3. CGEN's Register Transfer Language

CGEN uses a variant of GCC's Register Transfer Language as the basis for its CPU description language.

3.1 RTL Introduction

The description language, or RTL (11), needs to support the definition of all the architectural and implementation features of a CPU, as well as enough information for all intended applications. At present this is just the opcodes table and an ISA level simulator, but it is not intended that applications be restricted to these two areas. The goal is having an application independent description of the CPU. In the end that's a lot to ask for from one language. Certainly gate level specification of a CPU is not attempted!

The syntax of the language is inspired by GCC's RTL and by the Scheme programming language, theoretically taking the best of both. To what extent that is true, and to what extent that is sufficient inspiration is certainly open to discussion. In actuality, there isn't much difference here from GCC's RTL that is attributable to being Scheme-ish. One important Scheme-derived concept is arbitrary precision of constants. Sign or zero extension of constants in GCC has always been a source of problems. In CGEN'S RTL constants have modes and there are both signed and unsigned modes.

Here is a graphical layout of the hierarchy of elements of a ‘.cpu’ file.

                           architecture
                          /            \
                    cpu-family1        cpu-family2  ...
                      /     \            /      \
                machine1   machine2  machine3   ...
                 /   \
             model1  model2  ...

Each of these elements is explained in more detail below. The architecture is one of ‘sparc’, ‘m32r’, etc. Within the ‘sparc’ architecture, cpu-family might be ‘sparc32’, ‘sparc64’, etc. Within the ‘sparc32’ CPU family, the machine might be ‘sparc-v8’, ‘sparclite’, etc. Within the ‘sparc-v8’ machine classification, model might be ‘hypersparc’, ‘supersparc’, etc.

Instructions form their own hierarchy as each instruction may be supported by more than one machine. Also, some architectures can handle more than one instruction set on one chip (e.g. ARM).

                     isa
                      |
                 instruction
                    /   \	   
             operand1  operand2  ... 
                |         |
         hw1+ifield1   hw2+ifield2  ...

Each of these elements is explained in more detail below.

3.2 Trade-offs

While CGEN is written in Scheme, this is not a requirement. The description language should be considered absent of any particular implementation, though certainly some things were done to simplify reading ‘.cpu’ files with Scheme. Scheme related choices have been made in areas that have no serious impact on the usefulness of the CPU description language. Places where that is not the case need to be revisited, though there currently are no known ones.

One place where the Scheme implementation influenced the design of CGEN's RTL is in the handling of modes. The Scheme implementation was simplified by treating modes as an explicit argument, rather than as an optional suffix of the operation name. For example, compare (add SI dr sr) in CGEN versus (add:SI dr sr) in GCC RTL. The mode is treated as optional so a shorthand form of (add dr sr) works.

3.3 Rules and notes

A few basic guidelines for all entries:

• Names must be valid Scheme symbols.
• Comments are used, for example, to comment the generated C code (12).
• Comments may be any number of lines, though generally succinct comments are preferable(13).
• Everything is case sensitive.(14)
• While "_" is a valid character to use in symbols, "-" is preferred
• Hex numbers are written using Scheme's notation. Write 255 in hex as #xff, not 0xff. One can also use #bNNN to write boolean values. E.g. #b111 == 7.
• Except for the ‘comment’ and ‘attrs’ fields and unless otherwise specified all fields must be present.
• Symbols used to be allowed anywhere a string can be used. This is what earlier versions of Guile supported. Guile is more strict now, so this relaxation is gone. The reverse is generally not allowed, strings can't be used in place of symbols.
• Use ‘()’ or ‘#f’ to indicate “not specified”, unless otherwise specified. This is not necessary for ‘define-foo’ elements, one can just elide the entry, but it is useful for ‘define-*-foo’ that take a fixed number of arguments. E.g., ‘define-normal-ifield’. Whether to use ‘()’ or ‘#f’ is largely a matter of style.

3.4 RTL Versions

CGEN has minimal support for making changes to the language without breaking existing ports. We do not put much effort into this because over time it can become unmaintainable, but for some changes it is useful to have a temporary window in which older versions are supported.

3.4.1 Specifying the RTL version

Specify the version of RTL that your cpu description was written to with ‘define-rtl-version’.

Syntax:

(define-rtl-version major-version minor-version)

When setting the RTL version, it must be the first thing done in the description file or the behaviour is undefined. This includes using or defining pmacros, the RTL version must be set first. After the RTL version is set, if it is changed the behavior is undefined.

Note that one can still set it to the same version multiple times. This is useful when the description is spread among several files, and one is debugging/testing files individually.

The default RTL version, if ‘define-rtl-version’ is elided, is 0.7.

The latest RTL version is 0.9:

(define-rtl-version 0 9)

Every increment in major and minor versions is generally non-upward compatible (otherwise the version would not have been incremented - CGEN does not keep support for older versions long).

3.4.2 List of supported RTL versions

CGEN currently supports the following RTL versions.

• 0.7 (define-rtl-version 0 7)

  This is the original RTL version. It is the default if no version is specified. It is supported by CGEN versions 1.0, 1.1, and the current development tree. Support for it will probably be removed for the CGEN 1.2 release.

• 0.8 (define-rtl-version 0 8)

  This version changed the syntax for defining keywords. See section Keywords. The ‘print-name’ field was renamed to ‘enum-prefix’ and the ‘prefix’ field was renamed to ‘name-prefix’.

  Previous syntax:

  (define-keyword (name keyword-name) (comment "description") (attrs attribute-list) (mode mode-name) (print-name "prefix-for-enum-values-with-trailing-dash") (prefix "prefix-for-names-in-string-table") (values value-list) )

  New syntax:

  (define-keyword (name keyword-name) (comment "description") (attrs attribute-list) (mode mode-name) (enum-prefix "prefix-for-enum-values") (name-prefix "prefix-for-names-in-string-table") (values value-list) )

  Note that ‘print-name’ has been replaced with ‘enum-prefix’ and ‘prefix’ has been replaced with ‘name-prefix’.

  Furthermore, there is also a difference between the behavior of ‘print-name’ and ‘enum-prefix’. When computing complete enum names with ‘print-name’, CGEN adds a ‘-’ between the prefix and the enum name. CGEN does not insert a ‘-’ with ‘enum-prefix’.

• 0.9 (define-rtl-version 0 9)

  This version changed the prefix of pmacros from ‘.’ to ‘%’. ‘.pmacro’ is changed to ‘%pmacro’.

3.5 Top level conditionals

CGEN supports conditionally defining objects through the use of ‘if’ and some specialized predicates. These must appear at the “top level”, i.e., not inside any other expression, except ‘begin’.

The following predicates are supported:

• (keep-isa? (isa-list)) Return “true” if any ISA in ‘isa-list’ is being kept. This is controlled by the ‘-i’ option.
• (keep-mach? (machine-list)) Return “true” if any machine in ‘machine-list’ is being kept. This is controlled by the ‘-m’ option.
• (application-is? application) Return “true” if the current application generator is ‘application’. Currently valid applications are:
  • DOC: ISA documentation generated from the description file(s).
  • OPCODES: Binutils support
  • GAS-TEST: GAS testsuite generator
  • SID-SIMULATOR: The SID simulator
  • SIMULATOR: The GNU simulator
  • SIM-TEST: The GNU simulator testsuite generator
  • TESTSUITE: The CGEN testsuite
• (rtl-version-equal? major minor) Return “true” if the RTL version specified by the ‘.cpu’ file is ‘major.minor’.
• (rtl-version-at-least? major minor) Return “true” if the RTL version specified by the ‘.cpu’ file is at least ‘major.minor’.

Here's an example from the CGEN testsuite. It is used to write some wrappers around a few builtin pmacros that are independent of the pmacro prefix character.

(if (rtl-version-at-least? 0 9)
    (begin
      (define-pmacro /begin %begin)
      (define-pmacro /print %print)
      (define-pmacro /dump %dump))
    (begin
      (define-pmacro /begin .begin)
      (define-pmacro /print .print)
      (define-pmacro /dump .dump)))

Here's an example from the ‘SH’ cpu description.

(if (keep-isa? (compact))
    (include "sh64-compact.cpu"))

(if (keep-isa? (media))
    (include "sh64-media.cpu"))

3.6 Definitions

Each entry has the same format: (define-foo arg1 arg2 ...), where ‘foo’ designates the type of entry (e.g. define-insn). In the general case each argument is a name/value pair expressed as (name value). (*Note: Another style in common use is `:name value' and doesn't require parentheses. Maybe that would be a better way to go here. The current style is easier to construct from macros though.)

While the general case is flexible, it also is excessively verbose in the normal case. To reduce this verbosity, a second version of most define-foo's, generally named ‘define-normal-foo’ or ‘define-simple-foo’, exist that takes a fixed number of positional arguments. With pmacros they can be even shortened further to just their acronym. E.g. ‘define-normal-ifield’ -> ‘dnf’. Ports are free to write their own preprocessor macros to simplify things further as desired. See sections titled “Simplification macros” later in this chapter.

3.7 Attributes

Attributes are used throughout for specifying various properties. For portability reasons attributes can only have 32 bit integral values (signed or unsigned).

There are four kinds of attributes: boolean, integer, enumerated, and bitset. Boolean attributes can be achieved via others, but they occur frequently enough that they are special cased (and one bit can be used to record them). Bitset attributes are a useful simplification when one wants to indicate an object can be in one of many states (e.g. an instruction may be supported by multiple machines).

String attributes might be a useful addition. Another useful addition might be functional attributes (the attribute is computed at run-time - currently all attributes are computed at compile time). One way to implement functional attributes would be to record the attributes as byte-code and lazily evaluate them, caching the results as appropriate. The syntax has been done to not preclude either as an upward compatible extension.

Attributes must be defined before they can be used. There are several predefined attributes for entry types that need them (instruction field, hardware, operand, and instruction). Predefined attributes are documented in each relevant section.

In C applications an enum is created that defines all the attributes. Applications that wish to have some architecture independent-ness need the attribute to have the same value across all architectures. This is achieved by giving the attribute the INDEX attribute (15), which specifies the enum value must be fixed across all architectures.

Convention requires attribute names consist of uppercase letters, numbers, "-", and "_", and must begin with a letter. To be consistent with Scheme, "-" is preferred over "_".

3.7.1 Boolean Attributes

Boolean attributes are defined with:

(define-attr
  (type boolean)
  (for user-list)
  (name attribute-name)
  (comment "attribute comment")
  (attrs attribute-attributes)
  (values #f #t)
  (default #f)
)

The default value of boolean attributes is always false. This can be relaxed, but it's one extra complication that is currently unnecessary. Boolean attributes are specified in either of two forms: (NAME expr), NAME, and !NAME. The first form is the canonical form. The latter two are shorthand versions. NAME means "true" and !NAME means "false". ‘expr’ is either #f or #t.

user-list is a space separated list of entry types that will use the attribute. Possible values are: ‘attr’, ‘enum’, ‘cpu’, ‘mach’, ‘model’, ‘ifield’, ‘hardware’, ‘operand’, ‘insn’ and ‘macro-insn’. If omitted all are considered users of the attribute.

The values and default fields if provided must have the indicated values. Usually these fields are elided.

3.7.2 Integer Attributes

Integer attributes are defined with:

(define-attr
  (type integer)
  (for user-list)
  (name attribute-name)
  (comment "attribute comment")
  (attrs attribute-attributes)
  (default integer-value)
)

If omitted, the default is 0.

Integer attributes are specified with (NAME value).

3.7.3 Enumerated Attributes

Enumerated attributes are the same as integer attributes except the range of possible values is restricted and each value has a name. Enumerated attributes are defined with

(define-attr
  (type enum)
  (for user-list)
  (name attribute-name)
  (comment "attribute comment")
  (attrs attribute-attributes)
  (values enum-value1 enum-value2 ...)
  (default default-enum-value)
)

If omitted, the default is the first entry in values.

Enum attributes are specified with (NAME enum-value).

3.7.4 Bitset Attributes

Bitset attributes are for situations where you want to indicate something is a subset of a small set of possibilities. The MACH attribute uses this for example to allow specifying which of the various machines support a particular insn. (*Note: At present the maximum number of possibilities is 32. This is an implementation restriction which can be relaxed, but there's currently no rush.)

Bitset attributes are defined with:

(define-attr
  (type bitset)
  (for user-list)
  (name attribute-name)
  (comment "attribute comment")
  (attrs attribute-attributes)
  (values enum-value1 enum-value2 ...)
  (default default-value1 default-value2 ...)
)

The default values must be from the specified values. The default must be provided, it may not be omitted.

Bitset attributes are specified with (NAME val1 val2 ...).

For backward compatibility they may also be specified with (NAME val1,val2,...) or (NAME "val1,val2,..."), there must be no spaces in “val1,val2,...” and each value must be a valid Scheme symbol. Use of (NAME val1,val2,...) is deprecated, and support for it will go away at some point.

3.8 Architecture variants

The base architecture and its variants are described in four parts: define-arch, define-isa, define-cpu, and define-mach.

3.8.1 define-arch

define-arch describes the overall architecture, and must be present.

The syntax of define-arch is:

(define-arch
  (name architecture-name) ; e.g. m32r
  (comment "description")  ; e.g. "Mitsubishi M32R"
  (attrs attribute-list)
  (default-alignment aligned|unaligned|forced)
  (insn-lsb0? #f|#t)
  (machs mach-name-list)
  (isas isa-name-list)
)

3.8.1.1 default-alignment

Specify the default alignment to use when fetching data (and instructions) from memory. At present this can't be overridden, but support can be added if necessary. The default is aligned.

3.8.1.2 insn-lsb0?

Specifies whether the most significant or least significant bit in a word is bit number 0. Generally this should conform to the convention in the architecture manual. This is independent of endianness and is an architecture wide specification. There is no support for using different bit numbering conventions within an architecture.

Instruction fields are always numbered beginning with the most significant bit. That is, the `start' of a field is always its most significant bit. For example, a 4 bit field in the uppermost bits of a 32 bit instruction would have a start/length of (31 4) when insn-lsb0? = #t, and (0 4) when insn-lsb0? = #f.

3.8.1.3 mach-name-list

The list of names of machines in the architecture. There should be one entry for each define-mach.

3.8.1.4 isa-name-list

The list of names of instruction sets in the architecture. There must be one for each define-isa. An example of an architecture with more than one is the ARM which has a 32 bit instruction set and a 16 bit "Thumb" instruction set (the sizes here refer to instruction size).

3.8.2 define-isa

define-isa describes aspects of the instruction set. A minimum of one ISA must be defined.

The syntax of define-isa is:

(define-isa
  (name isa-name)
  (comment "description")
  (attrs attribute-list)
  (default-insn-word-bitsize n)
  (default-insn-bitsize n)
  (base-insn-bitsize n)
  ; (decode-assist (b0 b1 b2 ...)) ; generally unnecessary
  (liw-insns n)
  (parallel-insns n)
  (condition ifield-name expr)
  (setup-semantics expr)
  ; (decode-splits decode-split-list) ; support temporarily disabled
  ; ??? missing here are fetch/execute specs
)

3.8.2.1 default-insn-word-bitsize

Specifies the default size of an instruction word in bits. This affects the numbering of field bits in words beyond the base instruction. See section Instruction fields, for more information.

3.8.2.2 default-insn-bitsize

The default size of an instruction in bits. It is generally the size of the smallest instruction. It is used when parsing instruction fields. It is also used by the disassembler to know how many bytes to skip for unrecognized instructions.

3.8.2.3 base-insn-bitsize

The minimum size of an instruction, in bits, to fetch during execution. If the architecture has a variable length instruction set, this is the size of the initial word to fetch. There is no need to specify the maximum length of an instruction, that can be computed from the instructions. Examples:

i386

8

M68k

16

SPARC

32

M32R

32

The M32R case is interesting because instructions can be 16 or 32 bits. However instructions on 32 bit boundaries can always be fetched 32 bits at a time as 16 bit instructions always come in pairs.

3.8.2.4 decode-assist

Override CGEN's heuristics about which bits to initially use to decode instructions in a simulator. For example on the SPARC these are bits: 31 30 24 23 22 21 20 19. The entire decoder can be machine generated, so this field is entirely optional. Since the heuristics are quite good, you should only use this field if you have evidence that you can pick a better set, in which case the CGEN developers would like to hear from you!

??? It might be useful to provide greater control, but this is sufficient for now.

It is okay if the opcode bits are over-specified for some instructions. It is also okay if the opcode bits are under-specified for some instructions. The machine generated decoder will properly handle both these situations. Just pick a useful number of bits that distinguishes most instructions. It is usually best to not pick more than 8 bits to keep the size of the initial decode table down.

Bit numbering is defined by the insn-lsb0? field.

3.8.2.5 liw-insns

The number of instructions the CPU always fetches at once. This is intended for architectures like the M32R, and does not refer to a CPU's ability to pre-fetch instructions. The default is 1.

3.8.2.6 parallel-insns

The maximum number of instructions the CPU can execute in parallel. The default is 1.

??? Rename this to max-parallel-insns?

3.8.2.7 condition

Some architectures like ARM and ARC conditionally execute every instruction based on the condition specified by one instruction field. The condition spec exists to support these architectures. ifield-name is the name of the instruction field denoting the condition and expression is an RTL expressions that returns the value of the condition (false=zero, true=non-zero).

3.8.2.8 setup-semantics

Specify a statement to be performed prior to executing particular instructions. This is used, for example, on the ARM where the value of the program counter (general register 15) is a function of the instruction (it is either pc+8 or pc+12, depending on the instruction).

3.8.2.9 decode-splits

Specify a list of field names and values to split instructions up by. This is used, for example, on the ARM where the behavior of some instructions is quite different when the destination register is r15 (the pc).

The syntax is:

(decode-splits
  (ifield1-name
   constraints
   ((split1-name (value1 value2 ...)) (split2-name ...)))
  (ifield2-name
   ...)
)

constraints is work-in-progress and should be () for now.

One copy of each instruction satisfying constraint is made for each specified split. The semantics of each copy are then simplified based on the known values of the specified instruction field.

3.8.3 define-cpu

define-cpu defines a “CPU family” which is a programmer specified collection of related machines. What constitutes a family is work-in-progress however it is intended to distinguish things like sparc32 vs sparc64. Machines in a family are sufficiently similar that the simulator semantic code can handle any differences at run time. At least that's the current idea. A minimum of one CPU family must be defined. (16)

The syntax of define-cpu is:

(define-cpu
  (name cpu-name)
  (comment "description")
  (attrs attribute-list)
  (endian big|little|either)
  (insn-endian big|little|either)
  (data-endian big|little|either)
  (float-endian big|little|either)
  (word-bitsize n)
  (insn-chunk-bitsize n)
  (parallel-insns n)
  (file-transform transformation)
)

3.8.3.1 endian

The endianness of the architecture is one of three values: big, little and either.

An architecture may have multiple endiannesses, including one for each of: instructions, integers, and floats (not that that's intended to be the complete list). These are specified with insn-endian, data-endian, and float-endian respectively.

Possible values for insn-endian are: big, little, and either. If missing, the value is taken from endian.

Possible values for data-endian and float-endian are: big, big-words, little, little-words and either. If big-words then each word is little-endian. If little-words then each word is big-endian. If missing, the value is taken from endian.

??? Support for these is work-in-progress. All forms are recognized by the ‘.cpu’ file reader, but not all are supported internally.

3.8.3.2 word-bitsize

The number of bits in a word. In GCC, this is BITS_PER_WORD.

3.8.3.3 insn-chunk-bitsize

The number of bits in an instruction word chunk, for purposes of per-chunk endianness conversion. The default is zero, meaning no chunking is required.

