UseCompressedOops kick-in threshold different for G1 and CMS

I am ruing a 64 bit Java 1.8 Hotspot JVM by Oracle. I have been trying to wrap my head around a difference in behavior of JVM to kick in compressed object pointers when different GC mechanisms are used. For example:

$ java -XX:+UseConcMarkSweepGC -XX:+PrintFlagsFinal -Xms32766m -Xmx32766m

bool UseCompressedClassPointers        := true        {lp64_product}
bool UseCompressedOops                 := true        {lp64_product}

$ java -XX:+UseConcMarkSweepGC -XX:+PrintFlagsFinal -Xms32767m -Xmx32767m

bool UseCompressedClassPointers        = false        {lp64_product}
bool UseCompressedOops                 = false        {lp64_product}

$ java -XX:+UseG1GC -XX:+PrintFlagsFinal -Xms32736m -Xmx32736m

bool UseCompressedClassPointers        := true        {lp64_product}
bool UseCompressedOops                 := true        {lp64_product}

$ java -XX:+G1GC -XX:+PrintFlagsFinal -Xms32737m -Xmx32737m

bool UseCompressedClassPointers        = false        {lp64_product}
bool UseCompressedOops                 = false        {lp64_product}

I have tried changing a few other G1GC knobs but caot get it compressed pointer optimization to kick in for heap sizes above 32736 MB. I am trying to understand what controls this threshold for different GC algorithms.