AMD x86 manual Minimize Floating-Point-to-Integer Conversions, Example 1 Fast

AMD Athlon™ Processor x86 Code Optimization 22007E/0 — November 1999

Minimize Floating-Point-to-Integer Conversions

C++, C, and Fortran define floating-point-to-integer conversions as truncating. This creates a problem because the active rounding mode in an application is typically round-to-nearest- even. The classical way to do a double-to-int conversion therefore works as follows:

Example 1 (Fast):

SUB	[I], EDX		;trunc(X)=rndint(X)-correction
FLD	QWORD PTR	[X]	;load double to be converted
FSTCW	[SAVE_CW]		;save current FPU control word
MOVZX	EAX, WORD	PTR[SAVE_CW];retrieve control word
OR	EAX, 0C00h		;rounding control field = truncate
MOV	WORD PTR [NEW_CW], AX ;new FPU control word
FLDCW	[NEW_CW]		;load new FPU control word
FISTP	DWORD PTR	[I]	;do double->int conversion
FLDCW	[SAVE_CW]		;restore original control word

The AMD Athlon processor contains special acceleration hardware to execute such code as quickly as possible. In most situations, the above code is therefore the fastest way to perform floating-point-to-integer conversion and the conversion is compliant both with programming language standards and the IEEE-754 standard.

According to the recommendations for inlining (see “Always Inline Functions with Fewer than 25 Machine Instructions” on page 72), the above code should not be put into a separate subroutine (e.g., ftol). It should rather be inlined into the main code.

In some codes, floating-point numbers are converted to an integer and the result is immediately converted back to floating-point. In such cases, the FRNDINT instruction should be used for maximum performance instead of FISTP in the code above. FRNDINT delivers the integral result directly to an FPU register in floating-point form, which is faster than first using FISTP to store the integer result and then converting it back to floating-point with FILD.

If there are multiple, consecutive floating-point-to-integer conversions, the cost of FLDCW operations should be minimized by saving the current FPU control word, forcing the

100 Minimize Floating-Point-to-Integer Conversions