nanjit-opcodes.md

nanojit OpCodes

In Nanojit, LIR is the source language for compilation to machine code. LIR stands for low-level intermediate representation.  It is a typed assembly language. The LIR instruction set is best learnt by reading nanojit/LIRopcode.tbl.  Code for manipulating LIR is in nanojit/LIR.h.

LIR OpCodes

Opcodes use type-indicators suffixes that are loosely based on C type names:

• • 'c': "char", ie. 8-bit integer
• • 's': "short", ie. 16-bit integer
• • 'i': "int", ie. 32-bit integer
• • 'q': "quad", ie. 64-bit integer
• • 'u': "unsigned", is used as a prefix on integer type-indicators when necessary
• • 'f': "float", ie. 32-bit floating point value
• -'f4': "float4", ie. 128-bit SIMD value containing 4 single-precision floating point values
• • 'd': "double", ie. 64-bit floating point value
• • 'p': "pointer", ie. an int on 32-bit machines, a quad on 64-bit machines

'p' opcodes are all aliases of int and quad opcodes, they're given in LIR.h and chosen according to the platform pointer size.

Certain opcodes aren't supported on all platforms

Miscellaneous operations

Opcode	Todo	Return Type	Featured	Description
start	Op0	V		start of a fragment
regfence	Op0	V		A register fence causes no code to be generated, but it affects register allocation so that no registers are live when it is reached.
unreachable	Op0	V		indicate that this location cannot be reached (no live regs here)
skip	Sk	V		links code chunks
parami	P	I	32-bit	load an int parameter (register or stack location)
paramq	P	Q	64-bit	load a quad parameter (register or stack location)
allocp	IorF	P		allocate stack space (result is an address)
reti	Op1	V		return an int
retq	Op1	V	64-bit	return a quad
retd	Op1	V		return a double
retf	Op1	V		return a float
retf4	Op1	V		return a float4
livei	Op1	V		extend live range of an int
liveq	Op1	V	64-bit	extend live range of a quad
lived	Op1	V		extend live range of a double
livef	Op1	V		extend live range of a float
livef4	Op1	V		extend live range of a float4
file	Op1	V		[VTune] source filename for debug symbols
line	Op1	V		[VTune] source line number for debug symbols
pc	Op1	V		[Shark] record the machine address of this instruction
comment	Op1	V		a comment shown, on its own line, in LIR dumps
safe	Safe	V		deoptimization safepoint
endsafe	Safe	V		deoptimization safepoint

Loads and stores

Opcode	Todo	Return Type	Featured	Description
ldc2i	Ld	I		load char and sign-extend to an int
lds2i	Ld	I		load short and sign-extend to an int
lduc2ui	Ld	I		load unsigned char and zero-extend to an unsigned int
ldus2ui	Ld	I		load unsigned short and zero-extend to an unsigned int
ldi	Ld	I		load int
ldq	Ld	Q	64-bit	load quad
ldd	Ld	D		load double
ldf	Ld	F		load float
ldf2d	Ld	D		load float and extend to a double
ldf4	Ld	F4		load float4 (SIMD, 4 floats)
sti2c	St	V		store int truncated to char
sti2s	St	V		store int truncated to short
sti	St	V		store int
stq	St	V	64-bit	store quad
std	St	V		store double
std2f	St	V
stf	St	V
stf4	St	V

Calls

Opcode	Todo	Return Type	Featured	Description
callv	C	V		call subroutine that returns void
calli	C	I		call subroutine that returns an int
callq	C	Q	64-bit	call subroutine that returns a quad
calld	C	D		call subroutine that returns a double
callf	C	F		call subroutine that returns a float
callf4	C	F4		call subroutine that returns a float4

Branches and labels

'jt' and 'jf' must be adjacent so that (op ^ 1) gives the opposite one. Static assertions in LIR.h check this requirement.

Opcode	Todo	Return Type	Featured	Description
j	Op2	V		jump always
jt	Op2	V		jump if true
jf	Op2	V		jump if false
jtbl	Jtbl	V		jump to address in table
label	Op0	V		a jump target (no machine code is emitted for this)

Guards

'xt' and 'xf' must be adjacent so that (op ^ 1) gives the opposite one. Static assertions in LIR.h check this requirement.

OP_UN (align_guards)

Opcode	Todo	Return Type	Featured	Description
x	Op2	V		exit always
xt	Op2	V		exit if true
xf	Op2	V		exit if false
xbarrier	Op2	V		A LIR_xbarrier cause no code to be generated, but it acts like a never-taken guard in that it inhibits certain optimisations, such as dead stack store elimination.

Immediates

Opcode	Todo	Return Type	Featured	Description
immi	IorF	I		int immediate
immq	QorD	Q	64-bit	quad immediate
immd	QorD	D		double immediate
immf	IorF	F		float immediate
immf4	F4	F4		float4 immediate

Comparisons

All comparisons return an int: 0 on failure and 1 on success. Within each type group, order must be preserved so that, except for eq*, (op ^ 1) gives the opposite one (eg. lt ^ 1 == gt). eq* must have odd numbers for this to work. They must also remain contiguous so that opcode range checking works correctly. Static assertions in LIR.h check these requirements.

