1 Notes on new codegen (Sept 09)
\r
2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\r
5 - Top-level SRT threading is a bit ugly
\r
7 - Add type/newtype for CmmModule = [CmmGroup] -- A module
\r
8 CmmGroup = [CmmTop] -- A .o file
\r
9 CmmTop = Proc | Data -- A procedure or data
\r
11 - This is a *change*: currently a CmmGroup is one function's-worth of code
\r
12 regardless of SplitObjs. Question: can we *always* generate M.o if there
\r
13 is just one element in the list (rather than M/M1.o, M/M2.o etc)
\r
16 type CmmZ = GenCmm CmmStatic CmmInfo (CmmStackInfo, CmmGraph)
\r
18 type CmmZ = GenCmm CmmStatic (CmmInfo, CmmStackInfo) CmmGraph
\r
19 -- And perhaps take opportunity to prune CmmInfo?
\r
21 - Clarify which fields of CmmInfo are still used
\r
22 - Maybe get rid of CmmFormals arg of CmmProc in all versions?
\r
24 - We aren't sure whether cmmToRawCmm is actively used by the new pipeline; check
\r
25 And what does CmmBuildInfoTables do?!
\r
27 - Nuke CmmZipUtil, move zipPreds into ZipCfg
\r
29 - Pull out Areas into its own module
\r
30 Parameterise AreaMap
\r
32 type SubArea = (Area, ByteOff, ByteWidth)
\r
33 ByteOff should not be defined in SMRep -- that is too high up the hierarchy
\r
35 - Think about a non-flattened representation?
\r
38 * Use record fields for LastCall!
\r
39 * cml_ret_off should be a ByteOff
\r
41 LastCall (which has a successor) and
\r
42 LastJump (which does not, includes return?)
\r
43 - does not have cml_cont, cml_ret_args, cml_ret_off
\r
46 - expands into save/MidForeignCall/restore/goto
\r
47 - like any LastCall, target of the call gets an info table
\r
49 - JD: remind self of what goes wrong if you turn off the
\r
50 liveness of the update frame
\r
52 - Garbage-collect http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/CPS
\r
53 moving good stuff into
\r
54 http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/NewCodeGenPipeline
\r
57 - We believe that all of CmmProcPointZ.addProcPointProtocols is dead. What
\r
58 goes wrong if we simply never call it?
\r
60 - Something fishy in CmmStackLayout.hs
\r
61 * In particular, 'getAreaSize' returns an AreaMap, but we *know* the width of
\r
62 LocalRegs, so it'd be better to return FiniteMap AreaId ByteWidth
\r
63 * setSuccSPs looks fishy. Rather than lookin in procPoints, it could
\r
64 just lookup the block in areaSize which, after all, has a binding
\r
65 for precisely successors of calls. All other blocks (including proc
\r
66 points that are not successors of a call, we think) can be treated
\r
67 uniformly: zero-size Area, and use inSP.
\r
71 CmmProcPoint (Michael Adams)
\r
73 HscMain.optionallyConvertAndOrCPS
\r
75 DynFlags: -fconvert-to-zipper-and-back, -frun-cps, -frun-cpsz
\r
79 Consider this program, which has a diamond control flow,
\r
80 with a call on one branch
\r
83 if b then { ... f(x) ...; q=5; goto J }
\r
84 else { ...; q=7; goto J }
\r
87 then the join point J is a "proc-point". So, is 'p' passed to J
\r
88 as a parameter? Or, if 'p' was saved on the stack anyway, perhaps
\r
89 to keep it alive across the call to h(), maybe 'p' gets communicated
\r
90 to J that way. This is an awkward choice. (We think that we currently
\r
91 never pass variables to join points via arguments.)
\r
93 Furthermore, there is *no way* to pass q to J in a register (other
\r
94 than a paramter register).
\r
96 What we want is to do register allocation across the whole caboodle.
\r
97 Then we could drop all the code that deals with the above awkward
\r
98 decisions about spilling variables across proc-points.
\r
100 Note that J doesn't need an info table.
\r
102 What we really want is for each Block to have an optional info table.
\r
103 To do that, we need to be polymorphic over first nodes.
\r
105 Figuring out proc-points
\r
106 ~~~~~~~~~~~~~~~~~~~~~~~~
\r
107 Proc-points are identified by
\r
108 CmmProcPointZ.minimalProcPointSet/extendPPSet Although there isn't
\r
109 that much code, JD thinks that it could be done much more nicely using
\r
110 a dominator analysis, using the Dataflow Engine.
