real Ints.
480 type HeapOffset = Int -- ToDo: remove
482 type VirtualHeapOffset = HeapOffset
483 type VirtualSpOffset = Int
485 type HpRelOffset = HeapOffset
486 type SpRelOffset = Int
489 %************************************************************************
491 \subsection[MagicId]{@MagicIds@: registers and such}
493 %************************************************************************
497 = BaseReg -- mentioned only in nativeGen
499 -- Argument and return registers
500 | VanillaReg -- pointers, unboxed ints and chars
502 FastInt -- its number (1 .. mAX_Vanilla_REG)
504 | FloatReg -- single-precision floating-point registers
505 FastInt -- its number (1 .. mAX_Float_REG)
507 | DoubleReg -- double-precision floating-point registers
508 FastInt -- its number (1 .. mAX_Double_REG)
511 | Sp -- Stack ptr; points to last occupied stack location.
512 | Su -- Stack update frame pointer
513 | SpLim -- Stack limit
514 | Hp -- Heap ptr; points to last occupied heap location.
515 | HpLim -- Heap limit register
516 | CurCostCentre -- current cost centre register.
517 | VoidReg -- see "VoidPrim" type; just a placeholder;
518 -- no actual register
519 | LongReg -- long int registers (64-bit, really)
520 PrimRep -- Int64Rep or Word64Rep
521 FastInt -- its number (1 .. mAX_Long_REG)
523 | CurrentTSO -- pointer to current thread's TSO
524 | CurrentNursery -- pointer to allocation area
525 | HpAlloc -- allocation count for heap check failure
528 node = VanillaReg PtrRep (_ILIT 1) -- A convenient alias for Node
529 tagreg = VanillaReg WordRep (_ILIT 2) -- A convenient alias for TagReg
534 We need magical @Eq@ because @VanillaReg@s come in multiple flavors.
537 instance Eq MagicId where
538 reg1 == reg2 = tag reg1 ==# tag reg2
540 tag BaseReg = (_ILIT(0) :: FastInt)
546 tag CurCostCentre = _ILIT(6)
547 tag VoidReg = _ILIT(7)
549 tag (VanillaReg _ i) = _ILIT(8) +# i
551 tag (FloatReg i) = _ILIT(8) +# maxv +# i
552 tag (DoubleReg i) = _ILIT(8) +# maxv +# maxf +# i
553 tag (LongReg _ i) = _ILIT(8) +# maxv +# maxf +# maxd +# i
555 maxv = iUnbox mAX_Vanilla_REG
556 maxf = iUnbox mAX_Float_REG
557 maxd = iUnbox mAX_Double_REG
560 Returns True for any register that {\em potentially} dies across
561 C calls (or anything near equivalent). We just say @True@ and
562 let the (machine-specific) registering macros sort things out...
565 isVolatileReg :: MagicId -> Bool
566 isVolatileReg any = True