[SOLVED] Force clang not to unroll a loop

Force clang not to unroll a loop

I have a loop of the form

for (int i = 0; i < SOME_CONSTANT; ++i) {
    // do some stuff involving `i`.
}

which clang is partially unrolling to do two passes at a time. I need it to not be unrolled at all.

One option is to add an inline assembly barrier to prevent clang/LLVM from seeing how i is modified:

for (int i = 0; i < SOME_CONSTANT; ++i) {
    asm volatile("" : "=r"(i) : "0"(i));
    // do some stuff involving `i`.
}

but I'm wondering if there is a clearer way without resorting to such hacks.

Solution

clang is quite aggressive with vectorization, even at "-O2" - and you may need to minimize that as well, so I'll add them here. The following pragma only affects the next loop statement:

_Pragma("clang loop unroll(disable) vectorize(disable)")
for (int i = 0; i < n; ++i) {
    // do some stuff involving `i`.
}

I've needed this for my own purposes, and spent some time chasing up version numbers that support these options. This requires: __clang_major__.__clang_minor__ >= 3.7, which is ancient, and corresponds to Apple's hacked Xcode release version: __apple_build_version__ >= (7000072))

FYI, (__GNUC__ >= 8), provides the loop pragma: _Pragma("GCC unroll (0)"); and (__GNUC__ >= 14) finally provides: _Pragma("GCC novector") - earlier versions need to use optimization options, or various inline asm tricks.