Files
sx/issues/0124-large-stack-array-aggregate-ops-crash-llvm.md
agra 837b5d375f fix(0124): large stack arrays lower to in-place access, not first-class values
Two lowering sites materialized a local array as a whole LLVM value;
the legalizer scalarizes each such op into one SelectionDAG node per
element, and at ~64K elements the DAG combiner segfaults
(DAGCombiner::visitMERGE_VALUES → ReplaceAllUsesWith).

- lowerVarDecl: an array-typed `---` initializer emits NO store — the
  slot stays uninitialized instead of receiving a whole-array undef
  store. The tuple zero-init carve-out stays; non-array `---` keeps
  the undef store. The interp is unchanged either way (slots start
  .undef).
- lowerIndexExpr: element reads on an array with addressable storage
  GEP the storage and load one element — the general-expression
  sibling of 0110's lowerFor fix — without value-lowering the object
  (a dead whole-array load would still reach the DAG). Storage-less
  arrays keep the index_get fallback.

Sibling shape filed as 0125: any_to_string's per-array-type arms still
pass the array by value, so a 64K+ array type + any {} print crashes.

Regression: examples/0055-basic-large-stack-array.sx (sx build
segfaulted pre-fix). 22 .ir snapshots re-pinned: removed undef stores
and ig.tmp spills, in-place gep+load (instruction-shape-only churn,
reviewed).
2026-06-12 08:19:20 +03:00

109 lines
4.8 KiB
Markdown

# RESOLVED — 0124: 64K+ stack arrays emit whole-aggregate load/store ops that segfault LLVM
> **RESOLVED** (2026-06-12). Root cause: two lowering sites materialized
> a local array as a first-class LLVM value, which the legalizer
> scalarizes into one SelectionDAG node per element. Fix: (1)
> `lowerVarDecl` (src/ir/lower/stmt.zig) emits NO store for an
> array-typed `---` initializer — the slot stays uninitialized instead
> of receiving a whole-array undef store (tuple zero-init carve-out
> kept; non-array `---` keeps the undef store); (2) `lowerIndexExpr`
> (src/ir/lower/expr.zig) reads elements of an array with addressable
> storage via `index_gep` on the storage + a single-element load — the
> general-expression sibling of 0110's `lowerFor` fix — without
> value-lowering the object (a dead whole-array load would still reach
> the DAG). Storage-less arrays (rvalues, by-value params) keep the
> `index_get` fallback. Residual sibling shapes filed as issue 0125
> (`any_to_string`'s per-array-type arms pass the array by value — any
> 64K+ array type + any `{}` print still crashes).
> Regression test: `examples/0055-basic-large-stack-array.sx`
> ([65536]u8 write/read loops + [131072]s64 first/last — `sx build`
> segfaulted pre-fix). 22 `.ir` snapshots re-pinned (removed undef
> stores / `ig.tmp` spills → in-place gep+load; reviewed
> instruction-shape-only). Gates: zig build test 426/426, suite
> 592/592, distribution repo 14/14.
## Symptom
Declaring a large (~64KB+) stack array in a function reachable from
`main` crashes the compiler during native emission — a segfault inside
libLLVM, not a diagnostic.
- **Observed**: `Segmentation fault at address 0x16b...` (a stack
address) under `sx build`, inside
`DAGCombiner::visitMERGE_VALUES``SelectionDAG::ReplaceAllUsesWith`
(via `LLVMTargetMachineEmitToFile`, src/ir/emit_llvm.zig:2894).
- **Expected**: the program compiles; the array lives in the frame and
is accessed in place.
The crash threshold is DAG-shape dependent, not a clean size boundary
(`[65535]u8` and `[65537]u8` compile, `[65536]u8`, `[66000]u8`,
`[131072]u8` crash), because the real problem is the SelectionDAG
node count: lowering materializes the array as a FIRST-CLASS LLVM
value, and the legalizer scalarizes each whole-aggregate op into one
node per element. Two emission shapes produce such ops:
1. `buf : [N]u8 = ---;` stores a whole-array undef constant
(`store [N x i8] undef, ptr %alloca`) — a store of nothing, for an
explicitly-uninitialized local.
2. `buf[i]` reads on a local array lower as `index_get` on the array
VALUE: load the entire array as an SSA value, spill it to an
`ig.tmp` alloca, GEP one element (the general-expression sibling of
resolved issue 0110, which fixed only `lowerFor`'s element fetch).
Besides the crash, this copies N bytes to read 1.
Each shape crashes llc in isolation on the dumped IR; with both
replaced by in-place access the module compiles.
## Reproduction
```sx
#import "modules/std.sx";
f :: (fd: s32) {
buf : [65536]u8 = ---;
if buf[0] > 0 { out("x\n"); }
}
main :: () -> s32 {
f(1);
return 0;
}
```
Observed at master 7f2b8b5: `sx build` segfaults in libLLVM with the
stack trace above. `sx ir` shows the two whole-aggregate ops:
```llvm
%alloca1 = alloca [65536 x i8], align 1
store [65536 x i8] undef, ptr %alloca1, align 1
%load = load [65536 x i8], ptr %alloca1, align 1
%ig.tmp = alloca [65536 x i8], align 1
store [65536 x i8] %load, ptr %ig.tmp, align 1
%ig.ptr = getelementptr [65536 x i8], ptr %ig.tmp, i64 0, i64 0
```
## Investigation prompt
Two lowering sites produce the whole-aggregate ops; fix both:
1. `src/ir/lower/stmt.zig` `lowerVarDecl` (annotated branch): a
`.undef_literal` initializer falls through to
`lowerExpr(val)``constUndef(array type)``store`. `---` means
explicitly uninitialized — emit NO store at all (keep the existing
tuple zero-init carve-out above it).
2. `src/ir/lower/expr.zig` `lowerIndexExpr`: when the indexed object
is an array with addressable storage (`getExprAlloca` hit, same
guard as 0110's `lowerFor` fix), emit `index_gep` on the storage +
a single-element `load` instead of `index_get` on the loaded array
value. Storage-less arrays (rvalues) keep the `index_get` fallback.
The object must NOT be lowered as a value on the storage path or
the dead whole-array `load` still reaches the DAG.
Verification: the repro builds and runs (prints nothing or `x`
depending on stack garbage — gate on exit 0 of the build, not the
read); `[65535]`/`[65537]`/`[131072]` variants all build. Pin a
regression example that builds AND deterministically runs (write
before read). `zig build && zig build test`,
`bash tests/run_examples.sh` green; expect `.ir` snapshot churn from
removed undef stores and the new gep+load shape — re-pin and review.