| Age | Commit message (Collapse) | Author |
|---|
|
|
|
* Change AnnotationSeq underlying from List to Seq
It was nothing but pointless copying.
* Make propagateAnnotations faster
There was lots of expensive logic for very little benefit.
|
|
* split big Emitter to submodules.
* fix all deprecated warning.
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
|
|
|
|
* Absorb WRef into Reference
* Absorb WSubField into SubField
* Absorb WSubIndex into SubIndex
* Absorb WSubAccess into SubAccess
* Absorb WDefInstance into DefInstance
------------------------- API CHANGE SEVERITY --------------------------
This is projected to not break source-level compatibility with any known
user code. However, it will break *binary* compatibility with all
existing user FIRRTL passes, as is generally allowed with major
releases of FIRRTL.
--------------------------- DESCRIPTION --------------------------------
Previously, there were several nodes in WIR.scala that had a one-to-one
correspondance with existing nodes in the standard firrtl.ir hierarchy.
These nodes would have a case class resembling the corresponding
standard IR node, but with the addition of one or more "analysis"
fields.
Since these fields (such as kind) represent helpful info that can be
invalidated or set to Unknown (e.g. UnknownKind for Kind), it does not
cause any issues to simply include these fields on any in-memory
representation of FIRRTL IR. Although other systems for tracking FIRRTL
analyses have evolved over time, the ubiquity of pattern-matching on
these fields has lead most core and custom transforms to be written
against WIR, rather than IR.
This PR unifies the IRs by adding the fields that would be in an
"augmented" WIR node directly into the corresponding IR node; i.e., the
"type" and "kind" fields from WRef are added directly to the definition
of the Reference case class, while these "repetitive" WIR case classes
are removed entirely.
-------------------- SOURCE-COMPATIBILITY ADAPTERS ---------------------
Several object methods are added to WIR.scala to maintain
source-compatiblity for passes that used WIR. These objects define
factory methods and unapply methods, so passes that relied on implicit
case class factories or pattern matching for the removed WIR types will
remain perfectly source-compatible. However, these do not guarantee
compatibility at the binary level.
The types of the removed WIR case classes are also added as type aliases
to the top-level firrtl package, which allows code that relies on
explicit constructor calls or reflection to retain source-compatibility.
Finally, additional explicit factory methods are added to the companion
objects of the newly-augmented IR case classes, which allows user code
to avoid having to specify any of the new analysis fields. Existing code
that created non-WIR IR nodes will be able to continue using the
previous factory signatures, which will cause all omitted analysis
fields to be set to Unknown.
---------------------- UNMITIGATED API CHANGES -------------------------
While passes that used WIR will be source-compatible with this change,
there is one significant change that affects any pass currently using
non-WIR IR: the signatures of pattern-matching cases for Reference,
SubField, SubIndex, SubAccess, and DefInstance must change to
accommodate the extra fields.
This cannot be worked at the API level due to restrictions on unapply
overloading, but it could theoretically be solved with macros or other
static rewriting. However, only four core transforms (RemoveProto,
ToWorkingIR, Dedup, and RemoveChirrtl) use non-WIR IR, and it is
expected that no user code currently relies on it, so the expected
migration strategy is simply to change the small fraction of code
relying on these nodes.
|
|
|
|
(#1463)
* Explicitly initialize firrtl.stage.Forms to prevent multi-thread collisions
See https://github.com/freechipsproject/firrtl/issues/1462.
Convert `lazy val` members of firrtl.stage.Forms to plan `val`s.
Reference firrtl.stage.Forms in sufficient locations to ensure the object is initialized before its members are accessed.
* Respond to comments - make _dummyForms private.
* Move Forms initialization to package object.
* Merge with master
|
|
Signed-off-by: Schuyler Eldridge <schuyler.eldridge@ibm.com>
|
|
* Add sbt-scalafix
* Add scalafix guide to README
* Remove Unused Import
* Remove deprecated procedure syntax
|
|
This adds FirrtlStage, a reimplementation of the original FIRRTL
Driver as a Stage. This updates the original firrtl.options package to
implement FirrtlStage (namely, TransformLike is added) along with
FirrtlMain. Finally, the original FIRRTL Driver is converted to a
compatibility wrapper around FirrtlStage.
For background, Stage and Phase form the basis of the Chisel/FIRRTL
Hardware Compiler Framework (HCF). A Phase is a class that performs a
mathematical transformation on an AnnotationSeq (in effect, a
generalization of a FIRRTL transform). Curtly, a Stage is a Phase that
also provides a user interface for generating annotations. By their
construction, Phases are designed to be composed sequentially into a
transformation pipeline.
This modifies the existing options package (which provides
Stage/Phase) to build out a type hierarchy around Stage/Phase. This
adds TransformLike[A] which implements a mathematical transformation
over some type A. Additionally, and as an interface between different
TransformLikes, this adds Translator[A, B] which extends
TransformLike[A], but does an internal transformation over type B.
This is used to interface Phases with the existing FIRRTL compiler.
This adds a runTransform method to Phase that, like
Transform.runTransform, will automatically detect deleted Annotations
and generate DeletedAnnotations.
The new FirrtlStage, a reimplementation of FIRRTL's Driver, is added
as a Stage composed of the following Phases:
1. AddDefaults - add default annotations
2. AddImplicitEmitter - adds an implicit emitter derived from the
compiler
3. Checks - sanity check the AnnotationSeq
4. AddCircuit - convert FIRRTL input files/sources to circuits
5. AddImplicitOutputFile - add a default output file
6. Compiler - run the FIRRTL compiler
7. WriteEmitted - write any emitted modules/circuits to files
The Driver is converted to a compatibility layer that replicates old
Driver behavior. This is implemented by first using new toAnnotation
methods for CommonOptions and FirrtlExecutionOptions that enable
AnnotationSeq generation. Second, the generated AnnotationSeq is
preprocessed and sent to FirrtlStage. The resulting Phase order is
then:
1. AddImplicitAnnotationFile - adds a default annotation file
2. AddImplicitFirrtlFile - adds a default FIRRTL file using top name
3. AddImplicitOutputFile - adds an output file from top name
4. AddImplicitEmitter - adds a default emitter derived from a
compiler and any split modules command line option
5. FirrtlStage - the aforementioned new FirrtlStage
Finally, the output AnnotationSeq is then viewed as a
FirrtlExecutionResult. This compatibility layer enables uninterrupted
usage of old Driver infrastructure, e.g., FirrtlExecutionOptions and
CommonOptions can still be mutated directly and used to run the
Driver.
This results in differing behavior between the new FirrtlStage and the
old Driver, specifically:
- FirrtlStage makes a clear delineation between a "compiler" and an
"emitter". These are defined using separate options. A compiler is
"-X/--compiler", while an emitter is one of "-E/--emit-circuit" or
"-e/--emit-modules".
- Related to the above, the "-fsm/--split-modules" has been removed
from the FirrtlStage. This option is confusing once an implicit
emitter is removed. It is also unclear how this should be handled
once the user can specify multiple emitters, e.g., which emitter
should "--split-modules" apply to?
- WriteOutputAnnotations will, by default, not write
DeletedAnnotations to the output file.
- The old top name ("-tn/--top-name") option has been removed from
FirrtlStage. This option is really a means to communicate what
input and output files are as opposed to anything associated with
the circuit name. This option is preserved for the Driver
compatibility layer.
Additionally, this changes existing transform scheduling to work for
emitters (which subclass Transform). Previously, one emitter was
explicitly scheduled at the end of all transforms for a given
compiler. Additional emitters could be added, but they would be
scheduled as transforms. This fixes this to rely on transform
scheduling for all emitters. In slightly more detail:
1. The explicit emitter is removed from Compiler.compile
2. An explicit emitter is added to Compiler.compileAndEmit
3. Compiler.mergeTransforms will schedule emitters as late as
possible, i.e., all emitters will occur after transforms that
output their input form.
4. All AddImplicitEmitter phases (DriverCompatibility and normal)
will add RunFirrtlTransformAnnotations to add implicit emitters
The FIRRTL fat jar utilities are changed to point at FirrtlStage and not
at the Driver. This has backwards incompatibility issues for users
that are using the utilities directly, e.g., Rocket Chip.
The Logger has been updated with methods for setting options based on
an AnnotationSeq. This migrates the Logger to use AnnotationSeq as
input parameters, e.g., for makeScope. Old-style methods are left in
place and deprecated. However, the Logger is not itself a Stage.
The options of Logger Annotations are included in the base Shell and
Stage is updated to wrap its Phases in a Logger scope.
Additionally, this changes any code that does option parsing to always
prepend an annotation as opposed to appending an annotation. This is
faster, but standardizing on this has implications for dealing with
the parallel compilation annotation ordering.
A Shell will now put the initial annotations first (in the order the
user specified) and then place all annotations generating from parsing
after that. This adds a test case to verify this behavior.
Discovered custom transforms (via `RunFirrtlTransformAnnotation`s) are
discovered by the compiler phase in a user-specified order, but are
stored in reverse order to more efficiently prepend (as opposed to
append) to a list. This now reverses the transform order before
execution to preserve backwards compatibility of custom transform
ordering.
The Compiler phase also generates one deleted annotation for each
`RunFirrtlTransformAnnotation`. These are also reversed.
Miscellaneous small changes:
- Split main method of Stage into StageMain class
- Only mix in HasScoptOptions into Annotation companion objects (h/t
@jackkoenig)
- Store Compiler in CompilerAnnotation
- CompilerNameAnnotation -> CompilerAnnotation
- Make Emitter abstract in outputSuffix (move out of FirrtlOptions)
- Add DriverCompatibility.AddImplicitOutputFile that will add an
output file annotation based on the presence of a
TopNameAnnotation. This is important for compatibility with the
old Driver.
- Cleanup Scaladoc
- Refactor CircuitOption to be abstract in "toCircuit" that converts
the option to a FirrtlCircuitAnnotation. This allows more of the
conversion steps to be moved out of AddCircuit and into the actual
annotation.
- Add WriteDeletedAnnotation to module WriteOutputAnnotations
- A method for accessing a FirrtlExecutionResultView is exposed in
FIRRTL's DriverCompatibilityLayer
- Using "--top-name/-tn" or "--split-modules/-fsm" with FirrtlStage
generates an error indicating that this option is no longer
supported
- Using FirrtlStage without at least one emitter will generate a
warning
- Use vals for emitter in Compiler subclasses (these are used to
build RunFirrtlTransformAnnotations and the object should be
stable for comparisons)
- Fixes to tests that use LowTransformSpec instead of
MiddleTransformSpec. (SimpleTransformSpec is dumb and won't
schedule transforms correctly. If you rely on an emitter, you need
to use the right transform spec to test your transform if you're
relying on an emitter.)
Signed-off-by: Schuyler Eldridge <schuyler.eldridge@ibm.com>
|
|
TargetDirAnnotation was moved from firrtl to firrtl.stage. However, this
is only aliased as a val in the firrtl package object. This also needs to
be type aliased for matching against a type. This fixes a bug I ran across
in the visualizer.
Signed-off-by: Schuyler Eldridge <schuyler.eldridge@ibm.com>
|
|
Signed-off-by: Schuyler Eldridge <schuyler.eldridge@ibm.com>
|
|
- Old Annotation renamed to deprecated LegacyAnnotation
- Annotation is now a trait that can be extended
- New JsonProtocol for Annotation [de]serialization
- Replace AnnotationMap with AnnotationSeq
- Deprecate Transform.getMyAnnotations
- Update Transforms
- Turn on deprecation warnings
- Remove deprecated Driver.compile
- Make AnnotationTests abstract with Legacy and Json subclasses
- Add functionality to convert LegacyAnnotations of built-in annos
This will give a noisy warning and is more of a best effort than a
robust solution.
Fixes #475 Closes #609
|
