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 * Move top and tail calls to runCmmContFlowOpts from HscMain to CmmCps.cpsTop
\r
12 (and rename the latter!)
\r
14 * "Remove redundant reloads" in CmmSpillReload should be redundant; since
\r
15 insertLateReloads is now gone, every reload is reloading a live variable.
\r
18 * Sink and inline S(RegSlot(x)) = e in precisely the same way that we
\r
19 sink and inline x = e
\r
21 * Stack layout is very like register assignment: find non-conflicting assigments.
\r
22 In particular we can use colouring or linear scan (etc).
\r
24 We'd fine-grain interference (on a word by word basis) to get maximum overlap.
\r
25 But that may make very big interference graphs. So linear scan might be
\r
28 NB: linear scan does on-the-fly live range splitting.
\r
30 * When stubbing dead slots be careful not to write into an area that
\r
31 overlaps with an area that's in use. So stubbing needs to *follow*
\r
37 In CmmNode, consider spliting CmmCall into two: call and jump
\r
39 Notes on new codegen (Aug 10)
\r
40 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\r
43 - We insert spills for variables before the stack check! This is the reason for
\r
44 some fishy code in StgCmmHeap.entryHeapCheck where we are doing some strange
\r
45 things to fix up the stack pointer before GC calls/jumps.
\r
47 The reason spills are inserted before the sp check is that at the entry to a
\r
48 function we always store the parameters passed in registers to local variables.
\r
49 The spill pass simply inserts spills at variable definitions. We instead should
\r
50 sink the spills so that we can avoid spilling them on branches that never
\r
53 This will fix the spill before stack check problem but only really as a side
\r
54 effect. A 'real fix' probably requires making the spiller know about sp checks.
\r
56 EZY: I don't understand this comment. David Terei, can you clarify?
\r
58 - Proc points pass all arguments on the stack, adding more code and
\r
59 slowing down things a lot. We either need to fix this or even better
\r
60 would be to get rid of proc points.
\r
62 - CmmInfo.cmmToRawCmm uses Old.Cmm, so it is called after converting Cmm.Cmm to
\r
63 Old.Cmm. We should abstract it to work on both representations, it needs only to
\r
64 convert a CmmInfoTable to [CmmStatic].
\r
66 - The MkGraph currenty uses a different semantics for <*> than Hoopl. Maybe
\r
67 we could convert codeGen/StgCmm* clients to the Hoopl's semantics?
\r
68 It's all deeply unsatisfactory.
\r
70 - Improve performance of Hoopl.
\r
72 A nofib comparison of -fasm vs -fnewcodegen nofib compilation parameters
\r
73 (using the same ghc-cmm branch +libraries compiled by the old codegenerator)
\r
74 is at http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghchoopl.txt
\r
75 - the code produced is 10.9% slower, the compilation is +118% slower!
\r
77 The same comparison with ghc-head with zip representation is at
\r
78 http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghczip.txt
\r
79 - the code produced is 11.7% slower, the compilation is +78% slower.
\r
81 When compiling nofib, ghc-cmm + libraries compiled with -fnew-codegen
\r
82 is 23.7% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.hooplghcoldgen.txt).
\r
83 When compiling nofib, ghc-head + libraries compiled with -fnew-codegen
\r
84 is 31.4% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.zipghcoldgen.txt).
\r
86 So we generate a bit better code, but it takes us longer!
\r
88 EZY: Also importantly, Hoopl uses dramatically more memory than the
\r
91 - Are all blockToNodeList and blockOfNodeList really needed? Maybe we could
\r
92 splice blocks instead?
\r
94 In the CmmContFlowOpt.blockConcat, using Dataflow seems too clumsy. Still,
\r
95 a block catenation function would be probably nicer than blockToNodeList
\r
96 / blockOfNodeList combo.
\r
98 - lowerSafeForeignCall seems too lowlevel. Just use Dataflow. After that
\r
99 delete splitEntrySeq from HooplUtils.
\r
101 - manifestSP seems to touch a lot of the graph representation. It is
\r
102 also slow for CmmSwitch nodes O(block_nodes * switch_statements).
\r
103 Maybe rewrite manifestSP to use Dataflow?
\r
105 - Sort out Label, LabelMap, LabelSet versus BlockId, BlockEnv, BlockSet
\r
106 dichotomy. Mostly this means global replace, but we also need to make
\r
107 Label an instance of Outputable (probably in the Outputable module).
\r
109 - NB that CmmProcPoint line 283 has a hack that works around a GADT-related
\r
112 - SDM (2010-02-26) can we remove the Foreign constructor from Convention?
\r
113 Reason: we never generate code for a function with the Foreign
\r
114 calling convention, and the code for calling foreign calls is generated
\r
116 - AsmCodeGen has a generic Cmm optimiser; move this into new pipeline
\r
117 EZY (2011-04-16): The mini-inliner has been generalized and ported,
\r
118 but the constant folding and other optimizations need to still be
\r
121 - AsmCodeGen has post-native-cg branch eliminator (shortCutBranches);
\r
122 we ultimately want to share this with the Cmm branch eliminator.
\r
124 - At the moment, references to global registers like Hp are "lowered"
\r
125 late (in CgUtils.fixStgRegisters). We should do this early, in the
\r
126 new native codegen, much in the way that we lower calling conventions.
\r
127 Might need to be a bit sophisticated about aliasing.
\r
129 - Question: currently we lift procpoints to become separate
\r
130 CmmProcs. Do we still want to do this?
\r
132 NB: and advantage of continuing to do this is that
\r
133 we can do common-proc elimination!
\r
135 - Move to new Cmm rep:
\r
136 * Make native CG consume New Cmm;
\r
137 * Convert Old Cmm->New Cmm to keep old path alive
\r
138 * Produce New Cmm when reading in .cmm files
\r
140 - Consider module names
\r
142 - Top-level SRT threading is a bit ugly
\r
144 - Add type/newtype for CmmModule = [CmmGroup] -- A module
\r
145 CmmGroup = [CmmTop] -- A .o file
\r
146 CmmTop = Proc | Data -- A procedure or data
\r
148 - This is a *change*: currently a CmmGroup is one function's-worth of code
\r
149 regardless of SplitObjs. Question: can we *always* generate M.o if there
\r
150 is just one element in the list (rather than M/M1.o, M/M2.o etc)
\r
154 - See "CAFs" below; we want to totally refactor the way SRTs are calculated
\r
156 - Pull out Areas into its own module
\r
157 Parameterise AreaMap (note there are type synonyms in CmmStackLayout!)
\r
158 Add ByteWidth = Int
\r
159 type SubArea = (Area, ByteOff, ByteWidth)
\r
160 ByteOff should not be defined in SMRep -- that is too high up the hierarchy
\r
162 - SMRep should not be imported by any module in cmm/! Make it so.
\r
163 -- ByteOff etc ==> CmmExpr
\r
164 -- rET_SMALL etc ==> CmmInfo
\r
165 Check that there are no other imports from codeGen in cmm/
\r
167 - If you eliminate a label by branch chain elimination,
\r
168 what happens if there's an Area associated with that label?
\r
170 - Think about a non-flattened representation?
\r
173 * Use record fields for LastCall!
\r
174 * cml_ret_off should be a ByteOff
\r
176 LastCall (which has a successor) and
\r
177 LastJump (which does not, includes return?)
\r
178 - does not have cml_cont, cml_ret_args, cml_ret_off
\r
181 - expands into save/MidForeignCall/restore/goto
\r
182 - like any LastCall, target of the call gets an info table
\r
184 - JD: remind self of what goes wrong if you turn off the
\r
185 liveness of the update frame
\r
187 - Garbage-collect http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/CPS
\r
188 moving good stuff into
\r
189 http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/NewCodeGenPipeline
\r
192 - We believe that all of CmmProcPoint.addProcPointProtocols is dead. What
\r
193 goes wrong if we simply never call it?
