function pad min - final pass over

This commit is contained in:
Nikita Smith
2025-06-11 14:56:07 -07:00
committed by Ryan Fleury
parent bbb825adf4
commit b19be0574c
5 changed files with 79 additions and 181 deletions
+52 -165
View File
@@ -2437,145 +2437,6 @@ THREAD_POOL_TASK_FUNC(lnk_patch_comdats_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);
U64List func_list = {0};
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;
}
}
}
u64_list_push(scratch.arena, &func_list, symbol_idx);
}
}
U64 *sect_offset_map_counts = push_array(scratch.arena, U64, obj->header.section_count_no_null);
PairU64 **sect_offset_map = push_array(scratch.arena, PairU64 *, obj->header.section_count_no_null);
for (U64 sect_idx = 0; sect_idx < obj->header.section_count_no_null; sect_idx += 1) {
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;
}
PairU64 *offset_map = push_array(scratch.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);
}
sect_offset_map_counts[sect_idx] = offset_map_count;
sect_offset_map[sect_idx] = offset_map;
}
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) {
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 = sect_offset_map_counts[sect_idx];
PairU64 *offset_map = sect_offset_map[sect_idx];
if (offset_map_count > 0) {
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;
}
}
}
}
}
for (U64Node *func_n = func_list.first; func_n != 0; func_n = func_n->next) {
COFF_ParsedSymbol symbol = lnk_parsed_symbol_from_coff_symbol_idx(obj, func_n->data);
U64 offset_map_count = sect_offset_map_counts[symbol.section_number-1];
PairU64 *offset_map = sect_offset_map[symbol.section_number-1];
U64 offset_idx = pair_u64_nearest_v0(offset_map, offset_map_count, symbol.value);
AssertAlways(offset_idx < offset_map_count);
U32 *value_ptr = obj->header.is_big_obj ? &((COFF_Symbol32 *)symbol.raw_symbol)->value : &((COFF_Symbol16 *)symbol.raw_symbol)->value;
*value_ptr = offset_map[offset_idx].v1 + (*value_ptr - offset_map[offset_idx].v0);
}
ProfEnd();
scratch_end(scratch);
}
internal int
lnk_section_contrib_ptr_is_before(void *raw_a, void *raw_b)
{
@@ -3139,6 +3000,25 @@ lnk_section_definition_is_before(void *raw_a, void *raw_b)
return u64_compar_is_before(&input_idx_a, &input_idx_b);
}
internal
THREAD_POOL_TASK_FUNC(lnk_flag_hotpatch_contribs_task)
{
LNK_BuildImageTask *task = raw_task;
U64 obj_idx = task_id;
LNK_Obj *obj = task->objs[obj_idx];
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);
COFF_SymbolValueInterpType interp = coff_interp_symbol(symbol.section_number, symbol.value, symbol.storage_class);
if (interp == COFF_SymbolValueInterp_Regular && COFF_SymbolType_IsFunc(symbol.type)) {
COFF_SectionHeader *section_header = lnk_coff_section_header_from_section_number(obj, symbol.section_number);
LNK_SectionContrib *sc = task->sect_map[obj_idx][symbol.section_number-1];
sc->hotpatch = !!(section_header->flags & COFF_SectionFlag_CntCode);
}
}
}
internal void
lnk_push_coff_symbols_from_data(Arena *arena, LNK_SymbolList *symbol_list, String8 data, LNK_SymbolArray obj_symbols)
{
@@ -3659,8 +3539,7 @@ THREAD_POOL_TASK_FUNC(lnk_patch_section_symbols_task)
int
lnk_base_reloc_page_compar(const void *raw_a, const void *raw_b)
{
const LNK_BaseRelocPage *a = raw_a;
const LNK_BaseRelocPage *b = raw_b;
const LNK_BaseRelocPage *a = raw_a, *b = raw_b;
return u64_compar(&a->voff, &b->voff);
}
@@ -4083,38 +3962,46 @@ lnk_build_image(TP_Arena *arena, TP_Context *tp, LNK_Config *config, LNK_SymbolT
ProfBegin("Define And Count Sections");
TP_Temp temp = tp_temp_begin(arena);
// init hash tables for gathering section definitions
ProfBegin("Init Hash Tables For Gathering Section Definitions");
task.u.gather_sects.defns = push_array(arena->v[0], HashTable *, tp->worker_count);
