- blockExpr: call rvalue on void result for unlabeled blocks, matching
upstream AstGen.zig:2431. This was causing a missing STORE_NODE when
empty blocks like {} were used as struct field values with pointer RL.
- identifierExpr: call rvalueNoCoercePreRef after LOAD in local_ptr case,
matching upstream AstGen.zig:8453-8454.
- Implement AST_NODE_ARRAY_MULT (** operator) with ArrayMul payload,
matching upstream AstGen.zig:774-785.
- Enable parser_test.zig and astgen_test.zig corpus tests.
- Enable combined corpus test (all 5 files pass).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Port switchExprErrUnion optimization for both catch and if patterns,
fix missing rvalue call in decl table lookup, fix identAsString
ordering for underscore error captures, and fill value_placeholder
with 0xaa to match upstream undefined pattern.
Enables corpus build.zig test.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add error capture scope to orelseCatchExpr (catch |err| now creates
a ScopeLocalVal for the captured error variable)
- Add @panic, @errorName, @field builtins
- Increase blockExprStmts scope arrays from 64 to 128 entries
(build.zig has 93 var decls in a single block)
corpus build.zig still skipped: needs switchExprErrUnion optimization
(catch |err| switch(err) pattern).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Valgrind doesn't support AVX-512 instructions (EVEX prefix 0x62).
The zig CC generates them for large struct copies on native x86_64
targets even at -O0 (e.g. vmovdqu64 with zmm registers).
Previously only avx512f was subtracted, which was insufficient —
the .evex512 feature (and other AVX-512 sub-features) also need
to be disabled to prevent EVEX-encoded 512-bit instructions.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Implement several interconnected features for function declarations:
- noalias_bits: Track which parameters have the noalias keyword by setting
corresponding bits in a uint32_t (supports up to 32 parameters)
- is_var_args: Detect ellipsis3 (...) token in parameter list
- is_noinline/has_inline_keyword: Detect noinline/inline modifiers
- callconv handling: Extract callconv_expr from fn_proto variants, create
cc_gz sub-block, emit builtin_value for explicit callconv() or inline
- func_fancy instruction: When any of cc_ref, is_var_args, noalias_bits,
or is_noinline are present, emit func_fancy instead of func/func_inferred
with the appropriate FuncFancy payload layout
- fn_var_args: Track in AstGenCtx for body code that checks it
- BuiltinValue constants: Add all ZIR_BUILTIN_VALUE_* defines to zir.h
- addBuiltinValue helper: Emit extended builtin_value instructions
Generic tracking (any_param_used, ret_ty_is_generic, ret_body_param_refs)
is not yet implemented as it requires is_used_or_discarded support in
ScopeLocalVal scope lookups.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Move rvalue calls inside builtinCall (all builtins now call rvalue
internally, matching upstream) and remove outer rvalue wrap from
call site
- Add rlResultTypeForCast that errors when no result type is available,
used by @bitCast, @intCast, @truncate, @ptrCast, @enumFromInt
- Fix @import to compute res_ty from result location instead of
hardcoding ZIR_REF_NONE
- Fix @embedFile to evaluate operand with coerced_ty=slice_const_u8_type
- Fix @cInclude/simpleCBuiltin to check c_import scope and use
comptimeExpr with coerced_ty=slice_const_u8_type
- Fix @cImport to pass actual block_result to ensure_result_used instead
of hardcoded ZIR_REF_VOID_VALUE
Not fixed: Issue 14 (ptrCast nested pointer cast collapsing) — upstream
routes @ptrCast through a dedicated ptrCast() function that walks nested
pointer cast builtins. Currently uses simple typeCast path only.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Port two missing features from upstream AstGen.zig varDecl:
1. Add reachableExprComptime (AstGen.zig:418-438) which wraps init
expressions in comptimeExpr when force_comptime is set, and checks
for noreturn results. Replace plain exprRl calls in all three varDecl
paths (const rvalue, const alloc, var) with reachableExprComptime.
2. Extract comptime_token by scanning backwards from mut_token (matching
Ast.zig fullVarDeclComponents). For const path, set force_comptime to
wrap init in comptime block. For var path, use comptime_token to set
is_comptime which selects alloc_comptime_mut/alloc_inferred_comptime_mut
tags and sets maybe_comptime on the scope.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Port two missing features from upstream AstGen.zig:
1. Handle identifier-named tests (decltest): when the token after `test`
is an identifier, set decl_id to DECL_ID_DECLTEST and record the
identifier string as the test name. Upstream performs full scope
resolution for validation which is skipped here.