\r
195 - Something fishy in CmmStackLayout.hs
\r
196 * In particular, 'getAreaSize' returns an AreaMap, but we *know* the width of
\r
197 LocalRegs, so it'd be better to return FiniteMap AreaId ByteWidth
\r
198 * setSuccSPs looks fishy. Rather than lookin in procPoints, it could
\r
199 just lookup the block in areaSize which, after all, has a binding
\r
200 for precisely successors of calls. All other blocks (including proc
\r
201 points that are not successors of a call, we think) can be treated
\r
202 uniformly: zero-size Area, and use inSP.
\r
205 - Currently AsmCodeGen top level calls AsmCodeGen.cmmToCmm, which is a small
\r
206 C-- optimiser. It has quite a lot of boilerplate folding code in AsmCodeGen
\r
207 (cmmBlockConFold, cmmStmtConFold, cmmExprConFold), before calling out to
\r
208 CmmOpt. ToDo: see what optimisations are being done; and do them before
\r
211 - Modularise the CPS pipeline; instead of ...; A;B;C; ...
\r
214 - Most of HscMain.tryNewCodeGen does not belong in HscMain. Instead
\r
217 processCmm [including generating "raw" cmm]
\r
223 - If we stick CAF and stack liveness info on a LastCall node (not LastRet/Jump)
\r
224 then all CAF and stack liveness stuff be completed before we split
\r
225 into separate C procedures.
\r
228 compute and attach liveness into to LastCall
\r
229 right at end, split, cvt to old rep
\r
230 [must split before cvt, because old rep is not expressive enough]
\r
233 when old rep disappears,
\r
234 move the whole splitting game into the C back end *only*
\r
235 (guided by the procpoint set)
