[PATCH] D137384: [MC][LoongArch] Fix needsRelocateWithSymbol() implementation

[Jinyang He via Phabricator via llvm-commits]

MQ-mengqing added a comment.

In D137384#3916546 <https://reviews.llvm.org/D137384#3916546>, @tangyouling wrote:

> In D137384#3910952 <https://reviews.llvm.org/D137384#3910952>, @MQ-mengqing wrote:
>
>> In D137384#3907086 <https://reviews.llvm.org/D137384#3907086>, @xry111 wrote:
>>
>>> Hmm, BFD linker does not support `R_LARCH_GOT_PC_{HI12,LO20}` with a non-zero addend:
>>>
>>>   case R_LARCH_GOT_PC_HI20:
>>>   case R_LARCH_GOT_HI20:
>>>     /* Calc got offset.  */
>>>       {
>>>         unresolved_reloc = false;
>>>         BFD_ASSERT (rel->r_addend == 0);
>>>
>>> Do we need to make a custom implementation which returns false for `R_LARCH_32_PCREL`, `R_LARCH_PCALA*`, and `R_LARCH_B*`, true for others?
>>
>> The procedure happens before linking. AFAIK, the way of using sec+offset is like RELA while the way of using sym is like REL (, of course it seems the REL is a subset of the RELA).
>> Relocation symbols that can replace sym with sec+offset are generally local and its offset from sec can be determined at assembly procedure.
>> If relax which may decrease instructions is not used, the offset would not be changed later. So it is different from the Relocation of RISCV. Replacing the sym with sec+offset effectively reduces the number of syms.
>> The GOT method is used to access global symbols mostly, can they enter this function?
>
> Yes, it happens before linking.
>
> I found that only `R_LARCH_32`, `R_LARCH_64`, ` R_LARCH_32_PCREL`, `R_LARCH_PCALA*`, and `R_LARCH_B*` enter the function at compile time (some others like GOT don't enter it).
> Modified it to the following, which also passed the test (when testing asan),
>
>   bool needsRelocateWithSymbol(const MCSymbol &Sym,
>                                unsigned Type) const override {
>     switch (Type) {
>     default:
>       return true;
>   
>     case ELF::R_LARCH_32:
>     case ELF::R_LARCH_64:
>     case ELF::R_LARCH_B16:
>     case ELF::R_LARCH_B21:
>     case ELF::R_LARCH_B26:
>     case ELF::R_LARCH_PCALA_HI20:
>     case ELF::R_LARCH_PCALA_LO12:
>     case ELF::R_LARCH_PCALA64_LO20:
>     case ELF::R_LARCH_PCALA64_HI12:
>     case ELF::R_LARCH_32_PCREL:
>       return false;
>     }
>   }
>
> Which do you think is more acceptable, using a generic implementation or using a custom implementation?



In D137384#3916546 <https://reviews.llvm.org/D137384#3916546>, @tangyouling wrote:

> In D137384#3910952 <https://reviews.llvm.org/D137384#3910952>, @MQ-mengqing wrote:
>
>> In D137384#3907086 <https://reviews.llvm.org/D137384#3907086>, @xry111 wrote:
>>
>>> Hmm, BFD linker does not support `R_LARCH_GOT_PC_{HI12,LO20}` with a non-zero addend:
>>>
>>>   case R_LARCH_GOT_PC_HI20:
>>>   case R_LARCH_GOT_HI20:
>>>     /* Calc got offset.  */
>>>       {
>>>         unresolved_reloc = false;
>>>         BFD_ASSERT (rel->r_addend == 0);
>>>
>>> Do we need to make a custom implementation which returns false for `R_LARCH_32_PCREL`, `R_LARCH_PCALA*`, and `R_LARCH_B*`, true for others?
>>
>> The procedure happens before linking. AFAIK, the way of using sec+offset is like RELA while the way of using sym is like REL (, of course it seems the REL is a subset of the RELA).
>> Relocation symbols that can replace sym with sec+offset are generally local and its offset from sec can be determined at assembly procedure.
>> If relax which may decrease instructions is not used, the offset would not be changed later. So it is different from the Relocation of RISCV. Replacing the sym with sec+offset effectively reduces the number of syms.
>> The GOT method is used to access global symbols mostly, can they enter this function?
>
> Yes, it happens before linking.
>
> I found that only `R_LARCH_32`, `R_LARCH_64`, ` R_LARCH_32_PCREL`, `R_LARCH_PCALA*`, and `R_LARCH_B*` enter the function at compile time (some others like GOT don't enter it).
> Modified it to the following, which also passed the test (when testing asan),
>
>   bool needsRelocateWithSymbol(const MCSymbol &Sym,
>                                unsigned Type) const override {
>     switch (Type) {
>     default:
>       return true;
>   
>     case ELF::R_LARCH_32:
>     case ELF::R_LARCH_64:
>     case ELF::R_LARCH_B16:
>     case ELF::R_LARCH_B21:
>     case ELF::R_LARCH_B26:
>     case ELF::R_LARCH_PCALA_HI20:
>     case ELF::R_LARCH_PCALA_LO12:
>     case ELF::R_LARCH_PCALA64_LO20:
>     case ELF::R_LARCH_PCALA64_HI12:
>     case ELF::R_LARCH_32_PCREL:
>       return false;
>     }
>   }
>
> Which do you think is more acceptable, using a generic implementation or using a custom implementation?

Hi, All,

Frankly the custom implementation is more appreciated at present. Although I want to use the generic implementation but binutils cannot guarantee the user action as follows,

  Code:
      la.got $r0, sym
      sym:
  Reloc:
      0000000000000000  000000010000004b R_LARCH_GOT_PC_HI20    0000000000000000 .text + 8
      0000000000000004  000000010000004c R_LARCH_GOT_PC_LO12    0000000000000000 .text + 8

When sym is local, binutils-ld failed here. As relax may be added in the future, thus, I think it's appropriate to use custom implementation.

Other ideas?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D137384/new/

https://reviews.llvm.org/D137384