Place instructions in a doubly linked list and point to the first and last
instruction in an EBB.
Provide an iterator for all the EBBs too. This doesn't reflect the layout
order, but simply the order blocks were created.
These two enums must have identical variants. One is generated from the
instruction formats in meta/cretonne/formats.py, the other defines the contents
of an instruction.
Emit a conversion from InstructionData to InstructionFormat which also serves
to verify the correspondence. Rustc will error is the match is not complete.
Avoid gathering too much code in repr.rs.
The `entities` module contains entity reference types, and the `instructions`
module contains instruction opcodes and formats.
This is a no-payload enum which will have the same variants as InstructionData.
This makes it possible to talk about the format of an instruction without
actually creating an InstructionData instance.
We have a two-level type system: OperandKinds and ValueTypes. The value types
only apply to value operands, but there are many more kinds of operands:
immediate numbers, condition codes, basic block references, etc.
Values that are defined together are represented as a singly linked list. These
lists appear in:
- Instructions with multiple result values. The first result value is special, and the
following results form a linked list of Def extended_value table entries.
- EBB arguments are represented as a linked list of Argument extended_value
table entries. The EbbData struct has pointers to the first and last argument
to allow fast insertion at both ends.
Add a Values iterator type whicih can enumerate both kinds of value lists.
When Function serves as a container for IL entities, use the Index trait to
translate a reference class to a Data object.
Works for:
- StackSlot -> StackSlotData
- Inst -> InstructionData
libctonfile -> libreader.
This library will only provide .cton file reading/parsing services which are
not needed after deployment.
Code for writing .cton files lives in the main cretonne library because it is
fairly small, and because it is useful for extracting test cases from a
deployed library.
Add a stack slot array to repr::Function, use repr::StackSlot to reference them.
Parse stack slot declarations in the function preamble, add them to the
function.
Add a new `Context` struct which keeps track of mappings between identifiers
used in the file and real references.
Don't use assertions to enforce the limits on SIMD lanes in a type, Type is too
fundamental for that. Instead, the Vector-forming by() method returns an
Optional<Type>, and None if the requested SIMD vector is not valid.
Same for Type::half_vector().
These are bitwise exact conversions from string to immediates, implementing the
inverse of the Display trait.
Only accept hexadecimal floating point numbers to avoid issues with rounding
when converting decimal numbers to binary.
The src/tools directory contains the cretonne-tools crate which will build
binaries for testing cretonne.
The src/libctonfile directory contains the ctonfile library crate which
provides reading and writing of .cton files.
Use a simple quadratically probed, open addressed hash table. We could use a
parfect hash function, but it would take longer to compute in Python, and this
is not in the critical path performancewise.
A function owns instructions and extended basic blocks. References to these
entities are implemented as opaque structs indexing into the functions internal
tables. This avoids fighting Rust's ownership checking and it also makes
references 4 bytes on all platforms.
SSA values are identified similarly, but with an optimization for the first
value produced by an instruction. Very few instructions will produce more than
one value, and there is an extended value table for those.
Generate nice doc comments for the Opcode enum variants that 'cargo doc'
will pick up.
Include a `Display` trait implementation that prints the lower
snake-case version of the opcode name.
Clarify the textual encoding of floating point numbers.
Don't allow decimal floating point since conversion to/from binary can produce
rounding problems on some (buggy) systems.
Add new immediate types for floating point and vector immediates.
Use new immediates to define the constant value instructions in meta.
Split the fconst instruction into two: f32const and f64const. This prevents
confusion about the interpretation of 64 immediate bits when generating an f32
constant.
Add an immvector ImmediateType. This immediate type is variable length, and
provides all the bits of a SIMD vector directly.
Jakob Stoklund Olesen
Jakob Stoklund Olesen