1 -----------------------------------------------------------------------------
5 -- (c) The University of Glasgow 2004-2006
7 -----------------------------------------------------------------------------
11 GenCmmTop(..), CmmTop,
12 GenBasicBlock(..), CmmBasicBlock, blockId, blockStmts,
13 CmmStmt(..), CmmActuals, CmmFormals,
15 CmmStatic(..), Section(..),
16 CmmExpr(..), cmmExprRep,
17 CmmReg(..), cmmRegRep,
18 CmmLit(..), cmmLitRep,
19 LocalReg(..), localRegRep,
21 GlobalReg(..), globalRegRep,
23 node, nodeReg, spReg, hpReg,
26 #include "HsVersions.h"
36 -----------------------------------------------------------------------------
37 -- Cmm, CmmTop, CmmBasicBlock
38 -----------------------------------------------------------------------------
40 -- A file is a list of top-level chunks. These may be arbitrarily
41 -- re-orderd during code generation.
43 -- GenCmm is abstracted over
44 -- (a) the type of static data elements
45 -- (b) the contents of a basic block.
46 -- We expect there to be two main instances of this type:
47 -- (a) Plain C--, i.e. populated with CmmLit and CmmExpr respectively,
48 -- (b) Native code, populated with instructions
50 newtype GenCmm d i = Cmm [GenCmmTop d i]
52 type Cmm = GenCmm CmmStatic CmmStmt
54 -- A top-level chunk, abstracted over the type of the contents of
55 -- the basic blocks (Cmm or instructions are the likely instantiations).
58 [d] -- Info table, may be empty
59 CLabel -- Used to generate both info & entry labels
60 [LocalReg] -- Argument locals live on entry (C-- procedure params)
61 [GenBasicBlock i] -- Code, may be empty. The first block is
62 -- the entry point. The order is otherwise initially
63 -- unimportant, but at some point the code gen will
66 -- the BlockId of the first block does not give rise
67 -- to a label. To jump to the first block in a Proc,
68 -- use the appropriate CLabel.
71 | CmmData Section [d] -- constant values only
73 type CmmTop = GenCmmTop CmmStatic CmmStmt
75 -- A basic block containing a single label, at the beginning.
76 -- The list of basic blocks in a top-level code block may be re-ordered.
77 -- Fall-through is not allowed: there must be an explicit jump at the
78 -- end of each basic block, but the code generator might rearrange basic
79 -- blocks in order to turn some jumps into fallthroughs.
81 data GenBasicBlock i = BasicBlock BlockId [i]
82 -- ToDo: Julian suggests that we might need to annotate this type
83 -- with the out & in edges in the graph, i.e. two * [BlockId]. This
84 -- information can be derived from the contents, but it might be
85 -- helpful to cache it here.
87 type CmmBasicBlock = GenBasicBlock CmmStmt
89 blockId :: GenBasicBlock i -> BlockId
90 -- The branch block id is that of the first block in
91 -- the branch, which is that branch's entry point
92 blockId (BasicBlock blk_id _ ) = blk_id
94 blockStmts :: GenBasicBlock i -> [i]
95 blockStmts (BasicBlock _ stmts) = stmts
98 -----------------------------------------------------------------------------
100 -- A "statement". Note that all branches are explicit: there are no
101 -- control transfers to computed addresses, except when transfering
102 -- control to a new function.
103 -----------------------------------------------------------------------------
107 | CmmComment FastString
109 | CmmAssign CmmReg CmmExpr -- Assign to register
111 | CmmStore CmmExpr CmmExpr -- Assign to memory location. Size is
112 -- given by cmmExprRep of the rhs.
114 | CmmCall -- A foreign call, with
116 CmmFormals -- zero or more results
117 CmmActuals -- zero or more arguments
118 (Maybe [GlobalReg]) -- Global regs that may need to be saved
119 -- if they will be clobbered by the call.
120 -- Nothing <=> save *all* globals that
121 -- might be clobbered.
