summaryrefslogtreecommitdiff
path: root/compiler/GHC/CmmToLlvm
Commit message (Collapse)AuthorAgeFilesLines
* Add fused multiply-add instructionssheaf2023-05-111-2/+26
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | This patch adds eight new primops that fuse a multiplication and an addition or subtraction: - `{fmadd,fmsub,fnmadd,fnmsub}{Float,Double}#` fmadd x y z is x * y + z, computed with a single rounding step. This patch implements code generation for these primops in the following backends: - X86, AArch64 and PowerPC NCG, - LLVM - C WASM uses the C implementation. The primops are unsupported in the JavaScript backend. The following constant folding rules are also provided: - compute a * b + c when a, b, c are all literals, - x * y + 0 ==> x * y, - ±1 * y + z ==> z ± y and x * ±1 + z ==> z ± x. NB: the constant folding rules incorrectly handle signed zero. This is a known limitation with GHC's floating-point constant folding rules (#21227), which we hope to resolve in the future.
* Allow LLVM backend to use HDoc for faster file generation.Alex Mason2023-03-213-40/+47
| | | | | | Also remove the MetaStmt constructor from LlvmStatement and places the annotations into the Store statement. Includes “Implement a workaround for -no-asm-shortcutting bug“ (https://gitlab.haskell.org/ghc/ghc/-/commit/2fda9e0df886cc551e2cd6b9c2a384192bdc3045)
* Enable tables next to code for LoongArch64lrzlin2023-02-031-1/+24
|
* Cmm: track the type of global registerssheaf2023-01-313-35/+35
| | | | | | | | | | | | This patch tracks the type of Cmm global registers. This is needed in order to lint uses of polymorphic registers, such as SIMD vector registers that can be used both for floating-point and integer values. This changes allows us to refactor VanillaReg to not store VGcPtr, as that information is instead stored in the type of the usage of the register. Fixes #22297
* llvmGen: Fix relaxed orderingBen Gamari2022-12-241-1/+1
| | | | | | | Previously I used LLVM's `unordered` ordering for the C11 `relaxed` ordering. However, this is wrong and should rather use the LLVM `monotonic` ordering. Fixes #22640
* llvm: Respect memory specified orderingsBen Gamari2022-12-151-7/+15
|
* cmm: Introduce MemoryOrderingsBen Gamari2022-12-151-7/+7
|
* compiler: remove unused MO_U_MulMayOfloCheng Shao2022-11-281-3/+0
| | | | We actually only emit MO_S_MulMayOflo and never emit MO_U_MulMayOflo anywhere.
* Minor SDoc-related cleanupKrzysztof Gogolewski2022-10-281-1/+1
| | | | | | | | | | | * Rename pprCLabel to pprCLabelStyle, and use the name pprCLabel for a function using CStyle (analogous to pprAsmLabel) * Move LabelStyle to the CLabel module, it no longer needs to be in Outputable. * Move calls to 'text' right next to literals, to make sure the text/str rule is triggered. * Remove FastString/String roundtrip in Tc.Deriv.Generate * Introduce showSDocForUser', which abstracts over a pattern in GHCi.UI
* Avoid Data.List.group; prefer Data.List.NonEmpty.groupBodigrim2022-09-281-9/+9
| | | | | This allows to avoid further partiality, e. g., map head . group is replaced by map NE.head . NE.group, and there are less panic calls.
* Clean up some. In particular:M Farkas-Dyck2022-09-171-10/+4
| | | | | | | | | | • Delete some dead code, largely under `GHC.Utils`. • Clean up a few definitions in `GHC.Utils.(Misc, Monad)`. • Clean up `GHC.Types.SrcLoc`. • Derive stock `Functor, Foldable, Traversable` for more types. • Derive more instances for newtypes. Bump haddock submodule.
* Fix typosEric Lindblad2022-09-141-1/+1
| | | | | | | This fixes various typos and spelling mistakes in the compiler. Fixes #21891
* Minor SDoc cleanupKrzysztof Gogolewski2022-09-073-11/+11
| | | | | | | Change calls to renderWithContext with showSDocOneLine; it's more efficient and explanatory. Remove polyPatSig (unused)
* Remove label style from printing contextKrzysztof Gogolewski2022-08-261-8/+1
| | | | | | | | | | | | Previously, the SDocContext used for code generation contained information whether the labels should use Asm or C style. However, at every individual call site, this is known statically. This removes the parameter to 'PprCode' and replaces every 'pdoc' used to print a label in code style with 'pprCLabel' or 'pprAsmLabel'. The OutputableP instance is now used only for dumps. The output of T15155 changes, it now uses the Asm style (which is faithful to what actually happens).
* CmmToLlvm: Don't aliasify builtin LLVM variablesBen Gamari2022-08-161-4/+16
| | | | | | | | | Our aliasification logic would previously turn builtin LLVM variables into aliases, which apparently confuses LLVM. This manifested in initializers failing to be emitted, resulting in many profiling failures with the LLVM backend. Fixes #22019.
* cmm: Remove unused ReadOnlyData16Cheng Shao2022-08-041-3/+0
| | | | We don't actually emit rodata16 sections anywhere.
* cmm: Eliminate orphan Outputable instancesBen Gamari2022-07-162-5/+3
| | | | | | | Here we reorganize `GHC.Cmm` to eliminate the orphan `Outputable` and `OutputableP` instances for the Cmm AST. This makes it significantly easier to use the Cmm pretty-printers in tracing output without incurring module import cycles.
* Change `Backend` type and remove direct dependencieswip/backend-as-recordNorman Ramsey2022-05-212-4/+0
| | | | | | | | | | | | | | | | | | | With this change, `Backend` becomes an abstract type (there are no more exposed value constructors). Decisions that were formerly made by asking "is the current back end equal to (or different from) this named value constructor?" are now made by interrogating the back end about its properties, which are functions exported by `GHC.Driver.Backend`. There is a description of how to migrate code using `Backend` in the user guide. Clients using the GHC API can find a backdoor to access the Backend datatype in GHC.Driver.Backend.Internal. Bumps haddock submodule. Fixes #20927
* Don't store LlvmConfig into DynFlagsSylvain Henry2022-05-172-50/+113
| | | | | | | | | | | | | | | | | | | | | LlvmConfig contains information read from llvm-passes and llvm-targets files in GHC's top directory. Reading these files is done only when needed (i.e. when the LLVM backend is used) and cached for the whole compiler session. This patch changes the way this is done: - Split LlvmConfig into LlvmConfig and LlvmConfigCache - Store LlvmConfigCache in HscEnv instead of DynFlags: there is no good reason to store it in DynFlags. As it is fixed per session, we store it in the session state instead (HscEnv). - Initializing LlvmConfigCache required some changes to driver functions such as newHscEnv. I've used the opportunity to untangle initHscEnv from initGhcMonad (in top-level GHC module) and to move it to GHC.Driver.Main, close to newHscEnv. - I've also made `cmmPipeline` independent of HscEnv in order to remove the call to newHscEnv in regalloc_unit_tests.
* Refactor handling of global initializersBen Gamari2022-04-011-1/+44
| | | | | | | | | | | | | | | | | | | | | | | | | | | GHC uses global initializers for a number of things including cost-center registration, info-table provenance registration, and setup of foreign exports. Previously, the global initializer arrays which referenced these initializers would live in the object file of the C stub, which would then be merged into the main object file of the module. Unfortunately, this approach is no longer tenable with the move to Clang/LLVM on Windows (see #21019). Specifically, lld's PE backend does not support object merging (that is, the -r flag). Instead we are now rather packaging a module's object files into a static library. However, this is problematic in the case of initializers as there are no references to the C stub object in the archive, meaning that the linker may drop the object from the final link. This patch refactors our handling of global initializers to instead place initializer arrays within the object file of the module to which they belong. We do this by introducing a Cmm data declaration containing the initializer array in the module's Cmm stream. While the initializer functions themselves remain in separate C stub objects, the reference from the module's object ensures that they are not dropped from the final link. In service of #21068.
* Fix all invalid haddock comments in the compilerZubin Duggal2022-03-291-2/+2
| | | | Fixes #20935 and #20924
* Fix some notesMatthew Pickering2022-02-081-1/+1
|
* llvmGen: Handle unaligned loads/storesBen Gamari2022-02-041-7/+7
| | | | | This allows us to produce valid code for indexWord8ArrayAs*# on platforms that lack unaligned memory access.
* Introduce alignment to CmmStoreBen Gamari2022-02-041-44/+55
|
* Introduce alignment in CmmLoadBen Gamari2022-02-041-32/+50
|
* Fix a few Note inconsistenciesBen Gamari2022-02-012-2/+1
|
* Consistently upper-case "Note ["Ben Gamari2022-02-012-2/+2
| | | | | | This was achieved with git ls-tree --name-only HEAD -r | xargs sed -i -e 's/note \[/Note \[/g'
* Rip out remaining SPARC supportBen Gamari2022-01-291-2/+2
|
* CmmToLlvm: rename LCGConfig -> LlvmCgConfigdoyougnu2021-12-145-30/+30
| | | | | | | | | | | | | | | | | | | | CmmToLlvm: renamce lcgPlatform -> llvmCgPlatform CmmToLlvm: rename lcgContext -> llvmCgContext CmmToLlvm: rename lcgFillUndefWithGarbage CmmToLlvm: rename lcgSplitSections CmmToLlvm: lcgBmiVersion -> llvmCgBmiVersion CmmToLlvm: lcgLlvmVersion -> llvmCgLlvmVersion CmmToLlvm: lcgDoWarn -> llvmCgDoWarn CmmToLlvm: lcgLlvmConfig -> llvmCgLlvmConfig CmmToLlvm: llvmCgPlatformMisc --> llvmCgLlvmTarget
* CmmToLlvm.Config: strictify LlvmConfig fielddoyougnu2021-12-141-2/+3
|
* LLVM.CodeGen: use fast-string literalsdoyougnu2021-12-141-34/+145
| | | | | | | That is remove factorization of common strings and string building code for the LLVM code gen ops. Replace these with string literals to obey the FastString rewrite rule in GHC.Data.FastString and compute the string length at compile time
* CmmToLlvm: Remove DynFlags, add LlvmCgConfigdoyougnu2021-12-145-131/+138
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | CodeOutput: LCGConfig, add handshake initLCGConfig Add two modules: GHC.CmmToLlvm.Config -- to hold the Llvm code gen config GHC.Driver.Config.CmmToLlvm -- for initialization, other utils CmmToLlvm: remove HasDynFlags, add LlvmConfig CmmToLlvm: add lcgContext to LCGConfig CmmToLlvm.Base: DynFlags --> LCGConfig Llvm: absorb LlvmOpts into LCGConfig CmmToLlvm.Ppr: swap DynFlags --> LCGConfig CmmToLlvm.CodeGen: swap DynFlags --> LCGConfig CmmToLlvm.CodeGen: swap DynFlags --> LCGConfig CmmToLlvm.Data: swap LlvmOpts --> LCGConfig CmmToLlvm: swap DynFlags --> LCGConfig CmmToLlvm: move LlvmVersion to CmmToLlvm.Config Additionally: - refactor Config and initConfig to hold LlvmVersion - push IO needed to get LlvmVersion to boundary between Cmm and LLvm code generation - remove redundant imports, this is much cleaner! CmmToLlvm.Config: store platformMisc_llvmTarget instead of all of platformMisc
* Avoid GHC_STAGE and other include bitsJohn Ericson2021-11-051-1/+1
| | | | | | | | | We should strive to make our includes in terms of the RTS as much as possible. One place there that is not possible, the llvm version, we make a new tiny header Stage numbers are somewhat arbitrary, if we simple need a newer RTS, we should say so.
* Let LLVM and C handle > native size arithmeticJohn Ericson2021-09-081-29/+31
| | | | | | NCG needs to call slow FFI functions where we "borrow" the C compiler's implementation, but there is no reason why we need to do that for LLVM, or the unregisterized backend where everything is via C anyways!
