The problem of assembly language.

§1. General implementation. This section does just one thing: compiles invocations of assembly-language opcodes. 

void CAssembly::initialise(code_generator *gtr) {
    METHOD_ADD(gtr, INVOKE_OPCODE_MTID, CAssembly::invoke_opcode);
    METHOD_ADD(gtr, ASSEMBLY_MARKER_MTID, CAssembly::assembly_marker);
}

typedef struct C_generation_assembly_data {
    struct dictionary *opcodes_used;
} C_generation_assembly_data;

void CAssembly::initialise_data(code_generation *gen) {
    C_GEN_DATA(asmdata.opcodes_used) = NULL;
}

void CAssembly::begin(code_generation *gen) {
    CAssembly::initialise_data(gen);
}

void CAssembly::end(code_generation *gen) {
}

§2. Inter is for the most part fully specified and cross-platform, but assembly language is the big hole in that. It is legal for Inter code to contain almost anything which purports to be assembly language. For example, the following code will successfully build as part of an Inter kit: 

    [ Peculiar x;
        @bandersnatch x;
    ];

Kit code, and also material included in (- and -) brackets in I7 source text, can claim to use assembly language opcodes with any names it likes. No checking is done that these are "real" opcodes. (Spoilers: @bandersnatch is not.)

The point of this is that different final targets support different sets of assembly language. This was always true for Inform 6 code (after around 2000, anyway), because the Z and Glulx virtual machines had different assembly languages: @split_window exists for Z but not Glulx, @atan exists for Glulx but not Z, for example.

So each different final generator needs to make its own decision about what assembly language opcodes to provide, and what they will do. In theory, we could make an entirely new assembly language for C. But in practice that would just make the standard Inform kits, such as BasicInformKit, impossible to support on C, because those kits make quite heavy use of opcodes from Z/Glulx.

We will instead:

In this way, we obtain both compatibility with the Inform kits, enabling us to compile works of IF to C, and also extensibility.

§3. Each different opcode we see will be matched up to a C_supported_opcode giving it some metadata: we will gather these into a dictionary so that names of opcodes can quickly be resolved to their metadata structures.

That dictionary will begin with (1) about 60 standard supported opcodes, but then may pick up (2) a few others such as @bandersnatch, if kits do something non-standard.

So now we define some very minimal metadata on our opcodes. Each opcode will, when used, be followed by a number of operands, which we number from 1: 

    @fmod a b rem quot;
        1 2 3   4

This opcode, which performs floating-point division with remainder, reads in operands 1 and 2, and writes results out to operands 3 and 4. In the following, store_this_operand[3] and store_this_operand[4] would be TRUE, while store_this_operand[1] and store_this_operand[2] would be FALSE. (In fact, this is an outlier, because it is the only opcode we support which has more than one store operand. But in principle we could have many.)

Glulx assembly language also allows variable numbers of arguments to some opcodes, or "varargs". For example: 

    @glk 4 _vararg_count ret;
       1 2             3

Here operand 3 is a store, and operands 1 and 2 are read in. But operand 2 is special in that it is a count of additional operands which are found on the stack rather than in the body of the instruction. For example, 

    @glk 4 6 ret;

would provide @glk with seven operands to read in: the one in the instruction itself, 4, and then the top 6 items on the stack.

Because of this, an operand holding a variable-argument count is special. There can be at most one for any opcode; vararg_operand is -1 if there isn't one, but for @glk, vararg_operand would be 2. 

typedef struct C_supported_opcode {
    struct text_stream *name;  including the opening @ character
    int store_this_operand[MAX_OPERANDS_IN_INTER_ASSEMBLY];
    int vararg_operand;  position of _vararg_count operand, or -1 if none
    int speculative;  i.e., not part of the standard supported set
    CLASS_DEFINITION
} C_supported_opcode;

§4. On creation, a C_supported_opcode is automatically added to the dictionary: 

