The Builder provides a convenient interface for inserting instructions
into an extended basic block.
The bulk of the builder methods are generated automatically from the
meta language instruction descriptions.
Still TODO: Keep track of an insertion position.
In instruction formats that have multiple results AND boxed storage,
place the second_result field outside the boxed storage. The 16-byte
instruction format has room for opcode, type, second_result in the first
8 bytes, and the boxed pointer in the last 8 bytes.
This provides a simpler implementation of the second_result() and
second_result_mut() InstructionData methods.
The Rust type is usually the camel-cased name of the operand kind, but
there are variations, so allow an explicit rust_type='IntCC' when
defining operand kinds.
These refer to external functions and function signatures declared in
the preamble. Since we're already using the type names 'Signature' and
'Function', these entity references don't folow the usual EntityData /
Entity naming convention.
Entity references in instruction format operands also have member names
in the InstructionData struct. Track them the same way we track immediate operand member names.
Value operands still go in the arg / args members.
Enable syntax: iadd(x, y) which creates an Apply node.
Enable syntax: z << iadd(x, y) which creates a Def node.
Add an XForm class which represents source and destination patterns as
RTL lists.
The isplit_lohi instruction breaks an integer into two halves. This will
typically be used to get the two halves of an `i64` value on 32-bit
CPUs.
The iconcat_lohi is the reverse operation. It reconstructs the `i64`
from the low and high bits.
The # is a more conventional prefix for hexadecimal, and when ISA
information is not available, there may be a decimal number in front
which would be confusing.
So prefer [1#10c] for the ISA-less encoding format. Here '1' is decimal
and '#10c' is hexadecimal.
When constructing the Flags object from the Builder, don't consume it,
but take a reference instead.
This makes it possible for the parser to accept multiple 'set' lines and
apply them to different ISA specifications.
Use the new ISA predicate numbering to emit ISA predicate instructions in the
encoding tables.
Properly decode the ISA predicate number in RISC-V and add tests for RV32M iwth
and without 'supports_m' enabled.
Move all the byte-sized settings to the front of the byte-vector, and add a
mechanism for assigning numbers to predicates that have no name as well as
predicates from the parent settings group.
This way, all the boolean predicates that are used by a target ISA appear as a
contiguous bit-vector that is a suffix of the settings byte-vector. This
bit-vector can then be indexed linearly when resolving ISA predicates on
encodings.
Add a numbered_predicate() method to the generated Flags structs that can read
a predicate by number dynamically.
This is just the basic 'imul' the M instruction set also has mulh/mulhu which
yield the high bits of a multiplication, and there are div/rem instructions to
be implemented.
These instructions are gated by the use_m predicate, but ISA predicates are not
completely implemented yet.
Three predicates affect each extension:
- supports_m determines whether the target CPU supports the instruction set.
- enable_m determines if the instructions should be used, assuming they're
available.
- use_m is the predicate used to actually use the instructions.
Define data types for the level 1 and level 2 hash tables. These data types are
generic over the offset integer type so they can be twice as compact for
typically small ISAs.
Use these new types when generating encoding hash tables.
Emit both level 1 and level 2 hash tables.
Define generic functions that perform lookups in the encoding tables.
Implement the TargetIsa::encode() method for RISC-V using these building
blocks.
Compute the u16 representation of encoding lists and emit a big table
concatenating all of them. Use the UniqueSeqTable to share some table space
between CPU modes.
We need to generate hash tables keyed by types, so the Python scripts need to
know the index used to represent types in Rust code.
To enforce this, add a new gen_types.py script which generates constant
definitions for the ir/types module.
Also generate constants for common SIMD vector sizes.
Amend build script to generate an encodings-<isa>.rs file for each target ISA.
Emit a function that can evaluate instruction predicates.
Describe the 3-level tables used for representing insrruction encoding tables.
Add Python classes representing the tables.
The generated code is incomplete and not used anywhere yet.
When generating Rust code for an instruction predicate, call the corresponding
function in the predicates module, using a qualified name.
We don't have methods corresponding to the predicates.
This collects all of the leaf predicates that go into a compound predicate.
Current leaf predicates are:
- Settings for ISA predicates, and
- FieldPredicates for instruction predicates.
Add new instruction predicates to support the 'I' encoding recipe: IsSignedInt,
IsUnsignedInt used to test that an immediate operand is in the allowed range.
Each InstructionFormat instance gets data members corresponding to its immediate
operands, so the can be referred to as BinaryImm.imm, for example.
This will be used to construct instruction predicates.
Usually an instruction firmat has only a single immediate operand called 'imm',
or 'cond' if it is one of the condigtion codes. Add a 'default_member' field to
ImmediateKind to keep track of this default member name in the InstructionData
struct.