3.8.3.4 parallel-insns

This is the same as the parallel-insns spec of define-isa. It allows a CPU family to override the value.

3.8.3.5 file-transform

Specify the file name transformation of generated code.

Each generated file has a named related to the ISA or CPU family. Sometimes generated code needs to know the name of another generated file (e.g. #include's). At present file-transform specifies the suffix.

For example, M32R/x generated files have an `x' suffix, as in ‘cpux.h’ for the ‘cpu.h’ header. This is indicated with (file-transform "x").

??? Ideally generated code wouldn't need to know anything about file names. This breaks down for #include's. It can be fixed with symlinks or other means.

3.8.4 define-mach

define-mach defines a distinct variant of a CPU. It currently has a one-to-one correspondence with BFD's "mach number". A minimum of one mach must be defined.

The syntax of define-mach is:

(define-mach
  (name mach-name)
  (comment "description")
  (attrs attribute-list)
  (cpu cpu-family-name)
  (bfd-name "bfd-name")
  (isas isa-name-list)
)

3.8.4.1 bfd-name

The name of the mach as used by BFD. If not specified the name of the mach is used.

3.8.4.2 isas

List of names of ISA's the machine supports.

3.9 Model variants

For each `machine', as defined here, there is one or more `models'. There must be at least one model for each machine. (*Note: There could be a default, but requiring one doesn't involve that much extra typing and forces the programmer to at least think about such things.)

(define-model
  (name model-name)
  (comment "description")
  (attrs attribute-list)
  (mach machine-name)
  (state (variable-name-1 variable-mode-1) ...)
  (unit name "comment" (attributes)
	issue done state inputs outputs profile)
)

3.9.1 mach

The name of the machine the model is an implementation of.

3.9.2 state

A list of variable-name/mode pairs for recording global function unit state. For example on the M32R the value is (state (h-gr UINT)) and is a bitmask of which register(s) are the targets of loads and thus subject to load stalls.

3.9.3 unit

Specifies a function unit. Any number of function units may be specified. The u-exec unit must be specified as it is the default.

The syntax is:

  (unit name "comment" (attributes)
     issue done state inputs outputs profile)

‘issue’ is the number of operations that may be in progress. It originates from GCC function unit specification. In general the value should be 1.

‘done’ is the latency of the unit. The value is the number of cycles until the result is ready.

‘state’ has the same syntax as the global model `state' and is a list of variable-name/mode pairs.

‘inputs’ is a list of inputs to the function unit. Each element is (operand-name mode default-value).

‘outputs’ is a list of outputs of the function unit. Each element is (operand-name mode default-value).

‘profile’ is an rtl-code sequence that performs function unit modeling. At present the only possible value is () meaning invoke a user supplied function named <cpu>_model_<mach>_<unit>.

The current function unit specification is a first pass in order to achieve something that moderately works for the intended purpose (cycle counting on the simulator). Something more elaborate is on the todo list but there is currently no schedule for it. The new specification must try to be application independent. Some known applications are: cycle counting in the simulator, code scheduling in a compiler, and code scheduling in a JIT simulator (where speed of analysis can be more important than getting an optimum schedule).

The inputs/outputs fields are how elements in the semantic code are mapped to function units. Each input and output has a name that corresponds with the name of the operand in the semantics. Where there is no correspondence, a mapping can be made in the unit specification of the instruction (see the subsection titled “Timing”).

Another way to achieve the correspondence is to create separate function units that contain the desired input/output names. For example on the M32R the u-exec unit is defined as:

(unit u-exec "Execution Unit" ()
   1 1 ; issue done
   () ; state
   ((sr INT -1) (sr2 INT -1)) ; inputs
   ((dr INT -1)) ; outputs
   () ; profile action (default)
)

This handles instructions that use sr, sr2 and dr as operands. A second function unit called ‘u-cmp’ is defined as:

(unit u-cmp "Compare Unit" ()
   1 1 ; issue done
   () ; state
   ((src1 INT -1) (src2 INT -1)) ; inputs
   () ; outputs
   () ; profile action (default)
)

This handles instructions that use src1 and src2 as operands. The organization of units is arbitrary. On the M32R, src1/src2 instructions are typically compare instructions so a separate function unit was created for them. Current limitations require that each hardware item behind the operands must be marked with the attribute PROFILE and the hardware item must not be scalar.

3.10 Hardware elements

The elements of hardware that make up a CPU are defined with define-hardware. Examples of hardware elements include registers, condition bits, immediate constants and memory.

Instruction fields that provide numerical values (“immediate constants”) aren't really elements of the hardware, but it simplifies things to think of them this way. Think of them as constant generators(17).

Hardware elements are defined with:

(define-hardware
  (name hardware-name)
  (comment "description")
  (attrs attribute-list)
  (semantic-name hardware-semantic-name)
  (type type-name type-arg1 type-arg2 ...)
  (indices index-type index-arg1 index-arg2 ...)
  (values values-type values-arg1 values-arg2 ...)
  (handlers handler1 handler2 ...)
  (get (args) expression)
  (set (args) expression)
  (layout layout-list)
)

The only required elements are ‘name’ and ‘type’. Convention requires ‘hardware-name’ begin with ‘h-’.

3.10.1 attrs

List of attributes. There are several predefined hardware attributes:

• - MACH

  A bitset attribute used to specify which machines have this hardware element. Do not specify the MACH attribute if the value is "all machs".

  Usage: (MACH mach1,mach2,...) There must be no spaces in “mach1,mach2,...”.

• - CACHE-ADDR

  A hint to the simulator semantic code generator to tell it it can record the address of a selected register in an array of registers. This speeds up simulation by moving the array computation to extraction time. This attribute is only useful to register arrays and cannot be specified with VIRTUAL (??? revisit).

• - PROFILE

  This attribute must be present for hardware elements to which references are profiled. Beware, this is work-in-progress. If you use this attribute it is likely you have to hack CGEN. (Please submit patches.)

• - VIRTUAL

  The hardware element doesn't require any storage. This is used when you want a value that is derived from some other value. If VIRTUAL is specified, get and set specs must be provided.

3.10.2 type

This is the type of hardware. Current values are: ‘pc’, ‘register’, ‘memory’, and ‘immediate’.

For ‘pc’, see See Program counter.

For registers the syntax is one of:

(register mode [(number)])
(register (mode bits) [(number)])

where ‘(number)’ is the number of registers and is optional. If omitted, the default is ‘(1)’. The second form is useful for describing registers with an odd (as in unusual) number of bits. mode for the second form must be one of ‘INT’ or ‘UINT’. Since these two modes don't have an implicit size, they cannot be used for the first form.

For memory the syntax is:

(memory mode (size))

where ‘(size)’ is the size of the memory in ‘mode’ units. In general ‘mode’ should be QI.

For immediates the syntax is one of

(immediate mode)
(immediate (mode bits))

The second form is for values for which a mode of that size doesn't exist. ‘mode’ for the second form must be one of INT or UINT. Since these two modes don't have an implicit size, they cannot be used for the first form.

??? There's no real reason why a mode like SI can't be used for odd-sized immediate values. The ‘bits’ field indicates the size and the ‘mode’ field indicates the mode in which the value will be used, as well as its signedness. This would allow removing INT/UINT for this purpose. On the other hand, a non-width specific mode allows applications to choose one (a simulator might prefer to store immediates in an `int' rather than, say, char if the specified mode was QI).

3.10.3 indices

Specify names for individual elements with the indices spec. It is only valid for registers with more than one element.

The syntax is:

(indices index-type arg1 arg2 ...)

where ‘index-type’ specifies the kind of index and ‘arg1 arg2 ...’ are arguments to ‘index-type’.

There are two supported values for ‘index-type’: keyword and extern-keyword. The difference is that indices defined with keyword are kept internal to the hardware element's definition and are not usable elsewhere, whereas extern-keyword specifies a set of indices defined elsewhere with define-keyword.

3.10.3.1 keyword

(indices keyword name-prefix ((name1 value1) (name2 value2) ...))

‘name-prefix’ is the assembler prefix common to each of the index names, and is added to name in the generated lookup table. For example, SPARC registers usually begin with ‘"%"’.

Each ‘(name value)’ pair maps a name with an index number. An index can be specified multiple times, for example, when a register has multiple names.

There may be gaps in the index list, e.g. for invalid/reserved registers.

No enum is defined for keywords defined this way. If you want an enum use ‘define-keyword’ and ‘extern-keyword’.

Example from Thumb:

(define-hardware 
  (name h-gr-t)
  (comment "Thumb's general purpose registers")
  (attrs (ISA thumb) VIRTUAL) ; ??? CACHE-ADDR should be doable
  (type register WI (8))
  (indices keyword ""
	   ((r0 0) (r1 1) (r2 2) (r3 3) (r4 4) (r5 5) (r6 6) (r7 7)))
  (get (regno) (reg h-gr regno))
  (set (regno newval) (set (reg h-gr regno) newval))
)

3.10.3.2 extern-keyword

(indices extern-keyword keyword-name)

Often one wants to make the keywords available for general use, i.e. to arbitrary tools. See section Keywords. When the collection of indices is defined with ‘define-keyword’ refer to it in the ‘indices’ field with ‘extern-keyword’.

Example from M32R:

(define-keyword
  (name gr-names)
  (enum-prefix H-GR-)
  (values (fp 13) (lr 14) (sp 15)
	  (r0 0) (r1 1) (r2 2) (r3 3) (r4 4) (r5 5) (r6 6) (r7 7)
	  (r8 8) (r9 9) (r10 10) (r11 11) (r12 12) (r13 13) (r14 14) (r15 15))
)

(define-hardware
  (name h-gr)
  (comment "general registers")
  (attrs PROFILE CACHE-ADDR)
  (type register WI (16))
  (indices extern-keyword gr-names)
)

3.10.4 values

Specify a list of valid values with the values spec.

