diff --git a/include/n64recomp.h b/include/n64recomp.h index 66ce06b..7c04c2e 100644 --- a/include/n64recomp.h +++ b/include/n64recomp.h @@ -86,6 +86,7 @@ namespace N64Recomp { bool executable = false; bool relocatable = false; // TODO is this needed? relocs being non-empty should be an equivalent check. bool has_mips32_relocs = false; + uint32_t gp_ram_addr = 0; }; struct ReferenceSection { diff --git a/src/analysis.cpp b/src/analysis.cpp index 5dfd955..e425ffb 100644 --- a/src/analysis.cpp +++ b/src/analysis.cpp @@ -16,15 +16,19 @@ struct RegState { uint32_t prev_addiu_vram; uint32_t prev_addu_vram; uint8_t prev_addend_reg; + // offset of lw rt,offset(gp) + uint32_t prev_got_offset; bool valid_lui; bool valid_addiu; bool valid_addend; + bool valid_got_offset; // For tracking a register that has been loaded from RAM uint32_t loaded_lw_vram; uint32_t loaded_addu_vram; uint32_t loaded_address; uint8_t loaded_addend_reg; bool valid_loaded; + bool valid_gp_loaded; RegState() = default; @@ -33,10 +37,12 @@ struct RegState { prev_addiu_vram = 0; prev_addu_vram = 0; prev_addend_reg = 0; + prev_got_offset = 0; valid_lui = false; valid_addiu = false; valid_addend = false; + valid_got_offset = false; loaded_lw_vram = 0; loaded_addu_vram = 0; @@ -44,6 +50,7 @@ struct RegState { loaded_addend_reg = 0; valid_loaded = false; + valid_gp_loaded = false; } }; @@ -109,6 +116,24 @@ bool analyze_instruction(const rabbitizer::InstructionCpu& instr, const N64Recom temp.valid_addend = true; temp.prev_addend_reg = addend_reg; temp.prev_addu_vram = instr.getVram(); + } else if (reg_states[rs].valid_got_offset != reg_states[rt].valid_got_offset) { + // Track which of the two registers has the valid got offset state and which is the addend + int valid_got_offset_reg = reg_states[rs].valid_got_offset ? rs : rt; + int addend_reg = reg_states[rs].valid_got_offset ? rt : rs; + + // Copy the got offset reg's state into the destination reg, then set the destination reg's addend to the other operand + temp = reg_states[valid_got_offset_reg]; + temp.valid_addend = true; + temp.prev_addend_reg = addend_reg; + temp.prev_addu_vram = instr.getVram(); + } else if (((rs == (int)RegId::GPR_O32_gp) || (rt == (int)RegId::GPR_O32_gp)) + && reg_states[rs].valid_gp_loaded != reg_states[rt].valid_gp_loaded) { + // `addu rd, rs, $gp` or `addu rd, $gp, rt` after valid $gp load, this is the last part of a $gp relative + // jump table call. Keep the register state intact. + int valid_gp_loaded_reg = reg_states[rs].valid_gp_loaded ? rs : rt; + int gp_reg = reg_states[rs].valid_gp_loaded ? rt : rs; + + temp = reg_states[valid_gp_loaded_reg]; } else { // Check if this is a move check_move(); @@ -178,6 +203,19 @@ bool analyze_instruction(const rabbitizer::InstructionCpu& instr, const N64Recom temp.loaded_addu_vram = reg_states[base].prev_addu_vram; } } + // If the base register has a valid GOT offset and a valid addend before this, then this may be a load from a $gp relative jump table + else if (reg_states[base].valid_got_offset && reg_states[base].valid_addend) { + temp.valid_gp_loaded = true; + temp.loaded_lw_vram = instr.getVram(); + temp.loaded_address = imm; // This address is relative for now, we'll calculate the absolute address later + temp.loaded_addend_reg = reg_states[base].prev_addend_reg; + temp.loaded_addu_vram = reg_states[base].prev_addu_vram; + temp.prev_got_offset = reg_states[base].prev_got_offset; + } else if (base == (int)RegId::GPR_O32_gp) { + // lw from the $gp register implies a read from the GOT + temp.prev_got_offset = imm; + temp.valid_got_offset = true; + } reg_states[rt] = temp; break; case InstrId::cpu_jr: @@ -194,6 +232,18 @@ bool analyze_instruction(const rabbitizer::InstructionCpu& instr, const N64Recom reg_states[rs].loaded_lw_vram, reg_states[rs].loaded_addu_vram, instr.getVram(), + std::nullopt, + std::vector{} + ); + } else if (reg_states[rs].valid_gp_loaded) { + stats.jump_tables.emplace_back( + reg_states[rs].loaded_address, + reg_states[rs].loaded_addend_reg, + 0, + reg_states[rs].loaded_lw_vram, + reg_states[rs].loaded_addu_vram, + instr.getVram(), + reg_states[rs].prev_got_offset, std::vector{} ); } else if (reg_states[rs].valid_lui && reg_states[rs].valid_addiu && !reg_states[rs].valid_addend && !reg_states[rs].valid_loaded) { @@ -236,6 +286,21 @@ bool N64Recomp::analyze_function(const N64Recomp::Context& context, const N64Rec } } + const Section* section = &context.sections[func.section_index]; + + // Calculate absolute addresses for jump tables that are relative to $gp + for (size_t i = 0; i < stats.jump_tables.size(); i++) { + JumpTable& cur_jtbl = stats.jump_tables[i]; + + if (cur_jtbl.got_offset.has_value()) { + uint32_t gp_ram_addr = section->gp_ram_addr; + uint32_t gp_rom_addr = gp_ram_addr + func.rom - func.vram; + uint32_t got_word = byteswap(*reinterpret_cast(&context.rom[gp_rom_addr + cur_jtbl.got_offset.value()])); + + cur_jtbl.vram += section->ram_addr + got_word; + } + } + // Sort jump tables by their address std::sort(stats.jump_tables.begin(), stats.jump_tables.end(), [](const JumpTable& a, const JumpTable& b) @@ -262,6 +327,13 @@ bool N64Recomp::analyze_function(const N64Recomp::Context& context, const N64Rec // TODO same as above uint32_t rom_addr = vram + func.rom - func.vram; uint32_t jtbl_word = byteswap(*reinterpret_cast(&context.rom[rom_addr])); + + if (cur_jtbl.got_offset.has_value()) { + // $gp relative jump tables have values that are offsets from $gp, + // convert those to absolute addresses + jtbl_word += section->gp_ram_addr; + } + // Check if the entry is a valid address in the current function if (jtbl_word < func.vram || jtbl_word > func.vram + func.words.size() * sizeof(func.words[0])) { // If it's not then this is the end of the jump table diff --git a/src/analysis.h b/src/analysis.h index eafd1e7..c8a92e8 100644 --- a/src/analysis.h +++ b/src/analysis.h @@ -14,10 +14,11 @@ namespace N64Recomp { uint32_t lw_vram; uint32_t addu_vram; uint32_t jr_vram; + std::optional got_offset; std::vector entries; - JumpTable(uint32_t vram, uint32_t addend_reg, uint32_t rom, uint32_t lw_vram, uint32_t addu_vram, uint32_t jr_vram, std::vector&& entries) - : vram(vram), addend_reg(addend_reg), rom(rom), lw_vram(lw_vram), addu_vram(addu_vram), jr_vram(jr_vram), entries(std::move(entries)) {} + JumpTable(uint32_t vram, uint32_t addend_reg, uint32_t rom, uint32_t lw_vram, uint32_t addu_vram, uint32_t jr_vram, std::optional got_offset, std::vector&& entries) + : vram(vram), addend_reg(addend_reg), rom(rom), lw_vram(lw_vram), addu_vram(addu_vram), jr_vram(jr_vram), got_offset(got_offset), entries(std::move(entries)) {} }; struct AbsoluteJump { diff --git a/src/config.cpp b/src/config.cpp index 4471923..39b8b7b 100644 --- a/src/config.cpp +++ b/src/config.cpp @@ -476,6 +476,7 @@ bool N64Recomp::Context::from_symbol_file(const std::filesystem::path& symbol_fi std::optional vram_addr = el["vram"].template value(); std::optional size = el["size"].template value(); std::optional name = el["name"].template value(); + std::optional gp_ram_addr = el["gp"].template value(); if (!rom_addr.has_value() || !vram_addr.has_value() || !size.has_value() || !name.has_value()) { throw toml::parse_error("Section entry missing required field(s)", el.source()); @@ -488,6 +489,7 @@ bool N64Recomp::Context::from_symbol_file(const std::filesystem::path& symbol_fi section.ram_addr = vram_addr.value(); section.size = size.value(); section.name = name.value(); + section.gp_ram_addr = gp_ram_addr.value_or(0); section.executable = true; // Read functions for the section.