Why would you combine two 32 bit integers into a 64 bit integer?

Question

Recently I saw how a compiler combined two 32 bit integers which were values of properties of a class and stored them as a 64 bit integer. My question now is, why is this done? What advantages are there when combining integers?

for example if we had the following properties of a class

class FooBar { int x = 1; int y = 100; }

so instead of

i32 = 00000001 i32 = 01100100

We get:

i64 = 0000000101100100

Why would you combine them?

My guess is that the register uses 64 bits. It's using just one register to store both instead of using two registers. — R Sahu
– R Sahu, Commented Apr 22, 2017 at 17:55
CPUs don't generally have types. Something like int is a property of your program, not of the resulting machine code. It's not really appropriate to call this "combining", because the before and after are parts of completely different domains. — Kerrek SB
– Kerrek SB, Commented Apr 22, 2017 at 18:05
If the machine uses 64 bits register & mem system, then it's kind of a waste of space to use 32 bits variables, hence the merge of the two variable. The draw back, however is that operations on these variables will take a bit longer, as they'll have to be AND'ed with 0x00FF and 0xFF00 (+ shift) when you want to work with them, most probably — Maliafo
– Maliafo, Commented Apr 22, 2017 at 18:06
Can you give an example showing the actual "combining" you are referring to? For example, if you noted it in the generated assembly, provide the assembly. — BeeOnRope
– BeeOnRope, Commented Apr 22, 2017 at 19:14

1201ProgramAlarm · Accepted Answer · 2017-04-22 20:21:27Z

The existing (as I write this) answer and comments, while partially correct, miss the point of this optimization. It is to replace two instructions (working with 32-bit data) with one instruction (working with 64-bit data). This results in a very slight reduction in code size and probably execution time.

The compiler is initializing both variables with one 64-bit instruction (since they share consecutive memory addresses). The variables are separate, and will be accessed separately. No shifting or masking is needed.

This is frequently seen in constructors when many members are initialized. A frequent case is with zero initialization, where the compiler will zero out a register then use that one value to initialize multiple members, combining writes to consecutive memory addresses with a larger single write (for example, by writing a 16-bit short zero value instead of two 8-bit ones).

> No shifting or masking is needed. I disagree with that point. You must alias the memory space to use 4 byte alignment. See this link: software.intel.com/en-us/articles/… That said, I think you are correct that aliasing an address as 32 or 64 is near-"free" regarding performance. Nice answer.
@Sam The optimizer sees a 4 byte write to address A, followed immediately by a 4 byte write to address (A+4). These are combined into an 8 byte write to address A. No other changes to the code are necessary or made. If there were issues with alignment or aliasing this optimization would not be made.

Sam · Accepted Answer · 2017-04-22 18:46:53Z

I believe you're observing an optimization. Intel instructions such as PADDSW assume multiple packed operands.

https://en.wikipedia.org/wiki/X86_instruction_listings

There are also benefits in only using 1 entry in a 64bit architecture cache.

There is a cost to unpack if you only want one of the values, but I suspect whatever code optimizer is running estimates there are better savings pack the values.

It used to be normal to align all members of a C struct onto a word boundary. That is a single char and an int would not be packed, but aligned to the word size of the machine. So, struct { char, int} would have a sizeof(..) of 8 bytes. I'm guessing that situation flipped?

Very interesting.

You probably heared of webassembly right? They are doing it there. What you are saying is really interesting
The PAD… instructions are for SIMD. They packed values are all of the same type. There is no cost to unpack, as the individual values still have normal addresses.

Collectives™ on Stack Overflow

Why would you combine two 32 bit integers into a 64 bit integer?

2 Answers 2

4 Comments

2 Comments

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

4 Comments

2 Comments

Related