ret
.size GNAME(post_signal_tramp),.-GNAME(post_signal_tramp)
\f
- .end
+ .text
+ .align align_8byte,0x90
+ .global GNAME(fast_bzero)
+ .type GNAME(fast_bzero),@function
+
+GNAME(fast_bzero):
+ /* A fast routine for zero-filling blocks of memory that are
+ * guaranteed to start and end at a 4096-byte aligned address.
+ */
+ shr $6, %rsi /* Amount of 64-byte blocks to copy */
+ jz Lend /* If none, stop */
+ mov %rsi, %rcx /* Save start address */
+ movups %xmm7, -16(%rsp) /* Save XMM register */
+ xorps %xmm7, %xmm7 /* Zero the XMM register */
+ jmp Lloop
+ .align 16
+Lloop:
+
+ /* Copy the 16 zeroes from xmm7 to memory, 4 times. MOVNTDQ is the
+ * non-caching double-quadword moving variant, i.e. the memory areas
+ * we're touching are not fetched into the L1 cache, since we're just
+ * going to overwrite the memory soon anyway.
+ */
+ movntdq %xmm7, 0(%rdi)
+ movntdq %xmm7, 16(%rdi)
+ movntdq %xmm7, 32(%rdi)
+ movntdq %xmm7, 48(%rdi)
+
+ add $64, %rdi /* Advance pointer */
+ dec %rsi /* Decrement 64-byte block count */
+ jnz Lloop
+ mfence /* Ensure that the writes are globally visible, since
+ * MOVNTDQ is weakly ordered */
+ movups -16(%rsp), %xmm7 /* Restore the XMM register */
+ prefetcht0 0(%rcx) /* Prefetch the start of the block into cache,
+ * since it's likely to be used immediately. */
+Lend:
+ ret
+ .size GNAME(fast_bzero), .-GNAME(fast_bzero)
+
+\f
+ .end