- Feb 08, 2024
-
-
Here we move a good deal of the implementation of `base` into a new package, `ghc-internal` such that it can be evolved independently from the user-visible interfaces of `base`. While we want to isolate implementation from interfaces, naturally, we would like to avoid turning `base` into a mere set of module re-exports. However, this is a non-trivial undertaking for a variety of reasons: * `base` contains numerous known-key and wired-in things, requiring corresponding changes in the compiler * `base` contains a significant amount of C code and corresponding autoconf logic, which is very fragile and difficult to break apart * `base` has numerous import cycles, which are currently dealt with via carefully balanced `hs-boot` files * We must not break existing users To accomplish this migration, I tried the following approaches: * [Split-GHC.Base]: Break apart the GHC.Base knot to allow incremental migration of modules into ghc-internal: this knot is simply too intertwined to be easily pulled apart, especially given the rather tricky import cycles that it contains) * [Move-Core]: Moving the "core" connected component of base (roughly 150 modules) into ghc-internal. While the Haskell side of this seems tractable, the C dependencies are very subtle to break apart. * [Move-Incrementally]: 1. Move all of base into ghc-internal 2. Examine the module structure and begin moving obvious modules (e.g. leaves of the import graph) back into base 3. Examine the modules remaining in ghc-internal, refactor as necessary to facilitate further moves 4. Go to (2) iterate until the cost/benefit of further moves is insufficient to justify continuing 5. Rename the modules moved into ghc-internal to ensure that they don't overlap with those in base 6. For each module moved into ghc-internal, add a shim module to base with the declarations which should be exposed and any requisite Haddocks (thus guaranteeing that base will be insulated from changes in the export lists of modules in ghc-internal Here I am using the [Move-Incrementally] approach, which is empirically the least painful of the unpleasant options above Bumps haddock submodule. Metric Decrease: haddock.Cabal haddock.base Metric Increase: MultiComponentModulesRecomp T16875 size_hello_artifact -
This logic appears to be incorrect as it would drop any dependency which was not in a direct dependency of the package being linked. In the ghc-internals split this started to cause errors because `ghc-internal` is not a direct dependency of most packages, and hence important symbols to keep which are hard coded into the js runtime were getting dropped.
-
-
- Feb 06, 2024
-
-
-
This fixes a bug in 8db8d2fd, where we could lose track of acyclic components at the start of an unresolved cycle. We now ensure we never loose track of any of these components. As T24275 demonstrates, a "cyclic" SCC might not really be a true SCC: When viewed without boot files, we have a single SCC ``` [REC main:T24275B [main:T24275B {-# SOURCE #-}, main:T24275A {-# SOURCE #-}] main:T24275A [main:T24275A {-# SOURCE #-}]] ``` But with boot files this turns into ``` [NONREC main:T24275B {-# SOURCE #-} [], REC main:T24275B [main:T24275B {-# SOURCE #-}, main:T24275A {-# SOURCE #-}] main:T24275A {-# SOURCE #-} [main:T24275B], NONREC main:T24275A [main:T24275A {-# SOURCE #-}]] ``` Note that this is truly not an SCC, as no nodes are reachable from T24275B.hs-boot. However, we treat this entire group as a single "SCC" because it seems so when we analyse the graph without taking boot files into account. Indeed, we must return a single ResolvedCycle element in the BuildPlan for this as described in Note [Upsweep]. However, since after resolving this is not a true SCC anymore, `findCycle` fails to find a cycle and we have a sub-optimal error message as a result. To handle this, I extended `findCycle` to not assume its input is an SCC, and to try harder to find cycles in its input. Fixes #24275
-
This patch is a preparation for @a-binders implementation. The main changes are: * Skolemisation is now prepared to deal with @binders. See Note [Skolemisation overview] in GHC.Tc.Utils.Unify. Most of the action is in - Utils.Unify.matchExpectedFunTys - Gen.Pat.tcMatchPats - Gen.Expr.tcPolyExprCheck - Gen.Binds.tcPolyCheck Some accompanying refactoring: * I found that funTyConAppTy_maybe was doing a lot of allocation, and rejigged userTypeError_maybe to avoid calling it. -
* Make HsMatchContext and HsStmtContext be parameterised over the function name itself, rather than over the pass. See [mc_fun field of FunRhs] in Language.Haskell.Syntax.Expr - Replace types HsMatchContext GhcPs --> HsMatchContextPs HsMatchContext GhcRn --> HsMatchContextRn HsMatchContext GhcTc --> HsMatchContextRn (sic! not Tc) HsStmtContext GhcRn --> HsStmtContextRn - Kill off convertHsMatchCtxt * Split GHC.Tc.Type.BasicTypes.TcSigInfo so that TcCompleteSig (describing a complete user-supplied signature) is its own data type. - Split TcIdSigInfo(CompleteSig, PartialSig) into TcCompleteSig(CSig) TcPartialSig(PSig) - Use TcCompleteSig in tcPolyCheck, CheckGen - Rename types and data constructors: TcIdSigInfo --> TcIdSig TcPatSynInfo(TPSI) --> TcPatSynSig(PatSig) - Shuffle around helper functions: tcSigInfoName (moved to GHC.Tc.Types.BasicTypes) completeSigPolyId_maybe (moved to GHC.Tc.Types.BasicTypes) tcIdSigName (inlined and removed) tcIdSigLoc (introduced) - Rearrange the pattern match in chooseInferredQuantifiers * Rename functions and types: tcMatchesCase --> tcCaseMatches tcMatchesFun --> tcFunBindMatches tcMatchLambda --> tcLambdaMatches tcPats --> tcMatchPats matchActualFunTysRho --> matchActualFunTys matchActualFunTySigma --> matchActualFunTy * Add HasDebugCallStack constraints to: mkBigCoreVarTupTy, mkBigCoreTupTy, boxTy, mkPiTy, mkPiTys, splitAppTys, splitTyConAppNoView_maybe * Use `penv` from the outer context in the inner loop of GHC.Tc.Gen.Pat.tcMultiple * Move tcMkVisFunTy, tcMkInvisFunTy, tcMkScaledFunTys down the file, factor out and export tcMkScaledFunTy. * Move isPatSigCtxt down the file. * Formatting and comments Co-authored-by:Vladislav Zavialov <vlad.z.4096@gmail.com>
-
- Feb 05, 2024
-
-
-
This MR fixes #24251. See Note [Case-to-let for strictly-used binders] in GHC.Core.Opt.Simplify.Iteration, plus #24251, for lots of discussion. Final Nofib changes over 0.1%: +----------------------------------------- | imaginary/digits-of-e2 -2.16% | imaginary/rfib -0.15% | real/fluid -0.10% | real/gamteb -1.47% | real/gg -0.20% | real/maillist +0.19% | real/pic -0.23% | real/scs -0.43% | shootout/n-body -0.41% | shootout/spectral-norm -0.12% +======================================== | geom mean -0.05% Pleasingly, overall executable size is down by just over 1%. Compile times (in perf/compiler) wobble around a bit +/- 0.5%, but the geometric mean is -0.1% which seems good.
