| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
| |
Justification in #22231. Short form: In a demand like `1C1(C1(L))`
it was too easy to confuse which `1` belongs to which `C`. Now
that should be more obvious.
Fixes #22231
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch fixes #21286, by not unboxing dictionaries in
worker/wrapper (ever). The main payload is tiny:
* In `GHC.Core.Opt.DmdAnal.finaliseArgBoxities`, do not unbox
dictionaries in `get_dmd`. See Note [Do not unbox class dictionaries]
in that module
* I also found that imported wrappers were being fruitlessly
specialised, so I fixed that too, in canSpecImport.
See Note [Specialising imported functions] point (2).
In doing due diligence in the testsuite I fixed a number of
other things:
* Improve Note [Specialising unfoldings] in GHC.Core.Unfold.Make,
and Note [Inline specialisations] in GHC.Core.Opt.Specialise,
and remove duplication between the two. The new Note describes
how we specialise functions with an INLINABLE pragma.
And simplify the defn of `spec_unf` in `GHC.Core.Opt.Specialise.specCalls`.
* Improve Note [Worker/wrapper for INLINABLE functions] in
GHC.Core.Opt.WorkWrap.
And (critially) make an actual change which is to propagate the
user-written pragma from the original function to the wrapper; see
`mkStrWrapperInlinePrag`.
* Write new Note [Specialising imported functions] in
GHC.Core.Opt.Specialise
All this has a big effect on some compile times. This is
compiler/perf, showing only changes over 1%:
Metrics: compile_time/bytes allocated
-------------------------------------
LargeRecord(normal) -50.2% GOOD
ManyConstructors(normal) +1.0%
MultiLayerModulesTH_OneShot(normal) +2.6%
PmSeriesG(normal) -1.1%
T10547(normal) -1.2%
T11195(normal) -1.2%
T11276(normal) -1.0%
T11303b(normal) -1.6%
T11545(normal) -1.4%
T11822(normal) -1.3%
T12150(optasm) -1.0%
T12234(optasm) -1.2%
T13056(optasm) -9.3% GOOD
T13253(normal) -3.8% GOOD
T15164(normal) -3.6% GOOD
T16190(normal) -2.1%
T16577(normal) -2.8% GOOD
T16875(normal) -1.6%
T17836(normal) +2.2%
T17977b(normal) -1.0%
T18223(normal) -33.3% GOOD
T18282(normal) -3.4% GOOD
T18304(normal) -1.4%
T18698a(normal) -1.4% GOOD
T18698b(normal) -1.3% GOOD
T19695(normal) -2.5% GOOD
T5837(normal) -2.3%
T9630(normal) -33.0% GOOD
WWRec(normal) -9.7% GOOD
hard_hole_fits(normal) -2.1% GOOD
hie002(normal) +1.6%
geo. mean -2.2%
minimum -50.2%
maximum +2.6%
I diligently investigated some of the big drops.
* Caused by not doing w/w for dictionaries:
T13056, T15164, WWRec, T18223
* Caused by not fruitlessly specialising wrappers
LargeRecord, T9630
For runtimes, here is perf/should+_run:
Metrics: runtime/bytes allocated
--------------------------------
T12990(normal) -3.8%
T5205(normal) -1.3%
T9203(normal) -10.7% GOOD
haddock.Cabal(normal) +0.1%
haddock.base(normal) -1.1%
haddock.compiler(normal) -0.3%
lazy-bs-alloc(normal) -0.2%
------------------------------------------
geo. mean -0.3%
minimum -10.7%
maximum +0.1%
I did not investigate exactly what happens in T9203.
Nofib is a wash:
+-------------------------------++--+-----------+-----------+
| || | tsv (rel) | std. err. |
+===============================++==+===========+===========+
| real/anna || | -0.13% | 0.0% |
| real/fem || | +0.13% | 0.0% |
| real/fulsom || | -0.16% | 0.0% |
| real/lift || | -1.55% | 0.0% |
| real/reptile || | -0.11% | 0.0% |
| real/smallpt || | +0.51% | 0.0% |
| spectral/constraints || | +0.20% | 0.0% |
| spectral/dom-lt || | +1.80% | 0.0% |
| spectral/expert || | +0.33% | 0.0% |
+===============================++==+===========+===========+
| geom mean || | | |
+-------------------------------++--+-----------+-----------+
I spent quite some time investigating dom-lt, but it's pretty
complicated. See my note on !7847. Conclusion: it's just a delicate
inlining interaction, and we have plenty of those.
Metric Decrease:
LargeRecord
T13056
T13253
T15164
T16577
T18223
T18282
T18698a
T18698b
T19695
T9630
WWRec
hard_hole_fits
T9203
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In #21717 we saw a reportedly unsound strictness signature due to an unsound
definition of plusSubDmd on Calls. This patch contains a description and the fix
to the unsoundness as outlined in `Note [Call SubDemand vs. evaluation Demand]`.
This fix means we also get rid of the special handling of `-fpedantic-bottoms`
in eta-reduction. Thanks to less strict and actually sound strictness results,
we will no longer eta-reduce the problematic cases in the first place, even
without `-fpedantic-bottoms`.
So fixing the unsoundness also makes our eta-reduction code simpler with less
hacks to explain. But there is another, more unfortunate side-effect:
We *unfix* #21085, but fortunately we have a new fix ready:
See `Note [mkCall and plusSubDmd]`.
There's another change:
I decided to make `Note [SubDemand denotes at least one evaluation]` a lot
simpler by using `plusSubDmd` (instead of `lubPlusSubDmd`) even if both argument
demands are lazy. That leads to less precise results, but in turn rids ourselves
from the need for 4 different `OpMode`s and the complication of
`Note [Manual specialisation of lub*Dmd/plus*Dmd]`. The result is simpler code
that is in line with the paper draft on Demand Analysis.
