IL Display

The IL display subsystem produces a human-readable text dump of the entire Intermediate Language graph. It is compiled from EDG's il_to_str.c (source path /dvs/p4/build/sw/rel/gpgpu/toolkit/r13.0/compiler/drivers/compiler/edg/EDG_6.6/src/il_to_str.c, confirmed by an assertion at line 6175 in form_float_constant). The display code occupies address range 0x5EB290--0x603A00 in the binary (roughly 90KB), with the main dispatch functions at 0x5EC600--0x5F7FD0 and formatting helpers continuing through 0x6039E0.

Activation is via the il_display CLI flag (flag index 10 in the boolean flag table), which triggers display_il_file after the frontend completes parsing. The output goes to stdout through an indirectable callback mechanism (qword_126F980). When active, every IL entry in every memory region is printed with labeled fields, 25-column-aligned formatting, and scope/address annotations.

Key Facts

Top-Level Control Flow

display_il_file (sub_5F7DF0)
│
│  printf("Display of IL file \"%s\", produced by the compilation of \"%s\"\n",
│         il_file_name, source_file_name)
│
├── display_il_header (sub_5F76B0)
│   │  dword_126FA30 = 1                              // file-scope mode
│   │  puts("\n\nIntermediate language for memory region 1 (file scope):")
│   │  puts("\nil_header:")
│   │  ... 30+ header fields ...
│   │
│   └── walk_file_scope_il(display_il_entry, ...)      // sub_60E4F0
│       └── display_il_entry (sub_5F4930)              // callback per entity
│
└── for region = 2 .. dword_126EC80:
        dword_126FA30 = 0                              // function-scope mode
        // lookup function name from scope table
        printf("\n\nIntermediate language for memory region %ld (function \"%s\"):\n",
               region, function_name)
        walk_routine_scope_il(region, display_il_entry, ...)   // sub_610200
        └── display_il_entry (sub_5F4930)              // callback per entity

Memory region 1 is always file-scope (global declarations, types, templates). Regions 2+ correspond to individual function bodies. The scope table at qword_126EB90 maps each region index to its owning scope entry; the display code checks scope.kind == 17 (sck_function) and extracts the routine name for the banner.

IL Header Fields

display_il_header (sub_5F76B0) prints the translation-unit-level metadata stored in BSS at 0x126EB60--0x126EBF8:

After printing the header, display_il_header calls walk_file_scope_il (sub_60E4F0) with display_il_entry as the per-entity callback. This iterates every IL entry in file-scope region 1.

The Main Dispatcher: display_il_entry

display_il_entry (sub_5F4930, 1,686 lines) is the central per-entity display function. It receives an entry pointer and an entry kind byte, and dispatches to the appropriate per-kind display function.

Transparent (Inline) Kinds

The first switch handles kinds that are displayed inline by their parent and should not appear as standalone entries. These return immediately without output:

Transparent kinds (early return):
  4   routine_type_supplement      15  (reserved)
  5   routine_type_extra           19  try_supplement
  14  (reserved)                   20  asm_supplement
  27  template_parameter_suppl     34  (reserved)
  37  (reserved)                   38  (reserved)
  46  (reserved)                   47  (reserved)
  48  (reserved)                   75  (reserved)
  76  (reserved)                   77  (reserved)
  78  (reserved)                   81  (reserved)

Entry Header Line

For non-transparent kinds, the dispatcher prints a scope-annotated header:

file-scope type@7f3a4b200100
func-scope variable@7f3a4b300200

The scope prefix comes from dword_126FA30 (1 = "file-scope", 0 = "func-scope"). The kind name is looked up from off_E6DD80[kind_byte]. The address is the raw entry pointer value. For entries in function-scope regions while dword_126FA30 == 1, a warning "**NON FILE SCOPE PTR**" is emitted.

Dispatch Table

The second switch dispatches to specialized display functions:

Per-Kind Display Functions

source_file_entry (Kind 1)

Displayed inline in the dispatcher. Fields:

source_corresp (Shared Prefix)

All named entities (variable, routine, type, field, label, namespace, template_param) share a source_corresp sub-record, printed by display_source_corresp (sub_5EDF40, 170 lines). This is the first thing displayed for each such entity:

source_corresp:
  name:                    foo
  unmangled_name_or_mangled_encoding: _Z3foov
  decl_position.seq:       42
  decl_position.column:    5
  name_references:         name_reference@7f3a...
  is_class_member:         TRUE
  access:                  public
  parent_scope:            file-scope scope@7f3a...
  enclosing_routine:       NULL
  referenced:              TRUE
  needed:                  TRUE
  name_linkage:            external

Fields displayed by display_source_corresp:

type (Kind 6)

display_type (sub_5F06B0, 1,033 lines) handles all 22 type kinds. After calling display_source_corresp, it prints common type fields then switches on the type kind byte at offset +132:

Common type fields:

Type kind switch (offset +132):

variable (Kind 7)

display_variable (sub_5EE500, 614 lines) is one of the most field-heavy display functions. After display_source_corresp, it prints:

CUDA-specific variable fields:

Boolean flags (approximately 50 flags spanning bytes 144--208):

is_weak, is_weakref, is_gnu_alias, has_gnu_used_attribute, has_gnu_abi_tag_attribute, is_not_common, is_common, has_internal_linkage_attribute, asm_name_is_valid, used, address_taken, is_parameter, is_parameter_pack, is_pack_element, is_enhanced_for_iterator, initializer_in_class, constant_valued, is_thread_local, extends_lifetime, is_template_param_object, compiler_generated, is_in_class_specialization, is_handler_param, is_this_parameter, referenced_non_locally, modified_within_try_block, is_template_variable, is_prototype_instantiation, is_nonreal, is_specialized, specialized_with_old_syntax, explicit_instantiation, class_explicitly_instantiated, explicit_do_not_instantiate, param_value_has_been_changed, param_used_more_than_once, is_anonymous_parent_object, is_member_constant, is_constexpr, declared_constinit, is_inline, suppress_inline_definition, superseded_external, has_variably_modified_type, is_vla, is_compound_literal, has_explicit_initializer, has_parenthesized_initializer, has_direct_braced_initializer, has_flexible_array_initializer, declared_with_auto_type_specifier, declared_with_decltype_auto, declared_with_class_template_placeholder

routine (Kind 11)

display_routine (sub_5EF1A0, ~1,160 lines) is the single largest per-kind display function. After display_source_corresp:

CUDA-specific routine flags (byte 182):

C99-specific fields (displayed when dword_126EBA8 == 1 and std_version > 199900):

fp_contract, fenv_access, cx_limited_range -- pragma state values from off_A6F460 (4 entries).

Boolean flags (approximately 60 flags spanning bytes 176--191):

address_taken, is_virtual, overrides_base_member, pure_virtual, final, override, covariant_return_virtual_override, is_inline, is_declared_constexpr, is_constexpr, is_constexpr_intrinsic, compiler_generated, defined, called, is_explicit_constructor, is_explicit_conversion_function, is_trivial_default_constructor, is_trivial_copy_function, is_trivial_destructor, is_initializer_list_ctor, is_delegating_ctor, is_inheriting_ctor, assignment_to_this_done, is_prototype_instantiation, is_template_function, is_specialized, specialized_with_old_syntax, explicit_instantiation, class_explicitly_instantiated, explicit_do_not_instantiate, has_nodiscard_attribute, never_throws, is_in_class_specialization, never_inline, is_pure, is_initialization_routine, is_finalization_routine, is_weak, is_weakref, is_gnu_alias, is_ifunc, has_gnu_used_attribute, has_gnu_abi_tag_attribute, in_gnu_abi_tag_namespace, allocates_memory, no_instrument_function, no_check_memory_usage, always_inline, gnu_c89_inline, implicit_alias, has_internal_linkage_attribute, contains_try_block, contains_local_class_type, superseded_external, defined_in_friend_decl, contains_statement_expression, inline_in_class_definition, is_lambda_body, is_defaulted, is_deleted, contains_local_static_variable, is_raw_literal_operator, is_tls_init_routine, has_deducible_return_type, has_deduced_return_type, contains_generic_lambda, is_coroutine, is_top_level_in_mem_region, friend_defined_in_instantiation, is_ineligible, definition_needed, defined_outside_of_parent, trailing_requires_clause

expr_node (Kind 13)

display_expr_node (sub_5ECFE0, 534 lines) handles 36 expression node kinds. Common expression fields are printed first:

Expression kind switch (offset +24):

Every expression case ends with dump_source_position("position", ...) to record the source location.

statement (Kind 21)

display_statement (sub_5EC600, 328 lines) handles 26 statement kinds. Common fields first:

Statement kind switch (offset +32):

Coroutine statement (case 9) displays the full C++20 coroutine lowering structure:

traits, handle, promise, init_await_resume, this_param_copy,
paramter_copies, final_suspend_label, initial_suspend_call,
final_suspend_call, unhandled_exception_call, get_return_object_call,
new_routine, delete_routine, ...

