This document presents a state machine for a parser for escape and control sequences, suitable for use in a VT emulator. It is claimed to have two important properties:
Completeness – it specifies the actions and transitions for every incoming character for every state of the parser. In particular, it covers the behaviour of the C0 controls (0/0 to 1/15) and characters 10/0 to 15/15 for every state.
Completeness does not mean that this state diagram contains all the information you need to write a terminal emulator! There are no details here of the mapping of character sets or of cursor movement behaviour, to name just two examples. However, it does specify how every incoming character affects the parser’s state.
Correctness – if you were to feed this parser a stream of characters that is random or deliberately pathological, it is claimed that this parser will exhibit the same visible behaviour as any one of DEC’s 8-bit ANSI-compatible terminals, from VT220 to VT525. A VT100-series parser would be simpler still, as the VT100 only supports 7-bit characters.
During the discussion of this design, I will mention some real terminal emulators by name. This is for comparative purposes only and no criticism is intended of decisions made by the authors or maintainers of the applications referred to. After all, I am not presenting a product to the world, merely an ideal software model and I am free to ignore efficiency!
In this document, “VT500” is used as shorthand for the VT500 series of terminals, the VT510, VT520 and VT525.
All of DEC’s terminals from the VT100 onward are compatible with ANSI X3.64-1979, “Additional Controls for Use with American National Standard Code for Information Interchange”, hereafter referred to just as X3.64. However, X3.64 defines many implementation-dependent features and error conditions without defining recovery procedures. A sample of these is given below; a more detailed treatment appears later.
Occurrences of characters 00-1F or 7F-FF in an escape sequence or control sequence is an error condition whose recovery is not specified.
For control sequences, the maximum length of parameter string is defined by implementation.
For control sequences, occurrences of a parameter character after an intermediate character is an error condition.
A terminal is a closed box that doesn’t normally report errors in its input stream to the host, so it must define a recovery procedure for all the circumstances left undefined by X3.64. DEC defined the recoveries for their terminals, so emulators should match these exactly¹.
The UML State Diagram should be readable to anyone who has seen a picture of a state machine before, but here are some notes on reading it.
Rounded boxes are states. A horizontal line separates the name of the state from the event list. The event list contains several event/action pairs, as shown below.
entry / osc_start
event 01-17,19,1C-1F / ignore
event 20-7F / osc_put
exit / osc_end
The entry event happens when a state is first entered. The events list incoming symbols which cause an action to take place while remaining in that state. The action associated with the exit event happens when an incoming symbol causes a transition from this state to another state (or even back to the same state). When going from one state to another, the actions take place in this order:
States with grey backgrounds are duplicates of a state described more fully somewhere else on the diagram. They are present to prevent too many lines crossing. Transitions shown with grey lines serve as reminders that certain events cause a change of state from anywhere.
All events in this diagram are the hex values of the incoming bytes. Although we are used to presenting
sequences in the form
ESC [ 3 m for readability, using hex values instead of ASCII
characters emphasises that these sequences are not dependent on the character encoding in use.
Section numbers (e.g. §3.5.6) refer to ANSI X3.64.
There are no explicit actions shown for incoming codes in the GR area (A0-FF). In all states, these codes are treated identically to GL codes 20-7F. This behaviour is suggested in Appendix H of X3.64:
In an 8-bit code, the bit combinations of columns 10 to 15 (except 10/0 and 15/15) are permitted to represent: - parameters, intermediates, and finals of a control sequence; - the contents of a control string; - the operand of a single-shift character. In these situations, the bit combinations in the range 10/1 to 15/14 have the same meanings as the corresponding bit combinations in the range 2/1 to 7/14.
This is the initial state of the parser, and the state used to consume all characters other than components of escape and control sequences.
