3 * CmmActuals -> [CmmActual] similary CmmFormals
\r
5 * Possible refactoring: Nuke AGraph in favour of
\r
6 mkIfThenElse :: Expr -> Graph -> Graph -> FCode Graph
\r
8 mkIfThenElse :: HasUniques m => Expr -> Graph -> Graph -> m Graph
\r
9 (Remmber that the .cmm file parser must use this function)
\r
11 or parameterise FCode over its envt; the CgState part seem useful for both
\r
13 * Move top and tail calls to runCmmContFlowOpts from HscMain to CmmCps.cpsTop
\r
14 (and rename the latter!)
\r
16 * "Remove redundant reloads" in CmmSpillReload should be redundant; since
\r
17 insertLateReloads is now gone, every reload is reloading a live variable.
\r
20 * Sink and inline S(RegSlot(x)) = e in precisely the same way that we
\r
21 sink and inline x = e
\r
23 * Stack layout is very like register assignment: find non-conflicting assigments.
\r
24 In particular we can use colouring or linear scan (etc).
\r
26 We'd fine-grain interference (on a word by word basis) to get maximum overlap.
\r
27 But that may make very big interference graphs. So linear scan might be
\r
30 NB: linear scan does on-the-fly live range splitting.
\r
32 * When stubbing dead slots be careful not to write into an area that
\r
33 overlaps with an area that's in use. So stubbing needs to *follow*
\r
39 In CmmNode, consider spliting CmmCall into two: call and jump
\r
41 Notes on new codegen (Aug 10)
\r
42 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\r
45 - We insert spills for variables before the stack check! This is the reason for
\r
46 some fishy code in StgCmmHeap.entryHeapCheck where we are doing some strange
\r
47 things to fix up the stack pointer before GC calls/jumps.
\r
49 The reason spills are inserted before the sp check is that at the entry to a
\r
50 function we always store the parameters passed in registers to local variables.
\r
51 The spill pass simply inserts spills at variable definitions. We instead should
\r
52 sink the spills so that we can avoid spilling them on branches that never
\r
55 This will fix the spill before stack check problem but only really as a side
\r
56 effect. A 'real fix' probably requires making the spiller know about sp checks.
\r
58 EZY: I don't understand this comment. David Terei, can you clarify?
\r
60 - Proc points pass all arguments on the stack, adding more code and
\r
61 slowing down things a lot. We either need to fix this or even better
\r
62 would be to get rid of proc points.
\r
64 - CmmInfo.cmmToRawCmm uses Old.Cmm, so it is called after converting Cmm.Cmm to
\r
65 Old.Cmm. We should abstract it to work on both representations, it needs only to
\r
66 convert a CmmInfoTable to [CmmStatic].
\r
68 - The MkGraph currenty uses a different semantics for <*> than Hoopl. Maybe
\r
69 we could convert codeGen/StgCmm* clients to the Hoopl's semantics?
\r
70 It's all deeply unsatisfactory.
\r
72 - Improve performance of Hoopl.
\r
74 A nofib comparison of -fasm vs -fnewcodegen nofib compilation parameters
\r
75 (using the same ghc-cmm branch +libraries compiled by the old codegenerator)
\r
76 is at http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghchoopl.txt
\r
77 - the code produced is 10.9% slower, the compilation is +118% slower!
\r
79 The same comparison with ghc-head with zip representation is at
\r
80 http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghczip.txt
\r
81 - the code produced is 11.7% slower, the compilation is +78% slower.
\r
83 When compiling nofib, ghc-cmm + libraries compiled with -fnew-codegen
\r
84 is 23.7% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.hooplghcoldgen.txt).
\r
85 When compiling nofib, ghc-head + libraries compiled with -fnew-codegen
\r
86 is 31.4% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.zipghcoldgen.txt).
\r
88 So we generate a bit better code, but it takes us longer!
\r
90 EZY: Also importantly, Hoopl uses dramatically more memory than the
\r
93 - Are all blockToNodeList and blockOfNodeList really needed? Maybe we could
\r
94 splice blocks instead?
\r
96 In the CmmContFlowOpt.blockConcat, using Dataflow seems too clumsy. Still,
\r
97 a block catenation function would be probably nicer than blockToNodeList
\r
98 / blockOfNodeList combo.
\r
100 - lowerSafeForeignCall seems too lowlevel. Just use Dataflow. After that
\r
101 delete splitEntrySeq from HooplUtils.
\r
103 - manifestSP seems to touch a lot of the graph representation. It is
\r
104 also slow for CmmSwitch nodes O(block_nodes * switch_statements).
\r
105 Maybe rewrite manifestSP to use Dataflow?
\r
107 - Sort out Label, LabelMap, LabelSet versus BlockId, BlockEnv, BlockSet
\r
108 dichotomy. Mostly this means global replace, but we also need to make
\r
109 Label an instance of Outputable (probably in the Outputable module).
\r
111 - NB that CmmProcPoint line 283 has a hack that works around a GADT-related
\r
114 - SDM (2010-02-26) can we remove the Foreign constructor from Convention?
\r
115 Reason: we never generate code for a function with the Foreign
\r
116 calling convention, and the code for calling foreign calls is generated
\r
118 - AsmCodeGen has a generic Cmm optimiser; move this into new pipeline
\r
119 EZY (2011-04-16): The mini-inliner has been generalized and ported,
\r
120 but the constant folding and other optimizations need to still be
\r
123 - AsmCodeGen has post-native-cg branch eliminator (shortCutBranches);
\r
124 we ultimately want to share this with the Cmm branch eliminator.
\r
126 - At the moment, references to global registers like Hp are "lowered"
\r
127 late (in CgUtils.fixStgRegisters). We should do this early, in the
\r
128 new native codegen, much in the way that we lower calling conventions.
\r
129 Might need to be a bit sophisticated about aliasing.
\r
131 - Question: currently we lift procpoints to become separate
\r
132 CmmProcs. Do we still want to do this?
\r
134 NB: and advantage of continuing to do this is that
\r
135 we can do common-proc elimination!
\r
137 - Move to new Cmm rep:
\r
138 * Make native CG consume New Cmm;
\r
139 * Convert Old Cmm->New Cmm to keep old path alive
\r
140 * Produce New Cmm when reading in .cmm files
\r
142 - Consider module names
\r
144 - Top-level SRT threading is a bit ugly
\r
146 - Add type/newtype for CmmModule = [CmmGroup] -- A module
\r
147 CmmGroup = [CmmTop] -- A .o file
\r
148 CmmTop = Proc | Data -- A procedure or data
\r
150 - This is a *change*: currently a CmmGroup is one function's-worth of code
\r
151 regardless of SplitObjs. Question: can we *always* generate M.o if there
\r
152 is just one element in the list (rather than M/M1.o, M/M2.o etc)
\r
156 - See "CAFs" below; we want to totally refactor the way SRTs are calculated
\r
158 - Pull out Areas into its own module
\r
159 Parameterise AreaMap (note there are type synonyms in CmmStackLayout!)
\r
160 Add ByteWidth = Int
\r
161 type SubArea = (Area, ByteOff, ByteWidth)
\r
162 ByteOff should not be defined in SMRep -- that is too high up the hierarchy
\r
164 - SMRep should not be imported by any module in cmm/! Make it so.
\r
165 -- ByteOff etc ==> CmmExpr
\r
166 -- rET_SMALL etc ==> CmmInfo
\r
167 Check that there are no other imports from codeGen in cmm/
\r
169 - If you eliminate a label by branch chain elimination,
\r
170 what happens if there's an Area associated with that label?
\r
172 - Think about a non-flattened representation?
\r
175 * Use record fields for LastCall!
\r
176 * cml_ret_off should be a ByteOff
\r
178 LastCall (which has a successor) and
\r
179 LastJump (which does not, includes return?)
\r
180 - does not have cml_cont, cml_ret_args, cml_ret_off
\r
183 - expands into save/MidForeignCall/restore/goto
\r
184 - like any LastCall, target of the call gets an info table
\r
186 - JD: remind self of what goes wrong if you turn off the
\r
187 liveness of the update frame
\r
189 - Garbage-collect http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/CPS
\r
190 moving good stuff into
\r
191 http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/NewCodeGenPipeline
\r
194 - We believe that all of CmmProcPoint.addProcPointProtocols is dead. What
\r
195 goes wrong if we simply never call it?
\r
197 - Something fishy in CmmStackLayout.hs
\r
198 * In particular, 'getAreaSize' returns an AreaMap, but we *know* the width of
\r
199 LocalRegs, so it'd be better to return FiniteMap AreaId ByteWidth
\r
200 * setSuccSPs looks fishy. Rather than lookin in procPoints, it could
\r
201 just lookup the block in areaSize which, after all, has a binding
\r
202 for precisely successors of calls. All other blocks (including proc
\r
203 points that are not successors of a call, we think) can be treated
\r
204 uniformly: zero-size Area, and use inSP.
\r
207 - Currently AsmCodeGen top level calls AsmCodeGen.cmmToCmm, which is a small
\r
208 C-- optimiser. It has quite a lot of boilerplate folding code in AsmCodeGen
\r
209 (cmmBlockConFold, cmmStmtConFold, cmmExprConFold), before calling out to
\r
210 CmmOpt. ToDo: see what optimisations are being done; and do them before
\r
213 - Modularise the CPS pipeline; instead of ...; A;B;C; ...
\r
216 - Most of HscMain.tryNewCodeGen does not belong in HscMain. Instead
\r
219 processCmm [including generating "raw" cmm]
\r
225 - If we stick CAF and stack liveness info on a LastCall node (not LastRet/Jump)
\r
226 then all CAF and stack liveness stuff be completed before we split
\r
227 into separate C procedures.
\r
230 compute and attach liveness into to LastCall
\r
231 right at end, split, cvt to old rep
\r
232 [must split before cvt, because old rep is not expressive enough]
\r
235 when old rep disappears,
\r
236 move the whole splitting game into the C back end *only*
\r
237 (guided by the procpoint set)
