Skip to content

Latest commit

 

History

History
834 lines (644 loc) · 28.1 KB

Borland C++ decompilation.md

File metadata and controls

834 lines (644 loc) · 28.1 KB

Local variables

DX First 8-bit variable declared if no other function is called
Second 16-bit variable declared if no other function is called
[bp-1] First 8-bit variable declared otherwise
SI First 16-bit variable declared
DI Second 16-bit variable declared if other functions are called

Example:

ASM Declaration sequence in C
SI int near *var_1;
[bp-1] char var_2;
[bp-2] char var_3;
  • Local enum variables with underlying 1-byte types are always word-aligned, regardless of the value of -a.

Grouping

Any structures or classes that contain more than a single scalar-type member are grouped according to their declaration order, and placed after (that is, further away from BP) than all scalar-type variables. This means that it's not possible to bundle a set of variables with the same meaning into a structure (e.g. pointers to all 4 VRAM planes) if a scalar-type variable is placed inbetween two of these structure instances on the stack: Those structure instances would be grouped and always placed next to each other, no matter where the scalar-type variable is declared in relation to them.

Signedness

MOV al, var
MOV ah, 0
var is unsigned char
MOV al, var
CBW
var is char, AX is int

Integer arithmetic

ADD r8, imm8 For non-AL registers, imm8 must be static_cast to uint8_t. Otherwise, the addition is done on AL and then MOVed to r8.
ADD [m8], imm8 Only achievable through a C++ method operating on a member?
MOV AX, [m16]
ADD AL, [m8]
Same – [m16] must be returned from an inlined function to avoid the optimization of it being directly shortened to 8 bits.
MOV AL, [m8]
ADD AL, imm8
MOV [m8], AL
Opposite; not an inlined function
CWD
SUB AX, DX
SAR AX, 1
AX / 2, AX is int
MOV [new_var], AX
CWD
XOR AX, DX
SUB AX, DX
abs(AX), defined in <stdlib.h>. AX is int
  • When bit-testing a variable with a 16-bit mask via & in a conditional expression, the TEST is optimized to cover just the high or low byte, if possible:
    long v = 0xFFFFFFFF; // Works regardless of size or signedness
    char       b00 = (v & 0x00000001) != 0; // TEST  BYTE PTR [v + 0], 1
    char       b08 = (v & 0x00000100) != 0; // TEST  BYTE PTR [v + 1], 1
    char       b16 = (v & 0x00010000) != 0; // TEST DWORD PTR [v + 0], 0x00010000
    char       b24 = (v & 0x01000000) != 0; // TEST DWORD PTR [v + 0], 0x01000000
    char b00_to_15 = (v & 0x0000FFFF) != 0; // TEST  WORD PTR [v + 0], 0xFFFF
    char b16_to_31 = (v & 0xFFFF0000) != 0; // TEST DWORD PTR [v + 0], 0xFFFF0000
    char b08_to_23 = (v & 0x00FFFF00) != 0; // TEST DWORD PTR [v + 0], 0x00FFFF00

Arithmetic on a register after assigning it to a variable?

Assignment is part of the C expression. If it's a comparison, that comparison must be spelled out to silence the Possibly incorrect assignment warning.

CALL somefunc
MOV ??, AX
OR AX, AX
JNZ ↑
while(( ?? = somefunc() ) != NULL)

SUB ??, imm vs. ADD ??, -imm

SUB means that ?? is unsigned. Might require suffixing imm with u in case it's part of an arithmetic expression that was promoted to int.

Comparisons

  • Any comparison of a register with a literal 0 is optimized to OR reg, reg followed by a conditional jump, no matter how many calculations and inlined functions are involved. Any CMP reg, 0 instructions must have either come from assembly, or referred to a pointer at address 0:

    extern void near *address_0; // Public symbol at near address 0
    register int i;
    
    if(i != reinterpret_cast<int>(address_0)) {
      // ↑ Will emit `CMP reg, 0`
    }
  • CMP instructions not followed by jumps correspond to empty if statements:

    if(foo > 100) { // CMP foo, 100
    }
    bar = 8;        // MOV bar, 8

Pointer arithmetic

  • Using parentheses or subscripts in an offset calculation implies a certain order of operations, which can greatly impact the generated code:

    char far *plane = reinterpret_cast<char __seg *>(0xA800);
    int y = (17 * (640 / 8)); // MOV DX, 1360
    int x = 4;                // MOV CX, 4
    
    // LES BX, [plane]
    // ADD BX, DX
    // ADD BX, CX
    // MOV AL, ES:[BX]
    _AL = *(plane + y + x);
    _AL = *(y + plane + x);
    
    // LES BX, [plane]
    // ADD BX, CX      ; CX and DX swapped, compared to the one above
    // ADD BX, DX
    // MOV AL, ES:[BX]
    _AL = *(y + (plane + x));
    
    // MOV BX, DX
    // ADD BX, CX
    // MOV ES, WORD PTR [plane + 2]
    // ADD BX, WORD PTR [plane]
    // MOV AL, ES:[BX]
    _AL = *(plane + (y + x));
    _AL = plane[y + x];

Floating-point arithmetic

  • Since the x87 FPU can only load from memory, all temporary results of arithmetic are spilled to one single compiler-generated variable (fpu_tmp) on the stack, which is reused across all of the function:

    MOV  AX, myint
    INC  AX
    MOV  fpu_tmp, ax
    FILD fpu_tmp
    FSTP ret
    float ret = (myint + 1)
  • The same fpu_tmp variable is also used as the destination for FNSTSW, used in comparisons.

  • On the stack, fpu_tmp is placed after all variables declared at the beginning of the function.

  • Performing arithmetic or comparisons between float and double variables always FLDs the float first, before emitting the corresponding FPU instruction for the double, regardless of how the variables are placed in the expression. The instruction order only matches the expression order for literals:

    char ret;
    float f;
    double d;
    
    ret = (f > d);     // FLD f,     FCOMP d
    ret = (d > f);     // FLD f,     FCOMP d
    
    ret = (d > 3.14f); // FLD d,     FCOMP 3.14f
    ret = (3.14f > d); // FLD 3.14f, FCOMP d
    ret = (f > 3.14);  // FLD f,     FCOMP 3.14 + 4
    ret = (3.14 > f);  // FLD 3.14,  FCOMP f + 4

Assignments

MOV ???, [SI+????] Only achievable through pointer arithmetic?
  • When assigning to an array element at a variable or non-0 index, the array element address is typically evaluated before the expression to be assigned. But when assigning

    • the result of any arithmetic expression of a 16-bit type
    • to an element of a far array of a 16-bit type,

    the expression will be evaluated first, if its signedness differs from that of the array:

    int far *s;
    unsigned int far *u;
    int s1, s2;
    unsigned int u1, u2;
    
    s[1] = (s1 | s2); // LES BX, [s]; MOV AX, s1; OR AX, s2; MOV ES:[BX+2], AX
    s[1] = (s1 | u2); // MOV AX, s1; OR AX, u2; LES BX, [s]; MOV ES:[BX+2], AX
    s[1] = (u1 | u2); // MOV AX, u1; OR AX, u2; LES BX, [s]; MOV ES:[BX+2], AX
    
    u[1] = (s1 | s2); // MOV AX, s1; OR AX, s2; LES BX, [u]; MOV ES:[BX+2], AX
    u[1] = (s1 | u2); // LES BX, [u]; MOV AX, s1; OR AX, u2; MOV ES:[BX+2], AX
    u[1] = (u1 | u2); // LES BX, [u]; MOV AX, u1; OR AX, u2; MOV ES:[BX+2], AX
  • Assigning AX to multiple variables in a row also indicates multiple assignment in C:

    // Applies to all storage durations
    int a, b, c;
    
    a = 0;  // MOV [a], 0
    b = 0;  // MOV [b], 0
    c = 0;  // MOV [c], 0
    
    a = b = c = 0; // XOR AX, AX;  MOV [c], AX; MOV [b], AX;  MOV [a], AX;
                   // Note the opposite order of variables!
  • For trivially copyable structures, copy assignments are optimized to an equivalent of memcpy():

    Structure size (no flags) -G
    1 via AL via AL
    2 via AX via AX
    3 SCOPY@ via AX and AL
    4 via DX:AX via DX:AX
    5, 7, 9 SCOPY@ via AX and AL
    6, 8 SCOPY@ via AX
    10, 12, 14, … SCOPY@ REP MOVSW
    11, 13, 15, … SCOPY@ REP MOVSW, MOVSB

    (With the -3 flag, EAX is used instead of DX:AX in the 4-byte case, but everything else stays the same.)

    Breaking triviality by overloading operator = in any of the structure members also breaks this optimization. In some cases, it might be possible to recreate it, by simulating triviality in an overloaded copy assignment operator inside the class in question:

    struct Nontrivial {
      nontrivial_char_t e[100];
      // Functions containing local classes aren't expanded inline, so...
      struct Trivial {
        char e[100];
      };
    
      void operator =(const Nontrivial &other) {
        reinterpret_cast<Trivial &>(*this) = (
            reinterpret_cast<const Trivial &>(other)
        );
      }
    };

    However, this only generates identical code to the original optimization if passing the other parameter can be inlined, which isn't always the case.

Function pointers

Type syntax (cf. platform.h):

… near function … far function
Near pointer to… int (near *near nn_t)() int (far *near fn_t)()
Far pointer to… int (near *far  nf_t)() int (far *far  ff_t)()

Calling conventions can be added before the *.

switch statements

  • Sequence of the individual cases is identical in both C and ASM

  • Multiple cases with the same offset in the table, to code that doesn't return? Code was compiled with -O

  • Having no more than 3 cases (4 with an additional default) generates comparison/branching code instead of a jump table. The comparisons will be sorted in ascending order of the case values, while the individual branch bodies still match their order given in the code:

    switch(foo) {            // MOV AX, foo
    default: foo = 0; break; // CMP AX, 10; JZ @@case_10
    case 30: foo = 3; break; // CMP AX, 20; JZ @@case_20
    case 10: foo = 1; break; // CMP AX, 30; JZ @@case_30
    case 20: foo = 2; break; // MOV foo, 0
    }                        // JMP @@after_switch
                             // @@case_30: MOV foo, 3; JMP @@after_switch
                             // @@case_10: MOV foo, 1; JMP @@after_switch
                             // @@case_20: MOV foo, 2;
                             // @@after_switch:
  • With the -G (Generate for speed) option, complicated switch statements that require both value and jump tables are compiled to a binary search with regular conditional branches:

    switch(foo) {
    case 0x4B: /* […] */ break;
    case 0x4D: /* […] */ break;
    case 0x11: /* […] */ break;
    case 0x1F: /* […] */ break;
    case 0x20: /* […] */ break;
    case 0x17: /* […] */ break;
    case 0x26: /* […] */ break;
    case 0x19: /* […] */ break;
    case 0x01: /* […] */ break;
    case 0x1C: /* […] */ break;
    }

    Resulting ASM:

    @@switch:
      MOV AX, foo
      CMP AX, 1Fh
      JZ  @@case_1Fh
      JG  @@GT_1Fh
      CMP AX, 17h
      JZ  @@case_17h
      JG  @@GT_17h_LT_1Fh
      CMP AX, 01h
      JZ  @@case_01h
      CMP AX, 11h
      JZ  @@case_11h
      JMP @@no_case_found
    
    @@GT_17h_LT_1Fh:
      CMP AX, 1Ch
      JZ  @@case_1Ch
      JMP @@no_case_found
    
    @@GT_1Fh:
      CMP AX, 4Bh
      JZ  @@case_4Bh
      JG  @@GT_4Bh
      CMP AX, 20h
      JZ  @@case_
      CMP AX, 26h
      JZ  @@case_26h
      JMP @@no_case_found
    
    @@GT_4Bh:
      CMP AX, 4Dh
      JZ  @@case_4Dh
      JMP @@no_case_found

Function calls

NOP insertion

Happens for every far call to outside of the current translation unit, even if both the caller and callee end up being linked into the same code segment.

Certainty: Seems like there might be a way around that, apart from temporarily spelling out these calls in ASM until both functions are compiled as part of the same translation unit. Found nothing so far, though.

Pushing byte arguments to functions

Borland C++ just pushes the entire word. Will cause IDA to mis-identify certain local variables as words when they aren't.

Pushing pointers

Passing far pointers to subscripted array elements requires code to calculate the offset. Turbo C++ emits this calculation (and not the PUSH itself) either before or after the segment is PUSHed. If either

  1. the pointer is near, or
  2. the parameter is a near or far const *,

the segment argument is always pushed immediately, before evaluating the offset:

#pragma option -ml

struct s100 {
  char c[100];
};

extern s100 structs[5];

void __cdecl process_mut(s100 *element);
void __cdecl process_const(const s100 *element);

void foo(int i) {
  process_mut((s100 near *)(&structs[i]));   // PUSH DS; (AX = offset); PUSH AX;
  process_mut((s100 far *)(&structs[i]));    // (AX = offset); PUSH DS; PUSH AX;
  process_const((s100 near *)(&structs[i])); // PUSH DS; (AX = offset); PUSH AX;
  process_const((s100 far *)(&structs[i]));  // PUSH DS; (AX = offset); PUSH AX;
}

Flags

-G (Generate for speed)

  • Replaces

    ENTER <stack size>, 0

    with

    PUSH BP
    MOV  BP, SP
    SUB  SP, <stack size>

-Z (Suppress register reloads)

  • The tracked contents of ES are reset after a conditional statement. If the original code had more LES instructions than necessary, this indicates a specific layout of conditional branches:

    struct foo {
      char a, b;
    
      char es_not_reset();
      char es_reset();
    };
    
    char foo::es_not_reset() {
      return (
        a    // LES BX, [bp+this]
        && b // `this` still remembered in ES, not reloaded
      );
    }
    
    char foo::es_reset() {
      if(a) return 1; // LES BX, [bp+this]
      // Tracked contents of ES are reset
      if(b) return 1; // LES BX, [bp+this]
      return 0;
    }

    This also applies to divisors stored in BX.

-3 (80386 Instructions)

  • 32-bit function return values are stored in DX:AX even with this option enabled. Assigning such a returned value generates different instructions if the signedness of the return type differs from the signedness of the target variable:

    /*    */ long ret_signed(void)   { return 0x12345678; }
    unsigned long ret_unsigned(void) { return 0x12345678; }
    
    void foo(void) {
      long s;
      unsigned long u;
    
      s = ret_signed();   // MOV WORD PTR [s+2], DX; MOV WORD PTR [s+0], AX;
      s = ret_unsigned(); // PUSH DX; PUSH AX; POP EAX; MOV DWORD PTR [s], EAX;
      u = ret_signed();   // PUSH DX; PUSH AX; POP EAX; MOV DWORD PTR [u], EAX;
      u = ret_unsigned(); // MOV WORD PTR [u+2], DX; MOV WORD PTR [u+0], AX;
    }

    Without -3, the two-MOV WORD PTR variant is generated in all four cases.

-3 (80386 Instructions) + -Z (Suppress register reloads)

Bundles two consecutive 16-bit function parameters into a single 32-bit one, passed via a single 32-bit PUSH. Currently confirmed to happen for literals and structure members whose memory layout matches the parameter list and calling convention. Signedness doesn't matter.

Won't happen for two consecutive 8-bit parameters, and can be circumvented by casting a near pointer to a 16-bit integer and back.

// Works for all storage durations
struct {          int  x, y; } p;
struct { unsigned int  x, y; } q;

void __cdecl foo_c(char x, char y);
void __cdecl foo_s(int x, int y);
void __cdecl foo_u(unsigned int x, unsigned int y);

foo_s(640, 400); // PUSH LARGE 1900280h
foo_u(640, 400); // PUSH LARGE 1900280h
foo_s(p.x, p.y); // PUSH LARGE [p]
foo_u(p.x, p.y); // PUSH LARGE [p]
foo_s(q.x, q.y); // PUSH LARGE [p]
foo_u(q.x, q.y); // PUSH LARGE [p]
foo_c(100, 200); // PUSH 200; PUSH 100

// PUSH [p.x]; PUSH [p.y];
foo_u(*reinterpret_cast<int near *>(reinterpret_cast<unsigned int>(&p.x)), p.y);
foo_s(*reinterpret_cast<int near *>(reinterpret_cast<unsigned int>(&p.x)), p.y);

-O (Optimize jumps)

Inhibited by:

  • identical variable declarations within more than one scope – the optimizer will only merge the code after the last ASM reference to that declared variable. Yes, even though the emitted ASM would be identical:

    if(a) {
        int v = set_v();
        do_something_else();
        use(v);
    } else if(b) {
        // Second declaration of [v]. Even though it's assigned to the same stack
        // offset, the second `PUSH c` call will still be emitted separately.
        // Thus, jump optimization only reuses the `CALL use` instruction.
        // Move the `int v;` declaration to the beginning of the function to
        // avoid this.
        int v = set_v();
        use(v);
    }
  • distinct instances of assignments of local variables in registers to itself

  • inlined calls to empty functions

-O also merges the ADD SP, imm8 or POP CX stack-clearing instructions after successive __cdecl function calls into a single one with their combined parameter size after the final function call in such a series. Declaring a local variable after a function call, with or without assigning a value, will interrupt such a series and force a stack-clearing instruction after the final function call before the declaration.

