EDG Diagnostic Identifier Catalogue

The EDG frontend embedded in CICC v13.0 carries a second diagnostic naming layer separate from the numeric rule IDs surfaced in SARIF output. Roughly 4,539 snake_case lowercase string identifiers live in the binary's read-only string segment, organised into prefix families that mirror EDG's internal diagnostic source tree. These identifiers are the symbolic names EDG itself uses when an internal lookup needs to materialise a diagnostic record — the numeric ID at offset +176 of the diagnostic record (see Diagnostics & Optimization Remarks) is derived from these names at table-build time.

This catalogue documents the identifier scheme, prefix taxonomy, suppression mechanics, and deprecation behaviour so a reimplementation can pre-allocate the same name space and produce byte-compatible SARIF output. It is the symbolic counterpart to the numeric -W##### flag inventory.

Why the symbolic names matter

A faithful reimplementation has to satisfy three constraints that the numeric ID alone cannot express:

1. SARIF stability across versions. EDG's SARIF emitter writes ruleId as the numeric form, but downstream tools (CI dashboards, lint aggregators) frequently key on the symbolic name pulled out of the rendered message text. NVIDIA's nv_diag_* pragma family also accepts symbolic suppression in some configurations.
2. #pragma diag_suppress keyed lookups. The dispatch function sub_67D520 consults a per-source-range suppression map. The map key derived from a #pragma diag_suppress directive is the symbolic identifier hashed through the same path used at diagnostic emission time. Two reimplementations cannot produce matching suppression behaviour unless they agree on the identifier set.
3. Deprecation cascades. Several identifiers (abi_tag_ignored, attribute_deprecated_with_message, the entire c23_*_deprecated set) trigger a second diagnostic to be enqueued as a child note. The child's identifier is fixed by the parent's name, not by the parent's numeric ID. Reordering or renaming the parent breaks the child's "see also" linkage.

Identifier Grammar

Every identifier in the catalogue obeys a uniform grammar that downstream parsers can rely on:

identifier := family_prefix "_" descriptor_tail
family_prefix := "abi" | "align" | "ambig" | "ambiguous" | "attr" | "attribute"
              |  "bad" | "c11" | "c23" | "cli" | "cl"
              |  "cppcli" | "cppcx" | "constexpr" | "cuda" | "cxx"
              |  "default" | "exp" | "ifc" | "invalid" | "ms"
              |  "nv" | "nv_abi" | "nv_atomic" | "nv_diag"
              |  "no" | "nonstd" | "omp" | "openmp" | "pragma"
              |  "ptr" | "strict" | "template" | "type" | "vector" | ...
descriptor_tail := snake_case description, may include digits and embedded keywords

The family prefix is itself sometimes a two-segment compound (nv_abi_pragma_*, nv_atomic_load_*, nv_diag_*) — the parser must allow the longest-match prefix rather than splitting on the first underscore.

Prefix Family Inventory

The 200+ prefix families fall into roughly nine functional clusters. The tables below list the most populated families with representative members and the role they play in dispatch.

Language Standard / Dialect Diagnostics

The cl_* family is unique: these diagnostics fire during command-line parsing before the EDG diagnostic dispatch is fully initialised, so they bypass the normal suppression path. They are always emitted via the early-exit error path in lgenfe_main.

Microsoft-Compatible Language Modes

The cl_cppcli_only_in_microsoft_cplusplus and cl_cppcx_and_cppcli identifiers are gating diagnostics emitted when these modes are requested outside Microsoft compatibility — they are cl_* family members rather than cppcli_* or cppcx_* because they fire during CLI parsing.

CUDA / NVIDIA-Specific Diagnostics

The six nv_diag_* strings (nv_diag_default, nv_diag_error, nv_diag_once, nv_diag_remark, nv_diag_suppress, nv_diag_warning) plus the sibling nv_diagnostic are not diagnostic identifiers — they are the pragma directive keywords recognised by the lexer. The parser reads #pragma nv_diag_suppress <id> and uses <id> (which may be a numeric value or a symbolic name from this catalogue) as the suppression key.

Pragma Handling

The #pragma diagnostic push / #pragma diagnostic pop pair is parsed separately and produces the literal error message no '#pragma diagnostic push' was found to match this 'diagnostic pop' rather than a snake_case identifier — these are among the few diagnostics with no catalogue entry.

Template and Concept Diagnostics

The two spellings ambiguous_ (long form) and ambig_ (abbreviated) are not synonyms — they refer to different diagnostic record types. ambig_suppresses_copy_asgn_decl is emitted as a child note attached to a parent ambiguous_assignment_operator diagnostic. The abbreviated form indicates "this diagnostic exists only as a sub-entry of a longer diagnostic chain". Reimplementations that fold the two into one prefix will produce different child-list structures than the original.

