Many C and C++ programmers have never seen bit fields.
Bit fields are a feature of the C and C++ language which completely hide what is often called "bit bashing".
Bit bashing is the manipulation of individual bits. Bit bashing goes to the very core of the C language. Remember that C is a high level assembly language, as we argue in Section 1 of this book. And C is the (later) language in which Unix was implemented and indeed, C was developed specifically to implement Unix.
Since an operating system directly interfaces with hardware - the C language grew to have features to aid Unix implementers.
With that said, consider this WARNING: the ordering of bits in a bit field is not guaranteed to be the same on different platforms and even between different compilers on the same platform.
Bit fields are implemented within a struct
by appending a colon plus
a number after the declaration of integer types.
For example:
struct BF {
unsigned char a : 1;
unsigned char b : 2;
unsigned char c : 5;
};
The above declares a struct
whose size is 1 byte. Members of the struct
are a
, b
and c
which are 1, 2 and 5 bits in size, respectively.
Consider a data structure for which there will be potentially millions of
instances in RAM. Or, perhaps billions of instances on disc. Suppose you
need 8 boolean members in every instance. The C++ standard does not
define the size of a bool
instead leaving it to be implementation
dependent. Some implementations equate bool
to int
, four bytes in
length. Some implement bool
with a char
, or 1 byte in length.
Let's assume the smallest case and equate a bool
with char
. Our
struct
, for which there may be millions or billions of instances
requires 8 bool
so therefore 8 bytes. Times millions or billions.
Ouch.
Bit fields can come to your aid here by using a single bit per boolean value. In the best case, 8 bytes collapse to 1 byte. In a worse case, 8 x 4 = 32 bytes collapsed into 1.
Before we examine using bit fields, let's look at what life would be like without them.
Let's assume we're working with a byte that is comprised of three
fields laid out as in struct BF
above. That is, a one, two and
five bit field inside one byte.
Without bit fields, we would have to write this code to clear a
to zero:
void ClearA(unsigned char * byte) {
*byte &= ~1;
}
This function takes the address of the byte containing the a
,
b
and c
portions.
Good programming practice would check byte
against NULL
or nullptr
.
The ~
operator is a bitwise negation. All the bits in the
value are flipped from 0 to 1 or 1 to 0. ~1
in an unsigned
char will produce 0xFE
, or all ones except for bit 0. and
ing
this value to *byte
ensures that its bit 0 is 0 and all other
bits are left alone.
In assembly language, written naively, this would look like this:
ClearA: ldrb w1, [x0] // 1
mov w2, 1 // 2
mvn w2, w2 // 3
and w1, w1, w2 // 4
strb w1, [x0] // 5
ret // 6
x30
does not have to be backed up or restored as this function is a "leaf."
Line 3
uses the instruction mvn
to flip all the bits in w2
.
This code completely tracks the C / C++ code.
We have no obligation to follow the C / C++ code exactly. Instead we could write:
ClearA: ldrb w1, [x0] // 1
and w1, w1, 0xFE // 2
strb w1, [x0] // 3
ret // 4
Here, the 0xFE
literal takes the place of lines 2 and 3
in the previous
version. We do this by pre-computing what the mov
and mvn
would have
produced.
For setting the a
bit, we would do this:
void SetA(unsigned char * byte) {
*byte |= 1;
}
This is an anomaly for bit bashing. In almost all cases when setting bit values, the bits must be cleared first because an or instruction is responsible for setting any 1 bits to 1.
It is important you get that when needing to set a number of bits to a
specific value, those bit must be cleared first so that an orr
can do
the right thing.
In this case, it is a single bit we're setting so we can just or it in.
In assembly language:
SetA: ldrb w1, [x0] // 1
orr w1, w1, 1 // 2
strb w1, [x0] // 3
ret // 4
orr
is one of several or instructions in AARCH64. It is the one that maps
most closely to |
in C and C++.
Moving onto the b
field, things begin to get a little more interesting.
To clear the b
field we might do this in C | C++.
void ClearB(unsigned char * byte) {
*byte &= ~6;
}
This could naively be written as:
ClearB: ldrb w1, [x0] // 1
mov w2, 6 // 2
mvn w2, w2 // 3
and w1, w1, w2 // 4
strb w1, [x0] // 5
ret // 6
This code is essentially the same as the naive version of ClearA
given
above. Once again, we can pre-compute the results of lines 2 and 3
to
make:
ClearB: ldrb w1, [x0] // 1
and w1, w1, 0xF9 // 2
strb w1, [x0] // 3
ret // 4
Turning to setting b
, the code gets a little more complicated as for
the first time, we have to accept a parameter for the value to place into
b
. And, b
is more than one bit.
void SetB(unsigned char * byte, unsigned char value) { // 1
value &= 3; // ensures only bits 0 and 1 can be set // 2
*byte &= ~6; // clears bits 1 and 2 in byte // 3
*byte |= (value << 1); // stores bits 0 and 1 into bits 2 and 3 // 4
} // 5
Line 2
is necessary to prevent stray 1's from being or'ed into *byte
.
