Commit 2a662bdb authored by Simon Marlow's avatar Simon Marlow
Browse files

do a pass over this file, remove obsolete stuff

parent 01386d38
......@@ -24,15 +24,6 @@ More notes (June 11)
* Check in ClosureInfo:
-- NB: Results here should line up with the results of SMRep.rtsClosureType
* "Remove redundant reloads" in CmmSpillReload should be redundant; since
insertLateReloads is now gone, every reload is reloading a live variable.
Test and nuke.
* When stubbing dead slots be careful not to write into an area that
overlaps with an area that's in use. So stubbing needs to *follow*
stack layout.
More notes (May 11)
~~~~~~~~~~~~~~~~~~~
In CmmNode, consider spliting CmmCall into two: call and jump
......@@ -41,67 +32,16 @@ Notes on new codegen (Aug 10)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Things to do:
- We insert spills for variables before the stack check! This is the reason for
some fishy code in StgCmmHeap.entryHeapCheck where we are doing some strange
things to fix up the stack pointer before GC calls/jumps.
The reason spills are inserted before the sp check is that at the entry to a
function we always store the parameters passed in registers to local variables.
The spill pass simply inserts spills at variable definitions. We instead should
sink the spills so that we can avoid spilling them on branches that never
reload them.
This will fix the spill before stack check problem but only really as a side
effect. A 'real fix' probably requires making the spiller know about sp checks.
EZY: I don't understand this comment. David Terei, can you clarify?
- Proc points pass all arguments on the stack, adding more code and
slowing down things a lot. We either need to fix this or even better
would be to get rid of proc points.
- CmmInfo.cmmToRawCmm uses Old.Cmm, so it is called after converting Cmm.Cmm to
Old.Cmm. We should abstract it to work on both representations, it needs only to
convert a CmmInfoTable to [CmmStatic].
- Improve performance of Hoopl.
A nofib comparison of -fasm vs -fnewcodegen nofib compilation parameters
(using the same ghc-cmm branch +libraries compiled by the old codegenerator)
is at http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghchoopl.txt
- the code produced is 10.9% slower, the compilation is +118% slower!
The same comparison with ghc-head with zip representation is at
http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.oldghczip.txt
- the code produced is 11.7% slower, the compilation is +78% slower.
When compiling nofib, ghc-cmm + libraries compiled with -fnew-codegen
is 23.7% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.hooplghcoldgen.txt).
When compiling nofib, ghc-head + libraries compiled with -fnew-codegen
is 31.4% slower (http://fox.auryn.cz/msrc/0517_hoopl/32bit.oldghcoldgen.zipghcoldgen.txt).
EZY: Also importantly, Hoopl uses dramatically more memory than the
old code generator.
- Are all blockToNodeList and blockOfNodeList really needed? Maybe we could
splice blocks instead?
- lowerSafeForeignCall seems too lowlevel. Just use Dataflow. After that
delete splitEntrySeq from HooplUtils.
- manifestSP seems to touch a lot of the graph representation. It is
also slow for CmmSwitch nodes O(block_nodes * switch_statements).
Maybe rewrite manifestSP to use Dataflow?
- Sort out Label, LabelMap, LabelSet versus BlockId, BlockEnv, BlockSet
dichotomy. Mostly this means global replace, but we also need to make
Label an instance of Outputable (probably in the Outputable module).
EZY: We should use Label, since that's the terminology Hoopl uses.
- NB that CmmProcPoint line 283 has a hack that works around a GADT-related
bug in 6.10.
- AsmCodeGen has a generic Cmm optimiser; move this into new pipeline
EZY (2011-04-16): The mini-inliner has been generalized and ported,
but the constant folding and other optimizations need to still be
......@@ -115,90 +55,25 @@ Things to do:
new native codegen, much in the way that we lower calling conventions.
Might need to be a bit sophisticated about aliasing.
- Question: currently we lift procpoints to become separate
CmmProcs. Do we still want to do this?
NB: and advantage of continuing to do this is that
we can do common-proc elimination!
- Move to new Cmm rep:
* Make native CG consume New Cmm;
* Convert Old Cmm->New Cmm to keep old path alive
* Produce New Cmm when reading in .cmm files
- Consider module names
- Top-level SRT threading is a bit ugly
- See "CAFs" below; we want to totally refactor the way SRTs are calculated
- Pull out Areas into its own module
Parameterise AreaMap (note there are type synonyms in CmmStackLayout!)
Add ByteWidth = Int
type SubArea = (Area, ByteOff, ByteWidth)
ByteOff should not be defined in SMRep -- that is too high up the hierarchy
- SMRep should not be imported by any module in cmm/! Make it so.
-- ByteOff etc ==> CmmExpr
-- rET_SMALL etc ==> CmmInfo
Check that there are no other imports from codeGen in cmm/
- If you eliminate a label by branch chain elimination,
what happens if there's an Area associated with that label?