OP_UN (align_eqi)

Opcode	Todo	Return Type	Featured	Description
eqi	Op2	I		int equality
lti	Op2	I		signed int less-than
gti	Op2	I		signed int greater-than
lei	Op2	I		signed int less-than-or-equal
gei	Op2	I		signed int greater-than-or-equal
ltui	Op2	I		unsigned int less-than
gtui	Op2	I		unsigned int greater-than
leui	Op2	I		unsigned int less-than-or-equal
geui	Op2	I		unsigned int greater-than-or-equal

OP_UN_64(align_eqq)

Opcode	Todo	Return Type	Featured	Description
eqq	Op2	I	64-bit	quad equality
ltq	Op2	I	64-bit	signed quad less-than
gtq	Op2	I	64-bit	signed quad greater-than
leq	Op2	I	64-bit	signed quad less-than-or-equal
geq	Op2	I	64-bit	signed quad greater-than-or-equal
ltuq	Op2	I	64-bit	unsigned quad less-than
gtuq	Op2	I	64-bit	unsigned quad greater-than
leuq	Op2	I	64-bit	unsigned quad less-than-or-equal
geuq	Op2	I	64-bit	unsigned quad greater-than-or-equal

OP_UN_64(align_eqd)

Opcode	Todo	Return Type	Featured	Description
eqd	Op2	I		double equality
ltd	Op2	I		double less-than
gtd	Op2	I		double greater-than
led	Op2	I		double less-than-or-equal
ged	Op2	I		double greater-than-or-equal

OP_UN (align_eqf)

Opcode	Todo	Return Type	Featured	Description
eqf	Op2	I		float equality
ltf	Op2	I		float less-than
gtf	Op2	I		float greater-than
lef	Op2	I		float less-than-or-equal
gef	Op2	I		float greater-than-or-equal
eqf4	Op2	I		float4 equality

Note: we don't do lt/gt/le/ge comparisons on float4 values

Arithmetic

Opcode	Todo	Return Type	Featured	Description
negi	Op1	I		negate int
addi	Op2	I		add int
subi	Op2	I		subtract int
muli	Op2	I		multiply int
divi	Op2	I	32-bit X86	divide int
modi	Op1	I	32-bit X86	modulo int. LIR_modi is a hack. It's only used on i386/X64. The operand is the result of a LIR_divi because on i386/X64 div and mod results are computed by the same instruction.
noti	Op1	I		bitwise-NOT int
andi	Op2	I		bitwise-AND int
ori	Op2	I		bitwise-OR int
xori	Op2	I		bitwise-XOR int
lshi	Op2	I		left shift int. For all three integer shift operations, only the bottom five bits of the second operand are used, and they are treated as unsigned. This matches x86 semantics.
rshi	Op2	I		right shift int (>>)
rshui	Op2	I		right shift unsigned int (>>>)
addq	Op2	Q	64-bit	add quad
subq	Op2	Q	64-bit	subtract quad
mulq	Op2	Q	64-bit X86	multiply quad
divq	Op2	Q	64-bit X86	divide quad
modq	Op1	Q	64-bit X86	modulo quad. LIR_modq is a hack. It's only used on i386/X64. The operand is the result of a LIR_divq because on i386/X64 div and mod results are computed by the same instruction.
andq	Op2	Q	64-bit	bitwise-AND quad
orq	Op2	Q	64-bit	bitwise-OR quad
xorq	Op2	Q	64-bit	bitwise-XOR quad
lshq	Op2	Q	64-bit	left shift quad; 2nd operand is an int. For all three quad shift operations, only the bottom six bits of the second operand are used, and they are treated as unsigned. This matches x86-64 semantics.
rshq	Op2	Q	64-bit	right shift quad; 2nd operand is an int
rshuq	Op2	Q	64-bit	right shift unsigned quad; 2nd operand is an int
negd	Op1	D		negate double
absd	Op1	D		absolute value of double
sqrtd	Op1	D		sqrt double
addd	Op2	D		add double
subd	Op2	D		subtract double
muld	Op2	D		multiply double
divd	Op2	D		divide double
modd	Op2	D		modulo double. LIR_modd is just a place-holder opcode, ie. the back-ends cannot generate code for it. It's used in TraceMonkey briefly but is always demoted to a LIR_modl or converted to a function call before Nanojit has to do anything serious with it.
negf	Op1	F		negate float
absf	Op1	F		absolute value of float
sqrtf	Op1	F		sqrt float
addf	Op2	F		add float
subf	Op2	F		subtract float
mulf	Op2	F		multiply float
divf	Op2	F		divide float
negf4	Op1	F4		negate float4
absf4	Op1	F4		absolute value of float4
sqrtf4	Op1	F4		sqrt float4
addf4	Op2	F4		add float4
subf4	Op2	F4		subtract float4
mulf4	Op2	F4		multiply float4
divf4	Op2	F4		divide float4
recipf	Op1	F		float reciprocal
rsqrtf	Op1	F		float reciprocal square root
minf	Op2	F		float min
maxf	Op2	F		float max
cmpgtf4	Op2	F4		float4.isGreater
cmpltf4	Op2	F4		float4.isLess
cmpgef4	Op2	F4		float4.isGreaterOrEqual
cmplef4	Op2	F4		float4.isLessOrEqual
cmpeqf4	Op2	F4		float4.isEqual
cmpnef4	Op2	F4		float4.isNotEqual
recipf4	Op1	F4		float4 reciprocal
rsqrtf4	Op1	F4		float4 reciprocal square root
minf4	Op2	F4		float4 min
maxf4	Op2	F4		float4 max
dotf4	Op2	F		4-component dot product
dotf3	Op2	F		3-component dot product
dotf2	Op2	F		2-component dot product
cmovi	Op3	I		conditional move int
cmovq	Op3	Q	64-bit	conditional move quad
cmovd	Op3	D		conditional move double
cmovf	Op3	F		conditional move float
cmovf4	Op3	F4		conditional move float4

