diff options
| author | Ersikan <julien.philippon@epitech.eu> | 2021-03-12 23:46:51 +0100 |
|---|---|---|
| committer | Andrew Kelley <andrew@ziglang.org> | 2021-12-21 11:33:12 -0800 |
| commit | e15a267668bae168f2bba06cc3706c54bc062522 (patch) | |
| tree | b84e6f17eda1e677dbb7a9adcf630741ae94c05b | |
| parent | 44061cd760cba603ddc2afa10d882ce71c5d389b (diff) | |
| download | zig-e15a267668bae168f2bba06cc3706c54bc062522.tar.gz zig-e15a267668bae168f2bba06cc3706c54bc062522.zip | |
elf: Put constant data in the .rodata section
Allocate a new program header and a new section to accomodate the read-only data
section ".rodata".
Separate TextBlock into multiple TextBlockList, to separate decl in different
sections.
If a Decl is not a function, it is added to the .rodata section.
| -rw-r--r-- | src/link/Elf.zig | 146 |
1 files changed, 106 insertions, 40 deletions
diff --git a/src/link/Elf.zig b/src/link/Elf.zig index 24f8a02b95..992332be36 100644 --- a/src/link/Elf.zig +++ b/src/link/Elf.zig @@ -54,13 +54,14 @@ phdr_load_re_index: ?u16 = null, /// The index into the program headers of the global offset table. /// It needs PT_LOAD and Read flags. phdr_got_index: ?u16 = null, +/// The index into the program headers of a PT_LOAD program header with Read flag +phdr_load_ro_index: ?u16 = null, entry_addr: ?u64 = null, debug_strtab: std.ArrayListUnmanaged(u8) = std.ArrayListUnmanaged(u8){}, shstrtab: std.ArrayListUnmanaged(u8) = std.ArrayListUnmanaged(u8){}, shstrtab_index: ?u16 = null, -text_section_index: ?u16 = null, symtab_section_index: ?u16 = null, got_section_index: ?u16 = null, debug_info_section_index: ?u16 = null, @@ -115,8 +116,8 @@ error_flags: File.ErrorFlags = File.ErrorFlags{}, /// overcapacity can be negative. A simple way to have negative overcapacity is to /// allocate a fresh text block, which will have ideal capacity, and then grow it /// by 1 byte. It will then have -1 overcapacity. -text_block_free_list: std.ArrayListUnmanaged(*TextBlock) = .{}, -last_text_block: ?*TextBlock = null, +text_block_list: TextBlockList = .{}, +rodata_block_list: TextBlockList = .{}, /// A list of `SrcFn` whose Line Number Programs have surplus capacity. /// This is the same concept as `text_block_free_list`; see those doc comments. @@ -204,6 +205,14 @@ pub const TextBlock = struct { } }; +/// A list of text blocks in a specific section +const TextBlockList = struct { + free_list: std.ArrayListUnmanaged(*TextBlock) = .{}, + last_block: ?*TextBlock = null, + phdr_index: ?u16 = null, + section_index: ?u16 = null, +}; + pub const Export = struct { sym_index: ?u32 = null, }; @@ -314,7 +323,8 @@ pub fn deinit(self: *Elf) void { self.global_symbol_free_list.deinit(self.base.allocator); self.local_symbol_free_list.deinit(self.base.allocator); self.offset_table_free_list.deinit(self.base.allocator); - self.text_block_free_list.deinit(self.base.allocator); + self.text_block_list.free_list.deinit(self.base.allocator); + self.rodata_block_list.free_list.deinit(self.base.allocator); self.dbg_line_fn_free_list.deinit(self.base.allocator); self.dbg_info_decl_free_list.deinit(self.base.allocator); self.offset_table.deinit(self.base.allocator); @@ -450,6 +460,7 @@ pub fn populateMissingMetadata(self: *Elf) !void { const ptr_size: u8 = self.ptrWidthBytes(); if (self.phdr_load_re_index == null) { self.phdr_load_re_index = @intCast(u16, self.program_headers.items.len); + self.text_block_list.phdr_index = self.phdr_load_re_index; const file_size = self.base.options.program_code_size_hint; const p_align = 0x1000; const off = self.findFreeSpace(file_size, p_align); @@ -492,6 +503,29 @@ pub fn populateMissingMetadata(self: *Elf) !void { }); self.phdr_table_dirty = true; } + if (self.phdr_load_ro_index == null) { + self.phdr_load_ro_index = @intCast(u16, self.program_headers.items.len); + self.rodata_block_list.phdr_index = self.phdr_load_ro_index; + // TODO Find a hint about how much data need to be in rodata ? + const file_size = 1024; + // Same reason as for GOT + const p_align = if (self.base.options.target.os.tag == .linux) 0x1000 else @as(u16, ptr_size); + const off = self.findFreeSpace(file_size, p_align); + log.debug("found PT_LOAD free space 0x{x} to 0x{x}\n", .{ off, off + file_size }); + // TODO Same as for GOT + const rodata_addr: u32 = if (self.base.options.target.cpu.arch.ptrBitWidth() >= 32) 0x6000000 else 0xD000; + try self.program_headers.append(self.base.allocator, .{ + .p_type = elf.PT_LOAD, + .p_offset = off, + .p_filesz = file_size, + .p_vaddr = rodata_addr, + .p_paddr = rodata_addr, + .p_memsz = file_size, + .p_align = p_align, + .p_flags = elf.PF_R, + }); + self.phdr_table_dirty = true; + } if (self.shstrtab_index == null) { self.shstrtab_index = @intCast(u16, self.sections.items.len); assert(self.shstrtab.items.len == 0); @@ -513,8 +547,8 @@ pub fn populateMissingMetadata(self: *Elf) !void { self.shstrtab_dirty = true; self.shdr_table_dirty = true; } - if (self.text_section_index == null) { - self.text_section_index = @intCast(u16, self.sections.items.len); + if (self.text_block_list.section_index == null) { + self.text_block_list.section_index = @intCast(u16, self.sections.items.len); const phdr = &self.program_headers.items[self.phdr_load_re_index.?]; try self.sections.append(self.base.allocator, .{ @@ -549,6 +583,24 @@ pub fn populateMissingMetadata(self: *Elf) !void { }); self.shdr_table_dirty = true; } + if (self.rodata_block_list.section_index == null) { + self.rodata_block_list.section_index = @intCast(u16, self.sections.items.len); + const phdr = &self.program_headers.items[self.phdr_load_ro_index.?]; + + try self.sections.append(self.base.allocator, .{ + .sh_name = try self.makeString(".rodata"), + .sh_type = elf.SHT_PROGBITS, + .sh_flags = elf.SHF_ALLOC, + .sh_addr = phdr.p_vaddr, + .sh_offset = phdr.p_offset, + .sh_size = phdr.p_filesz, + .sh_link = 0, + .sh_info = 0, + .sh_addralign = phdr.p_align, + .sh_entsize = 0, + }); + self.shdr_table_dirty = true; + } if (self.symtab_section_index == null) { self.symtab_section_index = @intCast(u16, self.sections.items.len); const min_align: u16 = if (small_ptr) @alignOf(elf.Elf32_Sym) else @alignOf(elf.Elf64_Sym); @@ -1927,17 +1979,17 @@ fn writeElfHeader(self: *Elf) !void { try self.base.file.?.pwriteAll(hdr_buf[0..index], 0); } -fn freeTextBlock(self: *Elf, text_block: *TextBlock) void { +fn freeTextBlock(self: *Elf, block_list: *TextBlockList, text_block: *TextBlock) void { var already_have_free_list_node = false; { var i: usize = 0; // TODO turn text_block_free_list into a hash map - while (i < self.text_block_free_list.items.len) { - if (self.text_block_free_list.items[i] == text_block) { - _ = self.text_block_free_list.swapRemove(i); + while (i < block_list.free_list.items.len) { + if (block_list.free_list.items[i] == text_block) { + _ = block_list.free_list.swapRemove(i); continue; } - if (self.text_block_free_list.items[i] == text_block.prev) { + if (block_list.free_list.items[i] == text_block.prev) { already_have_free_list_node = true; } i += 1; @@ -1945,9 +1997,9 @@ fn freeTextBlock(self: *Elf, text_block: *TextBlock) void { } // TODO process free list for dbg info just like we do above for vaddrs - if (self.last_text_block == text_block) { + if (block_list.last_block == text_block) { // TODO shrink the .text section size here - self.last_text_block = text_block.prev; + block_list.last_block = text_block.prev; } if (self.dbg_info_decl_first == text_block) { self.dbg_info_decl_first = text_block.dbg_info_next; @@ -1963,7 +2015,7 @@ fn freeTextBlock(self: *Elf, text_block: *TextBlock) void { if (!already_have_free_list_node and prev.freeListEligible(self.*)) { // The free list is heuristics, it doesn't have to be perfect, so we can // ignore the OOM here. - self.text_block_free_list.append(self.base.allocator, prev) catch {}; + block_list.free_list.append(self.base.allocator, prev) catch {}; } } else { text_block.prev = null; @@ -1990,25 +2042,24 @@ fn freeTextBlock(self: *Elf, text_block: *TextBlock) void { } } -fn shrinkTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64) void { +fn shrinkTextBlock(self: *Elf, block_list: *TextBlockList, text_block: *TextBlock, new_block_size: u64) void { _ = self; + _ = block_list; _ = text_block; _ = new_block_size; - // TODO check the new capacity, and if it crosses the size threshold into a big enough - // capacity, insert a free list node for it. } -fn growTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64, alignment: u64) !u64 { +fn growTextBlock(self: *Elf, block_list: *TextBlockList, text_block: *TextBlock, new_block_size: u64, alignment: u64) !u64 { const sym = self.local_symbols.items[text_block.local_sym_index]; const align_ok = mem.alignBackwardGeneric(u64, sym.st_value, alignment) == sym.st_value; const need_realloc = !align_ok or new_block_size > text_block.capacity(self.*); if (!need_realloc) return sym.st_value; - return self.allocateTextBlock(text_block, new_block_size, alignment); + return self.allocateTextBlock(block_list, text_block, new_block_size, alignment); } -fn allocateTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64, alignment: u64) !u64 { - const phdr = &self.program_headers.items[self.phdr_load_re_index.?]; - const shdr = &self.sections.items[self.text_section_index.?]; +fn allocateTextBlock(self: *Elf, block_list: *TextBlockList, text_block: *TextBlock, new_block_size: u64, alignment: u64) !u64 { + const phdr = &self.program_headers.items[block_list.phdr_index.?]; + const shdr = &self.sections.items[block_list.section_index.?]; const new_block_ideal_capacity = padToIdeal(new_block_size); // We use these to indicate our intention to update metadata, placing the new block, @@ -2023,8 +2074,8 @@ fn allocateTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64, al // The list is unordered. We'll just take the first thing that works. const vaddr = blk: { var i: usize = 0; - while (i < self.text_block_free_list.items.len) { - const big_block = self.text_block_free_list.items[i]; + while (i < block_list.free_list.items.len) { + const big_block = block_list.free_list.items[i]; // We now have a pointer to a live text block that has too much capacity. // Is it enough that we could fit this new text block? const sym = self.local_symbols.items[big_block.local_sym_index]; @@ -2039,7 +2090,7 @@ fn allocateTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64, al // should be deleted because the block that it points to has grown to take up // more of the extra capacity. if (!big_block.freeListEligible(self.