WIP function pad min

This commit is contained in:
Nikita Smith
2025-06-10 10:03:48 -07:00
committed by Ryan Fleury
parent 9d93eda3ed
commit 5fb0d978b9
7 changed files with 211 additions and 64 deletions
+26 -10
View File
@@ -8,16 +8,12 @@ safe_cast_u16x(U64 x)
return (U16)x;
}
////////////////////////////////
internal U64
u128_mod64(U128 a, U64 b)
{
return a.u64[1] % b;
}
////////////////////////////////
internal Version
make_version(U64 major, U64 minor)
{
@@ -45,8 +41,6 @@ version_compar(Version a, Version b)
return cmp;
}
////////////////////////////////
internal ISectOff
isect_off(U32 isect, U32 off)
{
@@ -54,8 +48,6 @@ isect_off(U32 isect, U32 off)
return result;
}
////////////////////////////////
internal int
u16_compar(const void *raw_a, const void *raw_b)
{
@@ -123,7 +115,6 @@ u64_compar_is_before(void *raw_a, void *raw_b)
return is_before;
}
internal int
u8_is_before(void *raw_a, void *raw_b)
{
@@ -212,8 +203,33 @@ pair_u64_compar_v1(const void *raw_a, const void *raw_b)
return u64_compar(&a->v1, &b->v1);
}
internal U64
pair_u64_nearest_v0(PairU64 *arr, U64 count, U64 v)
{
U64 result = max_U64;
////////////////////////////////
if (count > 1 && arr[0].v0 <= v && v < arr[count-1].v0) {
U64 l = 0;
U64 r = count - 1;
for (; l <= r; ) {
U64 m = l + (r - l) / 2;
if (arr[m].v0 == v) {
return m;
} else if (arr[m].v0 < v) {
l = m + 1;
} else {
r = m - 1;
}
}
result = l;
} else if (count == 1 && arr[0].v0 == v) {
result = 0;
} else if (count > 0 && v >= arr[count-1].v0) {
result = count-1;
}
return result;
}
internal void
str8_list_concat_in_place_array(String8List *list, String8List *arr, U64 count)
+163 -32
View File
@@ -1490,6 +1490,12 @@ lnk_build_link_context(TP_Context *tp, TP_Arena *tp_arena, LNK_Config *config)
for (U64 obj_idx = 0; obj_idx < obj_node_arr.count; obj_idx += 1) {
if (obj_node_arr.v[obj_idx].data.header.machine != COFF_MachineType_Unknown) {
config->machine = obj_node_arr.v[obj_idx].data.header.machine;
if (config->do_function_pad_min == LNK_SwitchState_Yes && config->function_pad_min == 0) {
config->function_pad_min = push_array(tp_arena->v[0], U64, 1);
*config->function_pad_min = lnk_get_default_function_pad_min(config->machine);
}
break;
}
}
@@ -2355,12 +2361,12 @@ THREAD_POOL_TASK_FUNC(lnk_gather_section_contribs_task)
if (sc_chunk) {
U16 sc_align = coff_align_size_from_section_flags(sect_header->flags);
sc = lnk_section_contrib_chunk_push_atomic(sc_chunk, 1);
sc->node.next = 0;
sc->node.string = str8_substr(obj->data, rng_1u64(sect_header->foff, sect_header->foff + sect_header->fsize));
sc->data_list = &sc->node;
sc->align = sc_align == 0 ? task->default_align : sc_align;
sc->u.obj_idx = obj_idx;
sc->u.obj_sect_idx = sect_idx;
sc->first_data_node.next = 0;
sc->first_data_node.string = str8_substr(obj->data, rng_1u64(sect_header->foff, sect_header->foff + sect_header->fsize));
sc->last_data_node = &sc->first_data_node;
sc->align = sc_align == 0 ? task->default_align : sc_align;
sc->u.obj_idx = obj_idx;
sc->u.obj_sect_idx = sect_idx;
}
temp_end(temp);
@@ -2396,6 +2402,122 @@ THREAD_POOL_TASK_FUNC(lnk_set_comdat_leaders_task)
ProfEnd();
}
internal
THREAD_POOL_TASK_FUNC(lnk_split_func_contribs_task)
{
Temp scratch = scratch_begin(&arena, 1);
LNK_BuildImageTask *task = raw_task;
U64 obj_idx = task_id;
LNK_Obj *obj = task->objs[obj_idx];
String8 string_table = str8_substr(obj->data, obj->header.string_table_range);
ProfBeginV("%S", obj->path);
COFF_ParsedSymbol symbol;
for (U64 symbol_idx = 0; symbol_idx < obj->header.symbol_count; symbol_idx += (1 + symbol.aux_symbol_count)) {
symbol = lnk_parsed_symbol_from_coff_symbol_idx(obj, symbol_idx);
// is this a function symbol?
COFF_SymbolValueInterpType interp = coff_interp_symbol(symbol.section_number, symbol.value, symbol.storage_class);
if (interp == COFF_SymbolValueInterp_Regular && COFF_SymbolType_IsFunc(symbol.type)) {
if (symbol.section_number == 0 || symbol.section_number > obj->header.section_count_no_null) {
lnk_error_obj(LNK_Error_IllData, obj, "out ouf bounds section index in symbol \"%S (%u)\"", symbol.name, symbol.section_number);
}
COFF_SectionHeader *section_header = lnk_coff_section_header_from_section_number(obj, symbol.section_number);
if (symbol.value > section_header->fsize) {
lnk_error_obj(LNK_Error_IllData, obj, "out of bounds section offset in symbol \"%S (%u)\"", symbol.name, symbol.value);
}
if (~section_header->flags & COFF_SectionFlag_CntCode) {
String8 section_name = coff_name_from_section_header(string_table, section_header);
lnk_error_obj(LNK_Error_IllData, obj, "symbol %S (No. 0x%x) has a function type but points into section that is not declared as code %S (No. 0x%x)",
symbol.name, symbol_idx, section_name, symbol.section_number);
continue;
}
if (symbol.value > 0) {
// find chunk that is near symbol
LNK_SectionContrib *sc = task->sect_map[obj_idx][symbol.section_number-1];
String8Node *current = &sc->first_data_node;
U64 offset_cursor = 0;
for (String8Node *c = current; c != 0; c = c->next) {
if (offset_cursor + c->string.size >= symbol.value) {
current = c;
break;
}
offset_cursor += c->string.size;
}
if (offset_cursor < symbol.value) {
// bifurcate chunk at symbol offset
U64 split_pos = symbol.value - offset_cursor;
String8 left = str8_substr(current->string, rng_1u64(0, split_pos));
String8 right = str8_substr(current->string, rng_1u64(split_pos, current->string.size));
// update split node data
current->string = left;
// create new data node
String8Node *split_node = push_array(arena, String8Node, 1);
split_node->string = right;
// insert split node after current node
split_node->next = current->next;
current->next = split_node;
if (sc->last_data_node == current) {
sc->last_data_node = split_node;
}
}
}
}
}
for (U64 sect_idx = 0; sect_idx < obj->header.section_count_no_null; sect_idx += 1) {
COFF_SectionHeader *section_header = lnk_coff_section_header_from_section_number(obj, sect_idx+1);
if (~section_header->flags & COFF_SectionFlag_CntCode) { continue; }
COFF_RelocInfo reloc_info = coff_reloc_info_from_section_header(obj->data, section_header);
COFF_Reloc *relocs = (COFF_Reloc *)(obj->data.str + reloc_info.array_off);
LNK_SectionContrib *sc = task->sect_map[obj_idx][sect_idx];
U64 offset_map_count = 0;
for (String8Node *data_n = &sc->first_data_node; data_n != 0; data_n = data_n->next) {
offset_map_count += 1;
}
if (offset_map_count <= 1) { continue; }
Temp temp = temp_begin(scratch.arena);
PairU64 *offset_map = push_array(temp.arena, PairU64, offset_map_count);
U64 data_node_idx = 0;
U64 prev_cursor_offset = 0;
U64 new_cursor_offset = 0;
for (String8Node *data_n = &sc->first_data_node; data_n != 0; data_n = data_n->next, data_node_idx += 1) {
offset_map[data_node_idx].v0 = prev_cursor_offset;
offset_map[data_node_idx].v1 = new_cursor_offset;
prev_cursor_offset += data_n->string.size;
new_cursor_offset += Max(task->function_pad_min, data_n->string.size);
}
for (U64 reloc_idx = 0; reloc_idx < reloc_info.count; reloc_idx += 1) {
COFF_Reloc *reloc = &relocs[reloc_idx];
U64 offset_idx = pair_u64_nearest_v0(offset_map, offset_map_count, reloc->apply_off);
if (offset_idx < offset_map_count) {
reloc->apply_off = offset_map[offset_idx].v1 + (reloc->apply_off - offset_map[offset_idx].v0);
} else {
InvalidPath;
}
}
temp_end(temp);
}
// TODO: update symbols with new offsets
ProfEnd();
scratch_end(scratch);
}
internal int
lnk_section_contrib_ptr_is_before(void *raw_a, void *raw_b)
{
@@ -3883,12 +4005,13 @@ lnk_build_image(TP_Arena *arena, TP_Context *tp, LNK_Config *config, LNK_SymbolT
LNK_Section *common_block_sect = lnk_section_table_search(sectab, str8_lit(".bss"), PE_BSS_SECTION_FLAGS);
LNK_BuildImageTask task = {0};
task.symtab = symtab;
task.sectab = sectab;
task.default_align = coff_default_align_from_machine(config->machine);
task.objs_count = objs_count;
task.objs = objs;
task.sect_map = 0;
task.symtab = symtab;
task.sectab = sectab;
task.function_pad_min = *config->function_pad_min;
task.default_align = coff_default_align_from_machine(config->machine);
task.objs_count = objs_count;
task.objs = objs;
task.sect_map = 0;
ProfBegin("Remove Associative Sections");
tp_for_parallel(tp, 0, objs_count, lnk_remove_associative_sections_task, &task);
@@ -3977,7 +4100,7 @@ lnk_build_image(TP_Arena *arena, TP_Context *tp, LNK_Config *config, LNK_SymbolT
LNK_SectionContribChunk *contrib_chunk = 0;
hash_table_search_string_raw(task.contribs_ht, defn_name_with_flags, &contrib_chunk);
if (!contrib_chunk) {
contrib_chunk = lnk_section_contrib_chunk_list_push_chunk(sectab->arena, &sect->contribs, sect_defn->contribs_count, sort_idx);
contrib_chunk = lnk_section_contrib_chunk_list_push_chunk(arena->v[0], &sect->contribs, sect_defn->contribs_count, sort_idx);
hash_table_push_string_raw(sectab->arena, task.contribs_ht, defn_name_with_flags, contrib_chunk);
}
@@ -4001,6 +4124,12 @@ lnk_build_image(TP_Arena *arena, TP_Context *tp, LNK_Config *config, LNK_SymbolT
tp_for_parallel(tp, 0, objs_count, lnk_gather_section_contribs_task, &task);
ProfEnd();
if (config->do_function_pad_min == LNK_SwitchState_Yes) {
ProfBegin("Split Code Sections");
tp_for_parallel(tp, arena, objs_count, lnk_split_func_contribs_task, &task);
ProfEnd();
}
// ensure determinism by sorting section contribs in chunks by input index
{
ProfBegin("Sort Section Contribs");
@@ -4084,7 +4213,7 @@ lnk_build_image(TP_Arena *arena, TP_Context *tp, LNK_Config *config, LNK_SymbolT
radsort(common_block_contribs, common_block_contribs_count, lnk_common_block_contrib_is_before);
// compute .bss virtual size - this marks start of the common block
lnk_finalize_section_layout(common_block_sect, config->file_align);
lnk_finalize_section_layout(common_block_sect, config->file_align, *config->function_pad_min);
U64 common_block_cursor = common_block_sect->vsize;
// compute and assign offsets into the common block
@@ -4100,12 +4229,12 @@ lnk_build_image(TP_Arena *arena, TP_Context *tp, LNK_Config *config, LNK_SymbolT
// append common block's contribution
LNK_SectionContribChunk *common_block_chunk = lnk_section_contrib_chunk_list_push_chunk(sectab->arena, &common_block_sect->contribs, 1, str8(0,0));
LNK_SectionContrib *common_block_sc = lnk_section_contrib_chunk_push(common_block_chunk, 1);
common_block_sc->u.obj_idx = max_U32;
common_block_sc->u.obj_sect_idx = max_U32;
common_block_sc->align = 1;
common_block_sc->node.next = 0;
common_block_sc->node.string = str8(0, common_block_cursor - common_block_sect->vsize);
common_block_sc->data_list = &common_block_sc->node;
common_block_sc->u.obj_idx = max_U32;
common_block_sc->u.obj_sect_idx = max_U32;
common_block_sc->align = 1;
common_block_sc->first_data_node.next = 0;
common_block_sc->first_data_node.string = str8(0, common_block_cursor - common_block_sect->vsize);
common_block_sc->last_data_node = &common_block_sc->first_data_node;
ProfEnd();
}
@@ -4124,7 +4253,7 @@ lnk_build_image(TP_Arena *arena, TP_Context *tp, LNK_Config *config, LNK_SymbolT
// assign contribs offsets, sizes, and section indices
