1 Notes on new codegen (Sept 09)
\r
2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\r
6 - CmmContFlowOpt.runCmmContFlowOptZs is not called!
\r
7 - Why is runCmmOpts called from HscMain? Seems too "high up".
\r
8 In fact HscMain calls (runCmmOpts cmmCfgOptsZ) which is what
\r
9 runCmmContFlowOptZs does. Tidy up!
\r
12 - AsmCodeGen has a generic Cmm optimiser; move this into new pipeline
\r
14 - AsmCodeGen has post-native-cg branch elimiator (shortCutBranches);
\r
15 we ultimately want to share this with the Cmm branch eliminator.
\r
17 - At the moment, references to global registers like Hp are "lowered"
\r
18 late (in AsmCodeGen.fixAssignTop and cmmToCmm). We should do this
\r
19 early, in the new native codegen, much in the way that we lower
\r
20 calling conventions. Might need to be a bit sophisticated about
\r
23 - Refactor Cmm so that it contains only shared stuff
\r
24 Add a module MoribundCmm which contains stuff from
\r
25 Cmm for old code gen path
\r
27 - Question: currently we lift procpoints to become separate
\r
28 CmmProcs. Do we still want to do this?
\r
30 NB: and advantage of continuing to do this is that
\r
31 we can do common-proc elimination!
\r
33 - Move to new Cmm rep:
\r
34 * Make native CG consume New Cmm;
\r
35 * Convert Old Cmm->New Cmm to keep old path alive
\r
36 * Produce New Cmm when reading in .cmm files
\r
38 - Consider module names
\r
40 - Top-level SRT threading is a bit ugly
\r
42 - Add type/newtype for CmmModule = [CmmGroup] -- A module
\r
43 CmmGroup = [CmmTop] -- A .o file
\r
44 CmmTop = Proc | Data -- A procedure or data
\r
46 - This is a *change*: currently a CmmGroup is one function's-worth of code
\r
47 regardless of SplitObjs. Question: can we *always* generate M.o if there
\r
48 is just one element in the list (rather than M/M1.o, M/M2.o etc)
\r
52 - See "CAFs" below; we want to totally refactor the way SRTs are calculated
\r
55 type CmmZ = GenCmm CmmStatic CmmInfo (CmmStackInfo, CmmGraph)
\r
57 type CmmZ = GenCmm CmmStatic (CmmInfo, CmmStackInfo) CmmGraph
\r
58 -- And perhaps take opportunity to prune CmmInfo?
\r
60 - Clarify which fields of CmmInfo are still used
\r
61 - Maybe get rid of CmmFormals arg of CmmProc in all versions?
\r
63 - We aren't sure whether cmmToRawCmm is actively used by the new pipeline; check
\r
64 And what does CmmBuildInfoTables do?!
\r
66 - Nuke CmmZipUtil, move zipPreds into ZipCfg
\r
68 - Pull out Areas into its own module
\r
69 Parameterise AreaMap
\r
71 type SubArea = (Area, ByteOff, ByteWidth)
\r
72 ByteOff should not be defined in SMRep -- that is too high up the hierarchy
\r
74 - SMRep should not be imported by any module in cmm/! Make it so.
\r
75 -- ByteOff etc ==> CmmExpr
\r
76 -- rET_SMALL etc ==> CmmInfo
\r
77 Check that there are no other imports from codeGen in cmm/
\r
79 - Think about a non-flattened representation?
\r
82 * Use record fields for LastCall!
\r
83 * cml_ret_off should be a ByteOff
\r
85 LastCall (which has a successor) and
\r
86 LastJump (which does not, includes return?)
\r
87 - does not have cml_cont, cml_ret_args, cml_ret_off
\r
90 - expands into save/MidForeignCall/restore/goto
\r
91 - like any LastCall, target of the call gets an info table
\r
93 - JD: remind self of what goes wrong if you turn off the
\r
94 liveness of the update frame
\r
96 - Garbage-collect http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/CPS
\r
97 moving good stuff into
\r
98 http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/NewCodeGenPipeline
\r
101 - We believe that all of CmmProcPointZ.addProcPointProtocols is dead. What
\r
102 goes wrong if we simply never call it?
