4.2.2Load-Execute Floating-Point Instructions with Floating-Point Operands
Optimization
❖When performing floating-point computations using floating-point (not integer) source operands, use load-execute instructions instead of discrete load and execute instructions.
Application
This optimization applies to:
•32-bit software
•64-bit software
Rationale
Using load-execute floating-point instructions that take floating-point operands improves performance for the following reasons:
•Denser code allows more work to be held in the instruction cache.
•Denser code generates fewer internal macro-ops, allowing the floating-point scheduler to hold more work, which increases the chances of extracting parallelism from the code.
Example
Avoid code like this, which uses discrete load and execute instructions:
movss xmm0, [float_var1] movss xmm12, [float_var2] mulss xmm0, xmm12
Instead, use code like this, which uses a load-execute floating-point instruction:
movss xmm0, [float_var1] mulss xmm0, [float_var2]
4.2.3Load-Execute Floating-Point Instructions with Integer Operands
Optimization
❖Avoid x87 load-execute floating-point instructions that take integer operands (FIADD, FICOM, FICOMP, FIDIV, FIDIVR, FIMUL, FISUB, and FISUBR). When performing floating-point computations using integer source operands, use discrete load (FILD) and execute instructions instead.
