AMD x86 manual Take Advantage of Write Combining, Use 3DNow! Instructions

AMD Athlon™ Processor x86 Code Optimization

Avoid Load-Execute Floating-Point Instructions with Integer Operands

22007E/0 — November 1999

	Do not use load-execute floating-point instructions with integer
TOP	operands. The floating-point load-execute instructions with
	integer operands are VectorPath and generate two OPs in a
✩
	cycle, while the discrete equivalent enables a third DirectPath
	instruction to be decoded in the same cycle.
Take Advantage of Write Combining
	This guideline applies only to operating system, device driver,
TOP	an d B IO S p rog ram m ers . I n o rd e r t o i m p rove sy st em
	performance, the AMD Athlon processor aggressively combines
✩
	multiple memory-write cycles of any data size that address
	locations within a 64-byte cache line aligned write buffer.
	See Appendix C, “Implementation of Write Combining” on
	page 155 for more details.
Use 3DNow!™ Instructions
	Unless accuracy requirements dictate otherwise, perform
TOP	floating-point computations using the 3DNow! instructions
	instead of x87 instructions. The SIMD nature of 3DNow!
✩
	instructions achieves twice the number of FLOPs that are
	achieved through x87 instructions. 3DNow! instructions also
	provide for a flat register file instead of the stack-based
	approach of x87 instructions.
	See Table 23 on page 217 for a list of 3DNow! instructions. For
	information about instruction usage, see the 3DNow!™
	Technology Manual, order# 21928.
Avoid Branches Dependent on Random Data
	Avoid data-dependent branches around a single instruction .
TOP	Data-dependent branches acting upon basically random data
	can cause the branch prediction logic to mispredict the branch
✩
	about 50% of the time. Design branch-free alternative code

sequences, which results in shorter average execution time.

See “Avoid Branches Dependent on Random Data” on page 57 for more details.

10 Group II Optimizations —Secondary Optimizations