The syntax is identical to the syntax for indices. It is only valid for immediates.

Example from sparc64:

(define-hardware
  (name h-p)
  (comment "prediction bit")
  (attrs (MACH64))
  (type immediate (UINT 1))
  (values keyword "" (("" 0) (",pf" 0) (",pt" 1)))
)

3.10.5 handlers

The handlers spec is an escape hatch for indicating when a programmer supplied routine must be called to perform a function.

The syntax is:

‘(handlers (handler-name1 "function_name1")
                (handler-name2 "function_name2")
                ...)’

‘handler-name’ must be one of parse or print. How ‘function_name’ is used is application specific, but in general it is the name of a function to call. The only application that uses this at present is Opcodes. See the Opcodes documentation for a description of each function's expected prototype.

3.10.6 get

Specify special processing to be performed when a value is read with the get spec.

The syntax for scalar registers is:

‘(get () (expression))’

The syntax for vector registers is:

‘(get (index) (expression))’

expression is an RTL expression that computes the value to return. The mode of the result must be the mode of the register.

index is the name of the index as it appears in expression.

At present, sequence, parallel, do-count and case expressions are not allowed here.

3.10.7 set

Specify special processing to be performed when a value is written with the set spec.

The syntax for scalar registers is:

‘(set (newval) (expression))’

The syntax for vector registers is:

‘(set (index newval) (expression))’

expression is an RTL expression that stores newval in the register. This may involve storing values in other registers as well. expression must be one of set, if, sequence, or case.

index is the name of the index as it appears in expression.

3.10.8 layout

For specific hardware elements, specifying a layout is an alternative to providing getter/setter specs.

At present this applies to only ‘register’ hardware elements, but not the ‘pc’.

Some registers are a collection of bits with different meanings. It is often useful to define each field of such a register as its own register. The ‘layout’ spec can then be used to build up the outer register from the individual register fields.

The fields are written from least to most significant. Each field is either the name of another hardware register, or a list of (value length) to specify hardwired bits.

A typical example is a “flags” register. Here is an example for a fictitious flags register. It is eight bits wide, with the lower four bits having defined values, and the upper four bits hardwired to zero.

(dsh h-cf "carry flag"    () (register BI))
(dsh h-sf "sign flag"     () (register BI))
(dsh h-of "overflow flag" () (register BI))
(dsh h-zf "zero flag"     () (register BI))
(define-hardware
  (name flags)
  (type register QI)
  (layout (h-cf h-sf h-of h-zf (0 4)))
)

3.10.9 Predefined hardware elements

Several hardware types are predefined:

h-uint

unsigned integer

h-sint

signed integer

h-memory

main memory, where “main” is loosely defined

h-addr

data address (data only)

h-iaddr

instruction address (instructions only)

3.10.10 Program counter

The program counter must be defined and is not a builtin. If get/set specs are not required, define it as:

(dnh h-pc "program counter" (PC) (pc) () () ())

If get/set specs are required, define it as:

(define-hardware
  (name h-pc)
  (comment "<ARCH> program counter")
  (attrs PC)
  (type pc)
  (get () <insert get code here>)
  (set (newval) <insert set code here>)
)

If the architecture has multiple instruction sets, all must be specified. If they're not, the default is the first one which is often not what you want. Here's an example from ‘arm.cpu’:

(define-hardware
  (name h-pc)
  (comment "ARM program counter (h-gr reg 15)")
  (attrs PC (ISA arm,thumb))
  (type pc)
  (set (newval)
       (if (reg h-tbit)
	   (set (raw-reg SI h-pc) (and newval -2))
	   (set (raw-reg SI h-pc) (and newval -4))))
)

3.10.11 Simplification macros

To simplify ‘.cpu’ files several pmacros are provided.

The define-normal-hardware pmacro (with alias dnh) takes a fixed set of positional arguments for the typical hardware element. The syntax is:

(dnh name comment attributes type indices values handlers)

Example:

(dnh h-gr "general registers"
     () ; attributes
     (register WI (16))
     (keyword "" ((fp 13) (sp 15) (lr 14)
                  (r0 0) (r1 1) (r2 2) (r3 3)
                  (r4 4) (r5 5) (r6 6) (r7 7)
                  (r8 8) (r9 9) (r10 10) (r11 11)
                  (r12 12) (r13 13) (r14 14) (r15 15)))
     () ()
)

This defines an array of 16 registers of mode WI ("word int"). The names of the registers are r0...r15, and registers 13, 14 and 15 also have the names fp, lr and sp respectively.

Scalar registers with no special requirements occur frequently. Macro define-simple-hardware (with alias dsh) is identical to dnh except does not include the indices, values, or handlers specs.

(dsh h-ibit "interrupt enable bit" () (register BI))

3.11 Instruction fields

Instruction fields (ifields) define the raw bitfields of each instruction. Minimal semantic meaning is attributed to them. Support is provided for mapping to and from the raw bit pattern and the usable contents, and other simple manipulations. (18)

Instruction fields must be uniquely named within an instruction set, but different instruction sets (ISAs) may have ifields with the same name.

The syntax for defining instruction fields is:

(define-ifield
  (name field-name)
  (comment "description")
  (attrs attribute-list)
  (word-offset word-offset-in-bits)
  (word-length word-length-in-bits)
  (start starting-bit-number)
  (length number-of-bits)
  (follows ifield-name)
  (mode mode-name)
  (encode (value pc) (rtx to describe encoding))
  (decode (value pc) (rtx to describe decoding))
)

The required elements are: ‘name’, ‘start’, ‘length’. (19)

Convention requires ‘field-name’ begin with ‘f-’.

3.11.1 attrs

There are several predefined instruction field attributes:

PCREL-ADDR

The field contains a PC relative address. Various CPUs have various offsets from the PC from which the address is calculated. This is specified in the encode and decode sections.

ABS-ADDR

The field contains an absolute address.

SIGN-OPT

The field has an optional sign. It is sign-extended during extraction. Allowable values are -2^(n-1) to (2^n)-1.

RESERVED

The field is marked as “reserved” by the architecture. This is an informational attribute. Tools may use it to validate programs, either statically or dynamically.

VIRTUAL

The field does not directly contribute to the instruction's value. This is used to simplify semantic or assembler descriptions where a field's value is based on other values. Multi-ifields are always virtual.

3.11.2 word-offset

The offset in bits from the start of the instruction to the word containing the field. This must be a multiple of eight.

Either both of ‘word-offset’ and ‘word-length’ must be specified or neither of them must be specified. The presence of ‘word-offset’ means the long form of specifying the field's position is being used. If absent then the short form is being used and the value for ‘word-offset’ is encoded in ‘start’.

3.11.3 word-length

The length in bits of the word containing the field. This must be a multiple of eight.

3.11.4 start

The bit number of the field's most significant bit in the instruction. Bit numbering is determined by the insn-lsb0? field of define-arch.

If using the long form of specifying the field's position (i.e., ‘word-offset’ is specified) then this value is the value within the containing word. If using the short form then this value includes the word offset. See the Porting document for more info (see section Writing define-ifield).

3.11.5 length

The number of bits in the field. The field must be contiguous. For non-contiguous instruction fields use “multi-ifields”.

3.11.6 follows

Optional. Experimental. This should not be used for the specification of RISC-like architectures. It is an experiment in supporting CISC-like architectures. The argument is the name of the ifield or operand that immediately precedes this one. In general the argument is an "anyof" operand. The follows spec allows subsequent ifields to “float”.

3.11.7 mode

The mode the value is to be interpreted in. Usually this is INT or UINT.

The ‘length’ field specifies the number of bits in the field, and the ‘mode’ field indicates the mode in which the value will be used, as well as its signedness. This would allow removing INT/UINT for this purpose. On the other hand, a non-width specific mode allows applications to choose one (a simulator might prefer to store immediates in an `int' rather than, say, char if the specified mode was QI).

3.11.8 encode

An expression to apply to convert from usable values to raw field values. The syntax is (encode (value pc) expression) or more generally (encode ((<mode1> value) (IAI pc)) <expression>), where <mode1> is the mode of the “incoming” value, and <expression> is an rtx to convert value to something that can be stored in the field.

Example:

(encode ((SF value) (IAI pc))
	(cond WI
	      ((eq value (const SF 1.0)) (const 0))
	      ((eq value (const SF 0.5)) (const 1))
	      ((eq value (const SF -1.0)) (const 2))
	      ((eq value (const SF 2.0)) (const 3))
	      (else (error "invalid floating point value for field foo"))))

In this example four floating point immediate values are represented in a field of two bits. The above might be expanded to a series of `if' statements or the generator could determine a `switch' statement is more appropriate.

3.11.9 decode

An expression to apply to convert from raw field values to usable values. The syntax is (decode (value pc) expression) or more generally (decode ((<mode1> value) (IAI pc)) <expression>), where <mode1> is the mode of the “incoming” value, and <expression> is an rtx to convert value to something usable.

Example:

(decode ((WI value) (IAI pc))
	(cond SF
	      ((eq value 0) (const SF 1.0))
	      ((eq value 1) (const SF 0.5))
	      ((eq value 2) (const SF -1.0))
	      ((eq value 3) (const SF 2.0))))

There's no need to provide an error case as presumably value would never have an invalid value, though certainly one could provide an error case if one wanted to.

3.11.10 Non-contiguous fields

Non-contiguous fields (e.g. sparc64's 16 bit displacement field) are built on top of support for contiguous fields. The syntax for defining such fields is:

(define-multi-ifield
  (name field-name)
  (comment "description")
  (attrs attribute-list)
  (mode mode-name)
  (subfields field1-name field2-name ...)
  (insert (code to set each subfield))
  (extract (code to set field from subfields))
  (encode (value pc) (rtx to describe encoding))
  (decode (value pc) (rtx to describe decoding))
)

The required elements are: ‘name’, ‘subfields’.