I left the abandoned idea in `Note [Unrealised opportunity in plusDmd]` for
posterity. The fallout in terms of regressions is negligible, as the testsuite
and NoFib shows.
```
Program Allocs Instrs
--------------------------------------------------------------------------------
hidden +0.2% -0.2%
linear -0.0% -0.7%
--------------------------------------------------------------------------------
Min -0.0% -0.7%
Max +0.2% +0.0%
Geometric Mean +0.0% -0.0%
```
Fixes #21717.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This MR fixes #21694, #21755. It also makes sure that #21948 and
fix to #21694.
* For #21694 the underlying problem was that we were calling arityType
on an expression that had free join points. This is a Bad Bad Idea.
See Note [No free join points in arityType].
* To make "no free join points in arityType" work out I had to avoid
trying to use eta-expansion for runRW#. This entailed a few changes
in the Simplifier's treatment of runRW#. See
GHC.Core.Opt.Simplify.Iteration Note [No eta-expansion in runRW#]
* I also made andArityType work correctly with -fpedantic-bottoms;
see Note [Combining case branches: andWithTail].
* Rewrote Note [Combining case branches: optimistic one-shot-ness]
* arityType previously treated join points differently to other
let-bindings. This patch makes them unform; arityType analyses
the RHS of all bindings to get its ArityType, and extends am_sigs.
I realised that, now we have am_sigs giving the ArityType for
let-bound Ids, we don't need the (pre-dating) special code in
arityType for join points. But instead we need to extend the env for
Rec bindings, which weren't doing before. More uniform now. See
Note [arityType for let-bindings].
This meant we could get rid of ae_joins, and in fact get rid of
EtaExpandArity altogether. Simpler.
* And finally, it was the strange treatment of join-point Ids in
arityType (involving a fake ABot type) that led to a serious bug:
#21755. Fixed by this refactoring, which treats them uniformly;
but without breaking #18328.
In fact, the arity for recursive join bindings is pretty tricky;
see the long Note [Arity for recursive join bindings]
in GHC.Core.Opt.Simplify.Utils. That led to more refactoring,
including deciding that an Id could have an Arity that is bigger
than its JoinArity; see Note [Invariants on join points], item
2(b) in GHC.Core
* Make sure that the "demand threshold" for join points in DmdAnal
is no bigger than the join-arity. In GHC.Core.Opt.DmdAnal see
Note [Demand signatures are computed for a threshold arity based on idArity]
* I moved GHC.Core.Utils.exprIsDeadEnd into GHC.Core.Opt.Arity,
where it more properly belongs.
* Remove an old, redundant hack in FloatOut. The old Note was
Note [Bottoming floats: eta expansion] in GHC.Core.Opt.SetLevels.
Compile time improves very slightly on average:
Metrics: compile_time/bytes allocated
---------------------------------------------------------------------------------------
T18223(normal) ghc/alloc 725,808,720 747,839,216 +3.0% BAD
T6048(optasm) ghc/alloc 105,006,104 101,599,472 -3.2% GOOD
geo. mean -0.2%
minimum -3.2%
maximum +3.0%
For some reason Windows was better
T10421(normal) ghc/alloc 125,888,360 124,129,168 -1.4% GOOD
T18140(normal) ghc/alloc 85,974,520 83,884,224 -2.4% GOOD
T18698b(normal) ghc/alloc 236,764,568 234,077,288 -1.1% GOOD
T18923(normal) ghc/alloc 75,660,528 73,994,512 -2.2% GOOD
T6048(optasm) ghc/alloc 112,232,512 108,182,520 -3.6% GOOD
geo. mean -0.6%
I had a quick look at T18223 but it is knee deep in coercions and
the size of everything looks similar before and after. I decided
to accept that 3% increase in exchange for goodness elsewhere.
Metric Decrease:
T10421
T18140
T18698b
T18923
T6048
Metric Increase:
T18223
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We used to put OtherCon unfoldings on lambda binders of workers
and sometimes also join points/specializations with with the
assumption that since the wrapper would force these arguments
once we execute the RHS they would indeed be in WHNF.
This was wrong for reasons detailed in #21472. So now we purge
evaluated unfoldings from *all* lambda binders.
This fixes #21472, but at the cost of sometimes not using as efficient a
calling convention. It can also change inlining behaviour as some
occurances will no longer look like value arguments when they did
before.
As consequence we also change how we compute CBV information for
arguments slightly. We now *always* determine the CBV convention
for arguments during tidy. Earlier in the pipeline we merely mark
functions as candidates for having their arguments treated as CBV.
As before the process is described in the relevant notes:
Note [CBV Function Ids]
Note [Attaching CBV Marks to ids]
Note [Never put `OtherCon` unfoldigns on lambda binders]
-------------------------
Metric Decrease:
T12425
T13035
T18223
T18223
T18923
MultiLayerModulesTH_OneShot
Metric Increase:
WWRec
-------------------------
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Similar to the fix to #20836 in CorePrep, we now track the set of enclosing
recursive binders in the SimplEnv and SimpleOptEnv.
See Note [Eta reduction in recursive RHSs] for details.
I also updated Note [Arity robustness] with the insights Simon and I had in a
call discussing the issue.
Fixes #21652.
Unfortunately, we get a 5% ghc/alloc regression in T16577. That is due to
additional eta reduction in GHC.Read.choose1 and the resulting ANF-isation
of a large list literal at the top-level that didn't happen before (presumably
because it was too interesting to float to the top-level). There's not much we
can do about that.