- Think about a non-flattened representation?
- LastCall:
* Use record fields for LastCall!
* cml_ret_off should be a ByteOff
* Split into
LastCall (which has a successor) and
LastJump (which does not, includes return?)
- does not have cml_cont, cml_ret_args, cml_ret_off
LastForeignCall
- safe!
- expands into save/MidForeignCall/restore/goto
- like any LastCall, target of the call gets an info table
- JD: remind self of what goes wrong if you turn off the
liveness of the update frame
- Garbage-collect http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/CPS
moving good stuff into
http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/NewCodeGenPipeline
- We believe that all of CmmProcPoint.addProcPointProtocols is dead. What
goes wrong if we simply never call it?
- Something fishy in CmmStackLayout.hs
* In particular, 'getAreaSize' returns an AreaMap, but we *know* the width of
LocalRegs, so it'd be better to return FiniteMap AreaId ByteWidth
* setSuccSPs looks fishy. Rather than lookin in procPoints, it could
just lookup the block in areaSize which, after all, has a binding
for precisely successors of calls. All other blocks (including proc
points that are not successors of a call, we think) can be treated
uniformly: zero-size Area, and use inSP.
- Currently AsmCodeGen top level calls AsmCodeGen.cmmToCmm, which is a small
C-- optimiser. It has quite a lot of boilerplate folding code in AsmCodeGen
(cmmBlockConFold, cmmStmtConFold, cmmExprConFold), before calling out to
CmmOpt. ToDo: see what optimisations are being done; and do them before
AsmCodeGen.
- Modularise the CPS pipeline; instead of ...; A;B;C; ...
use ..; ABC; ....
- Most of HscMain.tryNewCodeGen does not belong in HscMain. Instead
if new_cg then
StgCmm.codeGen
processCmm [including generating "raw" cmm]
else
CodeGen.codeGen
cmmToRawCmm
- If we stick CAF and stack liveness info on a LastCall node (not LastRet/Jump)
then all CAF and stack liveness stuff be completed before we split
into separate C procedures.
......@@ -271,9 +146,6 @@ ClosureInfo.lhs
Modules in cmm/
----------------------------------------------------
-------- Testing stuff ------------
DynFlags: -frun-cpsz
-------- Moribund stuff ------------
OldCmm.hs Definition of flowgraph of old representation
Imports some data types from (new) Cmm
......@@ -316,93 +188,6 @@ PprC.hs Pretty print Cmm in C syntax
CLabel.hs CLabel
BlockId.hs BlockId, BlockEnv, BlockSet
----------------------------------------------------
Top-level structure
----------------------------------------------------
* New codgen called in HscMain.hscGenHardCode, by calling HscMain.tryNewCodeGen,
enabled by -fnew-codegen (Opt_TryNewCodeGen)
THEN it calls CmmInfo.cmmToRawCmm to lay out the details of info tables
type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
* HscMain.tryNewCodeGen
- STG->Cmm: StgCmm.codeGen (new codegen)
- Optimize and CPS: CmmPipeline.cmmPipeline
- Convert: CmmCvt.cmmOfZgraph (convert to old rep) very self contained
* StgCmm.hs The new STG -> Cmm conversion code generator
Lots of modules StgCmmXXX
----------------------------------------------------
CmmPipeline.cmmPipeline The new pipeline
----------------------------------------------------
CmmPipeline.cmmPipeline:
1. Do control flow optimization
2. Do cpsTop for each procedures separately
3. Build SRT representation; this spans multiple procedures
(unless split-objs)
4. Do control flow optimization on all resulting procedures
cpsTop:
* CmmCommonBlockElim.elimCommonBlocks:
eliminate common blocks
* CmmProcPoint.minimalProcPointSet
identify proc-points
no change to graph
* CmmProcPoint.addProcPointProtocols
something to do with the MA optimisation
probably entirely unnecessary
* Spill and reload:
- CmmSpillReload.dualLivenessWithInsertion
insert spills/reloads across
LastCalls, and
Branches to proc-points
Now sink those reloads (and other instructions):
- CmmSpillReload.rewriteAssignments
- CmmSpillReload.removeDeadAssignmentsAndReloads
* CmmStackLayout.stubSlotsOnDeath
debug only: zero out dead slots when they die
* Stack layout
- CmmStackLayout.lifeSlotAnal:
find which sub-areas are live on entry to each block
- CmmStackLayout.layout
Lay out the stack, returning an AreaMap
type AreaMap = FiniteMap Area ByteOff
-- Byte offset of the oldest byte of the Area,
-- relative to the oldest byte of the Old Area
- CmmStackLayout.manifestSP
Manifest the stack pointer
* Split into separate procedures
- CmmProcPoint.procPointAnalysis
Given set of proc points (computed earlier by
CmmProcPoint.minimalProcPointSet) find which blocks
are reachable from each
Each block should be reachable from *one* proc point, so
the blocks reachable from P are the internal nodes of
the final procedure P
NB: if we the earlier analysis had produced too few proc-points
we should nevertheless be fine by code duplication; but
that is not implemented
- CmmProcPoint.splitAtProcPoints
Using this info, split into separate procedures
- CmmBuildInfoTables.setInfoTableStackMap
Attach stack maps to each info table
----------------------------------------------------
Proc-points
......@@ -498,116 +283,3 @@ a dominator analysis, using the Dataflow Engine.