2. Add `within_fn` field to AstGenCtx (mirrors AstGen.within_fn). Save,
set to true, and restore in both testDecl and fnDecl. This flag
propagates to maybe_generic on namespace scopes for container
declarations inside function/test bodies.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Port two missing features from upstream AstGen.zig ret() function:
1. Add any_defer_node field to GenZir (AstGen.zig:11812) to track
whether we're inside a defer expression. Set it in defer body
generation and propagate via makeSubBlock. retExpr now checks
this field and errors with "cannot return from defer expression"
(AstGen.zig:8127-8135). Also reorder retExpr checks to match
upstream: fn_block null check first, then any_defer_node check,
then emitDbgNode.
2. Add reachableExpr wrapper (AstGen.zig:408-416) that calls exprRl
and checks refIsNoReturn to detect unreachable code. Use it in
retExpr instead of plain exprRl for the return operand
(AstGen.zig:8185-8186). nameStratExpr is left as TODO since
containerDecl does not yet accept a name_strategy parameter.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add two fixes from audit of ptrTypeExpr against upstream AstGen.zig ptrType:
1. Reject `[*c]allowzero T` with a compile error matching upstream
(AstGen.zig:3840-3842). C pointers always allow address zero, so
the allowzero modifier is invalid on them.
2. Save source_offset/source_line/source_column before typeExpr and
restore them before evaluating each trailing expression (sentinel,
addrspace, align). This ensures correct debug info source locations
matching upstream (AstGen.zig:3844-3846, 3859-3861, 3876-3878,
3885-3887).
Issue 3 (addrspace RL using addBuiltinValue) is skipped as
addBuiltinValue is not yet implemented.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Port assignShift (AstGen.zig:3786) and assignShiftSat (AstGen.zig:3812)
from upstream, handling <<=, >>=, and <<|= operators as both statements
in blockExprStmts and expressions in exprRl. Previously these fell
through to SET_ERROR.
Add grouped_expression unwrapping loop in blockExprStmts (matching
AstGen.zig:2569-2630) so that parenthesized statements like `(x += 1)`
are correctly dispatched to assignment handlers instead of going through
the default unusedResultExpr path.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Fix three issues in the RL annotation pre-pass (rlExpr):
1. Label detection for `inline while`/`inline for` now accounts for
the `keyword_inline` token before checking for `identifier colon`,
matching upstream fullWhileComponents/fullForComponents logic.
2. `assign_destructure` now recurses into variable nodes and the value
expression with RL_RI_NONE, matching upstream behavior instead of
returning false without visiting sub-expressions.
3. `rlTokenIdentEqual` now handles @"..."-quoted identifiers by comparing
the quoted content rather than stopping at the `@` character, which
previously caused all @-quoted identifiers to compare as equal.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The exprRl function was wrapping blockExprExpr's return value in an
extra rvalue() call, but blockExprExpr already applies rvalue internally
for labeled blocks when need_result_rvalue=true. The upstream expr()
function at AstGen.zig:991 returns blockExpr's result directly without
extra rvalue wrapping. This could produce duplicate coercion/store
instructions for non-trivial result locations.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Port two missing checks from upstream AstGen.zig rvalueInner to the C
rvalue function:
1. isAlwaysVoid (Zir.zig:1343-1608): When the result refers to an
instruction that always produces void (e.g., dbg_stmt, store_node,
export, memcpy, etc.), replace the result with void_value before
proceeding. This prevents emitting unnecessary type coercions or
stores on always-void instructions.
2. endsWithNoReturn (AstGen.zig:11068): When the current GenZir block
ends with a noreturn instruction, return the result immediately
without emitting any rvalue instructions. This avoids emitting dead
ZIR instructions after noreturn.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Port the \u{NNNNNN} unicode escape parsing from upstream Zig's
string_literal.zig:parseEscapeSequence into both strLitAsString
(string literal decoding with UTF-8 encoding) and char_literal
(codepoint value extraction). Without this, \u escapes fell through
to the default branch which wrote a literal 'u' character, producing
incorrect ZIR string bytes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
BuiltinCall.Flags has ensure_result_used at bit 1, not bit 3 like
Call/FieldCall. Separate the case to use the correct bit.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three bugs found by auditing against upstream AstGen.zig/AstRlAnnotate.zig:
1. rlExpr: defer was recursing into nd.rhs (always 0) instead of nd.lhs
(the actual deferred expression), so the RL annotation pass never
visited defer bodies.