\r
237 ----------------------------------------------------
\r
239 ----------------------------------------------------
\r
241 -------- Testing stuff ------------
\r
242 HscMain.optionallyConvertAndOrCPS
\r
244 DynFlags: -fconvert-to-zipper-and-back, -frun-cpsz
\r
246 -------- Moribund stuff ------------
\r
247 OldCmm.hs Definition of flowgraph of old representation
\r
248 OldCmmUtil.hs Utilites that operates mostly on on CmmStmt
\r
249 OldPprCmm.hs Pretty print for CmmStmt, GenBasicBlock and ListGraph
\r
250 CmmCvt.hs Conversion between old and new Cmm reps
\r
251 CmmOpt.hs Hopefully-redundant optimiser
\r
253 -------- Stuff to keep ------------
\r
254 CmmCPS.hs Driver for new pipeline
\r
256 CmmLive.hs Liveness analysis, dead code elim
\r
257 CmmProcPoint.hs Identifying and splitting out proc-points
\r
259 CmmSpillReload.hs Save and restore across calls
\r
261 CmmCommonBlockElim.hs Common block elim
\r
262 CmmContFlowOpt.hs Other optimisations (branch-chain, merging)
\r
264 CmmBuildInfoTables.hs New info-table
\r
265 CmmStackLayout.hs and stack layout
\r
267 CmmInfo.hs Defn of InfoTables, and conversion to exact byte layout
\r
269 ---------- Cmm data types --------------
\r
270 Cmm.hs Cmm instantiations of dataflow graph framework
\r
271 MkGraph.hs Interface for building Cmm for codeGen/Stg*.hs modules
\r
273 CmmDecl.hs Shared Cmm types of both representations
\r
274 CmmExpr.hs Type of Cmm expression
\r
275 CmmType.hs Type of Cmm types and their widths
\r
276 CmmMachOp.hs MachOp type and accompanying utilities
\r
281 PprC.hs Pretty print Cmm in C syntax
\r
282 PprCmm.hs Pretty printer for CmmGraph.
\r
283 PprCmmDecl.hs Pretty printer for common Cmm types.
\r
284 PprCmmExpr.hs Pretty printer for Cmm expressions.
\r
287 BlockId.hs BlockId, BlockEnv, BlockSet
\r
289 ----------------------------------------------------
\r
290 Top-level structure
\r
291 ----------------------------------------------------
\r
293 * New codgen called in HscMain.hscGenHardCode, by calling HscMain.tryNewCodeGen,
\r
294 enabled by -fnew-codegen (Opt_TryNewCodeGen)
\r
296 THEN it calls CmmInfo.cmmToRawCmm to lay out the details of info tables
\r
297 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
298 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
300 * HscMain.tryNewCodeGen
\r
301 - STG->Cmm: StgCmm.codeGen (new codegen)
\r
302 - Optimise: CmmContFlowOpt (simple optimisations, very self contained)
\r
303 - Cps convert: CmmCPS.protoCmmCPS
\r
304 - Optimise: CmmContFlowOpt again
\r
305 - Convert: CmmCvt.cmmOfZgraph (convert to old rep) very self contained
\r
307 * StgCmm.hs The new STG -> Cmm conversion code generator
\r
308 Lots of modules StgCmmXXX
\r
311 ----------------------------------------------------
\r
312 CmmCPS.protoCmmCPS The new pipeline
\r
313 ----------------------------------------------------
\r
315 CmmCPS.protoCmmCPS:
\r
316 1. Do cpsTop for each procedures separately
\r
317 2. Build SRT representation; this spans multiple procedures
\r
318 (unless split-objs)
\r
321 * CmmCommonBlockElim.elimCommonBlocks:
\r
322 eliminate common blocks
\r
324 * CmmProcPoint.minimalProcPointSet
\r
325 identify proc-points
\r
328 * CmmProcPoint.addProcPointProtocols
\r
329 something to do with the MA optimisation
\r
330 probably entirely unnecessary
\r
332 * Spill and reload:
\r
333 - CmmSpillReload.dualLivenessWithInsertion
\r
334 insert spills/reloads across
\r
336 Branches to proc-points
\r
337 Now sink those reloads (and other instructions):
\r
338 - CmmSpillReload.rewriteAssignments
\r
339 - CmmSpillReload.removeDeadAssignmentsAndReloads
\r
341 * CmmStackLayout.stubSlotsOnDeath
\r
342 debug only: zero out dead slots when they die
\r
345 - CmmStackLayout.lifeSlotAnal:
\r
346 find which sub-areas are live on entry to each block
\r
348 - CmmStackLayout.layout
\r
349 Lay out the stack, returning an AreaMap
\r
350 type AreaMap = FiniteMap Area ByteOff
\r
351 -- Byte offset of the oldest byte of the Area,
\r
352 -- relative to the oldest byte of the Old Area
\r
354 - CmmStackLayout.manifestSP
\r
355 Manifest the stack pointer
\r
357 * Split into separate procedures
\r
358 - CmmProcPoint.procPointAnalysis
\r
359 Given set of proc points, which blocks are reachable from each
\r
360 Claim: too few proc-points => code duplication, but program still works??
\r
362 - CmmProcPoint.splitAtProcPoints
\r
363 Using this info, split into separate procedures
\r
365 - CmmBuildInfoTables.setInfoTableStackMap
\r
366 Attach stack maps to each info table
\r
369 ----------------------------------------------------
\r
371 ----------------------------------------------------
\r
373 Consider this program, which has a diamond control flow,
\r
374 with a call on one branch
\r
377 if b then { ... f(x) ...; q=5; goto J }
\r
378 else { ...; q=7; goto J }
\r
381 then the join point J is a "proc-point". So, is 'p' passed to J
\r
382 as a parameter? Or, if 'p' was saved on the stack anyway, perhaps
\r
383 to keep it alive across the call to h(), maybe 'p' gets communicated
\r
384 to J that way. This is an awkward choice. (We think that we currently
\r
385 never pass variables to join points via arguments.)
