LLVM optimisation bug or undefined behaviour?

c, clang, llvm

Solution

I am quite sure this is an optimizer bug. It repro's in LLVM-2.7 and LLVM-3.1, the only versions I have access to.

I posted a bug to the LLVM Bugzilla.

The bug is demonstrated by this SSCCE:

#include <stdio.h>

unsigned long int * table;

void foo( unsigned long int q )
{
  unsigned long int i,j,e;

  e = 0;
  for (i = 1; i <= 256; i *= q)
    e++;
  e--;

  for (i = 0; i < q; i++)
    for (j = 0; j < e; j++)
      ((unsigned char*)(table) + 13 )[i*e + j] = 0;   // bug here
}

int main() {
    unsigned long int v[8];
    int i;
    memset(v, 1, sizeof(v));

    table = v;
    foo(2);

    for(i=0; i<sizeof(v); i++) {
        printf("%d", ((unsigned char*)v)[i]);
    }
    puts("");
    return 0;
}

It should print

1111111111111000000000000000011111111111111111111111111111111111

under GCC and "clang -O0". The incorrect output observed with LLVM is

0000000011111111111110000000011111111111111111111111111111111111

Thanks for noticing this!

Problem

While compiling a larger project with clang I stumbled upon an irritating bug. Consider the following small example: ``` unsigned long int * * fee(); void foo( unsigned long int q ) { unsigned long int i,j,k,e; unsigned long int pows[7]; unsigned long int * * table; e = 0; for (i = 1; i <= 256; i *= q) pows[e++] = i; pows[e--] = i; table = fee(); // need to set table to something unknown // here, otherwise the compiler optimises // parts of the loops below away // (and no bug occurs) for (i = 0; i < q; i++) for (j = 0; j < e; j++) ((unsigned char*)(*table) + 5 )[i*e + j] = 0; // bug here } ``` To the best of my knowledge this code does not violate the C standard in any way, although the last line seems awkward (in the actual project, code like this appears due to excessive use of preprocessor macros). Compiling this with clang (version 3.1 or higher) at optimisation level -O1 or higher results in code writing to the wrong position in memory. The crucial parts of the assembly file produced by clang/LLVM read as follows: (This is GAS syntax, so to those of you who are used to Intel: Beware!) ``` [...] callq _fee leaq 6(%rbx), %r8 ## at this point, %rbx == e-1 xorl %edx, %edx LBB0_4: [...] movq %r8, %rsi imulq %rdx, %rsi incq %rdx LBB0_6: movq (%rax), %rcx ## %rax == fee() movb $0, (%rcx,%rsi) incq %rsi [conditional jumps back to LBB0_6 resp. LBB0_4] [...] ``` In other words, the instructions do ``` (*table)[i*(e+5) + j] = 0; ``` instead of the last line written above. The choice of `+ 5` is arbitrary, adding (or subtracting) other integers results in the same behaviour. So - is this a bug in LLVM's optimisation or is there undefined behaviour going on here? Edit: Note also that the bug disappears if I leave out the cast `(unsigned char*)` in the last line. In general, the bug appears to be quite sensitive to any changes.

Original source