C_supported_opcode *CAssembly::new_opcode(code_generation *gen, text_stream *name,
    int s1, int s2, int va) {
    C_supported_opcode *opc = CREATE(C_supported_opcode);
    opc->speculative = FALSE;
    opc->name = Str::duplicate(name);
    for (int i=0; i<MAX_OPERANDS_IN_INTER_ASSEMBLY; i++) opc->store_this_operand[i] = FALSE;
    if (s1 >= 1) opc->store_this_operand[s1] = TRUE;
    if (s2 >= 1) opc->store_this_operand[s2] = TRUE;
    opc->vararg_operand = va;
    Dictionaries::create(C_GEN_DATA(asmdata.opcodes_used), name);
    Dictionaries::write_value(C_GEN_DATA(asmdata.opcodes_used), name, opc);
    return opc;
}

§5. When the generator encounters an opcode called name which seems to be used with operand_count operands, it calls the following function to find the corresponding metadata. Note that this always returns a valid C_supported_opcode, because even if a completely unexpected name is encountered, the above mechanism will just create a meaning for it. 

C_supported_opcode *CAssembly::find_opcode(code_generation *gen, text_stream *name,
    int operand_count) {
    if (C_GEN_DATA(asmdata.opcodes_used) == NULL) {
        C_GEN_DATA(asmdata.opcodes_used) = Dictionaries::new(256, FALSE);
        Stock with the basics5.1;
    }
    C_supported_opcode *opc;
    if (Dictionaries::find(C_GEN_DATA(asmdata.opcodes_used), name)) {
        opc = Dictionaries::read_value(C_GEN_DATA(asmdata.opcodes_used), name);
    } else {
        Add a speculative new opcode to the dictionary5.2;
    }
    return opc;
}

§5.1. Stock with the basics5.1 = 

    CAssembly::new_opcode(gen, I"@acos",             2, -1, -1);
    CAssembly::new_opcode(gen, I"@add",              3, -1, -1);
    CAssembly::new_opcode(gen, I"@aload",            3, -1, -1);
    CAssembly::new_opcode(gen, I"@aloadb",           3, -1, -1);
    CAssembly::new_opcode(gen, I"@aloads",           3, -1, -1);
    CAssembly::new_opcode(gen, I"@asin",             2, -1, -1);
    CAssembly::new_opcode(gen, I"@atan",             2, -1, -1);
    CAssembly::new_opcode(gen, I"@binarysearch",     8, -1, -1);
    CAssembly::new_opcode(gen, I"@call",             3, -1,  2);
    CAssembly::new_opcode(gen, I"@ceil",             2, -1, -1);
    CAssembly::new_opcode(gen, I"@copy",             2, -1, -1);
    CAssembly::new_opcode(gen, I"@cos",              2, -1, -1);
    CAssembly::new_opcode(gen, I"@div",              3, -1, -1);
    CAssembly::new_opcode(gen, I"@exp",              2, -1, -1);
    CAssembly::new_opcode(gen, I"@fadd",             3, -1, -1);
    CAssembly::new_opcode(gen, I"@fdiv",             3, -1, -1);
    CAssembly::new_opcode(gen, I"@floor",            2, -1, -1);
    CAssembly::new_opcode(gen, I"@fmod",             3,  4, -1);
    CAssembly::new_opcode(gen, I"@fmul",             3, -1, -1);
    CAssembly::new_opcode(gen, I"@fsub",             3, -1, -1);
    CAssembly::new_opcode(gen, I"@ftonumn",          2, -1, -1);
    CAssembly::new_opcode(gen, I"@ftonumz",          2, -1, -1);
    CAssembly::new_opcode(gen, I"@gestalt",          3, -1, -1);
    CAssembly::new_opcode(gen, I"@glk",              3, -1,  2);
    CAssembly::new_opcode(gen, I"@hasundo",          1, -1, -1);
    CAssembly::new_opcode(gen, I"@jeq",             -1, -1, -1);
    CAssembly::new_opcode(gen, I"@jfeq",            -1, -1, -1);
    CAssembly::new_opcode(gen, I"@jfge",            -1, -1, -1);
    CAssembly::new_opcode(gen, I"@jflt",            -1, -1, -1);
    CAssembly::new_opcode(gen, I"@jisinf",          -1, -1, -1);
    CAssembly::new_opcode(gen, I"@jisnan",          -1, -1, -1);
    CAssembly::new_opcode(gen, I"@jleu",            -1, -1, -1);
    CAssembly::new_opcode(gen, I"@jnz",             -1, -1, -1);
    CAssembly::new_opcode(gen, I"@jz",              -1, -1, -1);
    CAssembly::new_opcode(gen, I"@log",              2, -1, -1);
    CAssembly::new_opcode(gen, I"@malloc",          -1, -1, -1);
    CAssembly::new_opcode(gen, I"@mcopy",           -1, -1, -1);
    CAssembly::new_opcode(gen, I"@mzero",           -1, -1, -1);
    CAssembly::new_opcode(gen, I"@mfree",           -1, -1, -1);
    CAssembly::new_opcode(gen, I"@mod",              3, -1, -1);
    CAssembly::new_opcode(gen, I"@mul",              3, -1, -1);
    CAssembly::new_opcode(gen, I"@neg",              2, -1, -1);
    CAssembly::new_opcode(gen, I"@nop",             -1, -1, -1);
    CAssembly::new_opcode(gen, I"@numtof",           2, -1, -1);
    CAssembly::new_opcode(gen, I"@pow",              3, -1, -1);
    CAssembly::new_opcode(gen, I"@quit",            -1, -1, -1);
    CAssembly::new_opcode(gen, I"@random",           2, -1, -1);
    CAssembly::new_opcode(gen, I"@restart",         -1, -1, -1);
    CAssembly::new_opcode(gen, I"@restore",          2, -1, -1);
    CAssembly::new_opcode(gen, I"@restoreundo",      1, -1, -1);
    CAssembly::new_opcode(gen, I"@return",          -1, -1, -1);
    CAssembly::new_opcode(gen, I"@save",             2, -1, -1);
    CAssembly::new_opcode(gen, I"@saveundo",         1, -1, -1);
    CAssembly::new_opcode(gen, I"@setiosys",        -1, -1, -1);
    CAssembly::new_opcode(gen, I"@setrandom",       -1, -1, -1);
    CAssembly::new_opcode(gen, I"@shiftl",           3, -1, -1);
    CAssembly::new_opcode(gen, I"@sin",              2, -1, -1);
    CAssembly::new_opcode(gen, I"@sqrt",             2, -1, -1);
    CAssembly::new_opcode(gen, I"@streamchar",      -1, -1, -1);
    CAssembly::new_opcode(gen, I"@streamnum",       -1, -1, -1);
    CAssembly::new_opcode(gen, I"@streamunichar",   -1, -1, -1);
    CAssembly::new_opcode(gen, I"@sub",              3, -1, -1);
    CAssembly::new_opcode(gen, I"@tan",              2, -1, -1);
    CAssembly::new_opcode(gen, I"@ushiftr",          3, -1, -1);
    CAssembly::new_opcode(gen, I"@verify",           1, -1, -1);

