0.9.13.3:

[sbcl.git] / src / runtime / x86-64-assem.S

diff --git a/src/runtime/x86-64-assem.S b/src/runtime/x86-64-assem.S
index c654861..92e0716 100644 (file)
--- a/src/runtime/x86-64-assem.S
+++ b/src/runtime/x86-64-assem.S
@@ -60,6 +60,10 @@
        .global GNAME(call_into_c)
        .type   GNAME(call_into_c),@function
 GNAME(call_into_c):
+        /* ABI requires that the direction flag be clear on function
+         * entry and exit. */
+        cld
+        
        push    %rbp            # Save old frame pointer.
        mov     %rsp,%rbp       # Establish new frame.
 
@@ -166,8 +170,9 @@ Lcall:
        
        /* If the function returned multiple values, it will return to
           this point.  Lose them */
+       jnc     LsingleValue    
        mov     %rbx, %rsp
-       /* A singled value function returns here */
+LsingleValue:  
 
 /* Restore the stack, in case there was a stack change. */
        pop     %rsp            # c-sp
@@ -179,6 +184,10 @@ Lcall:
        pop     %r12
        pop     %rbx
 
+        /* ABI requires that the direction flag be clear on function
+         * entry and exit. */
+        cld
+        
 /* FIXME Restore the NPX state. */
 
        /* return value is already in rax where lisp expects it */
@@ -282,11 +291,7 @@ GNAME(closure_tramp):
        .align  align_8byte
 GNAME(fun_end_breakpoint_guts):
        /* Multiple Value return */
-       jmp     multiple_value_return
-       /* the above jmp is only 2 bytes long, we need to add a nop for 
-        * padding since the single value return convention jumps to original
-        * return address + 3 bytes */
-       nop
+       jc      multiple_value_return
        /* Single value return: The eventual return will now use the
           multiple values return convention but with a return values
           count of one. */
@@ -344,4 +349,44 @@ GNAME(post_signal_tramp):
        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)
+
+        .end