* Add Word64#/Int64# primopsSylvain Henry2021-07-151-0/+31
| | | | | | | | | | | | | | | | | | | | | | | Word64#/Int64# are only used on 32-bit architectures. Before this patch, operations on these types were directly using the FFI. Now we use real primops that are then lowered into ccalls. The advantage of doing this is that we can now perform constant folding on Word64#/Int64# (#19024). Most of this work was done by John Ericson in !3658. However this patch doesn't go as far as e.g. changing Word64 to always be using Word64#. Noticeable performance improvements T9203(normal) run/alloc 89870808.0 66662456.0 -25.8% GOOD haddock.Cabal(normal) run/alloc 14215777340.8 12780374172.0 -10.1% GOOD haddock.base(normal) run/alloc 15420020877.6 13643834480.0 -11.5% GOOD Metric Decrease: T9203 haddock.Cabal haddock.base
* Set min LLVM version to 9 and make version checking use a non-inclusive upperZubin Duggal2021-06-201-7/+9
| | | | | | | bound. We use a non-inclusive upper bound so that setting the upper bound to 13 for example means that all 12.x versions are accepted.
* Make Logger independent of DynFlagsSylvain Henry2021-06-072-24/+18
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Introduce LogFlags as a independent subset of DynFlags used for logging. As a consequence in many places we don't have to pass both Logger and DynFlags anymore. The main reason for this refactoring is that I want to refactor the systools interfaces: for now many systools functions use DynFlags both to use the Logger and to fetch their parameters (e.g. ldInputs for the linker). I'm interested in refactoring the way they fetch their parameters (i.e. use dedicated XxxOpts data types instead of DynFlags) for #19877. But if I did this refactoring before refactoring the Logger, we would have duplicate parameters (e.g. ldInputs from DynFlags and linkerInputs from LinkerOpts). Hence this patch first. Some flags don't really belong to LogFlags because they are subsystem specific (e.g. most DumpFlags). For example -ddump-asm should better be passed in NCGConfig somehow. This patch doesn't fix this tight coupling: the dump flags are part of the UI but they are passed all the way down for example to infer the file name for the dumps. Because LogFlags are a subset of the DynFlags, we must update the former when the latter changes (not so often). As a consequence we now use accessors to read/write DynFlags in HscEnv instead of using `hsc_dflags` directly. In the process I've also made some subsystems less dependent on DynFlags: - CmmToAsm: by passing some missing flags via NCGConfig (see new fields in GHC.CmmToAsm.Config) - Core.Opt.*: - by passing -dinline-check value into UnfoldingOpts - by fixing some Core passes interfaces (e.g. CallArity, FloatIn) that took DynFlags argument for no good reason. - as a side-effect GHC.Core.Opt.Pipeline.doCorePass is much less convoluted.
* Cmm: fix sinking after suspendThreadSylvain Henry2021-05-191-0/+3
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Suppose a safe call: myCall(x,y,z) It is lowered into three unsafe calls in Cmm: r = suspendThread(...); myCall(x,y,z); resumeThread(r); Consider the following situation for myCall arguments: x = Sp[..] -- stack y = Hp[..] -- heap z = R1 -- global register r = suspendThread(...); myCall(x,y,z); resumeThread(r); The sink pass assumes that unsafe calls clobber memory (heap and stack), hence x and y assignments are not sunk after `suspendThread`. The sink pass also correctly handles global register clobbering for all unsafe calls, except `suspendThread`! `suspendThread` is special because it releases the capability the thread is running on. Hence the sink pass must also take into account global registers that are mapped into memory (in the capability). In the example above, we could get: r = suspendThread(...); z = R1 myCall(x,y,z); resumeThread(r); But this transformation isn't valid if R1 is (BaseReg->rR1) as BaseReg is invalid between suspendThread and resumeThread. This caused argument corruption at least with the C backend ("unregisterised") in #19237. Fix #19237
* Remove useless {-# LANGUAGE CPP #-} pragmasSylvain Henry2021-05-124-6/+6
|
* Fully remove HsVersions.hSylvain Henry2021-05-125-9/+0
| | | | | | | | | | Replace uses of WARN macro with calls to: warnPprTrace :: Bool -> SDoc -> a -> a Remove the now unused HsVersions.h Bump haddock submodule
* Replace CPP assertions with Haskell functionsSylvain Henry2021-05-121-7/+10
| | | | | | | | | | | | | | | There is no reason to use CPP. __LINE__ and __FILE__ macros are now better replaced with GHC's CallStack. As a bonus, assert error messages now contain more information (function name, column). Here is the mapping table (HasCallStack omitted): * ASSERT: assert :: Bool -> a -> a * MASSERT: massert :: Bool -> m () * ASSERTM: assertM :: m Bool -> m () * ASSERT2: assertPpr :: Bool -> SDoc -> a -> a * MASSERT2: massertPpr :: Bool -> SDoc -> m () * ASSERTM2: assertPprM :: m Bool -> SDoc -> m ()
* Ensure assert from Control.Exception isn't usedSylvain Henry2021-05-121-1/+1
|
* Enable tables next to code for riscv64Andreas Schwab2021-04-191-1/+22
| | | | | This requires adding another rewrite to the mangler, to avoid generating PLT entries.
