3 * Possible refactoring: Nuke AGraph in favour of
\r
4 mkIfThenElse :: Expr -> Graph -> Graph -> FCode Graph
\r
6 mkIfThenElse :: HasUniques m => Expr -> Graph -> Graph -> m Graph
\r
7 (Remmber that the .cmm file parser must use this function)
\r
9 or parameterise FCode over its envt; the CgState part seem useful for both
\r
11 * "Remove redundant reloads" in CmmSpillReload should be redundant; since
\r
12 insertLateReloads is now gone, every reload is reloading a live variable.
\r
15 * Sink and inline S(RegSlot(x)) = e in precisely the same way that we
\r
16 sink and inline x = e
\r
18 * Stack layout is very like register assignment: find non-conflicting assigments.
\r
19 In particular we can use colouring or linear scan (etc).
\r
21 We'd fine-grain interference (on a word by word basis) to get maximum overlap.
\r
22 But that may make very big interference graphs. So linear scan might be
\r
25 NB: linear scan does on-the-fly live range splitting.
\r
27 * When stubbing dead slots be careful not to write into an area that
\r
28 overlaps with an area that's in use. So stubbing needs to *follow*
\r
34 In CmmNode, consider spliting CmmCall into two: call and jump
\r
36 Notes on new codegen (Aug 10)
\r
37 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\r
40 - We insert spills for variables before the stack check! This is the reason for
\r
41 some fishy code in StgCmmHeap.entryHeapCheck where we are doing some strange
\r
42 things to fix up the stack pointer before GC calls/jumps.
\r
44 The reason spills are inserted before the sp check is that at the entry to a
\r
45 function we always store the parameters passed in registers to local variables.
\r
46 The spill pass simply inserts spills at variable definitions. We instead should
\r
47 sink the spills so that we can avoid spilling them on branches that never
\r
50 This will fix the spill before stack check problem but only really as a side
\r
51 effect. A 'real fix' probably requires making the spiller know about sp checks.
\r
53 EZY: I don't understand this comment. David Terei, can you clarify?
\r
55 - Proc points pass all arguments on the stack, adding more code and
\r
56 slowing down things a lot. We either need to fix this or even better
\r
57 would be to get rid of proc points.
\r
59 - CmmInfo.cmmToRawCmm uses Old.Cmm, so it is called after converting Cmm.Cmm to
\r
60 Old.Cmm. We should abstract it to work on both representations, it needs only to
\r
61 convert a CmmInfoTable to [CmmStatic].
\r
63 - The MkGraph currenty uses a different semantics for <*> than Hoopl. Maybe
\r
64 we could convert codeGen/StgCmm* clients to the Hoopl's semantics?
\r
65 It's all deeply unsatisfactory.
\r
67 - Improve performance of Hoopl.
\r
69 A nofib comparison of -fasm vs -fnewcodegen nofib compilation parameters
\r
70 (using the same ghc-cmm branch +libraries compiled by the old codegenerator)
\r
71 is at http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghchoopl.txt
\r
72 - the code produced is 10.9% slower, the compilation is +118% slower!
\r
74 The same comparison with ghc-head with zip representation is at
\r
75 http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghczip.txt
\r
76 - the code produced is 11.7% slower, the compilation is +78% slower.
\r
78 When compiling nofib, ghc-cmm + libraries compiled with -fnew-codegen
\r
79 is 23.7% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.hooplghcoldgen.txt).
\r
80 When compiling nofib, ghc-head + libraries compiled with -fnew-codegen
\r
81 is 31.4% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.zipghcoldgen.txt).
\r
83 So we generate a bit better code, but it takes us longer!
\r
85 EZY: Also importantly, Hoopl uses dramatically more memory than the
\r
88 - Are all blockToNodeList and blockOfNodeList really needed? Maybe we could
\r
89 splice blocks instead?
\r
91 In the CmmContFlowOpt.blockConcat, using Dataflow seems too clumsy. Still,
\r
92 a block catenation function would be probably nicer than blockToNodeList
\r
93 / blockOfNodeList combo.