GL characters (20 to 7F) are printed. I have included 20 (SP) and 7F (DEL) in this area, although both codes have special behaviour. If a 94-character set is mapped into GL, 20 will cause a space to be displayed, and 7F will be ignored. When a 96-character set is mapped into GL, both 20 and 7F may cause a character to be displayed. Later models of the VT220 included the DEC Multinational Character Set (MCS), which has 94 characters in its supplemental set (i.e. the characters supplied in addition to ASCII), so terminals only claiming VT220 compatibility can always ignore 7F. The VT320 introduced ISO Latin-1, which has 96 characters in its supplemental set, so emulators with a VT320 compatibility mode need to treat 7F as a printable character.
This state is entered whenever the C0 control ESC is received. This will immediately cancel any escape sequence, control sequence or control string in progress. If an escape sequence or control sequence was in progress, “cancel” means that the sequence will have no effect, because the final character that determines the control function (in conjunction with any intermediates) will not have been received. However, the ESC that cancels a control string may occur after the control function has been determined and the following string has had some effect on terminal state. For example, some soft characters may already have been defined. Cancelling a control string does not undo these effects.
A control string that started with DCS, OSC, PM or APC is usually terminated by the C1 control ST
(String Terminator). In a 7-bit environment, ST will be represented by
ESC \ (1B 5C).
However, receiving the ESC character will “cancel” the control string, so the ST control function that is
invoked by the arrival of the following “\” is essentially a “no-op” function. Does this point seem like
pure trivia? Maybe, but I worried for ages about whether the control string recogniser needed a one character
lookahead in order to know whether
ESC \ was going to terminate it. The actual solution
became clear when I was using ReGIS on a VT330: sending ESC immediately caused the graphics output cursor
to disappear from the screen, so I knew that the control string had already finished before the “\” arrived.
Many of the clues that enabled me to derive this state diagram have been as subtle as that.
This state is entered when an intermediate character arrives in an escape sequence. Escape sequences have no parameters, so the control function to be invoked is determined by the intermediate and final characters. In this parser there is just one escape intermediate, and the parser uses the collect action to remember intermediate characters as they arrive, for processing by the esc_dispatch action when the final character arrives. An alternate approach (and the one adopted by xterm) is to have multiple copies of this state and choose the next appropriate one as each intermediate character arrives. I think that this alternate approach is merely an optimisation; the approach presented here doesn’t require any more states if the repertoire of supported control functions increases.
This state is only split from the escape state because certain escape sequences are the 7-bit representations of C1 controls that change the state of the parser. Without these “compatibility sequences”, there could just be one escape state to collect intermediates and dispatch the sequence when a final character was received.
This state is entered when the control function CSI is recognised, in 7-bit or 8-bit form. This state
will only deal with the
first character of a control sequence, because the characters 3C-3F can only appear as the first character
of a control sequence, if they appear at all. Strictly speaking, X3.64 says that the entire string is “subject
to private or experimental interpretation” if the first character is one of 3C-3F, which allows sequences like
CSI ?::<? F, but Digital’s
terminals only ever used one private-marker character at a time. As far as I am aware, only characters 3D (=), 3E (>) and 3F (?) were
used by Digital.
C0 controls are executed immediately during the recognition of a control sequence. C1 controls will cancel the sequence and then be executed. I imagine this treatment of C1 controls is prompted by the consideration that the 7-bit (ESC Fe) and 8-bit representations of C1 controls should act in the same way. When the first character of the 7-bit representation, ESC, is received, it will cancel the control sequence, so the 8-bit representation should do so as well.
This state is entered when a parameter character is recognised in a control sequence. It then recognises other parameter characters until an intermediate or final character appears. Further occurrences of the private-marker characters 3C-3F or the character 3A, which has no standardised meaning, will cause transition to the csi ignore state.
This state is entered when an intermediate character is recognised in a control sequence. It then recognises other intermediate characters until a final character appears. If any more parameter characters appear, this is an error condition which will cause a transition to the csi ignore state.
Neither X3.64 nor Digital defined any control sequences with more than one intermediate character, although X3.64 doesn’t place any limit on the possible number.