§5.2. Speculative opcodes cannot store and cannot have varargs. Also, since they are not part of our supported set, there's no code here to implement them. Instead we predeclare a function and simply assume that the user will have written this function somewhere and linked it to us. For example, we might predeclare this: 

    void i7_opcode_bandersnatch(i7process_t *proc, i7word_t v1);

Add a speculative new opcode to the dictionary5.2 = 

    opc = CAssembly::new_opcode(gen, name, -1, -1, -1);
    opc->speculative = TRUE;
    segmentation_pos saved = CodeGen::select(gen, c_predeclarations_I7CGS);
    text_stream *OUT = CodeGen::current(gen);
    WRITE("void ");
    CNamespace::mangle_opcode(OUT, name);
    WRITE("(i7process_t *proc");
    for (int i=1; i<=operand_count; i++) WRITE(", i7word_t v%d", i);
    WRITE(");\n");
    CodeGen::deselect(gen, saved);

§6. We finally have enough infrastructure to invoke a general assembly-language instruction found in our Inter. 

void CAssembly::invoke_opcode(code_generator *gtr, code_generation *gen,
    text_stream *opcode, int operand_count, inter_tree_node **operands,
    inter_tree_node *label, int label_sense) {
    C_supported_opcode *opc = CAssembly::find_opcode(gen, opcode, operand_count);
    text_stream *OUT = CodeGen::current(gen);
    if (label_sense != NOT_APPLICABLE) Begin a branch instruction6.1;
    int push_store[MAX_OPERANDS_IN_INTER_ASSEMBLY];
    for (int i=0; i<MAX_OPERANDS_IN_INTER_ASSEMBLY; i++) push_store[i] = FALSE;
    Generate a function call6.3;
    if (label_sense != NOT_APPLICABLE) End a branch instruction6.2;
    Push any stored results which need to end up on the stack6.4;
    WRITE(";\n");
}

§6.1. Begin a branch instruction6.1 = 

    WRITE("if (");

§6.2. End a branch instruction6.2 = 

    if (label_sense == FALSE) WRITE(" == FALSE");
    WRITE(") goto ");
    if (label == NULL) internal_error("no branch label");
    Vanilla::node(gen, label);

§6.3. Each instruction becomes a function call to the function implementing the opcode in question, except that @return becomes the C statement return. If the opcode has N operands then the function has N+1 arguments, since the first is always the process pointer.

It may seem to compile slow code if we turn instructions into function calls, but

Generate a function call6.3 = 

    if (Str::eq(opcode, I"@return")) {
        WRITE("return (");
    } else {
        CNamespace::mangle_opcode(OUT, opcode); WRITE("(proc");
        if (operand_count > 0) WRITE(", ");
    }
    for (int operand = 1; operand <= operand_count; operand++) {
        if (operand > 1) WRITE(", ");
        TEMPORARY_TEXT(O)
        CodeGen::select_temporary(gen, O);
        Vanilla::node(gen, operands[operand-1]);
        CodeGen::deselect_temporary(gen);
        if (opc->store_this_operand[operand]) Generate a store operand6.3.2
        else Generate a regular operand6.3.1;
        DISCARD_TEXT(O)
    }
    WRITE(")");

§6.3.1. The argument for a regular operand will have type i7word_t, so we have to compile something of that type here.

The special operand notation sp is a pseudo-variable meaning "the top of the stack", so if we see that then we compile that to a pull: note that i7_pull returns an i7word_t.

Generate a regular operand6.3.1 = 

    if (Str::eq(O, I"sp")) {
        WRITE("i7_pull(proc)");
    } else {
        WRITE("%S", O);
    }

§6.3.2. The argument for a store operand will have type i7word_t *, so now we have to make a pointer.

Again, sp is a pseudo-variable meaning "the top of the stack", but this time we have to push, not pull, and that's something we can't do until the function has returned — the function will create the value we need to push. We get around this by compiling a pointer to some temporary memory.

Finally, assembly also allows 0 as a special value for a store operand, and this means "throw the value away". We don't want to incur a C compiler warning by attempting to write 0 in a pointer context, so we pass it as NULL instead.

Generate a store operand6.3.2 = 

    if (Str::eq(O, I"sp")) {
        WRITE("&(proc->state.tmp[%d])", operand);
        push_store[operand] = TRUE;
    } else if (Str::eq(O, I"0")) {
        WRITE("NULL");
    } else {
        WRITE("&%S", O);
    }

§6.4. That may leave a few stored results stranded in temporary workspace, so:

Push any stored results which need to end up on the stack6.4 = 

    for (int operand = 1; operand <= operand_count; operand++)
        if (push_store[operand])
            WRITE("; i7_push(proc, proc->state.tmp[%d])", operand);

§7. And where does the special operand sp come from? From here: 

void CAssembly::assembly_marker(code_generator *gtr, code_generation *gen, inter_ti marker) {
    text_stream *OUT = CodeGen::current(gen);
    switch (marker) {
        case ASM_SP_ASMMARKER: WRITE("sp"); break;
        default:
            WRITE_TO(STDERR, "Unsupported assembly marker is '%d'\n", marker);
            internal_error("unsupported assembly marker in C");
    }
}

§8. call. That does everything except to implement the standard set of opcodes, which must be done with about 60 functions in the C library.

This is not the place to specify what Glulx opcodes do. See Andrew Plotkin's documentation on the Glulx virtual machine.

Most of the opcodes we support are defined below, but see also C Input-Output Model for @glk, and see C Arithmetic for the plethora of mathematical operations such as @fmul.