The field name "paramter_copies" (missing the 'e' in "parameter") is a typo preserved verbatim from the EDG source. This confirms the display strings originate from Edison Design Group's own il_to_str.c -- a reimplementation would spell it correctly.

scope (Kind 23)

display_scope (sub_5F2140, 177 lines) handles 9 scope kinds:

Common scope fields: next, parent, kind

Boolean flags: do_not_free_memory_region, is_constexpr_routine, is_stmt_expr_block, is_placeholder_scope, needed_walk_done

Child entity lists: assoc_block, lifetime, constants, types, variables, nonstatic_variables, labels, routines, asm_entries, scopes

Conditional lists (controlled by bitmask tests on scope kind):

// Bitmask 0x20044 = bits 2+6+17 = sck_block + sck_class_struct_union + sck_function
// Bitmask 0x9     = bits 0+3    = sck_file + sck_namespace

if ((1LL << kind) & 0x20044) {
    // display: namespaces, using_declarations, using_directives
}
if ((1LL << kind) & 0x9) {
    // display: namespaces, using_declarations, using_directives
}
// Also: dynamic_inits, local_static_variable_inits (function/block scopes)
//       expr_node_refs, scope_refs, vla_dimensions (function scope + C mode)
//       pragmas, hidden_names, templates, source_sequence_list, src_seq_sublist_list

constant (Kind 2)

display_constant (sub_5F2720, 605 lines) handles 16 constant kinds. After display_source_corresp, common fields include next, type, orig_type, expr, and approximately 25 boolean flags.

Constant kind switch (offset +148):

dynamic_init (Kind 30)

display_dynamic_init (sub_5F37F0, 248 lines) handles 9 dynamic initialization kinds:

Common fields: next, variable, destructor, lifetime, next_in_destruction_list, unordered, init_expr_lifetime, and approximately 20 boolean flags including static_temp, follows_an_exec_statement, inside_conditional_expression, has_temporary_lifetime, is_constructor_init, is_freeing_of_storage_on_exception, overlaps_temps_in_inner_lifetime, is_reused_value, is_creation_of_initializer_list_object, master_entry.

class_info (Kind 39)

display_class_type_supplement (sub_5F4030, 366 lines) is not dispatched directly from the kind table but called by display_type when the type kind is class/struct/union (kinds 9/10/11). It prints the class supplement record:

Formatting Infrastructure

25-Column Field Labels

dump_field_label (sub_5EB2A0, 22 lines) is the universal field label formatter. It prints "field_name:" then pads with spaces to column 25. If the label plus colon exceeds 24 characters, it prints a newline first to avoid misalignment:

storage_class:           static
alignment:               16
is_constexpr:            TRUE

This produces the consistent columnar output visible in all IL dumps.

Boolean Fields

dump_field_bool (sub_5EB450, 25 lines) prints a label and "TRUE" or "FALSE":

is_virtual:              TRUE
pure_virtual:            FALSE

Source Position Fields

dump_source_position (sub_5EB4E0, 82 lines) prints position as two sub-fields when the position is non-zero (seq != 0 or column != 0):

position.seq:            42
position.column:         5

Reads a 32-bit sequence number at *position and a 16-bit column at *(position + 4).

IL Pointer Annotations

dump_il_entity_pointer (sub_5EB8B0, 99 lines) is the most comprehensive pointer formatter. For each IL entity pointer, it prints:

1. Scope prefix: "file-scope" or "func-scope" (from bit 0 of the entry prefix byte at entry_ptr - 8)
2. Kind name: from off_E6DD80[kind_byte]
3. Hex address: @%lx
4. Entity name (kind-dependent):
   • Kinds with name at offset +8 (bitmask 0x2000000010001984): prints the name string
   • Kind 12 (label): prints "label " prefix + name
   • Kind 6 (type): calls qualified name formatter
   • Kind 2 (constant): calls type display
   • Kind 0x40 (seq_number_lookup): prints qualified name from offset +0
   • Kind with bit 36 set: prints qualified name from offset +40, plus "in" context from +56

primary_source_file:     file-scope source_file_entry@7f3a4b100020 "test.cu"
main_routine:            file-scope routine@7f3a4b200100 "main"

The variant dump_il_string_pointer (sub_5EB670) prints the same format but includes the string value from the pointed-to entry. A scope mismatch (e.g., function-scope pointer found during file-scope display) triggers a "**NON FILE SCOPE PTR**" warning.

Entity List Display

display_entity_list (sub_5EC450, 87 lines) walks a linked list of entity pointers and prints each with scope/kind/address annotations:

entities:                file-scope variable@7f3a... "x"
                         func-scope variable@7f3a... "y"

It follows the next link at offset 0 of each list node until NULL.

String Literal Display

dump_string_value (sub_5EB300, 41 lines) prints string values with proper escape handling:

• NULL pointers print "NULL"
• Non-printable characters are printed as octal escapes (\OOO)
• Backslash and double-quote are backslash-escaped (\\, \")
• The octal mask width is controlled by dword_126E49C (CHAR_BIT equivalent, typically 8)

file_name:               "test.cu"
full_name:               "/home/user/project/test.cu"

Float Constant Formatting

form_float_constant (sub_5F7FD0, 302 lines) handles float-to-string conversion with EDG-specific formatting. An assertion at line 6175 guards against buffer overflow (63-byte limit).

