The IA-32 EL used a two-phase (later three-phase) approach: initially it quickly translated every piece of code at a basic block level, adding certain instrumentation for detecting hot code; then hot code was dynamically optimized at a super-block level, and the optimized translated code replaced cold code on the fly.[2] Later interpretation engine was added that allowed to avoid altogether translation of code executed just a few times - cold non-optimized translation became thus the second phase, and hot optimized translation became the third phase. IA-32 Execution Layer supported self-modifying code, and could even optimize it quite well.
Part of the software is under the LGPL and part is under an Intel proprietary license.[3]