| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
It turns out that some important native debugging/profiling tools (e.g.
perf) rely only on symbol tables for function name resolution (as
opposed to using DWARF DIEs). However, previously GHC would emit
temporary symbols (e.g. `.La42b`) to identify module-internal
entities. Such symbols are dropped during linking and therefore not
visible to runtime tools (in addition to having rather un-helpful unique
names). For instance, `perf report` would often end up attributing all
cost to the libc `frame_dummy` symbol since Haskell code was no covered
by any proper symbol (see #17605).
We now rather follow the model of C compilers and emit
descriptively-named local symbols for module internal things. Since this
will increase object file size this behavior can be disabled with the
`-fno-expose-all-symbols` flag.
With this `perf record` can finally be used against Haskell executables.
Even more, with `-g3` `perf annotate` provides inline source code.
|
| |
|
|
|
|
|
| |
Currently we're passing a indexWord8OffAddr# type function to
utf8DecodeLazy# which then passes it on to utf8DecodeChar#. By passing one
of utf8DecodeCharAddr# or utf8DecodeCharByteArray# instead we benefit from
the inlining and specialization already done for those.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
There are multiple reasons we want this:
- Fewer allocations: ByteString has 3 fields, ShortByteString just has one.
- ByteString memory is pinned:
- This can cause fragmentation issues (see for example #13110) but also
- makes using FastStrings in compact regions impossible.
Metric Decrease:
T5837
T12150
T12234
T12425
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is consistent with the interface file behavior where we omit
HsNoCafRefs annotations with -fomit-interface-pragmas (implied by -O0).
ModDetails and ModIface are just different representations of the same
thing, so they really need to be in sync. This patch does the right
thing and does not need too much explanation, but here's an example of a
problem not doing this causes in !2842:
-- MyInteger.hs
module MyInteger
( MyInteger (MyInteger)
, ToMyInteger (toMyInteger)
) where
newtype MyInteger = MyInteger Integer
class ToMyInteger a where
toMyInteger :: a -> MyInteger
instance ToMyInteger Integer where
toMyInteger = MyInteger {- . succ -}
-- Main.hs
module Main
( main
) where
import MyInteger (MyInteger (MyInteger), toMyInteger)
main :: IO ()
main = do
let (MyInteger i) = (id . toMyInteger) (41 :: Integer)
print i
If I build this with -O0, without this fix, we generate a ModDetails with
accurate LFInfo for toMyInteger (MyInteger.$fToMyIntegerInteger) which says that
it's a LFReEntrant with arity 1. This means in the use site (Main) we tag the
value:
R3 = MyInteger.$fToMyIntegerInteger_closure + 1;
R2 = GHC.Base.id_closure;
R1 = GHC.Base.._closure;
Sp = Sp - 16;
call stg_ap_ppp_fast(R4, R3, R2, R1) args: 24, res: 0, upd: 24;
Now we change the definition by uncommenting the `succ` part and it becomes a thunk:
MyInteger.$fToMyIntegerInteger [InlPrag=INLINE (sat-args=0)]
:: MyInteger.ToMyInteger GHC.Integer.Type.Integer
[GblId[DFunId(nt)]] =
{} \u [] $ctoMyInteger_rEA;
and its LFInfo is now LFThunk. This change in LFInfo makes a difference in the
use site: we can no longer tag it.
But becuase the interface fingerprint does not change (because ModIface does not
change) we don't rebuild Main and tag the thunk.
(1.2% increase in allocations when building T12545 on armv7 because we
generate more code without CafInfos)
Metric Increase:
T12545
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch makes mapType use the successful idiom described
in TyCoRep
Note [Specialising foldType]
I have not yet changed any functions to use mapType, though there
may be some suitable candidates.
This patch should be a no-op in terms of functionality but,
because it inlines the mapper itself, I'm hoping that there may
be some modest perf improvements.
Metric Decrease:
T5631
T5642
T3064
T9020
T14683
hie002
haddock.Cabal
haddock.base
haddock.compiler
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Right now, in the output of -ddump-timings to a file, you can't tell what the units are:
```
CodeGen [TemplateTestImports]: alloc=22454880 time=14.597
```
I believe bytes/milliseconds are the correct units, but confirmation would be appreciated. I'm basing it off of this snippet from `withTiming'`:
```
when (verbosity dflags >= 2 && prtimings == PrintTimings)
$ liftIO $ logInfo dflags (defaultUserStyle dflags)
(text "!!!" <+> what <> colon <+> text "finished in"
<+> doublePrec 2 time
<+> text "milliseconds"
<> comma
<+> text "allocated"
<+> doublePrec 3 (realToFrac alloc / 1024 / 1024)
<+> text "megabytes")
```
which implies time is in milliseconds, and allocations in bytes (which divided by 1024 would be KB, and again would be MB)
|
| |
|
|
|
| |
The test operator "==" is a Bash extension and produces a wrong result
if /bin/sh is not Bash.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Remove an invalid assumption about GC checking what_next field. The GC
doesn't care about what_next at all, if a TSO is reachable then all
its pointers are followed (other than global_tso, which is only
followed by compacting GC).
