Inline Assembler

D, being a systems programming language, provides an inline assembler. The inline assembler is standardized for D implementations across the same CPU family, for example, the Intel Pentium inline assembler for a Win32 D compiler will be syntax compatible with the inline assembler for Linux running on an Intel Pentium.

Implementations of D on different architectures, however, are free to innovate upon the memory model, function call/return conventions, argument passing conventions, etc.

This document describes the x86 and x86_64 implementations of the inline assembler. The inline assembler platform support that a compiler provides is indicated by the D_InlineAsm_X86 and D_InlineAsm_X86_64 version identifiers, respectively.

Asm statement

AsmStatement:
    asm FunctionAttributesopt { AsmInstructionListopt }

    AsmInstructionList:
        AsmInstruction ;
        AsmInstruction ; AsmInstructionList
    

Assembler instructions must be located inside an asm block. Like functions, asm statements must be anotated with adequate function attributes to be compatible with the caller. Asm statements attributes must be explicitly defined, they are not infered.

void func1() pure nothrow @safe @nogc
{
    asm pure nothrow @trusted @nogc
    {}
}

void func2() @safe @nogc
{
    asm @nogc // Error: asm statement is assumed to be @system - mark it with '@trusted' if it is not
    {}
}

Asm instruction

AsmInstruction:
    Identifier : AsmInstruction
    align IntegerExpression
    even
    naked
    db Operands
    ds Operands
    di Operands
    dl Operands
    df Operands
    dd Operands
    de Operands
    db StringLiteral
    ds StringLiteral
    di StringLiteral
    dl StringLiteral
    dw StringLiteral
    dq StringLiteral
    Opcode
    Opcode Operands

Operands:
    Operand
    Operand , Operands

Labels

Assembler instructions can be labeled just like other statements. They can be the target of goto statements. For example:

void *pc;
asm
{
    call L1          ;
  L1:                ;
    pop  EBX         ;
    mov  pc[EBP],EBX ; // pc now points to code at L1
}

align IntegerExpression

IntegerExpression:
    IntegerLiteral
    Identifier

Causes the assembler to emit NOP instructions to align the next assembler instruction on an IntegerExpression boundary. IntegerExpression must evaluate at compile time to an integer that is a power of 2.

Aligning the start of a loop body can sometimes have a dramatic effect on the execution speed.

even

Causes the assembler to emit NOP instructions to align the next assembler instruction on an even boundary.

naked

Causes the compiler to not generate the function prolog and epilog sequences. This means such is the responsibility of inline assembly programmer, and is normally used when the entire function is to be written in assembler.

db, ds, di, dl, df, dd, de

These pseudo ops are for inserting raw data directly into the code. db is for bytes, ds is for 16 bit words, di is for 32 bit words, dl is for 64 bit words, df is for 32 bit floats, dd is for 64 bit doubles, and de is for 80 bit extended reals. Each can have multiple operands. If an operand is a string literal, it is as if there were length operands, where length is the number of characters in the string. One character is used per operand. For example:

asm
{
    db 5,6,0x83;   // insert bytes 0x05, 0x06, and 0x83 into code
    ds 0x1234;     // insert bytes 0x34, 0x12
    di 0x1234;     // insert bytes 0x34, 0x12, 0x00, 0x00
    dl 0x1234;     // insert bytes 0x34, 0x12, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
    df 1.234;      // insert float 1.234
    dd 1.234;      // insert double 1.234
    de 1.234;      // insert real 1.234
    db "abc";      // insert bytes 0x61, 0x62, and 0x63
    ds "abc";      // insert bytes 0x61, 0x00, 0x62, 0x00, 0x63, 0x00
}

Opcodes

A list of supported opcodes is at the end.

The following registers are supported. Register names are always in upper case.

Register:
    AL AH AX EAX
    BL BH BX EBX
    CL CH CX ECX
    DL DH DX EDX
    BP EBP
    SP ESP
    DI EDI
    SI ESI
    ES CS SS DS GS FS
    CR0 CR2 CR3 CR4
    DR0 DR1 DR2 DR3 DR6 DR7
    TR3 TR4 TR5 TR6 TR7
    ST
    ST(0) ST(1) ST(2) ST(3) ST(4) ST(5) ST(6) ST(7)
    MM0  MM1  MM2  MM3  MM4  MM5  MM6  MM7
    XMM0 XMM1 XMM2 XMM3 XMM4 XMM5 XMM6 XMM7

x86_64 adds these additional registers.

Register64:
    RAX  RBX  RCX  RDX
    BPL  RBP
    SPL  RSP
    DIL  RDI
    SIL  RSI
    R8B  R8W  R8D  R8
    R9B  R9W  R9D  R9
    R10B R10W R10D R10
    R11B R11W R11D R11
    R12B R12W R12D R12
    R13B R13W R13D R13
    R14B R14W R14D R14
    R15B R15W R15D R15
    XMM8 XMM9 XMM10 XMM11 XMM12 XMM13 XMM14 XMM15
    YMM0 YMM1 YMM2  YMM3  YMM4  YMM5  YMM6  YMM7
    YMM8 YMM9 YMM10 YMM11 YMM12 YMM13 YMM14 YMM15

Special Cases

lock, rep, repe, repne, repnz, repz

These prefix instructions do not appear in the same statement as the instructions they prefix; they appear in their own statement. For example:

asm
{
    rep   ;
    movsb ;
}

pause

This opcode is not supported by the assembler, instead use

asm
{
    rep  ;
    nop  ;
}

which produces the same result.

floating point ops

Use the two operand form of the instruction format;

fdiv ST(1);     // wrong
fmul ST;        // wrong
fdiv ST,ST(1);  // right
fmul ST,ST(0);  // right

Operands

Operand:
    AsmExp

AsmExp:
    AsmLogOrExp
    AsmLogOrExp ? AsmExp : AsmExp

AsmLogOrExp:
    AsmLogAndExp
    AsmLogOrExp || AsmLogAndExp

AsmLogAndExp:
    AsmOrExp
    AsmLogAndExp && AsmOrExp

AsmOrExp:
    AsmXorExp
    AsmOrExp | AsmXorExp

AsmXorExp:
    AsmAndExp
    AsmXorExp ^ AsmAndExp

AsmAndExp:
    AsmEqualExp
    AsmAndExp & AsmEqualExp

AsmEqualExp:
    AsmRelExp
    AsmEqualExp == AsmRelExp
    AsmEqualExp != AsmRelExp

AsmRelExp:
    AsmShiftExp
    AsmRelExp < AsmShiftExp
    AsmRelExp <= AsmShiftExp
    AsmRelExp > AsmShiftExp
    AsmRelExp >= AsmShiftExp

AsmShiftExp:
    AsmAddExp
    AsmShiftExp << AsmAddExp
    AsmShiftExp >> AsmAddExp
    AsmShiftExp >>> AsmAddExp

AsmAddExp:
    AsmMulExp
    AsmAddExp + AsmMulExp
    AsmAddExp - AsmMulExp

AsmMulExp:
    AsmBrExp
    AsmMulExp * AsmBrExp
    AsmMulExp / AsmBrExp
    AsmMulExp % AsmBrExp

AsmBrExp:
    AsmUnaExp
    AsmBrExp [ AsmExp ]

AsmUnaExp:
    AsmTypePrefix AsmExp
    offsetof AsmExp
    seg AsmExp
    + AsmUnaExp
    - AsmUnaExp
    ! AsmUnaExp
    ~ AsmUnaExp
    AsmPrimaryExp

AsmPrimaryExp:
    IntegerLiteral
    FloatLiteral
    __LOCAL_SIZE
    $
    Register
    Register : AsmExp
    Register64
    Register64 : AsmExp
    DotIdentifier
    this

DotIdentifier:
    Identifier
    Identifier . DotIdentifier

The operand syntax more or less follows the Intel CPU documentation conventions. In particular, the convention is that for two operand instructions the source is the right operand and the destination is the left operand. The syntax differs from that of Intel's in order to be compatible with the D language tokenizer and to simplify parsing.

The seg means load the segment number that the symbol is in. This is not relevant for flat model code. Instead, do a move from the relevant segment register.

Operand Types

AsmTypePrefix:
    near ptr
    far ptr
    byte ptr
    short ptr
    int ptr
    word ptr
    dword ptr
    qword ptr
    float ptr
    double ptr
    real ptr

In cases where the operand size is ambiguous, as in:

add [EAX],3     ;

it can be disambiguated by using an AsmTypePrefix:

add  byte ptr [EAX],3 ;
add  int ptr [EAX],7  ;

far ptr is not relevant for flat model code.