This state is used to consume remaining characters of a control sequence that is still being recognised, but has already been disregarded as malformed. This state will only exit when a final character is recognised, at which point it transitions to ground state without dispatching the control function. This state may be entered because:
C0 controls will still be executed while a control sequence is being ignored.
This state is entered when the control function DCS is recognised, in 7-bit or 8-bit form. X3.64 doesn’t define any structure for device control strings, but Digital made them appear like control sequences followed by a data string, with a form and length dependent on the control function. This state is only used to recognise the first character of the control string, mirroring the csi entry state.
C0 controls other than CAN, SUB and ESC are not executed while recognising the first part of a device control string.
This state is entered when a parameter character is recognised in a device control string. It then recognises other parameter characters until an intermediate or final character appears. Occurrences of the private-marker characters 3C-3F or the undefined character 3A will cause a transition to the dcs ignore state.
This state is entered when an intermediate character is recognised in a device control string. It then recognises other intermediate characters until a final character appears. If any more parameter characters appear, this is an error condition which will cause a transition to the dcs ignore state.
This state is a shortcut for writing state machines for all possible device control strings into the main parser. When a final character has been recognised in a device control string, this state will establish a channel to a handler for the appropriate control function, and then pass all subsequent characters through to this alternate handler, until the data string is terminated (usually by recognising the ST control function).
This state has an exit action so that the control function handler can be informed when the data string has come to an end. This is so that the last soft character in a DECDLD string can be completed when there is no other means of knowing that its definition has ended, for example.
This state is used to consume remaining characters of a device control string that is still being recognised, but has already been disregarded as malformed. This state will only exit when the control function ST is recognised, at which point it transitions to ground state. This state may be entered because:
These conditions are only errors in the first part of the control string, until a final character has been recognised. The data string that follows is not checked by this parser.
This state is entered when the control function OSC (Operating System Command) is recognised. On entry it prepares an external parser for OSC strings and passes all printable characters to a handler function. C0 controls other than CAN, SUB and ESC are ignored during reception of the control string.
The only control functions invoked by OSC strings are DECSIN (Set Icon Name) and DECSWT (Set Window Title), present on the multisession VT520 and VT525 terminals. Earlier terminals treat OSC in the same way as PM and APC, ignoring the entire control string.
The VT500 doesn’t define any function for these control strings, so this state ignores all received characters until the control function ST is recognised.
This isn’t a real state. It is used on the state diagram to show transitions that can occur from any state to some other state. These invariant transitions are:
On the VT220, VT420 and VT500, the C0 controls CAN and SUB cancel any escape sequence, control sequence or control string in progress and return to ground state. SUB will also display the error character, a reversed question mark, “”. The programmer’s information for the VT320 says that CAN and SUB “no longer” cancel these sequences, so there must have been a rethink when the VT420 was being designed.
All C1 controls cancel any escape sequence, control sequence or control string in progress and are executed. Control functions special to this parser, i.e. DCS, SOS, CSI, OSC, PM and APC, cause a transition to their appropriate states. All other C1 control functions (even those with no defined meaning), cause a transition to ground state.
On terminals earlier than the VT500, there would have been one other invariant action: the C0 control NUL was ignored on input to the terminal and would not take part in any processing. Its only purpose was as a time-fill character. However, the VT500 defines a control function DECNULM (Null Mode), which allows NUL to be passed to an attached printer. So in this parser, NUL is treated the same as other C0 controls.
An event may cause one of these actions to occur with or without a change of state.
The character or control is not processed. No observable difference in the terminal’s state would occur if the character that caused this action was not present in the input stream. (Therefore, this action can only occur within a state.)
This action only occurs in ground state. The current code should be mapped to a glyph according to the character set mappings and shift states in effect, and that glyph should be displayed. 20 (SP) and 7F (DEL) have special behaviour in later VT series, as described in ground.
The C0 or C1 control function should be executed, which may have any one of a variety of effects, including changing the cursor position, suspending or resuming communications or changing the shift states in effect. There are no parameters to this action.