for (U64 worker_idx = 0; worker_idx < tp->worker_count; worker_idx += 1) task.u.gather_sects.defns[worker_idx] = hash_table_init(arena->v[0], 128);
ProfEnd();
ProfBegin("Gather Section Definitions");
tp_for_parallel(tp, arena, objs_count, lnk_gather_section_definitions_task, &task);
ProfEnd();
ProfBegin("Merge Section Definitions Hash Tables");
for (U64 worker_idx = 1; worker_idx < tp->worker_count; worker_idx += 1) {
U64 sect_defns_count = task.u.gather_sects.defns[worker_idx]->count;
LNK_SectionDefinition **sect_defns = values_from_hash_table_raw(arena->v[0], task.u.gather_sects.defns[worker_idx]);
radsort(sect_defns, sect_defns_count, lnk_section_definition_is_before);
U64 sect_defns_count;
LNK_SectionDefinition **sect_defns;
{
for (U64 worker_idx = 1; worker_idx < tp->worker_count; worker_idx += 1) {
U64 sect_defns_count = task.u.gather_sects.defns[worker_idx]->count;
LNK_SectionDefinition **sect_defns = values_from_hash_table_raw(arena->v[0], task.u.gather_sects.defns[worker_idx]);
radsort(sect_defns, sect_defns_count, lnk_section_definition_is_before);
for (U64 defn_idx = 0; defn_idx < sect_defns_count; defn_idx += 1) {
LNK_SectionDefinition *defn = sect_defns[defn_idx];
String8 name_with_flags = lnk_make_name_with_flags(arena->v[0], defn->name, defn->flags);
LNK_SectionDefinition *main_defn = 0;
hash_table_search_string_raw(task.u.gather_sects.defns[0], name_with_flags, &main_defn);
if (main_defn == 0) {
main_defn = sect_defns[defn_idx];
hash_table_push_string_raw(arena->v[0], task.u.gather_sects.defns[0], name_with_flags, main_defn);
} else {
main_defn->contribs_count += sect_defns[defn_idx]->contribs_count;
for (U64 defn_idx = 0; defn_idx < sect_defns_count; defn_idx += 1) {
LNK_SectionDefinition *defn = sect_defns[defn_idx];
String8 name_with_flags = lnk_make_name_with_flags(arena->v[0], defn->name, defn->flags);
LNK_SectionDefinition *main_defn = 0;
hash_table_search_string_raw(task.u.gather_sects.defns[0], name_with_flags, &main_defn);
if (main_defn == 0) {
main_defn = sect_defns[defn_idx];
hash_table_push_string_raw(arena->v[0], task.u.gather_sects.defns[0], name_with_flags, main_defn);
} else {
if (lnk_section_definition_is_before(&sect_defns[defn_idx], &main_defn)) {
main_defn->obj = sect_defns[defn_idx]->obj;
main_defn->obj_sect_idx = sect_defns[defn_idx]->obj_sect_idx;
}
main_defn->contribs_count += sect_defns[defn_idx]->contribs_count;
}
}
}
sect_defns_count = task.u.gather_sects.defns[0]->count;
sect_defns = values_from_hash_table_raw(arena->v[0], task.u.gather_sects.defns[0]);
}
ProfEnd();
U64 sect_defns_count = task.u.gather_sects.defns[0]->count;
LNK_SectionDefinition **sect_defns = values_from_hash_table_raw(arena->v[0], task.u.gather_sects.defns[0]);
ProfBegin("Sort Sections Definitions");
radsort(sect_defns, sect_defns_count, lnk_section_definition_is_before);
ProfEnd();
@@ -4186,12 +4073,6 @@ 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");
@@ -4309,6 +4190,12 @@ lnk_build_image(TP_Arena *arena, TP_Context *tp, LNK_Config *config, LNK_SymbolT
lnk_section_table_merge(sectab, config->merge_list);
}
if (config->do_function_pad_min == LNK_SwitchState_Yes) {
ProfBegin("Flag Hotpatch Section Contribs");
tp_for_parallel(tp, arena, objs_count, lnk_flag_hotpatch_contribs_task, &task);
ProfEnd();
}
// 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, config->function_pad_min);
+3 -1
View File
@@ -229,7 +229,9 @@ THREAD_POOL_TASK_FUNC(lnk_obj_initer)
// Extract obj features from compile symbol in .debug$S
//
B8 hotpatch = 0;
{
if (header.machine == COFF_MachineType_X64) {
hotpatch = 1;
} else {
Temp scratch = scratch_begin(&arena, 1);
CV_Symbol comp_symbol = {0};
+8 -4
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, U64 function_pad_min)