\r
387 Furthermore, there is *no way* to pass q to J in a register (other
\r
388 than a parameter register).
\r
390 What we want is to do register allocation across the whole caboodle.
\r
391 Then we could drop all the code that deals with the above awkward
\r
392 decisions about spilling variables across proc-points.
\r
394 Note that J doesn't need an info table.
\r
396 What we really want is for each LastCall (not LastJump/Ret)
\r
397 to have an info table. Note that ProcPoints that are not successors
\r
398 of calls don't need an info table.
\r
400 Figuring out proc-points
\r
401 ~~~~~~~~~~~~~~~~~~~~~~~~
\r
402 Proc-points are identified by
\r
403 CmmProcPoint.minimalProcPointSet/extendPPSet Although there isn't
\r
404 that much code, JD thinks that it could be done much more nicely using
\r
405 a dominator analysis, using the Dataflow Engine.
\r
407 ----------------------------------------------------
\r
409 ----------------------------------------------------
\r
411 * The code for a procedure f may refer to either the *closure*
\r
412 or the *entry point* of another top-level procedure g.
\r
413 If f is live, then so is g. f's SRT must include g's closure.
\r
415 * The CLabel for the entry-point/closure reveals whether g is
\r
416 a CAF (or refers to CAFs). See the IdLabel constructor of CLabel.
\r
418 * The CAF-ness of the original top-level defininions is figured out
\r
419 (by TidyPgm) before we generate C--. This CafInfo is only set for
\r
420 top-level Ids; nested bindings stay with MayHaveCafRefs.
\r
422 * Currently an SRT contains (only) pointers to (top-level) closures.
\r
424 * Consider this Core code
\r
425 f = \x -> let g = \y -> ...x...y...h1...
\r
427 and suppose that h1, h2 have IdInfo of MayHaveCafRefs.
\r
428 Therefore, so will f, But g will not (since it's nested).
\r
430 This generates C-- roughly like this:
\r
431 f_closure: .word f_entry
\r
432 f_entry() [info-tbl-for-f] { ...jump g_entry...jump h2... }
\r
433 g_entry() [info-tbl-for-g] { ...jump h1... }
\r
435 Note that there is no top-level closure for g (only an info table).
\r
436 This fact (whether or not there is a top-level closure) is recorded
\r
437 in the InfoTable attached to the CmmProc for f, g
\r
439 Any out-of-Group references to an IdLabel goes to
\r
440 a Proc whose InfoTable says "I have a top-level closure".
\r
442 A CmmProc whose InfoTable says "I do not have a top-level
\r
443 closure" is referred to only from its own Group.
\r
445 * So: info-tbl-for-f must have an SRT that keeps h1,h2 alive
\r
446 info-tbl-for-g must have an SRT that keeps h1 (only) alive
\r
448 But if we just look for the free CAF refs, we get:
\r
452 So we need to do a transitive closure thing to flesh out
\r
453 f's keep-alive refs to include h1.
\r
455 * The SRT info is the C_SRT field of Cmm.ClosureTypeInfo in a
\r
456 CmmInfoTable attached to each CmmProc. CmmCPS.toTops actually does
\r
457 the attaching, right at the end of the pipeline. The C_SRT part
\r
458 gives offsets within a single, shared table of closure pointers.
\r
460 * DECIDED: we can generate SRTs based on the final Cmm program
\r
461 without knowledge of how it is generated.
\r
463 ----------------------------------------------------
\r
465 ----------------------------------------------------
\r
467 See Note [Foreign calls] in CmmNode! This explains that a safe
\r
468 foreign call must do this:
\r
470 push info table (on thread stack) to describe frame
\r
471 make call (via C stack)
\r
473 restore thread state
\r
474 and explains why this expansion must be done late in the day.
