[flang][OpenMP] Added support for lowering OpenMP taskwait construct #280
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
merges it onto a working branch.
fix the fallout from the pass manager changes in MLIR lapack work
[audit] add back changes that were lost with merge
Work on adjustable arrays.
- This requires analysis of local symbols.
- To do that requires getting the symbol table (scope), which required
threading the SemanticsContext for the main program.
- Added sorting and sorrted symbol tables to the PFT.
- Refactored temporaries.
- Misc. refactoring to follow LLVM coding standard, etc.
part 2 - implement the lower of variables
With these changes, I can compile code like srot.f.
A couple minor bug fixes
implement format for data transfer statements
update the lapack list
add a check and TODO
improve the folder for convert on bools DCE was broken. This adds back the Op traits to make it work correctly. Also tightens and repairs two tests, resp. Replace internal symbol map. The internal symbol map had served its purpose and needs to be replaced with a mapping structure that is more appropriate for Fortran variables, which can be composed of a number of runtime values. This new map will allow us to lower more complicated entities correctly by tracking their component values.
* Generate FIR for a group of multi-way branches - arithmetic if - assigned goto (needs assign statement, which is also implemented) - computed goto - select case * Review comment update.
fix compilation issue post merge remove reference to deleted header
Add fir::isa_real and fir::isa_complex and rework FirOpBuilder::isCharacter
…ap for
lowering. This piece primarily gets started in the direction of being able
to lower dynamic arrays and character types correctly.
There is still more work in fully leveraging this in the bridge, but all the
existing tests pass with this rewrite.
FIR changes:
This patch changes the FIR region operations to be more like those found in
the loop dialect. The changes are driven by the desire to be able to convert
FIR into a register ssa form. The changes are specifically:
- fir.loop has been replaced by fir.do_loop. The new form has the same
semantics (inclusive bounds, unordered), but can now carry ssa-values
around the loop and return them as results to the parent.
- fir.where has been replaced by fir.if. This new form is identical to
the current loop.if operation.
- fir.iterate_while has been added. This operation is very similar to
fir.loop with the addition that it requires a single loop-carried bool
value that signals an early-exit condition to the operation. While that
value is true, iterate_while will continue to iterate. When it becomes
false, the loop is exited.
- fir.result is the new terminator for the above ops to facilitate the
carrying of ssa-values through block arguments (phi nodes).
Change the syntax for fir.iterate_while to make it clear that the iterate condition is required.
Fixes bugs with lowering of fir.iterate_while to CFG. The lowered code produced is now as follows.
%0 = llvm.mlir.constant(1 : index) : !llvm.i64
%1 = llvm.mlir.constant(100 : index) : !llvm.i64
llvm.br ^bb1(%0, %arg0, %arg1 : !llvm.i64, !llvm.i1, !llvm.i32)
^bb1(%2: !llvm.i64, %3: !llvm.i1, %4: !llvm.i32): // 2 preds: ^bb0, ^bb2
%5 = llvm.icmp "slt" %2, %1 : !llvm.i64
%6 = llvm.and %5, %3 : !llvm.i1
llvm.cond_br %6, ^bb2, ^bb3
^bb2: // pred: ^bb1
%7 = ... : !llvm.i1
%8 = llvm.add %2, %0 : !llvm.i64
llvm.br ^bb1(%8, %7, %4 : !llvm.i64, !llvm.i1, !llvm.i32)
^bb3: // pred: ^bb1
[review comment] improve description text to make it more clear
[review] minor cleanups per review comments
1. Initialize scalar intrinsic types 2. Add test global inits. 3. Add TODO place holders for things that need to be implemented
speculatively. This eliminates creation of some empty fir.if blocks. Fix a bug in argument passing. When the expression is already a reference, we don't need or want to create a reference to the reference to pass the original reference. Instead just pass the reference itself. add FIXME for POINTER and ALLOCATABLE tighten some checks fix type mismatch on assignment misc fixes for LAPACK format bullet items
draft fix the lowering of adjusted arrays
Share pgmath description in lowering and folding. Simplify intrinsic lowering. Also add more intrinsics lowering: - ICHAR, ACHAR and plug all pgmath. Fix libpgmath linking regression after D78215 Rename isNotInMemory to needToMaterialize. Fix StaticMultimapView usage for clang
…lex of one precision to a complex of another. add a test. untabify Add a test and fix some bugs for F77 array lowering More work on LAPACK LAPACK: changes to get xerbla_array.f to compile clean. Remove the special case handling of LOGICAL and i1.
Implement subroutine calls with alternate returns.
…isions. move safe conversion to builder. rework FirOpBuilder for KindMapping. fixes all the internal conversion problems when lowering intrinsics as exposed by BLAS. fixes logical conversions add back a change lost in merge Start converting the speculative conversion of logical to bool to use lazy type coercion. Fix bugs that fall out of that conversion. fix compilation failure. function was merged twice. small cleanup on comments, missing cast remove stray change Wrap a CHARACTER literal in a box.
using the local tools setup. alternative fix. this one works on the Mac as well. Move the lib files around so they align with the include directory. Rework the cmake files and use the new styles there. Moving some code around. Renaming files to be more reflective of what they do currently. Refactoring comingled code into separate files.