Struct/Union/Class Member Offsets

To access members of an aggregate, given a pointer to the aggregate is in a register, use the .offsetof property of the qualified name of the member:

struct Foo { int a,b,c; }
int bar(Foo *f)
{
    asm
    {
        mov EBX,f                   ;
        mov EAX,Foo.b.offsetof[EBX] ;
    }
}
void main()
{
    Foo f = Foo(0, 2, 0);
    assert(bar(&f) == 2);
}

Alternatively, inside the scope of an aggregate, only the member name is needed:

struct Foo   // or class
{
    int a,b,c;
    int bar()
    {
        asm
        {
            mov EBX, this   ;
            mov EAX, b[EBX] ;
        }
    }
}
void main()
{
    Foo f = Foo(0, 2, 0);
    assert(f.bar() == 2);
}

Stack Variables

Stack variables (variables local to a function and allocated on the stack) are accessed via the name of the variable indexed by EBP:

int foo(int x)
{
    asm
    {
        mov EAX,x[EBP] ; // loads value of parameter x into EAX
        mov EAX,x      ; // does the same thing
    }
}

If the [EBP] is omitted, it is assumed for local variables. If naked is used, this no longer holds.

Special Symbols

$

Represents the program counter of the start of the next instruction. So,

jmp  $  ;

branches to the instruction following the jmp instruction. The $ can only appear as the target of a jmp or call instruction.

__LOCAL_SIZE

This gets replaced by the number of local bytes in the local stack frame. It is most handy when the naked is invoked and a custom stack frame is programmed.

Opcodes Supported

Opcodes

aaa	aad	aam	aas	adc
add	addpd	addps	addsd	addss
and	andnpd	andnps	andpd	andps
arpl	bound	bsf	bsr	bswap
bt	btc	btr	bts	call
cbw	cdq	clc	cld	clflush
cli	clts	cmc	cmova	cmovae
cmovb	cmovbe	cmovc	cmove	cmovg
cmovge	cmovl	cmovle	cmovna	cmovnae
cmovnb	cmovnbe	cmovnc	cmovne	cmovng
cmovnge	cmovnl	cmovnle	cmovno	cmovnp
cmovns	cmovnz	cmovo	cmovp	cmovpe
cmovpo	cmovs	cmovz	cmp	cmppd
cmpps	cmps	cmpsb	cmpsd	cmpss
cmpsw	cmpxchg	cmpxchg8b	cmpxchg16b
comisd	comiss
cpuid	cvtdq2pd	cvtdq2ps	cvtpd2dq	cvtpd2pi
cvtpd2ps	cvtpi2pd	cvtpi2ps	cvtps2dq	cvtps2pd
cvtps2pi	cvtsd2si	cvtsd2ss	cvtsi2sd	cvtsi2ss
cvtss2sd	cvtss2si	cvttpd2dq	cvttpd2pi	cvttps2dq
cvttps2pi	cvttsd2si	cvttss2si	cwd	cwde
da	daa	das	db	dd
de	dec	df	di	div
divpd	divps	divsd	divss	dl
dq	ds	dt	dw	emms
enter	f2xm1	fabs	fadd	faddp
fbld	fbstp	fchs	fclex	fcmovb
fcmovbe	fcmove	fcmovnb	fcmovnbe	fcmovne
fcmovnu	fcmovu	fcom	fcomi	fcomip
fcomp	fcompp	fcos	fdecstp	fdisi
fdiv	fdivp	fdivr	fdivrp	feni
ffree	fiadd	ficom	ficomp	fidiv
fidivr	fild	fimul	fincstp	finit
fist	fistp	fisub	fisubr	fld
fld1	fldcw	fldenv	fldl2e	fldl2t
fldlg2	fldln2	fldpi	fldz	fmul
fmulp	fnclex	fndisi	fneni	fninit
fnop	fnsave	fnstcw	fnstenv	fnstsw
fpatan	fprem	fprem1	fptan	frndint
frstor	fsave	fscale	fsetpm	fsin
fsincos	fsqrt	fst	fstcw	fstenv
fstp	fstsw	fsub	fsubp	fsubr
fsubrp	ftst	fucom	fucomi	fucomip
fucomp	fucompp	fwait	fxam	fxch
fxrstor	fxsave	fxtract	fyl2x	fyl2xp1
hlt	idiv	imul	in	inc
ins	insb	insd	insw	int
into	invd	invlpg	iret	iretd
iretq	ja	jae	jb	jbe
jc	jcxz	je	jecxz	jg
jge	jl	jle	jmp	jna
jnae	jnb	jnbe	jnc	jne
jng	jnge	jnl	jnle	jno
jnp	jns	jnz	jo	jp
jpe	jpo	js	jz	lahf
lar	ldmxcsr	lds	lea	leave
les	lfence	lfs	lgdt	lgs
lidt	lldt	lmsw	lock	lods
lodsb	lodsd	lodsw	loop	loope
loopne	loopnz	loopz	lsl	lss
ltr	maskmovdqu	maskmovq	maxpd	maxps
maxsd	maxss	mfence	minpd	minps
minsd	minss	mov	movapd	movaps
movd	movdq2q	movdqa	movdqu	movhlps
movhpd	movhps	movlhps	movlpd	movlps
movmskpd	movmskps	movntdq	movnti	movntpd
movntps	movntq	movq	movq2dq	movs
movsb	movsd	movss	movsw	movsx
movupd	movups	movzx	mul	mulpd
mulps	mulsd	mulss	neg	nop
not	or	orpd	orps	out
outs	outsb	outsd	outsw	packssdw
packsswb	packuswb	paddb	paddd	paddq
paddsb	paddsw	paddusb	paddusw	paddw
pand	pandn	pavgb	pavgw	pcmpeqb
pcmpeqd	pcmpeqw	pcmpgtb	pcmpgtd	pcmpgtw
pextrw	pinsrw	pmaddwd	pmaxsw	pmaxub
pminsw	pminub	pmovmskb	pmulhuw	pmulhw
pmullw	pmuludq	pop	popa	popad
popf	popfd	por	prefetchnta	prefetcht0
prefetcht1	prefetcht2	psadbw	pshufd	pshufhw
pshuflw	pshufw	pslld	pslldq	psllq
psllw	psrad	psraw	psrld	psrldq
psrlq	psrlw	psubb	psubd	psubq
psubsb	psubsw	psubusb	psubusw	psubw
punpckhbw	punpckhdq	punpckhqdq	punpckhwd	punpcklbw
punpckldq	punpcklqdq	punpcklwd	push	pusha
pushad	pushf	pushfd	pxor	rcl
rcpps	rcpss	rcr	rdmsr	rdpmc
rdtsc	rep	repe	repne	repnz
repz	ret	retf	rol	ror
rsm	rsqrtps	rsqrtss	sahf	sal
sar	sbb	scas	scasb	scasd
scasw	seta	setae	setb	setbe
setc	sete	setg	setge	setl
setle	setna	setnae	setnb	setnbe
setnc	setne	setng	setnge	setnl
setnle	setno	setnp	setns	setnz
seto	setp	setpe	setpo	sets
setz	sfence	sgdt	shl	shld
shr	shrd	shufpd	shufps	sidt
sldt	smsw	sqrtpd	sqrtps	sqrtsd
sqrtss	stc	std	sti	stmxcsr
stos	stosb	stosd	stosw	str
sub	subpd	subps	subsd	subss
syscall	sysenter	sysexit	sysret	test
ucomisd	ucomiss	ud2	unpckhpd	unpckhps
unpcklpd	unpcklps	verr	verw	wait
wbinvd	wrmsr	xadd	xchg	xlat
xlatb	xor	xorpd	xorps

Pentium 4 (Prescott) Opcodes Supported

Pentium 4 Opcodes

addsubpd	addsubps	fisttp	haddpd	haddps
hsubpd	hsubps	lddqu	monitor	movddup
movshdup	movsldup	mwait

AMD Opcodes Supported

AMD Opcodes

pavgusb	pf2id	pfacc	pfadd	pfcmpeq
pfcmpge	pfcmpgt	pfmax	pfmin	pfmul
pfnacc	pfpnacc	pfrcp	pfrcpit1	pfrcpit2
pfrsqit1	pfrsqrt	pfsub	pfsubr	pi2fd
pmulhrw	pswapd

SIMD

SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2 and AVX are supported.