- Remove checkSTACK in checkTSO: TSO stacks will be visited in
checkHeapChain, or checkLargeObjects etc.
- Add an assertion in checkTSO to check that the global_link field is
sane.
- Did some refactor to remove forward decls in checkGlobalTSOList and
added braces around single-statement if statements.
|
| |
|
|
| |
Update submodule: haddock
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Fixes #17785
Here's how the problem occurs:
- In generation 0 we have a TSO that is finished (i.e. it has no more
work to do or it is killed).
- The TSO only becomes reachable after collectDeadWeakPtrs().
- After collectDeadWeakPtrs() we switch to WeakDone phase where we don't
move TSOs to different lists anymore (like the next gen's thread list
or the resurrected_threads list).
- So the TSO will never be moved to a generation's thread list, but it
will be promoted to generation 1.
- Generation 1 collected via mark-compact, and because the TSO is
reachable it is marked, and its `global_link` field, which is bogus at
this point (because the TSO is not in a list), will be threaded.
- Chaos ensues.
In other words, when these conditions hold:
- A TSO is reachable only after collectDeadWeakPtrs()
- It's finished (what_next is ThreadComplete or ThreadKilled)
- It's retained by mark-compact collector (moving collector doesn't
evacuate the global_list field)
We end up doing random mutations on the heap because the TSO's
global_list field is not valid, but it still looks like a heap pointer
so we thread it during compacting GC.
The fix is simple: when we traverse old_threads lists to resurrect
unreachable threads the threads that won't be resurrected currently
stays on the old_threads lists. Those threads will never be visited
again by MarkWeak so we now reset the global_list fields. This way
compacting GC does not thread pointers to nowhere.
Testing
-------
The reproducer in #17785 is quite large and hard to build, because of
the dependencies, so I'm not adding a regression test.
In my testing the reproducer would take a less than 5 seconds to run,
and once in every ~5 runs would fail with a segfault or an assertion
error. In other cases it also fails with a test failure. Because the
tests never fail with the bug fix, assuming the code is correct, this
also means that this bug can sometimes lead to incorrect runtime
results.
After the fix I was able to run the reproducer repeatedly for about an
hour, with no runtime crashes or test failures.
To run the reproducer clone the git repo:
$ git clone https://github.com/osa1/streamly --branch ghc-segfault
Then clone primitive and atomic-primops from their git repos and point
to the clones in cabal.project.local. The project should then be
buildable using GHC HEAD. Run the executable `properties` with `+RTS -c
-DZ`.
In addition to the reproducer above I run the test suite using:
$ make slowtest EXTRA_HC_OPTS="-debug -with-rtsopts=-DS \
-with-rtsopts=-c +RTS -c -RTS" SKIPWAY='nonmoving nonmoving_thr'
This enables compacting GC always in both GHC when building the test
programs and when running the test programs, and also enables sanity
checking when running the test programs. These set of flags are not
compatible for all tests so there are some failures, but I got the same
set of failures with this patch compared to GHC HEAD.
|
| |
|
|
|
|
|
|
| |
This flag undoes the effect of a previous "-haddock" flag. Having both flags makes it easier
for build systems to enable Haddock parsing in a set of global flags, but then disable it locally for
specific targets (e.g., third-party packages whose comments don't pass the validation in the latest GHC).
I added the flag to expected-undocumented-flags.txt since `-haddock` was alreadyin that list.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch disentangles a bit more DynFlags from the native code
generator (CmmToAsm).
In more details:
- add a new NCGConfig datatype in GHC.CmmToAsm.Config which contains the
configuration of a native code generation session
- explicitly pass NCGConfig/Platform arguments when necessary
- as a consequence `sdocWithPlatform` is gone and there are only a few
`sdocWithDynFlags` left
- remove the use of `unsafeGlobalDynFlags` from GHC.CmmToAsm.CFG
- remove `sdocDebugLevel` (now we pass the debug level via NCGConfig)
There are still some places where DynFlags is used, especially because
of pretty-printing (CLabel), because of Cmm helpers (such as
`cmmExprType`) and because of `Outputable` instance for the
instructions. These are left for future refactoring as this patch is
already big.
|
| |
|
|
|
| |
Also fix the markup in the general note at the top of the module. Haddock
(usability trade-off), does not support multi-line emphasised text.
|
| |
|
|
| |
From the notes.ghc.drop list found using weeder in #17713
|
| |
|
|
|
|
|
| |
Currently, the names of cost centres must be quoted or
be lowercase identifiers.