for (LNK_SectionNode *sect_n = sectab->list.first; sect_n != 0; sect_n = sect_n->next) {
lnk_finalize_section_layout(&sect_n->data, config->file_align);
lnk_finalize_section_layout(&sect_n->data, config->file_align, *config->function_pad_min);
}
// remove empty sections
@@ -4238,11 +4367,12 @@ lnk_build_image(TP_Arena *arena, TP_Context *tp, LNK_Config *config, LNK_SymbolT
LNK_Section *reloc = lnk_section_table_push(sectab, str8_lit(".reloc"), PE_RELOC_SECTION_FLAGS);
LNK_SectionContribChunk *first_sc_chunk = lnk_section_contrib_chunk_list_push_chunk(sectab->arena, &reloc->contribs, 1, str8_zero());
LNK_SectionContrib *sc = lnk_section_contrib_chunk_push(first_sc_chunk, 1);
sc->data_list = base_relocs_data.first;
sc->align = 1;
sc->u.obj_idx = max_U32;
sc->first_data_node = *base_relocs_data.first;
sc->last_data_node = base_relocs_data.last;
sc->align = 1;
sc->u.obj_idx = max_U32;
lnk_finalize_section_layout(reloc, config->file_align);
lnk_finalize_section_layout(reloc, config->file_align, *config->function_pad_min);
lnk_assign_section_virtual_space(reloc, config->sect_align, &voff_cursor);
lnk_assign_section_index(reloc, sectab->next_sect_idx++);
@@ -4270,10 +4400,11 @@ lnk_build_image(TP_Arena *arena, TP_Context *tp, LNK_Config *config, LNK_SymbolT
LNK_SectionContribChunk *image_header_sc_chunk = lnk_section_contrib_chunk_list_push_chunk(sectab->arena, &image_header_sect->contribs, 1, str8_zero());
LNK_SectionContrib *image_header_sc = lnk_section_contrib_chunk_push(image_header_sc_chunk, 1);
image_header_sc->align = config->file_align;
image_header_sc->data_list = image_header_data.first;
image_header_sc->align = config->file_align;
image_header_sc->first_data_node = *image_header_data.first;
image_header_sc->last_data_node = image_header_data.last;
lnk_finalize_section_layout(image_header_sect, config->file_align);
lnk_finalize_section_layout(image_header_sect, config->file_align, *config->function_pad_min);
}
ProfBegin("Patch Section Symbols");
@@ -4319,7 +4450,7 @@ lnk_build_image(TP_Arena *arena, TP_Context *tp, LNK_Config *config, LNK_SymbolT
// copy contrib contents
{
U64 cursor = 0;
for (String8Node *data_n = sc->data_list; data_n != 0; data_n = data_n->next) {
for (String8Node *data_n = &sc->first_data_node; data_n != 0; data_n = data_n->next) {
Assert(sc->u.off + data_n->string.size <= sect->vsize);
MemoryCopy(image_data.str + sect->foff + sc->u.off + cursor, data_n->string.str, data_n->string.size);
cursor += data_n->string.size;
@@ -4876,9 +5007,9 @@ entry_point(CmdLine *cmdline)
{
Temp scratch = scratch_begin(0,0);
lnk_init_error_handler();
LNK_Config *config = lnk_config_from_argcv(scratch.arena, cmdline->argc, cmdline->argv);
TP_Context *tp = tp_alloc(scratch.arena, config->worker_count, config->max_worker_count, config->shared_thread_pool_name);
TP_Arena *tp_arena = tp_arena_alloc(tp);
LNK_Config *config = lnk_config_from_argcv(scratch.arena, cmdline->argc, cmdline->argv);
TP_Context *tp = tp_alloc(scratch.arena, config->worker_count, config->max_worker_count, config->shared_thread_pool_name);
TP_Arena *tp_arena = tp_arena_alloc(tp);
lnk_run(tp, tp_arena, config);
scratch_end(scratch);
}
+2 -1
View File
@@ -74,9 +74,10 @@ typedef struct
{
LNK_SymbolTable *symtab;
LNK_SectionTable *sectab;
U64 function_pad_min;
U64 default_align;
U64 objs_count;
LNK_Obj **objs;
LNK_Obj **objs;
LNK_SectionContrib ***sect_map;
HashTable *contribs_ht;
LNK_SectionArray image_sects;
+4 -4
View File
@@ -1202,12 +1202,12 @@ lnk_apply_cmd_option_to_config(Arena *arena, LNK_Config *config, String8 cmd_nam
case LNK_CmdSwitch_FunctionPadMin: {
if (value_strings.node_count == 0) {