-
- Feb 04, 2024
-
-
In the most common case, the comment queue is empty, so we can skip the work of processing it. This reduces allocations by about 10% in the parsing001 test. Metric Decrease: MultiLayerModulesRecomp parsing001
-
- Feb 03, 2024
-
-
Apoorv Ingle authored- Fixes #18324 #20020 #23147 #22788 #15598 #22086 #21206 - The change is detailed in - Note [Expanding HsDo with HsExpansion] in `GHC.Tc.Gen.Do` - Note [Doing HsExpansion in the Renamer vs Typechecker] in `GHC.Rename.Expr` expains the rational of doing expansions in type checker as opposed to in the renamer - Adds new datatypes: - `GHC.Hs.Expr.XXExprGhcRn`: new datatype makes this expansion work easier 1. Expansion bits for Expressions, Statements and Patterns in (`ExpandedThingRn`) 2. `PopErrCtxt` a special GhcRn Phase only artifcat to pop the previous error message in the error context stack - `GHC.Basic.Origin` now tracks the reason for expansion in case of Generated This is useful for type checking cf. `GHC.Tc.Gen.Expr.tcExpr` case for `HsLam` - Kills `HsExpansion` and `HsExpanded` as we have inlined them in `XXExprGhcRn` and `XXExprGhcTc` - Ensures warnings such as 1. Pattern match checks 2. Failable patterns 3. non-() return in body statements are preserved - Kill `HsMatchCtxt` in favor of `TcMatchAltChecker` - Testcases: * T18324 T20020 T23147 T22788 T15598 T22086 * T23147b (error message check), * DoubleMatch (match inside a match for pmc check) * pattern-fails (check pattern match with non-refutable pattern, eg. newtype) * Simple-rec (rec statements inside do statment) * T22788 (code snippet from #22788) * DoExpanion1 (Error messages for body statments) * DoExpansion2 (Error messages for bind statements) * DoExpansion3 (Error messages for let statements) Also repoint haddock to the right submodule so that the test (haddockHypsrcTest) pass Metric Increase 'compile_time/bytes allocated': T9020 The testcase is a pathalogical example of a `do`-block with many statements that do nothing. Given that we are expanding the statements into function binds, we will have to bear a (small) 2% cost upfront in the compiler to unroll the statements.
-
- Feb 01, 2024
-
-
New syntax for WARNING and DEPRECATED pragmas was added, namely namespace specifierss: namespace_spec ::= 'type' | 'data' | {- empty -} warning ::= warning_category namespace_spec namelist strings deprecation ::= namespace_spec namelist strings A new data type was introduced to represent these namespace specifiers: data NamespaceSpecifier = NoSpecifier | TypeNamespaceSpecifier (EpToken "type") | DataNamespaceSpecifier (EpToken "data") Extension field XWarning now contains this NamespaceSpecifier. lookupBindGroupOcc function was changed: it now takes NamespaceSpecifier and checks that the namespace of the found names matches the passed flag. With this change {-# WARNING data D "..." #-} pragma will only affect value namespace and {-# WARNING type D "..." #-} will only affect type namespace. The same logic is applicable to DEPRECATED pragmas. Finding duplicated warnings inside rnSrcWarnDecls now takes into consideration NamespaceSpecifier flag to allow warnings with the same names that refer to different namespaces. -
This commit enables the generic cmm optimizations in other NCGs to be run in the wasm backend as well, followed by a late cmm control-flow optimization pass. The added optimizations do catch some corner cases not handled by the pre-NCG cmm pipeline and are useful in generating smaller CFGs.
-
This commit explicitly disables the ncgPIC flag for the wasm32 target. The wasm backend doesn't support PIC for the time being.
-
This commit moves GHC.CmmToAsm.cmmToCmm to a standalone module, GHC.Cmm.GenericOpt. The main motivation is enabling this logic to be run in the wasm backend NCG code, which is defined in other modules that's imported by GHC.CmmToAsm, causing a cyclic dependency issue.
-
-
-
In hand-written Cmm it can sometimes be necessary to atomically load from memory deep within an expression (e.g. see the `CHECK_GC` macro). This MachOp provides a convenient way to do so without breaking the expression into multiple statements.
-
- Jan 31, 2024
-
-
Simon Peyton Jones authored
This fixes #24370 by making decomposeRuleLhs undertand dictionary /functions/ as well as plain /dictionaries/
-
A test *сс018* is attached (not sure about the naming convention though). Note that without the fix, the test fails with the *dodgy-foreign-imports* warning passed to stderr. The warning disappears after the fix. GHC shouldn't warn on imports of natural function pointers from C by value (which is feasible with CApiFFI), such as ```haskell foreign import capi "cc018.h value f" f :: FunPtr (Int -> IO ()) ``` where ```c void (*f)(int); ``` See a related real-world use-case [here](https://gitlab.com/daniel-casanueva/pcre-light/-/merge_requests/17). There, GHC warns on import of C function pointer `pcre_free`.