Line 3
is necessary to squash the existing target bits to zero prior
to being orr
'ed.
Notice value
is being shifted left by 1 bit as the b
field begins at
bit index 1.
In naive assembly language we could write this:
SetB: ldrb w3, [x0] // 1
and w1, w1, 3 // value &= 3 // 2
lsl w1, w1, 1 // 3
mov w2, 6 // 4
mvn w2, w2 // 5
and w3, w3, w2 // B is cleared // 6
orr w3, w3, w1 // 7
strb w3, [x0] // 8
ret // 9
The only interesting thing in this code is that we chose to perform the
left shift (lsl
) by one bit earlier in the code rather than later.
There is ill no side effect to changing this order.
lsl
means "left shift logical" which fills the right side recently
vacated bits with zero.
Now, we present a more sophisticated version of SetB
:
SetB: ldrb w3, [x0] // 1
bfi w3, w1, 1, 2 // copy bit 0..1 in w1 to bit 1..2 in w3 // 2
strb w3, [x0] // 3
ret // 4
Whoa. Nine instructions down to four! What the heck is bfi
?
bfi dst, src, start, width
copies width
bits starting at 0 in src
to bits starting at start
in dst
.
It obviates the need for line 2
in
the naive code because it plucks only bits 0 and 1 and no others from the
original value
of w1
.
The bfi
then internally does the shift appropriate to move
bit 0 of w1
to bit start
along with width - 1
subsequent bits. Finally, the shifted bits overwrite the same bits
in w3
.
Some might argue that instructions like bfi
(and ubfiz
described
below) is an example of ISA creep
where ISA's get
more and more cumbersome with the latest instructions du jure. This is
definitely true in the x86 ISA. Perhaps this is true in the AARCH64 ISA
as well, but certainly not to the extent of the x86.
Remember that the ARM family of processors are examples of RISC machines - reduced instruction set architectures.
Finally, we come to handling field c
. Recall c
is 5 bits long starting
at bit 3.
Clearing the bits in c
is easily accomplished:
void ClearC(unsigned char * byte) {
*byte &= 7; // squashes bits 3 to 7 to 0
}
This is optimally implemented using:
ClearC: ldrb w1, [x0] // 1
and w1, w1, 7 // 2
strb w1, [x0] // 3
ret // 4
As for setting the value of c
, we have this in C / C++:
void SetC(unsigned char * byte, unsigned char value) {
value &= 0x1F; // ensures only bits 0 to 4 can be set
*byte &= ~(0x1F << 3); // squashes correct bits in byte
*byte |= (value << 3); // or's in the bits at the right place
}
In naive assembly language, this function would look like this:
SetC: ldrb w3, [x0] // 1
mov w2, 0x1F // 2
and w1, w1, w2 // 3
lsl w1, w1, 3 // 4
lsl w2, w2, 3 // 5
mvn w2, w2 // 6
and w3, w3, w2 // 7
orr w3, w3, w1 // 8
strb w3, [x0] // 9
ret // 10
Lines 1 and 2
in the assembly language performs line 1
of the C code.
Line 4
shifts value
up to where c
starts. Line 5
similarly shifts
the mask up to where c
starts. Its bits are negated on line 6
. Line 7
squashes the upper five bits to zero followed by the orr
ing on line 8
.
A more sophisticated version of the assembly language, leveraging some fancy bit insertion / copying instructions, is far shorter.
SetC: ldrb w2, [x0] // put *byte into w2 // 1
ubfiz w1, w1, 3, 5 // zero new w1, copy bits 0..4 to 3..7 // 2
and w2, w2, 7 // preserve only 1st 3 bits in *byte // 3
orr w2, w2, w1 // or in value into *byte // 4
strb w2, [x0] // 5
ret // 6
Line 2
uses the instruction ubfiz
which means Unsigned Bit Field Insert
Zeroed. This instruction:
-
Zeros out a new copy of
value
(w1
), the destination and -
Copies 5 bits starting at bit 0 of the old
value
to bits 3 through 7 in the new version ofvalue
.
This one instruction does the work of lines 2, 3 and 4
in the naive version
of the assembly language.
Line 3
of the new assembly language replaces lines 4, 5 and 6
in the naive.
This works because the enlightened human saw an easier way to zero out *byte
except for the first 3 bits (where a
and b
live).
The remainder is as expected.
In this chapter we saw was life was like without bit fields. We saw that we had to implement our own bit bashing functions to do things like:
-
Ensure parameters are in the right range
-
Shift values around to line up with their destination
-
Zero out destination fields
-
Or in new values, having been shifted to the right position
and more.
We brushed upon the idea that bit bashing and bit fields are critical to directly interfacing with hardware but are also useful in decreasing the size of data structures in memory and on disc.
In Computer Science there is an eternal tension between space and time. The following is a law:
If you want something to go faster, it will cost more memory.
If you want to save memory, what you're doing will take more time.
This law shows up here... recall the example of where we wanted to save
memory by collapsing 8 bool
into 1 byte? To save that memory we will
slow down because accessing the right bits takes a couple of instructions
where overwriting a bool
implemented as an int
takes just one
instruction.