2. addEnsureResult: compile_error was missing from the noreturn
instruction list, causing spurious ensure_result_used instructions
to be emitted after @compileError calls.
3. blockExprExpr: force_comptime was derived from gz->is_comptime,
but upstream blockExpr always passes force_comptime=false to
labeledBlockExpr. This caused labeled blocks in comptime contexts
to incorrectly emit BLOCK_COMPTIME + BREAK_INLINE instead of
BLOCK + BREAK.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Fix two parser bugs found by auditing against upstream Parse.zig:
1. In parseTypeExpr's while case, the continue expression was parsed
inline as `eatToken(COLON) ? expectExpr : 0` which missed the
required parentheses. Use parseWhileContinueExpr(p) instead,
matching what parseWhileExpr already does.
2. In expectStatement, comptime blocks used parseBlock() which only
matches `{ ... }`. Use parseBlockExpr() to also recognize labeled
blocks like `comptime label: { ... }`.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Match Zig's Signedness enum values (unsigned=1, signed=0) and
reorder int_type struct fields to match Zig's layout:
[src_node, bit_count, signedness, pad].
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Inline index_inst at usage site to narrow scope, initialize
var_init_rl.ctx to RI_CTX_NONE (matching upstream default).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- whileExpr: emit emitDbgNode before condition evaluation to match
upstream AstGen.zig:6579. Fixes astgen_test.zig corpus (1 missing
DBG_STMT).
- Block expressions in exprRl: wrap blockExprExpr result with rvalue()
to handle result location storage (RL_PTR → STORE_NODE, etc.).
Fixes parser_test.zig inst_len to exact match.
- parser_test.zig corpus now has matching inst_len and all tags, but
has 1 int_type data signedness mismatch (pre-existing issue).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Typed struct init empty (SomeType{}) was returning the result directly
without going through rvalue(), missing STORE_NODE/STORE_TO_INFERRED_PTR/
COERCE_PTR_ELEM_TY+REF emissions when result location requires storage.
Reduces parser_test.zig corpus diff from 5 to 1 instruction.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- retExpr: check nodesNeedRl to use RL_PTR with ret_ptr/ret_load instead of
always RL_COERCED_TY with ret_node. Handle .always/.maybe error paths with
load from ptr when needed.
- Use typeExpr() instead of expr()/exprRl() for type sub-expressions in
optional_type, error_union, merge_error_sets, and array elem types in
structInitExpr/arrayInitExpr. This generates BLOCK_COMPTIME wrappers for
non-primitive type identifiers.
- arrayInitExpr: only use ARRAY_INIT_REF for RL_REF (not RL_REF_COERCED_TY),
and pass non-ref results through rvalue().
- slice_sentinel: emit SLICE_SENTINEL_TY and coerce sentinel to that type.
All slice variants: coerce start/end to usize.
- COERCE_PTR_ELEM_TY in rvalue for RL_REF_COERCED_TY.
- rvalueNoCoercePreRef for local variable references.
- structInitExprPtr/arrayInitExprPtr for RL_PTR with OPT_EU_BASE_PTR_INIT.
- Typed struct init: use RL_COERCED_TY with field type for init expressions.
Reduces parser_test.zig corpus diff from 225 to 5 instructions.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- @as builtin: propagate RL_TY with dest_type through exprRl instead of
evaluating operand with RL_NONE and manually emitting as_node. Matches
upstream AstGen.zig lines 8909-8920.
- rlResultType: add missing RL_REF_COERCED_TY case (elem_type extraction).
- continue handler: use AST_NODE_OFFSET_NONE for addBreak operand_src_node
instead of computing node offset. Upstream uses addBreak (not
addBreakWithSrcNode), which writes .none.
- varDecl: set init_rl.src_node = 0 for RL_PTR (upstream leaves
PtrResultLoc.src_node at default .none).
Enables astgen_test.zig corpus test — all corpus tests now pass.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
addInstruction() already returns idx + ZIR_REF_START_INDEX (a ref),
so the extra + ZIR_REF_START_INDEX on the inplace_arith_result_ty path
resulted in a double-offset (+248 instead of +124) being stored in
extra data for += and -= compound assignments.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
