General-Purpose Register Conventions

Register Status Use

GPR0 volatile In function prologs.
GPR1 dedicated Stack pointer.
GPR2 dedicated Table of Contents (TOC) pointer.
GPR3 volatile First word of a function's argument list;
 first word of a scalar function return.
GPR4 volatile Second word of a function's argument list;
 second word of a scalar function return.
GPR5 volatile Third word of a function's argument list.
GPR6 volatile Fourth word of a function's argument list.
GPR7 volatile Fifth word of a function's argument list.
GPR8 volatile Sixth word of a function's argument list.
GPR9 volatile Seventh word of a function's argument list.
GPR10 volatile Eighth word of a function's argument list.
GPR11 volatile In calls by pointer and as an environment pointer
 for languages that require it (for example, PASCAL).
GPR12 volatile For special exception handling required by certain
 languages and in glink code.
GPR13 reserved Reserved under 64-bit environment;
 not restored across system calls.
GPR14:GPR31 nonvolatile These registers must be preserved across
 a function call.

Vector Register Conventions

Register Status

VR0:V19 Volatile
VR20:VR31 Nonvolatile (extended ABI mode) their values are preserved
 across function calls

Addressing memory

There are many ways to reference data, to maintain support of
position-independent code the current implementations use GOT-indirect
addressing (Accessing data through the global offset table):
1. Define data in .data section
2. Load the address of data into register from the global offset table
   Use 32-bit offset (medium or large code model) to get maximum addressing 
   reach of 4 GB e.g. addis r7, r2, my_var@got@ha
                      ld r7, my_var@got@l(r7)
3. Use the address to load the value of data into register
   e.g. ld r3, 0(r7)
Refer to [2] for more information about referencing data

VSX instructions "lxvd2x/stxvd2x" are used to load and store data to
memory instead of VR instructions "lvx/stvx" as it produces a fewer
instructions "lvx/stvx" can be used to load/store data into storage
operands but additional instructions are needed to access unaligned
storage operands, refer to "6.4.1 Accessing Unaligned Storage Operands"
in [3] to see an example of accessing unaligned storage operands.
"lxvd2x/stxvd2x" can be used to load/store data into unaligned storage
operands but permuting is needed for loading and storing data in
little-endian mode VSX registers are defined with "X" suffix

Function Prologue

Big-endian systems only support ELFv1 ABI which requires the following
steps in the function prologue:
1. Write the "official procedure descriptor" in ".opd","aw" section
2. Write procedure description for .my_func in my_func label
3. Switch back to ".text" section for program code
4. Label the beginning of the code .my_func
Refer to [1] for more information
Little-endian systems are compatible with ELFv2 ABI, an example of
function prologue for ELFv2 ABI can be seen in [2]

[1] http://www.ibm.com/developerworks/linux/library/l-powasm1.html
[2] https://openpowerfoundation.org/?resource_lib=64-bit-elf-v2-abi-specification-power-architecture
[3] https://openpowerfoundation.org/?resource_lib=ibm-power-isa-version-2-07-b