Example:

(define-multi-ifield
  (name f-i20)
  (comment "20 bit unsigned")
  (attrs)
  (mode UINT)
  (subfields f-i20-4 f-i20-16)
  (insert (sequence ()
                    (set (ifield f-i20-4)  (srl (ifield f-i20) (const 16)))
                    (set (ifield f-i20-16) (and (ifield f-i20) (const #xffff)))
                    ))
  (extract (sequence ()
                     (set (ifield f-i20) (or (sll (ifield f-i20-4) (const 16))
                                             (ifield f-i20-16)))
                     ))
)

3.11.10.1 subfields

The names of the already defined fields that make up the multi-ifield.

3.11.10.2 insert

Code to set the subfields from the multi-ifield. All fields are referred to with (ifield <name>).

3.11.10.3 extract

Code to set the multi-ifield from the subfields. All fields are referred to with (ifield <name>).

3.11.11 Simplification macros

To simplify ‘.cpu’ files several pmacros are provided.

The define-normal-ifield pmacro (with alias dnf) takes a fixed set of positional arguments for the typical instruction field. The syntax is:

(dnf name comment attributes start length)

Example:

(dnf f-r1 "register r1" () 4 4)

This defines a field called ‘f-r1’ that is an unsigned field of 4 bits beginning at bit 4. All fields defined with dnf are unsigned.

The df pmacro adds mode, encode, and decode elements.

The syntax of df is:

(df name comment attributes start length mode encode decode)

Example:

(df f-disp8
    "disp8, slot unknown" (PCREL-ADDR)
    8 8 INT
    ((value pc) (sra WI (sub WI value (and WI pc (const -4))) (const 2)))
    ((value pc) (add WI (sll WI value (const 2)) (and WI pc (const -4)))))

This defines a field called ‘f-disp8’ that is a signed PC-relative address beginning at bit 8 of size 8 bits that is left shifted by 2.

The macro define-normal-multi-ifield (with alias dnmf) takes a fixed set of positional arguments for the typical multi-ifield. The syntax is:

(dnmf name comment attributes mode subfields insert extract)

The macro dsmf takes a fixed set of positional arguments for simple multi-ifields. The syntax is:

(dsmf name comment attributes mode subfields)

3.12 Enumerated constants

Enumerated constants (enums) are important enough in instruction set descriptions that they are given special treatment. Enums are defined with:

(define-enum
  (name enum-name)
  (comment "description")
  (attrs attribute-list)
  (prefix prefix)
  (values val1 val2 ...)
)

Enums in opcode fields are further enhanced by specifying the opcode field they are used in. This allows the enum's name to be specified in an instruction's format entry.

Instruction enums are defined with define-insn-enum:

(define-insn-enum
  (name enum-name)
  (comment "description")
  (attrs attribute-list)
  (ifield ifield-name)
  (prefix prefix)
  (values val1 val2 ...)
)

define-insn-enum is currently not provided, use define-normal-insn-enum instead. See define-normal-insn-enum.

3.12.1 prefix

Convention requires each enum value to be prefixed with the same text. Rather than specifying the prefix in each entry, it is specified once, here. Convention requires ‘prefix’ not contain any lowercase characters. You generally want to end ‘prefix’ with ‘-’ or ‘_’ as the complete name of each enum value is ‘prefix’ + ‘value-name’. The convention is to use ‘-’, though this convention is not adhered to as well as the other conventions.

The default value is ‘""’.

3.12.2 ifield

The name of the instruction field that the enum is intended for. This must be a simple ifield, not a multi-ifield.

3.12.3 values

A list of possible values. Each element has one of the following forms:

• name
• (name)
• (name value)
• (name - (attribute-list))
• (name value (attribute-list))

The syntax for numbers is Scheme's, so hex numbers are #xnnnn. A value of - means use the next value (previous value plus 1).

Enum values currently always have mode ‘INT’.

Example:

(values "a" ("b") ("c" #x12)
	("d" - (sanitize foo)) ("e" #x1234 (sanitize bar)))

3.12.4 Simplification macros

To simplify ‘.cpu’ files several pmacros are provided.

The define-normal-enum pmacro takes a fixed set of positional arguments for the typical enum. The syntax is:

(define-normal-enum name comment attrs prefix vals)

The define-normal-insn-enum pmacro takes a fixed set of positional arguments for the typical instruction enum. The syntax is:

(define-normal-insn-enum name comment attrs prefix ifield vals)

Example:

(dnf f-op1 "op1" () 0 4)
(define-normal-insn-enum insn-op1 "insn format enums" () OP1_ f-op1
  (.map .str (.iota 16))
)

This defines an instruction enum for field ‘f-op1’ with values OP1_0, OP1_1, ..., OP1_15. These values can be directly used in instruction format specs. This applies to “instruction enums” only. One can use normal enums in instruction format specs but one needs to explicitly specify the ifield, e.g. (f-op1 OP1_0).

3.13 Keywords

Keywords are like enums, See section Enumerated constants, but they also cause a table of names of each value to be generated. This is useful for things like registers where you want arbitrary tools to have access to the table of names.

The syntax for defining keywords changed from RTL version 0.7 to RTL version 0.8. See section RTL Versions.

RTL version 0.7 syntax:

(define-keyword
  (name keyword-name)
  (comment "description")
  (attrs attribute-list)
  (mode mode-name)
  (print-name "prefix-for-enum-values-without-trailing-dash")
  (prefix "prefix-for-names-in-string-table")
  (values value-list)
)

RTL version 0.8 syntax:

(define-keyword
  (name keyword-name)
  (comment "description")
  (attrs attribute-list)
  (mode mode-name)
  (enum-prefix "prefix-for-enum-values")
  (name-prefix "prefix-for-names-in-string-table")
  (values value-list)
)

Note that ‘print-name’ has been replaced with ‘enum-prefix’ and ‘prefix’ has been replaced with ‘name-prefix’.

Furthermore, there is also a difference between the behavior of ‘print-name’ and ‘enum-prefix’. When computing complete enum names with ‘print-name’, CGEN adds a ‘-’ between the prefix and the enum name. CGEN does not insert a ‘-’ with ‘enum-prefix’.

3.13.1 mode

This is the mode to reference and record the keyword's value in. The default is ‘INT’. It is normally not necessary to use something else.

3.13.2 print-name

NOTE: This is for RTL version 0.7 only.

This value plus a trailing ‘-’ is passed as the ‘prefix’ parameter when defining the corresponding enum. See section Enumerated constants.

Convention requires ‘print-name’ not contain any lowercase characters.

The default value is the keyword's name in uppercase.

3.13.3 prefix

NOTE: This is for RTL version 0.7 only.

‘prefix’ is the assembler prefix common to each of the index names, and is added to name in the generated lookup table. For example, SPARC registers usually begin with ‘"%"’. It is not added to the corresponding enum value names.

The default value is ‘""’.

3.13.4 enum-prefix

NOTE: This is for RTL version 0.8 and higher. You must specify the RTL version at the top of the description file.

This value is passed as the ‘prefix’ parameter when defining the corresponding enum. See section Enumerated constants.

NOTE: Unlike ‘print-name’ in RTL version ‘0.7’, ‘-’ is not appended when defining the corresponding enum.

Convention requires ‘enum-prefix’ not contain any lowercase characters.

The default value is the keyword's name in uppercase + ‘-’.

3.13.5 name-prefix

NOTE: This is for RTL version 0.8 and higher. You must specify the RTL version at the top of the description file.

‘name-prefix’ is the assembler prefix common to each of the index names, and is added to name in the generated lookup table. For example, SPARC registers usually begin with ‘"%"’. It is not added to the corresponding enum value names.

The default value is ‘""’.

3.13.6 values

The ‘values’ field has the same syntax as the ‘values’ field of ‘define-enum’. See Enum Values.

Example from M32R:

(define-keyword
  (name gr-names)
  (enum-prefix H-GR-)
  (values (fp 13) (lr 14) (sp 15)
	  (r0 0) (r1 1) (r2 2) (r3 3) (r4 4) (r5 5) (r6 6) (r7 7)
	  (r8 8) (r9 9) (r10 10) (r11 11) (r12 12) (r13 13) (r14 14) (r15 15))
)

Referencing enum values from this keyword in the .cpu file would use ‘H-GR-’ + ‘register-name’. E.g., H-GR-r12.

3.14 Instruction operands

Instruction operands provide:

• a layer between the assembler and the raw hardware description
• the main means of making an instruction's fields useful to the semantic code

Instruction operands must be uniquely named within an instruction set, but different instruction sets (ISAs) may have operands with the same name.

The syntax for defining an operand is:

(define-operand
  (name operand-name)
  (comment "description")
  (attrs attribute-list)
  (type hardware-element)
  (mode mode-name)
  (index instruction-field)
  (handlers handler-spec)
  (getter getter-spec)
  (setter setter-spec)
)

The required elements are: name, type, mode, and if type is not a scaler index.

3.14.1 name

This is the name of the operand as a Scheme symbol. The name choice is fairly important as it is used in instruction syntax entries, instruction format entries, and semantic expressions. It can't collide with symbols used in semantic expressions (e.g. and, set, etc).

The convention is that operands have no prefix (whereas ifields begin with ‘f-’ and hardware elements begin with ‘h-’). A prefix like ‘o-’ would avoid collisions with other semantic elements, but operands are used often enough that any prefix is a hassle.

Note that if you do decide to prefix operand names, e.g. use a style like ‘o-foo’, then you will need to remember to use the ‘${o-foo}’ form in the assembler syntax and not the ‘$o-foo’ form because the latter only takes alphanumeric characters. See syntax.

3.14.2 attrs

A list of attributes. In addition to attributes defined for the operand, an operand inherits the attributes of its instruction field. There are several predefined operand attributes:

NEGATIVE

The operand contains negative values (not used yet so definition is still nebulous.

RELAX

This operand contains the changeable field (usually a branch address) of a relaxable/relaxed instruction.

SEM-ONLY

Use the SEM-ONLY attribute for cases where the operand will only be used in semantic specification, and not assembly code specification. A typical example is condition codes.