\r
104 - Something fishy in CmmStackLayout.hs
\r
105 * In particular, 'getAreaSize' returns an AreaMap, but we *know* the width of
\r
106 LocalRegs, so it'd be better to return FiniteMap AreaId ByteWidth
\r
107 * setSuccSPs looks fishy. Rather than lookin in procPoints, it could
\r
108 just lookup the block in areaSize which, after all, has a binding
\r
109 for precisely successors of calls. All other blocks (including proc
\r
110 points that are not successors of a call, we think) can be treated
\r
111 uniformly: zero-size Area, and use inSP.
\r
114 - Currently AsmCodeGen top level calls AsmCodeGen.cmmToCmm, which is a small
\r
115 C-- optimiser. It has quite a lot of boilerplate folding code in AsmCodeGen
\r
116 (cmmBlockConFold, cmmStmtConFold, cmmExprConFold), before calling out to
\r
117 CmmOpt. ToDo: see what optimisations are being done; and do them before
\r
120 - Modularise the CPS pipeline; instead of ...; A;B;C; ...
\r
123 - Most of HscMain.tryNewCodeGen does not belong in HscMain. Instead
\r
126 processCmm [including generating "raw" cmm]
\r
132 - If we stick CAF and stack liveness info on a LastCall node (not LastRet/Jump)
\r
133 then all CAF and stack liveness stuff be completed before we split
\r
134 into separate C procedures.
\r
137 compute and attach liveness into to LastCall
\r
138 right at end, split, cvt to old rep
\r
139 [must split before cvt, because old rep is not expressive enough]
\r
142 when old rep disappears,
\r
143 move the whole splitting game into the C back end *only*
\r
144 (guided by the procpoint set)
\r
147 ----------------------------------------------------
\r
149 ----------------------------------------------------
\r
151 -------- Dead stuff ------------
\r
152 CmmProcPoint Dead: Michael Adams
\r
153 CmmCPS Dead: Michael Adams
\r
154 CmmCPSGen.hs Dead: Michael Adams
\r
155 CmmBrokenBlock.hs Dead: Michael Adams
\r
156 CmmLive.hs Dead: Michael Adams
\r
157 CmmProcPoint.hs Dead: Michael Adams
\r
158 Dataflow.hs Dead: Michael Adams
\r
159 StackColor.hs Norman?
\r
160 StackPlacements.hs Norman?
\r
162 HscMain.optionallyConvertAndOrCPS
\r
164 DynFlags: -fconvert-to-zipper-and-back, -frun-cps, -frun-cpsz
\r
166 -------- Moribund stuff ------------
\r
167 CmmCvt.hs Conversion between old and new Cmm reps
\r
168 CmmOpt.hs Hopefully-redundant optimiser
\r
169 CmmZipUtil.hs Only one function; move elsewhere
\r
171 -------- Stuff to keep ------------
\r
172 CmmCPSZ.hs Driver for new pipeline
\r
174 CmmLiveZ.hs Liveness analysis, dead code elim
\r
175 CmmProcPointZ.hs Identifying and splitting out proc-points
\r
177 CmmSpillReload.hs Save and restore across calls
\r
179 CmmCommonBlockElimZ.hs Common block elim
\r
180 CmmContFlowOpt.hs Other optimisations (branch-chain, merging)
\r
182 CmmBuildInfoTables.hs New info-table
\r
183 CmmStackLayout.hs and stack layout
\r
185 CmmInfo.hs Defn of InfoTables, and conversion to exact layout
\r
187 ---------- Cmm data types --------------
\r
188 ZipCfgCmmRep.hs Cmm instantiations of dataflow graph framework
\r
189 MkZipCfgCmm.hs Cmm instantiations of dataflow graph framework
\r
191 Cmm.hs Key module; a mix of old and new stuff
\r
192 so needs tidying up in due course
\r
197 PprC.hs Pretty print Cmm in C syntax
\r
198 PprCmm.hs Pretty printer for Cmm
\r
199 PprCmmZ.hs Additional stuff for zipper rep
\r
203 ---------- Dataflow modules --------------
\r
204 Goal: separate library; for now, separate directory
\r