lnk_finalize_section_layout(LNK_Section *sect, U64 file_align, U64 pad_size)
{
Temp scratch = scratch_begin(0,0);
@@ -341,19 +341,23 @@ lnk_finalize_section_layout(LNK_Section *sect, U64 file_align, U64 function_pad_
ProfEnd();
ProfBegin("Layout Contribs");
U64 min_sc_size = sect->flags & COFF_SectionFlag_CntCode ? function_pad_min : 0;
U64 cursor = 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];
// add section pad bytes
if (sc->hotpatch) {
cursor += pad_size;
}
// assign offset
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);
cursor += sc_size;
}
}
ProfEnd();
+2 -1
View File
@@ -22,6 +22,7 @@ typedef struct LNK_SectionContrib
};
} u;
U16 align; // contribution alignment in the image
B8 hotpatch;
} LNK_SectionContrib;
typedef struct LNK_SectionContribChunk
@@ -111,7 +112,7 @@ internal void lnk_section_table_merge(LNK_SectionTable *sectab, L
// --- Section Finalization ----------------------------------------------------
internal void lnk_finalize_section_layout (LNK_Section *sect, U64 file_align, U64 function_pad_min);
internal void lnk_finalize_section_layout (LNK_Section *sect, U64 file_align, U64 pad_size);
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);
+14 -10
View File
@@ -309,7 +309,7 @@ typedef enum
internal COFF_ObjSection *
t_push_text_section(COFF_ObjWriter *obj_writer, String8 data)
{
return coff_obj_writer_push_section(obj_writer, str8_lit(".text"), PE_TEXT_SECTION_FLAGS, data);
return coff_obj_writer_push_section(obj_writer, str8_lit(".text"), PE_TEXT_SECTION_FLAGS | COFF_SectionFlag_Align1Bytes, data);
}
internal COFF_ObjSection *
@@ -3360,17 +3360,21 @@ t_function_pad_min(void)
{
COFF_ObjWriter *obj_writer = coff_obj_writer_alloc(0, COFF_MachineType_X64);
U8 ret[] = { 0xc3, 0xc3, 0xc3 };
COFF_ObjSection *text_sect = t_push_text_section(obj_writer, str8_array_fixed(ret));
coff_obj_writer_push_symbol_extern_func(obj_writer, str8_lit("A"), 0, text_sect);
coff_obj_writer_push_symbol_extern_func(obj_writer, str8_lit("B"), 1, text_sect);
coff_obj_writer_push_symbol_extern_func(obj_writer, str8_lit("C"), 2, text_sect);
U8 ret[] = { 0xc3 };
COFF_ObjSection *text_sect_0 = t_push_text_section(obj_writer, str8_array_fixed(ret));
COFF_ObjSection *text_sect_1 = t_push_text_section(obj_writer, str8_array_fixed(ret));
COFF_ObjSection *text_sect_2 = t_push_text_section(obj_writer, str8_array_fixed(ret));
text_sect_0->flags |= COFF_SectionFlag_Align4Bytes;
text_sect_1->flags |= COFF_SectionFlag_Align2Bytes;
coff_obj_writer_push_symbol_extern_func(obj_writer, str8_lit("A"), 0, text_sect_0);
coff_obj_writer_push_symbol_extern_func(obj_writer, str8_lit("B"), 0, text_sect_1);
coff_obj_writer_push_symbol_extern_func(obj_writer, str8_lit("C"), 0, text_sect_2);
String8 obj = coff_obj_writer_serialize(scratch.arena, obj_writer);
coff_obj_writer_release(&obj_writer);
if (!t_write_file(str8_lit("funcs.obj"), obj)) { goto exit; }
}
int linker_exit_code = t_invoke_linkerf("/subsystem:console /entry:A /functionpadmin:8 /out:a.exe funcs.obj");
int linker_exit_code = t_invoke_linkerf("/subsystem:console /entry:A /functionpadmin:1 /out:a.exe funcs.obj");
if (linker_exit_code != 0) { goto exit; }
String8 exe = t_read_file(scratch.arena, str8_lit("a.exe"));
@@ -3382,9 +3386,9 @@ t_function_pad_min(void)
String8 text_data = str8_substr(exe, rng_1u64(text_sect->foff, text_sect->foff + text_sect->vsize));
U8 expected_text[] = {
0xc3, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc,
0xc3, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc,
0xc3, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc
0xcc, 0xcc, 0xcc, 0xcc, 0xc3,
0xcc, 0xcc, 0xcc, 0xc3,
0xcc, 0xc3,
};
if (!str8_match(text_data, str8_array_fixed(expected_text), 0)) { goto exit; }