To begin, here is a @call, which performs a function call to a perhaps computed address: 

void i7_opcode_call(i7process_t *proc, i7word_t fn_ref, i7word_t varargc, i7word_t *z);

void i7_opcode_call(i7process_t *proc, i7word_t fn_ref, i7word_t varargc, i7word_t *z) {
    i7word_t args[10]; for (int i=0; i<10; i++) args[i] = 0;
    for (int i=0; i<varargc; i++) args[i] = i7_pull(proc);
    i7word_t rv = i7_gen_call(proc, fn_ref, args, varargc);
    if (z) *z = rv;
}

§9. copy. Though it doesn't look it, this is one of the main ways Glulx assembly language programs push or pull to the stack — @copy sp x pulls the stack to x; @copy sp 0 pops the stack; @copy x sp pushes x to the stack. But all of that is handled by the general mechanism above. 

void i7_opcode_copy(i7process_t *proc, i7word_t x, i7word_t *y);

void i7_opcode_copy(i7process_t *proc, i7word_t x, i7word_t *y) {
    if (y) *y = x;
}

§10. aload, aloads, aloadb. 

void i7_opcode_aload(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z);
void i7_opcode_aloads(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z);
void i7_opcode_aloadb(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z);

void i7_opcode_aload(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z) {
    if (z) *z = i7_read_word(proc, x, y);
}

void i7_opcode_aloads(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z) {
    if (z) *z = i7_read_sword(proc, x, y);
}

void i7_opcode_aloadb(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z) {
    if (z) *z = i7_read_byte(proc, x+y);
}

§11. ushiftr, shiftl. 

void i7_opcode_shiftl(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z);
void i7_opcode_ushiftr(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z);

void i7_opcode_shiftl(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z) {
    i7word_t value = 0;
    if ((y >= 0) && (y < 32)) value = (x << y);
    if (z) *z = value;
}

void i7_opcode_ushiftr(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z) {
    i7word_t value = 0;
    if ((y >= 0) && (y < 32)) value = (x >> y);
    if (z) *z = value;
}

§12. jeq, jleu, jnz, jz. These are branch opcodes and return an int.

The implementation of @jleu here is modelled on the one from dumb-glulxe. Writing code like *((i7word_t *) &ux) = x really makes you proud to have chosen C as your programming language, but it works. 

int i7_opcode_jeq(i7process_t *proc, i7word_t x, i7word_t y);
int i7_opcode_jleu(i7process_t *proc, i7word_t x, i7word_t y);
int i7_opcode_jnz(i7process_t *proc, i7word_t x);
int i7_opcode_jz(i7process_t *proc, i7word_t x);

int i7_opcode_jeq(i7process_t *proc, i7word_t x, i7word_t y) {
    if (x == y) return 1;
    return 0;
}

int i7_opcode_jleu(i7process_t *proc, i7word_t x, i7word_t y) {
    unsigned_i7word_t ux, uy;
    *((i7word_t *) &ux) = x; *((i7word_t *) &uy) = y;
    if (ux <= uy) return 1;
    return 0;
}

int i7_opcode_jnz(i7process_t *proc, i7word_t x) {
    if (x != 0) return 1;
    return 0;
}

int i7_opcode_jz(i7process_t *proc, i7word_t x) {
    if (x == 0) return 1;
    return 0;
}

§13. nop, quit, verify. There is no real meaning for @verify in this situation: it's supposed to check the checksum for the contents of a virtual machine, to protect against the (entirely likely) scenario of a floppy disk sector going bad in 1983. So we unconditionally store the "okay" result. 

void i7_opcode_nop(i7process_t *proc);
void i7_opcode_quit(i7process_t *proc);
void i7_opcode_verify(i7process_t *proc, i7word_t *z);

void i7_opcode_nop(i7process_t *proc) {
}

void i7_opcode_quit(i7process_t *proc) {
    i7_fatal_exit(proc);
}

void i7_opcode_verify(i7process_t *proc, i7word_t *z) {
    if (z) *z = 0;
}

§14. restoreundo, saveundo, hasundo, discardundo. This all works, but we do something pretty inelegant to support @restoreundo: we insert a call to a (presumably kit-based) function called DealWithUndo, provided this exists. This is done because we are unable safely to follow the proper Glulx specification. In principle, after a @restoreundo succeeds, execution immediately continues from the position in the program where the @saveundo occurred. For a while the implementation here imitated this by using longjmp and setjmp, but it all proved very fragile because of the difficulty of storing setjmp positions safely in memory.

Correspondingly, our implementation of @saveundo always stores the result value 0. The result value 1 would indicate that execution had switched there from a successful @restoreundo: but, as noted, that never happens. 

void i7_opcode_restoreundo(i7process_t *proc, i7word_t *x);
void i7_opcode_saveundo(i7process_t *proc, i7word_t *x);
void i7_opcode_hasundo(i7process_t *proc, i7word_t *x);
void i7_opcode_discardundo(i7process_t *proc);

#ifdef i7_mgl_DealWithUndo
i7word_t i7_fn_DealWithUndo(i7process_t *proc);
#endif

void i7_opcode_restoreundo(i7process_t *proc, i7word_t *x) {
    if (i7_has_snapshot(proc)) {
        i7_restore_snapshot(proc);
        if (x) *x = 0;
        #ifdef i7_mgl_DealWithUndo
        i7_fn_DealWithUndo(proc);
        #endif
    } else {
        if (x) *x = 1;
    }
}

void i7_opcode_saveundo(i7process_t *proc, i7word_t *x) {
    i7_save_snapshot(proc);
    if (x) *x = 0;
}

void i7_opcode_hasundo(i7process_t *proc, i7word_t *x) {
    i7word_t rv = 0; if (i7_has_snapshot(proc)) rv = 1;
    if (x) *x = rv;
}

void i7_opcode_discardundo(i7process_t *proc) {
    i7_destroy_latest_snapshot(proc);
}

§15. restart, restore, save. For the moment, at least, we intentionally do not implement these. It seems likely that anyone using C to run interactive fiction is doing so in a wider framework where saved states will work differently from the traditional model of asking the user for a filename and then saving data out to a binary file of that name in the current working directory. Better to do nothing here, and let users handle this themselves.

Similar considerations apply to @restart. The intention of this opcode is essentially to reboot the virtual machine and start over: here, though, we have a real machine. It's easy enough to reinitialise the process state, but not so simple to restart execution as if from a clean process start. 

void i7_opcode_restart(i7process_t *proc);
void i7_opcode_restore(i7process_t *proc, i7word_t x, i7word_t *y);
void i7_opcode_save(i7process_t *proc, i7word_t x, i7word_t *y);

void i7_opcode_restart(i7process_t *proc) {
    printf("(RESTART is not implemented on this C program.)\n");
}

void i7_opcode_restore(i7process_t *proc, i7word_t x, i7word_t *y) {
    printf("(RESTORE is not implemented on this C program.)\n");
}

void i7_opcode_save(i7process_t *proc, i7word_t x, i7word_t *y) {
    printf("(SAVE is not implemented on this C program.)\n");
}

§16. streamchar, streamnum, streamunichar. 

void i7_opcode_streamnum(i7process_t *proc, i7word_t x);
void i7_opcode_streamchar(i7process_t *proc, i7word_t x);
void i7_opcode_streamunichar(i7process_t *proc, i7word_t x);

void i7_opcode_streamnum(i7process_t *proc, i7word_t x) {
    i7_print_decimal(proc, x);
}

void i7_opcode_streamchar(i7process_t *proc, i7word_t x) {
    i7_print_char(proc, x);
}

void i7_opcode_streamunichar(i7process_t *proc, i7word_t x) {
    i7_print_char(proc, x);
}

§17. binarysearch. This is a Grand Imperial Hotel among Glulx opcodes, with 8 operands, only the last of which is a store. It performs a binary search on a block of structures known to be sorted already. It has a nice general-purpose look but was devised so that command verbs could be looked up quickly in dictionary tables when interactive fiction is being played: that's the only use which the standard Inform kits make of it.

The elegant implementation here comes from Andrew Plotkin's reference code for glulxe, a Glulx interpreter. options is a bitmap of the bits defined below. In the only use the standard Inform kits make of this opcode, options will be just serop_KeyIndirect, but keysize will be more than 4, so that the elaborate speed optimisation for keys of size 1, 2 and 4, and thus keybuf, are never used. But we may as well have the full functionality here. 

