[Object][RISCV] Fix R_RISCV_SET6 and R_RISCV_SUB6 relocations resolution

[luismarques]

Summary

These relocations had a suspicious resolution logic, given their name. This patch
makes the resolution match the LLD one, which makes more sense.

[jrtc27]

Yes that looks rather better. Especially SET6 which was previously an ADD6 (although that aspect was likely harmless as nothing sensible would put something there just to be clobbered by a SET6).

[HsiangKai]

LGTM. Do you have any test case for this patch?

[luismarques]

In D70396#1751017, @HsiangKai wrote:

LGTM. Do you have any test case for this patch?

There is one now: D70541.

[HsiangKai]

Thanks for your patch.

llvm/lib/Object/RelocationResolver.cpp

static uint64_t resolveRISCV(RelocationRef R, uint64_t S, uint64_t A) {
   switch (R.getType()) {
   case ELF::R_RISCV_NONE:
     return A;
   case ELF::R_RISCV_32:
     return (S + RA) & 0xFFFFFFFF;
   case ELF::R_RISCV_64:
     return S + RA;
   case ELF::R_RISCV_SET6:
-    return (A + (S + RA)) & 0xFF;
+    return (A & 0xC0) | ((S + RA) & 0x3F);
   case ELF::R_RISCV_SUB6:
-    return (A - (S + RA)) & 0xFF;
+    return (A & 0xC0) | (((A & 0x3F) - (S + RA)) & 0x3F);
   case ELF::R_RISCV_ADD8:
     return (A + (S + RA)) & 0xFF;
   case ELF::R_RISCV_SUB8:
     return (A - (S + RA)) & 0xFF;
   case ELF::R_RISCV_ADD16:
     return (A + (S + RA)) & 0xFFFF;
   case ELF::R_RISCV_SUB16:
     return (A - (S + RA)) & 0xFFFF;