\r
95 - lowerSafeForeignCall seems too lowlevel. Just use Dataflow. After that
\r
96 delete splitEntrySeq from HooplUtils.
\r
98 - manifestSP seems to touch a lot of the graph representation. It is
\r
99 also slow for CmmSwitch nodes O(block_nodes * switch_statements).
\r
100 Maybe rewrite manifestSP to use Dataflow?
\r
102 - Sort out Label, LabelMap, LabelSet versus BlockId, BlockEnv, BlockSet
\r
103 dichotomy. Mostly this means global replace, but we also need to make
\r
104 Label an instance of Outputable (probably in the Outputable module).
\r
106 - NB that CmmProcPoint line 283 has a hack that works around a GADT-related
\r
109 - SDM (2010-02-26) can we remove the Foreign constructor from Convention?
\r
110 Reason: we never generate code for a function with the Foreign
\r
111 calling convention, and the code for calling foreign calls is generated
\r
113 - AsmCodeGen has a generic Cmm optimiser; move this into new pipeline
\r
114 EZY (2011-04-16): The mini-inliner has been generalized and ported,
\r
115 but the constant folding and other optimizations need to still be
\r
118 - AsmCodeGen has post-native-cg branch eliminator (shortCutBranches);
\r
119 we ultimately want to share this with the Cmm branch eliminator.
\r
121 - At the moment, references to global registers like Hp are "lowered"
\r
122 late (in CgUtils.fixStgRegisters). We should do this early, in the
\r
123 new native codegen, much in the way that we lower calling conventions.
\r
124 Might need to be a bit sophisticated about aliasing.
\r
126 - Question: currently we lift procpoints to become separate
\r
127 CmmProcs. Do we still want to do this?
\r
129 NB: and advantage of continuing to do this is that
\r
130 we can do common-proc elimination!
\r
132 - Move to new Cmm rep:
\r
133 * Make native CG consume New Cmm;
\r
134 * Convert Old Cmm->New Cmm to keep old path alive
\r
135 * Produce New Cmm when reading in .cmm files
\r
137 - Consider module names
\r
139 - Top-level SRT threading is a bit ugly
\r
141 - Add type/newtype for CmmModule = [CmmGroup] -- A module
\r
142 CmmGroup = [CmmTop] -- A .o file
\r
143 CmmTop = Proc | Data -- A procedure or data
\r
145 - This is a *change*: currently a CmmGroup is one function's-worth of code
\r
146 regardless of SplitObjs. Question: can we *always* generate M.o if there
\r
147 is just one element in the list (rather than M/M1.o, M/M2.o etc)
\r
151 - See "CAFs" below; we want to totally refactor the way SRTs are calculated
\r
153 - Pull out Areas into its own module
\r
154 Parameterise AreaMap (note there are type synonyms in CmmStackLayout!)
\r
155 Add ByteWidth = Int
\r
156 type SubArea = (Area, ByteOff, ByteWidth)
\r
157 ByteOff should not be defined in SMRep -- that is too high up the hierarchy
\r
159 - SMRep should not be imported by any module in cmm/! Make it so.
\r
160 -- ByteOff etc ==> CmmExpr
\r
161 -- rET_SMALL etc ==> CmmInfo
\r
162 Check that there are no other imports from codeGen in cmm/
\r
164 - If you eliminate a label by branch chain elimination,
\r
165 what happens if there's an Area associated with that label?
\r
167 - Think about a non-flattened representation?
\r
170 * Use record fields for LastCall!
\r
171 * cml_ret_off should be a ByteOff
\r
173 LastCall (which has a successor) and
\r
174 LastJump (which does not, includes return?)
\r
175 - does not have cml_cont, cml_ret_args, cml_ret_off
\r
178 - expands into save/MidForeignCall/restore/goto
\r
179 - like any LastCall, target of the call gets an info table
\r
181 - JD: remind self of what goes wrong if you turn off the
\r
182 liveness of the update frame
\r
184 - Garbage-collect http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/CPS
\r
185 moving good stuff into
\r
186 http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/NewCodeGenPipeline
\r
189 - We believe that all of CmmProcPoint.addProcPointProtocols is dead. What
\r
190 goes wrong if we simply never call it?
\r
192 - Something fishy in CmmStackLayout.hs
\r
193 * In particular, 'getAreaSize' returns an AreaMap, but we *know* the width of
\r
194 LocalRegs, so it'd be better to return FiniteMap AreaId ByteWidth
\r
195 * setSuccSPs looks fishy. Rather than lookin in procPoints, it could
\r
196 just lookup the block in areaSize which, after all, has a binding
\r
197 for precisely successors of calls. All other blocks (including proc
\r
198 points that are not successors of a call, we think) can be treated
\r
199 uniformly: zero-size Area, and use inSP.
\r
202 - Currently AsmCodeGen top level calls AsmCodeGen.cmmToCmm, which is a small
\r
203 C-- optimiser. It has quite a lot of boilerplate folding code in AsmCodeGen
\r
204 (cmmBlockConFold, cmmStmtConFold, cmmExprConFold), before calling out to
\r
205 CmmOpt. ToDo: see what optimisations are being done; and do them before
\r
208 - Modularise the CPS pipeline; instead of ...; A;B;C; ...
\r
211 - Most of HscMain.tryNewCodeGen does not belong in HscMain. Instead
\r
214 processCmm [including generating "raw" cmm]
\r
220 - If we stick CAF and stack liveness info on a LastCall node (not LastRet/Jump)
\r
221 then all CAF and stack liveness stuff be completed before we split
\r
222 into separate C procedures.
\r
225 compute and attach liveness into to LastCall
\r
226 right at end, split, cvt to old rep
\r
227 [must split before cvt, because old rep is not expressive enough]
\r
230 when old rep disappears,
\r
231 move the whole splitting game into the C back end *only*
\r
232 (guided by the procpoint set)
\r
234 ----------------------------------------------------
\r
236 ----------------------------------------------------
\r
238 -------- Testing stuff ------------
\r
239 HscMain.optionallyConvertAndOrCPS
\r
241 DynFlags: -fconvert-to-zipper-and-back, -frun-cpsz
\r
243 -------- Moribund stuff ------------
\r
244 OldCmm.hs Definition of flowgraph of old representation
\r
245 OldCmmUtil.hs Utilites that operates mostly on on CmmStmt
\r
246 OldPprCmm.hs Pretty print for CmmStmt, GenBasicBlock and ListGraph
\r
247 CmmCvt.hs Conversion between old and new Cmm reps
\r
248 CmmOpt.hs Hopefully-redundant optimiser
\r
250 -------- Stuff to keep ------------
\r
251 CmmPipeline.hs Driver for new pipeline
\r
253 CmmLive.hs Liveness analysis, dead code elim
\r
254 CmmProcPoint.hs Identifying and splitting out proc-points
\r
256 CmmSpillReload.hs Save and restore across calls
\r
258 CmmCommonBlockElim.hs Common block elim
\r
259 CmmContFlowOpt.hs Other optimisations (branch-chain, merging)
\r
261 CmmBuildInfoTables.hs New info-table
\r
262 CmmStackLayout.hs and stack layout
\r
264 CmmInfo.hs Defn of InfoTables, and conversion to exact byte layout
\r
266 ---------- Cmm data types --------------
\r
267 Cmm.hs Cmm instantiations of dataflow graph framework
\r
268 MkGraph.hs Interface for building Cmm for codeGen/Stg*.hs modules
\r
270 CmmDecl.hs Shared Cmm types of both representations
\r
271 CmmExpr.hs Type of Cmm expression
\r
272 CmmType.hs Type of Cmm types and their widths
\r
273 CmmMachOp.hs MachOp type and accompanying utilities
\r
278 PprC.hs Pretty print Cmm in C syntax
\r
279 PprCmm.hs Pretty printer for CmmGraph.
\r
280 PprCmmDecl.hs Pretty printer for common Cmm types.
\r
281 PprCmmExpr.hs Pretty printer for Cmm expressions.
\r
284 BlockId.hs BlockId, BlockEnv, BlockSet
\r
286 ----------------------------------------------------
\r
287 Top-level structure
\r
288 ----------------------------------------------------
\r
290 * New codgen called in HscMain.hscGenHardCode, by calling HscMain.tryNewCodeGen,
\r
291 enabled by -fnew-codegen (Opt_TryNewCodeGen)
\r
293 THEN it calls CmmInfo.cmmToRawCmm to lay out the details of info tables
\r
294 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
295 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
297 * HscMain.tryNewCodeGen
\r
298 - STG->Cmm: StgCmm.codeGen (new codegen)
\r
299 - Optimize and CPS: CmmPipeline.cmmPipeline
\r
300 - Convert: CmmCvt.cmmOfZgraph (convert to old rep) very self contained
\r
302 * StgCmm.hs The new STG -> Cmm conversion code generator
\r
303 Lots of modules StgCmmXXX
\r
306 ----------------------------------------------------
\r
307 CmmPipeline.cmmPipeline The new pipeline
\r
308 ----------------------------------------------------
\r
310 CmmPipeline.cmmPipeline:
\r
311 1. Do control flow optimization
\r
312 2. Do cpsTop for each procedures separately
\r
313 3. Build SRT representation; this spans multiple procedures
\r
314 (unless split-objs)
\r
315 4. Do control flow optimization on all resulting procedures
\r
318 * CmmCommonBlockElim.elimCommonBlocks:
\r
319 eliminate common blocks
\r
321 * CmmProcPoint.minimalProcPointSet
\r
322 identify proc-points
\r
325 * CmmProcPoint.addProcPointProtocols
\r
326 something to do with the MA optimisation
\r
327 probably entirely unnecessary
\r
329 * Spill and reload:
\r
330 - CmmSpillReload.dualLivenessWithInsertion
\r
331 insert spills/reloads across
\r
333 Branches to proc-points
\r
334 Now sink those reloads (and other instructions):
\r
335 - CmmSpillReload.rewriteAssignments
\r
336 - CmmSpillReload.removeDeadAssignmentsAndReloads
\r
338 * CmmStackLayout.stubSlotsOnDeath
\r
339 debug only: zero out dead slots when they die
\r
342 - CmmStackLayout.lifeSlotAnal:
\r
343 find which sub-areas are live on entry to each block
\r
345 - CmmStackLayout.layout
\r
346 Lay out the stack, returning an AreaMap
\r
347 type AreaMap = FiniteMap Area ByteOff
\r
348 -- Byte offset of the oldest byte of the Area,
\r
349 -- relative to the oldest byte of the Old Area
\r
351 - CmmStackLayout.manifestSP
\r
352 Manifest the stack pointer
\r
354 * Split into separate procedures
\r
355 - CmmProcPoint.procPointAnalysis
\r
356 Given set of proc points, which blocks are reachable from each
\r
357 Claim: too few proc-points => code duplication, but program still works??
\r
359 - CmmProcPoint.splitAtProcPoints
\r
360 Using this info, split into separate procedures
\r
362 - CmmBuildInfoTables.setInfoTableStackMap
\r
363 Attach stack maps to each info table
\r
366 ----------------------------------------------------
\r
368 ----------------------------------------------------
\r
370 Consider this program, which has a diamond control flow,
\r
371 with a call on one branch
\r
374 if b then { ... f(x) ...; q=5; goto J }
\r
375 else { ...; q=7; goto J }
\r
378 then the join point J is a "proc-point". So, is 'p' passed to J
\r
379 as a parameter? Or, if 'p' was saved on the stack anyway, perhaps
\r
380 to keep it alive across the call to h(), maybe 'p' gets communicated
\r
381 to J that way. This is an awkward choice. (We think that we currently
\r
382 never pass variables to join points via arguments.)
