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
13 * "Remove redundant reloads" in CmmSpillReload should be redundant; since
\r
14 insertLateReloads is now gone, every reload is reloading a live variable.
\r
17 * Sink and inline S(RegSlot(x)) = e in precisely the same way that we
\r
18 sink and inline x = e
\r
20 * Stack layout is very like register assignment: find non-conflicting assigments.
\r
21 In particular we can use colouring or linear scan (etc).
\r
23 We'd fine-grain interference (on a word by word basis) to get maximum overlap.
\r
24 But that may make very big interference graphs. So linear scan might be
\r
27 NB: linear scan does on-the-fly live range splitting.
\r
29 * When stubbing dead slots be careful not to write into an area that
\r
30 overlaps with an area that's in use. So stubbing needs to *follow*
\r
36 In CmmNode, consider spliting CmmCall into two: call and jump
\r
38 Notes on new codegen (Aug 10)
\r
39 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\r
42 - We insert spills for variables before the stack check! This is the reason for
\r
43 some fishy code in StgCmmHeap.entryHeapCheck where we are doing some strange
\r
44 things to fix up the stack pointer before GC calls/jumps.
\r
46 The reason spills are inserted before the sp check is that at the entry to a
\r
47 function we always store the parameters passed in registers to local variables.
\r
48 The spill pass simply inserts spills at variable definitions. We instead should
\r
49 sink the spills so that we can avoid spilling them on branches that never
\r
52 This will fix the spill before stack check problem but only really as a side
\r
53 effect. A 'real fix' probably requires making the spiller know about sp checks.
\r
55 EZY: I don't understand this comment. David Terei, can you clarify?
\r
57 - Proc points pass all arguments on the stack, adding more code and
\r
58 slowing down things a lot. We either need to fix this or even better
\r
59 would be to get rid of proc points.
\r
61 - CmmInfo.cmmToRawCmm uses Old.Cmm, so it is called after converting Cmm.Cmm to
\r
62 Old.Cmm. We should abstract it to work on both representations, it needs only to
\r
63 convert a CmmInfoTable to [CmmStatic].
\r
65 - The MkGraph currenty uses a different semantics for <*> than Hoopl. Maybe
\r
66 we could convert codeGen/StgCmm* clients to the Hoopl's semantics?
\r
67 It's all deeply unsatisfactory.
\r
69 - Improve performance of Hoopl.
\r
71 A nofib comparison of -fasm vs -fnewcodegen nofib compilation parameters
\r
72 (using the same ghc-cmm branch +libraries compiled by the old codegenerator)
\r
73 is at http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghchoopl.txt
\r
74 - the code produced is 10.9% slower, the compilation is +118% slower!
\r
76 The same comparison with ghc-head with zip representation is at
\r
77 http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghczip.txt
\r
78 - the code produced is 11.7% slower, the compilation is +78% slower.
\r
80 When compiling nofib, ghc-cmm + libraries compiled with -fnew-codegen
\r
81 is 23.7% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.hooplghcoldgen.txt).
\r
82 When compiling nofib, ghc-head + libraries compiled with -fnew-codegen
\r
83 is 31.4% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.zipghcoldgen.txt).
\r
85 So we generate a bit better code, but it takes us longer!
\r
87 EZY: Also importantly, Hoopl uses dramatically more memory than the
\r
90 - Are all blockToNodeList and blockOfNodeList really needed? Maybe we could
\r
91 splice blocks instead?
\r
93 In the CmmContFlowOpt.blockConcat, using Dataflow seems too clumsy. Still,
\r
94 a block catenation function would be probably nicer than blockToNodeList
\r
95 / blockOfNodeList combo.
\r
97 - lowerSafeForeignCall seems too lowlevel. Just use Dataflow. After that
\r
98 delete splitEntrySeq from HooplUtils.
\r
100 - manifestSP seems to touch a lot of the graph representation. It is
\r
101 also slow for CmmSwitch nodes O(block_nodes * switch_statements).
\r
102 Maybe rewrite manifestSP to use Dataflow?
\r
104 - Sort out Label, LabelMap, LabelSet versus BlockId, BlockEnv, BlockSet
\r
105 dichotomy. Mostly this means global replace, but we also need to make
\r
106 Label an instance of Outputable (probably in the Outputable module).
\r
108 - NB that CmmProcPoint line 283 has a hack that works around a GADT-related
\r
111 - SDM (2010-02-26) can we remove the Foreign constructor from Convention?
\r
112 Reason: we never generate code for a function with the Foreign
\r
113 calling convention, and the code for calling foreign calls is generated
\r
115 - AsmCodeGen has a generic Cmm optimiser; move this into new pipeline
\r
116 EZY (2011-04-16): The mini-inliner has been generalized and ported,
\r
117 but the constant folding and other optimizations need to still be
\r
120 - AsmCodeGen has post-native-cg branch eliminator (shortCutBranches);
\r
121 we ultimately want to share this with the Cmm branch eliminator.
\r
123 - At the moment, references to global registers like Hp are "lowered"
\r
124 late (in CgUtils.fixStgRegisters). We should do this early, in the
\r
125 new native codegen, much in the way that we lower calling conventions.
\r
126 Might need to be a bit sophisticated about aliasing.
\r
128 - Question: currently we lift procpoints to become separate
\r
129 CmmProcs. Do we still want to do this?
\r
131 NB: and advantage of continuing to do this is that
\r
132 we can do common-proc elimination!
\r
134 - Move to new Cmm rep:
\r
135 * Make native CG consume New Cmm;
\r
136 * Convert Old Cmm->New Cmm to keep old path alive
\r
137 * Produce New Cmm when reading in .cmm files
\r
139 - Consider module names
\r
141 - Top-level SRT threading is a bit ugly
\r
143 - Add type/newtype for CmmModule = [CmmGroup] -- A module
\r
144 CmmGroup = [CmmTop] -- A .o file
\r
145 CmmTop = Proc | Data -- A procedure or data
\r
147 - This is a *change*: currently a CmmGroup is one function's-worth of code
\r
148 regardless of SplitObjs. Question: can we *always* generate M.o if there
\r
149 is just one element in the list (rather than M/M1.o, M/M2.o etc)
\r
153 - See "CAFs" below; we want to totally refactor the way SRTs are calculated
\r
155 - Pull out Areas into its own module
\r
156 Parameterise AreaMap (note there are type synonyms in CmmStackLayout!)
\r
157 Add ByteWidth = Int
\r
158 type SubArea = (Area, ByteOff, ByteWidth)
\r
159 ByteOff should not be defined in SMRep -- that is too high up the hierarchy
\r
161 - SMRep should not be imported by any module in cmm/! Make it so.
\r
162 -- ByteOff etc ==> CmmExpr
\r
163 -- rET_SMALL etc ==> CmmInfo
\r
164 Check that there are no other imports from codeGen in cmm/
\r
166 - If you eliminate a label by branch chain elimination,
\r
167 what happens if there's an Area associated with that label?
\r
169 - Think about a non-flattened representation?
\r
172 * Use record fields for LastCall!
\r
173 * cml_ret_off should be a ByteOff
\r
175 LastCall (which has a successor) and
\r
176 LastJump (which does not, includes return?)
\r
177 - does not have cml_cont, cml_ret_args, cml_ret_off
\r
180 - expands into save/MidForeignCall/restore/goto
\r
181 - like any LastCall, target of the call gets an info table
\r
183 - JD: remind self of what goes wrong if you turn off the
\r
184 liveness of the update frame
\r
186 - Garbage-collect http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/CPS
\r
187 moving good stuff into
\r
188 http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/NewCodeGenPipeline
\r
191 - We believe that all of CmmProcPoint.addProcPointProtocols is dead. What
\r
192 goes wrong if we simply never call it?