To refer to a hardware element in semantic code one must either use an operand or one of reg/mem/const. Operands generally exist to map instruction fields to the selected hardware element and are easier to use in semantic code than referring to the hardware element directly (e.g. sr is easier to type and read than (reg h-gr <index>)). Example:

  (dnop condbit "condition bit" (SEM-ONLY) h-cond f-nil)

f-nil is the value to use when there is no instruction field

3.14.3 type

The hardware element this operand applies to. This must be the name of a hardware element.

3.14.4 mode

The mode the value is to be interpreted in.

3.14.5 index

The index of the hardware element. This is used to mate the hardware element with the instruction field that selects it, and must be the name of an ifield entry. (*Note: The index may be other things besides ifields in the future.) It must not be a multi-ifield, currently.

3.14.6 handlers

Sometimes it's necessary to escape to C to parse assembler, or print a value. This field is an escape hatch to implement this. The syntax is:

(handlers handler-spec)

where handler-spec is one or more of:

(parse "function_suffix") – a call to function parse_<function_suffix> is generated.

(print "function_suffix") – a call to function print_<function_suffix> is generated.

These functions are intended to be provided in a separate ‘.opc’ file. The prototype of a parse function depends on the hardware type. See ‘cpu/*.opc’ for examples.

For integer it is:

static const char *
parse_foo (CGEN_CPU_DESC cd,
	   const char **strp,
	   int opindex,
	   unsigned long *valuep);

cd is the result of <arch>_cgen_cpu_open. strp is a pointer to a pointer to the assembler and is updated by the function. opindex is ???. valuep is a pointer to where to record the parsed value. If a relocation is needed, it is queued with a call to ???. Queued relocations are processed after the instruction has been parsed.

The result is an error message or NULL if successful.

The prototype of a print function depends on the hardware type. See ‘cpu/*.opc’ for examples. For integers it is:

void print_foo (CGEN_CPU_DESC cd,
                PTR dis_info,
                long value,
                unsigned int attrs,
                bfd_vma pc,
                int length);

‘cd’ is the result of <arch>_cgen_cpu_open. ‘ptr’ is the `info' argument to print_insn_<arch>. ‘value’ is the value to be printed. ‘attrs’ is the set of boolean attributes. ‘pc’ is the PC value of the instruction. ‘length’ is the length of the instruction.

Actual printing is done by calling ((disassemble_info *) dis_info)->fprintf_func.

3.14.7 Simplification macros

To simplify ‘.cpu’ files several pmacros are provided.

The define-normal-operand) pmacro (with alias dno) takes a fixed set of positional arguments for the typical operand.

There is also the dnop pmacro, it is an alias of dno.

The syntax of dno is:

(dno name comment attrs type index)

Example:

(dno sr "source register" () h-gr f-r2)

This defines an operand name ‘sr’ that is an ‘h-gr’ register indexed by the ‘f-r2’ ifield.

3.15 Derived operands

Derived operands are an experiment in supporting the addressing modes of CISC-like architectures. Addressing modes are difficult to support as they essentially increase the number of instructions in the architecture by an order of magnitude. Defining all the variants requires something in addition to the RISC-like architecture support. The theory is that since CISC-like instructions are basically "normal" instructions with complex operands the place to add the necessary support is in the operands.

Two kinds of operands exist to support CISC-like cpus, and they work together. “derived-operands” describe one variant of a complex argument, and “anyof” operands group them together.

The syntax for defining derived operands is:

(define-derived-operand
  (name operand-name)
  (comment "description")
  (attrs attribute-list)
  (mode mode-name)
  (args arg1-operand-name arg2-operand-name ...)
  (syntax "syntax")
  (base-ifield ifield-name)
  (encoding (+ arg1-operand-name arg2-operand-name ...))
  (ifield-assertion expression)
  (getter expression)
  (setter expression)
)

The syntax for defining anyof operands is:

(define-anyof-operand
  (name operand-name)
  (comment "description")
  (attrs attribute-list)
  (mode mode-name)
  (base-ifield ifield-name)
  (choices derived-operand1-name derived-operand2-name ...)
)

3.15.1 mode

The name of the mode of the operand.

3.15.2 args

List of names of operands the derived operand uses. The operands must already be defined. The argument operands can be any kind of operand: normal, derived, anyof.

3.15.3 syntax

Assembler syntax of the operand.

??? This part needs more work. Addressing mode specification in assembler needn't be localized to the vicinity of the operand.

3.15.4 base-ifield

The name of the instruction field common to all related derived operands. Here related means "used by the same `anyof' operand".

3.15.5 encoding

The machine encoding of the operand.

3.15.6 ifield-assertion

An assertion of what values any instruction fields will or will not have in the containing instruction.

The syntax of the assertion is a restricted subset of RTL. It may only contain ‘andif’, ‘eq’, ‘ne’, and may only use scalar instruction fields (20) and integers as operands. Furthermore, ifields must be specified in the first operand of ‘eq’, ‘ne’.

As a degenerate case, a single non-zero integer, is also supported, meaning the assertion passes.

In addition, the assertion may also use ‘member’.

Syntax: (member ifield-name (number-list value1 [value2 ...])) (21)

The result of ‘member’ is one if the value of the ifield is a member of the list (value1 [value2 ...]). Otherwise the result is zero.

If the result of the assertion is non-zero, the assertion passes. Otherwise it fails, and the instruction is not selected for that particular bit pattern.

3.15.7 getter

RTL expression to get the value of the operand. All operands refered to must be specified in args.

3.15.8 setter

RTL expression to set the value of the operand. All operands refered to must be specified in args. Use newval to refer to the value to be set.

3.15.9 choices

For anyof operands, the names of the derived operands. The operand may be "any of" the specified choices.

3.16 Instructions

Each instruction in the instruction set has an entry in the description file. (22)

Instructions must be uniquely named within an instruction set, but different instruction sets (ISAs) may have instructions with the same name.

The syntax for defining an instruction is:

(define-insn
  (name insn-name)
  (comment "description")
  (attrs attribute-list)
  (syntax "assembler syntax")
  (format (+ field-list))
  (ifield-assertion expression)
  (semantics expression)
  (timing timing-data)
)

The required elements are: name, ???.

Instructions specific to a particular cpu variant are denoted as such with the MACH attribute.

Possible additions for the future:

• a field to describe a final constraint for determining a match
• choosing the output from a set of choices

3.16.1 attrs

A list of attributes, for which there are several predefined instruction attributes:

MACH

A bitset attribute used to specify which machines have this hardware element. Do not specify the MACH attribute if the value is for all machines.

Usage: (MACH mach1,mach2,...)

There must be no spaces in “mach1,mach2,...”.

UNCOND-CTI

The instruction is an unconditional “control transfer instruction”.

(*Note: This attribute is derived from the semantic code. However if the computed value is wrong (dunno if it ever will be) the value can be overridden by explicitly mentioning it.)

COND-CTI

The instruction is an conditional "control transfer instruction".

(*Note: This attribute is derived from the semantic code. However if the computed value is wrong (dunno if it ever will be) the value can be overridden by explicitly mentioning it.)

SKIP-CTI

The instruction can cause one or more insns to be skipped. This is derived from the semantic code.

DELAY-SLOT

The instruction has one or more delay slots. This is derived from the semantic code.

RELAXABLE

The instruction has one or more identical variants. The assembler tries this one first and then the relaxation phases switches to larger ones as necessary.

RELAXED

The instruction is a non-minimal variant of a relaxable instruction. It is avoided by the assembler in the first pass.

ALIAS

Internal attribute set for macro-instructions that are an alias for one real insn.

NO-DIS

For macro-instructions, don't use during disassembly.

3.16.2 syntax

This is a character string consisting of raw characters and operands. Fields are denoted by $operand or ${operand}. The ${operand} form is required if the operand name contains non-alphanumeric characters. If a ‘$’ is required in the syntax, it is specified with ‘\$’. If a ‘\’ is required in the syntax, it is specified with ‘\\’.

At most one white-space character may be present and it must be a blank separating the instruction mnemonic from the operands. This doesn't restrict the user's assembler, this is just a description file restriction to separate the mnemonic from the operands(23). Note that the assembler will accept multiple spaces in the assembler code after the mnemonic and between operands as expected.

Operands can refer to registers, constants, and whatever else is necessary.

Instruction mnemonics can take operands. For example, on the SPARC a branch instruction can take ,a as an argument to indicate the instruction is being annulled (e.g. bge$a $disp22).

3.16.3 format

This is a complete list of fields that specify the instruction. At present it must be prefaced with + to allow for future additions. Reserved bits must also be specified, gaps are not allowed. The ordering of the fields is not important.

Format elements can be any of:

• an instruction field name with an integer, e.g. (f-op1 4)
• an instruction field name with an enum, e.g. (f-op1 OP1_4)
• an instruction field enum, e.g. OP1_4
• an operand name, e.g. dr

3.16.4 ifield-assertion

This is an expression with a boolean result that is run as the final part of instruction decoding to verify a match.

The syntax of the assertion is a restricted subset of RTL. See ifield-assertion-rtl.

3.16.5 semantics

This field provides a mathematical description of what the instruction does. Its syntax is GCC RTL-like on purpose since GCC's RTL is well known by the intended audience. However, it is not intended that it be precisely GCC RTL.

Obviously there are some instructions that are difficult if not impossible to provide a description for (e.g. I/O instructions). Rather than create a new semantic function for each quirky operation, escape hatches to C are provided to handle all such cases. The c-code, c-call and c-raw-call semantic functions provide an escape-hatch to invoke C code to perform the operation. See section Expressions.

3.16.6 timing

A list of entries for each function unit the instruction uses on each machine that supports the instruction. The default function unit is the u-exec unit.