\r
477 - Every foreign call is represented as a middle node
\r
479 - *Unsafe* foreign calls are simply "fat machine instructions"
\r
480 and are passed along to the native code generator
\r
482 - *Safe* foreign calls are "lowered" to unsafe calls by wrapping
\r
483 them in the above save/restore sequence. This step is done
\r
484 very late in the pipeline, just before handing to the native
\r
487 This lowering is done by BuildInfoTables.lowerSafeForeignCalls
\r
490 NEW PLAN for foreign calls:
\r
491 - Unsafe foreign calls remain as a middle node (fat machine instruction)
\r
492 Even the parameter passing is not lowered (just as machine instrs
\r
495 - Initially, safe foreign calls appear as LastCalls with
\r
498 ----------------------------------------------------
\r
499 Cmm representations
\r
500 ----------------------------------------------------
\r
503 The type [GenCmm d h g] represents a whole module,
\r
504 ** one list element per .o file **
\r
505 Without SplitObjs, the list has exactly one element
\r
507 newtype GenCmm d h g = Cmm [GenCmmTop d h g] -- A whole .o file
\r
508 data GenCmmTop d h g
\r
509 = CmmProc h g -- One procedure, graph d
\r
510 | CmmData <stuff> [d] -- Initialised data, items d
\r
512 Old and new piplines use different representations
\r
513 (CmmCvt.hs converts between the two)
\r
517 OLD BACK END representations (OldCmm.hs):
\r
518 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
520 newtype ListGraph i = ListGraph [GenBasicBlock i]
\r
522 data CmmStmt = Assign | Store | Return etc -- OLD BACK END ONLY
\r
525 Once the info tables are laid out, we replace CmmInfo with [CmmStatic]
\r
526 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
527 which represents the info tables as data, that should
\r
528 immediately precede the code
\r
531 NEW BACK END representations
\r
532 * Uses Hoopl library, a zero-boot package
\r
533 * CmmNode defines a node of a flow graph.
\r
534 * Cmm defines CmmGraph, CmmTop, Cmm
\r
535 - CmmGraph is a closed/closed graph + an entry node.
\r
537 data CmmGraph = CmmGraph { g_entry :: BlockId
\r
538 , g_graph :: Graph CmmNode C C }
\r
540 - CmmTop is a top level chunk, specialization of GenCmmTop from CmmDecl.hs
\r
541 with CmmGraph as a flow graph.
\r
542 - Cmm is a collection of CmmTops.
\r
544 type Cmm = GenCmm CmmStatic CmmTopInfo CmmGraph
\r
545 type CmmTop = GenCmmTop CmmStatic CmmTopInfo CmmGraph
\r
547 - CmmTop uses CmmTopInfo, which is a CmmInfoTable and CmmStackInfo
\r
549 data CmmTopInfo = TopInfo {info_tbl :: CmmInfoTable, stack_info :: CmmStackInfo}
\r
553 data CmmStackInfo = StackInfo {arg_space :: ByteOff, updfr_space :: Maybe ByteOff}
\r
555 * arg_space = SP offset on entry
\r
556 * updfr_space space = SP offset on exit
\r
557 Once the staci is manifested, we could drom CmmStackInfo, ie. get
\r
558 GenCmm CmmStatic CmmInfoTable CmmGraph, but we do not do that currently.
\r
561 * MkGraph.hs: smart constructors for Cmm.hs
\r
562 Beware, the CmmAGraph defined here does not use AGraph from Hoopl,
\r
563 as CmmAGraph can be opened or closed at exit, See the notes in that module.
\r
567 CmmDecl.hs - GenCmm and GenCmmTop types
\r
568 CmmExpr.hs - defines the Cmm expression types
\r
569 - CmmExpr, CmmReg, CmmLit, LocalReg, GlobalReg
\r
570 - Area, AreaId etc (separate module?)
\r
571 CmmType.hs - CmmType, Width etc (saparate module?)
\r
572 CmmMachOp.hs - MachOp and CallishMachOp types
\r
574 BlockId.hs defines BlockId, BlockEnv, BlockSet
\r