Avoidance of changes which might cause existing games to misbehave; and
Minimisation of features which would require updates to the Inform Designer’s Manual.
Since the first release of the Inform 6.3 compiler, the Inform 6 library has been split into a separate project, maintained at https://gitlab.com/DavidGriffith/inform6lib.
The compiler can now handle switch cases which are expressions, rather than just looking for bare literals and symbols. The expression must still evaluate to a constant, but now parentheses and constant-folded arithmetic are handled:
Constant CONST = 5;
! These have always worked.
switch (x) {
0: return 0;
1: return 1;
-2: return -2;
}
! These now also work.
switch (x) {
(0): return 0;
(-(1)): return -1;
(CONST): return 5;
(CONST+1): return 6;
}
For backwards compatibility, the expression must be wrapped in parens, so -(1): is not a valid case. Lists
of expressions are also supported. Expression parsing applies as long as the first value is wrapped in parens.
Wrapping the entire list in parens also works:
switch (x) {
1, 2, 3: return 0; ! old style
(4), (CONST), (CONST+1): return 1; ! new style
(10), CONST+6, CONST+7: return 2; ! this also works
(20, CONST+16, CONST+17): return 3; ! as does this
}
Note that the to keyword does not support expressions. You cannot say (CONST) to (CONST+5):
as a case. Also, case expressions only work within a switch statement. Top-level action cases must still be bare
action names.
Inform identifiers can now be any length, and the entire identifier is significant. Dictionary words can also be any length. The DICT_WORD_SIZE limit still applies, but now dictionary words are silently trimmed to DICT_WORD_SIZE. For Glulx, DICT_WORD_SIZE can now be increased without limit.
Arbitrary bytes and words can be compiled into the game, using two new statements:
@ -> BYTE BYTE BYTE ...; @ --> WORD WORD WORD ...;The given bytes or words are directly copied out into the function. (Words are two bytes in Z-code, and four bytes in Glulx.) The compiler assumes that the data forms valid opcodes, but does nothing to verify this. Bytes must be numeric constants, while words are either numeric constants or symbols, such as the name of a function.
A new setting exists to omit the symbol names table, $OMIT_SYMBOL_TABLE. The symbol names table contains the names of all properties, attributes, fake actions, and arrays as strings, and is generally only used by debug library code and debug veneer error messages. When $OMIT_SYMBOL_TABLE=1 is set:
A new setting $ZCODE_MAX_INLINE_STRING has been added to determine how long a string can be and still be compiled to a @print opcode, rather than be added to the string segment and compiled to a @print_paddr opcode. This setting has a default value of 32, which matches the previous behaviour of the compiler, where this limit was hard-coded at 32 characters.
The Abbreviate directive now accepts abbreviations of any length.
The -u option, which computes abbreviations, can now generate abbreviations of any length.
Inform is now able to correctly set the plural flag on long dictionary words (e.g. 'directions//p'). However, due to the way Inform 7 has defined plural kind names in the past, fixing this will affect the parsing of Inform 7 games if the output Inform 6 code is then compiled with a version of Inform 6 that fixes this issue. As a result, there is a new setting $LONG_DICT_FLAG_BUG, which defaults to 1. The new behaviour is only enabled if this setting is set to 0.
Flags for dictionary words now include setting the NOUN flag, and also provides a way to explicitly not set a flag. The possible choices are:
The --trace PROPS and --trace SYMDEF options now display the line number that each property or symbol is defined at.
The --trace ASM=2 option now shows backpatch markers as a short descriptive string, rather than as a number.
The statement print "^"; now compiles to a single opcode (@new_line for Z-code, or @streamchar 10 for Glulx) rather than printing a one character string.
For Glulx, with strict mode turned off, print (char) X; compiles to either @streamchar X or @streamunichar X, depending on whether X is a compile-time constant less than 256, or not.
Grammar table lines entries which have no verb are now omitted. When this occurs a warning is printed, as this most likely indicates an oversight in the game's source code.
Error messages about invalid tokens are now more informative.
Inform now handles line breaks itself, rather than relying on the C standard library. This gives consistent handling of Windows and Unix style line breaks on all platforms.
The output file "gametext.txt" now includes the Z-code or Glulx version being compiled to.
The Z-Machine opcodes added in revision 1.1 of the Z-Machine Specification Standard, set_true_colour and buffer_screen, are now supported.
The Glulx build-in function random(x) built-in function now follows the DM4 specification: if x is positive, the function returns the result of 1+(@random x); if zero or negative, @setrandom x is called.
In several places (local variable declarations, action names and the Switches directive) the compiler would accept quoted strings and then ignore the quotes. This is now an error.
The case of a property having too many entries is now always an error, and is checked correctly in the case of compiling to Z-code V3.
An unclosed double quote at the end of a source file no longer causes the compiler to hang.
A number of issues relating to error messages giving incorrect information have been fixed by improving how the compiler handles backtracking through parsed symbols in some tricky cases.
The Z-code version of the veneer function Box__Routine (which is used in the implementation of the box statement) now contains a check to prevent calling the @set_cursor opcode with negative co-ordinates.
