[thesyncim]

Loads can already be at i+e, e+i or i-e for a loop-invariant e. This allows the same for stores, which covers row-wise code like:

for j := 0; j < n; j++ { b[i*ldb+j] = a[i*lda+j] }        // gonum dlacpy
for j := 0; j < k; j++ { c[i*ldc+j] -= work[i*ldwork+j] } // gonum dlarfb
for x := range w { line[off+x] = v }

Every store to a given slice has to use the same offset, so each iteration still writes its own element. Two different offsets are rejected with a new reason. For loads of a stored slice, the distance is now measured from the store offset instead of from i. The same offset means the same element. A constant positive distance in a forward loop is loaded before the chunk's stores, as today. A negative distance, a non-constant one, or a loop that counts down is rejected. The store gets its own window, cut at the offset, so the vector loop never writes an element the scalar loop doesn't; the length check and the overlap window include the offset. The if-select and copy-builtin paths take the offset too, and a constant offset that puts the first stored index below zero is rejected, as mauri870#2 does for loads.

With loopvec -json -methods (I read every added loop), std goes from 28 to 29 vectorized loops (an image/jpeg fill), gonum from 161 to 177 (dlacpy, dlaset, dlascl and dlarfb rows, plus some testlapack helpers), and goav1 from 81 to 86. TSVC s423 and s424 now vectorize and match their goldens. The generated code for loops that were already vectorized doesn't change.

offset_store.txt runs 9 shapes over 14 sizes, plus short slices that must panic and aliased views that must fall back, and both builds have to print the same output. offset_store_no_rewrite.txt pins the rejected cases.