Here are a few lines from the map file for a system built with --opt_level=2 ...

LINKER GENERATED COPY TABLES

__TI_cinit_table @ 00011234 records: 2, size/record: 8, table size: 16
	.data: load addr=00011248, load size=000002f0 bytes, run addr=20003fa8, run size=000002e8 bytes, compression=copy
	.bss: load addr=00011538, load size=00000008 bytes, run addr=20004290, run size=00000170 bytes, compression=zero_init

Here are the same lines from the map file for the same system built the same way, except with --opt_level=4 ...

LINKER GENERATED COPY TABLES

__TI_cinit_table @ 00010744 records: 2, size/record: 8, table size: 16
	.data: load addr=00010758, load size=0000038c bytes, run addr=20003fa8, run size=00000384 bytes, compression=copy
	.bss: load addr=00010ae8, load size=00000008 bytes, run addr=20004330, run size=000000d0 bytes, compression=zero_init

Compare the load size numbers. Notice how --opt_level=4 requires more memory. This is because some variables change how they are initialized. Under --opt_level=2, these variables are in .bss and are filled with 0 by the compression=zero_init method. This is a method which takes very little memory to implement. Under --opt_level=4, some of these variables change from being in .bss to .data. These variables are no longer initialized by zero fill, but by being explicitly initialized to 0 with entries in .cinit. Such entries take more memory.

The test case which demonstrates the problem, and how to build it, requires a long description. For brevity, it is not included here, but in other parts of this record which are not publicly exposed.