We recently heavily reduced the memory consumption of the callstack representation within the compiler however once compilation is complete we then use the old inefficient representation of the callstacks.

If we pull up the memory flamegraph of executing rollup-base-public with a cached artifact, we can then see that 96% of the peak memory usage comes from just deserializing the program artifact. Of this 1.6GB (52.6%) comes from the debug info as opposed to 1.3GB of actual bytecode (43.6%).

We should perform a similar optimization to the debug data stored in the program artifacts to minimise their memory footprint.