Reusing translated code can remove part of a repeated launch. It does not eliminate every source of startup work or guarantee the same benefit for every application.
The default has an important exception
The 0.4.4 announcement enables compressed DynaCache by default with a 2GB limit, but excludes programs using BOX64_DYNAREC_DIRTY. It also introduces the Python configurator for application profiles. The documented box64 -dl command lists cache details; box64 -dc removes obsolete entries. The latter changes the cache, while the former is useful for inspection.
Translation is only one part of launch time
Our invented example splits a cold start into 3 seconds of translation and 7 seconds of other work: 10 seconds total. If a compatible cache replaces that translation with a 1-second load, the next launch takes 8 seconds, a 20% reduction. Those numbers are arithmetic, not Box64 measurements.
The remaining work can include library loading, file reads, application initialization and graphics preparation. DynaCache is not a GPU shader cache, even if both can make repeated launches feel quicker. Conflating them makes a slow startup difficult to diagnose.
Keep a known-working profile
Before removing an old override, record the current profile and the problem it originally addressed. Test one change at a time with the same program version and inputs. A newer default can make a workaround unnecessary, but a release-wide optimization does not prove that every local compatibility setting is obsolete.
Compare both cold and repeated starts, and check that the program still produces correct output or reaches the same playable state. Note the backend and host hardware. A demonstration on one ARM64 machine cannot rank ARM64, RISC-V and LoongArch as architectures when their CPUs and GPUs differ. The useful result is a repeatable improvement on the machine you actually use.
September 8: include the DIRTY cache exclusion, distinguish translation cache from shader cache, and replace blanket tuning advice with controlled comparisons.