Clang produces crashing code with -nostdlib

I am playing around with setting up my own runtime environment for an executable, and I can't get clang (v3.4-1ubuntu1, target: x86_64-pc-linux-gnu) to produce an executable that doesn't segfault. I have reduced the problem to the following:

If I have a file crt1.c that does nothing except satisfy the linker requirement for a _start symbol:

_start(char *arguments, ...)

Then I compile it with clang -nostdlib crt1.c, it produces the following executable (from objdump -d a.out):

a.out:     file format elf64-x86-64

Disassembly of section .text:

0000000000400150 <_start>:
  400150:   55                      push   %rbp
  400151:   48 89 e5                mov    %rsp,%rbp
  400154:   48 81 ec f0 00 00 00    sub    $0xf0,%rsp
  40015b:   84 c0                   test   %al,%al
  40015d:   0f 29 bd 30 ff ff ff    movaps %xmm7,-0xd0(%rbp)
  400164:   0f 29 b5 20 ff ff ff    movaps %xmm6,-0xe0(%rbp)
  40016b:   0f 29 ad 10 ff ff ff    movaps %xmm5,-0xf0(%rbp)
  400172:   0f 29 a5 00 ff ff ff    movaps %xmm4,-0x100(%rbp)
  400179:   0f 29 9d f0 fe ff ff    movaps %xmm3,-0x110(%rbp)
  400180:   0f 29 95 e0 fe ff ff    movaps %xmm2,-0x120(%rbp)
  400187:   0f 29 8d d0 fe ff ff    movaps %xmm1,-0x130(%rbp)
  40018e:   0f 29 85 c0 fe ff ff    movaps %xmm0,-0x140(%rbp)
  400195:   48 89 bd b8 fe ff ff    mov    %rdi,-0x148(%rbp)
  40019c:   4c 89 8d b0 fe ff ff    mov    %r9,-0x150(%rbp)
  4001a3:   4c 89 85 a8 fe ff ff    mov    %r8,-0x158(%rbp)
  4001aa:   48 89 8d a0 fe ff ff    mov    %rcx,-0x160(%rbp)
  4001b1:   48 89 95 98 fe ff ff    mov    %rdx,-0x168(%rbp)
  4001b8:   48 89 b5 90 fe ff ff    mov    %rsi,-0x170(%rbp)
  4001bf:   0f 84 5b 00 00 00       je     400220 <_start+0xd0>
  4001c5:   0f 28 85 c0 fe ff ff    movaps -0x140(%rbp),%xmm0
  4001cc:   0f 29 85 70 ff ff ff    movaps %xmm0,-0x90(%rbp)
  4001d3:   0f 28 8d d0 fe ff ff    movaps -0x130(%rbp),%xmm1
  4001da:   0f 29 4d 80             movaps %xmm1,-0x80(%rbp)
  4001de:   0f 28 95 e0 fe ff ff    movaps -0x120(%rbp),%xmm2
  4001e5:   0f 29 55 90             movaps %xmm2,-0x70(%rbp)
  4001e9:   0f 28 9d f0 fe ff ff    movaps -0x110(%rbp),%xmm3
  4001f0:   0f 29 5d a0             movaps %xmm3,-0x60(%rbp)
  4001f4:   0f 28 a5 00 ff ff ff    movaps -0x100(%rbp),%xmm4
  4001fb:   0f 29 65 b0             movaps %xmm4,-0x50(%rbp)
  4001ff:   0f 28 ad 10 ff ff ff    movaps -0xf0(%rbp),%xmm5
  400206:   0f 29 6d c0             movaps %xmm5,-0x40(%rbp)
  40020a:   0f 28 b5 20 ff ff ff    movaps -0xe0(%rbp),%xmm6
  400211:   0f 29 75 d0             movaps %xmm6,-0x30(%rbp)
  400215:   0f 28 bd 30 ff ff ff    movaps -0xd0(%rbp),%xmm7
  40021c:   0f 29 7d e0             movaps %xmm7,-0x20(%rbp)
  400220:   48 8b 85 b0 fe ff ff    mov    -0x150(%rbp),%rax
  400227:   48 89 85 68 ff ff ff    mov    %rax,-0x98(%rbp)
  40022e:   48 8b 8d a8 fe ff ff    mov    -0x158(%rbp),%rcx
  400235:   48 89 8d 60 ff ff ff    mov    %rcx,-0xa0(%rbp)
  40023c:   48 8b 95 a0 fe ff ff    mov    -0x160(%rbp),%rdx
  400243:   48 89 95 58 ff ff ff    mov    %rdx,-0xa8(%rbp)
  40024a:   48 8b b5 98 fe ff ff    mov    -0x168(%rbp),%rsi
  400251:   48 89 b5 50 ff ff ff    mov    %rsi,-0xb0(%rbp)
  400258:   48 8b bd 90 fe ff ff    mov    -0x170(%rbp),%rdi
  40025f:   48 89 bd 48 ff ff ff    mov    %rdi,-0xb8(%rbp)
  400266:   4c 8b 85 b8 fe ff ff    mov    -0x148(%rbp),%r8
  40026d:   4c 89 45 f8             mov    %r8,-0x8(%rbp)
  400271:   48 81 c4 f0 00 00 00    add    $0xf0,%rsp
  400278:   5d                      pop    %rbp
  400279:   c3                      retq   

The executable crashes with a segmentation fault at the instruction at address 40015d--the one that saves away %xmm7. I don't know why clang is saving these away, gcc produces no such instructions.

The value passed in %rbp is 7fffffffe588, which is not 16-byte aligned, which I guess in some sense explains the segmentation fault. But how would I get this to work? Get it to suppress those save instructions? Get it to align the rbp pointer somehow?

EDIT: I guess this problem comes down to the fact that the code that clang is producing assumes that %rsp is going to be 16-byte aligned. Is that a valid assumption to make? Why is it not true in this example?


EDIT: found a solution - skip to see below...

First, ensure that stack (%rsp) is 16-byte aligned:

pushq  %rbp
movq   %rsp, %rbp
andq   $-0x10, %rsp  ; rsp = rsp & 0xffffffffffffff0

This is problematic, as it's normally the caller's responsibility to ensure %rsp is 16-byte aligned, as %rbp + 16.n might not be on a 16-byte boundary. So perhaps movq %rsp, %rbp should appear after the alignment of %rsp.

sub $0xf0, %rsp allocates 0xf0 byte of stack space; 0xf0 being a multiple of (16). If %rsp is not 16-byte aligned, movaps %xmm7, -0xd0(%rbp) => movaps %xmm7, 0x20(%rsp). In other words, the SSE register is stored at %rsp + 32. If not aligned, this raises a 'general protection exception', i.e., a segfault.

Another issue you might encounter are reads/writes to -0x170(%rbp) => that's -0x80(%rsp), which is either on (or past? I might be out in my offsets) the boundary of the red zone. As this is a leaf function, you are free to use it, but not write past it.

Note: if your function was called, you should subtract another (8) bytes from %rsp to ensure 16-byte alignment. This will in turn affect the offsets for (%ebp).

I'm not fact checking with the ABI standard here, I may have made some mistakes; so it might be best to check with the x86-64 SysV ABI (section 3.2).

SOLUTION: compiling the function above with the -mstackrealign flag explicitly aligns %rsp to a 16-byte boundary. I'm using clang on OS X, which is basically the same as x86-64 SysV (x86-64 ELF / Linux) with respect to calling conventions and alignment requirements:

clang -nostdlib -mstackrealign -c crt1.c

0000000000000000        pushq   %rbp
0000000000000001        movq    %rsp, %rbp
0000000000000004        andq    $0xfffffffffffffff0, %rsp
000000000000000b        subq    $0x170, %rsp
0000000000000012        testb   %al, %al

BTW - this avoids the %rbp issue entirely by making all loads/stores relative to %rsp. Consequently, there is no use of the red zone - at least with Apple's LLVM 3.3 based clang.

Need Your Help

Arabic language is not supported in CentOS

php linux unicode centos arabic

I am not able to save Arabic text when I moved my project to a new server (CentOS 6). The coding and database structure are the same. But in my new server it is printing just like ????.

How can I take screenshots of webpages with Perl?


Is it possible to write a script in Perl that opens different URLs and saves a screenshot of each of them?