\r
194 - Something fishy in CmmStackLayout.hs
\r
195 * In particular, 'getAreaSize' returns an AreaMap, but we *know* the width of
\r
196 LocalRegs, so it'd be better to return FiniteMap AreaId ByteWidth
\r
197 * setSuccSPs looks fishy. Rather than lookin in procPoints, it could
\r
198 just lookup the block in areaSize which, after all, has a binding
\r
199 for precisely successors of calls. All other blocks (including proc
\r
200 points that are not successors of a call, we think) can be treated
\r
201 uniformly: zero-size Area, and use inSP.
\r
204 - Currently AsmCodeGen top level calls AsmCodeGen.cmmToCmm, which is a small
\r
205 C-- optimiser. It has quite a lot of boilerplate folding code in AsmCodeGen
\r
206 (cmmBlockConFold, cmmStmtConFold, cmmExprConFold), before calling out to
\r
207 CmmOpt. ToDo: see what optimisations are being done; and do them before
\r
210 - Modularise the CPS pipeline; instead of ...; A;B;C; ...
\r
213 - Most of HscMain.tryNewCodeGen does not belong in HscMain. Instead
\r
216 processCmm [including generating "raw" cmm]
\r
222 - If we stick CAF and stack liveness info on a LastCall node (not LastRet/Jump)
\r
223 then all CAF and stack liveness stuff be completed before we split
\r
224 into separate C procedures.
\r
227 compute and attach liveness into to LastCall
\r
228 right at end, split, cvt to old rep
\r
229 [must split before cvt, because old rep is not expressive enough]
\r
232 when old rep disappears,
\r
233 move the whole splitting game into the C back end *only*
\r
234 (guided by the procpoint set)
\r
236 ----------------------------------------------------
\r
238 ----------------------------------------------------
\r
240 -------- Testing stuff ------------
\r
241 HscMain.optionallyConvertAndOrCPS
\r
243 DynFlags: -fconvert-to-zipper-and-back, -frun-cpsz
\r
245 -------- Moribund stuff ------------
\r
246 OldCmm.hs Definition of flowgraph of old representation
\r
247 OldCmmUtil.hs Utilites that operates mostly on on CmmStmt
\r
248 OldPprCmm.hs Pretty print for CmmStmt, GenBasicBlock and ListGraph
\r
249 CmmCvt.hs Conversion between old and new Cmm reps
\r
250 CmmOpt.hs Hopefully-redundant optimiser
\r
252 -------- Stuff to keep ------------
\r
253 CmmCPS.hs Driver for new pipeline
\r
255 CmmLive.hs Liveness analysis, dead code elim
\r
256 CmmProcPoint.hs Identifying and splitting out proc-points
\r
258 CmmSpillReload.hs Save and restore across calls
\r
260 CmmCommonBlockElim.hs Common block elim
\r
261 CmmContFlowOpt.hs Other optimisations (branch-chain, merging)
\r
263 CmmBuildInfoTables.hs New info-table
\r
264 CmmStackLayout.hs and stack layout
\r
266 CmmInfo.hs Defn of InfoTables, and conversion to exact byte layout
\r
268 ---------- Cmm data types --------------
\r
269 Cmm.hs Cmm instantiations of dataflow graph framework
\r
270 MkGraph.hs Interface for building Cmm for codeGen/Stg*.hs modules
\r
272 CmmDecl.hs Shared Cmm types of both representations
\r
273 CmmExpr.hs Type of Cmm expression
\r
274 CmmType.hs Type of Cmm types and their widths
\r
275 CmmMachOp.hs MachOp type and accompanying utilities
\r
280 PprC.hs Pretty print Cmm in C syntax
\r
281 PprCmm.hs Pretty printer for CmmGraph.
\r
282 PprCmmDecl.hs Pretty printer for common Cmm types.
\r
283 PprCmmExpr.hs Pretty printer for Cmm expressions.
\r
286 BlockId.hs BlockId, BlockEnv, BlockSet
\r
288 ----------------------------------------------------
\r
289 Top-level structure
\r
290 ----------------------------------------------------
\r
292 * New codgen called in HscMain.hscGenHardCode, by calling HscMain.tryNewCodeGen,
\r
293 enabled by -fnew-codegen (Opt_TryNewCodeGen)
\r
295 THEN it calls CmmInfo.cmmToRawCmm to lay out the details of info tables
\r
296 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
297 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
299 * HscMain.tryNewCodeGen
\r
300 - STG->Cmm: StgCmm.codeGen (new codegen)
\r
301 - Optimise: CmmContFlowOpt (simple optimisations, very self contained)
\r
302 - Cps convert: CmmCPS.protoCmmCPS
\r
303 - Optimise: CmmContFlowOpt again
\r
304 - Convert: CmmCvt.cmmOfZgraph (convert to old rep) very self contained
\r
306 * StgCmm.hs The new STG -> Cmm conversion code generator
\r
307 Lots of modules StgCmmXXX
\r
310 ----------------------------------------------------
\r
311 CmmCPS.protoCmmCPS The new pipeline
\r
312 ----------------------------------------------------
\r
314 CmmCPS.protoCmmCPS:
\r
315 1. Do cpsTop for each procedures separately
\r
316 2. Build SRT representation; this spans multiple procedures
\r
317 (unless split-objs)
\r
320 * CmmCommonBlockElim.elimCommonBlocks:
\r
321 eliminate common blocks
\r
323 * CmmProcPoint.minimalProcPointSet
\r
324 identify proc-points
\r
327 * CmmProcPoint.addProcPointProtocols
\r
328 something to do with the MA optimisation
\r
329 probably entirely unnecessary
\r
331 * Spill and reload:
\r
332 - CmmSpillReload.dualLivenessWithInsertion
\r
333 insert spills/reloads across
\r
335 Branches to proc-points
\r
336 Now sink those reloads (and other instructions):
\r
337 - CmmSpillReload.rewriteAssignments
\r
338 - CmmSpillReload.removeDeadAssignmentsAndReloads
\r
340 * CmmStackLayout.stubSlotsOnDeath
\r
341 debug only: zero out dead slots when they die
\r
344 - CmmStackLayout.lifeSlotAnal:
\r
345 find which sub-areas are live on entry to each block
\r
347 - CmmStackLayout.layout
\r
348 Lay out the stack, returning an AreaMap
\r
349 type AreaMap = FiniteMap Area ByteOff
\r
350 -- Byte offset of the oldest byte of the Area,
\r
351 -- relative to the oldest byte of the Old Area
\r
353 - CmmStackLayout.manifestSP
\r
354 Manifest the stack pointer
\r
356 * Split into separate procedures
\r
357 - CmmProcPoint.procPointAnalysis
\r
358 Given set of proc points, which blocks are reachable from each
\r
359 Claim: too few proc-points => code duplication, but program still works??
\r
361 - CmmProcPoint.splitAtProcPoints
\r
362 Using this info, split into separate procedures
\r
364 - CmmBuildInfoTables.setInfoTableStackMap
\r
365 Attach stack maps to each info table
\r
368 ----------------------------------------------------
\r
370 ----------------------------------------------------
\r
372 Consider this program, which has a diamond control flow,
\r
373 with a call on one branch
\r
376 if b then { ... f(x) ...; q=5; goto J }
\r
377 else { ...; q=7; goto J }
\r
380 then the join point J is a "proc-point". So, is 'p' passed to J
\r
381 as a parameter? Or, if 'p' was saved on the stack anyway, perhaps
\r
382 to keep it alive across the call to h(), maybe 'p' gets communicated
\r
383 to J that way. This is an awkward choice. (We think that we currently
\r
384 never pass variables to join points via arguments.)