#define serop_KeyIndirect        1
#define serop_ZeroKeyTerminates  2
#define serop_ReturnIndex        4
void i7_opcode_binarysearch(i7process_t *proc, i7word_t key, i7word_t keysize,
    i7word_t start, i7word_t structsize, i7word_t numstructs, i7word_t keyoffset,
    i7word_t options, i7word_t *s1);

void i7_opcode_binarysearch(i7process_t *proc, i7word_t key, i7word_t keysize,
    i7word_t start, i7word_t structsize, i7word_t numstructs, i7word_t keyoffset,
    i7word_t options, i7word_t *s1) {

    if (s1 == NULL) return; /* Do not spend any time if the result is to be ignored */


    /* If the key size is 4 or fewer, copy it directly into the keybuf array */
    unsigned char keybuf[4];
    if (options & serop_KeyIndirect) {
        if (keysize <= 4)
            for (int ix=0; ix<keysize; ix++)
                keybuf[ix] = i7_read_byte(proc, key + ix);
    } else {
        switch (keysize) {
            case 4:
                keybuf[0] = I7BYTE_0(key); keybuf[1] = I7BYTE_1(key);
                keybuf[2] = I7BYTE_2(key); keybuf[3] = I7BYTE_3(key); break;
            case 2:
                keybuf[0] = I7BYTE_0(key); keybuf[1] = I7BYTE_1(key); break;
            case 1:
                keybuf[0] = key; break;
        }
    }

    i7word_t bot = 0, top = numstructs; /* Initial search range, including bot but not top */
    while (bot < top) { /* I.e., while the search range is not empty */
        /* Find the structure at the midpoint of the search range */
        i7word_t val = (top+bot) / 2;
        i7word_t addr = start + val * structsize;

        /* Compute cmp = 0 if the key matches this, -1 if it precedes, 1 if it follows */
        int cmp = 0;
        if (keysize <= 4) {
            for (int ix=0; (!cmp) && ix<keysize; ix++) {
                unsigned char byte = i7_read_byte(proc, addr + keyoffset + ix);
                unsigned char byte2 = keybuf[ix];
                if (byte < byte2) cmp = -1;
                else if (byte > byte2) cmp = 1;
            }
        } else {
            for (int ix=0; (!cmp) && ix<keysize; ix++) {
                unsigned char byte  = i7_read_byte(proc, addr + keyoffset + ix);
                unsigned char byte2 = i7_read_byte(proc, key + ix);
                if (byte < byte2) cmp = -1;
                else if (byte > byte2) cmp = 1;
            }
        }

        if (cmp == 0) {
            /* Success! */
            if (options & serop_ReturnIndex) *s1 = val; else *s1 = addr;
            return;
        }

        if (cmp < 0) bot = val+1; /* Chop search range to the second half */
        else top = val; /* Chop search range to the first half */
    }

    /* Failure! */
    if (options & serop_ReturnIndex) *s1 = -1; else *s1 = 0;
}

§18. mcopy, mzero, malloc, mfree. A Glulx assembly opcode is provided for fast memory copies, which we must implement. We're choosing not to implement the Glulx @malloc or @mfree opcodes for now, but that will surely need to change in due course. (When that does change, we will need also to change @gestalt.) 

void i7_opcode_mcopy(i7process_t *proc, i7word_t x, i7word_t y, i7word_t z);
void i7_opcode_mzero(i7process_t *proc, i7word_t x, i7word_t y);
void i7_opcode_malloc(i7process_t *proc, i7word_t x, i7word_t y);
void i7_opcode_mfree(i7process_t *proc, i7word_t x);

void i7_opcode_mcopy(i7process_t *proc, i7word_t x, i7word_t y, i7word_t z) {
    if (z < y)
        for (i7word_t i=0; i<x; i++)
            i7_write_byte(proc, z+i, i7_read_byte(proc, y+i));
    else
        for (i7word_t i=x-1; i>=0; i--)
            i7_write_byte(proc, z+i, i7_read_byte(proc, y+i));
}

void i7_opcode_mzero(i7process_t *proc, i7word_t x, i7word_t y) {
    for (i7word_t i=0; i<x; i++) i7_write_byte(proc, y+i, 0);
}

void i7_opcode_malloc(i7process_t *proc, i7word_t x, i7word_t y) {
    printf("Unimplemented: i7_opcode_malloc.\n");
    i7_fatal_exit(proc);
}

void i7_opcode_mfree(i7process_t *proc, i7word_t x) {
    printf("Unimplemented: i7_opcode_mfree.\n");
    i7_fatal_exit(proc);
}

§19. random, setrandom. Note that the random(...) function built in to Inform is just a name for the @random opcode, so we define that here too.

