Skip to content
Francois Bedard edited this page Apr 25, 2017 · 9 revisions

Working with small data section

DISCLAIMER: These notes are applicable to upcoming ARC GCC releases ONLY, starting with 2017.09 release.

Small data available access range

Data Type Range #Elements Size
char [-256,255] 512 512 bytes
short [-256,510] 383 766 bytes
int [-256,1020] 319 1276 bytes

The lower limit depends on the possibility to access byte-aligned datum, hence, it is hard connected to the range of s9 short immediate (i.e., -256). Any other access can be done using address-scaling feature of the load/store instructions.

The number of elements which we can fit in sdata section highly depends on the data alignment properties. For example if we use only 4 byte datum, 1 byte aligned, we can fit up to 128 elements in the section.

Sdata and address scaling mode discussion

To increase the access range for small data area, the compiler will use scaled addresses whenever it is possible. Thus, we can extend theoretically this range to [-1024,1020] for 32-bit data (e.g., int) if type aligned (i.e., 4-bytes). However, as we cannot be 100% sure we address only 32-bit/4-byte aligned data, we need to consider the worst case which is byte-aligned data. Thus, the effective range is [-256,255], with possibilities to access 16-bit aligned data (e.g., short) up to 510, and 32-bit aligned data (e.g., int or long long) up to 1020. While the lower limit remains set to -256. This is because we set the linker script defined variable __SDATA_BEGIN__ with an offset of 0x100. However, this rule can be overwritten by using a custom linker script.

Controlling what goes into SDATA segment.

Automatic for global data

Global data smaller than a given number in bytes can be placed into the sdata section. The number of bytes can be controlled via -G option.

For ARC, by default this number is set to 8 whenever we have double load/store operations available (i.e., ARC HS architecture), otherwise to 4.

For example, a 8 bytes setting will allow us to place into sdata the following variables:

char gA[8];
short gB[4];
int gC[2];
long long gD;

Notable exceptions:

  • Volatile global data will not be placed into sdata section when -mno-volatile-cache option is used;
  • Strings and functions never end in small data area;
  • Weak variables as well not;
  • No constant will end in small data area as those one, we would like to place into ROM.

Using named sections

Another way to control which data goes into small data area is to use named sections like this:

int a __attribute__((section (".sdata"))) = 1;
int b __attribute__((section (".sbss")));

Variables a and b will go into sdata/sbss sections without checking the data type size against the -G value. Thus, we can always control which data is accesses via gp register by setting -G0 and using named sections to the desired ones. Using named section we can also place into sdata static data.