210 CmmTx.hs Transactions
\r
211 OptimizationFuel.hs Fuel
\r
212 BlockId.hs BlockId, BlockEnv, BlockSet
\r
216 ----------------------------------------------------
\r
217 Top-level structure
\r
218 ----------------------------------------------------
\r
220 * New codgen called in HscMain.hscGenHardCode, by calling HscMain.tryNewCodeGen,
\r
221 enabled by -fnew-codegen (Opt_TryNewCodeGen)
\r
223 THEN it calls CmmInfo.cmmToRawCmm to lay out the details of info tables
\r
224 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
225 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
227 * HscMain.tryNewCodeGen
\r
228 - STG->Cmm: StgCmm.codeGen (new codegen)
\r
229 - Optimise: CmmContFlowOpt (simple optimisations, very self contained)
\r
230 - Cps convert: CmmCPSZ.protoCmmCPSZ
\r
231 - Optimise: CmmContFlowOpt again
\r
232 - Convert: CmmCvt.cmmOfZgraph (convert to old rep) very self contained
\r
234 * StgCmm.hs The new STG -> Cmm conversion code generator
\r
235 Lots of modules StgCmmXXX
\r
238 ----------------------------------------------------
\r
239 CmmCPSZ.protoCmmCPSZ The new pipeline
\r
240 ----------------------------------------------------
\r
242 CmmCPSZprotoCmmCPSZ:
\r
243 1. Do cpsTop for each procedures separately
\r
244 2. Build SRT representation; this spans multiple procedures
\r
245 (unless split-objs)
\r
248 * CmmCommonBlockElimZ.elimCommonBlocks:
\r
249 eliminate common blocks
\r
251 * CmmProcPointZ.minimalProcPointSet
\r
252 identify proc-points
\r
255 * CmmProcPointZ.addProcPointProtocols
\r
256 something to do with the MA optimisation
\r
257 probably entirely unnecessary
\r
259 * Spill and reload:
\r
260 - CmmSpillReload.dualLivenessWithInsertion
\r
261 insert spills/reloads across
\r
263 Branches to proc-points
\r
264 Now sink those reloads:
\r
265 - CmmSpillReload.insertLateReloads
\r
266 - CmmSpillReload.removeDeadAssignmentsAndReloads
\r
268 * CmmStackLayout.stubSlotsOnDeath
\r
269 debug only: zero out dead slots when they die
\r
272 - CmmStackLayout.lifeSlotAnal:
\r
273 find which sub-areas are live on entry to each block
\r
275 - CmmStackLayout.layout
\r
276 Lay out the stack, returning an AreaMap
\r
277 type AreaMap = FiniteMap Area ByteOff
\r
278 -- Byte offset of the oldest byte of the Area,
\r
279 -- relative to the oldest byte of the Old Area
\r
281 - CmmStackLayout.manifestSP
\r
282 Manifest the stack pointer
\r
284 * Split into separate procedures
\r
285 - CmmProcPointZ.procPointAnalysis
\r
286 Given set of proc points, which blocks are reachable from each
\r
287 Claim: too few proc-points => code duplication, but program still works??
\r
289 - CmmProcPointZ.splitAtProcPoints
\r
290 Using this info, split into separate procedures
\r
292 - CmmBuildInfoTables.setInfoTableStackMap
\r
293 Attach stack maps to each info table
\r
296 ----------------------------------------------------
\r
298 ----------------------------------------------------
\r
300 Consider this program, which has a diamond control flow,
\r
301 with a call on one branch
\r
304 if b then { ... f(x) ...; q=5; goto J }
\r
305 else { ...; q=7; goto J }
\r
308 then the join point J is a "proc-point". So, is 'p' passed to J
\r
309 as a parameter? Or, if 'p' was saved on the stack anyway, perhaps
\r
310 to keep it alive across the call to h(), maybe 'p' gets communicated
\r
311 to J that way. This is an awkward choice. (We think that we currently
\r
312 never pass variables to join points via arguments.)
\r
314 Furthermore, there is *no way* to pass q to J in a register (other
\r
315 than a paramter register).
\r
317 What we want is to do register allocation across the whole caboodle.
\r
318 Then we could drop all the code that deals with the above awkward
\r
319 decisions about spilling variables across proc-points.
\r
321 Note that J doesn't need an info table.
\r
323 What we really want is for each LastCall (not LastJump/Ret)
\r
324 to have an info table. Note that ProcPoints that are not successors
\r
325 of calls don't need an info table.
\r
327 Figuring out proc-points
\r
328 ~~~~~~~~~~~~~~~~~~~~~~~~
\r
329 Proc-points are identified by
\r
330 CmmProcPointZ.minimalProcPointSet/extendPPSet Although there isn't
\r
331 that much code, JD thinks that it could be done much more nicely using
\r
332 a dominator analysis, using the Dataflow Engine.
\r
334 ----------------------------------------------------
\r
336 ----------------------------------------------------
\r
338 * The code for a procedure f may refer to either the *closure*
\r
339 or the *entry point* of another top-level procedure g.
\r
340 If f is live, then so is g. f's SRT must include g's closure.
\r
342 * The CLabel for the entry-point/closure reveals whether g is
\r
343 a CAF (or refers to CAFs). See the IdLabel constructor of CLabel.
\r
345 * The CAF-ness of the original top-level defininions is figured out
\r
346 (by TidyPgm) before we generate C--. This CafInfo is only set for
\r
347 top-level Ids; nested bindings stay with MayHaveCafRefs.
\r
349 * Currently an SRT contains (only) pointers to (top-level) closures.
\r
351 * Consider this Core code
\r
352 f = \x -> let g = \y -> ...x...y...h1...
\r
354 and suppose that h1, h2 have IdInfo of MayHaveCafRefs.
\r
355 Therefore, so will f, But g will not (since it's nested).
\r
357 This generates C-- roughly like this:
\r
358 f_closure: .word f_entry
\r
359 f_entry() [info-tbl-for-f] { ...jump g_entry...jump h2... }
\r
360 g_entry() [info-tbl-for-g] { ...jump h1... }
\r
362 Note that there is no top-level closure for g (only an info table).
\r
363 This fact (whether or not there is a top-level closure) is recorded
\r
364 in the InfoTable attached to the CmmProc for f, g
\r
366 Any out-of-Group references to an IdLabel goes to
\r
367 a Proc whose InfoTable says "I have a top-level closure".
\r
369 A CmmProc whose InfoTable says "I do not have a top-level
\r
370 closure" is referred to only from its own Group.
\r
372 * So: info-tbl-for-f must have an SRT that keeps h1,h2 alive
\r
373 info-tbl-for-g must have an SRT that keeps h1 (only) alive
\r
375 But if we just look for the free CAF refs, we get:
\r
379 So we need to do a transitive closure thing to flesh out
\r
380 f's keep-alive refs to include h1.
\r
382 * The SRT info is the C_SRT field of Cmm.ClosureTypeInfo in a
\r
383 CmmInfoTable attached to each CmmProc. CmmCPSZ.toTops actually does
\r
384 the attaching, right at the end of the pipeline. The C_SRT part
\r
385 gives offsets within a single, shared table of closure pointers.
\r
387 * DECIDED: we can generate SRTs based on the final Cmm program
\r
388 without knowledge of how it is generated.
\r
390 ----------------------------------------------------
\r
392 ----------------------------------------------------
\r
394 See Note [Foreign calls] in ZipCfgCmmRep! This explains that a safe
\r
395 foreign call must do this:
\r
397 push info table (on thread stack) to describe frame
\r
398 make call (via C stack)
\r
400 restore thread state
\r
401 and explains why this expansion must be done late in the day.
