Relax a GOT load into a PC-relative address materialization

riscv-non-isa/riscv-elf-psabi-doc#397

elf/arch-riscv.cc

@@ -364,6 +364,32 @@ void InputSection<E>::apply_reloc_alloc(Context<E> &ctx, u8 *base) {
       break;
     }
     case R_RISCV_GOT_HI20:
+      if (ctx.arg.relax &&
+          sym.is_pcrel_linktime_const(ctx) &&
+          i + 2 < rels.size() &&
+          rels[i + 1].r_type == R_RISCV_PCREL_LO12_I &&
+          rels[i + 1].r_offset == rels[i].r_offset + 4 &&
+          file.symbols[rels[i + 1].r_sym]->value == r_offset &&
+          rels[i + 2].r_type == R_RISCV_RELAX) {
+        // AUIPC + LD is used to load a value from a GOT slot. If the
+        // value is actually a PC-relative link-time constant, we can
+        // rewrite AUIPC + LD with AUIPC + ADDI to eliminate the memory
+        // load.
+        i64 val = S + A - P;
+        i64 rd1 = get_rd(rel.r_offset);
+        i64 rd2 = get_rd(rel.r_offset + 4);
+
+        if ((i32)val == val && rd1 == rd2) {
+          // auipc <reg>, %hi20(val)
+          write_utype(loc, val);
+
+          // addi <reg>, <reg>, %lo12(val)
+          *(ul32 *)(loc + 4) = 0b0010011 | (rd1 << 15) | (rd1 << 7);
+          write_itype(loc + 4, val);
+          i += 2;
+          break;
+        }
+      }
       write_utype(loc, G + GOT + A - P);
       break;
     case R_RISCV_TLS_GOT_HI20: