-#if !CHIP_HAS_WH64()
-
- ahead32 = CACHE_LINE_SIZE_IN_WORDS;
-
- /* We already prefetched the first and last cache lines, so
- * we only need to do more prefetching if we are storing
- * to more than two cache lines.
- */
- if (n32 > CACHE_LINE_SIZE_IN_WORDS * 2) {
- int i;
-
- /* Prefetch the next several cache lines.
- * This is the setup code for the software-pipelined
- * loop below.
- */
-#define MAX_PREFETCH 5
- ahead32 = n32 & -CACHE_LINE_SIZE_IN_WORDS;
- if (ahead32 > MAX_PREFETCH * CACHE_LINE_SIZE_IN_WORDS)
- ahead32 = MAX_PREFETCH * CACHE_LINE_SIZE_IN_WORDS;
-
- for (i = CACHE_LINE_SIZE_IN_WORDS;
- i < ahead32; i += CACHE_LINE_SIZE_IN_WORDS)
- __insn_prefetch(&out32[i]);
- }
-
- if (n32 > ahead32) {
- while (1) {
- int j;
-
- /* Prefetch by reading one word several cache lines
- * ahead. Since loads are non-blocking this will
- * cause the full cache line to be read while we are
- * finishing earlier cache lines. Using a store
- * here causes microarchitectural performance
- * problems where a victimizing store miss goes to
- * the head of the retry FIFO and locks the pipe for
- * a few cycles. So a few subsequent stores in this
- * loop go into the retry FIFO, and then later
- * stores see other stores to the same cache line
- * are already in the retry FIFO and themselves go
- * into the retry FIFO, filling it up and grinding
- * to a halt waiting for the original miss to be
- * satisfied.
- */
- __insn_prefetch(&out32[ahead32]);
-
-#if CACHE_LINE_SIZE_IN_WORDS % 4 != 0
-#error "Unhandled CACHE_LINE_SIZE_IN_WORDS"
-#endif
-
- n32 -= CACHE_LINE_SIZE_IN_WORDS;
-
- /* Save icache space by only partially unrolling
- * this loop.
- */
- for (j = CACHE_LINE_SIZE_IN_WORDS / 4; j > 0; j--) {
- *out32++ = v32;
- *out32++ = v32;
- *out32++ = v32;
- *out32++ = v32;
- }
-
- /* To save compiled code size, reuse this loop even
- * when we run out of prefetching to do by dropping
- * ahead32 down.
- */
- if (n32 <= ahead32) {
- /* Not even a full cache line left,
- * so stop now.
- */
- if (n32 < CACHE_LINE_SIZE_IN_WORDS)
- break;
-
- /* Choose a small enough value that we don't
- * prefetch past the end. There's no sense
- * in touching cache lines we don't have to.
- */
- ahead32 = CACHE_LINE_SIZE_IN_WORDS - 1;
- }
- }
- }
-
-#else /* CHIP_HAS_WH64() */
-