\r
239 ----------------------------------------------------
\r
241 ----------------------------------------------------
\r
243 -------- Testing stuff ------------
\r
244 HscMain.optionallyConvertAndOrCPS
\r
246 DynFlags: -fconvert-to-zipper-and-back, -frun-cpsz
\r
248 -------- Moribund stuff ------------
\r
249 OldCmm.hs Definition of flowgraph of old representation
\r
250 OldCmmUtil.hs Utilites that operates mostly on on CmmStmt
\r
251 OldPprCmm.hs Pretty print for CmmStmt, GenBasicBlock and ListGraph
\r
252 CmmCvt.hs Conversion between old and new Cmm reps
\r
253 CmmOpt.hs Hopefully-redundant optimiser
\r
255 -------- Stuff to keep ------------
\r
256 CmmCPS.hs Driver for new pipeline
\r
258 CmmLive.hs Liveness analysis, dead code elim
\r
259 CmmProcPoint.hs Identifying and splitting out proc-points
\r
261 CmmSpillReload.hs Save and restore across calls
\r
263 CmmCommonBlockElim.hs Common block elim
\r
264 CmmContFlowOpt.hs Other optimisations (branch-chain, merging)
\r
266 CmmBuildInfoTables.hs New info-table
\r
267 CmmStackLayout.hs and stack layout
\r
269 CmmInfo.hs Defn of InfoTables, and conversion to exact byte layout
\r
271 ---------- Cmm data types --------------
\r
272 Cmm.hs Cmm instantiations of dataflow graph framework
\r
273 MkGraph.hs Interface for building Cmm for codeGen/Stg*.hs modules
\r
275 CmmDecl.hs Shared Cmm types of both representations
\r
276 CmmExpr.hs Type of Cmm expression
\r
277 CmmType.hs Type of Cmm types and their widths
\r
278 CmmMachOp.hs MachOp type and accompanying utilities
\r
283 PprC.hs Pretty print Cmm in C syntax
\r
284 PprCmm.hs Pretty printer for CmmGraph.
\r
285 PprCmmDecl.hs Pretty printer for common Cmm types.
\r
286 PprCmmExpr.hs Pretty printer for Cmm expressions.
\r
289 BlockId.hs BlockId, BlockEnv, BlockSet
\r
291 ----------------------------------------------------
\r
292 Top-level structure
\r
293 ----------------------------------------------------
\r
295 * New codgen called in HscMain.hscGenHardCode, by calling HscMain.tryNewCodeGen,
\r
296 enabled by -fnew-codegen (Opt_TryNewCodeGen)
\r
298 THEN it calls CmmInfo.cmmToRawCmm to lay out the details of info tables
\r
299 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
300 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
302 * HscMain.tryNewCodeGen
\r
303 - STG->Cmm: StgCmm.codeGen (new codegen)
\r
304 - Optimise: CmmContFlowOpt (simple optimisations, very self contained)
\r
305 - Cps convert: CmmCPS.protoCmmCPS
\r
306 - Optimise: CmmContFlowOpt again
\r
307 - Convert: CmmCvt.cmmOfZgraph (convert to old rep) very self contained
\r
309 * StgCmm.hs The new STG -> Cmm conversion code generator
\r
310 Lots of modules StgCmmXXX
\r
313 ----------------------------------------------------
\r
314 CmmCPS.protoCmmCPS The new pipeline
\r
315 ----------------------------------------------------
\r
317 CmmCPS.protoCmmCPS:
\r
318 1. Do cpsTop for each procedures separately
\r
319 2. Build SRT representation; this spans multiple procedures
\r
320 (unless split-objs)
\r
323 * CmmCommonBlockElim.elimCommonBlocks:
\r
324 eliminate common blocks
\r
326 * CmmProcPoint.minimalProcPointSet
\r
327 identify proc-points
\r
330 * CmmProcPoint.addProcPointProtocols
\r
331 something to do with the MA optimisation
\r
332 probably entirely unnecessary
\r
334 * Spill and reload:
\r
335 - CmmSpillReload.dualLivenessWithInsertion
\r
336 insert spills/reloads across
\r
338 Branches to proc-points
\r
339 Now sink those reloads (and other instructions):
\r
340 - CmmSpillReload.rewriteAssignments
\r
341 - CmmSpillReload.removeDeadAssignmentsAndReloads
\r
343 * CmmStackLayout.stubSlotsOnDeath
\r
344 debug only: zero out dead slots when they die
\r
347 - CmmStackLayout.lifeSlotAnal:
\r
348 find which sub-areas are live on entry to each block
\r
350 - CmmStackLayout.layout
\r
351 Lay out the stack, returning an AreaMap
\r
352 type AreaMap = FiniteMap Area ByteOff
\r
353 -- Byte offset of the oldest byte of the Area,
\r
354 -- relative to the oldest byte of the Old Area
\r
356 - CmmStackLayout.manifestSP
\r
357 Manifest the stack pointer
\r
359 * Split into separate procedures
\r
360 - CmmProcPoint.procPointAnalysis
\r
361 Given set of proc points, which blocks are reachable from each
\r
362 Claim: too few proc-points => code duplication, but program still works??
\r
364 - CmmProcPoint.splitAtProcPoints
\r
365 Using this info, split into separate procedures
\r
367 - CmmBuildInfoTables.setInfoTableStackMap
\r
368 Attach stack maps to each info table
\r
371 ----------------------------------------------------
\r
373 ----------------------------------------------------
\r
375 Consider this program, which has a diamond control flow,
\r
376 with a call on one branch
\r
379 if b then { ... f(x) ...; q=5; goto J }
\r
380 else { ...; q=7; goto J }
\r
383 then the join point J is a "proc-point". So, is 'p' passed to J
\r
384 as a parameter? Or, if 'p' was saved on the stack anyway, perhaps
\r
385 to keep it alive across the call to h(), maybe 'p' gets communicated
\r
386 to J that way. This is an awkward choice. (We think that we currently
\r
387 never pass variables to join points via arguments.)
\r
389 Furthermore, there is *no way* to pass q to J in a register (other
\r
390 than a parameter register).
\r
392 What we want is to do register allocation across the whole caboodle.
\r
393 Then we could drop all the code that deals with the above awkward
\r
394 decisions about spilling variables across proc-points.
\r
396 Note that J doesn't need an info table.
\r
398 What we really want is for each LastCall (not LastJump/Ret)
\r
399 to have an info table. Note that ProcPoints that are not successors
\r
400 of calls don't need an info table.
\r
402 Figuring out proc-points
\r
403 ~~~~~~~~~~~~~~~~~~~~~~~~
\r
404 Proc-points are identified by
\r
405 CmmProcPoint.minimalProcPointSet/extendPPSet Although there isn't
\r
406 that much code, JD thinks that it could be done much more nicely using
\r
407 a dominator analysis, using the Dataflow Engine.
\r
409 ----------------------------------------------------
\r
411 ----------------------------------------------------
\r
413 * The code for a procedure f may refer to either the *closure*
\r
414 or the *entry point* of another top-level procedure g.
\r
415 If f is live, then so is g. f's SRT must include g's closure.
\r
417 * The CLabel for the entry-point/closure reveals whether g is
\r
418 a CAF (or refers to CAFs). See the IdLabel constructor of CLabel.
\r
420 * The CAF-ness of the original top-level defininions is figured out
\r
421 (by TidyPgm) before we generate C--. This CafInfo is only set for
\r
422 top-level Ids; nested bindings stay with MayHaveCafRefs.
\r
424 * Currently an SRT contains (only) pointers to (top-level) closures.
\r
426 * Consider this Core code
\r
427 f = \x -> let g = \y -> ...x...y...h1...
\r
429 and suppose that h1, h2 have IdInfo of MayHaveCafRefs.
\r
430 Therefore, so will f, But g will not (since it's nested).
\r
432 This generates C-- roughly like this:
\r
433 f_closure: .word f_entry
\r
434 f_entry() [info-tbl-for-f] { ...jump g_entry...jump h2... }
\r
435 g_entry() [info-tbl-for-g] { ...jump h1... }
\r
437 Note that there is no top-level closure for g (only an info table).
\r
438 This fact (whether or not there is a top-level closure) is recorded
\r
439 in the InfoTable attached to the CmmProc for f, g
\r
441 Any out-of-Group references to an IdLabel goes to
\r
442 a Proc whose InfoTable says "I have a top-level closure".
\r
444 A CmmProc whose InfoTable says "I do not have a top-level
\r
445 closure" is referred to only from its own Group.
\r
447 * So: info-tbl-for-f must have an SRT that keeps h1,h2 alive
\r
448 info-tbl-for-g must have an SRT that keeps h1 (only) alive
\r
450 But if we just look for the free CAF refs, we get:
\r
454 So we need to do a transitive closure thing to flesh out
\r
455 f's keep-alive refs to include h1.
\r
457 * The SRT info is the C_SRT field of Cmm.ClosureTypeInfo in a
\r
458 CmmInfoTable attached to each CmmProc. CmmCPS.toTops actually does
\r
459 the attaching, right at the end of the pipeline. The C_SRT part
\r
460 gives offsets within a single, shared table of closure pointers.
\r
462 * DECIDED: we can generate SRTs based on the final Cmm program
\r
463 without knowledge of how it is generated.
\r
465 ----------------------------------------------------
\r
467 ----------------------------------------------------
\r
469 See Note [Foreign calls] in CmmNode! This explains that a safe
\r
470 foreign call must do this:
\r
472 push info table (on thread stack) to describe frame
\r
473 make call (via C stack)
\r
475 restore thread state
\r
476 and explains why this expansion must be done late in the day.