123 | CmmBranch BlockId -- branch to another BB in this fn
125 | CmmCondBranch CmmExpr BlockId -- conditional branch
127 | CmmSwitch CmmExpr [Maybe BlockId] -- Table branch
128 -- The scrutinee is zero-based;
129 -- zero -> first block
130 -- one -> second block etc
131 -- Undefined outside range, and when there's a Nothing
133 | CmmJump CmmExpr -- Jump to another function,
134 CmmActuals -- with these parameters.
136 | CmmReturn -- Return from a function,
137 CmmActuals -- with these return values.
139 type CmmActuals = [(CmmExpr,MachHint)]
140 type CmmFormals = [(CmmReg,MachHint)]
146 One possible problem with the above type is that the only way to do a
147 non-local conditional jump is to encode it as a branch to a block that
148 contains a single jump. This leads to inefficient code in the back end.
150 One possible way to fix this would be:
154 | CmmJump CmmBranchDest
155 | CmmCondJump CmmExpr CmmBranchDest
160 | NonLocal CmmExpr [LocalReg]
164 + one fewer constructors in CmmStmt
165 + allows both cond branch and switch to jump to non-local destinations
169 - not strictly necessary: can already encode as branch+jump
170 - not always possible to implement any better in the back end
171 - could do the optimisation in the back end (but then plat-specific?)
172 - C-- doesn't have it
173 - back-end optimisation might be more general (jump shortcutting)
175 So we'll stick with the way it is, and add the optimisation to the NCG.
178 -----------------------------------------------------------------------------
181 -- The target of a CmmCall.
182 -----------------------------------------------------------------------------
185 = CmmForeignCall -- Call to a foreign function
186 CmmExpr -- literal label <=> static call
187 -- other expression <=> dynamic call
188 CCallConv -- The calling convention
190 | CmmPrim -- Call to a "primitive" (eg. sin, cos)
191 CallishMachOp -- These might be implemented as inline
192 -- code by the backend.
194 -----------------------------------------------------------------------------
196 -- An expression. Expressions have no side effects.
197 -----------------------------------------------------------------------------
200 = CmmLit CmmLit -- Literal
201 | CmmLoad CmmExpr MachRep -- Read memory location
202 | CmmReg CmmReg -- Contents of register
203 | CmmMachOp MachOp [CmmExpr] -- Machine operation (+, -, *, etc.)
204 | CmmRegOff CmmReg Int
206 -- ** is shorthand only, meaning **
207 -- CmmMachOp (MO_S_Add rep (CmmReg reg) (CmmLit (CmmInt i rep)))
208 -- where rep = cmmRegRep reg
211 cmmExprRep :: CmmExpr -> MachRep
212 cmmExprRep (CmmLit lit) = cmmLitRep lit
213 cmmExprRep (CmmLoad _ rep) = rep
214 cmmExprRep (CmmReg reg) = cmmRegRep reg
215 cmmExprRep (CmmMachOp op _) = resultRepOfMachOp op
216 cmmExprRep (CmmRegOff reg _) = cmmRegRep reg
220 | CmmGlobal GlobalReg
223 cmmRegRep :: CmmReg -> MachRep
224 cmmRegRep (CmmLocal reg) = localRegRep reg
225 cmmRegRep (CmmGlobal reg) = globalRegRep reg
228 = LocalReg !Unique MachRep
230 instance Eq LocalReg where
231 (LocalReg u1 _) == (LocalReg u2 _) = u1 == u2
233 instance Uniquable LocalReg where
234 getUnique (LocalReg uniq _) = uniq
236 localRegRep :: LocalReg -> MachRep
237 localRegRep (LocalReg _ rep) = rep
240 = CmmInt Integer MachRep
241 -- Interpretation: the 2's complement representation of the value
242 -- is truncated to the specified size. This is easier than trying
243 -- to keep the value within range, because we don't know whether
244 -- it will be used as a signed or unsigned value (the MachRep doesn't
245 -- distinguish between signed & unsigned).