Fixes #17916.
|
| | |
|
| | |
|
| |
|
|
| |
The submodule has already been bumped to contain the fix.
|
| |
|
|
|
|
|
|
|
| |
Before this patch, tcDataDefn used to call tcLookupTcTyCon twice in a row:
1. in bindTyClTyVars itself
2. in the continuation passed to it
Now bindTyClTyVars passes the TcTyCon to the continuation, making
the second lookup unnecessary.
|
| |
|
|
|
|
|
| |
As noted in #17912, `open` system calls were `safe` rather than
`interruptible`. Consequently, the program could not be interrupted with
SIGINT if stuck in a slow open operation. Fix this by marking
`c_safe_open` as interruptible.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch is just refactoring: no change in
behaviour.
I removed the rather complicated
checkConstraintsTcS
checkTvConstraintsTcS
in favour of simpler functions
emitImplicationTcS
emitTvImplicationTcS
pushLevelNoWorkList
The last of these is a little strange, but overall
it's much better I think.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Ticket #17841 showed that we can get a kind error
in a class signature, but lack an enclosing implication
that binds its skolems.
This patch
* Adds the wrapping implication: the new call to
checkTvConstraints in tcClassDecl1
* Simplifies the API to checkTvConstraints, which
was not otherwise called at all.
* Simplifies TcErrors.report_unsolved by *not*
initialising the TidyEnv from the typechecker lexical
envt. It's enough to do so from the free vars of the
unsolved constraints; and we get silly renamings if
we add variables twice: once from the lexical scope
and once from the implication constraint.
|
| |
|
|
| |
Closes #16144.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
This patch improves the way that CSE combines identical
alternatives. See #17901.
I'm still not happy about the duplication between CSE.combineAlts
and GHC.Core.Utils.combineIdenticalAlts; see the Notes with those
functions. But this patch is a step forward.
Metric Decrease:
T12425
T5642
|
| |
|
|
|
| |
No change in functionality. Just seems tidier (and signficantly more
efficient) to deal with ticks directly than to call stripTicksTopE.
|
| |
|
|
|
| |
nonmovingSweep already clears the bitmap in the sweep loop. There is no
reason to do so a second time.
|
| |
|
|
|
|
|
|
|
| |
The non-moving collector would previously walk the entire filled segment
list during the preparatory pause. However, this is far more work than
is strictly necessary. We can rather get away with merely collecting the
allocators' filled segment list heads and process the lists themselves
during the concurrent phase. This can significantly reduce the maximum
gen1 GC pause time in programs with high rates of long-lived allocations.
|
| | |
|
| |
|
|
|
|
|
|
|
| |
Previously it was possible to override the stage0 C compiler via `CC_STAGE0`,
but you couldn't override `ld` or `ar` in stage0. This change allows overriding them
by setting `LD_STAGE0` or `AR_STAGE0`, respectively.
Our team uses this feature internally to take more control of our GHC build
and make it run more hermetically.
|
| |
|
|
| |
s/pgmo/opti
|
| |
|
|
| |
Document and use simpler rules for the ghc-gmp.h header.
|
| |
|
|
| |
Shake's "**" wildcard doesn't match absolute root. We must use "//" instead.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
Ticket #17590 pointed out a bug in the way the linter dealt with
type lets, exposed by the new uniqAway story.
The fix is described in Note [Linting type lets]. I ended up
putting the in-scope Ids in a different env field, le_ids,
rather than (as before) sneaking them into the TCvSubst.
Surprisingly tiresome, but done.
Metric Decrease:
hie002
|
| |
|
| |
Previously two integer-simple jobs declared the same test environment. One (the nightly job) was built in the perf way, the other in the validate way. Consequently they had appreciably different performance characteristics, causing in the nightly job to spuriously fail with performance changes.
|
| |
|
|
| |
Use a push option instead of tagging.
|
| | |
|
| |
|
|
|
|
| |
I wanted to fix the dangling comment in `isDllName` ("This is the cause
of #", #8696 is already mentioned earlier). I took the opportunity to
change the function name to better reflect what it does.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The code in Rules.Register responsible for finding all the build artifacts
that Cabal installs when registering a library (static/shared libs, .hi files,
...) was looking in the wrong place. This patch fixes that logic and makes sure
we gather all those artifacts in a list to declare that the rule for a given
`.conf` file, our proxy for "Hadrian, please install this package in the package
db for this stage", also produces those artifacts under the said package
database.
We also were completely missing some logic to declare that the check-* programs
have dependencies besides their source code, at least when testing an in-tree
compiler.
Finally, this patch also removes redundant packages from 'testsuitePackages',
since they should already be covered by the stage<N>Packages lists from
Settings.Default.
With this patch, after a complete build and freezing stage 1, a change to
`compiler/parser/Parser.y` results in rebuilding the ghc lib, reinstalling it,
and rebuilding the few programs that depend on it, _including_ `check-ppr` and
`check-api-annotations` (therefore fixing #17273).
|
| |
|
|
|
|
|
|
|
|
|
|
| |
The unsafeCoerce# patch requires that unsafeCoerce# has
a compulsory unfolding that is always available. So we have
to be careful to expose compulsory unfoldings unconditionally
and consistently.
We didn't get this quite right: #17871. This patch fixes
it. No real surprises here.
See Note [Always expose compulsory unfoldings] in GHC.Iface.Tidy
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* Do not print `join` in ApplictiveStmt, unless ppr-debug
* Print parens around multiple parallel binds
When ApplicativeDo is enabled, the renamer analyses the statements of a
`do` block and in certain cases marks them as needing to be rewritten
using `join`.
For example, if you have:
```
foo = do
a <- e1
b <- e2
doSomething a b
```
it will be desugared into:
```
foo = join (doSomething <$> e1 <*> e2)
```
After renaming but before desugaring the expression is stored
essentially as:
```
foo = do
[will need join] (a <- e1 | b <- e2)
[no return] doSomething a b
```
Before this change, the pretty printer would print a call to `join`,
even though it is not needed at this stage at all. The expression will be
actually rewritten into one using join only at desugaring, at which
point a literal call to join will be inserted.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
After ApplicatveDo strips the last `return` during renaming, the pretty
printer has to restore it. However, if the return was followed by `$`,
the dollar was stripped too and not restored.
For example, the last stamement in:
```
foo = do
x <- ...
...
return $ f x
```
would be printed as:
```
return f x
```
This commit preserved the dolar, so it becomes:
```
return $ f x
```
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Issue #17880 demonstrates that `DeriveFunctor`-generated code is
surprisingly fragile when rank-_n_ types are involved. The culprit is
that `$fmap` (the algorithm used to generate `fmap` implementations)
was too keen on applying arguments with rank-_n_ types to lambdas,
which fail to typecheck more often than not.
In this patch, I change `$fmap` (both the specification and the
implementation) to produce code that avoids creating as many lambdas,
avoiding problems when rank-_n_ field types arise.
See the comments titled "Functor instances" in `TcGenFunctor` for a
more detailed description. Not only does this fix #17880, but it also
ensures that the code that `DeriveFunctor` generates will continue
to work after simplified subsumption is implemented (see #17775).
What is truly amazing is that #17880 is actually a regression
(introduced in GHC 7.6.3) caused by commit
49ca2a37bef18aa57235ff1dbbf1cc0434979b1e, the fix #7436. Prior to
that commit, the version of `$fmap` that was used was almost
identical to the one used in this patch! Why did that commit change
`$fmap` then? It was to avoid severe performance issues that would
arise for recursive `fmap` implementations, such as in the example
below:
```hs
data List a = Nil | Cons a (List a) deriving Functor
-- ===>
instance Functor List where
fmap f Nil = Nil
fmap f (Cons x xs) = Cons (f x) (fmap (\y -> f y) xs)
```
The fact that `\y -> f y` was eta expanded caused significant
performance overheads. Commit
49ca2a37bef18aa57235ff1dbbf1cc0434979b1e fixed this performance
issue, but it went too far. As a result, this patch partially
reverts 49ca2a37bef18aa57235ff1dbbf1cc0434979b1e.
To ensure that the performance issues pre-#7436 do not resurface,
I have taken some precautionary measures:
* I have added a special case to `$fmap` for situations where the
last type variable in an application of some type occurs directly.
If this special case fires, we avoid creating a lambda expression.
This ensures that we generate
`fmap f (Cons x xs) = Cons (f x) (fmap f xs)` in the derived
`Functor List` instance above. For more details, see
`Note [Avoid unnecessary eta expansion in derived fmap implementations]`
in `TcGenFunctor`.
* I have added a `T7436b` test case to ensure that the performance
of this derived `Functor List`-style code does not regress.
When implementing this, I discovered that `$replace`, the algorithm
which generates implementations of `(<$)`, has a special case that is
very similar to the `$fmap` special case described above. `$replace`
marked this special case with a custom `Replacer` data type, which
was a bit overkill. In order to use the same machinery for both
`Functor` methods, I ripped out `Replacer` and instead implemented
a simple way to detect the special case. See the updated commentary
in `Note [Deriving <$]` for more details.
|
| |
|
|
|
|
|
|
| |
Should make `member` queries faster and avoid messing up with missing
`nubSort`.
Metric Increase:
hie002
|