Metric Increase:
T16577
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is a large collection of changes all relating to eta
reduction, originally triggered by #18993, but there followed
a long saga.
Specifics:
* Move state-hack stuff from GHC.Types.Id (where it never belonged)
to GHC.Core.Opt.Arity (which seems much more appropriate).
* Add a crucial mkCast in the Cast case of
GHC.Core.Opt.Arity.eta_expand; helps with T18223
* Add clarifying notes about eta-reducing to PAPs.
See Note [Do not eta reduce PAPs]
* I moved tryEtaReduce from GHC.Core.Utils to GHC.Core.Opt.Arity,
where it properly belongs. See Note [Eta reduce PAPs]
* In GHC.Core.Opt.Simplify.Utils.tryEtaExpandRhs, pull out the code for
when eta-expansion is wanted, to make wantEtaExpansion, and all that
same function in GHC.Core.Opt.Simplify.simplStableUnfolding. It was
previously inconsistent, but it's doing the same thing.
* I did a substantial refactor of ArityType; see Note [ArityType].
This allowed me to do away with the somewhat mysterious takeOneShots;
more generally it allows arityType to describe the function, leaving
its clients to decide how to use that information.
I made ArityType abstract, so that clients have to use functions
to access it.
* Make GHC.Core.Opt.Simplify.Utils.rebuildLam (was stupidly called
mkLam before) aware of the floats that the simplifier builds up, so
that it can still do eta-reduction even if there are some floats.
(Previously that would not happen.) That means passing the floats
to rebuildLam, and an extra check when eta-reducting (etaFloatOk).
* In GHC.Core.Opt.Simplify.Utils.tryEtaExpandRhs, make use of call-info
in the idDemandInfo of the binder, as well as the CallArity info. The
occurrence analyser did this but we were failing to take advantage here.
In the end I moved the heavy lifting to GHC.Core.Opt.Arity.findRhsArity;
see Note [Combining arityType with demand info], and functions
idDemandOneShots and combineWithDemandOneShots.
(These changes partly drove my refactoring of ArityType.)
* In GHC.Core.Opt.Arity.findRhsArity
* I'm now taking account of the demand on the binder to give
extra one-shot info. E.g. if the fn is always called with two
args, we can give better one-shot info on the binders
than if we just look at the RHS.
* Don't do any fixpointing in the non-recursive
case -- simple short cut.
* Trim arity inside the loop. See Note [Trim arity inside the loop]
* Make SimpleOpt respect the eta-reduction flag
(Some associated refactoring here.)
* I made the CallCtxt which the Simplifier uses distinguish between
recursive and non-recursive right-hand sides.
data CallCtxt = ... | RhsCtxt RecFlag | ...
It affects only one thing:
- We call an RHS context interesting only if it is non-recursive
see Note [RHS of lets] in GHC.Core.Unfold
* Remove eta-reduction in GHC.CoreToStg.Prep, a welcome simplification.
See Note [No eta reduction needed in rhsToBody] in GHC.CoreToStg.Prep.
Other incidental changes
* Fix a fairly long-standing outright bug in the ApplyToVal case of
GHC.Core.Opt.Simplify.mkDupableContWithDmds. I was failing to take the
tail of 'dmds' in the recursive call, which meant the demands were All
Wrong. I have no idea why this has not caused problems before now.
* Delete dead function GHC.Core.Opt.Simplify.Utils.contIsRhsOrArg
Metrics: compile_time/bytes allocated
Test Metric Baseline New value Change
---------------------------------------------------------------------------------------
MultiLayerModulesTH_OneShot(normal) ghc/alloc 2,743,297,692 2,619,762,992 -4.5% GOOD
T18223(normal) ghc/alloc 1,103,161,360 972,415,992 -11.9% GOOD
T3064(normal) ghc/alloc 201,222,500 184,085,360 -8.5% GOOD
T8095(normal) ghc/alloc 3,216,292,528 3,254,416,960 +1.2%
T9630(normal) ghc/alloc 1,514,131,032 1,557,719,312 +2.9% BAD
parsing001(normal) ghc/alloc 530,409,812 525,077,696 -1.0%
geo. mean -0.1%
Nofib:
Program Size Allocs Runtime Elapsed TotalMem
--------------------------------------------------------------------------------
banner +0.0% +0.4% -8.9% -8.7% 0.0%
exact-reals +0.0% -7.4% -36.3% -37.4% 0.0%
fannkuch-redux +0.0% -0.1% -1.0% -1.0% 0.0%
fft2 -0.1% -0.2% -17.8% -19.2% 0.0%
fluid +0.0% -1.3% -2.1% -2.1% 0.0%
gg -0.0% +2.2% -0.2% -0.1% 0.0%
spectral-norm +0.1% -0.2% 0.0% 0.0% 0.0%
tak +0.0% -0.3% -9.8% -9.8% 0.0%
x2n1 +0.0% -0.2% -3.2% -3.2% 0.0%
--------------------------------------------------------------------------------
Min -3.5% -7.4% -58.7% -59.9% 0.0%
Max +0.1% +2.2% +32.9% +32.9% 0.0%
Geometric Mean -0.0% -0.1% -14.2% -14.8% -0.0%
Metric Decrease:
MultiLayerModulesTH_OneShot
T18223
T3064
T15185
T14766
Metric Increase:
T9630
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
See the new `Note [SubDemand denotes at least one evaluation]`.
A demand `n :* sd` on a let binder `x=e` now means
> "`x` was evaluated `n` times and in any program trace it is evaluated, `e` is
> evaluated deeply in sub-demand `sd`."