\r
404 - Every foreign call is represented as a middle node
\r
406 - *Unsafe* foreign calls are simply "fat machine instructions"
\r
407 and are passed along to the native code generator
\r
409 - *Safe* foreign calls are "lowered" to unsafe calls by wrapping
\r
410 them in the above save/restore sequence. This step is done
\r
411 very late in the pipeline, just before handing to the native
\r
414 This lowering is done by BuildInfoTables.lowerSafeForeignCalls
\r
417 NEW PLAN for foreign calls:
\r
418 - Unsafe foreign calls remain as a middle node (fat machine instruction)
\r
419 Even the parameter passing is not lowered (just as machine instrs
\r
422 - Initially, safe foreign calls appear as LastCalls with
\r
425 ----------------------------------------------------
\r
426 Cmm representations
\r
427 ----------------------------------------------------
\r
430 The type [GenCmm d h g] represents a whole module,
\r
431 ** one list element per .o file **
\r
432 Without SplitObjs, the list has exactly one element
\r
434 newtype GenCmm d h g = Cmm [GenCmmTop d h g] -- A whole .o file
\r
435 data GenCmmTop d h g
\r
436 = CmmProc h g -- One procedure, graph d
\r
437 | CmmData <stuff> [d] -- Initialised data, items d
\r
439 Old and new piplines use different representations
\r
440 (CmmCvt.hs converts between the two)
\r
444 OLD BACK END representations (Cmm.hs):
\r
445 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
447 newtype ListGraph i = ListGraph [GenBasicBlock i]
\r
449 data CmmStmt = Assign | Store | Return etc -- OLD BACK END ONLY
\r
452 Once the info tables are laid out, we replace CmmInfo with [CmmStatic]
\r
453 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
454 which represents the info tables as data, that should
\r
455 immediately precede the code
\r
458 NEW BACK END representations
\r
459 * Not Cmm-specific at all
\r
460 ZipCfg.hs defines Graph, LGraph, FGraph,
\r
461 ZHead, ZTail, ZBlock ...
\r
463 classes LastNode, HavingSuccessors
\r
465 MkZipCfg.hs: AGraph: building graphs
\r
467 * ZipCfgCmmRep: instantiates ZipCfg for Cmm
\r
468 data Middle = ...CmmExpr...
\r
469 data Last = ...CmmExpr...
\r
470 type CmmGraph = Graph Middle Last
\r
472 type CmmZ = GenCmm CmmStatic CmmInfo (CmmStackInfo, CmmGraph)
\r
473 type CmmStackInfo = (ByteOff, Maybe ByteOff)
\r
474 -- (SP offset on entry, update frame space = SP offset on exit)
\r
475 -- The new codegen produces CmmZ, but once the stack is
\r
476 -- manifested we can drop that in favour of
\r
477 -- GenCmm CmmStatic CmmInfo CmmGraph
\r
481 - CmmInfo: partly used by NEW
\r
482 - CmmFormals: not used at all PERHAPS NOT EVEN BY OLD PIPELINE!
\r
484 * MkZipCfgCmm.hs: smart constructors for ZipCfgCmmRep
\r
485 Depends on (a) MkZipCfg (Cmm-independent)
\r
486 (b) ZipCfgCmmRep (Cmm-specific)
\r
490 CmmExpr.hs defines the Cmm expression types
\r
491 - CmmExpr, CmmReg, Width, CmmLit, LocalReg, GlobalReg
\r
492 - CmmType, Width etc (saparate module?)
\r
493 - MachOp (separate module?)
\r
494 - Area, AreaId etc (separate module?)
\r
496 BlockId.hs defines BlockId, BlockEnv, BlockSet
\r
502 * Transactions indicate whether or not the result changes: CmmTx
\r
503 type Tx a = a -> TxRes a
\r
504 data TxRes a = TxRes ChangeFlag a
\r