Three nested loops result in erroneous control flow

[tomiaijo]

While testing simple OpenCL image kernels through the HSAIL LLVM I noticed that three nested loops cause the outer loop to branch incorrectly. More specifically, the loop variable is incremented every other iteration of the outer loop. Attached is a simple OpenCL kernel that reproduces the issue. The issue persists up to -O2, but is not present when using -O3.

For this kernel, the HSAIL LLVM create the following outer loop structure (complete HSAIL attached):

…
@BB0_2:
    mov_b32 $s5, 0;
    not_b1  $c1, $c1;
    cbr_b1  $c1, @BB0_3;
    br  @BB0_7;
…
@BB0_7:
    add_u32 $s4, $s4, 1;
    cmp_ne_b1_s32   $c2, $s4, $s0;
    cbr_b1  $c2, @BB0_2;

The boolean c1 is negated on each iteration, but only every other iteration increments the outer loop variable $s4. It seems that the not clause is related to a check that asserts that the outer loop is executed at all (in this case, that x_size > -1).

Pull request #29 fixes this simple case, but more thorough control flow analysis is required to detect all variable uses inside loops.

kernel.zip