The "any time it is evaluated" premise is what this patch adds. As a result,
we get better nested strictness. For example (T21081)
```hs
f :: (Bool, Bool) -> (Bool, Bool)
f pr = (case pr of (a,b) -> a /= b, True)
-- before: <MP(L,L)>
-- after: <MP(SL,SL)>
g :: Int -> (Bool, Bool)
g x = let y = let z = odd x in (z,z) in f y
```
The change in demand signature "before" to "after" allows us to case-bind `z`
here.
Similarly good things happen for the `sd` in call sub-demands `Cn(sd)`, which
allows for more eta-reduction (which is only sound with `-fno-pedantic-bottoms`,
albeit).
We also fix #21085, a surprising inconsistency with `Poly` to `Call` sub-demand
expansion.
In an attempt to fix a regression caused by less inlining due to eta-reduction
in T15426, I eta-expanded the definition of `elemIndex` and `elemIndices`, thus
fixing #21345 on the go.
The main point of this patch is that it fixes #21081 and #21133.
Annoyingly, I discovered that more precise demand signatures for join points can
transform a program into a lazier program if that join point gets floated to the
top-level, see #21392. There is no simple fix at the moment, but !5349 might.
Thus, we accept a ~5% regression in `MultiLayerModulesTH_OneShot`, where #21392
bites us in `addListToUniqDSet`. T21392 reliably reproduces the issue.
Surprisingly, ghc/alloc perf on Windows improves much more than on other jobs, by
0.4% in the geometric mean and by 2% in T16875.
Metric Increase:
MultiLayerModulesTH_OneShot
Metric Decrease:
T16875
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously, we let `Unboxed` win in `lubBoxity`, which is unsoundly optimistic
in terms ob Boxity analysis. "Unsoundly" in the sense that we sometimes unbox
parameters that we better shouldn't unbox. Examples are #18907 and T19871.absent.
Until now, we thought that this hack pulled its weight becuase it worked around
some shortcomings of the phase separation between Boxity analysis and CPR
analysis. But it is a gross hack which caused regressions itself that needed all
kinds of fixes and workarounds. See for example #20767. It became impossible to
work with in !7599, so I want to remove it.
For example, at the moment, `lubDmd B dmd` will not unbox `dmd`,
but `lubDmd A dmd` will. Given that `B` is supposed to be the bottom element of
the lattice, it's hardly justifiable to get a better demand when `lub`bing with
`A`.
The consequence of letting `Boxed` win in `lubBoxity` is that we *would* regress
#2387, #16040 and parts of #5075 and T19871.sumIO, until Boxity and CPR
are able to communicate better. Fortunately, that is not the case since I could
tweak the other source of optimism in Boxity analysis that is described in
`Note [Unboxed demand on function bodies returning small products]` so that
we *recursively* assume unboxed demands on function bodies returning small
products. See the updated Note.
`Note [Boxity for bottoming functions]` describes why we need bottoming
functions to have signatures that say that they deeply unbox their arguments.
In so doing, I had to tweak `finaliseArgBoxities` so that it will never unbox
recursive data constructors. This is in line with our handling of them in CPR.
I updated `Note [Which types are unboxed?]` to reflect that.
In turn we fix #21119, #20767, #18907, T19871.absent and get a much simpler
implementation (at least to think about). We can also drop the very ad-hoc
definition of `deferAfterPreciseException` and its Note in favor of the
simple, intuitive definition we used to have.
Metric Decrease:
T16875
T18223
T18698a
T18698b
hard_hole_fits
Metric Increase:
LargeRecord
MultiComponentModulesRecomp
T15703
T8095
T9872d
Out of all the regresions, only the one in T9872d doesn't vanish in a perf
build, where the compiler is bootstrapped with -O2 and thus SpecConstr.
Reason for regressions:
* T9872d is due to `ty_co_subst` taking its `LiftingContext` boxed.
That is because the context is passed to a function argument, for
example in `liftCoSubstTyVarBndrUsing`.
* In T15703, LargeRecord and T8095, we get a bit more allocations in
`expand_syn` and `piResultTys`, because a `TCvSubst` isn't unboxed.
In both cases that guards against reboxing in some code paths.
* The same is true for MultiComponentModulesRecomp, where we get less unboxing
in `GHC.Unit.Finder.$wfindInstalledHomeModule`. In a perf build, allocations
actually *improve* by over 4%!
Results on NoFib:
--------------------------------------------------------------------------------
Program Allocs Instrs
--------------------------------------------------------------------------------
awards -0.4% +0.3%
cacheprof -0.3% +2.4%
fft -1.5% -5.1%
fibheaps +1.2% +0.8%
fluid -0.3% -0.1%
ida +0.4% +0.9%
k-nucleotide +0.4% -0.1%
last-piece +10.5% +13.9%
lift -4.4% +3.5%
mandel2 -99.7% -99.8%
mate -0.4% +3.6%
parser -1.0% +0.1%
puzzle -11.6% +6.5%
reverse-complem -3.0% +2.0%
scs -0.5% +0.1%
sphere -0.4% -0.2%
wave4main -8.2% -0.3%
--------------------------------------------------------------------------------
Summary excludes mandel2 because of excessive bias
Min -11.6% -5.1%
Max +10.5% +13.9%
Geometric Mean -0.2% +0.3%
--------------------------------------------------------------------------------
Not bad for a bug fix.
The regression in `last-piece` could become a win if SpecConstr would work on
non-recursive functions. The regression in `fibheaps` is due to
`Note [Reboxed crud for bottoming calls]`, e.g., #21128.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This does three major things:
* Enforce the invariant that all strict fields must contain tagged
pointers.
* Try to predict the tag on bindings in order to omit tag checks.
* Allows functions to pass arguments unlifted (call-by-value).
