basic character assignement end-to-end regression test#2
Open
jeanPerier wants to merge 124 commits intojpr-easier-loweringfrom
Open
basic character assignement end-to-end regression test#2jeanPerier wants to merge 124 commits intojpr-easier-loweringfrom
jeanPerier wants to merge 124 commits intojpr-easier-loweringfrom
Conversation
adding types - expand sequence type to support Fortran descriptor information add code to lower function type signatures to LLVM types add support and lowering for an if-then-else construct add support and lowering for SELECT X operations add address arithmetic to canonicalize changes to start adding new concepts to FIR start adding top-level special forms convert the build system to work as an in-tree build finish draft of pretty parsers and printers Conflicts: CMakeLists.txt
add "How To Lower" doc changes to put optional attributes in and position them before the type allow unquoted identifiers turn off warnings on Mac
The goal is to abstract which library is targeted while lowering the calls. What matters in this commit is the exposed interface and the dynamic data structure that holds the runtime description. The static description of the runtime used to feed this API is a dummy description. The long term goal is to share this knowledge with the folding code that uses runtime. SPOT.
Initial PR: schweitzpgi#2 * std::string -> llvm::StringRef * removed const& around mlir::Type * Added comment about optional argument non handling
convert MemToReg to template form add converter stubs for all FIR ops move the bridge from FIR to standard to be a transformation pass
mlir::ModuleManager is not required anymore, mlir::ModuleOp can directly be used as a mlir::SymbolTable to querry symbols. The mlir::ModuleManager is usefull to create distinct temp names from a variable name, but it does not sync its symbole table automatically with the module it manages. Keeping a ModuleManager in a high level context seems to be too risky. For instance, there was currently a bug where function names were directly created in the Module but looked-up in the ModuleManager that could not find them.
add complex primitives; start lowering some FIR to LLVM IR
add FirEndOp converter
add lowering for mulc and divc
add real primitives to FIR to support extended precision types add float primitives for !fir.real types that are not supported in std dialect add option for user control of KIND mappings
+ Add a level of indirection before the facility that describes the runtime
functions that implement Fortran intrinsic functions. This level allows
to select a code generation function depending on the intrinsic name.
This is because not all Fortran intrinsic functions map to runtime functions
and even when they do, they may not map one to one and may require generating
some boilerplate.
+ Completly hide the implementation using an owning pointer. It did not bring
anything to define implemenation related class in the header. So only
expose a simple interface.
+ All code implementation code is reorganised in three classes:
-> RuntimeStaticDescription: allows constexpr description of the runtime libraries.
-> MathRuntimeLibrary: maps a reified representation of runtime functions to generic
intrinsic names. It is dynamically built from RuntimeStaticDescription.
-> IntrinsicLibrary::Implementation: define functions to generate FIR+MLIR for
intrinsic functions. Defines a dispatch table between generic names and such
generator functions. Holds a MathRuntimeLibrary to allow generator function to
generate runtime calls.
+ Get rid of non constexpr global objects and make IntrinsicLibrary immutable once
it is built. This is to harden this utility for use in multithreaded contexts.
+ Add the possibility to generate FIR+MLIR code for intrinsic calls inside wrappers
to keep the code readable and also keep better track of Fortran intrinsics which
may be usefull/necessary in future optimization passes.
Address review comments
* Type -> TypeCode to avoid confusion with mlir::Type
* Remove resultType member from IntrinsicLibrary::Implementation::Context
* std::optional -> llvm::Optional
identified structs; box ops; add complex changes from 8c33745
…ls can now use the `fir.call` operation.
Update FIR language document.
create Transforms folder clone the MLIR CSE pass for modification
* Provide a class to help generating fir to
manipulate complex in fir. It hides InsertValueOp and
ExtractValueOp details.
* Implement complex compare
* Replace all mlir float basic ops (e.g. mlir::AddFOp) by their
FIR equivalent. This is because mlir::AddFOp and such do not
support fir real types operands and we cannot ensure all real
types are lowered to mlir real types (e.g. fir.real<16> currently).
Fir real ops support both fir and mlir real operands.
* Prevent re-loading of a same symbol while lowering a same symbol.
This especially important for complex because the expression
analysis may have already re-written some complex ops as real
ops piece-wise. This increase the occurance of symbols.
Fortran gives us the guarantee we should not have to reload a var
while evaluating an expression (see F2018 section 10.1.4).
Return statement were not returning any value for
functions. The function generating return opertion was lacking knowledge
about the result symbol to return a value. This commit solves this by:
+ Adding a pointer to the root of the parse tree being lowered in
AnalysisData.
+ Restructuring the functions handling flat::ReturnOp to handle
main, subroutine and functions differently and proagate
the information needed to emit the related return ops
+ For functions, get the symbol result form FunctionSubprogram parse tree
node, load its value and return the value.
The knowledege of the parse tree root that can be flattened is now contained in
a variant in AnalysisData so there is no need to pass the tree node in
`CreateFlatFIR`. The template and its explicit instantiations can be removed
and replaced by a `std::visit` inside the single `CreateFlatFIR` entry point.
Conflicts:
lib/burnside/bridge.cc
+ This is a primitive implementation that only works with
implicit interface and when the arguments are variables (i.e. symbols).
+ Strip out the part that are not specific to maths runtime into runtime.cc and runtime.h so that they can be used by future other runtime description. (RuntimeStaticDescription and TypeCode) + Provide a new StaticMultimapView that can provide constexpr multimap search functions over a static constexpr tables. This is to provide a way to handle runtime description without needing to dynamicaaly instanciate a map and hold it somewhere. For now, it is not used on the math runtime so that we can compare both strategies.
It uses what was made generic from the maths intrinsic runtime. The main differences is that it use an enum based key to sort the functions instead of math function names and also that it use the newly created constexpr multimap to avoid having to create a dynamic map to handle it. The runtime described so far is experimental. Conflicts: lib/burnside/bridge.cc
add primitives for REAL and COMPLEX comparison and unary negation Fix TypeCodeVector GCC compile issues. GCC has trouble with template specialization in classes. lowering of select and select-rank ops Implement MAX and MIN lowering Lower to MERGE + comparisons to avoid introducing branches. add primitives for REAL and COMPLEX comparison and unary negation lowering of select and select-rank ops Use SymbolRef in Burnside bridge add pure calls to CSE fix caching bug in mem2reg select case on ints remove fir processing from bbc; finish adding fir.len type Add a rough draft of a name mangling scheme. This is unofficial and a mere starting point. Rework fir.field_index and fie.len_param_index operations; fix bbc link issue work on coordinate lowering; add internal name generator add fir.constant
…time failures with the latest Xcode tools.
runtime.def path was wrong after directory merry-round and rather than fixing the pass I took the oportunity to replace it and remove old code. The new representation used the constexpr runtime description `RuntimeStaticDescription` that was already defined.
Make mono build-able again
Facilitate OpBuilder manipulations by introducing the OpBuilderWrapper class that automatically add the location and help removing the boilerplate to create intreger constants. Create a LoopCreator class that inherits this class and helps generating simple (one nesting level) loops. Create a ChararacterOpsCreator that also inherits OpBuilderWrapper and provides character copy/padding implementation.
Its interface remains unchanged, but it now inherits the recently added OpBuilderHandler class to make some op creation easier. Also remove the implementation from the header into OpBuilder.cpp so that the include file is cleaner. ComplexHandler.h is deleted.
jeanPerier
commented
Jan 16, 2020
| # distributed under the License is distributed on an "AS IS" BASIS, | ||
| # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. |
Owner
Author
There was a problem hiding this comment.
Note to myself: change licence if you push this.
jeanPerier
force-pushed
the
jpr-easier-lowering
branch
from
c22b98d to
0fcec98
Compare
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
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.
No description provided.