[Git][ghc/ghc][wip/andreask/opt_cmm_sink_sets] 18 commits: Move core flattening algorithm to Core.Unify

Andreas Klebinger gitlab at gitlab.haskell.org
Tue Dec 8 11:51:16 UTC 2020 


Andreas Klebinger pushed to branch wip/andreask/opt_cmm_sink_sets at Glasgow Haskell Compiler / GHC


Commits:
72a87fbc by Richard Eisenberg at 2020-12-01T19:57:41-05:00
Move core flattening algorithm to Core.Unify

This sets the stage for a later change, where this
algorithm will be needed from GHC.Core.InstEnv.

This commit also splits GHC.Core.Map into
GHC.Core.Map.Type and GHC.Core.Map.Expr,
in order to avoid module import cycles
with GHC.Core.

- - - - -
0dd45d0a by Richard Eisenberg at 2020-12-01T19:57:41-05:00
Bump the # of commits searched for perf baseline

The previous value of 75 meant that a feature branch with
more than 75 commits would get spurious CI passes.

This affects #18692, but does not fix that ticket, because
if a baseline cannot be found, we should fail, not succeed.

- - - - -
8bb52d91 by Richard Eisenberg at 2020-12-01T19:57:41-05:00
Remove flattening variables

This patch redesigns the flattener to simplify type family applications
directly instead of using flattening meta-variables and skolems. The key new
innovation is the CanEqLHS type and the new CEqCan constraint (Ct). A CanEqLHS
is either a type variable or exactly-saturated type family application; either
can now be rewritten using a CEqCan constraint in the inert set.

Because the flattener no longer reduces all type family applications to
variables, there was some performance degradation if a lengthy type family
application is now flattened over and over (not making progress). To
compensate, this patch contains some extra optimizations in the flattener,
leading to a number of performance improvements.

Close #18875.
Close #18910.

There are many extra parts of the compiler that had to be affected in writing
this patch:

* The family-application cache (formerly the flat-cache) sometimes stores
  coercions built from Given inerts. When these inerts get kicked out, we must
  kick out from the cache as well. (This was, I believe, true previously, but
  somehow never caused trouble.) Kicking out from the cache requires adding a
  filterTM function to TrieMap.

* This patch obviates the need to distinguish "blocking" coercion holes from
  non-blocking ones (which, previously, arose from CFunEqCans). There is thus
  some simplification around coercion holes.

* Extra commentary throughout parts of the code I read through, to preserve
  the knowledge I gained while working.

* A change in the pure unifier around unifying skolems with other types.
  Unifying a skolem now leads to SurelyApart, not MaybeApart, as documented
  in Note [Binding when looking up instances] in GHC.Core.InstEnv.

* Some more use of MCoercion where appropriate.

* Previously, class-instance lookup automatically noticed that e.g. C Int was
  a "unifier" to a target [W] C (F Bool), because the F Bool was flattened to
  a variable. Now, a little more care must be taken around checking for
  unifying instances.

* Previously, tcSplitTyConApp_maybe would split (Eq a => a). This is silly,
  because (=>) is not a tycon in Haskell. Fixed now, but there are some
  knock-on changes in e.g. TrieMap code and in the canonicaliser.

* New function anyFreeVarsOf{Type,Co} to check whether a free variable
  satisfies a certain predicate.

* Type synonyms now remember whether or not they are "forgetful"; a forgetful
  synonym drops at least one argument. This is useful when flattening; see
  flattenView.

* The pattern-match completeness checker invokes the solver. This invocation
  might need to look through newtypes when checking representational equality.
  Thus, the desugarer needs to keep track of the in-scope variables to know
  what newtype constructors are in scope. I bet this bug was around before but
  never noticed.

* Extra-constraints wildcards are no longer simplified before printing.
  See Note [Do not simplify ConstraintHoles] in GHC.Tc.Solver.

* Whether or not there are Given equalities has become slightly subtler.
  See the new HasGivenEqs datatype.

* Note [Type variable cycles in Givens] in GHC.Tc.Solver.Canonical
  explains a significant new wrinkle in the new approach.

* See Note [What might match later?] in GHC.Tc.Solver.Interact, which
  explains the fix to #18910.

* The inert_count field of InertCans wasn't actually used, so I removed
  it.

Though I (Richard) did the implementation, Simon PJ was very involved
in design and review.

This updates the Haddock submodule to avoid #18932 by adding
a type signature.

-------------------------
Metric Decrease:
    T12227
    T5030
    T9872a
    T9872b
    T9872c
Metric Increase:
    T9872d
-------------------------

- - - - -
d66660ba by Richard Eisenberg at 2020-12-01T19:57:41-05:00
Rename the flattener to become the rewriter.

Now that flattening doesn't produce flattening variables,
it's not really flattening anything: it's rewriting. This
change also means that the rewriter can no longer be confused
the core flattener (in GHC.Core.Unify), which is sometimes used
during type-checking.

- - - - -
add0aeae by Ben Gamari at 2020-12-01T19:58:17-05:00
rts: Introduce mmapAnonForLinker

Previously most of the uses of mmapForLinker were mapping anonymous
memory, resulting in a great deal of unnecessary repetition. Factor this
out into a new helper.

Also fixes a few places where error checking was missing or suboptimal.

- - - - -
97d71646 by Ben Gamari at 2020-12-01T19:58:17-05:00
rts/linker: Introduce munmapForLinker

Consolidates munmap calls to ensure consistent error handling.

- - - - -
d8872af0 by Ben Gamari at 2020-12-01T19:58:18-05:00
rts/Linker: Introduce Windows implementations for mmapForLinker, et al.

- - - - -
c35d0e03 by Ben Gamari at 2020-12-01T19:58:18-05:00
rts/m32: Introduce NEEDS_M32 macro

Instead of relying on RTS_LINKER_USE_MMAP

- - - - -
41c64eb5 by Ben Gamari at 2020-12-01T19:58:18-05:00
rts/linker: Use m32 to allocate symbol extras in PEi386

- - - - -
e0b08c5f by Ben Gamari at 2020-12-03T13:01:47-05:00
gitlab-ci: Fix copy-paste error

Also be more consistent in quoting.

- - - - -
33ec3a06 by Ben Gamari at 2020-12-03T23:11:31-05:00
gitlab-ci: Run linters through ci.sh

Ensuring that the right toolchain is used.

- - - - -
4a437bc1 by Shayne Fletcher at 2020-12-05T09:06:38-05:00
Fix bad span calculations of post qualified imports

- - - - -
8fac4b93 by Ben Gamari at 2020-12-05T09:07:13-05:00
testsuite: Add a test for #18923

- - - - -
2f693c82 by Andreas Klebinger at 2020-12-08T12:50:57+01:00
CodeGen: Make folds User/DefinerOfRegs INLINEABLE.

Reduces allocation for the test case I was looking at by about 1.2%.
Mostly from avoiding allocation of some folding functions which turn
into let-no-escape bindings which just reuse their environment instead.

We also force inlining in a few key places in CmmSink which helps a bit
more.

- - - - -
8ff748af by Andreas Klebinger at 2020-12-08T12:50:57+01:00
CmmSink: Force inlining of foldRegsDefd

Helps avoid allocating the folding function. Improves
perf for T3294 by about 1%.

- - - - -
0b8b8a8d by Andreas Klebinger at 2020-12-08T12:50:57+01:00
Cmm: Make a few types and utility function slightly stricter.

About 0.6% reduction in allocations for the code I was looking at.

Not a huge difference but no need to throw away performance.

- - - - -
1c841f9c by Andreas Klebinger at 2020-12-08T12:50:57+01:00
Cmm.Sink: Optimize retaining of assignments, live sets.

Sinking requires us to track live local regs after each
cmm statement. We used to do this via "Set LocalReg".

However we can replace this with a solution based on IntSet
which is overall more efficient without losing much. The thing
we lose is width of the variables, which isn't used by the sinking
pass anyway.

I also reworked how we keep assignments to regs mentioned in
skipped assignments. I put the details into
Note [Keeping assignemnts mentioned in skipped RHSs].

The gist of it is instead of keeping track of it via the use count
which is a `IntMap Int` we now use the live regs set (IntSet) which
is quite a bit faster.

I think it also matches the semantics a lot better. The skipped
(not discarded) assignment does in fact keep the regs on it's rhs
alive so keeping track of this in the live set seems like the clearer
solution as well.

Improves allocations for T3294 by yet another 1%.

- - - - -
acc75450 by Andreas Klebinger at 2020-12-08T12:50:57+01:00
GHC.Cmm.Opt: Be stricter in results.

Optimization either returns Nothing if nothing is to be done or
`Just <cmmExpr>` otherwise. There is no point in being lazy in
`cmmExpr`. We usually inspect this element so the thunk gets forced
not long after.

We might eliminate it as dead code once in a blue moon but that's
not a case worth optimizing for.

Overall the impact of this is rather low. As Cmm.Opt doesn't allocate
much (compared to the rest of GHC) to begin with.

- - - - -


25 changed files:

- .gitlab-ci.yml
- .gitlab/ci.sh
- compiler/GHC/Cmm/CommonBlockElim.hs
- compiler/GHC/Cmm/Dataflow/Label.hs
- compiler/GHC/Cmm/Expr.hs
- + compiler/GHC/Cmm/LRegSet.hs
- compiler/GHC/Cmm/Liveness.hs
- compiler/GHC/Cmm/Node.hs
- compiler/GHC/Cmm/Opt.hs
- compiler/GHC/Cmm/Sink.hs
- compiler/GHC/Cmm/Utils.hs
- compiler/GHC/Core/Coercion.hs
- compiler/GHC/Core/Coercion.hs-boot
- compiler/GHC/Core/Coercion/Axiom.hs
- compiler/GHC/Core/Coercion/Opt.hs
- compiler/GHC/Core/FamInstEnv.hs
- compiler/GHC/Core/InstEnv.hs
- + compiler/GHC/Core/Map/Expr.hs
- compiler/GHC/Core/Map.hs → compiler/GHC/Core/Map/Type.hs
- compiler/GHC/Core/Opt/CSE.hs
- compiler/GHC/Core/TyCo/FVs.hs
- compiler/GHC/Core/TyCo/Rep.hs
- compiler/GHC/Core/TyCon.hs
- compiler/GHC/Core/TyCon/Env.hs
- compiler/GHC/Core/Type.hs


The diff was not included because it is too large.


View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/23e1050e1a9abaa58e2216143c5165efec9ff31c...acc7545008639b9fd23bc4e55217d6d3e4b00c96

-- 
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/23e1050e1a9abaa58e2216143c5165efec9ff31c...acc7545008639b9fd23bc4e55217d6d3e4b00c96
You're receiving this email because of your account on gitlab.haskell.org.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-commits/attachments/20201208/900edd34/attachment.html>

More information about the ghc-commits mailing list