The former is "simply" achieved by wrapping any constructor allocations with
a case which will evaluate the respective strict bindings.
The prediction is done by a new data flow analysis based on the STG
representation of a program. This also helps us to avoid generating
redudant cases for the above invariant.
StrictWorkers are created by W/W directly and SpecConstr indirectly.
See the Note [Strict Worker Ids]
Other minor changes:
* Add StgUtil module containing a few functions needed by, but
not specific to the tag analysis.
-------------------------
Metric Decrease:
T12545
T18698b
T18140
T18923
LargeRecord
Metric Increase:
LargeRecord
ManyAlternatives
ManyConstructors
T10421
T12425
T12707
T13035
T13056
T13253
T13253-spj
T13379
T15164
T18282
T18304
T18698a
T1969
T20049
T3294
T4801
T5321FD
T5321Fun
T783
T9233
T9675
T9961
T19695
WWRec
-------------------------
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch fixes some abundant reboxing of `DynFlags` in
`GHC.HsToCore.Match.Literal.warnAboutOverflowedLit` (which was the topic
of #19407) by introducing a Boxity analysis to GHC, done as part of demand
analysis. This allows to accurately capture ad-hoc unboxing decisions previously
made in worker/wrapper in demand analysis now, where the boxity info can
propagate through demand signatures.
See the new `Note [Boxity analysis]`. The actual fix for #19407 is described in
`Note [No lazy, Unboxed demand in demand signature]`, but
`Note [Finalising boxity for demand signature]` is probably a better entry-point.
To support the fix for #19407, I had to change (what was)
`Note [Add demands for strict constructors]` a bit
(now `Note [Unboxing evaluated arguments]`). In particular, we now take care of
it in `finaliseBoxity` (which is only called from demand analaysis) instead of
`wantToUnboxArg`.
I also had to resurrect `Note [Product demands for function body]` and rename
it to `Note [Unboxed demand on function bodies returning small products]` to
avoid huge regressions in `join004` and `join007`, thereby fixing #4267 again.
See the updated Note for details.
A nice side-effect is that the worker/wrapper transformation no longer needs to
look at strictness info and other bits such as `InsideInlineableFun` flags
(needed for `Note [Do not unbox class dictionaries]`) at all. It simply collects
boxity info from argument demands and interprets them with a severely simplified
`wantToUnboxArg`. All the smartness is in `finaliseBoxity`, which could be moved
to DmdAnal completely, if it wasn't for the call to `dubiousDataConInstArgTys`
which would be awkward to export.
I spent some time figuring out the reason for why `T16197` failed prior to my
amendments to `Note [Unboxing evaluated arguments]`. After having it figured
out, I minimised it a bit and added `T16197b`, which simply compares computed
strictness signatures and thus should be far simpler to eyeball.
The 12% ghc/alloc regression in T11545 is because of the additional `Boxity`
field in `Poly` and `Prod` that results in more allocation during `lubSubDmd`
and `plusSubDmd`. I made sure in the ticky profiles that the number of calls
to those functions stayed the same. We can bear such an increase here, as we
recently improved it by -68% (in b760c1f).
T18698* regress slightly because there is more unboxing of dictionaries
happening and that causes Lint (mostly) to allocate more.
Fixes #19871, #19407, #4267, #16859, #18907 and #13331.
Metric Increase:
T11545
T18698a
T18698b
Metric Decrease:
T12425
T16577
T18223
T18282
T4267
T9961
|
|
|
|
|
|
|
|
|
| |
T17516 allocations increase by 48% because Integer's predicates are
inlined in some Ord instance methods. These methods become too big to be
inlined while they probably should: this is tracked in #20516.
Metric Increase:
T17516
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch enables worker/wrapper for nested constructed products, as described
in `Note [Nested CPR]`. The machinery for expressing Nested CPR was already
there, since !5054. Worker/wrapper is equipped to exploit Nested CPR annotations
since !5338. CPR analysis already handles applications in batches since !5753.
This patch just needs to flip a few more switches:
1. In `cprTransformDataConWork`, we need to look at the field expressions
and their `CprType`s to see whether the evaluation of the expressions
terminates quickly (= is in HNF) or if they are put in strict fields.
If that is the case, then we retain their CPR info and may unbox nestedly
later on. More details in `Note [Nested CPR]`.
2. Enable nested `ConCPR` signatures in `GHC.Types.Cpr`.
3. In the `asConCpr` call in `GHC.Core.Opt.WorkWrap.Utils`, pass CPR info of
fields to the `Unbox`.
4. Instead of giving CPR signatures to DataCon workers and wrappers, we now have
`cprTransformDataConWork` for workers and treat wrappers by analysing their
unfolding. As a result, the code from GHC.Types.Id.Make went away completely.
5. I deactivated worker/wrappering for recursive DataCons and wrote a function
`isRecDataCon` to detect them. We really don't want to give `repeat` or
`replicate` the Nested CPR property.
See Note [CPR for recursive data structures] for which kind of recursive
DataCons we target.
6. Fix a couple of tests and their outputs.
I also documented that CPR can destroy sharing and lead to asymptotic increase
in allocations (which is tracked by #13331/#19326) in
`Note [CPR for data structures can destroy sharing]`.
Nofib results:
```
--------------------------------------------------------------------------------
Program Allocs Instrs
--------------------------------------------------------------------------------
ben-raytrace -3.1% -0.4%
binary-trees +0.8% -2.9%
digits-of-e2 +5.8% +1.2%
event +0.8% -2.1%
fannkuch-redux +0.0% -1.4%
fish 0.0% -1.5%
gamteb -1.4% -0.3%
mkhprog +1.4% +0.8%
multiplier +0.0% -1.9%
pic -0.6% -0.1%
reptile -20.9% -17.8%
wave4main +4.8% +0.4%
x2n1 -100.0% -7.6%
--------------------------------------------------------------------------------
Min -95.0% -17.8%
Max +5.8% +1.2%
Geometric Mean -2.9% -0.4%
```
The huge wins in x2n1 (loopy list) and reptile (see #19970) are due to
refraining from unboxing (:). Other benchmarks like digits-of-e2 or wave4main
regress because of that. Ultimately there are no great improvements due to
Nested CPR alone, but at least it's a win.
Binary sizes decrease by 0.6%.
There are a significant number of metric decreases. The most notable ones (>1%):
```
ManyAlternatives(normal) ghc/alloc 771656002.7 762187472.0 -1.2%
ManyConstructors(normal) ghc/alloc 4191073418.7 4114369216.0 -1.8%
MultiLayerModules(normal) ghc/alloc 3095678333.3 3128720704.0 +1.1%
PmSeriesG(normal) ghc/alloc 50096429.3 51495664.0 +2.8%
PmSeriesS(normal) ghc/alloc 63512989.3 64681600.0 +1.8%
PmSeriesV(normal) ghc/alloc 62575424.0 63767208.0 +1.9%
T10547(normal) ghc/alloc 29347469.3 29944240.0 +2.0%
T11303b(normal) ghc/alloc 46018752.0 47367576.0 +2.9%
T12150(optasm) ghc/alloc 81660890.7 82547696.0 +1.1%
T12234(optasm) ghc/alloc 59451253.3 60357952.0 +1.5%
T12545(normal) ghc/alloc 1705216250.7 1751278952.0 +2.7%
T12707(normal) ghc/alloc 981000472.0 968489800.0 -1.3% GOOD
T13056(optasm) ghc/alloc 389322664.0 372495160.0 -4.3% GOOD
T13253(normal) ghc/alloc 337174229.3 341954576.0 +1.4%
T13701(normal) ghc/alloc 2381455173.3 2439790328.0 +2.4% BAD
T14052(ghci) ghc/alloc 2162530642.7 2139108784.0 -1.1%
T14683(normal) ghc/alloc 3049744728.0 2977535064.0 -2.4% GOOD
T14697(normal) ghc/alloc 362980213.3 369304512.0 +1.7%
T15164(normal) ghc/alloc 1323102752.0 1307480600.0 -1.2%
T15304(normal) ghc/alloc 1304607429.3 1291024568.0 -1.0%
T16190(normal) ghc/alloc 281450410.7 284878048.0 +1.2%
T16577(normal) ghc/alloc 7984960789.3 7811668768.0 -2.2% GOOD
T17516(normal) ghc/alloc 1171051192.0 1153649664.0 -1.5%
T17836(normal) ghc/alloc 1115569746.7 1098197592.0 -1.6%
T17836b(normal) ghc/alloc 54322597.3 55518216.0 +2.2%
T17977(normal) ghc/alloc 47071754.7 48403408.0 +2.8%
T17977b(normal) ghc/alloc 42579133.3 43977392.0 +3.3%
T18923(normal) ghc/alloc 71764237.3 72566240.0 +1.1%
T1969(normal) ghc/alloc 784821002.7 773971776.0 -1.4% GOOD
T3294(normal) ghc/alloc 1634913973.3 1614323584.0 -1.3% GOOD
T4801(normal) ghc/alloc 295619648.0 292776440.0 -1.0%
T5321FD(normal) ghc/alloc 278827858.7 276067280.0 -1.0%
T5631(normal) ghc/alloc 586618202.7 577579960.0 -1.5%
T5642(normal) ghc/alloc 494923048.0 487927208.0 -1.4%
T5837(normal) ghc/alloc 37758061.3 39261608.0 +4.0%
T9020(optasm) ghc/alloc 257362077.3 254672416.0 -1.0%
T9198(normal) ghc/alloc 49313365.3 50603936.0 +2.6% BAD
T9233(normal) ghc/alloc 704944258.7 685692712.0 -2.7% GOOD
T9630(normal) ghc/alloc 1476621560.0 1455192784.0 -1.5%
T9675(optasm) ghc/alloc 443183173.3 433859696.0 -2.1% GOOD
T9872a(normal) ghc/alloc 1720926653.3 1693190072.0 -1.6% GOOD
T9872b(normal) ghc/alloc 2185618061.3 2162277568.0 -1.1% GOOD
T9872c(normal) ghc/alloc 1765842405.3 1733618088.0 -1.8% GOOD
TcPlugin_RewritePerf(normal) ghc/alloc 2388882730.7 2365504696.0 -1.0%
WWRec(normal) ghc/alloc 607073186.7 597512216.0 -1.6%
T9203(normal) run/alloc 107284064.0 102881832.0 -4.1%
haddock.Cabal(normal) run/alloc 24025329589.3 23768382560.0 -1.1%
haddock.base(normal) run/alloc 25660521653.3 25370321824.0 -1.1%
haddock.compiler(normal) run/alloc 74064171706.7 73358712280.0 -1.0%
```
The biggest exception to the rule is T13701 which seems to fluctuate as usual
(not unlike T12545). T14697 has a similar quality, being a generated
multi-module test. T5837 is small enough that it similarly doesn't measure
anything significant besides module loading overhead.
T13253 simply does one additional round of Simplification due to Nested CPR.
There are also some apparent regressions in T9198, T12234 and PmSeriesG that we
(@mpickering and I) were simply unable to reproduce locally. @mpickering tried
to run the CI script in a local Docker container and actually found that T9198
and PmSeriesG *improved*. In MRs that were rebased on top this one, like !4229,
I did not experience such increases. Let's not get hung up on these regression
tests, they were meant to test for asymptotic regressions.
