3 * Bake in ByteOff to ParamLocation and ArgumentFormat
\r
4 CmmActuals -> [CmmActual] similary CmmFormals
\r
6 * Possible refactoring: Nuke AGraph in favour of
\r
7 mkIfThenElse :: Expr -> Graph -> Graph -> FCode Graph
\r
9 mkIfThenElse :: HasUniques m => Expr -> Graph -> Graph -> m Graph
\r
10 (Remmber that the .cmm file parser must use this function)
\r
12 or parameterise FCode over its envt; the CgState part seem useful for both
\r
14 * Move top and tail calls to runCmmContFlowOpts from HscMain to CmmCps.cpsTop
\r
15 (and rename the latter!)
\r
17 * "Remove redundant reloads" in CmmSpillReload should be redundant; since
\r
18 insertLateReloads is now gone, every reload is reloading a live variable.
\r
21 * Sink and inline S(RegSlot(x)) = e in precisely the same way that we
\r
22 sink and inline x = e
\r
24 * Stack layout is very like register assignment: find non-conflicting assigments.
\r
25 In particular we can use colouring or linear scan (etc).
\r
27 We'd fine-grain interference (on a word by word basis) to get maximum overlap.
\r
28 But that may make very big interference graphs. So linear scan might be
\r
31 NB: linear scan does on-the-fly live range splitting.
\r
33 * When stubbing dead slots be careful not to write into an area that
\r
34 overlaps with an area that's in use. So stubbing needs to *follow*
\r
40 In CmmNode, consider spliting CmmCall into two: call and jump
\r
42 Notes on new codegen (Aug 10)
\r
43 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\r
46 - We insert spills for variables before the stack check! This is the reason for
\r
47 some fishy code in StgCmmHeap.entryHeapCheck where we are doing some strange
\r
48 things to fix up the stack pointer before GC calls/jumps.
\r
50 The reason spills are inserted before the sp check is that at the entry to a
\r
51 function we always store the parameters passed in registers to local variables.
\r
52 The spill pass simply inserts spills at variable definitions. We instead should
\r
53 sink the spills so that we can avoid spilling them on branches that never
\r
56 This will fix the spill before stack check problem but only really as a side
\r
57 effect. A 'real fix' probably requires making the spiller know about sp checks.
\r
59 EZY: I don't understand this comment. David Terei, can you clarify?
\r
61 - Proc points pass all arguments on the stack, adding more code and
\r
62 slowing down things a lot. We either need to fix this or even better
\r
63 would be to get rid of proc points.
\r
65 - CmmInfo.cmmToRawCmm uses Old.Cmm, so it is called after converting Cmm.Cmm to
\r
66 Old.Cmm. We should abstract it to work on both representations, it needs only to
\r
67 convert a CmmInfoTable to [CmmStatic].
\r
69 - The MkGraph currenty uses a different semantics for <*> than Hoopl. Maybe
\r
70 we could convert codeGen/StgCmm* clients to the Hoopl's semantics?
\r
71 It's all deeply unsatisfactory.
\r
73 - Improve performance of Hoopl.
\r
75 A nofib comparison of -fasm vs -fnewcodegen nofib compilation parameters
\r
76 (using the same ghc-cmm branch +libraries compiled by the old codegenerator)
\r
77 is at http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghchoopl.txt
\r
78 - the code produced is 10.9% slower, the compilation is +118% slower!
\r
80 The same comparison with ghc-head with zip representation is at
\r
81 http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghczip.txt
\r
82 - the code produced is 11.7% slower, the compilation is +78% slower.
\r
84 When compiling nofib, ghc-cmm + libraries compiled with -fnew-codegen
\r
85 is 23.7% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.hooplghcoldgen.txt).
\r
86 When compiling nofib, ghc-head + libraries compiled with -fnew-codegen
\r
87 is 31.4% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.zipghcoldgen.txt).
\r
89 So we generate a bit better code, but it takes us longer!
\r
91 EZY: Also importantly, Hoopl uses dramatically more memory than the
\r
94 - Are all blockToNodeList and blockOfNodeList really needed? Maybe we could
\r
95 splice blocks instead?
\r
97 In the CmmContFlowOpt.blockConcat, using Dataflow seems too clumsy. Still,
\r
98 a block catenation function would be probably nicer than blockToNodeList
\r
99 / blockOfNodeList combo.
\r
101 - lowerSafeForeignCall seems too lowlevel. Just use Dataflow. After that
\r
102 delete splitEntrySeq from HooplUtils.
\r
104 - manifestSP seems to touch a lot of the graph representation. It is
\r
105 also slow for CmmSwitch nodes O(block_nodes * switch_statements).
\r
106 Maybe rewrite manifestSP to use Dataflow?
\r
108 - Sort out Label, LabelMap, LabelSet versus BlockId, BlockEnv, BlockSet
\r
109 dichotomy. Mostly this means global replace, but we also need to make
\r
110 Label an instance of Outputable (probably in the Outputable module).
\r
112 - NB that CmmProcPoint line 283 has a hack that works around a GADT-related
\r
115 - SDM (2010-02-26) can we remove the Foreign constructor from Convention?
\r
116 Reason: we never generate code for a function with the Foreign
\r
117 calling convention, and the code for calling foreign calls is generated
\r
119 - AsmCodeGen has a generic Cmm optimiser; move this into new pipeline
\r
120 EZY (2011-04-16): The mini-inliner has been generalized and ported,
\r
121 but the constant folding and other optimizations need to still be
\r
124 - AsmCodeGen has post-native-cg branch eliminator (shortCutBranches);
\r
125 we ultimately want to share this with the Cmm branch eliminator.
\r
127 - At the moment, references to global registers like Hp are "lowered"
\r
128 late (in CgUtils.fixStgRegisters). We should do this early, in the
\r
129 new native codegen, much in the way that we lower calling conventions.
\r
130 Might need to be a bit sophisticated about aliasing.
\r
132 - Question: currently we lift procpoints to become separate
\r
133 CmmProcs. Do we still want to do this?
\r
135 NB: and advantage of continuing to do this is that
\r
136 we can do common-proc elimination!
\r
138 - Move to new Cmm rep:
\r
139 * Make native CG consume New Cmm;
\r
140 * Convert Old Cmm->New Cmm to keep old path alive
\r
141 * Produce New Cmm when reading in .cmm files
\r
143 - Consider module names
\r
145 - Top-level SRT threading is a bit ugly
\r
147 - Add type/newtype for CmmModule = [CmmGroup] -- A module
\r
148 CmmGroup = [CmmTop] -- A .o file
\r
149 CmmTop = Proc | Data -- A procedure or data
\r
151 - This is a *change*: currently a CmmGroup is one function's-worth of code
\r
152 regardless of SplitObjs. Question: can we *always* generate M.o if there
\r
153 is just one element in the list (rather than M/M1.o, M/M2.o etc)
\r
157 - See "CAFs" below; we want to totally refactor the way SRTs are calculated
\r
159 - Pull out Areas into its own module
\r
160 Parameterise AreaMap (note there are type synonyms in CmmStackLayout!)
\r
161 Add ByteWidth = Int
\r
162 type SubArea = (Area, ByteOff, ByteWidth)
\r
163 ByteOff should not be defined in SMRep -- that is too high up the hierarchy
\r
165 - SMRep should not be imported by any module in cmm/! Make it so.
\r
166 -- ByteOff etc ==> CmmExpr
\r
167 -- rET_SMALL etc ==> CmmInfo
\r
168 Check that there are no other imports from codeGen in cmm/
\r
170 - If you eliminate a label by branch chain elimination,
\r
171 what happens if there's an Area associated with that label?
\r
173 - Think about a non-flattened representation?
\r
176 * Use record fields for LastCall!
\r
177 * cml_ret_off should be a ByteOff
\r
179 LastCall (which has a successor) and
\r
180 LastJump (which does not, includes return?)
\r
181 - does not have cml_cont, cml_ret_args, cml_ret_off
\r
184 - expands into save/MidForeignCall/restore/goto
\r
185 - like any LastCall, target of the call gets an info table
\r
187 - JD: remind self of what goes wrong if you turn off the
\r
188 liveness of the update frame
\r
190 - Garbage-collect http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/CPS
\r
191 moving good stuff into
\r
192 http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/NewCodeGenPipeline
\r
195 - We believe that all of CmmProcPoint.addProcPointProtocols is dead. What
\r
196 goes wrong if we simply never call it?
\r
198 - Something fishy in CmmStackLayout.hs
\r
199 * In particular, 'getAreaSize' returns an AreaMap, but we *know* the width of
\r
200 LocalRegs, so it'd be better to return FiniteMap AreaId ByteWidth
\r
201 * setSuccSPs looks fishy. Rather than lookin in procPoints, it could
\r
202 just lookup the block in areaSize which, after all, has a binding
\r
203 for precisely successors of calls. All other blocks (including proc
\r
204 points that are not successors of a call, we think) can be treated
\r
205 uniformly: zero-size Area, and use inSP.
\r
208 - Currently AsmCodeGen top level calls AsmCodeGen.cmmToCmm, which is a small
\r
209 C-- optimiser. It has quite a lot of boilerplate folding code in AsmCodeGen
\r
210 (cmmBlockConFold, cmmStmtConFold, cmmExprConFold), before calling out to
\r
211 CmmOpt. ToDo: see what optimisations are being done; and do them before
\r
214 - Modularise the CPS pipeline; instead of ...; A;B;C; ...
\r
217 - Most of HscMain.tryNewCodeGen does not belong in HscMain. Instead
\r
220 processCmm [including generating "raw" cmm]
\r
226 - If we stick CAF and stack liveness info on a LastCall node (not LastRet/Jump)
\r
227 then all CAF and stack liveness stuff be completed before we split
\r
228 into separate C procedures.
\r
231 compute and attach liveness into to LastCall
\r
232 right at end, split, cvt to old rep
\r
233 [must split before cvt, because old rep is not expressive enough]
\r
236 when old rep disappears,
\r
237 move the whole splitting game into the C back end *only*
\r
238 (guided by the procpoint set)
\r
240 ----------------------------------------------------
\r
242 ----------------------------------------------------
\r
244 -------- Testing stuff ------------
\r
245 HscMain.optionallyConvertAndOrCPS
\r
247 DynFlags: -fconvert-to-zipper-and-back, -frun-cpsz
\r
249 -------- Moribund stuff ------------
\r
250 OldCmm.hs Definition of flowgraph of old representation
\r
251 OldCmmUtil.hs Utilites that operates mostly on on CmmStmt
\r
252 OldPprCmm.hs Pretty print for CmmStmt, GenBasicBlock and ListGraph
\r
253 CmmCvt.hs Conversion between old and new Cmm reps
\r
254 CmmOpt.hs Hopefully-redundant optimiser
\r
256 -------- Stuff to keep ------------
\r
257 CmmCPS.hs Driver for new pipeline
\r
259 CmmLive.hs Liveness analysis, dead code elim
\r
260 CmmProcPoint.hs Identifying and splitting out proc-points
\r
262 CmmSpillReload.hs Save and restore across calls
\r
264 CmmCommonBlockElim.hs Common block elim
\r
265 CmmContFlowOpt.hs Other optimisations (branch-chain, merging)
\r
267 CmmBuildInfoTables.hs New info-table
\r
268 CmmStackLayout.hs and stack layout
\r
270 CmmInfo.hs Defn of InfoTables, and conversion to exact byte layout
\r
272 ---------- Cmm data types --------------
\r
273 Cmm.hs Cmm instantiations of dataflow graph framework
\r
274 MkGraph.hs Interface for building Cmm for codeGen/Stg*.hs modules
\r
276 CmmDecl.hs Shared Cmm types of both representations
\r
277 CmmExpr.hs Type of Cmm expression
\r
278 CmmType.hs Type of Cmm types and their widths
\r
279 CmmMachOp.hs MachOp type and accompanying utilities
\r
284 PprC.hs Pretty print Cmm in C syntax
\r
285 PprCmm.hs Pretty printer for CmmGraph.
\r
286 PprCmmDecl.hs Pretty printer for common Cmm types.
\r
287 PprCmmExpr.hs Pretty printer for Cmm expressions.
\r
290 BlockId.hs BlockId, BlockEnv, BlockSet
\r
292 ----------------------------------------------------
\r
293 Top-level structure
\r
294 ----------------------------------------------------
\r
296 * New codgen called in HscMain.hscGenHardCode, by calling HscMain.tryNewCodeGen,
\r
297 enabled by -fnew-codegen (Opt_TryNewCodeGen)
\r
299 THEN it calls CmmInfo.cmmToRawCmm to lay out the details of info tables
\r
300 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
301 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
303 * HscMain.tryNewCodeGen
\r
304 - STG->Cmm: StgCmm.codeGen (new codegen)
\r
305 - Optimise: CmmContFlowOpt (simple optimisations, very self contained)
\r
306 - Cps convert: CmmCPS.protoCmmCPS
\r
307 - Optimise: CmmContFlowOpt again
\r
308 - Convert: CmmCvt.cmmOfZgraph (convert to old rep) very self contained
\r
310 * StgCmm.hs The new STG -> Cmm conversion code generator
\r
311 Lots of modules StgCmmXXX
\r
314 ----------------------------------------------------
\r
315 CmmCPS.protoCmmCPS The new pipeline
\r
316 ----------------------------------------------------
\r
318 CmmCPS.protoCmmCPS:
\r
319 1. Do cpsTop for each procedures separately
\r
320 2. Build SRT representation; this spans multiple procedures
\r
321 (unless split-objs)
\r
324 * CmmCommonBlockElim.elimCommonBlocks:
\r
325 eliminate common blocks
\r
327 * CmmProcPoint.minimalProcPointSet
\r
328 identify proc-points
\r
331 * CmmProcPoint.addProcPointProtocols
\r
332 something to do with the MA optimisation
\r
333 probably entirely unnecessary
\r
335 * Spill and reload:
\r
336 - CmmSpillReload.dualLivenessWithInsertion
\r
337 insert spills/reloads across
\r
339 Branches to proc-points
\r
340 Now sink those reloads (and other instructions):
\r
341 - CmmSpillReload.rewriteAssignments
\r
342 - CmmSpillReload.removeDeadAssignmentsAndReloads
\r
344 * CmmStackLayout.stubSlotsOnDeath
\r
345 debug only: zero out dead slots when they die
\r
348 - CmmStackLayout.lifeSlotAnal:
\r
349 find which sub-areas are live on entry to each block
\r
351 - CmmStackLayout.layout
\r
352 Lay out the stack, returning an AreaMap
\r
353 type AreaMap = FiniteMap Area ByteOff
\r
354 -- Byte offset of the oldest byte of the Area,
\r
355 -- relative to the oldest byte of the Old Area
\r
357 - CmmStackLayout.manifestSP
\r
358 Manifest the stack pointer
\r
360 * Split into separate procedures
\r
361 - CmmProcPoint.procPointAnalysis
\r
362 Given set of proc points, which blocks are reachable from each
\r
363 Claim: too few proc-points => code duplication, but program still works??
\r
365 - CmmProcPoint.splitAtProcPoints
\r
366 Using this info, split into separate procedures
\r
368 - CmmBuildInfoTables.setInfoTableStackMap
\r
369 Attach stack maps to each info table
\r
372 ----------------------------------------------------
\r
374 ----------------------------------------------------
\r
376 Consider this program, which has a diamond control flow,
\r
377 with a call on one branch
\r
380 if b then { ... f(x) ...; q=5; goto J }
\r
381 else { ...; q=7; goto J }
\r
384 then the join point J is a "proc-point". So, is 'p' passed to J
\r
385 as a parameter? Or, if 'p' was saved on the stack anyway, perhaps
\r
386 to keep it alive across the call to h(), maybe 'p' gets communicated
\r
387 to J that way. This is an awkward choice. (We think that we currently
\r
388 never pass variables to join points via arguments.)
\r
390 Furthermore, there is *no way* to pass q to J in a register (other
\r
391 than a parameter register).
\r
393 What we want is to do register allocation across the whole caboodle.
\r
394 Then we could drop all the code that deals with the above awkward
\r
395 decisions about spilling variables across proc-points.
\r
397 Note that J doesn't need an info table.
\r
399 What we really want is for each LastCall (not LastJump/Ret)
\r
400 to have an info table. Note that ProcPoints that are not successors
\r
401 of calls don't need an info table.
\r
403 Figuring out proc-points
\r
404 ~~~~~~~~~~~~~~~~~~~~~~~~
\r
405 Proc-points are identified by
\r
406 CmmProcPoint.minimalProcPointSet/extendPPSet Although there isn't
\r
407 that much code, JD thinks that it could be done much more nicely using
\r
408 a dominator analysis, using the Dataflow Engine.
\r
410 ----------------------------------------------------
\r
412 ----------------------------------------------------
\r
414 * The code for a procedure f may refer to either the *closure*
\r
415 or the *entry point* of another top-level procedure g.
\r
416 If f is live, then so is g. f's SRT must include g's closure.
\r
418 * The CLabel for the entry-point/closure reveals whether g is
\r
419 a CAF (or refers to CAFs). See the IdLabel constructor of CLabel.
\r
421 * The CAF-ness of the original top-level defininions is figured out
\r
422 (by TidyPgm) before we generate C--. This CafInfo is only set for
\r
423 top-level Ids; nested bindings stay with MayHaveCafRefs.
\r
425 * Currently an SRT contains (only) pointers to (top-level) closures.
\r
427 * Consider this Core code
\r
428 f = \x -> let g = \y -> ...x...y...h1...
\r
430 and suppose that h1, h2 have IdInfo of MayHaveCafRefs.
\r
431 Therefore, so will f, But g will not (since it's nested).
\r
433 This generates C-- roughly like this:
\r
434 f_closure: .word f_entry
\r
435 f_entry() [info-tbl-for-f] { ...jump g_entry...jump h2... }
\r
436 g_entry() [info-tbl-for-g] { ...jump h1... }
\r
438 Note that there is no top-level closure for g (only an info table).
\r
439 This fact (whether or not there is a top-level closure) is recorded
\r
440 in the InfoTable attached to the CmmProc for f, g
\r
442 Any out-of-Group references to an IdLabel goes to
\r
443 a Proc whose InfoTable says "I have a top-level closure".
\r
445 A CmmProc whose InfoTable says "I do not have a top-level
\r
446 closure" is referred to only from its own Group.
\r
448 * So: info-tbl-for-f must have an SRT that keeps h1,h2 alive
\r
449 info-tbl-for-g must have an SRT that keeps h1 (only) alive
\r
451 But if we just look for the free CAF refs, we get:
\r
455 So we need to do a transitive closure thing to flesh out
\r
456 f's keep-alive refs to include h1.
\r
458 * The SRT info is the C_SRT field of Cmm.ClosureTypeInfo in a
\r
459 CmmInfoTable attached to each CmmProc. CmmCPS.toTops actually does
\r
460 the attaching, right at the end of the pipeline. The C_SRT part
\r
461 gives offsets within a single, shared table of closure pointers.
\r
463 * DECIDED: we can generate SRTs based on the final Cmm program
\r
464 without knowledge of how it is generated.
\r
466 ----------------------------------------------------
\r
468 ----------------------------------------------------
\r
470 See Note [Foreign calls] in CmmNode! This explains that a safe
\r
471 foreign call must do this:
\r
473 push info table (on thread stack) to describe frame
\r
474 make call (via C stack)
\r
476 restore thread state
\r
477 and explains why this expansion must be done late in the day.
\r
480 - Every foreign call is represented as a middle node
\r
482 - *Unsafe* foreign calls are simply "fat machine instructions"
\r
483 and are passed along to the native code generator
\r
485 - *Safe* foreign calls are "lowered" to unsafe calls by wrapping
\r
486 them in the above save/restore sequence. This step is done
\r
487 very late in the pipeline, just before handing to the native
\r
490 This lowering is done by BuildInfoTables.lowerSafeForeignCalls
\r
493 NEW PLAN for foreign calls:
\r
494 - Unsafe foreign calls remain as a middle node (fat machine instruction)
\r
495 Even the parameter passing is not lowered (just as machine instrs
\r
498 - Initially, safe foreign calls appear as LastCalls with
\r
501 ----------------------------------------------------
\r
502 Cmm representations
\r
503 ----------------------------------------------------
\r
506 The type [GenCmm d h g] represents a whole module,
\r
507 ** one list element per .o file **
\r
508 Without SplitObjs, the list has exactly one element
\r
510 newtype GenCmm d h g = Cmm [GenCmmTop d h g] -- A whole .o file
\r
511 data GenCmmTop d h g
\r
512 = CmmProc h g -- One procedure, graph d
\r
513 | CmmData <stuff> [d] -- Initialised data, items d
\r
515 Old and new piplines use different representations
\r
516 (CmmCvt.hs converts between the two)
\r
520 OLD BACK END representations (OldCmm.hs):
\r
521 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
523 newtype ListGraph i = ListGraph [GenBasicBlock i]
\r
525 data CmmStmt = Assign | Store | Return etc -- OLD BACK END ONLY
\r
528 Once the info tables are laid out, we replace CmmInfo with [CmmStatic]
\r
529 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
530 which represents the info tables as data, that should
\r
531 immediately precede the code
\r
534 NEW BACK END representations
\r
535 * Uses Hoopl library, a zero-boot package
\r
536 * CmmNode defines a node of a flow graph.
\r
537 * Cmm defines CmmGraph, CmmTop, Cmm
\r
538 - CmmGraph is a closed/closed graph + an entry node.
\r
540 data CmmGraph = CmmGraph { g_entry :: BlockId
\r
541 , g_graph :: Graph CmmNode C C }
\r
543 - CmmTop is a top level chunk, specialization of GenCmmTop from CmmDecl.hs
\r
544 with CmmGraph as a flow graph.
\r
545 - Cmm is a collection of CmmTops.
\r
547 type Cmm = GenCmm CmmStatic CmmTopInfo CmmGraph
\r
548 type CmmTop = GenCmmTop CmmStatic CmmTopInfo CmmGraph
\r
550 - CmmTop uses CmmTopInfo, which is a CmmInfoTable and CmmStackInfo
\r
552 data CmmTopInfo = TopInfo {info_tbl :: CmmInfoTable, stack_info :: CmmStackInfo}
\r
556 data CmmStackInfo = StackInfo {arg_space :: ByteOff, updfr_space :: Maybe ByteOff}
\r
558 * arg_space = SP offset on entry
\r
559 * updfr_space space = SP offset on exit
\r
560 Once the staci is manifested, we could drom CmmStackInfo, ie. get
\r
561 GenCmm CmmStatic CmmInfoTable CmmGraph, but we do not do that currently.
\r
564 * MkGraph.hs: smart constructors for Cmm.hs
\r
565 Beware, the CmmAGraph defined here does not use AGraph from Hoopl,
\r
566 as CmmAGraph can be opened or closed at exit, See the notes in that module.
\r
570 CmmDecl.hs - GenCmm and GenCmmTop types
\r
571 CmmExpr.hs - defines the Cmm expression types
\r
572 - CmmExpr, CmmReg, CmmLit, LocalReg, GlobalReg
\r
573 - Area, AreaId etc (separate module?)
\r
574 CmmType.hs - CmmType, Width etc (saparate module?)
\r
575 CmmMachOp.hs - MachOp and CallishMachOp types
\r
577 BlockId.hs defines BlockId, BlockEnv, BlockSet
\r