The syntax is:

(model-name (unit name (direction unit-var-name1 insn-operand-name1)
                       (direction unit-var-name2 insn-operand-name2)
                       ...
                       (cycles cycle-count))

direction/unit-var-name/insn-operand-name mappings are optional. They map unit inputs/outputs to semantic elements. The direction specifier can be in or out mapping the name of a unit input or output, respectively, to an insn operand.

cycles overrides the done value (latency) of the function unit and is optional.

3.16.7 Simplification macros

To simplify ‘.cpu’ files several pmacros are provided.

The define-normal-insn pmacro (with alias dni) takes a fixed set of positional arguments for the typical instruction.

The syntax of dni is:

(dni name comment attrs syntax format semantics timing)

Example:

(dni addi "add 8 bit signed immediate"
     ()
     "addi $dr,$simm8"
     (+ OP1_4 dr simm8)
     (set dr (add dr simm8))
     ()
)

3.17 Macro-instructions

Macro-instructions are for the assembler side of things and are not used by the simulator.

Macro-instructions must be uniquely named within an instruction set, but different instruction sets (ISAs) may have macro-instructions with the same name.

The syntax for defining a macro-instruction is:

(define-macro-insn
  (name macro-insn-name)
  (comment "description")
  (attrs attribute-list)
  (syntax "assembler syntax")
  (expansions expansion-spec)
)

3.17.1 syntax

Syntax of the macro-instruction. This has the same value as the syntax field in define-insn.

3.17.2 expansions

An expression to emit code for the instruction. This is intended to be general in nature, allowing tests to be done at runtime that choose the form of the expansion. Currently the only supported form is:

(emit insn arg1 arg2 ...)

where insn is the name of an instruction defined with define-insn and argn is the set of operands to insn's syntax. Each argument is mapped in order to one operand in insn's syntax and may be any of:

• operand specified in syntax
• (operand value)

3.17.3 Simplification macros

To simplify ‘.cpu’ files several pmacros are provided.

The define-normal-macro-insn) pmacro (with alias dnmi) takes a fixed set of positional arguments for the typical macro-instruction.

The syntax of dnmi is:

(dnmi name comment attrs syntax expansion)

Example:

(dni st-minus "st-" ()
     "st $src1,$src2"
     (+ OP1_2 OP2_7 src1 src2)
     (sequence ((WI new-src2))
	       (set new-src2 (sub src2 (const 4)))
	       (set (mem WI new-src2) src1)
	       (set src2 new-src2))
     ()
)

(dnmi push "push" ()
  "push $src1"
  (emit st-minus src1 (src2 15)) ; "st %0,sp"
)

In this example, the st-minus instruction is a general store-and-decrement instruction and push is a specialized version of it that uses the stack pointer.

3.18 Modes

Modes provide a simple and succinct way of specifying data types.

(*Note: Should more complex types will be needed (e.g. structs? unions?), these can be handled by extending the definition of a mode to encompass them.)

Modes are similar to their usage in GCC, but there are some differences:

• modes for boolean values (i.e. bits) are also supported as they are useful
• integer modes exist in signed and unsigned versions
• constants have modes

Currently supported modes are:

VOID

VOIDmode in GCC.

DFLT

Indicate the default mode is wanted, the value of which depends on context. This is a pseudo-mode and never appears in generated code.

BI

Boolean zero/one

QI,HI,SI,DI

Same as GCC.

QI is an 8 bit quantity ("quarter int"). HI is a 16 bit quantity ("half int"). SI is a 32 bit quantity ("single int"). DI is a 64 bit quantity ("double int").

In cases where signedness matters, these modes are signed.

UQI,UHI,USI,UDI

Unsigned versions of QI,HI,SI,DI.

These modes do not appear in semantic RTL. Instead, the RTL function specifies the signedness of its operands where necessary. To a cpu, a 32 bit register is a 32 bit register. Ditto for when the 32 bit quantity lives in memory. It's only in how it is subsequently used or interpreted that signedness might come into play. When signedness comes into play on the chip, it's explicitly specified in the operation, _not_ in the data. Ergo from this perspective Umodes don't belong in .cpu files. This is the perspective to use when writing .cpu files.

WI,UWI

word int, unsigned word int (word_mode in gcc). These are aliases for the real mode, typically either SI or DI.

SF,DF,XF,TF

Same as GCC.

SF is a 32 bit IEEE float ("single float"). DF is a 64 bit IEEE float ("double float"). XF is either an 80 or 96 bit IEEE float ("extended float"). (*Note: XF values on m68k and i386 are different so may wish to give them different names). TF is a 128 bit IEEE float.

AI

Address integer

IAI

Instruction address integer

INT,UINT

Varying width int/unsigned-int. The width is specified by context, usually in an instruction field definition.

3.19 Expressions

The syntax of CGEN's RTL expressions (or rtx) basically follows that of GCC's RTL.

The handling of modes is different to simplify the implementation. Implementation shouldn't necessarily drive design, but it was a useful simplification. Still, it needs to be reviewed. The difference is that in GCC (function:MODE arg1 ...) is written in CGEN as (function MODE arg1 ...). Note the space after ‘function’.

GCC RTL allows flags to be recorded with RTL (e.g. MEM_VOLATILE_P). This is supported in CGEN RTL by prefixing each RTL function's arguments with an optional list of modifiers: (function (#:mod1 #:mod2) MODE arg1 ...). The list is a set of modifier names prefixed with '#:'. They can take arguments. ??? Modifiers are supported by the RTL traversing code, but no use is made of them yet.

The mode may be elided if it can be deduced from the operands. For example, while the full form of add is ‘(add () MODE arg1 arg2)’, it may be written as ‘(add arg1 arg2)’, with the mode being taken from the mode of ‘arg1’. The fully specified version is called the “canonical” form.

The currently defined semantic functions are:

(set mode destination source)

Assign ‘source’ to ‘destination’ reference in mode ‘mode’.

(set-quiet mode destination source)

Assign ‘source’ to ‘destination’ referenced in mode ‘mode’, but do not print any tracing message.

(reg mode hw-name [index])

Return an `operand' of hardware element ‘hw-name’ in mode ‘mode’. If ‘hw-name’ is an array, ‘index’ selects which register.

(raw-reg mode hw-name [index])

Return an `operand' of hardware element ‘hw-name’ in mode ‘mode’, bypassing any get or set specs of the register. If ‘hw-name’ is an array, ‘index’ selects which register. This cannot be used with virtual registers (those specified with the ‘VIRTUAL’ attribute).

raw-reg is most often used in get and set specs of a register: if it weren't read and write operations would infinitely recurse.

(mem mode address)

Return an `operand' of memory referenced at ‘address’ in mode ‘mode’.

(const mode value)

Return an `operand' of constant ‘value’ in mode ‘mode’.

(enum mode value-name)

Return an `operand' of constant ‘value-name’ in mode ‘mode’. The value must be from a previously defined enum.

(subword mode value word-num)

Return part of ‘value’. Which part is determined by ‘mode’ and ‘word-num’. There are three cases.

If ‘mode’ is the same size as the mode of ‘value’, ‘word-num’ must be ‘0’ and the result is ‘value’ recast in the new mode. There is no change in the bits of ‘value’, they're just interpreted in a possibly different mode. This is most often used to interpret an integer value as a float and vice versa.

If ‘mode’ is smaller than the mode of ‘value’, ‘value’ is divided into N pieces and ‘word-num’ picks which piece. All pieces have the size of ‘mode’ except possibly the last. If the last piece has a different size, it cannot be referenced. Word number 0 is the most significant word, regardless of endianness.

If ‘mode’ is larger than the mode of ‘value’, ‘value’ is interpreted in the larger mode with the upper most significant bits treated as garbage (their value is assumed to be unimportant to the context in which the value will be used). ‘word-num’ must be ‘0’.

(join out-mode in-mode arg1 . arg-rest)

Concatenate ‘arg1[,arg2[,...]]’ to create a value of mode ‘out-mode’. ‘arg1’ becomes the most significant part of the result. Each argument is interpreted in mode ‘in-mode’. ‘in-mode’ must evenly divide ‘out-mode’.

(sequence mode ((mode1 local1) ...) expr1 ...)

Execute ‘expr1’, ‘expr2’, etc. sequentially. At least one expression must be specified, even if the result mode is ‘VOID’.

The result, if non-void-mode, is the value of the last expression.

‘mode’ is the mode of the result. If ‘mode’ is elided it is set to ‘VOID’ (void mode).

`((mode1 local1) ...)' is a set of local variables.

(parallel mode empty expr1 ...)

Execute ‘expr1’, ‘expr2’, etc. in parallel. All inputs are read before any output is written. At least one expression must be specified.

‘empty’ must be ‘()’ and is present for consistency with ‘sequence’.

‘mode’ must be ‘VOID’ (void mode), or it can be elided.

(do-count mode iteration-variable number-of-iterations expr1 ...)

This is a simple looping operation. Execute ‘expr1’, ‘expr2’, etc. the specified number of times. At least one expression must be specified.

‘iteration-variable’ will contain the iteration number and is available for use in expressions. It has mode ‘INT’. It's value will be 0 ... ‘number-of-iterations’ - 1.

‘number-of-iterations’ is an rtl expression of mode INT (or a compatible mode). It is computed once and may not be modified inside the loop.

‘mode’ must be ‘VOID’ (void mode), or it can be elided.

(unop mode operand)

Perform a unary arithmetic operation.

‘unop’ is one of neg, abs, inv, not, zflag, nflag. zflag returns a bit indicating if ‘operand’ is zero. nflag returns a bit indicating if ‘operand’ is negative. inv returns the bitwise complement of ‘operand’, whereas not returns its logical negation.

(binop mode operand1 operand2)

Perform a binary arithmetic operation.

‘binop’ is one of add, sub, and, or, xor, mul, div, udiv, mod, umod.

(binop-with-bit mode operand1 operand2 operand3)

Same as ‘binop’, except taking 3 operands. The third operand is always a single bit.

‘binop-with-bit’ is one of addc, addc-cflag, addc-oflag, subc, subc-cflag, subc-oflag.

