[SPARC64] Turn the delay slot of R_SPARC_TLS_LDM_CALL to nop if it sets %o0

GCC might put the argument-setting instruction to R_SPARC_TLS_LDM_CALL
in its delay slot as an unmarked regular instruction.
If we're optimizing the call into a nop then we need to detect and
turn the delay slot into nop too as needed.

Author: koachan

src/arch-sparc64.cc

@@ -429,6 +429,13 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
         *(ub32 *)loc |= bits(addr + A - P, 31, 2);
       } else {
         *(ub32 *)loc = 0x0100'0000; // nop
+
+        // If the argument to __tls_get_addr is set in the delay slot of the call,
+        // turn the delay slot instruction into a nop too.
+        u32 ds_rd = bits(*(ub32 *)(loc + 4), 29, 25);
+        if (ds_rd == 8) { // rd is %o0
+          *(ub32 *)(loc + 4) = 0x0100'0000; // nop
+        }
       }
       break;
     case R_SPARC_TLS_LDO_HIX22: