3 In CmmNode, consider spliting CmmCall into two: call and jump
\r
5 Notes on new codegen (Aug 10)
\r
6 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\r
9 - We insert spills for variables before the stack check! This is the reason for
\r
10 some fishy code in StgCmmHeap.entryHeapCheck where we are doing some strange
\r
11 things to fix up the stack pointer before GC calls/jumps.
\r
13 The reason spills are inserted before the sp check is that at the entry to a
\r
14 function we always store the parameters passed in registers to local variables.
\r
15 The spill pass simply inserts spills at variable definitions. We instead should
\r
16 sink the spills so that we can avoid spilling them on branches that never
\r
19 This will fix the spill before stack check problem but only really as a side
\r
20 effect. A 'real fix' probably requires making the spiller know about sp checks.
\r
22 EZY: I don't understand this comment. David Terei, can you clarify?
\r
24 - Proc points pass all arguments on the stack, adding more code and
\r
25 slowing down things a lot. We either need to fix this or even better
\r
26 would be to get rid of proc points.
\r
28 - CmmInfo.cmmToRawCmm uses Old.Cmm, so it is called after converting Cmm.Cmm to
\r
29 Old.Cmm. We should abstract it to work on both representations, it needs only to
\r
30 convert a CmmInfoTable to [CmmStatic].
\r
32 - The MkGraph currenty uses a different semantics for <*> than Hoopl. Maybe
\r
33 we could convert codeGen/StgCmm* clients to the Hoopl's semantics?
\r
34 It's all deeply unsatisfactory.
\r
36 - Improve performance of Hoopl.
\r
38 A nofib comparison of -fasm vs -fnewcodegen nofib compilation parameters
\r
39 (using the same ghc-cmm branch +libraries compiled by the old codegenerator)
\r
40 is at http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghchoopl.txt
\r
41 - the code produced is 10.9% slower, the compilation is +118% slower!
\r
43 The same comparison with ghc-head with zip representation is at
\r
44 http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghczip.txt
\r
45 - the code produced is 11.7% slower, the compilation is +78% slower.
\r
47 When compiling nofib, ghc-cmm + libraries compiled with -fnew-codegen
\r
48 is 23.7% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.hooplghcoldgen.txt).
\r
49 When compiling nofib, ghc-head + libraries compiled with -fnew-codegen
\r
50 is 31.4% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.zipghcoldgen.txt).
\r
52 So we generate a bit better code, but it takes us longer!
\r
54 EZY: Also importantly, Hoopl uses dramatically more memory than the
\r
57 - Are all blockToNodeList and blockOfNodeList really needed? Maybe we could
\r
58 splice blocks instead?
\r
60 In the CmmContFlowOpt.blockConcat, using Dataflow seems too clumsy. Still,
\r
61 a block catenation function would be probably nicer than blockToNodeList
\r
62 / blockOfNodeList combo.
\r
64 - lowerSafeForeignCall seems too lowlevel. Just use Dataflow. After that
\r
65 delete splitEntrySeq from HooplUtils.
\r
67 - manifestSP seems to touch a lot of the graph representation. It is
\r
68 also slow for CmmSwitch nodes O(block_nodes * switch_statements).
\r
69 Maybe rewrite manifestSP to use Dataflow?
\r
71 - Sort out Label, LabelMap, LabelSet versus BlockId, BlockEnv, BlockSet
\r
72 dichotomy. Mostly this means global replace, but we also need to make
\r
73 Label an instance of Outputable (probably in the Outputable module).
\r
75 - NB that CmmProcPoint line 283 has a hack that works around a GADT-related
\r
78 - SDM (2010-02-26) can we remove the Foreign constructor from Convention?
\r
79 Reason: we never generate code for a function with the Foreign
\r
80 calling convention, and the code for calling foreign calls is generated
\r
82 - AsmCodeGen has a generic Cmm optimiser; move this into new pipeline
\r
83 EZY (2011-04-16): The mini-inliner has been generalized and ported,
\r
84 but the constant folding and other optimizations need to still be
\r
87 - AsmCodeGen has post-native-cg branch eliminator (shortCutBranches);
\r
88 we ultimately want to share this with the Cmm branch eliminator.
\r
90 - At the moment, references to global registers like Hp are "lowered"
\r
91 late (in CgUtils.fixStgRegisters). We should do this early, in the
\r
92 new native codegen, much in the way that we lower calling conventions.
\r
93 Might need to be a bit sophisticated about aliasing.
\r
95 - Question: currently we lift procpoints to become separate
\r
96 CmmProcs. Do we still want to do this?
\r
98 NB: and advantage of continuing to do this is that
\r
99 we can do common-proc elimination!
\r
101 - Move to new Cmm rep:
\r
102 * Make native CG consume New Cmm;
\r
103 * Convert Old Cmm->New Cmm to keep old path alive
\r
104 * Produce New Cmm when reading in .cmm files
\r
106 - Consider module names
\r
108 - Top-level SRT threading is a bit ugly
\r
110 - Add type/newtype for CmmModule = [CmmGroup] -- A module
\r
111 CmmGroup = [CmmTop] -- A .o file
\r
112 CmmTop = Proc | Data -- A procedure or data
\r
114 - This is a *change*: currently a CmmGroup is one function's-worth of code
\r
115 regardless of SplitObjs. Question: can we *always* generate M.o if there
\r
116 is just one element in the list (rather than M/M1.o, M/M2.o etc)
\r
120 - See "CAFs" below; we want to totally refactor the way SRTs are calculated
\r
122 - Pull out Areas into its own module
\r
123 Parameterise AreaMap (note there are type synonyms in CmmStackLayout!)
\r
124 Add ByteWidth = Int
\r
125 type SubArea = (Area, ByteOff, ByteWidth)
\r
126 ByteOff should not be defined in SMRep -- that is too high up the hierarchy
\r
128 - SMRep should not be imported by any module in cmm/! Make it so.
\r
129 -- ByteOff etc ==> CmmExpr
\r
130 -- rET_SMALL etc ==> CmmInfo
\r
131 Check that there are no other imports from codeGen in cmm/
\r
133 - If you eliminate a label by branch chain elimination,
\r
134 what happens if there's an Area associated with that label?
\r
136 - Think about a non-flattened representation?
\r
139 * Use record fields for LastCall!
\r
140 * cml_ret_off should be a ByteOff
\r
142 LastCall (which has a successor) and
\r
143 LastJump (which does not, includes return?)
\r
144 - does not have cml_cont, cml_ret_args, cml_ret_off
\r
147 - expands into save/MidForeignCall/restore/goto
\r
148 - like any LastCall, target of the call gets an info table
\r
150 - JD: remind self of what goes wrong if you turn off the
\r
151 liveness of the update frame
\r
153 - Garbage-collect http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/CPS
\r
154 moving good stuff into
\r
155 http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/NewCodeGenPipeline
\r
158 - We believe that all of CmmProcPoint.addProcPointProtocols is dead. What
\r
159 goes wrong if we simply never call it?
\r
161 - Something fishy in CmmStackLayout.hs
\r
162 * In particular, 'getAreaSize' returns an AreaMap, but we *know* the width of
\r
163 LocalRegs, so it'd be better to return FiniteMap AreaId ByteWidth
\r
164 * setSuccSPs looks fishy. Rather than lookin in procPoints, it could
\r
165 just lookup the block in areaSize which, after all, has a binding
\r
166 for precisely successors of calls. All other blocks (including proc
\r
167 points that are not successors of a call, we think) can be treated
\r
168 uniformly: zero-size Area, and use inSP.
\r
171 - Currently AsmCodeGen top level calls AsmCodeGen.cmmToCmm, which is a small
\r
172 C-- optimiser. It has quite a lot of boilerplate folding code in AsmCodeGen
\r
173 (cmmBlockConFold, cmmStmtConFold, cmmExprConFold), before calling out to
\r
174 CmmOpt. ToDo: see what optimisations are being done; and do them before
\r
177 - Modularise the CPS pipeline; instead of ...; A;B;C; ...
\r
180 - Most of HscMain.tryNewCodeGen does not belong in HscMain. Instead
\r
183 processCmm [including generating "raw" cmm]
\r
189 - If we stick CAF and stack liveness info on a LastCall node (not LastRet/Jump)
\r
190 then all CAF and stack liveness stuff be completed before we split
\r
191 into separate C procedures.
\r
194 compute and attach liveness into to LastCall
\r
195 right at end, split, cvt to old rep
\r
196 [must split before cvt, because old rep is not expressive enough]
\r
199 when old rep disappears,
\r
200 move the whole splitting game into the C back end *only*
\r
201 (guided by the procpoint set)
\r
203 ----------------------------------------------------
\r
205 ----------------------------------------------------
\r
207 -------- Testing stuff ------------
\r
208 HscMain.optionallyConvertAndOrCPS
\r
210 DynFlags: -fconvert-to-zipper-and-back, -frun-cpsz
\r
212 -------- Moribund stuff ------------
\r
213 OldCmm.hs Definition of flowgraph of old representation
\r
214 OldCmmUtil.hs Utilites that operates mostly on on CmmStmt
\r
215 OldPprCmm.hs Pretty print for CmmStmt, GenBasicBlock and ListGraph
\r
216 CmmCvt.hs Conversion between old and new Cmm reps
\r
217 CmmOpt.hs Hopefully-redundant optimiser
\r
219 -------- Stuff to keep ------------
\r
220 CmmCPS.hs Driver for new pipeline
\r
222 CmmLive.hs Liveness analysis, dead code elim
\r
223 CmmProcPoint.hs Identifying and splitting out proc-points
\r
225 CmmSpillReload.hs Save and restore across calls
\r
227 CmmCommonBlockElim.hs Common block elim
\r
228 CmmContFlowOpt.hs Other optimisations (branch-chain, merging)
\r
230 CmmBuildInfoTables.hs New info-table
\r
231 CmmStackLayout.hs and stack layout
\r
233 CmmInfo.hs Defn of InfoTables, and conversion to exact byte layout
\r
235 ---------- Cmm data types --------------
\r
236 Cmm.hs Cmm instantiations of dataflow graph framework
\r
237 MkGraph.hs Interface for building Cmm for codeGen/Stg*.hs modules
\r
239 CmmDecl.hs Shared Cmm types of both representations
\r
240 CmmExpr.hs Type of Cmm expression
\r
241 CmmType.hs Type of Cmm types and their widths
\r
242 CmmMachOp.hs MachOp type and accompanying utilities
\r
247 PprC.hs Pretty print Cmm in C syntax
\r
248 PprCmm.hs Pretty printer for CmmGraph.
\r
249 PprCmmDecl.hs Pretty printer for common Cmm types.
\r
250 PprCmmExpr.hs Pretty printer for Cmm expressions.
\r
253 BlockId.hs BlockId, BlockEnv, BlockSet
\r
255 ----------------------------------------------------
\r
256 Top-level structure
\r
257 ----------------------------------------------------
\r
259 * New codgen called in HscMain.hscGenHardCode, by calling HscMain.tryNewCodeGen,
\r
260 enabled by -fnew-codegen (Opt_TryNewCodeGen)
\r
262 THEN it calls CmmInfo.cmmToRawCmm to lay out the details of info tables
\r
263 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
264 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
266 * HscMain.tryNewCodeGen
\r
267 - STG->Cmm: StgCmm.codeGen (new codegen)
\r
268 - Optimise: CmmContFlowOpt (simple optimisations, very self contained)
\r
269 - Cps convert: CmmCPS.protoCmmCPS
\r
270 - Optimise: CmmContFlowOpt again
\r
271 - Convert: CmmCvt.cmmOfZgraph (convert to old rep) very self contained
\r
273 * StgCmm.hs The new STG -> Cmm conversion code generator
\r
274 Lots of modules StgCmmXXX
\r
277 ----------------------------------------------------
\r
278 CmmCPS.protoCmmCPS The new pipeline
\r
279 ----------------------------------------------------
\r
281 CmmCPS.protoCmmCPS:
\r
282 1. Do cpsTop for each procedures separately
\r
283 2. Build SRT representation; this spans multiple procedures
\r
284 (unless split-objs)
\r
287 * CmmCommonBlockElim.elimCommonBlocks:
\r
288 eliminate common blocks
\r
290 * CmmProcPoint.minimalProcPointSet
\r
291 identify proc-points
\r
294 * CmmProcPoint.addProcPointProtocols
\r
295 something to do with the MA optimisation
\r
296 probably entirely unnecessary
\r
298 * Spill and reload:
\r
299 - CmmSpillReload.dualLivenessWithInsertion
\r
300 insert spills/reloads across
\r
302 Branches to proc-points
\r
303 Now sink those reloads (and other instructions):
\r
304 - CmmSpillReload.rewriteAssignments
\r
305 - CmmSpillReload.removeDeadAssignmentsAndReloads
\r
307 * CmmStackLayout.stubSlotsOnDeath
\r
308 debug only: zero out dead slots when they die
\r
311 - CmmStackLayout.lifeSlotAnal:
\r
312 find which sub-areas are live on entry to each block
\r
314 - CmmStackLayout.layout
\r
315 Lay out the stack, returning an AreaMap
\r
316 type AreaMap = FiniteMap Area ByteOff
\r
317 -- Byte offset of the oldest byte of the Area,
\r
318 -- relative to the oldest byte of the Old Area
\r
320 - CmmStackLayout.manifestSP
\r
321 Manifest the stack pointer
\r
323 * Split into separate procedures
\r
324 - CmmProcPoint.procPointAnalysis
\r
325 Given set of proc points, which blocks are reachable from each
\r
326 Claim: too few proc-points => code duplication, but program still works??
\r
328 - CmmProcPoint.splitAtProcPoints
\r
329 Using this info, split into separate procedures
\r
331 - CmmBuildInfoTables.setInfoTableStackMap
\r
332 Attach stack maps to each info table
\r
335 ----------------------------------------------------
\r
337 ----------------------------------------------------
\r
339 Consider this program, which has a diamond control flow,
\r
340 with a call on one branch
\r
343 if b then { ... f(x) ...; q=5; goto J }
\r
344 else { ...; q=7; goto J }
\r
347 then the join point J is a "proc-point". So, is 'p' passed to J
\r
348 as a parameter? Or, if 'p' was saved on the stack anyway, perhaps
\r
349 to keep it alive across the call to h(), maybe 'p' gets communicated
\r
350 to J that way. This is an awkward choice. (We think that we currently
\r
351 never pass variables to join points via arguments.)
\r
353 Furthermore, there is *no way* to pass q to J in a register (other
\r
354 than a parameter register).
\r
356 What we want is to do register allocation across the whole caboodle.
\r
357 Then we could drop all the code that deals with the above awkward
\r
358 decisions about spilling variables across proc-points.
\r
360 Note that J doesn't need an info table.
\r
362 What we really want is for each LastCall (not LastJump/Ret)
\r
363 to have an info table. Note that ProcPoints that are not successors
\r
364 of calls don't need an info table.
\r
366 Figuring out proc-points
\r
367 ~~~~~~~~~~~~~~~~~~~~~~~~
\r
368 Proc-points are identified by
\r
369 CmmProcPoint.minimalProcPointSet/extendPPSet Although there isn't
\r
370 that much code, JD thinks that it could be done much more nicely using
\r
371 a dominator analysis, using the Dataflow Engine.
\r
373 ----------------------------------------------------
\r
375 ----------------------------------------------------
\r
377 * The code for a procedure f may refer to either the *closure*
\r
378 or the *entry point* of another top-level procedure g.
\r
379 If f is live, then so is g. f's SRT must include g's closure.
\r
381 * The CLabel for the entry-point/closure reveals whether g is
\r
382 a CAF (or refers to CAFs). See the IdLabel constructor of CLabel.
\r
384 * The CAF-ness of the original top-level defininions is figured out
\r
385 (by TidyPgm) before we generate C--. This CafInfo is only set for
\r
386 top-level Ids; nested bindings stay with MayHaveCafRefs.
\r
388 * Currently an SRT contains (only) pointers to (top-level) closures.
\r
390 * Consider this Core code
\r
391 f = \x -> let g = \y -> ...x...y...h1...
\r
393 and suppose that h1, h2 have IdInfo of MayHaveCafRefs.
\r
394 Therefore, so will f, But g will not (since it's nested).
\r
396 This generates C-- roughly like this:
\r
397 f_closure: .word f_entry
\r
398 f_entry() [info-tbl-for-f] { ...jump g_entry...jump h2... }
\r
399 g_entry() [info-tbl-for-g] { ...jump h1... }
\r
401 Note that there is no top-level closure for g (only an info table).
\r
402 This fact (whether or not there is a top-level closure) is recorded
\r
403 in the InfoTable attached to the CmmProc for f, g
\r
405 Any out-of-Group references to an IdLabel goes to
\r
406 a Proc whose InfoTable says "I have a top-level closure".
\r
408 A CmmProc whose InfoTable says "I do not have a top-level
\r
409 closure" is referred to only from its own Group.
\r
411 * So: info-tbl-for-f must have an SRT that keeps h1,h2 alive
\r
412 info-tbl-for-g must have an SRT that keeps h1 (only) alive
\r
414 But if we just look for the free CAF refs, we get:
\r
418 So we need to do a transitive closure thing to flesh out
\r
419 f's keep-alive refs to include h1.
\r
421 * The SRT info is the C_SRT field of Cmm.ClosureTypeInfo in a
\r
422 CmmInfoTable attached to each CmmProc. CmmCPS.toTops actually does
\r
423 the attaching, right at the end of the pipeline. The C_SRT part
\r
424 gives offsets within a single, shared table of closure pointers.