Attributes and ABI

constexpr / consteval Diagnostics

The constexpr_* family is the single largest functional cluster (111 identifiers). It mirrors the structure of the EDG constant evaluator's error states one-to-one: each evaluation failure path has its own identifier, and identifiers ending in _pos (constexpr_allocation_pos) are companion records that carry the position of an earlier note rather than freestanding diagnostics.

OpenMP and Parallelism

The omp_* strings are not diagnostic identifiers. They are predeclared OpenMP runtime function names that the EDG header emits when --openmp is enabled. They appear in the string table because the binary builds OpenMP support, but they belong to the runtime API layer rather than the diagnostic catalogue.

Module-Boundary (IFC) Diagnostics

The IFC family fires during module import when the imported BMI's table-of-contents disagrees with what the current translation unit expects. These are always fatal — there is no per-diagnostic suppression because an inconsistent module file would produce silent miscompilation.

Generic Bad/Invalid/Expected Diagnostics

The no_* family is the inverse of a normal diagnostic family: rather than reporting that something happened, these identifiers report that a feature is unavailable. They fire when source code uses a construct that has been disabled via --no_<feature> on the command line, and the corresponding identifier name typically embeds the flag name itself (no_aligned_new ↔ --no_aligned_new).

Severity Mapping

EDG severity is stored at offset +180 of the diagnostic record as a single byte. The identifier name does not directly encode severity — the same identifier can be emitted at different severities depending on context and -W* flags. The severity byte is written into the record by the allocator sub_67D610 after passing through a per-diagnostic override hook sub_67C4B0 that may reduce it (the if (sev <= 7u) guard around the hook prevents overrides from touching fatal severities 9..11). The dispatcher sub_6837D0 then routes on the stored value, treating 9..11 as fatal via (unsigned int)(severity) - 9 <= 2:

The SARIF level mapping is the ground truth (switch table inside sub_6837D0): case 4 → "remark", case 5 → "warning", case 9 → "catastrophe", case 11 → "internal_error", all other in-range values fall through to "error". Severity 2 records exist only as children attached via sub_6855B0(diag_num, …) (which internally calls sub_67D610(…, 2, …)); they are emitted alongside their parent but the SARIF emitter has no case 2u and sub_721090() aborts if a sev-2 record reaches the switch as a top-level diagnostic.

See Diagnostics for the SARIF / terminal mapping per value. Severity 9, 10, and 11 force the compiler to abort the current translation unit immediately after rendering. The dispatch logic in sub_6837D0 increments qword_4F074B0 (error count) and qword_4F074B8 (catastrophic count), and when their sum reaches unk_4F07478 (error limit, set by --error_limit), the localized message string at index 1508 ("error limit reached") is printed via fprintf("%s\n", sub_67C860(1508)) and sub_7235F0(9) aborts. The 1508 here is a string-table index used by sub_67C860, not a diagnostic record number — the error-limit branch fprintfs the rendered message directly and exits without going through sub_67D610 to allocate a numbered diagnostic.

Suppression Mechanics

Five suppression channels stack in priority order — once any channel suppresses a diagnostic, the lower-priority channels are not consulted:

1. Severity floor. byte_4F07481[0] holds the minimum severity to emit. Diagnostics below this threshold are dropped before any other check.
2. Duplicate dedup. byte_4CFFE80[4*errnum+2] carries per-numeric-ID bit flags; identical diagnostic numbers within the same source range are counted once and emitted once.
3. #pragma diag_suppress. sub_67D520 walks a per-source-position list of active suppressions installed by #pragma nv_diag_suppress. The pragma accepts both numeric IDs and the symbolic names from this catalogue.
4. Command-line -W flags. The flag table built at CLI parse time maps numeric IDs and prefix patterns (like -Wno-template-*) to severity overrides.
5. Error-limit cutoff. When the running error count reaches --error_limit, the dispatcher fprintfs the "error limit reached" message (localized string index 1508 via sub_67C860) and calls sub_7235F0(9) to abort. Further diagnostics are never queued because the process exits in this path.

The nv_diag_default pragma resets a previously-modified diagnostic to its compile-time default severity. The nv_diag_once pragma installs a once-per-translation-unit suppression after the first emission.

