rdi_from_dwarf: use sibling tag attributes to gather root-level tag offsets when present - 15x speedup in the gather on webkit DWARFs...

This commit is contained in:
Ryan Fleury
2026-06-19 20:11:50 -07:00
parent 9f82a6fe91
commit 15601d48d4
2 changed files with 133 additions and 104 deletions
+95 -95
View File
@@ -5,162 +5,162 @@ internal String8List
elf_dump_note(Arena *arena, String8 raw_notes, ELF_Class elf_class, ELF_MachineKind e_machine) elf_dump_note(Arena *arena, String8 raw_notes, ELF_Class elf_class, ELF_MachineKind e_machine)
{ {
Temp scratch = scratch_begin(&arena, 1); Temp scratch = scratch_begin(&arena, 1);
B32 is_bad_parse = 1; B32 is_bad_parse = 1;
String8List strings = {0}; String8List strings = {0};
U64 cursor = 0; U64 cursor = 0;
for (;cursor < raw_notes.size;) { for (;cursor < raw_notes.size;) {
U32 owner_size; U32 owner_size;
U64 owner_size_size = str8_deserial_read_struct(raw_notes, cursor, &owner_size); U64 owner_size_size = str8_deserial_read_struct(raw_notes, cursor, &owner_size);
if (owner_size_size == 0) { goto exit; } if (owner_size_size == 0) { goto exit; }
cursor += owner_size_size; cursor += owner_size_size;
U32 desc_size; U32 desc_size;
U64 desc_size_size = str8_deserial_read_struct(raw_notes, cursor, &desc_size); U64 desc_size_size = str8_deserial_read_struct(raw_notes, cursor, &desc_size);
if (desc_size_size == 0) { goto exit; } if (desc_size_size == 0) { goto exit; }
cursor += desc_size_size; cursor += desc_size_size;
ELF_NoteType note_type; ELF_NoteType note_type;
U64 type_size = str8_deserial_read_struct(raw_notes, cursor, &note_type); U64 type_size = str8_deserial_read_struct(raw_notes, cursor, &note_type);
if (type_size == 0) { goto exit; } if (type_size == 0) { goto exit; }
cursor += type_size; cursor += type_size;
if (cursor + owner_size > raw_notes.size) { goto exit; } if (cursor + owner_size > raw_notes.size) { goto exit; }
String8 owner = str8_cstring_capped(raw_notes.str + cursor, raw_notes.str + cursor + owner_size); String8 owner = str8_cstring_capped(raw_notes.str + cursor, raw_notes.str + cursor + owner_size);
cursor += owner_size; cursor += owner_size;
if (cursor + desc_size > raw_notes.size) { goto exit; } if (cursor + desc_size > raw_notes.size) { goto exit; }
String8 raw_desc = str8_substr(raw_notes, r1u64(cursor, cursor + desc_size)); String8 raw_desc = str8_substr(raw_notes, r1u64(cursor, cursor + desc_size));
cursor += desc_size; cursor += desc_size;
cursor = AlignPow2(cursor, 4); cursor = AlignPow2(cursor, 4);
String8List desc_fmt = {0}; String8List desc_fmt = {0};
String8 note_type_str = {0}; String8 note_type_str = {0};
if (str8_match(owner, str8_lit("GNU"), StringMatchFlag_CaseInsensitive)) { if (str8_match(owner, str8_lit("GNU"), StringMatchFlag_CaseInsensitive)) {
// format description // format description
switch (note_type) { switch (note_type) {
case GNU_NoteType_Abi: { case GNU_NoteType_Abi: {
U64 desc_cursor = 0; U64 desc_cursor = 0;
GNU_AbiTag os = 0; GNU_AbiTag os = 0;
U64 os_size = str8_deserial_read_struct(raw_desc, desc_cursor, &os); U64 os_size = str8_deserial_read_struct(raw_desc, desc_cursor, &os);
if (os_size == 0) { goto exit; } if (os_size == 0) { goto exit; }
cursor += os_size; cursor += os_size;
U32 major = 0; U32 major = 0;
U64 major_size = str8_deserial_read_struct(raw_desc, desc_cursor, &major); U64 major_size = str8_deserial_read_struct(raw_desc, desc_cursor, &major);
if (major_size == 0) { goto exit; } if (major_size == 0) { goto exit; }
cursor += major_size; cursor += major_size;
U32 minor = 0; U32 minor = 0;
U64 minor_size = str8_deserial_read_struct(raw_desc, desc_cursor, &minor); U64 minor_size = str8_deserial_read_struct(raw_desc, desc_cursor, &minor);
if (minor_size == 0) { goto exit; } if (minor_size == 0) { goto exit; }
cursor += minor_size; cursor += minor_size;
U32 sub_minor = 0; U32 sub_minor = 0;
U64 sub_minor_size = str8_deserial_read_struct(raw_desc, desc_cursor, &sub_minor); U64 sub_minor_size = str8_deserial_read_struct(raw_desc, desc_cursor, &sub_minor);
if (sub_minor_size == 0) { goto exit; } if (sub_minor_size == 0) { goto exit; }
cursor += sub_minor_size; cursor += sub_minor_size;
String8 os_str = gnu_string_from_abi_tag(os); String8 os_str = gnu_string_from_abi_tag(os);
if (os_str.size == 0) os_str = str8f(scratch.arena, "0x%x", os); if (os_str.size == 0) os_str = str8f(scratch.arena, "0x%x", os);
str8_list_pushf(scratch.arena, &desc_fmt, "OS: %S, ABI: %u.%u.%u", os_str, major, minor, sub_minor); str8_list_pushf(scratch.arena, &desc_fmt, "OS: %S, ABI: %u.%u.%u", os_str, major, minor, sub_minor);
} break; } break;
case GNU_NoteType_BuildId: { case GNU_NoteType_BuildId: {
String8List build_id = {0}; String8List build_id = {0};
for EachIndex(desc_cursor, desc_size) { for EachIndex(desc_cursor, desc_size) {
U8 v = 0; U8 v = 0;
U64 v_size = str8_deserial_read_struct(raw_desc, desc_cursor, &v); U64 v_size = str8_deserial_read_struct(raw_desc, desc_cursor, &v);
if (v_size == 0) { goto exit; } if (v_size == 0) { goto exit; }
desc_cursor += v_size; desc_cursor += v_size;
str8_list_pushf(scratch.arena, &build_id, "%02x", v); str8_list_pushf(scratch.arena, &build_id, "%02x", v);
}
String8 build_id_str = str8_list_join(scratch.arena, &build_id, 0);
str8_list_pushf(scratch.arena, &desc_fmt, "Build ID: %S", build_id_str);
} break;
case GNU_NoteType_PropertyType0: {
U64 align = elf_class == ELF_Class_64 ? 8 : 4;
for (U64 desc_cursor = 0; desc_cursor < raw_desc.size; ) {
GNU_Property type = 0;
U64 type_size = str8_deserial_read_struct(raw_desc, desc_cursor, &type);
if (type_size == 0) { goto exit; }
desc_cursor += type_size;
U32 size = 0;
U64 size_size = str8_deserial_read_struct(raw_desc, desc_cursor, &size);
if (size_size == 0) { goto exit; }
desc_cursor += size_size;
U32 flags = 0;
if (size == 4) {
U64 flags_size = str8_deserial_read_struct(raw_desc, desc_cursor, &flags);
if (flags_size == 0) { goto exit; }
desc_cursor += flags_size;
} }
String8 build_id_str = str8_list_join(scratch.arena, &build_id, 0);
switch (e_machine) { str8_list_pushf(scratch.arena, &desc_fmt, "Build ID: %S", build_id_str);
case ELF_MachineKind_None: break; } break;
case ELF_MachineKind_X86_64: { case GNU_NoteType_PropertyType0: {
String8 features = gnu_string_from_property_flags_x86(scratch.arena, type, flags); U64 align = elf_class == ELF_Class_64 ? 8 : 4;
str8_list_pushf(scratch.arena, &desc_fmt, "x86 features: %S", features); for (U64 desc_cursor = 0; desc_cursor < raw_desc.size; ) {
} break; GNU_Property type = 0;
default: NotImplemented; break; U64 type_size = str8_deserial_read_struct(raw_desc, desc_cursor, &type);
if (type_size == 0) { goto exit; }
desc_cursor += type_size;
U32 size = 0;
U64 size_size = str8_deserial_read_struct(raw_desc, desc_cursor, &size);
if (size_size == 0) { goto exit; }
desc_cursor += size_size;
U32 flags = 0;
if (size == 4) {
U64 flags_size = str8_deserial_read_struct(raw_desc, desc_cursor, &flags);
if (flags_size == 0) { goto exit; }
desc_cursor += flags_size;
}
switch (e_machine) {
case ELF_MachineKind_None: break;
case ELF_MachineKind_X86_64: {
String8 features = gnu_string_from_property_flags_x86(scratch.arena, type, flags);
str8_list_pushf(scratch.arena, &desc_fmt, "x86 features: %S", features);
} break;
default:{}break;
}
desc_cursor = AlignPow2(desc_cursor, align);
} }
} break;
desc_cursor = AlignPow2(desc_cursor, align); default:{}break;
}
} break;
default: NotImplemented; break;
} }
note_type_str = gnu_string_from_note_type(note_type); note_type_str = gnu_string_from_note_type(note_type);
} else if (str8_match(owner, str8_lit("stapsdt"), StringMatchFlag_CaseInsensitive)) { } else if (str8_match(owner, str8_lit("stapsdt"), StringMatchFlag_CaseInsensitive)) {
if (note_type == ELF_NoteType_STapSdt) { if (note_type == ELF_NoteType_STapSdt) {
U64 desc_cursor = 0; U64 desc_cursor = 0;
U64 addr_size = elf_class == ELF_Class_64 ? 8 : 4; U64 addr_size = elf_class == ELF_Class_64 ? 8 : 4;
U64 pc = 0; U64 pc = 0;
U64 pc_size = str8_deserial_read(raw_desc, desc_cursor, &pc, addr_size, addr_size); U64 pc_size = str8_deserial_read(raw_desc, desc_cursor, &pc, addr_size, addr_size);
if (pc_size == 0) { goto exit; } if (pc_size == 0) { goto exit; }
desc_cursor += pc_size; desc_cursor += pc_size;
U64 base_addr = 0; U64 base_addr = 0;
U64 base_addr_size = str8_deserial_read(raw_desc, desc_cursor, &base_addr, addr_size, addr_size); U64 base_addr_size = str8_deserial_read(raw_desc, desc_cursor, &base_addr, addr_size, addr_size);
if (base_addr_size == 0) { goto exit; } if (base_addr_size == 0) { goto exit; }
desc_cursor += base_addr_size; desc_cursor += base_addr_size;
U64 semaphore = 0; U64 semaphore = 0;
U64 semaphore_size = str8_deserial_read(raw_desc, desc_cursor, &semaphore, addr_size, addr_size); U64 semaphore_size = str8_deserial_read(raw_desc, desc_cursor, &semaphore, addr_size, addr_size);
if (semaphore_size == 0) { goto exit; } if (semaphore_size == 0) { goto exit; }
desc_cursor += semaphore_size; desc_cursor += semaphore_size;
String8 provider = str8_cstring_capped(raw_desc.str + desc_cursor, raw_desc.str + raw_desc.size); String8 provider = str8_cstring_capped(raw_desc.str + desc_cursor, raw_desc.str + raw_desc.size);
desc_cursor += provider.size + 1; desc_cursor += provider.size + 1;
if (desc_cursor > raw_desc.size) { goto exit; } if (desc_cursor > raw_desc.size) { goto exit; }
String8 probe = str8_cstring_capped(raw_desc.str + desc_cursor, raw_desc.str + raw_desc.size); String8 probe = str8_cstring_capped(raw_desc.str + desc_cursor, raw_desc.str + raw_desc.size);
desc_cursor += probe.size + 1; desc_cursor += probe.size + 1;
if (desc_cursor > raw_desc.size) { goto exit; } if (desc_cursor > raw_desc.size) { goto exit; }
String8 args = str8_cstring_capped(raw_desc.str + desc_cursor, raw_desc.str + raw_desc.size); String8 args = str8_cstring_capped(raw_desc.str + desc_cursor, raw_desc.str + raw_desc.size);
desc_cursor += args.size + 1; desc_cursor += args.size + 1;
if (desc_cursor > raw_desc.size) { goto exit; } if (desc_cursor > raw_desc.size) { goto exit; }
str8_list_pushf(scratch.arena, &desc_fmt, "Provider: %S", provider); str8_list_pushf(scratch.arena, &desc_fmt, "Provider: %S", provider);
str8_list_pushf(scratch.arena, &desc_fmt, "Probe: %S", probe); str8_list_pushf(scratch.arena, &desc_fmt, "Probe: %S", probe);
str8_list_pushf(scratch.arena, &desc_fmt, "PC: 0x%I64x", pc); str8_list_pushf(scratch.arena, &desc_fmt, "PC: 0x%I64x", pc);
str8_list_pushf(scratch.arena, &desc_fmt, "Base: 0x%I64x", base_addr); str8_list_pushf(scratch.arena, &desc_fmt, "Base: 0x%I64x", base_addr);
str8_list_pushf(scratch.arena, &desc_fmt, "Semaphore: 0x%I64x", semaphore); str8_list_pushf(scratch.arena, &desc_fmt, "Semaphore: 0x%I64x", semaphore);
str8_list_pushf(scratch.arena, &desc_fmt, "Arguments: %S", args); str8_list_pushf(scratch.arena, &desc_fmt, "Arguments: %S", args);
note_type_str = str8_lit("NT_STAPSDT"); note_type_str = str8_lit("NT_STAPSDT");
} }
} }
if (note_type_str.size == 0) note_type_str = str8f(scratch.arena, "0x%x", note_type); if (note_type_str.size == 0) note_type_str = str8f(scratch.arena, "0x%x", note_type);
str8_list_pushf(arena, &strings, "{"); str8_list_pushf(arena, &strings, "{");
str8_list_pushf(arena, &strings, " Owner: %S", owner); str8_list_pushf(arena, &strings, " Owner: %S", owner);
str8_list_pushf(arena, &strings, " Data Size: 0x%x", desc_size); str8_list_pushf(arena, &strings, " Data Size: 0x%x", desc_size);
@@ -173,9 +173,9 @@ elf_dump_note(Arena *arena, String8 raw_notes, ELF_Class elf_class, ELF_MachineK
} }
str8_list_pushf(arena, &strings, "}"); str8_list_pushf(arena, &strings, "}");
} }
is_bad_parse = 0; is_bad_parse = 0;
exit:; exit:;
if (is_bad_parse) { if (is_bad_parse) {
str8_list_pushf(arena, &strings, "ERROR: unable to parse data @ 0x%Ix64", cursor); str8_list_pushf(arena, &strings, "ERROR: unable to parse data @ 0x%Ix64", cursor);
} }
@@ -187,10 +187,10 @@ internal String8List
elf_dump(Arena *arena, String8 raw_elf, ELF_DumpSubsetFlags flags) elf_dump(Arena *arena, String8 raw_elf, ELF_DumpSubsetFlags flags)
{ {
Temp scratch = scratch_begin(&arena, 1); Temp scratch = scratch_begin(&arena, 1);
String8List strings = {0}; String8List strings = {0};
ELF_Bin elf = elf_bin_from_data(scratch.arena, raw_elf); ELF_Bin elf = elf_bin_from_data(scratch.arena, raw_elf);
if (flags & ELF_DumpSubsetFlag_Note) { if (flags & ELF_DumpSubsetFlag_Note) {
for EachIndex(sect_idx, elf.shdrs.count) { for EachIndex(sect_idx, elf.shdrs.count) {
ELF_Shdr64 *shdr = &elf.shdrs.v[sect_idx]; ELF_Shdr64 *shdr = &elf.shdrs.v[sect_idx];
@@ -205,11 +205,11 @@ elf_dump(Arena *arena, String8 raw_elf, ELF_DumpSubsetFlags flags)
} }
} }
} }
String8 out = str8_list_join(arena, &strings, &(StringJoin){.sep=str8_lit("\n"), .post=str8_lit("\n")}); String8 out = str8_list_join(arena, &strings, &(StringJoin){.sep=str8_lit("\n"), .post=str8_lit("\n")});
String8List result = {0}; String8List result = {0};
str8_list_push(arena, &result, out); str8_list_push(arena, &result, out);
scratch_end(scratch); scratch_end(scratch);
return result; return result;
} }
+38 -9
View File
@@ -1002,14 +1002,36 @@ d2r_convert(Arena *arena, D2R_ConvertParams *params)
DW2_Tag tag = {0}; DW2_Tag tag = {0};
off += dw2_read_tag(scratch3.arena, raw, unit_parse_ctx, raw->sec[DW_SectionKind_Info].data, off, &tag); off += dw2_read_tag(scratch3.arena, raw, unit_parse_ctx, raw->sec[DW_SectionKind_Info].data, off, &tag);
//- rjf: do tree navigations //- rjf: look for sibling attribute fast path
if(tag.has_children) B32 sibling_encoded = 0;
U64 sibling_off = 0;
for EachNode(n, DW2_AttribNode, tag.attribs.first)
{ {
depth += 1; if(n->v.attrib_kind == DW_AttribKind_Sibling)
{
sibling_off = dw2_reference_info_off_from_form_val(unit_parse_ctx, &n->v.val);
sibling_encoded = 1;
break;
}
} }
if(tag.kind == DW_TagKind_Null)
//- rjf: if we have a sibling link, follow it -> depth stays the same
if(sibling_encoded && sibling_off >= off)
{ {
depth -= 1; off = sibling_off;
}
//- rjf: otherwise, do tree navigations...
else
{
if(tag.has_children)
{
depth += 1;
}
if(tag.kind == DW_TagKind_Null)
{
depth -= 1;
}
} }
scratch_end(scratch3); scratch_end(scratch3);
@@ -1149,7 +1171,7 @@ d2r_convert(Arena *arena, D2R_ConvertParams *params)
for EachInRange(unit_idx, range) for EachInRange(unit_idx, range)
{ {
Rng1U64 unit_info_tag_range = unit_info_tag_ranges[unit_idx]; Rng1U64 unit_info_tag_range = unit_info_tag_ranges[unit_idx];
unit_deduped_tag_maps[unit_idx].slots_count = dim_1u64(unit_info_tag_range) / 256 + 1; unit_deduped_tag_maps[unit_idx].slots_count = dim_1u64(unit_info_tag_range) / 32 + 1;
unit_deduped_tag_maps[unit_idx].slots = push_array(scratch.arena, D2R_UnitDedupedTagNode *, unit_deduped_tag_maps[unit_idx].slots_count); unit_deduped_tag_maps[unit_idx].slots = push_array(scratch.arena, D2R_UnitDedupedTagNode *, unit_deduped_tag_maps[unit_idx].slots_count);
} }
lane_sync(); lane_sync();
@@ -1168,6 +1190,7 @@ d2r_convert(Arena *arena, D2R_ConvertParams *params)
{ {
break; break;
} }
ProfBegin("gather unique tags work");
Temp work_scratch = scratch_begin(&scratch.arena, 1); Temp work_scratch = scratch_begin(&scratch.arena, 1);
//- rjf: unpack work //- rjf: unpack work
@@ -1188,6 +1211,7 @@ d2r_convert(Arena *arena, D2R_ConvertParams *params)
U64 container_ancestor_info_off; U64 container_ancestor_info_off;
U64 hash_seed; U64 hash_seed;
}; };
U64 tags_to_dedup_count = 0;
D2R_TagNode *first_tag_to_dedup = 0; D2R_TagNode *first_tag_to_dedup = 0;
D2R_TagNode *last_tag_to_dedup = 0; D2R_TagNode *last_tag_to_dedup = 0;
for(U64 root_tag_idx = origin_unit_root_tag_idx_range.min; for(U64 root_tag_idx = origin_unit_root_tag_idx_range.min;
@@ -1265,6 +1289,7 @@ d2r_convert(Arena *arena, D2R_ConvertParams *params)
n->container_ancestor_info_off = container_ancestor_info_off; n->container_ancestor_info_off = container_ancestor_info_off;
n->hash_seed = top_parent ? top_parent->hash_seed : 0; n->hash_seed = top_parent ? top_parent->hash_seed : 0;
SLLQueuePush(first_tag_to_dedup, last_tag_to_dedup, n); SLLQueuePush(first_tag_to_dedup, last_tag_to_dedup, n);
tags_to_dedup_count += 1;
} }
// rjf: compute hash seed for this tag // rjf: compute hash seed for this tag
@@ -1321,6 +1346,7 @@ d2r_convert(Arena *arena, D2R_ConvertParams *params)
} }
//- rjf: hash all tags we need to deduplicate & gather //- rjf: hash all tags we need to deduplicate & gather
U64 total_tag_count_this_work = 0;
for(D2R_TagNode *tag_n = first_tag_to_dedup; tag_n != 0; tag_n = tag_n->next) for(D2R_TagNode *tag_n = first_tag_to_dedup; tag_n != 0; tag_n = tag_n->next)
{ {
Temp dedup_root_scratch = scratch_begin(&scratch.arena, 1); Temp dedup_root_scratch = scratch_begin(&scratch.arena, 1);
@@ -1331,6 +1357,7 @@ d2r_convert(Arena *arena, D2R_ConvertParams *params)
//- rjf: hash all tags & dependency tag trees //- rjf: hash all tags & dependency tag trees
U64 hash = hash_seed; U64 hash = hash_seed;
ProfScope("hash tag trees & dependencies")
{ {
typedef struct TagTask TagTask; typedef struct TagTask TagTask;
struct TagTask struct TagTask
@@ -1373,6 +1400,7 @@ d2r_convert(Arena *arena, D2R_ConvertParams *params)
// rjf: read tag // rjf: read tag
DW2_Tag tag = {0}; DW2_Tag tag = {0};
t_off += dw2_read_tag(tag_scratch.arena, raw, unit_parse_ctx, raw->sec[DW_SectionKind_Info].data, t_off, &tag); t_off += dw2_read_tag(tag_scratch.arena, raw, unit_parse_ctx, raw->sec[DW_SectionKind_Info].data, t_off, &tag);
total_tag_count_this_work += 1;
// rjf: determine if tag should be skipped from hashing // rjf: determine if tag should be skipped from hashing
B32 should_skip_tag = (tag.kind == DW_TagKind_LexicalBlock || B32 should_skip_tag = (tag.kind == DW_TagKind_LexicalBlock ||
@@ -1624,7 +1652,8 @@ d2r_convert(Arena *arena, D2R_ConvertParams *params)
} }
scratch_end(dedup_root_scratch); scratch_end(dedup_root_scratch);
} }
// ProfMsg("work_idx %I64u: %I64u tags encountered", work_idx, total_tag_count_this_work);
ProfEnd();
scratch_end(work_scratch); scratch_end(work_scratch);
} }
lane_sync(); lane_sync();
@@ -2896,7 +2925,7 @@ d2r_convert(Arena *arena, D2R_ConvertParams *params)
{ {
U64 voff_base = lopc_attrib->val.addr - base_vaddr; U64 voff_base = lopc_attrib->val.addr - base_vaddr;
U64 voff_opl = 0; U64 voff_opl = 0;
if(dw_attrib_class_from_form_kind(unit_parse_ctx->version, unit_parse_ctx->exts, hipc_attrib->val.kind) & (1<<DW_AttribClass_Address)) if(dw_attrib_class_from_form_kind(unit_parse_ctx->version, unit_parse_ctx->exts, hipc_attrib->val.kind) & DW_AttribClass_Address)
{ {
voff_opl = voff_base + hipc_attrib->val.u128.u64[0]; voff_opl = voff_base + hipc_attrib->val.u128.u64[0];
} }
@@ -3078,7 +3107,7 @@ d2r_convert(Arena *arena, D2R_ConvertParams *params)
{ {
U64 voff_base = lopc_attrib->val.addr - base_vaddr; U64 voff_base = lopc_attrib->val.addr - base_vaddr;
U64 voff_opl = 0; U64 voff_opl = 0;
if(dw_attrib_class_from_form_kind(unit_parse_ctx->version, unit_parse_ctx->exts, hipc_attrib->val.kind) & (1<<DW_AttribClass_Address)) if(dw_attrib_class_from_form_kind(unit_parse_ctx->version, unit_parse_ctx->exts, hipc_attrib->val.kind) & DW_AttribClass_Address)
{ {
voff_opl = hipc_attrib->val.addr; voff_opl = hipc_attrib->val.addr;
} }