\r
384 Furthermore, there is *no way* to pass q to J in a register (other
\r
385 than a parameter register).
\r
387 What we want is to do register allocation across the whole caboodle.
\r
388 Then we could drop all the code that deals with the above awkward
\r
389 decisions about spilling variables across proc-points.
\r
391 Note that J doesn't need an info table.
\r
393 What we really want is for each LastCall (not LastJump/Ret)
\r
394 to have an info table. Note that ProcPoints that are not successors
\r
395 of calls don't need an info table.
\r
397 Figuring out proc-points
\r
398 ~~~~~~~~~~~~~~~~~~~~~~~~
\r
399 Proc-points are identified by
\r
400 CmmProcPoint.minimalProcPointSet/extendPPSet Although there isn't
\r
401 that much code, JD thinks that it could be done much more nicely using
\r
402 a dominator analysis, using the Dataflow Engine.
\r
404 ----------------------------------------------------
\r
406 ----------------------------------------------------
\r
408 * The code for a procedure f may refer to either the *closure*
\r
409 or the *entry point* of another top-level procedure g.
\r
410 If f is live, then so is g. f's SRT must include g's closure.
\r
412 * The CLabel for the entry-point/closure reveals whether g is
\r
413 a CAF (or refers to CAFs). See the IdLabel constructor of CLabel.
\r
415 * The CAF-ness of the original top-level defininions is figured out
\r
416 (by TidyPgm) before we generate C--. This CafInfo is only set for
\r
417 top-level Ids; nested bindings stay with MayHaveCafRefs.
\r
419 * Currently an SRT contains (only) pointers to (top-level) closures.
\r
421 * Consider this Core code
\r
422 f = \x -> let g = \y -> ...x...y...h1...
\r
424 and suppose that h1, h2 have IdInfo of MayHaveCafRefs.
\r
425 Therefore, so will f, But g will not (since it's nested).
\r
427 This generates C-- roughly like this:
\r
428 f_closure: .word f_entry
\r
429 f_entry() [info-tbl-for-f] { ...jump g_entry...jump h2... }
\r
430 g_entry() [info-tbl-for-g] { ...jump h1... }
\r
432 Note that there is no top-level closure for g (only an info table).
\r
433 This fact (whether or not there is a top-level closure) is recorded
\r
434 in the InfoTable attached to the CmmProc for f, g
\r
436 Any out-of-Group references to an IdLabel goes to
\r
437 a Proc whose InfoTable says "I have a top-level closure".
\r
439 A CmmProc whose InfoTable says "I do not have a top-level
\r
440 closure" is referred to only from its own Group.
\r
442 * So: info-tbl-for-f must have an SRT that keeps h1,h2 alive
\r
443 info-tbl-for-g must have an SRT that keeps h1 (only) alive
\r
445 But if we just look for the free CAF refs, we get:
\r
449 So we need to do a transitive closure thing to flesh out
\r
450 f's keep-alive refs to include h1.
\r
452 * The SRT info is the C_SRT field of Cmm.ClosureTypeInfo in a
\r
453 CmmInfoTable attached to each CmmProc. CmmPipeline.toTops actually does
\r
454 the attaching, right at the end of the pipeline. The C_SRT part
\r
455 gives offsets within a single, shared table of closure pointers.
\r
457 * DECIDED: we can generate SRTs based on the final Cmm program
\r
458 without knowledge of how it is generated.
\r
460 ----------------------------------------------------
\r
462 ----------------------------------------------------
\r
464 See Note [Foreign calls] in CmmNode! This explains that a safe
\r
465 foreign call must do this:
\r
467 push info table (on thread stack) to describe frame
\r
468 make call (via C stack)
\r
470 restore thread state
\r
471 and explains why this expansion must be done late in the day.