This action causes the current private flag, intermediate characters, final character and parameters to be forgotten. This occurs on entry
to the escape, csi entry and dcs entry states, so that erroneous sequences
CSI 3 ; 1 CSI 2 J are handled correctly.
The private marker or intermediate character should be stored for later use in selecting a control function to be executed when a final character arrives. X3.64 doesn’t place any limit on the number of intermediate characters allowed before a final character, although it doesn’t define any control sequences with more than one. Digital defined escape sequences with two intermediate characters, and control sequences and device control strings with one. If more than two intermediate characters arrive, the parser can just flag this so that the dispatch can be turned into a null operation.
This action collects the characters of a parameter string for a control sequence or device control sequence and builds a list of parameters. The characters processed by this action are the digits 0-9 (codes 30-39) and the semicolon (code 3B). The semicolon separates parameters. There is no limit to the number of characters in a parameter string, although a maximum of 16 parameters need be stored. If more than 16 parameters arrive, all the extra parameters are silently ignored.
The VT500 Programmer Information is inconsistent regarding the maximum value that a parameter can take. In section 18.104.22.168 of EK-VT520-RM it says that “any parameter greater than 9999 (decimal) is set to 9999 (decimal)”. However, in the description of DECSR (Secure Reset), its parameter is allowed to range from 0 to 16383. Because individual control functions need to make sure that numeric parameters are within specific limits, the supported maximum is not critical, but it must be at least 16383.
Most control functions support default values for their parameters. The default value for a parameter is given by either leaving the parameter blank, or specifying a value of zero. Judging by previous threads on the newsgroup comp.terminals, this causes some confusion, with the occasional assertion that zero is the default parameter value for control functions. This is not the case: many control functions have a default value of 1, one (GSM) has a default value of 100, and some have no default. However, in all cases the default value is represented by either zero or a blank value.
In the standard ECMA-48, which can be considered X3.64’s successor², there is a distinction between a parameter with an empty value (representing the default value), and one that has the value zero. There used to be a mode, ZDM (Zero Default Mode), in which the two cases were treated identically, but that is now deprecated in the fifth edition (1991). Although a VT500 parser needs to treat both empty and zero parameters as representing the default, it is worth considering future extensions by distinguishing them internally.
The final character of an escape sequence has arrived, so determined the control function to be executed from the intermediate character(s) and final character, and execute it. The intermediate characters are available because collect stored them as they arrived.
A final character has arrived, so determine the control function to be executed from private marker, intermediate character(s) and final character, and execute it, passing in the parameter list. The private marker and intermediate characters are available because collect stored them as they arrived.
Digital mostly used private markers to extend the parameters of existing X3.64-defined control functions, while keeping a similar meaning. A few examples are shown in the table below.
|No Private Marker||With Private Marker|
|SM, Set ANSI Modes||SM, Set Digital Private Modes|
|ED, Erase in Display||DECSED, Selective Erase in Display|
|CPR, Cursor Position Report||DECXCPR, Extended Cursor Position Report|
In the cases above, csi_dispatch needn’t know about the private marker at all, as long as it is passed along to the control function when it is executed. However, the VT500 has a single case where the use of a private marker selects an entirely different control function (DECSTBM, Set Top and Bottom Margins and DECPCTERM, Enter/Exit PCTerm or Scancode Mode), so this action needs to use the private marker in its choice. xterm takes the same approach for efficiency, even though it doesn’t support DECPCTERM.
The selected control function will have access to the list of parameters, which it will use some or all
of. If more parameters are supplied than the control function requires, only the earliest parameters will be
used and the rest will be ignored. If too few parameters are supplied, default values will be used. If the
control function has no default values, defaulted parameters will be ignored; this may result in the control
function having no effect. For example, if the SM (Set Mode)
control function is invoked with the sequence
CSI 2;0;5 h, the second parameter will be ignored because
SM has no default value.
This action is invoked when a final character arrives in the first part of a device control string. It determines the control function from the private marker, intermediate character(s) and final character, and executes it, passing in the parameter list. It also selects a handler function for the rest of the characters in the control string. This handler function will be called by the put action for every character in the control string as it arrives.
This way of handling device control strings has been selected because it allows the simple plugging-in of extra parsers as functionality is added. Support for a fairly simple control string like DECDLD (Downline Load) could be added into the main parser if soft characters were required, but the main parser is no place for complicated protocols like ReGIS.
This action passes characters from the data string part of a device control string to a handler that has previously been selected by the hook action. C0 controls are also passed to the handler.
When a device control string is terminated by ST, CAN, SUB or ESC, this action calls the previously selected handler function with an “end of data” parameter. This allows the handler to finish neatly.
When the control function OSC (Operating System Command) is recognised, this action initializes an external parser (the “OSC Handler”) to handle the characters from the control string. OSC control strings are not structured in the same way as device control strings, so there is no choice of parsers.
This action passes characters from the control string to the OSC Handler as they arrive. There is therefore no need to buffer characters until the end of the control string is recognised.
This action is called when the OSC string is terminated by ST, CAN, SUB or ESC, to allow the OSC handler to finish neatly.
As I said above, X3.64 deliberately leaves some decisions to implementers. It doesn’t define recovery from error conditions, and some limits are implementation dependent. The following sections define DEC’s method of coping with all of these sections of the standard.
X3.64: §2.2.3 Classes of Bit Combinations
Example: The format of an Escape sequence as defined in ANSI X3.41-1974 and used in this standard is: ESC I...I F [...] (4) The occurrence of characters in the inclusive ranges of 0/0 to 1/15 and 7/15 to 15/15 is an error condition whose recovery is not specified.
DEC: The C0 controls 00-1F are executed, 7F is ignored, C1 controls 80-9F are executed (cancelling the escape sequence) and GR codes A0-FF are treated as GL codes 20-7F.
X3.64: §3.2.1 Software Control Strings
The opening delimiters for the software strings are: Name Mnemonic Operating System Command OSC Privacy Message PM Application Program Command APC The string is terminated by the occurrence of a String Terminator (see 3.2.3). The occurrence of other control characters and/or characters from columns 10 to 15 within such a string are error conditions whose recovery is not specified by this standard.
DEC: None of Digital’s terminals define any meaning for received Privacy Message or Application Program Command strings. C0 controls, GL characters and GR characters are ignored. C1 controls will cancel the sequence.
The VT500 defines two Operating System Commands, DECSIN (Set Icon Name) and DECSWT (Set Window Title). C0 controls are ignored in these sequences. C1 controls will cancel the sequence.
X3.64: §3.2.2 Device Control Strings
These strings take the form of the introducer character DCS followed by one or more bit combinations representing the function. These may be characters from columns 2 to 7, excluding 7/15 (2/0 to 7/14) [...] The occurrence of other control characters and/or characters from columns 10 to 15 within such a string are error conditions whose recovery is not specified by this standard.
DEC: On Digital’s terminals, device control strings have structure that is not specified by X3.64. The first section of a device control string has the same structure as a control sequence, up to what would be the final character of a control sequence. At that point, the device function is determined from the intermediate and final characters, and the rest of the string has a meaning specific to the selected device function. In the first section, C0 controls are ignored. In the second section, C0 controls may or may not have a meaning to the device function. If they have no meaning for the device function (e.g. when defining soft characters with DECDLD) they will be silently ignored. If they have a meaning for the device function (e.g. when in ReGIS mode), they will be acted upon.
X3.64: §3.5 Control Sequence Functions
The general form of a control sequence function is as follows: CSI P...P I...I F [...] (2) P...P is called the “parameter string.” The minimum length is zero, and the maximum length is defined by the implementation. However, all bit combinations are from 3/0 to 3/15 inclusive. [...] (5) The occurrence of bit combinations from columns 0 and 1 (0/0 to 1/15), from columns 8 to 15 (8/0 to 15/15), or position 7/15 in control sequences are error conditions whose recovery is not specified by this standard.