The build-cabal test improves by 1.2% in -O0.
Metric Increase:
T10421
T12234
T12545
T13035
T13056
T13701
T14697
T18923
T5837
T9198
Metric Decrease:
ManyConstructors
T12545
T12707
T13056
T14683
T16577
T18223
T1969
T3294
T9203
T9233
T9675
T9872a
T9872b
T9872c
T9961
TcPlugin_RewritePerf
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This commit:
commit c6faa42bfb954445c09c5680afd4fb875ef03758
Author: Simon Peyton Jones <simonpj@microsoft.com>
Date: Mon Mar 9 10:20:42 2020 +0000
Avoid useless w/w split
This patch is just a tidy-up for the post-strictness-analysis
worker wrapper split. Consider
f x = x
Strictnesss analysis does not lead to a w/w split, so the
obvious thing is to leave it 100% alone. But actually, because
the RHS is small, we ended up adding a StableUnfolding for it.
There is some reason to do this if we choose /not/ do to w/w
on the grounds that the function is small. See
Note [Don't w/w inline small non-loop-breaker things]
But there is no reason if we would not have done w/w anyway.
This patch just moves the conditional to later. Easy.
turns out to have a bug in it. Instead of /moving/ the conditional,
I /duplicated/ it. Then in a subsequent unrelated tidy-up
(087ac4eb) I removed the second (redundant) test!
This patch does what I originally intended.
There is also a small refactoring in GHC.Core.Unfold, to make the
code clearer, but with no change in behaviour.
It does, however, have a generally good effect on compile times,
because we aren't dealing with so many silly stable unfoldings.
Here are the non-zero changes:
Metrics: compile_time/bytes allocated
-------------------------------------
Baseline
Test Metric value New value Change
---------------------------------------------------------------------------
ManyAlternatives(normal) ghc/alloc 791969344.0 792665048.0 +0.1%
ManyConstructors(normal) ghc/alloc 4351126824.0 4358303528.0 +0.2%
PmSeriesG(normal) ghc/alloc 50362552.0 50482208.0 +0.2%
PmSeriesS(normal) ghc/alloc 63733024.0 63619912.0 -0.2%
T10421(normal) ghc/alloc 121224624.0 119695448.0 -1.3% GOOD
T10421a(normal) ghc/alloc 85256392.0 83714224.0 -1.8%
T10547(normal) ghc/alloc 29253072.0 29258256.0 +0.0%
T10858(normal) ghc/alloc 189343152.0 187972328.0 -0.7%
T11195(normal) ghc/alloc 281208248.0 279727584.0 -0.5%
T11276(normal) ghc/alloc 141966952.0 142046224.0 +0.1%
T11303b(normal) ghc/alloc 46228360.0 46259024.0 +0.1%
T11545(normal) ghc/alloc 2663128768.0 2667412656.0 +0.2%
T11822(normal) ghc/alloc 138686944.0 138760176.0 +0.1%
T12227(normal) ghc/alloc 482836000.0 475421056.0 -1.5% GOOD
T12234(optasm) ghc/alloc 60710520.0 60781808.0 +0.1%
T12425(optasm) ghc/alloc 104089000.0 104022424.0 -0.1%
T12545(normal) ghc/alloc 1711759416.0 1705711528.0 -0.4%
T12707(normal) ghc/alloc 991541120.0 991921776.0 +0.0%
T13035(normal) ghc/alloc 108199872.0 108370704.0 +0.2%
T13056(optasm) ghc/alloc 414642544.0 412580384.0 -0.5%
T13253(normal) ghc/alloc 361701272.0 355838624.0 -1.6%
T13253-spj(normal) ghc/alloc 157710168.0 157397768.0 -0.2%
T13379(normal) ghc/alloc 370984400.0 371345888.0 +0.1%
T13701(normal) ghc/alloc 2439764144.0 2441351984.0 +0.1%
T14052(ghci) ghc/alloc 2154090896.0 2156671400.0 +0.1%
T15164(normal) ghc/alloc 1478517688.0 1440317696.0 -2.6% GOOD
T15630(normal) ghc/alloc 178053912.0 172489808.0 -3.1%
T16577(normal) ghc/alloc 7859948896.0 7854524080.0 -0.1%
T17516(normal) ghc/alloc 1271520128.0 1202096488.0 -5.5% GOOD
T17836(normal) ghc/alloc 1123320632.0 1123922480.0 +0.1%
T17836b(normal) ghc/alloc 54526280.0 54576776.0 +0.1%
T17977b(normal) ghc/alloc 42706752.0 42730544.0 +0.1%
T18140(normal) ghc/alloc 108834568.0 108693816.0 -0.1%
T18223(normal) ghc/alloc 5539629264.0 5579500872.0 +0.7%
T18304(normal) ghc/alloc 97589720.0 97196944.0 -0.4%
T18478(normal) ghc/alloc 770755472.0 771232888.0 +0.1%
T18698a(normal) ghc/alloc 408691160.0 374364992.0 -8.4% GOOD
T18698b(normal) ghc/alloc 492419768.0 458809408.0 -6.8% GOOD
T18923(normal) ghc/alloc 72177032.0 71368824.0 -1.1%
T1969(normal) ghc/alloc 803523496.0 804655112.0 +0.1%
T3064(normal) ghc/alloc 198411784.0 198608512.0 +0.1%
T4801(normal) ghc/alloc 312416688.0 312874976.0 +0.1%
T5321Fun(normal) ghc/alloc 325230680.0 325474448.0 +0.1%
T5631(normal) ghc/alloc 592064448.0 593518968.0 +0.2%
T5837(normal) ghc/alloc 37691496.0 37710904.0 +0.1%
T783(normal) ghc/alloc 404629536.0 405064432.0 +0.1%
T9020(optasm) ghc/alloc 266004608.0 266375592.0 +0.1%
T9198(normal) ghc/alloc 49221336.0 49268648.0 +0.1%
T9233(normal) ghc/alloc 913464984.0 742680256.0 -18.7% GOOD
T9675(optasm) ghc/alloc 552296608.0 466322000.0 -15.6% GOOD
T9872a(normal) ghc/alloc 1789910616.0 1793924472.0 +0.2%
T9872b(normal) ghc/alloc 2315141376.0 2310338056.0 -0.2%
