Intel® 64 and IA-32 Architectures Optimization Reference Manual

GENERAL OPTIMIZATION GUIDELINEStranscendental performance with these techniques by choosing the desired numericprecision and the size of the look-up table, and by taking advantage of theparallelism of the SSE and the SSE2 instructions.3.8.2 Floating-point Modes and ExceptionsWhen working with floating-point numbers, high-speed microprocessors frequentlymust deal with situations that need special handling in hardware or code.3.8.2.1 Floating-point ExceptionsThe most frequent cause of performance degradation is the use of masked floatingpointexception conditions such as:• arithmetic overflow• arithmetic underflow• denormalized operandRefer to Chapter 4 of Intel® 64 and IA-32 Architectures Software Developer’sManual, Volume 1, for definitions of overflow, underflow and denormal exceptions.Denormalized floating-point numbers impact performance in two ways:• directly when are used as operands• indirectly when are produced as a result of an underflow situationIf a floating-point application never underflows, the denormals can only come fromfloating-point constants.User/Source Coding Rule 19. (H impact, ML generality) Denormalizedfloating-point constants should be avoided as much as possible.Denormal and arithmetic underflow exceptions can occur during the execution of x87instructions or SSE/SSE2/SSE3 instructions. Processors based on Intel NetBurstmicroarchitecture handle these exceptions more efficiently when executingSSE/SSE2/SSE3 instructions and when speed is more important than complying withthe IEEE standard. The following paragraphs give recommendations on how to optimizeyour code to reduce performance degradations related to floating-point exceptions.3.8.2.2 Dealing with floating-point exceptions in x87 FPU codeEvery special situation listed in Section 3.8.2.1, “Floating-point Exceptions,” is costlyin terms of performance. For that reason, x87 FPU code should be written to avoidthese situations.3-79