config->function_pad_min = 0; // :function_pad_min
config->function_pad_min = 0;
} else {
local_persist U64 function_pad_min;
lnk_cmd_switch_parse_u64(obj_path, lib_path, cmd_switch, value_strings, &function_pad_min, LNK_ParseU64Flag_CheckUnder32bit);
config->function_pad_min = &function_pad_min;
config->function_pad_min = push_array(arena, U64, 1);
lnk_cmd_switch_parse_u64(obj_path, lib_path, cmd_switch, value_strings, config->function_pad_min, LNK_ParseU64Flag_CheckUnder32bit);
}
config->do_function_pad_min = LNK_SwitchState_Yes;
} break;
case LNK_CmdSwitch_Heap: {
+1
View File
@@ -324,6 +324,7 @@ typedef struct LNK_Config
U64 worker_count;
U64 max_worker_count;
String8 shared_thread_pool_name;
LNK_SwitchState do_function_pad_min;
U64 *function_pad_min;
U64 *manifest_resource_id;
B32 no_default_libs;
+10 -13
View File
@@ -319,7 +319,7 @@ lnk_section_contrib_chunk_is_before(void *raw_a, void *raw_b)
}
internal void
lnk_finalize_section_layout(LNK_Section *sect, U64 file_align)
lnk_finalize_section_layout(LNK_Section *sect, U64 file_align, U64 function_pad_min)
{
Temp scratch = scratch_begin(0,0);
@@ -341,22 +341,19 @@ lnk_finalize_section_layout(LNK_Section *sect, U64 file_align)
ProfEnd();
ProfBegin("Layout Contribs");
U64 cursor = 0;
U64 min_sc_size = sect->flags & COFF_SectionFlag_CntCode ? function_pad_min : 0;
U64 cursor = 0;
for (LNK_SectionContribChunk *sc_chunk = sect->contribs.first; sc_chunk != 0; sc_chunk = sc_chunk->next) {
for (U64 sc_idx = 0; sc_idx < sc_chunk->count; sc_idx += 1) {
LNK_SectionContrib *sc = sc_chunk->v[sc_idx];
cursor = AlignPow2(cursor, sc->align);
// store section contribution start offset
U64 sc_off = cursor;
// compute contrib size
U64 sc_size = lnk_size_from_section_contrib(sc);
cursor += sc_size;
// assign offset and size
sc->u.off = sc_off;
cursor = AlignPow2(cursor, sc->align);
sc->u.off = cursor;
// advance cursor
U64 sc_size = lnk_size_from_section_contrib(sc);
cursor += Max(min_sc_size, sc_size);
}
}
ProfEnd();
@@ -403,7 +400,7 @@ internal U64
lnk_size_from_section_contrib(LNK_SectionContrib *sc)
{
U64 size = 0;
for (String8Node *n = sc->data_list; n != 0; n = n->next) {
for (String8Node *n = &sc->first_data_node; n != 0; n = n->next) {
size += n->string.size;
}
return size;
+5 -4
View File
@@ -5,14 +5,15 @@
typedef struct LNK_SectionContrib
{
String8Node node; // most contributions require at least one data node, so preallocate it here
String8Node *data_list; // list of data nodes that contribute to final section
String8Node first_data_node; // most contributions require at least one data node, so preallocate it here
String8Node *last_data_node; // list of data nodes that contribute to final section
union {
// used before section layout is finalized
// used to sort sections to get deterministic output
struct {
U32 obj_idx; // index of the input obj that contributes to the image section
U32 obj_sect_idx; // index into contributing obj's section table
};
// used after section layout is finalized
struct {
U32 off; // contribution offset within the image section
@@ -110,7 +111,7 @@ internal void lnk_section_table_merge(LNK_SectionTable *sectab, L
// --- Section Finalization ----------------------------------------------------
internal void lnk_finalize_section_layout (LNK_Section *sect, U64 file_align);
internal void lnk_finalize_section_layout (LNK_Section *sect, U64 file_align, U64 function_pad_min);
internal void lnk_assign_section_index (LNK_Section *sect, U64 sect_idx);
internal void lnk_assign_section_virtual_space(LNK_Section *sect, U64 sect_align, U64 *voff_cursor);
internal void lnk_assign_section_file_space (LNK_Section *sect, U64 *foff_cursor);