\r
479 - Every foreign call is represented as a middle node
\r
481 - *Unsafe* foreign calls are simply "fat machine instructions"
\r
482 and are passed along to the native code generator
\r
484 - *Safe* foreign calls are "lowered" to unsafe calls by wrapping
\r
485 them in the above save/restore sequence. This step is done
\r
486 very late in the pipeline, just before handing to the native
\r
489 This lowering is done by BuildInfoTables.lowerSafeForeignCalls
\r
492 NEW PLAN for foreign calls:
\r
493 - Unsafe foreign calls remain as a middle node (fat machine instruction)
\r
494 Even the parameter passing is not lowered (just as machine instrs
\r
497 - Initially, safe foreign calls appear as LastCalls with
\r
500 ----------------------------------------------------
\r
501 Cmm representations
\r
502 ----------------------------------------------------
\r
505 The type [GenCmm d h g] represents a whole module,
\r
506 ** one list element per .o file **
\r
507 Without SplitObjs, the list has exactly one element
\r
509 newtype GenCmm d h g = Cmm [GenCmmTop d h g] -- A whole .o file
\r
510 data GenCmmTop d h g
\r
511 = CmmProc h g -- One procedure, graph d
\r
512 | CmmData <stuff> [d] -- Initialised data, items d
\r
514 Old and new piplines use different representations
\r
515 (CmmCvt.hs converts between the two)
\r
519 OLD BACK END representations (OldCmm.hs):
\r
520 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
522 newtype ListGraph i = ListGraph [GenBasicBlock i]
\r
524 data CmmStmt = Assign | Store | Return etc -- OLD BACK END ONLY
\r
527 Once the info tables are laid out, we replace CmmInfo with [CmmStatic]
\r
528 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
529 which represents the info tables as data, that should
\r
530 immediately precede the code
\r
533 NEW BACK END representations
\r
534 * Uses Hoopl library, a zero-boot package
\r
535 * CmmNode defines a node of a flow graph.
\r
536 * Cmm defines CmmGraph, CmmTop, Cmm
\r
537 - CmmGraph is a closed/closed graph + an entry node.
\r
539 data CmmGraph = CmmGraph { g_entry :: BlockId
\r
540 , g_graph :: Graph CmmNode C C }
\r
542 - CmmTop is a top level chunk, specialization of GenCmmTop from CmmDecl.hs
\r
543 with CmmGraph as a flow graph.
\r
544 - Cmm is a collection of CmmTops.
\r
546 type Cmm = GenCmm CmmStatic CmmTopInfo CmmGraph
\r
547 type CmmTop = GenCmmTop CmmStatic CmmTopInfo CmmGraph
\r
549 - CmmTop uses CmmTopInfo, which is a CmmInfoTable and CmmStackInfo
\r
551 data CmmTopInfo = TopInfo {info_tbl :: CmmInfoTable, stack_info :: CmmStackInfo}
\r
555 data CmmStackInfo = StackInfo {arg_space :: ByteOff, updfr_space :: Maybe ByteOff}
\r
557 * arg_space = SP offset on entry
\r
558 * updfr_space space = SP offset on exit
\r
559 Once the staci is manifested, we could drom CmmStackInfo, ie. get
\r
560 GenCmm CmmStatic CmmInfoTable CmmGraph, but we do not do that currently.
\r
563 * MkGraph.hs: smart constructors for Cmm.hs
\r
564 Beware, the CmmAGraph defined here does not use AGraph from Hoopl,
\r
565 as CmmAGraph can be opened or closed at exit, See the notes in that module.
\r
569 CmmDecl.hs - GenCmm and GenCmmTop types
\r
570 CmmExpr.hs - defines the Cmm expression types
\r
571 - CmmExpr, CmmReg, CmmLit, LocalReg, GlobalReg
\r
572 - Area, AreaId etc (separate module?)
\r
573 CmmType.hs - CmmType, Width etc (saparate module?)
\r
574 CmmMachOp.hs - MachOp and CallishMachOp types
\r
576 BlockId.hs defines BlockId, BlockEnv, BlockSet
\r