Deprecation Behaviour

Identifiers ending in _deprecated, _ignored, or starting with c23_*_deprecated participate in a uniform deprecation chain. When the parent diagnostic fires:

1. The parent is emitted at severity 4 or 5 (remark / warning) with the deprecation message text.
2. A child note is enqueued in note_list with severity 2 (sub-note, attached via sub_6855B0) carrying a "use X instead" hint. The child's identifier is derived from the parent by suffix replacement.
3. If --Werror=deprecated is in effect, the parent is promoted to severity 7 (error) and the child note remains at severity 2 (notes are not promoted alongside their parent in the current dispatcher).

The attribute_deprecated_with_message identifier is special: it accepts a custom message string supplied via [[deprecated("...")]] and embeds it into the rendered output via format placeholder %s. The identifier itself is fixed; only the message text varies.

QUIRKs

QUIRK 1: Two namespaces with identical-looking prefixes. The omp_* and strict_f* identifiers look like diagnostic identifiers (they match the snake_case pattern) but belong to entirely different subsystems. omp_* are OpenMP runtime function names emitted into predeclared headers; strict_f* are LLVM strict-FP intrinsic operation names appearing in SelectionDAG lowering tables. Treating either as a diagnostic identifier — passing either to #pragma nv_diag_suppress — will produce a "diagnostic number not found" error from sub_67D520 because the build-time identifier-to-number resolution skips both prefixes. The 4,539 count given in this page is for the raw string match; the true diagnostic count is closer to 4,200 once these two namespaces are excluded.

QUIRK 2: ambig_* records are not standalone diagnostics. The five ambig_* identifiers (ambig_literal_operator, ambig_suppresses_copy_asgn_decl, ambig_suppresses_copy_ctor_decl, and two related entries) are emitted only as child notes of a parent ambiguous_* diagnostic. They have their own diagnostic numbers and can be individually suppressed by number, but suppressing the parent silently suppresses all ambig_* children attached to it — even if the children are configured as nv_diag_error and would otherwise be promoted. The interaction is one-directional: promoting an ambig_* child has no effect on the parent's severity.

QUIRK 3: The _pos suffix marks position-only continuation records. Several constexpr_* identifiers come in pairs: constexpr_allocation_mismatch is the primary diagnostic carrying the error message, and constexpr_allocation_pos is a continuation record carrying only a source-location annotation pointing at the original allocation site. The continuation record's +180 severity byte is always 2 (note), and its +96 has_source_location is always nonzero. Reimplementations that emit only the primary diagnostic without the _pos companion produce technically-valid output but lose the secondary source-position annotation that IDEs rely on to draw the "see previous allocation here" link. The _pos records are essentially severity-2 placeholders whose entire purpose is to carry a second source location that the EDG diagnostic record format has no other slot for.

Reimplementation Checklist

A reimplementation that aims for byte-identical SARIF output must, in order:

1. Lift the identifier set. Extract the snake_case lowercase strings from the binary's read-only string table at the addresses CICC v13.0 places them. The total cardinality is 4,539 raw matches; the diagnostic-only subset (excluding omp_* runtime names and strict_f* LLVM intrinsics) is ~4,200.
2. Preserve prefix grammar. The longest-match prefix rule matters: nv_abi_pragma_bad_format parses as family nv_abi_pragma with tail bad_format, not as nv + abi_pragma_bad_format. The suppression key derivation depends on getting the family split right.
3. Assign numeric IDs deterministically. EDG's identifier-to-number resolution runs once at init and assigns sequential numeric IDs in a fixed traversal order. Changing the order changes every ruleId in SARIF output and every -W##### flag.
4. Wire the ambig_* ↔ ambiguous_* parent-child linkages. The five abbreviated identifiers must be attached to their long-form parents in child_list, not emitted as top-level diagnostics.
5. Wire the _pos continuation records. Every constexpr_*_pos identifier must be attached as a severity-2 note to the corresponding primary constexpr_* diagnostic, carrying only a source location.
6. Implement the suppression stack. All five priority levels (severity floor, dedup, #pragma diag_suppress, -W flags, error-limit cutoff) must be present and ordered correctly. Missing any level produces visible behavioural differences.
7. Honour the deprecation child-emission rule. Identifiers that emit a deprecation warning must enqueue a severity-2 sub-note (via sub_6855B0, which internally allocates a severity-2 record through sub_67D610) with the matching "use X instead" identifier. The --Werror=deprecated promotion lifts the parent from severity 4/5 (remark/warning) to severity 7 (error); the child note stays at severity 2 because the dispatcher promotes only the parent record.

Cross-References

• Diagnostics & Optimization Remarks — record layout, dispatch architecture, terminal and SARIF renderers.
• CICC Pipeline — where EDG diagnostic emission sits in the overall compilation flow.
• Methodology — how the string table was mined and how identifiers were grouped.