\r
426 * DECIDED: we can generate SRTs based on the final Cmm program
\r
427 without knowledge of how it is generated.
\r
429 ----------------------------------------------------
\r
431 ----------------------------------------------------
\r
433 See Note [Foreign calls] in CmmNode! This explains that a safe
\r
434 foreign call must do this:
\r
436 push info table (on thread stack) to describe frame
\r
437 make call (via C stack)
\r
439 restore thread state
\r
440 and explains why this expansion must be done late in the day.
\r
443 - Every foreign call is represented as a middle node
\r
445 - *Unsafe* foreign calls are simply "fat machine instructions"
\r
446 and are passed along to the native code generator
\r
448 - *Safe* foreign calls are "lowered" to unsafe calls by wrapping
\r
449 them in the above save/restore sequence. This step is done
\r
450 very late in the pipeline, just before handing to the native
\r
453 This lowering is done by BuildInfoTables.lowerSafeForeignCalls
\r
456 NEW PLAN for foreign calls:
\r
457 - Unsafe foreign calls remain as a middle node (fat machine instruction)
\r
458 Even the parameter passing is not lowered (just as machine instrs
\r
461 - Initially, safe foreign calls appear as LastCalls with
\r
464 ----------------------------------------------------
\r
465 Cmm representations
\r
466 ----------------------------------------------------
\r
469 The type [GenCmm d h g] represents a whole module,
\r
470 ** one list element per .o file **
\r
471 Without SplitObjs, the list has exactly one element
\r
473 newtype GenCmm d h g = Cmm [GenCmmTop d h g] -- A whole .o file
\r
474 data GenCmmTop d h g
\r
475 = CmmProc h g -- One procedure, graph d
\r
476 | CmmData <stuff> [d] -- Initialised data, items d
\r
478 Old and new piplines use different representations
\r
479 (CmmCvt.hs converts between the two)
\r
483 OLD BACK END representations (OldCmm.hs):
\r
484 type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
\r
486 newtype ListGraph i = ListGraph [GenBasicBlock i]
\r
488 data CmmStmt = Assign | Store | Return etc -- OLD BACK END ONLY
\r
491 Once the info tables are laid out, we replace CmmInfo with [CmmStatic]
\r
492 type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
\r
493 which represents the info tables as data, that should
\r
494 immediately precede the code
\r
497 NEW BACK END representations
\r
498 * Uses Hoopl library, a zero-boot package
\r
499 * CmmNode defines a node of a flow graph.
\r
500 * Cmm defines CmmGraph, CmmTop, Cmm
\r
501 - CmmGraph is a closed/closed graph + an entry node.
\r
503 data CmmGraph = CmmGraph { g_entry :: BlockId
\r
504 , g_graph :: Graph CmmNode C C }
\r
506 - CmmTop is a top level chunk, specialization of GenCmmTop from CmmDecl.hs
\r
507 with CmmGraph as a flow graph.
\r
508 - Cmm is a collection of CmmTops.
\r
510 type Cmm = GenCmm CmmStatic CmmTopInfo CmmGraph
\r
511 type CmmTop = GenCmmTop CmmStatic CmmTopInfo CmmGraph
\r
513 - CmmTop uses CmmTopInfo, which is a CmmInfoTable and CmmStackInfo
\r
515 data CmmTopInfo = TopInfo {info_tbl :: CmmInfoTable, stack_info :: CmmStackInfo}
\r
519 data CmmStackInfo = StackInfo {arg_space :: ByteOff, updfr_space :: Maybe ByteOff}
\r
521 * arg_space = SP offset on entry
\r
522 * updfr_space space = SP offset on exit
\r
523 Once the staci is manifested, we could drom CmmStackInfo, ie. get
\r
524 GenCmm CmmStatic CmmInfoTable CmmGraph, but we do not do that currently.
\r
527 * MkGraph.hs: smart constructors for Cmm.hs
\r
528 Beware, the CmmAGraph defined here does not use AGraph from Hoopl,
\r
529 as CmmAGraph can be opened or closed at exit, See the notes in that module.
\r
533 CmmDecl.hs - GenCmm and GenCmmTop types
\r
534 CmmExpr.hs - defines the Cmm expression types
\r
535 - CmmExpr, CmmReg, CmmLit, LocalReg, GlobalReg
\r
536 - Area, AreaId etc (separate module?)
\r
537 CmmType.hs - CmmType, Width etc (saparate module?)
\r
538 CmmMachOp.hs - MachOp and CallishMachOp types
\r
540 BlockId.hs defines BlockId, BlockEnv, BlockSet
\r