LoongArch hi64 calculation error and PCALA should set plt if sym is imported

[MQ-mengqing]

I find when I test mold in ArchLinux, something wrongs.

1,hi64 calculation error

The [63:32] bits should be PC-relative, too. The current codes seems only return val-relative.

diff --git a/elf/arch-loongarch.cc b/elf/arch-loongarch.cc
index b5bd0878..4bc60f57 100644
--- a/elf/arch-loongarch.cc
+++ b/elf/arch-loongarch.cc
@@ -66,12 +66,11 @@ static u64 hi64(u64 val, u64 pc) {
   //
   // Compensating all the sign-extensions is a bit complicated.
   bool x = val & 0x800;
-  bool y = (page(val + 0x800) - page(pc)) & 0x8000'0000;
-
-  if (x && !y)
-    return val - 0x1'0000'0000;
-  if (!x && y)
-    return val + 0x1'0000'0000;
+  val = page(val) - page(pc);
+  if (x)
+    val += 0x1000 - 0x1'0000'0000;
+  if (val & 0x8000'0000)
+    val += 0x1'0000'0000;
   return val;
 }

[1] https://github.com/llvm/llvm-project/blob/main/lld/ELF/Arch/LoongArch.cpp#L158
[2] https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=bfd/elfnn-loongarch.c;#l2304

2, PCALA should set plt if sym is imported

PCALA used for a long call in -mcmodel=extreme mode. The sequence is "la.pcrel $rxx, $ryy, sym; jirl $ra, $rxx, 0"
So if the sym is imported, we also need add NEED_PLT.

@@ -587,6 +586,7 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
       scan_dyn_absrel(ctx, sym, rel);
       break;
     case R_LARCH_B26:
+    case R_LARCH_PCALA_HI20:
       if (sym.is_imported)
         sym.flags |= NEEDS_PLT;
       break;
@@ -620,7 +620,6 @@ void InputSection<E>::scan_relocations(Context<E> &ctx) {
     case R_LARCH_ABS_LO12:
     case R_LARCH_ABS64_LO20:
     case R_LARCH_ABS64_HI12:
-    case R_LARCH_PCALA_HI20:
     case R_LARCH_PCALA_LO12:
     case R_LARCH_PCALA64_LO20:
     case R_LARCH_PCALA64_HI12:

[rui314]

Thanks. Should be fixed in the above commit.

[rui314]

Slightly off-topic, but as I pointed out several times, I think you guys want to define a "no check" version of R_LARCH_PCALA_HI20 relocation. Currently, the same relocation is used for the usual code model and the extreme code model, so we can't do overflow check for R_LARCH_PCALA_HI20, meaning that we can't report a relocation overflow error from the linker for that relocation.

I removed the relocation overflow check in the above commit. It's not ideal though.

[MQ-mengqing]

Thanks for you pointing out it. I have reported it to our internal group.
BTW, we had found that if PCALA_HI20 and PCALA64_{LO20,HI12} are not in the same 4k-page may cause relocation errors. (The edge case is (-2 + 4N)G ). I hope we may fix later by add relocations.