Float kind suffixes:

Special value handling:

• NaN: __builtin_nanf(""), __builtin_nan(""), etc. (when compiler version > 30299)
• Infinity: __builtin_huge_valf() or (__extension__ 0x1.0p<exp>f)
• Division form: (f/0.0f) or (f/(0,0.0f)) (C++ vs C modes, selected by dword_126E1D8/dword_126E1E8)
• User-defined literals: (funcname("string_value")) form

Data Tables Referenced

The display subsystem relies on approximately 20 string-to-enum lookup tables in the .rodata segment:

All tables use the same bounds-checking pattern:

const char *name = "**BAD STORAGE CLASS**";
if ((unsigned char)value <= 6u)
    name = storage_class_names[value];
puts(name);

Out-of-range values produce "**BAD <KIND>**" sentinel strings, which serve as diagnostic markers for corrupted IL.

Global State

Helper Functions (0x5F8000--0x6039E0)

The display subsystem includes approximately 50 additional helper functions in the address range beyond the main dispatchers:

The "paramter_copies" Typo

The coroutine statement display (case 9 in display_statement) prints the field label "paramter_copies" -- missing the 'e' in "parameter." This typo is present in the compiled binary's string table and originates from Edison Design Group's source code. It serves as strong provenance evidence: a clean-room reimplementation would not reproduce this exact spelling error, confirming that cudafe++ links genuine EDG il_to_str.c object code.

Complete Call Graph

display_il_file (sub_5F7DF0) ─── TOP LEVEL
├── display_il_header (sub_5F76B0)
│   ├── init_output_context (sub_5F9040)
│   ├── dump_il_entity_pointer (sub_5EB8B0) ×30+ for header fields
│   ├── dump_field_bool (sub_5EB450) ×15+ for header booleans
│   ├── dump_string (sub_5EB790)
│   └── walk_file_scope_il (sub_60E4F0)
│       └── display_il_entry (sub_5F4930) ─── callback per entity
│
└── [loop over regions 2..N]
    └── walk_routine_scope_il (sub_610200)
        └── display_il_entry (sub_5F4930) ─── callback per entity

display_il_entry (sub_5F4930) ─── MAIN DISPATCHER
├── display_source_corresp (sub_5EDF40) ─── shared by named entities
├── display_statement (sub_5EC600) ─── case 0x15
│   ├── display_coroutine_info (sub_600870)
│   └── display_for_range (sub_600740)
├── display_expr_node (sub_5ECFE0) ─── case 0x0D
│   ├── display_expr_operator (sub_5FE8B0)
│   ├── display_expr_operator_name (sub_5FB270)
│   └── display_expr_details (sub_5FBCD0)
├── display_variable (sub_5EE500) ─── case 0x07
│   └── display_init_kind (sub_5EBB50)
├── display_routine (sub_5EF1A0) ─── case 0x0B
│   └── display_template_arg_list (sub_5EBF60 / sub_5FA0D0)
├── display_type (sub_5F06B0) ─── case 0x06
│   ├── display_class_supplement (sub_5F4030)
│   ├── display_int_type_kind (sub_5F9110)
│   └── display_float_type_kind (sub_5F93D0)
├── display_scope (sub_5F2140) ─── case 0x17
├── display_constant (sub_5F2720) ─── case 0x02
│   ├── display_integer_constant (sub_602F20)
│   └── display_float_constant (sub_5FCAF0)
├── display_dynamic_init (sub_5F37F0) ─── case 0x1E
├── display_name_reference (sub_5EBC60) ─── case 0x3E
└── display_entity_list (sub_5EC450) ─── multiple cases

display_single_entity (sub_5F7D50) ─── TARGETED DISPLAY
├── entity_lookup (sub_73D400)
├── resolve_entity (sub_7377D0)
├── get_entity_kind (sub_5C64C0)
├── init_output_context (sub_5F9040)
└── display_il_entry (sub_5F4930)

Relationship to Other Subsystems

The IL display subsystem is read-only: it never modifies the IL graph. It shares the same entry walker functions used by the IL Tree Walking framework (walk_file_scope_il = sub_60E4F0, walk_routine_scope_il = sub_610200) and the Keep-in-IL mark phase, but passes display_il_entry as the callback instead of a transformation function.

The IL Allocation subsystem provides dump_il_table_stats (sub_5E99D0), which dumps allocation counters rather than IL content -- a complementary diagnostic activated separately.

The field offsets printed by the display functions serve as ground truth for the IL Overview entry kind table and the Entity Node Layout documentation.