When computing the base addresses of an array slice to make a
descriptor, codegen generated two LLVM GEPs. The first to compute
the address of the base character element, and a second one to
compute the substring base inside that element.
The previous code did not care about getting the result of the first
GEP right: it used the base array LLVM type as the result type.
This used to work when opaque pointer were not enabled (the actual GEP
result type was probably applied in some later pass). But with opaque
pointers, the second GEP ends-up computing an offset of len*<LLVM array
type> instead of len*<character width>. A previous attempt to fix the
issue was done in D129079, but it does not cover the cases where the
array slice contains subcomponents before the substring
(e.g: array(:)%char_field(5:10)).
This patch fix the issue by computing the actual GEP result type in
codegen. There is also enough knowledge now so that a single GEP can be
generated instead of two.
Differential Revision: https://reviews.llvm.org/D129481
FirOpBuilder takes a fir::KindMapping reference. When the getKindMapping()
call is made inside the ctor call, the lifetime of this reference may
be as short as the ctor call (at least with when building flang in
release mode with clang 8). This can cause segfaults when later using
the FirOpBuilder.
Ensure the kindMap passed to the FirOpBuilder ctor is the same as the
FirOpBuilder.
Differential Revision: https://reviews.llvm.org/D129494
Flang C++ Style Guide tells us to avoid .has_value() in the predicate
expressions of control flow statements. I am treating ternary
expressions as control flow statements for the purpose of this patch.
Differential Revision: https://reviews.llvm.org/D128622
This patch is part of the upstreaming effort from fir-dev branch.
This is the last patch for the upstreaming effort.
Reviewed By: jeanPerier
Differential Revision: https://reviews.llvm.org/D129187
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
When addressing a substring of a character array, codegen emits two
GEPs: one for to compute the address of the base element, and a second
one to address the first characters from that element.
The first GEP still returns the LLVM array type (if the FIR array type could be
translated to an array type. Therefore) so zero
indexes must be added to the second GEP in this case to cover for the
Fortran array dimensions before inserting the susbtring offset index.
Surprisingly, the previous code worked ok when MLIR emits none opaque
pointers. But with opaque pointers, the two GEPs are folded in an
invalid GEP where the substring offset becomes an offset for the outer
array dimension.
Note that I tried to fix the issue by modifying the first GEP to return the
element type, but this still gave bad results (here something might be
wrong with opaque pointer in MLIR or LLVM).
Differential Revision: https://reviews.llvm.org/D129079
This patch just make the code more similar
in each conversion.
This patch is part of the upstreaming effort from fir-dev branch.
Reviewed By: jeanPerier
Differential Revision: https://reviews.llvm.org/D129071
- Add verifiers that determine if an Op requires type parameters or
not and checks that the correct number of parameters is specified.
This patch is part of the upstreaming effort from fir-dev branch.
Reviewed By: PeteSteinfeld
Differential Revision: https://reviews.llvm.org/D128828
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Added new -lower-math-early option that defaults to 'true' that matches
the current math lowering scheme. If set to 'false', the intrinsic math
operations will be lowered to MLIR operations, which should potentially
enable more MLIR optimizations, or libm calls, if there is no corresponding
MLIR operation exists or if "precise" mode is requested.
The generated math MLIR operations are then converted to LLVM dialect
during codegen phase.
The -lower-math-early option is not exposed to users currently. I plan to
get rid of the "early" lowering completely, when "late" lowering
is robust enough to support all math intrinsics that are currently
supported via pgmath. So "late" mode will become default and -lower-math-early
option will not be needed. This will effectively eliminate the mandatory
dependency on pgmath in Fortran lowering, but this is WIP.
Differential Revision: https://reviews.llvm.org/D128385
This patch restores several calls to Optional::value() in preference
to Optional::operator*.
The Flang C++ Style Guide tells us to use x.value() where no presence
test is obviously protecting a *x reference to the contents.
Differential Revision: https://reviews.llvm.org/D128590
This patch clean up some code for upstreaming and add couple of
missing tests that were left in fir-dev.
This patch is part of the upstreaming effort from fir-dev branch.
Reviewed By: jeanPerier, PeteSteinfeld
Differential Revision: https://reviews.llvm.org/D128258
Co-authored-by: Jean Perier <jperier@nvidia.com>
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Add more diagnostics to fir.coordinate_of to provide better checking
that the IR is sane.
This patch is part of the upstreaming effort from fir-dev branch.
Reviewed By: PeteSteinfeld
Differential Revision: https://reviews.llvm.org/D128255
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
The upstreamed code was not incrementing the sliceOffset in multiples
of 3. This issue is fixed by using Offsets and incrementing by 3 during
every iteration.
In the conversion pattern, we were comparing the definingOp of an
operand with an FIR::UndefOp. Use LLVM::UndefOp for conversion.
Reviewed By: clementval, Leporacanthicus
Differential Revision: https://reviews.llvm.org/D128017
This patch is part of the upstreaming effort from fir-dev branch.
It also ensures all descriptors created inline complies with LBOUND
requirement that the lower bound is `1` when the related dimension
extent is zero.
This patch is part of the upstreaming effort from fir-dev branch.
Reviewed By: jeanPerier, PeteSteinfeld
Differential Revision: https://reviews.llvm.org/D128047
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Co-authored-by: Jean Perier <jperier@nvidia.com>
Fix some mismatch in format used in the file and reduce the diff with fir-dev
to be able to finish the upstreaming on this file.
This patch is part of the upstreaming effort from fir-dev branch.
Reviewed By: PeteSteinfeld, kiranchandramohan
Differential Revision: https://reviews.llvm.org/D127849
FIR models Fortran intrinsic types with deliberate KIND values. Like
Fortran, COMPLEX and REAL have related KINDs in FIR. Lowering now
converts REAL types to floating point (MLIR) up front. This patch moves
the code to convert from FIR RealType to MLIR FloatType out of codegen
and into the builder, allowing FIR ComplexTypes to have their element
type returned as an MLIR FloatType.
We should consider whether to replace fir::ComplexType with
mlir::ComplexType at some point. I believe these types are presently
used to convey distinctins in the target ABIs in the Tilikum bridge
however.
This patch is part of the upstreaming effort from fir-dev branch.
Reviewed By: PeteSteinfeld
Differential Revision: https://reviews.llvm.org/D127636
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
This patch is part of the upstreaming effort from fir-dev branch.
Reviewed By: jeanPerier, PeteSteinfeld
Differential Revision: https://reviews.llvm.org/D127634
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
The pass was raising TODOs when a function both had a fir.boxproc<> argument
and a fir.type<> argument (even if the fir.type<> did not contain a
fir.boxproc itself).
Prevent the TODO from firing when a fir.type<> does not actually contain
a fir.boxproc. Add the location for the remaining TODO (it will be
needed when procedure pointer components are supported in lowering).
FYI, I actually tried to just implement the TODO, but I there is a funny
issue. When creating the new fir::RecordType, since the name and context
are the same as the type being translated, fir::RecordType:get just
returns the existing type, and there is no way to change it (finalize()
does nothing since it is already finalized). So this will require to add
the ability to mutate the existing type, and I am not sure what are the
MLIR constraints here, so I escaped and left the TODO for that case.
This patch is part of the upstreaming effort from fir-dev branch.
Reviewed By: jeanPerier, PeteSteinfeld
Differential Revision: https://reviews.llvm.org/D127633
Co-authored-by: Jean Perier <jperier@nvidia.com>
Remove a backwards dependence from Optimizer -> Lower by moving Todo.h
to the optimizer and out of lowering.
This patch is part of the upstreaming effort from fir-dev branch.
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Reviewed By: jeanPerier
Differential Revision: https://reviews.llvm.org/D127292
Simply add a source and target materialization handler that do nothing
and that override the default handlers that would add illegal
LLVM::DialectCastOp otherwise.
This is the simplest workaround, but not an actual fix, something may be
inconsistent after D82831 (most likely fir lowering to llvm happens in a
way that mlir infrastructure is not expecting in D82831).
Here is a minimal reproducer of what the issue was:
```
func @foop(%a : !fir.real<4>) -> ()
func @bar(%a : !fir.real<2>) {
%1 = fir.convert %a : (!fir.real<2>) -> !fir.real<4>
call @foop(%1) : (!fir.real<4>) -> ()
return
}
```
tco -o - output was:
```
error: 'llvm.mlir.cast' op type must be non-index integer types, float types, or vector of mentioned types.
llvm.func @foop(!llvm.float)
llvm.func @bar(%arg0: !llvm.half) {
%0 = llvm.fpext %arg0 : !llvm.half to !llvm.float
%1 = llvm.mlir.cast %0 : !llvm.float to !fir.real<4>
llvm.call @foop(%1) : (!fir.real<4>) -> ()
llvm.return
}
```
This patch disable the introduction of the llvm.mlir.cast and preserve the previous behavior.
Also fixes https://github.com/llvm/llvm-project/issues/55210.
Note: This is part of upstreaming from the fir-dev branch of
https://github.com/flang-compiler/f18-llvm-project.
Reviewed By: awarzynski
Differential Revision: https://reviews.llvm.org/D127212
Co-authored-by: Jean Perier <jperier@nvidia.com>
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
The previous XArrayCoorOp conversion forgot to change getting the
operands from OpAdaptor for upper bound and step of slice. This leads to
the fail of incompatible of types of codegen when slices are index type.
Reviewed By: kiranchandramohan, schweitz
Differential Revision: https://reviews.llvm.org/D125967
This patch basically extends https://reviews.llvm.org/D122008 with
support for MacOSX/Darwin.
To facilitate this, I've added `MacOSX` to the list of supported OSes in
Target.cpp. Flang already supports `Darwin` and it doesn't really do
anything OS-specific there (it could probably safely skip checking the
OS for now).
Note that generating executables remains hidden behind the
`-flang-experimental-exec` flag. Also, we don't need to add `-lm` on
MacOSX as `libm` is effectively included in `libSystem` (which is linked
in unconditionally).
Differential Revision: https://reviews.llvm.org/D125628
For arrays without a constant interior or arrays of character with
dynamic length arrays, the data types are converted to a pointer to the
element type, so the scale size of the constant extents needs to be
counted. The previous AllocaOp conversion does not consider the arrays
of character with dynamic length arrays, and the previous AllocMemOp
conversion does not consider arrays without a constant interior. This
fixes them and refactors the code so that it can be shared. Also add
the test cases.
Reviewed By: Jean Perier
Differential Revision: https://reviews.llvm.org/D124766
This patch adds lowering support for atomic read and write constructs.
Also added is pointer modelling code to allow FIR pointer like types to
be inferred and converted while lowering.
Reviewed By: kiranchandramohan
Differential Revision: https://reviews.llvm.org/D122725
Co-authored-by: Kiran Chandramohan <kiran.chandramohan@arm.com>
Fix https://github.com/flang-compiler/f18-llvm-project/issues/1416.
The `constRows` variable was being decremented too soon, causing the
last constant interior dimension extent being used to multiply the GEP
offset. This lead to wrong address computation and caused segfaults.
Note: also upstream fir.embox tests that can be upstreamed.
Differential Revision: https://reviews.llvm.org/D123130
Do not use the shift of a fir.embox to set lower bounds if there is
a fir.slice operand. This matches Fortran semantics where lower bounds
of array sections are ones.
Note that in case there is a fir.slice, the array shift may be provided
because it is used to calculate the origin/base address of an array slice.
Add a TODO for substring codegen since I noticed it was not upstreamed
yet and would cause some program to silently compile incorrectly.
Differential Revision: https://reviews.llvm.org/D123123
This commit restructures how TypeID is implemented to ideally avoid
the current problems related to shared libraries. This is done by changing
the "implicit" fallback path to use the name of the type, instead of using
a static template variable (which breaks shared libraries). The major downside to this
is that it adds some additional initialization costs for the implicit path. Given the
use of type names for uniqueness in the fallback, we also no longer allow types
defined in anonymous namespaces to have an implicit TypeID. To simplify defining
an ID for these classes, a new `MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID` macro
was added to allow for explicitly defining a TypeID directly on an internal class.
To help identify when types are using the fallback, `-debug-only=typeid` can be
used to log which types are using implicit ids.
This change generally only requires changes to the test passes, which are all defined
in anonymous namespaces, and thus can't use the fallback any longer.
Differential Revision: https://reviews.llvm.org/D122775
These are mostly small changes to make the code a bit clearer and more
consistent. Summary of changes:
* add missing namespace qualifiers (that's the preference in Flang)
* replace const member methods with static methods (to avoid passing
the *this pointer unnecessarily)
* rename `currentObjTy` (current object type) as `cpnTy` (component
type) - the latter feels more fitting
* remove redundant `return failure();` calls (` return
mlir::emitError` gives the same result)
* updated a few comments
Differential Revision: https://reviews.llvm.org/D122799
This patch addes some global initialization and global
box initialization tests.
This patch is part of the upstreaming effort from fir-dev branch.
Reviewed By: schweitz
Differential Revision: https://reviews.llvm.org/D122881
Co-authored-by: Jean Perier <jperier@nvidia.com>
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Follow-up of https://reviews.llvm.org/D121488 to ensure all descriptors
created inline complies with LBOUND requirement that the lower bound is
`1` when the related dimension extent is zero.
Both fir.xrebox and fir.xembox codegen is updated to enforce this
constraint.
Also upstream the "normalized lower bound" attribute that was added in fir-dev
since embox codegen was upstreamed, it is conflicting with this patch
otherwise.
Differential Revision: https://reviews.llvm.org/D122419
In FIR, we want to wrap function pointers in a special box known as a
boxproc value. Fortran has a limited form of dynamic scoping
[https://tinyurl.com/2p8v2hw7] between "host procedures" and "internal
procedures". There are a number of implementations possible.
Boxproc typed values abstract away the implementation details of when a
function pointer can be passed directly (as a raw address) and when a
function pointer has to account for the presence of a dynamic scope.
When lowering Fortran syntax to FIR, all function pointers are emboxed
as boxproc values.
When creating LLVM IR, we must strip away the abstraction and produce
low-level LLVM "assembly" code. This patch implements that
transformation as converting the boxproc values to either raw function
pointers or executable trampolines on the stack as needed. The
trampoline then captures the dynamic scope context within an executable
thunk that can be passed instead of the function's raw address.
Some extra handling is required for Fortran functions that return a
character value to deal with LEN values here.
Some of the code in Bridge.cpp and ConvertExpr.cpp and be re-arranged to
faciliate the upstreaming effort.
This patch is part of the upstreaming effort from fir-dev branch.
Reviewed By: jeanPerier, PeteSteinfeld
Differential Revision: https://reviews.llvm.org/D122223
Co-authored-by: mleair <leairmark@gmail.com>
Co-authored-by: Jean Perier <jperier@nvidia.com>
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Co-authored-by: V Donaldson <vdonaldson@nvidia.com>
Co-authored-by: Kiran Chandramohan <kiran.chandramohan@arm.com>
This patch adds the OpenMP conversion patterns to the FIR to LLVM
dialect lowering pass in Codegen. Appropriate legalization
conditions are also added. This ensures that a mix of FIR and OpenMP
dialects can be lowered to LLVM and OpenMP dialects. Also adds two
tests.
This is part of the upstreaming effort from the fir-dev branch in [1].
[1] https://github.com/flang-compiler/f18-llvm-project
Reviewed By: clementval, peixin
Differential Revision: https://reviews.llvm.org/D121793
Co-authored-by: Sourabh Singh Tomar <SourabhSingh.Tomar@amd.com>
Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
This removes any potential confusion with the `getType` accessors
which correspond to SSA results of an operation, and makes it
clear what the intent is (i.e. to represent the type of the function).
Differential Revision: https://reviews.llvm.org/D121762
Currently, CGOps.h and FIROps.h contain `using namespace mlir;`. Every
file that includes one of these header files (directly and transitively)
will have the MLIR namespace enabled. With name-clashes within
sub-projects (LLVM and MLIR, MLIR and Flang), this is not desired. Also,
it is not possible to "un-use" a namespace once it is "used". Instead,
we should try to limit `using namespace` to implementation files (i.e.
*.cpp).
This patch removes `using namespace mlir;` from header files and adjusts
other files accordingly. In header and TableGen files, extra namespace
qualifier is added when referring to symbols defined in MLIR. Similar
approach is adopted in source files that didn't require many changes. In
files that would require a lot of changes, `using namespace mlir;` is
added instead.
Differential Revision: https://reviews.llvm.org/D120897
The front-end and the runtime are currently using the unix logical
representation, but lowering was not. These inconsistencies could
caused issues.
The only place that defines what the logical representation is in
lowering is the translation from FIR to LLVM (FIR is agnostic to the
actual representation). More precisely, the LLVM implementation of
`fir.convert` between `i1` and `fir.logcial` is what defines the
representation:
- `fir.convert` from `i1` to `fir.logical` defines the `.true.` and `.false.`
canonical representations
- `fir.convert` from `fir.logical` to `i1` decides what the test for
truth is.
Unix representation is:
- .true. canonical integer representation is 1
- .false. canonical integer representation is 0
- the test for truth is "integer representation != 0"
For the record, the previous representation that was used was in
codegen was:
- .true. canonical integer representation is -1 (all bits 1)
- .false. canonical integer representation is 0
- the test for truth is "integer representation lowest bit == 1"
Differential Revision: https://reviews.llvm.org/D121200
The current StandardToLLVM conversion patterns only really handle
the Func dialect. The pass itself adds patterns for Arithmetic/CFToLLVM, but
those should be/will be split out in a followup. This commit focuses solely
on being an NFC rename.
Aside from the directory change, the pattern and pass creation API have been renamed:
* populateStdToLLVMFuncOpConversionPattern -> populateFuncToLLVMFuncOpConversionPattern
* populateStdToLLVMConversionPatterns -> populateFuncToLLVMConversionPatterns
* createLowerToLLVMPass -> createConvertFuncToLLVMPass
Differential Revision: https://reviews.llvm.org/D120778
