question about support relaxation of loongarch about mold

[rui314]

Could you share more details about your problem? What are the relocations you are trying to handle?

[ywgrit]

Could you share more details about your problem? What are the relocations you are trying to handle?

Sure.
I'd like to replace the original call to InputSection::apply_reloc_[no]alloc with InputSection::apply_relax. Where apply_relax is a new member function added to the InputSection class that handles relaxation including pcalau12i/addi -> pcaddi.
The copy_chunks (where the relaxation occurs) is followed by apply_relocate(new routine), which invokes OutputSection::apply_relocate to relocate all InputSections.

In the current hypothesis, the calling relationship is as follows

copy_chunks->chunk.copy_buf->OutputSection::write_to->InputSection::write_to->InputSection::apply_relax

apply_relocate->OutputSection::apply_relocate->InputSection::apply_reloc_[no[alloc 

All of these changes are made so that all InputSections have been relocated before relaxation.

[rui314]

Have you read https://github.com/rui314/mold/blob/main/elf/arch-riscv.cc? I believe LoongArch's relaxation is modeled after the RISC-V psABI's, so maybe you can do the same as RISC-V?

[ywgrit]

Have you read https://github.com/rui314/mold/blob/main/elf/arch-riscv.cc? I believe LoongArch's relaxation is modeled after the RISC-V psABI's, so maybe you can do the same as RISC-V?

Thanks.
I've searched for this file before based on "arg.relax", but there is only one case processed in riscv and that case doesn't delete the instruction bytes, so that case doesn't cause the symval to change. But for the case of relaxation that deletes instructions(i.e., pcalau12i/addi -> pcaddi in loongarch), it may cause some of the symbol values to be changed. If relocation is done where the symbols are used before these relaxations are applied, then after the relaxation is applied, the value of the symbols are changed and the relocation is invalidated. That's why maybe we need to do the relaxation before the relocation.

[rui314]

We do support the RISC-V relaxations that remove bytes. Please take a look at riscv_resize_sections in that file.

[ywgrit]

We do support the RISC-V relaxations that remove bytes. Please take a look at riscv_resize_sections in that file.

Sorry for my omission. Thank you, Rui.