diff --git a/issues/0125-any-to-string-array-arms-by-value.md b/issues/0125-any-to-string-array-arms-by-value.md new file mode 100644 index 0000000..1e49088 --- /dev/null +++ b/issues/0125-any-to-string-array-arms-by-value.md @@ -0,0 +1,90 @@ +# 0125 — any_to_string's array arms materialize every interned array type by value + +## Symptom + +A program that (a) interns any large (~64KB+) array type and (b) uses +`{}` formatting anywhere — `print("{}\n", 5)` of a plain int is enough — +crashes `sx build` inside libLLVM (`DAGCombiner::visitMERGE_VALUES` → +`SelectionDAG::ReplaceAllUsesWith`), and makes `sx run` (-O0) take ~18s +to compile a trivial file. The two triggers are independent: the array +need never be printed, sliced, or passed anywhere near the format call. + +- **Observed**: segfault under `sx build`; multi-second compiles under + `sx run`. +- **Expected**: formatting an int is unaffected by an unrelated large + array type; printing the array itself formats in place. + +Root cause shape: `any_to_string`'s comptime type-switch +(library/modules/std/fmt.sx, `case array:` arm) expands one arm per +interned array type, and each arm is +`array_to_string(cast(type) val)`: + +1. the `cast(type) val` unbox loads the WHOLE array from the Any + payload pointer (`coerceFromI64`, src/ir/emit_llvm.zig ~2240, + `ua.load`), +2. the call passes the array BY VALUE to the `array_to_string` mono, +3. the mono spills its by-value param to an alloca and (since the + param is an SSA value, not addressable storage) reads elements via + `index_get` on the value — copy-whole-array per element. + +LLVM's legalizer scalarizes each whole-aggregate op into one +SelectionDAG node per element; at ~64K elements the DAG combiner +recurses to death (the sibling of issue 0124, which fixed the +local-variable shapes: `---` undef store and index reads on +addressable storage). + +## Reproduction + +```sx +#import "modules/std.sx"; + +f :: () { + buf : [65536]u8 = ---; + buf[0] = 1; + out(string.{ ptr = @buf[0], len = 1 }); +} + +main :: () -> s32 { + f(); + print("{}\n", 5); + return 0; +} +``` + +Observed (with 0124's fix in place): `sx build` segfaults in libLLVM; +`sx ir` shows the giant arm inside `@any_to_string`: + +```llvm +%ua.load = load [65536 x i8], ptr %ua.ptr, align 1 +%call = call { ptr, i64 } @array_to_string__AR_65536_u8(ptr %0, [65536 x i8] %ua.load) +``` + +## Investigation prompt + +The fix needs the array formatting chain to never materialize the +array as a first-class value. The Any payload for an array IS a +pointer to its storage (that is what `coerceFromI64` intToPtr+loads), +so the arm has everything it needs to format in place. Plausible +routes, most contained first: + +1. Lower the `case array:` arm to a slice view: box the payload + pointer + the array's element count as a `[]elem` and call + `slice_to_string` (slices unbox as a 16-byte {ptr,len} — no giant + ops). Needs the element type at arm-expansion time — the comptime + type-switch already has the concrete array TypeId in hand; an + `element_type(T)`-style comptime accessor may need to be added for + the sx-level spelling, or the arm can be synthesized in the + compiler where both pieces are known. +2. Teach `array_to_string :: (a: $T)` monos (and the unbox `cast`) an + indirect ABI for array-typed params — bigger blast radius: touches + call emission, param spills, and many `.ir` snapshots. + +Suspected files: src/ir/lower/comptime.zig / lower/call.zig (the +type-switch arm expansion and `cast(type)` lowering), +src/ir/emit_llvm.zig `coerceFromI64`, +library/modules/std/fmt.sx (`any_to_string`, `array_to_string`). + +Verification: the repro builds and runs printing `5`; printing the +array itself (`print("{}\n", buf)` on a small array) still renders +element lists (pinned by 0101/0904 et al.); `zig build test` and +`bash tests/run_examples.sh` green; the repro pinned as an example.