Intel® 64 and IA-32 Architectures Optimization Reference Manual

GENERAL OPTIMIZATION GUIDELINESThere are basically three ways to reduce the impact of overflow/underflow situationswith x87 FPU code:• Choose floating-point data types that are large enough to accommodate resultswithout generating arithmetic overflow and underflow exceptions.• Scale the range of operands/results to reduce as much as possible the number ofarithmetic overflow/underflow situations.• Keep intermediate results on the x87 FPU register stack until the final resultshave been computed and stored in memory. Overflow or underflow is less likelyto happen when intermediate results are kept in the x87 FPU stack (this isbecause data on the stack is stored in double extended-precision format andoverflow/underflow conditions are detected accordingly).• Denormalized floating-point constants (which are read-only, and hence neverchange) should be avoided and replaced, if possible, with zeros of the same sign.3.8.2.3 Floating-point Exceptions in SSE/SSE2/SSE3 CodeMost special situations that involve masked floating-point exceptions are handledefficiently in hardware. When a masked overflow exception occurs while executingSSE/SSE2/SSE3 code, processor hardware can handles it without performancepenalty.Underflow exceptions and denormalized source operands are usually treatedaccording to the IEEE 754 specification, but this can incur significant performancedelay. If a programmer is willing to trade pure IEEE 754 compliance for speed, twonon-IEEE 754 compliant modes are provided to speed situations where underflowsand input are frequent: FTZ mode and DAZ mode.When the FTZ mode is enabled, an underflow result is automatically converted to azero with the correct sign. Although this behavior is not compliant with IEEE 754, it isprovided for use in applications where performance is more important than IEEE 754compliance. Since denormal results are not produced when the FTZ mode is enabled,the only denormal floating-point numbers that can be encountered in FTZ mode arethe ones specified as constants (read only).The DAZ mode is provided to handle denormal source operands efficiently whenrunning a SIMD floating-point application. When the DAZ mode is enabled, inputdenormals are treated as zeros with the same sign. Enabling the DAZ mode is theway to deal with denormal floating-point constants when performance is the objective.If departing from the IEEE 754 specification is acceptable and performance is critical,run SSE/SSE2/SSE3 applications with FTZ and DAZ modes enabled.NOTEThe DAZ mode is available with both the SSE and SSE2 extensions,although the speed improvement expected from this mode is fullyrealized only in SSE code.3-80