If I understand what you are trying to do, then I suggest to delay the decision until the emit stage, when the offset is known. So instead of deciding early in loadConLNodesTuple_create, have it create a "smart" loadConLNode that can emit either addis+ld or *ld*, depending on the offset. dl On 2/13/26 4:52 AM, Suchismith Roy wrote:
Hi All
I am currently working on JDK-8216554 , which is related to optimise loading of constants in c2 compilation by determining the size of TOC in post_alloc stage. This is for PowerPC Currently there is a default code that emits combination of addis+ld instruction even in possible cases where only ld instruction can be emitted. There is an alternate code based on small size of TOC that emits only ld, but we are not taking it.
I tried to do a gdb debug, i am unable to get the constant table size at post_alloc stage. It is only after the emit stage we get the size of TOC.
Does that mean we cannot ever get accurate size in post_alloc stage ?
I also tried a test case and forced the c2 compiler emit ld only.
Interestingly, the global size of TOC was large and calculated much later, and macroassembler puts in addis+addi .
So the case of large offset, even if miscalculated in post_alloc stage , is handled in some way.
Does that mean we can just keep the conservative path of emitting Addis+addi ? Or maybe we can even remove the helper function in post_alloc stage.
Open to your suggestions. Will provide a few debug logs too as per your request.
Thanks Suchismith Roy