DEC: There is no limit to the number of characters in the parameter string, but a maximum of 16 parameters will be processed. All parameters beyond the 16th will be silently ignored. In a control sequence the C0 controls 00-1F are executed, 7F is ignored, C1 controls 80-9F are executed (cancelling the control sequence) and GR codes A0-FF are treated as GL codes 20-7F.
X3.64: §3.5.1 Parameter Values
The bit combination 3/10 is reserved for future standardization.
DEC: Character 3A (ASCII colon) will cause the control sequence to be ignored, though not cancelled. All characters up to a valid final character will be collected, but then no action will take place.
X3.64: §3.5.5 Selective Parameters
The maximum number of parameters in a selective Control Sequence is implementation-defined, as is the order of performance and the effect of conflicting or unusual combinations.
DEC: There can be a maximum of 16 parameters in
a control sequence. All parameters beyond the 16th will
be silently ignored. Parameters are processed from first to last, with
conflicts simply resolved by allowing later parameters to override the
effects of earlier ones. This means that the sequence
CSI 7;0 m
will not set the reverse video attribute, because the parameter ‘7’ is cancelled
by the ‘0’.
X3.64: §3.5.6 Structure of Control Sequences
For both numeric and selective parameters the complete control sequence structure is: CSI P11...P1m 3/11 P21...P2m 3/11 ... ... 3/11 Pn1...Pnm I...I F If P11 is 3/0 to 3/11, inclusive, the parameter string is interpreted according to the standard format described below. If P11 is 3/12 to 3/15, inclusive, the entire parameter string is subject to private or experimental interpretation. [...] (2) Px1...Pxm is a numeric or selective parameter. Pxy is 3/0 to 3/9 inclusive for standardized parameters. Occurrences of 3/12 to 3/15 inclusive are undefined. [...] NOTE: The occurrence of bit combinations in the following inclusive ranges: 0/0 to 1/15, or 7/15 to 15/15, or the occurrence of a P after an I has been encountered is an error condition whose recovery is not specified by this standard.
DEC: Occurrences of 3C to 3F in positions other than the first character of the parameter string cause the entire control sequence to be ignored. In a control sequence the C0 controls 00-1F are executed, 7F is ignored, C1 controls 80-9F are executed (cancelling the control sequence) and GR codes A0-FF are treated as GL codes 20-7F.
As of 2005, Josh Haberman has implemented this parser in C and placed it in the public domain. You will also need Ruby to create the parser tables at compile time. It’s on GitHub.
If you have any questions about this document, please send them to me, no matter how trivial you think they are. Even if the answer is already stated here, it may need clarification (or writing in bigger letters!) If you try to write the parser for a terminal emulator from this specification and you find you need a leap of logic, I’ve not done my job properly, and I’d like to hear about it.
It is debatable how far it is necessary to go with making an emulator match the error-recovery behaviour of the terminal, for two reasons. Firstly, for the practical reason that information on error recovery isn’t contained in DEC’s terminal manuals and discovering it means taking detailed and seemingly-endless notes about the terminal’s behaviour when certain bizarre sequences are sent to it. (OK, I’ve done that!)
Secondly, how often would erroneous sequences be sent
to the terminal anyway? I would answer this by saying that people who write applications for terminals
don’t always read the manuals and may rely on some observed behaviour of the terminal without realising
that they are seeing the effects of error recovery. It appears to be common knowledge among emulator
writers (and their critics) that the sequence
CSI 2 LF C moves
the cursor two columns right and one row down. How many realise that this behaviour is not specified in
X3.64, but just happens to have been the error recovery chosen by the designers of the VT100?
The lesson I take from this is that if you’re going to emulate a real terminal, you should match
all observable behaviour.
With its first edition having been published in 1976, ECMA-48 “Control Functions for Coded Character Sets” predates ANSI X3.64 and has been updated for longer. As ECMA make their standards available free of charge, I find it surprising that anyone ever bothered claiming conformance with ANSI X3.64.