* alternate return test
…ing calls. More work is needed here. review suggestions Remove the positive step check. Fortran has decrementing loops. fix typo update the LAPACK issues list Dominance.h moved fix breakage in tests
LEN_TRIM inline implementation + make character length index type - Use fir::iter_while operation to provide a scalar implementation of LEN_TRIM that is inlined. It is just a loop that goes from the right most character to the first one and stop as soon as a character is not a blank. - Change getLengthType to return the indexType. It makes loop manipulation of character easier. A few functions that expected i64 for the length are modified to support whatever integer type getLengthType is returning. Conflicts: flang/lib/Lower/FIRBuilder.cpp Conflicts: flang/LAPACK-bugs.txt
…really fix the lowering of procedure calls and their types.
To improve build times, lop off the f18 tool's dependence on Lowering. Moved the functionality to bbc, added some option support, and am able to run 3 of the 4 tests. Marked the last as expected fail. Conflicts: flang/tools/f18/CMakeLists.txt Remove raw_ostream references from Decimal
…initializers test
move towards making conversions more formal in the bridge Removes some old aliases. Use the long names. rename function
* I/O condition specifier branching * review update
Follow convention to capitalize Fortran_main.c
This patch tries to gather the actual requirements of what is expected for upcoming patches.
This merges remaining parts of: https://reviews.llvm.org/D83659
[flang][OpenMP] downstream merge after 3d0b760
…oduces some new types and operations that distinguish the runtime properties of array shape and array slicing operations. Additionally, we refine the array_coor and embox operations to use these new values. fix bug in pretty printer. add test. review comment - fix assert text add an assert to catch cases when shape is missing, etc.
Refine the IR as it relates to shapes and slices of arrays. This intr…
[Porting] Port a sample test in not-test/ to test/Fir.
fix standard option handling
…tation. Theextension is not presently used. However, the plan will be to experiment and see if it helps simplify lowering.
…traits Differential Revision: https://reviews.llvm.org/D83522
…traits Differential Revision: https://reviews.llvm.org/D83522
[flang] Adopt NoRegionArguments (WhereOp) and ParentOneOf (ResultOp) traits
Extend !fir.char to improve disambiguation of CHARACTER type
This patch adds lowering support for OpenMP-4.5 taskwait construct to OpenMP Dialect operations.
|
|
||
| program taskwait | ||
|
|
||
| integer :: a,b,c |
There was a problem hiding this comment.
Nit: Is this the usual alignment?
There was a problem hiding this comment.
The intention here, is to have some code/operation at least. This helps in seeing the actual lowering operation placed in correct place(However this test doesn't bother/check's what's above and below).
flang/test/Lower/omp-taskwait.f90
Outdated
| !$OMP TASKWAIT | ||
| !FIRDialect: omp.taskwait | ||
| !LLVMIRDialect: omp.taskwait | ||
| !LLVMIR: %{{.*}} = call i32 @__kmpc_omp_taskwait(%struct.ident_t* @1, i32 %omp_global_thread_num) |
There was a problem hiding this comment.
Is it better to be less specific about variable names (@1, omp_global_thread_num) here?
There was a problem hiding this comment.
Sure, Thanks for pointing this out.
I'll do it barrier test too. :)
| @@ -41,4 +41,19 @@ TEST_F(OpenMPLoweringTest, Barrier) { | |||
| EXPECT_EQ(succeeded(barrierOp.verify()), true); | |||
| } | |||
|
|
|||
| TEST_F(OpenMPLoweringTest, TaskWait) { | |||
There was a problem hiding this comment.
Is it possible to share some code between these tests?
There was a problem hiding this comment.
It can be done(by clustering directives in one test and check) but IMO that will defeat the purpose of unit test. right ?
The common code that can be shared across all is already in fixture.
kiranchandramohan
approved these changes
Jul 16, 2020
|
Thanks for reviewing this :). I'm raising review on Phabricator, so that community can also have a look. Share concerns if any! |
schweitzpgi
approved these changes
Jul 16, 2020
|
Please go ahead and merge it. I'll push in the phab review soon. |
|
Merged. |
SouraVX
added a commit
to llvm/llvm-project
that referenced
this pull request
Jul 17, 2020
Summary: This patch lower `!OMP TASKWAIT` construct from PFT to OpenMPDialect operations. Construct is lowered with conformance to OpenMP 4.5 spec. Patch is carved out of following approved PR: flang-compiler#280 Reviewed By: kiranchandramohan, clementval Differential Revision: https://reviews.llvm.org/D83983
arichardson
pushed a commit
to arichardson/llvm-project
that referenced
this pull request
Mar 21, 2021
Summary: This patch lower `!OMP TASKWAIT` construct from PFT to OpenMPDialect operations. Construct is lowered with conformance to OpenMP 4.5 spec. Patch is carved out of following approved PR: flang-compiler/f18-llvm-project#280 Reviewed By: kiranchandramohan, clementval Differential Revision: https://reviews.llvm.org/D83983
mem-frob
pushed a commit
to draperlaboratory/hope-llvm-project
that referenced
this pull request
Oct 7, 2022
Summary: This patch lower `!OMP TASKWAIT` construct from PFT to OpenMPDialect operations. Construct is lowered with conformance to OpenMP 4.5 spec. Patch is carved out of following approved PR: flang-compiler/f18-llvm-project#280 Reviewed By: kiranchandramohan, clementval Differential Revision: https://reviews.llvm.org/D83983
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This patch adds lowering support for OpenMP-4.5 taskwait construct to
OpenMP Dialect operations.
omp-taskwait.f90this file remain downstream, rest of the patch can be upstreamed once reviewed.Please review/approve this PR, don't merge it. As discussed we'll merge/review upstream first.