[lld] [lld][LoongArch] Relax call36/tail36: R_LARCH_CALL36 (PR #123576)

[WÁNG Xuěruì via llvm-commits]

================
@@ -830,6 +832,37 @@ static void relaxPCHi20Lo12(Ctx &ctx, const InputSection &sec, size_t i,
   remove = 4;
 }
 
+// Relax code sequence.
+// From:
+//   pcaddu18i $ra, %call36(foo)
+//   jirl $ra, $ra, 0
+// To:
+//   b/bl foo
+static void relaxCall36(Ctx &ctx, const InputSection &sec, size_t i,
+                        uint64_t loc, Relocation &r, uint32_t &remove) {
+  const uint64_t dest =
+      (r.expr == R_PLT_PC ? r.sym->getPltVA(ctx) : r.sym->getVA(ctx)) +
+      r.addend;
+
+  const int64_t displace = dest - loc;
+  // Check if the displace aligns 4 bytes or exceeds the range of b[l].
+  if ((displace & 0x3) != 0 || !isInt<28>(displace))
+    return;
+
+  const uint32_t nextInsn = read32le(sec.content().data() + r.offset + 4);
----------------
xen0n wrote:

> After further consideration and revisiting the manual again(https://github.com/loongson/la-abi-specs/blob/release/laelf.adoc), the generated `pcaddu18i` and `jirl` by GCC and Clang must be adjacent.
> 
> As for the interoperability you mentioned, `lld` handles instruction sequences generated by `gas`. And `gas` should not generate non-contiguous instruction sequences either. Therefore, we think that adding a check in `lld` is unnecessary.
> 
> Additionally, if `jirl` check is added here, other relaxations (such as R_LARCH_{PCALA,GOT_PC}_{HI20,LO12}, etc) would also need to check the instructions, which contradicts the principle of avoiding instruction checks unless absolutely necessary.

I'm not objecting to the current revision *per se*, but rather *behavior consistency*, no matter whether the input is well-formed or not -- the so-called "bug-for-bug compatibility". This means:

* either we add a check for `jirl` here in LLD and exactly replicate `ld.bfd` behavior, or
* we leave the code as-is, and remove the check at `ld.bfd` side, with your comment here as justification,

so eventually we're going to have 2 consistent implementations of LoongArch ELF psABI and happy users. What do you think here?

https://github.com/llvm/llvm-project/pull/123576