\r
386 Furthermore, there is *no way* to pass q to J in a register (other
\r
387 than a parameter register).
\r
389 What we want is to do register allocation across the whole caboodle.
\r
390 Then we could drop all the code that deals with the above awkward
\r
391 decisions about spilling variables across proc-points.
\r
393 Note that J doesn't need an info table.
\r
395 What we really want is for each LastCall (not LastJump/Ret)
\r
396 to have an info table. Note that ProcPoints that are not successors
\r
397 of calls don't need an info table.
\r
399 Figuring out proc-points
\r
400 ~~~~~~~~~~~~~~~~~~~~~~~~
\r
401 Proc-points are identified by
\r
402 CmmProcPoint.minimalProcPointSet/extendPPSet Although there isn't
\r
403 that much code, JD thinks that it could be done much more nicely using
\r
404 a dominator analysis, using the Dataflow Engine.
\r
406 ----------------------------------------------------
\r
408 ----------------------------------------------------
\r
410 * The code for a procedure f may refer to either the *closure*
\r
411 or the *entry point* of another top-level procedure g.
\r
412 If f is live, then so is g. f's SRT must include g's closure.
\r
414 * The CLabel for the entry-point/closure reveals whether g is
\r
415 a CAF (or refers to CAFs). See the IdLabel constructor of CLabel.
\r
417 * The CAF-ness of the original top-level defininions is figured out
\r
418 (by TidyPgm) before we generate C--. This CafInfo is only set for
\r
419 top-level Ids; nested bindings stay with MayHaveCafRefs.
\r
421 * Currently an SRT contains (only) pointers to (top-level) closures.
\r
423 * Consider this Core code
\r
424 f = \x -> let g = \y -> ...x...y...h1...
\r
426 and suppose that h1, h2 have IdInfo of MayHaveCafRefs.
\r
427 Therefore, so will f, But g will not (since it's nested).
\r
429 This generates C-- roughly like this:
\r
430 f_closure: .word f_entry
\r
431 f_entry() [info-tbl-for-f] { ...jump g_entry...jump h2... }
\r
432 g_entry() [info-tbl-for-g] { ...jump h1... }
\r
434 Note that there is no top-level closure for g (only an info table).
\r
435 This fact (whether or not there is a top-level closure) is recorded
\r
436 in the InfoTable attached to the CmmProc for f, g
\r
438 Any out-of-Group references to an IdLabel goes to
\r
439 a Proc whose InfoTable says "I have a top-level closure".
\r
441 A CmmProc whose InfoTable says "I do not have a top-level
\r
442 closure" is referred to only from its own Group.
\r
444 * So: info-tbl-for-f must have an SRT that keeps h1,h2 alive
\r
445 info-tbl-for-g must have an SRT that keeps h1 (only) alive
\r
447 But if we just look for the free CAF refs, we get:
\r
451 So we need to do a transitive closure thing to flesh out
\r
452 f's keep-alive refs to include h1.
\r
454 * The SRT info is the C_SRT field of Cmm.ClosureTypeInfo in a
\r
455 CmmInfoTable attached to each CmmProc. CmmCPS.toTops actually does
\r
456 the attaching, right at the end of the pipeline. The C_SRT part
\r
457 gives offsets within a single, shared table of closure pointers.
\r
459 * DECIDED: we can generate SRTs based on the final Cmm program
\r
460 without knowledge of how it is generated.
\r
462 ----------------------------------------------------
\r
464 ----------------------------------------------------
\r
466 See Note [Foreign calls] in CmmNode! This explains that a safe
\r
467 foreign call must do this:
\r
469 push info table (on thread stack) to describe frame
\r
470 make call (via C stack)
\r
472 restore thread state
\r
473 and explains why this expansion must be done late in the day.
\r
476 - Every foreign call is represented as a middle node
\r
478 - *Unsafe* foreign calls are simply "fat machine instructions"
\r
479 and are passed along to the native code generator
\r
481 - *Safe* foreign calls are "lowered" to unsafe calls by wrapping
\r
482 them in the above save/restore sequence. This step is done
\r
483 very late in the pipeline, just before handing to the native
\r
486 This lowering is done by BuildInfoTables.lowerSafeForeignCalls
\r
489 NEW PLAN for foreign calls:
\r
490 - Unsafe foreign calls remain as a middle node (fat machine instruction)
\r
491 Even the parameter passing is not lowered (just as machine instrs
\r
494 - Initially, safe foreign calls appear as LastCalls with
\r
497 ----------------------------------------------------
\r
498 Cmm representations
\r
499 ----------------------------------------------------
\r
502 The type [GenCmm d h g] represents a whole module,
\r
503 ** one list element per .o file **
\r
504 Without SplitObjs, the list has exactly one element
\r
506 newtype GenCmm d h g = Cmm [GenCmmTop d h g] -- A whole .o file
\r
507 data GenCmmTop d h g
\r
508 = CmmProc h g -- One procedure, graph d
\r
509 | CmmData <stuff> [d] -- Initialised data, items d
\r
511 Old and new piplines use different representations
\r
512 (CmmCvt.hs converts between the two)
\r
516 OLD BACK END representations (OldCmm.hs):
\r
517 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
519 newtype ListGraph i = ListGraph [GenBasicBlock i]
\r
521 data CmmStmt = Assign | Store | Return etc -- OLD BACK END ONLY
\r
524 Once the info tables are laid out, we replace CmmInfo with [CmmStatic]
\r
525 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
526 which represents the info tables as data, that should
\r
527 immediately precede the code
\r
530 NEW BACK END representations
\r
531 * Uses Hoopl library, a zero-boot package
\r
532 * CmmNode defines a node of a flow graph.
\r
533 * Cmm defines CmmGraph, CmmTop, Cmm
\r
534 - CmmGraph is a closed/closed graph + an entry node.
\r
536 data CmmGraph = CmmGraph { g_entry :: BlockId
\r
537 , g_graph :: Graph CmmNode C C }
\r
539 - CmmTop is a top level chunk, specialization of GenCmmTop from CmmDecl.hs
\r
540 with CmmGraph as a flow graph.
\r
541 - Cmm is a collection of CmmTops.
\r
543 type Cmm = GenCmm CmmStatic CmmTopInfo CmmGraph
\r
544 type CmmTop = GenCmmTop CmmStatic CmmTopInfo CmmGraph
\r
546 - CmmTop uses CmmTopInfo, which is a CmmInfoTable and CmmStackInfo
\r
548 data CmmTopInfo = TopInfo {info_tbl :: CmmInfoTable, stack_info :: CmmStackInfo}
\r
552 data CmmStackInfo = StackInfo {arg_space :: ByteOff, updfr_space :: Maybe ByteOff}
\r
554 * arg_space = SP offset on entry
\r
555 * updfr_space space = SP offset on exit
\r
556 Once the staci is manifested, we could drom CmmStackInfo, ie. get
\r
557 GenCmm CmmStatic CmmInfoTable CmmGraph, but we do not do that currently.
\r
560 * MkGraph.hs: smart constructors for Cmm.hs
\r
561 Beware, the CmmAGraph defined here does not use AGraph from Hoopl,
\r
562 as CmmAGraph can be opened or closed at exit, See the notes in that module.
\r
566 CmmDecl.hs - GenCmm and GenCmmTop types
\r
567 CmmExpr.hs - defines the Cmm expression types
\r
568 - CmmExpr, CmmReg, CmmLit, LocalReg, GlobalReg
\r
569 - Area, AreaId etc (separate module?)
\r
570 CmmType.hs - CmmType, Width etc (saparate module?)
\r
571 CmmMachOp.hs - MachOp and CallishMachOp types
\r
573 BlockId.hs defines BlockId, BlockEnv, BlockSet
\r