Intel® 64 and IA-32 Architectures Optimization Reference Manual

APPENDIX DSTACK ALIGNMENTThis appendix details on the alignment of the stacks of data for Streaming SIMDExtensions and Streaming SIMD Extensions 2.D.1 STACK FRAMESThis section describes the stack alignment conventions for both ESP-based (normal),and EDP-based (debug) stack frames. A stack frame is a contiguous block of memoryallocated to a function for its local memory needs. It contains space for the function’sparameters, return address, local variables, register spills, parameters needing to bepassed to other functions that a stack frame may call, and possibly others. It is typicallydelineated in memory by a stack frame pointer (ESP) that points to the base ofthe frame for the function and from which all data are referenced via appropriateoffsets. The convention on Intel 64 and IA-32 is to use the ESP register as the stackframe pointer for normal optimized code, and to use EDP in place of ESP when debuginformation must be kept. Debuggers use the EDP register to find the informationabout the function via the stack frame.It is important to ensure that the stack frame is aligned to a 16-byte boundary uponfunction entry to keep local __m128 data, parameters, and XMM register spill locationsaligned throughout a function invocation.The Intel C++ Compiler for Win32*Systems supports conventions presented here help to prevent memory referencesfrom incurring penalties due to misaligned data by keeping them aligned to 16-byteboundaries. In addition, this scheme supports improved alignment for __m64 anddouble type data by enforcing that these 64-bit data items are at least eight-bytealigned (they will now be 16-byte aligned).For variables allocated in the stack frame, the compiler cannot guarantee the base ofthe variable is aligned unless it also ensures that the stack frame itself is 16-bytealigned. Previous software conventions, as implemented in most compilers, onlyensure that individual stack frames are 4-byte aligned. Therefore, a function calledfrom a Microsoft-compiled function, for example, can only assume that the framepointer it used is 4-byte aligned.Earlier versions of the Intel C++ Compiler for Win32 Systems have attempted toprovide 8-byte aligned stack frames by dynamically adjusting the stack framepointer in the prologue of main and preserving 8-byte alignment of the functions itcompiles. This technique is limited in its applicability for the following reasons:• The main function must be compiled by the Intel C++ Compiler.• There may be no functions in the call tree compiled by some other compiler (asmight be the case for routines registered as callbacks).• Support is not provided for proper alignment of parameters.D-1