Note: The following are deprecated:

• add-cflag, replaced with addc-cflag
• add-oflag, replaced with addc-oflag
• sub-cflag, replaced with subc-cflag
• sub-cflag, replaced with subc-oflag

(shiftop mode operand1 operand2)

Perform a shift operation. ‘operand1’ is shifted (or rotated) by the amount specified in ‘operand2’.

‘shiftop’ is one of sll, srl, sra, ror, rol.

‘mode’ must match the mode of ‘operand1’. The mode of ‘operand1’ may be any integral mode. The mode of ‘operand2’ may be any integral mode, and need not match the mode of ‘operand1’.

It is an error if ‘operand2’ is negative or greater than or equal to the size of ‘operand1’. If the architecture handles negative or large shift amounts, that needs to be handled in the surrounding RTL.

(andif mode operand1 operand2)

Evaluate ‘operand1’. If it evaluates to zero the result is zero, and ‘operand2’ is not evaluated. If ‘operand1’ evaluates to non-zero, then evaluate ‘operand2’. If it evaluates to non-zero the result is one, otherwise the result is zero.

The mode of the result is ‘BI’. ‘mode’ is generally elided or is ‘BI’.

(orif mode operand1 operand2)

Evaluate ‘operand1’. If it evaluates to non-zero the result is one, and ‘operand2’ is not evaluated. If ‘operand1’ evaluates to zero, then evaluate ‘operand2’. If it evaluates to non-zero the result is one, otherwise the result is zero.

The mode of the result is ‘BI’. ‘mode’ is generally elided or is ‘BI’.

(integer-convop mode operand)

Perform an integer mode->mode conversion operation.

‘integer-convop’ is one of:

• ext Sign-extend ‘operand’, which must have an integer mode narrower than ‘mode’, which also must be an integer mode.
• zext Zero-extend ‘operand’, which must have an integer mode narrower than ‘mode’, which also must be an integer mode.
• trunc Truncate ‘operand’, which must have an integer mode wider than ‘mode’, which also must be an integer mode.

(float-convop mode how operand)

Perform a mode->mode conversion operation involving a floating point value.

Conversions involving floating point values need to specify how things like truncation will be performed, e.g., the rounding mode. ‘how’ is an rtx of mode ‘INT’ that specifies how the conversion will be performed. The interpretation of ‘how’ is architecture-dependent, except that a value of zero has a specific meaning: If a particular floating-point conversion can only be done one way, or if the conversion is to be done the “default” way, specify zero for ‘how’. What “the default way” is is application-dependent.

‘float-convop’ is one of:

• fext Extend ‘operand’, which must have a floating point mode narrower than ‘mode’, which also must be a floating point mode.
• ftrunc Truncate ‘operand’, which must have a floating point mode wider than ‘mode’, which also must be a floating point mode.
• float Convert ‘operand’, which must have an integer mode, to a floating point value of mode ‘mode’. ‘operand’ is treated as a signed integer.
• ufloat Convert ‘operand’, which must have an integer mode, to a floating point value of mode ‘mode’. ‘operand’ is treated as an unsigned integer.
• fix Convert ‘operand’, which must have a floating point mode, to a signed integer of mode ‘mode’.
• ufix Convert ‘operand’, which must have a floating point mode, to an unsigned integer of mode ‘mode’.

An enum is defined that specifies several predefined rounding modes.

(define-enum (name fpconv-kind) (comment "builtin floating point conversion kinds") (attrs VIRTUAL) ;; let app provide def'n instead of each cpu's desc.h (prefix FPCONV-) (values ((DEFAULT 0) (TIES-TO-EVEN 1) (TIES-TO-AWAY 2) (TOWARD-ZERO 3) (TOWARD-POSITIVE 4) (TOWARD-NEGATIVE 5))) )

(cmpop mode operand1 operand2)

Perform a comparison.

‘cmpop’ is one of eq, ne, lt, le, gt, ge, ltu, leu, gtu, geu.

If the comparison succeeds the result is one, otherwise the result is zero. The mode of the result is ‘BI’.

(mathop mode operand)

Perform a mathematical operation.

‘mathop’ is one of sqrt, cos, sin.

(*nan mode operand)

Return a boolean indicating if ‘operand’ is a NaN. ‘mode’ must be a floating point mode. There are three versions.

• (nan operand) Test whether ‘operand’ is any kind of NaN.
• (qnan operand) Test whether ‘operand’ is a quiet NaN.
• (snan operand) Test whether ‘operand’ is a signalling NaN.

(if mode condition then [else])

Standard if statement.

‘condition’ is any arithmetic expression. If the value is non-zero the ‘then’ part is executed. Otherwise, the ‘else’ part is executed (if present).

‘mode’ is the mode of the result, not of ‘condition’. If ‘mode’ is not VOID (void mode), ‘else’ must be present. When the result is used, ‘mode’ must specified, and not be VOID.

(cond mode (condition1 expr1a ...) (...) [(else exprNa...)])

From Scheme: keep testing conditions until one succeeds, and then process the associated expressions.

(case mode test ((case1 ..) expr1a ..) (..) [(else exprNa ..)])

From Scheme: Compare ‘test’ with ‘case1’, ‘case2’, etc. and process the associated expressions.

(c-code mode "C expression")

An escape hook to insert arbitrary C code. ‘mode’ must the compatible with the result of “C expression”.

(c-call mode symbol operand1 operand2 ...)

An escape hook to emit a subroutine call to function named ‘symbol’ passing operands ‘operand1’, ‘operand2’, etc. An implicit first argument of current_cpu is passed to ‘symbol’. ‘mode’ is the mode of the result. Be aware that ‘symbol’ will be restricted by reserved words in the C programming language and by existing symbols in the generated code.

(c-raw-call mode symbol operand1 operand2 ...)

Same as c-call: except there is no implicit current_cpu first argument. ‘mode’ is the mode of the result.

(clobber mode object)

Indicate that ‘object’ is written in mode ‘mode’, without saying how. This could be useful in conjunction with the C escape hooks.

(delay mode num expr)

Indicate that there are ‘num’ delay slots in the processing of ‘expr’. When using this rtx in instruction semantics, CGEN will infer that the instruction has the DELAY-SLOT attribute.

(delay num expr)

In older "sim" simulators, indicates that there are ‘num’ delay slots in the processing of ‘expr’. When using this rtx in instruction semantics, CGEN will infer that the instruction has the DELAY-SLOT attribute.

In newer "sid" simulators, evaluates to the writeback queue for hardware operand ‘expr’, at ‘num’ instruction cycles in the future. ‘expr’ must be a hardware operand in this case.

For example, (set (delay 3 pc) (+ pc 1)) will schedule write to the ‘pc’ register in the writeback phase of the 3rd instruction after the current. Alternatively, (set gr1 (delay 3 gr2)) will immediately update the ‘gr1’ register with the latest write to the ‘gr2’ register scheduled between the present and 3 instructions in the future. (delay 0 ...) refers to the writeback phase of the current instruction.

This effect is modeled with a circular buffer of "write stacks" for each hardware element (register banks get a single stack). The size of the circular buffer is calculated from the uses of (delay ...) rtxs. When a delayed write occurs, the simulator pushes the write onto the appropriate write stack in the "future" of the circular buffer for the written-to hardware element. At the end of each instruction cycle, the simulator executes all writes in all write stacks for the time slice just ending. When a delayed read (essentially a pipeline bypass) occurs, the simulator looks ahead in the circular buffer for any writes scheduled in the future write stack. If it doesn't find one, it progressively backs off towards the "current" instruction cycle's write stack, and if it still finds no scheduled writes then it returns the current state of the CPU. Thus while delayed writes are fast, delayed reads are potentially slower in a simulator with long pipelines and very large register banks.

(annul yes?)

Annul the following instruction if ‘yes?’ is non-zero. This rtx is an experiment and will probably change.

(skip yes?)

Skip the next instruction if ‘yes?’ is non-zero. This rtx is an experiment and will probably change.

(symbol name)

Return a symbol with value ‘name’, for use in attribute processing. This is equivalent to ‘quote’ in Scheme but ‘quote’ sounds too jargonish.

(int-attr mode object attr-name)

Return the value of attribute ‘attr-name’ in mode ‘mode’. ‘object’ must currently be ‘(current-insn)’, the current instruction, or ‘(current-mach)’, the current machine. The attribute's value must be representable as an integer.

(eq-attr mode object attr-name value)

Return non-zero if the value of attribute ‘attr-name’ of object ‘object’ is ‘value’.

NOTE: List values of ‘value’ may be changed to allow use the ‘number-list’ rtx function. If ‘value’ is a list return “true” if the attribute is any of the listed values. But this is not implemented yet.

(index-of operand)

Return the index of ‘operand’. For registers this is the register number.

(regno operand)

Same as index-of, but improves readability for registers.

(error mode message)

Emit an error message from CGEN RTL. Error message is specified by ‘message’.

(nop)

A no-op.

(ifield field-name)

Return the value of field ‘field-name’. ‘field-name’ must be a field in the instruction.

Operands can be any of:

• an operand defined in the description file
• a register reference, created with (reg mode [index])
• a memory reference, created with (mem mode address)
• a constant, created with (const mode value)
• a `sequence' local variable
• a `do-count' iteration variable
• another expression

The ‘symbol’ in a c-call or c-raw-call function is currently the name of a C function or macro that is invoked by the generated semantic code.

3.20 Macro-expressions

Macro RTL expressions are a way to not have to always specify a mode for every expression (and sub-expression thereof). Whereas the formal way to specify, say, an add is (add SI arg1 arg2) if SI is the default mode of `arg1' then this can be simply written as (add arg1 arg2). This gets expanded to (add DFLT arg1 arg2) where DFLT means “default mode”.

It might be possible to replace macro expressions with preprocessor macros, however for the nonce there is no plan to do this.

This document was generated by Doug Evans on January, 28 2010 using texi2html 1.78.