\r
474 - Every foreign call is represented as a middle node
\r
476 - *Unsafe* foreign calls are simply "fat machine instructions"
\r
477 and are passed along to the native code generator
\r
479 - *Safe* foreign calls are "lowered" to unsafe calls by wrapping
\r
480 them in the above save/restore sequence. This step is done
\r
481 very late in the pipeline, just before handing to the native
\r
484 This lowering is done by BuildInfoTables.lowerSafeForeignCalls
\r
487 NEW PLAN for foreign calls:
\r
488 - Unsafe foreign calls remain as a middle node (fat machine instruction)
\r
489 Even the parameter passing is not lowered (just as machine instrs
\r
492 - Initially, safe foreign calls appear as LastCalls with
\r
495 ----------------------------------------------------
\r
496 Cmm representations
\r
497 ----------------------------------------------------
\r
500 The type [GenCmm d h g] represents a whole module,
\r
501 ** one list element per .o file **
\r
502 Without SplitObjs, the list has exactly one element
\r
504 newtype GenCmm d h g = Cmm [GenCmmTop d h g] -- A whole .o file
\r
505 data GenCmmTop d h g
\r
506 = CmmProc h g -- One procedure, graph d
\r
507 | CmmData <stuff> [d] -- Initialised data, items d
\r
509 Old and new piplines use different representations
\r
510 (CmmCvt.hs converts between the two)
\r
514 OLD BACK END representations (OldCmm.hs):
\r
515 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
517 newtype ListGraph i = ListGraph [GenBasicBlock i]
\r
519 data CmmStmt = Assign | Store | Return etc -- OLD BACK END ONLY
\r
522 Once the info tables are laid out, we replace CmmInfo with [CmmStatic]
\r
523 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
524 which represents the info tables as data, that should
\r
525 immediately precede the code
\r
528 NEW BACK END representations
\r
529 * Uses Hoopl library, a zero-boot package
\r
530 * CmmNode defines a node of a flow graph.
\r
531 * Cmm defines CmmGraph, CmmTop, Cmm
\r
532 - CmmGraph is a closed/closed graph + an entry node.
\r
534 data CmmGraph = CmmGraph { g_entry :: BlockId
\r
535 , g_graph :: Graph CmmNode C C }
\r
537 - CmmTop is a top level chunk, specialization of GenCmmTop from CmmDecl.hs
\r
538 with CmmGraph as a flow graph.
\r
539 - Cmm is a collection of CmmTops.
\r
541 type Cmm = GenCmm CmmStatic CmmTopInfo CmmGraph
\r
542 type CmmTop = GenCmmTop CmmStatic CmmTopInfo CmmGraph
\r
544 - CmmTop uses CmmTopInfo, which is a CmmInfoTable and CmmStackInfo
\r
546 data CmmTopInfo = TopInfo {info_tbl :: CmmInfoTable, stack_info :: CmmStackInfo}
\r
550 data CmmStackInfo = StackInfo {arg_space :: ByteOff, updfr_space :: Maybe ByteOff}
\r
552 * arg_space = SP offset on entry
\r
553 * updfr_space space = SP offset on exit
\r
554 Once the staci is manifested, we could drom CmmStackInfo, ie. get
\r
555 GenCmm CmmStatic CmmInfoTable CmmGraph, but we do not do that currently.
\r
558 * MkGraph.hs: smart constructors for Cmm.hs
\r
559 Beware, the CmmAGraph defined here does not use AGraph from Hoopl,
\r
560 as CmmAGraph can be opened or closed at exit, See the notes in that module.
\r
564 CmmDecl.hs - GenCmm and GenCmmTop types
\r
565 CmmExpr.hs - defines the Cmm expression types
\r
566 - CmmExpr, CmmReg, CmmLit, LocalReg, GlobalReg
\r
567 - Area, AreaId etc (separate module?)
\r
568 CmmType.hs - CmmType, Width etc (saparate module?)
\r
569 CmmMachOp.hs - MachOp and CallishMachOp types
\r
571 BlockId.hs defines BlockId, BlockEnv, BlockSet
\r