[Flang][OpenMP] Important issues to be fixed prior to removing experimental status of OpenMP

[kiranchandramohan]

TODOs for unsupported OpenMP 5.2 features

• conditional modifier for lastprivate [flang][OpenMP] Parse lastprivate modifier, add TODO to lowering #110568
• iterator modifier for affinity, depend, from, map, to (@kparzysz)
• scan directive : [Flang][OpenMP][Sema] Adding parsing and semantic support for scan directive. #102792 @anchuraj
• inscan modifier for reduction (@Leporacanthicus)
• task modifier for reduction clause (@Leporacanthicus)
• in_reduction clause for task and target (@Leporacanthicus)
• task_reduction clause for taskgroup (@Leporacanthicus)
• reduction, in_reduction clause for taskloop (@kaviya2510) [FLANG][OpenMP][Taskloop] - Add testcase for reduction and in_reduction clause in taskloop construct #139704
• close, present, ompx_hold map-type modifiers for map clause (@kparzysz) [flang][OpenMP] Parsing support for map type modifiers #111860
• none parameter for defaultmap clause (Currently caught in Semantics giving an error - leaving as is for now @Leporacanthicus)
• declare mapper directive @Leporacanthicus [Flang][OMP]Add support for DECLARE MAPPER parsing and semantics #115160
• metadirective (@kparzysz)
  • [flang][OpenMP] Make parsing of trait properties more context-sensitive #122900
  • [flang][OpenMP] Parse WHEN, OTHERWISE, MATCH clauses plus METADIRECTIVE #121817 (approved)
  • [flang][OpenMP] Semantic checks for context selectors #123243
  • [flang][OpenMP] Parse METADIRECTIVE in specification part #123397 (approved)
  • [flang][OpenMP] Handle directive arguments in OmpDirectiveSpecifier #124278
• declare variant directive : ARM @kiranchandramohan WIP [Flang][OpenMP] Add frontend support for declare variant #130578
• collapse of imperfectly nested loops for simd, worksharing, distribute and taskloop. Currently there is a semantics error.
• affinity clause for task (@kparzysz) [flang][OpenMP] Parse AFFINITY clause, lowering not supported yet #113485
• detach clause for task (@NimishMishra): [flang][OpenMP] Add parsing support for Task detach #112312, [mlir][llvm] Translation support for task detach #116601, [flang][OpenMP] Add lowering support for task detach #119128, [Flang] [OpenMP] Add semantic checks for detach clause in task #119172
• cancel construct : support for taskloop (@kiranktp )
• mutexinoutset for depend clause (@kparzysz) [flang][OpenMP] Add frontend support for INOUTSET and MUTEXINOUTSET #114895
• ancestor modifier for device clause of target construct
• depend and nowait clause for taskwait (@Leporacanthicus)
• acq_rel, acquire, release clauses for atomic construct
• depend clause extensions to support iterators and depend objects with depobj construct (@kparzysz) [flang][OpenMP] Parsing support for iterator in DEPEND clause #113622, [flang][OpenMP] Parser support for DEPOBJ plus DEPEND, DESTROY, UPDATE #114074
• Combined constructs master taskloop, parallelmaster, parallel master taskloop, master taskloop simd, parallel master taskloop simd @kiranchandramohan [Flang][OpenMP] Add frontend support for directives involving master #113893
• iterators in map clause and data motion clauses of target update (@kparzysz)
• present argument for defaultmap ( @agozillon )
• align clause on allocate directive : ARM [FLANG][OpenMP]Add support for ALIGN clause on OMP ALLOCATE #120791
• allocator and align modifiers for allocate clause : ARM [Flang][OpenMP]Add tests for align and allocator in allocate clauses #121356
• adjust_args and append_args clauses for declare variant directive : ARM
• dispatch construct : [Flang][OpenMP]Add parsing support for DISPATCH construct #121982 @Leporacanthicus
• Assumption directives : [FLANG][OpenMP]Add frontend support for ASSUME and ASSUMES #120770 @Leporacanthicus
• nothing directive (@kparzysz) [flang][OpenMP] Frontend support for NOTHING directive #120606
• error directive : ARM [flang][OpenMP]Add support for OpenMP ERROR directive #119582
• Loop transformation constructs @kiranchandramohan The parsing of these constructs work and a TODO is emitted, but semantic checks should be updated to allow loop transformation constructs in place of Fortran do loops.
• scope directive @kiranchandramohan Add parser+semantics support for scope construct #113700
• strict modifier for grainsize and num_tasks clause @kiranchandramohan [Flang][OpenMP] Add parser support for grainsize and num_tasks clause #113136
• thread_limit clause for target directive. Frontend and Lowering support is present. Translation support is probably missing (@skatrak).
• has_device_addr clause for target construct. Frontend and Lowering support is present. Translation support is probably missing. (@kparzysz) [flang][OpenMP] Implement HAS_DEVICE_ADDR clause #128568
• interop directive : [Swateesh- AMD] [flang][openmp] Adds Parser and Semantic Support for Interop Construct, and Init and Use Clauses. #120584
• compare, fail clauses : ARM Compare support: [flang][OpenMP]Add parsing and semantics support for ATOMIC COMPARE #117032 and 1False positive: "Undefined or garbage value returned to caller" #18683
• seq_cst clause for flush construct seq_cst is allowed in Flush since OpenMP 5.1. #114072
• inoutset argument for depend clause (@kparzysz) [flang][OpenMP] Add frontend support for INOUTSET and MUTEXINOUTSET #114895
• Allow copyprivate, nowait on the directive (in addition to end directive) (@Thirumalai-Shaktivel, [Flang][OpenMP] Allow copyprivate and nowait on the directive clauses #127769)
• Deprecation of destroy with depobj (@kparzysz) [flang][OpenMP] Deprecation message for DESTROY with no argument #114988
• step modifier for linear. Deprecation (@kparzysz) [flang][OpenMP] Rework LINEAR clause #119278
• Deprecation of modifiers without comma separators (@kparzysz) [flang][OpenMP] Use new modifier infrastructure for MAP/FROM/TO clauses #117447
• enter as synonym for to clause in declare target . to clause is deprecated.
• declare mapper directive extension to accept iterator-modifier and the present map-type-modifier (@kparzysz)
• allocators construct. allocate directive for ALLOCATE statement is deprecated : ARM
• dispatch construct has optional end directive : ARM
• otherwise clause as synonym for default on metadirective (@kparzysz) ([flang][OpenMP] Parse WHEN, OTHERWISE, MATCH clauses plus METADIRECTIVE #121817)
• allocate and firstprivate for scope directive : ARM (@Leporacanthicus) [flang][OpenMP]Add TODO checking for scope allocate and firstprivate #116801
• target enter and exit data has default map type
• doacross clause as synonym for depend in some situations (@kparzysz) [flang][OpenMP] Parse DOACROSS clause #115396 [flang][OpenMP] Semantic checks for DOACROSS clause #115397

TODOs for unsupported OpenMP 4.5 features

TODOs for unsupported OpenMP 4.0 features

• declare reduction @Leporacanthicus

Must fix issues from the tracker

A subset of issues from https://github.com/llvm/llvm-project/issues?q=is%3Aopen+is%3Aissue+label%3Aflang%3Aopenmp

• Atomic with array constructor [Flang][OpenMP] ICE with "LLVM ERROR: not implemented" (omp atomic) #104526 @Thirumalai-Shaktivel [Flang] Check if two ArrayConstructor's are Equal #121181
• Atomic write to complex [flang][OpenMP] atomic write on complex #113479 @NimishMishra [mlir][OpenMP] Add __atomic_store to AtomicInfo #121055
• Error when capture-statement described in atomic capture construct contains a different type [flang][OpenMP] llvm-project/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp:8095: OpenMPIRBuilder::InsertPointTy llvm::OpenMPIRBuilder::createAtomicWrite(const llvm::OpenMPIRBuilder::LocationDescription &, llvm::OpenMPIRBuilder::AtomicOpValue &, llvm::Value *, llvm::AtomicOrdering): Assertion `(XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() || XElemTy->isPointerTy()) && "OMP atomic write expected a scalar type"' failed. #115652 @NimishMishra : [mlir][OpenMP] Add __atomic_store to AtomicInfo #121055
• Runtime segfault when allocatable threadprivate variable is used with copyin [flang][OpenMP] runtime segfault when an allocatable variable is used with copyin #113191 @kaviya2510 : https://github.com/kaviya2510/llvm-project/tree/copyin_allocatable
• Assertion failure using default clause with critical [flang][OpenMP] assertion failure using default clause with critical #112572
• Crash in OpenMPIRBuilder when trying to outline a parallel region with an infinite loop [flang][OpenMP] crash in OpenMPIRBuilder when trying to outline a parallel region with an infinite loop #112884 @tblah
• Problem building GAMESS (delayed privatisation of derived type) [Flang][OpenMP] Compilation error when derived type with procedure pointer appears within private attribute #109727
• dot product inside of omp.parallel (part of cp2k) [flang] compiler crashes in OptimizedBufferizationPass when dot_product is used inside parallel region #116954
• Privatisation with sections ICE [Flang][OpenMP] Assertion `success && "Privatization failed due to existing binding"' failed. (parallel sections directive) #135108 @tblah
• Common block name in omp flush [flang] [OpenMP] Common block name symbol is not found (with the prefix "Internal") #136583 @tblah
• threadprivate with pointer to character [Flang][OpenMP] Assertion `!lenParams.empty()' failed. #108136 @tblah
• [flang][OpenMP] DECLARE SIMD LINEAR implicitly declaring a variable #140754
• [flang][OpenMP] symbol resolution error for DECLARE SIMD UNIFORM #140741
• [flang][OpenMP] Missing pre-determined DSA for static duration variables #140732
• [flang][OpenMP] Lowering of OpenMP construct fails in a procedure with alternate entries #139071
• [Flang][OpenMP] compilation failed with "Basic Block in function '_QQmain' does not have terminator!" #138436 @tblah
• [Flang][OpenMP] Referring to an instruction in another function! #138102
• [Flang][OpenMP] only assertion-enabled flang can reach llvm_unreachable (cancel directive) #133685 @tblah
• [Flang][OpenMP] UNREACHABLE executed at llvm-project/flang/lib/Parser/parse-tree.cpp:297! #133678
• [Flang][OpenMP] Compilation abnormally terminates when the left-hand side of write-statement specified in atomic write construct is a variable of derived type #126451
• [Flang][OpenMP] private clause on target teams directive violates assertion #116428

Must fix issues not in tracker

• Default loop schedule modifier for worksharing-loop constructs without static schedule and ordered clause is nonmonotonic since OpenMP 5.0 @tblah
• Privatization issues in interfacing with OpenMPIRBuilder. https://discourse.llvm.org/t/rfc-openmp-fix-issue-in-mlir-to-llvmir-translation-for-delayed-privatisation/81225. Fixed by @tblah [mlir][OpenMP] rewrite conversion of privatisation for omp.parallel #111844
• Support for deferred task execution @tblah [mlir][OpenMP] Pack task private variables into a heap-allocated context struct #125307

Minor bugs:

• Crash with undeclared variable and implicit none: [flang][OpenMP]Flang crashes with implicit none and undeclared variables #131655
• Bad word ordering leads to assert in lowering: [flang][openmp]Incorrect order in openmp line causes assert #131659

[anchuraj]

inscan modifier support is WIP

[kiranchandramohan]

inscan modifier support is WIP

Feel free to update the tasklist if you or someone you know is working on it. The immediate requirement is to only add a TODO message after parsing and semantics.

[h-vetinari]

TODOs for unsupported OpenMP 5.2 features

I find it strange that a standard version released after flang got merged into LLVM is somehow a deciding factor for whether the whole compiler is considered incomplete or experimental.

It sounds like massive scope creep, in the sense that, when flang was proposed for inclusion, no-one could have required an unreleased standard as the benchmark for what counts as a full-fledged fortran compiler.

[kiranchandramohan]

TODOs for unsupported OpenMP 5.2 features

I find it strange that a standard version released after flang got merged into LLVM is somehow a deciding factor for whether the whole compiler is considered incomplete or experimental.

It sounds like massive scope creep, in the sense that, when flang was proposed for inclusion, no-one could have required an unreleased standard as the benchmark for what counts as a full-fledged fortran compiler.

We started with OpenMP 4.5 standard when it was the F18 project. Subsequently, we have added OpenMP 5.2 features to the implementation. What I would like to get to is to be able to make statements on what is supported and what is not supported. Since some OpenMP 5.2 features are added the only standard that we can aim to completely address is OpenMP 5.2.

Also, I am not saying we have to support all OpenMP 5.2 features. The requirement is only to issue messages and gracefully error that the feature is not supported as the rest of the compiler says for Fortran language features. Also, this is not a requirement for the whole of the compiler, it is only about the experimental nature of the OpenMP implementation.
Note: I will change the title of the ticket to reflect this.

[Leporacanthicus]

Additional tests and small OpenMP.cpp changes available to review here:
#111562

I will continue this work, but this seemed like a good place to stop for today, and may start reviewing this part.

[tblah]

Should we also include fixing (at least some) compiler crash tickets before removing the experimental warning?

[kiranchandramohan]

Should we also include fixing (at least some) compiler crash tickets before removing the experimental warning?

If there are important issues please add it to the following entry. I have not yet managed to go through all the issues in the tracker and hence that section is empty. May be I will populate with a subset of entries from issues with label flang:openmp.

Must fix issues from the tracker

[tblah]

If there are important issues please add it to the following entry.

I have added some, striking an opinionated balance between adding the whole issue tracker and missing important issues. If others disagree with my prioritization of those issues or would like to add more please go ahead.