Dumb question: what would be the cost of a compiler-based mitigation, ie. never using fused instructions (ie. complex instructions broken-down into several mic

Dumb question: what would be the cost of a compiler-based mitigation, ie. never using fused instructions (ie. complex instructions broken-down into several micro-instructions internally by the pipeline) but only micro-ops that won't use the u-ops cache ? This would have an obvious impact on the code size/cache efficiency, and on the number of registers used to store intermediate results, but how much would this kill performances, depending on the use-case ?