[sanguineman]

Problem

The text format section on memory instructions says the alignment defaults to the natural alignment, but the grammar right below it does not:

(§6.5.6 of the Release 3.0 PDF, 2026-10-28)

align_N (Talign_ in 6.3-text.instructions.spectec L184-L186) receives the access size N in bytes and should return the alignment exponent that memarg stores. The explicit case converts the byte value m to the exponent n, but the default case returns the byte value N unchanged. For example, i32.load uses Tmemarg_(4) (L192), so i32.load without align= gets ALIGN 4 instead of ALIGN 2.

Memarg_ok requires 2^n <= N/8 with N in bits, i.e. 2^n must not exceed the access size in bytes. With the default, n is the byte size itself, and 2^n <= n never holds. So every load and store that omits align= is invalid under this grammar, for example:

(module (memory 1)
  (func (drop (i32.load (i32.const 0)))))

The reference interpreter follows the prose: it uses the natural alignment as an exponent (lexer.mll L249-L256, e.g. opt a 2 for i32.load).

In Wasm 2.0, align_N returned the byte value and memarg_N converted it with a = 2^n (wg-2.0 text/instructions.rst L288-L295), so ε ⇒ N was correct there. The SpecTec version moved the conversion into Talign_, and #2107 fixed its explicit case to return n, but the default case was left as N.

Fix

Convert the default in the same way as the explicit case:

 grammar Talign_(N) : u32 =
   | "align=" m:Tu64 => n  -- if $(m = 2^n)
-  | eps => N
+  | eps => n  -- if $(N = 2^n)

The rendered grammar becomes ε ⇒ n if N = 2^n, matching the line above it.

This PR changes 6.3-text.instructions.spectec in wasm-3.0 and wasm-latest and updates the spectec/test-*/TEST.md outputs. No test is added, since the reference interpreter's parser already uses the natural alignment.

[rossberg]

Thanks! Can you also backport this to the 1.0 and 2.0 specs?

[sanguineman]

Thanks!
I looked into this, but I don't think there is anything to backport.
The 1.0 and 2.0 SpecTec specs (specification/wasm-1.0, specification/wasm-2.0) don't define the text format; it only exists in SpecTec from 3.0 on.
And the hand-written 1.0 and 2.0 documents are correct: there, align_N returns the byte value and memarg_N converts it with a = 2^n, so the default ε ⇒ N gives the right exponent (wg-1.0 L185-L192, wg-2.0 L288-L295).
Am I missing something?

[rossberg]

Ah, sorry, indeed we haven't recreated the text format for 1.0 and 2.0 in SpecTec yet.