T9872c(normal) ghc/alloc 1840422424.0 1841567224.0 +0.1%
T9872d(normal) ghc/alloc 556713248.0 556838432.0 +0.0%
T9961(normal) ghc/alloc 383809160.0 384601600.0 +0.2%
WWRec(normal) ghc/alloc 773751272.0 753949608.0 -2.6% GOOD
Residency goes down too:
Metrics: compile_time/max_bytes_used
------------------------------------
Baseline
Test Metric value New value Change
-----------------------------------------------------------
T10370(optasm) ghc/max 42058448.0 39481672.0 -6.1%
T11545(normal) ghc/max 43641392.0 43634752.0 -0.0%
T15304(normal) ghc/max 29895824.0 29439032.0 -1.5%
T15630(normal) ghc/max 8822568.0 8772328.0 -0.6%
T18698a(normal) ghc/max 13882536.0 13787112.0 -0.7%
T18698b(normal) ghc/max 14714112.0 13836408.0 -6.0%
T1969(normal) ghc/max 24724128.0 24733496.0 +0.0%
T3064(normal) ghc/max 14041152.0 14034768.0 -0.0%
T3294(normal) ghc/max 32769248.0 32760312.0 -0.0%
T9630(normal) ghc/max 41605120.0 41572184.0 -0.1%
T9675(optasm) ghc/max 18652296.0 17253480.0 -7.5%
Metric Decrease:
T10421
T12227
T15164
T17516
T18698a
T18698b
T9233
T9675
WWRec
Metric Increase:
T12545
|
|
|
|
|
|
|
|
|
| |
The update of the Outputable instance resulted in a slew of
documentation changes within Notes that used the old syntax.
The most important doc changes are to `Note [Demand notation]`
and the user's guide.
Fixes #19016.
|
|
|
|
|
|
| |
Both sub-demands encode the same information.
This is a trivial change and already affects a few regression tests
(e.g. `T5075`), so no separate regression test is necessary.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
As outlined in #18903, interleaving usage and strictness demands not
only means a more compact demand representation, but also allows us to
express demands that we weren't easily able to express before.
Call demands are *relative* in the sense that a call demand `Cn(cd)`
on `g` says "`g` is called `n` times. *Whenever `g` is called*, the
result is used according to `cd`". Example from #18903:
```hs
h :: Int -> Int
h m =
let g :: Int -> (Int,Int)
g 1 = (m, 0)
g n = (2 * n, 2 `div` n)
{-# NOINLINE g #-}
in case m of
1 -> 0
2 -> snd (g m)
_ -> uncurry (+) (g m)
```
Without the interleaved representation, we would just get `L` for the
strictness demand on `g`. Now we are able to express that whenever
`g` is called, its second component is used strictly in denoting `g`
by `1C1(P(1P(U),SP(U)))`. This would allow Nested CPR to unbox the
division, for example.
Fixes #18903.
While fixing regressions, I also discovered and fixed #18957.
Metric Decrease:
T13253-spj
|
|
|
|
|
|
|
| |
See Note [Exciting arity] why we emit the warning at all and why we only
do after the second iteration now.
Fixes #18937.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
As we found out in #18870, `andArityType` is not monotone, with
potentially severe consequences for termination of fixed-point
iteration. That showed in an abundance of "Exciting arity" DEBUG
messages that are emitted whenever we do more than one step in
fixed-point iteration.
The solution necessitates also recording `OneShotInfo` info for
`ABot` arity type. Thus we get the following definition for `ArityType`:
```
data ArityType = AT [OneShotInfo] Divergence
```
The majority of changes in this patch are the result of refactoring use
sites of `ArityType` to match the new definition.
The regression test `T18870` asserts that we indeed don't emit any DEBUG
output anymore for a function where we previously would have.
Similarly, there's a regression test `T18937` for #18937, which we
expect to be broken for now.
Fixes #18870.
|
|
|
|
|
| |
We didn't seem to test these old tests at all, judging from their
expected output.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In #18793, we saw a compelling example which requires us to look at
non-recursive let-bindings during arity analysis and unleash their arity
types at use sites.
After the refactoring in the previous patch, the needed change is quite
simple and very local to `arityType`'s defn for non-recurisve `Let`.
Apart from that, we had to get rid of the second item of
`Note [Dealing with bottoms]`, which was entirely a safety measure and
hindered optimistic fixed-point iteration.
Fixes #18793.
The following metric increases are all caused by this commit and a
result of the fact that we just do more work now:
Metric Increase:
T3294
T12545
T12707
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Summary:
My understanding is that ghc 7.10 should be buildable with the last 3 versions
of ghc, i.e 7.6, 7.8 and 7.10 itself.
Test Plan: x
Reviewers: austin
Reviewed By: austin
Subscribers: hvr, simonmar, ezyang, carter, thomie
Differential Revision: https://phabricator.haskell.org/D254
|
|
|