(CCing the list this time.)
On Nov 24, 2013, at 5:37 PM, Bálint Réczey <balint@xxxxxxxxxxxxxxx> wrote:
> I tried to trigger a crash using GCC and -ftrapv without success while
> the clang-compiled binary crashed as expected:
$ gcc -S -O2 test.c
$ mv test.s test.s.noftrapv
$ gcc -ftrapv -S -O2 test.c
$ diff test.s.noftrapv test.s
$
whereas:
$ clang -S -O2 test.c
$ mv test.s test.s.noftrapv
$ clang -ftrapv -S -O2 test.c
$ diff test.s.noftrapv test.s
15,21c15
< leaq L_.str(%rip), %rdi
< movl $-2147483550, %esi ## imm = 0xFFFFFFFF80000062
< xorb %al, %al
< callq _printf
< xorl %eax, %eax
< popq %rbp
< ret
---
> ud2
24,27d17
< .section __TEXT,__cstring,cstring_literals
< L_.str: ## @.str
< .asciz "res:%d\n"
and, if we make it so that the compiler can't do all the work at compile time:
$ cat test.c
#include <limits.h>
#include <stdio.h>
int
foo(int i)
{
int b = 100;
#ifdef TEST
if ((i + b) < i) {
printf("overflow!\n");
return 1;
}
#endif
i += b;
printf("res:%d\n",i);
return i;
}
$ gcc -S -O2 test.c
$ mv test.s test.s.noftrapv
$ gcc -ftrapv -S -O2 test.c
$ diff test.s.noftrapv test.s
$ clang -S -O2 test.c
$ mv test.s test.s.noftrapv
$ clang -ftrapv -S -O2 test.c
$ diff test.s.noftrapv test.s
20a21,22
> jo LBB0_1
> ## BB#2:
29a32,33
> LBB0_1:
> ud2
so GCC is apparently not actually doing overflow checks, while clang is - and "gcc" is, on my machine, actually llvm-gcc:
$ gcc --version
i686-apple-darwin11-llvm-gcc-4.2 (GCC) 4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2336.11.00)
Copyright (C) 2007 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
(you might not be surprised to hear that there's no button under my machine's trackpad - the trackpad is the one-and-only mouse button :-))
On SPARC, there's a TRAPV instruction that traps if the integer overflow condition code bit is set; on 32-bit x86, there's INTO, which traps if the overflow condition code bit is set, but it's illegal in 64-bit x86. JO jumps if the overflow bit is set in both 32-bit and 64-bit mode, and UD2 is an instruction that's defined to trap (in 32-bit and 64-bit mode), so jumping to a UD2 if the overflow flag is set is a valid "trap if overflow" instruction sequence even on x86-64.
I don't know why GCC isn't actually implementing -ftrapv on x86 but is implementing it on SPARC, however; it's a little more work on x86 than just sticking in TRAPVs on SPARC, but not *that* much more.