246 | CmmFloat Rational MachRep
247 | CmmLabel CLabel -- Address of label
248 | CmmLabelOff CLabel Int -- Address of label + byte offset
250 -- Due to limitations in the C backend, the following
251 -- MUST ONLY be used inside the info table indicated by label2
252 -- (label2 must be the info label), and label1 must be an
253 -- SRT, a slow entrypoint or a large bitmap (see the Mangler)
254 -- Don't use it at all unless tablesNextToCode.
255 -- It is also used inside the NCG during when generating
256 -- position-independent code.
257 | CmmLabelDiffOff CLabel CLabel Int -- label1 - label2 + offset
260 cmmLitRep :: CmmLit -> MachRep
261 cmmLitRep (CmmInt _ rep) = rep
262 cmmLitRep (CmmFloat _ rep) = rep
263 cmmLitRep (CmmLabel _) = wordRep
264 cmmLitRep (CmmLabelOff _ _) = wordRep
265 cmmLitRep (CmmLabelDiffOff _ _ _) = wordRep
267 -----------------------------------------------------------------------------
270 -- Local labels must be unique within a single compilation unit.
272 newtype BlockId = BlockId Unique
275 instance Uniquable BlockId where
276 getUnique (BlockId u) = u
278 -----------------------------------------------------------------------------
280 -----------------------------------------------------------------------------
286 | RelocatableReadOnlyData
288 | ReadOnlyData16 -- .rodata.cst16 on x86_64, 16-byte aligned
289 | OtherSection String
292 = CmmStaticLit CmmLit
293 -- a literal value, size given by cmmLitRep of the literal.
294 | CmmUninitialised Int
295 -- uninitialised data, N bytes long
297 -- align to next N-byte boundary (N must be a power of 2).
298 | CmmDataLabel CLabel
299 -- label the current position in this section.
301 -- string of 8-bit values only, not zero terminated.
303 -----------------------------------------------------------------------------
304 -- Global STG registers
305 -----------------------------------------------------------------------------
308 -- Argument and return registers
309 = VanillaReg -- pointers, unboxed ints and chars
310 {-# UNPACK #-} !Int -- its number
312 | FloatReg -- single-precision floating-point registers
313 {-# UNPACK #-} !Int -- its number
315 | DoubleReg -- double-precision floating-point registers
316 {-# UNPACK #-} !Int -- its number
318 | LongReg -- long int registers (64-bit, really)
319 {-# UNPACK #-} !Int -- its number
322 | Sp -- Stack ptr; points to last occupied stack location.
323 | SpLim -- Stack limit
324 | Hp -- Heap ptr; points to last occupied heap location.
325 | HpLim -- Heap limit register
326 | CurrentTSO -- pointer to current thread's TSO
327 | CurrentNursery -- pointer to allocation area
328 | HpAlloc -- allocation count for heap check failure
330 -- We keep the address of some commonly-called
331 -- functions in the register table, to keep code
333 | GCEnter1 -- stg_gc_enter_1
334 | GCFun -- stg_gc_fun
336 -- Base offset for the register table, used for accessing registers
337 -- which do not have real registers assigned to them. This register
338 -- will only appear after we have expanded GlobalReg into memory accesses
339 -- (where necessary) in the native code generator.
342 -- Base Register for PIC (position-independent code) calculations
343 -- Only used inside the native code generator. It's exact meaning differs
344 -- from platform to platform (see module PositionIndependentCode).
353 -- convenient aliases
354 spReg, hpReg, nodeReg :: CmmReg
357 nodeReg = CmmGlobal node
362 globalRegRep :: GlobalReg -> MachRep
363 globalRegRep (VanillaReg _) = wordRep
364 globalRegRep (FloatReg _) = F32
365 globalRegRep (DoubleReg _) = F64
366 globalRegRep (LongReg _) = I64
367 globalRegRep _ = wordRep