* llvmGen: Accept range of LLVM versionsBen Gamari2021-03-171-4/+8
| | | | | | | Previously we would support only one LLVM major version. Here we generalize this to accept a range, taking this range to be LLVM 10 to 11, as 11 is necessary for Apple M1 support. We also accept 12, as that is what apple ships with BigSur on the M1.
* Reimplement Stream in "yoneda" style for efficiencyMatthew Pickering2021-02-261-11/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | 'Stream' is implemented in the "yoneda" style for efficiency. By representing a stream in this manner 'fmap' and '>>=' operations are accumulated in the function parameters before being applied once when the stream is destroyed. In the old implementation each usage of 'mapM' and '>>=' would traverse the entire stream in order to apply the substitution at the leaves. It is well-known for free monads that this representation can improve performance, and the test results demonstrate this for GHC as well. The operation mapAccumL is not used in the compiler and can't be implemented efficiently because it requires destroying and rebuilding the stream. I removed one use of mapAccumL_ which has similar problems but the other use was difficult to remove. In the future it may be worth exploring whether the 'Stream' encoding could be modified further to capture the mapAccumL pattern, and likewise defer the passing of accumulation parameter until the stream is finally consumed. The >>= operation for 'Stream' was a hot-spot in the ticky profile for the "ManyConstructors" test which called the 'cg' function many times in "StgToCmm.hs" Metric Decrease: ManyConstructors
* Refactor LoggerSylvain Henry2021-02-132-7/+15
| | | | | | | | | | | | | | | | | | | | | Before this patch, the only way to override GHC's default logging behavior was to set `log_action`, `dump_action` and `trace_action` fields in DynFlags. This patch introduces a new Logger abstraction and stores it in HscEnv instead. This is part of #17957 (avoid storing state in DynFlags). DynFlags are duplicated and updated per-module (because of OPTIONS_GHC pragma), so we shouldn't store global state in them. This patch also fixes a race in parallel "--make" mode which updated the `generatedDumps` IORef concurrently. Bump haddock submodule The increase in MultilayerModules is tracked in #19293. Metric Increase: MultiLayerModules
* LLVM: fix sized shift primops (#19215)Sylvain Henry2021-01-221-10/+20
| | | | | Ensure that shift amount parameter has the same type as the parameter to shift.
* CmmToLlvm: Sign/Zero extend parameters for foreign callsStefan Schulze Frielinghaus2021-01-171-14/+17
| | | | | | | | For some architectures the C calling convention is that any integer shorter than 64 bits is replaced by its 64 bits representation using sign or zero extension. Fixes #19023.
* CmmToLlvm: Declare signature for memcmpwip/angerman/arm64Ben Gamari2020-11-241-2/+7
| | | | | | Otherwise `opt` fails with: error: use of undefined value '@memcmp$def'
* Add the proper HLint rules and remove redundant keywords from compilerHécate2020-11-011-16/+14
|
View Lorry Controller Status