Avoid string instructions in x86_64 memset

newlib/libc/machine/x86_64/memset.S

Index: newlib/libc/machine/x86_64/memset.S
diff --git a/newlib/libc/machine/x86_64/memset.S b/newlib/libc/machine/x86_64/memset.S
index c16cd58106ee87c78b82f640a273e63496ddcca4..a46dbfcbf6a63174045b4269f5055d95e73a673b 100644
--- a/newlib/libc/machine/x86_64/memset.S
+++ b/newlib/libc/machine/x86_64/memset.S
@@ -13,78 +13,119 @@
   .global SYM (memset)
   SOTYPE_FUNCTION(memset)
 
+ /*
+  * Avoid revealing the sandbox base address. In particular, we never read from
+  * r11, and avoid using the string instructions (stos) because they leave the
+  * full 64 bits in rdi.
+  */
+
 SYM (memset):
   movl    edi, r9d                /* Save return value */
   movl    esi, eax
   movl    edx, ecx
   cmpl    $16, edx
-  jb      byte_set
+  jb      .Lbyte_set
 
   movl    edi, r8d                /* Align on quad word boundary */
   andl    $7, r8d
-  jz      quadword_aligned
+  jz      .Lquadword_aligned
   movl    $8, ecx
   subl    r8d, ecx
   subl    ecx, edx
-  rep     stosb al, %nacl:(rdi), r15
+.Lheader_loop:
+  movb    al, %nacl:(r15, rdi)
+  inc     edi
+  dec     ecx
+  jnz     .Lheader_loop
+
   movl    edx, ecx
 
-quadword_aligned:
+.Lquadword_aligned:
   movabs  $0x0101010101010101, r8
   movzbl  sil, eax
   imul    r8, rax
   cmpl    $256, edx
-  jb      quadword_set
+  jb      .Lquadword_set
 
-  shrl    $7, ecx                 /* Store 128 bytes at a time with minimum cache polution */
+  shrl    $7, ecx                 /* Store 128 bytes at a time */
 
+  pushq   $0
+  movl    ebp, (rsp)
   .p2align 4
-loop:
-  movntiq rax, %nacl:     (r15,rdi)
-  movntiq rax, %nacl:   8 (r15,rdi)
-  movntiq rax, %nacl:  16 (r15,rdi)
-  movntiq rax, %nacl:  24 (r15,rdi)
-  movntiq rax, %nacl:  32 (r15,rdi)
-  movntiq rax, %nacl:  40 (r15,rdi)
-  movntiq rax, %nacl:  48 (r15,rdi)
-  movntiq rax, %nacl:  56 (r15,rdi)
-  movntiq rax, %nacl:  64 (r15,rdi)
-  movntiq rax, %nacl:  72 (r15,rdi)
-  movntiq rax, %nacl:  80 (r15,rdi)
-  movntiq rax, %nacl:  88 (r15,rdi)
-  movntiq rax, %nacl:  96 (r15,rdi)
-  movntiq rax, %nacl: 104 (r15,rdi)
-  movntiq rax, %nacl: 112 (r15,rdi)
-  movntiq rax, %nacl: 120 (r15,rdi)
+.Lquadword_aligned_loop:
+  naclrestbp edi, r15
+  movq    rax,     (rbp)
+  movq    rax, 8   (rbp)
+  movq    rax, 16  (rbp)
+  movq    rax, 24  (rbp)
+  movq    rax, 32  (rbp)
+  movq    rax, 40  (rbp)
+  movq    rax, 48  (rbp)
+  movq    rax, 56  (rbp)
+  movq    rax, 64  (rbp)
+  movq    rax, 72  (rbp)
+  movq    rax, 80  (rbp)
+  movq    rax, 88  (rbp)
+  movq    rax, 96  (rbp)
+  movq    rax, 104 (rbp)
+  movq    rax, 112 (rbp)
+  movq    rax, 120 (rbp)
 
   leal    128 (rdi), edi

[jvoung (off chromium), 2015/02/27 20:33:52]

Might have been able to add to ebp and then restore r15 together w/ a custom
bundle_lock/unlock, rather than go through edi (if you start the loop after
setting rbp to edi + 15)?

[Derek Schuff, 2015/02/27 21:03:19]

I like it. Done.

 
   dec     ecx
-  jnz     loop
+  jnz     .Lquadword_aligned_loop
+
+  popq    r8
+  naclrestbp r8d, r15
 
-  sfence
   movl    edx, ecx
   andl    $127, ecx
-  rep     stosb al, %nacl:(rdi), r15
+  jz      .Lfooter_loop_end
+.Lfooter_loop:
+  movb    al, %nacl:(r15, rdi)
+  inc     edi
+  dec     ecx
+  jnz      .Lfooter_loop
+.Lfooter_loop_end:
   movl    r9d, eax
   pop     r11
   nacljmp r11d, r15
 
 
-byte_set:
-  rep     stosb al, %nacl:(rdi), r15
+.Lbyte_set:
+  testl   ecx, ecx
+  jz      .Lbyte_set_end
+.Lbyte_set_loop:
+  movb    al, %nacl:(r15, rdi)
+  inc     edi
+  dec     ecx
+  jnz     .Lbyte_set_loop
+.Lbyte_set_end:
   movl    r9d, eax
   pop     r11
   nacljmp r11d, r15
 
 
-quadword_set:
+.Lquadword_set:
   shrl    $3, ecx
+  jz      .Lquadword_loop_end
   .p2align 4
-  rep     stosq rax, %nacl:(rdi), r15
+.Lquadword_loop:
+  movq    rax, %nacl:(r15, rdi)
+  add     $8, edi
+  dec     ecx
+  jnz     .Lquadword_loop
+.Lquadword_loop_end:
   movl    edx, ecx
   andl    $7, ecx
-  rep     stosb al, %nacl:(rdi), r15 /* Store the remaining bytes */
+  jz      .Lquadword_footer_end
+.Lquadword_footer:
+  movb    al, %nacl:(r15, rdi)
+  inc     edi
+  dec     ecx
+  jnz     .Lquadword_footer
+.Lquadword_footer_end:
   movl    r9d, eax
   pop     r11
   nacljmp r11d, r15