*)) { - _ = self.text_block_free_list.swapRemove(i); + _ = block_list.free_list.swapRemove(i); } else { i += 1; } @@ -2057,7 +2108,7 @@ fn allocateTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64, al free_list_removal = i; } break :blk new_start_vaddr; - } else if (self.last_text_block) |last| { + } else if (block_list.last_block) |last| { const sym = self.local_symbols.items[last.local_sym_index]; const ideal_capacity = padToIdeal(sym.st_size); const ideal_capacity_end_vaddr = sym.st_value + ideal_capacity; @@ -2077,7 +2128,7 @@ fn allocateTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64, al if (needed_size > text_capacity) { // Must move the entire text section. const new_offset = self.findFreeSpace(needed_size, 0x1000); - const text_size = if (self.last_text_block) |last| blk: { + const text_size = if (block_list.last_block) |last| blk: { const sym = self.local_symbols.items[last.local_sym_index]; break :blk (sym.st_value + sym.st_size) - phdr.p_vaddr; } else 0; @@ -2086,7 +2137,7 @@ fn allocateTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64, al shdr.sh_offset = new_offset; phdr.p_offset = new_offset; } - self.last_text_block = text_block; + block_list.last_block = text_block; shdr.sh_size = needed_size; phdr.p_memsz = needed_size; @@ -2124,11 +2175,19 @@ fn allocateTextBlock(self: *Elf, text_block: *TextBlock, new_block_size: u64, al text_block.next = null; } if (free_list_removal) |i| { - _ = self.text_block_free_list.swapRemove(i); + _ = block_list.free_list.swapRemove(i); } return vaddr; } +/// Get the block list corresponding to a specific decl +/// For example, if the decl is a function, it returns the list of the section .text +fn getDeclBlockList(self: *Elf, decl: *const Module.Decl) *TextBlockList { + // const is_fn = decl.val.tag() == .function; + const is_fn = decl.ty.zigTypeTag() == .Fn; + return if (is_fn) &self.text_block_list else &self.rodata_block_list; +} + pub fn allocateDeclIndexes(self: *Elf, decl: *Module.Decl) !void { if (self.llvm_object) |_| return; @@ -2172,8 +2231,10 @@ pub fn freeDecl(self: *Elf, decl: *Module.Decl) void { if (self.llvm_object) |llvm_object| return llvm_object.freeDecl(decl); } + const block_list = self.getDeclBlockList(decl); + // Appending to free lists is allowed to fail because the free lists are heuristics based anyway. - self.freeTextBlock(&decl.link.elf); + self.freeTextBlock(block_list, &decl.link.elf); if (decl.link.elf.local_sym_index != 0) { self.local_symbol_free_list.append(self.base.allocator, decl.link.elf.local_sym_index) catch {}; self.offset_table_free_list.append(self.base.allocator, decl.link.elf.offset_table_index) catch {}; @@ -2216,6 +2277,8 @@ fn deinitRelocs(gpa: Allocator, table: *File.DbgInfoTypeRelocsTable) void { fn updateDeclCode(self: *Elf, decl: *Module.Decl, code: []const u8, stt_bits: u8) !*elf.Elf64_Sym { const required_alignment = decl.ty.abiAlignment(self.base.options.target); + const block_list = self.getDeclBlockList(decl); + assert(decl.link.elf.local_sym_index != 0); // Caller forgot to allocateDeclIndexes() const local_sym = &self.local_symbols.items[decl.link.elf.local_sym_index]; if (local_sym.st_size != 0) { @@ -2223,7 +2286,7 @@ fn updateDeclCode(self: *Elf, decl: *Module.Decl, code: []const u8, stt_bits: u8 const need_realloc = code.len > capacity or !mem.isAlignedGeneric(u64, local_sym.st_value, required_alignment); if (need_realloc) { - const vaddr = try self.growTextBlock(&decl.link.elf, code.len, required_alignment); + const vaddr = try self.growTextBlock(block_list, &decl.link.elf, code.len, required_alignment); log.debug("growing {s} from 0x{x} to 0x{x}", .{ decl.name, local_sym.st_value, vaddr }); if (vaddr != local_sym.st_value) { local_sym.st_value = vaddr; @@ -2233,27 +2296,28 @@ fn updateDeclCode(self: *Elf, decl: *Module.Decl, code: []const u8, stt_bits: u8 try self.writeOffsetTableEntry(decl.link.elf.offset_table_index); } } else if (code.len < local_sym.st_size) { - self.shrinkTextBlock(&decl.link.elf, code.len); + self.shrinkTextBlock(block_list, &decl.link.elf, code.len); } local_sym.st_size = code.len; local_sym.st_name = try self.updateString(local_sym.st_name, mem.sliceTo(decl.name, 0)); local_sym.st_info = (elf.STB_LOCAL << 4) | stt_bits; local_sym.st_other = 0; - local_sym.st_shndx = self.text_section_index.?; + local_sym.st_shndx = block_list.section_index.?; // TODO this write could be avoided if no fields of the symbol were changed. try self.writeSymbol(decl.link.elf.local_sym_index); } else { const decl_name = mem.sliceTo(decl.name, 0); const name_str_index = try self.makeString(decl_name); - const vaddr = try self.allocateTextBlock(&decl.link.elf, code.len, required_alignment); + const vaddr = try self.allocateTextBlock(block_list, &decl.link.elf, code.len, required_alignment); + errdefer self.freeTextBlock(block_list, &decl.link.elf); log.debug("allocated text block for {s} at 0x{x}", .{ decl_name, vaddr }); - errdefer self.freeTextBlock(&decl.link.elf); + errdefer self.freeTextBlock(block_list, &decl.link.elf); local_sym.* = .{ .st_name = name_str_index, .st_info = (elf.STB_LOCAL << 4) | stt_bits, .st_other = 0, - .st_shndx = self.text_section_index.?, + .st_shndx = block_list.section_index.?, .st_value = vaddr, .st_size = code.len, }; @@ -2263,8 +2327,8 @@ fn updateDeclCode(self: *Elf, decl: *Module.Decl, code: []const u8, stt_bits: u8 try self.writeOffsetTableEntry(decl.link.elf.offset_table_index); } - const section_offset = local_sym.st_value - self.program_headers.items[self.phdr_load_re_index.?].p_vaddr; - const file_offset = self.sections.items[self.text_section_index.?].sh_offset + section_offset; + const section_offset = local_sym.st_value - self.program_headers.items[block_list.phdr_index.?].p_vaddr; + const file_offset = self.sections.items[block_list.section_index.?].sh_offset + section_offset; try self.base.file.?.pwriteAll(code, file_offset); return local_sym; @@ -2772,6 +2836,8 @@ pub fn updateDeclExports( if (decl.link.elf.local_sym_index == 0) return; const decl_sym = self.local_symbols.items[decl.link.elf.local_sym_index]; + const block_list = self.getDeclBlockList(decl); + for (exports) |exp| { if (exp.options.section) |section_name| { if (!mem.eql(u8, section_name, ".text")) { @@ -2808,7 +2874,7 @@ pub fn updateDeclExports( .st_name = try self.updateString(sym.st_name, exp.options.name), .st_info = (stb_bits << 4) | stt_bits, .st_other = 0, - .st_shndx = self.text_section_index.?, + .st_shndx = block_list.section_index.?, .st_value = decl_sym.st_value, .st_size = decl_sym.st_size, }; @@ -2822,7 +2888,7 @@ pub fn updateDeclExports( .st_name = name, .st_info = (stb_bits << 4) | stt_bits, .st_other = 0, - .st_shndx = self.text_section_index.?, + .st_shndx = block_list.section_index.?, .st_value = decl_sym.st_value, .st_size = decl_sym.st_size, }; |