-
- Jan 29, 2024
-
-
-
-
On aarch64 fma is always on so the +fma flag doesn't exist for that target. Hence no need to try and pass +fma to llvm. Fixes #24379
-
- Jan 26, 2024
-
-
- Jan 25, 2024
-
-
-
The GHC.Linker module is empty and unused, other than as a hack for the make build system. We can remove it now that make is long gone; the note is moved to GHC.Linker.Loader instead.
-
- Jan 24, 2024
-
-
We were emitting the wrong instructions for fused multiply-add operations on LLVM: - the instruction name is "llvm.fma.f32" or "llvm.fma.f64", not "fmadd" - LLVM does not support other instructions such as "fmsub"; instead we implement these by flipping signs of some arguments - the instruction is an LLVM intrinsic, which requires handling it like a normal function call instead of a machine instruction Fixes #24223 -
They don't make it through readFile' without explicitly setting the encoding. See ghc/ghc#17755
-
The wasm backend didn't properly make use of all Cmm global registers due to #24347. Now that it is fixed, this patch re-enables full register mapping for wasm32, and we can now generate smaller & faster wasm modules that doesn't always spill arguments onto the stack. Fixes #22460 #24152.
-
- Jan 20, 2024
-
-
- Jan 19, 2024
-
-
Go through compiler/ sources, and remove all BangPatterns (and other GHC2021 enabled extensions in these files).
-
- Jan 18, 2024
-
-
These bugs are all described in #24335 It's not easy to provoke the bug, hence no test case.
-
- Jan 17, 2024
-
-
Each phase of the simplifier typically rewrites the majority of ModGuts, so we want to be able to release the old ModGuts as soon as possible. `name_ppr_ctxt` lives throught the whole optimiser phase and it was retaining a reference to `ModGuts`, so we were failing to release the old `ModGuts` until the end of the phase (potentially doubling peak memory usage for that particular phase). This was discovered using eras profiling (#24332) Fixes #24328
-
- Jan 16, 2024
-
-
This MR makes three improvements to SpecConstr: see #24282 * It fixes an outright (and recently-introduced) bug in `betterPat`, which was wrongly forgetting to compare the lengths of the argument lists. * It enhances ConVal to inclue a boolean for work-free-ness, so that the envt can contain non-work-free constructor applications, so that we can do more: see Note [ConVal work-free-ness] * It rejigs `subsumePats` so that it doesn't reverse the list. This can make a difference because, when patterns overlap, we arbitrarily pick the first. There is no "right" way, but this retains the old pre-subsumePats behaviour, thereby "fixing" the regression in #24282. Nofib results +======================================== | spectral/ansi -21.14% | spectral/hartel/comp_lab_zift -0.12% | spectral/hartel/parstof +0.09% | spectral/last-piece -2.32% | spectral/multiplier +6.03% | spectral/para +0.60% | spectral/simple -0.26% +======================================== | geom mean -0.18% +---------------------------------------- The regression in `multiplier` is sad, but it simply replicates GHC's previous behaviour (e.g. GHC 9.6).
-
- Jan 15, 2024
-
-
When looking up a value, we want to be able to find both variables and record fields. So we should not use the lookupSameOccRn_maybe function, as we can't know ahead of time which record field namespace a record field with the given textual name will belong to. Fixes #24293
-
- Jan 14, 2024
-
-
Pmc: COMPLETE pragmas associated with Family TyCons should apply to representation TyCons as well (#24326) Fixes #24326.
-
- Jan 13, 2024
-
-
Fixes #24311
-
In #24296, we had a program in which we eta expanded away an error despite the presence of `-fpedantic-bottoms`. This was caused by turning called *at least once* lambdas into one-shot lambdas, while with `-fpedantic-bottoms` it is only sound to eta expand over lambdas that are called *exactly* once. An example can be found in `Note [Combining arity type with demand info]`. Fixes #24296.
-
- Jan 10, 2024
-
-
Also remove -Wnoncanonical-{monoid,monad}-instances from -Wcompat, since they are enabled by default. Refresh wcompat-warnings/ test with new -Wcompat warnings. Part of #24267 Co-authored-by:sheaf <sam.derbyshire@gmail.com>
-
-