  • Bug: Any emitted call to SCOPY@ will disable this feature of -O for all generated code in a translation unit that follows the SCOPY@ call.

    This can explain why a function might seem impossible to decompile with the wrong translation unit layout. If it

    • doesn't contain the stack-clearing optimization,
    • but does definitely contain optimized jumps,
    • which couldn't be reproduced with the slight jump optimization provided by -O- -y,

    the translation unit is simply missing a SCOPY@ before the function in question.

-y (Produce line number info)

Provides its own kind of slight jump optimization if combined with -O-. Yes, seriously. Might be required to decompile code that seems to contain both some of the jump optimizations from -O and the stack-clearing instructions after every function call from -O-.

Initialization

Any initialization of a variable with static storage duration (even a const one) that involves function calls (even those that would regularly inline) will emit a #pragma startup function to perform that initialization at runtime. This extends to C++ constructors, making macros the only way to initialize such variables with arithmetic expressions at compile time.

#define FOO(x) (x << 1)

inline char foo(const char x)  {
    return FOO(x);
}

const char static_storage[3] = { FOO(1), foo(2), FOO(3) };

Resulting ASM (abbreviated):

  .data
static_storage  db  2, 0, 6

  .code
@_STCON_$qv	proc	near
	push bp
	mov  bp, sp
	mov  static_storage[1], 4
	pop	 bp
	ret
@_STCON_$qv	endp

Padding bytes in code segments

  • Usually, padding 0x00 bytes are only emitted to word-align switch jump tables with -a2. Anywhere else, it typically indicates the start or end of a word-aligned SEGMENT compiled from assembly. There are two potential workarounds though:

    • The -WX option (Create DPMI application) will enforce word alignment for the code segment, at the cost of slightly different code generation in certain places. Since it also adds an additional INC BP instruction before PUSH BP, and an additional DEC BP instruction after POP BP, this option can only really be used in translation units with disabled stack frames (-k-).

    • #pragma codestring \x00 unconditionally emits a 0x00 byte. However, this won't respect different alignment requirements of surrounding translation units.

    Certainty: Reverse-engineering TCC.EXE confirmed that these are the only ways.

Memory segmentation

The segment and group a function will be emitted into can be controlled via #pragma option -zC / #pragma option -zP and #pragma codeseg. These mechanisms apply equally to function declarations and definitions. The active segment/group during a function's first reference will take precedence over any later segment/group the function shows up in:

#pragma option -zCfoo_TEXT -zPfoo

void bar(void);

#pragma codeseg baz_TEXT baz

// Despite the segment change in the line above, this function will still be
// put into `foo_TEXT`, the active segment during the first appearance of the
// function name.
void bar(void) {
}

// This function hasn't been declared yet, so it will go into `baz_TEXT` as
// expected.
void baz(void) {
}

When fixing up near references, the linker takes the actual flat/linear address and subtracts it from the base address of the reference's declared segment, assuming that the respective segment register is set to that specific segment at runtime. Therefore, incorrect segment declarations lead to incorrectly calculated offsets, and the linker can't realistically warn about such cases. There is the Fixup overflow error, but the linker only throws that one if the calculated distance exceeds 64 KiB and thus couldn't be expressed in a near reference to begin with.

Generally, it's better to just fix wrong segments at declaration time:

// Set the correct segment and group
#pragma codeseg bar_TEXT bar_correct_group

void near bar(void);

// Return to the default code segment
#pragma codeseg

However, there is a workaround if the intended near offset should simply be relative to the actual segment of the symbol: Declaring the identifier as far rather than near, and casting its implicit segment away for the near context. In the case of function pointers:

void far bar();
static nn_t qux = reinterpret_cast<nn_t>(bar);

This works because a far symbol always includes the segment it was emitted into. The cast simply reduces such a reference to its offset part within that segment.
This wrong declaration of bar() must, of course, not be #included into the translation unit that actually defines bar() as a near function, as it was intended. It can't also be local to an inlined function that's part of a public header, since those declarations seem to escape to the global scope there.

C++

In C++ mode, the value of a const scalar-type variable declared at global scope is always inlined, and not emitted into the data segment. Also, no externally visible symbol for the variable is emitted into the .OBJ file, even if the variable was not declared static. This makes such variables largely equivalent to #define macros.

Methods

Note the distinction between struct/class distance and method distance:

  • Declaring the type as near or far controls whether this is passed as a near or far pointer.
  • Declaring a method as near or far controls whether a method call generates a CALL near ptr or CALL far ptr instruction.

These can be freely combined, and one does not imply the other.

Inlining

Support for inlined functions is exclusive to C++ mode, with both top-level inline and class methods defined inside class declarations (obviously) not being supported in C mode. The compiler will inline every function defined in one of these ways, unless it contains one of these language constructs:

  • Loops (do, for, while, break, continue)
  • goto
  • switch and case
  • throw

If it doesn't, inlining is always worth a try to get rid of a potential macro, especially if all parameters are compile-time constants. There are a few further constructs that typically don't inline optimally though:

  • Assigning lvalues to value parameters (spills the value into a new compiler-generated local variable)
  • Assigning either rvalues or lvalues stored in registers to reference parameters (same)
  • Nested if statements – inlining will always generate a useless JMP SHORT $+2 at the end of the last branch.

Due to lazy code generation, values returned from inlined functions always preserve their type during the instruction generated for the return statement, and are only later cast into whatever the call site requires. Thus, certain type-mismatched instructions in arithmetic expressions can only be generated by returning one of the operands from an inlined function – unlike static_cast and reinterpret_cast, which can be optimized away.

Class methods inline to their ideal representation if all of these are true:

  • returns void || (returns *this && is at the first nesting level of inlining)
  • takes no parameters || takes only built-in, scalar-type parameters

Examples:

  • A class method (first nesting level) calling an overloaded operator (second nesting level) returning *this will generate (needless) instructions equivalent to MOV AX, *this. Thus, any overloaded =, +=, -=, etc. operator should always return void.

    Certainty: See the examples in 9d121c7. This is what allows us to use custom types with overloaded assignment operators, with the resulting code generation being indistinguishable from equivalent C preprocessor macros.

  • Returning anything else but void or *this will first store that result in AX, leading any branches at the call site to then refer to AX.

    Certainty: Maybe Borland (not Turbo) C++ has an optimization option against it?

Boilerplate for constructors defined outside the class declaration

struct MyClass {
  // Members…

  MyClass();
};

MyClass::MyClass() {
  // Initialization…
}

Resulting ASM:

; MyClass::MyClass(MyClass* this)
; Exact instructions differ depending on the memory model. Model-independent
; ASM instructions are in UPPERCASE.
@MyClass@$bctr$qv proc
  PUSH  BP
  MOV   BP, SP
  ; (saving SI and DI, if used in constructor code)
  ; (if this, 0)
  JNZ   @@ctor_code
  PUSH  sizeof(MyClass)
  CALL  @$bnew$qui      ; operator new(uint)
  POP   CX
  ; (this = value_returned_from_new)
  ; (if this)
  JZ    @@ret

@@ctor_code:
  ; Initialization…

@@ret:
  ; (retval = this)
  ; (restoring DI and SI, if used in constructor code)
  POP   BP
  RET
@MyClass@$bctr$qv endp
  • Arrays of nontrivially constructible objects are always constructed via _vector_new_. Conversely, no _vector_new_ = no array.

Limits of decompilability

MOV BX, SP-style functions, or others with no standard stack frame

These almost certainly weren't compiled from C. By disabling stack frames using #pragma option -k-, it might be possible to still get the exact same code out of Turbo C++ – even though it will most certainly look horrible, and barely more readable than assembly (or even less so), with tons of inline ASM and register pseudovariables. However, it's futile to even try if the function contains one of the following:

  • References to the SI or DI registers. In that case, Turbo C++ always inserts

    • a PUSH (SI|DI) at the beginning (after any PUSH BP; MOV BP, SP instructions and before anything else)
    • and a POP (SI|DI) before returning.

    Certainty: Confirmed through reverse-engineering TCC.EXE, no way around it.

Compiler bugs

  • Dereferencing a far pointer constructed from the _FS and _GS pseudoregisters emits wrong segment prefix opcodes – 0x46 (INC SI) and 0x4E (DEC SI) rather than the correct 0x64 and 0x65, respectively.

    Workaround: Not happening when compiling via TASM (-B on the command line, or #pragma inline).

  • Any emitted call to SCOPY@ will disable the stack cleanup optimization generated by -O for all generated code in a translation unit that follows the SCOPY@ call.