-
Notifications
You must be signed in to change notification settings - Fork 1k
AddressSanitizerZeroBasedShadow
To enable this mode, build all the sources with -fsanitize-address-zero-base-shadow, clang-only.
Instead of computing Shadow = (Mem >> 3) + kOffset we compute Shadow = (Mem >> 3), i.e. kOffset = 0 (see AddressSanitizerAlgorithm).
This mode is possible only if 1/8 of address space starting from the zero page is not occupied at startup and AddressSanitizer can map it for the shadow.
We are aware of only one situation where this works: Linux executables compiled with -fPIC or -fPIE and linked with -pie.
% cat procmaps.cc
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
int main() {
char buff[100];
sprintf(buff, "cat /proc/%d/maps", getpid());
system(buff);
}
Regular mode:
% clang -fsanitize=address ./procmaps.cc && ./a.out
00400000-00421000 r-xp 00000000 fc:00 1573310 /tmp/a.out
00620000-00621000 r--p 00020000 fc:00 1573310 /tmp/a.out
00621000-00622000 rw-p 00021000 fc:00 1573310 /tmp/a.out
00622000-0338e000 rw-p 00000000 00:00 0
ffffffff000-120000000000 rw-p 00000000 00:00 0 <<<<<<<<<<< LowShadow
120000000000-140000000000 ---p 00000000 00:00 0 <<<<<<<<< SHadowGap
140000000000-200000000000 rw-p 00000000 00:00 0 <<<<<<<<< HighShadow
7f1ebd5a9000-7f1ebd75e000 r-xp 00000000 fc:00 1208459 /lib/x86_64-linux-gnu/libc-2.15.so
....
7f1ebe1c2000-7f1ebe1c4000 rw-p 00023000 fc:00 1208920 /lib/x86_64-linux-gnu/ld-2.15.so
7fffccd4b000-7fffccd6c000 rw-p 00000000 00:00 0 [stack]
7fffccdff000-7fffcce00000 r-xp 00000000 00:00 0 [vdso]
Zero shadow offset mode (with -pie):
% clang -fsanitize=address -fsanitize-address-zero-base-shadow -fPIE -pie ./procmaps.cc && ./a.out
00040000-20000000000 ---p 00000000 00:00 0 <<<<<<<<< ShadowGap
20000000000-100000000000 rw-p 00000000 00:00 0 <<<<<<<<< HighShadow
7f343283c000-7f34329f1000 r-xp 00000000 fc:00 1208459 /lib/x86_64-linux-gnu/libc-2.15.so
7f3433679000-7f343367a000 rw-p 00022000 fc:00 1573310 /tmp/a.out
7f343367a000-7f34363e6000 rw-p 00000000 00:00 0
7fff79e32000-7fff79e53000 rw-p 00000000 00:00 0 [stack]
7fff79f1a000-7fff79f1b000 r-xp 00000000 00:00 0 [vdso]
Why is this worth the trouble? Zero shadow offset not just saves one arithmetic instruction, but also saves one register, which is otherwise occupied with the large constant. Compare these two:
% cat load.c
long Foo(long *a) { return *a; }
% clang -fsanitize=address -fsanitize-address-zero-base-shadow -O2 -c load.c && objdump -d load.o
0000000000000000 <Foo>:
0: 50 push %rax
1: 48 89 f8 mov %rdi,%rax
4: 48 c1 e8 03 shr $0x3,%rax
8: 80 38 00 cmpb $0x0,(%rax)
b: 75 05 jne 12 <Foo+0x12>
d: 48 8b 07 mov (%rdi),%rax
10: 5a pop %rdx
% clang -fsanitize=address -O2 -c load.c && objdump -d load.o
0000000000000000 <Foo>:
0: 50 push %rax
1: 48 89 f8 mov %rdi,%rax
4: 48 c1 e8 03 shr $0x3,%rax
8: 48 b9 00 00 00 00 00 movabs $0x100000000000,%rcx
f: 10 00 00
12: 48 09 c1 or %rax,%rcx
15: 80 39 00 cmpb $0x0,(%rcx)
18: 75 05 jne 1f <Foo+0x1f>
1a: 48 8b 07 mov (%rdi),%rax
1d: 5a pop %rdx
1e: c3 retq
The data below reflects the performance gain from zero-based shadow on SPEC CPU 2006.
<BR>
Measured LLVM r174590 (on Google Code) on Intel Xeon W3690 @3.47GHz
400.perlbench, 1200.00, 1179.00, 0.98
401.bzip2, 885.00, 821.00, 0.93
403.gcc, 744.00, 724.00, 0.97
429.mcf, 605.00, 579.00, 0.96
445.gobmk, 843.00, 809.00, 0.96
456.hmmer, 1305.00, 1236.00, 0.95
458.sjeng, 923.00, 898.00, 0.97
462.libquantum, 539.00, 532.00, 0.99
464.h264ref, 1269.00, 1194.00, 0.94
471.omnetpp, 629.00, 598.00, 0.95
473.astar, 676.00, 663.00, 0.98
483.xalancbmk, 493.00, 452.00, 0.92
433.milc, 712.00, 656.00, 0.92
444.namd, 637.00, 592.00, 0.93
447.dealII, 654.00, 631.00, 0.96
450.soplex, 391.00, 369.00, 0.94
453.povray, 458.00, 419.00, 0.91
470.lbm, 388.00, 384.00, 0.99
482.sphinx3, 995.00, 886.00, 0.89
We are experimenting with another way to achieve zero-based offset using -Wl,-Ttext-segment=0x40000000. TODO