\r
112 ----------------------------------------------------
\r
113 Top-level structure
\r
114 ----------------------------------------------------
\r
116 * New codgen called in HscMain.hscGenHardCode, by calling HscMain.tryNewCodeGen,
\r
117 enabled by -fnew-codegen (Opt_TryNewCodeGen)
\r
119 THEN it calls CmmInfo.cmmToRawCmm to lay out the details of info tables
\r
120 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
121 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
123 * HscMain.tryNewCodeGen
\r
124 - STG->Cmm: StgCmm.codeGen (new codegen)
\r
125 - Optimise: CmmContFlowOpt (simple optimisations, very self contained)
\r
126 - Cps convert: CmmCPSZ.protoCmmCPSZ
\r
127 - Optimise: CmmContFlowOpt again
\r
128 - Convert: CmmCvt.cmmOfZgraph (convert to old rep) very self contained
\r
130 * StgCmm.hs The new STG -> Cmm conversion code generator
\r
131 Lots of modules StgCmmXXX
\r
134 ----------------------------------------------------
\r
135 CmmCPSZ.protoCmmCPSZ The new pipeline
\r
136 ----------------------------------------------------
\r
138 CmmCPSZprotoCmmCPSZ:
\r
139 1. Do cpsTop for each procedures separately
\r
140 2. Build SRT representation; this spans multiple procedures
\r
141 (unless split-objs)
\r
144 * CmmCommonBlockElimZ.elimCommonBlocks:
\r
145 eliminate common blocks
\r
147 * CmmProcPointZ.minimalProcPointSet
\r
148 identify proc-points
\r
150 * CmmProcPointZ.addProcPointProtocols
\r
151 something to do with the MA optimisation
\r
152 probably entirely unnecessary
\r
155 * Spill and reload:
\r
156 - CmmSpillReload.dualLivenessWithInsertion
\r
157 insert spills/reloads across
\r
159 Branches to proc-points
\r
160 Now sink those reloads:
\r
161 - CmmSpillReload.insertLateReloads
\r
162 - CmmSpillReload.removeDeadAssignmentsAndReloads
\r
164 * CmmStackLayout.stubSlotsOnDeath
\r
165 debug only: zero out dead slots when they die
\r
168 - CmmStackLayout.lifeSlotAnal:
\r
169 find which sub-areas are live on entry to each block
\r
171 - CmmStackLayout.layout
\r
172 Lay out the stack, returning an AreaMap
\r
173 type AreaMap = FiniteMap Area ByteOff
\r
174 -- Byte offset of the oldest byte of the Area,
\r
175 -- relative to the oldest byte of the Old Area
\r
177 - CmmStackLayout.manifestSP
\r
178 Manifest the stack pointer
\r
180 * Split into separate procedures
\r
181 - CmmProcPointZ.procPointAnalysis
\r
182 Given set of proc points, which blocks are reachable from each
\r
184 - CmmProcPointZ.splitAtProcPoints
\r
185 Using this info, split into separate procedures
\r
187 ----------------------------------------------------
\r
189 ----------------------------------------------------
\r
191 * The code for a procedure f may refer to either the *closure*
\r
192 or the *entry point* of another top-level procedure g.
\r
193 If f is live, then so is g. f's SRT must include g's closure.
\r
195 * The CLabel for the entry-point/closure reveals whether g is
\r
196 a CAF (or refers to CAFs). See the IdLabell constructor of CLabel.
\r
198 * The CAF-ness of the original top-level defininions is figured out
\r
199 (by TidyPgm) before we generate C--. This CafInfo is only set for
\r
200 top-level Ids; nested bindings stay with NoCafRefs.
\r
202 * Currently an SRT contains (only) pointers to (top-level) closures.
\r
204 * Consider this Core code
\r
205 f = \x -> let g = \y -> ...x...y...h1...
\r
207 and suppose that h1, h2 have IdInfo of MayHaveCafRefs.
\r
208 Therefore, so will f, But g will not (since it's nested).
\r
210 This generates C-- roughly like this:
\r
211 f_closure: .word f_entry
\r
212 f_entry() [info-tbl-for-f] { ...jump g_entry...jump h2... }
\r
213 g_entry() [info-tbl-for-g] { ...jump h1 }
\r
215 Note that there is no top-level closure for g (only an info table).
\r
216 So: info-tbl-for-f must have an SRT that keeps h1,h2 alive
\r
217 info-tbl-for-g must have an SRT that keeps h1 (only) alive
\r
219 But if we just look for the free CAF refs, we get:
\r
223 So we need to do a transitive closure thing to flesh out
\r
224 f's keep-alive refs to include h1.
\r
226 * The SRT info is the C_SRT field of Cmm.ClosureTypeInfo in a
\r
227 CmmInfoTable attached to each CmmProc. CmmCPSZ.toTops actually does
\r
228 the attaching, right at the end of the pipeline. The C_SRT part
\r
229 gives offsets within a single, shared table of closure pointers.
\r
231 ----------------------------------------------------
\r
233 ----------------------------------------------------
\r
235 See Note [Foreign calls] in ZipCfgCmmRep! This explains that a safe
\r
236 foreign call must do this:
\r
238 push info table (on thread stack) to describe frame
\r
239 make call (via C stack)
\r
241 restore thread state
\r
242 and explains why this expansion must be done late in the day.
\r
245 - Every foreign call is represented as a middle node
\r
247 - *Unsafe* foreign calls are simply "fat machine instructions"
\r
248 and are passed along to the native code generator
\r
250 - *Safe* foreign calls are "lowered" to unsafe calls by wrapping
\r
251 them in the above save/restore sequence. This step is done
\r
252 very late in the pipeline, just before handing to the native
\r
255 This lowering is done by BuildInfoTables.lowerSafeForeignCalls
\r
258 NEW PLAN for foreign calls:
\r
259 - Unsafe foreign calls remain as a middle node (fat machine instruction)
\r
260 Even the parameter passing is not lowered (just as machine instrs
\r
263 - Initially, safe foreign calls appear as LastCalls with
\r
266 ----------------------------------------------------
\r
267 Cmm representations
\r
268 ----------------------------------------------------
\r
271 The type [GenCmm d h g] represents a whole module,
\r
272 ** one list element per .o file **
\r
273 Without SplitObjs, the list has exactly one element
\r
275 newtype GenCmm d h g = Cmm [GenCmmTop d h g] -- A whole .o file
\r
276 data GenCmmTop d h g
\r
277 = CmmProc h g -- One procedure, graph d
\r
278 | CmmData <stuff> [d] -- Initialised data, items d
\r
280 Old and new piplines use different representations
\r
281 (CmmCvt.hs converts between the two)
\r
285 OLD BACK END representations (Cmm.hs):
\r
286 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
288 newtype ListGraph i = ListGraph [GenBasicBlock i]
\r
290 data CmmStmt = Assign | Store | Return etc -- OLD BACK END ONLY
\r
293 Once the info tables are laid out, we replace CmmInfo with [CmmStatic]
\r
294 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
295 which represents the info tables as data, that should
\r
296 immediately precede the code
\r
299 NEW BACK END representations
\r
300 * Not Cmm-specific at all
\r
301 ZipCfg.hs defines Graph, LGraph, FGraph,
\r
302 ZHead, ZTail, ZBlock ...
\r
304 classes LastNode, HavingSuccessors
\r
306 MkZipCfg.hs: AGraph: building graphs
\r
308 * ZipCfgCmmRep: instantiates ZipCfg for Cmm
\r
309 data Middle = ...CmmExpr...
\r
310 data Last = ...CmmExpr...
\r
311 type CmmGraph = Graph Middle Last
\r
313 type CmmZ = GenCmm CmmStatic CmmInfo (CmmStackInfo, CmmGraph)
\r
314 type CmmStackInfo = (ByteOff, Maybe ByteOff)
\r
315 -- (SP offset on entry, update frame space = SP offset on exit)
\r
316 -- The new codegen produces CmmZ, but once the stack is
\r
317 -- manifested we can drop that in favour of
\r
318 -- GenCmm CmmStatic CmmInfo CmmGraph
\r
322 - CmmInfo: partly used by NEW
\r
323 - CmmFormals: not used at all PERHAPS NOT EVEN BY OLD PIPELINE!
\r
325 * MkZipCfgCmm.hs: smart constructors for ZipCfgCmmRep
\r
326 Depends on (a) MkZipCfg (Cmm-independent)
\r
327 (b) ZipCfgCmmRep (Cmm-specific)
\r
331 CmmExpr.hs defines the Cmm expression types
\r
332 - CmmExpr, CmmReg, Width, CmmLit, LocalReg, GlobalReg
\r
333 - CmmType, Width etc (saparate module?)
\r
334 - MachOp (separate module?)
\r
335 - Area, AreaId etc (separate module?)
\r
337 BlockId.hs defines BlockId, BlockEnv, BlockSet
\r
343 * Transactions indicate whether or not the result changes: CmmTx
\r
344 type Tx a = a -> TxRes a
\r
345 data TxRes a = TxRes ChangeFlag a
\r