We have no convincing need for a statistically good random number algorithm, but we do want cross-platform consistency in order that the test suite for Inform should behave equivalently on MacOS, Linux and Windows — at least when the generator is seeded with the same value. To that end, we borrow the algorithm used by the frotz Z-machine interpreter, which in turn is based on suggestions in the Z-machine standards document. 

i7rngseed_t i7_initial_rng_seed(void);
void i7_opcode_random(i7process_t *proc, i7word_t x, i7word_t *y);
void i7_opcode_setrandom(i7process_t *proc, i7word_t s);
i7word_t i7_random(i7process_t *proc, i7word_t x);

i7rngseed_t i7_initial_rng_seed(void) {
    i7rngseed_t seed;
    seed.A = 1;
    seed.interval = 0;
    seed.counter = 0;
    return seed;
}

void i7_opcode_random(i7process_t *proc, i7word_t x, i7word_t *y) {
    uint32_t rawvalue = 0;
    if (proc->state.seed.interval != 0) {
        rawvalue = proc->state.seed.counter++;
        if (proc->state.seed.counter == proc->state.seed.interval) proc->state.seed.counter = 0;
    } else {
        proc->state.seed.A = 0x015a4e35L * proc->state.seed.A + 1;
        rawvalue = (proc->state.seed.A >> 16) & 0x7fff;
    }
    uint32_t value;
    if (x == 0) value = rawvalue;
    else if (x >= 1) value = rawvalue % (uint32_t) (x);
    else value = -(rawvalue % (uint32_t) (-x));
    *y = (i7word_t) value;
}

void i7_opcode_setrandom(i7process_t *proc, i7word_t s) {
    if (s == 0) {
        proc->state.seed.A = (uint32_t) time(NULL);
        proc->state.seed.interval = 0;
    } else if (s < 1000) {
        proc->state.seed.interval = s;
        proc->state.seed.counter = 0;
    } else {
        proc->state.seed.A = s;
        proc->state.seed.interval = 0;
    }
}

i7word_t i7_random(i7process_t *proc, i7word_t x) {
    i7word_t r;
    i7_opcode_random(proc, x, &r);
    return r+1;
}

§20. setiosys. This opcode in principle allows a story file to select the input-output system it will use. But the Inform kits only use system 2, called Glk, and this is the only system we support, so we will simply ignore this. 

void i7_opcode_setiosys(i7process_t *proc, i7word_t x, i7word_t y);

void i7_opcode_setiosys(i7process_t *proc, i7word_t x, i7word_t y) {
}

§21. gestalt. This opcode allows a story file to ask the Glulx interpreter running it whether or not the interpreter can perform certain tasks. 

void i7_opcode_gestalt(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z);

void i7_opcode_gestalt(i7process_t *proc, i7word_t x, i7word_t y, i7word_t *z) {
    int r = 0;
    switch (x) {
        case 0: r = 0x00030103; break; /* Say that the Glulx version is v3.1.3 */
        case 1: r = 1;          break; /* Say that the interpreter version is 1 */
        case 2: r = 0;          break; /* We do not (yet) support @setmemsize */
        case 3: r = 1;          break; /* We do support UNDO */
        case 4: if (y == 2) r = 1;     /* We do support Glk */
                       else r = 0;     /* But not any other I/O system */
                break;
        case 5: r = 1;          break; /* We do support Unicode operations */
        case 6: r = 1;          break; /* We do support @mzero and @mcopy */
        case 7: r = 0;          break; /* We do not (yet) support @malloc or @mfree */
        case 8: r = 0;          break; /* Since we do not support @malloc */
        case 9: r = 0;          break; /* We do not support @accelfunc pr @accelparam */
        case 10: r = 0;         break; /* And therefore provide none of their accelerants */
        case 11: r = 1;         break; /* We do support floating-point maths operations */
        case 12: r = 1;         break; /* We do support @hasundo and @discardundo */
    }
    if (z) *z = r;
}