Conversions

rounding behavior of LIR_d2f is platform-specific

Platform	Asm code	Behavior
x86 w/ x87	FST32	uses current FP control word (default is rounding)
x86 w/ SSE	cvtsd2ss	according to MXCSR register (default is round to nearest)
x64 (SSE)	cvtsd2ss	according to MXCSR register (default is round to nearest)
others		not implemented yet

The rounding behavior of LIR_d2i is platform specific.

Platform	Asm code	Behavior
x86 w/ x87	fist	uses current FP control word (default is rounding)
x86 w/ SSE	cvttsd2si	performs round to zero (truncate)
x64 (SSE)	cvttsd2si	performs round to zero (truncate)
PowerPC		unsupported
ARM	ftosid	round to nearest
MIPS	trunc.w.d	performs round to zero (truncate)
SH4	frtc	performs round to zero (truncate)
SPARC	fdtoi	performs round to zero (truncate)

• round to zero examples: 1.9 -> 1, 1.1 -> 1, -1.1 -> -1, -1.9 -> -1
• round to nearest examples: 1.9 -> 2, 1.1 -> 1, -1.1 -> -1, -1.9 -> -2

Opcode	Todo	Return Type	Featured	Description
i2q	Op1	Q	64-bit	sign-extend int to quad
ui2uq	Op1	Q	64-bit	zero-extend unsigned int to unsigned quad
q2i	Op1	I	64-bit	truncate quad to int (removes the high 32 bits)
q2d	Op1	D	64-bit	convert quad to double
i2d	Op1	D		convert int to double
i2f	Op1	F		convert int to float
ui2d	Op1	D		convert unsigned int to double
ui2f	Op1	F		convert unsigned int to float
f2d	Op1	D		convert float to double
d2f	Op1	F		convert double to float (no exceptions raised)
d2i	Op1	I		convert double to int (no exceptions raised)
d2q	Op1	Q	64-bit X86	convert double to quad (no exceptions raised?)
f2i	Op1	I		convert float to int (no exceptions raised)
f2f4	Op1	F4		convert float to float4 (no exceptions raised) - essentially copies the float across all elements
ffff2f4	Op4	F4		convert float to float4 (no exceptions raised) - essentially copies the float across all elements
f4x	Op1	F		extract first float from a float4
f4y	Op1	F		extract second float from a float4
f4z	Op1	F		extract third float from a float4
f4w	Op1	F		extract fourth float from a float4
swzf4	Op1b	F4		swizzle float4 according to 8-bit selector
dasq	Op1	Q	64-bit	interpret the bits of a double as a quad
qasd	Op1	D	64-bit	interpret the bits of a quad as a double

Overflow arithmetic

These all exit if overflow occurred. The result is valid on either path.

Opcode	Todo	Return Type	Featured	Description
addxovi	Op3	I		add int and exit on overflow
subxovi	Op3	I		subtract int and exit on overflow
mulxovi	Op3	I		multiply int and exit on overflow

These all branch if overflow occurred. The result is valid on either path.

Opcode	Todo	Return Type	Featured	Description
addjovi	Op3	I		add int and branch on overflow
subjovi	Op3	I		subtract int and branch on overflow
muljovi	Op3	I		multiply int and branch on overflow
addjovq	Op3	Q	64-bit	add quad and branch on overflow
subjovq	Op3	Q	64-bit	subtract quad and branch on overflow

SoftFloat

Opcode	Todo	Return Type	Featured	Description
dlo2i	Op1	I	SF	get the low 32 bits of a double as an int
dhi2i	Op1	I	SF	get the high 32 bits of a double as an int
ii2d	Op2	D	SF	join two ints (1st arg is low bits, 2nd is high)
hcalli	Op1	I	SF	LIR_hcalli is a hack that's only used on 32-bit platforms that use SoftFloat. Its operand is always a LIR_calli, but one that specifies a function that returns a double. It indicates that the double result is returned via two 32-bit integer registers. The result is always used as the second operand of a LIR_ii2d.

Safepoint Polling

Opcode	Todo	Return Type	Featured	Description
memfence	Op0	V
brsavpc	Op2	V		branch and save pc
restorepc	Op0	V
pushstate	Op0	V
popstate	Op0	V