The veneer functions RA__Pr(), RL__Pr() and CP__Tab() are now correct for Z-code V3.
Errors in the declaration of arrays could sometimes cause the compiler to emit a large number of error messages, this is now fixed so that only the initial error is printed.
Invalid expressions like (a++ b), (a ~b), (a++ ~b), and (a++ --b) previously caused an internal compiler error, but now produce a sensible error message.
When computing abbreviations, the space character is now correctly treated as only occupying one byte, not four.
The argument supplied to the Z-machine opcode @jump is now interpreted correctly. Previously this was only done properly for the jump statement, not the opcode.
The dead code elimination logic now allows forward jumps to labels that would be otherwise unreachable, like this:
if (condition) {
jump MiddleLabel;
}
if (0) {
while (condition) {
...
.MiddleLabel;
...
}
}
The error that occurs when the compiler encounters a jump to an unknown label now includes the name of that label.
The various command line arguments that produce tracing or statistical information have been consolidated in a new argument: $!TRACEOPT or $!TRACEOPT=N. The Unix-style equivalent is --trace TRACEOPT or --trace TRACEOPT=N. You can also use $! by itself (or --helptrace) to list all available trace options. Trace option names are case-insensitive.
The optional N is always an integer. 0 always means off, 1 is the base level, 2 or more may increase verbosity. As a general rule, setting N to a high number is not an error; it just does the same thing as the highest supported number for that option. (This lets us add more verbosity levels to any option without worrying about compatibility errors.)
Four trace settings can be changed in mid-compile with the Trace directive. (This has always been true but now it is handled consistently.) Trace assembly, Trace expressions, Trace tokens and Trace linker are equivalent to --trace asm, --trace expr, --trace tokens and --trace linker, respectively. As with the command-line versions, you can optionally specify 0 to turn off that setting or 2 or more for more verbosity.
Four more trace directives do an immediate information dump: Trace dictionary, Trace objects, Trace symbols and Trace verbs. The command-line equivalents --trace dict, --trace objects, --trace symbols and --trace verbs do the same at the end of compilation.
The following single-letter options have been removed and replaced with trace options:
Some of the information that used to be under -a4 is now a separate trace setting, --trace BPATCH.
The -u option now just outputs computed abbreviations. If you want the verbose calculation information that it used to print, use --trace FINDABBREVS or --trace FINDABBREVS=2.
The compiler is now capable of dead code elimination, allowing it to:
Dead code elimination does mean that theoretically possible (though very odd) valid code will now result in a compilation error. For example,
if (DOSTUFF) {
while (condition) {
...
.MiddleLabel;
...
}
}
jump MiddleLabel; ! error
This will fail with a compilation error if DOSTUFF is defined as a constant zero. This optimization
can be turned off by setting the compiler setting $STRIP_UNREACHABLE_LABELS to zero (its default is one).
There are now further warnings if the compiler detects that the type used in certain expressions is not correct:
There is a new syntax for dynamic-string interpolations: "@(N)". The original syntax was "@NN", which is still supported. In the new syntax "N" can be a number, or an already defined numeric constant. As a result of this, under Glulx the limit on $MAX_DYNAMIC_STRINGS being less than 100 has been removed.
The constants #dictionary_table and #grammar_table are now available when compiling to Z-code, as well as Glulx.
The command line switch to display percentages (-p) now works with Glulx, and acts as an extension to the -z switch.
The double precision floating point related opcodes added to the Glulx 3.1.3 specification (that is, @numtod, @dtonumz, @dtonumn, @ftod, @dtof, @dceil, @dfloor, @dadd, @dsub, @dmul, @ddiv, @dmodr, @dmodq, @dsqrt, @dexp, @dlog, @dpow, @dsin, @dcos, @dtan, @dasin, @dacos, @datan, @datan2, @jdeq, @jdne, @jdlt, @jdle, @jdgt, @jdge, @jdisnan, @jdisinf) are now supported. In addition there are also two new macros @dload and @dstore, to load and store double precision floating point values.
The undo related opcodes added to the Glulx 3.1.3 specification (that is, @hasundo and @discardundo) are now supported.
The Version and Switches directives are now deprecated, and produce a warning if used. The recommended way to set the Z-code version and any other switches is to use the command line, or !% comments at the top of the source file.
The ability to compile Inform source code as modules, and then later link it, has been removed. This functionality always had significant restrictions and was never widely used. The -M and -U command line options have been removed, along with the +module_path option. The Link and Import directives now just produce errors.
The ability to use temporary files for internal storage (the -F switch and the +temporary_path option) have been removed.
The = sign in the Global directive is now optional, just as it is in the Constant directive. So all of these are now legal:
Constant c1 11; Constant c2 = 22; Global g1 33; Global g2 = 44;
The long deprecated forms of the Global directive that could define arrays have been removed. All these are now no longer legal:
Global array --> size; Global array data ...; Global array initial ...; Global array initstr ...;
Abbreviation directives now only add abbreviations to the output file if economy mode (-e) is set.
An array overrun when the $ZCODE_LESS_DICT_DATA setting is enabled has been fixed.
The logic that allows a test for the compiler version of the form #IfDef VN_1640 now requires what follows the "VN_" to be a number.