* DECIDED: we can generate SRTs based on the final Cmm program
without knowledge of how it is generated.
----------------------------------------------------
Foreign calls
----------------------------------------------------
See Note [Foreign calls] in CmmNode! This explains that a safe
foreign call must do this:
save thread state
push info table (on thread stack) to describe frame
make call (via C stack)
pop info table
restore thread state
and explains why this expansion must be done late in the day.
Hence,
- Every foreign call is represented as a middle node
- *Unsafe* foreign calls are simply "fat machine instructions"
and are passed along to the native code generator
- *Safe* foreign calls are "lowered" to unsafe calls by wrapping
them in the above save/restore sequence. This step is done
very late in the pipeline, just before handing to the native
code gen.
This lowering is done by BuildInfoTables.lowerSafeForeignCalls
NEW PLAN for foreign calls:
- Unsafe foreign calls remain as a middle node (fat machine instruction)
Even the parameter passing is not lowered (just as machine instrs
get arguments).
- Initially, safe foreign calls appear as LastCalls with
----------------------------------------------------
Cmm representations
----------------------------------------------------
* CmmDecl.hs
The type [GenCmm d h g] represents a whole module,
** one list element per .o file **
Without SplitObjs, the list has exactly one element
newtype GenCmm d h g = Cmm [GenCmmTop d h g] -- A whole .o file
data GenCmmTop d h g
= CmmProc h g -- One procedure, graph d
| CmmData <stuff> [d] -- Initialised data, items d
Old and new piplines use different representations
(CmmCvt.hs converts between the two)
-------------
OLD BACK END representations (OldCmm.hs):
type Cmm = GenCmm CmmStatic CmmInfo (ListGraph CmmStmt)
-- A whole module
newtype ListGraph i = ListGraph [GenBasicBlock i]
data CmmStmt = Assign | Store | Return etc -- OLD BACK END ONLY
Once the info tables are laid out, we replace CmmInfo with [CmmStatic]
type RawCmm = GenCmm CmmStatic [CmmStatic] (ListGraph CmmStmt)
which represents the info tables as data, that should
immediately precede the code
-------------
NEW BACK END representations
* Uses Hoopl library, a zero-boot package
* CmmNode defines a node of a flow graph.
* Cmm defines CmmGraph, CmmTop, Cmm
- CmmGraph is a closed/closed graph + an entry node.
data CmmGraph = CmmGraph { g_entry :: BlockId
, g_graph :: Graph CmmNode C C }
- CmmTop is a top level chunk, specialization of GenCmmTop from CmmDecl.hs
with CmmGraph as a flow graph.
- Cmm is a collection of CmmTops.
type Cmm = GenCmm CmmStatic CmmTopInfo CmmGraph
type CmmTop = GenCmmTop CmmStatic CmmTopInfo CmmGraph
- CmmTop uses CmmTopInfo, which is a CmmInfoTable and CmmStackInfo
data CmmTopInfo = TopInfo {info_tbl :: CmmInfoTable, stack_info :: CmmStackInfo}
- CmmStackInfo
data CmmStackInfo = StackInfo {arg_space :: ByteOff, updfr_space :: Maybe ByteOff}
* arg_space = SP offset on entry
* updfr_space space = SP offset on exit
Once the staci is manifested, we could drom CmmStackInfo, ie. get
GenCmm CmmStatic CmmInfoTable CmmGraph, but we do not do that currently.
* MkGraph.hs: smart constructors for Cmm.hs
Beware, the CmmAGraph defined here does not use AGraph from Hoopl,
as CmmAGraph can be opened or closed at exit, See the notes in that module.
-------------
* SHARED stuff
CmmDecl.hs - GenCmm and GenCmmTop types
CmmExpr.hs - defines the Cmm expression types
- CmmExpr, CmmReg, CmmLit, LocalReg, GlobalReg
- Area, AreaId etc (separate module?)
CmmType.hs - CmmType, Width etc (saparate module?)
CmmMachOp.hs - MachOp and CallishMachOp types
BlockId.hs defines BlockId, BlockEnv, BlockSet
-------------
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment